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 #include <trace/events/block.h>
  59
  60 #include "md.h"
  61 #include "raid5.h"
  62 #include "raid0.h"
  63 #include "bitmap.h"
  64
  65 #define cpu_to_group(cpu) cpu_to_node(cpu)
  66 #define ANY_GROUP NUMA_NO_NODE
  67
  68 static bool devices_handle_discard_safely = false;
  69 module_param(devices_handle_discard_safely, bool, 0644);
  70 MODULE_PARM_DESC(devices_handle_discard_safely,
  71                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  72 static struct workqueue_struct *raid5_wq;
  73
  74 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  75 {
  76         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  77         return &conf->stripe_hashtbl[hash];
  78 }
  79
  80 static inline int stripe_hash_locks_hash(sector_t sect)
  81 {
  82         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  83 }
  84
  85 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  86 {
  87         spin_lock_irq(conf->hash_locks + hash);
  88         spin_lock(&conf->device_lock);
  89 }
  90
  91 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  92 {
  93         spin_unlock(&conf->device_lock);
  94         spin_unlock_irq(conf->hash_locks + hash);
  95 }
  96
  97 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
  98 {
  99         int i;
 100         local_irq_disable();
 101         spin_lock(conf->hash_locks);
 102         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 103                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 104         spin_lock(&conf->device_lock);
 105 }
 106
 107 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 108 {
 109         int i;
 110         spin_unlock(&conf->device_lock);
 111         for (i = NR_STRIPE_HASH_LOCKS; i; i--)
 112                 spin_unlock(conf->hash_locks + i - 1);
 113         local_irq_enable();
 114 }
 115
 116 /* Find first data disk in a raid6 stripe */
 117 static inline int raid6_d0(struct stripe_head *sh)
 118 {
 119         if (sh->ddf_layout)
 120                 /* ddf always start from first device */
 121                 return 0;
 122         /* md starts just after Q block */
 123         if (sh->qd_idx == sh->disks - 1)
 124                 return 0;
 125         else
 126                 return sh->qd_idx + 1;
 127 }
 128 static inline int raid6_next_disk(int disk, int raid_disks)
 129 {
 130         disk++;
 131         return (disk < raid_disks) ? disk : 0;
 132 }
 133
 134 /* When walking through the disks in a raid5, starting at raid6_d0,
 135  * We need to map each disk to a 'slot', where the data disks are slot
 136  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 137  * is raid_disks-1.  This help does that mapping.
 138  */
 139 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 140                              int *count, int syndrome_disks)
 141 {
 142         int slot = *count;
 143
 144         if (sh->ddf_layout)
 145                 (*count)++;
 146         if (idx == sh->pd_idx)
 147                 return syndrome_disks;
 148         if (idx == sh->qd_idx)
 149                 return syndrome_disks + 1;
 150         if (!sh->ddf_layout)
 151                 (*count)++;
 152         return slot;
 153 }
 154
 155 static void return_io(struct bio_list *return_bi)
 156 {
 157         struct bio *bi;
 158         while ((bi = bio_list_pop(return_bi)) != NULL) {
 159                 bi->bi_iter.bi_size = 0;
 160                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 161                                          bi, 0);
 162                 bio_endio(bi);
 163         }
 164 }
 165
 166 static void print_raid5_conf (struct r5conf *conf);
 167
 168 static int stripe_operations_active(struct stripe_head *sh)
 169 {
 170         return sh->check_state || sh->reconstruct_state ||
 171                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 172                test_bit(STRIPE_COMPUTE_RUN, &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                 list_add_tail(&sh->lru, &group->handle_list);
 191                 group->stripes_cnt++;
 192                 sh->group = group;
 193         }
 194
 195         if (conf->worker_cnt_per_group == 0) {
 196                 md_wakeup_thread(conf->mddev->thread);
 197                 return;
 198         }
 199
 200         group = conf->worker_groups + cpu_to_group(sh->cpu);
 201
 202         group->workers[0].working = true;
 203         /* at least one worker should run to avoid race */
 204         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 205
 206         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 207         /* wakeup more workers */
 208         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 209                 if (group->workers[i].working == false) {
 210                         group->workers[i].working = true;
 211                         queue_work_on(sh->cpu, raid5_wq,
 212                                       &group->workers[i].work);
 213                         thread_cnt--;
 214                 }
 215         }
 216 }
 217
 218 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 219                               struct list_head *temp_inactive_list)
 220 {
 221         int i;
 222         int injournal = 0;      /* number of date pages with R5_InJournal */
 223
 224         BUG_ON(!list_empty(&sh->lru));
 225         BUG_ON(atomic_read(&conf->active_stripes)==0);
 226
 227         if (r5c_is_writeback(conf->log))
 228                 for (i = sh->disks; i--; )
 229                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 230                                 injournal++;
 231         /*
 232          * When quiesce in r5c write back, set STRIPE_HANDLE for stripes with
 233          * data in journal, so they are not released to cached lists
 234          */
 235         if (conf->quiesce && r5c_is_writeback(conf->log) &&
 236             !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0) {
 237                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 238                         r5c_make_stripe_write_out(sh);
 239                 set_bit(STRIPE_HANDLE, &sh->state);
 240         }
 241
 242         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 243                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 244                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 245                         list_add_tail(&sh->lru, &conf->delayed_list);
 246                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 247                            sh->bm_seq - conf->seq_write > 0)
 248                         list_add_tail(&sh->lru, &conf->bitmap_list);
 249                 else {
 250                         clear_bit(STRIPE_DELAYED, &sh->state);
 251                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 252                         if (conf->worker_cnt_per_group == 0) {
 253                                 list_add_tail(&sh->lru, &conf->handle_list);
 254                         } else {
 255                                 raid5_wakeup_stripe_thread(sh);
 256                                 return;
 257                         }
 258                 }
 259                 md_wakeup_thread(conf->mddev->thread);
 260         } else {
 261                 BUG_ON(stripe_operations_active(sh));
 262                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 263                         if (atomic_dec_return(&conf->preread_active_stripes)
 264                             < IO_THRESHOLD)
 265                                 md_wakeup_thread(conf->mddev->thread);
 266                 atomic_dec(&conf->active_stripes);
 267                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 268                         if (!r5c_is_writeback(conf->log))
 269                                 list_add_tail(&sh->lru, temp_inactive_list);
 270                         else {
 271                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 272                                 if (injournal == 0)
 273                                         list_add_tail(&sh->lru, temp_inactive_list);
 274                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 275                                         /* full stripe */
 276                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 277                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 278                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 279                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 280                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 281                                         r5c_check_cached_full_stripe(conf);
 282                                 } else {
 283                                         /* partial stripe */
 284                                         if (!test_and_set_bit(STRIPE_R5C_PARTIAL_STRIPE,
 285                                                               &sh->state))
 286                                                 atomic_inc(&conf->r5c_cached_partial_stripes);
 287                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 288                                 }
 289                         }
 290                 }
 291         }
 292 }
 293
 294 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 295                              struct list_head *temp_inactive_list)
 296 {
 297         if (atomic_dec_and_test(&sh->count))
 298                 do_release_stripe(conf, sh, temp_inactive_list);
 299 }
 300
 301 /*
 302  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 303  *
 304  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 305  * given time. Adding stripes only takes device lock, while deleting stripes
 306  * only takes hash lock.
 307  */
 308 static void release_inactive_stripe_list(struct r5conf *conf,
 309                                          struct list_head *temp_inactive_list,
 310                                          int hash)
 311 {
 312         int size;
 313         bool do_wakeup = false;
 314         unsigned long flags;
 315
 316         if (hash == NR_STRIPE_HASH_LOCKS) {
 317                 size = NR_STRIPE_HASH_LOCKS;
 318                 hash = NR_STRIPE_HASH_LOCKS - 1;
 319         } else
 320                 size = 1;
 321         while (size) {
 322                 struct list_head *list = &temp_inactive_list[size - 1];
 323
 324                 /*
 325                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 326                  * remove stripes from the list
 327                  */
 328                 if (!list_empty_careful(list)) {
 329                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 330                         if (list_empty(conf->inactive_list + hash) &&
 331                             !list_empty(list))
 332                                 atomic_dec(&conf->empty_inactive_list_nr);
 333                         list_splice_tail_init(list, conf->inactive_list + hash);
 334                         do_wakeup = true;
 335                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 336                 }
 337                 size--;
 338                 hash--;
 339         }
 340
 341         if (do_wakeup) {
 342                 wake_up(&conf->wait_for_stripe);
 343                 if (atomic_read(&conf->active_stripes) == 0)
 344                         wake_up(&conf->wait_for_quiescent);
 345                 if (conf->retry_read_aligned)
 346                         md_wakeup_thread(conf->mddev->thread);
 347         }
 348 }
 349
 350 /* should hold conf->device_lock already */
 351 static int release_stripe_list(struct r5conf *conf,
 352                                struct list_head *temp_inactive_list)
 353 {
 354         struct stripe_head *sh;
 355         int count = 0;
 356         struct llist_node *head;
 357
 358         head = llist_del_all(&conf->released_stripes);
 359         head = llist_reverse_order(head);
 360         while (head) {
 361                 int hash;
 362
 363                 sh = llist_entry(head, struct stripe_head, release_list);
 364                 head = llist_next(head);
 365                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 366                 smp_mb();
 367                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 368                 /*
 369                  * Don't worry the bit is set here, because if the bit is set
 370                  * again, the count is always > 1. This is true for
 371                  * STRIPE_ON_UNPLUG_LIST bit too.
 372                  */
 373                 hash = sh->hash_lock_index;
 374                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 375                 count++;
 376         }
 377
 378         return count;
 379 }
 380
 381 void raid5_release_stripe(struct stripe_head *sh)
 382 {
 383         struct r5conf *conf = sh->raid_conf;
 384         unsigned long flags;
 385         struct list_head list;
 386         int hash;
 387         bool wakeup;
 388
 389         /* Avoid release_list until the last reference.
 390          */
 391         if (atomic_add_unless(&sh->count, -1, 1))
 392                 return;
 393
 394         if (unlikely(!conf->mddev->thread) ||
 395                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 396                 goto slow_path;
 397         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 398         if (wakeup)
 399                 md_wakeup_thread(conf->mddev->thread);
 400         return;
 401 slow_path:
 402         local_irq_save(flags);
 403         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 404         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 405                 INIT_LIST_HEAD(&list);
 406                 hash = sh->hash_lock_index;
 407                 do_release_stripe(conf, sh, &list);
 408                 spin_unlock(&conf->device_lock);
 409                 release_inactive_stripe_list(conf, &list, hash);
 410         }
 411         local_irq_restore(flags);
 412 }
 413
 414 static inline void remove_hash(struct stripe_head *sh)
 415 {
 416         pr_debug("remove_hash(), stripe %llu\n",
 417                 (unsigned long long)sh->sector);
 418
 419         hlist_del_init(&sh->hash);
 420 }
 421
 422 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 423 {
 424         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 425
 426         pr_debug("insert_hash(), stripe %llu\n",
 427                 (unsigned long long)sh->sector);
 428
 429         hlist_add_head(&sh->hash, hp);
 430 }
 431
 432 /* find an idle stripe, make sure it is unhashed, and return it. */
 433 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 434 {
 435         struct stripe_head *sh = NULL;
 436         struct list_head *first;
 437
 438         if (list_empty(conf->inactive_list + hash))
 439                 goto out;
 440         first = (conf->inactive_list + hash)->next;
 441         sh = list_entry(first, struct stripe_head, lru);
 442         list_del_init(first);
 443         remove_hash(sh);
 444         atomic_inc(&conf->active_stripes);
 445         BUG_ON(hash != sh->hash_lock_index);
 446         if (list_empty(conf->inactive_list + hash))
 447                 atomic_inc(&conf->empty_inactive_list_nr);
 448 out:
 449         return sh;
 450 }
 451
 452 static void shrink_buffers(struct stripe_head *sh)
 453 {
 454         struct page *p;
 455         int i;
 456         int num = sh->raid_conf->pool_size;
 457
 458         for (i = 0; i < num ; i++) {
 459                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 460                 p = sh->dev[i].page;
 461                 if (!p)
 462                         continue;
 463                 sh->dev[i].page = NULL;
 464                 put_page(p);
 465         }
 466 }
 467
 468 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 469 {
 470         int i;
 471         int num = sh->raid_conf->pool_size;
 472
 473         for (i = 0; i < num; i++) {
 474                 struct page *page;
 475
 476                 if (!(page = alloc_page(gfp))) {
 477                         return 1;
 478                 }
 479                 sh->dev[i].page = page;
 480                 sh->dev[i].orig_page = page;
 481         }
 482         return 0;
 483 }
 484
 485 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 486 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 487                             struct stripe_head *sh);
 488
 489 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 490 {
 491         struct r5conf *conf = sh->raid_conf;
 492         int i, seq;
 493
 494         BUG_ON(atomic_read(&sh->count) != 0);
 495         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 496         BUG_ON(stripe_operations_active(sh));
 497         BUG_ON(sh->batch_head);
 498
 499         pr_debug("init_stripe called, stripe %llu\n",
 500                 (unsigned long long)sector);
 501 retry:
 502         seq = read_seqcount_begin(&conf->gen_lock);
 503         sh->generation = conf->generation - previous;
 504         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 505         sh->sector = sector;
 506         stripe_set_idx(sector, conf, previous, sh);
 507         sh->state = 0;
 508
 509         for (i = sh->disks; i--; ) {
 510                 struct r5dev *dev = &sh->dev[i];
 511
 512                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 513                     test_bit(R5_LOCKED, &dev->flags)) {
 514                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 515                                (unsigned long long)sh->sector, i, dev->toread,
 516                                dev->read, dev->towrite, dev->written,
 517                                test_bit(R5_LOCKED, &dev->flags));
 518                         WARN_ON(1);
 519                 }
 520                 dev->flags = 0;
 521                 raid5_build_block(sh, i, previous);
 522         }
 523         if (read_seqcount_retry(&conf->gen_lock, seq))
 524                 goto retry;
 525         sh->overwrite_disks = 0;
 526         insert_hash(conf, sh);
 527         sh->cpu = smp_processor_id();
 528         set_bit(STRIPE_BATCH_READY, &sh->state);
 529 }
 530
 531 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 532                                          short generation)
 533 {
 534         struct stripe_head *sh;
 535
 536         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 537         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 538                 if (sh->sector == sector && sh->generation == generation)
 539                         return sh;
 540         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 541         return NULL;
 542 }
 543
 544 /*
 545  * Need to check if array has failed when deciding whether to:
 546  *  - start an array
 547  *  - remove non-faulty devices
 548  *  - add a spare
 549  *  - allow a reshape
 550  * This determination is simple when no reshape is happening.
 551  * However if there is a reshape, we need to carefully check
 552  * both the before and after sections.
 553  * This is because some failed devices may only affect one
 554  * of the two sections, and some non-in_sync devices may
 555  * be insync in the section most affected by failed devices.
 556  */
 557 static int calc_degraded(struct r5conf *conf)
 558 {
 559         int degraded, degraded2;
 560         int i;
 561
 562         rcu_read_lock();
 563         degraded = 0;
 564         for (i = 0; i < conf->previous_raid_disks; i++) {
 565                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 566                 if (rdev && test_bit(Faulty, &rdev->flags))
 567                         rdev = rcu_dereference(conf->disks[i].replacement);
 568                 if (!rdev || test_bit(Faulty, &rdev->flags))
 569                         degraded++;
 570                 else if (test_bit(In_sync, &rdev->flags))
 571                         ;
 572                 else
 573                         /* not in-sync or faulty.
 574                          * If the reshape increases the number of devices,
 575                          * this is being recovered by the reshape, so
 576                          * this 'previous' section is not in_sync.
 577                          * If the number of devices is being reduced however,
 578                          * the device can only be part of the array if
 579                          * we are reverting a reshape, so this section will
 580                          * be in-sync.
 581                          */
 582                         if (conf->raid_disks >= conf->previous_raid_disks)
 583                                 degraded++;
 584         }
 585         rcu_read_unlock();
 586         if (conf->raid_disks == conf->previous_raid_disks)
 587                 return degraded;
 588         rcu_read_lock();
 589         degraded2 = 0;
 590         for (i = 0; i < conf->raid_disks; i++) {
 591                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 592                 if (rdev && test_bit(Faulty, &rdev->flags))
 593                         rdev = rcu_dereference(conf->disks[i].replacement);
 594                 if (!rdev || test_bit(Faulty, &rdev->flags))
 595                         degraded2++;
 596                 else if (test_bit(In_sync, &rdev->flags))
 597                         ;
 598                 else
 599                         /* not in-sync or faulty.
 600                          * If reshape increases the number of devices, this
 601                          * section has already been recovered, else it
 602                          * almost certainly hasn't.
 603                          */
 604                         if (conf->raid_disks <= conf->previous_raid_disks)
 605                                 degraded2++;
 606         }
 607         rcu_read_unlock();
 608         if (degraded2 > degraded)
 609                 return degraded2;
 610         return degraded;
 611 }
 612
 613 static int has_failed(struct r5conf *conf)
 614 {
 615         int degraded;
 616
 617         if (conf->mddev->reshape_position == MaxSector)
 618                 return conf->mddev->degraded > conf->max_degraded;
 619
 620         degraded = calc_degraded(conf);
 621         if (degraded > conf->max_degraded)
 622                 return 1;
 623         return 0;
 624 }
 625
 626 struct stripe_head *
 627 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 628                         int previous, int noblock, int noquiesce)
 629 {
 630         struct stripe_head *sh;
 631         int hash = stripe_hash_locks_hash(sector);
 632         int inc_empty_inactive_list_flag;
 633
 634         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 635
 636         spin_lock_irq(conf->hash_locks + hash);
 637
 638         do {
 639                 wait_event_lock_irq(conf->wait_for_quiescent,
 640                                     conf->quiesce == 0 || noquiesce,
 641                                     *(conf->hash_locks + hash));
 642                 sh = __find_stripe(conf, sector, conf->generation - previous);
 643                 if (!sh) {
 644                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 645                                 sh = get_free_stripe(conf, hash);
 646                                 if (!sh && !test_bit(R5_DID_ALLOC,
 647                                                      &conf->cache_state))
 648                                         set_bit(R5_ALLOC_MORE,
 649                                                 &conf->cache_state);
 650                         }
 651                         if (noblock && sh == NULL)
 652                                 break;
 653
 654                         r5c_check_stripe_cache_usage(conf);
 655                         if (!sh) {
 656                                 set_bit(R5_INACTIVE_BLOCKED,
 657                                         &conf->cache_state);
 658                                 r5l_wake_reclaim(conf->log, 0);
 659                                 wait_event_lock_irq(
 660                                         conf->wait_for_stripe,
 661                                         !list_empty(conf->inactive_list + hash) &&
 662                                         (atomic_read(&conf->active_stripes)
 663                                          < (conf->max_nr_stripes * 3 / 4)
 664                                          || !test_bit(R5_INACTIVE_BLOCKED,
 665                                                       &conf->cache_state)),
 666                                         *(conf->hash_locks + hash));
 667                                 clear_bit(R5_INACTIVE_BLOCKED,
 668                                           &conf->cache_state);
 669                         } else {
 670                                 init_stripe(sh, sector, previous);
 671                                 atomic_inc(&sh->count);
 672                         }
 673                 } else if (!atomic_inc_not_zero(&sh->count)) {
 674                         spin_lock(&conf->device_lock);
 675                         if (!atomic_read(&sh->count)) {
 676                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 677                                         atomic_inc(&conf->active_stripes);
 678                                 BUG_ON(list_empty(&sh->lru) &&
 679                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 680                                 inc_empty_inactive_list_flag = 0;
 681                                 if (!list_empty(conf->inactive_list + hash))
 682                                         inc_empty_inactive_list_flag = 1;
 683                                 list_del_init(&sh->lru);
 684                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 685                                         atomic_inc(&conf->empty_inactive_list_nr);
 686                                 if (sh->group) {
 687                                         sh->group->stripes_cnt--;
 688                                         sh->group = NULL;
 689                                 }
 690                         }
 691                         atomic_inc(&sh->count);
 692                         spin_unlock(&conf->device_lock);
 693                 }
 694         } while (sh == NULL);
 695
 696         spin_unlock_irq(conf->hash_locks + hash);
 697         return sh;
 698 }
 699
 700 static bool is_full_stripe_write(struct stripe_head *sh)
 701 {
 702         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 703         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 704 }
 705
 706 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 707 {
 708         local_irq_disable();
 709         if (sh1 > sh2) {
 710                 spin_lock(&sh2->stripe_lock);
 711                 spin_lock_nested(&sh1->stripe_lock, 1);
 712         } else {
 713                 spin_lock(&sh1->stripe_lock);
 714                 spin_lock_nested(&sh2->stripe_lock, 1);
 715         }
 716 }
 717
 718 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 719 {
 720         spin_unlock(&sh1->stripe_lock);
 721         spin_unlock(&sh2->stripe_lock);
 722         local_irq_enable();
 723 }
 724
 725 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 726 static bool stripe_can_batch(struct stripe_head *sh)
 727 {
 728         struct r5conf *conf = sh->raid_conf;
 729
 730         if (conf->log)
 731                 return false;
 732         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 733                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 734                 is_full_stripe_write(sh);
 735 }
 736
 737 /* we only do back search */
 738 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 739 {
 740         struct stripe_head *head;
 741         sector_t head_sector, tmp_sec;
 742         int hash;
 743         int dd_idx;
 744         int inc_empty_inactive_list_flag;
 745
 746         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 747         tmp_sec = sh->sector;
 748         if (!sector_div(tmp_sec, conf->chunk_sectors))
 749                 return;
 750         head_sector = sh->sector - STRIPE_SECTORS;
 751
 752         hash = stripe_hash_locks_hash(head_sector);
 753         spin_lock_irq(conf->hash_locks + hash);
 754         head = __find_stripe(conf, head_sector, conf->generation);
 755         if (head && !atomic_inc_not_zero(&head->count)) {
 756                 spin_lock(&conf->device_lock);
 757                 if (!atomic_read(&head->count)) {
 758                         if (!test_bit(STRIPE_HANDLE, &head->state))
 759                                 atomic_inc(&conf->active_stripes);
 760                         BUG_ON(list_empty(&head->lru) &&
 761                                !test_bit(STRIPE_EXPANDING, &head->state));
 762                         inc_empty_inactive_list_flag = 0;
 763                         if (!list_empty(conf->inactive_list + hash))
 764                                 inc_empty_inactive_list_flag = 1;
 765                         list_del_init(&head->lru);
 766                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 767                                 atomic_inc(&conf->empty_inactive_list_nr);
 768                         if (head->group) {
 769                                 head->group->stripes_cnt--;
 770                                 head->group = NULL;
 771                         }
 772                 }
 773                 atomic_inc(&head->count);
 774                 spin_unlock(&conf->device_lock);
 775         }
 776         spin_unlock_irq(conf->hash_locks + hash);
 777
 778         if (!head)
 779                 return;
 780         if (!stripe_can_batch(head))
 781                 goto out;
 782
 783         lock_two_stripes(head, sh);
 784         /* clear_batch_ready clear the flag */
 785         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 786                 goto unlock_out;
 787
 788         if (sh->batch_head)
 789                 goto unlock_out;
 790
 791         dd_idx = 0;
 792         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 793                 dd_idx++;
 794         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 795             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 796                 goto unlock_out;
 797
 798         if (head->batch_head) {
 799                 spin_lock(&head->batch_head->batch_lock);
 800                 /* This batch list is already running */
 801                 if (!stripe_can_batch(head)) {
 802                         spin_unlock(&head->batch_head->batch_lock);
 803                         goto unlock_out;
 804                 }
 805
 806                 /*
 807                  * at this point, head's BATCH_READY could be cleared, but we
 808                  * can still add the stripe to batch list
 809                  */
 810                 list_add(&sh->batch_list, &head->batch_list);
 811                 spin_unlock(&head->batch_head->batch_lock);
 812
 813                 sh->batch_head = head->batch_head;
 814         } else {
 815                 head->batch_head = head;
 816                 sh->batch_head = head->batch_head;
 817                 spin_lock(&head->batch_lock);
 818                 list_add_tail(&sh->batch_list, &head->batch_list);
 819                 spin_unlock(&head->batch_lock);
 820         }
 821
 822         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 823                 if (atomic_dec_return(&conf->preread_active_stripes)
 824                     < IO_THRESHOLD)
 825                         md_wakeup_thread(conf->mddev->thread);
 826
 827         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 828                 int seq = sh->bm_seq;
 829                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 830                     sh->batch_head->bm_seq > seq)
 831                         seq = sh->batch_head->bm_seq;
 832                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 833                 sh->batch_head->bm_seq = seq;
 834         }
 835
 836         atomic_inc(&sh->count);
 837 unlock_out:
 838         unlock_two_stripes(head, sh);
 839 out:
 840         raid5_release_stripe(head);
 841 }
 842
 843 /* Determine if 'data_offset' or 'new_data_offset' should be used
 844  * in this stripe_head.
 845  */
 846 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 847 {
 848         sector_t progress = conf->reshape_progress;
 849         /* Need a memory barrier to make sure we see the value
 850          * of conf->generation, or ->data_offset that was set before
 851          * reshape_progress was updated.
 852          */
 853         smp_rmb();
 854         if (progress == MaxSector)
 855                 return 0;
 856         if (sh->generation == conf->generation - 1)
 857                 return 0;
 858         /* We are in a reshape, and this is a new-generation stripe,
 859          * so use new_data_offset.
 860          */
 861         return 1;
 862 }
 863
 864 static void
 865 raid5_end_read_request(struct bio *bi);
 866 static void
 867 raid5_end_write_request(struct bio *bi);
 868
 869 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 870 {
 871         struct r5conf *conf = sh->raid_conf;
 872         int i, disks = sh->disks;
 873         struct stripe_head *head_sh = sh;
 874
 875         might_sleep();
 876
 877         if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
 878                 /* writing out phase */
 879                 if (s->waiting_extra_page)
 880                         return;
 881                 if (r5l_write_stripe(conf->log, sh) == 0)
 882                         return;
 883         } else {  /* caching phase */
 884                 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state)) {
 885                         r5c_cache_data(conf->log, sh, s);
 886                         return;
 887                 }
 888         }
 889
 890         for (i = disks; i--; ) {
 891                 int op, op_flags = 0;
 892                 int replace_only = 0;
 893                 struct bio *bi, *rbi;
 894                 struct md_rdev *rdev, *rrdev = NULL;
 895
 896                 sh = head_sh;
 897                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 898                         op = REQ_OP_WRITE;
 899                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 900                                 op_flags = REQ_FUA;
 901                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 902                                 op = REQ_OP_DISCARD;
 903                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 904                         op = REQ_OP_READ;
 905                 else if (test_and_clear_bit(R5_WantReplace,
 906                                             &sh->dev[i].flags)) {
 907                         op = REQ_OP_WRITE;
 908                         replace_only = 1;
 909                 } else
 910                         continue;
 911                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 912                         op_flags |= REQ_SYNC;
 913
 914 again:
 915                 bi = &sh->dev[i].req;
 916                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 917
 918                 rcu_read_lock();
 919                 rrdev = rcu_dereference(conf->disks[i].replacement);
 920                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 921                 rdev = rcu_dereference(conf->disks[i].rdev);
 922                 if (!rdev) {
 923                         rdev = rrdev;
 924                         rrdev = NULL;
 925                 }
 926                 if (op_is_write(op)) {
 927                         if (replace_only)
 928                                 rdev = NULL;
 929                         if (rdev == rrdev)
 930                                 /* We raced and saw duplicates */
 931                                 rrdev = NULL;
 932                 } else {
 933                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
 934                                 rdev = rrdev;
 935                         rrdev = NULL;
 936                 }
 937
 938                 if (rdev && test_bit(Faulty, &rdev->flags))
 939                         rdev = NULL;
 940                 if (rdev)
 941                         atomic_inc(&rdev->nr_pending);
 942                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 943                         rrdev = NULL;
 944                 if (rrdev)
 945                         atomic_inc(&rrdev->nr_pending);
 946                 rcu_read_unlock();
 947
 948                 /* We have already checked bad blocks for reads.  Now
 949                  * need to check for writes.  We never accept write errors
 950                  * on the replacement, so we don't to check rrdev.
 951                  */
 952                 while (op_is_write(op) && rdev &&
 953                        test_bit(WriteErrorSeen, &rdev->flags)) {
 954                         sector_t first_bad;
 955                         int bad_sectors;
 956                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 957                                               &first_bad, &bad_sectors);
 958                         if (!bad)
 959                                 break;
 960
 961                         if (bad < 0) {
 962                                 set_bit(BlockedBadBlocks, &rdev->flags);
 963                                 if (!conf->mddev->external &&
 964                                     conf->mddev->sb_flags) {
 965                                         /* It is very unlikely, but we might
 966                                          * still need to write out the
 967                                          * bad block log - better give it
 968                                          * a chance*/
 969                                         md_check_recovery(conf->mddev);
 970                                 }
 971                                 /*
 972                                  * Because md_wait_for_blocked_rdev
 973                                  * will dec nr_pending, we must
 974                                  * increment it first.
 975                                  */
 976                                 atomic_inc(&rdev->nr_pending);
 977                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 978                         } else {
 979                                 /* Acknowledged bad block - skip the write */
 980                                 rdev_dec_pending(rdev, conf->mddev);
 981                                 rdev = NULL;
 982                         }
 983                 }
 984
 985                 if (rdev) {
 986                         if (s->syncing || s->expanding || s->expanded
 987                             || s->replacing)
 988                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 989
 990                         set_bit(STRIPE_IO_STARTED, &sh->state);
 991
 992                         bi->bi_bdev = rdev->bdev;
 993                         bio_set_op_attrs(bi, op, op_flags);
 994                         bi->bi_end_io = op_is_write(op)
 995                                 ? raid5_end_write_request
 996                                 : raid5_end_read_request;
 997                         bi->bi_private = sh;
 998
 999                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1000                                 __func__, (unsigned long long)sh->sector,
1001                                 bi->bi_opf, i);
1002                         atomic_inc(&sh->count);
1003                         if (sh != head_sh)
1004                                 atomic_inc(&head_sh->count);
1005                         if (use_new_offset(conf, sh))
1006                                 bi->bi_iter.bi_sector = (sh->sector
1007                                                  + rdev->new_data_offset);
1008                         else
1009                                 bi->bi_iter.bi_sector = (sh->sector
1010                                                  + rdev->data_offset);
1011                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1012                                 bi->bi_opf |= REQ_NOMERGE;
1013
1014                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1015                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1016                         sh->dev[i].vec.bv_page = sh->dev[i].page;
1017                         bi->bi_vcnt = 1;
1018                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1019                         bi->bi_io_vec[0].bv_offset = 0;
1020                         bi->bi_iter.bi_size = STRIPE_SIZE;
1021                         /*
1022                          * If this is discard request, set bi_vcnt 0. We don't
1023                          * want to confuse SCSI because SCSI will replace payload
1024                          */
1025                         if (op == REQ_OP_DISCARD)
1026                                 bi->bi_vcnt = 0;
1027                         if (rrdev)
1028                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1029
1030                         if (conf->mddev->gendisk)
1031                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1032                                                       bi, disk_devt(conf->mddev->gendisk),
1033                                                       sh->dev[i].sector);
1034                         generic_make_request(bi);
1035                 }
1036                 if (rrdev) {
1037                         if (s->syncing || s->expanding || s->expanded
1038                             || s->replacing)
1039                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1040
1041                         set_bit(STRIPE_IO_STARTED, &sh->state);
1042
1043                         rbi->bi_bdev = rrdev->bdev;
1044                         bio_set_op_attrs(rbi, op, op_flags);
1045                         BUG_ON(!op_is_write(op));
1046                         rbi->bi_end_io = raid5_end_write_request;
1047                         rbi->bi_private = sh;
1048
1049                         pr_debug("%s: for %llu schedule op %d on "
1050                                  "replacement disc %d\n",
1051                                 __func__, (unsigned long long)sh->sector,
1052                                 rbi->bi_opf, i);
1053                         atomic_inc(&sh->count);
1054                         if (sh != head_sh)
1055                                 atomic_inc(&head_sh->count);
1056                         if (use_new_offset(conf, sh))
1057                                 rbi->bi_iter.bi_sector = (sh->sector
1058                                                   + rrdev->new_data_offset);
1059                         else
1060                                 rbi->bi_iter.bi_sector = (sh->sector
1061                                                   + rrdev->data_offset);
1062                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1063                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1064                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1065                         rbi->bi_vcnt = 1;
1066                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1067                         rbi->bi_io_vec[0].bv_offset = 0;
1068                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1069                         /*
1070                          * If this is discard request, set bi_vcnt 0. We don't
1071                          * want to confuse SCSI because SCSI will replace payload
1072                          */
1073                         if (op == REQ_OP_DISCARD)
1074                                 rbi->bi_vcnt = 0;
1075                         if (conf->mddev->gendisk)
1076                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1077                                                       rbi, disk_devt(conf->mddev->gendisk),
1078                                                       sh->dev[i].sector);
1079                         generic_make_request(rbi);
1080                 }
1081                 if (!rdev && !rrdev) {
1082                         if (op_is_write(op))
1083                                 set_bit(STRIPE_DEGRADED, &sh->state);
1084                         pr_debug("skip op %d on disc %d for sector %llu\n",
1085                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1086                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1087                         set_bit(STRIPE_HANDLE, &sh->state);
1088                 }
1089
1090                 if (!head_sh->batch_head)
1091                         continue;
1092                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1093                                       batch_list);
1094                 if (sh != head_sh)
1095                         goto again;
1096         }
1097 }
1098
1099 static struct dma_async_tx_descriptor *
1100 async_copy_data(int frombio, struct bio *bio, struct page **page,
1101         sector_t sector, struct dma_async_tx_descriptor *tx,
1102         struct stripe_head *sh, int no_skipcopy)
1103 {
1104         struct bio_vec bvl;
1105         struct bvec_iter iter;
1106         struct page *bio_page;
1107         int page_offset;
1108         struct async_submit_ctl submit;
1109         enum async_tx_flags flags = 0;
1110
1111         if (bio->bi_iter.bi_sector >= sector)
1112                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1113         else
1114                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1115
1116         if (frombio)
1117                 flags |= ASYNC_TX_FENCE;
1118         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1119
1120         bio_for_each_segment(bvl, bio, iter) {
1121                 int len = bvl.bv_len;
1122                 int clen;
1123                 int b_offset = 0;
1124
1125                 if (page_offset < 0) {
1126                         b_offset = -page_offset;
1127                         page_offset += b_offset;
1128                         len -= b_offset;
1129                 }
1130
1131                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1132                         clen = STRIPE_SIZE - page_offset;
1133                 else
1134                         clen = len;
1135
1136                 if (clen > 0) {
1137                         b_offset += bvl.bv_offset;
1138                         bio_page = bvl.bv_page;
1139                         if (frombio) {
1140                                 if (sh->raid_conf->skip_copy &&
1141                                     b_offset == 0 && page_offset == 0 &&
1142                                     clen == STRIPE_SIZE &&
1143                                     !no_skipcopy)
1144                                         *page = bio_page;
1145                                 else
1146                                         tx = async_memcpy(*page, bio_page, page_offset,
1147                                                   b_offset, clen, &submit);
1148                         } else
1149                                 tx = async_memcpy(bio_page, *page, b_offset,
1150                                                   page_offset, clen, &submit);
1151                 }
1152                 /* chain the operations */
1153                 submit.depend_tx = tx;
1154
1155                 if (clen < len) /* hit end of page */
1156                         break;
1157                 page_offset +=  len;
1158         }
1159
1160         return tx;
1161 }
1162
1163 static void ops_complete_biofill(void *stripe_head_ref)
1164 {
1165         struct stripe_head *sh = stripe_head_ref;
1166         struct bio_list return_bi = BIO_EMPTY_LIST;
1167         int i;
1168
1169         pr_debug("%s: stripe %llu\n", __func__,
1170                 (unsigned long long)sh->sector);
1171
1172         /* clear completed biofills */
1173         for (i = sh->disks; i--; ) {
1174                 struct r5dev *dev = &sh->dev[i];
1175
1176                 /* acknowledge completion of a biofill operation */
1177                 /* and check if we need to reply to a read request,
1178                  * new R5_Wantfill requests are held off until
1179                  * !STRIPE_BIOFILL_RUN
1180                  */
1181                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1182                         struct bio *rbi, *rbi2;
1183
1184                         BUG_ON(!dev->read);
1185                         rbi = dev->read;
1186                         dev->read = NULL;
1187                         while (rbi && rbi->bi_iter.bi_sector <
1188                                 dev->sector + STRIPE_SECTORS) {
1189                                 rbi2 = r5_next_bio(rbi, dev->sector);
1190                                 if (!raid5_dec_bi_active_stripes(rbi))
1191                                         bio_list_add(&return_bi, rbi);
1192                                 rbi = rbi2;
1193                         }
1194                 }
1195         }
1196         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1197
1198         return_io(&return_bi);
1199
1200         set_bit(STRIPE_HANDLE, &sh->state);
1201         raid5_release_stripe(sh);
1202 }
1203
1204 static void ops_run_biofill(struct stripe_head *sh)
1205 {
1206         struct dma_async_tx_descriptor *tx = NULL;
1207         struct async_submit_ctl submit;
1208         int i;
1209
1210         BUG_ON(sh->batch_head);
1211         pr_debug("%s: stripe %llu\n", __func__,
1212                 (unsigned long long)sh->sector);
1213
1214         for (i = sh->disks; i--; ) {
1215                 struct r5dev *dev = &sh->dev[i];
1216                 if (test_bit(R5_Wantfill, &dev->flags)) {
1217                         struct bio *rbi;
1218                         spin_lock_irq(&sh->stripe_lock);
1219                         dev->read = rbi = dev->toread;
1220                         dev->toread = NULL;
1221                         spin_unlock_irq(&sh->stripe_lock);
1222                         while (rbi && rbi->bi_iter.bi_sector <
1223                                 dev->sector + STRIPE_SECTORS) {
1224                                 tx = async_copy_data(0, rbi, &dev->page,
1225                                                      dev->sector, tx, sh, 0);
1226                                 rbi = r5_next_bio(rbi, dev->sector);
1227                         }
1228                 }
1229         }
1230
1231         atomic_inc(&sh->count);
1232         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1233         async_trigger_callback(&submit);
1234 }
1235
1236 static void mark_target_uptodate(struct stripe_head *sh, int target)
1237 {
1238         struct r5dev *tgt;
1239
1240         if (target < 0)
1241                 return;
1242
1243         tgt = &sh->dev[target];
1244         set_bit(R5_UPTODATE, &tgt->flags);
1245         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1246         clear_bit(R5_Wantcompute, &tgt->flags);
1247 }
1248
1249 static void ops_complete_compute(void *stripe_head_ref)
1250 {
1251         struct stripe_head *sh = stripe_head_ref;
1252
1253         pr_debug("%s: stripe %llu\n", __func__,
1254                 (unsigned long long)sh->sector);
1255
1256         /* mark the computed target(s) as uptodate */
1257         mark_target_uptodate(sh, sh->ops.target);
1258         mark_target_uptodate(sh, sh->ops.target2);
1259
1260         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1261         if (sh->check_state == check_state_compute_run)
1262                 sh->check_state = check_state_compute_result;
1263         set_bit(STRIPE_HANDLE, &sh->state);
1264         raid5_release_stripe(sh);
1265 }
1266
1267 /* return a pointer to the address conversion region of the scribble buffer */
1268 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1269                                  struct raid5_percpu *percpu, int i)
1270 {
1271         void *addr;
1272
1273         addr = flex_array_get(percpu->scribble, i);
1274         return addr + sizeof(struct page *) * (sh->disks + 2);
1275 }
1276
1277 /* return a pointer to the address conversion region of the scribble buffer */
1278 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1279 {
1280         void *addr;
1281
1282         addr = flex_array_get(percpu->scribble, i);
1283         return addr;
1284 }
1285
1286 static struct dma_async_tx_descriptor *
1287 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1288 {
1289         int disks = sh->disks;
1290         struct page **xor_srcs = to_addr_page(percpu, 0);
1291         int target = sh->ops.target;
1292         struct r5dev *tgt = &sh->dev[target];
1293         struct page *xor_dest = tgt->page;
1294         int count = 0;
1295         struct dma_async_tx_descriptor *tx;
1296         struct async_submit_ctl submit;
1297         int i;
1298
1299         BUG_ON(sh->batch_head);
1300
1301         pr_debug("%s: stripe %llu block: %d\n",
1302                 __func__, (unsigned long long)sh->sector, target);
1303         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1304
1305         for (i = disks; i--; )
1306                 if (i != target)
1307                         xor_srcs[count++] = sh->dev[i].page;
1308
1309         atomic_inc(&sh->count);
1310
1311         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1312                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1313         if (unlikely(count == 1))
1314                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1315         else
1316                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1317
1318         return tx;
1319 }
1320
1321 /* set_syndrome_sources - populate source buffers for gen_syndrome
1322  * @srcs - (struct page *) array of size sh->disks
1323  * @sh - stripe_head to parse
1324  *
1325  * Populates srcs in proper layout order for the stripe and returns the
1326  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1327  * destination buffer is recorded in srcs[count] and the Q destination
1328  * is recorded in srcs[count+1]].
1329  */
1330 static int set_syndrome_sources(struct page **srcs,
1331                                 struct stripe_head *sh,
1332                                 int srctype)
1333 {
1334         int disks = sh->disks;
1335         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1336         int d0_idx = raid6_d0(sh);
1337         int count;
1338         int i;
1339
1340         for (i = 0; i < disks; i++)
1341                 srcs[i] = NULL;
1342
1343         count = 0;
1344         i = d0_idx;
1345         do {
1346                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1347                 struct r5dev *dev = &sh->dev[i];
1348
1349                 if (i == sh->qd_idx || i == sh->pd_idx ||
1350                     (srctype == SYNDROME_SRC_ALL) ||
1351                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1352                      (test_bit(R5_Wantdrain, &dev->flags) ||
1353                       test_bit(R5_InJournal, &dev->flags))) ||
1354                     (srctype == SYNDROME_SRC_WRITTEN &&
1355                      dev->written)) {
1356                         if (test_bit(R5_InJournal, &dev->flags))
1357                                 srcs[slot] = sh->dev[i].orig_page;
1358                         else
1359                                 srcs[slot] = sh->dev[i].page;
1360                 }
1361                 i = raid6_next_disk(i, disks);
1362         } while (i != d0_idx);
1363
1364         return syndrome_disks;
1365 }
1366
1367 static struct dma_async_tx_descriptor *
1368 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1369 {
1370         int disks = sh->disks;
1371         struct page **blocks = to_addr_page(percpu, 0);
1372         int target;
1373         int qd_idx = sh->qd_idx;
1374         struct dma_async_tx_descriptor *tx;
1375         struct async_submit_ctl submit;
1376         struct r5dev *tgt;
1377         struct page *dest;
1378         int i;
1379         int count;
1380
1381         BUG_ON(sh->batch_head);
1382         if (sh->ops.target < 0)
1383                 target = sh->ops.target2;
1384         else if (sh->ops.target2 < 0)
1385                 target = sh->ops.target;
1386         else
1387                 /* we should only have one valid target */
1388                 BUG();
1389         BUG_ON(target < 0);
1390         pr_debug("%s: stripe %llu block: %d\n",
1391                 __func__, (unsigned long long)sh->sector, target);
1392
1393         tgt = &sh->dev[target];
1394         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1395         dest = tgt->page;
1396
1397         atomic_inc(&sh->count);
1398
1399         if (target == qd_idx) {
1400                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1401                 blocks[count] = NULL; /* regenerating p is not necessary */
1402                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1403                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1404                                   ops_complete_compute, sh,
1405                                   to_addr_conv(sh, percpu, 0));
1406                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1407         } else {
1408                 /* Compute any data- or p-drive using XOR */
1409                 count = 0;
1410                 for (i = disks; i-- ; ) {
1411                         if (i == target || i == qd_idx)
1412                                 continue;
1413                         blocks[count++] = sh->dev[i].page;
1414                 }
1415
1416                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1417                                   NULL, ops_complete_compute, sh,
1418                                   to_addr_conv(sh, percpu, 0));
1419                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1420         }
1421
1422         return tx;
1423 }
1424
1425 static struct dma_async_tx_descriptor *
1426 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1427 {
1428         int i, count, disks = sh->disks;
1429         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1430         int d0_idx = raid6_d0(sh);
1431         int faila = -1, failb = -1;
1432         int target = sh->ops.target;
1433         int target2 = sh->ops.target2;
1434         struct r5dev *tgt = &sh->dev[target];
1435         struct r5dev *tgt2 = &sh->dev[target2];
1436         struct dma_async_tx_descriptor *tx;
1437         struct page **blocks = to_addr_page(percpu, 0);
1438         struct async_submit_ctl submit;
1439
1440         BUG_ON(sh->batch_head);
1441         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1442                  __func__, (unsigned long long)sh->sector, target, target2);
1443         BUG_ON(target < 0 || target2 < 0);
1444         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1445         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1446
1447         /* we need to open-code set_syndrome_sources to handle the
1448          * slot number conversion for 'faila' and 'failb'
1449          */
1450         for (i = 0; i < disks ; i++)
1451                 blocks[i] = NULL;
1452         count = 0;
1453         i = d0_idx;
1454         do {
1455                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1456
1457                 blocks[slot] = sh->dev[i].page;
1458
1459                 if (i == target)
1460                         faila = slot;
1461                 if (i == target2)
1462                         failb = slot;
1463                 i = raid6_next_disk(i, disks);
1464         } while (i != d0_idx);
1465
1466         BUG_ON(faila == failb);
1467         if (failb < faila)
1468                 swap(faila, failb);
1469         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1470                  __func__, (unsigned long long)sh->sector, faila, failb);
1471
1472         atomic_inc(&sh->count);
1473
1474         if (failb == syndrome_disks+1) {
1475                 /* Q disk is one of the missing disks */
1476                 if (faila == syndrome_disks) {
1477                         /* Missing P+Q, just recompute */
1478                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1479                                           ops_complete_compute, sh,
1480                                           to_addr_conv(sh, percpu, 0));
1481                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1482                                                   STRIPE_SIZE, &submit);
1483                 } else {
1484                         struct page *dest;
1485                         int data_target;
1486                         int qd_idx = sh->qd_idx;
1487
1488                         /* Missing D+Q: recompute D from P, then recompute Q */
1489                         if (target == qd_idx)
1490                                 data_target = target2;
1491                         else
1492                                 data_target = target;
1493
1494                         count = 0;
1495                         for (i = disks; i-- ; ) {
1496                                 if (i == data_target || i == qd_idx)
1497                                         continue;
1498                                 blocks[count++] = sh->dev[i].page;
1499                         }
1500                         dest = sh->dev[data_target].page;
1501                         init_async_submit(&submit,
1502                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1503                                           NULL, NULL, NULL,
1504                                           to_addr_conv(sh, percpu, 0));
1505                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1506                                        &submit);
1507
1508                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1509                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1510                                           ops_complete_compute, sh,
1511                                           to_addr_conv(sh, percpu, 0));
1512                         return async_gen_syndrome(blocks, 0, count+2,
1513                                                   STRIPE_SIZE, &submit);
1514                 }
1515         } else {
1516                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1517                                   ops_complete_compute, sh,
1518                                   to_addr_conv(sh, percpu, 0));
1519                 if (failb == syndrome_disks) {
1520                         /* We're missing D+P. */
1521                         return async_raid6_datap_recov(syndrome_disks+2,
1522                                                        STRIPE_SIZE, faila,
1523                                                        blocks, &submit);
1524                 } else {
1525                         /* We're missing D+D. */
1526                         return async_raid6_2data_recov(syndrome_disks+2,
1527                                                        STRIPE_SIZE, faila, failb,
1528                                                        blocks, &submit);
1529                 }
1530         }
1531 }
1532
1533 static void ops_complete_prexor(void *stripe_head_ref)
1534 {
1535         struct stripe_head *sh = stripe_head_ref;
1536
1537         pr_debug("%s: stripe %llu\n", __func__,
1538                 (unsigned long long)sh->sector);
1539
1540         if (r5c_is_writeback(sh->raid_conf->log))
1541                 /*
1542                  * raid5-cache write back uses orig_page during prexor.
1543                  * After prexor, it is time to free orig_page
1544                  */
1545                 r5c_release_extra_page(sh);
1546 }
1547
1548 static struct dma_async_tx_descriptor *
1549 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1550                 struct dma_async_tx_descriptor *tx)
1551 {
1552         int disks = sh->disks;
1553         struct page **xor_srcs = to_addr_page(percpu, 0);
1554         int count = 0, pd_idx = sh->pd_idx, i;
1555         struct async_submit_ctl submit;
1556
1557         /* existing parity data subtracted */
1558         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1559
1560         BUG_ON(sh->batch_head);
1561         pr_debug("%s: stripe %llu\n", __func__,
1562                 (unsigned long long)sh->sector);
1563
1564         for (i = disks; i--; ) {
1565                 struct r5dev *dev = &sh->dev[i];
1566                 /* Only process blocks that are known to be uptodate */
1567                 if (test_bit(R5_InJournal, &dev->flags))
1568                         xor_srcs[count++] = dev->orig_page;
1569                 else if (test_bit(R5_Wantdrain, &dev->flags))
1570                         xor_srcs[count++] = dev->page;
1571         }
1572
1573         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1574                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1575         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1576
1577         return tx;
1578 }
1579
1580 static struct dma_async_tx_descriptor *
1581 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1582                 struct dma_async_tx_descriptor *tx)
1583 {
1584         struct page **blocks = to_addr_page(percpu, 0);
1585         int count;
1586         struct async_submit_ctl submit;
1587
1588         pr_debug("%s: stripe %llu\n", __func__,
1589                 (unsigned long long)sh->sector);
1590
1591         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1592
1593         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1594                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1595         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1596
1597         return tx;
1598 }
1599
1600 static struct dma_async_tx_descriptor *
1601 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1602 {
1603         struct r5conf *conf = sh->raid_conf;
1604         int disks = sh->disks;
1605         int i;
1606         struct stripe_head *head_sh = sh;
1607
1608         pr_debug("%s: stripe %llu\n", __func__,
1609                 (unsigned long long)sh->sector);
1610
1611         for (i = disks; i--; ) {
1612                 struct r5dev *dev;
1613                 struct bio *chosen;
1614
1615                 sh = head_sh;
1616                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1617                         struct bio *wbi;
1618
1619 again:
1620                         dev = &sh->dev[i];
1621                         /*
1622                          * clear R5_InJournal, so when rewriting a page in
1623                          * journal, it is not skipped by r5l_log_stripe()
1624                          */
1625                         clear_bit(R5_InJournal, &dev->flags);
1626                         spin_lock_irq(&sh->stripe_lock);
1627                         chosen = dev->towrite;
1628                         dev->towrite = NULL;
1629                         sh->overwrite_disks = 0;
1630                         BUG_ON(dev->written);
1631                         wbi = dev->written = chosen;
1632                         spin_unlock_irq(&sh->stripe_lock);
1633                         WARN_ON(dev->page != dev->orig_page);
1634
1635                         while (wbi && wbi->bi_iter.bi_sector <
1636                                 dev->sector + STRIPE_SECTORS) {
1637                                 if (wbi->bi_opf & REQ_FUA)
1638                                         set_bit(R5_WantFUA, &dev->flags);
1639                                 if (wbi->bi_opf & REQ_SYNC)
1640                                         set_bit(R5_SyncIO, &dev->flags);
1641                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1642                                         set_bit(R5_Discard, &dev->flags);
1643                                 else {
1644                                         tx = async_copy_data(1, wbi, &dev->page,
1645                                                              dev->sector, tx, sh,
1646                                                              r5c_is_writeback(conf->log));
1647                                         if (dev->page != dev->orig_page &&
1648                                             !r5c_is_writeback(conf->log)) {
1649                                                 set_bit(R5_SkipCopy, &dev->flags);
1650                                                 clear_bit(R5_UPTODATE, &dev->flags);
1651                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1652                                         }
1653                                 }
1654                                 wbi = r5_next_bio(wbi, dev->sector);
1655                         }
1656
1657                         if (head_sh->batch_head) {
1658                                 sh = list_first_entry(&sh->batch_list,
1659                                                       struct stripe_head,
1660                                                       batch_list);
1661                                 if (sh == head_sh)
1662                                         continue;
1663                                 goto again;
1664                         }
1665                 }
1666         }
1667
1668         return tx;
1669 }
1670
1671 static void ops_complete_reconstruct(void *stripe_head_ref)
1672 {
1673         struct stripe_head *sh = stripe_head_ref;
1674         int disks = sh->disks;
1675         int pd_idx = sh->pd_idx;
1676         int qd_idx = sh->qd_idx;
1677         int i;
1678         bool fua = false, sync = false, discard = false;
1679
1680         pr_debug("%s: stripe %llu\n", __func__,
1681                 (unsigned long long)sh->sector);
1682
1683         for (i = disks; i--; ) {
1684                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1685                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1686                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1687         }
1688
1689         for (i = disks; i--; ) {
1690                 struct r5dev *dev = &sh->dev[i];
1691
1692                 if (dev->written || i == pd_idx || i == qd_idx) {
1693                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1694                                 set_bit(R5_UPTODATE, &dev->flags);
1695                         if (fua)
1696                                 set_bit(R5_WantFUA, &dev->flags);
1697                         if (sync)
1698                                 set_bit(R5_SyncIO, &dev->flags);
1699                 }
1700         }
1701
1702         if (sh->reconstruct_state == reconstruct_state_drain_run)
1703                 sh->reconstruct_state = reconstruct_state_drain_result;
1704         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1705                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1706         else {
1707                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1708                 sh->reconstruct_state = reconstruct_state_result;
1709         }
1710
1711         set_bit(STRIPE_HANDLE, &sh->state);
1712         raid5_release_stripe(sh);
1713 }
1714
1715 static void
1716 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1717                      struct dma_async_tx_descriptor *tx)
1718 {
1719         int disks = sh->disks;
1720         struct page **xor_srcs;
1721         struct async_submit_ctl submit;
1722         int count, pd_idx = sh->pd_idx, i;
1723         struct page *xor_dest;
1724         int prexor = 0;
1725         unsigned long flags;
1726         int j = 0;
1727         struct stripe_head *head_sh = sh;
1728         int last_stripe;
1729
1730         pr_debug("%s: stripe %llu\n", __func__,
1731                 (unsigned long long)sh->sector);
1732
1733         for (i = 0; i < sh->disks; i++) {
1734                 if (pd_idx == i)
1735                         continue;
1736                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1737                         break;
1738         }
1739         if (i >= sh->disks) {
1740                 atomic_inc(&sh->count);
1741                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1742                 ops_complete_reconstruct(sh);
1743                 return;
1744         }
1745 again:
1746         count = 0;
1747         xor_srcs = to_addr_page(percpu, j);
1748         /* check if prexor is active which means only process blocks
1749          * that are part of a read-modify-write (written)
1750          */
1751         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1752                 prexor = 1;
1753                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1754                 for (i = disks; i--; ) {
1755                         struct r5dev *dev = &sh->dev[i];
1756                         if (head_sh->dev[i].written ||
1757                             test_bit(R5_InJournal, &head_sh->dev[i].flags))
1758                                 xor_srcs[count++] = dev->page;
1759                 }
1760         } else {
1761                 xor_dest = sh->dev[pd_idx].page;
1762                 for (i = disks; i--; ) {
1763                         struct r5dev *dev = &sh->dev[i];
1764                         if (i != pd_idx)
1765                                 xor_srcs[count++] = dev->page;
1766                 }
1767         }
1768
1769         /* 1/ if we prexor'd then the dest is reused as a source
1770          * 2/ if we did not prexor then we are redoing the parity
1771          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1772          * for the synchronous xor case
1773          */
1774         last_stripe = !head_sh->batch_head ||
1775                 list_first_entry(&sh->batch_list,
1776                                  struct stripe_head, batch_list) == head_sh;
1777         if (last_stripe) {
1778                 flags = ASYNC_TX_ACK |
1779                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1780
1781                 atomic_inc(&head_sh->count);
1782                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1783                                   to_addr_conv(sh, percpu, j));
1784         } else {
1785                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1786                 init_async_submit(&submit, flags, tx, NULL, NULL,
1787                                   to_addr_conv(sh, percpu, j));
1788         }
1789
1790         if (unlikely(count == 1))
1791                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1792         else
1793                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1794         if (!last_stripe) {
1795                 j++;
1796                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1797                                       batch_list);
1798                 goto again;
1799         }
1800 }
1801
1802 static void
1803 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1804                      struct dma_async_tx_descriptor *tx)
1805 {
1806         struct async_submit_ctl submit;
1807         struct page **blocks;
1808         int count, i, j = 0;
1809         struct stripe_head *head_sh = sh;
1810         int last_stripe;
1811         int synflags;
1812         unsigned long txflags;
1813
1814         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1815
1816         for (i = 0; i < sh->disks; i++) {
1817                 if (sh->pd_idx == i || sh->qd_idx == i)
1818                         continue;
1819                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1820                         break;
1821         }
1822         if (i >= sh->disks) {
1823                 atomic_inc(&sh->count);
1824                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1825                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1826                 ops_complete_reconstruct(sh);
1827                 return;
1828         }
1829
1830 again:
1831         blocks = to_addr_page(percpu, j);
1832
1833         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1834                 synflags = SYNDROME_SRC_WRITTEN;
1835                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1836         } else {
1837                 synflags = SYNDROME_SRC_ALL;
1838                 txflags = ASYNC_TX_ACK;
1839         }
1840
1841         count = set_syndrome_sources(blocks, sh, synflags);
1842         last_stripe = !head_sh->batch_head ||
1843                 list_first_entry(&sh->batch_list,
1844                                  struct stripe_head, batch_list) == head_sh;
1845
1846         if (last_stripe) {
1847                 atomic_inc(&head_sh->count);
1848                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1849                                   head_sh, to_addr_conv(sh, percpu, j));
1850         } else
1851                 init_async_submit(&submit, 0, tx, NULL, NULL,
1852                                   to_addr_conv(sh, percpu, j));
1853         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1854         if (!last_stripe) {
1855                 j++;
1856                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1857                                       batch_list);
1858                 goto again;
1859         }
1860 }
1861
1862 static void ops_complete_check(void *stripe_head_ref)
1863 {
1864         struct stripe_head *sh = stripe_head_ref;
1865
1866         pr_debug("%s: stripe %llu\n", __func__,
1867                 (unsigned long long)sh->sector);
1868
1869         sh->check_state = check_state_check_result;
1870         set_bit(STRIPE_HANDLE, &sh->state);
1871         raid5_release_stripe(sh);
1872 }
1873
1874 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1875 {
1876         int disks = sh->disks;
1877         int pd_idx = sh->pd_idx;
1878         int qd_idx = sh->qd_idx;
1879         struct page *xor_dest;
1880         struct page **xor_srcs = to_addr_page(percpu, 0);
1881         struct dma_async_tx_descriptor *tx;
1882         struct async_submit_ctl submit;
1883         int count;
1884         int i;
1885
1886         pr_debug("%s: stripe %llu\n", __func__,
1887                 (unsigned long long)sh->sector);
1888
1889         BUG_ON(sh->batch_head);
1890         count = 0;
1891         xor_dest = sh->dev[pd_idx].page;
1892         xor_srcs[count++] = xor_dest;
1893         for (i = disks; i--; ) {
1894                 if (i == pd_idx || i == qd_idx)
1895                         continue;
1896                 xor_srcs[count++] = sh->dev[i].page;
1897         }
1898
1899         init_async_submit(&submit, 0, NULL, NULL, NULL,
1900                           to_addr_conv(sh, percpu, 0));
1901         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1902                            &sh->ops.zero_sum_result, &submit);
1903
1904         atomic_inc(&sh->count);
1905         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1906         tx = async_trigger_callback(&submit);
1907 }
1908
1909 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1910 {
1911         struct page **srcs = to_addr_page(percpu, 0);
1912         struct async_submit_ctl submit;
1913         int count;
1914
1915         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1916                 (unsigned long long)sh->sector, checkp);
1917
1918         BUG_ON(sh->batch_head);
1919         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1920         if (!checkp)
1921                 srcs[count] = NULL;
1922
1923         atomic_inc(&sh->count);
1924         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1925                           sh, to_addr_conv(sh, percpu, 0));
1926         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1927                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1928 }
1929
1930 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1931 {
1932         int overlap_clear = 0, i, disks = sh->disks;
1933         struct dma_async_tx_descriptor *tx = NULL;
1934         struct r5conf *conf = sh->raid_conf;
1935         int level = conf->level;
1936         struct raid5_percpu *percpu;
1937         unsigned long cpu;
1938
1939         cpu = get_cpu();
1940         percpu = per_cpu_ptr(conf->percpu, cpu);
1941         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1942                 ops_run_biofill(sh);
1943                 overlap_clear++;
1944         }
1945
1946         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1947                 if (level < 6)
1948                         tx = ops_run_compute5(sh, percpu);
1949                 else {
1950                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1951                                 tx = ops_run_compute6_1(sh, percpu);
1952                         else
1953                                 tx = ops_run_compute6_2(sh, percpu);
1954                 }
1955                 /* terminate the chain if reconstruct is not set to be run */
1956                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1957                         async_tx_ack(tx);
1958         }
1959
1960         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1961                 if (level < 6)
1962                         tx = ops_run_prexor5(sh, percpu, tx);
1963                 else
1964                         tx = ops_run_prexor6(sh, percpu, tx);
1965         }
1966
1967         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1968                 tx = ops_run_biodrain(sh, tx);
1969                 overlap_clear++;
1970         }
1971
1972         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1973                 if (level < 6)
1974                         ops_run_reconstruct5(sh, percpu, tx);
1975                 else
1976                         ops_run_reconstruct6(sh, percpu, tx);
1977         }
1978
1979         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1980                 if (sh->check_state == check_state_run)
1981                         ops_run_check_p(sh, percpu);
1982                 else if (sh->check_state == check_state_run_q)
1983                         ops_run_check_pq(sh, percpu, 0);
1984                 else if (sh->check_state == check_state_run_pq)
1985                         ops_run_check_pq(sh, percpu, 1);
1986                 else
1987                         BUG();
1988         }
1989
1990         if (overlap_clear && !sh->batch_head)
1991                 for (i = disks; i--; ) {
1992                         struct r5dev *dev = &sh->dev[i];
1993                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1994                                 wake_up(&sh->raid_conf->wait_for_overlap);
1995                 }
1996         put_cpu();
1997 }
1998
1999 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2000         int disks)
2001 {
2002         struct stripe_head *sh;
2003         int i;
2004
2005         sh = kmem_cache_zalloc(sc, gfp);
2006         if (sh) {
2007                 spin_lock_init(&sh->stripe_lock);
2008                 spin_lock_init(&sh->batch_lock);
2009                 INIT_LIST_HEAD(&sh->batch_list);
2010                 INIT_LIST_HEAD(&sh->lru);
2011                 INIT_LIST_HEAD(&sh->r5c);
2012                 INIT_LIST_HEAD(&sh->log_list);
2013                 atomic_set(&sh->count, 1);
2014                 sh->log_start = MaxSector;
2015                 for (i = 0; i < disks; i++) {
2016                         struct r5dev *dev = &sh->dev[i];
2017
2018                         bio_init(&dev->req, &dev->vec, 1);
2019                         bio_init(&dev->rreq, &dev->rvec, 1);
2020                 }
2021         }
2022         return sh;
2023 }
2024 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2025 {
2026         struct stripe_head *sh;
2027
2028         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size);
2029         if (!sh)
2030                 return 0;
2031
2032         sh->raid_conf = conf;
2033
2034         if (grow_buffers(sh, gfp)) {
2035                 shrink_buffers(sh);
2036                 kmem_cache_free(conf->slab_cache, sh);
2037                 return 0;
2038         }
2039         sh->hash_lock_index =
2040                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2041         /* we just created an active stripe so... */
2042         atomic_inc(&conf->active_stripes);
2043
2044         raid5_release_stripe(sh);
2045         conf->max_nr_stripes++;
2046         return 1;
2047 }
2048
2049 static int grow_stripes(struct r5conf *conf, int num)
2050 {
2051         struct kmem_cache *sc;
2052         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2053
2054         if (conf->mddev->gendisk)
2055                 sprintf(conf->cache_name[0],
2056                         "raid%d-%s", conf->level, mdname(conf->mddev));
2057         else
2058                 sprintf(conf->cache_name[0],
2059                         "raid%d-%p", conf->level, conf->mddev);
2060         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2061
2062         conf->active_name = 0;
2063         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2064                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2065                                0, 0, NULL);
2066         if (!sc)
2067                 return 1;
2068         conf->slab_cache = sc;
2069         conf->pool_size = devs;
2070         while (num--)
2071                 if (!grow_one_stripe(conf, GFP_KERNEL))
2072                         return 1;
2073
2074         return 0;
2075 }
2076
2077 /**
2078  * scribble_len - return the required size of the scribble region
2079  * @num - total number of disks in the array
2080  *
2081  * The size must be enough to contain:
2082  * 1/ a struct page pointer for each device in the array +2
2083  * 2/ room to convert each entry in (1) to its corresponding dma
2084  *    (dma_map_page()) or page (page_address()) address.
2085  *
2086  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2087  * calculate over all devices (not just the data blocks), using zeros in place
2088  * of the P and Q blocks.
2089  */
2090 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2091 {
2092         struct flex_array *ret;
2093         size_t len;
2094
2095         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2096         ret = flex_array_alloc(len, cnt, flags);
2097         if (!ret)
2098                 return NULL;
2099         /* always prealloc all elements, so no locking is required */
2100         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2101                 flex_array_free(ret);
2102                 return NULL;
2103         }
2104         return ret;
2105 }
2106
2107 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2108 {
2109         unsigned long cpu;
2110         int err = 0;
2111
2112         /*
2113          * Never shrink. And mddev_suspend() could deadlock if this is called
2114          * from raid5d. In that case, scribble_disks and scribble_sectors
2115          * should equal to new_disks and new_sectors
2116          */
2117         if (conf->scribble_disks >= new_disks &&
2118             conf->scribble_sectors >= new_sectors)
2119                 return 0;
2120         mddev_suspend(conf->mddev);
2121         get_online_cpus();
2122         for_each_present_cpu(cpu) {
2123                 struct raid5_percpu *percpu;
2124                 struct flex_array *scribble;
2125
2126                 percpu = per_cpu_ptr(conf->percpu, cpu);
2127                 scribble = scribble_alloc(new_disks,
2128                                           new_sectors / STRIPE_SECTORS,
2129                                           GFP_NOIO);
2130
2131                 if (scribble) {
2132                         flex_array_free(percpu->scribble);
2133                         percpu->scribble = scribble;
2134                 } else {
2135                         err = -ENOMEM;
2136                         break;
2137                 }
2138         }
2139         put_online_cpus();
2140         mddev_resume(conf->mddev);
2141         if (!err) {
2142                 conf->scribble_disks = new_disks;
2143                 conf->scribble_sectors = new_sectors;
2144         }
2145         return err;
2146 }
2147
2148 static int resize_stripes(struct r5conf *conf, int newsize)
2149 {
2150         /* Make all the stripes able to hold 'newsize' devices.
2151          * New slots in each stripe get 'page' set to a new page.
2152          *
2153          * This happens in stages:
2154          * 1/ create a new kmem_cache and allocate the required number of
2155          *    stripe_heads.
2156          * 2/ gather all the old stripe_heads and transfer the pages across
2157          *    to the new stripe_heads.  This will have the side effect of
2158          *    freezing the array as once all stripe_heads have been collected,
2159          *    no IO will be possible.  Old stripe heads are freed once their
2160          *    pages have been transferred over, and the old kmem_cache is
2161          *    freed when all stripes are done.
2162          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2163          *    we simple return a failre status - no need to clean anything up.
2164          * 4/ allocate new pages for the new slots in the new stripe_heads.
2165          *    If this fails, we don't bother trying the shrink the
2166          *    stripe_heads down again, we just leave them as they are.
2167          *    As each stripe_head is processed the new one is released into
2168          *    active service.
2169          *
2170          * Once step2 is started, we cannot afford to wait for a write,
2171          * so we use GFP_NOIO allocations.
2172          */
2173         struct stripe_head *osh, *nsh;
2174         LIST_HEAD(newstripes);
2175         struct disk_info *ndisks;
2176         int err;
2177         struct kmem_cache *sc;
2178         int i;
2179         int hash, cnt;
2180
2181         if (newsize <= conf->pool_size)
2182                 return 0; /* never bother to shrink */
2183
2184         err = md_allow_write(conf->mddev);
2185         if (err)
2186                 return err;
2187
2188         /* Step 1 */
2189         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2190                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2191                                0, 0, NULL);
2192         if (!sc)
2193                 return -ENOMEM;
2194
2195         /* Need to ensure auto-resizing doesn't interfere */
2196         mutex_lock(&conf->cache_size_mutex);
2197
2198         for (i = conf->max_nr_stripes; i; i--) {
2199                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize);
2200                 if (!nsh)
2201                         break;
2202
2203                 nsh->raid_conf = conf;
2204                 list_add(&nsh->lru, &newstripes);
2205         }
2206         if (i) {
2207                 /* didn't get enough, give up */
2208                 while (!list_empty(&newstripes)) {
2209                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2210                         list_del(&nsh->lru);
2211                         kmem_cache_free(sc, nsh);
2212                 }
2213                 kmem_cache_destroy(sc);
2214                 mutex_unlock(&conf->cache_size_mutex);
2215                 return -ENOMEM;
2216         }
2217         /* Step 2 - Must use GFP_NOIO now.
2218          * OK, we have enough stripes, start collecting inactive
2219          * stripes and copying them over
2220          */
2221         hash = 0;
2222         cnt = 0;
2223         list_for_each_entry(nsh, &newstripes, lru) {
2224                 lock_device_hash_lock(conf, hash);
2225                 wait_event_cmd(conf->wait_for_stripe,
2226                                     !list_empty(conf->inactive_list + hash),
2227                                     unlock_device_hash_lock(conf, hash),
2228                                     lock_device_hash_lock(conf, hash));
2229                 osh = get_free_stripe(conf, hash);
2230                 unlock_device_hash_lock(conf, hash);
2231
2232                 for(i=0; i<conf->pool_size; i++) {
2233                         nsh->dev[i].page = osh->dev[i].page;
2234                         nsh->dev[i].orig_page = osh->dev[i].page;
2235                 }
2236                 nsh->hash_lock_index = hash;
2237                 kmem_cache_free(conf->slab_cache, osh);
2238                 cnt++;
2239                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2240                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2241                         hash++;
2242                         cnt = 0;
2243                 }
2244         }
2245         kmem_cache_destroy(conf->slab_cache);
2246
2247         /* Step 3.
2248          * At this point, we are holding all the stripes so the array
2249          * is completely stalled, so now is a good time to resize
2250          * conf->disks and the scribble region
2251          */
2252         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2253         if (ndisks) {
2254                 for (i = 0; i < conf->pool_size; i++)
2255                         ndisks[i] = conf->disks[i];
2256
2257                 for (i = conf->pool_size; i < newsize; i++) {
2258                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2259                         if (!ndisks[i].extra_page)
2260                                 err = -ENOMEM;
2261                 }
2262
2263                 if (err) {
2264                         for (i = conf->pool_size; i < newsize; i++)
2265                                 if (ndisks[i].extra_page)
2266                                         put_page(ndisks[i].extra_page);
2267                         kfree(ndisks);
2268                 } else {
2269                         kfree(conf->disks);
2270                         conf->disks = ndisks;
2271                 }
2272         } else
2273                 err = -ENOMEM;
2274
2275         mutex_unlock(&conf->cache_size_mutex);
2276         /* Step 4, return new stripes to service */
2277         while(!list_empty(&newstripes)) {
2278                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2279                 list_del_init(&nsh->lru);
2280
2281                 for (i=conf->raid_disks; i < newsize; i++)
2282                         if (nsh->dev[i].page == NULL) {
2283                                 struct page *p = alloc_page(GFP_NOIO);
2284                                 nsh->dev[i].page = p;
2285                                 nsh->dev[i].orig_page = p;
2286                                 if (!p)
2287                                         err = -ENOMEM;
2288                         }
2289                 raid5_release_stripe(nsh);
2290         }
2291         /* critical section pass, GFP_NOIO no longer needed */
2292
2293         conf->slab_cache = sc;
2294         conf->active_name = 1-conf->active_name;
2295         if (!err)
2296                 conf->pool_size = newsize;
2297         return err;
2298 }
2299
2300 static int drop_one_stripe(struct r5conf *conf)
2301 {
2302         struct stripe_head *sh;
2303         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2304
2305         spin_lock_irq(conf->hash_locks + hash);
2306         sh = get_free_stripe(conf, hash);
2307         spin_unlock_irq(conf->hash_locks + hash);
2308         if (!sh)
2309                 return 0;
2310         BUG_ON(atomic_read(&sh->count));
2311         shrink_buffers(sh);
2312         kmem_cache_free(conf->slab_cache, sh);
2313         atomic_dec(&conf->active_stripes);
2314         conf->max_nr_stripes--;
2315         return 1;
2316 }
2317
2318 static void shrink_stripes(struct r5conf *conf)
2319 {
2320         while (conf->max_nr_stripes &&
2321                drop_one_stripe(conf))
2322                 ;
2323
2324         kmem_cache_destroy(conf->slab_cache);
2325         conf->slab_cache = NULL;
2326 }
2327
2328 static void raid5_end_read_request(struct bio * bi)
2329 {
2330         struct stripe_head *sh = bi->bi_private;
2331         struct r5conf *conf = sh->raid_conf;
2332         int disks = sh->disks, i;
2333         char b[BDEVNAME_SIZE];
2334         struct md_rdev *rdev = NULL;
2335         sector_t s;
2336
2337         for (i=0 ; i<disks; i++)
2338                 if (bi == &sh->dev[i].req)
2339                         break;
2340
2341         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2342                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2343                 bi->bi_error);
2344         if (i == disks) {
2345                 bio_reset(bi);
2346                 BUG();
2347                 return;
2348         }
2349         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2350                 /* If replacement finished while this request was outstanding,
2351                  * 'replacement' might be NULL already.
2352                  * In that case it moved down to 'rdev'.
2353                  * rdev is not removed until all requests are finished.
2354                  */
2355                 rdev = conf->disks[i].replacement;
2356         if (!rdev)
2357                 rdev = conf->disks[i].rdev;
2358
2359         if (use_new_offset(conf, sh))
2360                 s = sh->sector + rdev->new_data_offset;
2361         else
2362                 s = sh->sector + rdev->data_offset;
2363         if (!bi->bi_error) {
2364                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2365                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2366                         /* Note that this cannot happen on a
2367                          * replacement device.  We just fail those on
2368                          * any error
2369                          */
2370                         pr_info_ratelimited(
2371                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2372                                 mdname(conf->mddev), STRIPE_SECTORS,
2373                                 (unsigned long long)s,
2374                                 bdevname(rdev->bdev, b));
2375                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2376                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2377                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2378                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2379                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2380
2381                 if (atomic_read(&rdev->read_errors))
2382                         atomic_set(&rdev->read_errors, 0);
2383         } else {
2384                 const char *bdn = bdevname(rdev->bdev, b);
2385                 int retry = 0;
2386                 int set_bad = 0;
2387
2388                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2389                 atomic_inc(&rdev->read_errors);
2390                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2391                         pr_warn_ratelimited(
2392                                 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2393                                 mdname(conf->mddev),
2394                                 (unsigned long long)s,
2395                                 bdn);
2396                 else if (conf->mddev->degraded >= conf->max_degraded) {
2397                         set_bad = 1;
2398                         pr_warn_ratelimited(
2399                                 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2400                                 mdname(conf->mddev),
2401                                 (unsigned long long)s,
2402                                 bdn);
2403                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2404                         /* Oh, no!!! */
2405                         set_bad = 1;
2406                         pr_warn_ratelimited(
2407                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2408                                 mdname(conf->mddev),
2409                                 (unsigned long long)s,
2410                                 bdn);
2411                 } else if (atomic_read(&rdev->read_errors)
2412                          > conf->max_nr_stripes)
2413                         pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2414                                mdname(conf->mddev), bdn);
2415                 else
2416                         retry = 1;
2417                 if (set_bad && test_bit(In_sync, &rdev->flags)
2418                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2419                         retry = 1;
2420                 if (retry)
2421                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2422                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2423                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2424                         } else
2425                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2426                 else {
2427                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2428                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2429                         if (!(set_bad
2430                               && test_bit(In_sync, &rdev->flags)
2431                               && rdev_set_badblocks(
2432                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2433                                 md_error(conf->mddev, rdev);
2434                 }
2435         }
2436         rdev_dec_pending(rdev, conf->mddev);
2437         bio_reset(bi);
2438         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2439         set_bit(STRIPE_HANDLE, &sh->state);
2440         raid5_release_stripe(sh);
2441 }
2442
2443 static void raid5_end_write_request(struct bio *bi)
2444 {
2445         struct stripe_head *sh = bi->bi_private;
2446         struct r5conf *conf = sh->raid_conf;
2447         int disks = sh->disks, i;
2448         struct md_rdev *uninitialized_var(rdev);
2449         sector_t first_bad;
2450         int bad_sectors;
2451         int replacement = 0;
2452
2453         for (i = 0 ; i < disks; i++) {
2454                 if (bi == &sh->dev[i].req) {
2455                         rdev = conf->disks[i].rdev;
2456                         break;
2457                 }
2458                 if (bi == &sh->dev[i].rreq) {
2459                         rdev = conf->disks[i].replacement;
2460                         if (rdev)
2461                                 replacement = 1;
2462                         else
2463                                 /* rdev was removed and 'replacement'
2464                                  * replaced it.  rdev is not removed
2465                                  * until all requests are finished.
2466                                  */
2467                                 rdev = conf->disks[i].rdev;
2468                         break;
2469                 }
2470         }
2471         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2472                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2473                 bi->bi_error);
2474         if (i == disks) {
2475                 bio_reset(bi);
2476                 BUG();
2477                 return;
2478         }
2479
2480         if (replacement) {
2481                 if (bi->bi_error)
2482                         md_error(conf->mddev, rdev);
2483                 else if (is_badblock(rdev, sh->sector,
2484                                      STRIPE_SECTORS,
2485                                      &first_bad, &bad_sectors))
2486                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2487         } else {
2488                 if (bi->bi_error) {
2489                         set_bit(STRIPE_DEGRADED, &sh->state);
2490                         set_bit(WriteErrorSeen, &rdev->flags);
2491                         set_bit(R5_WriteError, &sh->dev[i].flags);
2492                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2493                                 set_bit(MD_RECOVERY_NEEDED,
2494                                         &rdev->mddev->recovery);
2495                 } else if (is_badblock(rdev, sh->sector,
2496                                        STRIPE_SECTORS,
2497                                        &first_bad, &bad_sectors)) {
2498                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2499                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2500                                 /* That was a successful write so make
2501                                  * sure it looks like we already did
2502                                  * a re-write.
2503                                  */
2504                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2505                 }
2506         }
2507         rdev_dec_pending(rdev, conf->mddev);
2508
2509         if (sh->batch_head && bi->bi_error && !replacement)
2510                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2511
2512         bio_reset(bi);
2513         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2514                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2515         set_bit(STRIPE_HANDLE, &sh->state);
2516         raid5_release_stripe(sh);
2517
2518         if (sh->batch_head && sh != sh->batch_head)
2519                 raid5_release_stripe(sh->batch_head);
2520 }
2521
2522 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2523 {
2524         struct r5dev *dev = &sh->dev[i];
2525
2526         dev->flags = 0;
2527         dev->sector = raid5_compute_blocknr(sh, i, previous);
2528 }
2529
2530 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2531 {
2532         char b[BDEVNAME_SIZE];
2533         struct r5conf *conf = mddev->private;
2534         unsigned long flags;
2535         pr_debug("raid456: error called\n");
2536
2537         spin_lock_irqsave(&conf->device_lock, flags);
2538         clear_bit(In_sync, &rdev->flags);
2539         mddev->degraded = calc_degraded(conf);
2540         spin_unlock_irqrestore(&conf->device_lock, flags);
2541         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2542
2543         set_bit(Blocked, &rdev->flags);
2544         set_bit(Faulty, &rdev->flags);
2545         set_mask_bits(&mddev->sb_flags, 0,
2546                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2547         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2548                 "md/raid:%s: Operation continuing on %d devices.\n",
2549                 mdname(mddev),
2550                 bdevname(rdev->bdev, b),
2551                 mdname(mddev),
2552                 conf->raid_disks - mddev->degraded);
2553 }
2554
2555 /*
2556  * Input: a 'big' sector number,
2557  * Output: index of the data and parity disk, and the sector # in them.
2558  */
2559 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2560                               int previous, int *dd_idx,
2561                               struct stripe_head *sh)
2562 {
2563         sector_t stripe, stripe2;
2564         sector_t chunk_number;
2565         unsigned int chunk_offset;
2566         int pd_idx, qd_idx;
2567         int ddf_layout = 0;
2568         sector_t new_sector;
2569         int algorithm = previous ? conf->prev_algo
2570                                  : conf->algorithm;
2571         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2572                                          : conf->chunk_sectors;
2573         int raid_disks = previous ? conf->previous_raid_disks
2574                                   : conf->raid_disks;
2575         int data_disks = raid_disks - conf->max_degraded;
2576
2577         /* First compute the information on this sector */
2578
2579         /*
2580          * Compute the chunk number and the sector offset inside the chunk
2581          */
2582         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2583         chunk_number = r_sector;
2584
2585         /*
2586          * Compute the stripe number
2587          */
2588         stripe = chunk_number;
2589         *dd_idx = sector_div(stripe, data_disks);
2590         stripe2 = stripe;
2591         /*
2592          * Select the parity disk based on the user selected algorithm.
2593          */
2594         pd_idx = qd_idx = -1;
2595         switch(conf->level) {
2596         case 4:
2597                 pd_idx = data_disks;
2598                 break;
2599         case 5:
2600                 switch (algorithm) {
2601                 case ALGORITHM_LEFT_ASYMMETRIC:
2602                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2603                         if (*dd_idx >= pd_idx)
2604                                 (*dd_idx)++;
2605                         break;
2606                 case ALGORITHM_RIGHT_ASYMMETRIC:
2607                         pd_idx = sector_div(stripe2, raid_disks);
2608                         if (*dd_idx >= pd_idx)
2609                                 (*dd_idx)++;
2610                         break;
2611                 case ALGORITHM_LEFT_SYMMETRIC:
2612                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2613                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2614                         break;
2615                 case ALGORITHM_RIGHT_SYMMETRIC:
2616                         pd_idx = sector_div(stripe2, raid_disks);
2617                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2618                         break;
2619                 case ALGORITHM_PARITY_0:
2620                         pd_idx = 0;
2621                         (*dd_idx)++;
2622                         break;
2623                 case ALGORITHM_PARITY_N:
2624                         pd_idx = data_disks;
2625                         break;
2626                 default:
2627                         BUG();
2628                 }
2629                 break;
2630         case 6:
2631
2632                 switch (algorithm) {
2633                 case ALGORITHM_LEFT_ASYMMETRIC:
2634                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2635                         qd_idx = pd_idx + 1;
2636                         if (pd_idx == raid_disks-1) {
2637                                 (*dd_idx)++;    /* Q D D D P */
2638                                 qd_idx = 0;
2639                         } else if (*dd_idx >= pd_idx)
2640                                 (*dd_idx) += 2; /* D D P Q D */
2641                         break;
2642                 case ALGORITHM_RIGHT_ASYMMETRIC:
2643                         pd_idx = sector_div(stripe2, raid_disks);
2644                         qd_idx = pd_idx + 1;
2645                         if (pd_idx == raid_disks-1) {
2646                                 (*dd_idx)++;    /* Q D D D P */
2647                                 qd_idx = 0;
2648                         } else if (*dd_idx >= pd_idx)
2649                                 (*dd_idx) += 2; /* D D P Q D */
2650                         break;
2651                 case ALGORITHM_LEFT_SYMMETRIC:
2652                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2653                         qd_idx = (pd_idx + 1) % raid_disks;
2654                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2655                         break;
2656                 case ALGORITHM_RIGHT_SYMMETRIC:
2657                         pd_idx = sector_div(stripe2, raid_disks);
2658                         qd_idx = (pd_idx + 1) % raid_disks;
2659                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2660                         break;
2661
2662                 case ALGORITHM_PARITY_0:
2663                         pd_idx = 0;
2664                         qd_idx = 1;
2665                         (*dd_idx) += 2;
2666                         break;
2667                 case ALGORITHM_PARITY_N:
2668                         pd_idx = data_disks;
2669                         qd_idx = data_disks + 1;
2670                         break;
2671
2672                 case ALGORITHM_ROTATING_ZERO_RESTART:
2673                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2674                          * of blocks for computing Q is different.
2675                          */
2676                         pd_idx = sector_div(stripe2, raid_disks);
2677                         qd_idx = pd_idx + 1;
2678                         if (pd_idx == raid_disks-1) {
2679                                 (*dd_idx)++;    /* Q D D D P */
2680                                 qd_idx = 0;
2681                         } else if (*dd_idx >= pd_idx)
2682                                 (*dd_idx) += 2; /* D D P Q D */
2683                         ddf_layout = 1;
2684                         break;
2685
2686                 case ALGORITHM_ROTATING_N_RESTART:
2687                         /* Same a left_asymmetric, by first stripe is
2688                          * D D D P Q  rather than
2689                          * Q D D D P
2690                          */
2691                         stripe2 += 1;
2692                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2693                         qd_idx = pd_idx + 1;
2694                         if (pd_idx == raid_disks-1) {
2695                                 (*dd_idx)++;    /* Q D D D P */
2696                                 qd_idx = 0;
2697                         } else if (*dd_idx >= pd_idx)
2698                                 (*dd_idx) += 2; /* D D P Q D */
2699                         ddf_layout = 1;
2700                         break;
2701
2702                 case ALGORITHM_ROTATING_N_CONTINUE:
2703                         /* Same as left_symmetric but Q is before P */
2704                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2705                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2706                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2707                         ddf_layout = 1;
2708                         break;
2709
2710                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2711                         /* RAID5 left_asymmetric, with Q on last device */
2712                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2713                         if (*dd_idx >= pd_idx)
2714                                 (*dd_idx)++;
2715                         qd_idx = raid_disks - 1;
2716                         break;
2717
2718                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2719                         pd_idx = sector_div(stripe2, raid_disks-1);
2720                         if (*dd_idx >= pd_idx)
2721                                 (*dd_idx)++;
2722                         qd_idx = raid_disks - 1;
2723                         break;
2724
2725                 case ALGORITHM_LEFT_SYMMETRIC_6:
2726                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2727                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2728                         qd_idx = raid_disks - 1;
2729                         break;
2730
2731                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2732                         pd_idx = sector_div(stripe2, raid_disks-1);
2733                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2734                         qd_idx = raid_disks - 1;
2735                         break;
2736
2737                 case ALGORITHM_PARITY_0_6:
2738                         pd_idx = 0;
2739                         (*dd_idx)++;
2740                         qd_idx = raid_disks - 1;
2741                         break;
2742
2743                 default:
2744                         BUG();
2745                 }
2746                 break;
2747         }
2748
2749         if (sh) {
2750                 sh->pd_idx = pd_idx;
2751                 sh->qd_idx = qd_idx;
2752                 sh->ddf_layout = ddf_layout;
2753         }
2754         /*
2755          * Finally, compute the new sector number
2756          */
2757         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2758         return new_sector;
2759 }
2760
2761 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2762 {
2763         struct r5conf *conf = sh->raid_conf;
2764         int raid_disks = sh->disks;
2765         int data_disks = raid_disks - conf->max_degraded;
2766         sector_t new_sector = sh->sector, check;
2767         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2768                                          : conf->chunk_sectors;
2769         int algorithm = previous ? conf->prev_algo
2770                                  : conf->algorithm;
2771         sector_t stripe;
2772         int chunk_offset;
2773         sector_t chunk_number;
2774         int dummy1, dd_idx = i;
2775         sector_t r_sector;
2776         struct stripe_head sh2;
2777
2778         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2779         stripe = new_sector;
2780
2781         if (i == sh->pd_idx)
2782                 return 0;
2783         switch(conf->level) {
2784         case 4: break;
2785         case 5:
2786                 switch (algorithm) {
2787                 case ALGORITHM_LEFT_ASYMMETRIC:
2788                 case ALGORITHM_RIGHT_ASYMMETRIC:
2789                         if (i > sh->pd_idx)
2790                                 i--;
2791                         break;
2792                 case ALGORITHM_LEFT_SYMMETRIC:
2793                 case ALGORITHM_RIGHT_SYMMETRIC:
2794                         if (i < sh->pd_idx)
2795                                 i += raid_disks;
2796                         i -= (sh->pd_idx + 1);
2797                         break;
2798                 case ALGORITHM_PARITY_0:
2799                         i -= 1;
2800                         break;
2801                 case ALGORITHM_PARITY_N:
2802                         break;
2803                 default:
2804                         BUG();
2805                 }
2806                 break;
2807         case 6:
2808                 if (i == sh->qd_idx)
2809                         return 0; /* It is the Q disk */
2810                 switch (algorithm) {
2811                 case ALGORITHM_LEFT_ASYMMETRIC:
2812                 case ALGORITHM_RIGHT_ASYMMETRIC:
2813                 case ALGORITHM_ROTATING_ZERO_RESTART:
2814                 case ALGORITHM_ROTATING_N_RESTART:
2815                         if (sh->pd_idx == raid_disks-1)
2816                                 i--;    /* Q D D D P */
2817                         else if (i > sh->pd_idx)
2818                                 i -= 2; /* D D P Q D */
2819                         break;
2820                 case ALGORITHM_LEFT_SYMMETRIC:
2821                 case ALGORITHM_RIGHT_SYMMETRIC:
2822                         if (sh->pd_idx == raid_disks-1)
2823                                 i--; /* Q D D D P */
2824                         else {
2825                                 /* D D P Q D */
2826                                 if (i < sh->pd_idx)
2827                                         i += raid_disks;
2828                                 i -= (sh->pd_idx + 2);
2829                         }
2830                         break;
2831                 case ALGORITHM_PARITY_0:
2832                         i -= 2;
2833                         break;
2834                 case ALGORITHM_PARITY_N:
2835                         break;
2836                 case ALGORITHM_ROTATING_N_CONTINUE:
2837                         /* Like left_symmetric, but P is before Q */
2838                         if (sh->pd_idx == 0)
2839                                 i--;    /* P D D D Q */
2840                         else {
2841                                 /* D D Q P D */
2842                                 if (i < sh->pd_idx)
2843                                         i += raid_disks;
2844                                 i -= (sh->pd_idx + 1);
2845                         }
2846                         break;
2847                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2848                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2849                         if (i > sh->pd_idx)
2850                                 i--;
2851                         break;
2852                 case ALGORITHM_LEFT_SYMMETRIC_6:
2853                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2854                         if (i < sh->pd_idx)
2855                                 i += data_disks + 1;
2856                         i -= (sh->pd_idx + 1);
2857                         break;
2858                 case ALGORITHM_PARITY_0_6:
2859                         i -= 1;
2860                         break;
2861                 default:
2862                         BUG();
2863                 }
2864                 break;
2865         }
2866
2867         chunk_number = stripe * data_disks + i;
2868         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2869
2870         check = raid5_compute_sector(conf, r_sector,
2871                                      previous, &dummy1, &sh2);
2872         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2873                 || sh2.qd_idx != sh->qd_idx) {
2874                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
2875                         mdname(conf->mddev));
2876                 return 0;
2877         }
2878         return r_sector;
2879 }
2880
2881 static void
2882 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2883                          int rcw, int expand)
2884 {
2885         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2886         struct r5conf *conf = sh->raid_conf;
2887         int level = conf->level;
2888
2889         if (rcw) {
2890                 /*
2891                  * In some cases, handle_stripe_dirtying initially decided to
2892                  * run rmw and allocates extra page for prexor. However, rcw is
2893                  * cheaper later on. We need to free the extra page now,
2894                  * because we won't be able to do that in ops_complete_prexor().
2895                  */
2896                 r5c_release_extra_page(sh);
2897
2898                 for (i = disks; i--; ) {
2899                         struct r5dev *dev = &sh->dev[i];
2900
2901                         if (dev->towrite) {
2902                                 set_bit(R5_LOCKED, &dev->flags);
2903                                 set_bit(R5_Wantdrain, &dev->flags);
2904                                 if (!expand)
2905                                         clear_bit(R5_UPTODATE, &dev->flags);
2906                                 s->locked++;
2907                         } else if (test_bit(R5_InJournal, &dev->flags)) {
2908                                 set_bit(R5_LOCKED, &dev->flags);
2909                                 s->locked++;
2910                         }
2911                 }
2912                 /* if we are not expanding this is a proper write request, and
2913                  * there will be bios with new data to be drained into the
2914                  * stripe cache
2915                  */
2916                 if (!expand) {
2917                         if (!s->locked)
2918                                 /* False alarm, nothing to do */
2919                                 return;
2920                         sh->reconstruct_state = reconstruct_state_drain_run;
2921                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2922                 } else
2923                         sh->reconstruct_state = reconstruct_state_run;
2924
2925                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2926
2927                 if (s->locked + conf->max_degraded == disks)
2928                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2929                                 atomic_inc(&conf->pending_full_writes);
2930         } else {
2931                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2932                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2933                 BUG_ON(level == 6 &&
2934                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2935                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2936
2937                 for (i = disks; i--; ) {
2938                         struct r5dev *dev = &sh->dev[i];
2939                         if (i == pd_idx || i == qd_idx)
2940                                 continue;
2941
2942                         if (dev->towrite &&
2943                             (test_bit(R5_UPTODATE, &dev->flags) ||
2944                              test_bit(R5_Wantcompute, &dev->flags))) {
2945                                 set_bit(R5_Wantdrain, &dev->flags);
2946                                 set_bit(R5_LOCKED, &dev->flags);
2947                                 clear_bit(R5_UPTODATE, &dev->flags);
2948                                 s->locked++;
2949                         } else if (test_bit(R5_InJournal, &dev->flags)) {
2950                                 set_bit(R5_LOCKED, &dev->flags);
2951                                 s->locked++;
2952                         }
2953                 }
2954                 if (!s->locked)
2955                         /* False alarm - nothing to do */
2956                         return;
2957                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2958                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2959                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2960                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2961         }
2962
2963         /* keep the parity disk(s) locked while asynchronous operations
2964          * are in flight
2965          */
2966         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2967         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2968         s->locked++;
2969
2970         if (level == 6) {
2971                 int qd_idx = sh->qd_idx;
2972                 struct r5dev *dev = &sh->dev[qd_idx];
2973
2974                 set_bit(R5_LOCKED, &dev->flags);
2975                 clear_bit(R5_UPTODATE, &dev->flags);
2976                 s->locked++;
2977         }
2978
2979         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2980                 __func__, (unsigned long long)sh->sector,
2981                 s->locked, s->ops_request);
2982 }
2983
2984 /*
2985  * Each stripe/dev can have one or more bion attached.
2986  * toread/towrite point to the first in a chain.
2987  * The bi_next chain must be in order.
2988  */
2989 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2990                           int forwrite, int previous)
2991 {
2992         struct bio **bip;
2993         struct r5conf *conf = sh->raid_conf;
2994         int firstwrite=0;
2995
2996         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2997                 (unsigned long long)bi->bi_iter.bi_sector,
2998                 (unsigned long long)sh->sector);
2999
3000         /*
3001          * If several bio share a stripe. The bio bi_phys_segments acts as a
3002          * reference count to avoid race. The reference count should already be
3003          * increased before this function is called (for example, in
3004          * raid5_make_request()), so other bio sharing this stripe will not free the
3005          * stripe. If a stripe is owned by one stripe, the stripe lock will
3006          * protect it.
3007          */
3008         spin_lock_irq(&sh->stripe_lock);
3009         /* Don't allow new IO added to stripes in batch list */
3010         if (sh->batch_head)
3011                 goto overlap;
3012         if (forwrite) {
3013                 bip = &sh->dev[dd_idx].towrite;
3014                 if (*bip == NULL)
3015                         firstwrite = 1;
3016         } else
3017                 bip = &sh->dev[dd_idx].toread;
3018         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3019                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3020                         goto overlap;
3021                 bip = & (*bip)->bi_next;
3022         }
3023         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3024                 goto overlap;
3025
3026         if (!forwrite || previous)
3027                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3028
3029         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3030         if (*bip)
3031                 bi->bi_next = *bip;
3032         *bip = bi;
3033         raid5_inc_bi_active_stripes(bi);
3034
3035         if (forwrite) {
3036                 /* check if page is covered */
3037                 sector_t sector = sh->dev[dd_idx].sector;
3038                 for (bi=sh->dev[dd_idx].towrite;
3039                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3040                              bi && bi->bi_iter.bi_sector <= sector;
3041                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3042                         if (bio_end_sector(bi) >= sector)
3043                                 sector = bio_end_sector(bi);
3044                 }
3045                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3046                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3047                                 sh->overwrite_disks++;
3048         }
3049
3050         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3051                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3052                 (unsigned long long)sh->sector, dd_idx);
3053
3054         if (conf->mddev->bitmap && firstwrite) {
3055                 /* Cannot hold spinlock over bitmap_startwrite,
3056                  * but must ensure this isn't added to a batch until
3057                  * we have added to the bitmap and set bm_seq.
3058                  * So set STRIPE_BITMAP_PENDING to prevent
3059                  * batching.
3060                  * If multiple add_stripe_bio() calls race here they
3061                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3062                  * to complete "bitmap_startwrite" gets to set
3063                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3064                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3065                  * any more.
3066                  */
3067                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3068                 spin_unlock_irq(&sh->stripe_lock);
3069                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3070                                   STRIPE_SECTORS, 0);
3071                 spin_lock_irq(&sh->stripe_lock);
3072                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3073                 if (!sh->batch_head) {
3074                         sh->bm_seq = conf->seq_flush+1;
3075                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3076                 }
3077         }
3078         spin_unlock_irq(&sh->stripe_lock);
3079
3080         if (stripe_can_batch(sh))
3081                 stripe_add_to_batch_list(conf, sh);
3082         return 1;
3083
3084  overlap:
3085         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3086         spin_unlock_irq(&sh->stripe_lock);
3087         return 0;
3088 }
3089
3090 static void end_reshape(struct r5conf *conf);
3091
3092 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3093                             struct stripe_head *sh)
3094 {
3095         int sectors_per_chunk =
3096                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3097         int dd_idx;
3098         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3099         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3100
3101         raid5_compute_sector(conf,
3102                              stripe * (disks - conf->max_degraded)
3103                              *sectors_per_chunk + chunk_offset,
3104                              previous,
3105                              &dd_idx, sh);
3106 }
3107
3108 static void
3109 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3110                                 struct stripe_head_state *s, int disks,
3111                                 struct bio_list *return_bi)
3112 {
3113         int i;
3114         BUG_ON(sh->batch_head);
3115         for (i = disks; i--; ) {
3116                 struct bio *bi;
3117                 int bitmap_end = 0;
3118
3119                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3120                         struct md_rdev *rdev;
3121                         rcu_read_lock();
3122                         rdev = rcu_dereference(conf->disks[i].rdev);
3123                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3124                             !test_bit(Faulty, &rdev->flags))
3125                                 atomic_inc(&rdev->nr_pending);
3126                         else
3127                                 rdev = NULL;
3128                         rcu_read_unlock();
3129                         if (rdev) {
3130                                 if (!rdev_set_badblocks(
3131                                             rdev,
3132                                             sh->sector,
3133                                             STRIPE_SECTORS, 0))
3134                                         md_error(conf->mddev, rdev);
3135                                 rdev_dec_pending(rdev, conf->mddev);
3136                         }
3137                 }
3138                 spin_lock_irq(&sh->stripe_lock);
3139                 /* fail all writes first */
3140                 bi = sh->dev[i].towrite;
3141                 sh->dev[i].towrite = NULL;
3142                 sh->overwrite_disks = 0;
3143                 spin_unlock_irq(&sh->stripe_lock);
3144                 if (bi)
3145                         bitmap_end = 1;
3146
3147                 r5l_stripe_write_finished(sh);
3148
3149                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3150                         wake_up(&conf->wait_for_overlap);
3151
3152                 while (bi && bi->bi_iter.bi_sector <
3153                         sh->dev[i].sector + STRIPE_SECTORS) {
3154                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3155
3156                         bi->bi_error = -EIO;
3157                         if (!raid5_dec_bi_active_stripes(bi)) {
3158                                 md_write_end(conf->mddev);
3159                                 bio_list_add(return_bi, bi);
3160                         }
3161                         bi = nextbi;
3162                 }
3163                 if (bitmap_end)
3164                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3165                                 STRIPE_SECTORS, 0, 0);
3166                 bitmap_end = 0;
3167                 /* and fail all 'written' */
3168                 bi = sh->dev[i].written;
3169                 sh->dev[i].written = NULL;
3170                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3171                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3172                         sh->dev[i].page = sh->dev[i].orig_page;
3173                 }
3174
3175                 if (bi) bitmap_end = 1;
3176                 while (bi && bi->bi_iter.bi_sector <
3177                        sh->dev[i].sector + STRIPE_SECTORS) {
3178                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3179
3180                         bi->bi_error = -EIO;
3181                         if (!raid5_dec_bi_active_stripes(bi)) {
3182                                 md_write_end(conf->mddev);
3183                                 bio_list_add(return_bi, bi);
3184                         }
3185                         bi = bi2;
3186                 }
3187
3188                 /* fail any reads if this device is non-operational and
3189                  * the data has not reached the cache yet.
3190                  */
3191                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3192                     s->failed > conf->max_degraded &&
3193                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3194                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3195                         spin_lock_irq(&sh->stripe_lock);
3196                         bi = sh->dev[i].toread;
3197                         sh->dev[i].toread = NULL;
3198                         spin_unlock_irq(&sh->stripe_lock);
3199                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3200                                 wake_up(&conf->wait_for_overlap);
3201                         if (bi)
3202                                 s->to_read--;
3203                         while (bi && bi->bi_iter.bi_sector <
3204                                sh->dev[i].sector + STRIPE_SECTORS) {
3205                                 struct bio *nextbi =
3206                                         r5_next_bio(bi, sh->dev[i].sector);
3207
3208                                 bi->bi_error = -EIO;
3209                                 if (!raid5_dec_bi_active_stripes(bi))
3210                                         bio_list_add(return_bi, bi);
3211                                 bi = nextbi;
3212                         }
3213                 }
3214                 if (bitmap_end)
3215                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3216                                         STRIPE_SECTORS, 0, 0);
3217                 /* If we were in the middle of a write the parity block might
3218                  * still be locked - so just clear all R5_LOCKED flags
3219                  */
3220                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3221         }
3222         s->to_write = 0;
3223         s->written = 0;
3224
3225         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3226                 if (atomic_dec_and_test(&conf->pending_full_writes))
3227                         md_wakeup_thread(conf->mddev->thread);
3228 }
3229
3230 static void
3231 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3232                    struct stripe_head_state *s)
3233 {
3234         int abort = 0;
3235         int i;
3236
3237         BUG_ON(sh->batch_head);
3238         clear_bit(STRIPE_SYNCING, &sh->state);
3239         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3240                 wake_up(&conf->wait_for_overlap);
3241         s->syncing = 0;
3242         s->replacing = 0;
3243         /* There is nothing more to do for sync/check/repair.
3244          * Don't even need to abort as that is handled elsewhere
3245          * if needed, and not always wanted e.g. if there is a known
3246          * bad block here.
3247          * For recover/replace we need to record a bad block on all
3248          * non-sync devices, or abort the recovery
3249          */
3250         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3251                 /* During recovery devices cannot be removed, so
3252                  * locking and refcounting of rdevs is not needed
3253                  */
3254                 rcu_read_lock();
3255                 for (i = 0; i < conf->raid_disks; i++) {
3256                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3257                         if (rdev
3258                             && !test_bit(Faulty, &rdev->flags)
3259                             && !test_bit(In_sync, &rdev->flags)
3260                             && !rdev_set_badblocks(rdev, sh->sector,
3261                                                    STRIPE_SECTORS, 0))
3262                                 abort = 1;
3263                         rdev = rcu_dereference(conf->disks[i].replacement);
3264                         if (rdev
3265                             && !test_bit(Faulty, &rdev->flags)
3266                             && !test_bit(In_sync, &rdev->flags)
3267                             && !rdev_set_badblocks(rdev, sh->sector,
3268                                                    STRIPE_SECTORS, 0))
3269                                 abort = 1;
3270                 }
3271                 rcu_read_unlock();
3272                 if (abort)
3273                         conf->recovery_disabled =
3274                                 conf->mddev->recovery_disabled;
3275         }
3276         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3277 }
3278
3279 static int want_replace(struct stripe_head *sh, int disk_idx)
3280 {
3281         struct md_rdev *rdev;
3282         int rv = 0;
3283
3284         rcu_read_lock();
3285         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3286         if (rdev
3287             && !test_bit(Faulty, &rdev->flags)
3288             && !test_bit(In_sync, &rdev->flags)
3289             && (rdev->recovery_offset <= sh->sector
3290                 || rdev->mddev->recovery_cp <= sh->sector))
3291                 rv = 1;
3292         rcu_read_unlock();
3293         return rv;
3294 }
3295
3296 /* fetch_block - checks the given member device to see if its data needs
3297  * to be read or computed to satisfy a request.
3298  *
3299  * Returns 1 when no more member devices need to be checked, otherwise returns
3300  * 0 to tell the loop in handle_stripe_fill to continue
3301  */
3302
3303 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3304                            int disk_idx, int disks)
3305 {
3306         struct r5dev *dev = &sh->dev[disk_idx];
3307         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3308                                   &sh->dev[s->failed_num[1]] };
3309         int i;
3310
3311
3312         if (test_bit(R5_LOCKED, &dev->flags) ||
3313             test_bit(R5_UPTODATE, &dev->flags))
3314                 /* No point reading this as we already have it or have
3315                  * decided to get it.
3316                  */
3317                 return 0;
3318
3319         if (dev->toread ||
3320             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3321                 /* We need this block to directly satisfy a request */
3322                 return 1;
3323
3324         if (s->syncing || s->expanding ||
3325             (s->replacing && want_replace(sh, disk_idx)))
3326                 /* When syncing, or expanding we read everything.
3327                  * When replacing, we need the replaced block.
3328                  */
3329                 return 1;
3330
3331         if ((s->failed >= 1 && fdev[0]->toread) ||
3332             (s->failed >= 2 && fdev[1]->toread))
3333                 /* If we want to read from a failed device, then
3334                  * we need to actually read every other device.
3335                  */
3336                 return 1;
3337
3338         /* Sometimes neither read-modify-write nor reconstruct-write
3339          * cycles can work.  In those cases we read every block we
3340          * can.  Then the parity-update is certain to have enough to
3341          * work with.
3342          * This can only be a problem when we need to write something,
3343          * and some device has failed.  If either of those tests
3344          * fail we need look no further.
3345          */
3346         if (!s->failed || !s->to_write)
3347                 return 0;
3348
3349         if (test_bit(R5_Insync, &dev->flags) &&
3350             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3351                 /* Pre-reads at not permitted until after short delay
3352                  * to gather multiple requests.  However if this
3353                  * device is no Insync, the block could only be be computed
3354                  * and there is no need to delay that.
3355                  */
3356                 return 0;
3357
3358         for (i = 0; i < s->failed && i < 2; i++) {
3359                 if (fdev[i]->towrite &&
3360                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3361                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3362                         /* If we have a partial write to a failed
3363                          * device, then we will need to reconstruct
3364                          * the content of that device, so all other
3365                          * devices must be read.
3366                          */
3367                         return 1;
3368         }
3369
3370         /* If we are forced to do a reconstruct-write, either because
3371          * the current RAID6 implementation only supports that, or
3372          * or because parity cannot be trusted and we are currently
3373          * recovering it, there is extra need to be careful.
3374          * If one of the devices that we would need to read, because
3375          * it is not being overwritten (and maybe not written at all)
3376          * is missing/faulty, then we need to read everything we can.
3377          */
3378         if (sh->raid_conf->level != 6 &&
3379             sh->sector < sh->raid_conf->mddev->recovery_cp)
3380                 /* reconstruct-write isn't being forced */
3381                 return 0;
3382         for (i = 0; i < s->failed && i < 2; i++) {
3383                 if (s->failed_num[i] != sh->pd_idx &&
3384                     s->failed_num[i] != sh->qd_idx &&
3385                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3386                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3387                         return 1;
3388         }
3389
3390         return 0;
3391 }
3392
3393 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3394                        int disk_idx, int disks)
3395 {
3396         struct r5dev *dev = &sh->dev[disk_idx];
3397
3398         /* is the data in this block needed, and can we get it? */
3399         if (need_this_block(sh, s, disk_idx, disks)) {
3400                 /* we would like to get this block, possibly by computing it,
3401                  * otherwise read it if the backing disk is insync
3402                  */
3403                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3404                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3405                 BUG_ON(sh->batch_head);
3406                 if ((s->uptodate == disks - 1) &&
3407                     (s->failed && (disk_idx == s->failed_num[0] ||
3408                                    disk_idx == s->failed_num[1]))) {
3409                         /* have disk failed, and we're requested to fetch it;
3410                          * do compute it
3411                          */
3412                         pr_debug("Computing stripe %llu block %d\n",
3413                                (unsigned long long)sh->sector, disk_idx);
3414                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3415                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3416                         set_bit(R5_Wantcompute, &dev->flags);
3417                         sh->ops.target = disk_idx;
3418                         sh->ops.target2 = -1; /* no 2nd target */
3419                         s->req_compute = 1;
3420                         /* Careful: from this point on 'uptodate' is in the eye
3421                          * of raid_run_ops which services 'compute' operations
3422                          * before writes. R5_Wantcompute flags a block that will
3423                          * be R5_UPTODATE by the time it is needed for a
3424                          * subsequent operation.
3425                          */
3426                         s->uptodate++;
3427                         return 1;
3428                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3429                         /* Computing 2-failure is *very* expensive; only
3430                          * do it if failed >= 2
3431                          */
3432                         int other;
3433                         for (other = disks; other--; ) {
3434                                 if (other == disk_idx)
3435                                         continue;
3436                                 if (!test_bit(R5_UPTODATE,
3437                                       &sh->dev[other].flags))
3438                                         break;
3439                         }
3440                         BUG_ON(other < 0);
3441                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3442                                (unsigned long long)sh->sector,
3443                                disk_idx, other);
3444                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3445                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3446                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3447                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3448                         sh->ops.target = disk_idx;
3449                         sh->ops.target2 = other;
3450                         s->uptodate += 2;
3451                         s->req_compute = 1;
3452                         return 1;
3453                 } else if (test_bit(R5_Insync, &dev->flags)) {
3454                         set_bit(R5_LOCKED, &dev->flags);
3455                         set_bit(R5_Wantread, &dev->flags);
3456                         s->locked++;
3457                         pr_debug("Reading block %d (sync=%d)\n",
3458                                 disk_idx, s->syncing);
3459                 }
3460         }
3461
3462         return 0;
3463 }
3464
3465 /**
3466  * handle_stripe_fill - read or compute data to satisfy pending requests.
3467  */
3468 static void handle_stripe_fill(struct stripe_head *sh,
3469                                struct stripe_head_state *s,
3470                                int disks)
3471 {
3472         int i;
3473
3474         /* look for blocks to read/compute, skip this if a compute
3475          * is already in flight, or if the stripe contents are in the
3476          * midst of changing due to a write
3477          */
3478         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3479             !sh->reconstruct_state)
3480                 for (i = disks; i--; )
3481                         if (fetch_block(sh, s, i, disks))
3482                                 break;
3483         set_bit(STRIPE_HANDLE, &sh->state);
3484 }
3485
3486 static void break_stripe_batch_list(struct stripe_head *head_sh,
3487                                     unsigned long handle_flags);
3488 /* handle_stripe_clean_event
3489  * any written block on an uptodate or failed drive can be returned.
3490  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3491  * never LOCKED, so we don't need to test 'failed' directly.
3492  */
3493 static void handle_stripe_clean_event(struct r5conf *conf,
3494         struct stripe_head *sh, int disks, struct bio_list *return_bi)
3495 {
3496         int i;
3497         struct r5dev *dev;
3498         int discard_pending = 0;
3499         struct stripe_head *head_sh = sh;
3500         bool do_endio = false;
3501
3502         for (i = disks; i--; )
3503                 if (sh->dev[i].written) {
3504                         dev = &sh->dev[i];
3505                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3506                             (test_bit(R5_UPTODATE, &dev->flags) ||
3507                              test_bit(R5_Discard, &dev->flags) ||
3508                              test_bit(R5_SkipCopy, &dev->flags))) {
3509                                 /* We can return any write requests */
3510                                 struct bio *wbi, *wbi2;
3511                                 pr_debug("Return write for disc %d\n", i);
3512                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3513                                         clear_bit(R5_UPTODATE, &dev->flags);
3514                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3515                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3516                                 }
3517                                 do_endio = true;
3518
3519 returnbi:
3520                                 dev->page = dev->orig_page;
3521                                 wbi = dev->written;
3522                                 dev->written = NULL;
3523                                 while (wbi && wbi->bi_iter.bi_sector <
3524                                         dev->sector + STRIPE_SECTORS) {
3525                                         wbi2 = r5_next_bio(wbi, dev->sector);
3526                                         if (!raid5_dec_bi_active_stripes(wbi)) {
3527                                                 md_write_end(conf->mddev);
3528                                                 bio_list_add(return_bi, wbi);
3529                                         }
3530                                         wbi = wbi2;
3531                                 }
3532                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3533                                                 STRIPE_SECTORS,
3534                                          !test_bit(STRIPE_DEGRADED, &sh->state),
3535                                                 0);
3536                                 if (head_sh->batch_head) {
3537                                         sh = list_first_entry(&sh->batch_list,
3538                                                               struct stripe_head,
3539                                                               batch_list);
3540                                         if (sh != head_sh) {
3541                                                 dev = &sh->dev[i];
3542                                                 goto returnbi;
3543                                         }
3544                                 }
3545                                 sh = head_sh;
3546                                 dev = &sh->dev[i];
3547                         } else if (test_bit(R5_Discard, &dev->flags))
3548                                 discard_pending = 1;
3549                 }
3550
3551         r5l_stripe_write_finished(sh);
3552
3553         if (!discard_pending &&
3554             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3555                 int hash;
3556                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3557                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3558                 if (sh->qd_idx >= 0) {
3559                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3560                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3561                 }
3562                 /* now that discard is done we can proceed with any sync */
3563                 clear_bit(STRIPE_DISCARD, &sh->state);
3564                 /*
3565                  * SCSI discard will change some bio fields and the stripe has
3566                  * no updated data, so remove it from hash list and the stripe
3567                  * will be reinitialized
3568                  */
3569 unhash:
3570                 hash = sh->hash_lock_index;
3571                 spin_lock_irq(conf->hash_locks + hash);
3572                 remove_hash(sh);
3573                 spin_unlock_irq(conf->hash_locks + hash);
3574                 if (head_sh->batch_head) {
3575                         sh = list_first_entry(&sh->batch_list,
3576                                               struct stripe_head, batch_list);
3577                         if (sh != head_sh)
3578                                         goto unhash;
3579                 }
3580                 sh = head_sh;
3581
3582                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3583                         set_bit(STRIPE_HANDLE, &sh->state);
3584
3585         }
3586
3587         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3588                 if (atomic_dec_and_test(&conf->pending_full_writes))
3589                         md_wakeup_thread(conf->mddev->thread);
3590
3591         if (head_sh->batch_head && do_endio)
3592                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3593 }
3594
3595 static int handle_stripe_dirtying(struct r5conf *conf,
3596                                   struct stripe_head *sh,
3597                                   struct stripe_head_state *s,
3598                                   int disks)
3599 {
3600         int rmw = 0, rcw = 0, i;
3601         sector_t recovery_cp = conf->mddev->recovery_cp;
3602
3603         /* Check whether resync is now happening or should start.
3604          * If yes, then the array is dirty (after unclean shutdown or
3605          * initial creation), so parity in some stripes might be inconsistent.
3606          * In this case, we need to always do reconstruct-write, to ensure
3607          * that in case of drive failure or read-error correction, we
3608          * generate correct data from the parity.
3609          */
3610         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3611             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3612              s->failed == 0)) {
3613                 /* Calculate the real rcw later - for now make it
3614                  * look like rcw is cheaper
3615                  */
3616                 rcw = 1; rmw = 2;
3617                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3618                          conf->rmw_level, (unsigned long long)recovery_cp,
3619                          (unsigned long long)sh->sector);
3620         } else for (i = disks; i--; ) {
3621                 /* would I have to read this buffer for read_modify_write */
3622                 struct r5dev *dev = &sh->dev[i];
3623                 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx ||
3624                      test_bit(R5_InJournal, &dev->flags)) &&
3625                     !test_bit(R5_LOCKED, &dev->flags) &&
3626                     !((test_bit(R5_UPTODATE, &dev->flags) &&
3627                        (!test_bit(R5_InJournal, &dev->flags) ||
3628                         dev->page != dev->orig_page)) ||
3629                       test_bit(R5_Wantcompute, &dev->flags))) {
3630                         if (test_bit(R5_Insync, &dev->flags))
3631                                 rmw++;
3632                         else
3633                                 rmw += 2*disks;  /* cannot read it */
3634                 }
3635                 /* Would I have to read this buffer for reconstruct_write */
3636                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3637                     i != sh->pd_idx && i != sh->qd_idx &&
3638                     !test_bit(R5_LOCKED, &dev->flags) &&
3639                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3640                       test_bit(R5_InJournal, &dev->flags) ||
3641                       test_bit(R5_Wantcompute, &dev->flags))) {
3642                         if (test_bit(R5_Insync, &dev->flags))
3643                                 rcw++;
3644                         else
3645                                 rcw += 2*disks;
3646                 }
3647         }
3648
3649         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3650                 (unsigned long long)sh->sector, rmw, rcw);
3651         set_bit(STRIPE_HANDLE, &sh->state);
3652         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3653                 /* prefer read-modify-write, but need to get some data */
3654                 if (conf->mddev->queue)
3655                         blk_add_trace_msg(conf->mddev->queue,
3656                                           "raid5 rmw %llu %d",
3657                                           (unsigned long long)sh->sector, rmw);
3658                 for (i = disks; i--; ) {
3659                         struct r5dev *dev = &sh->dev[i];
3660                         if (test_bit(R5_InJournal, &dev->flags) &&
3661                             dev->page == dev->orig_page &&
3662                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3663                                 /* alloc page for prexor */
3664                                 struct page *p = alloc_page(GFP_NOIO);
3665
3666                                 if (p) {
3667                                         dev->orig_page = p;
3668                                         continue;
3669                                 }
3670
3671                                 /*
3672                                  * alloc_page() failed, try use
3673                                  * disk_info->extra_page
3674                                  */
3675                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3676                                                       &conf->cache_state)) {
3677                                         r5c_use_extra_page(sh);
3678                                         break;
3679                                 }
3680
3681                                 /* extra_page in use, add to delayed_list */
3682                                 set_bit(STRIPE_DELAYED, &sh->state);
3683                                 s->waiting_extra_page = 1;
3684                                 return -EAGAIN;
3685                         }
3686                 }
3687
3688                 for (i = disks; i--; ) {
3689                         struct r5dev *dev = &sh->dev[i];
3690                         if ((dev->towrite ||
3691                              i == sh->pd_idx || i == sh->qd_idx ||
3692                              test_bit(R5_InJournal, &dev->flags)) &&
3693                             !test_bit(R5_LOCKED, &dev->flags) &&
3694                             !((test_bit(R5_UPTODATE, &dev->flags) &&
3695                                (!test_bit(R5_InJournal, &dev->flags) ||
3696                                 dev->page != dev->orig_page)) ||
3697                               test_bit(R5_Wantcompute, &dev->flags)) &&
3698                             test_bit(R5_Insync, &dev->flags)) {
3699                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3700                                              &sh->state)) {
3701                                         pr_debug("Read_old block %d for r-m-w\n",
3702                                                  i);
3703                                         set_bit(R5_LOCKED, &dev->flags);
3704                                         set_bit(R5_Wantread, &dev->flags);
3705                                         s->locked++;
3706                                 } else {
3707                                         set_bit(STRIPE_DELAYED, &sh->state);
3708                                         set_bit(STRIPE_HANDLE, &sh->state);
3709                                 }
3710                         }
3711                 }
3712         }
3713         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3714                 /* want reconstruct write, but need to get some data */
3715                 int qread =0;
3716                 rcw = 0;
3717                 for (i = disks; i--; ) {
3718                         struct r5dev *dev = &sh->dev[i];
3719                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3720                             i != sh->pd_idx && i != sh->qd_idx &&
3721                             !test_bit(R5_LOCKED, &dev->flags) &&
3722                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3723                               test_bit(R5_InJournal, &dev->flags) ||
3724                               test_bit(R5_Wantcompute, &dev->flags))) {
3725                                 rcw++;
3726                                 if (test_bit(R5_Insync, &dev->flags) &&
3727                                     test_bit(STRIPE_PREREAD_ACTIVE,
3728                                              &sh->state)) {
3729                                         pr_debug("Read_old block "
3730                                                 "%d for Reconstruct\n", i);
3731                                         set_bit(R5_LOCKED, &dev->flags);
3732                                         set_bit(R5_Wantread, &dev->flags);
3733                                         s->locked++;
3734                                         qread++;
3735                                 } else {
3736                                         set_bit(STRIPE_DELAYED, &sh->state);
3737                                         set_bit(STRIPE_HANDLE, &sh->state);
3738                                 }
3739                         }
3740                 }
3741                 if (rcw && conf->mddev->queue)
3742                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3743                                           (unsigned long long)sh->sector,
3744                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3745         }
3746
3747         if (rcw > disks && rmw > disks &&
3748             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3749                 set_bit(STRIPE_DELAYED, &sh->state);
3750
3751         /* now if nothing is locked, and if we have enough data,
3752          * we can start a write request
3753          */
3754         /* since handle_stripe can be called at any time we need to handle the
3755          * case where a compute block operation has been submitted and then a
3756          * subsequent call wants to start a write request.  raid_run_ops only
3757          * handles the case where compute block and reconstruct are requested
3758          * simultaneously.  If this is not the case then new writes need to be
3759          * held off until the compute completes.
3760          */
3761         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3762             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3763              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3764                 schedule_reconstruction(sh, s, rcw == 0, 0);
3765         return 0;
3766 }
3767
3768 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3769                                 struct stripe_head_state *s, int disks)
3770 {
3771         struct r5dev *dev = NULL;
3772
3773         BUG_ON(sh->batch_head);
3774         set_bit(STRIPE_HANDLE, &sh->state);
3775
3776         switch (sh->check_state) {
3777         case check_state_idle:
3778                 /* start a new check operation if there are no failures */
3779                 if (s->failed == 0) {
3780                         BUG_ON(s->uptodate != disks);
3781                         sh->check_state = check_state_run;
3782                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3783                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3784                         s->uptodate--;
3785                         break;
3786                 }
3787                 dev = &sh->dev[s->failed_num[0]];
3788                 /* fall through */
3789         case check_state_compute_result:
3790                 sh->check_state = check_state_idle;
3791                 if (!dev)
3792                         dev = &sh->dev[sh->pd_idx];
3793
3794                 /* check that a write has not made the stripe insync */
3795                 if (test_bit(STRIPE_INSYNC, &sh->state))
3796                         break;
3797
3798                 /* either failed parity check, or recovery is happening */
3799                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3800                 BUG_ON(s->uptodate != disks);
3801
3802                 set_bit(R5_LOCKED, &dev->flags);
3803                 s->locked++;
3804                 set_bit(R5_Wantwrite, &dev->flags);
3805
3806                 clear_bit(STRIPE_DEGRADED, &sh->state);
3807                 set_bit(STRIPE_INSYNC, &sh->state);
3808                 break;
3809         case check_state_run:
3810                 break; /* we will be called again upon completion */
3811         case check_state_check_result:
3812                 sh->check_state = check_state_idle;
3813
3814                 /* if a failure occurred during the check operation, leave
3815                  * STRIPE_INSYNC not set and let the stripe be handled again
3816                  */
3817                 if (s->failed)
3818                         break;
3819
3820                 /* handle a successful check operation, if parity is correct
3821                  * we are done.  Otherwise update the mismatch count and repair
3822                  * parity if !MD_RECOVERY_CHECK
3823                  */
3824                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3825                         /* parity is correct (on disc,
3826                          * not in buffer any more)
3827                          */
3828                         set_bit(STRIPE_INSYNC, &sh->state);
3829                 else {
3830                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3831                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3832                                 /* don't try to repair!! */
3833                                 set_bit(STRIPE_INSYNC, &sh->state);
3834                         else {
3835                                 sh->check_state = check_state_compute_run;
3836                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3837                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3838                                 set_bit(R5_Wantcompute,
3839                                         &sh->dev[sh->pd_idx].flags);
3840                                 sh->ops.target = sh->pd_idx;
3841                                 sh->ops.target2 = -1;
3842                                 s->uptodate++;
3843                         }
3844                 }
3845                 break;
3846         case check_state_compute_run:
3847                 break;
3848         default:
3849                 pr_err("%s: unknown check_state: %d sector: %llu\n",
3850                        __func__, sh->check_state,
3851                        (unsigned long long) sh->sector);
3852                 BUG();
3853         }
3854 }
3855
3856 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3857                                   struct stripe_head_state *s,
3858                                   int disks)
3859 {
3860         int pd_idx = sh->pd_idx;
3861         int qd_idx = sh->qd_idx;
3862         struct r5dev *dev;
3863
3864         BUG_ON(sh->batch_head);
3865         set_bit(STRIPE_HANDLE, &sh->state);
3866
3867         BUG_ON(s->failed > 2);
3868
3869         /* Want to check and possibly repair P and Q.
3870          * However there could be one 'failed' device, in which
3871          * case we can only check one of them, possibly using the
3872          * other to generate missing data
3873          */
3874
3875         switch (sh->check_state) {
3876         case check_state_idle:
3877                 /* start a new check operation if there are < 2 failures */
3878                 if (s->failed == s->q_failed) {
3879                         /* The only possible failed device holds Q, so it
3880                          * makes sense to check P (If anything else were failed,
3881                          * we would have used P to recreate it).
3882                          */
3883                         sh->check_state = check_state_run;
3884                 }
3885                 if (!s->q_failed && s->failed < 2) {
3886                         /* Q is not failed, and we didn't use it to generate
3887                          * anything, so it makes sense to check it
3888                          */
3889                         if (sh->check_state == check_state_run)
3890                                 sh->check_state = check_state_run_pq;
3891                         else
3892                                 sh->check_state = check_state_run_q;
3893                 }
3894
3895                 /* discard potentially stale zero_sum_result */
3896                 sh->ops.zero_sum_result = 0;
3897
3898                 if (sh->check_state == check_state_run) {
3899                         /* async_xor_zero_sum destroys the contents of P */
3900                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3901                         s->uptodate--;
3902                 }
3903                 if (sh->check_state >= check_state_run &&
3904                     sh->check_state <= check_state_run_pq) {
3905                         /* async_syndrome_zero_sum preserves P and Q, so
3906                          * no need to mark them !uptodate here
3907                          */
3908                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3909                         break;
3910                 }
3911
3912                 /* we have 2-disk failure */
3913                 BUG_ON(s->failed != 2);
3914                 /* fall through */
3915         case check_state_compute_result:
3916                 sh->check_state = check_state_idle;
3917
3918                 /* check that a write has not made the stripe insync */
3919                 if (test_bit(STRIPE_INSYNC, &sh->state))
3920                         break;
3921
3922                 /* now write out any block on a failed drive,
3923                  * or P or Q if they were recomputed
3924                  */
3925                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3926                 if (s->failed == 2) {
3927                         dev = &sh->dev[s->failed_num[1]];
3928                         s->locked++;
3929                         set_bit(R5_LOCKED, &dev->flags);
3930                         set_bit(R5_Wantwrite, &dev->flags);
3931                 }
3932                 if (s->failed >= 1) {
3933                         dev = &sh->dev[s->failed_num[0]];
3934                         s->locked++;
3935                         set_bit(R5_LOCKED, &dev->flags);
3936                         set_bit(R5_Wantwrite, &dev->flags);
3937                 }
3938                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3939                         dev = &sh->dev[pd_idx];
3940                         s->locked++;
3941                         set_bit(R5_LOCKED, &dev->flags);
3942                         set_bit(R5_Wantwrite, &dev->flags);
3943                 }
3944                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3945                         dev = &sh->dev[qd_idx];
3946                         s->locked++;
3947                         set_bit(R5_LOCKED, &dev->flags);
3948                         set_bit(R5_Wantwrite, &dev->flags);
3949                 }
3950                 clear_bit(STRIPE_DEGRADED, &sh->state);
3951
3952                 set_bit(STRIPE_INSYNC, &sh->state);
3953                 break;
3954         case check_state_run:
3955         case check_state_run_q:
3956         case check_state_run_pq:
3957                 break; /* we will be called again upon completion */
3958         case check_state_check_result:
3959                 sh->check_state = check_state_idle;
3960
3961                 /* handle a successful check operation, if parity is correct
3962                  * we are done.  Otherwise update the mismatch count and repair
3963                  * parity if !MD_RECOVERY_CHECK
3964                  */
3965                 if (sh->ops.zero_sum_result == 0) {
3966                         /* both parities are correct */
3967                         if (!s->failed)
3968                                 set_bit(STRIPE_INSYNC, &sh->state);
3969                         else {
3970                                 /* in contrast to the raid5 case we can validate
3971                                  * parity, but still have a failure to write
3972                                  * back
3973                                  */
3974                                 sh->check_state = check_state_compute_result;
3975                                 /* Returning at this point means that we may go
3976                                  * off and bring p and/or q uptodate again so
3977                                  * we make sure to check zero_sum_result again
3978                                  * to verify if p or q need writeback
3979                                  */
3980                         }
3981                 } else {
3982                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3983                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3984                                 /* don't try to repair!! */
3985                                 set_bit(STRIPE_INSYNC, &sh->state);
3986                         else {
3987                                 int *target = &sh->ops.target;
3988
3989                                 sh->ops.target = -1;
3990                                 sh->ops.target2 = -1;
3991                                 sh->check_state = check_state_compute_run;
3992                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3993                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3994                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3995                                         set_bit(R5_Wantcompute,
3996                                                 &sh->dev[pd_idx].flags);
3997                                         *target = pd_idx;
3998                                         target = &sh->ops.target2;
3999                                         s->uptodate++;
4000                                 }
4001                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4002                                         set_bit(R5_Wantcompute,
4003                                                 &sh->dev[qd_idx].flags);
4004                                         *target = qd_idx;
4005                                         s->uptodate++;
4006                                 }
4007                         }
4008                 }
4009                 break;
4010         case check_state_compute_run:
4011                 break;
4012         default:
4013                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4014                         __func__, sh->check_state,
4015                         (unsigned long long) sh->sector);
4016                 BUG();
4017         }
4018 }
4019
4020 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4021 {
4022         int i;
4023
4024         /* We have read all the blocks in this stripe and now we need to
4025          * copy some of them into a target stripe for expand.
4026          */
4027         struct dma_async_tx_descriptor *tx = NULL;
4028         BUG_ON(sh->batch_head);
4029         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4030         for (i = 0; i < sh->disks; i++)
4031                 if (i != sh->pd_idx && i != sh->qd_idx) {
4032                         int dd_idx, j;
4033                         struct stripe_head *sh2;
4034                         struct async_submit_ctl submit;
4035
4036                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4037                         sector_t s = raid5_compute_sector(conf, bn, 0,
4038                                                           &dd_idx, NULL);
4039                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4040                         if (sh2 == NULL)
4041                                 /* so far only the early blocks of this stripe
4042                                  * have been requested.  When later blocks
4043                                  * get requested, we will try again
4044                                  */
4045                                 continue;
4046                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4047                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4048                                 /* must have already done this block */
4049                                 raid5_release_stripe(sh2);
4050                                 continue;
4051                         }
4052
4053                         /* place all the copies on one channel */
4054                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4055                         tx = async_memcpy(sh2->dev[dd_idx].page,
4056                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
4057                                           &submit);
4058
4059                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4060                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4061                         for (j = 0; j < conf->raid_disks; j++)
4062                                 if (j != sh2->pd_idx &&
4063                                     j != sh2->qd_idx &&
4064                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4065                                         break;
4066                         if (j == conf->raid_disks) {
4067                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4068                                 set_bit(STRIPE_HANDLE, &sh2->state);
4069                         }
4070                         raid5_release_stripe(sh2);
4071
4072                 }
4073         /* done submitting copies, wait for them to complete */
4074         async_tx_quiesce(&tx);
4075 }
4076
4077 /*
4078  * handle_stripe - do things to a stripe.
4079  *
4080  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4081  * state of various bits to see what needs to be done.
4082  * Possible results:
4083  *    return some read requests which now have data
4084  *    return some write requests which are safely on storage
4085  *    schedule a read on some buffers
4086  *    schedule a write of some buffers
4087  *    return confirmation of parity correctness
4088  *
4089  */
4090
4091 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4092 {
4093         struct r5conf *conf = sh->raid_conf;
4094         int disks = sh->disks;
4095         struct r5dev *dev;
4096         int i;
4097         int do_recovery = 0;
4098
4099         memset(s, 0, sizeof(*s));
4100
4101         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4102         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4103         s->failed_num[0] = -1;
4104         s->failed_num[1] = -1;
4105         s->log_failed = r5l_log_disk_error(conf);
4106
4107         /* Now to look around and see what can be done */
4108         rcu_read_lock();
4109         for (i=disks; i--; ) {
4110                 struct md_rdev *rdev;
4111                 sector_t first_bad;
4112                 int bad_sectors;
4113                 int is_bad = 0;
4114
4115                 dev = &sh->dev[i];
4116
4117                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4118                          i, dev->flags,
4119                          dev->toread, dev->towrite, dev->written);
4120                 /* maybe we can reply to a read
4121                  *
4122                  * new wantfill requests are only permitted while
4123                  * ops_complete_biofill is guaranteed to be inactive
4124                  */
4125                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4126                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4127                         set_bit(R5_Wantfill, &dev->flags);
4128
4129                 /* now count some things */
4130                 if (test_bit(R5_LOCKED, &dev->flags))
4131                         s->locked++;
4132                 if (test_bit(R5_UPTODATE, &dev->flags))
4133                         s->uptodate++;
4134                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4135                         s->compute++;
4136                         BUG_ON(s->compute > 2);
4137                 }
4138
4139                 if (test_bit(R5_Wantfill, &dev->flags))
4140                         s->to_fill++;
4141                 else if (dev->toread)
4142                         s->to_read++;
4143                 if (dev->towrite) {
4144                         s->to_write++;
4145                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4146                                 s->non_overwrite++;
4147                 }
4148                 if (dev->written)
4149                         s->written++;
4150                 /* Prefer to use the replacement for reads, but only
4151                  * if it is recovered enough and has no bad blocks.
4152                  */
4153                 rdev = rcu_dereference(conf->disks[i].replacement);
4154                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4155                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4156                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4157                                  &first_bad, &bad_sectors))
4158                         set_bit(R5_ReadRepl, &dev->flags);
4159                 else {
4160                         if (rdev && !test_bit(Faulty, &rdev->flags))
4161                                 set_bit(R5_NeedReplace, &dev->flags);
4162                         else
4163                                 clear_bit(R5_NeedReplace, &dev->flags);
4164                         rdev = rcu_dereference(conf->disks[i].rdev);
4165                         clear_bit(R5_ReadRepl, &dev->flags);
4166                 }
4167                 if (rdev && test_bit(Faulty, &rdev->flags))
4168                         rdev = NULL;
4169                 if (rdev) {
4170                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4171                                              &first_bad, &bad_sectors);
4172                         if (s->blocked_rdev == NULL
4173                             && (test_bit(Blocked, &rdev->flags)
4174                                 || is_bad < 0)) {
4175                                 if (is_bad < 0)
4176                                         set_bit(BlockedBadBlocks,
4177                                                 &rdev->flags);
4178                                 s->blocked_rdev = rdev;
4179                                 atomic_inc(&rdev->nr_pending);
4180                         }
4181                 }
4182                 clear_bit(R5_Insync, &dev->flags);
4183                 if (!rdev)
4184                         /* Not in-sync */;
4185                 else if (is_bad) {
4186                         /* also not in-sync */
4187                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4188                             test_bit(R5_UPTODATE, &dev->flags)) {
4189                                 /* treat as in-sync, but with a read error
4190                                  * which we can now try to correct
4191                                  */
4192                                 set_bit(R5_Insync, &dev->flags);
4193                                 set_bit(R5_ReadError, &dev->flags);
4194                         }
4195                 } else if (test_bit(In_sync, &rdev->flags))
4196                         set_bit(R5_Insync, &dev->flags);
4197                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4198                         /* in sync if before recovery_offset */
4199                         set_bit(R5_Insync, &dev->flags);
4200                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4201                          test_bit(R5_Expanded, &dev->flags))
4202                         /* If we've reshaped into here, we assume it is Insync.
4203                          * We will shortly update recovery_offset to make
4204                          * it official.
4205                          */
4206                         set_bit(R5_Insync, &dev->flags);
4207
4208                 if (test_bit(R5_WriteError, &dev->flags)) {
4209                         /* This flag does not apply to '.replacement'
4210                          * only to .rdev, so make sure to check that*/
4211                         struct md_rdev *rdev2 = rcu_dereference(
4212                                 conf->disks[i].rdev);
4213                         if (rdev2 == rdev)
4214                                 clear_bit(R5_Insync, &dev->flags);
4215                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4216                                 s->handle_bad_blocks = 1;
4217                                 atomic_inc(&rdev2->nr_pending);
4218                         } else
4219                                 clear_bit(R5_WriteError, &dev->flags);
4220                 }
4221                 if (test_bit(R5_MadeGood, &dev->flags)) {
4222                         /* This flag does not apply to '.replacement'
4223                          * only to .rdev, so make sure to check that*/
4224                         struct md_rdev *rdev2 = rcu_dereference(
4225                                 conf->disks[i].rdev);
4226                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4227                                 s->handle_bad_blocks = 1;
4228                                 atomic_inc(&rdev2->nr_pending);
4229                         } else
4230                                 clear_bit(R5_MadeGood, &dev->flags);
4231                 }
4232                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4233                         struct md_rdev *rdev2 = rcu_dereference(
4234                                 conf->disks[i].replacement);
4235                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4236                                 s->handle_bad_blocks = 1;
4237                                 atomic_inc(&rdev2->nr_pending);
4238                         } else
4239                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4240                 }
4241                 if (!test_bit(R5_Insync, &dev->flags)) {
4242                         /* The ReadError flag will just be confusing now */
4243                         clear_bit(R5_ReadError, &dev->flags);
4244                         clear_bit(R5_ReWrite, &dev->flags);
4245                 }
4246                 if (test_bit(R5_ReadError, &dev->flags))
4247                         clear_bit(R5_Insync, &dev->flags);
4248                 if (!test_bit(R5_Insync, &dev->flags)) {
4249                         if (s->failed < 2)
4250                                 s->failed_num[s->failed] = i;
4251                         s->failed++;
4252                         if (rdev && !test_bit(Faulty, &rdev->flags))
4253                                 do_recovery = 1;
4254                 }
4255
4256                 if (test_bit(R5_InJournal, &dev->flags))
4257                         s->injournal++;
4258                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4259                         s->just_cached++;
4260         }
4261         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4262                 /* If there is a failed device being replaced,
4263                  *     we must be recovering.
4264                  * else if we are after recovery_cp, we must be syncing
4265                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4266                  * else we can only be replacing
4267                  * sync and recovery both need to read all devices, and so
4268                  * use the same flag.
4269                  */
4270                 if (do_recovery ||
4271                     sh->sector >= conf->mddev->recovery_cp ||
4272                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4273                         s->syncing = 1;
4274                 else
4275                         s->replacing = 1;
4276         }
4277         rcu_read_unlock();
4278 }
4279
4280 static int clear_batch_ready(struct stripe_head *sh)
4281 {
4282         /* Return '1' if this is a member of batch, or
4283          * '0' if it is a lone stripe or a head which can now be
4284          * handled.
4285          */
4286         struct stripe_head *tmp;
4287         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4288                 return (sh->batch_head && sh->batch_head != sh);
4289         spin_lock(&sh->stripe_lock);
4290         if (!sh->batch_head) {
4291                 spin_unlock(&sh->stripe_lock);
4292                 return 0;
4293         }
4294
4295         /*
4296          * this stripe could be added to a batch list before we check
4297          * BATCH_READY, skips it
4298          */
4299         if (sh->batch_head != sh) {
4300                 spin_unlock(&sh->stripe_lock);
4301                 return 1;
4302         }
4303         spin_lock(&sh->batch_lock);
4304         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4305                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4306         spin_unlock(&sh->batch_lock);
4307         spin_unlock(&sh->stripe_lock);
4308
4309         /*
4310          * BATCH_READY is cleared, no new stripes can be added.
4311          * batch_list can be accessed without lock
4312          */
4313         return 0;
4314 }
4315
4316 static void break_stripe_batch_list(struct stripe_head *head_sh,
4317                                     unsigned long handle_flags)
4318 {
4319         struct stripe_head *sh, *next;
4320         int i;
4321         int do_wakeup = 0;
4322
4323         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4324
4325                 list_del_init(&sh->batch_list);
4326
4327                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4328                                           (1 << STRIPE_SYNCING) |
4329                                           (1 << STRIPE_REPLACED) |
4330                                           (1 << STRIPE_DELAYED) |
4331                                           (1 << STRIPE_BIT_DELAY) |
4332                                           (1 << STRIPE_FULL_WRITE) |
4333                                           (1 << STRIPE_BIOFILL_RUN) |
4334                                           (1 << STRIPE_COMPUTE_RUN)  |
4335                                           (1 << STRIPE_OPS_REQ_PENDING) |
4336                                           (1 << STRIPE_DISCARD) |
4337                                           (1 << STRIPE_BATCH_READY) |
4338                                           (1 << STRIPE_BATCH_ERR) |
4339                                           (1 << STRIPE_BITMAP_PENDING)),
4340                         "stripe state: %lx\n", sh->state);
4341                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4342                                               (1 << STRIPE_REPLACED)),
4343                         "head stripe state: %lx\n", head_sh->state);
4344
4345                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4346                                             (1 << STRIPE_PREREAD_ACTIVE) |
4347                                             (1 << STRIPE_DEGRADED)),
4348                               head_sh->state & (1 << STRIPE_INSYNC));
4349
4350                 sh->check_state = head_sh->check_state;
4351                 sh->reconstruct_state = head_sh->reconstruct_state;
4352                 for (i = 0; i < sh->disks; i++) {
4353                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4354                                 do_wakeup = 1;
4355                         sh->dev[i].flags = head_sh->dev[i].flags &
4356                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4357                 }
4358                 spin_lock_irq(&sh->stripe_lock);
4359                 sh->batch_head = NULL;
4360                 spin_unlock_irq(&sh->stripe_lock);
4361                 if (handle_flags == 0 ||
4362                     sh->state & handle_flags)
4363                         set_bit(STRIPE_HANDLE, &sh->state);
4364                 raid5_release_stripe(sh);
4365         }
4366         spin_lock_irq(&head_sh->stripe_lock);
4367         head_sh->batch_head = NULL;
4368         spin_unlock_irq(&head_sh->stripe_lock);
4369         for (i = 0; i < head_sh->disks; i++)
4370                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4371                         do_wakeup = 1;
4372         if (head_sh->state & handle_flags)
4373                 set_bit(STRIPE_HANDLE, &head_sh->state);
4374
4375         if (do_wakeup)
4376                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4377 }
4378
4379 static void handle_stripe(struct stripe_head *sh)
4380 {
4381         struct stripe_head_state s;
4382         struct r5conf *conf = sh->raid_conf;
4383         int i;
4384         int prexor;
4385         int disks = sh->disks;
4386         struct r5dev *pdev, *qdev;
4387
4388         clear_bit(STRIPE_HANDLE, &sh->state);
4389         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4390                 /* already being handled, ensure it gets handled
4391                  * again when current action finishes */
4392                 set_bit(STRIPE_HANDLE, &sh->state);
4393                 return;
4394         }
4395
4396         if (clear_batch_ready(sh) ) {
4397                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4398                 return;
4399         }
4400
4401         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4402                 break_stripe_batch_list(sh, 0);
4403
4404         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4405                 spin_lock(&sh->stripe_lock);
4406                 /* Cannot process 'sync' concurrently with 'discard' */
4407                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4408                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4409                         set_bit(STRIPE_SYNCING, &sh->state);
4410                         clear_bit(STRIPE_INSYNC, &sh->state);
4411                         clear_bit(STRIPE_REPLACED, &sh->state);
4412                 }
4413                 spin_unlock(&sh->stripe_lock);
4414         }
4415         clear_bit(STRIPE_DELAYED, &sh->state);
4416
4417         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4418                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4419                (unsigned long long)sh->sector, sh->state,
4420                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4421                sh->check_state, sh->reconstruct_state);
4422
4423         analyse_stripe(sh, &s);
4424
4425         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4426                 goto finish;
4427
4428         if (s.handle_bad_blocks) {
4429                 set_bit(STRIPE_HANDLE, &sh->state);
4430                 goto finish;
4431         }
4432
4433         if (unlikely(s.blocked_rdev)) {
4434                 if (s.syncing || s.expanding || s.expanded ||
4435                     s.replacing || s.to_write || s.written) {
4436                         set_bit(STRIPE_HANDLE, &sh->state);
4437                         goto finish;
4438                 }
4439                 /* There is nothing for the blocked_rdev to block */
4440                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4441                 s.blocked_rdev = NULL;
4442         }
4443
4444         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4445                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4446                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4447         }
4448
4449         pr_debug("locked=%d uptodate=%d to_read=%d"
4450                " to_write=%d failed=%d failed_num=%d,%d\n",
4451                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4452                s.failed_num[0], s.failed_num[1]);
4453         /* check if the array has lost more than max_degraded devices and,
4454          * if so, some requests might need to be failed.
4455          */
4456         if (s.failed > conf->max_degraded || s.log_failed) {
4457                 sh->check_state = 0;
4458                 sh->reconstruct_state = 0;
4459                 break_stripe_batch_list(sh, 0);
4460                 if (s.to_read+s.to_write+s.written)
4461                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4462                 if (s.syncing + s.replacing)
4463                         handle_failed_sync(conf, sh, &s);
4464         }
4465
4466         /* Now we check to see if any write operations have recently
4467          * completed
4468          */
4469         prexor = 0;
4470         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4471                 prexor = 1;
4472         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4473             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4474                 sh->reconstruct_state = reconstruct_state_idle;
4475
4476                 /* All the 'written' buffers and the parity block are ready to
4477                  * be written back to disk
4478                  */
4479                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4480                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4481                 BUG_ON(sh->qd_idx >= 0 &&
4482                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4483                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4484                 for (i = disks; i--; ) {
4485                         struct r5dev *dev = &sh->dev[i];
4486                         if (test_bit(R5_LOCKED, &dev->flags) &&
4487                                 (i == sh->pd_idx || i == sh->qd_idx ||
4488                                  dev->written || test_bit(R5_InJournal,
4489                                                           &dev->flags))) {
4490                                 pr_debug("Writing block %d\n", i);
4491                                 set_bit(R5_Wantwrite, &dev->flags);
4492                                 if (prexor)
4493                                         continue;
4494                                 if (s.failed > 1)
4495                                         continue;
4496                                 if (!test_bit(R5_Insync, &dev->flags) ||
4497                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4498                                      s.failed == 0))
4499                                         set_bit(STRIPE_INSYNC, &sh->state);
4500                         }
4501                 }
4502                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4503                         s.dec_preread_active = 1;
4504         }
4505
4506         /*
4507          * might be able to return some write requests if the parity blocks
4508          * are safe, or on a failed drive
4509          */
4510         pdev = &sh->dev[sh->pd_idx];
4511         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4512                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4513         qdev = &sh->dev[sh->qd_idx];
4514         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4515                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4516                 || conf->level < 6;
4517
4518         if (s.written &&
4519             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4520                              && !test_bit(R5_LOCKED, &pdev->flags)
4521                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4522                                  test_bit(R5_Discard, &pdev->flags))))) &&
4523             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4524                              && !test_bit(R5_LOCKED, &qdev->flags)
4525                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4526                                  test_bit(R5_Discard, &qdev->flags))))))
4527                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4528
4529         if (s.just_cached)
4530                 r5c_handle_cached_data_endio(conf, sh, disks, &s.return_bi);
4531         r5l_stripe_write_finished(sh);
4532
4533         /* Now we might consider reading some blocks, either to check/generate
4534          * parity, or to satisfy requests
4535          * or to load a block that is being partially written.
4536          */
4537         if (s.to_read || s.non_overwrite
4538             || (conf->level == 6 && s.to_write && s.failed)
4539             || (s.syncing && (s.uptodate + s.compute < disks))
4540             || s.replacing
4541             || s.expanding)
4542                 handle_stripe_fill(sh, &s, disks);
4543
4544         /*
4545          * When the stripe finishes full journal write cycle (write to journal
4546          * and raid disk), this is the clean up procedure so it is ready for
4547          * next operation.
4548          */
4549         r5c_finish_stripe_write_out(conf, sh, &s);
4550
4551         /*
4552          * Now to consider new write requests, cache write back and what else,
4553          * if anything should be read.  We do not handle new writes when:
4554          * 1/ A 'write' operation (copy+xor) is already in flight.
4555          * 2/ A 'check' operation is in flight, as it may clobber the parity
4556          *    block.
4557          * 3/ A r5c cache log write is in flight.
4558          */
4559
4560         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4561                 if (!r5c_is_writeback(conf->log)) {
4562                         if (s.to_write)
4563                                 handle_stripe_dirtying(conf, sh, &s, disks);
4564                 } else { /* write back cache */
4565                         int ret = 0;
4566
4567                         /* First, try handle writes in caching phase */
4568                         if (s.to_write)
4569                                 ret = r5c_try_caching_write(conf, sh, &s,
4570                                                             disks);
4571                         /*
4572                          * If caching phase failed: ret == -EAGAIN
4573                          *    OR
4574                          * stripe under reclaim: !caching && injournal
4575                          *
4576                          * fall back to handle_stripe_dirtying()
4577                          */
4578                         if (ret == -EAGAIN ||
4579                             /* stripe under reclaim: !caching && injournal */
4580                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4581                              s.injournal > 0)) {
4582                                 ret = handle_stripe_dirtying(conf, sh, &s,
4583                                                              disks);
4584                                 if (ret == -EAGAIN)
4585                                         goto finish;
4586                         }
4587                 }
4588         }
4589
4590         /* maybe we need to check and possibly fix the parity for this stripe
4591          * Any reads will already have been scheduled, so we just see if enough
4592          * data is available.  The parity check is held off while parity
4593          * dependent operations are in flight.
4594          */
4595         if (sh->check_state ||
4596             (s.syncing && s.locked == 0 &&
4597              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4598              !test_bit(STRIPE_INSYNC, &sh->state))) {
4599                 if (conf->level == 6)
4600                         handle_parity_checks6(conf, sh, &s, disks);
4601                 else
4602                         handle_parity_checks5(conf, sh, &s, disks);
4603         }
4604
4605         if ((s.replacing || s.syncing) && s.locked == 0
4606             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4607             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4608                 /* Write out to replacement devices where possible */
4609                 for (i = 0; i < conf->raid_disks; i++)
4610                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4611                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4612                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4613                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4614                                 s.locked++;
4615                         }
4616                 if (s.replacing)
4617                         set_bit(STRIPE_INSYNC, &sh->state);
4618                 set_bit(STRIPE_REPLACED, &sh->state);
4619         }
4620         if ((s.syncing || s.replacing) && s.locked == 0 &&
4621             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4622             test_bit(STRIPE_INSYNC, &sh->state)) {
4623                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4624                 clear_bit(STRIPE_SYNCING, &sh->state);
4625                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4626                         wake_up(&conf->wait_for_overlap);
4627         }
4628
4629         /* If the failed drives are just a ReadError, then we might need
4630          * to progress the repair/check process
4631          */
4632         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4633                 for (i = 0; i < s.failed; i++) {
4634                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4635                         if (test_bit(R5_ReadError, &dev->flags)
4636                             && !test_bit(R5_LOCKED, &dev->flags)
4637                             && test_bit(R5_UPTODATE, &dev->flags)
4638                                 ) {
4639                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4640                                         set_bit(R5_Wantwrite, &dev->flags);
4641                                         set_bit(R5_ReWrite, &dev->flags);
4642                                         set_bit(R5_LOCKED, &dev->flags);
4643                                         s.locked++;
4644                                 } else {
4645                                         /* let's read it back */
4646                                         set_bit(R5_Wantread, &dev->flags);
4647                                         set_bit(R5_LOCKED, &dev->flags);
4648                                         s.locked++;
4649                                 }
4650                         }
4651                 }
4652
4653         /* Finish reconstruct operations initiated by the expansion process */
4654         if (sh->reconstruct_state == reconstruct_state_result) {
4655                 struct stripe_head *sh_src
4656                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4657                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4658                         /* sh cannot be written until sh_src has been read.
4659                          * so arrange for sh to be delayed a little
4660                          */
4661                         set_bit(STRIPE_DELAYED, &sh->state);
4662                         set_bit(STRIPE_HANDLE, &sh->state);
4663                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4664                                               &sh_src->state))
4665                                 atomic_inc(&conf->preread_active_stripes);
4666                         raid5_release_stripe(sh_src);
4667                         goto finish;
4668                 }
4669                 if (sh_src)
4670                         raid5_release_stripe(sh_src);
4671
4672                 sh->reconstruct_state = reconstruct_state_idle;
4673                 clear_bit(STRIPE_EXPANDING, &sh->state);
4674                 for (i = conf->raid_disks; i--; ) {
4675                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4676                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4677                         s.locked++;
4678                 }
4679         }
4680
4681         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4682             !sh->reconstruct_state) {
4683                 /* Need to write out all blocks after computing parity */
4684                 sh->disks = conf->raid_disks;
4685                 stripe_set_idx(sh->sector, conf, 0, sh);
4686                 schedule_reconstruction(sh, &s, 1, 1);
4687         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4688                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4689                 atomic_dec(&conf->reshape_stripes);
4690                 wake_up(&conf->wait_for_overlap);
4691                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4692         }
4693
4694         if (s.expanding && s.locked == 0 &&
4695             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4696                 handle_stripe_expansion(conf, sh);
4697
4698 finish:
4699         /* wait for this device to become unblocked */
4700         if (unlikely(s.blocked_rdev)) {
4701                 if (conf->mddev->external)
4702                         md_wait_for_blocked_rdev(s.blocked_rdev,
4703                                                  conf->mddev);
4704                 else
4705                         /* Internal metadata will immediately
4706                          * be written by raid5d, so we don't
4707                          * need to wait here.
4708                          */
4709                         rdev_dec_pending(s.blocked_rdev,
4710                                          conf->mddev);
4711         }
4712
4713         if (s.handle_bad_blocks)
4714                 for (i = disks; i--; ) {
4715                         struct md_rdev *rdev;
4716                         struct r5dev *dev = &sh->dev[i];
4717                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4718                                 /* We own a safe reference to the rdev */
4719                                 rdev = conf->disks[i].rdev;
4720                                 if (!rdev_set_badblocks(rdev, sh->sector,
4721                                                         STRIPE_SECTORS, 0))
4722                                         md_error(conf->mddev, rdev);
4723                                 rdev_dec_pending(rdev, conf->mddev);
4724                         }
4725                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4726                                 rdev = conf->disks[i].rdev;
4727                                 rdev_clear_badblocks(rdev, sh->sector,
4728                                                      STRIPE_SECTORS, 0);
4729                                 rdev_dec_pending(rdev, conf->mddev);
4730                         }
4731                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4732                                 rdev = conf->disks[i].replacement;
4733                                 if (!rdev)
4734                                         /* rdev have been moved down */
4735                                         rdev = conf->disks[i].rdev;
4736                                 rdev_clear_badblocks(rdev, sh->sector,
4737                                                      STRIPE_SECTORS, 0);
4738                                 rdev_dec_pending(rdev, conf->mddev);
4739                         }
4740                 }
4741
4742         if (s.ops_request)
4743                 raid_run_ops(sh, s.ops_request);
4744
4745         ops_run_io(sh, &s);
4746
4747         if (s.dec_preread_active) {
4748                 /* We delay this until after ops_run_io so that if make_request
4749                  * is waiting on a flush, it won't continue until the writes
4750                  * have actually been submitted.
4751                  */
4752                 atomic_dec(&conf->preread_active_stripes);
4753                 if (atomic_read(&conf->preread_active_stripes) <
4754                     IO_THRESHOLD)
4755                         md_wakeup_thread(conf->mddev->thread);
4756         }
4757
4758         if (!bio_list_empty(&s.return_bi)) {
4759                 if (test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4760                         spin_lock_irq(&conf->device_lock);
4761                         bio_list_merge(&conf->return_bi, &s.return_bi);
4762                         spin_unlock_irq(&conf->device_lock);
4763                         md_wakeup_thread(conf->mddev->thread);
4764                 } else
4765                         return_io(&s.return_bi);
4766         }
4767
4768         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4769 }
4770
4771 static void raid5_activate_delayed(struct r5conf *conf)
4772 {
4773         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4774                 while (!list_empty(&conf->delayed_list)) {
4775                         struct list_head *l = conf->delayed_list.next;
4776                         struct stripe_head *sh;
4777                         sh = list_entry(l, struct stripe_head, lru);
4778                         list_del_init(l);
4779                         clear_bit(STRIPE_DELAYED, &sh->state);
4780                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4781                                 atomic_inc(&conf->preread_active_stripes);
4782                         list_add_tail(&sh->lru, &conf->hold_list);
4783                         raid5_wakeup_stripe_thread(sh);
4784                 }
4785         }
4786 }
4787
4788 static void activate_bit_delay(struct r5conf *conf,
4789         struct list_head *temp_inactive_list)
4790 {
4791         /* device_lock is held */
4792         struct list_head head;
4793         list_add(&head, &conf->bitmap_list);
4794         list_del_init(&conf->bitmap_list);
4795         while (!list_empty(&head)) {
4796                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4797                 int hash;
4798                 list_del_init(&sh->lru);
4799                 atomic_inc(&sh->count);
4800                 hash = sh->hash_lock_index;
4801                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4802         }
4803 }
4804
4805 static int raid5_congested(struct mddev *mddev, int bits)
4806 {
4807         struct r5conf *conf = mddev->private;
4808
4809         /* No difference between reads and writes.  Just check
4810          * how busy the stripe_cache is
4811          */
4812
4813         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4814                 return 1;
4815
4816         /* Also checks whether there is pressure on r5cache log space */
4817         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
4818                 return 1;
4819         if (conf->quiesce)
4820                 return 1;
4821         if (atomic_read(&conf->empty_inactive_list_nr))
4822                 return 1;
4823
4824         return 0;
4825 }
4826
4827 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4828 {
4829         struct r5conf *conf = mddev->private;
4830         sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4831         unsigned int chunk_sectors;
4832         unsigned int bio_sectors = bio_sectors(bio);
4833
4834         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4835         return  chunk_sectors >=
4836                 ((sector & (chunk_sectors - 1)) + bio_sectors);
4837 }
4838
4839 /*
4840  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
4841  *  later sampled by raid5d.
4842  */
4843 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4844 {
4845         unsigned long flags;
4846
4847         spin_lock_irqsave(&conf->device_lock, flags);
4848
4849         bi->bi_next = conf->retry_read_aligned_list;
4850         conf->retry_read_aligned_list = bi;
4851
4852         spin_unlock_irqrestore(&conf->device_lock, flags);
4853         md_wakeup_thread(conf->mddev->thread);
4854 }
4855
4856 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4857 {
4858         struct bio *bi;
4859
4860         bi = conf->retry_read_aligned;
4861         if (bi) {
4862                 conf->retry_read_aligned = NULL;
4863                 return bi;
4864         }
4865         bi = conf->retry_read_aligned_list;
4866         if(bi) {
4867                 conf->retry_read_aligned_list = bi->bi_next;
4868                 bi->bi_next = NULL;
4869                 /*
4870                  * this sets the active strip count to 1 and the processed
4871                  * strip count to zero (upper 8 bits)
4872                  */
4873                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4874         }
4875
4876         return bi;
4877 }
4878
4879 /*
4880  *  The "raid5_align_endio" should check if the read succeeded and if it
4881  *  did, call bio_endio on the original bio (having bio_put the new bio
4882  *  first).
4883  *  If the read failed..
4884  */
4885 static void raid5_align_endio(struct bio *bi)
4886 {
4887         struct bio* raid_bi  = bi->bi_private;
4888         struct mddev *mddev;
4889         struct r5conf *conf;
4890         struct md_rdev *rdev;
4891         int error = bi->bi_error;
4892
4893         bio_put(bi);
4894
4895         rdev = (void*)raid_bi->bi_next;
4896         raid_bi->bi_next = NULL;
4897         mddev = rdev->mddev;
4898         conf = mddev->private;
4899
4900         rdev_dec_pending(rdev, conf->mddev);
4901
4902         if (!error) {
4903                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4904                                          raid_bi, 0);
4905                 bio_endio(raid_bi);
4906                 if (atomic_dec_and_test(&conf->active_aligned_reads))
4907                         wake_up(&conf->wait_for_quiescent);
4908                 return;
4909         }
4910
4911         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4912
4913         add_bio_to_retry(raid_bi, conf);
4914 }
4915
4916 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4917 {
4918         struct r5conf *conf = mddev->private;
4919         int dd_idx;
4920         struct bio* align_bi;
4921         struct md_rdev *rdev;
4922         sector_t end_sector;
4923
4924         if (!in_chunk_boundary(mddev, raid_bio)) {
4925                 pr_debug("%s: non aligned\n", __func__);
4926                 return 0;
4927         }
4928         /*
4929          * use bio_clone_mddev to make a copy of the bio
4930          */
4931         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4932         if (!align_bi)
4933                 return 0;
4934         /*
4935          *   set bi_end_io to a new function, and set bi_private to the
4936          *     original bio.
4937          */
4938         align_bi->bi_end_io  = raid5_align_endio;
4939         align_bi->bi_private = raid_bio;
4940         /*
4941          *      compute position
4942          */
4943         align_bi->bi_iter.bi_sector =
4944                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4945                                      0, &dd_idx, NULL);
4946
4947         end_sector = bio_end_sector(align_bi);
4948         rcu_read_lock();
4949         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4950         if (!rdev || test_bit(Faulty, &rdev->flags) ||
4951             rdev->recovery_offset < end_sector) {
4952                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4953                 if (rdev &&
4954                     (test_bit(Faulty, &rdev->flags) ||
4955                     !(test_bit(In_sync, &rdev->flags) ||
4956                       rdev->recovery_offset >= end_sector)))
4957                         rdev = NULL;
4958         }
4959         if (rdev) {
4960                 sector_t first_bad;
4961                 int bad_sectors;
4962
4963                 atomic_inc(&rdev->nr_pending);
4964                 rcu_read_unlock();
4965                 raid_bio->bi_next = (void*)rdev;
4966                 align_bi->bi_bdev =  rdev->bdev;
4967                 bio_clear_flag(align_bi, BIO_SEG_VALID);
4968
4969                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4970                                 bio_sectors(align_bi),
4971                                 &first_bad, &bad_sectors)) {
4972                         bio_put(align_bi);
4973                         rdev_dec_pending(rdev, mddev);
4974                         return 0;
4975                 }
4976
4977                 /* No reshape active, so we can trust rdev->data_offset */
4978                 align_bi->bi_iter.bi_sector += rdev->data_offset;
4979
4980                 spin_lock_irq(&conf->device_lock);
4981                 wait_event_lock_irq(conf->wait_for_quiescent,
4982                                     conf->quiesce == 0,
4983                                     conf->device_lock);
4984                 atomic_inc(&conf->active_aligned_reads);
4985                 spin_unlock_irq(&conf->device_lock);
4986
4987                 if (mddev->gendisk)
4988                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4989                                               align_bi, disk_devt(mddev->gendisk),
4990                                               raid_bio->bi_iter.bi_sector);
4991                 generic_make_request(align_bi);
4992                 return 1;
4993         } else {
4994                 rcu_read_unlock();
4995                 bio_put(align_bi);
4996                 return 0;
4997         }
4998 }
4999
5000 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5001 {
5002         struct bio *split;
5003
5004         do {
5005                 sector_t sector = raid_bio->bi_iter.bi_sector;
5006                 unsigned chunk_sects = mddev->chunk_sectors;
5007                 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5008
5009                 if (sectors < bio_sectors(raid_bio)) {
5010                         split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
5011                         bio_chain(split, raid_bio);
5012                 } else
5013                         split = raid_bio;
5014
5015                 if (!raid5_read_one_chunk(mddev, split)) {
5016                         if (split != raid_bio)
5017                                 generic_make_request(raid_bio);
5018                         return split;
5019                 }
5020         } while (split != raid_bio);
5021
5022         return NULL;
5023 }
5024
5025 /* __get_priority_stripe - get the next stripe to process
5026  *
5027  * Full stripe writes are allowed to pass preread active stripes up until
5028  * the bypass_threshold is exceeded.  In general the bypass_count
5029  * increments when the handle_list is handled before the hold_list; however, it
5030  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5031  * stripe with in flight i/o.  The bypass_count will be reset when the
5032  * head of the hold_list has changed, i.e. the head was promoted to the
5033  * handle_list.
5034  */
5035 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5036 {
5037         struct stripe_head *sh = NULL, *tmp;
5038         struct list_head *handle_list = NULL;
5039         struct r5worker_group *wg = NULL;
5040
5041         if (conf->worker_cnt_per_group == 0) {
5042                 handle_list = &conf->handle_list;
5043         } else if (group != ANY_GROUP) {
5044                 handle_list = &conf->worker_groups[group].handle_list;
5045                 wg = &conf->worker_groups[group];
5046         } else {
5047                 int i;
5048                 for (i = 0; i < conf->group_cnt; i++) {
5049                         handle_list = &conf->worker_groups[i].handle_list;
5050                         wg = &conf->worker_groups[i];
5051                         if (!list_empty(handle_list))
5052                                 break;
5053                 }
5054         }
5055
5056         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5057                   __func__,
5058                   list_empty(handle_list) ? "empty" : "busy",
5059                   list_empty(&conf->hold_list) ? "empty" : "busy",
5060                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5061
5062         if (!list_empty(handle_list)) {
5063                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5064
5065                 if (list_empty(&conf->hold_list))
5066                         conf->bypass_count = 0;
5067                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5068                         if (conf->hold_list.next == conf->last_hold)
5069                                 conf->bypass_count++;
5070                         else {
5071                                 conf->last_hold = conf->hold_list.next;
5072                                 conf->bypass_count -= conf->bypass_threshold;
5073                                 if (conf->bypass_count < 0)
5074                                         conf->bypass_count = 0;
5075                         }
5076                 }
5077         } else if (!list_empty(&conf->hold_list) &&
5078                    ((conf->bypass_threshold &&
5079                      conf->bypass_count > conf->bypass_threshold) ||
5080                     atomic_read(&conf->pending_full_writes) == 0)) {
5081
5082                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5083                         if (conf->worker_cnt_per_group == 0 ||
5084                             group == ANY_GROUP ||
5085                             !cpu_online(tmp->cpu) ||
5086                             cpu_to_group(tmp->cpu) == group) {
5087                                 sh = tmp;
5088                                 break;
5089                         }
5090                 }
5091
5092                 if (sh) {
5093                         conf->bypass_count -= conf->bypass_threshold;
5094                         if (conf->bypass_count < 0)
5095                                 conf->bypass_count = 0;
5096                 }
5097                 wg = NULL;
5098         }
5099
5100         if (!sh)
5101                 return NULL;
5102
5103         if (wg) {
5104                 wg->stripes_cnt--;
5105                 sh->group = NULL;
5106         }
5107         list_del_init(&sh->lru);
5108         BUG_ON(atomic_inc_return(&sh->count) != 1);
5109         return sh;
5110 }
5111
5112 struct raid5_plug_cb {
5113         struct blk_plug_cb      cb;
5114         struct list_head        list;
5115         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5116 };
5117
5118 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5119 {
5120         struct raid5_plug_cb *cb = container_of(
5121                 blk_cb, struct raid5_plug_cb, cb);
5122         struct stripe_head *sh;
5123         struct mddev *mddev = cb->cb.data;
5124         struct r5conf *conf = mddev->private;
5125         int cnt = 0;
5126         int hash;
5127
5128         if (cb->list.next && !list_empty(&cb->list)) {
5129                 spin_lock_irq(&conf->device_lock);
5130                 while (!list_empty(&cb->list)) {
5131                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5132                         list_del_init(&sh->lru);
5133                         /*
5134                          * avoid race release_stripe_plug() sees
5135                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5136                          * is still in our list
5137                          */
5138                         smp_mb__before_atomic();
5139                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5140                         /*
5141                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5142                          * case, the count is always > 1 here
5143                          */
5144                         hash = sh->hash_lock_index;
5145                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5146                         cnt++;
5147                 }
5148                 spin_unlock_irq(&conf->device_lock);
5149         }
5150         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5151                                      NR_STRIPE_HASH_LOCKS);
5152         if (mddev->queue)
5153                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5154         kfree(cb);
5155 }
5156
5157 static void release_stripe_plug(struct mddev *mddev,
5158                                 struct stripe_head *sh)
5159 {
5160         struct blk_plug_cb *blk_cb = blk_check_plugged(
5161                 raid5_unplug, mddev,
5162                 sizeof(struct raid5_plug_cb));
5163         struct raid5_plug_cb *cb;
5164
5165         if (!blk_cb) {
5166                 raid5_release_stripe(sh);
5167                 return;
5168         }
5169
5170         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5171
5172         if (cb->list.next == NULL) {
5173                 int i;
5174                 INIT_LIST_HEAD(&cb->list);
5175                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5176                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5177         }
5178
5179         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5180                 list_add_tail(&sh->lru, &cb->list);
5181         else
5182                 raid5_release_stripe(sh);
5183 }
5184
5185 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5186 {
5187         struct r5conf *conf = mddev->private;
5188         sector_t logical_sector, last_sector;
5189         struct stripe_head *sh;
5190         int remaining;
5191         int stripe_sectors;
5192
5193         if (mddev->reshape_position != MaxSector)
5194                 /* Skip discard while reshape is happening */
5195                 return;
5196
5197         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5198         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5199
5200         bi->bi_next = NULL;
5201         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5202
5203         stripe_sectors = conf->chunk_sectors *
5204                 (conf->raid_disks - conf->max_degraded);
5205         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5206                                                stripe_sectors);
5207         sector_div(last_sector, stripe_sectors);
5208
5209         logical_sector *= conf->chunk_sectors;
5210         last_sector *= conf->chunk_sectors;
5211
5212         for (; logical_sector < last_sector;
5213              logical_sector += STRIPE_SECTORS) {
5214                 DEFINE_WAIT(w);
5215                 int d;
5216         again:
5217                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5218                 prepare_to_wait(&conf->wait_for_overlap, &w,
5219                                 TASK_UNINTERRUPTIBLE);
5220                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5221                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5222                         raid5_release_stripe(sh);
5223                         schedule();
5224                         goto again;
5225                 }
5226                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5227                 spin_lock_irq(&sh->stripe_lock);
5228                 for (d = 0; d < conf->raid_disks; d++) {
5229                         if (d == sh->pd_idx || d == sh->qd_idx)
5230                                 continue;
5231                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5232                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5233                                 spin_unlock_irq(&sh->stripe_lock);
5234                                 raid5_release_stripe(sh);
5235                                 schedule();
5236                                 goto again;
5237                         }
5238                 }
5239                 set_bit(STRIPE_DISCARD, &sh->state);
5240                 finish_wait(&conf->wait_for_overlap, &w);
5241                 sh->overwrite_disks = 0;
5242                 for (d = 0; d < conf->raid_disks; d++) {
5243                         if (d == sh->pd_idx || d == sh->qd_idx)
5244                                 continue;
5245                         sh->dev[d].towrite = bi;
5246                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5247                         raid5_inc_bi_active_stripes(bi);
5248                         sh->overwrite_disks++;
5249                 }
5250                 spin_unlock_irq(&sh->stripe_lock);
5251                 if (conf->mddev->bitmap) {
5252                         for (d = 0;
5253                              d < conf->raid_disks - conf->max_degraded;
5254                              d++)
5255                                 bitmap_startwrite(mddev->bitmap,
5256                                                   sh->sector,
5257                                                   STRIPE_SECTORS,
5258                                                   0);
5259                         sh->bm_seq = conf->seq_flush + 1;
5260                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5261                 }
5262
5263                 set_bit(STRIPE_HANDLE, &sh->state);
5264                 clear_bit(STRIPE_DELAYED, &sh->state);
5265                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5266                         atomic_inc(&conf->preread_active_stripes);
5267                 release_stripe_plug(mddev, sh);
5268         }
5269
5270         remaining = raid5_dec_bi_active_stripes(bi);
5271         if (remaining == 0) {
5272                 md_write_end(mddev);
5273                 bio_endio(bi);
5274         }
5275 }
5276
5277 static void raid5_make_request(struct mddev *mddev, struct bio * bi)
5278 {
5279         struct r5conf *conf = mddev->private;
5280         int dd_idx;
5281         sector_t new_sector;
5282         sector_t logical_sector, last_sector;
5283         struct stripe_head *sh;
5284         const int rw = bio_data_dir(bi);
5285         int remaining;
5286         DEFINE_WAIT(w);
5287         bool do_prepare;
5288         bool do_flush = false;
5289
5290         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5291                 int ret = r5l_handle_flush_request(conf->log, bi);
5292
5293                 if (ret == 0)
5294                         return;
5295                 if (ret == -ENODEV) {
5296                         md_flush_request(mddev, bi);
5297                         return;
5298                 }
5299                 /* ret == -EAGAIN, fallback */
5300                 /*
5301                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5302                  * we need to flush journal device
5303                  */
5304                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5305         }
5306
5307         md_write_start(mddev, bi);
5308
5309         /*
5310          * If array is degraded, better not do chunk aligned read because
5311          * later we might have to read it again in order to reconstruct
5312          * data on failed drives.
5313          */
5314         if (rw == READ && mddev->degraded == 0 &&
5315             !r5c_is_writeback(conf->log) &&
5316             mddev->reshape_position == MaxSector) {
5317                 bi = chunk_aligned_read(mddev, bi);
5318                 if (!bi)
5319                         return;
5320         }
5321
5322         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5323                 make_discard_request(mddev, bi);
5324                 return;
5325         }
5326
5327         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5328         last_sector = bio_end_sector(bi);
5329         bi->bi_next = NULL;
5330         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
5331
5332         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5333         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5334                 int previous;
5335                 int seq;
5336
5337                 do_prepare = false;
5338         retry:
5339                 seq = read_seqcount_begin(&conf->gen_lock);
5340                 previous = 0;
5341                 if (do_prepare)
5342                         prepare_to_wait(&conf->wait_for_overlap, &w,
5343                                 TASK_UNINTERRUPTIBLE);
5344                 if (unlikely(conf->reshape_progress != MaxSector)) {
5345                         /* spinlock is needed as reshape_progress may be
5346                          * 64bit on a 32bit platform, and so it might be
5347                          * possible to see a half-updated value
5348                          * Of course reshape_progress could change after
5349                          * the lock is dropped, so once we get a reference
5350                          * to the stripe that we think it is, we will have
5351                          * to check again.
5352                          */
5353                         spin_lock_irq(&conf->device_lock);
5354                         if (mddev->reshape_backwards
5355                             ? logical_sector < conf->reshape_progress
5356                             : logical_sector >= conf->reshape_progress) {
5357                                 previous = 1;
5358                         } else {
5359                                 if (mddev->reshape_backwards
5360                                     ? logical_sector < conf->reshape_safe
5361                                     : logical_sector >= conf->reshape_safe) {
5362                                         spin_unlock_irq(&conf->device_lock);
5363                                         schedule();
5364                                         do_prepare = true;
5365                                         goto retry;
5366                                 }
5367                         }
5368                         spin_unlock_irq(&conf->device_lock);
5369                 }
5370
5371                 new_sector = raid5_compute_sector(conf, logical_sector,
5372                                                   previous,
5373                                                   &dd_idx, NULL);
5374                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5375                         (unsigned long long)new_sector,
5376                         (unsigned long long)logical_sector);
5377
5378                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5379                                        (bi->bi_opf & REQ_RAHEAD), 0);
5380                 if (sh) {
5381                         if (unlikely(previous)) {
5382                                 /* expansion might have moved on while waiting for a
5383                                  * stripe, so we must do the range check again.
5384                                  * Expansion could still move past after this
5385                                  * test, but as we are holding a reference to
5386                                  * 'sh', we know that if that happens,
5387                                  *  STRIPE_EXPANDING will get set and the expansion
5388                                  * won't proceed until we finish with the stripe.
5389                                  */
5390                                 int must_retry = 0;
5391                                 spin_lock_irq(&conf->device_lock);
5392                                 if (mddev->reshape_backwards
5393                                     ? logical_sector >= conf->reshape_progress
5394                                     : logical_sector < conf->reshape_progress)
5395                                         /* mismatch, need to try again */
5396                                         must_retry = 1;
5397                                 spin_unlock_irq(&conf->device_lock);
5398                                 if (must_retry) {
5399                                         raid5_release_stripe(sh);
5400                                         schedule();
5401                                         do_prepare = true;
5402                                         goto retry;
5403                                 }
5404                         }
5405                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5406                                 /* Might have got the wrong stripe_head
5407                                  * by accident
5408                                  */
5409                                 raid5_release_stripe(sh);
5410                                 goto retry;
5411                         }
5412
5413                         if (rw == WRITE &&
5414                             logical_sector >= mddev->suspend_lo &&
5415                             logical_sector < mddev->suspend_hi) {
5416                                 raid5_release_stripe(sh);
5417                                 /* As the suspend_* range is controlled by
5418                                  * userspace, we want an interruptible
5419                                  * wait.
5420                                  */
5421                                 flush_signals(current);
5422                                 prepare_to_wait(&conf->wait_for_overlap,
5423                                                 &w, TASK_INTERRUPTIBLE);
5424                                 if (logical_sector >= mddev->suspend_lo &&
5425                                     logical_sector < mddev->suspend_hi) {
5426                                         schedule();
5427                                         do_prepare = true;
5428                                 }
5429                                 goto retry;
5430                         }
5431
5432                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5433                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5434                                 /* Stripe is busy expanding or
5435                                  * add failed due to overlap.  Flush everything
5436                                  * and wait a while
5437                                  */
5438                                 md_wakeup_thread(mddev->thread);
5439                                 raid5_release_stripe(sh);
5440                                 schedule();
5441                                 do_prepare = true;
5442                                 goto retry;
5443                         }
5444                         if (do_flush) {
5445                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5446                                 /* we only need flush for one stripe */
5447                                 do_flush = false;
5448                         }
5449
5450                         set_bit(STRIPE_HANDLE, &sh->state);
5451                         clear_bit(STRIPE_DELAYED, &sh->state);
5452                         if ((!sh->batch_head || sh == sh->batch_head) &&
5453                             (bi->bi_opf & REQ_SYNC) &&
5454                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5455                                 atomic_inc(&conf->preread_active_stripes);
5456                         release_stripe_plug(mddev, sh);
5457                 } else {
5458                         /* cannot get stripe for read-ahead, just give-up */
5459                         bi->bi_error = -EIO;
5460                         break;
5461                 }
5462         }
5463         finish_wait(&conf->wait_for_overlap, &w);
5464
5465         remaining = raid5_dec_bi_active_stripes(bi);
5466         if (remaining == 0) {
5467
5468                 if ( rw == WRITE )
5469                         md_write_end(mddev);
5470
5471                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5472                                          bi, 0);
5473                 bio_endio(bi);
5474         }
5475 }
5476
5477 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5478
5479 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5480 {
5481         /* reshaping is quite different to recovery/resync so it is
5482          * handled quite separately ... here.
5483          *
5484          * On each call to sync_request, we gather one chunk worth of
5485          * destination stripes and flag them as expanding.
5486          * Then we find all the source stripes and request reads.
5487          * As the reads complete, handle_stripe will copy the data
5488          * into the destination stripe and release that stripe.
5489          */
5490         struct r5conf *conf = mddev->private;
5491         struct stripe_head *sh;
5492         sector_t first_sector, last_sector;
5493         int raid_disks = conf->previous_raid_disks;
5494         int data_disks = raid_disks - conf->max_degraded;
5495         int new_data_disks = conf->raid_disks - conf->max_degraded;
5496         int i;
5497         int dd_idx;
5498         sector_t writepos, readpos, safepos;
5499         sector_t stripe_addr;
5500         int reshape_sectors;
5501         struct list_head stripes;
5502         sector_t retn;
5503
5504         if (sector_nr == 0) {
5505                 /* If restarting in the middle, skip the initial sectors */
5506                 if (mddev->reshape_backwards &&
5507                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5508                         sector_nr = raid5_size(mddev, 0, 0)
5509                                 - conf->reshape_progress;
5510                 } else if (mddev->reshape_backwards &&
5511                            conf->reshape_progress == MaxSector) {
5512                         /* shouldn't happen, but just in case, finish up.*/
5513                         sector_nr = MaxSector;
5514                 } else if (!mddev->reshape_backwards &&
5515                            conf->reshape_progress > 0)
5516                         sector_nr = conf->reshape_progress;
5517                 sector_div(sector_nr, new_data_disks);
5518                 if (sector_nr) {
5519                         mddev->curr_resync_completed = sector_nr;
5520                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5521                         *skipped = 1;
5522                         retn = sector_nr;
5523                         goto finish;
5524                 }
5525         }
5526
5527         /* We need to process a full chunk at a time.
5528          * If old and new chunk sizes differ, we need to process the
5529          * largest of these
5530          */
5531
5532         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5533
5534         /* We update the metadata at least every 10 seconds, or when
5535          * the data about to be copied would over-write the source of
5536          * the data at the front of the range.  i.e. one new_stripe
5537          * along from reshape_progress new_maps to after where
5538          * reshape_safe old_maps to
5539          */
5540         writepos = conf->reshape_progress;
5541         sector_div(writepos, new_data_disks);
5542         readpos = conf->reshape_progress;
5543         sector_div(readpos, data_disks);
5544         safepos = conf->reshape_safe;
5545         sector_div(safepos, data_disks);
5546         if (mddev->reshape_backwards) {
5547                 BUG_ON(writepos < reshape_sectors);
5548                 writepos -= reshape_sectors;
5549                 readpos += reshape_sectors;
5550                 safepos += reshape_sectors;
5551         } else {
5552                 writepos += reshape_sectors;
5553                 /* readpos and safepos are worst-case calculations.
5554                  * A negative number is overly pessimistic, and causes
5555                  * obvious problems for unsigned storage.  So clip to 0.
5556                  */
5557                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5558                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5559         }
5560
5561         /* Having calculated the 'writepos' possibly use it
5562          * to set 'stripe_addr' which is where we will write to.
5563          */
5564         if (mddev->reshape_backwards) {
5565                 BUG_ON(conf->reshape_progress == 0);
5566                 stripe_addr = writepos;
5567                 BUG_ON((mddev->dev_sectors &
5568                         ~((sector_t)reshape_sectors - 1))
5569                        - reshape_sectors - stripe_addr
5570                        != sector_nr);
5571         } else {
5572                 BUG_ON(writepos != sector_nr + reshape_sectors);
5573                 stripe_addr = sector_nr;
5574         }
5575
5576         /* 'writepos' is the most advanced device address we might write.
5577          * 'readpos' is the least advanced device address we might read.
5578          * 'safepos' is the least address recorded in the metadata as having
5579          *     been reshaped.
5580          * If there is a min_offset_diff, these are adjusted either by
5581          * increasing the safepos/readpos if diff is negative, or
5582          * increasing writepos if diff is positive.
5583          * If 'readpos' is then behind 'writepos', there is no way that we can
5584          * ensure safety in the face of a crash - that must be done by userspace
5585          * making a backup of the data.  So in that case there is no particular
5586          * rush to update metadata.
5587          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5588          * update the metadata to advance 'safepos' to match 'readpos' so that
5589          * we can be safe in the event of a crash.
5590          * So we insist on updating metadata if safepos is behind writepos and
5591          * readpos is beyond writepos.
5592          * In any case, update the metadata every 10 seconds.
5593          * Maybe that number should be configurable, but I'm not sure it is
5594          * worth it.... maybe it could be a multiple of safemode_delay???
5595          */
5596         if (conf->min_offset_diff < 0) {
5597                 safepos += -conf->min_offset_diff;
5598                 readpos += -conf->min_offset_diff;
5599         } else
5600                 writepos += conf->min_offset_diff;
5601
5602         if ((mddev->reshape_backwards
5603              ? (safepos > writepos && readpos < writepos)
5604              : (safepos < writepos && readpos > writepos)) ||
5605             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5606                 /* Cannot proceed until we've updated the superblock... */
5607                 wait_event(conf->wait_for_overlap,
5608                            atomic_read(&conf->reshape_stripes)==0
5609                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5610                 if (atomic_read(&conf->reshape_stripes) != 0)
5611                         return 0;
5612                 mddev->reshape_position = conf->reshape_progress;
5613                 mddev->curr_resync_completed = sector_nr;
5614                 conf->reshape_checkpoint = jiffies;
5615                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5616                 md_wakeup_thread(mddev->thread);
5617                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5618                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5619                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5620                         return 0;
5621                 spin_lock_irq(&conf->device_lock);
5622                 conf->reshape_safe = mddev->reshape_position;
5623                 spin_unlock_irq(&conf->device_lock);
5624                 wake_up(&conf->wait_for_overlap);
5625                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5626         }
5627
5628         INIT_LIST_HEAD(&stripes);
5629         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5630                 int j;
5631                 int skipped_disk = 0;
5632                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5633                 set_bit(STRIPE_EXPANDING, &sh->state);
5634                 atomic_inc(&conf->reshape_stripes);
5635                 /* If any of this stripe is beyond the end of the old
5636                  * array, then we need to zero those blocks
5637                  */
5638                 for (j=sh->disks; j--;) {
5639                         sector_t s;
5640                         if (j == sh->pd_idx)
5641                                 continue;
5642                         if (conf->level == 6 &&
5643                             j == sh->qd_idx)
5644                                 continue;
5645                         s = raid5_compute_blocknr(sh, j, 0);
5646                         if (s < raid5_size(mddev, 0, 0)) {
5647                                 skipped_disk = 1;
5648                                 continue;
5649                         }
5650                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5651                         set_bit(R5_Expanded, &sh->dev[j].flags);
5652                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5653                 }
5654                 if (!skipped_disk) {
5655                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5656                         set_bit(STRIPE_HANDLE, &sh->state);
5657                 }
5658                 list_add(&sh->lru, &stripes);
5659         }
5660         spin_lock_irq(&conf->device_lock);
5661         if (mddev->reshape_backwards)
5662                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5663         else
5664                 conf->reshape_progress += reshape_sectors * new_data_disks;
5665         spin_unlock_irq(&conf->device_lock);
5666         /* Ok, those stripe are ready. We can start scheduling
5667          * reads on the source stripes.
5668          * The source stripes are determined by mapping the first and last
5669          * block on the destination stripes.
5670          */
5671         first_sector =
5672                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5673                                      1, &dd_idx, NULL);
5674         last_sector =
5675                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5676                                             * new_data_disks - 1),
5677                                      1, &dd_idx, NULL);
5678         if (last_sector >= mddev->dev_sectors)
5679                 last_sector = mddev->dev_sectors - 1;
5680         while (first_sector <= last_sector) {
5681                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5682                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5683                 set_bit(STRIPE_HANDLE, &sh->state);
5684                 raid5_release_stripe(sh);
5685                 first_sector += STRIPE_SECTORS;
5686         }
5687         /* Now that the sources are clearly marked, we can release
5688          * the destination stripes
5689          */
5690         while (!list_empty(&stripes)) {
5691                 sh = list_entry(stripes.next, struct stripe_head, lru);
5692                 list_del_init(&sh->lru);
5693                 raid5_release_stripe(sh);
5694         }
5695         /* If this takes us to the resync_max point where we have to pause,
5696          * then we need to write out the superblock.
5697          */
5698         sector_nr += reshape_sectors;
5699         retn = reshape_sectors;
5700 finish:
5701         if (mddev->curr_resync_completed > mddev->resync_max ||
5702             (sector_nr - mddev->curr_resync_completed) * 2
5703             >= mddev->resync_max - mddev->curr_resync_completed) {
5704                 /* Cannot proceed until we've updated the superblock... */
5705                 wait_event(conf->wait_for_overlap,
5706                            atomic_read(&conf->reshape_stripes) == 0
5707                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5708                 if (atomic_read(&conf->reshape_stripes) != 0)
5709                         goto ret;
5710                 mddev->reshape_position = conf->reshape_progress;
5711                 mddev->curr_resync_completed = sector_nr;
5712                 conf->reshape_checkpoint = jiffies;
5713                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5714                 md_wakeup_thread(mddev->thread);
5715                 wait_event(mddev->sb_wait,
5716                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5717                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5718                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5719                         goto ret;
5720                 spin_lock_irq(&conf->device_lock);
5721                 conf->reshape_safe = mddev->reshape_position;
5722                 spin_unlock_irq(&conf->device_lock);
5723                 wake_up(&conf->wait_for_overlap);
5724                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5725         }
5726 ret:
5727         return retn;
5728 }
5729
5730 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
5731                                           int *skipped)
5732 {
5733         struct r5conf *conf = mddev->private;
5734         struct stripe_head *sh;
5735         sector_t max_sector = mddev->dev_sectors;
5736         sector_t sync_blocks;
5737         int still_degraded = 0;
5738         int i;
5739
5740         if (sector_nr >= max_sector) {
5741                 /* just being told to finish up .. nothing much to do */
5742
5743                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5744                         end_reshape(conf);
5745                         return 0;
5746                 }
5747
5748                 if (mddev->curr_resync < max_sector) /* aborted */
5749                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5750                                         &sync_blocks, 1);
5751                 else /* completed sync */
5752                         conf->fullsync = 0;
5753                 bitmap_close_sync(mddev->bitmap);
5754
5755                 return 0;
5756         }
5757
5758         /* Allow raid5_quiesce to complete */
5759         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5760
5761         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5762                 return reshape_request(mddev, sector_nr, skipped);
5763
5764         /* No need to check resync_max as we never do more than one
5765          * stripe, and as resync_max will always be on a chunk boundary,
5766          * if the check in md_do_sync didn't fire, there is no chance
5767          * of overstepping resync_max here
5768          */
5769
5770         /* if there is too many failed drives and we are trying
5771          * to resync, then assert that we are finished, because there is
5772          * nothing we can do.
5773          */
5774         if (mddev->degraded >= conf->max_degraded &&
5775             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5776                 sector_t rv = mddev->dev_sectors - sector_nr;
5777                 *skipped = 1;
5778                 return rv;
5779         }
5780         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5781             !conf->fullsync &&
5782             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5783             sync_blocks >= STRIPE_SECTORS) {
5784                 /* we can skip this block, and probably more */
5785                 sync_blocks /= STRIPE_SECTORS;
5786                 *skipped = 1;
5787                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5788         }
5789
5790         bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5791
5792         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5793         if (sh == NULL) {
5794                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5795                 /* make sure we don't swamp the stripe cache if someone else
5796                  * is trying to get access
5797                  */
5798                 schedule_timeout_uninterruptible(1);
5799         }
5800         /* Need to check if array will still be degraded after recovery/resync
5801          * Note in case of > 1 drive failures it's possible we're rebuilding
5802          * one drive while leaving another faulty drive in array.
5803          */
5804         rcu_read_lock();
5805         for (i = 0; i < conf->raid_disks; i++) {
5806                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5807
5808                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5809                         still_degraded = 1;
5810         }
5811         rcu_read_unlock();
5812
5813         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5814
5815         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5816         set_bit(STRIPE_HANDLE, &sh->state);
5817
5818         raid5_release_stripe(sh);
5819
5820         return STRIPE_SECTORS;
5821 }
5822
5823 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5824 {
5825         /* We may not be able to submit a whole bio at once as there
5826          * may not be enough stripe_heads available.
5827          * We cannot pre-allocate enough stripe_heads as we may need
5828          * more than exist in the cache (if we allow ever large chunks).
5829          * So we do one stripe head at a time and record in
5830          * ->bi_hw_segments how many have been done.
5831          *
5832          * We *know* that this entire raid_bio is in one chunk, so
5833          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5834          */
5835         struct stripe_head *sh;
5836         int dd_idx;
5837         sector_t sector, logical_sector, last_sector;
5838         int scnt = 0;
5839         int remaining;
5840         int handled = 0;
5841
5842         logical_sector = raid_bio->bi_iter.bi_sector &
5843                 ~((sector_t)STRIPE_SECTORS-1);
5844         sector = raid5_compute_sector(conf, logical_sector,
5845                                       0, &dd_idx, NULL);
5846         last_sector = bio_end_sector(raid_bio);
5847
5848         for (; logical_sector < last_sector;
5849              logical_sector += STRIPE_SECTORS,
5850                      sector += STRIPE_SECTORS,
5851                      scnt++) {
5852
5853                 if (scnt < raid5_bi_processed_stripes(raid_bio))
5854                         /* already done this stripe */
5855                         continue;
5856
5857                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5858
5859                 if (!sh) {
5860                         /* failed to get a stripe - must wait */
5861                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5862                         conf->retry_read_aligned = raid_bio;
5863                         return handled;
5864                 }
5865
5866                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5867                         raid5_release_stripe(sh);
5868                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5869                         conf->retry_read_aligned = raid_bio;
5870                         return handled;
5871                 }
5872
5873                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5874                 handle_stripe(sh);
5875                 raid5_release_stripe(sh);
5876                 handled++;
5877         }
5878         remaining = raid5_dec_bi_active_stripes(raid_bio);
5879         if (remaining == 0) {
5880                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5881                                          raid_bio, 0);
5882                 bio_endio(raid_bio);
5883         }
5884         if (atomic_dec_and_test(&conf->active_aligned_reads))
5885                 wake_up(&conf->wait_for_quiescent);
5886         return handled;
5887 }
5888
5889 static int handle_active_stripes(struct r5conf *conf, int group,
5890                                  struct r5worker *worker,
5891                                  struct list_head *temp_inactive_list)
5892 {
5893         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5894         int i, batch_size = 0, hash;
5895         bool release_inactive = false;
5896
5897         while (batch_size < MAX_STRIPE_BATCH &&
5898                         (sh = __get_priority_stripe(conf, group)) != NULL)
5899                 batch[batch_size++] = sh;
5900
5901         if (batch_size == 0) {
5902                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5903                         if (!list_empty(temp_inactive_list + i))
5904                                 break;
5905                 if (i == NR_STRIPE_HASH_LOCKS) {
5906                         spin_unlock_irq(&conf->device_lock);
5907                         r5l_flush_stripe_to_raid(conf->log);
5908                         spin_lock_irq(&conf->device_lock);
5909                         return batch_size;
5910                 }
5911                 release_inactive = true;
5912         }
5913         spin_unlock_irq(&conf->device_lock);
5914
5915         release_inactive_stripe_list(conf, temp_inactive_list,
5916                                      NR_STRIPE_HASH_LOCKS);
5917
5918         r5l_flush_stripe_to_raid(conf->log);
5919         if (release_inactive) {
5920                 spin_lock_irq(&conf->device_lock);
5921                 return 0;
5922         }
5923
5924         for (i = 0; i < batch_size; i++)
5925                 handle_stripe(batch[i]);
5926         r5l_write_stripe_run(conf->log);
5927
5928         cond_resched();
5929
5930         spin_lock_irq(&conf->device_lock);
5931         for (i = 0; i < batch_size; i++) {
5932                 hash = batch[i]->hash_lock_index;
5933                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5934         }
5935         return batch_size;
5936 }
5937
5938 static void raid5_do_work(struct work_struct *work)
5939 {
5940         struct r5worker *worker = container_of(work, struct r5worker, work);
5941         struct r5worker_group *group = worker->group;
5942         struct r5conf *conf = group->conf;
5943         int group_id = group - conf->worker_groups;
5944         int handled;
5945         struct blk_plug plug;
5946
5947         pr_debug("+++ raid5worker active\n");
5948
5949         blk_start_plug(&plug);
5950         handled = 0;
5951         spin_lock_irq(&conf->device_lock);
5952         while (1) {
5953                 int batch_size, released;
5954
5955                 released = release_stripe_list(conf, worker->temp_inactive_list);
5956
5957                 batch_size = handle_active_stripes(conf, group_id, worker,
5958                                                    worker->temp_inactive_list);
5959                 worker->working = false;
5960                 if (!batch_size && !released)
5961                         break;
5962                 handled += batch_size;
5963         }
5964         pr_debug("%d stripes handled\n", handled);
5965
5966         spin_unlock_irq(&conf->device_lock);
5967         blk_finish_plug(&plug);
5968
5969         pr_debug("--- raid5worker inactive\n");
5970 }
5971
5972 /*
5973  * This is our raid5 kernel thread.
5974  *
5975  * We scan the hash table for stripes which can be handled now.
5976  * During the scan, completed stripes are saved for us by the interrupt
5977  * handler, so that they will not have to wait for our next wakeup.
5978  */
5979 static void raid5d(struct md_thread *thread)
5980 {
5981         struct mddev *mddev = thread->mddev;
5982         struct r5conf *conf = mddev->private;
5983         int handled;
5984         struct blk_plug plug;
5985
5986         pr_debug("+++ raid5d active\n");
5987
5988         md_check_recovery(mddev);
5989
5990         if (!bio_list_empty(&conf->return_bi) &&
5991             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
5992                 struct bio_list tmp = BIO_EMPTY_LIST;
5993                 spin_lock_irq(&conf->device_lock);
5994                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
5995                         bio_list_merge(&tmp, &conf->return_bi);
5996                         bio_list_init(&conf->return_bi);
5997                 }
5998                 spin_unlock_irq(&conf->device_lock);
5999                 return_io(&tmp);
6000         }
6001
6002         blk_start_plug(&plug);
6003         handled = 0;
6004         spin_lock_irq(&conf->device_lock);
6005         while (1) {
6006                 struct bio *bio;
6007                 int batch_size, released;
6008
6009                 released = release_stripe_list(conf, conf->temp_inactive_list);
6010                 if (released)
6011                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6012
6013                 if (
6014                     !list_empty(&conf->bitmap_list)) {
6015                         /* Now is a good time to flush some bitmap updates */
6016                         conf->seq_flush++;
6017                         spin_unlock_irq(&conf->device_lock);
6018                         bitmap_unplug(mddev->bitmap);
6019                         spin_lock_irq(&conf->device_lock);
6020                         conf->seq_write = conf->seq_flush;
6021                         activate_bit_delay(conf, conf->temp_inactive_list);
6022                 }
6023                 raid5_activate_delayed(conf);
6024
6025                 while ((bio = remove_bio_from_retry(conf))) {
6026                         int ok;
6027                         spin_unlock_irq(&conf->device_lock);
6028                         ok = retry_aligned_read(conf, bio);
6029                         spin_lock_irq(&conf->device_lock);
6030                         if (!ok)
6031                                 break;
6032                         handled++;
6033                 }
6034
6035                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6036                                                    conf->temp_inactive_list);
6037                 if (!batch_size && !released)
6038                         break;
6039                 handled += batch_size;
6040
6041                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6042                         spin_unlock_irq(&conf->device_lock);
6043                         md_check_recovery(mddev);
6044                         spin_lock_irq(&conf->device_lock);
6045                 }
6046         }
6047         pr_debug("%d stripes handled\n", handled);
6048
6049         spin_unlock_irq(&conf->device_lock);
6050         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6051             mutex_trylock(&conf->cache_size_mutex)) {
6052                 grow_one_stripe(conf, __GFP_NOWARN);
6053                 /* Set flag even if allocation failed.  This helps
6054                  * slow down allocation requests when mem is short
6055                  */
6056                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6057                 mutex_unlock(&conf->cache_size_mutex);
6058         }
6059
6060         r5l_flush_stripe_to_raid(conf->log);
6061
6062         async_tx_issue_pending_all();
6063         blk_finish_plug(&plug);
6064
6065         pr_debug("--- raid5d inactive\n");
6066 }
6067
6068 static ssize_t
6069 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6070 {
6071         struct r5conf *conf;
6072         int ret = 0;
6073         spin_lock(&mddev->lock);
6074         conf = mddev->private;
6075         if (conf)
6076                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6077         spin_unlock(&mddev->lock);
6078         return ret;
6079 }
6080
6081 int
6082 raid5_set_cache_size(struct mddev *mddev, int size)
6083 {
6084         struct r5conf *conf = mddev->private;
6085         int err;
6086
6087         if (size <= 16 || size > 32768)
6088                 return -EINVAL;
6089
6090         conf->min_nr_stripes = size;
6091         mutex_lock(&conf->cache_size_mutex);
6092         while (size < conf->max_nr_stripes &&
6093                drop_one_stripe(conf))
6094                 ;
6095         mutex_unlock(&conf->cache_size_mutex);
6096
6097
6098         err = md_allow_write(mddev);
6099         if (err)
6100                 return err;
6101
6102         mutex_lock(&conf->cache_size_mutex);
6103         while (size > conf->max_nr_stripes)
6104                 if (!grow_one_stripe(conf, GFP_KERNEL))
6105                         break;
6106         mutex_unlock(&conf->cache_size_mutex);
6107
6108         return 0;
6109 }
6110 EXPORT_SYMBOL(raid5_set_cache_size);
6111
6112 static ssize_t
6113 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6114 {
6115         struct r5conf *conf;
6116         unsigned long new;
6117         int err;
6118
6119         if (len >= PAGE_SIZE)
6120                 return -EINVAL;
6121         if (kstrtoul(page, 10, &new))
6122                 return -EINVAL;
6123         err = mddev_lock(mddev);
6124         if (err)
6125                 return err;
6126         conf = mddev->private;
6127         if (!conf)
6128                 err = -ENODEV;
6129         else
6130                 err = raid5_set_cache_size(mddev, new);
6131         mddev_unlock(mddev);
6132
6133         return err ?: len;
6134 }
6135
6136 static struct md_sysfs_entry
6137 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6138                                 raid5_show_stripe_cache_size,
6139                                 raid5_store_stripe_cache_size);
6140
6141 static ssize_t
6142 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6143 {
6144         struct r5conf *conf = mddev->private;
6145         if (conf)
6146                 return sprintf(page, "%d\n", conf->rmw_level);
6147         else
6148                 return 0;
6149 }
6150
6151 static ssize_t
6152 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6153 {
6154         struct r5conf *conf = mddev->private;
6155         unsigned long new;
6156
6157         if (!conf)
6158                 return -ENODEV;
6159
6160         if (len >= PAGE_SIZE)
6161                 return -EINVAL;
6162
6163         if (kstrtoul(page, 10, &new))
6164                 return -EINVAL;
6165
6166         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6167                 return -EINVAL;
6168
6169         if (new != PARITY_DISABLE_RMW &&
6170             new != PARITY_ENABLE_RMW &&
6171             new != PARITY_PREFER_RMW)
6172                 return -EINVAL;
6173
6174         conf->rmw_level = new;
6175         return len;
6176 }
6177
6178 static struct md_sysfs_entry
6179 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6180                          raid5_show_rmw_level,
6181                          raid5_store_rmw_level);
6182
6183
6184 static ssize_t
6185 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6186 {
6187         struct r5conf *conf;
6188         int ret = 0;
6189         spin_lock(&mddev->lock);
6190         conf = mddev->private;
6191         if (conf)
6192                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6193         spin_unlock(&mddev->lock);
6194         return ret;
6195 }
6196
6197 static ssize_t
6198 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6199 {
6200         struct r5conf *conf;
6201         unsigned long new;
6202         int err;
6203
6204         if (len >= PAGE_SIZE)
6205                 return -EINVAL;
6206         if (kstrtoul(page, 10, &new))
6207                 return -EINVAL;
6208
6209         err = mddev_lock(mddev);
6210         if (err)
6211                 return err;
6212         conf = mddev->private;
6213         if (!conf)
6214                 err = -ENODEV;
6215         else if (new > conf->min_nr_stripes)
6216                 err = -EINVAL;
6217         else
6218                 conf->bypass_threshold = new;
6219         mddev_unlock(mddev);
6220         return err ?: len;
6221 }
6222
6223 static struct md_sysfs_entry
6224 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6225                                         S_IRUGO | S_IWUSR,
6226                                         raid5_show_preread_threshold,
6227                                         raid5_store_preread_threshold);
6228
6229 static ssize_t
6230 raid5_show_skip_copy(struct mddev *mddev, char *page)
6231 {
6232         struct r5conf *conf;
6233         int ret = 0;
6234         spin_lock(&mddev->lock);
6235         conf = mddev->private;
6236         if (conf)
6237                 ret = sprintf(page, "%d\n", conf->skip_copy);
6238         spin_unlock(&mddev->lock);
6239         return ret;
6240 }
6241
6242 static ssize_t
6243 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6244 {
6245         struct r5conf *conf;
6246         unsigned long new;
6247         int err;
6248
6249         if (len >= PAGE_SIZE)
6250                 return -EINVAL;
6251         if (kstrtoul(page, 10, &new))
6252                 return -EINVAL;
6253         new = !!new;
6254
6255         err = mddev_lock(mddev);
6256         if (err)
6257                 return err;
6258         conf = mddev->private;
6259         if (!conf)
6260                 err = -ENODEV;
6261         else if (new != conf->skip_copy) {
6262                 mddev_suspend(mddev);
6263                 conf->skip_copy = new;
6264                 if (new)
6265                         mddev->queue->backing_dev_info.capabilities |=
6266                                 BDI_CAP_STABLE_WRITES;
6267                 else
6268                         mddev->queue->backing_dev_info.capabilities &=
6269                                 ~BDI_CAP_STABLE_WRITES;
6270                 mddev_resume(mddev);
6271         }
6272         mddev_unlock(mddev);
6273         return err ?: len;
6274 }
6275
6276 static struct md_sysfs_entry
6277 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6278                                         raid5_show_skip_copy,
6279                                         raid5_store_skip_copy);
6280
6281 static ssize_t
6282 stripe_cache_active_show(struct mddev *mddev, char *page)
6283 {
6284         struct r5conf *conf = mddev->private;
6285         if (conf)
6286                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6287         else
6288                 return 0;
6289 }
6290
6291 static struct md_sysfs_entry
6292 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6293
6294 static ssize_t
6295 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6296 {
6297         struct r5conf *conf;
6298         int ret = 0;
6299         spin_lock(&mddev->lock);
6300         conf = mddev->private;
6301         if (conf)
6302                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6303         spin_unlock(&mddev->lock);
6304         return ret;
6305 }
6306
6307 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6308                                int *group_cnt,
6309                                int *worker_cnt_per_group,
6310                                struct r5worker_group **worker_groups);
6311 static ssize_t
6312 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6313 {
6314         struct r5conf *conf;
6315         unsigned long new;
6316         int err;
6317         struct r5worker_group *new_groups, *old_groups;
6318         int group_cnt, worker_cnt_per_group;
6319
6320         if (len >= PAGE_SIZE)
6321                 return -EINVAL;
6322         if (kstrtoul(page, 10, &new))
6323                 return -EINVAL;
6324
6325         err = mddev_lock(mddev);
6326         if (err)
6327                 return err;
6328         conf = mddev->private;
6329         if (!conf)
6330                 err = -ENODEV;
6331         else if (new != conf->worker_cnt_per_group) {
6332                 mddev_suspend(mddev);
6333
6334                 old_groups = conf->worker_groups;
6335                 if (old_groups)
6336                         flush_workqueue(raid5_wq);
6337
6338                 err = alloc_thread_groups(conf, new,
6339                                           &group_cnt, &worker_cnt_per_group,
6340                                           &new_groups);
6341                 if (!err) {
6342                         spin_lock_irq(&conf->device_lock);
6343                         conf->group_cnt = group_cnt;
6344                         conf->worker_cnt_per_group = worker_cnt_per_group;
6345                         conf->worker_groups = new_groups;
6346                         spin_unlock_irq(&conf->device_lock);
6347
6348                         if (old_groups)
6349                                 kfree(old_groups[0].workers);
6350                         kfree(old_groups);
6351                 }
6352                 mddev_resume(mddev);
6353         }
6354         mddev_unlock(mddev);
6355
6356         return err ?: len;
6357 }
6358
6359 static struct md_sysfs_entry
6360 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6361                                 raid5_show_group_thread_cnt,
6362                                 raid5_store_group_thread_cnt);
6363
6364 static struct attribute *raid5_attrs[] =  {
6365         &raid5_stripecache_size.attr,
6366         &raid5_stripecache_active.attr,
6367         &raid5_preread_bypass_threshold.attr,
6368         &raid5_group_thread_cnt.attr,
6369         &raid5_skip_copy.attr,
6370         &raid5_rmw_level.attr,
6371         &r5c_journal_mode.attr,
6372         NULL,
6373 };
6374 static struct attribute_group raid5_attrs_group = {
6375         .name = NULL,
6376         .attrs = raid5_attrs,
6377 };
6378
6379 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6380                                int *group_cnt,
6381                                int *worker_cnt_per_group,
6382                                struct r5worker_group **worker_groups)
6383 {
6384         int i, j, k;
6385         ssize_t size;
6386         struct r5worker *workers;
6387
6388         *worker_cnt_per_group = cnt;
6389         if (cnt == 0) {
6390                 *group_cnt = 0;
6391                 *worker_groups = NULL;
6392                 return 0;
6393         }
6394         *group_cnt = num_possible_nodes();
6395         size = sizeof(struct r5worker) * cnt;
6396         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6397         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6398                                 *group_cnt, GFP_NOIO);
6399         if (!*worker_groups || !workers) {
6400                 kfree(workers);
6401                 kfree(*worker_groups);
6402                 return -ENOMEM;
6403         }
6404
6405         for (i = 0; i < *group_cnt; i++) {
6406                 struct r5worker_group *group;
6407
6408                 group = &(*worker_groups)[i];
6409                 INIT_LIST_HEAD(&group->handle_list);
6410                 group->conf = conf;
6411                 group->workers = workers + i * cnt;
6412
6413                 for (j = 0; j < cnt; j++) {
6414                         struct r5worker *worker = group->workers + j;
6415                         worker->group = group;
6416                         INIT_WORK(&worker->work, raid5_do_work);
6417
6418                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6419                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6420                 }
6421         }
6422
6423         return 0;
6424 }
6425
6426 static void free_thread_groups(struct r5conf *conf)
6427 {
6428         if (conf->worker_groups)
6429                 kfree(conf->worker_groups[0].workers);
6430         kfree(conf->worker_groups);
6431         conf->worker_groups = NULL;
6432 }
6433
6434 static sector_t
6435 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6436 {
6437         struct r5conf *conf = mddev->private;
6438
6439         if (!sectors)
6440                 sectors = mddev->dev_sectors;
6441         if (!raid_disks)
6442                 /* size is defined by the smallest of previous and new size */
6443                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6444
6445         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6446         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6447         return sectors * (raid_disks - conf->max_degraded);
6448 }
6449
6450 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6451 {
6452         safe_put_page(percpu->spare_page);
6453         if (percpu->scribble)
6454                 flex_array_free(percpu->scribble);
6455         percpu->spare_page = NULL;
6456         percpu->scribble = NULL;
6457 }
6458
6459 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6460 {
6461         if (conf->level == 6 && !percpu->spare_page)
6462                 percpu->spare_page = alloc_page(GFP_KERNEL);
6463         if (!percpu->scribble)
6464                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6465                                                       conf->previous_raid_disks),
6466                                                   max(conf->chunk_sectors,
6467                                                       conf->prev_chunk_sectors)
6468                                                    / STRIPE_SECTORS,
6469                                                   GFP_KERNEL);
6470
6471         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6472                 free_scratch_buffer(conf, percpu);
6473                 return -ENOMEM;
6474         }
6475
6476         return 0;
6477 }
6478
6479 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6480 {
6481         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6482
6483         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6484         return 0;
6485 }
6486
6487 static void raid5_free_percpu(struct r5conf *conf)
6488 {
6489         if (!conf->percpu)
6490                 return;
6491
6492         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6493         free_percpu(conf->percpu);
6494 }
6495
6496 static void free_conf(struct r5conf *conf)
6497 {
6498         int i;
6499
6500         if (conf->log)
6501                 r5l_exit_log(conf->log);
6502         if (conf->shrinker.nr_deferred)
6503                 unregister_shrinker(&conf->shrinker);
6504
6505         free_thread_groups(conf);
6506         shrink_stripes(conf);
6507         raid5_free_percpu(conf);
6508         for (i = 0; i < conf->pool_size; i++)
6509                 if (conf->disks[i].extra_page)
6510                         put_page(conf->disks[i].extra_page);
6511         kfree(conf->disks);
6512         kfree(conf->stripe_hashtbl);
6513         kfree(conf);
6514 }
6515
6516 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6517 {
6518         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6519         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6520
6521         if (alloc_scratch_buffer(conf, percpu)) {
6522                 pr_warn("%s: failed memory allocation for cpu%u\n",
6523                         __func__, cpu);
6524                 return -ENOMEM;
6525         }
6526         return 0;
6527 }
6528
6529 static int raid5_alloc_percpu(struct r5conf *conf)
6530 {
6531         int err = 0;
6532
6533         conf->percpu = alloc_percpu(struct raid5_percpu);
6534         if (!conf->percpu)
6535                 return -ENOMEM;
6536
6537         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6538         if (!err) {
6539                 conf->scribble_disks = max(conf->raid_disks,
6540                         conf->previous_raid_disks);
6541                 conf->scribble_sectors = max(conf->chunk_sectors,
6542                         conf->prev_chunk_sectors);
6543         }
6544         return err;
6545 }
6546
6547 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6548                                       struct shrink_control *sc)
6549 {
6550         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6551         unsigned long ret = SHRINK_STOP;
6552
6553         if (mutex_trylock(&conf->cache_size_mutex)) {
6554                 ret= 0;
6555                 while (ret < sc->nr_to_scan &&
6556                        conf->max_nr_stripes > conf->min_nr_stripes) {
6557                         if (drop_one_stripe(conf) == 0) {
6558                                 ret = SHRINK_STOP;
6559                                 break;
6560                         }
6561                         ret++;
6562                 }
6563                 mutex_unlock(&conf->cache_size_mutex);
6564         }
6565         return ret;
6566 }
6567
6568 static unsigned long raid5_cache_count(struct shrinker *shrink,
6569                                        struct shrink_control *sc)
6570 {
6571         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6572
6573         if (conf->max_nr_stripes < conf->min_nr_stripes)
6574                 /* unlikely, but not impossible */
6575                 return 0;
6576         return conf->max_nr_stripes - conf->min_nr_stripes;
6577 }
6578
6579 static struct r5conf *setup_conf(struct mddev *mddev)
6580 {
6581         struct r5conf *conf;
6582         int raid_disk, memory, max_disks;
6583         struct md_rdev *rdev;
6584         struct disk_info *disk;
6585         char pers_name[6];
6586         int i;
6587         int group_cnt, worker_cnt_per_group;
6588         struct r5worker_group *new_group;
6589
6590         if (mddev->new_level != 5
6591             && mddev->new_level != 4
6592             && mddev->new_level != 6) {
6593                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6594                         mdname(mddev), mddev->new_level);
6595                 return ERR_PTR(-EIO);
6596         }
6597         if ((mddev->new_level == 5
6598              && !algorithm_valid_raid5(mddev->new_layout)) ||
6599             (mddev->new_level == 6
6600              && !algorithm_valid_raid6(mddev->new_layout))) {
6601                 pr_warn("md/raid:%s: layout %d not supported\n",
6602                         mdname(mddev), mddev->new_layout);
6603                 return ERR_PTR(-EIO);
6604         }
6605         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6606                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6607                         mdname(mddev), mddev->raid_disks);
6608                 return ERR_PTR(-EINVAL);
6609         }
6610
6611         if (!mddev->new_chunk_sectors ||
6612             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6613             !is_power_of_2(mddev->new_chunk_sectors)) {
6614                 pr_warn("md/raid:%s: invalid chunk size %d\n",
6615                         mdname(mddev), mddev->new_chunk_sectors << 9);
6616                 return ERR_PTR(-EINVAL);
6617         }
6618
6619         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6620         if (conf == NULL)
6621                 goto abort;
6622         /* Don't enable multi-threading by default*/
6623         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6624                                  &new_group)) {
6625                 conf->group_cnt = group_cnt;
6626                 conf->worker_cnt_per_group = worker_cnt_per_group;
6627                 conf->worker_groups = new_group;
6628         } else
6629                 goto abort;
6630         spin_lock_init(&conf->device_lock);
6631         seqcount_init(&conf->gen_lock);
6632         mutex_init(&conf->cache_size_mutex);
6633         init_waitqueue_head(&conf->wait_for_quiescent);
6634         init_waitqueue_head(&conf->wait_for_stripe);
6635         init_waitqueue_head(&conf->wait_for_overlap);
6636         INIT_LIST_HEAD(&conf->handle_list);
6637         INIT_LIST_HEAD(&conf->hold_list);
6638         INIT_LIST_HEAD(&conf->delayed_list);
6639         INIT_LIST_HEAD(&conf->bitmap_list);
6640         bio_list_init(&conf->return_bi);
6641         init_llist_head(&conf->released_stripes);
6642         atomic_set(&conf->active_stripes, 0);
6643         atomic_set(&conf->preread_active_stripes, 0);
6644         atomic_set(&conf->active_aligned_reads, 0);
6645         conf->bypass_threshold = BYPASS_THRESHOLD;
6646         conf->recovery_disabled = mddev->recovery_disabled - 1;
6647
6648         conf->raid_disks = mddev->raid_disks;
6649         if (mddev->reshape_position == MaxSector)
6650                 conf->previous_raid_disks = mddev->raid_disks;
6651         else
6652                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6653         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6654
6655         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6656                               GFP_KERNEL);
6657
6658         if (!conf->disks)
6659                 goto abort;
6660
6661         for (i = 0; i < max_disks; i++) {
6662                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6663                 if (!conf->disks[i].extra_page)
6664                         goto abort;
6665         }
6666
6667         conf->mddev = mddev;
6668
6669         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6670                 goto abort;
6671
6672         /* We init hash_locks[0] separately to that it can be used
6673          * as the reference lock in the spin_lock_nest_lock() call
6674          * in lock_all_device_hash_locks_irq in order to convince
6675          * lockdep that we know what we are doing.
6676          */
6677         spin_lock_init(conf->hash_locks);
6678         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6679                 spin_lock_init(conf->hash_locks + i);
6680
6681         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6682                 INIT_LIST_HEAD(conf->inactive_list + i);
6683
6684         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6685                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6686
6687         atomic_set(&conf->r5c_cached_full_stripes, 0);
6688         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6689         atomic_set(&conf->r5c_cached_partial_stripes, 0);
6690         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6691
6692         conf->level = mddev->new_level;
6693         conf->chunk_sectors = mddev->new_chunk_sectors;
6694         if (raid5_alloc_percpu(conf) != 0)
6695                 goto abort;
6696
6697         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6698
6699         rdev_for_each(rdev, mddev) {
6700                 raid_disk = rdev->raid_disk;
6701                 if (raid_disk >= max_disks
6702                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6703                         continue;
6704                 disk = conf->disks + raid_disk;
6705
6706                 if (test_bit(Replacement, &rdev->flags)) {
6707                         if (disk->replacement)
6708                                 goto abort;
6709                         disk->replacement = rdev;
6710                 } else {
6711                         if (disk->rdev)
6712                                 goto abort;
6713                         disk->rdev = rdev;
6714                 }
6715
6716                 if (test_bit(In_sync, &rdev->flags)) {
6717                         char b[BDEVNAME_SIZE];
6718                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
6719                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6720                 } else if (rdev->saved_raid_disk != raid_disk)
6721                         /* Cannot rely on bitmap to complete recovery */
6722                         conf->fullsync = 1;
6723         }
6724
6725         conf->level = mddev->new_level;
6726         if (conf->level == 6) {
6727                 conf->max_degraded = 2;
6728                 if (raid6_call.xor_syndrome)
6729                         conf->rmw_level = PARITY_ENABLE_RMW;
6730                 else
6731                         conf->rmw_level = PARITY_DISABLE_RMW;
6732         } else {
6733                 conf->max_degraded = 1;
6734                 conf->rmw_level = PARITY_ENABLE_RMW;
6735         }
6736         conf->algorithm = mddev->new_layout;
6737         conf->reshape_progress = mddev->reshape_position;
6738         if (conf->reshape_progress != MaxSector) {
6739                 conf->prev_chunk_sectors = mddev->chunk_sectors;
6740                 conf->prev_algo = mddev->layout;
6741         } else {
6742                 conf->prev_chunk_sectors = conf->chunk_sectors;
6743                 conf->prev_algo = conf->algorithm;
6744         }
6745
6746         conf->min_nr_stripes = NR_STRIPES;
6747         if (mddev->reshape_position != MaxSector) {
6748                 int stripes = max_t(int,
6749                         ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
6750                         ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
6751                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
6752                 if (conf->min_nr_stripes != NR_STRIPES)
6753                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
6754                                 mdname(mddev), conf->min_nr_stripes);
6755         }
6756         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6757                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6758         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6759         if (grow_stripes(conf, conf->min_nr_stripes)) {
6760                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
6761                         mdname(mddev), memory);
6762                 goto abort;
6763         } else
6764                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
6765         /*
6766          * Losing a stripe head costs more than the time to refill it,
6767          * it reduces the queue depth and so can hurt throughput.
6768          * So set it rather large, scaled by number of devices.
6769          */
6770         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6771         conf->shrinker.scan_objects = raid5_cache_scan;
6772         conf->shrinker.count_objects = raid5_cache_count;
6773         conf->shrinker.batch = 128;
6774         conf->shrinker.flags = 0;
6775         if (register_shrinker(&conf->shrinker)) {
6776                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
6777                         mdname(mddev));
6778                 goto abort;
6779         }
6780
6781         sprintf(pers_name, "raid%d", mddev->new_level);
6782         conf->thread = md_register_thread(raid5d, mddev, pers_name);
6783         if (!conf->thread) {
6784                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
6785                         mdname(mddev));
6786                 goto abort;
6787         }
6788
6789         return conf;
6790
6791  abort:
6792         if (conf) {
6793                 free_conf(conf);
6794                 return ERR_PTR(-EIO);
6795         } else
6796                 return ERR_PTR(-ENOMEM);
6797 }
6798
6799 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6800 {
6801         switch (algo) {
6802         case ALGORITHM_PARITY_0:
6803                 if (raid_disk < max_degraded)
6804                         return 1;
6805                 break;
6806         case ALGORITHM_PARITY_N:
6807                 if (raid_disk >= raid_disks - max_degraded)
6808                         return 1;
6809                 break;
6810         case ALGORITHM_PARITY_0_6:
6811                 if (raid_disk == 0 ||
6812                     raid_disk == raid_disks - 1)
6813                         return 1;
6814                 break;
6815         case ALGORITHM_LEFT_ASYMMETRIC_6:
6816         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6817         case ALGORITHM_LEFT_SYMMETRIC_6:
6818         case ALGORITHM_RIGHT_SYMMETRIC_6:
6819                 if (raid_disk == raid_disks - 1)
6820                         return 1;
6821         }
6822         return 0;
6823 }
6824
6825 static int raid5_run(struct mddev *mddev)
6826 {
6827         struct r5conf *conf;
6828         int working_disks = 0;
6829         int dirty_parity_disks = 0;
6830         struct md_rdev *rdev;
6831         struct md_rdev *journal_dev = NULL;
6832         sector_t reshape_offset = 0;
6833         int i;
6834         long long min_offset_diff = 0;
6835         int first = 1;
6836
6837         if (mddev->recovery_cp != MaxSector)
6838                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
6839                           mdname(mddev));
6840
6841         rdev_for_each(rdev, mddev) {
6842                 long long diff;
6843
6844                 if (test_bit(Journal, &rdev->flags)) {
6845                         journal_dev = rdev;
6846                         continue;
6847                 }
6848                 if (rdev->raid_disk < 0)
6849                         continue;
6850                 diff = (rdev->new_data_offset - rdev->data_offset);
6851                 if (first) {
6852                         min_offset_diff = diff;
6853                         first = 0;
6854                 } else if (mddev->reshape_backwards &&
6855                          diff < min_offset_diff)
6856                         min_offset_diff = diff;
6857                 else if (!mddev->reshape_backwards &&
6858                          diff > min_offset_diff)
6859                         min_offset_diff = diff;
6860         }
6861
6862         if (mddev->reshape_position != MaxSector) {
6863                 /* Check that we can continue the reshape.
6864                  * Difficulties arise if the stripe we would write to
6865                  * next is at or after the stripe we would read from next.
6866                  * For a reshape that changes the number of devices, this
6867                  * is only possible for a very short time, and mdadm makes
6868                  * sure that time appears to have past before assembling
6869                  * the array.  So we fail if that time hasn't passed.
6870                  * For a reshape that keeps the number of devices the same
6871                  * mdadm must be monitoring the reshape can keeping the
6872                  * critical areas read-only and backed up.  It will start
6873                  * the array in read-only mode, so we check for that.
6874                  */
6875                 sector_t here_new, here_old;
6876                 int old_disks;
6877                 int max_degraded = (mddev->level == 6 ? 2 : 1);
6878                 int chunk_sectors;
6879                 int new_data_disks;
6880
6881                 if (journal_dev) {
6882                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
6883                                 mdname(mddev));
6884                         return -EINVAL;
6885                 }
6886
6887                 if (mddev->new_level != mddev->level) {
6888                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
6889                                 mdname(mddev));
6890                         return -EINVAL;
6891                 }
6892                 old_disks = mddev->raid_disks - mddev->delta_disks;
6893                 /* reshape_position must be on a new-stripe boundary, and one
6894                  * further up in new geometry must map after here in old
6895                  * geometry.
6896                  * If the chunk sizes are different, then as we perform reshape
6897                  * in units of the largest of the two, reshape_position needs
6898                  * be a multiple of the largest chunk size times new data disks.
6899                  */
6900                 here_new = mddev->reshape_position;
6901                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6902                 new_data_disks = mddev->raid_disks - max_degraded;
6903                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6904                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
6905                                 mdname(mddev));
6906                         return -EINVAL;
6907                 }
6908                 reshape_offset = here_new * chunk_sectors;
6909                 /* here_new is the stripe we will write to */
6910                 here_old = mddev->reshape_position;
6911                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6912                 /* here_old is the first stripe that we might need to read
6913                  * from */
6914                 if (mddev->delta_disks == 0) {
6915                         /* We cannot be sure it is safe to start an in-place
6916                          * reshape.  It is only safe if user-space is monitoring
6917                          * and taking constant backups.
6918                          * mdadm always starts a situation like this in
6919                          * readonly mode so it can take control before
6920                          * allowing any writes.  So just check for that.
6921                          */
6922                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6923                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
6924                                 /* not really in-place - so OK */;
6925                         else if (mddev->ro == 0) {
6926                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
6927                                         mdname(mddev));
6928                                 return -EINVAL;
6929                         }
6930                 } else if (mddev->reshape_backwards
6931                     ? (here_new * chunk_sectors + min_offset_diff <=
6932                        here_old * chunk_sectors)
6933                     : (here_new * chunk_sectors >=
6934                        here_old * chunk_sectors + (-min_offset_diff))) {
6935                         /* Reading from the same stripe as writing to - bad */
6936                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
6937                                 mdname(mddev));
6938                         return -EINVAL;
6939                 }
6940                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
6941                 /* OK, we should be able to continue; */
6942         } else {
6943                 BUG_ON(mddev->level != mddev->new_level);
6944                 BUG_ON(mddev->layout != mddev->new_layout);
6945                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6946                 BUG_ON(mddev->delta_disks != 0);
6947         }
6948
6949         if (mddev->private == NULL)
6950                 conf = setup_conf(mddev);
6951         else
6952                 conf = mddev->private;
6953
6954         if (IS_ERR(conf))
6955                 return PTR_ERR(conf);
6956
6957         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
6958                 if (!journal_dev) {
6959                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
6960                                 mdname(mddev));
6961                         mddev->ro = 1;
6962                         set_disk_ro(mddev->gendisk, 1);
6963                 } else if (mddev->recovery_cp == MaxSector)
6964                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
6965         }
6966
6967         conf->min_offset_diff = min_offset_diff;
6968         mddev->thread = conf->thread;
6969         conf->thread = NULL;
6970         mddev->private = conf;
6971
6972         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6973              i++) {
6974                 rdev = conf->disks[i].rdev;
6975                 if (!rdev && conf->disks[i].replacement) {
6976                         /* The replacement is all we have yet */
6977                         rdev = conf->disks[i].replacement;
6978                         conf->disks[i].replacement = NULL;
6979                         clear_bit(Replacement, &rdev->flags);
6980                         conf->disks[i].rdev = rdev;
6981                 }
6982                 if (!rdev)
6983                         continue;
6984                 if (conf->disks[i].replacement &&
6985                     conf->reshape_progress != MaxSector) {
6986                         /* replacements and reshape simply do not mix. */
6987                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
6988                         goto abort;
6989                 }
6990                 if (test_bit(In_sync, &rdev->flags)) {
6991                         working_disks++;
6992                         continue;
6993                 }
6994                 /* This disc is not fully in-sync.  However if it
6995                  * just stored parity (beyond the recovery_offset),
6996                  * when we don't need to be concerned about the
6997                  * array being dirty.
6998                  * When reshape goes 'backwards', we never have
6999                  * partially completed devices, so we only need
7000                  * to worry about reshape going forwards.
7001                  */
7002                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7003                 if (mddev->major_version == 0 &&
7004                     mddev->minor_version > 90)
7005                         rdev->recovery_offset = reshape_offset;
7006
7007                 if (rdev->recovery_offset < reshape_offset) {
7008                         /* We need to check old and new layout */
7009                         if (!only_parity(rdev->raid_disk,
7010                                          conf->algorithm,
7011                                          conf->raid_disks,
7012                                          conf->max_degraded))
7013                                 continue;
7014                 }
7015                 if (!only_parity(rdev->raid_disk,
7016                                  conf->prev_algo,
7017                                  conf->previous_raid_disks,
7018                                  conf->max_degraded))
7019                         continue;
7020                 dirty_parity_disks++;
7021         }
7022
7023         /*
7024          * 0 for a fully functional array, 1 or 2 for a degraded array.
7025          */
7026         mddev->degraded = calc_degraded(conf);
7027
7028         if (has_failed(conf)) {
7029                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7030                         mdname(mddev), mddev->degraded, conf->raid_disks);
7031                 goto abort;
7032         }
7033
7034         /* device size must be a multiple of chunk size */
7035         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7036         mddev->resync_max_sectors = mddev->dev_sectors;
7037
7038         if (mddev->degraded > dirty_parity_disks &&
7039             mddev->recovery_cp != MaxSector) {
7040                 if (mddev->ok_start_degraded)
7041                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7042                                 mdname(mddev));
7043                 else {
7044                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7045                                 mdname(mddev));
7046                         goto abort;
7047                 }
7048         }
7049
7050         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7051                 mdname(mddev), conf->level,
7052                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7053                 mddev->new_layout);
7054
7055         print_raid5_conf(conf);
7056
7057         if (conf->reshape_progress != MaxSector) {
7058                 conf->reshape_safe = conf->reshape_progress;
7059                 atomic_set(&conf->reshape_stripes, 0);
7060                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7061                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7062                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7063                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7064                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7065                                                         "reshape");
7066         }
7067
7068         /* Ok, everything is just fine now */
7069         if (mddev->to_remove == &raid5_attrs_group)
7070                 mddev->to_remove = NULL;
7071         else if (mddev->kobj.sd &&
7072             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7073                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7074                         mdname(mddev));
7075         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7076
7077         if (mddev->queue) {
7078                 int chunk_size;
7079                 bool discard_supported = true;
7080                 /* read-ahead size must cover two whole stripes, which
7081                  * is 2 * (datadisks) * chunksize where 'n' is the
7082                  * number of raid devices
7083                  */
7084                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7085                 int stripe = data_disks *
7086                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7087                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7088                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7089
7090                 chunk_size = mddev->chunk_sectors << 9;
7091                 blk_queue_io_min(mddev->queue, chunk_size);
7092                 blk_queue_io_opt(mddev->queue, chunk_size *
7093                                  (conf->raid_disks - conf->max_degraded));
7094                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7095                 /*
7096                  * We can only discard a whole stripe. It doesn't make sense to
7097                  * discard data disk but write parity disk
7098                  */
7099                 stripe = stripe * PAGE_SIZE;
7100                 /* Round up to power of 2, as discard handling
7101                  * currently assumes that */
7102                 while ((stripe-1) & stripe)
7103                         stripe = (stripe | (stripe-1)) + 1;
7104                 mddev->queue->limits.discard_alignment = stripe;
7105                 mddev->queue->limits.discard_granularity = stripe;
7106
7107                 /*
7108                  * We use 16-bit counter of active stripes in bi_phys_segments
7109                  * (minus one for over-loaded initialization)
7110                  */
7111                 blk_queue_max_hw_sectors(mddev->queue, 0xfffe * STRIPE_SECTORS);
7112                 blk_queue_max_discard_sectors(mddev->queue,
7113                                               0xfffe * STRIPE_SECTORS);
7114
7115                 /*
7116                  * unaligned part of discard request will be ignored, so can't
7117                  * guarantee discard_zeroes_data
7118                  */
7119                 mddev->queue->limits.discard_zeroes_data = 0;
7120
7121                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7122
7123                 rdev_for_each(rdev, mddev) {
7124                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7125                                           rdev->data_offset << 9);
7126                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7127                                           rdev->new_data_offset << 9);
7128                         /*
7129                          * discard_zeroes_data is required, otherwise data
7130                          * could be lost. Consider a scenario: discard a stripe
7131                          * (the stripe could be inconsistent if
7132                          * discard_zeroes_data is 0); write one disk of the
7133                          * stripe (the stripe could be inconsistent again
7134                          * depending on which disks are used to calculate
7135                          * parity); the disk is broken; The stripe data of this
7136                          * disk is lost.
7137                          */
7138                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7139                             !bdev_get_queue(rdev->bdev)->
7140                                                 limits.discard_zeroes_data)
7141                                 discard_supported = false;
7142                         /* Unfortunately, discard_zeroes_data is not currently
7143                          * a guarantee - just a hint.  So we only allow DISCARD
7144                          * if the sysadmin has confirmed that only safe devices
7145                          * are in use by setting a module parameter.
7146                          */
7147                         if (!devices_handle_discard_safely) {
7148                                 if (discard_supported) {
7149                                         pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7150                                         pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7151                                 }
7152                                 discard_supported = false;
7153                         }
7154                 }
7155
7156                 if (discard_supported &&
7157                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7158                     mddev->queue->limits.discard_granularity >= stripe)
7159                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7160                                                 mddev->queue);
7161                 else
7162                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7163                                                 mddev->queue);
7164
7165                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7166         }
7167
7168         if (journal_dev) {
7169                 char b[BDEVNAME_SIZE];
7170
7171                 pr_debug("md/raid:%s: using device %s as journal\n",
7172                          mdname(mddev), bdevname(journal_dev->bdev, b));
7173                 if (r5l_init_log(conf, journal_dev))
7174                         goto abort;
7175         }
7176
7177         return 0;
7178 abort:
7179         md_unregister_thread(&mddev->thread);
7180         print_raid5_conf(conf);
7181         free_conf(conf);
7182         mddev->private = NULL;
7183         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7184         return -EIO;
7185 }
7186
7187 static void raid5_free(struct mddev *mddev, void *priv)
7188 {
7189         struct r5conf *conf = priv;
7190
7191         free_conf(conf);
7192         mddev->to_remove = &raid5_attrs_group;
7193 }
7194
7195 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7196 {
7197         struct r5conf *conf = mddev->private;
7198         int i;
7199
7200         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7201                 conf->chunk_sectors / 2, mddev->layout);
7202         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7203         rcu_read_lock();
7204         for (i = 0; i < conf->raid_disks; i++) {
7205                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7206                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7207         }
7208         rcu_read_unlock();
7209         seq_printf (seq, "]");
7210 }
7211
7212 static void print_raid5_conf (struct r5conf *conf)
7213 {
7214         int i;
7215         struct disk_info *tmp;
7216
7217         pr_debug("RAID conf printout:\n");
7218         if (!conf) {
7219                 pr_debug("(conf==NULL)\n");
7220                 return;
7221         }
7222         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7223                conf->raid_disks,
7224                conf->raid_disks - conf->mddev->degraded);
7225
7226         for (i = 0; i < conf->raid_disks; i++) {
7227                 char b[BDEVNAME_SIZE];
7228                 tmp = conf->disks + i;
7229                 if (tmp->rdev)
7230                         pr_debug(" disk %d, o:%d, dev:%s\n",
7231                                i, !test_bit(Faulty, &tmp->rdev->flags),
7232                                bdevname(tmp->rdev->bdev, b));
7233         }
7234 }
7235
7236 static int raid5_spare_active(struct mddev *mddev)
7237 {
7238         int i;
7239         struct r5conf *conf = mddev->private;
7240         struct disk_info *tmp;
7241         int count = 0;
7242         unsigned long flags;
7243
7244         for (i = 0; i < conf->raid_disks; i++) {
7245                 tmp = conf->disks + i;
7246                 if (tmp->replacement
7247                     && tmp->replacement->recovery_offset == MaxSector
7248                     && !test_bit(Faulty, &tmp->replacement->flags)
7249                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7250                         /* Replacement has just become active. */
7251                         if (!tmp->rdev
7252                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7253                                 count++;
7254                         if (tmp->rdev) {
7255                                 /* Replaced device not technically faulty,
7256                                  * but we need to be sure it gets removed
7257                                  * and never re-added.
7258                                  */
7259                                 set_bit(Faulty, &tmp->rdev->flags);
7260                                 sysfs_notify_dirent_safe(
7261                                         tmp->rdev->sysfs_state);
7262                         }
7263                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7264                 } else if (tmp->rdev
7265                     && tmp->rdev->recovery_offset == MaxSector
7266                     && !test_bit(Faulty, &tmp->rdev->flags)
7267                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7268                         count++;
7269                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7270                 }
7271         }
7272         spin_lock_irqsave(&conf->device_lock, flags);
7273         mddev->degraded = calc_degraded(conf);
7274         spin_unlock_irqrestore(&conf->device_lock, flags);
7275         print_raid5_conf(conf);
7276         return count;
7277 }
7278
7279 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7280 {
7281         struct r5conf *conf = mddev->private;
7282         int err = 0;
7283         int number = rdev->raid_disk;
7284         struct md_rdev **rdevp;
7285         struct disk_info *p = conf->disks + number;
7286
7287         print_raid5_conf(conf);
7288         if (test_bit(Journal, &rdev->flags) && conf->log) {
7289                 struct r5l_log *log;
7290                 /*
7291                  * we can't wait pending write here, as this is called in
7292                  * raid5d, wait will deadlock.
7293                  */
7294                 if (atomic_read(&mddev->writes_pending))
7295                         return -EBUSY;
7296                 log = conf->log;
7297                 conf->log = NULL;
7298                 synchronize_rcu();
7299                 r5l_exit_log(log);
7300                 return 0;
7301         }
7302         if (rdev == p->rdev)
7303                 rdevp = &p->rdev;
7304         else if (rdev == p->replacement)
7305                 rdevp = &p->replacement;
7306         else
7307                 return 0;
7308
7309         if (number >= conf->raid_disks &&
7310             conf->reshape_progress == MaxSector)
7311                 clear_bit(In_sync, &rdev->flags);
7312
7313         if (test_bit(In_sync, &rdev->flags) ||
7314             atomic_read(&rdev->nr_pending)) {
7315                 err = -EBUSY;
7316                 goto abort;
7317         }
7318         /* Only remove non-faulty devices if recovery
7319          * isn't possible.
7320          */
7321         if (!test_bit(Faulty, &rdev->flags) &&
7322             mddev->recovery_disabled != conf->recovery_disabled &&
7323             !has_failed(conf) &&
7324             (!p->replacement || p->replacement == rdev) &&
7325             number < conf->raid_disks) {
7326                 err = -EBUSY;
7327                 goto abort;
7328         }
7329         *rdevp = NULL;
7330         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7331                 synchronize_rcu();
7332                 if (atomic_read(&rdev->nr_pending)) {
7333                         /* lost the race, try later */
7334                         err = -EBUSY;
7335                         *rdevp = rdev;
7336                 }
7337         }
7338         if (p->replacement) {
7339                 /* We must have just cleared 'rdev' */
7340                 p->rdev = p->replacement;
7341                 clear_bit(Replacement, &p->replacement->flags);
7342                 smp_mb(); /* Make sure other CPUs may see both as identical
7343                            * but will never see neither - if they are careful
7344                            */
7345                 p->replacement = NULL;
7346                 clear_bit(WantReplacement, &rdev->flags);
7347         } else
7348                 /* We might have just removed the Replacement as faulty-
7349                  * clear the bit just in case
7350                  */
7351                 clear_bit(WantReplacement, &rdev->flags);
7352 abort:
7353
7354         print_raid5_conf(conf);
7355         return err;
7356 }
7357
7358 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7359 {
7360         struct r5conf *conf = mddev->private;
7361         int err = -EEXIST;
7362         int disk;
7363         struct disk_info *p;
7364         int first = 0;
7365         int last = conf->raid_disks - 1;
7366
7367         if (test_bit(Journal, &rdev->flags)) {
7368                 char b[BDEVNAME_SIZE];
7369                 if (conf->log)
7370                         return -EBUSY;
7371
7372                 rdev->raid_disk = 0;
7373                 /*
7374                  * The array is in readonly mode if journal is missing, so no
7375                  * write requests running. We should be safe
7376                  */
7377                 r5l_init_log(conf, rdev);
7378                 pr_debug("md/raid:%s: using device %s as journal\n",
7379                          mdname(mddev), bdevname(rdev->bdev, b));
7380                 return 0;
7381         }
7382         if (mddev->recovery_disabled == conf->recovery_disabled)
7383                 return -EBUSY;
7384
7385         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7386                 /* no point adding a device */
7387                 return -EINVAL;
7388
7389         if (rdev->raid_disk >= 0)
7390                 first = last = rdev->raid_disk;
7391
7392         /*
7393          * find the disk ... but prefer rdev->saved_raid_disk
7394          * if possible.
7395          */
7396         if (rdev->saved_raid_disk >= 0 &&
7397             rdev->saved_raid_disk >= first &&
7398             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7399                 first = rdev->saved_raid_disk;
7400
7401         for (disk = first; disk <= last; disk++) {
7402                 p = conf->disks + disk;
7403                 if (p->rdev == NULL) {
7404                         clear_bit(In_sync, &rdev->flags);
7405                         rdev->raid_disk = disk;
7406                         err = 0;
7407                         if (rdev->saved_raid_disk != disk)
7408                                 conf->fullsync = 1;
7409                         rcu_assign_pointer(p->rdev, rdev);
7410                         goto out;
7411                 }
7412         }
7413         for (disk = first; disk <= last; disk++) {
7414                 p = conf->disks + disk;
7415                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7416                     p->replacement == NULL) {
7417                         clear_bit(In_sync, &rdev->flags);
7418                         set_bit(Replacement, &rdev->flags);
7419                         rdev->raid_disk = disk;
7420                         err = 0;
7421                         conf->fullsync = 1;
7422                         rcu_assign_pointer(p->replacement, rdev);
7423                         break;
7424                 }
7425         }
7426 out:
7427         print_raid5_conf(conf);
7428         return err;
7429 }
7430
7431 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7432 {
7433         /* no resync is happening, and there is enough space
7434          * on all devices, so we can resize.
7435          * We need to make sure resync covers any new space.
7436          * If the array is shrinking we should possibly wait until
7437          * any io in the removed space completes, but it hardly seems
7438          * worth it.
7439          */
7440         sector_t newsize;
7441         struct r5conf *conf = mddev->private;
7442
7443         if (conf->log)
7444                 return -EINVAL;
7445         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7446         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7447         if (mddev->external_size &&
7448             mddev->array_sectors > newsize)
7449                 return -EINVAL;
7450         if (mddev->bitmap) {
7451                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7452                 if (ret)
7453                         return ret;
7454         }
7455         md_set_array_sectors(mddev, newsize);
7456         set_capacity(mddev->gendisk, mddev->array_sectors);
7457         revalidate_disk(mddev->gendisk);
7458         if (sectors > mddev->dev_sectors &&
7459             mddev->recovery_cp > mddev->dev_sectors) {
7460                 mddev->recovery_cp = mddev->dev_sectors;
7461                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7462         }
7463         mddev->dev_sectors = sectors;
7464         mddev->resync_max_sectors = sectors;
7465         return 0;
7466 }
7467
7468 static int check_stripe_cache(struct mddev *mddev)
7469 {
7470         /* Can only proceed if there are plenty of stripe_heads.
7471          * We need a minimum of one full stripe,, and for sensible progress
7472          * it is best to have about 4 times that.
7473          * If we require 4 times, then the default 256 4K stripe_heads will
7474          * allow for chunk sizes up to 256K, which is probably OK.
7475          * If the chunk size is greater, user-space should request more
7476          * stripe_heads first.
7477          */
7478         struct r5conf *conf = mddev->private;
7479         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7480             > conf->min_nr_stripes ||
7481             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7482             > conf->min_nr_stripes) {
7483                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7484                         mdname(mddev),
7485                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7486                          / STRIPE_SIZE)*4);
7487                 return 0;
7488         }
7489         return 1;
7490 }
7491
7492 static int check_reshape(struct mddev *mddev)
7493 {
7494         struct r5conf *conf = mddev->private;
7495
7496         if (conf->log)
7497                 return -EINVAL;
7498         if (mddev->delta_disks == 0 &&
7499             mddev->new_layout == mddev->layout &&
7500             mddev->new_chunk_sectors == mddev->chunk_sectors)
7501                 return 0; /* nothing to do */
7502         if (has_failed(conf))
7503                 return -EINVAL;
7504         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7505                 /* We might be able to shrink, but the devices must
7506                  * be made bigger first.
7507                  * For raid6, 4 is the minimum size.
7508                  * Otherwise 2 is the minimum
7509                  */
7510                 int min = 2;
7511                 if (mddev->level == 6)
7512                         min = 4;
7513                 if (mddev->raid_disks + mddev->delta_disks < min)
7514                         return -EINVAL;
7515         }
7516
7517         if (!check_stripe_cache(mddev))
7518                 return -ENOSPC;
7519
7520         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7521             mddev->delta_disks > 0)
7522                 if (resize_chunks(conf,
7523                                   conf->previous_raid_disks
7524                                   + max(0, mddev->delta_disks),
7525                                   max(mddev->new_chunk_sectors,
7526                                       mddev->chunk_sectors)
7527                             ) < 0)
7528                         return -ENOMEM;
7529         return resize_stripes(conf, (conf->previous_raid_disks
7530                                      + mddev->delta_disks));
7531 }
7532
7533 static int raid5_start_reshape(struct mddev *mddev)
7534 {
7535         struct r5conf *conf = mddev->private;
7536         struct md_rdev *rdev;
7537         int spares = 0;
7538         unsigned long flags;
7539
7540         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7541                 return -EBUSY;
7542
7543         if (!check_stripe_cache(mddev))
7544                 return -ENOSPC;
7545
7546         if (has_failed(conf))
7547                 return -EINVAL;
7548
7549         rdev_for_each(rdev, mddev) {
7550                 if (!test_bit(In_sync, &rdev->flags)
7551                     && !test_bit(Faulty, &rdev->flags))
7552                         spares++;
7553         }
7554
7555         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7556                 /* Not enough devices even to make a degraded array
7557                  * of that size
7558                  */
7559                 return -EINVAL;
7560
7561         /* Refuse to reduce size of the array.  Any reductions in
7562          * array size must be through explicit setting of array_size
7563          * attribute.
7564          */
7565         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7566             < mddev->array_sectors) {
7567                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7568                         mdname(mddev));
7569                 return -EINVAL;
7570         }
7571
7572         atomic_set(&conf->reshape_stripes, 0);
7573         spin_lock_irq(&conf->device_lock);
7574         write_seqcount_begin(&conf->gen_lock);
7575         conf->previous_raid_disks = conf->raid_disks;
7576         conf->raid_disks += mddev->delta_disks;
7577         conf->prev_chunk_sectors = conf->chunk_sectors;
7578         conf->chunk_sectors = mddev->new_chunk_sectors;
7579         conf->prev_algo = conf->algorithm;
7580         conf->algorithm = mddev->new_layout;
7581         conf->generation++;
7582         /* Code that selects data_offset needs to see the generation update
7583          * if reshape_progress has been set - so a memory barrier needed.
7584          */
7585         smp_mb();
7586         if (mddev->reshape_backwards)
7587                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7588         else
7589                 conf->reshape_progress = 0;
7590         conf->reshape_safe = conf->reshape_progress;
7591         write_seqcount_end(&conf->gen_lock);
7592         spin_unlock_irq(&conf->device_lock);
7593
7594         /* Now make sure any requests that proceeded on the assumption
7595          * the reshape wasn't running - like Discard or Read - have
7596          * completed.
7597          */
7598         mddev_suspend(mddev);
7599         mddev_resume(mddev);
7600
7601         /* Add some new drives, as many as will fit.
7602          * We know there are enough to make the newly sized array work.
7603          * Don't add devices if we are reducing the number of
7604          * devices in the array.  This is because it is not possible
7605          * to correctly record the "partially reconstructed" state of
7606          * such devices during the reshape and confusion could result.
7607          */
7608         if (mddev->delta_disks >= 0) {
7609                 rdev_for_each(rdev, mddev)
7610                         if (rdev->raid_disk < 0 &&
7611                             !test_bit(Faulty, &rdev->flags)) {
7612                                 if (raid5_add_disk(mddev, rdev) == 0) {
7613                                         if (rdev->raid_disk
7614                                             >= conf->previous_raid_disks)
7615                                                 set_bit(In_sync, &rdev->flags);
7616                                         else
7617                                                 rdev->recovery_offset = 0;
7618
7619                                         if (sysfs_link_rdev(mddev, rdev))
7620                                                 /* Failure here is OK */;
7621                                 }
7622                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7623                                    && !test_bit(Faulty, &rdev->flags)) {
7624                                 /* This is a spare that was manually added */
7625                                 set_bit(In_sync, &rdev->flags);
7626                         }
7627
7628                 /* When a reshape changes the number of devices,
7629                  * ->degraded is measured against the larger of the
7630                  * pre and post number of devices.
7631                  */
7632                 spin_lock_irqsave(&conf->device_lock, flags);
7633                 mddev->degraded = calc_degraded(conf);
7634                 spin_unlock_irqrestore(&conf->device_lock, flags);
7635         }
7636         mddev->raid_disks = conf->raid_disks;
7637         mddev->reshape_position = conf->reshape_progress;
7638         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7639
7640         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7641         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7642         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7643         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7644         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7645         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7646                                                 "reshape");
7647         if (!mddev->sync_thread) {
7648                 mddev->recovery = 0;
7649                 spin_lock_irq(&conf->device_lock);
7650                 write_seqcount_begin(&conf->gen_lock);
7651                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7652                 mddev->new_chunk_sectors =
7653                         conf->chunk_sectors = conf->prev_chunk_sectors;
7654                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7655                 rdev_for_each(rdev, mddev)
7656                         rdev->new_data_offset = rdev->data_offset;
7657                 smp_wmb();
7658                 conf->generation --;
7659                 conf->reshape_progress = MaxSector;
7660                 mddev->reshape_position = MaxSector;
7661                 write_seqcount_end(&conf->gen_lock);
7662                 spin_unlock_irq(&conf->device_lock);
7663                 return -EAGAIN;
7664         }
7665         conf->reshape_checkpoint = jiffies;
7666         md_wakeup_thread(mddev->sync_thread);
7667         md_new_event(mddev);
7668         return 0;
7669 }
7670
7671 /* This is called from the reshape thread and should make any
7672  * changes needed in 'conf'
7673  */
7674 static void end_reshape(struct r5conf *conf)
7675 {
7676
7677         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7678                 struct md_rdev *rdev;
7679
7680                 spin_lock_irq(&conf->device_lock);
7681                 conf->previous_raid_disks = conf->raid_disks;
7682                 rdev_for_each(rdev, conf->mddev)
7683                         rdev->data_offset = rdev->new_data_offset;
7684                 smp_wmb();
7685                 conf->reshape_progress = MaxSector;
7686                 conf->mddev->reshape_position = MaxSector;
7687                 spin_unlock_irq(&conf->device_lock);
7688                 wake_up(&conf->wait_for_overlap);
7689
7690                 /* read-ahead size must cover two whole stripes, which is
7691                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7692                  */
7693                 if (conf->mddev->queue) {
7694                         int data_disks = conf->raid_disks - conf->max_degraded;
7695                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7696                                                    / PAGE_SIZE);
7697                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7698                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7699                 }
7700         }
7701 }
7702
7703 /* This is called from the raid5d thread with mddev_lock held.
7704  * It makes config changes to the device.
7705  */
7706 static void raid5_finish_reshape(struct mddev *mddev)
7707 {
7708         struct r5conf *conf = mddev->private;
7709
7710         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7711
7712                 if (mddev->delta_disks > 0) {
7713                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7714                         if (mddev->queue) {
7715                                 set_capacity(mddev->gendisk, mddev->array_sectors);
7716                                 revalidate_disk(mddev->gendisk);
7717                         }
7718                 } else {
7719                         int d;
7720                         spin_lock_irq(&conf->device_lock);
7721                         mddev->degraded = calc_degraded(conf);
7722                         spin_unlock_irq(&conf->device_lock);
7723                         for (d = conf->raid_disks ;
7724                              d < conf->raid_disks - mddev->delta_disks;
7725                              d++) {
7726                                 struct md_rdev *rdev = conf->disks[d].rdev;
7727                                 if (rdev)
7728                                         clear_bit(In_sync, &rdev->flags);
7729                                 rdev = conf->disks[d].replacement;
7730                                 if (rdev)
7731                                         clear_bit(In_sync, &rdev->flags);
7732                         }
7733                 }
7734                 mddev->layout = conf->algorithm;
7735                 mddev->chunk_sectors = conf->chunk_sectors;
7736                 mddev->reshape_position = MaxSector;
7737                 mddev->delta_disks = 0;
7738                 mddev->reshape_backwards = 0;
7739         }
7740 }
7741
7742 static void raid5_quiesce(struct mddev *mddev, int state)
7743 {
7744         struct r5conf *conf = mddev->private;
7745
7746         switch(state) {
7747         case 2: /* resume for a suspend */
7748                 wake_up(&conf->wait_for_overlap);
7749                 break;
7750
7751         case 1: /* stop all writes */
7752                 lock_all_device_hash_locks_irq(conf);
7753                 /* '2' tells resync/reshape to pause so that all
7754                  * active stripes can drain
7755                  */
7756                 r5c_flush_cache(conf, INT_MAX);
7757                 conf->quiesce = 2;
7758                 wait_event_cmd(conf->wait_for_quiescent,
7759                                     atomic_read(&conf->active_stripes) == 0 &&
7760                                     atomic_read(&conf->active_aligned_reads) == 0,
7761                                     unlock_all_device_hash_locks_irq(conf),
7762                                     lock_all_device_hash_locks_irq(conf));
7763                 conf->quiesce = 1;
7764                 unlock_all_device_hash_locks_irq(conf);
7765                 /* allow reshape to continue */
7766                 wake_up(&conf->wait_for_overlap);
7767                 break;
7768
7769         case 0: /* re-enable writes */
7770                 lock_all_device_hash_locks_irq(conf);
7771                 conf->quiesce = 0;
7772                 wake_up(&conf->wait_for_quiescent);
7773                 wake_up(&conf->wait_for_overlap);
7774                 unlock_all_device_hash_locks_irq(conf);
7775                 break;
7776         }
7777         r5l_quiesce(conf->log, state);
7778 }
7779
7780 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7781 {
7782         struct r0conf *raid0_conf = mddev->private;
7783         sector_t sectors;
7784
7785         /* for raid0 takeover only one zone is supported */
7786         if (raid0_conf->nr_strip_zones > 1) {
7787                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7788                         mdname(mddev));
7789                 return ERR_PTR(-EINVAL);
7790         }
7791
7792         sectors = raid0_conf->strip_zone[0].zone_end;
7793         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7794         mddev->dev_sectors = sectors;
7795         mddev->new_level = level;
7796         mddev->new_layout = ALGORITHM_PARITY_N;
7797         mddev->new_chunk_sectors = mddev->chunk_sectors;
7798         mddev->raid_disks += 1;
7799         mddev->delta_disks = 1;
7800         /* make sure it will be not marked as dirty */
7801         mddev->recovery_cp = MaxSector;
7802
7803         return setup_conf(mddev);
7804 }
7805
7806 static void *raid5_takeover_raid1(struct mddev *mddev)
7807 {
7808         int chunksect;
7809         void *ret;
7810
7811         if (mddev->raid_disks != 2 ||
7812             mddev->degraded > 1)
7813                 return ERR_PTR(-EINVAL);
7814
7815         /* Should check if there are write-behind devices? */
7816
7817         chunksect = 64*2; /* 64K by default */
7818
7819         /* The array must be an exact multiple of chunksize */
7820         while (chunksect && (mddev->array_sectors & (chunksect-1)))
7821                 chunksect >>= 1;
7822
7823         if ((chunksect<<9) < STRIPE_SIZE)
7824                 /* array size does not allow a suitable chunk size */
7825                 return ERR_PTR(-EINVAL);
7826
7827         mddev->new_level = 5;
7828         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7829         mddev->new_chunk_sectors = chunksect;
7830
7831         ret = setup_conf(mddev);
7832         if (!IS_ERR_VALUE(ret))
7833                 clear_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
7834         return ret;
7835 }
7836
7837 static void *raid5_takeover_raid6(struct mddev *mddev)
7838 {
7839         int new_layout;
7840
7841         switch (mddev->layout) {
7842         case ALGORITHM_LEFT_ASYMMETRIC_6:
7843                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7844                 break;
7845         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7846                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7847                 break;
7848         case ALGORITHM_LEFT_SYMMETRIC_6:
7849                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7850                 break;
7851         case ALGORITHM_RIGHT_SYMMETRIC_6:
7852                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7853                 break;
7854         case ALGORITHM_PARITY_0_6:
7855                 new_layout = ALGORITHM_PARITY_0;
7856                 break;
7857         case ALGORITHM_PARITY_N:
7858                 new_layout = ALGORITHM_PARITY_N;
7859                 break;
7860         default:
7861                 return ERR_PTR(-EINVAL);
7862         }
7863         mddev->new_level = 5;
7864         mddev->new_layout = new_layout;
7865         mddev->delta_disks = -1;
7866         mddev->raid_disks -= 1;
7867         return setup_conf(mddev);
7868 }
7869
7870 static int raid5_check_reshape(struct mddev *mddev)
7871 {
7872         /* For a 2-drive array, the layout and chunk size can be changed
7873          * immediately as not restriping is needed.
7874          * For larger arrays we record the new value - after validation
7875          * to be used by a reshape pass.
7876          */
7877         struct r5conf *conf = mddev->private;
7878         int new_chunk = mddev->new_chunk_sectors;
7879
7880         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7881                 return -EINVAL;
7882         if (new_chunk > 0) {
7883                 if (!is_power_of_2(new_chunk))
7884                         return -EINVAL;
7885                 if (new_chunk < (PAGE_SIZE>>9))
7886                         return -EINVAL;
7887                 if (mddev->array_sectors & (new_chunk-1))
7888                         /* not factor of array size */
7889                         return -EINVAL;
7890         }
7891
7892         /* They look valid */
7893
7894         if (mddev->raid_disks == 2) {
7895                 /* can make the change immediately */
7896                 if (mddev->new_layout >= 0) {
7897                         conf->algorithm = mddev->new_layout;
7898                         mddev->layout = mddev->new_layout;
7899                 }
7900                 if (new_chunk > 0) {
7901                         conf->chunk_sectors = new_chunk ;
7902                         mddev->chunk_sectors = new_chunk;
7903                 }
7904                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7905                 md_wakeup_thread(mddev->thread);
7906         }
7907         return check_reshape(mddev);
7908 }
7909
7910 static int raid6_check_reshape(struct mddev *mddev)
7911 {
7912         int new_chunk = mddev->new_chunk_sectors;
7913
7914         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7915                 return -EINVAL;
7916         if (new_chunk > 0) {
7917                 if (!is_power_of_2(new_chunk))
7918                         return -EINVAL;
7919                 if (new_chunk < (PAGE_SIZE >> 9))
7920                         return -EINVAL;
7921                 if (mddev->array_sectors & (new_chunk-1))
7922                         /* not factor of array size */
7923                         return -EINVAL;
7924         }
7925
7926         /* They look valid */
7927         return check_reshape(mddev);
7928 }
7929
7930 static void *raid5_takeover(struct mddev *mddev)
7931 {
7932         /* raid5 can take over:
7933          *  raid0 - if there is only one strip zone - make it a raid4 layout
7934          *  raid1 - if there are two drives.  We need to know the chunk size
7935          *  raid4 - trivial - just use a raid4 layout.
7936          *  raid6 - Providing it is a *_6 layout
7937          */
7938         if (mddev->level == 0)
7939                 return raid45_takeover_raid0(mddev, 5);
7940         if (mddev->level == 1)
7941                 return raid5_takeover_raid1(mddev);
7942         if (mddev->level == 4) {
7943                 mddev->new_layout = ALGORITHM_PARITY_N;
7944                 mddev->new_level = 5;
7945                 return setup_conf(mddev);
7946         }
7947         if (mddev->level == 6)
7948                 return raid5_takeover_raid6(mddev);
7949
7950         return ERR_PTR(-EINVAL);
7951 }
7952
7953 static void *raid4_takeover(struct mddev *mddev)
7954 {
7955         /* raid4 can take over:
7956          *  raid0 - if there is only one strip zone
7957          *  raid5 - if layout is right
7958          */
7959         if (mddev->level == 0)
7960                 return raid45_takeover_raid0(mddev, 4);
7961         if (mddev->level == 5 &&
7962             mddev->layout == ALGORITHM_PARITY_N) {
7963                 mddev->new_layout = 0;
7964                 mddev->new_level = 4;
7965                 return setup_conf(mddev);
7966         }
7967         return ERR_PTR(-EINVAL);
7968 }
7969
7970 static struct md_personality raid5_personality;
7971
7972 static void *raid6_takeover(struct mddev *mddev)
7973 {
7974         /* Currently can only take over a raid5.  We map the
7975          * personality to an equivalent raid6 personality
7976          * with the Q block at the end.
7977          */
7978         int new_layout;
7979
7980         if (mddev->pers != &raid5_personality)
7981                 return ERR_PTR(-EINVAL);
7982         if (mddev->degraded > 1)
7983                 return ERR_PTR(-EINVAL);
7984         if (mddev->raid_disks > 253)
7985                 return ERR_PTR(-EINVAL);
7986         if (mddev->raid_disks < 3)
7987                 return ERR_PTR(-EINVAL);
7988
7989         switch (mddev->layout) {
7990         case ALGORITHM_LEFT_ASYMMETRIC:
7991                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7992                 break;
7993         case ALGORITHM_RIGHT_ASYMMETRIC:
7994                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7995                 break;
7996         case ALGORITHM_LEFT_SYMMETRIC:
7997                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7998                 break;
7999         case ALGORITHM_RIGHT_SYMMETRIC:
8000                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8001                 break;
8002         case ALGORITHM_PARITY_0:
8003                 new_layout = ALGORITHM_PARITY_0_6;
8004                 break;
8005         case ALGORITHM_PARITY_N:
8006                 new_layout = ALGORITHM_PARITY_N;
8007                 break;
8008         default:
8009                 return ERR_PTR(-EINVAL);
8010         }
8011         mddev->new_level = 6;
8012         mddev->new_layout = new_layout;
8013         mddev->delta_disks = 1;
8014         mddev->raid_disks += 1;
8015         return setup_conf(mddev);
8016 }
8017
8018 static struct md_personality raid6_personality =
8019 {
8020         .name           = "raid6",
8021         .level          = 6,
8022         .owner          = THIS_MODULE,
8023         .make_request   = raid5_make_request,
8024         .run            = raid5_run,
8025         .free           = raid5_free,
8026         .status         = raid5_status,
8027         .error_handler  = raid5_error,
8028         .hot_add_disk   = raid5_add_disk,
8029         .hot_remove_disk= raid5_remove_disk,
8030         .spare_active   = raid5_spare_active,
8031         .sync_request   = raid5_sync_request,
8032         .resize         = raid5_resize,
8033         .size           = raid5_size,
8034         .check_reshape  = raid6_check_reshape,
8035         .start_reshape  = raid5_start_reshape,
8036         .finish_reshape = raid5_finish_reshape,
8037         .quiesce        = raid5_quiesce,
8038         .takeover       = raid6_takeover,
8039         .congested      = raid5_congested,
8040 };
8041 static struct md_personality raid5_personality =
8042 {
8043         .name           = "raid5",
8044         .level          = 5,
8045         .owner          = THIS_MODULE,
8046         .make_request   = raid5_make_request,
8047         .run            = raid5_run,
8048         .free           = raid5_free,
8049         .status         = raid5_status,
8050         .error_handler  = raid5_error,
8051         .hot_add_disk   = raid5_add_disk,
8052         .hot_remove_disk= raid5_remove_disk,
8053         .spare_active   = raid5_spare_active,
8054         .sync_request   = raid5_sync_request,
8055         .resize         = raid5_resize,
8056         .size           = raid5_size,
8057         .check_reshape  = raid5_check_reshape,
8058         .start_reshape  = raid5_start_reshape,
8059         .finish_reshape = raid5_finish_reshape,
8060         .quiesce        = raid5_quiesce,
8061         .takeover       = raid5_takeover,
8062         .congested      = raid5_congested,
8063 };
8064
8065 static struct md_personality raid4_personality =
8066 {
8067         .name           = "raid4",
8068         .level          = 4,
8069         .owner          = THIS_MODULE,
8070         .make_request   = raid5_make_request,
8071         .run            = raid5_run,
8072         .free           = raid5_free,
8073         .status         = raid5_status,
8074         .error_handler  = raid5_error,
8075         .hot_add_disk   = raid5_add_disk,
8076         .hot_remove_disk= raid5_remove_disk,
8077         .spare_active   = raid5_spare_active,
8078         .sync_request   = raid5_sync_request,
8079         .resize         = raid5_resize,
8080         .size           = raid5_size,
8081         .check_reshape  = raid5_check_reshape,
8082         .start_reshape  = raid5_start_reshape,
8083         .finish_reshape = raid5_finish_reshape,
8084         .quiesce        = raid5_quiesce,
8085         .takeover       = raid4_takeover,
8086         .congested      = raid5_congested,
8087 };
8088
8089 static int __init raid5_init(void)
8090 {
8091         int ret;
8092
8093         raid5_wq = alloc_workqueue("raid5wq",
8094                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8095         if (!raid5_wq)
8096                 return -ENOMEM;
8097
8098         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8099                                       "md/raid5:prepare",
8100                                       raid456_cpu_up_prepare,
8101                                       raid456_cpu_dead);
8102         if (ret) {
8103                 destroy_workqueue(raid5_wq);
8104                 return ret;
8105         }
8106         register_md_personality(&raid6_personality);
8107         register_md_personality(&raid5_personality);
8108         register_md_personality(&raid4_personality);
8109         return 0;
8110 }
8111
8112 static void raid5_exit(void)
8113 {
8114         unregister_md_personality(&raid6_personality);
8115         unregister_md_personality(&raid5_personality);
8116         unregister_md_personality(&raid4_personality);
8117         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8118         destroy_workqueue(raid5_wq);
8119 }
8120
8121 module_init(raid5_init);
8122 module_exit(raid5_exit);
8123 MODULE_LICENSE("GPL");
8124 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8125 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8126 MODULE_ALIAS("md-raid5");
8127 MODULE_ALIAS("md-raid4");
8128 MODULE_ALIAS("md-level-5");
8129 MODULE_ALIAS("md-level-4");
8130 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8131 MODULE_ALIAS("md-raid6");
8132 MODULE_ALIAS("md-level-6");
8133
8134 /* This used to be two separate modules, they were: */
8135 MODULE_ALIAS("raid5");
8136 MODULE_ALIAS("raid6");