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 (conf->log || 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         set_bit(Faulty, &rdev->flags);
2685         clear_bit(In_sync, &rdev->flags);
2686         mddev->degraded = raid5_calc_degraded(conf);
2687         spin_unlock_irqrestore(&conf->device_lock, flags);
2688         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2689
2690         set_bit(Blocked, &rdev->flags);
2691         set_mask_bits(&mddev->sb_flags, 0,
2692                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2693         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2694                 "md/raid:%s: Operation continuing on %d devices.\n",
2695                 mdname(mddev),
2696                 bdevname(rdev->bdev, b),
2697                 mdname(mddev),
2698                 conf->raid_disks - mddev->degraded);
2699         r5c_update_on_rdev_error(mddev, rdev);
2700 }
2701
2702 /*
2703  * Input: a 'big' sector number,
2704  * Output: index of the data and parity disk, and the sector # in them.
2705  */
2706 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2707                               int previous, int *dd_idx,
2708                               struct stripe_head *sh)
2709 {
2710         sector_t stripe, stripe2;
2711         sector_t chunk_number;
2712         unsigned int chunk_offset;
2713         int pd_idx, qd_idx;
2714         int ddf_layout = 0;
2715         sector_t new_sector;
2716         int algorithm = previous ? conf->prev_algo
2717                                  : conf->algorithm;
2718         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2719                                          : conf->chunk_sectors;
2720         int raid_disks = previous ? conf->previous_raid_disks
2721                                   : conf->raid_disks;
2722         int data_disks = raid_disks - conf->max_degraded;
2723
2724         /* First compute the information on this sector */
2725
2726         /*
2727          * Compute the chunk number and the sector offset inside the chunk
2728          */
2729         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2730         chunk_number = r_sector;
2731
2732         /*
2733          * Compute the stripe number
2734          */
2735         stripe = chunk_number;
2736         *dd_idx = sector_div(stripe, data_disks);
2737         stripe2 = stripe;
2738         /*
2739          * Select the parity disk based on the user selected algorithm.
2740          */
2741         pd_idx = qd_idx = -1;
2742         switch(conf->level) {
2743         case 4:
2744                 pd_idx = data_disks;
2745                 break;
2746         case 5:
2747                 switch (algorithm) {
2748                 case ALGORITHM_LEFT_ASYMMETRIC:
2749                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2750                         if (*dd_idx >= pd_idx)
2751                                 (*dd_idx)++;
2752                         break;
2753                 case ALGORITHM_RIGHT_ASYMMETRIC:
2754                         pd_idx = sector_div(stripe2, raid_disks);
2755                         if (*dd_idx >= pd_idx)
2756                                 (*dd_idx)++;
2757                         break;
2758                 case ALGORITHM_LEFT_SYMMETRIC:
2759                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2760                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2761                         break;
2762                 case ALGORITHM_RIGHT_SYMMETRIC:
2763                         pd_idx = sector_div(stripe2, raid_disks);
2764                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2765                         break;
2766                 case ALGORITHM_PARITY_0:
2767                         pd_idx = 0;
2768                         (*dd_idx)++;
2769                         break;
2770                 case ALGORITHM_PARITY_N:
2771                         pd_idx = data_disks;
2772                         break;
2773                 default:
2774                         BUG();
2775                 }
2776                 break;
2777         case 6:
2778
2779                 switch (algorithm) {
2780                 case ALGORITHM_LEFT_ASYMMETRIC:
2781                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2782                         qd_idx = pd_idx + 1;
2783                         if (pd_idx == raid_disks-1) {
2784                                 (*dd_idx)++;    /* Q D D D P */
2785                                 qd_idx = 0;
2786                         } else if (*dd_idx >= pd_idx)
2787                                 (*dd_idx) += 2; /* D D P Q D */
2788                         break;
2789                 case ALGORITHM_RIGHT_ASYMMETRIC:
2790                         pd_idx = sector_div(stripe2, raid_disks);
2791                         qd_idx = pd_idx + 1;
2792                         if (pd_idx == raid_disks-1) {
2793                                 (*dd_idx)++;    /* Q D D D P */
2794                                 qd_idx = 0;
2795                         } else if (*dd_idx >= pd_idx)
2796                                 (*dd_idx) += 2; /* D D P Q D */
2797                         break;
2798                 case ALGORITHM_LEFT_SYMMETRIC:
2799                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2800                         qd_idx = (pd_idx + 1) % raid_disks;
2801                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2802                         break;
2803                 case ALGORITHM_RIGHT_SYMMETRIC:
2804                         pd_idx = sector_div(stripe2, raid_disks);
2805                         qd_idx = (pd_idx + 1) % raid_disks;
2806                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2807                         break;
2808
2809                 case ALGORITHM_PARITY_0:
2810                         pd_idx = 0;
2811                         qd_idx = 1;
2812                         (*dd_idx) += 2;
2813                         break;
2814                 case ALGORITHM_PARITY_N:
2815                         pd_idx = data_disks;
2816                         qd_idx = data_disks + 1;
2817                         break;
2818
2819                 case ALGORITHM_ROTATING_ZERO_RESTART:
2820                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2821                          * of blocks for computing Q is different.
2822                          */
2823                         pd_idx = sector_div(stripe2, raid_disks);
2824                         qd_idx = pd_idx + 1;
2825                         if (pd_idx == raid_disks-1) {
2826                                 (*dd_idx)++;    /* Q D D D P */
2827                                 qd_idx = 0;
2828                         } else if (*dd_idx >= pd_idx)
2829                                 (*dd_idx) += 2; /* D D P Q D */
2830                         ddf_layout = 1;
2831                         break;
2832
2833                 case ALGORITHM_ROTATING_N_RESTART:
2834                         /* Same a left_asymmetric, by first stripe is
2835                          * D D D P Q  rather than
2836                          * Q D D D P
2837                          */
2838                         stripe2 += 1;
2839                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2840                         qd_idx = pd_idx + 1;
2841                         if (pd_idx == raid_disks-1) {
2842                                 (*dd_idx)++;    /* Q D D D P */
2843                                 qd_idx = 0;
2844                         } else if (*dd_idx >= pd_idx)
2845                                 (*dd_idx) += 2; /* D D P Q D */
2846                         ddf_layout = 1;
2847                         break;
2848
2849                 case ALGORITHM_ROTATING_N_CONTINUE:
2850                         /* Same as left_symmetric but Q is before P */
2851                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2852                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2853                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2854                         ddf_layout = 1;
2855                         break;
2856
2857                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2858                         /* RAID5 left_asymmetric, with Q on last device */
2859                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2860                         if (*dd_idx >= pd_idx)
2861                                 (*dd_idx)++;
2862                         qd_idx = raid_disks - 1;
2863                         break;
2864
2865                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2866                         pd_idx = sector_div(stripe2, raid_disks-1);
2867                         if (*dd_idx >= pd_idx)
2868                                 (*dd_idx)++;
2869                         qd_idx = raid_disks - 1;
2870                         break;
2871
2872                 case ALGORITHM_LEFT_SYMMETRIC_6:
2873                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2874                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2875                         qd_idx = raid_disks - 1;
2876                         break;
2877
2878                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2879                         pd_idx = sector_div(stripe2, raid_disks-1);
2880                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2881                         qd_idx = raid_disks - 1;
2882                         break;
2883
2884                 case ALGORITHM_PARITY_0_6:
2885                         pd_idx = 0;
2886                         (*dd_idx)++;
2887                         qd_idx = raid_disks - 1;
2888                         break;
2889
2890                 default:
2891                         BUG();
2892                 }
2893                 break;
2894         }
2895
2896         if (sh) {
2897                 sh->pd_idx = pd_idx;
2898                 sh->qd_idx = qd_idx;
2899                 sh->ddf_layout = ddf_layout;
2900         }
2901         /*
2902          * Finally, compute the new sector number
2903          */
2904         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2905         return new_sector;
2906 }
2907
2908 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2909 {
2910         struct r5conf *conf = sh->raid_conf;
2911         int raid_disks = sh->disks;
2912         int data_disks = raid_disks - conf->max_degraded;
2913         sector_t new_sector = sh->sector, check;
2914         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2915                                          : conf->chunk_sectors;
2916         int algorithm = previous ? conf->prev_algo
2917                                  : conf->algorithm;
2918         sector_t stripe;
2919         int chunk_offset;
2920         sector_t chunk_number;
2921         int dummy1, dd_idx = i;
2922         sector_t r_sector;
2923         struct stripe_head sh2;
2924
2925         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2926         stripe = new_sector;
2927
2928         if (i == sh->pd_idx)
2929                 return 0;
2930         switch(conf->level) {
2931         case 4: break;
2932         case 5:
2933                 switch (algorithm) {
2934                 case ALGORITHM_LEFT_ASYMMETRIC:
2935                 case ALGORITHM_RIGHT_ASYMMETRIC:
2936                         if (i > sh->pd_idx)
2937                                 i--;
2938                         break;
2939                 case ALGORITHM_LEFT_SYMMETRIC:
2940                 case ALGORITHM_RIGHT_SYMMETRIC:
2941                         if (i < sh->pd_idx)
2942                                 i += raid_disks;
2943                         i -= (sh->pd_idx + 1);
2944                         break;
2945                 case ALGORITHM_PARITY_0:
2946                         i -= 1;
2947                         break;
2948                 case ALGORITHM_PARITY_N:
2949                         break;
2950                 default:
2951                         BUG();
2952                 }
2953                 break;
2954         case 6:
2955                 if (i == sh->qd_idx)
2956                         return 0; /* It is the Q disk */
2957                 switch (algorithm) {
2958                 case ALGORITHM_LEFT_ASYMMETRIC:
2959                 case ALGORITHM_RIGHT_ASYMMETRIC:
2960                 case ALGORITHM_ROTATING_ZERO_RESTART:
2961                 case ALGORITHM_ROTATING_N_RESTART:
2962                         if (sh->pd_idx == raid_disks-1)
2963                                 i--;    /* Q D D D P */
2964                         else if (i > sh->pd_idx)
2965                                 i -= 2; /* D D P Q D */
2966                         break;
2967                 case ALGORITHM_LEFT_SYMMETRIC:
2968                 case ALGORITHM_RIGHT_SYMMETRIC:
2969                         if (sh->pd_idx == raid_disks-1)
2970                                 i--; /* Q D D D P */
2971                         else {
2972                                 /* D D P Q D */
2973                                 if (i < sh->pd_idx)
2974                                         i += raid_disks;
2975                                 i -= (sh->pd_idx + 2);
2976                         }
2977                         break;
2978                 case ALGORITHM_PARITY_0:
2979                         i -= 2;
2980                         break;
2981                 case ALGORITHM_PARITY_N:
2982                         break;
2983                 case ALGORITHM_ROTATING_N_CONTINUE:
2984                         /* Like left_symmetric, but P is before Q */
2985                         if (sh->pd_idx == 0)
2986                                 i--;    /* P D D D Q */
2987                         else {
2988                                 /* D D Q P D */
2989                                 if (i < sh->pd_idx)
2990                                         i += raid_disks;
2991                                 i -= (sh->pd_idx + 1);
2992                         }
2993                         break;
2994                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2995                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2996                         if (i > sh->pd_idx)
2997                                 i--;
2998                         break;
2999                 case ALGORITHM_LEFT_SYMMETRIC_6:
3000                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3001                         if (i < sh->pd_idx)
3002                                 i += data_disks + 1;
3003                         i -= (sh->pd_idx + 1);
3004                         break;
3005                 case ALGORITHM_PARITY_0_6:
3006                         i -= 1;
3007                         break;
3008                 default:
3009                         BUG();
3010                 }
3011                 break;
3012         }
3013
3014         chunk_number = stripe * data_disks + i;
3015         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3016
3017         check = raid5_compute_sector(conf, r_sector,
3018                                      previous, &dummy1, &sh2);
3019         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3020                 || sh2.qd_idx != sh->qd_idx) {
3021                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3022                         mdname(conf->mddev));
3023                 return 0;
3024         }
3025         return r_sector;
3026 }
3027
3028 /*
3029  * There are cases where we want handle_stripe_dirtying() and
3030  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3031  *
3032  * This function checks whether we want to delay the towrite. Specifically,
3033  * we delay the towrite when:
3034  *
3035  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3036  *      stripe has data in journal (for other devices).
3037  *
3038  *      In this case, when reading data for the non-overwrite dev, it is
3039  *      necessary to handle complex rmw of write back cache (prexor with
3040  *      orig_page, and xor with page). To keep read path simple, we would
3041  *      like to flush data in journal to RAID disks first, so complex rmw
3042  *      is handled in the write patch (handle_stripe_dirtying).
3043  *
3044  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3045  *
3046  *      It is important to be able to flush all stripes in raid5-cache.
3047  *      Therefore, we need reserve some space on the journal device for
3048  *      these flushes. If flush operation includes pending writes to the
3049  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3050  *      for the flush out. If we exclude these pending writes from flush
3051  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3052  *      Therefore, excluding pending writes in these cases enables more
3053  *      efficient use of the journal device.
3054  *
3055  *      Note: To make sure the stripe makes progress, we only delay
3056  *      towrite for stripes with data already in journal (injournal > 0).
3057  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3058  *      no_space_stripes list.
3059  *
3060  *   3. during journal failure
3061  *      In journal failure, we try to flush all cached data to raid disks
3062  *      based on data in stripe cache. The array is read-only to upper
3063  *      layers, so we would skip all pending writes.
3064  *
3065  */
3066 static inline bool delay_towrite(struct r5conf *conf,
3067                                  struct r5dev *dev,
3068                                  struct stripe_head_state *s)
3069 {
3070         /* case 1 above */
3071         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3072             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3073                 return true;
3074         /* case 2 above */
3075         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3076             s->injournal > 0)
3077                 return true;
3078         /* case 3 above */
3079         if (s->log_failed && s->injournal)
3080                 return true;
3081         return false;
3082 }
3083
3084 static void
3085 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3086                          int rcw, int expand)
3087 {
3088         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3089         struct r5conf *conf = sh->raid_conf;
3090         int level = conf->level;
3091
3092         if (rcw) {
3093                 /*
3094                  * In some cases, handle_stripe_dirtying initially decided to
3095                  * run rmw and allocates extra page for prexor. However, rcw is
3096                  * cheaper later on. We need to free the extra page now,
3097                  * because we won't be able to do that in ops_complete_prexor().
3098                  */
3099                 r5c_release_extra_page(sh);
3100
3101                 for (i = disks; i--; ) {
3102                         struct r5dev *dev = &sh->dev[i];
3103
3104                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3105                                 set_bit(R5_LOCKED, &dev->flags);
3106                                 set_bit(R5_Wantdrain, &dev->flags);
3107                                 if (!expand)
3108                                         clear_bit(R5_UPTODATE, &dev->flags);
3109                                 s->locked++;
3110                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3111                                 set_bit(R5_LOCKED, &dev->flags);
3112                                 s->locked++;
3113                         }
3114                 }
3115                 /* if we are not expanding this is a proper write request, and
3116                  * there will be bios with new data to be drained into the
3117                  * stripe cache
3118                  */
3119                 if (!expand) {
3120                         if (!s->locked)
3121                                 /* False alarm, nothing to do */
3122                                 return;
3123                         sh->reconstruct_state = reconstruct_state_drain_run;
3124                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3125                 } else
3126                         sh->reconstruct_state = reconstruct_state_run;
3127
3128                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3129
3130                 if (s->locked + conf->max_degraded == disks)
3131                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3132                                 atomic_inc(&conf->pending_full_writes);
3133         } else {
3134                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3135                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3136                 BUG_ON(level == 6 &&
3137                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3138                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3139
3140                 for (i = disks; i--; ) {
3141                         struct r5dev *dev = &sh->dev[i];
3142                         if (i == pd_idx || i == qd_idx)
3143                                 continue;
3144
3145                         if (dev->towrite &&
3146                             (test_bit(R5_UPTODATE, &dev->flags) ||
3147                              test_bit(R5_Wantcompute, &dev->flags))) {
3148                                 set_bit(R5_Wantdrain, &dev->flags);
3149                                 set_bit(R5_LOCKED, &dev->flags);
3150                                 clear_bit(R5_UPTODATE, &dev->flags);
3151                                 s->locked++;
3152                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3153                                 set_bit(R5_LOCKED, &dev->flags);
3154                                 s->locked++;
3155                         }
3156                 }
3157                 if (!s->locked)
3158                         /* False alarm - nothing to do */
3159                         return;
3160                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3161                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3162                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3163                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3164         }
3165
3166         /* keep the parity disk(s) locked while asynchronous operations
3167          * are in flight
3168          */
3169         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3170         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3171         s->locked++;
3172
3173         if (level == 6) {
3174                 int qd_idx = sh->qd_idx;
3175                 struct r5dev *dev = &sh->dev[qd_idx];
3176
3177                 set_bit(R5_LOCKED, &dev->flags);
3178                 clear_bit(R5_UPTODATE, &dev->flags);
3179                 s->locked++;
3180         }
3181
3182         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3183             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3184             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3185             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3186                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3187
3188         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3189                 __func__, (unsigned long long)sh->sector,
3190                 s->locked, s->ops_request);
3191 }
3192
3193 /*
3194  * Each stripe/dev can have one or more bion attached.
3195  * toread/towrite point to the first in a chain.
3196  * The bi_next chain must be in order.
3197  */
3198 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3199                           int forwrite, int previous)
3200 {
3201         struct bio **bip;
3202         struct r5conf *conf = sh->raid_conf;
3203         int firstwrite=0;
3204
3205         pr_debug("adding bi b#%llu to stripe s#%llu\n",
3206                 (unsigned long long)bi->bi_iter.bi_sector,
3207                 (unsigned long long)sh->sector);
3208
3209         spin_lock_irq(&sh->stripe_lock);
3210         sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3211         /* Don't allow new IO added to stripes in batch list */
3212         if (sh->batch_head)
3213                 goto overlap;
3214         if (forwrite) {
3215                 bip = &sh->dev[dd_idx].towrite;
3216                 if (*bip == NULL)
3217                         firstwrite = 1;
3218         } else
3219                 bip = &sh->dev[dd_idx].toread;
3220         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3221                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3222                         goto overlap;
3223                 bip = & (*bip)->bi_next;
3224         }
3225         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3226                 goto overlap;
3227
3228         if (forwrite && raid5_has_ppl(conf)) {
3229                 /*
3230                  * With PPL only writes to consecutive data chunks within a
3231                  * stripe are allowed because for a single stripe_head we can
3232                  * only have one PPL entry at a time, which describes one data
3233                  * range. Not really an overlap, but wait_for_overlap can be
3234                  * used to handle this.
3235                  */
3236                 sector_t sector;
3237                 sector_t first = 0;
3238                 sector_t last = 0;
3239                 int count = 0;
3240                 int i;
3241
3242                 for (i = 0; i < sh->disks; i++) {
3243                         if (i != sh->pd_idx &&
3244                             (i == dd_idx || sh->dev[i].towrite)) {
3245                                 sector = sh->dev[i].sector;
3246                                 if (count == 0 || sector < first)
3247                                         first = sector;
3248                                 if (sector > last)
3249                                         last = sector;
3250                                 count++;
3251                         }
3252                 }
3253
3254                 if (first + conf->chunk_sectors * (count - 1) != last)
3255                         goto overlap;
3256         }
3257
3258         if (!forwrite || previous)
3259                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3260
3261         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3262         if (*bip)
3263                 bi->bi_next = *bip;
3264         *bip = bi;
3265         bio_inc_remaining(bi);
3266         md_write_inc(conf->mddev, bi);
3267
3268         if (forwrite) {
3269                 /* check if page is covered */
3270                 sector_t sector = sh->dev[dd_idx].sector;
3271                 for (bi=sh->dev[dd_idx].towrite;
3272                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3273                              bi && bi->bi_iter.bi_sector <= sector;
3274                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3275                         if (bio_end_sector(bi) >= sector)
3276                                 sector = bio_end_sector(bi);
3277                 }
3278                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3279                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3280                                 sh->overwrite_disks++;
3281         }
3282
3283         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3284                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3285                 (unsigned long long)sh->sector, dd_idx);
3286
3287         if (conf->mddev->bitmap && firstwrite) {
3288                 /* Cannot hold spinlock over bitmap_startwrite,
3289                  * but must ensure this isn't added to a batch until
3290                  * we have added to the bitmap and set bm_seq.
3291                  * So set STRIPE_BITMAP_PENDING to prevent
3292                  * batching.
3293                  * If multiple add_stripe_bio() calls race here they
3294                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3295                  * to complete "bitmap_startwrite" gets to set
3296                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3297                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3298                  * any more.
3299                  */
3300                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3301                 spin_unlock_irq(&sh->stripe_lock);
3302                 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3303                                      STRIPE_SECTORS, 0);
3304                 spin_lock_irq(&sh->stripe_lock);
3305                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3306                 if (!sh->batch_head) {
3307                         sh->bm_seq = conf->seq_flush+1;
3308                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3309                 }
3310         }
3311         spin_unlock_irq(&sh->stripe_lock);
3312
3313         if (stripe_can_batch(sh))
3314                 stripe_add_to_batch_list(conf, sh);
3315         return 1;
3316
3317  overlap:
3318         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3319         spin_unlock_irq(&sh->stripe_lock);
3320         return 0;
3321 }
3322
3323 static void end_reshape(struct r5conf *conf);
3324
3325 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3326                             struct stripe_head *sh)
3327 {
3328         int sectors_per_chunk =
3329                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3330         int dd_idx;
3331         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3332         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3333
3334         raid5_compute_sector(conf,
3335                              stripe * (disks - conf->max_degraded)
3336                              *sectors_per_chunk + chunk_offset,
3337                              previous,
3338                              &dd_idx, sh);
3339 }
3340
3341 static void
3342 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3343                      struct stripe_head_state *s, int disks)
3344 {
3345         int i;
3346         BUG_ON(sh->batch_head);
3347         for (i = disks; i--; ) {
3348                 struct bio *bi;
3349                 int bitmap_end = 0;
3350
3351                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3352                         struct md_rdev *rdev;
3353                         rcu_read_lock();
3354                         rdev = rcu_dereference(conf->disks[i].rdev);
3355                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3356                             !test_bit(Faulty, &rdev->flags))
3357                                 atomic_inc(&rdev->nr_pending);
3358                         else
3359                                 rdev = NULL;
3360                         rcu_read_unlock();
3361                         if (rdev) {
3362                                 if (!rdev_set_badblocks(
3363                                             rdev,
3364                                             sh->sector,
3365                                             STRIPE_SECTORS, 0))
3366                                         md_error(conf->mddev, rdev);
3367                                 rdev_dec_pending(rdev, conf->mddev);
3368                         }
3369                 }
3370                 spin_lock_irq(&sh->stripe_lock);
3371                 /* fail all writes first */
3372                 bi = sh->dev[i].towrite;
3373                 sh->dev[i].towrite = NULL;
3374                 sh->overwrite_disks = 0;
3375                 spin_unlock_irq(&sh->stripe_lock);
3376                 if (bi)
3377                         bitmap_end = 1;
3378
3379                 log_stripe_write_finished(sh);
3380
3381                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3382                         wake_up(&conf->wait_for_overlap);
3383
3384                 while (bi && bi->bi_iter.bi_sector <
3385                         sh->dev[i].sector + STRIPE_SECTORS) {
3386                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3387
3388                         md_write_end(conf->mddev);
3389                         bio_io_error(bi);
3390                         bi = nextbi;
3391                 }
3392                 if (bitmap_end)
3393                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3394                                            STRIPE_SECTORS, 0, 0);
3395                 bitmap_end = 0;
3396                 /* and fail all 'written' */
3397                 bi = sh->dev[i].written;
3398                 sh->dev[i].written = NULL;
3399                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3400                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3401                         sh->dev[i].page = sh->dev[i].orig_page;
3402                 }
3403
3404                 if (bi) bitmap_end = 1;
3405                 while (bi && bi->bi_iter.bi_sector <
3406                        sh->dev[i].sector + STRIPE_SECTORS) {
3407                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3408
3409                         md_write_end(conf->mddev);
3410                         bio_io_error(bi);
3411                         bi = bi2;
3412                 }
3413
3414                 /* fail any reads if this device is non-operational and
3415                  * the data has not reached the cache yet.
3416                  */
3417                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3418                     s->failed > conf->max_degraded &&
3419                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3420                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3421                         spin_lock_irq(&sh->stripe_lock);
3422                         bi = sh->dev[i].toread;
3423                         sh->dev[i].toread = NULL;
3424                         spin_unlock_irq(&sh->stripe_lock);
3425                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3426                                 wake_up(&conf->wait_for_overlap);
3427                         if (bi)
3428                                 s->to_read--;
3429                         while (bi && bi->bi_iter.bi_sector <
3430                                sh->dev[i].sector + STRIPE_SECTORS) {
3431                                 struct bio *nextbi =
3432                                         r5_next_bio(bi, sh->dev[i].sector);
3433
3434                                 bio_io_error(bi);
3435                                 bi = nextbi;
3436                         }
3437                 }
3438                 if (bitmap_end)
3439                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3440                                            STRIPE_SECTORS, 0, 0);
3441                 /* If we were in the middle of a write the parity block might
3442                  * still be locked - so just clear all R5_LOCKED flags
3443                  */
3444                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3445         }
3446         s->to_write = 0;
3447         s->written = 0;
3448
3449         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3450                 if (atomic_dec_and_test(&conf->pending_full_writes))
3451                         md_wakeup_thread(conf->mddev->thread);
3452 }
3453
3454 static void
3455 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3456                    struct stripe_head_state *s)
3457 {
3458         int abort = 0;
3459         int i;
3460
3461         BUG_ON(sh->batch_head);
3462         clear_bit(STRIPE_SYNCING, &sh->state);
3463         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3464                 wake_up(&conf->wait_for_overlap);
3465         s->syncing = 0;
3466         s->replacing = 0;
3467         /* There is nothing more to do for sync/check/repair.
3468          * Don't even need to abort as that is handled elsewhere
3469          * if needed, and not always wanted e.g. if there is a known
3470          * bad block here.
3471          * For recover/replace we need to record a bad block on all
3472          * non-sync devices, or abort the recovery
3473          */
3474         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3475                 /* During recovery devices cannot be removed, so
3476                  * locking and refcounting of rdevs is not needed
3477                  */
3478                 rcu_read_lock();
3479                 for (i = 0; i < conf->raid_disks; i++) {
3480                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3481                         if (rdev
3482                             && !test_bit(Faulty, &rdev->flags)
3483                             && !test_bit(In_sync, &rdev->flags)
3484                             && !rdev_set_badblocks(rdev, sh->sector,
3485                                                    STRIPE_SECTORS, 0))
3486                                 abort = 1;
3487                         rdev = rcu_dereference(conf->disks[i].replacement);
3488                         if (rdev
3489                             && !test_bit(Faulty, &rdev->flags)
3490                             && !test_bit(In_sync, &rdev->flags)
3491                             && !rdev_set_badblocks(rdev, sh->sector,
3492                                                    STRIPE_SECTORS, 0))
3493                                 abort = 1;
3494                 }
3495                 rcu_read_unlock();
3496                 if (abort)
3497                         conf->recovery_disabled =
3498                                 conf->mddev->recovery_disabled;
3499         }
3500         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3501 }
3502
3503 static int want_replace(struct stripe_head *sh, int disk_idx)
3504 {
3505         struct md_rdev *rdev;
3506         int rv = 0;
3507
3508         rcu_read_lock();
3509         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3510         if (rdev
3511             && !test_bit(Faulty, &rdev->flags)
3512             && !test_bit(In_sync, &rdev->flags)
3513             && (rdev->recovery_offset <= sh->sector
3514                 || rdev->mddev->recovery_cp <= sh->sector))
3515                 rv = 1;
3516         rcu_read_unlock();
3517         return rv;
3518 }
3519
3520 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3521                            int disk_idx, int disks)
3522 {
3523         struct r5dev *dev = &sh->dev[disk_idx];
3524         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3525                                   &sh->dev[s->failed_num[1]] };
3526         int i;
3527
3528
3529         if (test_bit(R5_LOCKED, &dev->flags) ||
3530             test_bit(R5_UPTODATE, &dev->flags))
3531                 /* No point reading this as we already have it or have
3532                  * decided to get it.
3533                  */
3534                 return 0;
3535
3536         if (dev->toread ||
3537             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3538                 /* We need this block to directly satisfy a request */
3539                 return 1;
3540
3541         if (s->syncing || s->expanding ||
3542             (s->replacing && want_replace(sh, disk_idx)))
3543                 /* When syncing, or expanding we read everything.
3544                  * When replacing, we need the replaced block.
3545                  */
3546                 return 1;
3547
3548         if ((s->failed >= 1 && fdev[0]->toread) ||
3549             (s->failed >= 2 && fdev[1]->toread))
3550                 /* If we want to read from a failed device, then
3551                  * we need to actually read every other device.
3552                  */
3553                 return 1;
3554
3555         /* Sometimes neither read-modify-write nor reconstruct-write
3556          * cycles can work.  In those cases we read every block we
3557          * can.  Then the parity-update is certain to have enough to
3558          * work with.
3559          * This can only be a problem when we need to write something,
3560          * and some device has failed.  If either of those tests
3561          * fail we need look no further.
3562          */
3563         if (!s->failed || !s->to_write)
3564                 return 0;
3565
3566         if (test_bit(R5_Insync, &dev->flags) &&
3567             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3568                 /* Pre-reads at not permitted until after short delay
3569                  * to gather multiple requests.  However if this
3570                  * device is no Insync, the block could only be computed
3571                  * and there is no need to delay that.
3572                  */
3573                 return 0;
3574
3575         for (i = 0; i < s->failed && i < 2; i++) {
3576                 if (fdev[i]->towrite &&
3577                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3578                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3579                         /* If we have a partial write to a failed
3580                          * device, then we will need to reconstruct
3581                          * the content of that device, so all other
3582                          * devices must be read.
3583                          */
3584                         return 1;
3585         }
3586
3587         /* If we are forced to do a reconstruct-write, either because
3588          * the current RAID6 implementation only supports that, or
3589          * because parity cannot be trusted and we are currently
3590          * recovering it, there is extra need to be careful.
3591          * If one of the devices that we would need to read, because
3592          * it is not being overwritten (and maybe not written at all)
3593          * is missing/faulty, then we need to read everything we can.
3594          */
3595         if (sh->raid_conf->level != 6 &&
3596             sh->sector < sh->raid_conf->mddev->recovery_cp)
3597                 /* reconstruct-write isn't being forced */
3598                 return 0;
3599         for (i = 0; i < s->failed && i < 2; i++) {
3600                 if (s->failed_num[i] != sh->pd_idx &&
3601                     s->failed_num[i] != sh->qd_idx &&
3602                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3603                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3604                         return 1;
3605         }
3606
3607         return 0;
3608 }
3609
3610 /* fetch_block - checks the given member device to see if its data needs
3611  * to be read or computed to satisfy a request.
3612  *
3613  * Returns 1 when no more member devices need to be checked, otherwise returns
3614  * 0 to tell the loop in handle_stripe_fill to continue
3615  */
3616 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3617                        int disk_idx, int disks)
3618 {
3619         struct r5dev *dev = &sh->dev[disk_idx];
3620
3621         /* is the data in this block needed, and can we get it? */
3622         if (need_this_block(sh, s, disk_idx, disks)) {
3623                 /* we would like to get this block, possibly by computing it,
3624                  * otherwise read it if the backing disk is insync
3625                  */
3626                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3627                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3628                 BUG_ON(sh->batch_head);
3629
3630                 /*
3631                  * In the raid6 case if the only non-uptodate disk is P
3632                  * then we already trusted P to compute the other failed
3633                  * drives. It is safe to compute rather than re-read P.
3634                  * In other cases we only compute blocks from failed
3635                  * devices, otherwise check/repair might fail to detect
3636                  * a real inconsistency.
3637                  */
3638
3639                 if ((s->uptodate == disks - 1) &&
3640                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3641                     (s->failed && (disk_idx == s->failed_num[0] ||
3642                                    disk_idx == s->failed_num[1])))) {
3643                         /* have disk failed, and we're requested to fetch it;
3644                          * do compute it
3645                          */
3646                         pr_debug("Computing stripe %llu block %d\n",
3647                                (unsigned long long)sh->sector, disk_idx);
3648                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3649                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3650                         set_bit(R5_Wantcompute, &dev->flags);
3651                         sh->ops.target = disk_idx;
3652                         sh->ops.target2 = -1; /* no 2nd target */
3653                         s->req_compute = 1;
3654                         /* Careful: from this point on 'uptodate' is in the eye
3655                          * of raid_run_ops which services 'compute' operations
3656                          * before writes. R5_Wantcompute flags a block that will
3657                          * be R5_UPTODATE by the time it is needed for a
3658                          * subsequent operation.
3659                          */
3660                         s->uptodate++;
3661                         return 1;
3662                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3663                         /* Computing 2-failure is *very* expensive; only
3664                          * do it if failed >= 2
3665                          */
3666                         int other;
3667                         for (other = disks; other--; ) {
3668                                 if (other == disk_idx)
3669                                         continue;
3670                                 if (!test_bit(R5_UPTODATE,
3671                                       &sh->dev[other].flags))
3672                                         break;
3673                         }
3674                         BUG_ON(other < 0);
3675                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3676                                (unsigned long long)sh->sector,
3677                                disk_idx, other);
3678                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3679                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3680                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3681                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3682                         sh->ops.target = disk_idx;
3683                         sh->ops.target2 = other;
3684                         s->uptodate += 2;
3685                         s->req_compute = 1;
3686                         return 1;
3687                 } else if (test_bit(R5_Insync, &dev->flags)) {
3688                         set_bit(R5_LOCKED, &dev->flags);
3689                         set_bit(R5_Wantread, &dev->flags);
3690                         s->locked++;
3691                         pr_debug("Reading block %d (sync=%d)\n",
3692                                 disk_idx, s->syncing);
3693                 }
3694         }
3695
3696         return 0;
3697 }
3698
3699 /**
3700  * handle_stripe_fill - read or compute data to satisfy pending requests.
3701  */
3702 static void handle_stripe_fill(struct stripe_head *sh,
3703                                struct stripe_head_state *s,
3704                                int disks)
3705 {
3706         int i;
3707
3708         /* look for blocks to read/compute, skip this if a compute
3709          * is already in flight, or if the stripe contents are in the
3710          * midst of changing due to a write
3711          */
3712         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3713             !sh->reconstruct_state) {
3714
3715                 /*
3716                  * For degraded stripe with data in journal, do not handle
3717                  * read requests yet, instead, flush the stripe to raid
3718                  * disks first, this avoids handling complex rmw of write
3719                  * back cache (prexor with orig_page, and then xor with
3720                  * page) in the read path
3721                  */
3722                 if (s->injournal && s->failed) {
3723                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3724                                 r5c_make_stripe_write_out(sh);
3725                         goto out;
3726                 }
3727
3728                 for (i = disks; i--; )
3729                         if (fetch_block(sh, s, i, disks))
3730                                 break;
3731         }
3732 out:
3733         set_bit(STRIPE_HANDLE, &sh->state);
3734 }
3735
3736 static void break_stripe_batch_list(struct stripe_head *head_sh,
3737                                     unsigned long handle_flags);
3738 /* handle_stripe_clean_event
3739  * any written block on an uptodate or failed drive can be returned.
3740  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3741  * never LOCKED, so we don't need to test 'failed' directly.
3742  */
3743 static void handle_stripe_clean_event(struct r5conf *conf,
3744         struct stripe_head *sh, int disks)
3745 {
3746         int i;
3747         struct r5dev *dev;
3748         int discard_pending = 0;
3749         struct stripe_head *head_sh = sh;
3750         bool do_endio = false;
3751
3752         for (i = disks; i--; )
3753                 if (sh->dev[i].written) {
3754                         dev = &sh->dev[i];
3755                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3756                             (test_bit(R5_UPTODATE, &dev->flags) ||
3757                              test_bit(R5_Discard, &dev->flags) ||
3758                              test_bit(R5_SkipCopy, &dev->flags))) {
3759                                 /* We can return any write requests */
3760                                 struct bio *wbi, *wbi2;
3761                                 pr_debug("Return write for disc %d\n", i);
3762                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3763                                         clear_bit(R5_UPTODATE, &dev->flags);
3764                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3765                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3766                                 }
3767                                 do_endio = true;
3768
3769 returnbi:
3770                                 dev->page = dev->orig_page;
3771                                 wbi = dev->written;
3772                                 dev->written = NULL;
3773                                 while (wbi && wbi->bi_iter.bi_sector <
3774                                         dev->sector + STRIPE_SECTORS) {
3775                                         wbi2 = r5_next_bio(wbi, dev->sector);
3776                                         md_write_end(conf->mddev);
3777                                         bio_endio(wbi);
3778                                         wbi = wbi2;
3779                                 }
3780                                 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3781                                                    STRIPE_SECTORS,
3782                                                    !test_bit(STRIPE_DEGRADED, &sh->state),
3783                                                    0);
3784                                 if (head_sh->batch_head) {
3785                                         sh = list_first_entry(&sh->batch_list,
3786                                                               struct stripe_head,
3787                                                               batch_list);
3788                                         if (sh != head_sh) {
3789                                                 dev = &sh->dev[i];
3790                                                 goto returnbi;
3791                                         }
3792                                 }
3793                                 sh = head_sh;
3794                                 dev = &sh->dev[i];
3795                         } else if (test_bit(R5_Discard, &dev->flags))
3796                                 discard_pending = 1;
3797                 }
3798
3799         log_stripe_write_finished(sh);
3800
3801         if (!discard_pending &&
3802             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3803                 int hash;
3804                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3805                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3806                 if (sh->qd_idx >= 0) {
3807                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3808                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3809                 }
3810                 /* now that discard is done we can proceed with any sync */
3811                 clear_bit(STRIPE_DISCARD, &sh->state);
3812                 /*
3813                  * SCSI discard will change some bio fields and the stripe has
3814                  * no updated data, so remove it from hash list and the stripe
3815                  * will be reinitialized
3816                  */
3817 unhash:
3818                 hash = sh->hash_lock_index;
3819                 spin_lock_irq(conf->hash_locks + hash);
3820                 remove_hash(sh);
3821                 spin_unlock_irq(conf->hash_locks + hash);
3822                 if (head_sh->batch_head) {
3823                         sh = list_first_entry(&sh->batch_list,
3824                                               struct stripe_head, batch_list);
3825                         if (sh != head_sh)
3826                                         goto unhash;
3827                 }
3828                 sh = head_sh;
3829
3830                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3831                         set_bit(STRIPE_HANDLE, &sh->state);
3832
3833         }
3834
3835         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3836                 if (atomic_dec_and_test(&conf->pending_full_writes))
3837                         md_wakeup_thread(conf->mddev->thread);
3838
3839         if (head_sh->batch_head && do_endio)
3840                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3841 }
3842
3843 /*
3844  * For RMW in write back cache, we need extra page in prexor to store the
3845  * old data. This page is stored in dev->orig_page.
3846  *
3847  * This function checks whether we have data for prexor. The exact logic
3848  * is:
3849  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3850  */
3851 static inline bool uptodate_for_rmw(struct r5dev *dev)
3852 {
3853         return (test_bit(R5_UPTODATE, &dev->flags)) &&
3854                 (!test_bit(R5_InJournal, &dev->flags) ||
3855                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3856 }
3857
3858 static int handle_stripe_dirtying(struct r5conf *conf,
3859                                   struct stripe_head *sh,
3860                                   struct stripe_head_state *s,
3861                                   int disks)
3862 {
3863         int rmw = 0, rcw = 0, i;
3864         sector_t recovery_cp = conf->mddev->recovery_cp;
3865
3866         /* Check whether resync is now happening or should start.
3867          * If yes, then the array is dirty (after unclean shutdown or
3868          * initial creation), so parity in some stripes might be inconsistent.
3869          * In this case, we need to always do reconstruct-write, to ensure
3870          * that in case of drive failure or read-error correction, we
3871          * generate correct data from the parity.
3872          */
3873         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3874             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3875              s->failed == 0)) {
3876                 /* Calculate the real rcw later - for now make it
3877                  * look like rcw is cheaper
3878                  */
3879                 rcw = 1; rmw = 2;
3880                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3881                          conf->rmw_level, (unsigned long long)recovery_cp,
3882                          (unsigned long long)sh->sector);
3883         } else for (i = disks; i--; ) {
3884                 /* would I have to read this buffer for read_modify_write */
3885                 struct r5dev *dev = &sh->dev[i];
3886                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3887                      i == sh->pd_idx || i == sh->qd_idx ||
3888                      test_bit(R5_InJournal, &dev->flags)) &&
3889                     !test_bit(R5_LOCKED, &dev->flags) &&
3890                     !(uptodate_for_rmw(dev) ||
3891                       test_bit(R5_Wantcompute, &dev->flags))) {
3892                         if (test_bit(R5_Insync, &dev->flags))
3893                                 rmw++;
3894                         else
3895                                 rmw += 2*disks;  /* cannot read it */
3896                 }
3897                 /* Would I have to read this buffer for reconstruct_write */
3898                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3899                     i != sh->pd_idx && i != sh->qd_idx &&
3900                     !test_bit(R5_LOCKED, &dev->flags) &&
3901                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3902                       test_bit(R5_Wantcompute, &dev->flags))) {
3903                         if (test_bit(R5_Insync, &dev->flags))
3904                                 rcw++;
3905                         else
3906                                 rcw += 2*disks;
3907                 }
3908         }
3909
3910         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3911                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
3912         set_bit(STRIPE_HANDLE, &sh->state);
3913         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3914                 /* prefer read-modify-write, but need to get some data */
3915                 if (conf->mddev->queue)
3916                         blk_add_trace_msg(conf->mddev->queue,
3917                                           "raid5 rmw %llu %d",
3918                                           (unsigned long long)sh->sector, rmw);
3919                 for (i = disks; i--; ) {
3920                         struct r5dev *dev = &sh->dev[i];
3921                         if (test_bit(R5_InJournal, &dev->flags) &&
3922                             dev->page == dev->orig_page &&
3923                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3924                                 /* alloc page for prexor */
3925                                 struct page *p = alloc_page(GFP_NOIO);
3926
3927                                 if (p) {
3928                                         dev->orig_page = p;
3929                                         continue;
3930                                 }
3931
3932                                 /*
3933                                  * alloc_page() failed, try use
3934                                  * disk_info->extra_page
3935                                  */
3936                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3937                                                       &conf->cache_state)) {
3938                                         r5c_use_extra_page(sh);
3939                                         break;
3940                                 }
3941
3942                                 /* extra_page in use, add to delayed_list */
3943                                 set_bit(STRIPE_DELAYED, &sh->state);
3944                                 s->waiting_extra_page = 1;
3945                                 return -EAGAIN;
3946                         }
3947                 }
3948
3949                 for (i = disks; i--; ) {
3950                         struct r5dev *dev = &sh->dev[i];
3951                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3952                              i == sh->pd_idx || i == sh->qd_idx ||
3953                              test_bit(R5_InJournal, &dev->flags)) &&
3954                             !test_bit(R5_LOCKED, &dev->flags) &&
3955                             !(uptodate_for_rmw(dev) ||
3956                               test_bit(R5_Wantcompute, &dev->flags)) &&
3957                             test_bit(R5_Insync, &dev->flags)) {
3958                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3959                                              &sh->state)) {
3960                                         pr_debug("Read_old block %d for r-m-w\n",
3961                                                  i);
3962                                         set_bit(R5_LOCKED, &dev->flags);
3963                                         set_bit(R5_Wantread, &dev->flags);
3964                                         s->locked++;
3965                                 } else {
3966                                         set_bit(STRIPE_DELAYED, &sh->state);
3967                                         set_bit(STRIPE_HANDLE, &sh->state);
3968                                 }
3969                         }
3970                 }
3971         }
3972         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3973                 /* want reconstruct write, but need to get some data */
3974                 int qread =0;
3975                 rcw = 0;
3976                 for (i = disks; i--; ) {
3977                         struct r5dev *dev = &sh->dev[i];
3978                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3979                             i != sh->pd_idx && i != sh->qd_idx &&
3980                             !test_bit(R5_LOCKED, &dev->flags) &&
3981                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3982                               test_bit(R5_Wantcompute, &dev->flags))) {
3983                                 rcw++;
3984                                 if (test_bit(R5_Insync, &dev->flags) &&
3985                                     test_bit(STRIPE_PREREAD_ACTIVE,
3986                                              &sh->state)) {
3987                                         pr_debug("Read_old block "
3988                                                 "%d for Reconstruct\n", i);
3989                                         set_bit(R5_LOCKED, &dev->flags);
3990                                         set_bit(R5_Wantread, &dev->flags);
3991                                         s->locked++;
3992                                         qread++;
3993                                 } else {
3994                                         set_bit(STRIPE_DELAYED, &sh->state);
3995                                         set_bit(STRIPE_HANDLE, &sh->state);
3996                                 }
3997                         }
3998                 }
3999                 if (rcw && conf->mddev->queue)
4000                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4001                                           (unsigned long long)sh->sector,
4002                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4003         }
4004
4005         if (rcw > disks && rmw > disks &&
4006             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4007                 set_bit(STRIPE_DELAYED, &sh->state);
4008
4009         /* now if nothing is locked, and if we have enough data,
4010          * we can start a write request
4011          */
4012         /* since handle_stripe can be called at any time we need to handle the
4013          * case where a compute block operation has been submitted and then a
4014          * subsequent call wants to start a write request.  raid_run_ops only
4015          * handles the case where compute block and reconstruct are requested
4016          * simultaneously.  If this is not the case then new writes need to be
4017          * held off until the compute completes.
4018          */
4019         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4020             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4021              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4022                 schedule_reconstruction(sh, s, rcw == 0, 0);
4023         return 0;
4024 }
4025
4026 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4027                                 struct stripe_head_state *s, int disks)
4028 {
4029         struct r5dev *dev = NULL;
4030
4031         BUG_ON(sh->batch_head);
4032         set_bit(STRIPE_HANDLE, &sh->state);
4033
4034         switch (sh->check_state) {
4035         case check_state_idle:
4036                 /* start a new check operation if there are no failures */
4037                 if (s->failed == 0) {
4038                         BUG_ON(s->uptodate != disks);
4039                         sh->check_state = check_state_run;
4040                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4041                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4042                         s->uptodate--;
4043                         break;
4044                 }
4045                 dev = &sh->dev[s->failed_num[0]];
4046                 /* fall through */
4047         case check_state_compute_result:
4048                 sh->check_state = check_state_idle;
4049                 if (!dev)
4050                         dev = &sh->dev[sh->pd_idx];
4051
4052                 /* check that a write has not made the stripe insync */
4053                 if (test_bit(STRIPE_INSYNC, &sh->state))
4054                         break;
4055
4056                 /* either failed parity check, or recovery is happening */
4057                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4058                 BUG_ON(s->uptodate != disks);
4059
4060                 set_bit(R5_LOCKED, &dev->flags);
4061                 s->locked++;
4062                 set_bit(R5_Wantwrite, &dev->flags);
4063
4064                 clear_bit(STRIPE_DEGRADED, &sh->state);
4065                 set_bit(STRIPE_INSYNC, &sh->state);
4066                 break;
4067         case check_state_run:
4068                 break; /* we will be called again upon completion */
4069         case check_state_check_result:
4070                 sh->check_state = check_state_idle;
4071
4072                 /* if a failure occurred during the check operation, leave
4073                  * STRIPE_INSYNC not set and let the stripe be handled again
4074                  */
4075                 if (s->failed)
4076                         break;
4077
4078                 /* handle a successful check operation, if parity is correct
4079                  * we are done.  Otherwise update the mismatch count and repair
4080                  * parity if !MD_RECOVERY_CHECK
4081                  */
4082                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4083                         /* parity is correct (on disc,
4084                          * not in buffer any more)
4085                          */
4086                         set_bit(STRIPE_INSYNC, &sh->state);
4087                 else {
4088                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4089                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4090                                 /* don't try to repair!! */
4091                                 set_bit(STRIPE_INSYNC, &sh->state);
4092                                 pr_warn_ratelimited("%s: mismatch sector in range "
4093                                                     "%llu-%llu\n", mdname(conf->mddev),
4094                                                     (unsigned long long) sh->sector,
4095                                                     (unsigned long long) sh->sector +
4096                                                     STRIPE_SECTORS);
4097                         } else {
4098                                 sh->check_state = check_state_compute_run;
4099                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4100                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4101                                 set_bit(R5_Wantcompute,
4102                                         &sh->dev[sh->pd_idx].flags);
4103                                 sh->ops.target = sh->pd_idx;
4104                                 sh->ops.target2 = -1;
4105                                 s->uptodate++;
4106                         }
4107                 }
4108                 break;
4109         case check_state_compute_run:
4110                 break;
4111         default:
4112                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4113                        __func__, sh->check_state,
4114                        (unsigned long long) sh->sector);
4115                 BUG();
4116         }
4117 }
4118
4119 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4120                                   struct stripe_head_state *s,
4121                                   int disks)
4122 {
4123         int pd_idx = sh->pd_idx;
4124         int qd_idx = sh->qd_idx;
4125         struct r5dev *dev;
4126
4127         BUG_ON(sh->batch_head);
4128         set_bit(STRIPE_HANDLE, &sh->state);
4129
4130         BUG_ON(s->failed > 2);
4131
4132         /* Want to check and possibly repair P and Q.
4133          * However there could be one 'failed' device, in which
4134          * case we can only check one of them, possibly using the
4135          * other to generate missing data
4136          */
4137
4138         switch (sh->check_state) {
4139         case check_state_idle:
4140                 /* start a new check operation if there are < 2 failures */
4141                 if (s->failed == s->q_failed) {
4142                         /* The only possible failed device holds Q, so it
4143                          * makes sense to check P (If anything else were failed,
4144                          * we would have used P to recreate it).
4145                          */
4146                         sh->check_state = check_state_run;
4147                 }
4148                 if (!s->q_failed && s->failed < 2) {
4149                         /* Q is not failed, and we didn't use it to generate
4150                          * anything, so it makes sense to check it
4151                          */
4152                         if (sh->check_state == check_state_run)
4153                                 sh->check_state = check_state_run_pq;
4154                         else
4155                                 sh->check_state = check_state_run_q;
4156                 }
4157
4158                 /* discard potentially stale zero_sum_result */
4159                 sh->ops.zero_sum_result = 0;
4160
4161                 if (sh->check_state == check_state_run) {
4162                         /* async_xor_zero_sum destroys the contents of P */
4163                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4164                         s->uptodate--;
4165                 }
4166                 if (sh->check_state >= check_state_run &&
4167                     sh->check_state <= check_state_run_pq) {
4168                         /* async_syndrome_zero_sum preserves P and Q, so
4169                          * no need to mark them !uptodate here
4170                          */
4171                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4172                         break;
4173                 }
4174
4175                 /* we have 2-disk failure */
4176                 BUG_ON(s->failed != 2);
4177                 /* fall through */
4178         case check_state_compute_result:
4179                 sh->check_state = check_state_idle;
4180
4181                 /* check that a write has not made the stripe insync */
4182                 if (test_bit(STRIPE_INSYNC, &sh->state))
4183                         break;
4184
4185                 /* now write out any block on a failed drive,
4186                  * or P or Q if they were recomputed
4187                  */
4188                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
4189                 if (s->failed == 2) {
4190                         dev = &sh->dev[s->failed_num[1]];
4191                         s->locked++;
4192                         set_bit(R5_LOCKED, &dev->flags);
4193                         set_bit(R5_Wantwrite, &dev->flags);
4194                 }
4195                 if (s->failed >= 1) {
4196                         dev = &sh->dev[s->failed_num[0]];
4197                         s->locked++;
4198                         set_bit(R5_LOCKED, &dev->flags);
4199                         set_bit(R5_Wantwrite, &dev->flags);
4200                 }
4201                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4202                         dev = &sh->dev[pd_idx];
4203                         s->locked++;
4204                         set_bit(R5_LOCKED, &dev->flags);
4205                         set_bit(R5_Wantwrite, &dev->flags);
4206                 }
4207                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4208                         dev = &sh->dev[qd_idx];
4209                         s->locked++;
4210                         set_bit(R5_LOCKED, &dev->flags);
4211                         set_bit(R5_Wantwrite, &dev->flags);
4212                 }
4213                 clear_bit(STRIPE_DEGRADED, &sh->state);
4214
4215                 set_bit(STRIPE_INSYNC, &sh->state);
4216                 break;
4217         case check_state_run:
4218         case check_state_run_q:
4219         case check_state_run_pq:
4220                 break; /* we will be called again upon completion */
4221         case check_state_check_result:
4222                 sh->check_state = check_state_idle;
4223
4224                 /* handle a successful check operation, if parity is correct
4225                  * we are done.  Otherwise update the mismatch count and repair
4226                  * parity if !MD_RECOVERY_CHECK
4227                  */
4228                 if (sh->ops.zero_sum_result == 0) {
4229                         /* both parities are correct */
4230                         if (!s->failed)
4231                                 set_bit(STRIPE_INSYNC, &sh->state);
4232                         else {
4233                                 /* in contrast to the raid5 case we can validate
4234                                  * parity, but still have a failure to write
4235                                  * back
4236                                  */
4237                                 sh->check_state = check_state_compute_result;
4238                                 /* Returning at this point means that we may go
4239                                  * off and bring p and/or q uptodate again so
4240                                  * we make sure to check zero_sum_result again
4241                                  * to verify if p or q need writeback
4242                                  */
4243                         }
4244                 } else {
4245                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4246                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4247                                 /* don't try to repair!! */
4248                                 set_bit(STRIPE_INSYNC, &sh->state);
4249                                 pr_warn_ratelimited("%s: mismatch sector in range "
4250                                                     "%llu-%llu\n", mdname(conf->mddev),
4251                                                     (unsigned long long) sh->sector,
4252                                                     (unsigned long long) sh->sector +
4253                                                     STRIPE_SECTORS);
4254                         } else {
4255                                 int *target = &sh->ops.target;
4256
4257                                 sh->ops.target = -1;
4258                                 sh->ops.target2 = -1;
4259                                 sh->check_state = check_state_compute_run;
4260                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4261                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4262                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4263                                         set_bit(R5_Wantcompute,
4264                                                 &sh->dev[pd_idx].flags);
4265                                         *target = pd_idx;
4266                                         target = &sh->ops.target2;
4267                                         s->uptodate++;
4268                                 }
4269                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4270                                         set_bit(R5_Wantcompute,
4271                                                 &sh->dev[qd_idx].flags);
4272                                         *target = qd_idx;
4273                                         s->uptodate++;
4274                                 }
4275                         }
4276                 }
4277                 break;
4278         case check_state_compute_run:
4279                 break;
4280         default:
4281                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4282                         __func__, sh->check_state,
4283                         (unsigned long long) sh->sector);
4284                 BUG();
4285         }
4286 }
4287
4288 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4289 {
4290         int i;
4291
4292         /* We have read all the blocks in this stripe and now we need to
4293          * copy some of them into a target stripe for expand.
4294          */
4295         struct dma_async_tx_descriptor *tx = NULL;
4296         BUG_ON(sh->batch_head);
4297         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4298         for (i = 0; i < sh->disks; i++)
4299                 if (i != sh->pd_idx && i != sh->qd_idx) {
4300                         int dd_idx, j;
4301                         struct stripe_head *sh2;
4302                         struct async_submit_ctl submit;
4303
4304                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4305                         sector_t s = raid5_compute_sector(conf, bn, 0,
4306                                                           &dd_idx, NULL);
4307                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4308                         if (sh2 == NULL)
4309                                 /* so far only the early blocks of this stripe
4310                                  * have been requested.  When later blocks
4311                                  * get requested, we will try again
4312                                  */
4313                                 continue;
4314                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4315                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4316                                 /* must have already done this block */
4317                                 raid5_release_stripe(sh2);
4318                                 continue;
4319                         }
4320
4321                         /* place all the copies on one channel */
4322                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4323                         tx = async_memcpy(sh2->dev[dd_idx].page,
4324                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
4325                                           &submit);
4326
4327                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4328                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4329                         for (j = 0; j < conf->raid_disks; j++)
4330                                 if (j != sh2->pd_idx &&
4331                                     j != sh2->qd_idx &&
4332                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4333                                         break;
4334                         if (j == conf->raid_disks) {
4335                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4336                                 set_bit(STRIPE_HANDLE, &sh2->state);
4337                         }
4338                         raid5_release_stripe(sh2);
4339
4340                 }
4341         /* done submitting copies, wait for them to complete */
4342         async_tx_quiesce(&tx);
4343 }
4344
4345 /*
4346  * handle_stripe - do things to a stripe.
4347  *
4348  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4349  * state of various bits to see what needs to be done.
4350  * Possible results:
4351  *    return some read requests which now have data
4352  *    return some write requests which are safely on storage
4353  *    schedule a read on some buffers
4354  *    schedule a write of some buffers
4355  *    return confirmation of parity correctness
4356  *
4357  */
4358
4359 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4360 {
4361         struct r5conf *conf = sh->raid_conf;
4362         int disks = sh->disks;
4363         struct r5dev *dev;
4364         int i;
4365         int do_recovery = 0;
4366
4367         memset(s, 0, sizeof(*s));
4368
4369         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4370         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4371         s->failed_num[0] = -1;
4372         s->failed_num[1] = -1;
4373         s->log_failed = r5l_log_disk_error(conf);
4374
4375         /* Now to look around and see what can be done */
4376         rcu_read_lock();
4377         for (i=disks; i--; ) {
4378                 struct md_rdev *rdev;
4379                 sector_t first_bad;
4380                 int bad_sectors;
4381                 int is_bad = 0;
4382
4383                 dev = &sh->dev[i];
4384
4385                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4386                          i, dev->flags,
4387                          dev->toread, dev->towrite, dev->written);
4388                 /* maybe we can reply to a read
4389                  *
4390                  * new wantfill requests are only permitted while
4391                  * ops_complete_biofill is guaranteed to be inactive
4392                  */
4393                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4394                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4395                         set_bit(R5_Wantfill, &dev->flags);
4396
4397                 /* now count some things */
4398                 if (test_bit(R5_LOCKED, &dev->flags))
4399                         s->locked++;
4400                 if (test_bit(R5_UPTODATE, &dev->flags))
4401                         s->uptodate++;
4402                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4403                         s->compute++;
4404                         BUG_ON(s->compute > 2);
4405                 }
4406
4407                 if (test_bit(R5_Wantfill, &dev->flags))
4408                         s->to_fill++;
4409                 else if (dev->toread)
4410                         s->to_read++;
4411                 if (dev->towrite) {
4412                         s->to_write++;
4413                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4414                                 s->non_overwrite++;
4415                 }
4416                 if (dev->written)
4417                         s->written++;
4418                 /* Prefer to use the replacement for reads, but only
4419                  * if it is recovered enough and has no bad blocks.
4420                  */
4421                 rdev = rcu_dereference(conf->disks[i].replacement);
4422                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4423                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4424                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4425                                  &first_bad, &bad_sectors))
4426                         set_bit(R5_ReadRepl, &dev->flags);
4427                 else {
4428                         if (rdev && !test_bit(Faulty, &rdev->flags))
4429                                 set_bit(R5_NeedReplace, &dev->flags);
4430                         else
4431                                 clear_bit(R5_NeedReplace, &dev->flags);
4432                         rdev = rcu_dereference(conf->disks[i].rdev);
4433                         clear_bit(R5_ReadRepl, &dev->flags);
4434                 }
4435                 if (rdev && test_bit(Faulty, &rdev->flags))
4436                         rdev = NULL;
4437                 if (rdev) {
4438                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4439                                              &first_bad, &bad_sectors);
4440                         if (s->blocked_rdev == NULL
4441                             && (test_bit(Blocked, &rdev->flags)
4442                                 || is_bad < 0)) {
4443                                 if (is_bad < 0)
4444                                         set_bit(BlockedBadBlocks,
4445                                                 &rdev->flags);
4446                                 s->blocked_rdev = rdev;
4447                                 atomic_inc(&rdev->nr_pending);
4448                         }
4449                 }
4450                 clear_bit(R5_Insync, &dev->flags);
4451                 if (!rdev)
4452                         /* Not in-sync */;
4453                 else if (is_bad) {
4454                         /* also not in-sync */
4455                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4456                             test_bit(R5_UPTODATE, &dev->flags)) {
4457                                 /* treat as in-sync, but with a read error
4458                                  * which we can now try to correct
4459                                  */
4460                                 set_bit(R5_Insync, &dev->flags);
4461                                 set_bit(R5_ReadError, &dev->flags);
4462                         }
4463                 } else if (test_bit(In_sync, &rdev->flags))
4464                         set_bit(R5_Insync, &dev->flags);
4465                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4466                         /* in sync if before recovery_offset */
4467                         set_bit(R5_Insync, &dev->flags);
4468                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4469                          test_bit(R5_Expanded, &dev->flags))
4470                         /* If we've reshaped into here, we assume it is Insync.
4471                          * We will shortly update recovery_offset to make
4472                          * it official.
4473                          */
4474                         set_bit(R5_Insync, &dev->flags);
4475
4476                 if (test_bit(R5_WriteError, &dev->flags)) {
4477                         /* This flag does not apply to '.replacement'
4478                          * only to .rdev, so make sure to check that*/
4479                         struct md_rdev *rdev2 = rcu_dereference(
4480                                 conf->disks[i].rdev);
4481                         if (rdev2 == rdev)
4482                                 clear_bit(R5_Insync, &dev->flags);
4483                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4484                                 s->handle_bad_blocks = 1;
4485                                 atomic_inc(&rdev2->nr_pending);
4486                         } else
4487                                 clear_bit(R5_WriteError, &dev->flags);
4488                 }
4489                 if (test_bit(R5_MadeGood, &dev->flags)) {
4490                         /* This flag does not apply to '.replacement'
4491                          * only to .rdev, so make sure to check that*/
4492                         struct md_rdev *rdev2 = rcu_dereference(
4493                                 conf->disks[i].rdev);
4494                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4495                                 s->handle_bad_blocks = 1;
4496                                 atomic_inc(&rdev2->nr_pending);
4497                         } else
4498                                 clear_bit(R5_MadeGood, &dev->flags);
4499                 }
4500                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4501                         struct md_rdev *rdev2 = rcu_dereference(
4502                                 conf->disks[i].replacement);
4503                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4504                                 s->handle_bad_blocks = 1;
4505                                 atomic_inc(&rdev2->nr_pending);
4506                         } else
4507                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4508                 }
4509                 if (!test_bit(R5_Insync, &dev->flags)) {
4510                         /* The ReadError flag will just be confusing now */
4511                         clear_bit(R5_ReadError, &dev->flags);
4512                         clear_bit(R5_ReWrite, &dev->flags);
4513                 }
4514                 if (test_bit(R5_ReadError, &dev->flags))
4515                         clear_bit(R5_Insync, &dev->flags);
4516                 if (!test_bit(R5_Insync, &dev->flags)) {
4517                         if (s->failed < 2)
4518                                 s->failed_num[s->failed] = i;
4519                         s->failed++;
4520                         if (rdev && !test_bit(Faulty, &rdev->flags))
4521                                 do_recovery = 1;
4522                         else if (!rdev) {
4523                                 rdev = rcu_dereference(
4524                                     conf->disks[i].replacement);
4525                                 if (rdev && !test_bit(Faulty, &rdev->flags))
4526                                         do_recovery = 1;
4527                         }
4528                 }
4529
4530                 if (test_bit(R5_InJournal, &dev->flags))
4531                         s->injournal++;
4532                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4533                         s->just_cached++;
4534         }
4535         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4536                 /* If there is a failed device being replaced,
4537                  *     we must be recovering.
4538                  * else if we are after recovery_cp, we must be syncing
4539                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4540                  * else we can only be replacing
4541                  * sync and recovery both need to read all devices, and so
4542                  * use the same flag.
4543                  */
4544                 if (do_recovery ||
4545                     sh->sector >= conf->mddev->recovery_cp ||
4546                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4547                         s->syncing = 1;
4548                 else
4549                         s->replacing = 1;
4550         }
4551         rcu_read_unlock();
4552 }
4553
4554 static int clear_batch_ready(struct stripe_head *sh)
4555 {
4556         /* Return '1' if this is a member of batch, or
4557          * '0' if it is a lone stripe or a head which can now be
4558          * handled.
4559          */
4560         struct stripe_head *tmp;
4561         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4562                 return (sh->batch_head && sh->batch_head != sh);
4563         spin_lock(&sh->stripe_lock);
4564         if (!sh->batch_head) {
4565                 spin_unlock(&sh->stripe_lock);
4566                 return 0;
4567         }
4568
4569         /*
4570          * this stripe could be added to a batch list before we check
4571          * BATCH_READY, skips it
4572          */
4573         if (sh->batch_head != sh) {
4574                 spin_unlock(&sh->stripe_lock);
4575                 return 1;
4576         }
4577         spin_lock(&sh->batch_lock);
4578         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4579                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4580         spin_unlock(&sh->batch_lock);
4581         spin_unlock(&sh->stripe_lock);
4582
4583         /*
4584          * BATCH_READY is cleared, no new stripes can be added.
4585          * batch_list can be accessed without lock
4586          */
4587         return 0;
4588 }
4589
4590 static void break_stripe_batch_list(struct stripe_head *head_sh,
4591                                     unsigned long handle_flags)
4592 {
4593         struct stripe_head *sh, *next;
4594         int i;
4595         int do_wakeup = 0;
4596
4597         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4598
4599                 list_del_init(&sh->batch_list);
4600
4601                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4602                                           (1 << STRIPE_SYNCING) |
4603                                           (1 << STRIPE_REPLACED) |
4604                                           (1 << STRIPE_DELAYED) |
4605                                           (1 << STRIPE_BIT_DELAY) |
4606                                           (1 << STRIPE_FULL_WRITE) |
4607                                           (1 << STRIPE_BIOFILL_RUN) |
4608                                           (1 << STRIPE_COMPUTE_RUN)  |
4609                                           (1 << STRIPE_OPS_REQ_PENDING) |
4610                                           (1 << STRIPE_DISCARD) |
4611                                           (1 << STRIPE_BATCH_READY) |
4612                                           (1 << STRIPE_BATCH_ERR) |
4613                                           (1 << STRIPE_BITMAP_PENDING)),
4614                         "stripe state: %lx\n", sh->state);
4615                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4616                                               (1 << STRIPE_REPLACED)),
4617                         "head stripe state: %lx\n", head_sh->state);
4618
4619                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4620                                             (1 << STRIPE_PREREAD_ACTIVE) |
4621                                             (1 << STRIPE_DEGRADED) |
4622                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4623                               head_sh->state & (1 << STRIPE_INSYNC));
4624
4625                 sh->check_state = head_sh->check_state;
4626                 sh->reconstruct_state = head_sh->reconstruct_state;
4627                 spin_lock_irq(&sh->stripe_lock);
4628                 sh->batch_head = NULL;
4629                 spin_unlock_irq(&sh->stripe_lock);
4630                 for (i = 0; i < sh->disks; i++) {
4631                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4632                                 do_wakeup = 1;
4633                         sh->dev[i].flags = head_sh->dev[i].flags &
4634                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4635                 }
4636                 if (handle_flags == 0 ||
4637                     sh->state & handle_flags)
4638                         set_bit(STRIPE_HANDLE, &sh->state);
4639                 raid5_release_stripe(sh);
4640         }
4641         spin_lock_irq(&head_sh->stripe_lock);
4642         head_sh->batch_head = NULL;
4643         spin_unlock_irq(&head_sh->stripe_lock);
4644         for (i = 0; i < head_sh->disks; i++)
4645                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4646                         do_wakeup = 1;
4647         if (head_sh->state & handle_flags)
4648                 set_bit(STRIPE_HANDLE, &head_sh->state);
4649
4650         if (do_wakeup)
4651                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4652 }
4653
4654 static void handle_stripe(struct stripe_head *sh)
4655 {
4656         struct stripe_head_state s;
4657         struct r5conf *conf = sh->raid_conf;
4658         int i;
4659         int prexor;
4660         int disks = sh->disks;
4661         struct r5dev *pdev, *qdev;
4662
4663         clear_bit(STRIPE_HANDLE, &sh->state);
4664         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4665                 /* already being handled, ensure it gets handled
4666                  * again when current action finishes */
4667                 set_bit(STRIPE_HANDLE, &sh->state);
4668                 return;
4669         }
4670
4671         if (clear_batch_ready(sh) ) {
4672                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4673                 return;
4674         }
4675
4676         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4677                 break_stripe_batch_list(sh, 0);
4678
4679         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4680                 spin_lock(&sh->stripe_lock);
4681                 /*
4682                  * Cannot process 'sync' concurrently with 'discard'.
4683                  * Flush data in r5cache before 'sync'.
4684                  */
4685                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4686                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4687                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4688                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4689                         set_bit(STRIPE_SYNCING, &sh->state);
4690                         clear_bit(STRIPE_INSYNC, &sh->state);
4691                         clear_bit(STRIPE_REPLACED, &sh->state);
4692                 }
4693                 spin_unlock(&sh->stripe_lock);
4694         }
4695         clear_bit(STRIPE_DELAYED, &sh->state);
4696
4697         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4698                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4699                (unsigned long long)sh->sector, sh->state,
4700                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4701                sh->check_state, sh->reconstruct_state);
4702
4703         analyse_stripe(sh, &s);
4704
4705         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4706                 goto finish;
4707
4708         if (s.handle_bad_blocks ||
4709             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4710                 set_bit(STRIPE_HANDLE, &sh->state);
4711                 goto finish;
4712         }
4713
4714         if (unlikely(s.blocked_rdev)) {
4715                 if (s.syncing || s.expanding || s.expanded ||
4716                     s.replacing || s.to_write || s.written) {
4717                         set_bit(STRIPE_HANDLE, &sh->state);
4718                         goto finish;
4719                 }
4720                 /* There is nothing for the blocked_rdev to block */
4721                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4722                 s.blocked_rdev = NULL;
4723         }
4724
4725         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4726                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4727                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4728         }
4729
4730         pr_debug("locked=%d uptodate=%d to_read=%d"
4731                " to_write=%d failed=%d failed_num=%d,%d\n",
4732                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4733                s.failed_num[0], s.failed_num[1]);
4734         /*
4735          * check if the array has lost more than max_degraded devices and,
4736          * if so, some requests might need to be failed.
4737          *
4738          * When journal device failed (log_failed), we will only process
4739          * the stripe if there is data need write to raid disks
4740          */
4741         if (s.failed > conf->max_degraded ||
4742             (s.log_failed && s.injournal == 0)) {
4743                 sh->check_state = 0;
4744                 sh->reconstruct_state = 0;
4745                 break_stripe_batch_list(sh, 0);
4746                 if (s.to_read+s.to_write+s.written)
4747                         handle_failed_stripe(conf, sh, &s, disks);
4748                 if (s.syncing + s.replacing)
4749                         handle_failed_sync(conf, sh, &s);
4750         }
4751
4752         /* Now we check to see if any write operations have recently
4753          * completed
4754          */
4755         prexor = 0;
4756         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4757                 prexor = 1;
4758         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4759             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4760                 sh->reconstruct_state = reconstruct_state_idle;
4761
4762                 /* All the 'written' buffers and the parity block are ready to
4763                  * be written back to disk
4764                  */
4765                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4766                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4767                 BUG_ON(sh->qd_idx >= 0 &&
4768                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4769                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4770                 for (i = disks; i--; ) {
4771                         struct r5dev *dev = &sh->dev[i];
4772                         if (test_bit(R5_LOCKED, &dev->flags) &&
4773                                 (i == sh->pd_idx || i == sh->qd_idx ||
4774                                  dev->written || test_bit(R5_InJournal,
4775                                                           &dev->flags))) {
4776                                 pr_debug("Writing block %d\n", i);
4777                                 set_bit(R5_Wantwrite, &dev->flags);
4778                                 if (prexor)
4779                                         continue;
4780                                 if (s.failed > 1)
4781                                         continue;
4782                                 if (!test_bit(R5_Insync, &dev->flags) ||
4783                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4784                                      s.failed == 0))
4785                                         set_bit(STRIPE_INSYNC, &sh->state);
4786                         }
4787                 }
4788                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4789                         s.dec_preread_active = 1;
4790         }
4791
4792         /*
4793          * might be able to return some write requests if the parity blocks
4794          * are safe, or on a failed drive
4795          */
4796         pdev = &sh->dev[sh->pd_idx];
4797         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4798                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4799         qdev = &sh->dev[sh->qd_idx];
4800         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4801                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4802                 || conf->level < 6;
4803
4804         if (s.written &&
4805             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4806                              && !test_bit(R5_LOCKED, &pdev->flags)
4807                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4808                                  test_bit(R5_Discard, &pdev->flags))))) &&
4809             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4810                              && !test_bit(R5_LOCKED, &qdev->flags)
4811                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4812                                  test_bit(R5_Discard, &qdev->flags))))))
4813                 handle_stripe_clean_event(conf, sh, disks);
4814
4815         if (s.just_cached)
4816                 r5c_handle_cached_data_endio(conf, sh, disks);
4817         log_stripe_write_finished(sh);
4818
4819         /* Now we might consider reading some blocks, either to check/generate
4820          * parity, or to satisfy requests
4821          * or to load a block that is being partially written.
4822          */
4823         if (s.to_read || s.non_overwrite
4824             || (conf->level == 6 && s.to_write && s.failed)
4825             || (s.syncing && (s.uptodate + s.compute < disks))
4826             || s.replacing
4827             || s.expanding)
4828                 handle_stripe_fill(sh, &s, disks);
4829
4830         /*
4831          * When the stripe finishes full journal write cycle (write to journal
4832          * and raid disk), this is the clean up procedure so it is ready for
4833          * next operation.
4834          */
4835         r5c_finish_stripe_write_out(conf, sh, &s);
4836
4837         /*
4838          * Now to consider new write requests, cache write back and what else,
4839          * if anything should be read.  We do not handle new writes when:
4840          * 1/ A 'write' operation (copy+xor) is already in flight.
4841          * 2/ A 'check' operation is in flight, as it may clobber the parity
4842          *    block.
4843          * 3/ A r5c cache log write is in flight.
4844          */
4845
4846         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4847                 if (!r5c_is_writeback(conf->log)) {
4848                         if (s.to_write)
4849                                 handle_stripe_dirtying(conf, sh, &s, disks);
4850                 } else { /* write back cache */
4851                         int ret = 0;
4852
4853                         /* First, try handle writes in caching phase */
4854                         if (s.to_write)
4855                                 ret = r5c_try_caching_write(conf, sh, &s,
4856                                                             disks);
4857                         /*
4858                          * If caching phase failed: ret == -EAGAIN
4859                          *    OR
4860                          * stripe under reclaim: !caching && injournal
4861                          *
4862                          * fall back to handle_stripe_dirtying()
4863                          */
4864                         if (ret == -EAGAIN ||
4865                             /* stripe under reclaim: !caching && injournal */
4866                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4867                              s.injournal > 0)) {
4868                                 ret = handle_stripe_dirtying(conf, sh, &s,
4869                                                              disks);
4870                                 if (ret == -EAGAIN)
4871                                         goto finish;
4872                         }
4873                 }
4874         }
4875
4876         /* maybe we need to check and possibly fix the parity for this stripe
4877          * Any reads will already have been scheduled, so we just see if enough
4878          * data is available.  The parity check is held off while parity
4879          * dependent operations are in flight.
4880          */
4881         if (sh->check_state ||
4882             (s.syncing && s.locked == 0 &&
4883              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4884              !test_bit(STRIPE_INSYNC, &sh->state))) {
4885                 if (conf->level == 6)
4886                         handle_parity_checks6(conf, sh, &s, disks);
4887                 else
4888                         handle_parity_checks5(conf, sh, &s, disks);
4889         }
4890
4891         if ((s.replacing || s.syncing) && s.locked == 0
4892             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4893             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4894                 /* Write out to replacement devices where possible */
4895                 for (i = 0; i < conf->raid_disks; i++)
4896                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4897                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4898                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4899                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4900                                 s.locked++;
4901                         }
4902                 if (s.replacing)
4903                         set_bit(STRIPE_INSYNC, &sh->state);
4904                 set_bit(STRIPE_REPLACED, &sh->state);
4905         }
4906         if ((s.syncing || s.replacing) && s.locked == 0 &&
4907             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4908             test_bit(STRIPE_INSYNC, &sh->state)) {
4909                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4910                 clear_bit(STRIPE_SYNCING, &sh->state);
4911                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4912                         wake_up(&conf->wait_for_overlap);
4913         }
4914
4915         /* If the failed drives are just a ReadError, then we might need
4916          * to progress the repair/check process
4917          */
4918         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4919                 for (i = 0; i < s.failed; i++) {
4920                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4921                         if (test_bit(R5_ReadError, &dev->flags)
4922                             && !test_bit(R5_LOCKED, &dev->flags)
4923                             && test_bit(R5_UPTODATE, &dev->flags)
4924                                 ) {
4925                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4926                                         set_bit(R5_Wantwrite, &dev->flags);
4927                                         set_bit(R5_ReWrite, &dev->flags);
4928                                         set_bit(R5_LOCKED, &dev->flags);
4929                                         s.locked++;
4930                                 } else {
4931                                         /* let's read it back */
4932                                         set_bit(R5_Wantread, &dev->flags);
4933                                         set_bit(R5_LOCKED, &dev->flags);
4934                                         s.locked++;
4935                                 }
4936                         }
4937                 }
4938
4939         /* Finish reconstruct operations initiated by the expansion process */
4940         if (sh->reconstruct_state == reconstruct_state_result) {
4941                 struct stripe_head *sh_src
4942                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4943                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4944                         /* sh cannot be written until sh_src has been read.
4945                          * so arrange for sh to be delayed a little
4946                          */
4947                         set_bit(STRIPE_DELAYED, &sh->state);
4948                         set_bit(STRIPE_HANDLE, &sh->state);
4949                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4950                                               &sh_src->state))
4951                                 atomic_inc(&conf->preread_active_stripes);
4952                         raid5_release_stripe(sh_src);
4953                         goto finish;
4954                 }
4955                 if (sh_src)
4956                         raid5_release_stripe(sh_src);
4957
4958                 sh->reconstruct_state = reconstruct_state_idle;
4959                 clear_bit(STRIPE_EXPANDING, &sh->state);
4960                 for (i = conf->raid_disks; i--; ) {
4961                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4962                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4963                         s.locked++;
4964                 }
4965         }
4966
4967         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4968             !sh->reconstruct_state) {
4969                 /* Need to write out all blocks after computing parity */
4970                 sh->disks = conf->raid_disks;
4971                 stripe_set_idx(sh->sector, conf, 0, sh);
4972                 schedule_reconstruction(sh, &s, 1, 1);
4973         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4974                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4975                 atomic_dec(&conf->reshape_stripes);
4976                 wake_up(&conf->wait_for_overlap);
4977                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4978         }
4979
4980         if (s.expanding && s.locked == 0 &&
4981             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4982                 handle_stripe_expansion(conf, sh);
4983
4984 finish:
4985         /* wait for this device to become unblocked */
4986         if (unlikely(s.blocked_rdev)) {
4987                 if (conf->mddev->external)
4988                         md_wait_for_blocked_rdev(s.blocked_rdev,
4989                                                  conf->mddev);
4990                 else
4991                         /* Internal metadata will immediately
4992                          * be written by raid5d, so we don't
4993                          * need to wait here.
4994                          */
4995                         rdev_dec_pending(s.blocked_rdev,
4996                                          conf->mddev);
4997         }
4998
4999         if (s.handle_bad_blocks)
5000                 for (i = disks; i--; ) {
5001                         struct md_rdev *rdev;
5002                         struct r5dev *dev = &sh->dev[i];
5003                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5004                                 /* We own a safe reference to the rdev */
5005                                 rdev = conf->disks[i].rdev;
5006                                 if (!rdev_set_badblocks(rdev, sh->sector,
5007                                                         STRIPE_SECTORS, 0))
5008                                         md_error(conf->mddev, rdev);
5009                                 rdev_dec_pending(rdev, conf->mddev);
5010                         }
5011                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5012                                 rdev = conf->disks[i].rdev;
5013                                 rdev_clear_badblocks(rdev, sh->sector,
5014                                                      STRIPE_SECTORS, 0);
5015                                 rdev_dec_pending(rdev, conf->mddev);
5016                         }
5017                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5018                                 rdev = conf->disks[i].replacement;
5019                                 if (!rdev)
5020                                         /* rdev have been moved down */
5021                                         rdev = conf->disks[i].rdev;
5022                                 rdev_clear_badblocks(rdev, sh->sector,
5023                                                      STRIPE_SECTORS, 0);
5024                                 rdev_dec_pending(rdev, conf->mddev);
5025                         }
5026                 }
5027
5028         if (s.ops_request)
5029                 raid_run_ops(sh, s.ops_request);
5030
5031         ops_run_io(sh, &s);
5032
5033         if (s.dec_preread_active) {
5034                 /* We delay this until after ops_run_io so that if make_request
5035                  * is waiting on a flush, it won't continue until the writes
5036                  * have actually been submitted.
5037                  */
5038                 atomic_dec(&conf->preread_active_stripes);
5039                 if (atomic_read(&conf->preread_active_stripes) <
5040                     IO_THRESHOLD)
5041                         md_wakeup_thread(conf->mddev->thread);
5042         }
5043
5044         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5045 }
5046
5047 static void raid5_activate_delayed(struct r5conf *conf)
5048 {
5049         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5050                 while (!list_empty(&conf->delayed_list)) {
5051                         struct list_head *l = conf->delayed_list.next;
5052                         struct stripe_head *sh;
5053                         sh = list_entry(l, struct stripe_head, lru);
5054                         list_del_init(l);
5055                         clear_bit(STRIPE_DELAYED, &sh->state);
5056                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5057                                 atomic_inc(&conf->preread_active_stripes);
5058                         list_add_tail(&sh->lru, &conf->hold_list);
5059                         raid5_wakeup_stripe_thread(sh);
5060                 }
5061         }
5062 }
5063
5064 static void activate_bit_delay(struct r5conf *conf,
5065         struct list_head *temp_inactive_list)
5066 {
5067         /* device_lock is held */
5068         struct list_head head;
5069         list_add(&head, &conf->bitmap_list);
5070         list_del_init(&conf->bitmap_list);
5071         while (!list_empty(&head)) {
5072                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5073                 int hash;
5074                 list_del_init(&sh->lru);
5075                 atomic_inc(&sh->count);
5076                 hash = sh->hash_lock_index;
5077                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5078         }
5079 }
5080
5081 static int raid5_congested(struct mddev *mddev, int bits)
5082 {
5083         struct r5conf *conf = mddev->private;
5084
5085         /* No difference between reads and writes.  Just check
5086          * how busy the stripe_cache is
5087          */
5088
5089         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5090                 return 1;
5091
5092         /* Also checks whether there is pressure on r5cache log space */
5093         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5094                 return 1;
5095         if (conf->quiesce)
5096                 return 1;
5097         if (atomic_read(&conf->empty_inactive_list_nr))
5098                 return 1;
5099
5100         return 0;
5101 }
5102
5103 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5104 {
5105         struct r5conf *conf = mddev->private;
5106         sector_t sector = bio->bi_iter.bi_sector;
5107         unsigned int chunk_sectors;
5108         unsigned int bio_sectors = bio_sectors(bio);
5109
5110         WARN_ON_ONCE(bio->bi_partno);
5111
5112         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5113         return  chunk_sectors >=
5114                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5115 }
5116
5117 /*
5118  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5119  *  later sampled by raid5d.
5120  */
5121 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5122 {
5123         unsigned long flags;
5124
5125         spin_lock_irqsave(&conf->device_lock, flags);
5126
5127         bi->bi_next = conf->retry_read_aligned_list;
5128         conf->retry_read_aligned_list = bi;
5129
5130         spin_unlock_irqrestore(&conf->device_lock, flags);
5131         md_wakeup_thread(conf->mddev->thread);
5132 }
5133
5134 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5135                                          unsigned int *offset)
5136 {
5137         struct bio *bi;
5138
5139         bi = conf->retry_read_aligned;
5140         if (bi) {
5141                 *offset = conf->retry_read_offset;
5142                 conf->retry_read_aligned = NULL;
5143                 return bi;
5144         }
5145         bi = conf->retry_read_aligned_list;
5146         if(bi) {
5147                 conf->retry_read_aligned_list = bi->bi_next;
5148                 bi->bi_next = NULL;
5149                 *offset = 0;
5150         }
5151
5152         return bi;
5153 }
5154
5155 /*
5156  *  The "raid5_align_endio" should check if the read succeeded and if it
5157  *  did, call bio_endio on the original bio (having bio_put the new bio
5158  *  first).
5159  *  If the read failed..
5160  */
5161 static void raid5_align_endio(struct bio *bi)
5162 {
5163         struct bio* raid_bi  = bi->bi_private;
5164         struct mddev *mddev;
5165         struct r5conf *conf;
5166         struct md_rdev *rdev;
5167         blk_status_t error = bi->bi_status;
5168
5169         bio_put(bi);
5170
5171         rdev = (void*)raid_bi->bi_next;
5172         raid_bi->bi_next = NULL;
5173         mddev = rdev->mddev;
5174         conf = mddev->private;
5175
5176         rdev_dec_pending(rdev, conf->mddev);
5177
5178         if (!error) {
5179                 bio_endio(raid_bi);
5180                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5181                         wake_up(&conf->wait_for_quiescent);
5182                 return;
5183         }
5184
5185         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5186
5187         add_bio_to_retry(raid_bi, conf);
5188 }
5189
5190 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5191 {
5192         struct r5conf *conf = mddev->private;
5193         int dd_idx;
5194         struct bio* align_bi;
5195         struct md_rdev *rdev;
5196         sector_t end_sector;
5197
5198         if (!in_chunk_boundary(mddev, raid_bio)) {
5199                 pr_debug("%s: non aligned\n", __func__);
5200                 return 0;
5201         }
5202         /*
5203          * use bio_clone_fast to make a copy of the bio
5204          */
5205         align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5206         if (!align_bi)
5207                 return 0;
5208         /*
5209          *   set bi_end_io to a new function, and set bi_private to the
5210          *     original bio.
5211          */
5212         align_bi->bi_end_io  = raid5_align_endio;
5213         align_bi->bi_private = raid_bio;
5214         /*
5215          *      compute position
5216          */
5217         align_bi->bi_iter.bi_sector =
5218                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5219                                      0, &dd_idx, NULL);
5220
5221         end_sector = bio_end_sector(align_bi);
5222         rcu_read_lock();
5223         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5224         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5225             rdev->recovery_offset < end_sector) {
5226                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5227                 if (rdev &&
5228                     (test_bit(Faulty, &rdev->flags) ||
5229                     !(test_bit(In_sync, &rdev->flags) ||
5230                       rdev->recovery_offset >= end_sector)))
5231                         rdev = NULL;
5232         }
5233
5234         if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5235                 rcu_read_unlock();
5236                 bio_put(align_bi);
5237                 return 0;
5238         }
5239
5240         if (rdev) {
5241                 sector_t first_bad;
5242                 int bad_sectors;
5243
5244                 atomic_inc(&rdev->nr_pending);
5245                 rcu_read_unlock();
5246                 raid_bio->bi_next = (void*)rdev;
5247                 bio_set_dev(align_bi, rdev->bdev);
5248                 bio_clear_flag(align_bi, BIO_SEG_VALID);
5249
5250                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5251                                 bio_sectors(align_bi),
5252                                 &first_bad, &bad_sectors)) {
5253                         bio_put(align_bi);
5254                         rdev_dec_pending(rdev, mddev);
5255                         return 0;
5256                 }
5257
5258                 /* No reshape active, so we can trust rdev->data_offset */
5259                 align_bi->bi_iter.bi_sector += rdev->data_offset;
5260
5261                 spin_lock_irq(&conf->device_lock);
5262                 wait_event_lock_irq(conf->wait_for_quiescent,
5263                                     conf->quiesce == 0,
5264                                     conf->device_lock);
5265                 atomic_inc(&conf->active_aligned_reads);
5266                 spin_unlock_irq(&conf->device_lock);
5267
5268                 if (mddev->gendisk)
5269                         trace_block_bio_remap(align_bi->bi_disk->queue,
5270                                               align_bi, disk_devt(mddev->gendisk),
5271                                               raid_bio->bi_iter.bi_sector);
5272                 generic_make_request(align_bi);
5273                 return 1;
5274         } else {
5275                 rcu_read_unlock();
5276                 bio_put(align_bi);
5277                 return 0;
5278         }
5279 }
5280
5281 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5282 {
5283         struct bio *split;
5284         sector_t sector = raid_bio->bi_iter.bi_sector;
5285         unsigned chunk_sects = mddev->chunk_sectors;
5286         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5287
5288         if (sectors < bio_sectors(raid_bio)) {
5289                 struct r5conf *conf = mddev->private;
5290                 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5291                 bio_chain(split, raid_bio);
5292                 generic_make_request(raid_bio);
5293                 raid_bio = split;
5294         }
5295
5296         if (!raid5_read_one_chunk(mddev, raid_bio))
5297                 return raid_bio;
5298
5299         return NULL;
5300 }
5301
5302 /* __get_priority_stripe - get the next stripe to process
5303  *
5304  * Full stripe writes are allowed to pass preread active stripes up until
5305  * the bypass_threshold is exceeded.  In general the bypass_count
5306  * increments when the handle_list is handled before the hold_list; however, it
5307  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5308  * stripe with in flight i/o.  The bypass_count will be reset when the
5309  * head of the hold_list has changed, i.e. the head was promoted to the
5310  * handle_list.
5311  */
5312 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5313 {
5314         struct stripe_head *sh, *tmp;
5315         struct list_head *handle_list = NULL;
5316         struct r5worker_group *wg;
5317         bool second_try = !r5c_is_writeback(conf->log) &&
5318                 !r5l_log_disk_error(conf);
5319         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5320                 r5l_log_disk_error(conf);
5321
5322 again:
5323         wg = NULL;
5324         sh = NULL;
5325         if (conf->worker_cnt_per_group == 0) {
5326                 handle_list = try_loprio ? &conf->loprio_list :
5327                                         &conf->handle_list;
5328         } else if (group != ANY_GROUP) {
5329                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5330                                 &conf->worker_groups[group].handle_list;
5331                 wg = &conf->worker_groups[group];
5332         } else {
5333                 int i;
5334                 for (i = 0; i < conf->group_cnt; i++) {
5335                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5336                                 &conf->worker_groups[i].handle_list;
5337                         wg = &conf->worker_groups[i];
5338                         if (!list_empty(handle_list))
5339                                 break;
5340                 }
5341         }
5342
5343         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5344                   __func__,
5345                   list_empty(handle_list) ? "empty" : "busy",
5346                   list_empty(&conf->hold_list) ? "empty" : "busy",
5347                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5348
5349         if (!list_empty(handle_list)) {
5350                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5351
5352                 if (list_empty(&conf->hold_list))
5353                         conf->bypass_count = 0;
5354                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5355                         if (conf->hold_list.next == conf->last_hold)
5356                                 conf->bypass_count++;
5357                         else {
5358                                 conf->last_hold = conf->hold_list.next;
5359                                 conf->bypass_count -= conf->bypass_threshold;
5360                                 if (conf->bypass_count < 0)
5361                                         conf->bypass_count = 0;
5362                         }
5363                 }
5364         } else if (!list_empty(&conf->hold_list) &&
5365                    ((conf->bypass_threshold &&
5366                      conf->bypass_count > conf->bypass_threshold) ||
5367                     atomic_read(&conf->pending_full_writes) == 0)) {
5368
5369                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5370                         if (conf->worker_cnt_per_group == 0 ||
5371                             group == ANY_GROUP ||
5372                             !cpu_online(tmp->cpu) ||
5373                             cpu_to_group(tmp->cpu) == group) {
5374                                 sh = tmp;
5375                                 break;
5376                         }
5377                 }
5378
5379                 if (sh) {
5380                         conf->bypass_count -= conf->bypass_threshold;
5381                         if (conf->bypass_count < 0)
5382                                 conf->bypass_count = 0;
5383                 }
5384                 wg = NULL;
5385         }
5386
5387         if (!sh) {
5388                 if (second_try)
5389                         return NULL;
5390                 second_try = true;
5391                 try_loprio = !try_loprio;
5392                 goto again;
5393         }
5394
5395         if (wg) {
5396                 wg->stripes_cnt--;
5397                 sh->group = NULL;
5398         }
5399         list_del_init(&sh->lru);
5400         BUG_ON(atomic_inc_return(&sh->count) != 1);
5401         return sh;
5402 }
5403
5404 struct raid5_plug_cb {
5405         struct blk_plug_cb      cb;
5406         struct list_head        list;
5407         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5408 };
5409
5410 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5411 {
5412         struct raid5_plug_cb *cb = container_of(
5413                 blk_cb, struct raid5_plug_cb, cb);
5414         struct stripe_head *sh;
5415         struct mddev *mddev = cb->cb.data;
5416         struct r5conf *conf = mddev->private;
5417         int cnt = 0;
5418         int hash;
5419
5420         if (cb->list.next && !list_empty(&cb->list)) {
5421                 spin_lock_irq(&conf->device_lock);
5422                 while (!list_empty(&cb->list)) {
5423                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5424                         list_del_init(&sh->lru);
5425                         /*
5426                          * avoid race release_stripe_plug() sees
5427                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5428                          * is still in our list
5429                          */
5430                         smp_mb__before_atomic();
5431                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5432                         /*
5433                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5434                          * case, the count is always > 1 here
5435                          */
5436                         hash = sh->hash_lock_index;
5437                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5438                         cnt++;
5439                 }
5440                 spin_unlock_irq(&conf->device_lock);
5441         }
5442         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5443                                      NR_STRIPE_HASH_LOCKS);
5444         if (mddev->queue)
5445                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5446         kfree(cb);
5447 }
5448
5449 static void release_stripe_plug(struct mddev *mddev,
5450                                 struct stripe_head *sh)
5451 {
5452         struct blk_plug_cb *blk_cb = blk_check_plugged(
5453                 raid5_unplug, mddev,
5454                 sizeof(struct raid5_plug_cb));
5455         struct raid5_plug_cb *cb;
5456
5457         if (!blk_cb) {
5458                 raid5_release_stripe(sh);
5459                 return;
5460         }
5461
5462         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5463
5464         if (cb->list.next == NULL) {
5465                 int i;
5466                 INIT_LIST_HEAD(&cb->list);
5467                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5468                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5469         }
5470
5471         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5472                 list_add_tail(&sh->lru, &cb->list);
5473         else
5474                 raid5_release_stripe(sh);
5475 }
5476
5477 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5478 {
5479         struct r5conf *conf = mddev->private;
5480         sector_t logical_sector, last_sector;
5481         struct stripe_head *sh;
5482         int stripe_sectors;
5483
5484         if (mddev->reshape_position != MaxSector)
5485                 /* Skip discard while reshape is happening */
5486                 return;
5487
5488         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5489         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5490
5491         bi->bi_next = NULL;
5492
5493         stripe_sectors = conf->chunk_sectors *
5494                 (conf->raid_disks - conf->max_degraded);
5495         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5496                                                stripe_sectors);
5497         sector_div(last_sector, stripe_sectors);
5498
5499         logical_sector *= conf->chunk_sectors;
5500         last_sector *= conf->chunk_sectors;
5501
5502         for (; logical_sector < last_sector;
5503              logical_sector += STRIPE_SECTORS) {
5504                 DEFINE_WAIT(w);
5505                 int d;
5506         again:
5507                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5508                 prepare_to_wait(&conf->wait_for_overlap, &w,
5509                                 TASK_UNINTERRUPTIBLE);
5510                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5511                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5512                         raid5_release_stripe(sh);
5513                         schedule();
5514                         goto again;
5515                 }
5516                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5517                 spin_lock_irq(&sh->stripe_lock);
5518                 for (d = 0; d < conf->raid_disks; d++) {
5519                         if (d == sh->pd_idx || d == sh->qd_idx)
5520                                 continue;
5521                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5522                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5523                                 spin_unlock_irq(&sh->stripe_lock);
5524                                 raid5_release_stripe(sh);
5525                                 schedule();
5526                                 goto again;
5527                         }
5528                 }
5529                 set_bit(STRIPE_DISCARD, &sh->state);
5530                 finish_wait(&conf->wait_for_overlap, &w);
5531                 sh->overwrite_disks = 0;
5532                 for (d = 0; d < conf->raid_disks; d++) {
5533                         if (d == sh->pd_idx || d == sh->qd_idx)
5534                                 continue;
5535                         sh->dev[d].towrite = bi;
5536                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5537                         bio_inc_remaining(bi);
5538                         md_write_inc(mddev, bi);
5539                         sh->overwrite_disks++;
5540                 }
5541                 spin_unlock_irq(&sh->stripe_lock);
5542                 if (conf->mddev->bitmap) {
5543                         for (d = 0;
5544                              d < conf->raid_disks - conf->max_degraded;
5545                              d++)
5546                                 md_bitmap_startwrite(mddev->bitmap,
5547                                                      sh->sector,
5548                                                      STRIPE_SECTORS,
5549                                                      0);
5550                         sh->bm_seq = conf->seq_flush + 1;
5551                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5552                 }
5553
5554                 set_bit(STRIPE_HANDLE, &sh->state);
5555                 clear_bit(STRIPE_DELAYED, &sh->state);
5556                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5557                         atomic_inc(&conf->preread_active_stripes);
5558                 release_stripe_plug(mddev, sh);
5559         }
5560
5561         bio_endio(bi);
5562 }
5563
5564 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5565 {
5566         struct r5conf *conf = mddev->private;
5567         int dd_idx;
5568         sector_t new_sector;
5569         sector_t logical_sector, last_sector;
5570         struct stripe_head *sh;
5571         const int rw = bio_data_dir(bi);
5572         DEFINE_WAIT(w);
5573         bool do_prepare;
5574         bool do_flush = false;
5575
5576         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5577                 int ret = log_handle_flush_request(conf, bi);
5578
5579                 if (ret == 0)
5580                         return true;
5581                 if (ret == -ENODEV) {
5582                         md_flush_request(mddev, bi);
5583                         return true;
5584                 }
5585                 /* ret == -EAGAIN, fallback */
5586                 /*
5587                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5588                  * we need to flush journal device
5589                  */
5590                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5591         }
5592
5593         if (!md_write_start(mddev, bi))
5594                 return false;
5595         /*
5596          * If array is degraded, better not do chunk aligned read because
5597          * later we might have to read it again in order to reconstruct
5598          * data on failed drives.
5599          */
5600         if (rw == READ && mddev->degraded == 0 &&
5601             mddev->reshape_position == MaxSector) {
5602                 bi = chunk_aligned_read(mddev, bi);
5603                 if (!bi)
5604                         return true;
5605         }
5606
5607         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5608                 make_discard_request(mddev, bi);
5609                 md_write_end(mddev);
5610                 return true;
5611         }
5612
5613         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5614         last_sector = bio_end_sector(bi);
5615         bi->bi_next = NULL;
5616
5617         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5618         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5619                 int previous;
5620                 int seq;
5621
5622                 do_prepare = false;
5623         retry:
5624                 seq = read_seqcount_begin(&conf->gen_lock);
5625                 previous = 0;
5626                 if (do_prepare)
5627                         prepare_to_wait(&conf->wait_for_overlap, &w,
5628                                 TASK_UNINTERRUPTIBLE);
5629                 if (unlikely(conf->reshape_progress != MaxSector)) {
5630                         /* spinlock is needed as reshape_progress may be
5631                          * 64bit on a 32bit platform, and so it might be
5632                          * possible to see a half-updated value
5633                          * Of course reshape_progress could change after
5634                          * the lock is dropped, so once we get a reference
5635                          * to the stripe that we think it is, we will have
5636                          * to check again.
5637                          */
5638                         spin_lock_irq(&conf->device_lock);
5639                         if (mddev->reshape_backwards
5640                             ? logical_sector < conf->reshape_progress
5641                             : logical_sector >= conf->reshape_progress) {
5642                                 previous = 1;
5643                         } else {
5644                                 if (mddev->reshape_backwards
5645                                     ? logical_sector < conf->reshape_safe
5646                                     : logical_sector >= conf->reshape_safe) {
5647                                         spin_unlock_irq(&conf->device_lock);
5648                                         schedule();
5649                                         do_prepare = true;
5650                                         goto retry;
5651                                 }
5652                         }
5653                         spin_unlock_irq(&conf->device_lock);
5654                 }
5655
5656                 new_sector = raid5_compute_sector(conf, logical_sector,
5657                                                   previous,
5658                                                   &dd_idx, NULL);
5659                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5660                         (unsigned long long)new_sector,
5661                         (unsigned long long)logical_sector);
5662
5663                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5664                                        (bi->bi_opf & REQ_RAHEAD), 0);
5665                 if (sh) {
5666                         if (unlikely(previous)) {
5667                                 /* expansion might have moved on while waiting for a
5668                                  * stripe, so we must do the range check again.
5669                                  * Expansion could still move past after this
5670                                  * test, but as we are holding a reference to
5671                                  * 'sh', we know that if that happens,
5672                                  *  STRIPE_EXPANDING will get set and the expansion
5673                                  * won't proceed until we finish with the stripe.
5674                                  */
5675                                 int must_retry = 0;
5676                                 spin_lock_irq(&conf->device_lock);
5677                                 if (mddev->reshape_backwards
5678                                     ? logical_sector >= conf->reshape_progress
5679                                     : logical_sector < conf->reshape_progress)
5680                                         /* mismatch, need to try again */
5681                                         must_retry = 1;
5682                                 spin_unlock_irq(&conf->device_lock);
5683                                 if (must_retry) {
5684                                         raid5_release_stripe(sh);
5685                                         schedule();
5686                                         do_prepare = true;
5687                                         goto retry;
5688                                 }
5689                         }
5690                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5691                                 /* Might have got the wrong stripe_head
5692                                  * by accident
5693                                  */
5694                                 raid5_release_stripe(sh);
5695                                 goto retry;
5696                         }
5697
5698                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5699                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5700                                 /* Stripe is busy expanding or
5701                                  * add failed due to overlap.  Flush everything
5702                                  * and wait a while
5703                                  */
5704                                 md_wakeup_thread(mddev->thread);
5705                                 raid5_release_stripe(sh);
5706                                 schedule();
5707                                 do_prepare = true;
5708                                 goto retry;
5709                         }
5710                         if (do_flush) {
5711                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5712                                 /* we only need flush for one stripe */
5713                                 do_flush = false;
5714                         }
5715
5716                         set_bit(STRIPE_HANDLE, &sh->state);
5717                         clear_bit(STRIPE_DELAYED, &sh->state);
5718                         if ((!sh->batch_head || sh == sh->batch_head) &&
5719                             (bi->bi_opf & REQ_SYNC) &&
5720                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5721                                 atomic_inc(&conf->preread_active_stripes);
5722                         release_stripe_plug(mddev, sh);
5723                 } else {
5724                         /* cannot get stripe for read-ahead, just give-up */
5725                         bi->bi_status = BLK_STS_IOERR;
5726                         break;
5727                 }
5728         }
5729         finish_wait(&conf->wait_for_overlap, &w);
5730
5731         if (rw == WRITE)
5732                 md_write_end(mddev);
5733         bio_endio(bi);
5734         return true;
5735 }
5736
5737 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5738
5739 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5740 {
5741         /* reshaping is quite different to recovery/resync so it is
5742          * handled quite separately ... here.
5743          *
5744          * On each call to sync_request, we gather one chunk worth of
5745          * destination stripes and flag them as expanding.
5746          * Then we find all the source stripes and request reads.
5747          * As the reads complete, handle_stripe will copy the data
5748          * into the destination stripe and release that stripe.
5749          */
5750         struct r5conf *conf = mddev->private;
5751         struct stripe_head *sh;
5752         struct md_rdev *rdev;
5753         sector_t first_sector, last_sector;
5754         int raid_disks = conf->previous_raid_disks;
5755         int data_disks = raid_disks - conf->max_degraded;
5756         int new_data_disks = conf->raid_disks - conf->max_degraded;
5757         int i;
5758         int dd_idx;
5759         sector_t writepos, readpos, safepos;
5760         sector_t stripe_addr;
5761         int reshape_sectors;
5762         struct list_head stripes;
5763         sector_t retn;
5764
5765         if (sector_nr == 0) {
5766                 /* If restarting in the middle, skip the initial sectors */
5767                 if (mddev->reshape_backwards &&
5768                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5769                         sector_nr = raid5_size(mddev, 0, 0)
5770                                 - conf->reshape_progress;
5771                 } else if (mddev->reshape_backwards &&
5772                            conf->reshape_progress == MaxSector) {
5773                         /* shouldn't happen, but just in case, finish up.*/
5774                         sector_nr = MaxSector;
5775                 } else if (!mddev->reshape_backwards &&
5776                            conf->reshape_progress > 0)
5777                         sector_nr = conf->reshape_progress;
5778                 sector_div(sector_nr, new_data_disks);
5779                 if (sector_nr) {
5780                         mddev->curr_resync_completed = sector_nr;
5781                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5782                         *skipped = 1;
5783                         retn = sector_nr;
5784                         goto finish;
5785                 }
5786         }
5787
5788         /* We need to process a full chunk at a time.
5789          * If old and new chunk sizes differ, we need to process the
5790          * largest of these
5791          */
5792
5793         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5794
5795         /* We update the metadata at least every 10 seconds, or when
5796          * the data about to be copied would over-write the source of
5797          * the data at the front of the range.  i.e. one new_stripe
5798          * along from reshape_progress new_maps to after where
5799          * reshape_safe old_maps to
5800          */
5801         writepos = conf->reshape_progress;
5802         sector_div(writepos, new_data_disks);
5803         readpos = conf->reshape_progress;
5804         sector_div(readpos, data_disks);
5805         safepos = conf->reshape_safe;
5806         sector_div(safepos, data_disks);
5807         if (mddev->reshape_backwards) {
5808                 BUG_ON(writepos < reshape_sectors);
5809                 writepos -= reshape_sectors;
5810                 readpos += reshape_sectors;
5811                 safepos += reshape_sectors;
5812         } else {
5813                 writepos += reshape_sectors;
5814                 /* readpos and safepos are worst-case calculations.
5815                  * A negative number is overly pessimistic, and causes
5816                  * obvious problems for unsigned storage.  So clip to 0.
5817                  */
5818                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5819                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5820         }
5821
5822         /* Having calculated the 'writepos' possibly use it
5823          * to set 'stripe_addr' which is where we will write to.
5824          */
5825         if (mddev->reshape_backwards) {
5826                 BUG_ON(conf->reshape_progress == 0);
5827                 stripe_addr = writepos;
5828                 BUG_ON((mddev->dev_sectors &
5829                         ~((sector_t)reshape_sectors - 1))
5830                        - reshape_sectors - stripe_addr
5831                        != sector_nr);
5832         } else {
5833                 BUG_ON(writepos != sector_nr + reshape_sectors);
5834                 stripe_addr = sector_nr;
5835         }
5836
5837         /* 'writepos' is the most advanced device address we might write.
5838          * 'readpos' is the least advanced device address we might read.
5839          * 'safepos' is the least address recorded in the metadata as having
5840          *     been reshaped.
5841          * If there is a min_offset_diff, these are adjusted either by
5842          * increasing the safepos/readpos if diff is negative, or
5843          * increasing writepos if diff is positive.
5844          * If 'readpos' is then behind 'writepos', there is no way that we can
5845          * ensure safety in the face of a crash - that must be done by userspace
5846          * making a backup of the data.  So in that case there is no particular
5847          * rush to update metadata.
5848          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5849          * update the metadata to advance 'safepos' to match 'readpos' so that
5850          * we can be safe in the event of a crash.
5851          * So we insist on updating metadata if safepos is behind writepos and
5852          * readpos is beyond writepos.
5853          * In any case, update the metadata every 10 seconds.
5854          * Maybe that number should be configurable, but I'm not sure it is
5855          * worth it.... maybe it could be a multiple of safemode_delay???
5856          */
5857         if (conf->min_offset_diff < 0) {
5858                 safepos += -conf->min_offset_diff;
5859                 readpos += -conf->min_offset_diff;
5860         } else
5861                 writepos += conf->min_offset_diff;
5862
5863         if ((mddev->reshape_backwards
5864              ? (safepos > writepos && readpos < writepos)
5865              : (safepos < writepos && readpos > writepos)) ||
5866             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5867                 /* Cannot proceed until we've updated the superblock... */
5868                 wait_event(conf->wait_for_overlap,
5869                            atomic_read(&conf->reshape_stripes)==0
5870                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5871                 if (atomic_read(&conf->reshape_stripes) != 0)
5872                         return 0;
5873                 mddev->reshape_position = conf->reshape_progress;
5874                 mddev->curr_resync_completed = sector_nr;
5875                 if (!mddev->reshape_backwards)
5876                         /* Can update recovery_offset */
5877                         rdev_for_each(rdev, mddev)
5878                                 if (rdev->raid_disk >= 0 &&
5879                                     !test_bit(Journal, &rdev->flags) &&
5880                                     !test_bit(In_sync, &rdev->flags) &&
5881                                     rdev->recovery_offset < sector_nr)
5882                                         rdev->recovery_offset = sector_nr;
5883
5884                 conf->reshape_checkpoint = jiffies;
5885                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5886                 md_wakeup_thread(mddev->thread);
5887                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5888                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5889                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5890                         return 0;
5891                 spin_lock_irq(&conf->device_lock);
5892                 conf->reshape_safe = mddev->reshape_position;
5893                 spin_unlock_irq(&conf->device_lock);
5894                 wake_up(&conf->wait_for_overlap);
5895                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5896         }
5897
5898         INIT_LIST_HEAD(&stripes);
5899         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5900                 int j;
5901                 int skipped_disk = 0;
5902                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5903                 set_bit(STRIPE_EXPANDING, &sh->state);
5904                 atomic_inc(&conf->reshape_stripes);
5905                 /* If any of this stripe is beyond the end of the old
5906                  * array, then we need to zero those blocks
5907                  */
5908                 for (j=sh->disks; j--;) {
5909                         sector_t s;
5910                         if (j == sh->pd_idx)
5911                                 continue;
5912                         if (conf->level == 6 &&
5913                             j == sh->qd_idx)
5914                                 continue;
5915                         s = raid5_compute_blocknr(sh, j, 0);
5916                         if (s < raid5_size(mddev, 0, 0)) {
5917                                 skipped_disk = 1;
5918                                 continue;
5919                         }
5920                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5921                         set_bit(R5_Expanded, &sh->dev[j].flags);
5922                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5923                 }
5924                 if (!skipped_disk) {
5925                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5926                         set_bit(STRIPE_HANDLE, &sh->state);
5927                 }
5928                 list_add(&sh->lru, &stripes);
5929         }
5930         spin_lock_irq(&conf->device_lock);
5931         if (mddev->reshape_backwards)
5932                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5933         else
5934                 conf->reshape_progress += reshape_sectors * new_data_disks;
5935         spin_unlock_irq(&conf->device_lock);
5936         /* Ok, those stripe are ready. We can start scheduling
5937          * reads on the source stripes.
5938          * The source stripes are determined by mapping the first and last
5939          * block on the destination stripes.
5940          */
5941         first_sector =
5942                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5943                                      1, &dd_idx, NULL);
5944         last_sector =
5945                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5946                                             * new_data_disks - 1),
5947                                      1, &dd_idx, NULL);
5948         if (last_sector >= mddev->dev_sectors)
5949                 last_sector = mddev->dev_sectors - 1;
5950         while (first_sector <= last_sector) {
5951                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5952                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5953                 set_bit(STRIPE_HANDLE, &sh->state);
5954                 raid5_release_stripe(sh);
5955                 first_sector += STRIPE_SECTORS;
5956         }
5957         /* Now that the sources are clearly marked, we can release
5958          * the destination stripes
5959          */
5960         while (!list_empty(&stripes)) {
5961                 sh = list_entry(stripes.next, struct stripe_head, lru);
5962                 list_del_init(&sh->lru);
5963                 raid5_release_stripe(sh);
5964         }
5965         /* If this takes us to the resync_max point where we have to pause,
5966          * then we need to write out the superblock.
5967          */
5968         sector_nr += reshape_sectors;
5969         retn = reshape_sectors;
5970 finish:
5971         if (mddev->curr_resync_completed > mddev->resync_max ||
5972             (sector_nr - mddev->curr_resync_completed) * 2
5973             >= mddev->resync_max - mddev->curr_resync_completed) {
5974                 /* Cannot proceed until we've updated the superblock... */
5975                 wait_event(conf->wait_for_overlap,
5976                            atomic_read(&conf->reshape_stripes) == 0
5977                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5978                 if (atomic_read(&conf->reshape_stripes) != 0)
5979                         goto ret;
5980                 mddev->reshape_position = conf->reshape_progress;
5981                 mddev->curr_resync_completed = sector_nr;
5982                 if (!mddev->reshape_backwards)
5983                         /* Can update recovery_offset */
5984                         rdev_for_each(rdev, mddev)
5985                                 if (rdev->raid_disk >= 0 &&
5986                                     !test_bit(Journal, &rdev->flags) &&
5987                                     !test_bit(In_sync, &rdev->flags) &&
5988                                     rdev->recovery_offset < sector_nr)
5989                                         rdev->recovery_offset = sector_nr;
5990                 conf->reshape_checkpoint = jiffies;
5991                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5992                 md_wakeup_thread(mddev->thread);
5993                 wait_event(mddev->sb_wait,
5994                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5995                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5996                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5997                         goto ret;
5998                 spin_lock_irq(&conf->device_lock);
5999                 conf->reshape_safe = mddev->reshape_position;
6000                 spin_unlock_irq(&conf->device_lock);
6001                 wake_up(&conf->wait_for_overlap);
6002                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6003         }
6004 ret:
6005         return retn;
6006 }
6007
6008 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6009                                           int *skipped)
6010 {
6011         struct r5conf *conf = mddev->private;
6012         struct stripe_head *sh;
6013         sector_t max_sector = mddev->dev_sectors;
6014         sector_t sync_blocks;
6015         int still_degraded = 0;
6016         int i;
6017
6018         if (sector_nr >= max_sector) {
6019                 /* just being told to finish up .. nothing much to do */
6020
6021                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6022                         end_reshape(conf);
6023                         return 0;
6024                 }
6025
6026                 if (mddev->curr_resync < max_sector) /* aborted */
6027                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6028                                            &sync_blocks, 1);
6029                 else /* completed sync */
6030                         conf->fullsync = 0;
6031                 md_bitmap_close_sync(mddev->bitmap);
6032
6033                 return 0;
6034         }
6035
6036         /* Allow raid5_quiesce to complete */
6037         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6038
6039         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6040                 return reshape_request(mddev, sector_nr, skipped);
6041
6042         /* No need to check resync_max as we never do more than one
6043          * stripe, and as resync_max will always be on a chunk boundary,
6044          * if the check in md_do_sync didn't fire, there is no chance
6045          * of overstepping resync_max here
6046          */
6047
6048         /* if there is too many failed drives and we are trying
6049          * to resync, then assert that we are finished, because there is
6050          * nothing we can do.
6051          */
6052         if (mddev->degraded >= conf->max_degraded &&
6053             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6054                 sector_t rv = mddev->dev_sectors - sector_nr;
6055                 *skipped = 1;
6056                 return rv;
6057         }
6058         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6059             !conf->fullsync &&
6060             !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6061             sync_blocks >= STRIPE_SECTORS) {
6062                 /* we can skip this block, and probably more */
6063                 sync_blocks /= STRIPE_SECTORS;
6064                 *skipped = 1;
6065                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6066         }
6067
6068         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6069
6070         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6071         if (sh == NULL) {
6072                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6073                 /* make sure we don't swamp the stripe cache if someone else
6074                  * is trying to get access
6075                  */
6076                 schedule_timeout_uninterruptible(1);
6077         }
6078         /* Need to check if array will still be degraded after recovery/resync
6079          * Note in case of > 1 drive failures it's possible we're rebuilding
6080          * one drive while leaving another faulty drive in array.
6081          */
6082         rcu_read_lock();
6083         for (i = 0; i < conf->raid_disks; i++) {
6084                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6085
6086                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6087                         still_degraded = 1;
6088         }
6089         rcu_read_unlock();
6090
6091         md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6092
6093         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6094         set_bit(STRIPE_HANDLE, &sh->state);
6095
6096         raid5_release_stripe(sh);
6097
6098         return STRIPE_SECTORS;
6099 }
6100
6101 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6102                                unsigned int offset)
6103 {
6104         /* We may not be able to submit a whole bio at once as there
6105          * may not be enough stripe_heads available.
6106          * We cannot pre-allocate enough stripe_heads as we may need
6107          * more than exist in the cache (if we allow ever large chunks).
6108          * So we do one stripe head at a time and record in
6109          * ->bi_hw_segments how many have been done.
6110          *
6111          * We *know* that this entire raid_bio is in one chunk, so
6112          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6113          */
6114         struct stripe_head *sh;
6115         int dd_idx;
6116         sector_t sector, logical_sector, last_sector;
6117         int scnt = 0;
6118         int handled = 0;
6119
6120         logical_sector = raid_bio->bi_iter.bi_sector &
6121                 ~((sector_t)STRIPE_SECTORS-1);
6122         sector = raid5_compute_sector(conf, logical_sector,
6123                                       0, &dd_idx, NULL);
6124         last_sector = bio_end_sector(raid_bio);
6125
6126         for (; logical_sector < last_sector;
6127              logical_sector += STRIPE_SECTORS,
6128                      sector += STRIPE_SECTORS,
6129                      scnt++) {
6130
6131                 if (scnt < offset)
6132                         /* already done this stripe */
6133                         continue;
6134
6135                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6136
6137                 if (!sh) {
6138                         /* failed to get a stripe - must wait */
6139                         conf->retry_read_aligned = raid_bio;
6140                         conf->retry_read_offset = scnt;
6141                         return handled;
6142                 }
6143
6144                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6145                         raid5_release_stripe(sh);
6146                         conf->retry_read_aligned = raid_bio;
6147                         conf->retry_read_offset = scnt;
6148                         return handled;
6149                 }
6150
6151                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6152                 handle_stripe(sh);
6153                 raid5_release_stripe(sh);
6154                 handled++;
6155         }
6156
6157         bio_endio(raid_bio);
6158
6159         if (atomic_dec_and_test(&conf->active_aligned_reads))
6160                 wake_up(&conf->wait_for_quiescent);
6161         return handled;
6162 }
6163
6164 static int handle_active_stripes(struct r5conf *conf, int group,
6165                                  struct r5worker *worker,
6166                                  struct list_head *temp_inactive_list)
6167 {
6168         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6169         int i, batch_size = 0, hash;
6170         bool release_inactive = false;
6171
6172         while (batch_size < MAX_STRIPE_BATCH &&
6173                         (sh = __get_priority_stripe(conf, group)) != NULL)
6174                 batch[batch_size++] = sh;
6175
6176         if (batch_size == 0) {
6177                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6178                         if (!list_empty(temp_inactive_list + i))
6179                                 break;
6180                 if (i == NR_STRIPE_HASH_LOCKS) {
6181                         spin_unlock_irq(&conf->device_lock);
6182                         log_flush_stripe_to_raid(conf);
6183                         spin_lock_irq(&conf->device_lock);
6184                         return batch_size;
6185                 }
6186                 release_inactive = true;
6187         }
6188         spin_unlock_irq(&conf->device_lock);
6189
6190         release_inactive_stripe_list(conf, temp_inactive_list,
6191                                      NR_STRIPE_HASH_LOCKS);
6192
6193         r5l_flush_stripe_to_raid(conf->log);
6194         if (release_inactive) {
6195                 spin_lock_irq(&conf->device_lock);
6196                 return 0;
6197         }
6198
6199         for (i = 0; i < batch_size; i++)
6200                 handle_stripe(batch[i]);
6201         log_write_stripe_run(conf);
6202
6203         cond_resched();
6204
6205         spin_lock_irq(&conf->device_lock);
6206         for (i = 0; i < batch_size; i++) {
6207                 hash = batch[i]->hash_lock_index;
6208                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6209         }
6210         return batch_size;
6211 }
6212
6213 static void raid5_do_work(struct work_struct *work)
6214 {
6215         struct r5worker *worker = container_of(work, struct r5worker, work);
6216         struct r5worker_group *group = worker->group;
6217         struct r5conf *conf = group->conf;
6218         struct mddev *mddev = conf->mddev;
6219         int group_id = group - conf->worker_groups;
6220         int handled;
6221         struct blk_plug plug;
6222
6223         pr_debug("+++ raid5worker active\n");
6224
6225         blk_start_plug(&plug);
6226         handled = 0;
6227         spin_lock_irq(&conf->device_lock);
6228         while (1) {
6229                 int batch_size, released;
6230
6231                 released = release_stripe_list(conf, worker->temp_inactive_list);
6232
6233                 batch_size = handle_active_stripes(conf, group_id, worker,
6234                                                    worker->temp_inactive_list);
6235                 worker->working = false;
6236                 if (!batch_size && !released)
6237                         break;
6238                 handled += batch_size;
6239                 wait_event_lock_irq(mddev->sb_wait,
6240                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6241                         conf->device_lock);
6242         }
6243         pr_debug("%d stripes handled\n", handled);
6244
6245         spin_unlock_irq(&conf->device_lock);
6246
6247         flush_deferred_bios(conf);
6248
6249         r5l_flush_stripe_to_raid(conf->log);
6250
6251         async_tx_issue_pending_all();
6252         blk_finish_plug(&plug);
6253
6254         pr_debug("--- raid5worker inactive\n");
6255 }
6256
6257 /*
6258  * This is our raid5 kernel thread.
6259  *
6260  * We scan the hash table for stripes which can be handled now.
6261  * During the scan, completed stripes are saved for us by the interrupt
6262  * handler, so that they will not have to wait for our next wakeup.
6263  */
6264 static void raid5d(struct md_thread *thread)
6265 {
6266         struct mddev *mddev = thread->mddev;
6267         struct r5conf *conf = mddev->private;
6268         int handled;
6269         struct blk_plug plug;
6270
6271         pr_debug("+++ raid5d active\n");
6272
6273         md_check_recovery(mddev);
6274
6275         blk_start_plug(&plug);
6276         handled = 0;
6277         spin_lock_irq(&conf->device_lock);
6278         while (1) {
6279                 struct bio *bio;
6280                 int batch_size, released;
6281                 unsigned int offset;
6282
6283                 released = release_stripe_list(conf, conf->temp_inactive_list);
6284                 if (released)
6285                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6286
6287                 if (
6288                     !list_empty(&conf->bitmap_list)) {
6289                         /* Now is a good time to flush some bitmap updates */
6290                         conf->seq_flush++;
6291                         spin_unlock_irq(&conf->device_lock);
6292                         md_bitmap_unplug(mddev->bitmap);
6293                         spin_lock_irq(&conf->device_lock);
6294                         conf->seq_write = conf->seq_flush;
6295                         activate_bit_delay(conf, conf->temp_inactive_list);
6296                 }
6297                 raid5_activate_delayed(conf);
6298
6299                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6300                         int ok;
6301                         spin_unlock_irq(&conf->device_lock);
6302                         ok = retry_aligned_read(conf, bio, offset);
6303                         spin_lock_irq(&conf->device_lock);
6304                         if (!ok)
6305                                 break;
6306                         handled++;
6307                 }
6308
6309                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6310                                                    conf->temp_inactive_list);
6311                 if (!batch_size && !released)
6312                         break;
6313                 handled += batch_size;
6314
6315                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6316                         spin_unlock_irq(&conf->device_lock);
6317                         md_check_recovery(mddev);
6318                         spin_lock_irq(&conf->device_lock);
6319                 }
6320         }
6321         pr_debug("%d stripes handled\n", handled);
6322
6323         spin_unlock_irq(&conf->device_lock);
6324         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6325             mutex_trylock(&conf->cache_size_mutex)) {
6326                 grow_one_stripe(conf, __GFP_NOWARN);
6327                 /* Set flag even if allocation failed.  This helps
6328                  * slow down allocation requests when mem is short
6329                  */
6330                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6331                 mutex_unlock(&conf->cache_size_mutex);
6332         }
6333
6334         flush_deferred_bios(conf);
6335
6336         r5l_flush_stripe_to_raid(conf->log);
6337
6338         async_tx_issue_pending_all();
6339         blk_finish_plug(&plug);
6340
6341         pr_debug("--- raid5d inactive\n");
6342 }
6343
6344 static ssize_t
6345 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6346 {
6347         struct r5conf *conf;
6348         int ret = 0;
6349         spin_lock(&mddev->lock);
6350         conf = mddev->private;
6351         if (conf)
6352                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6353         spin_unlock(&mddev->lock);
6354         return ret;
6355 }
6356
6357 int
6358 raid5_set_cache_size(struct mddev *mddev, int size)
6359 {
6360         struct r5conf *conf = mddev->private;
6361
6362         if (size <= 16 || size > 32768)
6363                 return -EINVAL;
6364
6365         conf->min_nr_stripes = size;
6366         mutex_lock(&conf->cache_size_mutex);
6367         while (size < conf->max_nr_stripes &&
6368                drop_one_stripe(conf))
6369                 ;
6370         mutex_unlock(&conf->cache_size_mutex);
6371
6372         md_allow_write(mddev);
6373
6374         mutex_lock(&conf->cache_size_mutex);
6375         while (size > conf->max_nr_stripes)
6376                 if (!grow_one_stripe(conf, GFP_KERNEL))
6377                         break;
6378         mutex_unlock(&conf->cache_size_mutex);
6379
6380         return 0;
6381 }
6382 EXPORT_SYMBOL(raid5_set_cache_size);
6383
6384 static ssize_t
6385 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6386 {
6387         struct r5conf *conf;
6388         unsigned long new;
6389         int err;
6390
6391         if (len >= PAGE_SIZE)
6392                 return -EINVAL;
6393         if (kstrtoul(page, 10, &new))
6394                 return -EINVAL;
6395         err = mddev_lock(mddev);
6396         if (err)
6397                 return err;
6398         conf = mddev->private;
6399         if (!conf)
6400                 err = -ENODEV;
6401         else
6402                 err = raid5_set_cache_size(mddev, new);
6403         mddev_unlock(mddev);
6404
6405         return err ?: len;
6406 }
6407
6408 static struct md_sysfs_entry
6409 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6410                                 raid5_show_stripe_cache_size,
6411                                 raid5_store_stripe_cache_size);
6412
6413 static ssize_t
6414 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6415 {
6416         struct r5conf *conf = mddev->private;
6417         if (conf)
6418                 return sprintf(page, "%d\n", conf->rmw_level);
6419         else
6420                 return 0;
6421 }
6422
6423 static ssize_t
6424 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6425 {
6426         struct r5conf *conf = mddev->private;
6427         unsigned long new;
6428
6429         if (!conf)
6430                 return -ENODEV;
6431
6432         if (len >= PAGE_SIZE)
6433                 return -EINVAL;
6434
6435         if (kstrtoul(page, 10, &new))
6436                 return -EINVAL;
6437
6438         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6439                 return -EINVAL;
6440
6441         if (new != PARITY_DISABLE_RMW &&
6442             new != PARITY_ENABLE_RMW &&
6443             new != PARITY_PREFER_RMW)
6444                 return -EINVAL;
6445
6446         conf->rmw_level = new;
6447         return len;
6448 }
6449
6450 static struct md_sysfs_entry
6451 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6452                          raid5_show_rmw_level,
6453                          raid5_store_rmw_level);
6454
6455
6456 static ssize_t
6457 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6458 {
6459         struct r5conf *conf;
6460         int ret = 0;
6461         spin_lock(&mddev->lock);
6462         conf = mddev->private;
6463         if (conf)
6464                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6465         spin_unlock(&mddev->lock);
6466         return ret;
6467 }
6468
6469 static ssize_t
6470 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6471 {
6472         struct r5conf *conf;
6473         unsigned long new;
6474         int err;
6475
6476         if (len >= PAGE_SIZE)
6477                 return -EINVAL;
6478         if (kstrtoul(page, 10, &new))
6479                 return -EINVAL;
6480
6481         err = mddev_lock(mddev);
6482         if (err)
6483                 return err;
6484         conf = mddev->private;
6485         if (!conf)
6486                 err = -ENODEV;
6487         else if (new > conf->min_nr_stripes)
6488                 err = -EINVAL;
6489         else
6490                 conf->bypass_threshold = new;
6491         mddev_unlock(mddev);
6492         return err ?: len;
6493 }
6494
6495 static struct md_sysfs_entry
6496 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6497                                         S_IRUGO | S_IWUSR,
6498                                         raid5_show_preread_threshold,
6499                                         raid5_store_preread_threshold);
6500
6501 static ssize_t
6502 raid5_show_skip_copy(struct mddev *mddev, char *page)
6503 {
6504         struct r5conf *conf;
6505         int ret = 0;
6506         spin_lock(&mddev->lock);
6507         conf = mddev->private;
6508         if (conf)
6509                 ret = sprintf(page, "%d\n", conf->skip_copy);
6510         spin_unlock(&mddev->lock);
6511         return ret;
6512 }
6513
6514 static ssize_t
6515 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6516 {
6517         struct r5conf *conf;
6518         unsigned long new;
6519         int err;
6520
6521         if (len >= PAGE_SIZE)
6522                 return -EINVAL;
6523         if (kstrtoul(page, 10, &new))
6524                 return -EINVAL;
6525         new = !!new;
6526
6527         err = mddev_lock(mddev);
6528         if (err)
6529                 return err;
6530         conf = mddev->private;
6531         if (!conf)
6532                 err = -ENODEV;
6533         else if (new != conf->skip_copy) {
6534                 mddev_suspend(mddev);
6535                 conf->skip_copy = new;
6536                 if (new)
6537                         mddev->queue->backing_dev_info->capabilities |=
6538                                 BDI_CAP_STABLE_WRITES;
6539                 else
6540                         mddev->queue->backing_dev_info->capabilities &=
6541                                 ~BDI_CAP_STABLE_WRITES;
6542                 mddev_resume(mddev);
6543         }
6544         mddev_unlock(mddev);
6545         return err ?: len;
6546 }
6547
6548 static struct md_sysfs_entry
6549 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6550                                         raid5_show_skip_copy,
6551                                         raid5_store_skip_copy);
6552
6553 static ssize_t
6554 stripe_cache_active_show(struct mddev *mddev, char *page)
6555 {
6556         struct r5conf *conf = mddev->private;
6557         if (conf)
6558                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6559         else
6560                 return 0;
6561 }
6562
6563 static struct md_sysfs_entry
6564 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6565
6566 static ssize_t
6567 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6568 {
6569         struct r5conf *conf;
6570         int ret = 0;
6571         spin_lock(&mddev->lock);
6572         conf = mddev->private;
6573         if (conf)
6574                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6575         spin_unlock(&mddev->lock);
6576         return ret;
6577 }
6578
6579 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6580                                int *group_cnt,
6581                                int *worker_cnt_per_group,
6582                                struct r5worker_group **worker_groups);
6583 static ssize_t
6584 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6585 {
6586         struct r5conf *conf;
6587         unsigned int new;
6588         int err;
6589         struct r5worker_group *new_groups, *old_groups;
6590         int group_cnt, worker_cnt_per_group;
6591
6592         if (len >= PAGE_SIZE)
6593                 return -EINVAL;
6594         if (kstrtouint(page, 10, &new))
6595                 return -EINVAL;
6596         /* 8192 should be big enough */
6597         if (new > 8192)
6598                 return -EINVAL;
6599
6600         err = mddev_lock(mddev);
6601         if (err)
6602                 return err;
6603         conf = mddev->private;
6604         if (!conf)
6605                 err = -ENODEV;
6606         else if (new != conf->worker_cnt_per_group) {
6607                 mddev_suspend(mddev);
6608
6609                 old_groups = conf->worker_groups;
6610                 if (old_groups)
6611                         flush_workqueue(raid5_wq);
6612
6613                 err = alloc_thread_groups(conf, new,
6614                                           &group_cnt, &worker_cnt_per_group,
6615                                           &new_groups);
6616                 if (!err) {
6617                         spin_lock_irq(&conf->device_lock);
6618                         conf->group_cnt = group_cnt;
6619                         conf->worker_cnt_per_group = worker_cnt_per_group;
6620                         conf->worker_groups = new_groups;
6621                         spin_unlock_irq(&conf->device_lock);
6622
6623                         if (old_groups)
6624                                 kfree(old_groups[0].workers);
6625                         kfree(old_groups);
6626                 }
6627                 mddev_resume(mddev);
6628         }
6629         mddev_unlock(mddev);
6630
6631         return err ?: len;
6632 }
6633
6634 static struct md_sysfs_entry
6635 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6636                                 raid5_show_group_thread_cnt,
6637                                 raid5_store_group_thread_cnt);
6638
6639 static struct attribute *raid5_attrs[] =  {
6640         &raid5_stripecache_size.attr,
6641         &raid5_stripecache_active.attr,
6642         &raid5_preread_bypass_threshold.attr,
6643         &raid5_group_thread_cnt.attr,
6644         &raid5_skip_copy.attr,
6645         &raid5_rmw_level.attr,
6646         &r5c_journal_mode.attr,
6647         NULL,
6648 };
6649 static struct attribute_group raid5_attrs_group = {
6650         .name = NULL,
6651         .attrs = raid5_attrs,
6652 };
6653
6654 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6655                                int *group_cnt,
6656                                int *worker_cnt_per_group,
6657                                struct r5worker_group **worker_groups)
6658 {
6659         int i, j, k;
6660         ssize_t size;
6661         struct r5worker *workers;
6662
6663         *worker_cnt_per_group = cnt;
6664         if (cnt == 0) {
6665                 *group_cnt = 0;
6666                 *worker_groups = NULL;
6667                 return 0;
6668         }
6669         *group_cnt = num_possible_nodes();
6670         size = sizeof(struct r5worker) * cnt;
6671         workers = kcalloc(size, *group_cnt, GFP_NOIO);
6672         *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6673                                  GFP_NOIO);
6674         if (!*worker_groups || !workers) {
6675                 kfree(workers);
6676                 kfree(*worker_groups);
6677                 return -ENOMEM;
6678         }
6679
6680         for (i = 0; i < *group_cnt; i++) {
6681                 struct r5worker_group *group;
6682
6683                 group = &(*worker_groups)[i];
6684                 INIT_LIST_HEAD(&group->handle_list);
6685                 INIT_LIST_HEAD(&group->loprio_list);
6686                 group->conf = conf;
6687                 group->workers = workers + i * cnt;
6688
6689                 for (j = 0; j < cnt; j++) {
6690                         struct r5worker *worker = group->workers + j;
6691                         worker->group = group;
6692                         INIT_WORK(&worker->work, raid5_do_work);
6693
6694                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6695                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6696                 }
6697         }
6698
6699         return 0;
6700 }
6701
6702 static void free_thread_groups(struct r5conf *conf)
6703 {
6704         if (conf->worker_groups)
6705                 kfree(conf->worker_groups[0].workers);
6706         kfree(conf->worker_groups);
6707         conf->worker_groups = NULL;
6708 }
6709
6710 static sector_t
6711 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6712 {
6713         struct r5conf *conf = mddev->private;
6714
6715         if (!sectors)
6716                 sectors = mddev->dev_sectors;
6717         if (!raid_disks)
6718                 /* size is defined by the smallest of previous and new size */
6719                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6720
6721         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6722         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6723         return sectors * (raid_disks - conf->max_degraded);
6724 }
6725
6726 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6727 {
6728         safe_put_page(percpu->spare_page);
6729         if (percpu->scribble)
6730                 flex_array_free(percpu->scribble);
6731         percpu->spare_page = NULL;
6732         percpu->scribble = NULL;
6733 }
6734
6735 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6736 {
6737         if (conf->level == 6 && !percpu->spare_page)
6738                 percpu->spare_page = alloc_page(GFP_KERNEL);
6739         if (!percpu->scribble)
6740                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6741                                                       conf->previous_raid_disks),
6742                                                   max(conf->chunk_sectors,
6743                                                       conf->prev_chunk_sectors)
6744                                                    / STRIPE_SECTORS,
6745                                                   GFP_KERNEL);
6746
6747         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6748                 free_scratch_buffer(conf, percpu);
6749                 return -ENOMEM;
6750         }
6751
6752         return 0;
6753 }
6754
6755 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6756 {
6757         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6758
6759         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6760         return 0;
6761 }
6762
6763 static void raid5_free_percpu(struct r5conf *conf)
6764 {
6765         if (!conf->percpu)
6766                 return;
6767
6768         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6769         free_percpu(conf->percpu);
6770 }
6771
6772 static void free_conf(struct r5conf *conf)
6773 {
6774         int i;
6775
6776         log_exit(conf);
6777
6778         unregister_shrinker(&conf->shrinker);
6779         free_thread_groups(conf);
6780         shrink_stripes(conf);
6781         raid5_free_percpu(conf);
6782         for (i = 0; i < conf->pool_size; i++)
6783                 if (conf->disks[i].extra_page)
6784                         put_page(conf->disks[i].extra_page);
6785         kfree(conf->disks);
6786         bioset_exit(&conf->bio_split);
6787         kfree(conf->stripe_hashtbl);
6788         kfree(conf->pending_data);
6789         kfree(conf);
6790 }
6791
6792 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6793 {
6794         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6795         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6796
6797         if (alloc_scratch_buffer(conf, percpu)) {
6798                 pr_warn("%s: failed memory allocation for cpu%u\n",
6799                         __func__, cpu);
6800                 return -ENOMEM;
6801         }
6802         return 0;
6803 }
6804
6805 static int raid5_alloc_percpu(struct r5conf *conf)
6806 {
6807         int err = 0;
6808
6809         conf->percpu = alloc_percpu(struct raid5_percpu);
6810         if (!conf->percpu)
6811                 return -ENOMEM;
6812
6813         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6814         if (!err) {
6815                 conf->scribble_disks = max(conf->raid_disks,
6816                         conf->previous_raid_disks);
6817                 conf->scribble_sectors = max(conf->chunk_sectors,
6818                         conf->prev_chunk_sectors);
6819         }
6820         return err;
6821 }
6822
6823 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6824                                       struct shrink_control *sc)
6825 {
6826         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6827         unsigned long ret = SHRINK_STOP;
6828
6829         if (mutex_trylock(&conf->cache_size_mutex)) {
6830                 ret= 0;
6831                 while (ret < sc->nr_to_scan &&
6832                        conf->max_nr_stripes > conf->min_nr_stripes) {
6833                         if (drop_one_stripe(conf) == 0) {
6834                                 ret = SHRINK_STOP;
6835                                 break;
6836                         }
6837                         ret++;
6838                 }
6839                 mutex_unlock(&conf->cache_size_mutex);
6840         }
6841         return ret;
6842 }
6843
6844 static unsigned long raid5_cache_count(struct shrinker *shrink,
6845                                        struct shrink_control *sc)
6846 {
6847         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6848
6849         if (conf->max_nr_stripes < conf->min_nr_stripes)
6850                 /* unlikely, but not impossible */
6851                 return 0;
6852         return conf->max_nr_stripes - conf->min_nr_stripes;
6853 }
6854
6855 static struct r5conf *setup_conf(struct mddev *mddev)
6856 {
6857         struct r5conf *conf;
6858         int raid_disk, memory, max_disks;
6859         struct md_rdev *rdev;
6860         struct disk_info *disk;
6861         char pers_name[6];
6862         int i;
6863         int group_cnt, worker_cnt_per_group;
6864         struct r5worker_group *new_group;
6865         int ret;
6866
6867         if (mddev->new_level != 5
6868             && mddev->new_level != 4
6869             && mddev->new_level != 6) {
6870                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6871                         mdname(mddev), mddev->new_level);
6872                 return ERR_PTR(-EIO);
6873         }
6874         if ((mddev->new_level == 5
6875              && !algorithm_valid_raid5(mddev->new_layout)) ||
6876             (mddev->new_level == 6
6877              && !algorithm_valid_raid6(mddev->new_layout))) {
6878                 pr_warn("md/raid:%s: layout %d not supported\n",
6879                         mdname(mddev), mddev->new_layout);
6880                 return ERR_PTR(-EIO);
6881         }
6882         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6883                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6884                         mdname(mddev), mddev->raid_disks);
6885                 return ERR_PTR(-EINVAL);
6886         }
6887
6888         if (!mddev->new_chunk_sectors ||
6889             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6890             !is_power_of_2(mddev->new_chunk_sectors)) {
6891                 pr_warn("md/raid:%s: invalid chunk size %d\n",
6892                         mdname(mddev), mddev->new_chunk_sectors << 9);
6893                 return ERR_PTR(-EINVAL);
6894         }
6895
6896         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6897         if (conf == NULL)
6898                 goto abort;
6899         INIT_LIST_HEAD(&conf->free_list);
6900         INIT_LIST_HEAD(&conf->pending_list);
6901         conf->pending_data = kcalloc(PENDING_IO_MAX,
6902                                      sizeof(struct r5pending_data),
6903                                      GFP_KERNEL);
6904         if (!conf->pending_data)
6905                 goto abort;
6906         for (i = 0; i < PENDING_IO_MAX; i++)
6907                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6908         /* Don't enable multi-threading by default*/
6909         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6910                                  &new_group)) {
6911                 conf->group_cnt = group_cnt;
6912                 conf->worker_cnt_per_group = worker_cnt_per_group;
6913                 conf->worker_groups = new_group;
6914         } else
6915                 goto abort;
6916         spin_lock_init(&conf->device_lock);
6917         seqcount_init(&conf->gen_lock);
6918         mutex_init(&conf->cache_size_mutex);
6919         init_waitqueue_head(&conf->wait_for_quiescent);
6920         init_waitqueue_head(&conf->wait_for_stripe);
6921         init_waitqueue_head(&conf->wait_for_overlap);
6922         INIT_LIST_HEAD(&conf->handle_list);
6923         INIT_LIST_HEAD(&conf->loprio_list);
6924         INIT_LIST_HEAD(&conf->hold_list);
6925         INIT_LIST_HEAD(&conf->delayed_list);
6926         INIT_LIST_HEAD(&conf->bitmap_list);
6927         init_llist_head(&conf->released_stripes);
6928         atomic_set(&conf->active_stripes, 0);
6929         atomic_set(&conf->preread_active_stripes, 0);
6930         atomic_set(&conf->active_aligned_reads, 0);
6931         spin_lock_init(&conf->pending_bios_lock);
6932         conf->batch_bio_dispatch = true;
6933         rdev_for_each(rdev, mddev) {
6934                 if (test_bit(Journal, &rdev->flags))
6935                         continue;
6936                 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6937                         conf->batch_bio_dispatch = false;
6938                         break;
6939                 }
6940         }
6941
6942         conf->bypass_threshold = BYPASS_THRESHOLD;
6943         conf->recovery_disabled = mddev->recovery_disabled - 1;
6944
6945         conf->raid_disks = mddev->raid_disks;
6946         if (mddev->reshape_position == MaxSector)
6947                 conf->previous_raid_disks = mddev->raid_disks;
6948         else
6949                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6950         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6951
6952         conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
6953                               GFP_KERNEL);
6954
6955         if (!conf->disks)
6956                 goto abort;
6957
6958         for (i = 0; i < max_disks; i++) {
6959                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6960                 if (!conf->disks[i].extra_page)
6961                         goto abort;
6962         }
6963
6964         ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
6965         if (ret)
6966                 goto abort;
6967         conf->mddev = mddev;
6968
6969         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6970                 goto abort;
6971
6972         /* We init hash_locks[0] separately to that it can be used
6973          * as the reference lock in the spin_lock_nest_lock() call
6974          * in lock_all_device_hash_locks_irq in order to convince
6975          * lockdep that we know what we are doing.
6976          */
6977         spin_lock_init(conf->hash_locks);
6978         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6979                 spin_lock_init(conf->hash_locks + i);
6980
6981         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6982                 INIT_LIST_HEAD(conf->inactive_list + i);
6983
6984         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6985                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6986
6987         atomic_set(&conf->r5c_cached_full_stripes, 0);
6988         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6989         atomic_set(&conf->r5c_cached_partial_stripes, 0);
6990         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6991         atomic_set(&conf->r5c_flushing_full_stripes, 0);
6992         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
6993
6994         conf->level = mddev->new_level;
6995         conf->chunk_sectors = mddev->new_chunk_sectors;
6996         if (raid5_alloc_percpu(conf) != 0)
6997                 goto abort;
6998
6999         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7000
7001         rdev_for_each(rdev, mddev) {
7002                 raid_disk = rdev->raid_disk;
7003                 if (raid_disk >= max_disks
7004                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7005                         continue;
7006                 disk = conf->disks + raid_disk;
7007
7008                 if (test_bit(Replacement, &rdev->flags)) {
7009                         if (disk->replacement)
7010                                 goto abort;
7011                         disk->replacement = rdev;
7012                 } else {
7013                         if (disk->rdev)
7014                                 goto abort;
7015                         disk->rdev = rdev;
7016                 }
7017
7018                 if (test_bit(In_sync, &rdev->flags)) {
7019                         char b[BDEVNAME_SIZE];
7020                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7021                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7022                 } else if (rdev->saved_raid_disk != raid_disk)
7023                         /* Cannot rely on bitmap to complete recovery */
7024                         conf->fullsync = 1;
7025         }
7026
7027         conf->level = mddev->new_level;
7028         if (conf->level == 6) {
7029                 conf->max_degraded = 2;
7030                 if (raid6_call.xor_syndrome)
7031                         conf->rmw_level = PARITY_ENABLE_RMW;
7032                 else
7033                         conf->rmw_level = PARITY_DISABLE_RMW;
7034         } else {
7035                 conf->max_degraded = 1;
7036                 conf->rmw_level = PARITY_ENABLE_RMW;
7037         }
7038         conf->algorithm = mddev->new_layout;
7039         conf->reshape_progress = mddev->reshape_position;
7040         if (conf->reshape_progress != MaxSector) {
7041                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7042                 conf->prev_algo = mddev->layout;
7043         } else {
7044                 conf->prev_chunk_sectors = conf->chunk_sectors;
7045                 conf->prev_algo = conf->algorithm;
7046         }
7047
7048         conf->min_nr_stripes = NR_STRIPES;
7049         if (mddev->reshape_position != MaxSector) {
7050                 int stripes = max_t(int,
7051                         ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7052                         ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7053                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7054                 if (conf->min_nr_stripes != NR_STRIPES)
7055                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7056                                 mdname(mddev), conf->min_nr_stripes);
7057         }
7058         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7059                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7060         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7061         if (grow_stripes(conf, conf->min_nr_stripes)) {
7062                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7063                         mdname(mddev), memory);
7064                 goto abort;
7065         } else
7066                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7067         /*
7068          * Losing a stripe head costs more than the time to refill it,
7069          * it reduces the queue depth and so can hurt throughput.
7070          * So set it rather large, scaled by number of devices.
7071          */
7072         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7073         conf->shrinker.scan_objects = raid5_cache_scan;
7074         conf->shrinker.count_objects = raid5_cache_count;
7075         conf->shrinker.batch = 128;
7076         conf->shrinker.flags = 0;
7077         if (register_shrinker(&conf->shrinker)) {
7078                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7079                         mdname(mddev));
7080                 goto abort;
7081         }
7082
7083         sprintf(pers_name, "raid%d", mddev->new_level);
7084         conf->thread = md_register_thread(raid5d, mddev, pers_name);
7085         if (!conf->thread) {
7086                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7087                         mdname(mddev));
7088                 goto abort;
7089         }
7090
7091         return conf;
7092
7093  abort:
7094         if (conf) {
7095                 free_conf(conf);
7096                 return ERR_PTR(-EIO);
7097         } else
7098                 return ERR_PTR(-ENOMEM);
7099 }
7100
7101 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7102 {
7103         switch (algo) {
7104         case ALGORITHM_PARITY_0:
7105                 if (raid_disk < max_degraded)
7106                         return 1;
7107                 break;
7108         case ALGORITHM_PARITY_N:
7109                 if (raid_disk >= raid_disks - max_degraded)
7110                         return 1;
7111                 break;
7112         case ALGORITHM_PARITY_0_6:
7113                 if (raid_disk == 0 ||
7114                     raid_disk == raid_disks - 1)
7115                         return 1;
7116                 break;
7117         case ALGORITHM_LEFT_ASYMMETRIC_6:
7118         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7119         case ALGORITHM_LEFT_SYMMETRIC_6:
7120         case ALGORITHM_RIGHT_SYMMETRIC_6:
7121                 if (raid_disk == raid_disks - 1)
7122                         return 1;
7123         }
7124         return 0;
7125 }
7126
7127 static int raid5_run(struct mddev *mddev)
7128 {
7129         struct r5conf *conf;
7130         int working_disks = 0;
7131         int dirty_parity_disks = 0;
7132         struct md_rdev *rdev;
7133         struct md_rdev *journal_dev = NULL;
7134         sector_t reshape_offset = 0;
7135         int i;
7136         long long min_offset_diff = 0;
7137         int first = 1;
7138
7139         if (mddev_init_writes_pending(mddev) < 0)
7140                 return -ENOMEM;
7141
7142         if (mddev->recovery_cp != MaxSector)
7143                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7144                           mdname(mddev));
7145
7146         rdev_for_each(rdev, mddev) {
7147                 long long diff;
7148
7149                 if (test_bit(Journal, &rdev->flags)) {
7150                         journal_dev = rdev;
7151                         continue;
7152                 }
7153                 if (rdev->raid_disk < 0)
7154                         continue;
7155                 diff = (rdev->new_data_offset - rdev->data_offset);
7156                 if (first) {
7157                         min_offset_diff = diff;
7158                         first = 0;
7159                 } else if (mddev->reshape_backwards &&
7160                          diff < min_offset_diff)
7161                         min_offset_diff = diff;
7162                 else if (!mddev->reshape_backwards &&
7163                          diff > min_offset_diff)
7164                         min_offset_diff = diff;
7165         }
7166
7167         if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7168             (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7169                 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7170                           mdname(mddev));
7171                 return -EINVAL;
7172         }
7173
7174         if (mddev->reshape_position != MaxSector) {
7175                 /* Check that we can continue the reshape.
7176                  * Difficulties arise if the stripe we would write to
7177                  * next is at or after the stripe we would read from next.
7178                  * For a reshape that changes the number of devices, this
7179                  * is only possible for a very short time, and mdadm makes
7180                  * sure that time appears to have past before assembling
7181                  * the array.  So we fail if that time hasn't passed.
7182                  * For a reshape that keeps the number of devices the same
7183                  * mdadm must be monitoring the reshape can keeping the
7184                  * critical areas read-only and backed up.  It will start
7185                  * the array in read-only mode, so we check for that.
7186                  */
7187                 sector_t here_new, here_old;
7188                 int old_disks;
7189                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7190                 int chunk_sectors;
7191                 int new_data_disks;
7192
7193                 if (journal_dev) {
7194                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7195                                 mdname(mddev));
7196                         return -EINVAL;
7197                 }
7198
7199                 if (mddev->new_level != mddev->level) {
7200                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7201                                 mdname(mddev));
7202                         return -EINVAL;
7203                 }
7204                 old_disks = mddev->raid_disks - mddev->delta_disks;
7205                 /* reshape_position must be on a new-stripe boundary, and one
7206                  * further up in new geometry must map after here in old
7207                  * geometry.
7208                  * If the chunk sizes are different, then as we perform reshape
7209                  * in units of the largest of the two, reshape_position needs
7210                  * be a multiple of the largest chunk size times new data disks.
7211                  */
7212                 here_new = mddev->reshape_position;
7213                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7214                 new_data_disks = mddev->raid_disks - max_degraded;
7215                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7216                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7217                                 mdname(mddev));
7218                         return -EINVAL;
7219                 }
7220                 reshape_offset = here_new * chunk_sectors;
7221                 /* here_new is the stripe we will write to */
7222                 here_old = mddev->reshape_position;
7223                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7224                 /* here_old is the first stripe that we might need to read
7225                  * from */
7226                 if (mddev->delta_disks == 0) {
7227                         /* We cannot be sure it is safe to start an in-place
7228                          * reshape.  It is only safe if user-space is monitoring
7229                          * and taking constant backups.
7230                          * mdadm always starts a situation like this in
7231                          * readonly mode so it can take control before
7232                          * allowing any writes.  So just check for that.
7233                          */
7234                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7235                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7236                                 /* not really in-place - so OK */;
7237                         else if (mddev->ro == 0) {
7238                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7239                                         mdname(mddev));
7240                                 return -EINVAL;
7241                         }
7242                 } else if (mddev->reshape_backwards
7243                     ? (here_new * chunk_sectors + min_offset_diff <=
7244                        here_old * chunk_sectors)
7245                     : (here_new * chunk_sectors >=
7246                        here_old * chunk_sectors + (-min_offset_diff))) {
7247                         /* Reading from the same stripe as writing to - bad */
7248                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7249                                 mdname(mddev));
7250                         return -EINVAL;
7251                 }
7252                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7253                 /* OK, we should be able to continue; */
7254         } else {
7255                 BUG_ON(mddev->level != mddev->new_level);
7256                 BUG_ON(mddev->layout != mddev->new_layout);
7257                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7258                 BUG_ON(mddev->delta_disks != 0);
7259         }
7260
7261         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7262             test_bit(MD_HAS_PPL, &mddev->flags)) {
7263                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7264                         mdname(mddev));
7265                 clear_bit(MD_HAS_PPL, &mddev->flags);
7266                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7267         }
7268
7269         if (mddev->private == NULL)
7270                 conf = setup_conf(mddev);
7271         else
7272                 conf = mddev->private;
7273
7274         if (IS_ERR(conf))
7275                 return PTR_ERR(conf);
7276
7277         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7278                 if (!journal_dev) {
7279                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7280                                 mdname(mddev));
7281                         mddev->ro = 1;
7282                         set_disk_ro(mddev->gendisk, 1);
7283                 } else if (mddev->recovery_cp == MaxSector)
7284                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7285         }
7286
7287         conf->min_offset_diff = min_offset_diff;
7288         mddev->thread = conf->thread;
7289         conf->thread = NULL;
7290         mddev->private = conf;
7291
7292         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7293              i++) {
7294                 rdev = conf->disks[i].rdev;
7295                 if (!rdev && conf->disks[i].replacement) {
7296                         /* The replacement is all we have yet */
7297                         rdev = conf->disks[i].replacement;
7298                         conf->disks[i].replacement = NULL;
7299                         clear_bit(Replacement, &rdev->flags);
7300                         conf->disks[i].rdev = rdev;
7301                 }
7302                 if (!rdev)
7303                         continue;
7304                 if (conf->disks[i].replacement &&
7305                     conf->reshape_progress != MaxSector) {
7306                         /* replacements and reshape simply do not mix. */
7307                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7308                         goto abort;
7309                 }
7310                 if (test_bit(In_sync, &rdev->flags)) {
7311                         working_disks++;
7312                         continue;
7313                 }
7314                 /* This disc is not fully in-sync.  However if it
7315                  * just stored parity (beyond the recovery_offset),
7316                  * when we don't need to be concerned about the
7317                  * array being dirty.
7318                  * When reshape goes 'backwards', we never have
7319                  * partially completed devices, so we only need
7320                  * to worry about reshape going forwards.
7321                  */
7322                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7323                 if (mddev->major_version == 0 &&
7324                     mddev->minor_version > 90)
7325                         rdev->recovery_offset = reshape_offset;
7326
7327                 if (rdev->recovery_offset < reshape_offset) {
7328                         /* We need to check old and new layout */
7329                         if (!only_parity(rdev->raid_disk,
7330                                          conf->algorithm,
7331                                          conf->raid_disks,
7332                                          conf->max_degraded))
7333                                 continue;
7334                 }
7335                 if (!only_parity(rdev->raid_disk,
7336                                  conf->prev_algo,
7337                                  conf->previous_raid_disks,
7338                                  conf->max_degraded))
7339                         continue;
7340                 dirty_parity_disks++;
7341         }
7342
7343         /*
7344          * 0 for a fully functional array, 1 or 2 for a degraded array.
7345          */
7346         mddev->degraded = raid5_calc_degraded(conf);
7347
7348         if (has_failed(conf)) {
7349                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7350                         mdname(mddev), mddev->degraded, conf->raid_disks);
7351                 goto abort;
7352         }
7353
7354         /* device size must be a multiple of chunk size */
7355         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7356         mddev->resync_max_sectors = mddev->dev_sectors;
7357
7358         if (mddev->degraded > dirty_parity_disks &&
7359             mddev->recovery_cp != MaxSector) {
7360                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7361                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7362                                 mdname(mddev));
7363                 else if (mddev->ok_start_degraded)
7364                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7365                                 mdname(mddev));
7366                 else {
7367                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7368                                 mdname(mddev));
7369                         goto abort;
7370                 }
7371         }
7372
7373         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7374                 mdname(mddev), conf->level,
7375                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7376                 mddev->new_layout);
7377
7378         print_raid5_conf(conf);
7379
7380         if (conf->reshape_progress != MaxSector) {
7381                 conf->reshape_safe = conf->reshape_progress;
7382                 atomic_set(&conf->reshape_stripes, 0);
7383                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7384                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7385                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7386                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7387                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7388                                                         "reshape");
7389         }
7390
7391         /* Ok, everything is just fine now */
7392         if (mddev->to_remove == &raid5_attrs_group)
7393                 mddev->to_remove = NULL;
7394         else if (mddev->kobj.sd &&
7395             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7396                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7397                         mdname(mddev));
7398         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7399
7400         if (mddev->queue) {
7401                 int chunk_size;
7402                 /* read-ahead size must cover two whole stripes, which
7403                  * is 2 * (datadisks) * chunksize where 'n' is the
7404                  * number of raid devices
7405                  */
7406                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7407                 int stripe = data_disks *
7408                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7409                 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7410                         mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7411
7412                 chunk_size = mddev->chunk_sectors << 9;
7413                 blk_queue_io_min(mddev->queue, chunk_size);
7414                 blk_queue_io_opt(mddev->queue, chunk_size *
7415                                  (conf->raid_disks - conf->max_degraded));
7416                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7417                 /*
7418                  * We can only discard a whole stripe. It doesn't make sense to
7419                  * discard data disk but write parity disk
7420                  */
7421                 stripe = stripe * PAGE_SIZE;
7422                 /* Round up to power of 2, as discard handling
7423                  * currently assumes that */
7424                 while ((stripe-1) & stripe)
7425                         stripe = (stripe | (stripe-1)) + 1;
7426                 mddev->queue->limits.discard_alignment = stripe;
7427                 mddev->queue->limits.discard_granularity = stripe;
7428
7429                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7430                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7431
7432                 rdev_for_each(rdev, mddev) {
7433                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7434                                           rdev->data_offset << 9);
7435                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7436                                           rdev->new_data_offset << 9);
7437                 }
7438
7439                 /*
7440                  * zeroing is required, otherwise data
7441                  * could be lost. Consider a scenario: discard a stripe
7442                  * (the stripe could be inconsistent if
7443                  * discard_zeroes_data is 0); write one disk of the
7444                  * stripe (the stripe could be inconsistent again
7445                  * depending on which disks are used to calculate
7446                  * parity); the disk is broken; The stripe data of this
7447                  * disk is lost.
7448                  *
7449                  * We only allow DISCARD if the sysadmin has confirmed that
7450                  * only safe devices are in use by setting a module parameter.
7451                  * A better idea might be to turn DISCARD into WRITE_ZEROES
7452                  * requests, as that is required to be safe.
7453                  */
7454                 if (devices_handle_discard_safely &&
7455                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7456                     mddev->queue->limits.discard_granularity >= stripe)
7457                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7458                                                 mddev->queue);
7459                 else
7460                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7461                                                 mddev->queue);
7462
7463                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7464         }
7465
7466         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7467                 goto abort;
7468
7469         return 0;
7470 abort:
7471         md_unregister_thread(&mddev->thread);
7472         print_raid5_conf(conf);
7473         free_conf(conf);
7474         mddev->private = NULL;
7475         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7476         return -EIO;
7477 }
7478
7479 static void raid5_free(struct mddev *mddev, void *priv)
7480 {
7481         struct r5conf *conf = priv;
7482
7483         free_conf(conf);
7484         mddev->to_remove = &raid5_attrs_group;
7485 }
7486
7487 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7488 {
7489         struct r5conf *conf = mddev->private;
7490         int i;
7491
7492         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7493                 conf->chunk_sectors / 2, mddev->layout);
7494         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7495         rcu_read_lock();
7496         for (i = 0; i < conf->raid_disks; i++) {
7497                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7498                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7499         }
7500         rcu_read_unlock();
7501         seq_printf (seq, "]");
7502 }
7503
7504 static void print_raid5_conf (struct r5conf *conf)
7505 {
7506         int i;
7507         struct disk_info *tmp;
7508
7509         pr_debug("RAID conf printout:\n");
7510         if (!conf) {
7511                 pr_debug("(conf==NULL)\n");
7512                 return;
7513         }
7514         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7515                conf->raid_disks,
7516                conf->raid_disks - conf->mddev->degraded);
7517
7518         for (i = 0; i < conf->raid_disks; i++) {
7519                 char b[BDEVNAME_SIZE];
7520                 tmp = conf->disks + i;
7521                 if (tmp->rdev)
7522                         pr_debug(" disk %d, o:%d, dev:%s\n",
7523                                i, !test_bit(Faulty, &tmp->rdev->flags),
7524                                bdevname(tmp->rdev->bdev, b));
7525         }
7526 }
7527
7528 static int raid5_spare_active(struct mddev *mddev)
7529 {
7530         int i;
7531         struct r5conf *conf = mddev->private;
7532         struct disk_info *tmp;
7533         int count = 0;
7534         unsigned long flags;
7535
7536         for (i = 0; i < conf->raid_disks; i++) {
7537                 tmp = conf->disks + i;
7538                 if (tmp->replacement
7539                     && tmp->replacement->recovery_offset == MaxSector
7540                     && !test_bit(Faulty, &tmp->replacement->flags)
7541                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7542                         /* Replacement has just become active. */
7543                         if (!tmp->rdev
7544                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7545                                 count++;
7546                         if (tmp->rdev) {
7547                                 /* Replaced device not technically faulty,
7548                                  * but we need to be sure it gets removed
7549                                  * and never re-added.
7550                                  */
7551                                 set_bit(Faulty, &tmp->rdev->flags);
7552                                 sysfs_notify_dirent_safe(
7553                                         tmp->rdev->sysfs_state);
7554                         }
7555                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7556                 } else if (tmp->rdev
7557                     && tmp->rdev->recovery_offset == MaxSector
7558                     && !test_bit(Faulty, &tmp->rdev->flags)
7559                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7560                         count++;
7561                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7562                 }
7563         }
7564         spin_lock_irqsave(&conf->device_lock, flags);
7565         mddev->degraded = raid5_calc_degraded(conf);
7566         spin_unlock_irqrestore(&conf->device_lock, flags);
7567         print_raid5_conf(conf);
7568         return count;
7569 }
7570
7571 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7572 {
7573         struct r5conf *conf = mddev->private;
7574         int err = 0;
7575         int number = rdev->raid_disk;
7576         struct md_rdev **rdevp;
7577         struct disk_info *p = conf->disks + number;
7578
7579         print_raid5_conf(conf);
7580         if (test_bit(Journal, &rdev->flags) && conf->log) {
7581                 /*
7582                  * we can't wait pending write here, as this is called in
7583                  * raid5d, wait will deadlock.
7584                  * neilb: there is no locking about new writes here,
7585                  * so this cannot be safe.
7586                  */
7587                 if (atomic_read(&conf->active_stripes) ||
7588                     atomic_read(&conf->r5c_cached_full_stripes) ||
7589                     atomic_read(&conf->r5c_cached_partial_stripes)) {
7590                         return -EBUSY;
7591                 }
7592                 log_exit(conf);
7593                 return 0;
7594         }
7595         if (rdev == p->rdev)
7596                 rdevp = &p->rdev;
7597         else if (rdev == p->replacement)
7598                 rdevp = &p->replacement;
7599         else
7600                 return 0;
7601
7602         if (number >= conf->raid_disks &&
7603             conf->reshape_progress == MaxSector)
7604                 clear_bit(In_sync, &rdev->flags);
7605
7606         if (test_bit(In_sync, &rdev->flags) ||
7607             atomic_read(&rdev->nr_pending)) {
7608                 err = -EBUSY;
7609                 goto abort;
7610         }
7611         /* Only remove non-faulty devices if recovery
7612          * isn't possible.
7613          */
7614         if (!test_bit(Faulty, &rdev->flags) &&
7615             mddev->recovery_disabled != conf->recovery_disabled &&
7616             !has_failed(conf) &&
7617             (!p->replacement || p->replacement == rdev) &&
7618             number < conf->raid_disks) {
7619                 err = -EBUSY;
7620                 goto abort;
7621         }
7622         *rdevp = NULL;
7623         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7624                 synchronize_rcu();
7625                 if (atomic_read(&rdev->nr_pending)) {
7626                         /* lost the race, try later */
7627                         err = -EBUSY;
7628                         *rdevp = rdev;
7629                 }
7630         }
7631         if (!err) {
7632                 err = log_modify(conf, rdev, false);
7633                 if (err)
7634                         goto abort;
7635         }
7636         if (p->replacement) {
7637                 /* We must have just cleared 'rdev' */
7638                 p->rdev = p->replacement;
7639                 clear_bit(Replacement, &p->replacement->flags);
7640                 smp_mb(); /* Make sure other CPUs may see both as identical
7641                            * but will never see neither - if they are careful
7642                            */
7643                 p->replacement = NULL;
7644
7645                 if (!err)
7646                         err = log_modify(conf, p->rdev, true);
7647         }
7648
7649         clear_bit(WantReplacement, &rdev->flags);
7650 abort:
7651
7652         print_raid5_conf(conf);
7653         return err;
7654 }
7655
7656 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7657 {
7658         struct r5conf *conf = mddev->private;
7659         int err = -EEXIST;
7660         int disk;
7661         struct disk_info *p;
7662         int first = 0;
7663         int last = conf->raid_disks - 1;
7664
7665         if (test_bit(Journal, &rdev->flags)) {
7666                 if (conf->log)
7667                         return -EBUSY;
7668
7669                 rdev->raid_disk = 0;
7670                 /*
7671                  * The array is in readonly mode if journal is missing, so no
7672                  * write requests running. We should be safe
7673                  */
7674                 log_init(conf, rdev, false);
7675                 return 0;
7676         }
7677         if (mddev->recovery_disabled == conf->recovery_disabled)
7678                 return -EBUSY;
7679
7680         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7681                 /* no point adding a device */
7682                 return -EINVAL;
7683
7684         if (rdev->raid_disk >= 0)
7685                 first = last = rdev->raid_disk;
7686
7687         /*
7688          * find the disk ... but prefer rdev->saved_raid_disk
7689          * if possible.
7690          */
7691         if (rdev->saved_raid_disk >= 0 &&
7692             rdev->saved_raid_disk >= first &&
7693             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7694                 first = rdev->saved_raid_disk;
7695
7696         for (disk = first; disk <= last; disk++) {
7697                 p = conf->disks + disk;
7698                 if (p->rdev == NULL) {
7699                         clear_bit(In_sync, &rdev->flags);
7700                         rdev->raid_disk = disk;
7701                         if (rdev->saved_raid_disk != disk)
7702                                 conf->fullsync = 1;
7703                         rcu_assign_pointer(p->rdev, rdev);
7704
7705                         err = log_modify(conf, rdev, true);
7706
7707                         goto out;
7708                 }
7709         }
7710         for (disk = first; disk <= last; disk++) {
7711                 p = conf->disks + disk;
7712                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7713                     p->replacement == NULL) {
7714                         clear_bit(In_sync, &rdev->flags);
7715                         set_bit(Replacement, &rdev->flags);
7716                         rdev->raid_disk = disk;
7717                         err = 0;
7718                         conf->fullsync = 1;
7719                         rcu_assign_pointer(p->replacement, rdev);
7720                         break;
7721                 }
7722         }
7723 out:
7724         print_raid5_conf(conf);
7725         return err;
7726 }
7727
7728 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7729 {
7730         /* no resync is happening, and there is enough space
7731          * on all devices, so we can resize.
7732          * We need to make sure resync covers any new space.
7733          * If the array is shrinking we should possibly wait until
7734          * any io in the removed space completes, but it hardly seems
7735          * worth it.
7736          */
7737         sector_t newsize;
7738         struct r5conf *conf = mddev->private;
7739
7740         if (conf->log || raid5_has_ppl(conf))
7741                 return -EINVAL;
7742         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7743         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7744         if (mddev->external_size &&
7745             mddev->array_sectors > newsize)
7746                 return -EINVAL;
7747         if (mddev->bitmap) {
7748                 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7749                 if (ret)
7750                         return ret;
7751         }
7752         md_set_array_sectors(mddev, newsize);
7753         if (sectors > mddev->dev_sectors &&
7754             mddev->recovery_cp > mddev->dev_sectors) {
7755                 mddev->recovery_cp = mddev->dev_sectors;
7756                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7757         }
7758         mddev->dev_sectors = sectors;
7759         mddev->resync_max_sectors = sectors;
7760         return 0;
7761 }
7762
7763 static int check_stripe_cache(struct mddev *mddev)
7764 {
7765         /* Can only proceed if there are plenty of stripe_heads.
7766          * We need a minimum of one full stripe,, and for sensible progress
7767          * it is best to have about 4 times that.
7768          * If we require 4 times, then the default 256 4K stripe_heads will
7769          * allow for chunk sizes up to 256K, which is probably OK.
7770          * If the chunk size is greater, user-space should request more
7771          * stripe_heads first.
7772          */
7773         struct r5conf *conf = mddev->private;
7774         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7775             > conf->min_nr_stripes ||
7776             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7777             > conf->min_nr_stripes) {
7778                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7779                         mdname(mddev),
7780                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7781                          / STRIPE_SIZE)*4);
7782                 return 0;
7783         }
7784         return 1;
7785 }
7786
7787 static int check_reshape(struct mddev *mddev)
7788 {
7789         struct r5conf *conf = mddev->private;
7790
7791         if (conf->log || raid5_has_ppl(conf))
7792                 return -EINVAL;
7793         if (mddev->delta_disks == 0 &&
7794             mddev->new_layout == mddev->layout &&
7795             mddev->new_chunk_sectors == mddev->chunk_sectors)
7796                 return 0; /* nothing to do */
7797         if (has_failed(conf))
7798                 return -EINVAL;
7799         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7800                 /* We might be able to shrink, but the devices must
7801                  * be made bigger first.
7802                  * For raid6, 4 is the minimum size.
7803                  * Otherwise 2 is the minimum
7804                  */
7805                 int min = 2;
7806                 if (mddev->level == 6)
7807                         min = 4;
7808                 if (mddev->raid_disks + mddev->delta_disks < min)
7809                         return -EINVAL;
7810         }
7811
7812         if (!check_stripe_cache(mddev))
7813                 return -ENOSPC;
7814
7815         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7816             mddev->delta_disks > 0)
7817                 if (resize_chunks(conf,
7818                                   conf->previous_raid_disks
7819                                   + max(0, mddev->delta_disks),
7820                                   max(mddev->new_chunk_sectors,
7821                                       mddev->chunk_sectors)
7822                             ) < 0)
7823                         return -ENOMEM;
7824
7825         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7826                 return 0; /* never bother to shrink */
7827         return resize_stripes(conf, (conf->previous_raid_disks
7828                                      + mddev->delta_disks));
7829 }
7830
7831 static int raid5_start_reshape(struct mddev *mddev)
7832 {
7833         struct r5conf *conf = mddev->private;
7834         struct md_rdev *rdev;
7835         int spares = 0;
7836         unsigned long flags;
7837
7838         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7839                 return -EBUSY;
7840
7841         if (!check_stripe_cache(mddev))
7842                 return -ENOSPC;
7843
7844         if (has_failed(conf))
7845                 return -EINVAL;
7846
7847         rdev_for_each(rdev, mddev) {
7848                 if (!test_bit(In_sync, &rdev->flags)
7849                     && !test_bit(Faulty, &rdev->flags))
7850                         spares++;
7851         }
7852
7853         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7854                 /* Not enough devices even to make a degraded array
7855                  * of that size
7856                  */
7857                 return -EINVAL;
7858
7859         /* Refuse to reduce size of the array.  Any reductions in
7860          * array size must be through explicit setting of array_size
7861          * attribute.
7862          */
7863         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7864             < mddev->array_sectors) {
7865                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7866                         mdname(mddev));
7867                 return -EINVAL;
7868         }
7869
7870         atomic_set(&conf->reshape_stripes, 0);
7871         spin_lock_irq(&conf->device_lock);
7872         write_seqcount_begin(&conf->gen_lock);
7873         conf->previous_raid_disks = conf->raid_disks;
7874         conf->raid_disks += mddev->delta_disks;
7875         conf->prev_chunk_sectors = conf->chunk_sectors;
7876         conf->chunk_sectors = mddev->new_chunk_sectors;
7877         conf->prev_algo = conf->algorithm;
7878         conf->algorithm = mddev->new_layout;
7879         conf->generation++;
7880         /* Code that selects data_offset needs to see the generation update
7881          * if reshape_progress has been set - so a memory barrier needed.
7882          */
7883         smp_mb();
7884         if (mddev->reshape_backwards)
7885                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7886         else
7887                 conf->reshape_progress = 0;
7888         conf->reshape_safe = conf->reshape_progress;
7889         write_seqcount_end(&conf->gen_lock);
7890         spin_unlock_irq(&conf->device_lock);
7891
7892         /* Now make sure any requests that proceeded on the assumption
7893          * the reshape wasn't running - like Discard or Read - have
7894          * completed.
7895          */
7896         mddev_suspend(mddev);
7897         mddev_resume(mddev);
7898
7899         /* Add some new drives, as many as will fit.
7900          * We know there are enough to make the newly sized array work.
7901          * Don't add devices if we are reducing the number of
7902          * devices in the array.  This is because it is not possible
7903          * to correctly record the "partially reconstructed" state of
7904          * such devices during the reshape and confusion could result.
7905          */
7906         if (mddev->delta_disks >= 0) {
7907                 rdev_for_each(rdev, mddev)
7908                         if (rdev->raid_disk < 0 &&
7909                             !test_bit(Faulty, &rdev->flags)) {
7910                                 if (raid5_add_disk(mddev, rdev) == 0) {
7911                                         if (rdev->raid_disk
7912                                             >= conf->previous_raid_disks)
7913                                                 set_bit(In_sync, &rdev->flags);
7914                                         else
7915                                                 rdev->recovery_offset = 0;
7916
7917                                         if (sysfs_link_rdev(mddev, rdev))
7918                                                 /* Failure here is OK */;
7919                                 }
7920                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7921                                    && !test_bit(Faulty, &rdev->flags)) {
7922                                 /* This is a spare that was manually added */
7923                                 set_bit(In_sync, &rdev->flags);
7924                         }
7925
7926                 /* When a reshape changes the number of devices,
7927                  * ->degraded is measured against the larger of the
7928                  * pre and post number of devices.
7929                  */
7930                 spin_lock_irqsave(&conf->device_lock, flags);
7931                 mddev->degraded = raid5_calc_degraded(conf);
7932                 spin_unlock_irqrestore(&conf->device_lock, flags);
7933         }
7934         mddev->raid_disks = conf->raid_disks;
7935         mddev->reshape_position = conf->reshape_progress;
7936         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7937
7938         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7939         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7940         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7941         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7942         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7943         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7944                                                 "reshape");
7945         if (!mddev->sync_thread) {
7946                 mddev->recovery = 0;
7947                 spin_lock_irq(&conf->device_lock);
7948                 write_seqcount_begin(&conf->gen_lock);
7949                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7950                 mddev->new_chunk_sectors =
7951                         conf->chunk_sectors = conf->prev_chunk_sectors;
7952                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7953                 rdev_for_each(rdev, mddev)
7954                         rdev->new_data_offset = rdev->data_offset;
7955                 smp_wmb();
7956                 conf->generation --;
7957                 conf->reshape_progress = MaxSector;
7958                 mddev->reshape_position = MaxSector;
7959                 write_seqcount_end(&conf->gen_lock);
7960                 spin_unlock_irq(&conf->device_lock);
7961                 return -EAGAIN;
7962         }
7963         conf->reshape_checkpoint = jiffies;
7964         md_wakeup_thread(mddev->sync_thread);
7965         md_new_event(mddev);
7966         return 0;
7967 }
7968
7969 /* This is called from the reshape thread and should make any
7970  * changes needed in 'conf'
7971  */
7972 static void end_reshape(struct r5conf *conf)
7973 {
7974
7975         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7976                 struct md_rdev *rdev;
7977
7978                 spin_lock_irq(&conf->device_lock);
7979                 conf->previous_raid_disks = conf->raid_disks;
7980                 md_finish_reshape(conf->mddev);
7981                 smp_wmb();
7982                 conf->reshape_progress = MaxSector;
7983                 conf->mddev->reshape_position = MaxSector;
7984                 rdev_for_each(rdev, conf->mddev)
7985                         if (rdev->raid_disk >= 0 &&
7986                             !test_bit(Journal, &rdev->flags) &&
7987                             !test_bit(In_sync, &rdev->flags))
7988                                 rdev->recovery_offset = MaxSector;
7989                 spin_unlock_irq(&conf->device_lock);
7990                 wake_up(&conf->wait_for_overlap);
7991
7992                 /* read-ahead size must cover two whole stripes, which is
7993                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7994                  */
7995                 if (conf->mddev->queue) {
7996                         int data_disks = conf->raid_disks - conf->max_degraded;
7997                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7998                                                    / PAGE_SIZE);
7999                         if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8000                                 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8001                 }
8002         }
8003 }
8004
8005 /* This is called from the raid5d thread with mddev_lock held.
8006  * It makes config changes to the device.
8007  */
8008 static void raid5_finish_reshape(struct mddev *mddev)
8009 {
8010         struct r5conf *conf = mddev->private;
8011
8012         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8013
8014                 if (mddev->delta_disks <= 0) {
8015                         int d;
8016                         spin_lock_irq(&conf->device_lock);
8017                         mddev->degraded = raid5_calc_degraded(conf);
8018                         spin_unlock_irq(&conf->device_lock);
8019                         for (d = conf->raid_disks ;
8020                              d < conf->raid_disks - mddev->delta_disks;
8021                              d++) {
8022                                 struct md_rdev *rdev = conf->disks[d].rdev;
8023                                 if (rdev)
8024                                         clear_bit(In_sync, &rdev->flags);
8025                                 rdev = conf->disks[d].replacement;
8026                                 if (rdev)
8027                                         clear_bit(In_sync, &rdev->flags);
8028                         }
8029                 }
8030                 mddev->layout = conf->algorithm;
8031                 mddev->chunk_sectors = conf->chunk_sectors;
8032                 mddev->reshape_position = MaxSector;
8033                 mddev->delta_disks = 0;
8034                 mddev->reshape_backwards = 0;
8035         }
8036 }
8037
8038 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8039 {
8040         struct r5conf *conf = mddev->private;
8041
8042         if (quiesce) {
8043                 /* stop all writes */
8044                 lock_all_device_hash_locks_irq(conf);
8045                 /* '2' tells resync/reshape to pause so that all
8046                  * active stripes can drain
8047                  */
8048                 r5c_flush_cache(conf, INT_MAX);
8049                 conf->quiesce = 2;
8050                 wait_event_cmd(conf->wait_for_quiescent,
8051                                     atomic_read(&conf->active_stripes) == 0 &&
8052                                     atomic_read(&conf->active_aligned_reads) == 0,
8053                                     unlock_all_device_hash_locks_irq(conf),
8054                                     lock_all_device_hash_locks_irq(conf));
8055                 conf->quiesce = 1;
8056                 unlock_all_device_hash_locks_irq(conf);
8057                 /* allow reshape to continue */
8058                 wake_up(&conf->wait_for_overlap);
8059         } else {
8060                 /* re-enable writes */
8061                 lock_all_device_hash_locks_irq(conf);
8062                 conf->quiesce = 0;
8063                 wake_up(&conf->wait_for_quiescent);
8064                 wake_up(&conf->wait_for_overlap);
8065                 unlock_all_device_hash_locks_irq(conf);
8066         }
8067         log_quiesce(conf, quiesce);
8068 }
8069
8070 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8071 {
8072         struct r0conf *raid0_conf = mddev->private;
8073         sector_t sectors;
8074
8075         /* for raid0 takeover only one zone is supported */
8076         if (raid0_conf->nr_strip_zones > 1) {
8077                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8078                         mdname(mddev));
8079                 return ERR_PTR(-EINVAL);
8080         }
8081
8082         sectors = raid0_conf->strip_zone[0].zone_end;
8083         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8084         mddev->dev_sectors = sectors;
8085         mddev->new_level = level;
8086         mddev->new_layout = ALGORITHM_PARITY_N;
8087         mddev->new_chunk_sectors = mddev->chunk_sectors;
8088         mddev->raid_disks += 1;
8089         mddev->delta_disks = 1;
8090         /* make sure it will be not marked as dirty */
8091         mddev->recovery_cp = MaxSector;
8092
8093         return setup_conf(mddev);
8094 }
8095
8096 static void *raid5_takeover_raid1(struct mddev *mddev)
8097 {
8098         int chunksect;
8099         void *ret;
8100
8101         if (mddev->raid_disks != 2 ||
8102             mddev->degraded > 1)
8103                 return ERR_PTR(-EINVAL);
8104
8105         /* Should check if there are write-behind devices? */
8106
8107         chunksect = 64*2; /* 64K by default */
8108
8109         /* The array must be an exact multiple of chunksize */
8110         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8111                 chunksect >>= 1;
8112
8113         if ((chunksect<<9) < STRIPE_SIZE)
8114                 /* array size does not allow a suitable chunk size */
8115                 return ERR_PTR(-EINVAL);
8116
8117         mddev->new_level = 5;
8118         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8119         mddev->new_chunk_sectors = chunksect;
8120
8121         ret = setup_conf(mddev);
8122         if (!IS_ERR(ret))
8123                 mddev_clear_unsupported_flags(mddev,
8124                         UNSUPPORTED_MDDEV_FLAGS);
8125         return ret;
8126 }
8127
8128 static void *raid5_takeover_raid6(struct mddev *mddev)
8129 {
8130         int new_layout;
8131
8132         switch (mddev->layout) {
8133         case ALGORITHM_LEFT_ASYMMETRIC_6:
8134                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8135                 break;
8136         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8137                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8138                 break;
8139         case ALGORITHM_LEFT_SYMMETRIC_6:
8140                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8141                 break;
8142         case ALGORITHM_RIGHT_SYMMETRIC_6:
8143                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8144                 break;
8145         case ALGORITHM_PARITY_0_6:
8146                 new_layout = ALGORITHM_PARITY_0;
8147                 break;
8148         case ALGORITHM_PARITY_N:
8149                 new_layout = ALGORITHM_PARITY_N;
8150                 break;
8151         default:
8152                 return ERR_PTR(-EINVAL);
8153         }
8154         mddev->new_level = 5;
8155         mddev->new_layout = new_layout;
8156         mddev->delta_disks = -1;
8157         mddev->raid_disks -= 1;
8158         return setup_conf(mddev);
8159 }
8160
8161 static int raid5_check_reshape(struct mddev *mddev)
8162 {
8163         /* For a 2-drive array, the layout and chunk size can be changed
8164          * immediately as not restriping is needed.
8165          * For larger arrays we record the new value - after validation
8166          * to be used by a reshape pass.
8167          */
8168         struct r5conf *conf = mddev->private;
8169         int new_chunk = mddev->new_chunk_sectors;
8170
8171         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8172                 return -EINVAL;
8173         if (new_chunk > 0) {
8174                 if (!is_power_of_2(new_chunk))
8175                         return -EINVAL;
8176                 if (new_chunk < (PAGE_SIZE>>9))
8177                         return -EINVAL;
8178                 if (mddev->array_sectors & (new_chunk-1))
8179                         /* not factor of array size */
8180                         return -EINVAL;
8181         }
8182
8183         /* They look valid */
8184
8185         if (mddev->raid_disks == 2) {
8186                 /* can make the change immediately */
8187                 if (mddev->new_layout >= 0) {
8188                         conf->algorithm = mddev->new_layout;
8189                         mddev->layout = mddev->new_layout;
8190                 }
8191                 if (new_chunk > 0) {
8192                         conf->chunk_sectors = new_chunk ;
8193                         mddev->chunk_sectors = new_chunk;
8194                 }
8195                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8196                 md_wakeup_thread(mddev->thread);
8197         }
8198         return check_reshape(mddev);
8199 }
8200
8201 static int raid6_check_reshape(struct mddev *mddev)
8202 {
8203         int new_chunk = mddev->new_chunk_sectors;
8204
8205         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8206                 return -EINVAL;
8207         if (new_chunk > 0) {
8208                 if (!is_power_of_2(new_chunk))
8209                         return -EINVAL;
8210                 if (new_chunk < (PAGE_SIZE >> 9))
8211                         return -EINVAL;
8212                 if (mddev->array_sectors & (new_chunk-1))
8213                         /* not factor of array size */
8214                         return -EINVAL;
8215         }
8216
8217         /* They look valid */
8218         return check_reshape(mddev);
8219 }
8220
8221 static void *raid5_takeover(struct mddev *mddev)
8222 {
8223         /* raid5 can take over:
8224          *  raid0 - if there is only one strip zone - make it a raid4 layout
8225          *  raid1 - if there are two drives.  We need to know the chunk size
8226          *  raid4 - trivial - just use a raid4 layout.
8227          *  raid6 - Providing it is a *_6 layout
8228          */
8229         if (mddev->level == 0)
8230                 return raid45_takeover_raid0(mddev, 5);
8231         if (mddev->level == 1)
8232                 return raid5_takeover_raid1(mddev);
8233         if (mddev->level == 4) {
8234                 mddev->new_layout = ALGORITHM_PARITY_N;
8235                 mddev->new_level = 5;
8236                 return setup_conf(mddev);
8237         }
8238         if (mddev->level == 6)
8239                 return raid5_takeover_raid6(mddev);
8240
8241         return ERR_PTR(-EINVAL);
8242 }
8243
8244 static void *raid4_takeover(struct mddev *mddev)
8245 {
8246         /* raid4 can take over:
8247          *  raid0 - if there is only one strip zone
8248          *  raid5 - if layout is right
8249          */
8250         if (mddev->level == 0)
8251                 return raid45_takeover_raid0(mddev, 4);
8252         if (mddev->level == 5 &&
8253             mddev->layout == ALGORITHM_PARITY_N) {
8254                 mddev->new_layout = 0;
8255                 mddev->new_level = 4;
8256                 return setup_conf(mddev);
8257         }
8258         return ERR_PTR(-EINVAL);
8259 }
8260
8261 static struct md_personality raid5_personality;
8262
8263 static void *raid6_takeover(struct mddev *mddev)
8264 {
8265         /* Currently can only take over a raid5.  We map the
8266          * personality to an equivalent raid6 personality
8267          * with the Q block at the end.
8268          */
8269         int new_layout;
8270
8271         if (mddev->pers != &raid5_personality)
8272                 return ERR_PTR(-EINVAL);
8273         if (mddev->degraded > 1)
8274                 return ERR_PTR(-EINVAL);
8275         if (mddev->raid_disks > 253)
8276                 return ERR_PTR(-EINVAL);
8277         if (mddev->raid_disks < 3)
8278                 return ERR_PTR(-EINVAL);
8279
8280         switch (mddev->layout) {
8281         case ALGORITHM_LEFT_ASYMMETRIC:
8282                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8283                 break;
8284         case ALGORITHM_RIGHT_ASYMMETRIC:
8285                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8286                 break;
8287         case ALGORITHM_LEFT_SYMMETRIC:
8288                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8289                 break;
8290         case ALGORITHM_RIGHT_SYMMETRIC:
8291                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8292                 break;
8293         case ALGORITHM_PARITY_0:
8294                 new_layout = ALGORITHM_PARITY_0_6;
8295                 break;
8296         case ALGORITHM_PARITY_N:
8297                 new_layout = ALGORITHM_PARITY_N;
8298                 break;
8299         default:
8300                 return ERR_PTR(-EINVAL);
8301         }
8302         mddev->new_level = 6;
8303         mddev->new_layout = new_layout;
8304         mddev->delta_disks = 1;
8305         mddev->raid_disks += 1;
8306         return setup_conf(mddev);
8307 }
8308
8309 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8310 {
8311         struct r5conf *conf;
8312         int err;
8313
8314         err = mddev_lock(mddev);
8315         if (err)
8316                 return err;
8317         conf = mddev->private;
8318         if (!conf) {
8319                 mddev_unlock(mddev);
8320                 return -ENODEV;
8321         }
8322
8323         if (strncmp(buf, "ppl", 3) == 0) {
8324                 /* ppl only works with RAID 5 */
8325                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8326                         err = log_init(conf, NULL, true);
8327                         if (!err) {
8328                                 err = resize_stripes(conf, conf->pool_size);
8329                                 if (err)
8330                                         log_exit(conf);
8331                         }
8332                 } else
8333                         err = -EINVAL;
8334         } else if (strncmp(buf, "resync", 6) == 0) {
8335                 if (raid5_has_ppl(conf)) {
8336                         mddev_suspend(mddev);
8337                         log_exit(conf);
8338                         mddev_resume(mddev);
8339                         err = resize_stripes(conf, conf->pool_size);
8340                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8341                            r5l_log_disk_error(conf)) {
8342                         bool journal_dev_exists = false;
8343                         struct md_rdev *rdev;
8344
8345                         rdev_for_each(rdev, mddev)
8346                                 if (test_bit(Journal, &rdev->flags)) {
8347                                         journal_dev_exists = true;
8348                                         break;
8349                                 }
8350
8351                         if (!journal_dev_exists) {
8352                                 mddev_suspend(mddev);
8353                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8354                                 mddev_resume(mddev);
8355                         } else  /* need remove journal device first */
8356                                 err = -EBUSY;
8357                 } else
8358                         err = -EINVAL;
8359         } else {
8360                 err = -EINVAL;
8361         }
8362
8363         if (!err)
8364                 md_update_sb(mddev, 1);
8365
8366         mddev_unlock(mddev);
8367
8368         return err;
8369 }
8370
8371 static int raid5_start(struct mddev *mddev)
8372 {
8373         struct r5conf *conf = mddev->private;
8374
8375         return r5l_start(conf->log);
8376 }
8377
8378 static struct md_personality raid6_personality =
8379 {
8380         .name           = "raid6",
8381         .level          = 6,
8382         .owner          = THIS_MODULE,
8383         .make_request   = raid5_make_request,
8384         .run            = raid5_run,
8385         .start          = raid5_start,
8386         .free           = raid5_free,
8387         .status         = raid5_status,
8388         .error_handler  = raid5_error,
8389         .hot_add_disk   = raid5_add_disk,
8390         .hot_remove_disk= raid5_remove_disk,
8391         .spare_active   = raid5_spare_active,
8392         .sync_request   = raid5_sync_request,
8393         .resize         = raid5_resize,
8394         .size           = raid5_size,
8395         .check_reshape  = raid6_check_reshape,
8396         .start_reshape  = raid5_start_reshape,
8397         .finish_reshape = raid5_finish_reshape,
8398         .quiesce        = raid5_quiesce,
8399         .takeover       = raid6_takeover,
8400         .congested      = raid5_congested,
8401         .change_consistency_policy = raid5_change_consistency_policy,
8402 };
8403 static struct md_personality raid5_personality =
8404 {
8405         .name           = "raid5",
8406         .level          = 5,
8407         .owner          = THIS_MODULE,
8408         .make_request   = raid5_make_request,
8409         .run            = raid5_run,
8410         .start          = raid5_start,
8411         .free           = raid5_free,
8412         .status         = raid5_status,
8413         .error_handler  = raid5_error,
8414         .hot_add_disk   = raid5_add_disk,
8415         .hot_remove_disk= raid5_remove_disk,
8416         .spare_active   = raid5_spare_active,
8417         .sync_request   = raid5_sync_request,
8418         .resize         = raid5_resize,
8419         .size           = raid5_size,
8420         .check_reshape  = raid5_check_reshape,
8421         .start_reshape  = raid5_start_reshape,
8422         .finish_reshape = raid5_finish_reshape,
8423         .quiesce        = raid5_quiesce,
8424         .takeover       = raid5_takeover,
8425         .congested      = raid5_congested,
8426         .change_consistency_policy = raid5_change_consistency_policy,
8427 };
8428
8429 static struct md_personality raid4_personality =
8430 {
8431         .name           = "raid4",
8432         .level          = 4,
8433         .owner          = THIS_MODULE,
8434         .make_request   = raid5_make_request,
8435         .run            = raid5_run,
8436         .start          = raid5_start,
8437         .free           = raid5_free,
8438         .status         = raid5_status,
8439         .error_handler  = raid5_error,
8440         .hot_add_disk   = raid5_add_disk,
8441         .hot_remove_disk= raid5_remove_disk,
8442         .spare_active   = raid5_spare_active,
8443         .sync_request   = raid5_sync_request,
8444         .resize         = raid5_resize,
8445         .size           = raid5_size,
8446         .check_reshape  = raid5_check_reshape,
8447         .start_reshape  = raid5_start_reshape,
8448         .finish_reshape = raid5_finish_reshape,
8449         .quiesce        = raid5_quiesce,
8450         .takeover       = raid4_takeover,
8451         .congested      = raid5_congested,
8452         .change_consistency_policy = raid5_change_consistency_policy,
8453 };
8454
8455 static int __init raid5_init(void)
8456 {
8457         int ret;
8458
8459         raid5_wq = alloc_workqueue("raid5wq",
8460                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8461         if (!raid5_wq)
8462                 return -ENOMEM;
8463
8464         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8465                                       "md/raid5:prepare",
8466                                       raid456_cpu_up_prepare,
8467                                       raid456_cpu_dead);
8468         if (ret) {
8469                 destroy_workqueue(raid5_wq);
8470                 return ret;
8471         }
8472         register_md_personality(&raid6_personality);
8473         register_md_personality(&raid5_personality);
8474         register_md_personality(&raid4_personality);
8475         return 0;
8476 }
8477
8478 static void raid5_exit(void)
8479 {
8480         unregister_md_personality(&raid6_personality);
8481         unregister_md_personality(&raid5_personality);
8482         unregister_md_personality(&raid4_personality);
8483         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8484         destroy_workqueue(raid5_wq);
8485 }
8486
8487 module_init(raid5_init);
8488 module_exit(raid5_exit);
8489 MODULE_LICENSE("GPL");
8490 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8491 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8492 MODULE_ALIAS("md-raid5");
8493 MODULE_ALIAS("md-raid4");
8494 MODULE_ALIAS("md-level-5");
8495 MODULE_ALIAS("md-level-4");
8496 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8497 MODULE_ALIAS("md-raid6");
8498 MODULE_ALIAS("md-level-6");
8499
8500 /* This used to be two separate modules, they were: */
8501 MODULE_ALIAS("raid5");
8502 MODULE_ALIAS("raid6");