drivers/md/raid5.c

   1 /*
   2  * raid5.c : Multiple Devices driver for Linux
   3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   4  *         Copyright (C) 1999, 2000 Ingo Molnar
   5  *         Copyright (C) 2002, 2003 H. Peter Anvin
   6  *
   7  * RAID-4/5/6 management functions.
   8  * Thanks to Penguin Computing for making the RAID-6 development possible
   9  * by donating a test server!
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 /*
  22  * BITMAP UNPLUGGING:
  23  *
  24  * The sequencing for updating the bitmap reliably is a little
  25  * subtle (and I got it wrong the first time) so it deserves some
  26  * explanation.
  27  *
  28  * We group bitmap updates into batches.  Each batch has a number.
  29  * We may write out several batches at once, but that isn't very important.
  30  * conf->seq_write is the number of the last batch successfully written.
  31  * conf->seq_flush is the number of the last batch that was closed to
  32  *    new additions.
  33  * When we discover that we will need to write to any block in a stripe
  34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  35  * the number of the batch it will be in. This is seq_flush+1.
  36  * When we are ready to do a write, if that batch hasn't been written yet,
  37  *   we plug the array and queue the stripe for later.
  38  * When an unplug happens, we increment bm_flush, thus closing the current
  39  *   batch.
  40  * When we notice that bm_flush > bm_write, we write out all pending updates
  41  * to the bitmap, and advance bm_write to where bm_flush was.
  42  * This may occasionally write a bit out twice, but is sure never to
  43  * miss any bits.
  44  */
  45
  46 #include <linux/blkdev.h>
  47 #include <linux/kthread.h>
  48 #include <linux/raid/pq.h>
  49 #include <linux/async_tx.h>
  50 #include <linux/module.h>
  51 #include <linux/async.h>
  52 #include <linux/seq_file.h>
  53 #include <linux/cpu.h>
  54 #include <linux/slab.h>
  55 #include <linux/ratelimit.h>
  56 #include <linux/nodemask.h>
  57 #include <linux/flex_array.h>
  58
  59 #include <trace/events/block.h>
  60 #include <linux/list_sort.h>
  61
  62 #include "md.h"
  63 #include "raid5.h"
  64 #include "raid0.h"
  65 #include "md-bitmap.h"
  66 #include "raid5-log.h"
  67
  68 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  69
  70 #define cpu_to_group(cpu) cpu_to_node(cpu)
  71 #define ANY_GROUP NUMA_NO_NODE
  72
  73 static bool devices_handle_discard_safely = false;
  74 module_param(devices_handle_discard_safely, bool, 0644);
  75 MODULE_PARM_DESC(devices_handle_discard_safely,
  76                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  77 static struct workqueue_struct *raid5_wq;
  78
  79 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  80 {
  81         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  82         return &conf->stripe_hashtbl[hash];
  83 }
  84
  85 static inline int stripe_hash_locks_hash(sector_t sect)
  86 {
  87         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  88 }
  89
  90 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  91 {
  92         spin_lock_irq(conf->hash_locks + hash);
  93         spin_lock(&conf->device_lock);
  94 }
  95
  96 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  97 {
  98         spin_unlock(&conf->device_lock);
  99         spin_unlock_irq(conf->hash_locks + hash);
 100 }
 101
 102 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 103 {
 104         int i;
 105         spin_lock_irq(conf->hash_locks);
 106         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 107                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 108         spin_lock(&conf->device_lock);
 109 }
 110
 111 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 112 {
 113         int i;
 114         spin_unlock(&conf->device_lock);
 115         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 116                 spin_unlock(conf->hash_locks + i);
 117         spin_unlock_irq(conf->hash_locks);
 118 }
 119
 120 /* Find first data disk in a raid6 stripe */
 121 static inline int raid6_d0(struct stripe_head *sh)
 122 {
 123         if (sh->ddf_layout)
 124                 /* ddf always start from first device */
 125                 return 0;
 126         /* md starts just after Q block */
 127         if (sh->qd_idx == sh->disks - 1)
 128                 return 0;
 129         else
 130                 return sh->qd_idx + 1;
 131 }
 132 static inline int raid6_next_disk(int disk, int raid_disks)
 133 {
 134         disk++;
 135         return (disk < raid_disks) ? disk : 0;
 136 }
 137
 138 /* When walking through the disks in a raid5, starting at raid6_d0,
 139  * We need to map each disk to a 'slot', where the data disks are slot
 140  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 141  * is raid_disks-1.  This help does that mapping.
 142  */
 143 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 144                              int *count, int syndrome_disks)
 145 {
 146         int slot = *count;
 147
 148         if (sh->ddf_layout)
 149                 (*count)++;
 150         if (idx == sh->pd_idx)
 151                 return syndrome_disks;
 152         if (idx == sh->qd_idx)
 153                 return syndrome_disks + 1;
 154         if (!sh->ddf_layout)
 155                 (*count)++;
 156         return slot;
 157 }
 158
 159 static void print_raid5_conf (struct r5conf *conf);
 160
 161 static int stripe_operations_active(struct stripe_head *sh)
 162 {
 163         return sh->check_state || sh->reconstruct_state ||
 164                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 165                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 166 }
 167
 168 static bool stripe_is_lowprio(struct stripe_head *sh)
 169 {
 170         return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 171                 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 172                !test_bit(STRIPE_R5C_CACHING, &sh->state);
 173 }
 174
 175 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 176 {
 177         struct r5conf *conf = sh->raid_conf;
 178         struct r5worker_group *group;
 179         int thread_cnt;
 180         int i, cpu = sh->cpu;
 181
 182         if (!cpu_online(cpu)) {
 183                 cpu = cpumask_any(cpu_online_mask);
 184                 sh->cpu = cpu;
 185         }
 186
 187         if (list_empty(&sh->lru)) {
 188                 struct r5worker_group *group;
 189                 group = conf->worker_groups + cpu_to_group(cpu);
 190                 if (stripe_is_lowprio(sh))
 191                         list_add_tail(&sh->lru, &group->loprio_list);
 192                 else
 193                         list_add_tail(&sh->lru, &group->handle_list);
 194                 group->stripes_cnt++;
 195                 sh->group = group;
 196         }
 197
 198         if (conf->worker_cnt_per_group == 0) {
 199                 md_wakeup_thread(conf->mddev->thread);
 200                 return;
 201         }
 202
 203         group = conf->worker_groups + cpu_to_group(sh->cpu);
 204
 205         group->workers[0].working = true;
 206         /* at least one worker should run to avoid race */
 207         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 208
 209         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 210         /* wakeup more workers */
 211         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 212                 if (group->workers[i].working == false) {
 213                         group->workers[i].working = true;
 214                         queue_work_on(sh->cpu, raid5_wq,
 215                                       &group->workers[i].work);
 216                         thread_cnt--;
 217                 }
 218         }
 219 }
 220
 221 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 222                               struct list_head *temp_inactive_list)
 223 {
 224         int i;
 225         int injournal = 0;      /* number of date pages with R5_InJournal */
 226
 227         BUG_ON(!list_empty(&sh->lru));
 228         BUG_ON(atomic_read(&conf->active_stripes)==0);
 229
 230         if (r5c_is_writeback(conf->log))
 231                 for (i = sh->disks; i--; )
 232                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 233                                 injournal++;
 234         /*
 235          * In the following cases, the stripe cannot be released to cached
 236          * lists. Therefore, we make the stripe write out and set
 237          * STRIPE_HANDLE:
 238          *   1. when quiesce in r5c write back;
 239          *   2. when resync is requested fot the stripe.
 240          */
 241         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 242             (conf->quiesce && r5c_is_writeback(conf->log) &&
 243              !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 244                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 245                         r5c_make_stripe_write_out(sh);
 246                 set_bit(STRIPE_HANDLE, &sh->state);
 247         }
 248
 249         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 250                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 251                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 252                         list_add_tail(&sh->lru, &conf->delayed_list);
 253                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 254                            sh->bm_seq - conf->seq_write > 0)
 255                         list_add_tail(&sh->lru, &conf->bitmap_list);
 256                 else {
 257                         clear_bit(STRIPE_DELAYED, &sh->state);
 258                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 259                         if (conf->worker_cnt_per_group == 0) {
 260                                 if (stripe_is_lowprio(sh))
 261                                         list_add_tail(&sh->lru,
 262                                                         &conf->loprio_list);
 263                                 else
 264                                         list_add_tail(&sh->lru,
 265                                                         &conf->handle_list);
 266                         } else {
 267                                 raid5_wakeup_stripe_thread(sh);
 268                                 return;
 269                         }
 270                 }
 271                 md_wakeup_thread(conf->mddev->thread);
 272         } else {
 273                 BUG_ON(stripe_operations_active(sh));
 274                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 275                         if (atomic_dec_return(&conf->preread_active_stripes)
 276                             < IO_THRESHOLD)
 277                                 md_wakeup_thread(conf->mddev->thread);
 278                 atomic_dec(&conf->active_stripes);
 279                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 280                         if (!r5c_is_writeback(conf->log))
 281                                 list_add_tail(&sh->lru, temp_inactive_list);
 282                         else {
 283                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 284                                 if (injournal == 0)
 285                                         list_add_tail(&sh->lru, temp_inactive_list);
 286                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 287                                         /* full stripe */
 288                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 289                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 290                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 291                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 292                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 293                                         r5c_check_cached_full_stripe(conf);
 294                                 } else
 295                                         /*
 296                                          * STRIPE_R5C_PARTIAL_STRIPE is set in
 297                                          * r5c_try_caching_write(). No need to
 298                                          * set it again.
 299                                          */
 300                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 301                         }
 302                 }
 303         }
 304 }
 305
 306 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 307                              struct list_head *temp_inactive_list)
 308 {
 309         if (atomic_dec_and_test(&sh->count))
 310                 do_release_stripe(conf, sh, temp_inactive_list);
 311 }
 312
 313 /*
 314  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 315  *
 316  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 317  * given time. Adding stripes only takes device lock, while deleting stripes
 318  * only takes hash lock.
 319  */
 320 static void release_inactive_stripe_list(struct r5conf *conf,
 321                                          struct list_head *temp_inactive_list,
 322                                          int hash)
 323 {
 324         int size;
 325         bool do_wakeup = false;
 326         unsigned long flags;
 327
 328         if (hash == NR_STRIPE_HASH_LOCKS) {
 329                 size = NR_STRIPE_HASH_LOCKS;
 330                 hash = NR_STRIPE_HASH_LOCKS - 1;
 331         } else
 332                 size = 1;
 333         while (size) {
 334                 struct list_head *list = &temp_inactive_list[size - 1];
 335
 336                 /*
 337                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 338                  * remove stripes from the list
 339                  */
 340                 if (!list_empty_careful(list)) {
 341                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 342                         if (list_empty(conf->inactive_list + hash) &&
 343                             !list_empty(list))
 344                                 atomic_dec(&conf->empty_inactive_list_nr);
 345                         list_splice_tail_init(list, conf->inactive_list + hash);
 346                         do_wakeup = true;
 347                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 348                 }
 349                 size--;
 350                 hash--;
 351         }
 352
 353         if (do_wakeup) {
 354                 wake_up(&conf->wait_for_stripe);
 355                 if (atomic_read(&conf->active_stripes) == 0)
 356                         wake_up(&conf->wait_for_quiescent);
 357                 if (conf->retry_read_aligned)
 358                         md_wakeup_thread(conf->mddev->thread);
 359         }
 360 }
 361
 362 /* should hold conf->device_lock already */
 363 static int release_stripe_list(struct r5conf *conf,
 364                                struct list_head *temp_inactive_list)
 365 {
 366         struct stripe_head *sh, *t;
 367         int count = 0;
 368         struct llist_node *head;
 369
 370         head = llist_del_all(&conf->released_stripes);
 371         head = llist_reverse_order(head);
 372         llist_for_each_entry_safe(sh, t, head, release_list) {
 373                 int hash;
 374
 375                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 376                 smp_mb();
 377                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 378                 /*
 379                  * Don't worry the bit is set here, because if the bit is set
 380                  * again, the count is always > 1. This is true for
 381                  * STRIPE_ON_UNPLUG_LIST bit too.
 382                  */
 383                 hash = sh->hash_lock_index;
 384                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 385                 count++;
 386         }
 387
 388         return count;
 389 }
 390
 391 void raid5_release_stripe(struct stripe_head *sh)
 392 {
 393         struct r5conf *conf = sh->raid_conf;
 394         unsigned long flags;
 395         struct list_head list;
 396         int hash;
 397         bool wakeup;
 398
 399         /* Avoid release_list until the last reference.
 400          */
 401         if (atomic_add_unless(&sh->count, -1, 1))
 402                 return;
 403
 404         if (unlikely(!conf->mddev->thread) ||
 405                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 406                 goto slow_path;
 407         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 408         if (wakeup)
 409                 md_wakeup_thread(conf->mddev->thread);
 410         return;
 411 slow_path:
 412         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 413         if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
 414                 INIT_LIST_HEAD(&list);
 415                 hash = sh->hash_lock_index;
 416                 do_release_stripe(conf, sh, &list);
 417                 spin_unlock_irqrestore(&conf->device_lock, flags);
 418                 release_inactive_stripe_list(conf, &list, hash);
 419         }
 420 }
 421
 422 static inline void remove_hash(struct stripe_head *sh)
 423 {
 424         pr_debug("remove_hash(), stripe %llu\n",
 425                 (unsigned long long)sh->sector);
 426
 427         hlist_del_init(&sh->hash);
 428 }
 429
 430 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 431 {
 432         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 433
 434         pr_debug("insert_hash(), stripe %llu\n",
 435                 (unsigned long long)sh->sector);
 436
 437         hlist_add_head(&sh->hash, hp);
 438 }
 439
 440 /* find an idle stripe, make sure it is unhashed, and return it. */
 441 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 442 {
 443         struct stripe_head *sh = NULL;
 444         struct list_head *first;
 445
 446         if (list_empty(conf->inactive_list + hash))
 447                 goto out;
 448         first = (conf->inactive_list + hash)->next;
 449         sh = list_entry(first, struct stripe_head, lru);
 450         list_del_init(first);
 451         remove_hash(sh);
 452         atomic_inc(&conf->active_stripes);
 453         BUG_ON(hash != sh->hash_lock_index);
 454         if (list_empty(conf->inactive_list + hash))
 455                 atomic_inc(&conf->empty_inactive_list_nr);
 456 out:
 457         return sh;
 458 }
 459
 460 static void shrink_buffers(struct stripe_head *sh)
 461 {
 462         struct page *p;
 463         int i;
 464         int num = sh->raid_conf->pool_size;
 465
 466         for (i = 0; i < num ; i++) {
 467                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 468                 p = sh->dev[i].page;
 469                 if (!p)
 470                         continue;
 471                 sh->dev[i].page = NULL;
 472                 put_page(p);
 473         }
 474 }
 475
 476 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 477 {
 478         int i;
 479         int num = sh->raid_conf->pool_size;
 480
 481         for (i = 0; i < num; i++) {
 482                 struct page *page;
 483
 484                 if (!(page = alloc_page(gfp))) {
 485                         return 1;
 486                 }
 487                 sh->dev[i].page = page;
 488                 sh->dev[i].orig_page = page;
 489         }
 490
 491         return 0;
 492 }
 493
 494 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 495                             struct stripe_head *sh);
 496
 497 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 498 {
 499         struct r5conf *conf = sh->raid_conf;
 500         int i, seq;
 501
 502         BUG_ON(atomic_read(&sh->count) != 0);
 503         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 504         BUG_ON(stripe_operations_active(sh));
 505         BUG_ON(sh->batch_head);
 506
 507         pr_debug("init_stripe called, stripe %llu\n",
 508                 (unsigned long long)sector);
 509 retry:
 510         seq = read_seqcount_begin(&conf->gen_lock);
 511         sh->generation = conf->generation - previous;
 512         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 513         sh->sector = sector;
 514         stripe_set_idx(sector, conf, previous, sh);
 515         sh->state = 0;
 516
 517         for (i = sh->disks; i--; ) {
 518                 struct r5dev *dev = &sh->dev[i];
 519
 520                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 521                     test_bit(R5_LOCKED, &dev->flags)) {
 522                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 523                                (unsigned long long)sh->sector, i, dev->toread,
 524                                dev->read, dev->towrite, dev->written,
 525                                test_bit(R5_LOCKED, &dev->flags));
 526                         WARN_ON(1);
 527                 }
 528                 dev->flags = 0;
 529                 dev->sector = raid5_compute_blocknr(sh, i, previous);
 530         }
 531         if (read_seqcount_retry(&conf->gen_lock, seq))
 532                 goto retry;
 533         sh->overwrite_disks = 0;
 534         insert_hash(conf, sh);
 535         sh->cpu = smp_processor_id();
 536         set_bit(STRIPE_BATCH_READY, &sh->state);
 537 }
 538
 539 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 540                                          short generation)
 541 {
 542         struct stripe_head *sh;
 543
 544         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 545         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 546                 if (sh->sector == sector && sh->generation == generation)
 547                         return sh;
 548         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 549         return NULL;
 550 }
 551
 552 /*
 553  * Need to check if array has failed when deciding whether to:
 554  *  - start an array
 555  *  - remove non-faulty devices
 556  *  - add a spare
 557  *  - allow a reshape
 558  * This determination is simple when no reshape is happening.
 559  * However if there is a reshape, we need to carefully check
 560  * both the before and after sections.
 561  * This is because some failed devices may only affect one
 562  * of the two sections, and some non-in_sync devices may
 563  * be insync in the section most affected by failed devices.
 564  */
 565 int raid5_calc_degraded(struct r5conf *conf)
 566 {
 567         int degraded, degraded2;
 568         int i;
 569
 570         rcu_read_lock();
 571         degraded = 0;
 572         for (i = 0; i < conf->previous_raid_disks; i++) {
 573                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 574                 if (rdev && test_bit(Faulty, &rdev->flags))
 575                         rdev = rcu_dereference(conf->disks[i].replacement);
 576                 if (!rdev || test_bit(Faulty, &rdev->flags))
 577                         degraded++;
 578                 else if (test_bit(In_sync, &rdev->flags))
 579                         ;
 580                 else
 581                         /* not in-sync or faulty.
 582                          * If the reshape increases the number of devices,
 583                          * this is being recovered by the reshape, so
 584                          * this 'previous' section is not in_sync.
 585                          * If the number of devices is being reduced however,
 586                          * the device can only be part of the array if
 587                          * we are reverting a reshape, so this section will
 588                          * be in-sync.
 589                          */
 590                         if (conf->raid_disks >= conf->previous_raid_disks)
 591                                 degraded++;
 592         }
 593         rcu_read_unlock();
 594         if (conf->raid_disks == conf->previous_raid_disks)
 595                 return degraded;
 596         rcu_read_lock();
 597         degraded2 = 0;
 598         for (i = 0; i < conf->raid_disks; i++) {
 599                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 600                 if (rdev && test_bit(Faulty, &rdev->flags))
 601                         rdev = rcu_dereference(conf->disks[i].replacement);
 602                 if (!rdev || test_bit(Faulty, &rdev->flags))
 603                         degraded2++;
 604                 else if (test_bit(In_sync, &rdev->flags))
 605                         ;
 606                 else
 607                         /* not in-sync or faulty.
 608                          * If reshape increases the number of devices, this
 609                          * section has already been recovered, else it
 610                          * almost certainly hasn't.
 611                          */
 612                         if (conf->raid_disks <= conf->previous_raid_disks)
 613                                 degraded2++;
 614         }
 615         rcu_read_unlock();
 616         if (degraded2 > degraded)
 617                 return degraded2;
 618         return degraded;
 619 }
 620
 621 static int has_failed(struct r5conf *conf)
 622 {
 623         int degraded;
 624
 625         if (conf->mddev->reshape_position == MaxSector)
 626                 return conf->mddev->degraded > conf->max_degraded;
 627
 628         degraded = raid5_calc_degraded(conf);
 629         if (degraded > conf->max_degraded)
 630                 return 1;
 631         return 0;
 632 }
 633
 634 struct stripe_head *
 635 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 636                         int previous, int noblock, int noquiesce)
 637 {
 638         struct stripe_head *sh;
 639         int hash = stripe_hash_locks_hash(sector);
 640         int inc_empty_inactive_list_flag;
 641
 642         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 643
 644         spin_lock_irq(conf->hash_locks + hash);
 645
 646         do {
 647                 wait_event_lock_irq(conf->wait_for_quiescent,
 648                                     conf->quiesce == 0 || noquiesce,
 649                                     *(conf->hash_locks + hash));
 650                 sh = __find_stripe(conf, sector, conf->generation - previous);
 651                 if (!sh) {
 652                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 653                                 sh = get_free_stripe(conf, hash);
 654                                 if (!sh && !test_bit(R5_DID_ALLOC,
 655                                                      &conf->cache_state))
 656                                         set_bit(R5_ALLOC_MORE,
 657                                                 &conf->cache_state);
 658                         }
 659                         if (noblock && sh == NULL)
 660                                 break;
 661
 662                         r5c_check_stripe_cache_usage(conf);
 663                         if (!sh) {
 664                                 set_bit(R5_INACTIVE_BLOCKED,
 665                                         &conf->cache_state);
 666                                 r5l_wake_reclaim(conf->log, 0);
 667                                 wait_event_lock_irq(
 668                                         conf->wait_for_stripe,
 669                                         !list_empty(conf->inactive_list + hash) &&
 670                                         (atomic_read(&conf->active_stripes)
 671                                          < (conf->max_nr_stripes * 3 / 4)
 672                                          || !test_bit(R5_INACTIVE_BLOCKED,
 673                                                       &conf->cache_state)),
 674                                         *(conf->hash_locks + hash));
 675                                 clear_bit(R5_INACTIVE_BLOCKED,
 676                                           &conf->cache_state);
 677                         } else {
 678                                 init_stripe(sh, sector, previous);
 679                                 atomic_inc(&sh->count);
 680                         }
 681                 } else if (!atomic_inc_not_zero(&sh->count)) {
 682                         spin_lock(&conf->device_lock);
 683                         if (!atomic_read(&sh->count)) {
 684                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 685                                         atomic_inc(&conf->active_stripes);
 686                                 BUG_ON(list_empty(&sh->lru) &&
 687                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 688                                 inc_empty_inactive_list_flag = 0;
 689                                 if (!list_empty(conf->inactive_list + hash))
 690                                         inc_empty_inactive_list_flag = 1;
 691                                 list_del_init(&sh->lru);
 692                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 693                                         atomic_inc(&conf->empty_inactive_list_nr);
 694                                 if (sh->group) {
 695                                         sh->group->stripes_cnt--;
 696                                         sh->group = NULL;
 697                                 }
 698                         }
 699                         atomic_inc(&sh->count);
 700                         spin_unlock(&conf->device_lock);
 701                 }
 702         } while (sh == NULL);
 703
 704         spin_unlock_irq(conf->hash_locks + hash);
 705         return sh;
 706 }
 707
 708 static bool is_full_stripe_write(struct stripe_head *sh)
 709 {
 710         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 711         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 712 }
 713
 714 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 715 {
 716         if (sh1 > sh2) {
 717                 spin_lock_irq(&sh2->stripe_lock);
 718                 spin_lock_nested(&sh1->stripe_lock, 1);
 719         } else {
 720                 spin_lock_irq(&sh1->stripe_lock);
 721                 spin_lock_nested(&sh2->stripe_lock, 1);
 722         }
 723 }
 724
 725 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 726 {
 727         spin_unlock(&sh1->stripe_lock);
 728         spin_unlock_irq(&sh2->stripe_lock);
 729 }
 730
 731 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 732 static bool stripe_can_batch(struct stripe_head *sh)
 733 {
 734         struct r5conf *conf = sh->raid_conf;
 735
 736         if (raid5_has_log(conf) || raid5_has_ppl(conf))
 737                 return false;
 738         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 739                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 740                 is_full_stripe_write(sh);
 741 }
 742
 743 /* we only do back search */
 744 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 745 {
 746         struct stripe_head *head;
 747         sector_t head_sector, tmp_sec;
 748         int hash;
 749         int dd_idx;
 750         int inc_empty_inactive_list_flag;
 751
 752         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 753         tmp_sec = sh->sector;
 754         if (!sector_div(tmp_sec, conf->chunk_sectors))
 755                 return;
 756         head_sector = sh->sector - STRIPE_SECTORS;
 757
 758         hash = stripe_hash_locks_hash(head_sector);
 759         spin_lock_irq(conf->hash_locks + hash);
 760         head = __find_stripe(conf, head_sector, conf->generation);
 761         if (head && !atomic_inc_not_zero(&head->count)) {
 762                 spin_lock(&conf->device_lock);
 763                 if (!atomic_read(&head->count)) {
 764                         if (!test_bit(STRIPE_HANDLE, &head->state))
 765                                 atomic_inc(&conf->active_stripes);
 766                         BUG_ON(list_empty(&head->lru) &&
 767                                !test_bit(STRIPE_EXPANDING, &head->state));
 768                         inc_empty_inactive_list_flag = 0;
 769                         if (!list_empty(conf->inactive_list + hash))
 770                                 inc_empty_inactive_list_flag = 1;
 771                         list_del_init(&head->lru);
 772                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 773                                 atomic_inc(&conf->empty_inactive_list_nr);
 774                         if (head->group) {
 775                                 head->group->stripes_cnt--;
 776                                 head->group = NULL;
 777                         }
 778                 }
 779                 atomic_inc(&head->count);
 780                 spin_unlock(&conf->device_lock);
 781         }
 782         spin_unlock_irq(conf->hash_locks + hash);
 783
 784         if (!head)
 785                 return;
 786         if (!stripe_can_batch(head))
 787                 goto out;
 788
 789         lock_two_stripes(head, sh);
 790         /* clear_batch_ready clear the flag */
 791         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 792                 goto unlock_out;
 793
 794         if (sh->batch_head)
 795                 goto unlock_out;
 796
 797         dd_idx = 0;
 798         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 799                 dd_idx++;
 800         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 801             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 802                 goto unlock_out;
 803
 804         if (head->batch_head) {
 805                 spin_lock(&head->batch_head->batch_lock);
 806                 /* This batch list is already running */
 807                 if (!stripe_can_batch(head)) {
 808                         spin_unlock(&head->batch_head->batch_lock);
 809                         goto unlock_out;
 810                 }
 811                 /*
 812                  * We must assign batch_head of this stripe within the
 813                  * batch_lock, otherwise clear_batch_ready of batch head
 814                  * stripe could clear BATCH_READY bit of this stripe and
 815                  * this stripe->batch_head doesn't get assigned, which
 816                  * could confuse clear_batch_ready for this stripe
 817                  */
 818                 sh->batch_head = head->batch_head;
 819
 820                 /*
 821                  * at this point, head's BATCH_READY could be cleared, but we
 822                  * can still add the stripe to batch list
 823                  */
 824                 list_add(&sh->batch_list, &head->batch_list);
 825                 spin_unlock(&head->batch_head->batch_lock);
 826         } else {
 827                 head->batch_head = head;
 828                 sh->batch_head = head->batch_head;
 829                 spin_lock(&head->batch_lock);
 830                 list_add_tail(&sh->batch_list, &head->batch_list);
 831                 spin_unlock(&head->batch_lock);
 832         }
 833
 834         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 835                 if (atomic_dec_return(&conf->preread_active_stripes)
 836                     < IO_THRESHOLD)
 837                         md_wakeup_thread(conf->mddev->thread);
 838
 839         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 840                 int seq = sh->bm_seq;
 841                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 842                     sh->batch_head->bm_seq > seq)
 843                         seq = sh->batch_head->bm_seq;
 844                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 845                 sh->batch_head->bm_seq = seq;
 846         }
 847
 848         atomic_inc(&sh->count);
 849 unlock_out:
 850         unlock_two_stripes(head, sh);
 851 out:
 852         raid5_release_stripe(head);
 853 }
 854
 855 /* Determine if 'data_offset' or 'new_data_offset' should be used
 856  * in this stripe_head.
 857  */
 858 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 859 {
 860         sector_t progress = conf->reshape_progress;
 861         /* Need a memory barrier to make sure we see the value
 862          * of conf->generation, or ->data_offset that was set before
 863          * reshape_progress was updated.
 864          */
 865         smp_rmb();
 866         if (progress == MaxSector)
 867                 return 0;
 868         if (sh->generation == conf->generation - 1)
 869                 return 0;
 870         /* We are in a reshape, and this is a new-generation stripe,
 871          * so use new_data_offset.
 872          */
 873         return 1;
 874 }
 875
 876 static void dispatch_bio_list(struct bio_list *tmp)
 877 {
 878         struct bio *bio;
 879
 880         while ((bio = bio_list_pop(tmp)))
 881                 generic_make_request(bio);
 882 }
 883
 884 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
 885 {
 886         const struct r5pending_data *da = list_entry(a,
 887                                 struct r5pending_data, sibling);
 888         const struct r5pending_data *db = list_entry(b,
 889                                 struct r5pending_data, sibling);
 890         if (da->sector > db->sector)
 891                 return 1;
 892         if (da->sector < db->sector)
 893                 return -1;
 894         return 0;
 895 }
 896
 897 static void dispatch_defer_bios(struct r5conf *conf, int target,
 898                                 struct bio_list *list)
 899 {
 900         struct r5pending_data *data;
 901         struct list_head *first, *next = NULL;
 902         int cnt = 0;
 903
 904         if (conf->pending_data_cnt == 0)
 905                 return;
 906
 907         list_sort(NULL, &conf->pending_list, cmp_stripe);
 908
 909         first = conf->pending_list.next;
 910
 911         /* temporarily move the head */
 912         if (conf->next_pending_data)
 913                 list_move_tail(&conf->pending_list,
 914                                 &conf->next_pending_data->sibling);
 915
 916         while (!list_empty(&conf->pending_list)) {
 917                 data = list_first_entry(&conf->pending_list,
 918                         struct r5pending_data, sibling);
 919                 if (&data->sibling == first)
 920                         first = data->sibling.next;
 921                 next = data->sibling.next;
 922
 923                 bio_list_merge(list, &data->bios);
 924                 list_move(&data->sibling, &conf->free_list);
 925                 cnt++;
 926                 if (cnt >= target)
 927                         break;
 928         }
 929         conf->pending_data_cnt -= cnt;
 930         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
 931
 932         if (next != &conf->pending_list)
 933                 conf->next_pending_data = list_entry(next,
 934                                 struct r5pending_data, sibling);
 935         else
 936                 conf->next_pending_data = NULL;
 937         /* list isn't empty */
 938         if (first != &conf->pending_list)
 939                 list_move_tail(&conf->pending_list, first);
 940 }
 941
 942 static void flush_deferred_bios(struct r5conf *conf)
 943 {
 944         struct bio_list tmp = BIO_EMPTY_LIST;
 945
 946         if (conf->pending_data_cnt == 0)
 947                 return;
 948
 949         spin_lock(&conf->pending_bios_lock);
 950         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
 951         BUG_ON(conf->pending_data_cnt != 0);
 952         spin_unlock(&conf->pending_bios_lock);
 953
 954         dispatch_bio_list(&tmp);
 955 }
 956
 957 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
 958                                 struct bio_list *bios)
 959 {
 960         struct bio_list tmp = BIO_EMPTY_LIST;
 961         struct r5pending_data *ent;
 962
 963         spin_lock(&conf->pending_bios_lock);
 964         ent = list_first_entry(&conf->free_list, struct r5pending_data,
 965                                                         sibling);
 966         list_move_tail(&ent->sibling, &conf->pending_list);
 967         ent->sector = sector;
 968         bio_list_init(&ent->bios);
 969         bio_list_merge(&ent->bios, bios);
 970         conf->pending_data_cnt++;
 971         if (conf->pending_data_cnt >= PENDING_IO_MAX)
 972                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
 973
 974         spin_unlock(&conf->pending_bios_lock);
 975
 976         dispatch_bio_list(&tmp);
 977 }
 978
 979 static void
 980 raid5_end_read_request(struct bio *bi);
 981 static void
 982 raid5_end_write_request(struct bio *bi);
 983
 984 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 985 {
 986         struct r5conf *conf = sh->raid_conf;
 987         int i, disks = sh->disks;
 988         struct stripe_head *head_sh = sh;
 989         struct bio_list pending_bios = BIO_EMPTY_LIST;
 990         bool should_defer;
 991
 992         might_sleep();
 993
 994         if (log_stripe(sh, s) == 0)
 995                 return;
 996
 997         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
 998
 999         for (i = disks; i--; ) {
1000                 int op, op_flags = 0;
1001                 int replace_only = 0;
1002                 struct bio *bi, *rbi;
1003                 struct md_rdev *rdev, *rrdev = NULL;
1004
1005                 sh = head_sh;
1006                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1007                         op = REQ_OP_WRITE;
1008                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1009                                 op_flags = REQ_FUA;
1010                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1011                                 op = REQ_OP_DISCARD;
1012                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1013                         op = REQ_OP_READ;
1014                 else if (test_and_clear_bit(R5_WantReplace,
1015                                             &sh->dev[i].flags)) {
1016                         op = REQ_OP_WRITE;
1017                         replace_only = 1;
1018                 } else
1019                         continue;
1020                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1021                         op_flags |= REQ_SYNC;
1022
1023 again:
1024                 bi = &sh->dev[i].req;
1025                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1026
1027                 rcu_read_lock();
1028                 rrdev = rcu_dereference(conf->disks[i].replacement);
1029                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1030                 rdev = rcu_dereference(conf->disks[i].rdev);
1031                 if (!rdev) {
1032                         rdev = rrdev;
1033                         rrdev = NULL;
1034                 }
1035                 if (op_is_write(op)) {
1036                         if (replace_only)
1037                                 rdev = NULL;
1038                         if (rdev == rrdev)
1039                                 /* We raced and saw duplicates */
1040                                 rrdev = NULL;
1041                 } else {
1042                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1043                                 rdev = rrdev;
1044                         rrdev = NULL;
1045                 }
1046
1047                 if (rdev && test_bit(Faulty, &rdev->flags))
1048                         rdev = NULL;
1049                 if (rdev)
1050                         atomic_inc(&rdev->nr_pending);
1051                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1052                         rrdev = NULL;
1053                 if (rrdev)
1054                         atomic_inc(&rrdev->nr_pending);
1055                 rcu_read_unlock();
1056
1057                 /* We have already checked bad blocks for reads.  Now
1058                  * need to check for writes.  We never accept write errors
1059                  * on the replacement, so we don't to check rrdev.
1060                  */
1061                 while (op_is_write(op) && rdev &&
1062                        test_bit(WriteErrorSeen, &rdev->flags)) {
1063                         sector_t first_bad;
1064                         int bad_sectors;
1065                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1066                                               &first_bad, &bad_sectors);
1067                         if (!bad)
1068                                 break;
1069
1070                         if (bad < 0) {
1071                                 set_bit(BlockedBadBlocks, &rdev->flags);
1072                                 if (!conf->mddev->external &&
1073                                     conf->mddev->sb_flags) {
1074                                         /* It is very unlikely, but we might
1075                                          * still need to write out the
1076                                          * bad block log - better give it
1077                                          * a chance*/
1078                                         md_check_recovery(conf->mddev);
1079                                 }
1080                                 /*
1081                                  * Because md_wait_for_blocked_rdev
1082                                  * will dec nr_pending, we must
1083                                  * increment it first.
1084                                  */
1085                                 atomic_inc(&rdev->nr_pending);
1086                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1087                         } else {
1088                                 /* Acknowledged bad block - skip the write */
1089                                 rdev_dec_pending(rdev, conf->mddev);
1090                                 rdev = NULL;
1091                         }
1092                 }
1093
1094                 if (rdev) {
1095                         if (s->syncing || s->expanding || s->expanded
1096                             || s->replacing)
1097                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1098
1099                         set_bit(STRIPE_IO_STARTED, &sh->state);
1100
1101                         bio_set_dev(bi, rdev->bdev);
1102                         bio_set_op_attrs(bi, op, op_flags);
1103                         bi->bi_end_io = op_is_write(op)
1104                                 ? raid5_end_write_request
1105                                 : raid5_end_read_request;
1106                         bi->bi_private = sh;
1107
1108                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1109                                 __func__, (unsigned long long)sh->sector,
1110                                 bi->bi_opf, i);
1111                         atomic_inc(&sh->count);
1112                         if (sh != head_sh)
1113                                 atomic_inc(&head_sh->count);
1114                         if (use_new_offset(conf, sh))
1115                                 bi->bi_iter.bi_sector = (sh->sector
1116                                                  + rdev->new_data_offset);
1117                         else
1118                                 bi->bi_iter.bi_sector = (sh->sector
1119                                                  + rdev->data_offset);
1120                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1121                                 bi->bi_opf |= REQ_NOMERGE;
1122
1123                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1124                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1125
1126                         if (!op_is_write(op) &&
1127                             test_bit(R5_InJournal, &sh->dev[i].flags))
1128                                 /*
1129                                  * issuing read for a page in journal, this
1130                                  * must be preparing for prexor in rmw; read
1131                                  * the data into orig_page
1132                                  */
1133                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1134                         else
1135                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1136                         bi->bi_vcnt = 1;
1137                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1138                         bi->bi_io_vec[0].bv_offset = 0;
1139                         bi->bi_iter.bi_size = STRIPE_SIZE;
1140                         bi->bi_write_hint = sh->dev[i].write_hint;
1141                         if (!rrdev)
1142                                 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1143                         /*
1144                          * If this is discard request, set bi_vcnt 0. We don't
1145                          * want to confuse SCSI because SCSI will replace payload
1146                          */
1147                         if (op == REQ_OP_DISCARD)
1148                                 bi->bi_vcnt = 0;
1149                         if (rrdev)
1150                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1151
1152                         if (conf->mddev->gendisk)
1153                                 trace_block_bio_remap(bi->bi_disk->queue,
1154                                                       bi, disk_devt(conf->mddev->gendisk),
1155                                                       sh->dev[i].sector);
1156                         if (should_defer && op_is_write(op))
1157                                 bio_list_add(&pending_bios, bi);
1158                         else
1159                                 generic_make_request(bi);
1160                 }
1161                 if (rrdev) {
1162                         if (s->syncing || s->expanding || s->expanded
1163                             || s->replacing)
1164                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1165
1166                         set_bit(STRIPE_IO_STARTED, &sh->state);
1167
1168                         bio_set_dev(rbi, rrdev->bdev);
1169                         bio_set_op_attrs(rbi, op, op_flags);
1170                         BUG_ON(!op_is_write(op));
1171                         rbi->bi_end_io = raid5_end_write_request;
1172                         rbi->bi_private = sh;
1173
1174                         pr_debug("%s: for %llu schedule op %d on "
1175                                  "replacement disc %d\n",
1176                                 __func__, (unsigned long long)sh->sector,
1177                                 rbi->bi_opf, i);
1178                         atomic_inc(&sh->count);
1179                         if (sh != head_sh)
1180                                 atomic_inc(&head_sh->count);
1181                         if (use_new_offset(conf, sh))
1182                                 rbi->bi_iter.bi_sector = (sh->sector
1183                                                   + rrdev->new_data_offset);
1184                         else
1185                                 rbi->bi_iter.bi_sector = (sh->sector
1186                                                   + rrdev->data_offset);
1187                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1188                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1189                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1190                         rbi->bi_vcnt = 1;
1191                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1192                         rbi->bi_io_vec[0].bv_offset = 0;
1193                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1194                         rbi->bi_write_hint = sh->dev[i].write_hint;
1195                         sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1196                         /*
1197                          * If this is discard request, set bi_vcnt 0. We don't
1198                          * want to confuse SCSI because SCSI will replace payload
1199                          */
1200                         if (op == REQ_OP_DISCARD)
1201                                 rbi->bi_vcnt = 0;
1202                         if (conf->mddev->gendisk)
1203                                 trace_block_bio_remap(rbi->bi_disk->queue,
1204                                                       rbi, disk_devt(conf->mddev->gendisk),
1205                                                       sh->dev[i].sector);
1206                         if (should_defer && op_is_write(op))
1207                                 bio_list_add(&pending_bios, rbi);
1208                         else
1209                                 generic_make_request(rbi);
1210                 }
1211                 if (!rdev && !rrdev) {
1212                         if (op_is_write(op))
1213                                 set_bit(STRIPE_DEGRADED, &sh->state);
1214                         pr_debug("skip op %d on disc %d for sector %llu\n",
1215                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1216                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1217                         set_bit(STRIPE_HANDLE, &sh->state);
1218                 }
1219
1220                 if (!head_sh->batch_head)
1221                         continue;
1222                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1223                                       batch_list);
1224                 if (sh != head_sh)
1225                         goto again;
1226         }
1227
1228         if (should_defer && !bio_list_empty(&pending_bios))
1229                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1230 }
1231
1232 static struct dma_async_tx_descriptor *
1233 async_copy_data(int frombio, struct bio *bio, struct page **page,
1234         sector_t sector, struct dma_async_tx_descriptor *tx,
1235         struct stripe_head *sh, int no_skipcopy)
1236 {
1237         struct bio_vec bvl;
1238         struct bvec_iter iter;
1239         struct page *bio_page;
1240         int page_offset;
1241         struct async_submit_ctl submit;
1242         enum async_tx_flags flags = 0;
1243
1244         if (bio->bi_iter.bi_sector >= sector)
1245                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1246         else
1247                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1248
1249         if (frombio)
1250                 flags |= ASYNC_TX_FENCE;
1251         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1252
1253         bio_for_each_segment(bvl, bio, iter) {
1254                 int len = bvl.bv_len;
1255                 int clen;
1256                 int b_offset = 0;
1257
1258                 if (page_offset < 0) {
1259                         b_offset = -page_offset;
1260                         page_offset += b_offset;
1261                         len -= b_offset;
1262                 }
1263
1264                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1265                         clen = STRIPE_SIZE - page_offset;
1266                 else
1267                         clen = len;
1268
1269                 if (clen > 0) {
1270                         b_offset += bvl.bv_offset;
1271                         bio_page = bvl.bv_page;
1272                         if (frombio) {
1273                                 if (sh->raid_conf->skip_copy &&
1274                                     b_offset == 0 && page_offset == 0 &&
1275                                     clen == STRIPE_SIZE &&
1276                                     !no_skipcopy)
1277                                         *page = bio_page;
1278                                 else
1279                                         tx = async_memcpy(*page, bio_page, page_offset,
1280                                                   b_offset, clen, &submit);
1281                         } else
1282                                 tx = async_memcpy(bio_page, *page, b_offset,
1283                                                   page_offset, clen, &submit);
1284                 }
1285                 /* chain the operations */
1286                 submit.depend_tx = tx;
1287
1288                 if (clen < len) /* hit end of page */
1289                         break;
1290                 page_offset +=  len;
1291         }
1292
1293         return tx;
1294 }
1295
1296 static void ops_complete_biofill(void *stripe_head_ref)
1297 {
1298         struct stripe_head *sh = stripe_head_ref;
1299         int i;
1300
1301         pr_debug("%s: stripe %llu\n", __func__,
1302                 (unsigned long long)sh->sector);
1303
1304         /* clear completed biofills */
1305         for (i = sh->disks; i--; ) {
1306                 struct r5dev *dev = &sh->dev[i];
1307
1308                 /* acknowledge completion of a biofill operation */
1309                 /* and check if we need to reply to a read request,
1310                  * new R5_Wantfill requests are held off until
1311                  * !STRIPE_BIOFILL_RUN
1312                  */
1313                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1314                         struct bio *rbi, *rbi2;
1315
1316                         BUG_ON(!dev->read);
1317                         rbi = dev->read;
1318                         dev->read = NULL;
1319                         while (rbi && rbi->bi_iter.bi_sector <
1320                                 dev->sector + STRIPE_SECTORS) {
1321                                 rbi2 = r5_next_bio(rbi, dev->sector);
1322                                 bio_endio(rbi);
1323                                 rbi = rbi2;
1324                         }
1325                 }
1326         }
1327         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1328
1329         set_bit(STRIPE_HANDLE, &sh->state);
1330         raid5_release_stripe(sh);
1331 }
1332
1333 static void ops_run_biofill(struct stripe_head *sh)
1334 {
1335         struct dma_async_tx_descriptor *tx = NULL;
1336         struct async_submit_ctl submit;
1337         int i;
1338
1339         BUG_ON(sh->batch_head);
1340         pr_debug("%s: stripe %llu\n", __func__,
1341                 (unsigned long long)sh->sector);
1342
1343         for (i = sh->disks; i--; ) {
1344                 struct r5dev *dev = &sh->dev[i];
1345                 if (test_bit(R5_Wantfill, &dev->flags)) {
1346                         struct bio *rbi;
1347                         spin_lock_irq(&sh->stripe_lock);
1348                         dev->read = rbi = dev->toread;
1349                         dev->toread = NULL;
1350                         spin_unlock_irq(&sh->stripe_lock);
1351                         while (rbi && rbi->bi_iter.bi_sector <
1352                                 dev->sector + STRIPE_SECTORS) {
1353                                 tx = async_copy_data(0, rbi, &dev->page,
1354                                                      dev->sector, tx, sh, 0);
1355                                 rbi = r5_next_bio(rbi, dev->sector);
1356                         }
1357                 }
1358         }
1359
1360         atomic_inc(&sh->count);
1361         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1362         async_trigger_callback(&submit);
1363 }
1364
1365 static void mark_target_uptodate(struct stripe_head *sh, int target)
1366 {
1367         struct r5dev *tgt;
1368
1369         if (target < 0)
1370                 return;
1371
1372         tgt = &sh->dev[target];
1373         set_bit(R5_UPTODATE, &tgt->flags);
1374         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1375         clear_bit(R5_Wantcompute, &tgt->flags);
1376 }
1377
1378 static void ops_complete_compute(void *stripe_head_ref)
1379 {
1380         struct stripe_head *sh = stripe_head_ref;
1381
1382         pr_debug("%s: stripe %llu\n", __func__,
1383                 (unsigned long long)sh->sector);
1384
1385         /* mark the computed target(s) as uptodate */
1386         mark_target_uptodate(sh, sh->ops.target);
1387         mark_target_uptodate(sh, sh->ops.target2);
1388
1389         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1390         if (sh->check_state == check_state_compute_run)
1391                 sh->check_state = check_state_compute_result;
1392         set_bit(STRIPE_HANDLE, &sh->state);
1393         raid5_release_stripe(sh);
1394 }
1395
1396 /* return a pointer to the address conversion region of the scribble buffer */
1397 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1398                                  struct raid5_percpu *percpu, int i)
1399 {
1400         void *addr;
1401
1402         addr = flex_array_get(percpu->scribble, i);
1403         return addr + sizeof(struct page *) * (sh->disks + 2);
1404 }
1405
1406 /* return a pointer to the address conversion region of the scribble buffer */
1407 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1408 {
1409         void *addr;
1410
1411         addr = flex_array_get(percpu->scribble, i);
1412         return addr;
1413 }
1414
1415 static struct dma_async_tx_descriptor *
1416 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1417 {
1418         int disks = sh->disks;
1419         struct page **xor_srcs = to_addr_page(percpu, 0);
1420         int target = sh->ops.target;
1421         struct r5dev *tgt = &sh->dev[target];
1422         struct page *xor_dest = tgt->page;
1423         int count = 0;
1424         struct dma_async_tx_descriptor *tx;
1425         struct async_submit_ctl submit;
1426         int i;
1427
1428         BUG_ON(sh->batch_head);
1429
1430         pr_debug("%s: stripe %llu block: %d\n",
1431                 __func__, (unsigned long long)sh->sector, target);
1432         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1433
1434         for (i = disks; i--; )
1435                 if (i != target)
1436                         xor_srcs[count++] = sh->dev[i].page;
1437
1438         atomic_inc(&sh->count);
1439
1440         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1441                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1442         if (unlikely(count == 1))
1443                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1444         else
1445                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1446
1447         return tx;
1448 }
1449
1450 /* set_syndrome_sources - populate source buffers for gen_syndrome
1451  * @srcs - (struct page *) array of size sh->disks
1452  * @sh - stripe_head to parse
1453  *
1454  * Populates srcs in proper layout order for the stripe and returns the
1455  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1456  * destination buffer is recorded in srcs[count] and the Q destination
1457  * is recorded in srcs[count+1]].
1458  */
1459 static int set_syndrome_sources(struct page **srcs,
1460                                 struct stripe_head *sh,
1461                                 int srctype)
1462 {
1463         int disks = sh->disks;
1464         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1465         int d0_idx = raid6_d0(sh);
1466         int count;
1467         int i;
1468
1469         for (i = 0; i < disks; i++)
1470                 srcs[i] = NULL;
1471
1472         count = 0;
1473         i = d0_idx;
1474         do {
1475                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1476                 struct r5dev *dev = &sh->dev[i];
1477
1478                 if (i == sh->qd_idx || i == sh->pd_idx ||
1479                     (srctype == SYNDROME_SRC_ALL) ||
1480                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1481                      (test_bit(R5_Wantdrain, &dev->flags) ||
1482                       test_bit(R5_InJournal, &dev->flags))) ||
1483                     (srctype == SYNDROME_SRC_WRITTEN &&
1484                      (dev->written ||
1485                       test_bit(R5_InJournal, &dev->flags)))) {
1486                         if (test_bit(R5_InJournal, &dev->flags))
1487                                 srcs[slot] = sh->dev[i].orig_page;
1488                         else
1489                                 srcs[slot] = sh->dev[i].page;
1490                 }
1491                 i = raid6_next_disk(i, disks);
1492         } while (i != d0_idx);
1493
1494         return syndrome_disks;
1495 }
1496
1497 static struct dma_async_tx_descriptor *
1498 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1499 {
1500         int disks = sh->disks;
1501         struct page **blocks = to_addr_page(percpu, 0);
1502         int target;
1503         int qd_idx = sh->qd_idx;
1504         struct dma_async_tx_descriptor *tx;
1505         struct async_submit_ctl submit;
1506         struct r5dev *tgt;
1507         struct page *dest;
1508         int i;
1509         int count;
1510
1511         BUG_ON(sh->batch_head);
1512         if (sh->ops.target < 0)
1513                 target = sh->ops.target2;
1514         else if (sh->ops.target2 < 0)
1515                 target = sh->ops.target;
1516         else
1517                 /* we should only have one valid target */
1518                 BUG();
1519         BUG_ON(target < 0);
1520         pr_debug("%s: stripe %llu block: %d\n",
1521                 __func__, (unsigned long long)sh->sector, target);
1522
1523         tgt = &sh->dev[target];
1524         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1525         dest = tgt->page;
1526
1527         atomic_inc(&sh->count);
1528
1529         if (target == qd_idx) {
1530                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1531                 blocks[count] = NULL; /* regenerating p is not necessary */
1532                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1533                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1534                                   ops_complete_compute, sh,
1535                                   to_addr_conv(sh, percpu, 0));
1536                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1537         } else {
1538                 /* Compute any data- or p-drive using XOR */
1539                 count = 0;
1540                 for (i = disks; i-- ; ) {
1541                         if (i == target || i == qd_idx)
1542                                 continue;
1543                         blocks[count++] = sh->dev[i].page;
1544                 }
1545
1546                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1547                                   NULL, ops_complete_compute, sh,
1548                                   to_addr_conv(sh, percpu, 0));
1549                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1550         }
1551
1552         return tx;
1553 }
1554
1555 static struct dma_async_tx_descriptor *
1556 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1557 {
1558         int i, count, disks = sh->disks;
1559         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1560         int d0_idx = raid6_d0(sh);
1561         int faila = -1, failb = -1;
1562         int target = sh->ops.target;
1563         int target2 = sh->ops.target2;
1564         struct r5dev *tgt = &sh->dev[target];
1565         struct r5dev *tgt2 = &sh->dev[target2];
1566         struct dma_async_tx_descriptor *tx;
1567         struct page **blocks = to_addr_page(percpu, 0);
1568         struct async_submit_ctl submit;
1569
1570         BUG_ON(sh->batch_head);
1571         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1572                  __func__, (unsigned long long)sh->sector, target, target2);
1573         BUG_ON(target < 0 || target2 < 0);
1574         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1575         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1576
1577         /* we need to open-code set_syndrome_sources to handle the
1578          * slot number conversion for 'faila' and 'failb'
1579          */
1580         for (i = 0; i < disks ; i++)
1581                 blocks[i] = NULL;
1582         count = 0;
1583         i = d0_idx;
1584         do {
1585                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1586
1587                 blocks[slot] = sh->dev[i].page;
1588
1589                 if (i == target)
1590                         faila = slot;
1591                 if (i == target2)
1592                         failb = slot;
1593                 i = raid6_next_disk(i, disks);
1594         } while (i != d0_idx);
1595
1596         BUG_ON(faila == failb);
1597         if (failb < faila)
1598                 swap(faila, failb);
1599         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1600                  __func__, (unsigned long long)sh->sector, faila, failb);
1601
1602         atomic_inc(&sh->count);
1603
1604         if (failb == syndrome_disks+1) {
1605                 /* Q disk is one of the missing disks */
1606                 if (faila == syndrome_disks) {
1607                         /* Missing P+Q, just recompute */
1608                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1609                                           ops_complete_compute, sh,
1610                                           to_addr_conv(sh, percpu, 0));
1611                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1612                                                   STRIPE_SIZE, &submit);
1613                 } else {
1614                         struct page *dest;
1615                         int data_target;
1616                         int qd_idx = sh->qd_idx;
1617
1618                         /* Missing D+Q: recompute D from P, then recompute Q */
1619                         if (target == qd_idx)
1620                                 data_target = target2;
1621                         else
1622                                 data_target = target;
1623
1624                         count = 0;
1625                         for (i = disks; i-- ; ) {
1626                                 if (i == data_target || i == qd_idx)
1627                                         continue;
1628                                 blocks[count++] = sh->dev[i].page;
1629                         }
1630                         dest = sh->dev[data_target].page;
1631                         init_async_submit(&submit,
1632                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1633                                           NULL, NULL, NULL,
1634                                           to_addr_conv(sh, percpu, 0));
1635                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1636                                        &submit);
1637
1638                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1639                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1640                                           ops_complete_compute, sh,
1641                                           to_addr_conv(sh, percpu, 0));
1642                         return async_gen_syndrome(blocks, 0, count+2,
1643                                                   STRIPE_SIZE, &submit);
1644                 }
1645         } else {
1646                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1647                                   ops_complete_compute, sh,
1648                                   to_addr_conv(sh, percpu, 0));
1649                 if (failb == syndrome_disks) {
1650                         /* We're missing D+P. */
1651                         return async_raid6_datap_recov(syndrome_disks+2,
1652                                                        STRIPE_SIZE, faila,
1653                                                        blocks, &submit);
1654                 } else {
1655                         /* We're missing D+D. */
1656                         return async_raid6_2data_recov(syndrome_disks+2,
1657                                                        STRIPE_SIZE, faila, failb,
1658                                                        blocks, &submit);
1659                 }
1660         }
1661 }
1662
1663 static void ops_complete_prexor(void *stripe_head_ref)
1664 {
1665         struct stripe_head *sh = stripe_head_ref;
1666
1667         pr_debug("%s: stripe %llu\n", __func__,
1668                 (unsigned long long)sh->sector);
1669
1670         if (r5c_is_writeback(sh->raid_conf->log))
1671                 /*
1672                  * raid5-cache write back uses orig_page during prexor.
1673                  * After prexor, it is time to free orig_page
1674                  */
1675                 r5c_release_extra_page(sh);
1676 }
1677
1678 static struct dma_async_tx_descriptor *
1679 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1680                 struct dma_async_tx_descriptor *tx)
1681 {
1682         int disks = sh->disks;
1683         struct page **xor_srcs = to_addr_page(percpu, 0);
1684         int count = 0, pd_idx = sh->pd_idx, i;
1685         struct async_submit_ctl submit;
1686
1687         /* existing parity data subtracted */
1688         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1689
1690         BUG_ON(sh->batch_head);
1691         pr_debug("%s: stripe %llu\n", __func__,
1692                 (unsigned long long)sh->sector);
1693
1694         for (i = disks; i--; ) {
1695                 struct r5dev *dev = &sh->dev[i];
1696                 /* Only process blocks that are known to be uptodate */
1697                 if (test_bit(R5_InJournal, &dev->flags))
1698                         xor_srcs[count++] = dev->orig_page;
1699                 else if (test_bit(R5_Wantdrain, &dev->flags))
1700                         xor_srcs[count++] = dev->page;
1701         }
1702
1703         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1704                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1705         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1706
1707         return tx;
1708 }
1709
1710 static struct dma_async_tx_descriptor *
1711 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1712                 struct dma_async_tx_descriptor *tx)
1713 {
1714         struct page **blocks = to_addr_page(percpu, 0);
1715         int count;
1716         struct async_submit_ctl submit;
1717
1718         pr_debug("%s: stripe %llu\n", __func__,
1719                 (unsigned long long)sh->sector);
1720
1721         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1722
1723         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1724                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1725         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1726
1727         return tx;
1728 }
1729
1730 static struct dma_async_tx_descriptor *
1731 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1732 {
1733         struct r5conf *conf = sh->raid_conf;
1734         int disks = sh->disks;
1735         int i;
1736         struct stripe_head *head_sh = sh;
1737
1738         pr_debug("%s: stripe %llu\n", __func__,
1739                 (unsigned long long)sh->sector);
1740
1741         for (i = disks; i--; ) {
1742                 struct r5dev *dev;
1743                 struct bio *chosen;
1744
1745                 sh = head_sh;
1746                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1747                         struct bio *wbi;
1748
1749 again:
1750                         dev = &sh->dev[i];
1751                         /*
1752                          * clear R5_InJournal, so when rewriting a page in
1753                          * journal, it is not skipped by r5l_log_stripe()
1754                          */
1755                         clear_bit(R5_InJournal, &dev->flags);
1756                         spin_lock_irq(&sh->stripe_lock);
1757                         chosen = dev->towrite;
1758                         dev->towrite = NULL;
1759                         sh->overwrite_disks = 0;
1760                         BUG_ON(dev->written);
1761                         wbi = dev->written = chosen;
1762                         spin_unlock_irq(&sh->stripe_lock);
1763                         WARN_ON(dev->page != dev->orig_page);
1764
1765                         while (wbi && wbi->bi_iter.bi_sector <
1766                                 dev->sector + STRIPE_SECTORS) {
1767                                 if (wbi->bi_opf & REQ_FUA)
1768                                         set_bit(R5_WantFUA, &dev->flags);
1769                                 if (wbi->bi_opf & REQ_SYNC)
1770                                         set_bit(R5_SyncIO, &dev->flags);
1771                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1772                                         set_bit(R5_Discard, &dev->flags);
1773                                 else {
1774                                         tx = async_copy_data(1, wbi, &dev->page,
1775                                                              dev->sector, tx, sh,
1776                                                              r5c_is_writeback(conf->log));
1777                                         if (dev->page != dev->orig_page &&
1778                                             !r5c_is_writeback(conf->log)) {
1779                                                 set_bit(R5_SkipCopy, &dev->flags);
1780                                                 clear_bit(R5_UPTODATE, &dev->flags);
1781                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1782                                         }
1783                                 }
1784                                 wbi = r5_next_bio(wbi, dev->sector);
1785                         }
1786
1787                         if (head_sh->batch_head) {
1788                                 sh = list_first_entry(&sh->batch_list,
1789                                                       struct stripe_head,
1790                                                       batch_list);
1791                                 if (sh == head_sh)
1792                                         continue;
1793                                 goto again;
1794                         }
1795                 }
1796         }
1797
1798         return tx;
1799 }
1800
1801 static void ops_complete_reconstruct(void *stripe_head_ref)
1802 {
1803         struct stripe_head *sh = stripe_head_ref;
1804         int disks = sh->disks;
1805         int pd_idx = sh->pd_idx;
1806         int qd_idx = sh->qd_idx;
1807         int i;
1808         bool fua = false, sync = false, discard = false;
1809
1810         pr_debug("%s: stripe %llu\n", __func__,
1811                 (unsigned long long)sh->sector);
1812
1813         for (i = disks; i--; ) {
1814                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1815                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1816                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1817         }
1818
1819         for (i = disks; i--; ) {
1820                 struct r5dev *dev = &sh->dev[i];
1821
1822                 if (dev->written || i == pd_idx || i == qd_idx) {
1823                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1824                                 set_bit(R5_UPTODATE, &dev->flags);
1825                                 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1826                                         set_bit(R5_Expanded, &dev->flags);
1827                         }
1828                         if (fua)
1829                                 set_bit(R5_WantFUA, &dev->flags);
1830                         if (sync)
1831                                 set_bit(R5_SyncIO, &dev->flags);
1832                 }
1833         }
1834
1835         if (sh->reconstruct_state == reconstruct_state_drain_run)
1836                 sh->reconstruct_state = reconstruct_state_drain_result;
1837         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1838                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1839         else {
1840                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1841                 sh->reconstruct_state = reconstruct_state_result;
1842         }
1843
1844         set_bit(STRIPE_HANDLE, &sh->state);
1845         raid5_release_stripe(sh);
1846 }
1847
1848 static void
1849 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1850                      struct dma_async_tx_descriptor *tx)
1851 {
1852         int disks = sh->disks;
1853         struct page **xor_srcs;
1854         struct async_submit_ctl submit;
1855         int count, pd_idx = sh->pd_idx, i;
1856         struct page *xor_dest;
1857         int prexor = 0;
1858         unsigned long flags;
1859         int j = 0;
1860         struct stripe_head *head_sh = sh;
1861         int last_stripe;
1862
1863         pr_debug("%s: stripe %llu\n", __func__,
1864                 (unsigned long long)sh->sector);
1865
1866         for (i = 0; i < sh->disks; i++) {
1867                 if (pd_idx == i)
1868                         continue;
1869                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1870                         break;
1871         }
1872         if (i >= sh->disks) {
1873                 atomic_inc(&sh->count);
1874                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1875                 ops_complete_reconstruct(sh);
1876                 return;
1877         }
1878 again:
1879         count = 0;
1880         xor_srcs = to_addr_page(percpu, j);
1881         /* check if prexor is active which means only process blocks
1882          * that are part of a read-modify-write (written)
1883          */
1884         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1885                 prexor = 1;
1886                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1887                 for (i = disks; i--; ) {
1888                         struct r5dev *dev = &sh->dev[i];
1889                         if (head_sh->dev[i].written ||
1890                             test_bit(R5_InJournal, &head_sh->dev[i].flags))
1891                                 xor_srcs[count++] = dev->page;
1892                 }
1893         } else {
1894                 xor_dest = sh->dev[pd_idx].page;
1895                 for (i = disks; i--; ) {
1896                         struct r5dev *dev = &sh->dev[i];
1897                         if (i != pd_idx)
1898                                 xor_srcs[count++] = dev->page;
1899                 }
1900         }
1901
1902         /* 1/ if we prexor'd then the dest is reused as a source
1903          * 2/ if we did not prexor then we are redoing the parity
1904          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1905          * for the synchronous xor case
1906          */
1907         last_stripe = !head_sh->batch_head ||
1908                 list_first_entry(&sh->batch_list,
1909                                  struct stripe_head, batch_list) == head_sh;
1910         if (last_stripe) {
1911                 flags = ASYNC_TX_ACK |
1912                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1913
1914                 atomic_inc(&head_sh->count);
1915                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1916                                   to_addr_conv(sh, percpu, j));
1917         } else {
1918                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1919                 init_async_submit(&submit, flags, tx, NULL, NULL,
1920                                   to_addr_conv(sh, percpu, j));
1921         }
1922
1923         if (unlikely(count == 1))
1924                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1925         else
1926                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1927         if (!last_stripe) {
1928                 j++;
1929                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1930                                       batch_list);
1931                 goto again;
1932         }
1933 }
1934
1935 static void
1936 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1937                      struct dma_async_tx_descriptor *tx)
1938 {
1939         struct async_submit_ctl submit;
1940         struct page **blocks;
1941         int count, i, j = 0;
1942         struct stripe_head *head_sh = sh;
1943         int last_stripe;
1944         int synflags;
1945         unsigned long txflags;
1946
1947         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1948
1949         for (i = 0; i < sh->disks; i++) {
1950                 if (sh->pd_idx == i || sh->qd_idx == i)
1951                         continue;
1952                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1953                         break;
1954         }
1955         if (i >= sh->disks) {
1956                 atomic_inc(&sh->count);
1957                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1958                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1959                 ops_complete_reconstruct(sh);
1960                 return;
1961         }
1962
1963 again:
1964         blocks = to_addr_page(percpu, j);
1965
1966         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1967                 synflags = SYNDROME_SRC_WRITTEN;
1968                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1969         } else {
1970                 synflags = SYNDROME_SRC_ALL;
1971                 txflags = ASYNC_TX_ACK;
1972         }
1973
1974         count = set_syndrome_sources(blocks, sh, synflags);
1975         last_stripe = !head_sh->batch_head ||
1976                 list_first_entry(&sh->batch_list,
1977                                  struct stripe_head, batch_list) == head_sh;
1978
1979         if (last_stripe) {
1980                 atomic_inc(&head_sh->count);
1981                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1982                                   head_sh, to_addr_conv(sh, percpu, j));
1983         } else
1984                 init_async_submit(&submit, 0, tx, NULL, NULL,
1985                                   to_addr_conv(sh, percpu, j));
1986         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1987         if (!last_stripe) {
1988                 j++;
1989                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1990                                       batch_list);
1991                 goto again;
1992         }
1993 }
1994
1995 static void ops_complete_check(void *stripe_head_ref)
1996 {
1997         struct stripe_head *sh = stripe_head_ref;
1998
1999         pr_debug("%s: stripe %llu\n", __func__,
2000                 (unsigned long long)sh->sector);
2001
2002         sh->check_state = check_state_check_result;
2003         set_bit(STRIPE_HANDLE, &sh->state);
2004         raid5_release_stripe(sh);
2005 }
2006
2007 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2008 {
2009         int disks = sh->disks;
2010         int pd_idx = sh->pd_idx;
2011         int qd_idx = sh->qd_idx;
2012         struct page *xor_dest;
2013         struct page **xor_srcs = to_addr_page(percpu, 0);
2014         struct dma_async_tx_descriptor *tx;
2015         struct async_submit_ctl submit;
2016         int count;
2017         int i;
2018
2019         pr_debug("%s: stripe %llu\n", __func__,
2020                 (unsigned long long)sh->sector);
2021
2022         BUG_ON(sh->batch_head);
2023         count = 0;
2024         xor_dest = sh->dev[pd_idx].page;
2025         xor_srcs[count++] = xor_dest;
2026         for (i = disks; i--; ) {
2027                 if (i == pd_idx || i == qd_idx)
2028                         continue;
2029                 xor_srcs[count++] = sh->dev[i].page;
2030         }
2031
2032         init_async_submit(&submit, 0, NULL, NULL, NULL,
2033                           to_addr_conv(sh, percpu, 0));
2034         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2035                            &sh->ops.zero_sum_result, &submit);
2036
2037         atomic_inc(&sh->count);
2038         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2039         tx = async_trigger_callback(&submit);
2040 }
2041
2042 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2043 {
2044         struct page **srcs = to_addr_page(percpu, 0);
2045         struct async_submit_ctl submit;
2046         int count;
2047
2048         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2049                 (unsigned long long)sh->sector, checkp);
2050
2051         BUG_ON(sh->batch_head);
2052         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2053         if (!checkp)
2054                 srcs[count] = NULL;
2055
2056         atomic_inc(&sh->count);
2057         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2058                           sh, to_addr_conv(sh, percpu, 0));
2059         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2060                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2061 }
2062
2063 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2064 {
2065         int overlap_clear = 0, i, disks = sh->disks;
2066         struct dma_async_tx_descriptor *tx = NULL;
2067         struct r5conf *conf = sh->raid_conf;
2068         int level = conf->level;
2069         struct raid5_percpu *percpu;
2070         unsigned long cpu;
2071
2072         cpu = get_cpu();
2073         percpu = per_cpu_ptr(conf->percpu, cpu);
2074         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2075                 ops_run_biofill(sh);
2076                 overlap_clear++;
2077         }
2078
2079         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2080                 if (level < 6)
2081                         tx = ops_run_compute5(sh, percpu);
2082                 else {
2083                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
2084                                 tx = ops_run_compute6_1(sh, percpu);
2085                         else
2086                                 tx = ops_run_compute6_2(sh, percpu);
2087                 }
2088                 /* terminate the chain if reconstruct is not set to be run */
2089                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2090                         async_tx_ack(tx);
2091         }
2092
2093         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2094                 if (level < 6)
2095                         tx = ops_run_prexor5(sh, percpu, tx);
2096                 else
2097                         tx = ops_run_prexor6(sh, percpu, tx);
2098         }
2099
2100         if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2101                 tx = ops_run_partial_parity(sh, percpu, tx);
2102
2103         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2104                 tx = ops_run_biodrain(sh, tx);
2105                 overlap_clear++;
2106         }
2107
2108         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2109                 if (level < 6)
2110                         ops_run_reconstruct5(sh, percpu, tx);
2111                 else
2112                         ops_run_reconstruct6(sh, percpu, tx);
2113         }
2114
2115         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2116                 if (sh->check_state == check_state_run)
2117                         ops_run_check_p(sh, percpu);
2118                 else if (sh->check_state == check_state_run_q)
2119                         ops_run_check_pq(sh, percpu, 0);
2120                 else if (sh->check_state == check_state_run_pq)
2121                         ops_run_check_pq(sh, percpu, 1);
2122                 else
2123                         BUG();
2124         }
2125
2126         if (overlap_clear && !sh->batch_head)
2127                 for (i = disks; i--; ) {
2128                         struct r5dev *dev = &sh->dev[i];
2129                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2130                                 wake_up(&sh->raid_conf->wait_for_overlap);
2131                 }
2132         put_cpu();
2133 }
2134
2135 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2136 {
2137         if (sh->ppl_page)
2138                 __free_page(sh->ppl_page);
2139         kmem_cache_free(sc, sh);
2140 }
2141
2142 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2143         int disks, struct r5conf *conf)
2144 {
2145         struct stripe_head *sh;
2146         int i;
2147
2148         sh = kmem_cache_zalloc(sc, gfp);
2149         if (sh) {
2150                 spin_lock_init(&sh->stripe_lock);
2151                 spin_lock_init(&sh->batch_lock);
2152                 INIT_LIST_HEAD(&sh->batch_list);
2153                 INIT_LIST_HEAD(&sh->lru);
2154                 INIT_LIST_HEAD(&sh->r5c);
2155                 INIT_LIST_HEAD(&sh->log_list);
2156                 atomic_set(&sh->count, 1);
2157                 sh->raid_conf = conf;
2158                 sh->log_start = MaxSector;
2159                 for (i = 0; i < disks; i++) {
2160                         struct r5dev *dev = &sh->dev[i];
2161
2162                         bio_init(&dev->req, &dev->vec, 1);
2163                         bio_init(&dev->rreq, &dev->rvec, 1);
2164                 }
2165
2166                 if (raid5_has_ppl(conf)) {
2167                         sh->ppl_page = alloc_page(gfp);
2168                         if (!sh->ppl_page) {
2169                                 free_stripe(sc, sh);
2170                                 sh = NULL;
2171                         }
2172                 }
2173         }
2174         return sh;
2175 }
2176 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2177 {
2178         struct stripe_head *sh;
2179
2180         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2181         if (!sh)
2182                 return 0;
2183
2184         if (grow_buffers(sh, gfp)) {
2185                 shrink_buffers(sh);
2186                 free_stripe(conf->slab_cache, sh);
2187                 return 0;
2188         }
2189         sh->hash_lock_index =
2190                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2191         /* we just created an active stripe so... */
2192         atomic_inc(&conf->active_stripes);
2193
2194         raid5_release_stripe(sh);
2195         conf->max_nr_stripes++;
2196         return 1;
2197 }
2198
2199 static int grow_stripes(struct r5conf *conf, int num)
2200 {
2201         struct kmem_cache *sc;
2202         size_t namelen = sizeof(conf->cache_name[0]);
2203         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2204
2205         if (conf->mddev->gendisk)
2206                 snprintf(conf->cache_name[0], namelen,
2207                         "raid%d-%s", conf->level, mdname(conf->mddev));
2208         else
2209                 snprintf(conf->cache_name[0], namelen,
2210                         "raid%d-%p", conf->level, conf->mddev);
2211         snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2212
2213         conf->active_name = 0;
2214         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2215                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2216                                0, 0, NULL);
2217         if (!sc)
2218                 return 1;
2219         conf->slab_cache = sc;
2220         conf->pool_size = devs;
2221         while (num--)
2222                 if (!grow_one_stripe(conf, GFP_KERNEL))
2223                         return 1;
2224
2225         return 0;
2226 }
2227
2228 /**
2229  * scribble_len - return the required size of the scribble region
2230  * @num - total number of disks in the array
2231  *
2232  * The size must be enough to contain:
2233  * 1/ a struct page pointer for each device in the array +2
2234  * 2/ room to convert each entry in (1) to its corresponding dma
2235  *    (dma_map_page()) or page (page_address()) address.
2236  *
2237  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2238  * calculate over all devices (not just the data blocks), using zeros in place
2239  * of the P and Q blocks.
2240  */
2241 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2242 {
2243         struct flex_array *ret;
2244         size_t len;
2245
2246         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2247         ret = flex_array_alloc(len, cnt, flags);
2248         if (!ret)
2249                 return NULL;
2250         /* always prealloc all elements, so no locking is required */
2251         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2252                 flex_array_free(ret);
2253                 return NULL;
2254         }
2255         return ret;
2256 }
2257
2258 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2259 {
2260         unsigned long cpu;
2261         int err = 0;
2262
2263         /*
2264          * Never shrink. And mddev_suspend() could deadlock if this is called
2265          * from raid5d. In that case, scribble_disks and scribble_sectors
2266          * should equal to new_disks and new_sectors
2267          */
2268         if (conf->scribble_disks >= new_disks &&
2269             conf->scribble_sectors >= new_sectors)
2270                 return 0;
2271         mddev_suspend(conf->mddev);
2272         get_online_cpus();
2273         for_each_present_cpu(cpu) {
2274                 struct raid5_percpu *percpu;
2275                 struct flex_array *scribble;
2276
2277                 percpu = per_cpu_ptr(conf->percpu, cpu);
2278                 scribble = scribble_alloc(new_disks,
2279                                           new_sectors / STRIPE_SECTORS,
2280                                           GFP_NOIO);
2281
2282                 if (scribble) {
2283                         flex_array_free(percpu->scribble);
2284                         percpu->scribble = scribble;
2285                 } else {
2286                         err = -ENOMEM;
2287                         break;
2288                 }
2289         }
2290         put_online_cpus();
2291         mddev_resume(conf->mddev);
2292         if (!err) {
2293                 conf->scribble_disks = new_disks;
2294                 conf->scribble_sectors = new_sectors;
2295         }
2296         return err;
2297 }
2298
2299 static int resize_stripes(struct r5conf *conf, int newsize)
2300 {
2301         /* Make all the stripes able to hold 'newsize' devices.
2302          * New slots in each stripe get 'page' set to a new page.
2303          *
2304          * This happens in stages:
2305          * 1/ create a new kmem_cache and allocate the required number of
2306          *    stripe_heads.
2307          * 2/ gather all the old stripe_heads and transfer the pages across
2308          *    to the new stripe_heads.  This will have the side effect of
2309          *    freezing the array as once all stripe_heads have been collected,
2310          *    no IO will be possible.  Old stripe heads are freed once their
2311          *    pages have been transferred over, and the old kmem_cache is
2312          *    freed when all stripes are done.
2313          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2314          *    we simple return a failure status - no need to clean anything up.
2315          * 4/ allocate new pages for the new slots in the new stripe_heads.
2316          *    If this fails, we don't bother trying the shrink the
2317          *    stripe_heads down again, we just leave them as they are.
2318          *    As each stripe_head is processed the new one is released into
2319          *    active service.
2320          *
2321          * Once step2 is started, we cannot afford to wait for a write,
2322          * so we use GFP_NOIO allocations.
2323          */
2324         struct stripe_head *osh, *nsh;
2325         LIST_HEAD(newstripes);
2326         struct disk_info *ndisks;
2327         int err = 0;
2328         struct kmem_cache *sc;
2329         int i;
2330         int hash, cnt;
2331
2332         md_allow_write(conf->mddev);
2333
2334         /* Step 1 */
2335         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2336                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2337                                0, 0, NULL);
2338         if (!sc)
2339                 return -ENOMEM;
2340
2341         /* Need to ensure auto-resizing doesn't interfere */
2342         mutex_lock(&conf->cache_size_mutex);
2343
2344         for (i = conf->max_nr_stripes; i; i--) {
2345                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2346                 if (!nsh)
2347                         break;
2348
2349                 list_add(&nsh->lru, &newstripes);
2350         }
2351         if (i) {
2352                 /* didn't get enough, give up */
2353                 while (!list_empty(&newstripes)) {
2354                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2355                         list_del(&nsh->lru);
2356                         free_stripe(sc, nsh);
2357                 }
2358                 kmem_cache_destroy(sc);
2359                 mutex_unlock(&conf->cache_size_mutex);
2360                 return -ENOMEM;
2361         }
2362         /* Step 2 - Must use GFP_NOIO now.
2363          * OK, we have enough stripes, start collecting inactive
2364          * stripes and copying them over
2365          */
2366         hash = 0;
2367         cnt = 0;
2368         list_for_each_entry(nsh, &newstripes, lru) {
2369                 lock_device_hash_lock(conf, hash);
2370                 wait_event_cmd(conf->wait_for_stripe,
2371                                     !list_empty(conf->inactive_list + hash),
2372                                     unlock_device_hash_lock(conf, hash),
2373                                     lock_device_hash_lock(conf, hash));
2374                 osh = get_free_stripe(conf, hash);
2375                 unlock_device_hash_lock(conf, hash);
2376
2377                 for(i=0; i<conf->pool_size; i++) {
2378                         nsh->dev[i].page = osh->dev[i].page;
2379                         nsh->dev[i].orig_page = osh->dev[i].page;
2380                 }
2381                 nsh->hash_lock_index = hash;
2382                 free_stripe(conf->slab_cache, osh);
2383                 cnt++;
2384                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2385                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2386                         hash++;
2387                         cnt = 0;
2388                 }
2389         }
2390         kmem_cache_destroy(conf->slab_cache);
2391
2392         /* Step 3.
2393          * At this point, we are holding all the stripes so the array
2394          * is completely stalled, so now is a good time to resize
2395          * conf->disks and the scribble region
2396          */
2397         ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2398         if (ndisks) {
2399                 for (i = 0; i < conf->pool_size; i++)
2400                         ndisks[i] = conf->disks[i];
2401
2402                 for (i = conf->pool_size; i < newsize; i++) {
2403                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2404                         if (!ndisks[i].extra_page)
2405                                 err = -ENOMEM;
2406                 }
2407
2408                 if (err) {
2409                         for (i = conf->pool_size; i < newsize; i++)
2410                                 if (ndisks[i].extra_page)
2411                                         put_page(ndisks[i].extra_page);
2412                         kfree(ndisks);
2413                 } else {
2414                         kfree(conf->disks);
2415                         conf->disks = ndisks;
2416                 }
2417         } else
2418                 err = -ENOMEM;
2419
2420         mutex_unlock(&conf->cache_size_mutex);
2421
2422         conf->slab_cache = sc;
2423         conf->active_name = 1-conf->active_name;
2424
2425         /* Step 4, return new stripes to service */
2426         while(!list_empty(&newstripes)) {
2427                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2428                 list_del_init(&nsh->lru);
2429
2430                 for (i=conf->raid_disks; i < newsize; i++)
2431                         if (nsh->dev[i].page == NULL) {
2432                                 struct page *p = alloc_page(GFP_NOIO);
2433                                 nsh->dev[i].page = p;
2434                                 nsh->dev[i].orig_page = p;
2435                                 if (!p)
2436                                         err = -ENOMEM;
2437                         }
2438                 raid5_release_stripe(nsh);
2439         }
2440         /* critical section pass, GFP_NOIO no longer needed */
2441
2442         if (!err)
2443                 conf->pool_size = newsize;
2444         return err;
2445 }
2446
2447 static int drop_one_stripe(struct r5conf *conf)
2448 {
2449         struct stripe_head *sh;
2450         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2451
2452         spin_lock_irq(conf->hash_locks + hash);
2453         sh = get_free_stripe(conf, hash);
2454         spin_unlock_irq(conf->hash_locks + hash);
2455         if (!sh)
2456                 return 0;
2457         BUG_ON(atomic_read(&sh->count));
2458         shrink_buffers(sh);
2459         free_stripe(conf->slab_cache, sh);
2460         atomic_dec(&conf->active_stripes);
2461         conf->max_nr_stripes--;
2462         return 1;
2463 }
2464
2465 static void shrink_stripes(struct r5conf *conf)
2466 {
2467         while (conf->max_nr_stripes &&
2468                drop_one_stripe(conf))
2469                 ;
2470
2471         kmem_cache_destroy(conf->slab_cache);
2472         conf->slab_cache = NULL;
2473 }
2474
2475 static void raid5_end_read_request(struct bio * bi)
2476 {
2477         struct stripe_head *sh = bi->bi_private;
2478         struct r5conf *conf = sh->raid_conf;
2479         int disks = sh->disks, i;
2480         char b[BDEVNAME_SIZE];
2481         struct md_rdev *rdev = NULL;
2482         sector_t s;
2483
2484         for (i=0 ; i<disks; i++)
2485                 if (bi == &sh->dev[i].req)
2486                         break;
2487
2488         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2489                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2490                 bi->bi_status);
2491         if (i == disks) {
2492                 bio_reset(bi);
2493                 BUG();
2494                 return;
2495         }
2496         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2497                 /* If replacement finished while this request was outstanding,
2498                  * 'replacement' might be NULL already.
2499                  * In that case it moved down to 'rdev'.
2500                  * rdev is not removed until all requests are finished.
2501                  */
2502                 rdev = conf->disks[i].replacement;
2503         if (!rdev)
2504                 rdev = conf->disks[i].rdev;
2505
2506         if (use_new_offset(conf, sh))
2507                 s = sh->sector + rdev->new_data_offset;
2508         else
2509                 s = sh->sector + rdev->data_offset;
2510         if (!bi->bi_status) {
2511                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2512                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2513                         /* Note that this cannot happen on a
2514                          * replacement device.  We just fail those on
2515                          * any error
2516                          */
2517                         pr_info_ratelimited(
2518                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2519                                 mdname(conf->mddev), STRIPE_SECTORS,
2520                                 (unsigned long long)s,
2521                                 bdevname(rdev->bdev, b));
2522                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2523                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2524                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2525                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2526                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2527
2528                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2529                         /*
2530                          * end read for a page in journal, this
2531                          * must be preparing for prexor in rmw
2532                          */
2533                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2534
2535                 if (atomic_read(&rdev->read_errors))
2536                         atomic_set(&rdev->read_errors, 0);
2537         } else {
2538                 const char *bdn = bdevname(rdev->bdev, b);
2539                 int retry = 0;
2540                 int set_bad = 0;
2541
2542                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2543                 atomic_inc(&rdev->read_errors);
2544                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2545                         pr_warn_ratelimited(
2546                                 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2547                                 mdname(conf->mddev),
2548                                 (unsigned long long)s,
2549                                 bdn);
2550                 else if (conf->mddev->degraded >= conf->max_degraded) {
2551                         set_bad = 1;
2552                         pr_warn_ratelimited(
2553                                 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2554                                 mdname(conf->mddev),
2555                                 (unsigned long long)s,
2556                                 bdn);
2557                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2558                         /* Oh, no!!! */
2559                         set_bad = 1;
2560                         pr_warn_ratelimited(
2561                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2562                                 mdname(conf->mddev),
2563                                 (unsigned long long)s,
2564                                 bdn);
2565                 } else if (atomic_read(&rdev->read_errors)
2566                          > conf->max_nr_stripes)
2567                         pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2568                                mdname(conf->mddev), bdn);
2569                 else
2570                         retry = 1;
2571                 if (set_bad && test_bit(In_sync, &rdev->flags)
2572                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2573                         retry = 1;
2574                 if (retry)
2575                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2576                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2577                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2578                         } else
2579                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2580                 else {
2581                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2582                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2583                         if (!(set_bad
2584                               && test_bit(In_sync, &rdev->flags)
2585                               && rdev_set_badblocks(
2586                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2587                                 md_error(conf->mddev, rdev);
2588                 }
2589         }
2590         rdev_dec_pending(rdev, conf->mddev);
2591         bio_reset(bi);
2592         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2593         set_bit(STRIPE_HANDLE, &sh->state);
2594         raid5_release_stripe(sh);
2595 }
2596
2597 static void raid5_end_write_request(struct bio *bi)
2598 {
2599         struct stripe_head *sh = bi->bi_private;
2600         struct r5conf *conf = sh->raid_conf;
2601         int disks = sh->disks, i;
2602         struct md_rdev *uninitialized_var(rdev);
2603         sector_t first_bad;
2604         int bad_sectors;
2605         int replacement = 0;
2606
2607         for (i = 0 ; i < disks; i++) {
2608                 if (bi == &sh->dev[i].req) {
2609                         rdev = conf->disks[i].rdev;
2610                         break;
2611                 }
2612                 if (bi == &sh->dev[i].rreq) {
2613                         rdev = conf->disks[i].replacement;
2614                         if (rdev)
2615                                 replacement = 1;
2616                         else
2617                                 /* rdev was removed and 'replacement'
2618                                  * replaced it.  rdev is not removed
2619                                  * until all requests are finished.
2620                                  */
2621                                 rdev = conf->disks[i].rdev;
2622                         break;
2623                 }
2624         }
2625         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2626                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2627                 bi->bi_status);
2628         if (i == disks) {
2629                 bio_reset(bi);
2630                 BUG();
2631                 return;
2632         }
2633
2634         if (replacement) {
2635                 if (bi->bi_status)
2636                         md_error(conf->mddev, rdev);
2637                 else if (is_badblock(rdev, sh->sector,
2638                                      STRIPE_SECTORS,
2639                                      &first_bad, &bad_sectors))
2640                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2641         } else {
2642                 if (bi->bi_status) {
2643                         set_bit(STRIPE_DEGRADED, &sh->state);
2644                         set_bit(WriteErrorSeen, &rdev->flags);
2645                         set_bit(R5_WriteError, &sh->dev[i].flags);
2646                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2647                                 set_bit(MD_RECOVERY_NEEDED,
2648                                         &rdev->mddev->recovery);
2649                 } else if (is_badblock(rdev, sh->sector,
2650                                        STRIPE_SECTORS,
2651                                        &first_bad, &bad_sectors)) {
2652                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2653                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2654                                 /* That was a successful write so make
2655                                  * sure it looks like we already did
2656                                  * a re-write.
2657                                  */
2658                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2659                 }
2660         }
2661         rdev_dec_pending(rdev, conf->mddev);
2662
2663         if (sh->batch_head && bi->bi_status && !replacement)
2664                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2665
2666         bio_reset(bi);
2667         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2668                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2669         set_bit(STRIPE_HANDLE, &sh->state);
2670         raid5_release_stripe(sh);
2671
2672         if (sh->batch_head && sh != sh->batch_head)
2673                 raid5_release_stripe(sh->batch_head);
2674 }
2675
2676 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2677 {
2678         char b[BDEVNAME_SIZE];
2679         struct r5conf *conf = mddev->private;
2680         unsigned long flags;
2681         pr_debug("raid456: error called\n");
2682
2683         spin_lock_irqsave(&conf->device_lock, flags);
2684
2685         if (test_bit(In_sync, &rdev->flags) &&
2686             mddev->degraded == conf->max_degraded) {
2687                 /*
2688                  * Don't allow to achieve failed state
2689                  * Don't try to recover this device
2690                  */
2691                 conf->recovery_disabled = mddev->recovery_disabled;
2692                 spin_unlock_irqrestore(&conf->device_lock, flags);
2693                 return;
2694         }
2695
2696         set_bit(Faulty, &rdev->flags);
2697         clear_bit(In_sync, &rdev->flags);
2698         mddev->degraded = raid5_calc_degraded(conf);
2699         spin_unlock_irqrestore(&conf->device_lock, flags);
2700         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2701
2702         set_bit(Blocked, &rdev->flags);
2703         set_mask_bits(&mddev->sb_flags, 0,
2704                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2705         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2706                 "md/raid:%s: Operation continuing on %d devices.\n",
2707                 mdname(mddev),
2708                 bdevname(rdev->bdev, b),
2709                 mdname(mddev),
2710                 conf->raid_disks - mddev->degraded);
2711         r5c_update_on_rdev_error(mddev, rdev);
2712 }
2713
2714 /*
2715  * Input: a 'big' sector number,
2716  * Output: index of the data and parity disk, and the sector # in them.
2717  */
2718 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2719                               int previous, int *dd_idx,
2720                               struct stripe_head *sh)
2721 {
2722         sector_t stripe, stripe2;
2723         sector_t chunk_number;
2724         unsigned int chunk_offset;
2725         int pd_idx, qd_idx;
2726         int ddf_layout = 0;
2727         sector_t new_sector;
2728         int algorithm = previous ? conf->prev_algo
2729                                  : conf->algorithm;
2730         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2731                                          : conf->chunk_sectors;
2732         int raid_disks = previous ? conf->previous_raid_disks
2733                                   : conf->raid_disks;
2734         int data_disks = raid_disks - conf->max_degraded;
2735
2736         /* First compute the information on this sector */
2737
2738         /*
2739          * Compute the chunk number and the sector offset inside the chunk
2740          */
2741         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2742         chunk_number = r_sector;
2743
2744         /*
2745          * Compute the stripe number
2746          */
2747         stripe = chunk_number;
2748         *dd_idx = sector_div(stripe, data_disks);
2749         stripe2 = stripe;
2750         /*
2751          * Select the parity disk based on the user selected algorithm.
2752          */
2753         pd_idx = qd_idx = -1;
2754         switch(conf->level) {
2755         case 4:
2756                 pd_idx = data_disks;
2757                 break;
2758         case 5:
2759                 switch (algorithm) {
2760                 case ALGORITHM_LEFT_ASYMMETRIC:
2761                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2762                         if (*dd_idx >= pd_idx)
2763                                 (*dd_idx)++;
2764                         break;
2765                 case ALGORITHM_RIGHT_ASYMMETRIC:
2766                         pd_idx = sector_div(stripe2, raid_disks);
2767                         if (*dd_idx >= pd_idx)
2768                                 (*dd_idx)++;
2769                         break;
2770                 case ALGORITHM_LEFT_SYMMETRIC:
2771                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2772                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2773                         break;
2774                 case ALGORITHM_RIGHT_SYMMETRIC:
2775                         pd_idx = sector_div(stripe2, raid_disks);
2776                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2777                         break;
2778                 case ALGORITHM_PARITY_0:
2779                         pd_idx = 0;
2780                         (*dd_idx)++;
2781                         break;
2782                 case ALGORITHM_PARITY_N:
2783                         pd_idx = data_disks;
2784                         break;
2785                 default:
2786                         BUG();
2787                 }
2788                 break;
2789         case 6:
2790
2791                 switch (algorithm) {
2792                 case ALGORITHM_LEFT_ASYMMETRIC:
2793                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2794                         qd_idx = pd_idx + 1;
2795                         if (pd_idx == raid_disks-1) {
2796                                 (*dd_idx)++;    /* Q D D D P */
2797                                 qd_idx = 0;
2798                         } else if (*dd_idx >= pd_idx)
2799                                 (*dd_idx) += 2; /* D D P Q D */
2800                         break;
2801                 case ALGORITHM_RIGHT_ASYMMETRIC:
2802                         pd_idx = sector_div(stripe2, raid_disks);
2803                         qd_idx = pd_idx + 1;
2804                         if (pd_idx == raid_disks-1) {
2805                                 (*dd_idx)++;    /* Q D D D P */
2806                                 qd_idx = 0;
2807                         } else if (*dd_idx >= pd_idx)
2808                                 (*dd_idx) += 2; /* D D P Q D */
2809                         break;
2810                 case ALGORITHM_LEFT_SYMMETRIC:
2811                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2812                         qd_idx = (pd_idx + 1) % raid_disks;
2813                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2814                         break;
2815                 case ALGORITHM_RIGHT_SYMMETRIC:
2816                         pd_idx = sector_div(stripe2, raid_disks);
2817                         qd_idx = (pd_idx + 1) % raid_disks;
2818                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2819                         break;
2820
2821                 case ALGORITHM_PARITY_0:
2822                         pd_idx = 0;
2823                         qd_idx = 1;
2824                         (*dd_idx) += 2;
2825                         break;
2826                 case ALGORITHM_PARITY_N:
2827                         pd_idx = data_disks;
2828                         qd_idx = data_disks + 1;
2829                         break;
2830
2831                 case ALGORITHM_ROTATING_ZERO_RESTART:
2832                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2833                          * of blocks for computing Q is different.
2834                          */
2835                         pd_idx = sector_div(stripe2, raid_disks);
2836                         qd_idx = pd_idx + 1;
2837                         if (pd_idx == raid_disks-1) {
2838                                 (*dd_idx)++;    /* Q D D D P */
2839                                 qd_idx = 0;
2840                         } else if (*dd_idx >= pd_idx)
2841                                 (*dd_idx) += 2; /* D D P Q D */
2842                         ddf_layout = 1;
2843                         break;
2844
2845                 case ALGORITHM_ROTATING_N_RESTART:
2846                         /* Same a left_asymmetric, by first stripe is
2847                          * D D D P Q  rather than
2848                          * Q D D D P
2849                          */
2850                         stripe2 += 1;
2851                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2852                         qd_idx = pd_idx + 1;
2853                         if (pd_idx == raid_disks-1) {
2854                                 (*dd_idx)++;    /* Q D D D P */
2855                                 qd_idx = 0;
2856                         } else if (*dd_idx >= pd_idx)
2857                                 (*dd_idx) += 2; /* D D P Q D */
2858                         ddf_layout = 1;
2859                         break;
2860
2861                 case ALGORITHM_ROTATING_N_CONTINUE:
2862                         /* Same as left_symmetric but Q is before P */
2863                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2864                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2865                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2866                         ddf_layout = 1;
2867                         break;
2868
2869                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2870                         /* RAID5 left_asymmetric, with Q on last device */
2871                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2872                         if (*dd_idx >= pd_idx)
2873                                 (*dd_idx)++;
2874                         qd_idx = raid_disks - 1;
2875                         break;
2876
2877                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2878                         pd_idx = sector_div(stripe2, raid_disks-1);
2879                         if (*dd_idx >= pd_idx)
2880                                 (*dd_idx)++;
2881                         qd_idx = raid_disks - 1;
2882                         break;
2883
2884                 case ALGORITHM_LEFT_SYMMETRIC_6:
2885                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2886                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2887                         qd_idx = raid_disks - 1;
2888                         break;
2889
2890                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2891                         pd_idx = sector_div(stripe2, raid_disks-1);
2892                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2893                         qd_idx = raid_disks - 1;
2894                         break;
2895
2896                 case ALGORITHM_PARITY_0_6:
2897                         pd_idx = 0;
2898                         (*dd_idx)++;
2899                         qd_idx = raid_disks - 1;
2900                         break;
2901
2902                 default:
2903                         BUG();
2904                 }
2905                 break;
2906         }
2907
2908         if (sh) {
2909                 sh->pd_idx = pd_idx;
2910                 sh->qd_idx = qd_idx;
2911                 sh->ddf_layout = ddf_layout;
2912         }
2913         /*
2914          * Finally, compute the new sector number
2915          */
2916         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2917         return new_sector;
2918 }
2919
2920 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2921 {
2922         struct r5conf *conf = sh->raid_conf;
2923         int raid_disks = sh->disks;
2924         int data_disks = raid_disks - conf->max_degraded;
2925         sector_t new_sector = sh->sector, check;
2926         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2927                                          : conf->chunk_sectors;
2928         int algorithm = previous ? conf->prev_algo
2929                                  : conf->algorithm;
2930         sector_t stripe;
2931         int chunk_offset;
2932         sector_t chunk_number;
2933         int dummy1, dd_idx = i;
2934         sector_t r_sector;
2935         struct stripe_head sh2;
2936
2937         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2938         stripe = new_sector;
2939
2940         if (i == sh->pd_idx)
2941                 return 0;
2942         switch(conf->level) {
2943         case 4: break;
2944         case 5:
2945                 switch (algorithm) {
2946                 case ALGORITHM_LEFT_ASYMMETRIC:
2947                 case ALGORITHM_RIGHT_ASYMMETRIC:
2948                         if (i > sh->pd_idx)
2949                                 i--;
2950                         break;
2951                 case ALGORITHM_LEFT_SYMMETRIC:
2952                 case ALGORITHM_RIGHT_SYMMETRIC:
2953                         if (i < sh->pd_idx)
2954                                 i += raid_disks;
2955                         i -= (sh->pd_idx + 1);
2956                         break;
2957                 case ALGORITHM_PARITY_0:
2958                         i -= 1;
2959                         break;
2960                 case ALGORITHM_PARITY_N:
2961                         break;
2962                 default:
2963                         BUG();
2964                 }
2965                 break;
2966         case 6:
2967                 if (i == sh->qd_idx)
2968                         return 0; /* It is the Q disk */
2969                 switch (algorithm) {
2970                 case ALGORITHM_LEFT_ASYMMETRIC:
2971                 case ALGORITHM_RIGHT_ASYMMETRIC:
2972                 case ALGORITHM_ROTATING_ZERO_RESTART:
2973                 case ALGORITHM_ROTATING_N_RESTART:
2974                         if (sh->pd_idx == raid_disks-1)
2975                                 i--;    /* Q D D D P */
2976                         else if (i > sh->pd_idx)
2977                                 i -= 2; /* D D P Q D */
2978                         break;
2979                 case ALGORITHM_LEFT_SYMMETRIC:
2980                 case ALGORITHM_RIGHT_SYMMETRIC:
2981                         if (sh->pd_idx == raid_disks-1)
2982                                 i--; /* Q D D D P */
2983                         else {
2984                                 /* D D P Q D */
2985                                 if (i < sh->pd_idx)
2986                                         i += raid_disks;
2987                                 i -= (sh->pd_idx + 2);
2988                         }
2989                         break;
2990                 case ALGORITHM_PARITY_0:
2991                         i -= 2;
2992                         break;
2993                 case ALGORITHM_PARITY_N:
2994                         break;
2995                 case ALGORITHM_ROTATING_N_CONTINUE:
2996                         /* Like left_symmetric, but P is before Q */
2997                         if (sh->pd_idx == 0)
2998                                 i--;    /* P D D D Q */
2999                         else {
3000                                 /* D D Q P D */
3001                                 if (i < sh->pd_idx)
3002                                         i += raid_disks;
3003                                 i -= (sh->pd_idx + 1);
3004                         }
3005                         break;
3006                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3007                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3008                         if (i > sh->pd_idx)
3009                                 i--;
3010                         break;
3011                 case ALGORITHM_LEFT_SYMMETRIC_6:
3012                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3013                         if (i < sh->pd_idx)
3014                                 i += data_disks + 1;
3015                         i -= (sh->pd_idx + 1);
3016                         break;
3017                 case ALGORITHM_PARITY_0_6:
3018                         i -= 1;
3019                         break;
3020                 default:
3021                         BUG();
3022                 }
3023                 break;
3024         }
3025
3026         chunk_number = stripe * data_disks + i;
3027         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3028
3029         check = raid5_compute_sector(conf, r_sector,
3030                                      previous, &dummy1, &sh2);
3031         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3032                 || sh2.qd_idx != sh->qd_idx) {
3033                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3034                         mdname(conf->mddev));
3035                 return 0;
3036         }
3037         return r_sector;
3038 }
3039
3040 /*
3041  * There are cases where we want handle_stripe_dirtying() and
3042  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3043  *
3044  * This function checks whether we want to delay the towrite. Specifically,
3045  * we delay the towrite when:
3046  *
3047  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3048  *      stripe has data in journal (for other devices).
3049  *
3050  *      In this case, when reading data for the non-overwrite dev, it is
3051  *      necessary to handle complex rmw of write back cache (prexor with
3052  *      orig_page, and xor with page). To keep read path simple, we would
3053  *      like to flush data in journal to RAID disks first, so complex rmw
3054  *      is handled in the write patch (handle_stripe_dirtying).
3055  *
3056  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3057  *
3058  *      It is important to be able to flush all stripes in raid5-cache.
3059  *      Therefore, we need reserve some space on the journal device for
3060  *      these flushes. If flush operation includes pending writes to the
3061  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3062  *      for the flush out. If we exclude these pending writes from flush
3063  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3064  *      Therefore, excluding pending writes in these cases enables more
3065  *      efficient use of the journal device.
3066  *
3067  *      Note: To make sure the stripe makes progress, we only delay
3068  *      towrite for stripes with data already in journal (injournal > 0).
3069  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3070  *      no_space_stripes list.
3071  *
3072  *   3. during journal failure
3073  *      In journal failure, we try to flush all cached data to raid disks
3074  *      based on data in stripe cache. The array is read-only to upper
3075  *      layers, so we would skip all pending writes.
3076  *
3077  */
3078 static inline bool delay_towrite(struct r5conf *conf,
3079                                  struct r5dev *dev,
3080                                  struct stripe_head_state *s)
3081 {
3082         /* case 1 above */
3083         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3084             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3085                 return true;
3086         /* case 2 above */
3087         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3088             s->injournal > 0)
3089                 return true;
3090         /* case 3 above */
3091         if (s->log_failed && s->injournal)
3092                 return true;
3093         return false;
3094 }
3095
3096 static void
3097 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3098                          int rcw, int expand)
3099 {
3100         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3101         struct r5conf *conf = sh->raid_conf;
3102         int level = conf->level;
3103
3104         if (rcw) {
3105                 /*
3106                  * In some cases, handle_stripe_dirtying initially decided to
3107                  * run rmw and allocates extra page for prexor. However, rcw is
3108                  * cheaper later on. We need to free the extra page now,
3109                  * because we won't be able to do that in ops_complete_prexor().
3110                  */
3111                 r5c_release_extra_page(sh);
3112
3113                 for (i = disks; i--; ) {
3114                         struct r5dev *dev = &sh->dev[i];
3115
3116                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3117                                 set_bit(R5_LOCKED, &dev->flags);
3118                                 set_bit(R5_Wantdrain, &dev->flags);
3119                                 if (!expand)
3120                                         clear_bit(R5_UPTODATE, &dev->flags);
3121                                 s->locked++;
3122                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3123                                 set_bit(R5_LOCKED, &dev->flags);
3124                                 s->locked++;
3125                         }
3126                 }
3127                 /* if we are not expanding this is a proper write request, and
3128                  * there will be bios with new data to be drained into the
3129                  * stripe cache
3130                  */
3131                 if (!expand) {
3132                         if (!s->locked)
3133                                 /* False alarm, nothing to do */
3134                                 return;
3135                         sh->reconstruct_state = reconstruct_state_drain_run;
3136                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3137                 } else
3138                         sh->reconstruct_state = reconstruct_state_run;
3139
3140                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3141
3142                 if (s->locked + conf->max_degraded == disks)
3143                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3144                                 atomic_inc(&conf->pending_full_writes);
3145         } else {
3146                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3147                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3148                 BUG_ON(level == 6 &&
3149                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3150                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3151
3152                 for (i = disks; i--; ) {
3153                         struct r5dev *dev = &sh->dev[i];
3154                         if (i == pd_idx || i == qd_idx)
3155                                 continue;
3156
3157                         if (dev->towrite &&
3158                             (test_bit(R5_UPTODATE, &dev->flags) ||
3159                              test_bit(R5_Wantcompute, &dev->flags))) {
3160                                 set_bit(R5_Wantdrain, &dev->flags);
3161                                 set_bit(R5_LOCKED, &dev->flags);
3162                                 clear_bit(R5_UPTODATE, &dev->flags);
3163                                 s->locked++;
3164                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3165                                 set_bit(R5_LOCKED, &dev->flags);
3166                                 s->locked++;
3167                         }
3168                 }
3169                 if (!s->locked)
3170                         /* False alarm - nothing to do */
3171                         return;
3172                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3173                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3174                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3175                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3176         }
3177
3178         /* keep the parity disk(s) locked while asynchronous operations
3179          * are in flight
3180          */
3181         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3182         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3183         s->locked++;
3184
3185         if (level == 6) {
3186                 int qd_idx = sh->qd_idx;
3187                 struct r5dev *dev = &sh->dev[qd_idx];
3188
3189                 set_bit(R5_LOCKED, &dev->flags);
3190                 clear_bit(R5_UPTODATE, &dev->flags);
3191                 s->locked++;
3192         }
3193
3194         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3195             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3196             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3197             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3198                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3199
3200         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3201                 __func__, (unsigned long long)sh->sector,
3202                 s->locked, s->ops_request);
3203 }
3204
3205 /*
3206  * Each stripe/dev can have one or more bion attached.
3207  * toread/towrite point to the first in a chain.
3208  * The bi_next chain must be in order.
3209  */
3210 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3211                           int forwrite, int previous)
3212 {
3213         struct bio **bip;
3214         struct r5conf *conf = sh->raid_conf;
3215         int firstwrite=0;
3216
3217         pr_debug("adding bi b#%llu to stripe s#%llu\n",
3218                 (unsigned long long)bi->bi_iter.bi_sector,
3219                 (unsigned long long)sh->sector);
3220
3221         spin_lock_irq(&sh->stripe_lock);
3222         sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3223         /* Don't allow new IO added to stripes in batch list */
3224         if (sh->batch_head)
3225                 goto overlap;
3226         if (forwrite) {
3227                 bip = &sh->dev[dd_idx].towrite;
3228                 if (*bip == NULL)
3229                         firstwrite = 1;
3230         } else
3231                 bip = &sh->dev[dd_idx].toread;
3232         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3233                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3234                         goto overlap;
3235                 bip = & (*bip)->bi_next;
3236         }
3237         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3238                 goto overlap;
3239
3240         if (forwrite && raid5_has_ppl(conf)) {
3241                 /*
3242                  * With PPL only writes to consecutive data chunks within a
3243                  * stripe are allowed because for a single stripe_head we can
3244                  * only have one PPL entry at a time, which describes one data
3245                  * range. Not really an overlap, but wait_for_overlap can be
3246                  * used to handle this.
3247                  */
3248                 sector_t sector;
3249                 sector_t first = 0;
3250                 sector_t last = 0;
3251                 int count = 0;
3252                 int i;
3253
3254                 for (i = 0; i < sh->disks; i++) {
3255                         if (i != sh->pd_idx &&
3256                             (i == dd_idx || sh->dev[i].towrite)) {
3257                                 sector = sh->dev[i].sector;
3258                                 if (count == 0 || sector < first)
3259                                         first = sector;
3260                                 if (sector > last)
3261                                         last = sector;
3262                                 count++;
3263                         }
3264                 }
3265
3266                 if (first + conf->chunk_sectors * (count - 1) != last)
3267                         goto overlap;
3268         }
3269
3270         if (!forwrite || previous)
3271                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3272
3273         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3274         if (*bip)
3275                 bi->bi_next = *bip;
3276         *bip = bi;
3277         bio_inc_remaining(bi);
3278         md_write_inc(conf->mddev, bi);
3279
3280         if (forwrite) {
3281                 /* check if page is covered */
3282                 sector_t sector = sh->dev[dd_idx].sector;
3283                 for (bi=sh->dev[dd_idx].towrite;
3284                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3285                              bi && bi->bi_iter.bi_sector <= sector;
3286                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3287                         if (bio_end_sector(bi) >= sector)
3288                                 sector = bio_end_sector(bi);
3289                 }
3290                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3291                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3292                                 sh->overwrite_disks++;
3293         }
3294
3295         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3296                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3297                 (unsigned long long)sh->sector, dd_idx);
3298
3299         if (conf->mddev->bitmap && firstwrite) {
3300                 /* Cannot hold spinlock over bitmap_startwrite,
3301                  * but must ensure this isn't added to a batch until
3302                  * we have added to the bitmap and set bm_seq.
3303                  * So set STRIPE_BITMAP_PENDING to prevent
3304                  * batching.
3305                  * If multiple add_stripe_bio() calls race here they
3306                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3307                  * to complete "bitmap_startwrite" gets to set
3308                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3309                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3310                  * any more.
3311                  */
3312                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3313                 spin_unlock_irq(&sh->stripe_lock);
3314                 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3315                                      STRIPE_SECTORS, 0);
3316                 spin_lock_irq(&sh->stripe_lock);
3317                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3318                 if (!sh->batch_head) {
3319                         sh->bm_seq = conf->seq_flush+1;
3320                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3321                 }
3322         }
3323         spin_unlock_irq(&sh->stripe_lock);
3324
3325         if (stripe_can_batch(sh))
3326                 stripe_add_to_batch_list(conf, sh);
3327         return 1;
3328
3329  overlap:
3330         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3331         spin_unlock_irq(&sh->stripe_lock);
3332         return 0;
3333 }
3334
3335 static void end_reshape(struct r5conf *conf);
3336
3337 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3338                             struct stripe_head *sh)
3339 {
3340         int sectors_per_chunk =
3341                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3342         int dd_idx;
3343         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3344         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3345
3346         raid5_compute_sector(conf,
3347                              stripe * (disks - conf->max_degraded)
3348                              *sectors_per_chunk + chunk_offset,
3349                              previous,
3350                              &dd_idx, sh);
3351 }
3352
3353 static void
3354 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3355                      struct stripe_head_state *s, int disks)
3356 {
3357         int i;
3358         BUG_ON(sh->batch_head);
3359         for (i = disks; i--; ) {
3360                 struct bio *bi;
3361                 int bitmap_end = 0;
3362
3363                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3364                         struct md_rdev *rdev;
3365                         rcu_read_lock();
3366                         rdev = rcu_dereference(conf->disks[i].rdev);
3367                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3368                             !test_bit(Faulty, &rdev->flags))
3369                                 atomic_inc(&rdev->nr_pending);
3370                         else
3371                                 rdev = NULL;
3372                         rcu_read_unlock();
3373                         if (rdev) {
3374                                 if (!rdev_set_badblocks(
3375                                             rdev,
3376                                             sh->sector,
3377                                             STRIPE_SECTORS, 0))
3378                                         md_error(conf->mddev, rdev);
3379                                 rdev_dec_pending(rdev, conf->mddev);
3380                         }
3381                 }
3382                 spin_lock_irq(&sh->stripe_lock);
3383                 /* fail all writes first */
3384                 bi = sh->dev[i].towrite;
3385                 sh->dev[i].towrite = NULL;
3386                 sh->overwrite_disks = 0;
3387                 spin_unlock_irq(&sh->stripe_lock);
3388                 if (bi)
3389                         bitmap_end = 1;
3390
3391                 log_stripe_write_finished(sh);
3392
3393                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3394                         wake_up(&conf->wait_for_overlap);
3395
3396                 while (bi && bi->bi_iter.bi_sector <
3397                         sh->dev[i].sector + STRIPE_SECTORS) {
3398                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3399
3400                         md_write_end(conf->mddev);
3401                         bio_io_error(bi);
3402                         bi = nextbi;
3403                 }
3404                 if (bitmap_end)
3405                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3406                                            STRIPE_SECTORS, 0, 0);
3407                 bitmap_end = 0;
3408                 /* and fail all 'written' */
3409                 bi = sh->dev[i].written;
3410                 sh->dev[i].written = NULL;
3411                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3412                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3413                         sh->dev[i].page = sh->dev[i].orig_page;
3414                 }
3415
3416                 if (bi) bitmap_end = 1;
3417                 while (bi && bi->bi_iter.bi_sector <
3418                        sh->dev[i].sector + STRIPE_SECTORS) {
3419                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3420
3421                         md_write_end(conf->mddev);
3422                         bio_io_error(bi);
3423                         bi = bi2;
3424                 }
3425
3426                 /* fail any reads if this device is non-operational and
3427                  * the data has not reached the cache yet.
3428                  */
3429                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3430                     s->failed > conf->max_degraded &&
3431                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3432                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3433                         spin_lock_irq(&sh->stripe_lock);
3434                         bi = sh->dev[i].toread;
3435                         sh->dev[i].toread = NULL;
3436                         spin_unlock_irq(&sh->stripe_lock);
3437                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3438                                 wake_up(&conf->wait_for_overlap);
3439                         if (bi)
3440                                 s->to_read--;
3441                         while (bi && bi->bi_iter.bi_sector <
3442                                sh->dev[i].sector + STRIPE_SECTORS) {
3443                                 struct bio *nextbi =
3444                                         r5_next_bio(bi, sh->dev[i].sector);
3445
3446                                 bio_io_error(bi);
3447                                 bi = nextbi;
3448                         }
3449                 }
3450                 if (bitmap_end)
3451                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3452                                            STRIPE_SECTORS, 0, 0);
3453                 /* If we were in the middle of a write the parity block might
3454                  * still be locked - so just clear all R5_LOCKED flags
3455                  */
3456                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3457         }
3458         s->to_write = 0;
3459         s->written = 0;
3460
3461         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3462                 if (atomic_dec_and_test(&conf->pending_full_writes))
3463                         md_wakeup_thread(conf->mddev->thread);
3464 }
3465
3466 static void
3467 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3468                    struct stripe_head_state *s)
3469 {
3470         int abort = 0;
3471         int i;
3472
3473         BUG_ON(sh->batch_head);
3474         clear_bit(STRIPE_SYNCING, &sh->state);
3475         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3476                 wake_up(&conf->wait_for_overlap);
3477         s->syncing = 0;
3478         s->replacing = 0;
3479         /* There is nothing more to do for sync/check/repair.
3480          * Don't even need to abort as that is handled elsewhere
3481          * if needed, and not always wanted e.g. if there is a known
3482          * bad block here.
3483          * For recover/replace we need to record a bad block on all
3484          * non-sync devices, or abort the recovery
3485          */
3486         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3487                 /* During recovery devices cannot be removed, so
3488                  * locking and refcounting of rdevs is not needed
3489                  */
3490                 rcu_read_lock();
3491                 for (i = 0; i < conf->raid_disks; i++) {
3492                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3493                         if (rdev
3494                             && !test_bit(Faulty, &rdev->flags)
3495                             && !test_bit(In_sync, &rdev->flags)
3496                             && !rdev_set_badblocks(rdev, sh->sector,
3497                                                    STRIPE_SECTORS, 0))
3498                                 abort = 1;
3499                         rdev = rcu_dereference(conf->disks[i].replacement);
3500                         if (rdev
3501                             && !test_bit(Faulty, &rdev->flags)
3502                             && !test_bit(In_sync, &rdev->flags)
3503                             && !rdev_set_badblocks(rdev, sh->sector,
3504                                                    STRIPE_SECTORS, 0))
3505                                 abort = 1;
3506                 }
3507                 rcu_read_unlock();
3508                 if (abort)
3509                         conf->recovery_disabled =
3510                                 conf->mddev->recovery_disabled;
3511         }
3512         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3513 }
3514
3515 static int want_replace(struct stripe_head *sh, int disk_idx)
3516 {
3517         struct md_rdev *rdev;
3518         int rv = 0;
3519
3520         rcu_read_lock();
3521         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3522         if (rdev
3523             && !test_bit(Faulty, &rdev->flags)
3524             && !test_bit(In_sync, &rdev->flags)
3525             && (rdev->recovery_offset <= sh->sector
3526                 || rdev->mddev->recovery_cp <= sh->sector))
3527                 rv = 1;
3528         rcu_read_unlock();
3529         return rv;
3530 }
3531
3532 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3533                            int disk_idx, int disks)
3534 {
3535         struct r5dev *dev = &sh->dev[disk_idx];
3536         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3537                                   &sh->dev[s->failed_num[1]] };
3538         int i;
3539
3540
3541         if (test_bit(R5_LOCKED, &dev->flags) ||
3542             test_bit(R5_UPTODATE, &dev->flags))
3543                 /* No point reading this as we already have it or have
3544                  * decided to get it.
3545                  */
3546                 return 0;
3547
3548         if (dev->toread ||
3549             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3550                 /* We need this block to directly satisfy a request */
3551                 return 1;
3552
3553         if (s->syncing || s->expanding ||
3554             (s->replacing && want_replace(sh, disk_idx)))
3555                 /* When syncing, or expanding we read everything.
3556                  * When replacing, we need the replaced block.
3557                  */
3558                 return 1;
3559
3560         if ((s->failed >= 1 && fdev[0]->toread) ||
3561             (s->failed >= 2 && fdev[1]->toread))
3562                 /* If we want to read from a failed device, then
3563                  * we need to actually read every other device.
3564                  */
3565                 return 1;
3566
3567         /* Sometimes neither read-modify-write nor reconstruct-write
3568          * cycles can work.  In those cases we read every block we
3569          * can.  Then the parity-update is certain to have enough to
3570          * work with.
3571          * This can only be a problem when we need to write something,
3572          * and some device has failed.  If either of those tests
3573          * fail we need look no further.
3574          */
3575         if (!s->failed || !s->to_write)
3576                 return 0;
3577
3578         if (test_bit(R5_Insync, &dev->flags) &&
3579             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3580                 /* Pre-reads at not permitted until after short delay
3581                  * to gather multiple requests.  However if this
3582                  * device is no Insync, the block could only be computed
3583                  * and there is no need to delay that.
3584                  */
3585                 return 0;
3586
3587         for (i = 0; i < s->failed && i < 2; i++) {
3588                 if (fdev[i]->towrite &&
3589                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3590                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3591                         /* If we have a partial write to a failed
3592                          * device, then we will need to reconstruct
3593                          * the content of that device, so all other
3594                          * devices must be read.
3595                          */
3596                         return 1;
3597         }
3598
3599         /* If we are forced to do a reconstruct-write, either because
3600          * the current RAID6 implementation only supports that, or
3601          * because parity cannot be trusted and we are currently
3602          * recovering it, there is extra need to be careful.
3603          * If one of the devices that we would need to read, because
3604          * it is not being overwritten (and maybe not written at all)
3605          * is missing/faulty, then we need to read everything we can.
3606          */
3607         if (sh->raid_conf->level != 6 &&
3608             sh->sector < sh->raid_conf->mddev->recovery_cp)
3609                 /* reconstruct-write isn't being forced */
3610                 return 0;
3611         for (i = 0; i < s->failed && i < 2; i++) {
3612                 if (s->failed_num[i] != sh->pd_idx &&
3613                     s->failed_num[i] != sh->qd_idx &&
3614                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3615                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3616                         return 1;
3617         }
3618
3619         return 0;
3620 }
3621
3622 /* fetch_block - checks the given member device to see if its data needs
3623  * to be read or computed to satisfy a request.
3624  *
3625  * Returns 1 when no more member devices need to be checked, otherwise returns
3626  * 0 to tell the loop in handle_stripe_fill to continue
3627  */
3628 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3629                        int disk_idx, int disks)
3630 {
3631         struct r5dev *dev = &sh->dev[disk_idx];
3632
3633         /* is the data in this block needed, and can we get it? */
3634         if (need_this_block(sh, s, disk_idx, disks)) {
3635                 /* we would like to get this block, possibly by computing it,
3636                  * otherwise read it if the backing disk is insync
3637                  */
3638                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3639                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3640                 BUG_ON(sh->batch_head);
3641
3642                 /*
3643                  * In the raid6 case if the only non-uptodate disk is P
3644                  * then we already trusted P to compute the other failed
3645                  * drives. It is safe to compute rather than re-read P.
3646                  * In other cases we only compute blocks from failed
3647                  * devices, otherwise check/repair might fail to detect
3648                  * a real inconsistency.
3649                  */
3650
3651                 if ((s->uptodate == disks - 1) &&
3652                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3653                     (s->failed && (disk_idx == s->failed_num[0] ||
3654                                    disk_idx == s->failed_num[1])))) {
3655                         /* have disk failed, and we're requested to fetch it;
3656                          * do compute it
3657                          */
3658                         pr_debug("Computing stripe %llu block %d\n",
3659                                (unsigned long long)sh->sector, disk_idx);
3660                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3661                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3662                         set_bit(R5_Wantcompute, &dev->flags);
3663                         sh->ops.target = disk_idx;
3664                         sh->ops.target2 = -1; /* no 2nd target */
3665                         s->req_compute = 1;
3666                         /* Careful: from this point on 'uptodate' is in the eye
3667                          * of raid_run_ops which services 'compute' operations
3668                          * before writes. R5_Wantcompute flags a block that will
3669                          * be R5_UPTODATE by the time it is needed for a
3670                          * subsequent operation.
3671                          */
3672                         s->uptodate++;
3673                         return 1;
3674                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3675                         /* Computing 2-failure is *very* expensive; only
3676                          * do it if failed >= 2
3677                          */
3678                         int other;
3679                         for (other = disks; other--; ) {
3680                                 if (other == disk_idx)
3681                                         continue;
3682                                 if (!test_bit(R5_UPTODATE,
3683                                       &sh->dev[other].flags))
3684                                         break;
3685                         }
3686                         BUG_ON(other < 0);
3687                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3688                                (unsigned long long)sh->sector,
3689                                disk_idx, other);
3690                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3691                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3692                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3693                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3694                         sh->ops.target = disk_idx;
3695                         sh->ops.target2 = other;
3696                         s->uptodate += 2;
3697                         s->req_compute = 1;
3698                         return 1;
3699                 } else if (test_bit(R5_Insync, &dev->flags)) {
3700                         set_bit(R5_LOCKED, &dev->flags);
3701                         set_bit(R5_Wantread, &dev->flags);
3702                         s->locked++;
3703                         pr_debug("Reading block %d (sync=%d)\n",
3704                                 disk_idx, s->syncing);
3705                 }
3706         }
3707
3708         return 0;
3709 }
3710
3711 /**
3712  * handle_stripe_fill - read or compute data to satisfy pending requests.
3713  */
3714 static void handle_stripe_fill(struct stripe_head *sh,
3715                                struct stripe_head_state *s,
3716                                int disks)
3717 {
3718         int i;
3719
3720         /* look for blocks to read/compute, skip this if a compute
3721          * is already in flight, or if the stripe contents are in the
3722          * midst of changing due to a write
3723          */
3724         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3725             !sh->reconstruct_state) {
3726
3727                 /*
3728                  * For degraded stripe with data in journal, do not handle
3729                  * read requests yet, instead, flush the stripe to raid
3730                  * disks first, this avoids handling complex rmw of write
3731                  * back cache (prexor with orig_page, and then xor with
3732                  * page) in the read path
3733                  */
3734                 if (s->injournal && s->failed) {
3735                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3736                                 r5c_make_stripe_write_out(sh);
3737                         goto out;
3738                 }
3739
3740                 for (i = disks; i--; )
3741                         if (fetch_block(sh, s, i, disks))
3742                                 break;
3743         }
3744 out:
3745         set_bit(STRIPE_HANDLE, &sh->state);
3746 }
3747
3748 static void break_stripe_batch_list(struct stripe_head *head_sh,
3749                                     unsigned long handle_flags);
3750 /* handle_stripe_clean_event
3751  * any written block on an uptodate or failed drive can be returned.
3752  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3753  * never LOCKED, so we don't need to test 'failed' directly.
3754  */
3755 static void handle_stripe_clean_event(struct r5conf *conf,
3756         struct stripe_head *sh, int disks)
3757 {
3758         int i;
3759         struct r5dev *dev;
3760         int discard_pending = 0;
3761         struct stripe_head *head_sh = sh;
3762         bool do_endio = false;
3763
3764         for (i = disks; i--; )
3765                 if (sh->dev[i].written) {
3766                         dev = &sh->dev[i];
3767                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3768                             (test_bit(R5_UPTODATE, &dev->flags) ||
3769                              test_bit(R5_Discard, &dev->flags) ||
3770                              test_bit(R5_SkipCopy, &dev->flags))) {
3771                                 /* We can return any write requests */
3772                                 struct bio *wbi, *wbi2;
3773                                 pr_debug("Return write for disc %d\n", i);
3774                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3775                                         clear_bit(R5_UPTODATE, &dev->flags);
3776                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3777                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3778                                 }
3779                                 do_endio = true;
3780
3781 returnbi:
3782                                 dev->page = dev->orig_page;
3783                                 wbi = dev->written;
3784                                 dev->written = NULL;
3785                                 while (wbi && wbi->bi_iter.bi_sector <
3786                                         dev->sector + STRIPE_SECTORS) {
3787                                         wbi2 = r5_next_bio(wbi, dev->sector);
3788                                         md_write_end(conf->mddev);
3789                                         bio_endio(wbi);
3790                                         wbi = wbi2;
3791                                 }
3792                                 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3793                                                    STRIPE_SECTORS,
3794                                                    !test_bit(STRIPE_DEGRADED, &sh->state),
3795                                                    0);
3796                                 if (head_sh->batch_head) {
3797                                         sh = list_first_entry(&sh->batch_list,
3798                                                               struct stripe_head,
3799                                                               batch_list);
3800                                         if (sh != head_sh) {
3801                                                 dev = &sh->dev[i];
3802                                                 goto returnbi;
3803                                         }
3804                                 }
3805                                 sh = head_sh;
3806                                 dev = &sh->dev[i];
3807                         } else if (test_bit(R5_Discard, &dev->flags))
3808                                 discard_pending = 1;
3809                 }
3810
3811         log_stripe_write_finished(sh);
3812
3813         if (!discard_pending &&
3814             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3815                 int hash;
3816                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3817                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3818                 if (sh->qd_idx >= 0) {
3819                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3820                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3821                 }
3822                 /* now that discard is done we can proceed with any sync */
3823                 clear_bit(STRIPE_DISCARD, &sh->state);
3824                 /*
3825                  * SCSI discard will change some bio fields and the stripe has
3826                  * no updated data, so remove it from hash list and the stripe
3827                  * will be reinitialized
3828                  */
3829 unhash:
3830                 hash = sh->hash_lock_index;
3831                 spin_lock_irq(conf->hash_locks + hash);
3832                 remove_hash(sh);
3833                 spin_unlock_irq(conf->hash_locks + hash);
3834                 if (head_sh->batch_head) {
3835                         sh = list_first_entry(&sh->batch_list,
3836                                               struct stripe_head, batch_list);
3837                         if (sh != head_sh)
3838                                         goto unhash;
3839                 }
3840                 sh = head_sh;
3841
3842                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3843                         set_bit(STRIPE_HANDLE, &sh->state);
3844
3845         }
3846
3847         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3848                 if (atomic_dec_and_test(&conf->pending_full_writes))
3849                         md_wakeup_thread(conf->mddev->thread);
3850
3851         if (head_sh->batch_head && do_endio)
3852                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3853 }
3854
3855 /*
3856  * For RMW in write back cache, we need extra page in prexor to store the
3857  * old data. This page is stored in dev->orig_page.
3858  *
3859  * This function checks whether we have data for prexor. The exact logic
3860  * is:
3861  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3862  */
3863 static inline bool uptodate_for_rmw(struct r5dev *dev)
3864 {
3865         return (test_bit(R5_UPTODATE, &dev->flags)) &&
3866                 (!test_bit(R5_InJournal, &dev->flags) ||
3867                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3868 }
3869
3870 static int handle_stripe_dirtying(struct r5conf *conf,
3871                                   struct stripe_head *sh,
3872                                   struct stripe_head_state *s,
3873                                   int disks)
3874 {
3875         int rmw = 0, rcw = 0, i;
3876         sector_t recovery_cp = conf->mddev->recovery_cp;
3877
3878         /* Check whether resync is now happening or should start.
3879          * If yes, then the array is dirty (after unclean shutdown or
3880          * initial creation), so parity in some stripes might be inconsistent.
3881          * In this case, we need to always do reconstruct-write, to ensure
3882          * that in case of drive failure or read-error correction, we
3883          * generate correct data from the parity.
3884          */
3885         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3886             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3887              s->failed == 0)) {
3888                 /* Calculate the real rcw later - for now make it
3889                  * look like rcw is cheaper
3890                  */
3891                 rcw = 1; rmw = 2;
3892                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3893                          conf->rmw_level, (unsigned long long)recovery_cp,
3894                          (unsigned long long)sh->sector);
3895         } else for (i = disks; i--; ) {
3896                 /* would I have to read this buffer for read_modify_write */
3897                 struct r5dev *dev = &sh->dev[i];
3898                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3899                      i == sh->pd_idx || i == sh->qd_idx ||
3900                      test_bit(R5_InJournal, &dev->flags)) &&
3901                     !test_bit(R5_LOCKED, &dev->flags) &&
3902                     !(uptodate_for_rmw(dev) ||
3903                       test_bit(R5_Wantcompute, &dev->flags))) {
3904                         if (test_bit(R5_Insync, &dev->flags))
3905                                 rmw++;
3906                         else
3907                                 rmw += 2*disks;  /* cannot read it */
3908                 }
3909                 /* Would I have to read this buffer for reconstruct_write */
3910                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3911                     i != sh->pd_idx && i != sh->qd_idx &&
3912                     !test_bit(R5_LOCKED, &dev->flags) &&
3913                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3914                       test_bit(R5_Wantcompute, &dev->flags))) {
3915                         if (test_bit(R5_Insync, &dev->flags))
3916                                 rcw++;
3917                         else
3918                                 rcw += 2*disks;
3919                 }
3920         }
3921
3922         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3923                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
3924         set_bit(STRIPE_HANDLE, &sh->state);
3925         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3926                 /* prefer read-modify-write, but need to get some data */
3927                 if (conf->mddev->queue)
3928                         blk_add_trace_msg(conf->mddev->queue,
3929                                           "raid5 rmw %llu %d",
3930                                           (unsigned long long)sh->sector, rmw);
3931                 for (i = disks; i--; ) {
3932                         struct r5dev *dev = &sh->dev[i];
3933                         if (test_bit(R5_InJournal, &dev->flags) &&
3934                             dev->page == dev->orig_page &&
3935                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3936                                 /* alloc page for prexor */
3937                                 struct page *p = alloc_page(GFP_NOIO);
3938
3939                                 if (p) {
3940                                         dev->orig_page = p;
3941                                         continue;
3942                                 }
3943
3944                                 /*
3945                                  * alloc_page() failed, try use
3946                                  * disk_info->extra_page
3947                                  */
3948                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3949                                                       &conf->cache_state)) {
3950                                         r5c_use_extra_page(sh);
3951                                         break;
3952                                 }
3953
3954                                 /* extra_page in use, add to delayed_list */
3955                                 set_bit(STRIPE_DELAYED, &sh->state);
3956                                 s->waiting_extra_page = 1;
3957                                 return -EAGAIN;
3958                         }
3959                 }
3960
3961                 for (i = disks; i--; ) {
3962                         struct r5dev *dev = &sh->dev[i];
3963                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3964                              i == sh->pd_idx || i == sh->qd_idx ||
3965                              test_bit(R5_InJournal, &dev->flags)) &&
3966                             !test_bit(R5_LOCKED, &dev->flags) &&
3967                             !(uptodate_for_rmw(dev) ||
3968                               test_bit(R5_Wantcompute, &dev->flags)) &&
3969                             test_bit(R5_Insync, &dev->flags)) {
3970                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3971                                              &sh->state)) {
3972                                         pr_debug("Read_old block %d for r-m-w\n",
3973                                                  i);
3974                                         set_bit(R5_LOCKED, &dev->flags);
3975                                         set_bit(R5_Wantread, &dev->flags);
3976                                         s->locked++;
3977                                 } else {
3978                                         set_bit(STRIPE_DELAYED, &sh->state);
3979                                         set_bit(STRIPE_HANDLE, &sh->state);
3980                                 }
3981                         }
3982                 }
3983         }
3984         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3985                 /* want reconstruct write, but need to get some data */
3986                 int qread =0;
3987                 rcw = 0;
3988                 for (i = disks; i--; ) {
3989                         struct r5dev *dev = &sh->dev[i];
3990                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3991                             i != sh->pd_idx && i != sh->qd_idx &&
3992                             !test_bit(R5_LOCKED, &dev->flags) &&
3993                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3994                               test_bit(R5_Wantcompute, &dev->flags))) {
3995                                 rcw++;
3996                                 if (test_bit(R5_Insync, &dev->flags) &&
3997                                     test_bit(STRIPE_PREREAD_ACTIVE,
3998                                              &sh->state)) {
3999                                         pr_debug("Read_old block "
4000                                                 "%d for Reconstruct\n", i);
4001                                         set_bit(R5_LOCKED, &dev->flags);
4002                                         set_bit(R5_Wantread, &dev->flags);
4003                                         s->locked++;
4004                                         qread++;
4005                                 } else {
4006                                         set_bit(STRIPE_DELAYED, &sh->state);
4007                                         set_bit(STRIPE_HANDLE, &sh->state);
4008                                 }
4009                         }
4010                 }
4011                 if (rcw && conf->mddev->queue)
4012                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4013                                           (unsigned long long)sh->sector,
4014                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4015         }
4016
4017         if (rcw > disks && rmw > disks &&
4018             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4019                 set_bit(STRIPE_DELAYED, &sh->state);
4020
4021         /* now if nothing is locked, and if we have enough data,
4022          * we can start a write request
4023          */
4024         /* since handle_stripe can be called at any time we need to handle the
4025          * case where a compute block operation has been submitted and then a
4026          * subsequent call wants to start a write request.  raid_run_ops only
4027          * handles the case where compute block and reconstruct are requested
4028          * simultaneously.  If this is not the case then new writes need to be
4029          * held off until the compute completes.
4030          */
4031         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4032             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4033              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4034                 schedule_reconstruction(sh, s, rcw == 0, 0);
4035         return 0;
4036 }
4037
4038 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4039                                 struct stripe_head_state *s, int disks)
4040 {
4041         struct r5dev *dev = NULL;
4042
4043         BUG_ON(sh->batch_head);
4044         set_bit(STRIPE_HANDLE, &sh->state);
4045
4046         switch (sh->check_state) {
4047         case check_state_idle:
4048                 /* start a new check operation if there are no failures */
4049                 if (s->failed == 0) {
4050                         BUG_ON(s->uptodate != disks);
4051                         sh->check_state = check_state_run;
4052                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4053                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4054                         s->uptodate--;
4055                         break;
4056                 }
4057                 dev = &sh->dev[s->failed_num[0]];
4058                 /* fall through */
4059         case check_state_compute_result:
4060                 sh->check_state = check_state_idle;
4061                 if (!dev)
4062                         dev = &sh->dev[sh->pd_idx];
4063
4064                 /* check that a write has not made the stripe insync */
4065                 if (test_bit(STRIPE_INSYNC, &sh->state))
4066                         break;
4067
4068                 /* either failed parity check, or recovery is happening */
4069                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4070                 BUG_ON(s->uptodate != disks);
4071
4072                 set_bit(R5_LOCKED, &dev->flags);
4073                 s->locked++;
4074                 set_bit(R5_Wantwrite, &dev->flags);
4075
4076                 clear_bit(STRIPE_DEGRADED, &sh->state);
4077                 set_bit(STRIPE_INSYNC, &sh->state);
4078                 break;
4079         case check_state_run:
4080                 break; /* we will be called again upon completion */
4081         case check_state_check_result:
4082                 sh->check_state = check_state_idle;
4083
4084                 /* if a failure occurred during the check operation, leave
4085                  * STRIPE_INSYNC not set and let the stripe be handled again
4086                  */
4087                 if (s->failed)
4088                         break;
4089
4090                 /* handle a successful check operation, if parity is correct
4091                  * we are done.  Otherwise update the mismatch count and repair
4092                  * parity if !MD_RECOVERY_CHECK
4093                  */
4094                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4095                         /* parity is correct (on disc,
4096                          * not in buffer any more)
4097                          */
4098                         set_bit(STRIPE_INSYNC, &sh->state);
4099                 else {
4100                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4101                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4102                                 /* don't try to repair!! */
4103                                 set_bit(STRIPE_INSYNC, &sh->state);
4104                                 pr_warn_ratelimited("%s: mismatch sector in range "
4105                                                     "%llu-%llu\n", mdname(conf->mddev),
4106                                                     (unsigned long long) sh->sector,
4107                                                     (unsigned long long) sh->sector +
4108                                                     STRIPE_SECTORS);
4109                         } else {
4110                                 sh->check_state = check_state_compute_run;
4111                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4112                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4113                                 set_bit(R5_Wantcompute,
4114                                         &sh->dev[sh->pd_idx].flags);
4115                                 sh->ops.target = sh->pd_idx;
4116                                 sh->ops.target2 = -1;
4117                                 s->uptodate++;
4118                         }
4119                 }
4120                 break;
4121         case check_state_compute_run:
4122                 break;
4123         default:
4124                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4125                        __func__, sh->check_state,
4126                        (unsigned long long) sh->sector);
4127                 BUG();
4128         }
4129 }
4130
4131 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4132                                   struct stripe_head_state *s,
4133                                   int disks)
4134 {
4135         int pd_idx = sh->pd_idx;
4136         int qd_idx = sh->qd_idx;
4137         struct r5dev *dev;
4138
4139         BUG_ON(sh->batch_head);
4140         set_bit(STRIPE_HANDLE, &sh->state);
4141
4142         BUG_ON(s->failed > 2);
4143
4144         /* Want to check and possibly repair P and Q.
4145          * However there could be one 'failed' device, in which
4146          * case we can only check one of them, possibly using the
4147          * other to generate missing data
4148          */
4149
4150         switch (sh->check_state) {
4151         case check_state_idle:
4152                 /* start a new check operation if there are < 2 failures */
4153                 if (s->failed == s->q_failed) {
4154                         /* The only possible failed device holds Q, so it
4155                          * makes sense to check P (If anything else were failed,
4156                          * we would have used P to recreate it).
4157                          */
4158                         sh->check_state = check_state_run;
4159                 }
4160                 if (!s->q_failed && s->failed < 2) {
4161                         /* Q is not failed, and we didn't use it to generate
4162                          * anything, so it makes sense to check it
4163                          */
4164                         if (sh->check_state == check_state_run)
4165                                 sh->check_state = check_state_run_pq;
4166                         else
4167                                 sh->check_state = check_state_run_q;
4168                 }
4169
4170                 /* discard potentially stale zero_sum_result */
4171                 sh->ops.zero_sum_result = 0;
4172
4173                 if (sh->check_state == check_state_run) {
4174                         /* async_xor_zero_sum destroys the contents of P */
4175                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4176                         s->uptodate--;
4177                 }
4178                 if (sh->check_state >= check_state_run &&
4179                     sh->check_state <= check_state_run_pq) {
4180                         /* async_syndrome_zero_sum preserves P and Q, so
4181                          * no need to mark them !uptodate here
4182                          */
4183                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4184                         break;
4185                 }
4186
4187                 /* we have 2-disk failure */
4188                 BUG_ON(s->failed != 2);
4189                 /* fall through */
4190         case check_state_compute_result:
4191                 sh->check_state = check_state_idle;
4192
4193                 /* check that a write has not made the stripe insync */
4194                 if (test_bit(STRIPE_INSYNC, &sh->state))
4195                         break;
4196
4197                 /* now write out any block on a failed drive,
4198                  * or P or Q if they were recomputed
4199                  */
4200                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
4201                 if (s->failed == 2) {
4202                         dev = &sh->dev[s->failed_num[1]];
4203                         s->locked++;
4204                         set_bit(R5_LOCKED, &dev->flags);
4205                         set_bit(R5_Wantwrite, &dev->flags);
4206                 }
4207                 if (s->failed >= 1) {
4208                         dev = &sh->dev[s->failed_num[0]];
4209                         s->locked++;
4210                         set_bit(R5_LOCKED, &dev->flags);
4211                         set_bit(R5_Wantwrite, &dev->flags);
4212                 }
4213                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4214                         dev = &sh->dev[pd_idx];
4215                         s->locked++;
4216                         set_bit(R5_LOCKED, &dev->flags);
4217                         set_bit(R5_Wantwrite, &dev->flags);
4218                 }
4219                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4220                         dev = &sh->dev[qd_idx];
4221                         s->locked++;
4222                         set_bit(R5_LOCKED, &dev->flags);
4223                         set_bit(R5_Wantwrite, &dev->flags);
4224                 }
4225                 clear_bit(STRIPE_DEGRADED, &sh->state);
4226
4227                 set_bit(STRIPE_INSYNC, &sh->state);
4228                 break;
4229         case check_state_run:
4230         case check_state_run_q:
4231         case check_state_run_pq:
4232                 break; /* we will be called again upon completion */
4233         case check_state_check_result:
4234                 sh->check_state = check_state_idle;
4235
4236                 /* handle a successful check operation, if parity is correct
4237                  * we are done.  Otherwise update the mismatch count and repair
4238                  * parity if !MD_RECOVERY_CHECK
4239                  */
4240                 if (sh->ops.zero_sum_result == 0) {
4241                         /* both parities are correct */
4242                         if (!s->failed)
4243                                 set_bit(STRIPE_INSYNC, &sh->state);
4244                         else {
4245                                 /* in contrast to the raid5 case we can validate
4246                                  * parity, but still have a failure to write
4247                                  * back
4248                                  */
4249                                 sh->check_state = check_state_compute_result;
4250                                 /* Returning at this point means that we may go
4251                                  * off and bring p and/or q uptodate again so
4252                                  * we make sure to check zero_sum_result again
4253                                  * to verify if p or q need writeback
4254                                  */
4255                         }
4256                 } else {
4257                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4258                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4259                                 /* don't try to repair!! */
4260                                 set_bit(STRIPE_INSYNC, &sh->state);
4261                                 pr_warn_ratelimited("%s: mismatch sector in range "
4262                                                     "%llu-%llu\n", mdname(conf->mddev),
4263                                                     (unsigned long long) sh->sector,
4264                                                     (unsigned long long) sh->sector +
4265                                                     STRIPE_SECTORS);
4266                         } else {
4267                                 int *target = &sh->ops.target;
4268
4269                                 sh->ops.target = -1;
4270                                 sh->ops.target2 = -1;
4271                                 sh->check_state = check_state_compute_run;
4272                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4273                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4274                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4275                                         set_bit(R5_Wantcompute,
4276                                                 &sh->dev[pd_idx].flags);
4277                                         *target = pd_idx;
4278                                         target = &sh->ops.target2;
4279                                         s->uptodate++;
4280                                 }
4281                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4282                                         set_bit(R5_Wantcompute,
4283                                                 &sh->dev[qd_idx].flags);
4284                                         *target = qd_idx;
4285                                         s->uptodate++;
4286                                 }
4287                         }
4288                 }
4289                 break;
4290         case check_state_compute_run:
4291                 break;
4292         default:
4293                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4294                         __func__, sh->check_state,
4295                         (unsigned long long) sh->sector);
4296                 BUG();
4297         }
4298 }
4299
4300 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4301 {
4302         int i;
4303
4304         /* We have read all the blocks in this stripe and now we need to
4305          * copy some of them into a target stripe for expand.
4306          */
4307         struct dma_async_tx_descriptor *tx = NULL;
4308         BUG_ON(sh->batch_head);
4309         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4310         for (i = 0; i < sh->disks; i++)
4311                 if (i != sh->pd_idx && i != sh->qd_idx) {
4312                         int dd_idx, j;
4313                         struct stripe_head *sh2;
4314                         struct async_submit_ctl submit;
4315
4316                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4317                         sector_t s = raid5_compute_sector(conf, bn, 0,
4318                                                           &dd_idx, NULL);
4319                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4320                         if (sh2 == NULL)
4321                                 /* so far only the early blocks of this stripe
4322                                  * have been requested.  When later blocks
4323                                  * get requested, we will try again
4324                                  */
4325                                 continue;
4326                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4327                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4328                                 /* must have already done this block */
4329                                 raid5_release_stripe(sh2);
4330                                 continue;
4331                         }
4332
4333                         /* place all the copies on one channel */
4334                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4335                         tx = async_memcpy(sh2->dev[dd_idx].page,
4336                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
4337                                           &submit);
4338
4339                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4340                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4341                         for (j = 0; j < conf->raid_disks; j++)
4342                                 if (j != sh2->pd_idx &&
4343                                     j != sh2->qd_idx &&
4344                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4345                                         break;
4346                         if (j == conf->raid_disks) {
4347                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4348                                 set_bit(STRIPE_HANDLE, &sh2->state);
4349                         }
4350                         raid5_release_stripe(sh2);
4351
4352                 }
4353         /* done submitting copies, wait for them to complete */
4354         async_tx_quiesce(&tx);
4355 }
4356
4357 /*
4358  * handle_stripe - do things to a stripe.
4359  *
4360  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4361  * state of various bits to see what needs to be done.
4362  * Possible results:
4363  *    return some read requests which now have data
4364  *    return some write requests which are safely on storage
4365  *    schedule a read on some buffers
4366  *    schedule a write of some buffers
4367  *    return confirmation of parity correctness
4368  *
4369  */
4370
4371 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4372 {
4373         struct r5conf *conf = sh->raid_conf;
4374         int disks = sh->disks;
4375         struct r5dev *dev;
4376         int i;
4377         int do_recovery = 0;
4378
4379         memset(s, 0, sizeof(*s));
4380
4381         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4382         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4383         s->failed_num[0] = -1;
4384         s->failed_num[1] = -1;
4385         s->log_failed = r5l_log_disk_error(conf);
4386
4387         /* Now to look around and see what can be done */
4388         rcu_read_lock();
4389         for (i=disks; i--; ) {
4390                 struct md_rdev *rdev;
4391                 sector_t first_bad;
4392                 int bad_sectors;
4393                 int is_bad = 0;
4394
4395                 dev = &sh->dev[i];
4396
4397                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4398                          i, dev->flags,
4399                          dev->toread, dev->towrite, dev->written);
4400                 /* maybe we can reply to a read
4401                  *
4402                  * new wantfill requests are only permitted while
4403                  * ops_complete_biofill is guaranteed to be inactive
4404                  */
4405                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4406                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4407                         set_bit(R5_Wantfill, &dev->flags);
4408
4409                 /* now count some things */
4410                 if (test_bit(R5_LOCKED, &dev->flags))
4411                         s->locked++;
4412                 if (test_bit(R5_UPTODATE, &dev->flags))
4413                         s->uptodate++;
4414                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4415                         s->compute++;
4416                         BUG_ON(s->compute > 2);
4417                 }
4418
4419                 if (test_bit(R5_Wantfill, &dev->flags))
4420                         s->to_fill++;
4421                 else if (dev->toread)
4422                         s->to_read++;
4423                 if (dev->towrite) {
4424                         s->to_write++;
4425                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4426                                 s->non_overwrite++;
4427                 }
4428                 if (dev->written)
4429                         s->written++;
4430                 /* Prefer to use the replacement for reads, but only
4431                  * if it is recovered enough and has no bad blocks.
4432                  */
4433                 rdev = rcu_dereference(conf->disks[i].replacement);
4434                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4435                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4436                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4437                                  &first_bad, &bad_sectors))
4438                         set_bit(R5_ReadRepl, &dev->flags);
4439                 else {
4440                         if (rdev && !test_bit(Faulty, &rdev->flags))
4441                                 set_bit(R5_NeedReplace, &dev->flags);
4442                         else
4443                                 clear_bit(R5_NeedReplace, &dev->flags);
4444                         rdev = rcu_dereference(conf->disks[i].rdev);
4445                         clear_bit(R5_ReadRepl, &dev->flags);
4446                 }
4447                 if (rdev && test_bit(Faulty, &rdev->flags))
4448                         rdev = NULL;
4449                 if (rdev) {
4450                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4451                                              &first_bad, &bad_sectors);
4452                         if (s->blocked_rdev == NULL
4453                             && (test_bit(Blocked, &rdev->flags)
4454                                 || is_bad < 0)) {
4455                                 if (is_bad < 0)
4456                                         set_bit(BlockedBadBlocks,
4457                                                 &rdev->flags);
4458                                 s->blocked_rdev = rdev;
4459                                 atomic_inc(&rdev->nr_pending);
4460                         }
4461                 }
4462                 clear_bit(R5_Insync, &dev->flags);
4463                 if (!rdev)
4464                         /* Not in-sync */;
4465                 else if (is_bad) {
4466                         /* also not in-sync */
4467                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4468                             test_bit(R5_UPTODATE, &dev->flags)) {
4469                                 /* treat as in-sync, but with a read error
4470                                  * which we can now try to correct
4471                                  */
4472                                 set_bit(R5_Insync, &dev->flags);
4473                                 set_bit(R5_ReadError, &dev->flags);
4474                         }
4475                 } else if (test_bit(In_sync, &rdev->flags))
4476                         set_bit(R5_Insync, &dev->flags);
4477                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4478                         /* in sync if before recovery_offset */
4479                         set_bit(R5_Insync, &dev->flags);
4480                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4481                          test_bit(R5_Expanded, &dev->flags))
4482                         /* If we've reshaped into here, we assume it is Insync.
4483                          * We will shortly update recovery_offset to make
4484                          * it official.
4485                          */
4486                         set_bit(R5_Insync, &dev->flags);
4487
4488                 if (test_bit(R5_WriteError, &dev->flags)) {
4489                         /* This flag does not apply to '.replacement'
4490                          * only to .rdev, so make sure to check that*/
4491                         struct md_rdev *rdev2 = rcu_dereference(
4492                                 conf->disks[i].rdev);
4493                         if (rdev2 == rdev)
4494                                 clear_bit(R5_Insync, &dev->flags);
4495                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4496                                 s->handle_bad_blocks = 1;
4497                                 atomic_inc(&rdev2->nr_pending);
4498                         } else
4499                                 clear_bit(R5_WriteError, &dev->flags);
4500                 }
4501                 if (test_bit(R5_MadeGood, &dev->flags)) {
4502                         /* This flag does not apply to '.replacement'
4503                          * only to .rdev, so make sure to check that*/
4504                         struct md_rdev *rdev2 = rcu_dereference(
4505                                 conf->disks[i].rdev);
4506                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4507                                 s->handle_bad_blocks = 1;
4508                                 atomic_inc(&rdev2->nr_pending);
4509                         } else
4510                                 clear_bit(R5_MadeGood, &dev->flags);
4511                 }
4512                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4513                         struct md_rdev *rdev2 = rcu_dereference(
4514                                 conf->disks[i].replacement);
4515                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4516                                 s->handle_bad_blocks = 1;
4517                                 atomic_inc(&rdev2->nr_pending);
4518                         } else
4519                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4520                 }
4521                 if (!test_bit(R5_Insync, &dev->flags)) {
4522                         /* The ReadError flag will just be confusing now */
4523                         clear_bit(R5_ReadError, &dev->flags);
4524                         clear_bit(R5_ReWrite, &dev->flags);
4525                 }
4526                 if (test_bit(R5_ReadError, &dev->flags))
4527                         clear_bit(R5_Insync, &dev->flags);
4528                 if (!test_bit(R5_Insync, &dev->flags)) {
4529                         if (s->failed < 2)
4530                                 s->failed_num[s->failed] = i;
4531                         s->failed++;
4532                         if (rdev && !test_bit(Faulty, &rdev->flags))
4533                                 do_recovery = 1;
4534                         else if (!rdev) {
4535                                 rdev = rcu_dereference(
4536                                     conf->disks[i].replacement);
4537                                 if (rdev && !test_bit(Faulty, &rdev->flags))
4538                                         do_recovery = 1;
4539                         }
4540                 }
4541
4542                 if (test_bit(R5_InJournal, &dev->flags))
4543                         s->injournal++;
4544                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4545                         s->just_cached++;
4546         }
4547         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4548                 /* If there is a failed device being replaced,
4549                  *     we must be recovering.
4550                  * else if we are after recovery_cp, we must be syncing
4551                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4552                  * else we can only be replacing
4553                  * sync and recovery both need to read all devices, and so
4554                  * use the same flag.
4555                  */
4556                 if (do_recovery ||
4557                     sh->sector >= conf->mddev->recovery_cp ||
4558                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4559                         s->syncing = 1;
4560                 else
4561                         s->replacing = 1;
4562         }
4563         rcu_read_unlock();
4564 }
4565
4566 static int clear_batch_ready(struct stripe_head *sh)
4567 {
4568         /* Return '1' if this is a member of batch, or
4569          * '0' if it is a lone stripe or a head which can now be
4570          * handled.
4571          */
4572         struct stripe_head *tmp;
4573         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4574                 return (sh->batch_head && sh->batch_head != sh);
4575         spin_lock(&sh->stripe_lock);
4576         if (!sh->batch_head) {
4577                 spin_unlock(&sh->stripe_lock);
4578                 return 0;
4579         }
4580
4581         /*
4582          * this stripe could be added to a batch list before we check
4583          * BATCH_READY, skips it
4584          */
4585         if (sh->batch_head != sh) {
4586                 spin_unlock(&sh->stripe_lock);
4587                 return 1;
4588         }
4589         spin_lock(&sh->batch_lock);
4590         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4591                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4592         spin_unlock(&sh->batch_lock);
4593         spin_unlock(&sh->stripe_lock);
4594
4595         /*
4596          * BATCH_READY is cleared, no new stripes can be added.
4597          * batch_list can be accessed without lock
4598          */
4599         return 0;
4600 }
4601
4602 static void break_stripe_batch_list(struct stripe_head *head_sh,
4603                                     unsigned long handle_flags)
4604 {
4605         struct stripe_head *sh, *next;
4606         int i;
4607         int do_wakeup = 0;
4608
4609         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4610
4611                 list_del_init(&sh->batch_list);
4612
4613                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4614                                           (1 << STRIPE_SYNCING) |
4615                                           (1 << STRIPE_REPLACED) |
4616                                           (1 << STRIPE_DELAYED) |
4617                                           (1 << STRIPE_BIT_DELAY) |
4618                                           (1 << STRIPE_FULL_WRITE) |
4619                                           (1 << STRIPE_BIOFILL_RUN) |
4620                                           (1 << STRIPE_COMPUTE_RUN)  |
4621                                           (1 << STRIPE_OPS_REQ_PENDING) |
4622                                           (1 << STRIPE_DISCARD) |
4623                                           (1 << STRIPE_BATCH_READY) |
4624                                           (1 << STRIPE_BATCH_ERR) |
4625                                           (1 << STRIPE_BITMAP_PENDING)),
4626                         "stripe state: %lx\n", sh->state);
4627                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4628                                               (1 << STRIPE_REPLACED)),
4629                         "head stripe state: %lx\n", head_sh->state);
4630
4631                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4632                                             (1 << STRIPE_PREREAD_ACTIVE) |
4633                                             (1 << STRIPE_DEGRADED) |
4634                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4635                               head_sh->state & (1 << STRIPE_INSYNC));
4636
4637                 sh->check_state = head_sh->check_state;
4638                 sh->reconstruct_state = head_sh->reconstruct_state;
4639                 spin_lock_irq(&sh->stripe_lock);
4640                 sh->batch_head = NULL;
4641                 spin_unlock_irq(&sh->stripe_lock);
4642                 for (i = 0; i < sh->disks; i++) {
4643                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4644                                 do_wakeup = 1;
4645                         sh->dev[i].flags = head_sh->dev[i].flags &
4646                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4647                 }
4648                 if (handle_flags == 0 ||
4649                     sh->state & handle_flags)
4650                         set_bit(STRIPE_HANDLE, &sh->state);
4651                 raid5_release_stripe(sh);
4652         }
4653         spin_lock_irq(&head_sh->stripe_lock);
4654         head_sh->batch_head = NULL;
4655         spin_unlock_irq(&head_sh->stripe_lock);
4656         for (i = 0; i < head_sh->disks; i++)
4657                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4658                         do_wakeup = 1;
4659         if (head_sh->state & handle_flags)
4660                 set_bit(STRIPE_HANDLE, &head_sh->state);
4661
4662         if (do_wakeup)
4663                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4664 }
4665
4666 static void handle_stripe(struct stripe_head *sh)
4667 {
4668         struct stripe_head_state s;
4669         struct r5conf *conf = sh->raid_conf;
4670         int i;
4671         int prexor;
4672         int disks = sh->disks;
4673         struct r5dev *pdev, *qdev;
4674
4675         clear_bit(STRIPE_HANDLE, &sh->state);
4676         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4677                 /* already being handled, ensure it gets handled
4678                  * again when current action finishes */
4679                 set_bit(STRIPE_HANDLE, &sh->state);
4680                 return;
4681         }
4682
4683         if (clear_batch_ready(sh) ) {
4684                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4685                 return;
4686         }
4687
4688         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4689                 break_stripe_batch_list(sh, 0);
4690
4691         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4692                 spin_lock(&sh->stripe_lock);
4693                 /*
4694                  * Cannot process 'sync' concurrently with 'discard'.
4695                  * Flush data in r5cache before 'sync'.
4696                  */
4697                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4698                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4699                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4700                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4701                         set_bit(STRIPE_SYNCING, &sh->state);
4702                         clear_bit(STRIPE_INSYNC, &sh->state);
4703                         clear_bit(STRIPE_REPLACED, &sh->state);
4704                 }
4705                 spin_unlock(&sh->stripe_lock);
4706         }
4707         clear_bit(STRIPE_DELAYED, &sh->state);
4708
4709         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4710                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4711                (unsigned long long)sh->sector, sh->state,
4712                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4713                sh->check_state, sh->reconstruct_state);
4714
4715         analyse_stripe(sh, &s);
4716
4717         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4718                 goto finish;
4719
4720         if (s.handle_bad_blocks ||
4721             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4722                 set_bit(STRIPE_HANDLE, &sh->state);
4723                 goto finish;
4724         }
4725
4726         if (unlikely(s.blocked_rdev)) {
4727                 if (s.syncing || s.expanding || s.expanded ||
4728                     s.replacing || s.to_write || s.written) {
4729                         set_bit(STRIPE_HANDLE, &sh->state);
4730                         goto finish;
4731                 }
4732                 /* There is nothing for the blocked_rdev to block */
4733                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4734                 s.blocked_rdev = NULL;
4735         }
4736
4737         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4738                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4739                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4740         }
4741
4742         pr_debug("locked=%d uptodate=%d to_read=%d"
4743                " to_write=%d failed=%d failed_num=%d,%d\n",
4744                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4745                s.failed_num[0], s.failed_num[1]);
4746         /*
4747          * check if the array has lost more than max_degraded devices and,
4748          * if so, some requests might need to be failed.
4749          *
4750          * When journal device failed (log_failed), we will only process
4751          * the stripe if there is data need write to raid disks
4752          */
4753         if (s.failed > conf->max_degraded ||
4754             (s.log_failed && s.injournal == 0)) {
4755                 sh->check_state = 0;
4756                 sh->reconstruct_state = 0;
4757                 break_stripe_batch_list(sh, 0);
4758                 if (s.to_read+s.to_write+s.written)
4759                         handle_failed_stripe(conf, sh, &s, disks);
4760                 if (s.syncing + s.replacing)
4761                         handle_failed_sync(conf, sh, &s);
4762         }
4763
4764         /* Now we check to see if any write operations have recently
4765          * completed
4766          */
4767         prexor = 0;
4768         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4769                 prexor = 1;
4770         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4771             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4772                 sh->reconstruct_state = reconstruct_state_idle;
4773
4774                 /* All the 'written' buffers and the parity block are ready to
4775                  * be written back to disk
4776                  */
4777                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4778                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4779                 BUG_ON(sh->qd_idx >= 0 &&
4780                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4781                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4782                 for (i = disks; i--; ) {
4783                         struct r5dev *dev = &sh->dev[i];
4784                         if (test_bit(R5_LOCKED, &dev->flags) &&
4785                                 (i == sh->pd_idx || i == sh->qd_idx ||
4786                                  dev->written || test_bit(R5_InJournal,
4787                                                           &dev->flags))) {
4788                                 pr_debug("Writing block %d\n", i);
4789                                 set_bit(R5_Wantwrite, &dev->flags);
4790                                 if (prexor)
4791                                         continue;
4792                                 if (s.failed > 1)
4793                                         continue;
4794                                 if (!test_bit(R5_Insync, &dev->flags) ||
4795                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4796                                      s.failed == 0))
4797                                         set_bit(STRIPE_INSYNC, &sh->state);
4798                         }
4799                 }
4800                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4801                         s.dec_preread_active = 1;
4802         }
4803
4804         /*
4805          * might be able to return some write requests if the parity blocks
4806          * are safe, or on a failed drive
4807          */
4808         pdev = &sh->dev[sh->pd_idx];
4809         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4810                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4811         qdev = &sh->dev[sh->qd_idx];
4812         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4813                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4814                 || conf->level < 6;
4815
4816         if (s.written &&
4817             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4818                              && !test_bit(R5_LOCKED, &pdev->flags)
4819                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4820                                  test_bit(R5_Discard, &pdev->flags))))) &&
4821             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4822                              && !test_bit(R5_LOCKED, &qdev->flags)
4823                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4824                                  test_bit(R5_Discard, &qdev->flags))))))
4825                 handle_stripe_clean_event(conf, sh, disks);
4826
4827         if (s.just_cached)
4828                 r5c_handle_cached_data_endio(conf, sh, disks);
4829         log_stripe_write_finished(sh);
4830
4831         /* Now we might consider reading some blocks, either to check/generate
4832          * parity, or to satisfy requests
4833          * or to load a block that is being partially written.
4834          */
4835         if (s.to_read || s.non_overwrite
4836             || (conf->level == 6 && s.to_write && s.failed)
4837             || (s.syncing && (s.uptodate + s.compute < disks))
4838             || s.replacing
4839             || s.expanding)
4840                 handle_stripe_fill(sh, &s, disks);
4841
4842         /*
4843          * When the stripe finishes full journal write cycle (write to journal
4844          * and raid disk), this is the clean up procedure so it is ready for
4845          * next operation.
4846          */
4847         r5c_finish_stripe_write_out(conf, sh, &s);
4848
4849         /*
4850          * Now to consider new write requests, cache write back and what else,
4851          * if anything should be read.  We do not handle new writes when:
4852          * 1/ A 'write' operation (copy+xor) is already in flight.
4853          * 2/ A 'check' operation is in flight, as it may clobber the parity
4854          *    block.
4855          * 3/ A r5c cache log write is in flight.
4856          */
4857
4858         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4859                 if (!r5c_is_writeback(conf->log)) {
4860                         if (s.to_write)
4861                                 handle_stripe_dirtying(conf, sh, &s, disks);
4862                 } else { /* write back cache */
4863                         int ret = 0;
4864
4865                         /* First, try handle writes in caching phase */
4866                         if (s.to_write)
4867                                 ret = r5c_try_caching_write(conf, sh, &s,
4868                                                             disks);
4869                         /*
4870                          * If caching phase failed: ret == -EAGAIN
4871                          *    OR
4872                          * stripe under reclaim: !caching && injournal
4873                          *
4874                          * fall back to handle_stripe_dirtying()
4875                          */
4876                         if (ret == -EAGAIN ||
4877                             /* stripe under reclaim: !caching && injournal */
4878                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4879                              s.injournal > 0)) {
4880                                 ret = handle_stripe_dirtying(conf, sh, &s,
4881                                                              disks);
4882                                 if (ret == -EAGAIN)
4883                                         goto finish;
4884                         }
4885                 }
4886         }
4887
4888         /* maybe we need to check and possibly fix the parity for this stripe
4889          * Any reads will already have been scheduled, so we just see if enough
4890          * data is available.  The parity check is held off while parity
4891          * dependent operations are in flight.
4892          */
4893         if (sh->check_state ||
4894             (s.syncing && s.locked == 0 &&
4895              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4896              !test_bit(STRIPE_INSYNC, &sh->state))) {
4897                 if (conf->level == 6)
4898                         handle_parity_checks6(conf, sh, &s, disks);
4899                 else
4900                         handle_parity_checks5(conf, sh, &s, disks);
4901         }
4902
4903         if ((s.replacing || s.syncing) && s.locked == 0
4904             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4905             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4906                 /* Write out to replacement devices where possible */
4907                 for (i = 0; i < conf->raid_disks; i++)
4908                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4909                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4910                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4911                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4912                                 s.locked++;
4913                         }
4914                 if (s.replacing)
4915                         set_bit(STRIPE_INSYNC, &sh->state);
4916                 set_bit(STRIPE_REPLACED, &sh->state);
4917         }
4918         if ((s.syncing || s.replacing) && s.locked == 0 &&
4919             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4920             test_bit(STRIPE_INSYNC, &sh->state)) {
4921                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4922                 clear_bit(STRIPE_SYNCING, &sh->state);
4923                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4924                         wake_up(&conf->wait_for_overlap);
4925         }
4926
4927         /* If the failed drives are just a ReadError, then we might need
4928          * to progress the repair/check process
4929          */
4930         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4931                 for (i = 0; i < s.failed; i++) {
4932                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4933                         if (test_bit(R5_ReadError, &dev->flags)
4934                             && !test_bit(R5_LOCKED, &dev->flags)
4935                             && test_bit(R5_UPTODATE, &dev->flags)
4936                                 ) {
4937                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4938                                         set_bit(R5_Wantwrite, &dev->flags);
4939                                         set_bit(R5_ReWrite, &dev->flags);
4940                                         set_bit(R5_LOCKED, &dev->flags);
4941                                         s.locked++;
4942                                 } else {
4943                                         /* let's read it back */
4944                                         set_bit(R5_Wantread, &dev->flags);
4945                                         set_bit(R5_LOCKED, &dev->flags);
4946                                         s.locked++;
4947                                 }
4948                         }
4949                 }
4950
4951         /* Finish reconstruct operations initiated by the expansion process */
4952         if (sh->reconstruct_state == reconstruct_state_result) {
4953                 struct stripe_head *sh_src
4954                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4955                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4956                         /* sh cannot be written until sh_src has been read.
4957                          * so arrange for sh to be delayed a little
4958                          */
4959                         set_bit(STRIPE_DELAYED, &sh->state);
4960                         set_bit(STRIPE_HANDLE, &sh->state);
4961                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4962                                               &sh_src->state))
4963                                 atomic_inc(&conf->preread_active_stripes);
4964                         raid5_release_stripe(sh_src);
4965                         goto finish;
4966                 }
4967                 if (sh_src)
4968                         raid5_release_stripe(sh_src);
4969
4970                 sh->reconstruct_state = reconstruct_state_idle;
4971                 clear_bit(STRIPE_EXPANDING, &sh->state);
4972                 for (i = conf->raid_disks; i--; ) {
4973                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4974                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4975                         s.locked++;
4976                 }
4977         }
4978
4979         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4980             !sh->reconstruct_state) {
4981                 /* Need to write out all blocks after computing parity */
4982                 sh->disks = conf->raid_disks;
4983                 stripe_set_idx(sh->sector, conf, 0, sh);
4984                 schedule_reconstruction(sh, &s, 1, 1);
4985         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4986                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4987                 atomic_dec(&conf->reshape_stripes);
4988                 wake_up(&conf->wait_for_overlap);
4989                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4990         }
4991
4992         if (s.expanding && s.locked == 0 &&
4993             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4994                 handle_stripe_expansion(conf, sh);
4995
4996 finish:
4997         /* wait for this device to become unblocked */
4998         if (unlikely(s.blocked_rdev)) {
4999                 if (conf->mddev->external)
5000                         md_wait_for_blocked_rdev(s.blocked_rdev,
5001                                                  conf->mddev);
5002                 else
5003                         /* Internal metadata will immediately
5004                          * be written by raid5d, so we don't
5005                          * need to wait here.
5006                          */
5007                         rdev_dec_pending(s.blocked_rdev,
5008                                          conf->mddev);
5009         }
5010
5011         if (s.handle_bad_blocks)
5012                 for (i = disks; i--; ) {
5013                         struct md_rdev *rdev;
5014                         struct r5dev *dev = &sh->dev[i];
5015                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5016                                 /* We own a safe reference to the rdev */
5017                                 rdev = conf->disks[i].rdev;
5018                                 if (!rdev_set_badblocks(rdev, sh->sector,
5019                                                         STRIPE_SECTORS, 0))
5020                                         md_error(conf->mddev, rdev);
5021                                 rdev_dec_pending(rdev, conf->mddev);
5022                         }
5023                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5024                                 rdev = conf->disks[i].rdev;
5025                                 rdev_clear_badblocks(rdev, sh->sector,
5026                                                      STRIPE_SECTORS, 0);
5027                                 rdev_dec_pending(rdev, conf->mddev);
5028                         }
5029                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5030                                 rdev = conf->disks[i].replacement;
5031                                 if (!rdev)
5032                                         /* rdev have been moved down */
5033                                         rdev = conf->disks[i].rdev;
5034                                 rdev_clear_badblocks(rdev, sh->sector,
5035                                                      STRIPE_SECTORS, 0);
5036                                 rdev_dec_pending(rdev, conf->mddev);
5037                         }
5038                 }
5039
5040         if (s.ops_request)
5041                 raid_run_ops(sh, s.ops_request);
5042
5043         ops_run_io(sh, &s);
5044
5045         if (s.dec_preread_active) {
5046                 /* We delay this until after ops_run_io so that if make_request
5047                  * is waiting on a flush, it won't continue until the writes
5048                  * have actually been submitted.
5049                  */
5050                 atomic_dec(&conf->preread_active_stripes);
5051                 if (atomic_read(&conf->preread_active_stripes) <
5052                     IO_THRESHOLD)
5053                         md_wakeup_thread(conf->mddev->thread);
5054         }
5055
5056         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5057 }
5058
5059 static void raid5_activate_delayed(struct r5conf *conf)
5060 {
5061         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5062                 while (!list_empty(&conf->delayed_list)) {
5063                         struct list_head *l = conf->delayed_list.next;
5064                         struct stripe_head *sh;
5065                         sh = list_entry(l, struct stripe_head, lru);
5066                         list_del_init(l);
5067                         clear_bit(STRIPE_DELAYED, &sh->state);
5068                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5069                                 atomic_inc(&conf->preread_active_stripes);
5070                         list_add_tail(&sh->lru, &conf->hold_list);
5071                         raid5_wakeup_stripe_thread(sh);
5072                 }
5073         }
5074 }
5075
5076 static void activate_bit_delay(struct r5conf *conf,
5077         struct list_head *temp_inactive_list)
5078 {
5079         /* device_lock is held */
5080         struct list_head head;
5081         list_add(&head, &conf->bitmap_list);
5082         list_del_init(&conf->bitmap_list);
5083         while (!list_empty(&head)) {
5084                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5085                 int hash;
5086                 list_del_init(&sh->lru);
5087                 atomic_inc(&sh->count);
5088                 hash = sh->hash_lock_index;
5089                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5090         }
5091 }
5092
5093 static int raid5_congested(struct mddev *mddev, int bits)
5094 {
5095         struct r5conf *conf = mddev->private;
5096
5097         /* No difference between reads and writes.  Just check
5098          * how busy the stripe_cache is
5099          */
5100
5101         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5102                 return 1;
5103
5104         /* Also checks whether there is pressure on r5cache log space */
5105         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5106                 return 1;
5107         if (conf->quiesce)
5108                 return 1;
5109         if (atomic_read(&conf->empty_inactive_list_nr))
5110                 return 1;
5111
5112         return 0;
5113 }
5114
5115 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5116 {
5117         struct r5conf *conf = mddev->private;
5118         sector_t sector = bio->bi_iter.bi_sector;
5119         unsigned int chunk_sectors;
5120         unsigned int bio_sectors = bio_sectors(bio);
5121
5122         WARN_ON_ONCE(bio->bi_partno);
5123
5124         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5125         return  chunk_sectors >=
5126                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5127 }
5128
5129 /*
5130  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5131  *  later sampled by raid5d.
5132  */
5133 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5134 {
5135         unsigned long flags;
5136
5137         spin_lock_irqsave(&conf->device_lock, flags);
5138
5139         bi->bi_next = conf->retry_read_aligned_list;
5140         conf->retry_read_aligned_list = bi;
5141
5142         spin_unlock_irqrestore(&conf->device_lock, flags);
5143         md_wakeup_thread(conf->mddev->thread);
5144 }
5145
5146 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5147                                          unsigned int *offset)
5148 {
5149         struct bio *bi;
5150
5151         bi = conf->retry_read_aligned;
5152         if (bi) {
5153                 *offset = conf->retry_read_offset;
5154                 conf->retry_read_aligned = NULL;
5155                 return bi;
5156         }
5157         bi = conf->retry_read_aligned_list;
5158         if(bi) {
5159                 conf->retry_read_aligned_list = bi->bi_next;
5160                 bi->bi_next = NULL;
5161                 *offset = 0;
5162         }
5163
5164         return bi;
5165 }
5166
5167 /*
5168  *  The "raid5_align_endio" should check if the read succeeded and if it
5169  *  did, call bio_endio on the original bio (having bio_put the new bio
5170  *  first).
5171  *  If the read failed..
5172  */
5173 static void raid5_align_endio(struct bio *bi)
5174 {
5175         struct bio* raid_bi  = bi->bi_private;
5176         struct mddev *mddev;
5177         struct r5conf *conf;
5178         struct md_rdev *rdev;
5179         blk_status_t error = bi->bi_status;
5180
5181         bio_put(bi);
5182
5183         rdev = (void*)raid_bi->bi_next;
5184         raid_bi->bi_next = NULL;
5185         mddev = rdev->mddev;
5186         conf = mddev->private;
5187
5188         rdev_dec_pending(rdev, conf->mddev);
5189
5190         if (!error) {
5191                 bio_endio(raid_bi);
5192                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5193                         wake_up(&conf->wait_for_quiescent);
5194                 return;
5195         }
5196
5197         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5198
5199         add_bio_to_retry(raid_bi, conf);
5200 }
5201
5202 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5203 {
5204         struct r5conf *conf = mddev->private;
5205         int dd_idx;
5206         struct bio* align_bi;
5207         struct md_rdev *rdev;
5208         sector_t end_sector;
5209
5210         if (!in_chunk_boundary(mddev, raid_bio)) {
5211                 pr_debug("%s: non aligned\n", __func__);
5212                 return 0;
5213         }
5214         /*
5215          * use bio_clone_fast to make a copy of the bio
5216          */
5217         align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5218         if (!align_bi)
5219                 return 0;
5220         /*
5221          *   set bi_end_io to a new function, and set bi_private to the
5222          *     original bio.
5223          */
5224         align_bi->bi_end_io  = raid5_align_endio;
5225         align_bi->bi_private = raid_bio;
5226         /*
5227          *      compute position
5228          */
5229         align_bi->bi_iter.bi_sector =
5230                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5231                                      0, &dd_idx, NULL);
5232
5233         end_sector = bio_end_sector(align_bi);
5234         rcu_read_lock();
5235         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5236         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5237             rdev->recovery_offset < end_sector) {
5238                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5239                 if (rdev &&
5240                     (test_bit(Faulty, &rdev->flags) ||
5241                     !(test_bit(In_sync, &rdev->flags) ||
5242                       rdev->recovery_offset >= end_sector)))
5243                         rdev = NULL;
5244         }
5245
5246         if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5247                 rcu_read_unlock();
5248                 bio_put(align_bi);
5249                 return 0;
5250         }
5251
5252         if (rdev) {
5253                 sector_t first_bad;
5254                 int bad_sectors;
5255
5256                 atomic_inc(&rdev->nr_pending);
5257                 rcu_read_unlock();
5258                 raid_bio->bi_next = (void*)rdev;
5259                 bio_set_dev(align_bi, rdev->bdev);
5260                 bio_clear_flag(align_bi, BIO_SEG_VALID);
5261
5262                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5263                                 bio_sectors(align_bi),
5264                                 &first_bad, &bad_sectors)) {
5265                         bio_put(align_bi);
5266                         rdev_dec_pending(rdev, mddev);
5267                         return 0;
5268                 }
5269
5270                 /* No reshape active, so we can trust rdev->data_offset */
5271                 align_bi->bi_iter.bi_sector += rdev->data_offset;
5272
5273                 spin_lock_irq(&conf->device_lock);
5274                 wait_event_lock_irq(conf->wait_for_quiescent,
5275                                     conf->quiesce == 0,
5276                                     conf->device_lock);
5277                 atomic_inc(&conf->active_aligned_reads);
5278                 spin_unlock_irq(&conf->device_lock);
5279
5280                 if (mddev->gendisk)
5281                         trace_block_bio_remap(align_bi->bi_disk->queue,
5282                                               align_bi, disk_devt(mddev->gendisk),
5283                                               raid_bio->bi_iter.bi_sector);
5284                 generic_make_request(align_bi);
5285                 return 1;
5286         } else {
5287                 rcu_read_unlock();
5288                 bio_put(align_bi);
5289                 return 0;
5290         }
5291 }
5292
5293 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5294 {
5295         struct bio *split;
5296         sector_t sector = raid_bio->bi_iter.bi_sector;
5297         unsigned chunk_sects = mddev->chunk_sectors;
5298         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5299
5300         if (sectors < bio_sectors(raid_bio)) {
5301                 struct r5conf *conf = mddev->private;
5302                 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5303                 bio_chain(split, raid_bio);
5304                 generic_make_request(raid_bio);
5305                 raid_bio = split;
5306         }
5307
5308         if (!raid5_read_one_chunk(mddev, raid_bio))
5309                 return raid_bio;
5310
5311         return NULL;
5312 }
5313
5314 /* __get_priority_stripe - get the next stripe to process
5315  *
5316  * Full stripe writes are allowed to pass preread active stripes up until
5317  * the bypass_threshold is exceeded.  In general the bypass_count
5318  * increments when the handle_list is handled before the hold_list; however, it
5319  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5320  * stripe with in flight i/o.  The bypass_count will be reset when the
5321  * head of the hold_list has changed, i.e. the head was promoted to the
5322  * handle_list.
5323  */
5324 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5325 {
5326         struct stripe_head *sh, *tmp;
5327         struct list_head *handle_list = NULL;
5328         struct r5worker_group *wg;
5329         bool second_try = !r5c_is_writeback(conf->log) &&
5330                 !r5l_log_disk_error(conf);
5331         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5332                 r5l_log_disk_error(conf);
5333
5334 again:
5335         wg = NULL;
5336         sh = NULL;
5337         if (conf->worker_cnt_per_group == 0) {
5338                 handle_list = try_loprio ? &conf->loprio_list :
5339                                         &conf->handle_list;
5340         } else if (group != ANY_GROUP) {
5341                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5342                                 &conf->worker_groups[group].handle_list;
5343                 wg = &conf->worker_groups[group];
5344         } else {
5345                 int i;
5346                 for (i = 0; i < conf->group_cnt; i++) {
5347                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5348                                 &conf->worker_groups[i].handle_list;
5349                         wg = &conf->worker_groups[i];
5350                         if (!list_empty(handle_list))
5351                                 break;
5352                 }
5353         }
5354
5355         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5356                   __func__,
5357                   list_empty(handle_list) ? "empty" : "busy",
5358                   list_empty(&conf->hold_list) ? "empty" : "busy",
5359                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5360
5361         if (!list_empty(handle_list)) {
5362                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5363
5364                 if (list_empty(&conf->hold_list))
5365                         conf->bypass_count = 0;
5366                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5367                         if (conf->hold_list.next == conf->last_hold)
5368                                 conf->bypass_count++;
5369                         else {
5370                                 conf->last_hold = conf->hold_list.next;
5371                                 conf->bypass_count -= conf->bypass_threshold;
5372                                 if (conf->bypass_count < 0)
5373                                         conf->bypass_count = 0;
5374                         }
5375                 }
5376         } else if (!list_empty(&conf->hold_list) &&
5377                    ((conf->bypass_threshold &&
5378                      conf->bypass_count > conf->bypass_threshold) ||
5379                     atomic_read(&conf->pending_full_writes) == 0)) {
5380
5381                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5382                         if (conf->worker_cnt_per_group == 0 ||
5383                             group == ANY_GROUP ||
5384                             !cpu_online(tmp->cpu) ||
5385                             cpu_to_group(tmp->cpu) == group) {
5386                                 sh = tmp;
5387                                 break;
5388                         }
5389                 }
5390
5391                 if (sh) {
5392                         conf->bypass_count -= conf->bypass_threshold;
5393                         if (conf->bypass_count < 0)
5394                                 conf->bypass_count = 0;
5395                 }
5396                 wg = NULL;
5397         }
5398
5399         if (!sh) {
5400                 if (second_try)
5401                         return NULL;
5402                 second_try = true;
5403                 try_loprio = !try_loprio;
5404                 goto again;
5405         }
5406
5407         if (wg) {
5408                 wg->stripes_cnt--;
5409                 sh->group = NULL;
5410         }
5411         list_del_init(&sh->lru);
5412         BUG_ON(atomic_inc_return(&sh->count) != 1);
5413         return sh;
5414 }
5415
5416 struct raid5_plug_cb {
5417         struct blk_plug_cb      cb;
5418         struct list_head        list;
5419         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5420 };
5421
5422 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5423 {
5424         struct raid5_plug_cb *cb = container_of(
5425                 blk_cb, struct raid5_plug_cb, cb);
5426         struct stripe_head *sh;
5427         struct mddev *mddev = cb->cb.data;
5428         struct r5conf *conf = mddev->private;
5429         int cnt = 0;
5430         int hash;
5431
5432         if (cb->list.next && !list_empty(&cb->list)) {
5433                 spin_lock_irq(&conf->device_lock);
5434                 while (!list_empty(&cb->list)) {
5435                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5436                         list_del_init(&sh->lru);
5437                         /*
5438                          * avoid race release_stripe_plug() sees
5439                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5440                          * is still in our list
5441                          */
5442                         smp_mb__before_atomic();
5443                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5444                         /*
5445                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5446                          * case, the count is always > 1 here
5447                          */
5448                         hash = sh->hash_lock_index;
5449                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5450                         cnt++;
5451                 }
5452                 spin_unlock_irq(&conf->device_lock);
5453         }
5454         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5455                                      NR_STRIPE_HASH_LOCKS);
5456         if (mddev->queue)
5457                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5458         kfree(cb);
5459 }
5460
5461 static void release_stripe_plug(struct mddev *mddev,
5462                                 struct stripe_head *sh)
5463 {
5464         struct blk_plug_cb *blk_cb = blk_check_plugged(
5465                 raid5_unplug, mddev,
5466                 sizeof(struct raid5_plug_cb));
5467         struct raid5_plug_cb *cb;
5468
5469         if (!blk_cb) {
5470                 raid5_release_stripe(sh);
5471                 return;
5472         }
5473
5474         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5475
5476         if (cb->list.next == NULL) {
5477                 int i;
5478                 INIT_LIST_HEAD(&cb->list);
5479                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5480                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5481         }
5482
5483         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5484                 list_add_tail(&sh->lru, &cb->list);
5485         else
5486                 raid5_release_stripe(sh);
5487 }
5488
5489 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5490 {
5491         struct r5conf *conf = mddev->private;
5492         sector_t logical_sector, last_sector;
5493         struct stripe_head *sh;
5494         int stripe_sectors;
5495
5496         if (mddev->reshape_position != MaxSector)
5497                 /* Skip discard while reshape is happening */
5498                 return;
5499
5500         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5501         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5502
5503         bi->bi_next = NULL;
5504
5505         stripe_sectors = conf->chunk_sectors *
5506                 (conf->raid_disks - conf->max_degraded);
5507         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5508                                                stripe_sectors);
5509         sector_div(last_sector, stripe_sectors);
5510
5511         logical_sector *= conf->chunk_sectors;
5512         last_sector *= conf->chunk_sectors;
5513
5514         for (; logical_sector < last_sector;
5515              logical_sector += STRIPE_SECTORS) {
5516                 DEFINE_WAIT(w);
5517                 int d;
5518         again:
5519                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5520                 prepare_to_wait(&conf->wait_for_overlap, &w,
5521                                 TASK_UNINTERRUPTIBLE);
5522                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5523                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5524                         raid5_release_stripe(sh);
5525                         schedule();
5526                         goto again;
5527                 }
5528                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5529                 spin_lock_irq(&sh->stripe_lock);
5530                 for (d = 0; d < conf->raid_disks; d++) {
5531                         if (d == sh->pd_idx || d == sh->qd_idx)
5532                                 continue;
5533                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5534                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5535                                 spin_unlock_irq(&sh->stripe_lock);
5536                                 raid5_release_stripe(sh);
5537                                 schedule();
5538                                 goto again;
5539                         }
5540                 }
5541                 set_bit(STRIPE_DISCARD, &sh->state);
5542                 finish_wait(&conf->wait_for_overlap, &w);
5543                 sh->overwrite_disks = 0;
5544                 for (d = 0; d < conf->raid_disks; d++) {
5545                         if (d == sh->pd_idx || d == sh->qd_idx)
5546                                 continue;
5547                         sh->dev[d].towrite = bi;
5548                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5549                         bio_inc_remaining(bi);
5550                         md_write_inc(mddev, bi);
5551                         sh->overwrite_disks++;
5552                 }
5553                 spin_unlock_irq(&sh->stripe_lock);
5554                 if (conf->mddev->bitmap) {
5555                         for (d = 0;
5556                              d < conf->raid_disks - conf->max_degraded;
5557                              d++)
5558                                 md_bitmap_startwrite(mddev->bitmap,
5559                                                      sh->sector,
5560                                                      STRIPE_SECTORS,
5561                                                      0);
5562                         sh->bm_seq = conf->seq_flush + 1;
5563                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5564                 }
5565
5566                 set_bit(STRIPE_HANDLE, &sh->state);
5567                 clear_bit(STRIPE_DELAYED, &sh->state);
5568                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5569                         atomic_inc(&conf->preread_active_stripes);
5570                 release_stripe_plug(mddev, sh);
5571         }
5572
5573         bio_endio(bi);
5574 }
5575
5576 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5577 {
5578         struct r5conf *conf = mddev->private;
5579         int dd_idx;
5580         sector_t new_sector;
5581         sector_t logical_sector, last_sector;
5582         struct stripe_head *sh;
5583         const int rw = bio_data_dir(bi);
5584         DEFINE_WAIT(w);
5585         bool do_prepare;
5586         bool do_flush = false;
5587
5588         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5589                 int ret = log_handle_flush_request(conf, bi);
5590
5591                 if (ret == 0)
5592                         return true;
5593                 if (ret == -ENODEV) {
5594                         md_flush_request(mddev, bi);
5595                         return true;
5596                 }
5597                 /* ret == -EAGAIN, fallback */
5598                 /*
5599                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5600                  * we need to flush journal device
5601                  */
5602                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5603         }
5604
5605         if (!md_write_start(mddev, bi))
5606                 return false;
5607         /*
5608          * If array is degraded, better not do chunk aligned read because
5609          * later we might have to read it again in order to reconstruct
5610          * data on failed drives.
5611          */
5612         if (rw == READ && mddev->degraded == 0 &&
5613             mddev->reshape_position == MaxSector) {
5614                 bi = chunk_aligned_read(mddev, bi);
5615                 if (!bi)
5616                         return true;
5617         }
5618
5619         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5620                 make_discard_request(mddev, bi);
5621                 md_write_end(mddev);
5622                 return true;
5623         }
5624
5625         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5626         last_sector = bio_end_sector(bi);
5627         bi->bi_next = NULL;
5628
5629         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5630         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5631                 int previous;
5632                 int seq;
5633
5634                 do_prepare = false;
5635         retry:
5636                 seq = read_seqcount_begin(&conf->gen_lock);
5637                 previous = 0;
5638                 if (do_prepare)
5639                         prepare_to_wait(&conf->wait_for_overlap, &w,
5640                                 TASK_UNINTERRUPTIBLE);
5641                 if (unlikely(conf->reshape_progress != MaxSector)) {
5642                         /* spinlock is needed as reshape_progress may be
5643                          * 64bit on a 32bit platform, and so it might be
5644                          * possible to see a half-updated value
5645                          * Of course reshape_progress could change after
5646                          * the lock is dropped, so once we get a reference
5647                          * to the stripe that we think it is, we will have
5648                          * to check again.
5649                          */
5650                         spin_lock_irq(&conf->device_lock);
5651                         if (mddev->reshape_backwards
5652                             ? logical_sector < conf->reshape_progress
5653                             : logical_sector >= conf->reshape_progress) {
5654                                 previous = 1;
5655                         } else {
5656                                 if (mddev->reshape_backwards
5657                                     ? logical_sector < conf->reshape_safe
5658                                     : logical_sector >= conf->reshape_safe) {
5659                                         spin_unlock_irq(&conf->device_lock);
5660                                         schedule();
5661                                         do_prepare = true;
5662                                         goto retry;
5663                                 }
5664                         }
5665                         spin_unlock_irq(&conf->device_lock);
5666                 }
5667
5668                 new_sector = raid5_compute_sector(conf, logical_sector,
5669                                                   previous,
5670                                                   &dd_idx, NULL);
5671                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5672                         (unsigned long long)new_sector,
5673                         (unsigned long long)logical_sector);
5674
5675                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5676                                        (bi->bi_opf & REQ_RAHEAD), 0);
5677                 if (sh) {
5678                         if (unlikely(previous)) {
5679                                 /* expansion might have moved on while waiting for a
5680                                  * stripe, so we must do the range check again.
5681                                  * Expansion could still move past after this
5682                                  * test, but as we are holding a reference to
5683                                  * 'sh', we know that if that happens,
5684                                  *  STRIPE_EXPANDING will get set and the expansion
5685                                  * won't proceed until we finish with the stripe.
5686                                  */
5687                                 int must_retry = 0;
5688                                 spin_lock_irq(&conf->device_lock);
5689                                 if (mddev->reshape_backwards
5690                                     ? logical_sector >= conf->reshape_progress
5691                                     : logical_sector < conf->reshape_progress)
5692                                         /* mismatch, need to try again */
5693                                         must_retry = 1;
5694                                 spin_unlock_irq(&conf->device_lock);
5695                                 if (must_retry) {
5696                                         raid5_release_stripe(sh);
5697                                         schedule();
5698                                         do_prepare = true;
5699                                         goto retry;
5700                                 }
5701                         }
5702                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5703                                 /* Might have got the wrong stripe_head
5704                                  * by accident
5705                                  */
5706                                 raid5_release_stripe(sh);
5707                                 goto retry;
5708                         }
5709
5710                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5711                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5712                                 /* Stripe is busy expanding or
5713                                  * add failed due to overlap.  Flush everything
5714                                  * and wait a while
5715                                  */
5716                                 md_wakeup_thread(mddev->thread);
5717                                 raid5_release_stripe(sh);
5718                                 schedule();
5719                                 do_prepare = true;
5720                                 goto retry;
5721                         }
5722                         if (do_flush) {
5723                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5724                                 /* we only need flush for one stripe */
5725                                 do_flush = false;
5726                         }
5727
5728                         set_bit(STRIPE_HANDLE, &sh->state);
5729                         clear_bit(STRIPE_DELAYED, &sh->state);
5730                         if ((!sh->batch_head || sh == sh->batch_head) &&
5731                             (bi->bi_opf & REQ_SYNC) &&
5732                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5733                                 atomic_inc(&conf->preread_active_stripes);
5734                         release_stripe_plug(mddev, sh);
5735                 } else {
5736                         /* cannot get stripe for read-ahead, just give-up */
5737                         bi->bi_status = BLK_STS_IOERR;
5738                         break;
5739                 }
5740         }
5741         finish_wait(&conf->wait_for_overlap, &w);
5742
5743         if (rw == WRITE)
5744                 md_write_end(mddev);
5745         bio_endio(bi);
5746         return true;
5747 }
5748
5749 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5750
5751 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5752 {
5753         /* reshaping is quite different to recovery/resync so it is
5754          * handled quite separately ... here.
5755          *
5756          * On each call to sync_request, we gather one chunk worth of
5757          * destination stripes and flag them as expanding.
5758          * Then we find all the source stripes and request reads.
5759          * As the reads complete, handle_stripe will copy the data
5760          * into the destination stripe and release that stripe.
5761          */
5762         struct r5conf *conf = mddev->private;
5763         struct stripe_head *sh;
5764         struct md_rdev *rdev;
5765         sector_t first_sector, last_sector;
5766         int raid_disks = conf->previous_raid_disks;
5767         int data_disks = raid_disks - conf->max_degraded;
5768         int new_data_disks = conf->raid_disks - conf->max_degraded;
5769         int i;
5770         int dd_idx;
5771         sector_t writepos, readpos, safepos;
5772         sector_t stripe_addr;
5773         int reshape_sectors;
5774         struct list_head stripes;
5775         sector_t retn;
5776
5777         if (sector_nr == 0) {
5778                 /* If restarting in the middle, skip the initial sectors */
5779                 if (mddev->reshape_backwards &&
5780                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5781                         sector_nr = raid5_size(mddev, 0, 0)
5782                                 - conf->reshape_progress;
5783                 } else if (mddev->reshape_backwards &&
5784                            conf->reshape_progress == MaxSector) {
5785                         /* shouldn't happen, but just in case, finish up.*/
5786                         sector_nr = MaxSector;
5787                 } else if (!mddev->reshape_backwards &&
5788                            conf->reshape_progress > 0)
5789                         sector_nr = conf->reshape_progress;
5790                 sector_div(sector_nr, new_data_disks);
5791                 if (sector_nr) {
5792                         mddev->curr_resync_completed = sector_nr;
5793                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5794                         *skipped = 1;
5795                         retn = sector_nr;
5796                         goto finish;
5797                 }
5798         }
5799
5800         /* We need to process a full chunk at a time.
5801          * If old and new chunk sizes differ, we need to process the
5802          * largest of these
5803          */
5804
5805         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5806
5807         /* We update the metadata at least every 10 seconds, or when
5808          * the data about to be copied would over-write the source of
5809          * the data at the front of the range.  i.e. one new_stripe
5810          * along from reshape_progress new_maps to after where
5811          * reshape_safe old_maps to
5812          */
5813         writepos = conf->reshape_progress;
5814         sector_div(writepos, new_data_disks);
5815         readpos = conf->reshape_progress;
5816         sector_div(readpos, data_disks);
5817         safepos = conf->reshape_safe;
5818         sector_div(safepos, data_disks);
5819         if (mddev->reshape_backwards) {
5820                 BUG_ON(writepos < reshape_sectors);
5821                 writepos -= reshape_sectors;
5822                 readpos += reshape_sectors;
5823                 safepos += reshape_sectors;
5824         } else {
5825                 writepos += reshape_sectors;
5826                 /* readpos and safepos are worst-case calculations.
5827                  * A negative number is overly pessimistic, and causes
5828                  * obvious problems for unsigned storage.  So clip to 0.
5829                  */
5830                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5831                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5832         }
5833
5834         /* Having calculated the 'writepos' possibly use it
5835          * to set 'stripe_addr' which is where we will write to.
5836          */
5837         if (mddev->reshape_backwards) {
5838                 BUG_ON(conf->reshape_progress == 0);
5839                 stripe_addr = writepos;
5840                 BUG_ON((mddev->dev_sectors &
5841                         ~((sector_t)reshape_sectors - 1))
5842                        - reshape_sectors - stripe_addr
5843                        != sector_nr);
5844         } else {
5845                 BUG_ON(writepos != sector_nr + reshape_sectors);
5846                 stripe_addr = sector_nr;
5847         }
5848
5849         /* 'writepos' is the most advanced device address we might write.
5850          * 'readpos' is the least advanced device address we might read.
5851          * 'safepos' is the least address recorded in the metadata as having
5852          *     been reshaped.
5853          * If there is a min_offset_diff, these are adjusted either by
5854          * increasing the safepos/readpos if diff is negative, or
5855          * increasing writepos if diff is positive.
5856          * If 'readpos' is then behind 'writepos', there is no way that we can
5857          * ensure safety in the face of a crash - that must be done by userspace
5858          * making a backup of the data.  So in that case there is no particular
5859          * rush to update metadata.
5860          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5861          * update the metadata to advance 'safepos' to match 'readpos' so that
5862          * we can be safe in the event of a crash.
5863          * So we insist on updating metadata if safepos is behind writepos and
5864          * readpos is beyond writepos.
5865          * In any case, update the metadata every 10 seconds.
5866          * Maybe that number should be configurable, but I'm not sure it is
5867          * worth it.... maybe it could be a multiple of safemode_delay???
5868          */
5869         if (conf->min_offset_diff < 0) {
5870                 safepos += -conf->min_offset_diff;
5871                 readpos += -conf->min_offset_diff;
5872         } else
5873                 writepos += conf->min_offset_diff;
5874
5875         if ((mddev->reshape_backwards
5876              ? (safepos > writepos && readpos < writepos)
5877              : (safepos < writepos && readpos > writepos)) ||
5878             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5879                 /* Cannot proceed until we've updated the superblock... */
5880                 wait_event(conf->wait_for_overlap,
5881                            atomic_read(&conf->reshape_stripes)==0
5882                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5883                 if (atomic_read(&conf->reshape_stripes) != 0)
5884                         return 0;
5885                 mddev->reshape_position = conf->reshape_progress;
5886                 mddev->curr_resync_completed = sector_nr;
5887                 if (!mddev->reshape_backwards)
5888                         /* Can update recovery_offset */
5889                         rdev_for_each(rdev, mddev)
5890                                 if (rdev->raid_disk >= 0 &&
5891                                     !test_bit(Journal, &rdev->flags) &&
5892                                     !test_bit(In_sync, &rdev->flags) &&
5893                                     rdev->recovery_offset < sector_nr)
5894                                         rdev->recovery_offset = sector_nr;
5895
5896                 conf->reshape_checkpoint = jiffies;
5897                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5898                 md_wakeup_thread(mddev->thread);
5899                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5900                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5901                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5902                         return 0;
5903                 spin_lock_irq(&conf->device_lock);
5904                 conf->reshape_safe = mddev->reshape_position;
5905                 spin_unlock_irq(&conf->device_lock);
5906                 wake_up(&conf->wait_for_overlap);
5907                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5908         }
5909
5910         INIT_LIST_HEAD(&stripes);
5911         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5912                 int j;
5913                 int skipped_disk = 0;
5914                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5915                 set_bit(STRIPE_EXPANDING, &sh->state);
5916                 atomic_inc(&conf->reshape_stripes);
5917                 /* If any of this stripe is beyond the end of the old
5918                  * array, then we need to zero those blocks
5919                  */
5920                 for (j=sh->disks; j--;) {
5921                         sector_t s;
5922                         if (j == sh->pd_idx)
5923                                 continue;
5924                         if (conf->level == 6 &&
5925                             j == sh->qd_idx)
5926                                 continue;
5927                         s = raid5_compute_blocknr(sh, j, 0);
5928                         if (s < raid5_size(mddev, 0, 0)) {
5929                                 skipped_disk = 1;
5930                                 continue;
5931                         }
5932                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5933                         set_bit(R5_Expanded, &sh->dev[j].flags);
5934                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5935                 }
5936                 if (!skipped_disk) {
5937                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5938                         set_bit(STRIPE_HANDLE, &sh->state);
5939                 }
5940                 list_add(&sh->lru, &stripes);
5941         }
5942         spin_lock_irq(&conf->device_lock);
5943         if (mddev->reshape_backwards)
5944                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5945         else
5946                 conf->reshape_progress += reshape_sectors * new_data_disks;
5947         spin_unlock_irq(&conf->device_lock);
5948         /* Ok, those stripe are ready. We can start scheduling
5949          * reads on the source stripes.
5950          * The source stripes are determined by mapping the first and last
5951          * block on the destination stripes.
5952          */
5953         first_sector =
5954                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5955                                      1, &dd_idx, NULL);
5956         last_sector =
5957                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5958                                             * new_data_disks - 1),
5959                                      1, &dd_idx, NULL);
5960         if (last_sector >= mddev->dev_sectors)
5961                 last_sector = mddev->dev_sectors - 1;
5962         while (first_sector <= last_sector) {
5963                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5964                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5965                 set_bit(STRIPE_HANDLE, &sh->state);
5966                 raid5_release_stripe(sh);
5967                 first_sector += STRIPE_SECTORS;
5968         }
5969         /* Now that the sources are clearly marked, we can release
5970          * the destination stripes
5971          */
5972         while (!list_empty(&stripes)) {
5973                 sh = list_entry(stripes.next, struct stripe_head, lru);
5974                 list_del_init(&sh->lru);
5975                 raid5_release_stripe(sh);
5976         }
5977         /* If this takes us to the resync_max point where we have to pause,
5978          * then we need to write out the superblock.
5979          */
5980         sector_nr += reshape_sectors;
5981         retn = reshape_sectors;
5982 finish:
5983         if (mddev->curr_resync_completed > mddev->resync_max ||
5984             (sector_nr - mddev->curr_resync_completed) * 2
5985             >= mddev->resync_max - mddev->curr_resync_completed) {
5986                 /* Cannot proceed until we've updated the superblock... */
5987                 wait_event(conf->wait_for_overlap,
5988                            atomic_read(&conf->reshape_stripes) == 0
5989                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5990                 if (atomic_read(&conf->reshape_stripes) != 0)
5991                         goto ret;
5992                 mddev->reshape_position = conf->reshape_progress;
5993                 mddev->curr_resync_completed = sector_nr;
5994                 if (!mddev->reshape_backwards)
5995                         /* Can update recovery_offset */
5996                         rdev_for_each(rdev, mddev)
5997                                 if (rdev->raid_disk >= 0 &&
5998                                     !test_bit(Journal, &rdev->flags) &&
5999                                     !test_bit(In_sync, &rdev->flags) &&
6000                                     rdev->recovery_offset < sector_nr)
6001                                         rdev->recovery_offset = sector_nr;
6002                 conf->reshape_checkpoint = jiffies;
6003                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6004                 md_wakeup_thread(mddev->thread);
6005                 wait_event(mddev->sb_wait,
6006                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6007                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6008                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6009                         goto ret;
6010                 spin_lock_irq(&conf->device_lock);
6011                 conf->reshape_safe = mddev->reshape_position;
6012                 spin_unlock_irq(&conf->device_lock);
6013                 wake_up(&conf->wait_for_overlap);
6014                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6015         }
6016 ret:
6017         return retn;
6018 }
6019
6020 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6021                                           int *skipped)
6022 {
6023         struct r5conf *conf = mddev->private;
6024         struct stripe_head *sh;
6025         sector_t max_sector = mddev->dev_sectors;
6026         sector_t sync_blocks;
6027         int still_degraded = 0;
6028         int i;
6029
6030         if (sector_nr >= max_sector) {
6031                 /* just being told to finish up .. nothing much to do */
6032
6033                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6034                         end_reshape(conf);
6035                         return 0;
6036                 }
6037
6038                 if (mddev->curr_resync < max_sector) /* aborted */
6039                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6040                                            &sync_blocks, 1);
6041                 else /* completed sync */
6042                         conf->fullsync = 0;
6043                 md_bitmap_close_sync(mddev->bitmap);
6044
6045                 return 0;
6046         }
6047
6048         /* Allow raid5_quiesce to complete */
6049         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6050
6051         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6052                 return reshape_request(mddev, sector_nr, skipped);
6053
6054         /* No need to check resync_max as we never do more than one
6055          * stripe, and as resync_max will always be on a chunk boundary,
6056          * if the check in md_do_sync didn't fire, there is no chance
6057          * of overstepping resync_max here
6058          */
6059
6060         /* if there is too many failed drives and we are trying
6061          * to resync, then assert that we are finished, because there is
6062          * nothing we can do.
6063          */
6064         if (mddev->degraded >= conf->max_degraded &&
6065             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6066                 sector_t rv = mddev->dev_sectors - sector_nr;
6067                 *skipped = 1;
6068                 return rv;
6069         }
6070         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6071             !conf->fullsync &&
6072             !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6073             sync_blocks >= STRIPE_SECTORS) {
6074                 /* we can skip this block, and probably more */
6075                 sync_blocks /= STRIPE_SECTORS;
6076                 *skipped = 1;
6077                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6078         }
6079
6080         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6081
6082         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6083         if (sh == NULL) {
6084                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6085                 /* make sure we don't swamp the stripe cache if someone else
6086                  * is trying to get access
6087                  */
6088                 schedule_timeout_uninterruptible(1);
6089         }
6090         /* Need to check if array will still be degraded after recovery/resync
6091          * Note in case of > 1 drive failures it's possible we're rebuilding
6092          * one drive while leaving another faulty drive in array.
6093          */
6094         rcu_read_lock();
6095         for (i = 0; i < conf->raid_disks; i++) {
6096                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6097
6098                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6099                         still_degraded = 1;
6100         }
6101         rcu_read_unlock();
6102
6103         md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6104
6105         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6106         set_bit(STRIPE_HANDLE, &sh->state);
6107
6108         raid5_release_stripe(sh);
6109
6110         return STRIPE_SECTORS;
6111 }
6112
6113 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6114                                unsigned int offset)
6115 {
6116         /* We may not be able to submit a whole bio at once as there
6117          * may not be enough stripe_heads available.
6118          * We cannot pre-allocate enough stripe_heads as we may need
6119          * more than exist in the cache (if we allow ever large chunks).
6120          * So we do one stripe head at a time and record in
6121          * ->bi_hw_segments how many have been done.
6122          *
6123          * We *know* that this entire raid_bio is in one chunk, so
6124          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6125          */
6126         struct stripe_head *sh;
6127         int dd_idx;
6128         sector_t sector, logical_sector, last_sector;
6129         int scnt = 0;
6130         int handled = 0;
6131
6132         logical_sector = raid_bio->bi_iter.bi_sector &
6133                 ~((sector_t)STRIPE_SECTORS-1);
6134         sector = raid5_compute_sector(conf, logical_sector,
6135                                       0, &dd_idx, NULL);
6136         last_sector = bio_end_sector(raid_bio);
6137
6138         for (; logical_sector < last_sector;
6139              logical_sector += STRIPE_SECTORS,
6140                      sector += STRIPE_SECTORS,
6141                      scnt++) {
6142
6143                 if (scnt < offset)
6144                         /* already done this stripe */
6145                         continue;
6146
6147                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6148
6149                 if (!sh) {
6150                         /* failed to get a stripe - must wait */
6151                         conf->retry_read_aligned = raid_bio;
6152                         conf->retry_read_offset = scnt;
6153                         return handled;
6154                 }
6155
6156                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6157                         raid5_release_stripe(sh);
6158                         conf->retry_read_aligned = raid_bio;
6159                         conf->retry_read_offset = scnt;
6160                         return handled;
6161                 }
6162
6163                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6164                 handle_stripe(sh);
6165                 raid5_release_stripe(sh);
6166                 handled++;
6167         }
6168
6169         bio_endio(raid_bio);
6170
6171         if (atomic_dec_and_test(&conf->active_aligned_reads))
6172                 wake_up(&conf->wait_for_quiescent);
6173         return handled;
6174 }
6175
6176 static int handle_active_stripes(struct r5conf *conf, int group,
6177                                  struct r5worker *worker,
6178                                  struct list_head *temp_inactive_list)
6179 {
6180         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6181         int i, batch_size = 0, hash;
6182         bool release_inactive = false;
6183
6184         while (batch_size < MAX_STRIPE_BATCH &&
6185                         (sh = __get_priority_stripe(conf, group)) != NULL)
6186                 batch[batch_size++] = sh;
6187
6188         if (batch_size == 0) {
6189                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6190                         if (!list_empty(temp_inactive_list + i))
6191                                 break;
6192                 if (i == NR_STRIPE_HASH_LOCKS) {
6193                         spin_unlock_irq(&conf->device_lock);
6194                         log_flush_stripe_to_raid(conf);
6195                         spin_lock_irq(&conf->device_lock);
6196                         return batch_size;
6197                 }
6198                 release_inactive = true;
6199         }
6200         spin_unlock_irq(&conf->device_lock);
6201
6202         release_inactive_stripe_list(conf, temp_inactive_list,
6203                                      NR_STRIPE_HASH_LOCKS);
6204
6205         r5l_flush_stripe_to_raid(conf->log);
6206         if (release_inactive) {
6207                 spin_lock_irq(&conf->device_lock);
6208                 return 0;
6209         }
6210
6211         for (i = 0; i < batch_size; i++)
6212                 handle_stripe(batch[i]);
6213         log_write_stripe_run(conf);
6214
6215         cond_resched();
6216
6217         spin_lock_irq(&conf->device_lock);
6218         for (i = 0; i < batch_size; i++) {
6219                 hash = batch[i]->hash_lock_index;
6220                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6221         }
6222         return batch_size;
6223 }
6224
6225 static void raid5_do_work(struct work_struct *work)
6226 {
6227         struct r5worker *worker = container_of(work, struct r5worker, work);
6228         struct r5worker_group *group = worker->group;
6229         struct r5conf *conf = group->conf;
6230         struct mddev *mddev = conf->mddev;
6231         int group_id = group - conf->worker_groups;
6232         int handled;
6233         struct blk_plug plug;
6234
6235         pr_debug("+++ raid5worker active\n");
6236
6237         blk_start_plug(&plug);
6238         handled = 0;
6239         spin_lock_irq(&conf->device_lock);
6240         while (1) {
6241                 int batch_size, released;
6242
6243                 released = release_stripe_list(conf, worker->temp_inactive_list);
6244
6245                 batch_size = handle_active_stripes(conf, group_id, worker,
6246                                                    worker->temp_inactive_list);
6247                 worker->working = false;
6248                 if (!batch_size && !released)
6249                         break;
6250                 handled += batch_size;
6251                 wait_event_lock_irq(mddev->sb_wait,
6252                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6253                         conf->device_lock);
6254         }
6255         pr_debug("%d stripes handled\n", handled);
6256
6257         spin_unlock_irq(&conf->device_lock);
6258
6259         flush_deferred_bios(conf);
6260
6261         r5l_flush_stripe_to_raid(conf->log);
6262
6263         async_tx_issue_pending_all();
6264         blk_finish_plug(&plug);
6265
6266         pr_debug("--- raid5worker inactive\n");
6267 }
6268
6269 /*
6270  * This is our raid5 kernel thread.
6271  *
6272  * We scan the hash table for stripes which can be handled now.
6273  * During the scan, completed stripes are saved for us by the interrupt
6274  * handler, so that they will not have to wait for our next wakeup.
6275  */
6276 static void raid5d(struct md_thread *thread)
6277 {
6278         struct mddev *mddev = thread->mddev;
6279         struct r5conf *conf = mddev->private;
6280         int handled;
6281         struct blk_plug plug;
6282
6283         pr_debug("+++ raid5d active\n");
6284
6285         md_check_recovery(mddev);
6286
6287         blk_start_plug(&plug);
6288         handled = 0;
6289         spin_lock_irq(&conf->device_lock);
6290         while (1) {
6291                 struct bio *bio;
6292                 int batch_size, released;
6293                 unsigned int offset;
6294
6295                 released = release_stripe_list(conf, conf->temp_inactive_list);
6296                 if (released)
6297                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6298
6299                 if (
6300                     !list_empty(&conf->bitmap_list)) {
6301                         /* Now is a good time to flush some bitmap updates */
6302                         conf->seq_flush++;
6303                         spin_unlock_irq(&conf->device_lock);
6304                         md_bitmap_unplug(mddev->bitmap);
6305                         spin_lock_irq(&conf->device_lock);
6306                         conf->seq_write = conf->seq_flush;
6307                         activate_bit_delay(conf, conf->temp_inactive_list);
6308                 }
6309                 raid5_activate_delayed(conf);
6310
6311                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6312                         int ok;
6313                         spin_unlock_irq(&conf->device_lock);
6314                         ok = retry_aligned_read(conf, bio, offset);
6315                         spin_lock_irq(&conf->device_lock);
6316                         if (!ok)
6317                                 break;
6318                         handled++;
6319                 }
6320
6321                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6322                                                    conf->temp_inactive_list);
6323                 if (!batch_size && !released)
6324                         break;
6325                 handled += batch_size;
6326
6327                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6328                         spin_unlock_irq(&conf->device_lock);
6329                         md_check_recovery(mddev);
6330                         spin_lock_irq(&conf->device_lock);
6331                 }
6332         }
6333         pr_debug("%d stripes handled\n", handled);
6334
6335         spin_unlock_irq(&conf->device_lock);
6336         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6337             mutex_trylock(&conf->cache_size_mutex)) {
6338                 grow_one_stripe(conf, __GFP_NOWARN);
6339                 /* Set flag even if allocation failed.  This helps
6340                  * slow down allocation requests when mem is short
6341                  */
6342                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6343                 mutex_unlock(&conf->cache_size_mutex);
6344         }
6345
6346         flush_deferred_bios(conf);
6347
6348         r5l_flush_stripe_to_raid(conf->log);
6349
6350         async_tx_issue_pending_all();
6351         blk_finish_plug(&plug);
6352
6353         pr_debug("--- raid5d inactive\n");
6354 }
6355
6356 static ssize_t
6357 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6358 {
6359         struct r5conf *conf;
6360         int ret = 0;
6361         spin_lock(&mddev->lock);
6362         conf = mddev->private;
6363         if (conf)
6364                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6365         spin_unlock(&mddev->lock);
6366         return ret;
6367 }
6368
6369 int
6370 raid5_set_cache_size(struct mddev *mddev, int size)
6371 {
6372         int result = 0;
6373         struct r5conf *conf = mddev->private;
6374
6375         if (size <= 16 || size > 32768)
6376                 return -EINVAL;
6377
6378         conf->min_nr_stripes = size;
6379         mutex_lock(&conf->cache_size_mutex);
6380         while (size < conf->max_nr_stripes &&
6381                drop_one_stripe(conf))
6382                 ;
6383         mutex_unlock(&conf->cache_size_mutex);
6384
6385         md_allow_write(mddev);
6386
6387         mutex_lock(&conf->cache_size_mutex);
6388         while (size > conf->max_nr_stripes)
6389                 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6390                         conf->min_nr_stripes = conf->max_nr_stripes;
6391                         result = -ENOMEM;
6392                         break;
6393                 }
6394         mutex_unlock(&conf->cache_size_mutex);
6395
6396         return result;
6397 }
6398 EXPORT_SYMBOL(raid5_set_cache_size);
6399
6400 static ssize_t
6401 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6402 {
6403         struct r5conf *conf;
6404         unsigned long new;
6405         int err;
6406
6407         if (len >= PAGE_SIZE)
6408                 return -EINVAL;
6409         if (kstrtoul(page, 10, &new))
6410                 return -EINVAL;
6411         err = mddev_lock(mddev);
6412         if (err)
6413                 return err;
6414         conf = mddev->private;
6415         if (!conf)
6416                 err = -ENODEV;
6417         else
6418                 err = raid5_set_cache_size(mddev, new);
6419         mddev_unlock(mddev);
6420
6421         return err ?: len;
6422 }
6423
6424 static struct md_sysfs_entry
6425 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6426                                 raid5_show_stripe_cache_size,
6427                                 raid5_store_stripe_cache_size);
6428
6429 static ssize_t
6430 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6431 {
6432         struct r5conf *conf = mddev->private;
6433         if (conf)
6434                 return sprintf(page, "%d\n", conf->rmw_level);
6435         else
6436                 return 0;
6437 }
6438
6439 static ssize_t
6440 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6441 {
6442         struct r5conf *conf = mddev->private;
6443         unsigned long new;
6444
6445         if (!conf)
6446                 return -ENODEV;
6447
6448         if (len >= PAGE_SIZE)
6449                 return -EINVAL;
6450
6451         if (kstrtoul(page, 10, &new))
6452                 return -EINVAL;
6453
6454         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6455                 return -EINVAL;
6456
6457         if (new != PARITY_DISABLE_RMW &&
6458             new != PARITY_ENABLE_RMW &&
6459             new != PARITY_PREFER_RMW)
6460                 return -EINVAL;
6461
6462         conf->rmw_level = new;
6463         return len;
6464 }
6465
6466 static struct md_sysfs_entry
6467 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6468                          raid5_show_rmw_level,
6469                          raid5_store_rmw_level);
6470
6471
6472 static ssize_t
6473 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6474 {
6475         struct r5conf *conf;
6476         int ret = 0;
6477         spin_lock(&mddev->lock);
6478         conf = mddev->private;
6479         if (conf)
6480                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6481         spin_unlock(&mddev->lock);
6482         return ret;
6483 }
6484
6485 static ssize_t
6486 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6487 {
6488         struct r5conf *conf;
6489         unsigned long new;
6490         int err;
6491
6492         if (len >= PAGE_SIZE)
6493                 return -EINVAL;
6494         if (kstrtoul(page, 10, &new))
6495                 return -EINVAL;
6496
6497         err = mddev_lock(mddev);
6498         if (err)
6499                 return err;
6500         conf = mddev->private;
6501         if (!conf)
6502                 err = -ENODEV;
6503         else if (new > conf->min_nr_stripes)
6504                 err = -EINVAL;
6505         else
6506                 conf->bypass_threshold = new;
6507         mddev_unlock(mddev);
6508         return err ?: len;
6509 }
6510
6511 static struct md_sysfs_entry
6512 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6513                                         S_IRUGO | S_IWUSR,
6514                                         raid5_show_preread_threshold,
6515                                         raid5_store_preread_threshold);
6516
6517 static ssize_t
6518 raid5_show_skip_copy(struct mddev *mddev, char *page)
6519 {
6520         struct r5conf *conf;
6521         int ret = 0;
6522         spin_lock(&mddev->lock);
6523         conf = mddev->private;
6524         if (conf)
6525                 ret = sprintf(page, "%d\n", conf->skip_copy);
6526         spin_unlock(&mddev->lock);
6527         return ret;
6528 }
6529
6530 static ssize_t
6531 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6532 {
6533         struct r5conf *conf;
6534         unsigned long new;
6535         int err;
6536
6537         if (len >= PAGE_SIZE)
6538                 return -EINVAL;
6539         if (kstrtoul(page, 10, &new))
6540                 return -EINVAL;
6541         new = !!new;
6542
6543         err = mddev_lock(mddev);
6544         if (err)
6545                 return err;
6546         conf = mddev->private;
6547         if (!conf)
6548                 err = -ENODEV;
6549         else if (new != conf->skip_copy) {
6550                 mddev_suspend(mddev);
6551                 conf->skip_copy = new;
6552                 if (new)
6553                         mddev->queue->backing_dev_info->capabilities |=
6554                                 BDI_CAP_STABLE_WRITES;
6555                 else
6556                         mddev->queue->backing_dev_info->capabilities &=
6557                                 ~BDI_CAP_STABLE_WRITES;
6558                 mddev_resume(mddev);
6559         }
6560         mddev_unlock(mddev);
6561         return err ?: len;
6562 }
6563
6564 static struct md_sysfs_entry
6565 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6566                                         raid5_show_skip_copy,
6567                                         raid5_store_skip_copy);
6568
6569 static ssize_t
6570 stripe_cache_active_show(struct mddev *mddev, char *page)
6571 {
6572         struct r5conf *conf = mddev->private;
6573         if (conf)
6574                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6575         else
6576                 return 0;
6577 }
6578
6579 static struct md_sysfs_entry
6580 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6581
6582 static ssize_t
6583 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6584 {
6585         struct r5conf *conf;
6586         int ret = 0;
6587         spin_lock(&mddev->lock);
6588         conf = mddev->private;
6589         if (conf)
6590                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6591         spin_unlock(&mddev->lock);
6592         return ret;
6593 }
6594
6595 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6596                                int *group_cnt,
6597                                int *worker_cnt_per_group,
6598                                struct r5worker_group **worker_groups);
6599 static ssize_t
6600 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6601 {
6602         struct r5conf *conf;
6603         unsigned int new;
6604         int err;
6605         struct r5worker_group *new_groups, *old_groups;
6606         int group_cnt, worker_cnt_per_group;
6607
6608         if (len >= PAGE_SIZE)
6609                 return -EINVAL;
6610         if (kstrtouint(page, 10, &new))
6611                 return -EINVAL;
6612         /* 8192 should be big enough */
6613         if (new > 8192)
6614                 return -EINVAL;
6615
6616         err = mddev_lock(mddev);
6617         if (err)
6618                 return err;
6619         conf = mddev->private;
6620         if (!conf)
6621                 err = -ENODEV;
6622         else if (new != conf->worker_cnt_per_group) {
6623                 mddev_suspend(mddev);
6624
6625                 old_groups = conf->worker_groups;
6626                 if (old_groups)
6627                         flush_workqueue(raid5_wq);
6628
6629                 err = alloc_thread_groups(conf, new,
6630                                           &group_cnt, &worker_cnt_per_group,
6631                                           &new_groups);
6632                 if (!err) {
6633                         spin_lock_irq(&conf->device_lock);
6634                         conf->group_cnt = group_cnt;
6635                         conf->worker_cnt_per_group = worker_cnt_per_group;
6636                         conf->worker_groups = new_groups;
6637                         spin_unlock_irq(&conf->device_lock);
6638
6639                         if (old_groups)
6640                                 kfree(old_groups[0].workers);
6641                         kfree(old_groups);
6642                 }
6643                 mddev_resume(mddev);
6644         }
6645         mddev_unlock(mddev);
6646
6647         return err ?: len;
6648 }
6649
6650 static struct md_sysfs_entry
6651 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6652                                 raid5_show_group_thread_cnt,
6653                                 raid5_store_group_thread_cnt);
6654
6655 static struct attribute *raid5_attrs[] =  {
6656         &raid5_stripecache_size.attr,
6657         &raid5_stripecache_active.attr,
6658         &raid5_preread_bypass_threshold.attr,
6659         &raid5_group_thread_cnt.attr,
6660         &raid5_skip_copy.attr,
6661         &raid5_rmw_level.attr,
6662         &r5c_journal_mode.attr,
6663         NULL,
6664 };
6665 static struct attribute_group raid5_attrs_group = {
6666         .name = NULL,
6667         .attrs = raid5_attrs,
6668 };
6669
6670 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6671                                int *group_cnt,
6672                                int *worker_cnt_per_group,
6673                                struct r5worker_group **worker_groups)
6674 {
6675         int i, j, k;
6676         ssize_t size;
6677         struct r5worker *workers;
6678
6679         *worker_cnt_per_group = cnt;
6680         if (cnt == 0) {
6681                 *group_cnt = 0;
6682                 *worker_groups = NULL;
6683                 return 0;
6684         }
6685         *group_cnt = num_possible_nodes();
6686         size = sizeof(struct r5worker) * cnt;
6687         workers = kcalloc(size, *group_cnt, GFP_NOIO);
6688         *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6689                                  GFP_NOIO);
6690         if (!*worker_groups || !workers) {
6691                 kfree(workers);
6692                 kfree(*worker_groups);
6693                 return -ENOMEM;
6694         }
6695
6696         for (i = 0; i < *group_cnt; i++) {
6697                 struct r5worker_group *group;
6698
6699                 group = &(*worker_groups)[i];
6700                 INIT_LIST_HEAD(&group->handle_list);
6701                 INIT_LIST_HEAD(&group->loprio_list);
6702                 group->conf = conf;
6703                 group->workers = workers + i * cnt;
6704
6705                 for (j = 0; j < cnt; j++) {
6706                         struct r5worker *worker = group->workers + j;
6707                         worker->group = group;
6708                         INIT_WORK(&worker->work, raid5_do_work);
6709
6710                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6711                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6712                 }
6713         }
6714
6715         return 0;
6716 }
6717
6718 static void free_thread_groups(struct r5conf *conf)
6719 {
6720         if (conf->worker_groups)
6721                 kfree(conf->worker_groups[0].workers);
6722         kfree(conf->worker_groups);
6723         conf->worker_groups = NULL;
6724 }
6725
6726 static sector_t
6727 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6728 {
6729         struct r5conf *conf = mddev->private;
6730
6731         if (!sectors)
6732                 sectors = mddev->dev_sectors;
6733         if (!raid_disks)
6734                 /* size is defined by the smallest of previous and new size */
6735                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6736
6737         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6738         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6739         return sectors * (raid_disks - conf->max_degraded);
6740 }
6741
6742 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6743 {
6744         safe_put_page(percpu->spare_page);
6745         if (percpu->scribble)
6746                 flex_array_free(percpu->scribble);
6747         percpu->spare_page = NULL;
6748         percpu->scribble = NULL;
6749 }
6750
6751 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6752 {
6753         if (conf->level == 6 && !percpu->spare_page)
6754                 percpu->spare_page = alloc_page(GFP_KERNEL);
6755         if (!percpu->scribble)
6756                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6757                                                       conf->previous_raid_disks),
6758                                                   max(conf->chunk_sectors,
6759                                                       conf->prev_chunk_sectors)
6760                                                    / STRIPE_SECTORS,
6761                                                   GFP_KERNEL);
6762
6763         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6764                 free_scratch_buffer(conf, percpu);
6765                 return -ENOMEM;
6766         }
6767
6768         return 0;
6769 }
6770
6771 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6772 {
6773         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6774
6775         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6776         return 0;
6777 }
6778
6779 static void raid5_free_percpu(struct r5conf *conf)
6780 {
6781         if (!conf->percpu)
6782                 return;
6783
6784         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6785         free_percpu(conf->percpu);
6786 }
6787
6788 static void free_conf(struct r5conf *conf)
6789 {
6790         int i;
6791
6792         log_exit(conf);
6793
6794         unregister_shrinker(&conf->shrinker);
6795         free_thread_groups(conf);
6796         shrink_stripes(conf);
6797         raid5_free_percpu(conf);
6798         for (i = 0; i < conf->pool_size; i++)
6799                 if (conf->disks[i].extra_page)
6800                         put_page(conf->disks[i].extra_page);
6801         kfree(conf->disks);
6802         bioset_exit(&conf->bio_split);
6803         kfree(conf->stripe_hashtbl);
6804         kfree(conf->pending_data);
6805         kfree(conf);
6806 }
6807
6808 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6809 {
6810         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6811         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6812
6813         if (alloc_scratch_buffer(conf, percpu)) {
6814                 pr_warn("%s: failed memory allocation for cpu%u\n",
6815                         __func__, cpu);
6816                 return -ENOMEM;
6817         }
6818         return 0;
6819 }
6820
6821 static int raid5_alloc_percpu(struct r5conf *conf)
6822 {
6823         int err = 0;
6824
6825         conf->percpu = alloc_percpu(struct raid5_percpu);
6826         if (!conf->percpu)
6827                 return -ENOMEM;
6828
6829         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6830         if (!err) {
6831                 conf->scribble_disks = max(conf->raid_disks,
6832                         conf->previous_raid_disks);
6833                 conf->scribble_sectors = max(conf->chunk_sectors,
6834                         conf->prev_chunk_sectors);
6835         }
6836         return err;
6837 }
6838
6839 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6840                                       struct shrink_control *sc)
6841 {
6842         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6843         unsigned long ret = SHRINK_STOP;
6844
6845         if (mutex_trylock(&conf->cache_size_mutex)) {
6846                 ret= 0;
6847                 while (ret < sc->nr_to_scan &&
6848                        conf->max_nr_stripes > conf->min_nr_stripes) {
6849                         if (drop_one_stripe(conf) == 0) {
6850                                 ret = SHRINK_STOP;
6851                                 break;
6852                         }
6853                         ret++;
6854                 }
6855                 mutex_unlock(&conf->cache_size_mutex);
6856         }
6857         return ret;
6858 }
6859
6860 static unsigned long raid5_cache_count(struct shrinker *shrink,
6861                                        struct shrink_control *sc)
6862 {
6863         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6864
6865         if (conf->max_nr_stripes < conf->min_nr_stripes)
6866                 /* unlikely, but not impossible */
6867                 return 0;
6868         return conf->max_nr_stripes - conf->min_nr_stripes;
6869 }
6870
6871 static struct r5conf *setup_conf(struct mddev *mddev)
6872 {
6873         struct r5conf *conf;
6874         int raid_disk, memory, max_disks;
6875         struct md_rdev *rdev;
6876         struct disk_info *disk;
6877         char pers_name[6];
6878         int i;
6879         int group_cnt, worker_cnt_per_group;
6880         struct r5worker_group *new_group;
6881         int ret;
6882
6883         if (mddev->new_level != 5
6884             && mddev->new_level != 4
6885             && mddev->new_level != 6) {
6886                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6887                         mdname(mddev), mddev->new_level);
6888                 return ERR_PTR(-EIO);
6889         }
6890         if ((mddev->new_level == 5
6891              && !algorithm_valid_raid5(mddev->new_layout)) ||
6892             (mddev->new_level == 6
6893              && !algorithm_valid_raid6(mddev->new_layout))) {
6894                 pr_warn("md/raid:%s: layout %d not supported\n",
6895                         mdname(mddev), mddev->new_layout);
6896                 return ERR_PTR(-EIO);
6897         }
6898         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6899                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6900                         mdname(mddev), mddev->raid_disks);
6901                 return ERR_PTR(-EINVAL);
6902         }
6903
6904         if (!mddev->new_chunk_sectors ||
6905             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6906             !is_power_of_2(mddev->new_chunk_sectors)) {
6907                 pr_warn("md/raid:%s: invalid chunk size %d\n",
6908                         mdname(mddev), mddev->new_chunk_sectors << 9);
6909                 return ERR_PTR(-EINVAL);
6910         }
6911
6912         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6913         if (conf == NULL)
6914                 goto abort;
6915         INIT_LIST_HEAD(&conf->free_list);
6916         INIT_LIST_HEAD(&conf->pending_list);
6917         conf->pending_data = kcalloc(PENDING_IO_MAX,
6918                                      sizeof(struct r5pending_data),
6919                                      GFP_KERNEL);
6920         if (!conf->pending_data)
6921                 goto abort;
6922         for (i = 0; i < PENDING_IO_MAX; i++)
6923                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6924         /* Don't enable multi-threading by default*/
6925         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6926                                  &new_group)) {
6927                 conf->group_cnt = group_cnt;
6928                 conf->worker_cnt_per_group = worker_cnt_per_group;
6929                 conf->worker_groups = new_group;
6930         } else
6931                 goto abort;
6932         spin_lock_init(&conf->device_lock);
6933         seqcount_init(&conf->gen_lock);
6934         mutex_init(&conf->cache_size_mutex);
6935         init_waitqueue_head(&conf->wait_for_quiescent);
6936         init_waitqueue_head(&conf->wait_for_stripe);
6937         init_waitqueue_head(&conf->wait_for_overlap);
6938         INIT_LIST_HEAD(&conf->handle_list);
6939         INIT_LIST_HEAD(&conf->loprio_list);
6940         INIT_LIST_HEAD(&conf->hold_list);
6941         INIT_LIST_HEAD(&conf->delayed_list);
6942         INIT_LIST_HEAD(&conf->bitmap_list);
6943         init_llist_head(&conf->released_stripes);
6944         atomic_set(&conf->active_stripes, 0);
6945         atomic_set(&conf->preread_active_stripes, 0);
6946         atomic_set(&conf->active_aligned_reads, 0);
6947         spin_lock_init(&conf->pending_bios_lock);
6948         conf->batch_bio_dispatch = true;
6949         rdev_for_each(rdev, mddev) {
6950                 if (test_bit(Journal, &rdev->flags))
6951                         continue;
6952                 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6953                         conf->batch_bio_dispatch = false;
6954                         break;
6955                 }
6956         }
6957
6958         conf->bypass_threshold = BYPASS_THRESHOLD;
6959         conf->recovery_disabled = mddev->recovery_disabled - 1;
6960
6961         conf->raid_disks = mddev->raid_disks;
6962         if (mddev->reshape_position == MaxSector)
6963                 conf->previous_raid_disks = mddev->raid_disks;
6964         else
6965                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6966         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6967
6968         conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
6969                               GFP_KERNEL);
6970
6971         if (!conf->disks)
6972                 goto abort;
6973
6974         for (i = 0; i < max_disks; i++) {
6975                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6976                 if (!conf->disks[i].extra_page)
6977                         goto abort;
6978         }
6979
6980         ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
6981         if (ret)
6982                 goto abort;
6983         conf->mddev = mddev;
6984
6985         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6986                 goto abort;
6987
6988         /* We init hash_locks[0] separately to that it can be used
6989          * as the reference lock in the spin_lock_nest_lock() call
6990          * in lock_all_device_hash_locks_irq in order to convince
6991          * lockdep that we know what we are doing.
6992          */
6993         spin_lock_init(conf->hash_locks);
6994         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6995                 spin_lock_init(conf->hash_locks + i);
6996
6997         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6998                 INIT_LIST_HEAD(conf->inactive_list + i);
6999
7000         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7001                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7002
7003         atomic_set(&conf->r5c_cached_full_stripes, 0);
7004         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7005         atomic_set(&conf->r5c_cached_partial_stripes, 0);
7006         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7007         atomic_set(&conf->r5c_flushing_full_stripes, 0);
7008         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7009
7010         conf->level = mddev->new_level;
7011         conf->chunk_sectors = mddev->new_chunk_sectors;
7012         if (raid5_alloc_percpu(conf) != 0)
7013                 goto abort;
7014
7015         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7016
7017         rdev_for_each(rdev, mddev) {
7018                 raid_disk = rdev->raid_disk;
7019                 if (raid_disk >= max_disks
7020                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7021                         continue;
7022                 disk = conf->disks + raid_disk;
7023
7024                 if (test_bit(Replacement, &rdev->flags)) {
7025                         if (disk->replacement)
7026                                 goto abort;
7027                         disk->replacement = rdev;
7028                 } else {
7029                         if (disk->rdev)
7030                                 goto abort;
7031                         disk->rdev = rdev;
7032                 }
7033
7034                 if (test_bit(In_sync, &rdev->flags)) {
7035                         char b[BDEVNAME_SIZE];
7036                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7037                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7038                 } else if (rdev->saved_raid_disk != raid_disk)
7039                         /* Cannot rely on bitmap to complete recovery */
7040                         conf->fullsync = 1;
7041         }
7042
7043         conf->level = mddev->new_level;
7044         if (conf->level == 6) {
7045                 conf->max_degraded = 2;
7046                 if (raid6_call.xor_syndrome)
7047                         conf->rmw_level = PARITY_ENABLE_RMW;
7048                 else
7049                         conf->rmw_level = PARITY_DISABLE_RMW;
7050         } else {
7051                 conf->max_degraded = 1;
7052                 conf->rmw_level = PARITY_ENABLE_RMW;
7053         }
7054         conf->algorithm = mddev->new_layout;
7055         conf->reshape_progress = mddev->reshape_position;
7056         if (conf->reshape_progress != MaxSector) {
7057                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7058                 conf->prev_algo = mddev->layout;
7059         } else {
7060                 conf->prev_chunk_sectors = conf->chunk_sectors;
7061                 conf->prev_algo = conf->algorithm;
7062         }
7063
7064         conf->min_nr_stripes = NR_STRIPES;
7065         if (mddev->reshape_position != MaxSector) {
7066                 int stripes = max_t(int,
7067                         ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7068                         ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7069                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7070                 if (conf->min_nr_stripes != NR_STRIPES)
7071                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7072                                 mdname(mddev), conf->min_nr_stripes);
7073         }
7074         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7075                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7076         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7077         if (grow_stripes(conf, conf->min_nr_stripes)) {
7078                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7079                         mdname(mddev), memory);
7080                 goto abort;
7081         } else
7082                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7083         /*
7084          * Losing a stripe head costs more than the time to refill it,
7085          * it reduces the queue depth and so can hurt throughput.
7086          * So set it rather large, scaled by number of devices.
7087          */
7088         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7089         conf->shrinker.scan_objects = raid5_cache_scan;
7090         conf->shrinker.count_objects = raid5_cache_count;
7091         conf->shrinker.batch = 128;
7092         conf->shrinker.flags = 0;
7093         if (register_shrinker(&conf->shrinker)) {
7094                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7095                         mdname(mddev));
7096                 goto abort;
7097         }
7098
7099         sprintf(pers_name, "raid%d", mddev->new_level);
7100         conf->thread = md_register_thread(raid5d, mddev, pers_name);
7101         if (!conf->thread) {
7102                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7103                         mdname(mddev));
7104                 goto abort;
7105         }
7106
7107         return conf;
7108
7109  abort:
7110         if (conf) {
7111                 free_conf(conf);
7112                 return ERR_PTR(-EIO);
7113         } else
7114                 return ERR_PTR(-ENOMEM);
7115 }
7116
7117 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7118 {
7119         switch (algo) {
7120         case ALGORITHM_PARITY_0:
7121                 if (raid_disk < max_degraded)
7122                         return 1;
7123                 break;
7124         case ALGORITHM_PARITY_N:
7125                 if (raid_disk >= raid_disks - max_degraded)
7126                         return 1;
7127                 break;
7128         case ALGORITHM_PARITY_0_6:
7129                 if (raid_disk == 0 ||
7130                     raid_disk == raid_disks - 1)
7131                         return 1;
7132                 break;
7133         case ALGORITHM_LEFT_ASYMMETRIC_6:
7134         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7135         case ALGORITHM_LEFT_SYMMETRIC_6:
7136         case ALGORITHM_RIGHT_SYMMETRIC_6:
7137                 if (raid_disk == raid_disks - 1)
7138                         return 1;
7139         }
7140         return 0;
7141 }
7142
7143 static int raid5_run(struct mddev *mddev)
7144 {
7145         struct r5conf *conf;
7146         int working_disks = 0;
7147         int dirty_parity_disks = 0;
7148         struct md_rdev *rdev;
7149         struct md_rdev *journal_dev = NULL;
7150         sector_t reshape_offset = 0;
7151         int i;
7152         long long min_offset_diff = 0;
7153         int first = 1;
7154
7155         if (mddev_init_writes_pending(mddev) < 0)
7156                 return -ENOMEM;
7157
7158         if (mddev->recovery_cp != MaxSector)
7159                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7160                           mdname(mddev));
7161
7162         rdev_for_each(rdev, mddev) {
7163                 long long diff;
7164
7165                 if (test_bit(Journal, &rdev->flags)) {
7166                         journal_dev = rdev;
7167                         continue;
7168                 }
7169                 if (rdev->raid_disk < 0)
7170                         continue;
7171                 diff = (rdev->new_data_offset - rdev->data_offset);
7172                 if (first) {
7173                         min_offset_diff = diff;
7174                         first = 0;
7175                 } else if (mddev->reshape_backwards &&
7176                          diff < min_offset_diff)
7177                         min_offset_diff = diff;
7178                 else if (!mddev->reshape_backwards &&
7179                          diff > min_offset_diff)
7180                         min_offset_diff = diff;
7181         }
7182
7183         if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7184             (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7185                 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7186                           mdname(mddev));
7187                 return -EINVAL;
7188         }
7189
7190         if (mddev->reshape_position != MaxSector) {
7191                 /* Check that we can continue the reshape.
7192                  * Difficulties arise if the stripe we would write to
7193                  * next is at or after the stripe we would read from next.
7194                  * For a reshape that changes the number of devices, this
7195                  * is only possible for a very short time, and mdadm makes
7196                  * sure that time appears to have past before assembling
7197                  * the array.  So we fail if that time hasn't passed.
7198                  * For a reshape that keeps the number of devices the same
7199                  * mdadm must be monitoring the reshape can keeping the
7200                  * critical areas read-only and backed up.  It will start
7201                  * the array in read-only mode, so we check for that.
7202                  */
7203                 sector_t here_new, here_old;
7204                 int old_disks;
7205                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7206                 int chunk_sectors;
7207                 int new_data_disks;
7208
7209                 if (journal_dev) {
7210                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7211                                 mdname(mddev));
7212                         return -EINVAL;
7213                 }
7214
7215                 if (mddev->new_level != mddev->level) {
7216                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7217                                 mdname(mddev));
7218                         return -EINVAL;
7219                 }
7220                 old_disks = mddev->raid_disks - mddev->delta_disks;
7221                 /* reshape_position must be on a new-stripe boundary, and one
7222                  * further up in new geometry must map after here in old
7223                  * geometry.
7224                  * If the chunk sizes are different, then as we perform reshape
7225                  * in units of the largest of the two, reshape_position needs
7226                  * be a multiple of the largest chunk size times new data disks.
7227                  */
7228                 here_new = mddev->reshape_position;
7229                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7230                 new_data_disks = mddev->raid_disks - max_degraded;
7231                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7232                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7233                                 mdname(mddev));
7234                         return -EINVAL;
7235                 }
7236                 reshape_offset = here_new * chunk_sectors;
7237                 /* here_new is the stripe we will write to */
7238                 here_old = mddev->reshape_position;
7239                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7240                 /* here_old is the first stripe that we might need to read
7241                  * from */
7242                 if (mddev->delta_disks == 0) {
7243                         /* We cannot be sure it is safe to start an in-place
7244                          * reshape.  It is only safe if user-space is monitoring
7245                          * and taking constant backups.
7246                          * mdadm always starts a situation like this in
7247                          * readonly mode so it can take control before
7248                          * allowing any writes.  So just check for that.
7249                          */
7250                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7251                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7252                                 /* not really in-place - so OK */;
7253                         else if (mddev->ro == 0) {
7254                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7255                                         mdname(mddev));
7256                                 return -EINVAL;
7257                         }
7258                 } else if (mddev->reshape_backwards
7259                     ? (here_new * chunk_sectors + min_offset_diff <=
7260                        here_old * chunk_sectors)
7261                     : (here_new * chunk_sectors >=
7262                        here_old * chunk_sectors + (-min_offset_diff))) {
7263                         /* Reading from the same stripe as writing to - bad */
7264                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7265                                 mdname(mddev));
7266                         return -EINVAL;
7267                 }
7268                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7269                 /* OK, we should be able to continue; */
7270         } else {
7271                 BUG_ON(mddev->level != mddev->new_level);
7272                 BUG_ON(mddev->layout != mddev->new_layout);
7273                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7274                 BUG_ON(mddev->delta_disks != 0);
7275         }
7276
7277         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7278             test_bit(MD_HAS_PPL, &mddev->flags)) {
7279                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7280                         mdname(mddev));
7281                 clear_bit(MD_HAS_PPL, &mddev->flags);
7282                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7283         }
7284
7285         if (mddev->private == NULL)
7286                 conf = setup_conf(mddev);
7287         else
7288                 conf = mddev->private;
7289
7290         if (IS_ERR(conf))
7291                 return PTR_ERR(conf);
7292
7293         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7294                 if (!journal_dev) {
7295                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7296                                 mdname(mddev));
7297                         mddev->ro = 1;
7298                         set_disk_ro(mddev->gendisk, 1);
7299                 } else if (mddev->recovery_cp == MaxSector)
7300                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7301         }
7302
7303         conf->min_offset_diff = min_offset_diff;
7304         mddev->thread = conf->thread;
7305         conf->thread = NULL;
7306         mddev->private = conf;
7307
7308         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7309              i++) {
7310                 rdev = conf->disks[i].rdev;
7311                 if (!rdev && conf->disks[i].replacement) {
7312                         /* The replacement is all we have yet */
7313                         rdev = conf->disks[i].replacement;
7314                         conf->disks[i].replacement = NULL;
7315                         clear_bit(Replacement, &rdev->flags);
7316                         conf->disks[i].rdev = rdev;
7317                 }
7318                 if (!rdev)
7319                         continue;
7320                 if (conf->disks[i].replacement &&
7321                     conf->reshape_progress != MaxSector) {
7322                         /* replacements and reshape simply do not mix. */
7323                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7324                         goto abort;
7325                 }
7326                 if (test_bit(In_sync, &rdev->flags)) {
7327                         working_disks++;
7328                         continue;
7329                 }
7330                 /* This disc is not fully in-sync.  However if it
7331                  * just stored parity (beyond the recovery_offset),
7332                  * when we don't need to be concerned about the
7333                  * array being dirty.
7334                  * When reshape goes 'backwards', we never have
7335                  * partially completed devices, so we only need
7336                  * to worry about reshape going forwards.
7337                  */
7338                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7339                 if (mddev->major_version == 0 &&
7340                     mddev->minor_version > 90)
7341                         rdev->recovery_offset = reshape_offset;
7342
7343                 if (rdev->recovery_offset < reshape_offset) {
7344                         /* We need to check old and new layout */
7345                         if (!only_parity(rdev->raid_disk,
7346                                          conf->algorithm,
7347                                          conf->raid_disks,
7348                                          conf->max_degraded))
7349                                 continue;
7350                 }
7351                 if (!only_parity(rdev->raid_disk,
7352                                  conf->prev_algo,
7353                                  conf->previous_raid_disks,
7354                                  conf->max_degraded))
7355                         continue;
7356                 dirty_parity_disks++;
7357         }
7358
7359         /*
7360          * 0 for a fully functional array, 1 or 2 for a degraded array.
7361          */
7362         mddev->degraded = raid5_calc_degraded(conf);
7363
7364         if (has_failed(conf)) {
7365                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7366                         mdname(mddev), mddev->degraded, conf->raid_disks);
7367                 goto abort;
7368         }
7369
7370         /* device size must be a multiple of chunk size */
7371         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7372         mddev->resync_max_sectors = mddev->dev_sectors;
7373
7374         if (mddev->degraded > dirty_parity_disks &&
7375             mddev->recovery_cp != MaxSector) {
7376                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7377                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7378                                 mdname(mddev));
7379                 else if (mddev->ok_start_degraded)
7380                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7381                                 mdname(mddev));
7382                 else {
7383                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7384                                 mdname(mddev));
7385                         goto abort;
7386                 }
7387         }
7388
7389         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7390                 mdname(mddev), conf->level,
7391                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7392                 mddev->new_layout);
7393
7394         print_raid5_conf(conf);
7395
7396         if (conf->reshape_progress != MaxSector) {
7397                 conf->reshape_safe = conf->reshape_progress;
7398                 atomic_set(&conf->reshape_stripes, 0);
7399                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7400                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7401                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7402                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7403                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7404                                                         "reshape");
7405         }
7406
7407         /* Ok, everything is just fine now */
7408         if (mddev->to_remove == &raid5_attrs_group)
7409                 mddev->to_remove = NULL;
7410         else if (mddev->kobj.sd &&
7411             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7412                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7413                         mdname(mddev));
7414         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7415
7416         if (mddev->queue) {
7417                 int chunk_size;
7418                 /* read-ahead size must cover two whole stripes, which
7419                  * is 2 * (datadisks) * chunksize where 'n' is the
7420                  * number of raid devices
7421                  */
7422                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7423                 int stripe = data_disks *
7424                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7425                 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7426                         mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7427
7428                 chunk_size = mddev->chunk_sectors << 9;
7429                 blk_queue_io_min(mddev->queue, chunk_size);
7430                 blk_queue_io_opt(mddev->queue, chunk_size *
7431                                  (conf->raid_disks - conf->max_degraded));
7432                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7433                 /*
7434                  * We can only discard a whole stripe. It doesn't make sense to
7435                  * discard data disk but write parity disk
7436                  */
7437                 stripe = stripe * PAGE_SIZE;
7438                 /* Round up to power of 2, as discard handling
7439                  * currently assumes that */
7440                 while ((stripe-1) & stripe)
7441                         stripe = (stripe | (stripe-1)) + 1;
7442                 mddev->queue->limits.discard_alignment = stripe;
7443                 mddev->queue->limits.discard_granularity = stripe;
7444
7445                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7446                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7447
7448                 rdev_for_each(rdev, mddev) {
7449                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7450                                           rdev->data_offset << 9);
7451                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7452                                           rdev->new_data_offset << 9);
7453                 }
7454
7455                 /*
7456                  * zeroing is required, otherwise data
7457                  * could be lost. Consider a scenario: discard a stripe
7458                  * (the stripe could be inconsistent if
7459                  * discard_zeroes_data is 0); write one disk of the
7460                  * stripe (the stripe could be inconsistent again
7461                  * depending on which disks are used to calculate
7462                  * parity); the disk is broken; The stripe data of this
7463                  * disk is lost.
7464                  *
7465                  * We only allow DISCARD if the sysadmin has confirmed that
7466                  * only safe devices are in use by setting a module parameter.
7467                  * A better idea might be to turn DISCARD into WRITE_ZEROES
7468                  * requests, as that is required to be safe.
7469                  */
7470                 if (devices_handle_discard_safely &&
7471                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7472                     mddev->queue->limits.discard_granularity >= stripe)
7473                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7474                                                 mddev->queue);
7475                 else
7476                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7477                                                 mddev->queue);
7478
7479                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7480         }
7481
7482         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7483                 goto abort;
7484
7485         return 0;
7486 abort:
7487         md_unregister_thread(&mddev->thread);
7488         print_raid5_conf(conf);
7489         free_conf(conf);
7490         mddev->private = NULL;
7491         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7492         return -EIO;
7493 }
7494
7495 static void raid5_free(struct mddev *mddev, void *priv)
7496 {
7497         struct r5conf *conf = priv;
7498
7499         free_conf(conf);
7500         mddev->to_remove = &raid5_attrs_group;
7501 }
7502
7503 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7504 {
7505         struct r5conf *conf = mddev->private;
7506         int i;
7507
7508         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7509                 conf->chunk_sectors / 2, mddev->layout);
7510         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7511         rcu_read_lock();
7512         for (i = 0; i < conf->raid_disks; i++) {
7513                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7514                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7515         }
7516         rcu_read_unlock();
7517         seq_printf (seq, "]");
7518 }
7519
7520 static void print_raid5_conf (struct r5conf *conf)
7521 {
7522         int i;
7523         struct disk_info *tmp;
7524
7525         pr_debug("RAID conf printout:\n");
7526         if (!conf) {
7527                 pr_debug("(conf==NULL)\n");
7528                 return;
7529         }
7530         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7531                conf->raid_disks,
7532                conf->raid_disks - conf->mddev->degraded);
7533
7534         for (i = 0; i < conf->raid_disks; i++) {
7535                 char b[BDEVNAME_SIZE];
7536                 tmp = conf->disks + i;
7537                 if (tmp->rdev)
7538                         pr_debug(" disk %d, o:%d, dev:%s\n",
7539                                i, !test_bit(Faulty, &tmp->rdev->flags),
7540                                bdevname(tmp->rdev->bdev, b));
7541         }
7542 }
7543
7544 static int raid5_spare_active(struct mddev *mddev)
7545 {
7546         int i;
7547         struct r5conf *conf = mddev->private;
7548         struct disk_info *tmp;
7549         int count = 0;
7550         unsigned long flags;
7551
7552         for (i = 0; i < conf->raid_disks; i++) {
7553                 tmp = conf->disks + i;
7554                 if (tmp->replacement
7555                     && tmp->replacement->recovery_offset == MaxSector
7556                     && !test_bit(Faulty, &tmp->replacement->flags)
7557                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7558                         /* Replacement has just become active. */
7559                         if (!tmp->rdev
7560                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7561                                 count++;
7562                         if (tmp->rdev) {
7563                                 /* Replaced device not technically faulty,
7564                                  * but we need to be sure it gets removed
7565                                  * and never re-added.
7566                                  */
7567                                 set_bit(Faulty, &tmp->rdev->flags);
7568                                 sysfs_notify_dirent_safe(
7569                                         tmp->rdev->sysfs_state);
7570                         }
7571                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7572                 } else if (tmp->rdev
7573                     && tmp->rdev->recovery_offset == MaxSector
7574                     && !test_bit(Faulty, &tmp->rdev->flags)
7575                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7576                         count++;
7577                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7578                 }
7579         }
7580         spin_lock_irqsave(&conf->device_lock, flags);
7581         mddev->degraded = raid5_calc_degraded(conf);
7582         spin_unlock_irqrestore(&conf->device_lock, flags);
7583         print_raid5_conf(conf);
7584         return count;
7585 }
7586
7587 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7588 {
7589         struct r5conf *conf = mddev->private;
7590         int err = 0;
7591         int number = rdev->raid_disk;
7592         struct md_rdev **rdevp;
7593         struct disk_info *p = conf->disks + number;
7594
7595         print_raid5_conf(conf);
7596         if (test_bit(Journal, &rdev->flags) && conf->log) {
7597                 /*
7598                  * we can't wait pending write here, as this is called in
7599                  * raid5d, wait will deadlock.
7600                  * neilb: there is no locking about new writes here,
7601                  * so this cannot be safe.
7602                  */
7603                 if (atomic_read(&conf->active_stripes) ||
7604                     atomic_read(&conf->r5c_cached_full_stripes) ||
7605                     atomic_read(&conf->r5c_cached_partial_stripes)) {
7606                         return -EBUSY;
7607                 }
7608                 log_exit(conf);
7609                 return 0;
7610         }
7611         if (rdev == p->rdev)
7612                 rdevp = &p->rdev;
7613         else if (rdev == p->replacement)
7614                 rdevp = &p->replacement;
7615         else
7616                 return 0;
7617
7618         if (number >= conf->raid_disks &&
7619             conf->reshape_progress == MaxSector)
7620                 clear_bit(In_sync, &rdev->flags);
7621
7622         if (test_bit(In_sync, &rdev->flags) ||
7623             atomic_read(&rdev->nr_pending)) {
7624                 err = -EBUSY;
7625                 goto abort;
7626         }
7627         /* Only remove non-faulty devices if recovery
7628          * isn't possible.
7629          */
7630         if (!test_bit(Faulty, &rdev->flags) &&
7631             mddev->recovery_disabled != conf->recovery_disabled &&
7632             !has_failed(conf) &&
7633             (!p->replacement || p->replacement == rdev) &&
7634             number < conf->raid_disks) {
7635                 err = -EBUSY;
7636                 goto abort;
7637         }
7638         *rdevp = NULL;
7639         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7640                 synchronize_rcu();
7641                 if (atomic_read(&rdev->nr_pending)) {
7642                         /* lost the race, try later */
7643                         err = -EBUSY;
7644                         *rdevp = rdev;
7645                 }
7646         }
7647         if (!err) {
7648                 err = log_modify(conf, rdev, false);
7649                 if (err)
7650                         goto abort;
7651         }
7652         if (p->replacement) {
7653                 /* We must have just cleared 'rdev' */
7654                 p->rdev = p->replacement;
7655                 clear_bit(Replacement, &p->replacement->flags);
7656                 smp_mb(); /* Make sure other CPUs may see both as identical
7657                            * but will never see neither - if they are careful
7658                            */
7659                 p->replacement = NULL;
7660
7661                 if (!err)
7662                         err = log_modify(conf, p->rdev, true);
7663         }
7664
7665         clear_bit(WantReplacement, &rdev->flags);
7666 abort:
7667
7668         print_raid5_conf(conf);
7669         return err;
7670 }
7671
7672 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7673 {
7674         struct r5conf *conf = mddev->private;
7675         int err = -EEXIST;
7676         int disk;
7677         struct disk_info *p;
7678         int first = 0;
7679         int last = conf->raid_disks - 1;
7680
7681         if (test_bit(Journal, &rdev->flags)) {
7682                 if (conf->log)
7683                         return -EBUSY;
7684
7685                 rdev->raid_disk = 0;
7686                 /*
7687                  * The array is in readonly mode if journal is missing, so no
7688                  * write requests running. We should be safe
7689                  */
7690                 log_init(conf, rdev, false);
7691                 return 0;
7692         }
7693         if (mddev->recovery_disabled == conf->recovery_disabled)
7694                 return -EBUSY;
7695
7696         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7697                 /* no point adding a device */
7698                 return -EINVAL;
7699
7700         if (rdev->raid_disk >= 0)
7701                 first = last = rdev->raid_disk;
7702
7703         /*
7704          * find the disk ... but prefer rdev->saved_raid_disk
7705          * if possible.
7706          */
7707         if (rdev->saved_raid_disk >= 0 &&
7708             rdev->saved_raid_disk >= first &&
7709             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7710                 first = rdev->saved_raid_disk;
7711
7712         for (disk = first; disk <= last; disk++) {
7713                 p = conf->disks + disk;
7714                 if (p->rdev == NULL) {
7715                         clear_bit(In_sync, &rdev->flags);
7716                         rdev->raid_disk = disk;
7717                         if (rdev->saved_raid_disk != disk)
7718                                 conf->fullsync = 1;
7719                         rcu_assign_pointer(p->rdev, rdev);
7720
7721                         err = log_modify(conf, rdev, true);
7722
7723                         goto out;
7724                 }
7725         }
7726         for (disk = first; disk <= last; disk++) {
7727                 p = conf->disks + disk;
7728                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7729                     p->replacement == NULL) {
7730                         clear_bit(In_sync, &rdev->flags);
7731                         set_bit(Replacement, &rdev->flags);
7732                         rdev->raid_disk = disk;
7733                         err = 0;
7734                         conf->fullsync = 1;
7735                         rcu_assign_pointer(p->replacement, rdev);
7736                         break;
7737                 }
7738         }
7739 out:
7740         print_raid5_conf(conf);
7741         return err;
7742 }
7743
7744 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7745 {
7746         /* no resync is happening, and there is enough space
7747          * on all devices, so we can resize.
7748          * We need to make sure resync covers any new space.
7749          * If the array is shrinking we should possibly wait until
7750          * any io in the removed space completes, but it hardly seems
7751          * worth it.
7752          */
7753         sector_t newsize;
7754         struct r5conf *conf = mddev->private;
7755
7756         if (raid5_has_log(conf) || raid5_has_ppl(conf))
7757                 return -EINVAL;
7758         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7759         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7760         if (mddev->external_size &&
7761             mddev->array_sectors > newsize)
7762                 return -EINVAL;
7763         if (mddev->bitmap) {
7764                 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7765                 if (ret)
7766                         return ret;
7767         }
7768         md_set_array_sectors(mddev, newsize);
7769         if (sectors > mddev->dev_sectors &&
7770             mddev->recovery_cp > mddev->dev_sectors) {
7771                 mddev->recovery_cp = mddev->dev_sectors;
7772                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7773         }
7774         mddev->dev_sectors = sectors;
7775         mddev->resync_max_sectors = sectors;
7776         return 0;
7777 }
7778
7779 static int check_stripe_cache(struct mddev *mddev)
7780 {
7781         /* Can only proceed if there are plenty of stripe_heads.
7782          * We need a minimum of one full stripe,, and for sensible progress
7783          * it is best to have about 4 times that.
7784          * If we require 4 times, then the default 256 4K stripe_heads will
7785          * allow for chunk sizes up to 256K, which is probably OK.
7786          * If the chunk size is greater, user-space should request more
7787          * stripe_heads first.
7788          */
7789         struct r5conf *conf = mddev->private;
7790         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7791             > conf->min_nr_stripes ||
7792             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7793             > conf->min_nr_stripes) {
7794                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7795                         mdname(mddev),
7796                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7797                          / STRIPE_SIZE)*4);
7798                 return 0;
7799         }
7800         return 1;
7801 }
7802
7803 static int check_reshape(struct mddev *mddev)
7804 {
7805         struct r5conf *conf = mddev->private;
7806
7807         if (raid5_has_log(conf) || raid5_has_ppl(conf))
7808                 return -EINVAL;
7809         if (mddev->delta_disks == 0 &&
7810             mddev->new_layout == mddev->layout &&
7811             mddev->new_chunk_sectors == mddev->chunk_sectors)
7812                 return 0; /* nothing to do */
7813         if (has_failed(conf))
7814                 return -EINVAL;
7815         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7816                 /* We might be able to shrink, but the devices must
7817                  * be made bigger first.
7818                  * For raid6, 4 is the minimum size.
7819                  * Otherwise 2 is the minimum
7820                  */
7821                 int min = 2;
7822                 if (mddev->level == 6)
7823                         min = 4;
7824                 if (mddev->raid_disks + mddev->delta_disks < min)
7825                         return -EINVAL;
7826         }
7827
7828         if (!check_stripe_cache(mddev))
7829                 return -ENOSPC;
7830
7831         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7832             mddev->delta_disks > 0)
7833                 if (resize_chunks(conf,
7834                                   conf->previous_raid_disks
7835                                   + max(0, mddev->delta_disks),
7836                                   max(mddev->new_chunk_sectors,
7837                                       mddev->chunk_sectors)
7838                             ) < 0)
7839                         return -ENOMEM;
7840
7841         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7842                 return 0; /* never bother to shrink */
7843         return resize_stripes(conf, (conf->previous_raid_disks
7844                                      + mddev->delta_disks));
7845 }
7846
7847 static int raid5_start_reshape(struct mddev *mddev)
7848 {
7849         struct r5conf *conf = mddev->private;
7850         struct md_rdev *rdev;
7851         int spares = 0;
7852         unsigned long flags;
7853
7854         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7855                 return -EBUSY;
7856
7857         if (!check_stripe_cache(mddev))
7858                 return -ENOSPC;
7859
7860         if (has_failed(conf))
7861                 return -EINVAL;
7862
7863         rdev_for_each(rdev, mddev) {
7864                 if (!test_bit(In_sync, &rdev->flags)
7865                     && !test_bit(Faulty, &rdev->flags))
7866                         spares++;
7867         }
7868
7869         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7870                 /* Not enough devices even to make a degraded array
7871                  * of that size
7872                  */
7873                 return -EINVAL;
7874
7875         /* Refuse to reduce size of the array.  Any reductions in
7876          * array size must be through explicit setting of array_size
7877          * attribute.
7878          */
7879         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7880             < mddev->array_sectors) {
7881                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7882                         mdname(mddev));
7883                 return -EINVAL;
7884         }
7885
7886         atomic_set(&conf->reshape_stripes, 0);
7887         spin_lock_irq(&conf->device_lock);
7888         write_seqcount_begin(&conf->gen_lock);
7889         conf->previous_raid_disks = conf->raid_disks;
7890         conf->raid_disks += mddev->delta_disks;
7891         conf->prev_chunk_sectors = conf->chunk_sectors;
7892         conf->chunk_sectors = mddev->new_chunk_sectors;
7893         conf->prev_algo = conf->algorithm;
7894         conf->algorithm = mddev->new_layout;
7895         conf->generation++;
7896         /* Code that selects data_offset needs to see the generation update
7897          * if reshape_progress has been set - so a memory barrier needed.
7898          */
7899         smp_mb();
7900         if (mddev->reshape_backwards)
7901                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7902         else
7903                 conf->reshape_progress = 0;
7904         conf->reshape_safe = conf->reshape_progress;
7905         write_seqcount_end(&conf->gen_lock);
7906         spin_unlock_irq(&conf->device_lock);
7907
7908         /* Now make sure any requests that proceeded on the assumption
7909          * the reshape wasn't running - like Discard or Read - have
7910          * completed.
7911          */
7912         mddev_suspend(mddev);
7913         mddev_resume(mddev);
7914
7915         /* Add some new drives, as many as will fit.
7916          * We know there are enough to make the newly sized array work.
7917          * Don't add devices if we are reducing the number of
7918          * devices in the array.  This is because it is not possible
7919          * to correctly record the "partially reconstructed" state of
7920          * such devices during the reshape and confusion could result.
7921          */
7922         if (mddev->delta_disks >= 0) {
7923                 rdev_for_each(rdev, mddev)
7924                         if (rdev->raid_disk < 0 &&
7925                             !test_bit(Faulty, &rdev->flags)) {
7926                                 if (raid5_add_disk(mddev, rdev) == 0) {
7927                                         if (rdev->raid_disk
7928                                             >= conf->previous_raid_disks)
7929                                                 set_bit(In_sync, &rdev->flags);
7930                                         else
7931                                                 rdev->recovery_offset = 0;
7932
7933                                         if (sysfs_link_rdev(mddev, rdev))
7934                                                 /* Failure here is OK */;
7935                                 }
7936                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7937                                    && !test_bit(Faulty, &rdev->flags)) {
7938                                 /* This is a spare that was manually added */
7939                                 set_bit(In_sync, &rdev->flags);
7940                         }
7941
7942                 /* When a reshape changes the number of devices,
7943                  * ->degraded is measured against the larger of the
7944                  * pre and post number of devices.
7945                  */
7946                 spin_lock_irqsave(&conf->device_lock, flags);
7947                 mddev->degraded = raid5_calc_degraded(conf);
7948                 spin_unlock_irqrestore(&conf->device_lock, flags);
7949         }
7950         mddev->raid_disks = conf->raid_disks;
7951         mddev->reshape_position = conf->reshape_progress;
7952         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7953
7954         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7955         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7956         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7957         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7958         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7959         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7960                                                 "reshape");
7961         if (!mddev->sync_thread) {
7962                 mddev->recovery = 0;
7963                 spin_lock_irq(&conf->device_lock);
7964                 write_seqcount_begin(&conf->gen_lock);
7965                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7966                 mddev->new_chunk_sectors =
7967                         conf->chunk_sectors = conf->prev_chunk_sectors;
7968                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7969                 rdev_for_each(rdev, mddev)
7970                         rdev->new_data_offset = rdev->data_offset;
7971                 smp_wmb();
7972                 conf->generation --;
7973                 conf->reshape_progress = MaxSector;
7974                 mddev->reshape_position = MaxSector;
7975                 write_seqcount_end(&conf->gen_lock);
7976                 spin_unlock_irq(&conf->device_lock);
7977                 return -EAGAIN;
7978         }
7979         conf->reshape_checkpoint = jiffies;
7980         md_wakeup_thread(mddev->sync_thread);
7981         md_new_event(mddev);
7982         return 0;
7983 }
7984
7985 /* This is called from the reshape thread and should make any
7986  * changes needed in 'conf'
7987  */
7988 static void end_reshape(struct r5conf *conf)
7989 {
7990
7991         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7992                 struct md_rdev *rdev;
7993
7994                 spin_lock_irq(&conf->device_lock);
7995                 conf->previous_raid_disks = conf->raid_disks;
7996                 md_finish_reshape(conf->mddev);
7997                 smp_wmb();
7998                 conf->reshape_progress = MaxSector;
7999                 conf->mddev->reshape_position = MaxSector;
8000                 rdev_for_each(rdev, conf->mddev)
8001                         if (rdev->raid_disk >= 0 &&
8002                             !test_bit(Journal, &rdev->flags) &&
8003                             !test_bit(In_sync, &rdev->flags))
8004                                 rdev->recovery_offset = MaxSector;
8005                 spin_unlock_irq(&conf->device_lock);
8006                 wake_up(&conf->wait_for_overlap);
8007
8008                 /* read-ahead size must cover two whole stripes, which is
8009                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
8010                  */
8011                 if (conf->mddev->queue) {
8012                         int data_disks = conf->raid_disks - conf->max_degraded;
8013                         int stripe = data_disks * ((conf->chunk_sectors << 9)
8014                                                    / PAGE_SIZE);
8015                         if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8016                                 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8017                 }
8018         }
8019 }
8020
8021 /* This is called from the raid5d thread with mddev_lock held.
8022  * It makes config changes to the device.
8023  */
8024 static void raid5_finish_reshape(struct mddev *mddev)
8025 {
8026         struct r5conf *conf = mddev->private;
8027
8028         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8029
8030                 if (mddev->delta_disks <= 0) {
8031                         int d;
8032                         spin_lock_irq(&conf->device_lock);
8033                         mddev->degraded = raid5_calc_degraded(conf);
8034                         spin_unlock_irq(&conf->device_lock);
8035                         for (d = conf->raid_disks ;
8036                              d < conf->raid_disks - mddev->delta_disks;
8037                              d++) {
8038                                 struct md_rdev *rdev = conf->disks[d].rdev;
8039                                 if (rdev)
8040                                         clear_bit(In_sync, &rdev->flags);
8041                                 rdev = conf->disks[d].replacement;
8042                                 if (rdev)
8043                                         clear_bit(In_sync, &rdev->flags);
8044                         }
8045                 }
8046                 mddev->layout = conf->algorithm;
8047                 mddev->chunk_sectors = conf->chunk_sectors;
8048                 mddev->reshape_position = MaxSector;
8049                 mddev->delta_disks = 0;
8050                 mddev->reshape_backwards = 0;
8051         }
8052 }
8053
8054 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8055 {
8056         struct r5conf *conf = mddev->private;
8057
8058         if (quiesce) {
8059                 /* stop all writes */
8060                 lock_all_device_hash_locks_irq(conf);
8061                 /* '2' tells resync/reshape to pause so that all
8062                  * active stripes can drain
8063                  */
8064                 r5c_flush_cache(conf, INT_MAX);
8065                 conf->quiesce = 2;
8066                 wait_event_cmd(conf->wait_for_quiescent,
8067                                     atomic_read(&conf->active_stripes) == 0 &&
8068                                     atomic_read(&conf->active_aligned_reads) == 0,
8069                                     unlock_all_device_hash_locks_irq(conf),
8070                                     lock_all_device_hash_locks_irq(conf));
8071                 conf->quiesce = 1;
8072                 unlock_all_device_hash_locks_irq(conf);
8073                 /* allow reshape to continue */
8074                 wake_up(&conf->wait_for_overlap);
8075         } else {
8076                 /* re-enable writes */
8077                 lock_all_device_hash_locks_irq(conf);
8078                 conf->quiesce = 0;
8079                 wake_up(&conf->wait_for_quiescent);
8080                 wake_up(&conf->wait_for_overlap);
8081                 unlock_all_device_hash_locks_irq(conf);
8082         }
8083         log_quiesce(conf, quiesce);
8084 }
8085
8086 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8087 {
8088         struct r0conf *raid0_conf = mddev->private;
8089         sector_t sectors;
8090
8091         /* for raid0 takeover only one zone is supported */
8092         if (raid0_conf->nr_strip_zones > 1) {
8093                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8094                         mdname(mddev));
8095                 return ERR_PTR(-EINVAL);
8096         }
8097
8098         sectors = raid0_conf->strip_zone[0].zone_end;
8099         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8100         mddev->dev_sectors = sectors;
8101         mddev->new_level = level;
8102         mddev->new_layout = ALGORITHM_PARITY_N;
8103         mddev->new_chunk_sectors = mddev->chunk_sectors;
8104         mddev->raid_disks += 1;
8105         mddev->delta_disks = 1;
8106         /* make sure it will be not marked as dirty */
8107         mddev->recovery_cp = MaxSector;
8108
8109         return setup_conf(mddev);
8110 }
8111
8112 static void *raid5_takeover_raid1(struct mddev *mddev)
8113 {
8114         int chunksect;
8115         void *ret;
8116
8117         if (mddev->raid_disks != 2 ||
8118             mddev->degraded > 1)
8119                 return ERR_PTR(-EINVAL);
8120
8121         /* Should check if there are write-behind devices? */
8122
8123         chunksect = 64*2; /* 64K by default */
8124
8125         /* The array must be an exact multiple of chunksize */
8126         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8127                 chunksect >>= 1;
8128
8129         if ((chunksect<<9) < STRIPE_SIZE)
8130                 /* array size does not allow a suitable chunk size */
8131                 return ERR_PTR(-EINVAL);
8132
8133         mddev->new_level = 5;
8134         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8135         mddev->new_chunk_sectors = chunksect;
8136
8137         ret = setup_conf(mddev);
8138         if (!IS_ERR(ret))
8139                 mddev_clear_unsupported_flags(mddev,
8140                         UNSUPPORTED_MDDEV_FLAGS);
8141         return ret;
8142 }
8143
8144 static void *raid5_takeover_raid6(struct mddev *mddev)
8145 {
8146         int new_layout;
8147
8148         switch (mddev->layout) {
8149         case ALGORITHM_LEFT_ASYMMETRIC_6:
8150                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8151                 break;
8152         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8153                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8154                 break;
8155         case ALGORITHM_LEFT_SYMMETRIC_6:
8156                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8157                 break;
8158         case ALGORITHM_RIGHT_SYMMETRIC_6:
8159                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8160                 break;
8161         case ALGORITHM_PARITY_0_6:
8162                 new_layout = ALGORITHM_PARITY_0;
8163                 break;
8164         case ALGORITHM_PARITY_N:
8165                 new_layout = ALGORITHM_PARITY_N;
8166                 break;
8167         default:
8168                 return ERR_PTR(-EINVAL);
8169         }
8170         mddev->new_level = 5;
8171         mddev->new_layout = new_layout;
8172         mddev->delta_disks = -1;
8173         mddev->raid_disks -= 1;
8174         return setup_conf(mddev);
8175 }
8176
8177 static int raid5_check_reshape(struct mddev *mddev)
8178 {
8179         /* For a 2-drive array, the layout and chunk size can be changed
8180          * immediately as not restriping is needed.
8181          * For larger arrays we record the new value - after validation
8182          * to be used by a reshape pass.
8183          */
8184         struct r5conf *conf = mddev->private;
8185         int new_chunk = mddev->new_chunk_sectors;
8186
8187         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8188                 return -EINVAL;
8189         if (new_chunk > 0) {
8190                 if (!is_power_of_2(new_chunk))
8191                         return -EINVAL;
8192                 if (new_chunk < (PAGE_SIZE>>9))
8193                         return -EINVAL;
8194                 if (mddev->array_sectors & (new_chunk-1))
8195                         /* not factor of array size */
8196                         return -EINVAL;
8197         }
8198
8199         /* They look valid */
8200
8201         if (mddev->raid_disks == 2) {
8202                 /* can make the change immediately */
8203                 if (mddev->new_layout >= 0) {
8204                         conf->algorithm = mddev->new_layout;
8205                         mddev->layout = mddev->new_layout;
8206                 }
8207                 if (new_chunk > 0) {
8208                         conf->chunk_sectors = new_chunk ;
8209                         mddev->chunk_sectors = new_chunk;
8210                 }
8211                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8212                 md_wakeup_thread(mddev->thread);
8213         }
8214         return check_reshape(mddev);
8215 }
8216
8217 static int raid6_check_reshape(struct mddev *mddev)
8218 {
8219         int new_chunk = mddev->new_chunk_sectors;
8220
8221         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8222                 return -EINVAL;
8223         if (new_chunk > 0) {
8224                 if (!is_power_of_2(new_chunk))
8225                         return -EINVAL;
8226                 if (new_chunk < (PAGE_SIZE >> 9))
8227                         return -EINVAL;
8228                 if (mddev->array_sectors & (new_chunk-1))
8229                         /* not factor of array size */
8230                         return -EINVAL;
8231         }
8232
8233         /* They look valid */
8234         return check_reshape(mddev);
8235 }
8236
8237 static void *raid5_takeover(struct mddev *mddev)
8238 {
8239         /* raid5 can take over:
8240          *  raid0 - if there is only one strip zone - make it a raid4 layout
8241          *  raid1 - if there are two drives.  We need to know the chunk size
8242          *  raid4 - trivial - just use a raid4 layout.
8243          *  raid6 - Providing it is a *_6 layout
8244          */
8245         if (mddev->level == 0)
8246                 return raid45_takeover_raid0(mddev, 5);
8247         if (mddev->level == 1)
8248                 return raid5_takeover_raid1(mddev);
8249         if (mddev->level == 4) {
8250                 mddev->new_layout = ALGORITHM_PARITY_N;
8251                 mddev->new_level = 5;
8252                 return setup_conf(mddev);
8253         }
8254         if (mddev->level == 6)
8255                 return raid5_takeover_raid6(mddev);
8256
8257         return ERR_PTR(-EINVAL);
8258 }
8259
8260 static void *raid4_takeover(struct mddev *mddev)
8261 {
8262         /* raid4 can take over:
8263          *  raid0 - if there is only one strip zone
8264          *  raid5 - if layout is right
8265          */
8266         if (mddev->level == 0)
8267                 return raid45_takeover_raid0(mddev, 4);
8268         if (mddev->level == 5 &&
8269             mddev->layout == ALGORITHM_PARITY_N) {
8270                 mddev->new_layout = 0;
8271                 mddev->new_level = 4;
8272                 return setup_conf(mddev);
8273         }
8274         return ERR_PTR(-EINVAL);
8275 }
8276
8277 static struct md_personality raid5_personality;
8278
8279 static void *raid6_takeover(struct mddev *mddev)
8280 {
8281         /* Currently can only take over a raid5.  We map the
8282          * personality to an equivalent raid6 personality
8283          * with the Q block at the end.
8284          */
8285         int new_layout;
8286
8287         if (mddev->pers != &raid5_personality)
8288                 return ERR_PTR(-EINVAL);
8289         if (mddev->degraded > 1)
8290                 return ERR_PTR(-EINVAL);
8291         if (mddev->raid_disks > 253)
8292                 return ERR_PTR(-EINVAL);
8293         if (mddev->raid_disks < 3)
8294                 return ERR_PTR(-EINVAL);
8295
8296         switch (mddev->layout) {
8297         case ALGORITHM_LEFT_ASYMMETRIC:
8298                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8299                 break;
8300         case ALGORITHM_RIGHT_ASYMMETRIC:
8301                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8302                 break;
8303         case ALGORITHM_LEFT_SYMMETRIC:
8304                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8305                 break;
8306         case ALGORITHM_RIGHT_SYMMETRIC:
8307                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8308                 break;
8309         case ALGORITHM_PARITY_0:
8310                 new_layout = ALGORITHM_PARITY_0_6;
8311                 break;
8312         case ALGORITHM_PARITY_N:
8313                 new_layout = ALGORITHM_PARITY_N;
8314                 break;
8315         default:
8316                 return ERR_PTR(-EINVAL);
8317         }
8318         mddev->new_level = 6;
8319         mddev->new_layout = new_layout;
8320         mddev->delta_disks = 1;
8321         mddev->raid_disks += 1;
8322         return setup_conf(mddev);
8323 }
8324
8325 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8326 {
8327         struct r5conf *conf;
8328         int err;
8329
8330         err = mddev_lock(mddev);
8331         if (err)
8332                 return err;
8333         conf = mddev->private;
8334         if (!conf) {
8335                 mddev_unlock(mddev);
8336                 return -ENODEV;
8337         }
8338
8339         if (strncmp(buf, "ppl", 3) == 0) {
8340                 /* ppl only works with RAID 5 */
8341                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8342                         err = log_init(conf, NULL, true);
8343                         if (!err) {
8344                                 err = resize_stripes(conf, conf->pool_size);
8345                                 if (err)
8346                                         log_exit(conf);
8347                         }
8348                 } else
8349                         err = -EINVAL;
8350         } else if (strncmp(buf, "resync", 6) == 0) {
8351                 if (raid5_has_ppl(conf)) {
8352                         mddev_suspend(mddev);
8353                         log_exit(conf);
8354                         mddev_resume(mddev);
8355                         err = resize_stripes(conf, conf->pool_size);
8356                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8357                            r5l_log_disk_error(conf)) {
8358                         bool journal_dev_exists = false;
8359                         struct md_rdev *rdev;
8360
8361                         rdev_for_each(rdev, mddev)
8362                                 if (test_bit(Journal, &rdev->flags)) {
8363                                         journal_dev_exists = true;
8364                                         break;
8365                                 }
8366
8367                         if (!journal_dev_exists) {
8368                                 mddev_suspend(mddev);
8369                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8370                                 mddev_resume(mddev);
8371                         } else  /* need remove journal device first */
8372                                 err = -EBUSY;
8373                 } else
8374                         err = -EINVAL;
8375         } else {
8376                 err = -EINVAL;
8377         }
8378
8379         if (!err)
8380                 md_update_sb(mddev, 1);
8381
8382         mddev_unlock(mddev);
8383
8384         return err;
8385 }
8386
8387 static int raid5_start(struct mddev *mddev)
8388 {
8389         struct r5conf *conf = mddev->private;
8390
8391         return r5l_start(conf->log);
8392 }
8393
8394 static struct md_personality raid6_personality =
8395 {
8396         .name           = "raid6",
8397         .level          = 6,
8398         .owner          = THIS_MODULE,
8399         .make_request   = raid5_make_request,
8400         .run            = raid5_run,
8401         .start          = raid5_start,
8402         .free           = raid5_free,
8403         .status         = raid5_status,
8404         .error_handler  = raid5_error,
8405         .hot_add_disk   = raid5_add_disk,
8406         .hot_remove_disk= raid5_remove_disk,
8407         .spare_active   = raid5_spare_active,
8408         .sync_request   = raid5_sync_request,
8409         .resize         = raid5_resize,
8410         .size           = raid5_size,
8411         .check_reshape  = raid6_check_reshape,
8412         .start_reshape  = raid5_start_reshape,
8413         .finish_reshape = raid5_finish_reshape,
8414         .quiesce        = raid5_quiesce,
8415         .takeover       = raid6_takeover,
8416         .congested      = raid5_congested,
8417         .change_consistency_policy = raid5_change_consistency_policy,
8418 };
8419 static struct md_personality raid5_personality =
8420 {
8421         .name           = "raid5",
8422         .level          = 5,
8423         .owner          = THIS_MODULE,
8424         .make_request   = raid5_make_request,
8425         .run            = raid5_run,
8426         .start          = raid5_start,
8427         .free           = raid5_free,
8428         .status         = raid5_status,
8429         .error_handler  = raid5_error,
8430         .hot_add_disk   = raid5_add_disk,
8431         .hot_remove_disk= raid5_remove_disk,
8432         .spare_active   = raid5_spare_active,
8433         .sync_request   = raid5_sync_request,
8434         .resize         = raid5_resize,
8435         .size           = raid5_size,
8436         .check_reshape  = raid5_check_reshape,
8437         .start_reshape  = raid5_start_reshape,
8438         .finish_reshape = raid5_finish_reshape,
8439         .quiesce        = raid5_quiesce,
8440         .takeover       = raid5_takeover,
8441         .congested      = raid5_congested,
8442         .change_consistency_policy = raid5_change_consistency_policy,
8443 };
8444
8445 static struct md_personality raid4_personality =
8446 {
8447         .name           = "raid4",
8448         .level          = 4,
8449         .owner          = THIS_MODULE,
8450         .make_request   = raid5_make_request,
8451         .run            = raid5_run,
8452         .start          = raid5_start,
8453         .free           = raid5_free,
8454         .status         = raid5_status,
8455         .error_handler  = raid5_error,
8456         .hot_add_disk   = raid5_add_disk,
8457         .hot_remove_disk= raid5_remove_disk,
8458         .spare_active   = raid5_spare_active,
8459         .sync_request   = raid5_sync_request,
8460         .resize         = raid5_resize,
8461         .size           = raid5_size,
8462         .check_reshape  = raid5_check_reshape,
8463         .start_reshape  = raid5_start_reshape,
8464         .finish_reshape = raid5_finish_reshape,
8465         .quiesce        = raid5_quiesce,
8466         .takeover       = raid4_takeover,
8467         .congested      = raid5_congested,
8468         .change_consistency_policy = raid5_change_consistency_policy,
8469 };
8470
8471 static int __init raid5_init(void)
8472 {
8473         int ret;
8474
8475         raid5_wq = alloc_workqueue("raid5wq",
8476                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8477         if (!raid5_wq)
8478                 return -ENOMEM;
8479
8480         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8481                                       "md/raid5:prepare",
8482                                       raid456_cpu_up_prepare,
8483                                       raid456_cpu_dead);
8484         if (ret) {
8485                 destroy_workqueue(raid5_wq);
8486                 return ret;
8487         }
8488         register_md_personality(&raid6_personality);
8489         register_md_personality(&raid5_personality);
8490         register_md_personality(&raid4_personality);
8491         return 0;
8492 }
8493
8494 static void raid5_exit(void)
8495 {
8496         unregister_md_personality(&raid6_personality);
8497         unregister_md_personality(&raid5_personality);
8498         unregister_md_personality(&raid4_personality);
8499         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8500         destroy_workqueue(raid5_wq);
8501 }
8502
8503 module_init(raid5_init);
8504 module_exit(raid5_exit);
8505 MODULE_LICENSE("GPL");
8506 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8507 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8508 MODULE_ALIAS("md-raid5");
8509 MODULE_ALIAS("md-raid4");
8510 MODULE_ALIAS("md-level-5");
8511 MODULE_ALIAS("md-level-4");
8512 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8513 MODULE_ALIAS("md-raid6");
8514 MODULE_ALIAS("md-level-6");
8515
8516 /* This used to be two separate modules, they were: */
8517 MODULE_ALIAS("raid5");
8518 MODULE_ALIAS("raid6");