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