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 UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  66
  67 #define cpu_to_group(cpu) cpu_to_node(cpu)
  68 #define ANY_GROUP NUMA_NO_NODE
  69
  70 static bool devices_handle_discard_safely = false;
  71 module_param(devices_handle_discard_safely, bool, 0644);
  72 MODULE_PARM_DESC(devices_handle_discard_safely,
  73                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  74 static struct workqueue_struct *raid5_wq;
  75
  76 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  77 {
  78         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  79         return &conf->stripe_hashtbl[hash];
  80 }
  81
  82 static inline int stripe_hash_locks_hash(sector_t sect)
  83 {
  84         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  85 }
  86
  87 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  88 {
  89         spin_lock_irq(conf->hash_locks + hash);
  90         spin_lock(&conf->device_lock);
  91 }
  92
  93 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  94 {
  95         spin_unlock(&conf->device_lock);
  96         spin_unlock_irq(conf->hash_locks + hash);
  97 }
  98
  99 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 100 {
 101         int i;
 102         local_irq_disable();
 103         spin_lock(conf->hash_locks);
 104         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 105                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 106         spin_lock(&conf->device_lock);
 107 }
 108
 109 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 110 {
 111         int i;
 112         spin_unlock(&conf->device_lock);
 113         for (i = NR_STRIPE_HASH_LOCKS; i; i--)
 114                 spin_unlock(conf->hash_locks + i - 1);
 115         local_irq_enable();
 116 }
 117
 118 /* Find first data disk in a raid6 stripe */
 119 static inline int raid6_d0(struct stripe_head *sh)
 120 {
 121         if (sh->ddf_layout)
 122                 /* ddf always start from first device */
 123                 return 0;
 124         /* md starts just after Q block */
 125         if (sh->qd_idx == sh->disks - 1)
 126                 return 0;
 127         else
 128                 return sh->qd_idx + 1;
 129 }
 130 static inline int raid6_next_disk(int disk, int raid_disks)
 131 {
 132         disk++;
 133         return (disk < raid_disks) ? disk : 0;
 134 }
 135
 136 /* When walking through the disks in a raid5, starting at raid6_d0,
 137  * We need to map each disk to a 'slot', where the data disks are slot
 138  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 139  * is raid_disks-1.  This help does that mapping.
 140  */
 141 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 142                              int *count, int syndrome_disks)
 143 {
 144         int slot = *count;
 145
 146         if (sh->ddf_layout)
 147                 (*count)++;
 148         if (idx == sh->pd_idx)
 149                 return syndrome_disks;
 150         if (idx == sh->qd_idx)
 151                 return syndrome_disks + 1;
 152         if (!sh->ddf_layout)
 153                 (*count)++;
 154         return slot;
 155 }
 156
 157 static void return_io(struct bio_list *return_bi)
 158 {
 159         struct bio *bi;
 160         while ((bi = bio_list_pop(return_bi)) != NULL) {
 161                 bi->bi_iter.bi_size = 0;
 162                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 163                                          bi, 0);
 164                 bio_endio(bi);
 165         }
 166 }
 167
 168 static void print_raid5_conf (struct r5conf *conf);
 169
 170 static int stripe_operations_active(struct stripe_head *sh)
 171 {
 172         return sh->check_state || sh->reconstruct_state ||
 173                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 174                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 175 }
 176
 177 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 178 {
 179         struct r5conf *conf = sh->raid_conf;
 180         struct r5worker_group *group;
 181         int thread_cnt;
 182         int i, cpu = sh->cpu;
 183
 184         if (!cpu_online(cpu)) {
 185                 cpu = cpumask_any(cpu_online_mask);
 186                 sh->cpu = cpu;
 187         }
 188
 189         if (list_empty(&sh->lru)) {
 190                 struct r5worker_group *group;
 191                 group = conf->worker_groups + cpu_to_group(cpu);
 192                 list_add_tail(&sh->lru, &group->handle_list);
 193                 group->stripes_cnt++;
 194                 sh->group = group;
 195         }
 196
 197         if (conf->worker_cnt_per_group == 0) {
 198                 md_wakeup_thread(conf->mddev->thread);
 199                 return;
 200         }
 201
 202         group = conf->worker_groups + cpu_to_group(sh->cpu);
 203
 204         group->workers[0].working = true;
 205         /* at least one worker should run to avoid race */
 206         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 207
 208         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 209         /* wakeup more workers */
 210         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 211                 if (group->workers[i].working == false) {
 212                         group->workers[i].working = true;
 213                         queue_work_on(sh->cpu, raid5_wq,
 214                                       &group->workers[i].work);
 215                         thread_cnt--;
 216                 }
 217         }
 218 }
 219
 220 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 221                               struct list_head *temp_inactive_list)
 222 {
 223         int i;
 224         int injournal = 0;      /* number of date pages with R5_InJournal */
 225
 226         BUG_ON(!list_empty(&sh->lru));
 227         BUG_ON(atomic_read(&conf->active_stripes)==0);
 228
 229         if (r5c_is_writeback(conf->log))
 230                 for (i = sh->disks; i--; )
 231                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 232                                 injournal++;
 233         /*
 234          * When quiesce in r5c write back, set STRIPE_HANDLE for stripes with
 235          * data in journal, so they are not released to cached lists
 236          */
 237         if (conf->quiesce && r5c_is_writeback(conf->log) &&
 238             !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0) {
 239                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 240                         r5c_make_stripe_write_out(sh);
 241                 set_bit(STRIPE_HANDLE, &sh->state);
 242         }
 243
 244         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 245                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 246                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 247                         list_add_tail(&sh->lru, &conf->delayed_list);
 248                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 249                            sh->bm_seq - conf->seq_write > 0)
 250                         list_add_tail(&sh->lru, &conf->bitmap_list);
 251                 else {
 252                         clear_bit(STRIPE_DELAYED, &sh->state);
 253                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 254                         if (conf->worker_cnt_per_group == 0) {
 255                                 list_add_tail(&sh->lru, &conf->handle_list);
 256                         } else {
 257                                 raid5_wakeup_stripe_thread(sh);
 258                                 return;
 259                         }
 260                 }
 261                 md_wakeup_thread(conf->mddev->thread);
 262         } else {
 263                 BUG_ON(stripe_operations_active(sh));
 264                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 265                         if (atomic_dec_return(&conf->preread_active_stripes)
 266                             < IO_THRESHOLD)
 267                                 md_wakeup_thread(conf->mddev->thread);
 268                 atomic_dec(&conf->active_stripes);
 269                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 270                         if (!r5c_is_writeback(conf->log))
 271                                 list_add_tail(&sh->lru, temp_inactive_list);
 272                         else {
 273                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 274                                 if (injournal == 0)
 275                                         list_add_tail(&sh->lru, temp_inactive_list);
 276                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 277                                         /* full stripe */
 278                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 279                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 280                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 281                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 282                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 283                                         r5c_check_cached_full_stripe(conf);
 284                                 } else {
 285                                         /* partial stripe */
 286                                         if (!test_and_set_bit(STRIPE_R5C_PARTIAL_STRIPE,
 287                                                               &sh->state))
 288                                                 atomic_inc(&conf->r5c_cached_partial_stripes);
 289                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 290                                 }
 291                         }
 292                 }
 293         }
 294 }
 295
 296 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 297                              struct list_head *temp_inactive_list)
 298 {
 299         if (atomic_dec_and_test(&sh->count))
 300                 do_release_stripe(conf, sh, temp_inactive_list);
 301 }
 302
 303 /*
 304  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 305  *
 306  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 307  * given time. Adding stripes only takes device lock, while deleting stripes
 308  * only takes hash lock.
 309  */
 310 static void release_inactive_stripe_list(struct r5conf *conf,
 311                                          struct list_head *temp_inactive_list,
 312                                          int hash)
 313 {
 314         int size;
 315         bool do_wakeup = false;
 316         unsigned long flags;
 317
 318         if (hash == NR_STRIPE_HASH_LOCKS) {
 319                 size = NR_STRIPE_HASH_LOCKS;
 320                 hash = NR_STRIPE_HASH_LOCKS - 1;
 321         } else
 322                 size = 1;
 323         while (size) {
 324                 struct list_head *list = &temp_inactive_list[size - 1];
 325
 326                 /*
 327                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 328                  * remove stripes from the list
 329                  */
 330                 if (!list_empty_careful(list)) {
 331                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 332                         if (list_empty(conf->inactive_list + hash) &&
 333                             !list_empty(list))
 334                                 atomic_dec(&conf->empty_inactive_list_nr);
 335                         list_splice_tail_init(list, conf->inactive_list + hash);
 336                         do_wakeup = true;
 337                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 338                 }
 339                 size--;
 340                 hash--;
 341         }
 342
 343         if (do_wakeup) {
 344                 wake_up(&conf->wait_for_stripe);
 345                 if (atomic_read(&conf->active_stripes) == 0)
 346                         wake_up(&conf->wait_for_quiescent);
 347                 if (conf->retry_read_aligned)
 348                         md_wakeup_thread(conf->mddev->thread);
 349         }
 350 }
 351
 352 /* should hold conf->device_lock already */
 353 static int release_stripe_list(struct r5conf *conf,
 354                                struct list_head *temp_inactive_list)
 355 {
 356         struct stripe_head *sh;
 357         int count = 0;
 358         struct llist_node *head;
 359
 360         head = llist_del_all(&conf->released_stripes);
 361         head = llist_reverse_order(head);
 362         while (head) {
 363                 int hash;
 364
 365                 sh = llist_entry(head, struct stripe_head, release_list);
 366                 head = llist_next(head);
 367                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 368                 smp_mb();
 369                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 370                 /*
 371                  * Don't worry the bit is set here, because if the bit is set
 372                  * again, the count is always > 1. This is true for
 373                  * STRIPE_ON_UNPLUG_LIST bit too.
 374                  */
 375                 hash = sh->hash_lock_index;
 376                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 377                 count++;
 378         }
 379
 380         return count;
 381 }
 382
 383 void raid5_release_stripe(struct stripe_head *sh)
 384 {
 385         struct r5conf *conf = sh->raid_conf;
 386         unsigned long flags;
 387         struct list_head list;
 388         int hash;
 389         bool wakeup;
 390
 391         /* Avoid release_list until the last reference.
 392          */
 393         if (atomic_add_unless(&sh->count, -1, 1))
 394                 return;
 395
 396         if (unlikely(!conf->mddev->thread) ||
 397                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 398                 goto slow_path;
 399         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 400         if (wakeup)
 401                 md_wakeup_thread(conf->mddev->thread);
 402         return;
 403 slow_path:
 404         local_irq_save(flags);
 405         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 406         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 407                 INIT_LIST_HEAD(&list);
 408                 hash = sh->hash_lock_index;
 409                 do_release_stripe(conf, sh, &list);
 410                 spin_unlock(&conf->device_lock);
 411                 release_inactive_stripe_list(conf, &list, hash);
 412         }
 413         local_irq_restore(flags);
 414 }
 415
 416 static inline void remove_hash(struct stripe_head *sh)
 417 {
 418         pr_debug("remove_hash(), stripe %llu\n",
 419                 (unsigned long long)sh->sector);
 420
 421         hlist_del_init(&sh->hash);
 422 }
 423
 424 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 425 {
 426         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 427
 428         pr_debug("insert_hash(), stripe %llu\n",
 429                 (unsigned long long)sh->sector);
 430
 431         hlist_add_head(&sh->hash, hp);
 432 }
 433
 434 /* find an idle stripe, make sure it is unhashed, and return it. */
 435 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 436 {
 437         struct stripe_head *sh = NULL;
 438         struct list_head *first;
 439
 440         if (list_empty(conf->inactive_list + hash))
 441                 goto out;
 442         first = (conf->inactive_list + hash)->next;
 443         sh = list_entry(first, struct stripe_head, lru);
 444         list_del_init(first);
 445         remove_hash(sh);
 446         atomic_inc(&conf->active_stripes);
 447         BUG_ON(hash != sh->hash_lock_index);
 448         if (list_empty(conf->inactive_list + hash))
 449                 atomic_inc(&conf->empty_inactive_list_nr);
 450 out:
 451         return sh;
 452 }
 453
 454 static void shrink_buffers(struct stripe_head *sh)
 455 {
 456         struct page *p;
 457         int i;
 458         int num = sh->raid_conf->pool_size;
 459
 460         for (i = 0; i < num ; i++) {
 461                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 462                 p = sh->dev[i].page;
 463                 if (!p)
 464                         continue;
 465                 sh->dev[i].page = NULL;
 466                 put_page(p);
 467         }
 468 }
 469
 470 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 471 {
 472         int i;
 473         int num = sh->raid_conf->pool_size;
 474
 475         for (i = 0; i < num; i++) {
 476                 struct page *page;
 477
 478                 if (!(page = alloc_page(gfp))) {
 479                         return 1;
 480                 }
 481                 sh->dev[i].page = page;
 482                 sh->dev[i].orig_page = page;
 483         }
 484         return 0;
 485 }
 486
 487 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 488 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 489                             struct stripe_head *sh);
 490
 491 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 492 {
 493         struct r5conf *conf = sh->raid_conf;
 494         int i, seq;
 495
 496         BUG_ON(atomic_read(&sh->count) != 0);
 497         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 498         BUG_ON(stripe_operations_active(sh));
 499         BUG_ON(sh->batch_head);
 500
 501         pr_debug("init_stripe called, stripe %llu\n",
 502                 (unsigned long long)sector);
 503 retry:
 504         seq = read_seqcount_begin(&conf->gen_lock);
 505         sh->generation = conf->generation - previous;
 506         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 507         sh->sector = sector;
 508         stripe_set_idx(sector, conf, previous, sh);
 509         sh->state = 0;
 510
 511         for (i = sh->disks; i--; ) {
 512                 struct r5dev *dev = &sh->dev[i];
 513
 514                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 515                     test_bit(R5_LOCKED, &dev->flags)) {
 516                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 517                                (unsigned long long)sh->sector, i, dev->toread,
 518                                dev->read, dev->towrite, dev->written,
 519                                test_bit(R5_LOCKED, &dev->flags));
 520                         WARN_ON(1);
 521                 }
 522                 dev->flags = 0;
 523                 raid5_build_block(sh, i, previous);
 524         }
 525         if (read_seqcount_retry(&conf->gen_lock, seq))
 526                 goto retry;
 527         sh->overwrite_disks = 0;
 528         insert_hash(conf, sh);
 529         sh->cpu = smp_processor_id();
 530         set_bit(STRIPE_BATCH_READY, &sh->state);
 531 }
 532
 533 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 534                                          short generation)
 535 {
 536         struct stripe_head *sh;
 537
 538         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 539         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 540                 if (sh->sector == sector && sh->generation == generation)
 541                         return sh;
 542         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 543         return NULL;
 544 }
 545
 546 /*
 547  * Need to check if array has failed when deciding whether to:
 548  *  - start an array
 549  *  - remove non-faulty devices
 550  *  - add a spare
 551  *  - allow a reshape
 552  * This determination is simple when no reshape is happening.
 553  * However if there is a reshape, we need to carefully check
 554  * both the before and after sections.
 555  * This is because some failed devices may only affect one
 556  * of the two sections, and some non-in_sync devices may
 557  * be insync in the section most affected by failed devices.
 558  */
 559 static int calc_degraded(struct r5conf *conf)
 560 {
 561         int degraded, degraded2;
 562         int i;
 563
 564         rcu_read_lock();
 565         degraded = 0;
 566         for (i = 0; i < conf->previous_raid_disks; i++) {
 567                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 568                 if (rdev && test_bit(Faulty, &rdev->flags))
 569                         rdev = rcu_dereference(conf->disks[i].replacement);
 570                 if (!rdev || test_bit(Faulty, &rdev->flags))
 571                         degraded++;
 572                 else if (test_bit(In_sync, &rdev->flags))
 573                         ;
 574                 else
 575                         /* not in-sync or faulty.
 576                          * If the reshape increases the number of devices,
 577                          * this is being recovered by the reshape, so
 578                          * this 'previous' section is not in_sync.
 579                          * If the number of devices is being reduced however,
 580                          * the device can only be part of the array if
 581                          * we are reverting a reshape, so this section will
 582                          * be in-sync.
 583                          */
 584                         if (conf->raid_disks >= conf->previous_raid_disks)
 585                                 degraded++;
 586         }
 587         rcu_read_unlock();
 588         if (conf->raid_disks == conf->previous_raid_disks)
 589                 return degraded;
 590         rcu_read_lock();
 591         degraded2 = 0;
 592         for (i = 0; i < conf->raid_disks; i++) {
 593                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 594                 if (rdev && test_bit(Faulty, &rdev->flags))
 595                         rdev = rcu_dereference(conf->disks[i].replacement);
 596                 if (!rdev || test_bit(Faulty, &rdev->flags))
 597                         degraded2++;
 598                 else if (test_bit(In_sync, &rdev->flags))
 599                         ;
 600                 else
 601                         /* not in-sync or faulty.
 602                          * If reshape increases the number of devices, this
 603                          * section has already been recovered, else it
 604                          * almost certainly hasn't.
 605                          */
 606                         if (conf->raid_disks <= conf->previous_raid_disks)
 607                                 degraded2++;
 608         }
 609         rcu_read_unlock();
 610         if (degraded2 > degraded)
 611                 return degraded2;
 612         return degraded;
 613 }
 614
 615 static int has_failed(struct r5conf *conf)
 616 {
 617         int degraded;
 618
 619         if (conf->mddev->reshape_position == MaxSector)
 620                 return conf->mddev->degraded > conf->max_degraded;
 621
 622         degraded = calc_degraded(conf);
 623         if (degraded > conf->max_degraded)
 624                 return 1;
 625         return 0;
 626 }
 627
 628 struct stripe_head *
 629 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 630                         int previous, int noblock, int noquiesce)
 631 {
 632         struct stripe_head *sh;
 633         int hash = stripe_hash_locks_hash(sector);
 634         int inc_empty_inactive_list_flag;
 635
 636         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 637
 638         spin_lock_irq(conf->hash_locks + hash);
 639
 640         do {
 641                 wait_event_lock_irq(conf->wait_for_quiescent,
 642                                     conf->quiesce == 0 || noquiesce,
 643                                     *(conf->hash_locks + hash));
 644                 sh = __find_stripe(conf, sector, conf->generation - previous);
 645                 if (!sh) {
 646                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 647                                 sh = get_free_stripe(conf, hash);
 648                                 if (!sh && !test_bit(R5_DID_ALLOC,
 649                                                      &conf->cache_state))
 650                                         set_bit(R5_ALLOC_MORE,
 651                                                 &conf->cache_state);
 652                         }
 653                         if (noblock && sh == NULL)
 654                                 break;
 655
 656                         r5c_check_stripe_cache_usage(conf);
 657                         if (!sh) {
 658                                 set_bit(R5_INACTIVE_BLOCKED,
 659                                         &conf->cache_state);
 660                                 r5l_wake_reclaim(conf->log, 0);
 661                                 wait_event_lock_irq(
 662                                         conf->wait_for_stripe,
 663                                         !list_empty(conf->inactive_list + hash) &&
 664                                         (atomic_read(&conf->active_stripes)
 665                                          < (conf->max_nr_stripes * 3 / 4)
 666                                          || !test_bit(R5_INACTIVE_BLOCKED,
 667                                                       &conf->cache_state)),
 668                                         *(conf->hash_locks + hash));
 669                                 clear_bit(R5_INACTIVE_BLOCKED,
 670                                           &conf->cache_state);
 671                         } else {
 672                                 init_stripe(sh, sector, previous);
 673                                 atomic_inc(&sh->count);
 674                         }
 675                 } else if (!atomic_inc_not_zero(&sh->count)) {
 676                         spin_lock(&conf->device_lock);
 677                         if (!atomic_read(&sh->count)) {
 678                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 679                                         atomic_inc(&conf->active_stripes);
 680                                 BUG_ON(list_empty(&sh->lru) &&
 681                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 682                                 inc_empty_inactive_list_flag = 0;
 683                                 if (!list_empty(conf->inactive_list + hash))
 684                                         inc_empty_inactive_list_flag = 1;
 685                                 list_del_init(&sh->lru);
 686                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 687                                         atomic_inc(&conf->empty_inactive_list_nr);
 688                                 if (sh->group) {
 689                                         sh->group->stripes_cnt--;
 690                                         sh->group = NULL;
 691                                 }
 692                         }
 693                         atomic_inc(&sh->count);
 694                         spin_unlock(&conf->device_lock);
 695                 }
 696         } while (sh == NULL);
 697
 698         spin_unlock_irq(conf->hash_locks + hash);
 699         return sh;
 700 }
 701
 702 static bool is_full_stripe_write(struct stripe_head *sh)
 703 {
 704         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 705         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 706 }
 707
 708 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 709 {
 710         local_irq_disable();
 711         if (sh1 > sh2) {
 712                 spin_lock(&sh2->stripe_lock);
 713                 spin_lock_nested(&sh1->stripe_lock, 1);
 714         } else {
 715                 spin_lock(&sh1->stripe_lock);
 716                 spin_lock_nested(&sh2->stripe_lock, 1);
 717         }
 718 }
 719
 720 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 721 {
 722         spin_unlock(&sh1->stripe_lock);
 723         spin_unlock(&sh2->stripe_lock);
 724         local_irq_enable();
 725 }
 726
 727 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 728 static bool stripe_can_batch(struct stripe_head *sh)
 729 {
 730         struct r5conf *conf = sh->raid_conf;
 731
 732         if (conf->log)
 733                 return false;
 734         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 735                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 736                 is_full_stripe_write(sh);
 737 }
 738
 739 /* we only do back search */
 740 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 741 {
 742         struct stripe_head *head;
 743         sector_t head_sector, tmp_sec;
 744         int hash;
 745         int dd_idx;
 746         int inc_empty_inactive_list_flag;
 747
 748         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 749         tmp_sec = sh->sector;
 750         if (!sector_div(tmp_sec, conf->chunk_sectors))
 751                 return;
 752         head_sector = sh->sector - STRIPE_SECTORS;
 753
 754         hash = stripe_hash_locks_hash(head_sector);
 755         spin_lock_irq(conf->hash_locks + hash);
 756         head = __find_stripe(conf, head_sector, conf->generation);
 757         if (head && !atomic_inc_not_zero(&head->count)) {
 758                 spin_lock(&conf->device_lock);
 759                 if (!atomic_read(&head->count)) {
 760                         if (!test_bit(STRIPE_HANDLE, &head->state))
 761                                 atomic_inc(&conf->active_stripes);
 762                         BUG_ON(list_empty(&head->lru) &&
 763                                !test_bit(STRIPE_EXPANDING, &head->state));
 764                         inc_empty_inactive_list_flag = 0;
 765                         if (!list_empty(conf->inactive_list + hash))
 766                                 inc_empty_inactive_list_flag = 1;
 767                         list_del_init(&head->lru);
 768                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 769                                 atomic_inc(&conf->empty_inactive_list_nr);
 770                         if (head->group) {
 771                                 head->group->stripes_cnt--;
 772                                 head->group = NULL;
 773                         }
 774                 }
 775                 atomic_inc(&head->count);
 776                 spin_unlock(&conf->device_lock);
 777         }
 778         spin_unlock_irq(conf->hash_locks + hash);
 779
 780         if (!head)
 781                 return;
 782         if (!stripe_can_batch(head))
 783                 goto out;
 784
 785         lock_two_stripes(head, sh);
 786         /* clear_batch_ready clear the flag */
 787         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 788                 goto unlock_out;
 789
 790         if (sh->batch_head)
 791                 goto unlock_out;
 792
 793         dd_idx = 0;
 794         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 795                 dd_idx++;
 796         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 797             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 798                 goto unlock_out;
 799
 800         if (head->batch_head) {
 801                 spin_lock(&head->batch_head->batch_lock);
 802                 /* This batch list is already running */
 803                 if (!stripe_can_batch(head)) {
 804                         spin_unlock(&head->batch_head->batch_lock);
 805                         goto unlock_out;
 806                 }
 807
 808                 /*
 809                  * at this point, head's BATCH_READY could be cleared, but we
 810                  * can still add the stripe to batch list
 811                  */
 812                 list_add(&sh->batch_list, &head->batch_list);
 813                 spin_unlock(&head->batch_head->batch_lock);
 814
 815                 sh->batch_head = head->batch_head;
 816         } else {
 817                 head->batch_head = head;
 818                 sh->batch_head = head->batch_head;
 819                 spin_lock(&head->batch_lock);
 820                 list_add_tail(&sh->batch_list, &head->batch_list);
 821                 spin_unlock(&head->batch_lock);
 822         }
 823
 824         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 825                 if (atomic_dec_return(&conf->preread_active_stripes)
 826                     < IO_THRESHOLD)
 827                         md_wakeup_thread(conf->mddev->thread);
 828
 829         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 830                 int seq = sh->bm_seq;
 831                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 832                     sh->batch_head->bm_seq > seq)
 833                         seq = sh->batch_head->bm_seq;
 834                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 835                 sh->batch_head->bm_seq = seq;
 836         }
 837
 838         atomic_inc(&sh->count);
 839 unlock_out:
 840         unlock_two_stripes(head, sh);
 841 out:
 842         raid5_release_stripe(head);
 843 }
 844
 845 /* Determine if 'data_offset' or 'new_data_offset' should be used
 846  * in this stripe_head.
 847  */
 848 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 849 {
 850         sector_t progress = conf->reshape_progress;
 851         /* Need a memory barrier to make sure we see the value
 852          * of conf->generation, or ->data_offset that was set before
 853          * reshape_progress was updated.
 854          */
 855         smp_rmb();
 856         if (progress == MaxSector)
 857                 return 0;
 858         if (sh->generation == conf->generation - 1)
 859                 return 0;
 860         /* We are in a reshape, and this is a new-generation stripe,
 861          * so use new_data_offset.
 862          */
 863         return 1;
 864 }
 865
 866 static void
 867 raid5_end_read_request(struct bio *bi);
 868 static void
 869 raid5_end_write_request(struct bio *bi);
 870
 871 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 872 {
 873         struct r5conf *conf = sh->raid_conf;
 874         int i, disks = sh->disks;
 875         struct stripe_head *head_sh = sh;
 876
 877         might_sleep();
 878
 879         if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
 880                 /* writing out phase */
 881                 if (s->waiting_extra_page)
 882                         return;
 883                 if (r5l_write_stripe(conf->log, sh) == 0)
 884                         return;
 885         } else {  /* caching phase */
 886                 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state)) {
 887                         r5c_cache_data(conf->log, sh, s);
 888                         return;
 889                 }
 890         }
 891
 892         for (i = disks; i--; ) {
 893                 int op, op_flags = 0;
 894                 int replace_only = 0;
 895                 struct bio *bi, *rbi;
 896                 struct md_rdev *rdev, *rrdev = NULL;
 897
 898                 sh = head_sh;
 899                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 900                         op = REQ_OP_WRITE;
 901                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 902                                 op_flags = REQ_FUA;
 903                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 904                                 op = REQ_OP_DISCARD;
 905                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 906                         op = REQ_OP_READ;
 907                 else if (test_and_clear_bit(R5_WantReplace,
 908                                             &sh->dev[i].flags)) {
 909                         op = REQ_OP_WRITE;
 910                         replace_only = 1;
 911                 } else
 912                         continue;
 913                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 914                         op_flags |= REQ_SYNC;
 915
 916 again:
 917                 bi = &sh->dev[i].req;
 918                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 919
 920                 rcu_read_lock();
 921                 rrdev = rcu_dereference(conf->disks[i].replacement);
 922                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 923                 rdev = rcu_dereference(conf->disks[i].rdev);
 924                 if (!rdev) {
 925                         rdev = rrdev;
 926                         rrdev = NULL;
 927                 }
 928                 if (op_is_write(op)) {
 929                         if (replace_only)
 930                                 rdev = NULL;
 931                         if (rdev == rrdev)
 932                                 /* We raced and saw duplicates */
 933                                 rrdev = NULL;
 934                 } else {
 935                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
 936                                 rdev = rrdev;
 937                         rrdev = NULL;
 938                 }
 939
 940                 if (rdev && test_bit(Faulty, &rdev->flags))
 941                         rdev = NULL;
 942                 if (rdev)
 943                         atomic_inc(&rdev->nr_pending);
 944                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 945                         rrdev = NULL;
 946                 if (rrdev)
 947                         atomic_inc(&rrdev->nr_pending);
 948                 rcu_read_unlock();
 949
 950                 /* We have already checked bad blocks for reads.  Now
 951                  * need to check for writes.  We never accept write errors
 952                  * on the replacement, so we don't to check rrdev.
 953                  */
 954                 while (op_is_write(op) && rdev &&
 955                        test_bit(WriteErrorSeen, &rdev->flags)) {
 956                         sector_t first_bad;
 957                         int bad_sectors;
 958                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 959                                               &first_bad, &bad_sectors);
 960                         if (!bad)
 961                                 break;
 962
 963                         if (bad < 0) {
 964                                 set_bit(BlockedBadBlocks, &rdev->flags);
 965                                 if (!conf->mddev->external &&
 966                                     conf->mddev->sb_flags) {
 967                                         /* It is very unlikely, but we might
 968                                          * still need to write out the
 969                                          * bad block log - better give it
 970                                          * a chance*/
 971                                         md_check_recovery(conf->mddev);
 972                                 }
 973                                 /*
 974                                  * Because md_wait_for_blocked_rdev
 975                                  * will dec nr_pending, we must
 976                                  * increment it first.
 977                                  */
 978                                 atomic_inc(&rdev->nr_pending);
 979                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 980                         } else {
 981                                 /* Acknowledged bad block - skip the write */
 982                                 rdev_dec_pending(rdev, conf->mddev);
 983                                 rdev = NULL;
 984                         }
 985                 }
 986
 987                 if (rdev) {
 988                         if (s->syncing || s->expanding || s->expanded
 989                             || s->replacing)
 990                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 991
 992                         set_bit(STRIPE_IO_STARTED, &sh->state);
 993
 994                         bi->bi_bdev = rdev->bdev;
 995                         bio_set_op_attrs(bi, op, op_flags);
 996                         bi->bi_end_io = op_is_write(op)
 997                                 ? raid5_end_write_request
 998                                 : raid5_end_read_request;
 999                         bi->bi_private = sh;
1000
1001                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1002                                 __func__, (unsigned long long)sh->sector,
1003                                 bi->bi_opf, i);
1004                         atomic_inc(&sh->count);
1005                         if (sh != head_sh)
1006                                 atomic_inc(&head_sh->count);
1007                         if (use_new_offset(conf, sh))
1008                                 bi->bi_iter.bi_sector = (sh->sector
1009                                                  + rdev->new_data_offset);
1010                         else
1011                                 bi->bi_iter.bi_sector = (sh->sector
1012                                                  + rdev->data_offset);
1013                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1014                                 bi->bi_opf |= REQ_NOMERGE;
1015
1016                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1017                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1018                         sh->dev[i].vec.bv_page = sh->dev[i].page;
1019                         bi->bi_vcnt = 1;
1020                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1021                         bi->bi_io_vec[0].bv_offset = 0;
1022                         bi->bi_iter.bi_size = STRIPE_SIZE;
1023                         /*
1024                          * If this is discard request, set bi_vcnt 0. We don't
1025                          * want to confuse SCSI because SCSI will replace payload
1026                          */
1027                         if (op == REQ_OP_DISCARD)
1028                                 bi->bi_vcnt = 0;
1029                         if (rrdev)
1030                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1031
1032                         if (conf->mddev->gendisk)
1033                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1034                                                       bi, disk_devt(conf->mddev->gendisk),
1035                                                       sh->dev[i].sector);
1036                         generic_make_request(bi);
1037                 }
1038                 if (rrdev) {
1039                         if (s->syncing || s->expanding || s->expanded
1040                             || s->replacing)
1041                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1042
1043                         set_bit(STRIPE_IO_STARTED, &sh->state);
1044
1045                         rbi->bi_bdev = rrdev->bdev;
1046                         bio_set_op_attrs(rbi, op, op_flags);
1047                         BUG_ON(!op_is_write(op));
1048                         rbi->bi_end_io = raid5_end_write_request;
1049                         rbi->bi_private = sh;
1050
1051                         pr_debug("%s: for %llu schedule op %d on "
1052                                  "replacement disc %d\n",
1053                                 __func__, (unsigned long long)sh->sector,
1054                                 rbi->bi_opf, i);
1055                         atomic_inc(&sh->count);
1056                         if (sh != head_sh)
1057                                 atomic_inc(&head_sh->count);
1058                         if (use_new_offset(conf, sh))
1059                                 rbi->bi_iter.bi_sector = (sh->sector
1060                                                   + rrdev->new_data_offset);
1061                         else
1062                                 rbi->bi_iter.bi_sector = (sh->sector
1063                                                   + rrdev->data_offset);
1064                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1065                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1066                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1067                         rbi->bi_vcnt = 1;
1068                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1069                         rbi->bi_io_vec[0].bv_offset = 0;
1070                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1071                         /*
1072                          * If this is discard request, set bi_vcnt 0. We don't
1073                          * want to confuse SCSI because SCSI will replace payload
1074                          */
1075                         if (op == REQ_OP_DISCARD)
1076                                 rbi->bi_vcnt = 0;
1077                         if (conf->mddev->gendisk)
1078                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1079                                                       rbi, disk_devt(conf->mddev->gendisk),
1080                                                       sh->dev[i].sector);
1081                         generic_make_request(rbi);
1082                 }
1083                 if (!rdev && !rrdev) {
1084                         if (op_is_write(op))
1085                                 set_bit(STRIPE_DEGRADED, &sh->state);
1086                         pr_debug("skip op %d on disc %d for sector %llu\n",
1087                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1088                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1089                         set_bit(STRIPE_HANDLE, &sh->state);
1090                 }
1091
1092                 if (!head_sh->batch_head)
1093                         continue;
1094                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1095                                       batch_list);
1096                 if (sh != head_sh)
1097                         goto again;
1098         }
1099 }
1100
1101 static struct dma_async_tx_descriptor *
1102 async_copy_data(int frombio, struct bio *bio, struct page **page,
1103         sector_t sector, struct dma_async_tx_descriptor *tx,
1104         struct stripe_head *sh, int no_skipcopy)
1105 {
1106         struct bio_vec bvl;
1107         struct bvec_iter iter;
1108         struct page *bio_page;
1109         int page_offset;
1110         struct async_submit_ctl submit;
1111         enum async_tx_flags flags = 0;
1112
1113         if (bio->bi_iter.bi_sector >= sector)
1114                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1115         else
1116                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1117
1118         if (frombio)
1119                 flags |= ASYNC_TX_FENCE;
1120         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1121
1122         bio_for_each_segment(bvl, bio, iter) {
1123                 int len = bvl.bv_len;
1124                 int clen;
1125                 int b_offset = 0;
1126
1127                 if (page_offset < 0) {
1128                         b_offset = -page_offset;
1129                         page_offset += b_offset;
1130                         len -= b_offset;
1131                 }
1132
1133                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1134                         clen = STRIPE_SIZE - page_offset;
1135                 else
1136                         clen = len;
1137
1138                 if (clen > 0) {
1139                         b_offset += bvl.bv_offset;
1140                         bio_page = bvl.bv_page;
1141                         if (frombio) {
1142                                 if (sh->raid_conf->skip_copy &&
1143                                     b_offset == 0 && page_offset == 0 &&
1144                                     clen == STRIPE_SIZE &&
1145                                     !no_skipcopy)
1146                                         *page = bio_page;
1147                                 else
1148                                         tx = async_memcpy(*page, bio_page, page_offset,
1149                                                   b_offset, clen, &submit);
1150                         } else
1151                                 tx = async_memcpy(bio_page, *page, b_offset,
1152                                                   page_offset, clen, &submit);
1153                 }
1154                 /* chain the operations */
1155                 submit.depend_tx = tx;
1156
1157                 if (clen < len) /* hit end of page */
1158                         break;
1159                 page_offset +=  len;
1160         }
1161
1162         return tx;
1163 }
1164
1165 static void ops_complete_biofill(void *stripe_head_ref)
1166 {
1167         struct stripe_head *sh = stripe_head_ref;
1168         struct bio_list return_bi = BIO_EMPTY_LIST;
1169         int i;
1170
1171         pr_debug("%s: stripe %llu\n", __func__,
1172                 (unsigned long long)sh->sector);
1173
1174         /* clear completed biofills */
1175         for (i = sh->disks; i--; ) {
1176                 struct r5dev *dev = &sh->dev[i];
1177
1178                 /* acknowledge completion of a biofill operation */
1179                 /* and check if we need to reply to a read request,
1180                  * new R5_Wantfill requests are held off until
1181                  * !STRIPE_BIOFILL_RUN
1182                  */
1183                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1184                         struct bio *rbi, *rbi2;
1185
1186                         BUG_ON(!dev->read);
1187                         rbi = dev->read;
1188                         dev->read = NULL;
1189                         while (rbi && rbi->bi_iter.bi_sector <
1190                                 dev->sector + STRIPE_SECTORS) {
1191                                 rbi2 = r5_next_bio(rbi, dev->sector);
1192                                 if (!raid5_dec_bi_active_stripes(rbi))
1193                                         bio_list_add(&return_bi, rbi);
1194                                 rbi = rbi2;
1195                         }
1196                 }
1197         }
1198         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1199
1200         return_io(&return_bi);
1201
1202         set_bit(STRIPE_HANDLE, &sh->state);
1203         raid5_release_stripe(sh);
1204 }
1205
1206 static void ops_run_biofill(struct stripe_head *sh)
1207 {
1208         struct dma_async_tx_descriptor *tx = NULL;
1209         struct async_submit_ctl submit;
1210         int i;
1211
1212         BUG_ON(sh->batch_head);
1213         pr_debug("%s: stripe %llu\n", __func__,
1214                 (unsigned long long)sh->sector);
1215
1216         for (i = sh->disks; i--; ) {
1217                 struct r5dev *dev = &sh->dev[i];
1218                 if (test_bit(R5_Wantfill, &dev->flags)) {
1219                         struct bio *rbi;
1220                         spin_lock_irq(&sh->stripe_lock);
1221                         dev->read = rbi = dev->toread;
1222                         dev->toread = NULL;
1223                         spin_unlock_irq(&sh->stripe_lock);
1224                         while (rbi && rbi->bi_iter.bi_sector <
1225                                 dev->sector + STRIPE_SECTORS) {
1226                                 tx = async_copy_data(0, rbi, &dev->page,
1227                                                      dev->sector, tx, sh, 0);
1228                                 rbi = r5_next_bio(rbi, dev->sector);
1229                         }
1230                 }
1231         }
1232
1233         atomic_inc(&sh->count);
1234         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1235         async_trigger_callback(&submit);
1236 }
1237
1238 static void mark_target_uptodate(struct stripe_head *sh, int target)
1239 {
1240         struct r5dev *tgt;
1241
1242         if (target < 0)
1243                 return;
1244
1245         tgt = &sh->dev[target];
1246         set_bit(R5_UPTODATE, &tgt->flags);
1247         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1248         clear_bit(R5_Wantcompute, &tgt->flags);
1249 }
1250
1251 static void ops_complete_compute(void *stripe_head_ref)
1252 {
1253         struct stripe_head *sh = stripe_head_ref;
1254
1255         pr_debug("%s: stripe %llu\n", __func__,
1256                 (unsigned long long)sh->sector);
1257
1258         /* mark the computed target(s) as uptodate */
1259         mark_target_uptodate(sh, sh->ops.target);
1260         mark_target_uptodate(sh, sh->ops.target2);
1261
1262         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1263         if (sh->check_state == check_state_compute_run)
1264                 sh->check_state = check_state_compute_result;
1265         set_bit(STRIPE_HANDLE, &sh->state);
1266         raid5_release_stripe(sh);
1267 }
1268
1269 /* return a pointer to the address conversion region of the scribble buffer */
1270 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1271                                  struct raid5_percpu *percpu, int i)
1272 {
1273         void *addr;
1274
1275         addr = flex_array_get(percpu->scribble, i);
1276         return addr + sizeof(struct page *) * (sh->disks + 2);
1277 }
1278
1279 /* return a pointer to the address conversion region of the scribble buffer */
1280 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1281 {
1282         void *addr;
1283
1284         addr = flex_array_get(percpu->scribble, i);
1285         return addr;
1286 }
1287
1288 static struct dma_async_tx_descriptor *
1289 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1290 {
1291         int disks = sh->disks;
1292         struct page **xor_srcs = to_addr_page(percpu, 0);
1293         int target = sh->ops.target;
1294         struct r5dev *tgt = &sh->dev[target];
1295         struct page *xor_dest = tgt->page;
1296         int count = 0;
1297         struct dma_async_tx_descriptor *tx;
1298         struct async_submit_ctl submit;
1299         int i;
1300
1301         BUG_ON(sh->batch_head);
1302
1303         pr_debug("%s: stripe %llu block: %d\n",
1304                 __func__, (unsigned long long)sh->sector, target);
1305         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1306
1307         for (i = disks; i--; )
1308                 if (i != target)
1309                         xor_srcs[count++] = sh->dev[i].page;
1310
1311         atomic_inc(&sh->count);
1312
1313         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1314                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1315         if (unlikely(count == 1))
1316                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1317         else
1318                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1319
1320         return tx;
1321 }
1322
1323 /* set_syndrome_sources - populate source buffers for gen_syndrome
1324  * @srcs - (struct page *) array of size sh->disks
1325  * @sh - stripe_head to parse
1326  *
1327  * Populates srcs in proper layout order for the stripe and returns the
1328  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1329  * destination buffer is recorded in srcs[count] and the Q destination
1330  * is recorded in srcs[count+1]].
1331  */
1332 static int set_syndrome_sources(struct page **srcs,
1333                                 struct stripe_head *sh,
1334                                 int srctype)
1335 {
1336         int disks = sh->disks;
1337         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1338         int d0_idx = raid6_d0(sh);
1339         int count;
1340         int i;
1341
1342         for (i = 0; i < disks; i++)
1343                 srcs[i] = NULL;
1344
1345         count = 0;
1346         i = d0_idx;
1347         do {
1348                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1349                 struct r5dev *dev = &sh->dev[i];
1350
1351                 if (i == sh->qd_idx || i == sh->pd_idx ||
1352                     (srctype == SYNDROME_SRC_ALL) ||
1353                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1354                      (test_bit(R5_Wantdrain, &dev->flags) ||
1355                       test_bit(R5_InJournal, &dev->flags))) ||
1356                     (srctype == SYNDROME_SRC_WRITTEN &&
1357                      dev->written)) {
1358                         if (test_bit(R5_InJournal, &dev->flags))
1359                                 srcs[slot] = sh->dev[i].orig_page;
1360                         else
1361                                 srcs[slot] = sh->dev[i].page;
1362                 }
1363                 i = raid6_next_disk(i, disks);
1364         } while (i != d0_idx);
1365
1366         return syndrome_disks;
1367 }
1368
1369 static struct dma_async_tx_descriptor *
1370 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1371 {
1372         int disks = sh->disks;
1373         struct page **blocks = to_addr_page(percpu, 0);
1374         int target;
1375         int qd_idx = sh->qd_idx;
1376         struct dma_async_tx_descriptor *tx;
1377         struct async_submit_ctl submit;
1378         struct r5dev *tgt;
1379         struct page *dest;
1380         int i;
1381         int count;
1382
1383         BUG_ON(sh->batch_head);
1384         if (sh->ops.target < 0)
1385                 target = sh->ops.target2;
1386         else if (sh->ops.target2 < 0)
1387                 target = sh->ops.target;
1388         else
1389                 /* we should only have one valid target */
1390                 BUG();
1391         BUG_ON(target < 0);
1392         pr_debug("%s: stripe %llu block: %d\n",
1393                 __func__, (unsigned long long)sh->sector, target);
1394
1395         tgt = &sh->dev[target];
1396         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1397         dest = tgt->page;
1398
1399         atomic_inc(&sh->count);
1400
1401         if (target == qd_idx) {
1402                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1403                 blocks[count] = NULL; /* regenerating p is not necessary */
1404                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1405                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1406                                   ops_complete_compute, sh,
1407                                   to_addr_conv(sh, percpu, 0));
1408                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1409         } else {
1410                 /* Compute any data- or p-drive using XOR */
1411                 count = 0;
1412                 for (i = disks; i-- ; ) {
1413                         if (i == target || i == qd_idx)
1414                                 continue;
1415                         blocks[count++] = sh->dev[i].page;
1416                 }
1417
1418                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1419                                   NULL, ops_complete_compute, sh,
1420                                   to_addr_conv(sh, percpu, 0));
1421                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1422         }
1423
1424         return tx;
1425 }
1426
1427 static struct dma_async_tx_descriptor *
1428 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1429 {
1430         int i, count, disks = sh->disks;
1431         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1432         int d0_idx = raid6_d0(sh);
1433         int faila = -1, failb = -1;
1434         int target = sh->ops.target;
1435         int target2 = sh->ops.target2;
1436         struct r5dev *tgt = &sh->dev[target];
1437         struct r5dev *tgt2 = &sh->dev[target2];
1438         struct dma_async_tx_descriptor *tx;
1439         struct page **blocks = to_addr_page(percpu, 0);
1440         struct async_submit_ctl submit;
1441
1442         BUG_ON(sh->batch_head);
1443         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1444                  __func__, (unsigned long long)sh->sector, target, target2);
1445         BUG_ON(target < 0 || target2 < 0);
1446         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1447         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1448
1449         /* we need to open-code set_syndrome_sources to handle the
1450          * slot number conversion for 'faila' and 'failb'
1451          */
1452         for (i = 0; i < disks ; i++)
1453                 blocks[i] = NULL;
1454         count = 0;
1455         i = d0_idx;
1456         do {
1457                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1458
1459                 blocks[slot] = sh->dev[i].page;
1460
1461                 if (i == target)
1462                         faila = slot;
1463                 if (i == target2)
1464                         failb = slot;
1465                 i = raid6_next_disk(i, disks);
1466         } while (i != d0_idx);
1467
1468         BUG_ON(faila == failb);
1469         if (failb < faila)
1470                 swap(faila, failb);
1471         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1472                  __func__, (unsigned long long)sh->sector, faila, failb);
1473
1474         atomic_inc(&sh->count);
1475
1476         if (failb == syndrome_disks+1) {
1477                 /* Q disk is one of the missing disks */
1478                 if (faila == syndrome_disks) {
1479                         /* Missing P+Q, just recompute */
1480                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1481                                           ops_complete_compute, sh,
1482                                           to_addr_conv(sh, percpu, 0));
1483                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1484                                                   STRIPE_SIZE, &submit);
1485                 } else {
1486                         struct page *dest;
1487                         int data_target;
1488                         int qd_idx = sh->qd_idx;
1489
1490                         /* Missing D+Q: recompute D from P, then recompute Q */
1491                         if (target == qd_idx)
1492                                 data_target = target2;
1493                         else
1494                                 data_target = target;
1495
1496                         count = 0;
1497                         for (i = disks; i-- ; ) {
1498                                 if (i == data_target || i == qd_idx)
1499                                         continue;
1500                                 blocks[count++] = sh->dev[i].page;
1501                         }
1502                         dest = sh->dev[data_target].page;
1503                         init_async_submit(&submit,
1504                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1505                                           NULL, NULL, NULL,
1506                                           to_addr_conv(sh, percpu, 0));
1507                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1508                                        &submit);
1509
1510                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1511                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1512                                           ops_complete_compute, sh,
1513                                           to_addr_conv(sh, percpu, 0));
1514                         return async_gen_syndrome(blocks, 0, count+2,
1515                                                   STRIPE_SIZE, &submit);
1516                 }
1517         } else {
1518                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1519                                   ops_complete_compute, sh,
1520                                   to_addr_conv(sh, percpu, 0));
1521                 if (failb == syndrome_disks) {
1522                         /* We're missing D+P. */
1523                         return async_raid6_datap_recov(syndrome_disks+2,
1524                                                        STRIPE_SIZE, faila,
1525                                                        blocks, &submit);
1526                 } else {
1527                         /* We're missing D+D. */
1528                         return async_raid6_2data_recov(syndrome_disks+2,
1529                                                        STRIPE_SIZE, faila, failb,
1530                                                        blocks, &submit);
1531                 }
1532         }
1533 }
1534
1535 static void ops_complete_prexor(void *stripe_head_ref)
1536 {
1537         struct stripe_head *sh = stripe_head_ref;
1538
1539         pr_debug("%s: stripe %llu\n", __func__,
1540                 (unsigned long long)sh->sector);
1541
1542         if (r5c_is_writeback(sh->raid_conf->log))
1543                 /*
1544                  * raid5-cache write back uses orig_page during prexor.
1545                  * After prexor, it is time to free orig_page
1546                  */
1547                 r5c_release_extra_page(sh);
1548 }
1549
1550 static struct dma_async_tx_descriptor *
1551 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1552                 struct dma_async_tx_descriptor *tx)
1553 {
1554         int disks = sh->disks;
1555         struct page **xor_srcs = to_addr_page(percpu, 0);
1556         int count = 0, pd_idx = sh->pd_idx, i;
1557         struct async_submit_ctl submit;
1558
1559         /* existing parity data subtracted */
1560         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1561
1562         BUG_ON(sh->batch_head);
1563         pr_debug("%s: stripe %llu\n", __func__,
1564                 (unsigned long long)sh->sector);
1565
1566         for (i = disks; i--; ) {
1567                 struct r5dev *dev = &sh->dev[i];
1568                 /* Only process blocks that are known to be uptodate */
1569                 if (test_bit(R5_InJournal, &dev->flags))
1570                         xor_srcs[count++] = dev->orig_page;
1571                 else if (test_bit(R5_Wantdrain, &dev->flags))
1572                         xor_srcs[count++] = dev->page;
1573         }
1574
1575         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1576                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1577         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1578
1579         return tx;
1580 }
1581
1582 static struct dma_async_tx_descriptor *
1583 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1584                 struct dma_async_tx_descriptor *tx)
1585 {
1586         struct page **blocks = to_addr_page(percpu, 0);
1587         int count;
1588         struct async_submit_ctl submit;
1589
1590         pr_debug("%s: stripe %llu\n", __func__,
1591                 (unsigned long long)sh->sector);
1592
1593         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1594
1595         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1596                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1597         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1598
1599         return tx;
1600 }
1601
1602 static struct dma_async_tx_descriptor *
1603 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1604 {
1605         struct r5conf *conf = sh->raid_conf;
1606         int disks = sh->disks;
1607         int i;
1608         struct stripe_head *head_sh = sh;
1609
1610         pr_debug("%s: stripe %llu\n", __func__,
1611                 (unsigned long long)sh->sector);
1612
1613         for (i = disks; i--; ) {
1614                 struct r5dev *dev;
1615                 struct bio *chosen;
1616
1617                 sh = head_sh;
1618                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1619                         struct bio *wbi;
1620
1621 again:
1622                         dev = &sh->dev[i];
1623                         /*
1624                          * clear R5_InJournal, so when rewriting a page in
1625                          * journal, it is not skipped by r5l_log_stripe()
1626                          */
1627                         clear_bit(R5_InJournal, &dev->flags);
1628                         spin_lock_irq(&sh->stripe_lock);
1629                         chosen = dev->towrite;
1630                         dev->towrite = NULL;
1631                         sh->overwrite_disks = 0;
1632                         BUG_ON(dev->written);
1633                         wbi = dev->written = chosen;
1634                         spin_unlock_irq(&sh->stripe_lock);
1635                         WARN_ON(dev->page != dev->orig_page);
1636
1637                         while (wbi && wbi->bi_iter.bi_sector <
1638                                 dev->sector + STRIPE_SECTORS) {
1639                                 if (wbi->bi_opf & REQ_FUA)
1640                                         set_bit(R5_WantFUA, &dev->flags);
1641                                 if (wbi->bi_opf & REQ_SYNC)
1642                                         set_bit(R5_SyncIO, &dev->flags);
1643                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1644                                         set_bit(R5_Discard, &dev->flags);
1645                                 else {
1646                                         tx = async_copy_data(1, wbi, &dev->page,
1647                                                              dev->sector, tx, sh,
1648                                                              r5c_is_writeback(conf->log));
1649                                         if (dev->page != dev->orig_page &&
1650                                             !r5c_is_writeback(conf->log)) {
1651                                                 set_bit(R5_SkipCopy, &dev->flags);
1652                                                 clear_bit(R5_UPTODATE, &dev->flags);
1653                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1654                                         }
1655                                 }
1656                                 wbi = r5_next_bio(wbi, dev->sector);
1657                         }
1658
1659                         if (head_sh->batch_head) {
1660                                 sh = list_first_entry(&sh->batch_list,
1661                                                       struct stripe_head,
1662                                                       batch_list);
1663                                 if (sh == head_sh)
1664                                         continue;
1665                                 goto again;
1666                         }
1667                 }
1668         }
1669
1670         return tx;
1671 }
1672
1673 static void ops_complete_reconstruct(void *stripe_head_ref)
1674 {
1675         struct stripe_head *sh = stripe_head_ref;
1676         int disks = sh->disks;
1677         int pd_idx = sh->pd_idx;
1678         int qd_idx = sh->qd_idx;
1679         int i;
1680         bool fua = false, sync = false, discard = false;
1681
1682         pr_debug("%s: stripe %llu\n", __func__,
1683                 (unsigned long long)sh->sector);
1684
1685         for (i = disks; i--; ) {
1686                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1687                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1688                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1689         }
1690
1691         for (i = disks; i--; ) {
1692                 struct r5dev *dev = &sh->dev[i];
1693
1694                 if (dev->written || i == pd_idx || i == qd_idx) {
1695                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1696                                 set_bit(R5_UPTODATE, &dev->flags);
1697                         if (fua)
1698                                 set_bit(R5_WantFUA, &dev->flags);
1699                         if (sync)
1700                                 set_bit(R5_SyncIO, &dev->flags);
1701                 }
1702         }
1703
1704         if (sh->reconstruct_state == reconstruct_state_drain_run)
1705                 sh->reconstruct_state = reconstruct_state_drain_result;
1706         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1707                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1708         else {
1709                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1710                 sh->reconstruct_state = reconstruct_state_result;
1711         }
1712
1713         set_bit(STRIPE_HANDLE, &sh->state);
1714         raid5_release_stripe(sh);
1715 }
1716
1717 static void
1718 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1719                      struct dma_async_tx_descriptor *tx)
1720 {
1721         int disks = sh->disks;
1722         struct page **xor_srcs;
1723         struct async_submit_ctl submit;
1724         int count, pd_idx = sh->pd_idx, i;
1725         struct page *xor_dest;
1726         int prexor = 0;
1727         unsigned long flags;
1728         int j = 0;
1729         struct stripe_head *head_sh = sh;
1730         int last_stripe;
1731
1732         pr_debug("%s: stripe %llu\n", __func__,
1733                 (unsigned long long)sh->sector);
1734
1735         for (i = 0; i < sh->disks; i++) {
1736                 if (pd_idx == i)
1737                         continue;
1738                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1739                         break;
1740         }
1741         if (i >= sh->disks) {
1742                 atomic_inc(&sh->count);
1743                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1744                 ops_complete_reconstruct(sh);
1745                 return;
1746         }
1747 again:
1748         count = 0;
1749         xor_srcs = to_addr_page(percpu, j);
1750         /* check if prexor is active which means only process blocks
1751          * that are part of a read-modify-write (written)
1752          */
1753         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1754                 prexor = 1;
1755                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1756                 for (i = disks; i--; ) {
1757                         struct r5dev *dev = &sh->dev[i];
1758                         if (head_sh->dev[i].written ||
1759                             test_bit(R5_InJournal, &head_sh->dev[i].flags))
1760                                 xor_srcs[count++] = dev->page;
1761                 }
1762         } else {
1763                 xor_dest = sh->dev[pd_idx].page;
1764                 for (i = disks; i--; ) {
1765                         struct r5dev *dev = &sh->dev[i];
1766                         if (i != pd_idx)
1767                                 xor_srcs[count++] = dev->page;
1768                 }
1769         }
1770
1771         /* 1/ if we prexor'd then the dest is reused as a source
1772          * 2/ if we did not prexor then we are redoing the parity
1773          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1774          * for the synchronous xor case
1775          */
1776         last_stripe = !head_sh->batch_head ||
1777                 list_first_entry(&sh->batch_list,
1778                                  struct stripe_head, batch_list) == head_sh;
1779         if (last_stripe) {
1780                 flags = ASYNC_TX_ACK |
1781                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1782
1783                 atomic_inc(&head_sh->count);
1784                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1785                                   to_addr_conv(sh, percpu, j));
1786         } else {
1787                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1788                 init_async_submit(&submit, flags, tx, NULL, NULL,
1789                                   to_addr_conv(sh, percpu, j));
1790         }
1791
1792         if (unlikely(count == 1))
1793                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1794         else
1795                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1796         if (!last_stripe) {
1797                 j++;
1798                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1799                                       batch_list);
1800                 goto again;
1801         }
1802 }
1803
1804 static void
1805 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1806                      struct dma_async_tx_descriptor *tx)
1807 {
1808         struct async_submit_ctl submit;
1809         struct page **blocks;
1810         int count, i, j = 0;
1811         struct stripe_head *head_sh = sh;
1812         int last_stripe;
1813         int synflags;
1814         unsigned long txflags;
1815
1816         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1817
1818         for (i = 0; i < sh->disks; i++) {
1819                 if (sh->pd_idx == i || sh->qd_idx == i)
1820                         continue;
1821                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1822                         break;
1823         }
1824         if (i >= sh->disks) {
1825                 atomic_inc(&sh->count);
1826                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1827                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1828                 ops_complete_reconstruct(sh);
1829                 return;
1830         }
1831
1832 again:
1833         blocks = to_addr_page(percpu, j);
1834
1835         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1836                 synflags = SYNDROME_SRC_WRITTEN;
1837                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1838         } else {
1839                 synflags = SYNDROME_SRC_ALL;
1840                 txflags = ASYNC_TX_ACK;
1841         }
1842
1843         count = set_syndrome_sources(blocks, sh, synflags);
1844         last_stripe = !head_sh->batch_head ||
1845                 list_first_entry(&sh->batch_list,
1846                                  struct stripe_head, batch_list) == head_sh;
1847
1848         if (last_stripe) {
1849                 atomic_inc(&head_sh->count);
1850                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1851                                   head_sh, to_addr_conv(sh, percpu, j));
1852         } else
1853                 init_async_submit(&submit, 0, tx, NULL, NULL,
1854                                   to_addr_conv(sh, percpu, j));
1855         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1856         if (!last_stripe) {
1857                 j++;
1858                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1859                                       batch_list);
1860                 goto again;
1861         }
1862 }
1863
1864 static void ops_complete_check(void *stripe_head_ref)
1865 {
1866         struct stripe_head *sh = stripe_head_ref;
1867
1868         pr_debug("%s: stripe %llu\n", __func__,
1869                 (unsigned long long)sh->sector);
1870
1871         sh->check_state = check_state_check_result;
1872         set_bit(STRIPE_HANDLE, &sh->state);
1873         raid5_release_stripe(sh);
1874 }
1875
1876 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1877 {
1878         int disks = sh->disks;
1879         int pd_idx = sh->pd_idx;
1880         int qd_idx = sh->qd_idx;
1881         struct page *xor_dest;
1882         struct page **xor_srcs = to_addr_page(percpu, 0);
1883         struct dma_async_tx_descriptor *tx;
1884         struct async_submit_ctl submit;
1885         int count;
1886         int i;
1887
1888         pr_debug("%s: stripe %llu\n", __func__,
1889                 (unsigned long long)sh->sector);
1890
1891         BUG_ON(sh->batch_head);
1892         count = 0;
1893         xor_dest = sh->dev[pd_idx].page;
1894         xor_srcs[count++] = xor_dest;
1895         for (i = disks; i--; ) {
1896                 if (i == pd_idx || i == qd_idx)
1897                         continue;
1898                 xor_srcs[count++] = sh->dev[i].page;
1899         }
1900
1901         init_async_submit(&submit, 0, NULL, NULL, NULL,
1902                           to_addr_conv(sh, percpu, 0));
1903         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1904                            &sh->ops.zero_sum_result, &submit);
1905
1906         atomic_inc(&sh->count);
1907         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1908         tx = async_trigger_callback(&submit);
1909 }
1910
1911 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1912 {
1913         struct page **srcs = to_addr_page(percpu, 0);
1914         struct async_submit_ctl submit;
1915         int count;
1916
1917         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1918                 (unsigned long long)sh->sector, checkp);
1919
1920         BUG_ON(sh->batch_head);
1921         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1922         if (!checkp)
1923                 srcs[count] = NULL;
1924
1925         atomic_inc(&sh->count);
1926         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1927                           sh, to_addr_conv(sh, percpu, 0));
1928         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1929                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1930 }
1931
1932 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1933 {
1934         int overlap_clear = 0, i, disks = sh->disks;
1935         struct dma_async_tx_descriptor *tx = NULL;
1936         struct r5conf *conf = sh->raid_conf;
1937         int level = conf->level;
1938         struct raid5_percpu *percpu;
1939         unsigned long cpu;
1940
1941         cpu = get_cpu();
1942         percpu = per_cpu_ptr(conf->percpu, cpu);
1943         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1944                 ops_run_biofill(sh);
1945                 overlap_clear++;
1946         }
1947
1948         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1949                 if (level < 6)
1950                         tx = ops_run_compute5(sh, percpu);
1951                 else {
1952                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1953                                 tx = ops_run_compute6_1(sh, percpu);
1954                         else
1955                                 tx = ops_run_compute6_2(sh, percpu);
1956                 }
1957                 /* terminate the chain if reconstruct is not set to be run */
1958                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1959                         async_tx_ack(tx);
1960         }
1961
1962         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1963                 if (level < 6)
1964                         tx = ops_run_prexor5(sh, percpu, tx);
1965                 else
1966                         tx = ops_run_prexor6(sh, percpu, tx);
1967         }
1968
1969         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1970                 tx = ops_run_biodrain(sh, tx);
1971                 overlap_clear++;
1972         }
1973
1974         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1975                 if (level < 6)
1976                         ops_run_reconstruct5(sh, percpu, tx);
1977                 else
1978                         ops_run_reconstruct6(sh, percpu, tx);
1979         }
1980
1981         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1982                 if (sh->check_state == check_state_run)
1983                         ops_run_check_p(sh, percpu);
1984                 else if (sh->check_state == check_state_run_q)
1985                         ops_run_check_pq(sh, percpu, 0);
1986                 else if (sh->check_state == check_state_run_pq)
1987                         ops_run_check_pq(sh, percpu, 1);
1988                 else
1989                         BUG();
1990         }
1991
1992         if (overlap_clear && !sh->batch_head)
1993                 for (i = disks; i--; ) {
1994                         struct r5dev *dev = &sh->dev[i];
1995                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1996                                 wake_up(&sh->raid_conf->wait_for_overlap);
1997                 }
1998         put_cpu();
1999 }
2000
2001 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2002         int disks)
2003 {
2004         struct stripe_head *sh;
2005         int i;
2006
2007         sh = kmem_cache_zalloc(sc, gfp);
2008         if (sh) {
2009                 spin_lock_init(&sh->stripe_lock);
2010                 spin_lock_init(&sh->batch_lock);
2011                 INIT_LIST_HEAD(&sh->batch_list);
2012                 INIT_LIST_HEAD(&sh->lru);
2013                 INIT_LIST_HEAD(&sh->r5c);
2014                 INIT_LIST_HEAD(&sh->log_list);
2015                 atomic_set(&sh->count, 1);
2016                 sh->log_start = MaxSector;
2017                 for (i = 0; i < disks; i++) {
2018                         struct r5dev *dev = &sh->dev[i];
2019
2020                         bio_init(&dev->req, &dev->vec, 1);
2021                         bio_init(&dev->rreq, &dev->rvec, 1);
2022                 }
2023         }
2024         return sh;
2025 }
2026 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2027 {
2028         struct stripe_head *sh;
2029
2030         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size);
2031         if (!sh)
2032                 return 0;
2033
2034         sh->raid_conf = conf;
2035
2036         if (grow_buffers(sh, gfp)) {
2037                 shrink_buffers(sh);
2038                 kmem_cache_free(conf->slab_cache, sh);
2039                 return 0;
2040         }
2041         sh->hash_lock_index =
2042                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2043         /* we just created an active stripe so... */
2044         atomic_inc(&conf->active_stripes);
2045
2046         raid5_release_stripe(sh);
2047         conf->max_nr_stripes++;
2048         return 1;
2049 }
2050
2051 static int grow_stripes(struct r5conf *conf, int num)
2052 {
2053         struct kmem_cache *sc;
2054         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2055
2056         if (conf->mddev->gendisk)
2057                 sprintf(conf->cache_name[0],
2058                         "raid%d-%s", conf->level, mdname(conf->mddev));
2059         else
2060                 sprintf(conf->cache_name[0],
2061                         "raid%d-%p", conf->level, conf->mddev);
2062         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2063
2064         conf->active_name = 0;
2065         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2066                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2067                                0, 0, NULL);
2068         if (!sc)
2069                 return 1;
2070         conf->slab_cache = sc;
2071         conf->pool_size = devs;
2072         while (num--)
2073                 if (!grow_one_stripe(conf, GFP_KERNEL))
2074                         return 1;
2075
2076         return 0;
2077 }
2078
2079 /**
2080  * scribble_len - return the required size of the scribble region
2081  * @num - total number of disks in the array
2082  *
2083  * The size must be enough to contain:
2084  * 1/ a struct page pointer for each device in the array +2
2085  * 2/ room to convert each entry in (1) to its corresponding dma
2086  *    (dma_map_page()) or page (page_address()) address.
2087  *
2088  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2089  * calculate over all devices (not just the data blocks), using zeros in place
2090  * of the P and Q blocks.
2091  */
2092 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2093 {
2094         struct flex_array *ret;
2095         size_t len;
2096
2097         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2098         ret = flex_array_alloc(len, cnt, flags);
2099         if (!ret)
2100                 return NULL;
2101         /* always prealloc all elements, so no locking is required */
2102         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2103                 flex_array_free(ret);
2104                 return NULL;
2105         }
2106         return ret;
2107 }
2108
2109 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2110 {
2111         unsigned long cpu;
2112         int err = 0;
2113
2114         /*
2115          * Never shrink. And mddev_suspend() could deadlock if this is called
2116          * from raid5d. In that case, scribble_disks and scribble_sectors
2117          * should equal to new_disks and new_sectors
2118          */
2119         if (conf->scribble_disks >= new_disks &&
2120             conf->scribble_sectors >= new_sectors)
2121                 return 0;
2122         mddev_suspend(conf->mddev);
2123         get_online_cpus();
2124         for_each_present_cpu(cpu) {
2125                 struct raid5_percpu *percpu;
2126                 struct flex_array *scribble;
2127
2128                 percpu = per_cpu_ptr(conf->percpu, cpu);
2129                 scribble = scribble_alloc(new_disks,
2130                                           new_sectors / STRIPE_SECTORS,
2131                                           GFP_NOIO);
2132
2133                 if (scribble) {
2134                         flex_array_free(percpu->scribble);
2135                         percpu->scribble = scribble;
2136                 } else {
2137                         err = -ENOMEM;
2138                         break;
2139                 }
2140         }
2141         put_online_cpus();
2142         mddev_resume(conf->mddev);
2143         if (!err) {
2144                 conf->scribble_disks = new_disks;
2145                 conf->scribble_sectors = new_sectors;
2146         }
2147         return err;
2148 }
2149
2150 static int resize_stripes(struct r5conf *conf, int newsize)
2151 {
2152         /* Make all the stripes able to hold 'newsize' devices.
2153          * New slots in each stripe get 'page' set to a new page.
2154          *
2155          * This happens in stages:
2156          * 1/ create a new kmem_cache and allocate the required number of
2157          *    stripe_heads.
2158          * 2/ gather all the old stripe_heads and transfer the pages across
2159          *    to the new stripe_heads.  This will have the side effect of
2160          *    freezing the array as once all stripe_heads have been collected,
2161          *    no IO will be possible.  Old stripe heads are freed once their
2162          *    pages have been transferred over, and the old kmem_cache is
2163          *    freed when all stripes are done.
2164          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2165          *    we simple return a failre status - no need to clean anything up.
2166          * 4/ allocate new pages for the new slots in the new stripe_heads.
2167          *    If this fails, we don't bother trying the shrink the
2168          *    stripe_heads down again, we just leave them as they are.
2169          *    As each stripe_head is processed the new one is released into
2170          *    active service.
2171          *
2172          * Once step2 is started, we cannot afford to wait for a write,
2173          * so we use GFP_NOIO allocations.
2174          */
2175         struct stripe_head *osh, *nsh;
2176         LIST_HEAD(newstripes);
2177         struct disk_info *ndisks;
2178         int err;
2179         struct kmem_cache *sc;
2180         int i;
2181         int hash, cnt;
2182
2183         if (newsize <= conf->pool_size)
2184                 return 0; /* never bother to shrink */
2185
2186         err = md_allow_write(conf->mddev);
2187         if (err)
2188                 return err;
2189
2190         /* Step 1 */
2191         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2192                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2193                                0, 0, NULL);
2194         if (!sc)
2195                 return -ENOMEM;
2196
2197         /* Need to ensure auto-resizing doesn't interfere */
2198         mutex_lock(&conf->cache_size_mutex);
2199
2200         for (i = conf->max_nr_stripes; i; i--) {
2201                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize);
2202                 if (!nsh)
2203                         break;
2204
2205                 nsh->raid_conf = conf;
2206                 list_add(&nsh->lru, &newstripes);
2207         }
2208         if (i) {
2209                 /* didn't get enough, give up */
2210                 while (!list_empty(&newstripes)) {
2211                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2212                         list_del(&nsh->lru);
2213                         kmem_cache_free(sc, nsh);
2214                 }
2215                 kmem_cache_destroy(sc);
2216                 mutex_unlock(&conf->cache_size_mutex);
2217                 return -ENOMEM;
2218         }
2219         /* Step 2 - Must use GFP_NOIO now.
2220          * OK, we have enough stripes, start collecting inactive
2221          * stripes and copying them over
2222          */
2223         hash = 0;
2224         cnt = 0;
2225         list_for_each_entry(nsh, &newstripes, lru) {
2226                 lock_device_hash_lock(conf, hash);
2227                 wait_event_cmd(conf->wait_for_stripe,
2228                                     !list_empty(conf->inactive_list + hash),
2229                                     unlock_device_hash_lock(conf, hash),
2230                                     lock_device_hash_lock(conf, hash));
2231                 osh = get_free_stripe(conf, hash);
2232                 unlock_device_hash_lock(conf, hash);
2233
2234                 for(i=0; i<conf->pool_size; i++) {
2235                         nsh->dev[i].page = osh->dev[i].page;
2236                         nsh->dev[i].orig_page = osh->dev[i].page;
2237                 }
2238                 nsh->hash_lock_index = hash;
2239                 kmem_cache_free(conf->slab_cache, osh);
2240                 cnt++;
2241                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2242                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2243                         hash++;
2244                         cnt = 0;
2245                 }
2246         }
2247         kmem_cache_destroy(conf->slab_cache);
2248
2249         /* Step 3.
2250          * At this point, we are holding all the stripes so the array
2251          * is completely stalled, so now is a good time to resize
2252          * conf->disks and the scribble region
2253          */
2254         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2255         if (ndisks) {
2256                 for (i = 0; i < conf->pool_size; i++)
2257                         ndisks[i] = conf->disks[i];
2258
2259                 for (i = conf->pool_size; i < newsize; i++) {
2260                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2261                         if (!ndisks[i].extra_page)
2262                                 err = -ENOMEM;
2263                 }
2264
2265                 if (err) {
2266                         for (i = conf->pool_size; i < newsize; i++)
2267                                 if (ndisks[i].extra_page)
2268                                         put_page(ndisks[i].extra_page);
2269                         kfree(ndisks);
2270                 } else {
2271                         kfree(conf->disks);
2272                         conf->disks = ndisks;
2273                 }
2274         } else
2275                 err = -ENOMEM;
2276
2277         mutex_unlock(&conf->cache_size_mutex);
2278         /* Step 4, return new stripes to service */
2279         while(!list_empty(&newstripes)) {
2280                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2281                 list_del_init(&nsh->lru);
2282
2283                 for (i=conf->raid_disks; i < newsize; i++)
2284                         if (nsh->dev[i].page == NULL) {
2285                                 struct page *p = alloc_page(GFP_NOIO);
2286                                 nsh->dev[i].page = p;
2287                                 nsh->dev[i].orig_page = p;
2288                                 if (!p)
2289                                         err = -ENOMEM;
2290                         }
2291                 raid5_release_stripe(nsh);
2292         }
2293         /* critical section pass, GFP_NOIO no longer needed */
2294
2295         conf->slab_cache = sc;
2296         conf->active_name = 1-conf->active_name;
2297         if (!err)
2298                 conf->pool_size = newsize;
2299         return err;
2300 }
2301
2302 static int drop_one_stripe(struct r5conf *conf)
2303 {
2304         struct stripe_head *sh;
2305         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2306
2307         spin_lock_irq(conf->hash_locks + hash);
2308         sh = get_free_stripe(conf, hash);
2309         spin_unlock_irq(conf->hash_locks + hash);
2310         if (!sh)
2311                 return 0;
2312         BUG_ON(atomic_read(&sh->count));
2313         shrink_buffers(sh);
2314         kmem_cache_free(conf->slab_cache, sh);
2315         atomic_dec(&conf->active_stripes);
2316         conf->max_nr_stripes--;
2317         return 1;
2318 }
2319
2320 static void shrink_stripes(struct r5conf *conf)
2321 {
2322         while (conf->max_nr_stripes &&
2323                drop_one_stripe(conf))
2324                 ;
2325
2326         kmem_cache_destroy(conf->slab_cache);
2327         conf->slab_cache = NULL;
2328 }
2329
2330 static void raid5_end_read_request(struct bio * bi)
2331 {
2332         struct stripe_head *sh = bi->bi_private;
2333         struct r5conf *conf = sh->raid_conf;
2334         int disks = sh->disks, i;
2335         char b[BDEVNAME_SIZE];
2336         struct md_rdev *rdev = NULL;
2337         sector_t s;
2338
2339         for (i=0 ; i<disks; i++)
2340                 if (bi == &sh->dev[i].req)
2341                         break;
2342
2343         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2344                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2345                 bi->bi_error);
2346         if (i == disks) {
2347                 bio_reset(bi);
2348                 BUG();
2349                 return;
2350         }
2351         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2352                 /* If replacement finished while this request was outstanding,
2353                  * 'replacement' might be NULL already.
2354                  * In that case it moved down to 'rdev'.
2355                  * rdev is not removed until all requests are finished.
2356                  */
2357                 rdev = conf->disks[i].replacement;
2358         if (!rdev)
2359                 rdev = conf->disks[i].rdev;
2360
2361         if (use_new_offset(conf, sh))
2362                 s = sh->sector + rdev->new_data_offset;
2363         else
2364                 s = sh->sector + rdev->data_offset;
2365         if (!bi->bi_error) {
2366                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2367                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2368                         /* Note that this cannot happen on a
2369                          * replacement device.  We just fail those on
2370                          * any error
2371                          */
2372                         pr_info_ratelimited(
2373                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2374                                 mdname(conf->mddev), STRIPE_SECTORS,
2375                                 (unsigned long long)s,
2376                                 bdevname(rdev->bdev, b));
2377                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2378                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2379                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2380                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2381                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2382
2383                 if (atomic_read(&rdev->read_errors))
2384                         atomic_set(&rdev->read_errors, 0);
2385         } else {
2386                 const char *bdn = bdevname(rdev->bdev, b);
2387                 int retry = 0;
2388                 int set_bad = 0;
2389
2390                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2391                 atomic_inc(&rdev->read_errors);
2392                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2393                         pr_warn_ratelimited(
2394                                 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2395                                 mdname(conf->mddev),
2396                                 (unsigned long long)s,
2397                                 bdn);
2398                 else if (conf->mddev->degraded >= conf->max_degraded) {
2399                         set_bad = 1;
2400                         pr_warn_ratelimited(
2401                                 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2402                                 mdname(conf->mddev),
2403                                 (unsigned long long)s,
2404                                 bdn);
2405                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2406                         /* Oh, no!!! */
2407                         set_bad = 1;
2408                         pr_warn_ratelimited(
2409                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2410                                 mdname(conf->mddev),
2411                                 (unsigned long long)s,
2412                                 bdn);
2413                 } else if (atomic_read(&rdev->read_errors)
2414                          > conf->max_nr_stripes)
2415                         pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2416                                mdname(conf->mddev), bdn);
2417                 else
2418                         retry = 1;
2419                 if (set_bad && test_bit(In_sync, &rdev->flags)
2420                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2421                         retry = 1;
2422                 if (retry)
2423                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2424                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2425                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2426                         } else
2427                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2428                 else {
2429                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2430                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2431                         if (!(set_bad
2432                               && test_bit(In_sync, &rdev->flags)
2433                               && rdev_set_badblocks(
2434                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2435                                 md_error(conf->mddev, rdev);
2436                 }
2437         }
2438         rdev_dec_pending(rdev, conf->mddev);
2439         bio_reset(bi);
2440         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2441         set_bit(STRIPE_HANDLE, &sh->state);
2442         raid5_release_stripe(sh);
2443 }
2444
2445 static void raid5_end_write_request(struct bio *bi)
2446 {
2447         struct stripe_head *sh = bi->bi_private;
2448         struct r5conf *conf = sh->raid_conf;
2449         int disks = sh->disks, i;
2450         struct md_rdev *uninitialized_var(rdev);
2451         sector_t first_bad;
2452         int bad_sectors;
2453         int replacement = 0;
2454
2455         for (i = 0 ; i < disks; i++) {
2456                 if (bi == &sh->dev[i].req) {
2457                         rdev = conf->disks[i].rdev;
2458                         break;
2459                 }
2460                 if (bi == &sh->dev[i].rreq) {
2461                         rdev = conf->disks[i].replacement;
2462                         if (rdev)
2463                                 replacement = 1;
2464                         else
2465                                 /* rdev was removed and 'replacement'
2466                                  * replaced it.  rdev is not removed
2467                                  * until all requests are finished.
2468                                  */
2469                                 rdev = conf->disks[i].rdev;
2470                         break;
2471                 }
2472         }
2473         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2474                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2475                 bi->bi_error);
2476         if (i == disks) {
2477                 bio_reset(bi);
2478                 BUG();
2479                 return;
2480         }
2481
2482         if (replacement) {
2483                 if (bi->bi_error)
2484                         md_error(conf->mddev, rdev);
2485                 else if (is_badblock(rdev, sh->sector,
2486                                      STRIPE_SECTORS,
2487                                      &first_bad, &bad_sectors))
2488                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2489         } else {
2490                 if (bi->bi_error) {
2491                         set_bit(STRIPE_DEGRADED, &sh->state);
2492                         set_bit(WriteErrorSeen, &rdev->flags);
2493                         set_bit(R5_WriteError, &sh->dev[i].flags);
2494                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2495                                 set_bit(MD_RECOVERY_NEEDED,
2496                                         &rdev->mddev->recovery);
2497                 } else if (is_badblock(rdev, sh->sector,
2498                                        STRIPE_SECTORS,
2499                                        &first_bad, &bad_sectors)) {
2500                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2501                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2502                                 /* That was a successful write so make
2503                                  * sure it looks like we already did
2504                                  * a re-write.
2505                                  */
2506                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2507                 }
2508         }
2509         rdev_dec_pending(rdev, conf->mddev);
2510
2511         if (sh->batch_head && bi->bi_error && !replacement)
2512                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2513
2514         bio_reset(bi);
2515         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2516                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2517         set_bit(STRIPE_HANDLE, &sh->state);
2518         raid5_release_stripe(sh);
2519
2520         if (sh->batch_head && sh != sh->batch_head)
2521                 raid5_release_stripe(sh->batch_head);
2522 }
2523
2524 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2525 {
2526         struct r5dev *dev = &sh->dev[i];
2527
2528         dev->flags = 0;
2529         dev->sector = raid5_compute_blocknr(sh, i, previous);
2530 }
2531
2532 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2533 {
2534         char b[BDEVNAME_SIZE];
2535         struct r5conf *conf = mddev->private;
2536         unsigned long flags;
2537         pr_debug("raid456: error called\n");
2538
2539         spin_lock_irqsave(&conf->device_lock, flags);
2540         clear_bit(In_sync, &rdev->flags);
2541         mddev->degraded = calc_degraded(conf);
2542         spin_unlock_irqrestore(&conf->device_lock, flags);
2543         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2544
2545         set_bit(Blocked, &rdev->flags);
2546         set_bit(Faulty, &rdev->flags);
2547         set_mask_bits(&mddev->sb_flags, 0,
2548                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2549         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2550                 "md/raid:%s: Operation continuing on %d devices.\n",
2551                 mdname(mddev),
2552                 bdevname(rdev->bdev, b),
2553                 mdname(mddev),
2554                 conf->raid_disks - mddev->degraded);
2555 }
2556
2557 /*
2558  * Input: a 'big' sector number,
2559  * Output: index of the data and parity disk, and the sector # in them.
2560  */
2561 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2562                               int previous, int *dd_idx,
2563                               struct stripe_head *sh)
2564 {
2565         sector_t stripe, stripe2;
2566         sector_t chunk_number;
2567         unsigned int chunk_offset;
2568         int pd_idx, qd_idx;
2569         int ddf_layout = 0;
2570         sector_t new_sector;
2571         int algorithm = previous ? conf->prev_algo
2572                                  : conf->algorithm;
2573         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2574                                          : conf->chunk_sectors;
2575         int raid_disks = previous ? conf->previous_raid_disks
2576                                   : conf->raid_disks;
2577         int data_disks = raid_disks - conf->max_degraded;
2578
2579         /* First compute the information on this sector */
2580
2581         /*
2582          * Compute the chunk number and the sector offset inside the chunk
2583          */
2584         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2585         chunk_number = r_sector;
2586
2587         /*
2588          * Compute the stripe number
2589          */
2590         stripe = chunk_number;
2591         *dd_idx = sector_div(stripe, data_disks);
2592         stripe2 = stripe;
2593         /*
2594          * Select the parity disk based on the user selected algorithm.
2595          */
2596         pd_idx = qd_idx = -1;
2597         switch(conf->level) {
2598         case 4:
2599                 pd_idx = data_disks;
2600                 break;
2601         case 5:
2602                 switch (algorithm) {
2603                 case ALGORITHM_LEFT_ASYMMETRIC:
2604                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2605                         if (*dd_idx >= pd_idx)
2606                                 (*dd_idx)++;
2607                         break;
2608                 case ALGORITHM_RIGHT_ASYMMETRIC:
2609                         pd_idx = sector_div(stripe2, raid_disks);
2610                         if (*dd_idx >= pd_idx)
2611                                 (*dd_idx)++;
2612                         break;
2613                 case ALGORITHM_LEFT_SYMMETRIC:
2614                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2615                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2616                         break;
2617                 case ALGORITHM_RIGHT_SYMMETRIC:
2618                         pd_idx = sector_div(stripe2, raid_disks);
2619                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2620                         break;
2621                 case ALGORITHM_PARITY_0:
2622                         pd_idx = 0;
2623                         (*dd_idx)++;
2624                         break;
2625                 case ALGORITHM_PARITY_N:
2626                         pd_idx = data_disks;
2627                         break;
2628                 default:
2629                         BUG();
2630                 }
2631                 break;
2632         case 6:
2633
2634                 switch (algorithm) {
2635                 case ALGORITHM_LEFT_ASYMMETRIC:
2636                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2637                         qd_idx = pd_idx + 1;
2638                         if (pd_idx == raid_disks-1) {
2639                                 (*dd_idx)++;    /* Q D D D P */
2640                                 qd_idx = 0;
2641                         } else if (*dd_idx >= pd_idx)
2642                                 (*dd_idx) += 2; /* D D P Q D */
2643                         break;
2644                 case ALGORITHM_RIGHT_ASYMMETRIC:
2645                         pd_idx = sector_div(stripe2, raid_disks);
2646                         qd_idx = pd_idx + 1;
2647                         if (pd_idx == raid_disks-1) {
2648                                 (*dd_idx)++;    /* Q D D D P */
2649                                 qd_idx = 0;
2650                         } else if (*dd_idx >= pd_idx)
2651                                 (*dd_idx) += 2; /* D D P Q D */
2652                         break;
2653                 case ALGORITHM_LEFT_SYMMETRIC:
2654                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2655                         qd_idx = (pd_idx + 1) % raid_disks;
2656                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2657                         break;
2658                 case ALGORITHM_RIGHT_SYMMETRIC:
2659                         pd_idx = sector_div(stripe2, raid_disks);
2660                         qd_idx = (pd_idx + 1) % raid_disks;
2661                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2662                         break;
2663
2664                 case ALGORITHM_PARITY_0:
2665                         pd_idx = 0;
2666                         qd_idx = 1;
2667                         (*dd_idx) += 2;
2668                         break;
2669                 case ALGORITHM_PARITY_N:
2670                         pd_idx = data_disks;
2671                         qd_idx = data_disks + 1;
2672                         break;
2673
2674                 case ALGORITHM_ROTATING_ZERO_RESTART:
2675                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2676                          * of blocks for computing Q is different.
2677                          */
2678                         pd_idx = sector_div(stripe2, raid_disks);
2679                         qd_idx = pd_idx + 1;
2680                         if (pd_idx == raid_disks-1) {
2681                                 (*dd_idx)++;    /* Q D D D P */
2682                                 qd_idx = 0;
2683                         } else if (*dd_idx >= pd_idx)
2684                                 (*dd_idx) += 2; /* D D P Q D */
2685                         ddf_layout = 1;
2686                         break;
2687
2688                 case ALGORITHM_ROTATING_N_RESTART:
2689                         /* Same a left_asymmetric, by first stripe is
2690                          * D D D P Q  rather than
2691                          * Q D D D P
2692                          */
2693                         stripe2 += 1;
2694                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2695                         qd_idx = pd_idx + 1;
2696                         if (pd_idx == raid_disks-1) {
2697                                 (*dd_idx)++;    /* Q D D D P */
2698                                 qd_idx = 0;
2699                         } else if (*dd_idx >= pd_idx)
2700                                 (*dd_idx) += 2; /* D D P Q D */
2701                         ddf_layout = 1;
2702                         break;
2703
2704                 case ALGORITHM_ROTATING_N_CONTINUE:
2705                         /* Same as left_symmetric but Q is before P */
2706                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2707                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2708                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2709                         ddf_layout = 1;
2710                         break;
2711
2712                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2713                         /* RAID5 left_asymmetric, with Q on last device */
2714                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2715                         if (*dd_idx >= pd_idx)
2716                                 (*dd_idx)++;
2717                         qd_idx = raid_disks - 1;
2718                         break;
2719
2720                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2721                         pd_idx = sector_div(stripe2, raid_disks-1);
2722                         if (*dd_idx >= pd_idx)
2723                                 (*dd_idx)++;
2724                         qd_idx = raid_disks - 1;
2725                         break;
2726
2727                 case ALGORITHM_LEFT_SYMMETRIC_6:
2728                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2729                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2730                         qd_idx = raid_disks - 1;
2731                         break;
2732
2733                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2734                         pd_idx = sector_div(stripe2, raid_disks-1);
2735                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2736                         qd_idx = raid_disks - 1;
2737                         break;
2738
2739                 case ALGORITHM_PARITY_0_6:
2740                         pd_idx = 0;
2741                         (*dd_idx)++;
2742                         qd_idx = raid_disks - 1;
2743                         break;
2744
2745                 default:
2746                         BUG();
2747                 }
2748                 break;
2749         }
2750
2751         if (sh) {
2752                 sh->pd_idx = pd_idx;
2753                 sh->qd_idx = qd_idx;
2754                 sh->ddf_layout = ddf_layout;
2755         }
2756         /*
2757          * Finally, compute the new sector number
2758          */
2759         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2760         return new_sector;
2761 }
2762
2763 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2764 {
2765         struct r5conf *conf = sh->raid_conf;
2766         int raid_disks = sh->disks;
2767         int data_disks = raid_disks - conf->max_degraded;
2768         sector_t new_sector = sh->sector, check;
2769         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2770                                          : conf->chunk_sectors;
2771         int algorithm = previous ? conf->prev_algo
2772                                  : conf->algorithm;
2773         sector_t stripe;
2774         int chunk_offset;
2775         sector_t chunk_number;
2776         int dummy1, dd_idx = i;
2777         sector_t r_sector;
2778         struct stripe_head sh2;
2779
2780         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2781         stripe = new_sector;
2782
2783         if (i == sh->pd_idx)
2784                 return 0;
2785         switch(conf->level) {
2786         case 4: break;
2787         case 5:
2788                 switch (algorithm) {
2789                 case ALGORITHM_LEFT_ASYMMETRIC:
2790                 case ALGORITHM_RIGHT_ASYMMETRIC:
2791                         if (i > sh->pd_idx)
2792                                 i--;
2793                         break;
2794                 case ALGORITHM_LEFT_SYMMETRIC:
2795                 case ALGORITHM_RIGHT_SYMMETRIC:
2796                         if (i < sh->pd_idx)
2797                                 i += raid_disks;
2798                         i -= (sh->pd_idx + 1);
2799                         break;
2800                 case ALGORITHM_PARITY_0:
2801                         i -= 1;
2802                         break;
2803                 case ALGORITHM_PARITY_N:
2804                         break;
2805                 default:
2806                         BUG();
2807                 }
2808                 break;
2809         case 6:
2810                 if (i == sh->qd_idx)
2811                         return 0; /* It is the Q disk */
2812                 switch (algorithm) {
2813                 case ALGORITHM_LEFT_ASYMMETRIC:
2814                 case ALGORITHM_RIGHT_ASYMMETRIC:
2815                 case ALGORITHM_ROTATING_ZERO_RESTART:
2816                 case ALGORITHM_ROTATING_N_RESTART:
2817                         if (sh->pd_idx == raid_disks-1)
2818                                 i--;    /* Q D D D P */
2819                         else if (i > sh->pd_idx)
2820                                 i -= 2; /* D D P Q D */
2821                         break;
2822                 case ALGORITHM_LEFT_SYMMETRIC:
2823                 case ALGORITHM_RIGHT_SYMMETRIC:
2824                         if (sh->pd_idx == raid_disks-1)
2825                                 i--; /* Q D D D P */
2826                         else {
2827                                 /* D D P Q D */
2828                                 if (i < sh->pd_idx)
2829                                         i += raid_disks;
2830                                 i -= (sh->pd_idx + 2);
2831                         }
2832                         break;
2833                 case ALGORITHM_PARITY_0:
2834                         i -= 2;
2835                         break;
2836                 case ALGORITHM_PARITY_N:
2837                         break;
2838                 case ALGORITHM_ROTATING_N_CONTINUE:
2839                         /* Like left_symmetric, but P is before Q */
2840                         if (sh->pd_idx == 0)
2841                                 i--;    /* P D D D Q */
2842                         else {
2843                                 /* D D Q P D */
2844                                 if (i < sh->pd_idx)
2845                                         i += raid_disks;
2846                                 i -= (sh->pd_idx + 1);
2847                         }
2848                         break;
2849                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2850                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2851                         if (i > sh->pd_idx)
2852                                 i--;
2853                         break;
2854                 case ALGORITHM_LEFT_SYMMETRIC_6:
2855                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2856                         if (i < sh->pd_idx)
2857                                 i += data_disks + 1;
2858                         i -= (sh->pd_idx + 1);
2859                         break;
2860                 case ALGORITHM_PARITY_0_6:
2861                         i -= 1;
2862                         break;
2863                 default:
2864                         BUG();
2865                 }
2866                 break;
2867         }
2868
2869         chunk_number = stripe * data_disks + i;
2870         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2871
2872         check = raid5_compute_sector(conf, r_sector,
2873                                      previous, &dummy1, &sh2);
2874         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2875                 || sh2.qd_idx != sh->qd_idx) {
2876                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
2877                         mdname(conf->mddev));
2878                 return 0;
2879         }
2880         return r_sector;
2881 }
2882
2883 static void
2884 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2885                          int rcw, int expand)
2886 {
2887         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2888         struct r5conf *conf = sh->raid_conf;
2889         int level = conf->level;
2890
2891         if (rcw) {
2892                 /*
2893                  * In some cases, handle_stripe_dirtying initially decided to
2894                  * run rmw and allocates extra page for prexor. However, rcw is
2895                  * cheaper later on. We need to free the extra page now,
2896                  * because we won't be able to do that in ops_complete_prexor().
2897                  */
2898                 r5c_release_extra_page(sh);
2899
2900                 for (i = disks; i--; ) {
2901                         struct r5dev *dev = &sh->dev[i];
2902
2903                         if (dev->towrite) {
2904                                 set_bit(R5_LOCKED, &dev->flags);
2905                                 set_bit(R5_Wantdrain, &dev->flags);
2906                                 if (!expand)
2907                                         clear_bit(R5_UPTODATE, &dev->flags);
2908                                 s->locked++;
2909                         } else if (test_bit(R5_InJournal, &dev->flags)) {
2910                                 set_bit(R5_LOCKED, &dev->flags);
2911                                 s->locked++;
2912                         }
2913                 }
2914                 /* if we are not expanding this is a proper write request, and
2915                  * there will be bios with new data to be drained into the
2916                  * stripe cache
2917                  */
2918                 if (!expand) {
2919                         if (!s->locked)
2920                                 /* False alarm, nothing to do */
2921                                 return;
2922                         sh->reconstruct_state = reconstruct_state_drain_run;
2923                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2924                 } else
2925                         sh->reconstruct_state = reconstruct_state_run;
2926
2927                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2928
2929                 if (s->locked + conf->max_degraded == disks)
2930                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2931                                 atomic_inc(&conf->pending_full_writes);
2932         } else {
2933                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2934                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2935                 BUG_ON(level == 6 &&
2936                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2937                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2938
2939                 for (i = disks; i--; ) {
2940                         struct r5dev *dev = &sh->dev[i];
2941                         if (i == pd_idx || i == qd_idx)
2942                                 continue;
2943
2944                         if (dev->towrite &&
2945                             (test_bit(R5_UPTODATE, &dev->flags) ||
2946                              test_bit(R5_Wantcompute, &dev->flags))) {
2947                                 set_bit(R5_Wantdrain, &dev->flags);
2948                                 set_bit(R5_LOCKED, &dev->flags);
2949                                 clear_bit(R5_UPTODATE, &dev->flags);
2950                                 s->locked++;
2951                         } else if (test_bit(R5_InJournal, &dev->flags)) {
2952                                 set_bit(R5_LOCKED, &dev->flags);
2953                                 s->locked++;
2954                         }
2955                 }
2956                 if (!s->locked)
2957                         /* False alarm - nothing to do */
2958                         return;
2959                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2960                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2961                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2962                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2963         }
2964
2965         /* keep the parity disk(s) locked while asynchronous operations
2966          * are in flight
2967          */
2968         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2969         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2970         s->locked++;
2971
2972         if (level == 6) {
2973                 int qd_idx = sh->qd_idx;
2974                 struct r5dev *dev = &sh->dev[qd_idx];
2975
2976                 set_bit(R5_LOCKED, &dev->flags);
2977                 clear_bit(R5_UPTODATE, &dev->flags);
2978                 s->locked++;
2979         }
2980
2981         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2982                 __func__, (unsigned long long)sh->sector,
2983                 s->locked, s->ops_request);
2984 }
2985
2986 /*
2987  * Each stripe/dev can have one or more bion attached.
2988  * toread/towrite point to the first in a chain.
2989  * The bi_next chain must be in order.
2990  */
2991 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2992                           int forwrite, int previous)
2993 {
2994         struct bio **bip;
2995         struct r5conf *conf = sh->raid_conf;
2996         int firstwrite=0;
2997
2998         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2999                 (unsigned long long)bi->bi_iter.bi_sector,
3000                 (unsigned long long)sh->sector);
3001
3002         /*
3003          * If several bio share a stripe. The bio bi_phys_segments acts as a
3004          * reference count to avoid race. The reference count should already be
3005          * increased before this function is called (for example, in
3006          * raid5_make_request()), so other bio sharing this stripe will not free the
3007          * stripe. If a stripe is owned by one stripe, the stripe lock will
3008          * protect it.
3009          */
3010         spin_lock_irq(&sh->stripe_lock);
3011         /* Don't allow new IO added to stripes in batch list */
3012         if (sh->batch_head)
3013                 goto overlap;
3014         if (forwrite) {
3015                 bip = &sh->dev[dd_idx].towrite;
3016                 if (*bip == NULL)
3017                         firstwrite = 1;
3018         } else
3019                 bip = &sh->dev[dd_idx].toread;
3020         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3021                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3022                         goto overlap;
3023                 bip = & (*bip)->bi_next;
3024         }
3025         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3026                 goto overlap;
3027
3028         if (!forwrite || previous)
3029                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3030
3031         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3032         if (*bip)
3033                 bi->bi_next = *bip;
3034         *bip = bi;
3035         raid5_inc_bi_active_stripes(bi);
3036
3037         if (forwrite) {
3038                 /* check if page is covered */
3039                 sector_t sector = sh->dev[dd_idx].sector;
3040                 for (bi=sh->dev[dd_idx].towrite;
3041                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3042                              bi && bi->bi_iter.bi_sector <= sector;
3043                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3044                         if (bio_end_sector(bi) >= sector)
3045                                 sector = bio_end_sector(bi);
3046                 }
3047                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3048                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3049                                 sh->overwrite_disks++;
3050         }
3051
3052         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3053                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3054                 (unsigned long long)sh->sector, dd_idx);
3055
3056         if (conf->mddev->bitmap && firstwrite) {
3057                 /* Cannot hold spinlock over bitmap_startwrite,
3058                  * but must ensure this isn't added to a batch until
3059                  * we have added to the bitmap and set bm_seq.
3060                  * So set STRIPE_BITMAP_PENDING to prevent
3061                  * batching.
3062                  * If multiple add_stripe_bio() calls race here they
3063                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3064                  * to complete "bitmap_startwrite" gets to set
3065                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3066                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3067                  * any more.
3068                  */
3069                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3070                 spin_unlock_irq(&sh->stripe_lock);
3071                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3072                                   STRIPE_SECTORS, 0);
3073                 spin_lock_irq(&sh->stripe_lock);
3074                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3075                 if (!sh->batch_head) {
3076                         sh->bm_seq = conf->seq_flush+1;
3077                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3078                 }
3079         }
3080         spin_unlock_irq(&sh->stripe_lock);
3081
3082         if (stripe_can_batch(sh))
3083                 stripe_add_to_batch_list(conf, sh);
3084         return 1;
3085
3086  overlap:
3087         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3088         spin_unlock_irq(&sh->stripe_lock);
3089         return 0;
3090 }
3091
3092 static void end_reshape(struct r5conf *conf);
3093
3094 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3095                             struct stripe_head *sh)
3096 {
3097         int sectors_per_chunk =
3098                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3099         int dd_idx;
3100         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3101         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3102
3103         raid5_compute_sector(conf,
3104                              stripe * (disks - conf->max_degraded)
3105                              *sectors_per_chunk + chunk_offset,
3106                              previous,
3107                              &dd_idx, sh);
3108 }
3109
3110 static void
3111 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3112                                 struct stripe_head_state *s, int disks,
3113                                 struct bio_list *return_bi)
3114 {
3115         int i;
3116         BUG_ON(sh->batch_head);
3117         for (i = disks; i--; ) {
3118                 struct bio *bi;
3119                 int bitmap_end = 0;
3120
3121                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3122                         struct md_rdev *rdev;
3123                         rcu_read_lock();
3124                         rdev = rcu_dereference(conf->disks[i].rdev);
3125                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3126                             !test_bit(Faulty, &rdev->flags))
3127                                 atomic_inc(&rdev->nr_pending);
3128                         else
3129                                 rdev = NULL;
3130                         rcu_read_unlock();
3131                         if (rdev) {
3132                                 if (!rdev_set_badblocks(
3133                                             rdev,
3134                                             sh->sector,
3135                                             STRIPE_SECTORS, 0))
3136                                         md_error(conf->mddev, rdev);
3137                                 rdev_dec_pending(rdev, conf->mddev);
3138                         }
3139                 }
3140                 spin_lock_irq(&sh->stripe_lock);
3141                 /* fail all writes first */
3142                 bi = sh->dev[i].towrite;
3143                 sh->dev[i].towrite = NULL;
3144                 sh->overwrite_disks = 0;
3145                 spin_unlock_irq(&sh->stripe_lock);
3146                 if (bi)
3147                         bitmap_end = 1;
3148
3149                 r5l_stripe_write_finished(sh);
3150
3151                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3152                         wake_up(&conf->wait_for_overlap);
3153
3154                 while (bi && bi->bi_iter.bi_sector <
3155                         sh->dev[i].sector + STRIPE_SECTORS) {
3156                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3157
3158                         bi->bi_error = -EIO;
3159                         if (!raid5_dec_bi_active_stripes(bi)) {
3160                                 md_write_end(conf->mddev);
3161                                 bio_list_add(return_bi, bi);
3162                         }
3163                         bi = nextbi;
3164                 }
3165                 if (bitmap_end)
3166                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3167                                 STRIPE_SECTORS, 0, 0);
3168                 bitmap_end = 0;
3169                 /* and fail all 'written' */
3170                 bi = sh->dev[i].written;
3171                 sh->dev[i].written = NULL;
3172                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3173                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3174                         sh->dev[i].page = sh->dev[i].orig_page;
3175                 }
3176
3177                 if (bi) bitmap_end = 1;
3178                 while (bi && bi->bi_iter.bi_sector <
3179                        sh->dev[i].sector + STRIPE_SECTORS) {
3180                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3181
3182                         bi->bi_error = -EIO;
3183                         if (!raid5_dec_bi_active_stripes(bi)) {
3184                                 md_write_end(conf->mddev);
3185                                 bio_list_add(return_bi, bi);
3186                         }
3187                         bi = bi2;
3188                 }
3189
3190                 /* fail any reads if this device is non-operational and
3191                  * the data has not reached the cache yet.
3192                  */
3193                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3194                     s->failed > conf->max_degraded &&
3195                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3196                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3197                         spin_lock_irq(&sh->stripe_lock);
3198                         bi = sh->dev[i].toread;
3199                         sh->dev[i].toread = NULL;
3200                         spin_unlock_irq(&sh->stripe_lock);
3201                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3202                                 wake_up(&conf->wait_for_overlap);
3203                         if (bi)
3204                                 s->to_read--;
3205                         while (bi && bi->bi_iter.bi_sector <
3206                                sh->dev[i].sector + STRIPE_SECTORS) {
3207                                 struct bio *nextbi =
3208                                         r5_next_bio(bi, sh->dev[i].sector);
3209
3210                                 bi->bi_error = -EIO;
3211                                 if (!raid5_dec_bi_active_stripes(bi))
3212                                         bio_list_add(return_bi, bi);
3213                                 bi = nextbi;
3214                         }
3215                 }
3216                 if (bitmap_end)
3217                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3218                                         STRIPE_SECTORS, 0, 0);
3219                 /* If we were in the middle of a write the parity block might
3220                  * still be locked - so just clear all R5_LOCKED flags
3221                  */
3222                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3223         }
3224         s->to_write = 0;
3225         s->written = 0;
3226
3227         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3228                 if (atomic_dec_and_test(&conf->pending_full_writes))
3229                         md_wakeup_thread(conf->mddev->thread);
3230 }
3231
3232 static void
3233 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3234                    struct stripe_head_state *s)
3235 {
3236         int abort = 0;
3237         int i;
3238
3239         BUG_ON(sh->batch_head);
3240         clear_bit(STRIPE_SYNCING, &sh->state);
3241         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3242                 wake_up(&conf->wait_for_overlap);
3243         s->syncing = 0;
3244         s->replacing = 0;
3245         /* There is nothing more to do for sync/check/repair.
3246          * Don't even need to abort as that is handled elsewhere
3247          * if needed, and not always wanted e.g. if there is a known
3248          * bad block here.
3249          * For recover/replace we need to record a bad block on all
3250          * non-sync devices, or abort the recovery
3251          */
3252         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3253                 /* During recovery devices cannot be removed, so
3254                  * locking and refcounting of rdevs is not needed
3255                  */
3256                 rcu_read_lock();
3257                 for (i = 0; i < conf->raid_disks; i++) {
3258                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3259                         if (rdev
3260                             && !test_bit(Faulty, &rdev->flags)
3261                             && !test_bit(In_sync, &rdev->flags)
3262                             && !rdev_set_badblocks(rdev, sh->sector,
3263                                                    STRIPE_SECTORS, 0))
3264                                 abort = 1;
3265                         rdev = rcu_dereference(conf->disks[i].replacement);
3266                         if (rdev
3267                             && !test_bit(Faulty, &rdev->flags)
3268                             && !test_bit(In_sync, &rdev->flags)
3269                             && !rdev_set_badblocks(rdev, sh->sector,
3270                                                    STRIPE_SECTORS, 0))
3271                                 abort = 1;
3272                 }
3273                 rcu_read_unlock();
3274                 if (abort)
3275                         conf->recovery_disabled =
3276                                 conf->mddev->recovery_disabled;
3277         }
3278         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3279 }
3280
3281 static int want_replace(struct stripe_head *sh, int disk_idx)
3282 {
3283         struct md_rdev *rdev;
3284         int rv = 0;
3285
3286         rcu_read_lock();
3287         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3288         if (rdev
3289             && !test_bit(Faulty, &rdev->flags)
3290             && !test_bit(In_sync, &rdev->flags)
3291             && (rdev->recovery_offset <= sh->sector
3292                 || rdev->mddev->recovery_cp <= sh->sector))
3293                 rv = 1;
3294         rcu_read_unlock();
3295         return rv;
3296 }
3297
3298 /* fetch_block - checks the given member device to see if its data needs
3299  * to be read or computed to satisfy a request.
3300  *
3301  * Returns 1 when no more member devices need to be checked, otherwise returns
3302  * 0 to tell the loop in handle_stripe_fill to continue
3303  */
3304
3305 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3306                            int disk_idx, int disks)
3307 {
3308         struct r5dev *dev = &sh->dev[disk_idx];
3309         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3310                                   &sh->dev[s->failed_num[1]] };
3311         int i;
3312
3313
3314         if (test_bit(R5_LOCKED, &dev->flags) ||
3315             test_bit(R5_UPTODATE, &dev->flags))
3316                 /* No point reading this as we already have it or have
3317                  * decided to get it.
3318                  */
3319                 return 0;
3320
3321         if (dev->toread ||
3322             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3323                 /* We need this block to directly satisfy a request */
3324                 return 1;
3325
3326         if (s->syncing || s->expanding ||
3327             (s->replacing && want_replace(sh, disk_idx)))
3328                 /* When syncing, or expanding we read everything.
3329                  * When replacing, we need the replaced block.
3330                  */
3331                 return 1;
3332
3333         if ((s->failed >= 1 && fdev[0]->toread) ||
3334             (s->failed >= 2 && fdev[1]->toread))
3335                 /* If we want to read from a failed device, then
3336                  * we need to actually read every other device.
3337                  */
3338                 return 1;
3339
3340         /* Sometimes neither read-modify-write nor reconstruct-write
3341          * cycles can work.  In those cases we read every block we
3342          * can.  Then the parity-update is certain to have enough to
3343          * work with.
3344          * This can only be a problem when we need to write something,
3345          * and some device has failed.  If either of those tests
3346          * fail we need look no further.
3347          */
3348         if (!s->failed || !s->to_write)
3349                 return 0;
3350
3351         if (test_bit(R5_Insync, &dev->flags) &&
3352             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3353                 /* Pre-reads at not permitted until after short delay
3354                  * to gather multiple requests.  However if this
3355                  * device is no Insync, the block could only be be computed
3356                  * and there is no need to delay that.
3357                  */
3358                 return 0;
3359
3360         for (i = 0; i < s->failed && i < 2; i++) {
3361                 if (fdev[i]->towrite &&
3362                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3363                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3364                         /* If we have a partial write to a failed
3365                          * device, then we will need to reconstruct
3366                          * the content of that device, so all other
3367                          * devices must be read.
3368                          */
3369                         return 1;
3370         }
3371
3372         /* If we are forced to do a reconstruct-write, either because
3373          * the current RAID6 implementation only supports that, or
3374          * or because parity cannot be trusted and we are currently
3375          * recovering it, there is extra need to be careful.
3376          * If one of the devices that we would need to read, because
3377          * it is not being overwritten (and maybe not written at all)
3378          * is missing/faulty, then we need to read everything we can.
3379          */
3380         if (sh->raid_conf->level != 6 &&
3381             sh->sector < sh->raid_conf->mddev->recovery_cp)
3382                 /* reconstruct-write isn't being forced */
3383                 return 0;
3384         for (i = 0; i < s->failed && i < 2; i++) {
3385                 if (s->failed_num[i] != sh->pd_idx &&
3386                     s->failed_num[i] != sh->qd_idx &&
3387                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3388                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3389                         return 1;
3390         }
3391
3392         return 0;
3393 }
3394
3395 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3396                        int disk_idx, int disks)
3397 {
3398         struct r5dev *dev = &sh->dev[disk_idx];
3399
3400         /* is the data in this block needed, and can we get it? */
3401         if (need_this_block(sh, s, disk_idx, disks)) {
3402                 /* we would like to get this block, possibly by computing it,
3403                  * otherwise read it if the backing disk is insync
3404                  */
3405                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3406                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3407                 BUG_ON(sh->batch_head);
3408                 if ((s->uptodate == disks - 1) &&
3409                     (s->failed && (disk_idx == s->failed_num[0] ||
3410                                    disk_idx == s->failed_num[1]))) {
3411                         /* have disk failed, and we're requested to fetch it;
3412                          * do compute it
3413                          */
3414                         pr_debug("Computing stripe %llu block %d\n",
3415                                (unsigned long long)sh->sector, disk_idx);
3416                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3417                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3418                         set_bit(R5_Wantcompute, &dev->flags);
3419                         sh->ops.target = disk_idx;
3420                         sh->ops.target2 = -1; /* no 2nd target */
3421                         s->req_compute = 1;
3422                         /* Careful: from this point on 'uptodate' is in the eye
3423                          * of raid_run_ops which services 'compute' operations
3424                          * before writes. R5_Wantcompute flags a block that will
3425                          * be R5_UPTODATE by the time it is needed for a
3426                          * subsequent operation.
3427                          */
3428                         s->uptodate++;
3429                         return 1;
3430                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3431                         /* Computing 2-failure is *very* expensive; only
3432                          * do it if failed >= 2
3433                          */
3434                         int other;
3435                         for (other = disks; other--; ) {
3436                                 if (other == disk_idx)
3437                                         continue;
3438                                 if (!test_bit(R5_UPTODATE,
3439                                       &sh->dev[other].flags))
3440                                         break;
3441                         }
3442                         BUG_ON(other < 0);
3443                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3444                                (unsigned long long)sh->sector,
3445                                disk_idx, other);
3446                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3447                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3448                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3449                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3450                         sh->ops.target = disk_idx;
3451                         sh->ops.target2 = other;
3452                         s->uptodate += 2;
3453                         s->req_compute = 1;
3454                         return 1;
3455                 } else if (test_bit(R5_Insync, &dev->flags)) {
3456                         set_bit(R5_LOCKED, &dev->flags);
3457                         set_bit(R5_Wantread, &dev->flags);
3458                         s->locked++;
3459                         pr_debug("Reading block %d (sync=%d)\n",
3460                                 disk_idx, s->syncing);
3461                 }
3462         }
3463
3464         return 0;
3465 }
3466
3467 /**
3468  * handle_stripe_fill - read or compute data to satisfy pending requests.
3469  */
3470 static void handle_stripe_fill(struct stripe_head *sh,
3471                                struct stripe_head_state *s,
3472                                int disks)
3473 {
3474         int i;
3475
3476         /* look for blocks to read/compute, skip this if a compute
3477          * is already in flight, or if the stripe contents are in the
3478          * midst of changing due to a write
3479          */
3480         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3481             !sh->reconstruct_state)
3482                 for (i = disks; i--; )
3483                         if (fetch_block(sh, s, i, disks))
3484                                 break;
3485         set_bit(STRIPE_HANDLE, &sh->state);
3486 }
3487
3488 static void break_stripe_batch_list(struct stripe_head *head_sh,
3489                                     unsigned long handle_flags);
3490 /* handle_stripe_clean_event
3491  * any written block on an uptodate or failed drive can be returned.
3492  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3493  * never LOCKED, so we don't need to test 'failed' directly.
3494  */
3495 static void handle_stripe_clean_event(struct r5conf *conf,
3496         struct stripe_head *sh, int disks, struct bio_list *return_bi)
3497 {
3498         int i;
3499         struct r5dev *dev;
3500         int discard_pending = 0;
3501         struct stripe_head *head_sh = sh;
3502         bool do_endio = false;
3503
3504         for (i = disks; i--; )
3505                 if (sh->dev[i].written) {
3506                         dev = &sh->dev[i];
3507                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3508                             (test_bit(R5_UPTODATE, &dev->flags) ||
3509                              test_bit(R5_Discard, &dev->flags) ||
3510                              test_bit(R5_SkipCopy, &dev->flags))) {
3511                                 /* We can return any write requests */
3512                                 struct bio *wbi, *wbi2;
3513                                 pr_debug("Return write for disc %d\n", i);
3514                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3515                                         clear_bit(R5_UPTODATE, &dev->flags);
3516                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3517                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3518                                 }
3519                                 do_endio = true;
3520
3521 returnbi:
3522                                 dev->page = dev->orig_page;
3523                                 wbi = dev->written;
3524                                 dev->written = NULL;
3525                                 while (wbi && wbi->bi_iter.bi_sector <
3526                                         dev->sector + STRIPE_SECTORS) {
3527                                         wbi2 = r5_next_bio(wbi, dev->sector);
3528                                         if (!raid5_dec_bi_active_stripes(wbi)) {
3529                                                 md_write_end(conf->mddev);
3530                                                 bio_list_add(return_bi, wbi);
3531                                         }
3532                                         wbi = wbi2;
3533                                 }
3534                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3535                                                 STRIPE_SECTORS,
3536                                          !test_bit(STRIPE_DEGRADED, &sh->state),
3537                                                 0);
3538                                 if (head_sh->batch_head) {
3539                                         sh = list_first_entry(&sh->batch_list,
3540                                                               struct stripe_head,
3541                                                               batch_list);
3542                                         if (sh != head_sh) {
3543                                                 dev = &sh->dev[i];
3544                                                 goto returnbi;
3545                                         }
3546                                 }
3547                                 sh = head_sh;
3548                                 dev = &sh->dev[i];
3549                         } else if (test_bit(R5_Discard, &dev->flags))
3550                                 discard_pending = 1;
3551                 }
3552
3553         r5l_stripe_write_finished(sh);
3554
3555         if (!discard_pending &&
3556             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3557                 int hash;
3558                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3559                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3560                 if (sh->qd_idx >= 0) {
3561                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3562                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3563                 }
3564                 /* now that discard is done we can proceed with any sync */
3565                 clear_bit(STRIPE_DISCARD, &sh->state);
3566                 /*
3567                  * SCSI discard will change some bio fields and the stripe has
3568                  * no updated data, so remove it from hash list and the stripe
3569                  * will be reinitialized
3570                  */
3571 unhash:
3572                 hash = sh->hash_lock_index;
3573                 spin_lock_irq(conf->hash_locks + hash);
3574                 remove_hash(sh);
3575                 spin_unlock_irq(conf->hash_locks + hash);
3576                 if (head_sh->batch_head) {
3577                         sh = list_first_entry(&sh->batch_list,
3578                                               struct stripe_head, batch_list);
3579                         if (sh != head_sh)
3580                                         goto unhash;
3581                 }
3582                 sh = head_sh;
3583
3584                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3585                         set_bit(STRIPE_HANDLE, &sh->state);
3586
3587         }
3588
3589         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3590                 if (atomic_dec_and_test(&conf->pending_full_writes))
3591                         md_wakeup_thread(conf->mddev->thread);
3592
3593         if (head_sh->batch_head && do_endio)
3594                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3595 }
3596
3597 static int handle_stripe_dirtying(struct r5conf *conf,
3598                                   struct stripe_head *sh,
3599                                   struct stripe_head_state *s,
3600                                   int disks)
3601 {
3602         int rmw = 0, rcw = 0, i;
3603         sector_t recovery_cp = conf->mddev->recovery_cp;
3604
3605         /* Check whether resync is now happening or should start.
3606          * If yes, then the array is dirty (after unclean shutdown or
3607          * initial creation), so parity in some stripes might be inconsistent.
3608          * In this case, we need to always do reconstruct-write, to ensure
3609          * that in case of drive failure or read-error correction, we
3610          * generate correct data from the parity.
3611          */
3612         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3613             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3614              s->failed == 0)) {
3615                 /* Calculate the real rcw later - for now make it
3616                  * look like rcw is cheaper
3617                  */
3618                 rcw = 1; rmw = 2;
3619                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3620                          conf->rmw_level, (unsigned long long)recovery_cp,
3621                          (unsigned long long)sh->sector);
3622         } else for (i = disks; i--; ) {
3623                 /* would I have to read this buffer for read_modify_write */
3624                 struct r5dev *dev = &sh->dev[i];
3625                 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx ||
3626                      test_bit(R5_InJournal, &dev->flags)) &&
3627                     !test_bit(R5_LOCKED, &dev->flags) &&
3628                     !((test_bit(R5_UPTODATE, &dev->flags) &&
3629                        (!test_bit(R5_InJournal, &dev->flags) ||
3630                         dev->page != dev->orig_page)) ||
3631                       test_bit(R5_Wantcompute, &dev->flags))) {
3632                         if (test_bit(R5_Insync, &dev->flags))
3633                                 rmw++;
3634                         else
3635                                 rmw += 2*disks;  /* cannot read it */
3636                 }
3637                 /* Would I have to read this buffer for reconstruct_write */
3638                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3639                     i != sh->pd_idx && i != sh->qd_idx &&
3640                     !test_bit(R5_LOCKED, &dev->flags) &&
3641                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3642                       test_bit(R5_InJournal, &dev->flags) ||
3643                       test_bit(R5_Wantcompute, &dev->flags))) {
3644                         if (test_bit(R5_Insync, &dev->flags))
3645                                 rcw++;
3646                         else
3647                                 rcw += 2*disks;
3648                 }
3649         }
3650
3651         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3652                 (unsigned long long)sh->sector, rmw, rcw);
3653         set_bit(STRIPE_HANDLE, &sh->state);
3654         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3655                 /* prefer read-modify-write, but need to get some data */
3656                 if (conf->mddev->queue)
3657                         blk_add_trace_msg(conf->mddev->queue,
3658                                           "raid5 rmw %llu %d",
3659                                           (unsigned long long)sh->sector, rmw);
3660                 for (i = disks; i--; ) {
3661                         struct r5dev *dev = &sh->dev[i];
3662                         if (test_bit(R5_InJournal, &dev->flags) &&
3663                             dev->page == dev->orig_page &&
3664                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3665                                 /* alloc page for prexor */
3666                                 struct page *p = alloc_page(GFP_NOIO);
3667
3668                                 if (p) {
3669                                         dev->orig_page = p;
3670                                         continue;
3671                                 }
3672
3673                                 /*
3674                                  * alloc_page() failed, try use
3675                                  * disk_info->extra_page
3676                                  */
3677                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3678                                                       &conf->cache_state)) {
3679                                         r5c_use_extra_page(sh);
3680                                         break;
3681                                 }
3682
3683                                 /* extra_page in use, add to delayed_list */
3684                                 set_bit(STRIPE_DELAYED, &sh->state);
3685                                 s->waiting_extra_page = 1;
3686                                 return -EAGAIN;
3687                         }
3688                 }
3689
3690                 for (i = disks; i--; ) {
3691                         struct r5dev *dev = &sh->dev[i];
3692                         if ((dev->towrite ||
3693                              i == sh->pd_idx || i == sh->qd_idx ||
3694                              test_bit(R5_InJournal, &dev->flags)) &&
3695                             !test_bit(R5_LOCKED, &dev->flags) &&
3696                             !((test_bit(R5_UPTODATE, &dev->flags) &&
3697                                (!test_bit(R5_InJournal, &dev->flags) ||
3698                                 dev->page != dev->orig_page)) ||
3699                               test_bit(R5_Wantcompute, &dev->flags)) &&
3700                             test_bit(R5_Insync, &dev->flags)) {
3701                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3702                                              &sh->state)) {
3703                                         pr_debug("Read_old block %d for r-m-w\n",
3704                                                  i);
3705                                         set_bit(R5_LOCKED, &dev->flags);
3706                                         set_bit(R5_Wantread, &dev->flags);
3707                                         s->locked++;
3708                                 } else {
3709                                         set_bit(STRIPE_DELAYED, &sh->state);
3710                                         set_bit(STRIPE_HANDLE, &sh->state);
3711                                 }
3712                         }
3713                 }
3714         }
3715         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3716                 /* want reconstruct write, but need to get some data */
3717                 int qread =0;
3718                 rcw = 0;
3719                 for (i = disks; i--; ) {
3720                         struct r5dev *dev = &sh->dev[i];
3721                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3722                             i != sh->pd_idx && i != sh->qd_idx &&
3723                             !test_bit(R5_LOCKED, &dev->flags) &&
3724                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3725                               test_bit(R5_InJournal, &dev->flags) ||
3726                               test_bit(R5_Wantcompute, &dev->flags))) {
3727                                 rcw++;
3728                                 if (test_bit(R5_Insync, &dev->flags) &&
3729                                     test_bit(STRIPE_PREREAD_ACTIVE,
3730                                              &sh->state)) {
3731                                         pr_debug("Read_old block "
3732                                                 "%d for Reconstruct\n", i);
3733                                         set_bit(R5_LOCKED, &dev->flags);
3734                                         set_bit(R5_Wantread, &dev->flags);
3735                                         s->locked++;
3736                                         qread++;
3737                                 } else {
3738                                         set_bit(STRIPE_DELAYED, &sh->state);
3739                                         set_bit(STRIPE_HANDLE, &sh->state);
3740                                 }
3741                         }
3742                 }
3743                 if (rcw && conf->mddev->queue)
3744                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3745                                           (unsigned long long)sh->sector,
3746                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3747         }
3748
3749         if (rcw > disks && rmw > disks &&
3750             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3751                 set_bit(STRIPE_DELAYED, &sh->state);
3752
3753         /* now if nothing is locked, and if we have enough data,
3754          * we can start a write request
3755          */
3756         /* since handle_stripe can be called at any time we need to handle the
3757          * case where a compute block operation has been submitted and then a
3758          * subsequent call wants to start a write request.  raid_run_ops only
3759          * handles the case where compute block and reconstruct are requested
3760          * simultaneously.  If this is not the case then new writes need to be
3761          * held off until the compute completes.
3762          */
3763         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3764             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3765              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3766                 schedule_reconstruction(sh, s, rcw == 0, 0);
3767         return 0;
3768 }
3769
3770 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3771                                 struct stripe_head_state *s, int disks)
3772 {
3773         struct r5dev *dev = NULL;
3774
3775         BUG_ON(sh->batch_head);
3776         set_bit(STRIPE_HANDLE, &sh->state);
3777
3778         switch (sh->check_state) {
3779         case check_state_idle:
3780                 /* start a new check operation if there are no failures */
3781                 if (s->failed == 0) {
3782                         BUG_ON(s->uptodate != disks);
3783                         sh->check_state = check_state_run;
3784                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3785                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3786                         s->uptodate--;
3787                         break;
3788                 }
3789                 dev = &sh->dev[s->failed_num[0]];
3790                 /* fall through */
3791         case check_state_compute_result:
3792                 sh->check_state = check_state_idle;
3793                 if (!dev)
3794                         dev = &sh->dev[sh->pd_idx];
3795
3796                 /* check that a write has not made the stripe insync */
3797                 if (test_bit(STRIPE_INSYNC, &sh->state))
3798                         break;
3799
3800                 /* either failed parity check, or recovery is happening */
3801                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3802                 BUG_ON(s->uptodate != disks);
3803
3804                 set_bit(R5_LOCKED, &dev->flags);
3805                 s->locked++;
3806                 set_bit(R5_Wantwrite, &dev->flags);
3807
3808                 clear_bit(STRIPE_DEGRADED, &sh->state);
3809                 set_bit(STRIPE_INSYNC, &sh->state);
3810                 break;
3811         case check_state_run:
3812                 break; /* we will be called again upon completion */
3813         case check_state_check_result:
3814                 sh->check_state = check_state_idle;
3815
3816                 /* if a failure occurred during the check operation, leave
3817                  * STRIPE_INSYNC not set and let the stripe be handled again
3818                  */
3819                 if (s->failed)
3820                         break;
3821
3822                 /* handle a successful check operation, if parity is correct
3823                  * we are done.  Otherwise update the mismatch count and repair
3824                  * parity if !MD_RECOVERY_CHECK
3825                  */
3826                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3827                         /* parity is correct (on disc,
3828                          * not in buffer any more)
3829                          */
3830                         set_bit(STRIPE_INSYNC, &sh->state);
3831                 else {
3832                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3833                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3834                                 /* don't try to repair!! */
3835                                 set_bit(STRIPE_INSYNC, &sh->state);
3836                         else {
3837                                 sh->check_state = check_state_compute_run;
3838                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3839                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3840                                 set_bit(R5_Wantcompute,
3841                                         &sh->dev[sh->pd_idx].flags);
3842                                 sh->ops.target = sh->pd_idx;
3843                                 sh->ops.target2 = -1;
3844                                 s->uptodate++;
3845                         }
3846                 }
3847                 break;
3848         case check_state_compute_run:
3849                 break;
3850         default:
3851                 pr_err("%s: unknown check_state: %d sector: %llu\n",
3852                        __func__, sh->check_state,
3853                        (unsigned long long) sh->sector);
3854                 BUG();
3855         }
3856 }
3857
3858 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3859                                   struct stripe_head_state *s,
3860                                   int disks)
3861 {
3862         int pd_idx = sh->pd_idx;
3863         int qd_idx = sh->qd_idx;
3864         struct r5dev *dev;
3865
3866         BUG_ON(sh->batch_head);
3867         set_bit(STRIPE_HANDLE, &sh->state);
3868
3869         BUG_ON(s->failed > 2);
3870
3871         /* Want to check and possibly repair P and Q.
3872          * However there could be one 'failed' device, in which
3873          * case we can only check one of them, possibly using the
3874          * other to generate missing data
3875          */
3876
3877         switch (sh->check_state) {
3878         case check_state_idle:
3879                 /* start a new check operation if there are < 2 failures */
3880                 if (s->failed == s->q_failed) {
3881                         /* The only possible failed device holds Q, so it
3882                          * makes sense to check P (If anything else were failed,
3883                          * we would have used P to recreate it).
3884                          */
3885                         sh->check_state = check_state_run;
3886                 }
3887                 if (!s->q_failed && s->failed < 2) {
3888                         /* Q is not failed, and we didn't use it to generate
3889                          * anything, so it makes sense to check it
3890                          */
3891                         if (sh->check_state == check_state_run)
3892                                 sh->check_state = check_state_run_pq;
3893                         else
3894                                 sh->check_state = check_state_run_q;
3895                 }
3896
3897                 /* discard potentially stale zero_sum_result */
3898                 sh->ops.zero_sum_result = 0;
3899
3900                 if (sh->check_state == check_state_run) {
3901                         /* async_xor_zero_sum destroys the contents of P */
3902                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3903                         s->uptodate--;
3904                 }
3905                 if (sh->check_state >= check_state_run &&
3906                     sh->check_state <= check_state_run_pq) {
3907                         /* async_syndrome_zero_sum preserves P and Q, so
3908                          * no need to mark them !uptodate here
3909                          */
3910                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3911                         break;
3912                 }
3913
3914                 /* we have 2-disk failure */
3915                 BUG_ON(s->failed != 2);
3916                 /* fall through */
3917         case check_state_compute_result:
3918                 sh->check_state = check_state_idle;
3919
3920                 /* check that a write has not made the stripe insync */
3921                 if (test_bit(STRIPE_INSYNC, &sh->state))
3922                         break;
3923
3924                 /* now write out any block on a failed drive,
3925                  * or P or Q if they were recomputed
3926                  */
3927                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3928                 if (s->failed == 2) {
3929                         dev = &sh->dev[s->failed_num[1]];
3930                         s->locked++;
3931                         set_bit(R5_LOCKED, &dev->flags);
3932                         set_bit(R5_Wantwrite, &dev->flags);
3933                 }
3934                 if (s->failed >= 1) {
3935                         dev = &sh->dev[s->failed_num[0]];
3936                         s->locked++;
3937                         set_bit(R5_LOCKED, &dev->flags);
3938                         set_bit(R5_Wantwrite, &dev->flags);
3939                 }
3940                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3941                         dev = &sh->dev[pd_idx];
3942                         s->locked++;
3943                         set_bit(R5_LOCKED, &dev->flags);
3944                         set_bit(R5_Wantwrite, &dev->flags);
3945                 }
3946                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3947                         dev = &sh->dev[qd_idx];
3948                         s->locked++;
3949                         set_bit(R5_LOCKED, &dev->flags);
3950                         set_bit(R5_Wantwrite, &dev->flags);
3951                 }
3952                 clear_bit(STRIPE_DEGRADED, &sh->state);
3953
3954                 set_bit(STRIPE_INSYNC, &sh->state);
3955                 break;
3956         case check_state_run:
3957         case check_state_run_q:
3958         case check_state_run_pq:
3959                 break; /* we will be called again upon completion */
3960         case check_state_check_result:
3961                 sh->check_state = check_state_idle;
3962
3963                 /* handle a successful check operation, if parity is correct
3964                  * we are done.  Otherwise update the mismatch count and repair
3965                  * parity if !MD_RECOVERY_CHECK
3966                  */
3967                 if (sh->ops.zero_sum_result == 0) {
3968                         /* both parities are correct */
3969                         if (!s->failed)
3970                                 set_bit(STRIPE_INSYNC, &sh->state);
3971                         else {
3972                                 /* in contrast to the raid5 case we can validate
3973                                  * parity, but still have a failure to write
3974                                  * back
3975                                  */
3976                                 sh->check_state = check_state_compute_result;
3977                                 /* Returning at this point means that we may go
3978                                  * off and bring p and/or q uptodate again so
3979                                  * we make sure to check zero_sum_result again
3980                                  * to verify if p or q need writeback
3981                                  */
3982                         }
3983                 } else {
3984                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3985                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3986                                 /* don't try to repair!! */
3987                                 set_bit(STRIPE_INSYNC, &sh->state);
3988                         else {
3989                                 int *target = &sh->ops.target;
3990
3991                                 sh->ops.target = -1;
3992                                 sh->ops.target2 = -1;
3993                                 sh->check_state = check_state_compute_run;
3994                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3995                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3996                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3997                                         set_bit(R5_Wantcompute,
3998                                                 &sh->dev[pd_idx].flags);
3999                                         *target = pd_idx;
4000                                         target = &sh->ops.target2;
4001                                         s->uptodate++;
4002                                 }
4003                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4004                                         set_bit(R5_Wantcompute,
4005                                                 &sh->dev[qd_idx].flags);
4006                                         *target = qd_idx;
4007                                         s->uptodate++;
4008                                 }
4009                         }
4010                 }
4011                 break;
4012         case check_state_compute_run:
4013                 break;
4014         default:
4015                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4016                         __func__, sh->check_state,
4017                         (unsigned long long) sh->sector);
4018                 BUG();
4019         }
4020 }
4021
4022 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4023 {
4024         int i;
4025
4026         /* We have read all the blocks in this stripe and now we need to
4027          * copy some of them into a target stripe for expand.
4028          */
4029         struct dma_async_tx_descriptor *tx = NULL;
4030         BUG_ON(sh->batch_head);
4031         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4032         for (i = 0; i < sh->disks; i++)
4033                 if (i != sh->pd_idx && i != sh->qd_idx) {
4034                         int dd_idx, j;
4035                         struct stripe_head *sh2;
4036                         struct async_submit_ctl submit;
4037
4038                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4039                         sector_t s = raid5_compute_sector(conf, bn, 0,
4040                                                           &dd_idx, NULL);
4041                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4042                         if (sh2 == NULL)
4043                                 /* so far only the early blocks of this stripe
4044                                  * have been requested.  When later blocks
4045                                  * get requested, we will try again
4046                                  */
4047                                 continue;
4048                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4049                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4050                                 /* must have already done this block */
4051                                 raid5_release_stripe(sh2);
4052                                 continue;
4053                         }
4054
4055                         /* place all the copies on one channel */
4056                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4057                         tx = async_memcpy(sh2->dev[dd_idx].page,
4058                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
4059                                           &submit);
4060
4061                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4062                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4063                         for (j = 0; j < conf->raid_disks; j++)
4064                                 if (j != sh2->pd_idx &&
4065                                     j != sh2->qd_idx &&
4066                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4067                                         break;
4068                         if (j == conf->raid_disks) {
4069                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4070                                 set_bit(STRIPE_HANDLE, &sh2->state);
4071                         }
4072                         raid5_release_stripe(sh2);
4073
4074                 }
4075         /* done submitting copies, wait for them to complete */
4076         async_tx_quiesce(&tx);
4077 }
4078
4079 /*
4080  * handle_stripe - do things to a stripe.
4081  *
4082  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4083  * state of various bits to see what needs to be done.
4084  * Possible results:
4085  *    return some read requests which now have data
4086  *    return some write requests which are safely on storage
4087  *    schedule a read on some buffers
4088  *    schedule a write of some buffers
4089  *    return confirmation of parity correctness
4090  *
4091  */
4092
4093 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4094 {
4095         struct r5conf *conf = sh->raid_conf;
4096         int disks = sh->disks;
4097         struct r5dev *dev;
4098         int i;
4099         int do_recovery = 0;
4100
4101         memset(s, 0, sizeof(*s));
4102
4103         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4104         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4105         s->failed_num[0] = -1;
4106         s->failed_num[1] = -1;
4107         s->log_failed = r5l_log_disk_error(conf);
4108
4109         /* Now to look around and see what can be done */
4110         rcu_read_lock();
4111         for (i=disks; i--; ) {
4112                 struct md_rdev *rdev;
4113                 sector_t first_bad;
4114                 int bad_sectors;
4115                 int is_bad = 0;
4116
4117                 dev = &sh->dev[i];
4118
4119                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4120                          i, dev->flags,
4121                          dev->toread, dev->towrite, dev->written);
4122                 /* maybe we can reply to a read
4123                  *
4124                  * new wantfill requests are only permitted while
4125                  * ops_complete_biofill is guaranteed to be inactive
4126                  */
4127                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4128                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4129                         set_bit(R5_Wantfill, &dev->flags);
4130
4131                 /* now count some things */
4132                 if (test_bit(R5_LOCKED, &dev->flags))
4133                         s->locked++;
4134                 if (test_bit(R5_UPTODATE, &dev->flags))
4135                         s->uptodate++;
4136                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4137                         s->compute++;
4138                         BUG_ON(s->compute > 2);
4139                 }
4140
4141                 if (test_bit(R5_Wantfill, &dev->flags))
4142                         s->to_fill++;
4143                 else if (dev->toread)
4144                         s->to_read++;
4145                 if (dev->towrite) {
4146                         s->to_write++;
4147                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4148                                 s->non_overwrite++;
4149                 }
4150                 if (dev->written)
4151                         s->written++;
4152                 /* Prefer to use the replacement for reads, but only
4153                  * if it is recovered enough and has no bad blocks.
4154                  */
4155                 rdev = rcu_dereference(conf->disks[i].replacement);
4156                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4157                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4158                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4159                                  &first_bad, &bad_sectors))
4160                         set_bit(R5_ReadRepl, &dev->flags);
4161                 else {
4162                         if (rdev && !test_bit(Faulty, &rdev->flags))
4163                                 set_bit(R5_NeedReplace, &dev->flags);
4164                         else
4165                                 clear_bit(R5_NeedReplace, &dev->flags);
4166                         rdev = rcu_dereference(conf->disks[i].rdev);
4167                         clear_bit(R5_ReadRepl, &dev->flags);
4168                 }
4169                 if (rdev && test_bit(Faulty, &rdev->flags))
4170                         rdev = NULL;
4171                 if (rdev) {
4172                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4173                                              &first_bad, &bad_sectors);
4174                         if (s->blocked_rdev == NULL
4175                             && (test_bit(Blocked, &rdev->flags)
4176                                 || is_bad < 0)) {
4177                                 if (is_bad < 0)
4178                                         set_bit(BlockedBadBlocks,
4179                                                 &rdev->flags);
4180                                 s->blocked_rdev = rdev;
4181                                 atomic_inc(&rdev->nr_pending);
4182                         }
4183                 }
4184                 clear_bit(R5_Insync, &dev->flags);
4185                 if (!rdev)
4186                         /* Not in-sync */;
4187                 else if (is_bad) {
4188                         /* also not in-sync */
4189                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4190                             test_bit(R5_UPTODATE, &dev->flags)) {
4191                                 /* treat as in-sync, but with a read error
4192                                  * which we can now try to correct
4193                                  */
4194                                 set_bit(R5_Insync, &dev->flags);
4195                                 set_bit(R5_ReadError, &dev->flags);
4196                         }
4197                 } else if (test_bit(In_sync, &rdev->flags))
4198                         set_bit(R5_Insync, &dev->flags);
4199                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4200                         /* in sync if before recovery_offset */
4201                         set_bit(R5_Insync, &dev->flags);
4202                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4203                          test_bit(R5_Expanded, &dev->flags))
4204                         /* If we've reshaped into here, we assume it is Insync.
4205                          * We will shortly update recovery_offset to make
4206                          * it official.
4207                          */
4208                         set_bit(R5_Insync, &dev->flags);
4209
4210                 if (test_bit(R5_WriteError, &dev->flags)) {
4211                         /* This flag does not apply to '.replacement'
4212                          * only to .rdev, so make sure to check that*/
4213                         struct md_rdev *rdev2 = rcu_dereference(
4214                                 conf->disks[i].rdev);
4215                         if (rdev2 == rdev)
4216                                 clear_bit(R5_Insync, &dev->flags);
4217                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4218                                 s->handle_bad_blocks = 1;
4219                                 atomic_inc(&rdev2->nr_pending);
4220                         } else
4221                                 clear_bit(R5_WriteError, &dev->flags);
4222                 }
4223                 if (test_bit(R5_MadeGood, &dev->flags)) {
4224                         /* This flag does not apply to '.replacement'
4225                          * only to .rdev, so make sure to check that*/
4226                         struct md_rdev *rdev2 = rcu_dereference(
4227                                 conf->disks[i].rdev);
4228                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4229                                 s->handle_bad_blocks = 1;
4230                                 atomic_inc(&rdev2->nr_pending);
4231                         } else
4232                                 clear_bit(R5_MadeGood, &dev->flags);
4233                 }
4234                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4235                         struct md_rdev *rdev2 = rcu_dereference(
4236                                 conf->disks[i].replacement);
4237                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4238                                 s->handle_bad_blocks = 1;
4239                                 atomic_inc(&rdev2->nr_pending);
4240                         } else
4241                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4242                 }
4243                 if (!test_bit(R5_Insync, &dev->flags)) {
4244                         /* The ReadError flag will just be confusing now */
4245                         clear_bit(R5_ReadError, &dev->flags);
4246                         clear_bit(R5_ReWrite, &dev->flags);
4247                 }
4248                 if (test_bit(R5_ReadError, &dev->flags))
4249                         clear_bit(R5_Insync, &dev->flags);
4250                 if (!test_bit(R5_Insync, &dev->flags)) {
4251                         if (s->failed < 2)
4252                                 s->failed_num[s->failed] = i;
4253                         s->failed++;
4254                         if (rdev && !test_bit(Faulty, &rdev->flags))
4255                                 do_recovery = 1;
4256                 }
4257
4258                 if (test_bit(R5_InJournal, &dev->flags))
4259                         s->injournal++;
4260                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4261                         s->just_cached++;
4262         }
4263         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4264                 /* If there is a failed device being replaced,
4265                  *     we must be recovering.
4266                  * else if we are after recovery_cp, we must be syncing
4267                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4268                  * else we can only be replacing
4269                  * sync and recovery both need to read all devices, and so
4270                  * use the same flag.
4271                  */
4272                 if (do_recovery ||
4273                     sh->sector >= conf->mddev->recovery_cp ||
4274                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4275                         s->syncing = 1;
4276                 else
4277                         s->replacing = 1;
4278         }
4279         rcu_read_unlock();
4280 }
4281
4282 static int clear_batch_ready(struct stripe_head *sh)
4283 {
4284         /* Return '1' if this is a member of batch, or
4285          * '0' if it is a lone stripe or a head which can now be
4286          * handled.
4287          */
4288         struct stripe_head *tmp;
4289         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4290                 return (sh->batch_head && sh->batch_head != sh);
4291         spin_lock(&sh->stripe_lock);
4292         if (!sh->batch_head) {
4293                 spin_unlock(&sh->stripe_lock);
4294                 return 0;
4295         }
4296
4297         /*
4298          * this stripe could be added to a batch list before we check
4299          * BATCH_READY, skips it
4300          */
4301         if (sh->batch_head != sh) {
4302                 spin_unlock(&sh->stripe_lock);
4303                 return 1;
4304         }
4305         spin_lock(&sh->batch_lock);
4306         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4307                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4308         spin_unlock(&sh->batch_lock);
4309         spin_unlock(&sh->stripe_lock);
4310
4311         /*
4312          * BATCH_READY is cleared, no new stripes can be added.
4313          * batch_list can be accessed without lock
4314          */
4315         return 0;
4316 }
4317
4318 static void break_stripe_batch_list(struct stripe_head *head_sh,
4319                                     unsigned long handle_flags)
4320 {
4321         struct stripe_head *sh, *next;
4322         int i;
4323         int do_wakeup = 0;
4324
4325         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4326
4327                 list_del_init(&sh->batch_list);
4328
4329                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4330                                           (1 << STRIPE_SYNCING) |
4331                                           (1 << STRIPE_REPLACED) |
4332                                           (1 << STRIPE_DELAYED) |
4333                                           (1 << STRIPE_BIT_DELAY) |
4334                                           (1 << STRIPE_FULL_WRITE) |
4335                                           (1 << STRIPE_BIOFILL_RUN) |
4336                                           (1 << STRIPE_COMPUTE_RUN)  |
4337                                           (1 << STRIPE_OPS_REQ_PENDING) |
4338                                           (1 << STRIPE_DISCARD) |
4339                                           (1 << STRIPE_BATCH_READY) |
4340                                           (1 << STRIPE_BATCH_ERR) |
4341                                           (1 << STRIPE_BITMAP_PENDING)),
4342                         "stripe state: %lx\n", sh->state);
4343                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4344                                               (1 << STRIPE_REPLACED)),
4345                         "head stripe state: %lx\n", head_sh->state);
4346
4347                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4348                                             (1 << STRIPE_PREREAD_ACTIVE) |
4349                                             (1 << STRIPE_DEGRADED)),
4350                               head_sh->state & (1 << STRIPE_INSYNC));
4351
4352                 sh->check_state = head_sh->check_state;
4353                 sh->reconstruct_state = head_sh->reconstruct_state;
4354                 for (i = 0; i < sh->disks; i++) {
4355                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4356                                 do_wakeup = 1;
4357                         sh->dev[i].flags = head_sh->dev[i].flags &
4358                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4359                 }
4360                 spin_lock_irq(&sh->stripe_lock);
4361                 sh->batch_head = NULL;
4362                 spin_unlock_irq(&sh->stripe_lock);
4363                 if (handle_flags == 0 ||
4364                     sh->state & handle_flags)
4365                         set_bit(STRIPE_HANDLE, &sh->state);
4366                 raid5_release_stripe(sh);
4367         }
4368         spin_lock_irq(&head_sh->stripe_lock);
4369         head_sh->batch_head = NULL;
4370         spin_unlock_irq(&head_sh->stripe_lock);
4371         for (i = 0; i < head_sh->disks; i++)
4372                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4373                         do_wakeup = 1;
4374         if (head_sh->state & handle_flags)
4375                 set_bit(STRIPE_HANDLE, &head_sh->state);
4376
4377         if (do_wakeup)
4378                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4379 }
4380
4381 static void handle_stripe(struct stripe_head *sh)
4382 {
4383         struct stripe_head_state s;
4384         struct r5conf *conf = sh->raid_conf;
4385         int i;
4386         int prexor;
4387         int disks = sh->disks;
4388         struct r5dev *pdev, *qdev;
4389
4390         clear_bit(STRIPE_HANDLE, &sh->state);
4391         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4392                 /* already being handled, ensure it gets handled
4393                  * again when current action finishes */
4394                 set_bit(STRIPE_HANDLE, &sh->state);
4395                 return;
4396         }
4397
4398         if (clear_batch_ready(sh) ) {
4399                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4400                 return;
4401         }
4402
4403         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4404                 break_stripe_batch_list(sh, 0);
4405
4406         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4407                 spin_lock(&sh->stripe_lock);
4408                 /* Cannot process 'sync' concurrently with 'discard' */
4409                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4410                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4411                         set_bit(STRIPE_SYNCING, &sh->state);
4412                         clear_bit(STRIPE_INSYNC, &sh->state);
4413                         clear_bit(STRIPE_REPLACED, &sh->state);
4414                 }
4415                 spin_unlock(&sh->stripe_lock);
4416         }
4417         clear_bit(STRIPE_DELAYED, &sh->state);
4418
4419         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4420                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4421                (unsigned long long)sh->sector, sh->state,
4422                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4423                sh->check_state, sh->reconstruct_state);
4424
4425         analyse_stripe(sh, &s);
4426
4427         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4428                 goto finish;
4429
4430         if (s.handle_bad_blocks) {
4431                 set_bit(STRIPE_HANDLE, &sh->state);
4432                 goto finish;
4433         }
4434
4435         if (unlikely(s.blocked_rdev)) {
4436                 if (s.syncing || s.expanding || s.expanded ||
4437                     s.replacing || s.to_write || s.written) {
4438                         set_bit(STRIPE_HANDLE, &sh->state);
4439                         goto finish;
4440                 }
4441                 /* There is nothing for the blocked_rdev to block */
4442                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4443                 s.blocked_rdev = NULL;
4444         }
4445
4446         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4447                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4448                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4449         }
4450
4451         pr_debug("locked=%d uptodate=%d to_read=%d"
4452                " to_write=%d failed=%d failed_num=%d,%d\n",
4453                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4454                s.failed_num[0], s.failed_num[1]);
4455         /* check if the array has lost more than max_degraded devices and,
4456          * if so, some requests might need to be failed.
4457          */
4458         if (s.failed > conf->max_degraded || s.log_failed) {
4459                 sh->check_state = 0;
4460                 sh->reconstruct_state = 0;
4461                 break_stripe_batch_list(sh, 0);
4462                 if (s.to_read+s.to_write+s.written)
4463                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4464                 if (s.syncing + s.replacing)
4465                         handle_failed_sync(conf, sh, &s);
4466         }
4467
4468         /* Now we check to see if any write operations have recently
4469          * completed
4470          */
4471         prexor = 0;
4472         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4473                 prexor = 1;
4474         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4475             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4476                 sh->reconstruct_state = reconstruct_state_idle;
4477
4478                 /* All the 'written' buffers and the parity block are ready to
4479                  * be written back to disk
4480                  */
4481                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4482                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4483                 BUG_ON(sh->qd_idx >= 0 &&
4484                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4485                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4486                 for (i = disks; i--; ) {
4487                         struct r5dev *dev = &sh->dev[i];
4488                         if (test_bit(R5_LOCKED, &dev->flags) &&
4489                                 (i == sh->pd_idx || i == sh->qd_idx ||
4490                                  dev->written || test_bit(R5_InJournal,
4491                                                           &dev->flags))) {
4492                                 pr_debug("Writing block %d\n", i);
4493                                 set_bit(R5_Wantwrite, &dev->flags);
4494                                 if (prexor)
4495                                         continue;
4496                                 if (s.failed > 1)
4497                                         continue;
4498                                 if (!test_bit(R5_Insync, &dev->flags) ||
4499                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4500                                      s.failed == 0))
4501                                         set_bit(STRIPE_INSYNC, &sh->state);
4502                         }
4503                 }
4504                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4505                         s.dec_preread_active = 1;
4506         }
4507
4508         /*
4509          * might be able to return some write requests if the parity blocks
4510          * are safe, or on a failed drive
4511          */
4512         pdev = &sh->dev[sh->pd_idx];
4513         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4514                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4515         qdev = &sh->dev[sh->qd_idx];
4516         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4517                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4518                 || conf->level < 6;
4519
4520         if (s.written &&
4521             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4522                              && !test_bit(R5_LOCKED, &pdev->flags)
4523                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4524                                  test_bit(R5_Discard, &pdev->flags))))) &&
4525             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4526                              && !test_bit(R5_LOCKED, &qdev->flags)
4527                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4528                                  test_bit(R5_Discard, &qdev->flags))))))
4529                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4530
4531         if (s.just_cached)
4532                 r5c_handle_cached_data_endio(conf, sh, disks, &s.return_bi);
4533         r5l_stripe_write_finished(sh);
4534
4535         /* Now we might consider reading some blocks, either to check/generate
4536          * parity, or to satisfy requests
4537          * or to load a block that is being partially written.
4538          */
4539         if (s.to_read || s.non_overwrite
4540             || (conf->level == 6 && s.to_write && s.failed)
4541             || (s.syncing && (s.uptodate + s.compute < disks))
4542             || s.replacing
4543             || s.expanding)
4544                 handle_stripe_fill(sh, &s, disks);
4545
4546         /*
4547          * When the stripe finishes full journal write cycle (write to journal
4548          * and raid disk), this is the clean up procedure so it is ready for
4549          * next operation.
4550          */
4551         r5c_finish_stripe_write_out(conf, sh, &s);
4552
4553         /*
4554          * Now to consider new write requests, cache write back and what else,
4555          * if anything should be read.  We do not handle new writes when:
4556          * 1/ A 'write' operation (copy+xor) is already in flight.
4557          * 2/ A 'check' operation is in flight, as it may clobber the parity
4558          *    block.
4559          * 3/ A r5c cache log write is in flight.
4560          */
4561
4562         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4563                 if (!r5c_is_writeback(conf->log)) {
4564                         if (s.to_write)
4565                                 handle_stripe_dirtying(conf, sh, &s, disks);
4566                 } else { /* write back cache */
4567                         int ret = 0;
4568
4569                         /* First, try handle writes in caching phase */
4570                         if (s.to_write)
4571                                 ret = r5c_try_caching_write(conf, sh, &s,
4572                                                             disks);
4573                         /*
4574                          * If caching phase failed: ret == -EAGAIN
4575                          *    OR
4576                          * stripe under reclaim: !caching && injournal
4577                          *
4578                          * fall back to handle_stripe_dirtying()
4579                          */
4580                         if (ret == -EAGAIN ||
4581                             /* stripe under reclaim: !caching && injournal */
4582                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4583                              s.injournal > 0)) {
4584                                 ret = handle_stripe_dirtying(conf, sh, &s,
4585                                                              disks);
4586                                 if (ret == -EAGAIN)
4587                                         goto finish;
4588                         }
4589                 }
4590         }
4591
4592         /* maybe we need to check and possibly fix the parity for this stripe
4593          * Any reads will already have been scheduled, so we just see if enough
4594          * data is available.  The parity check is held off while parity
4595          * dependent operations are in flight.
4596          */
4597         if (sh->check_state ||
4598             (s.syncing && s.locked == 0 &&
4599              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4600              !test_bit(STRIPE_INSYNC, &sh->state))) {
4601                 if (conf->level == 6)
4602                         handle_parity_checks6(conf, sh, &s, disks);
4603                 else
4604                         handle_parity_checks5(conf, sh, &s, disks);
4605         }
4606
4607         if ((s.replacing || s.syncing) && s.locked == 0
4608             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4609             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4610                 /* Write out to replacement devices where possible */
4611                 for (i = 0; i < conf->raid_disks; i++)
4612                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4613                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4614                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4615                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4616                                 s.locked++;
4617                         }
4618                 if (s.replacing)
4619                         set_bit(STRIPE_INSYNC, &sh->state);
4620                 set_bit(STRIPE_REPLACED, &sh->state);
4621         }
4622         if ((s.syncing || s.replacing) && s.locked == 0 &&
4623             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4624             test_bit(STRIPE_INSYNC, &sh->state)) {
4625                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4626                 clear_bit(STRIPE_SYNCING, &sh->state);
4627                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4628                         wake_up(&conf->wait_for_overlap);
4629         }
4630
4631         /* If the failed drives are just a ReadError, then we might need
4632          * to progress the repair/check process
4633          */
4634         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4635                 for (i = 0; i < s.failed; i++) {
4636                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4637                         if (test_bit(R5_ReadError, &dev->flags)
4638                             && !test_bit(R5_LOCKED, &dev->flags)
4639                             && test_bit(R5_UPTODATE, &dev->flags)
4640                                 ) {
4641                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4642                                         set_bit(R5_Wantwrite, &dev->flags);
4643                                         set_bit(R5_ReWrite, &dev->flags);
4644                                         set_bit(R5_LOCKED, &dev->flags);
4645                                         s.locked++;
4646                                 } else {
4647                                         /* let's read it back */
4648                                         set_bit(R5_Wantread, &dev->flags);
4649                                         set_bit(R5_LOCKED, &dev->flags);
4650                                         s.locked++;
4651                                 }
4652                         }
4653                 }
4654
4655         /* Finish reconstruct operations initiated by the expansion process */
4656         if (sh->reconstruct_state == reconstruct_state_result) {
4657                 struct stripe_head *sh_src
4658                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4659                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4660                         /* sh cannot be written until sh_src has been read.
4661                          * so arrange for sh to be delayed a little
4662                          */
4663                         set_bit(STRIPE_DELAYED, &sh->state);
4664                         set_bit(STRIPE_HANDLE, &sh->state);
4665                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4666                                               &sh_src->state))
4667                                 atomic_inc(&conf->preread_active_stripes);
4668                         raid5_release_stripe(sh_src);
4669                         goto finish;
4670                 }
4671                 if (sh_src)
4672                         raid5_release_stripe(sh_src);
4673
4674                 sh->reconstruct_state = reconstruct_state_idle;
4675                 clear_bit(STRIPE_EXPANDING, &sh->state);
4676                 for (i = conf->raid_disks; i--; ) {
4677                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4678                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4679                         s.locked++;
4680                 }
4681         }
4682
4683         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4684             !sh->reconstruct_state) {
4685                 /* Need to write out all blocks after computing parity */
4686                 sh->disks = conf->raid_disks;
4687                 stripe_set_idx(sh->sector, conf, 0, sh);
4688                 schedule_reconstruction(sh, &s, 1, 1);
4689         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4690                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4691                 atomic_dec(&conf->reshape_stripes);
4692                 wake_up(&conf->wait_for_overlap);
4693                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4694         }
4695
4696         if (s.expanding && s.locked == 0 &&
4697             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4698                 handle_stripe_expansion(conf, sh);
4699
4700 finish:
4701         /* wait for this device to become unblocked */
4702         if (unlikely(s.blocked_rdev)) {
4703                 if (conf->mddev->external)
4704                         md_wait_for_blocked_rdev(s.blocked_rdev,
4705                                                  conf->mddev);
4706                 else
4707                         /* Internal metadata will immediately
4708                          * be written by raid5d, so we don't
4709                          * need to wait here.
4710                          */
4711                         rdev_dec_pending(s.blocked_rdev,
4712                                          conf->mddev);
4713         }
4714
4715         if (s.handle_bad_blocks)
4716                 for (i = disks; i--; ) {
4717                         struct md_rdev *rdev;
4718                         struct r5dev *dev = &sh->dev[i];
4719                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4720                                 /* We own a safe reference to the rdev */
4721                                 rdev = conf->disks[i].rdev;
4722                                 if (!rdev_set_badblocks(rdev, sh->sector,
4723                                                         STRIPE_SECTORS, 0))
4724                                         md_error(conf->mddev, rdev);
4725                                 rdev_dec_pending(rdev, conf->mddev);
4726                         }
4727                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4728                                 rdev = conf->disks[i].rdev;
4729                                 rdev_clear_badblocks(rdev, sh->sector,
4730                                                      STRIPE_SECTORS, 0);
4731                                 rdev_dec_pending(rdev, conf->mddev);
4732                         }
4733                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4734                                 rdev = conf->disks[i].replacement;
4735                                 if (!rdev)
4736                                         /* rdev have been moved down */
4737                                         rdev = conf->disks[i].rdev;
4738                                 rdev_clear_badblocks(rdev, sh->sector,
4739                                                      STRIPE_SECTORS, 0);
4740                                 rdev_dec_pending(rdev, conf->mddev);
4741                         }
4742                 }
4743
4744         if (s.ops_request)
4745                 raid_run_ops(sh, s.ops_request);
4746
4747         ops_run_io(sh, &s);
4748
4749         if (s.dec_preread_active) {
4750                 /* We delay this until after ops_run_io so that if make_request
4751                  * is waiting on a flush, it won't continue until the writes
4752                  * have actually been submitted.
4753                  */
4754                 atomic_dec(&conf->preread_active_stripes);
4755                 if (atomic_read(&conf->preread_active_stripes) <
4756                     IO_THRESHOLD)
4757                         md_wakeup_thread(conf->mddev->thread);
4758         }
4759
4760         if (!bio_list_empty(&s.return_bi)) {
4761                 if (test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4762                         spin_lock_irq(&conf->device_lock);
4763                         bio_list_merge(&conf->return_bi, &s.return_bi);
4764                         spin_unlock_irq(&conf->device_lock);
4765                         md_wakeup_thread(conf->mddev->thread);
4766                 } else
4767                         return_io(&s.return_bi);
4768         }
4769
4770         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4771 }
4772
4773 static void raid5_activate_delayed(struct r5conf *conf)
4774 {
4775         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4776                 while (!list_empty(&conf->delayed_list)) {
4777                         struct list_head *l = conf->delayed_list.next;
4778                         struct stripe_head *sh;
4779                         sh = list_entry(l, struct stripe_head, lru);
4780                         list_del_init(l);
4781                         clear_bit(STRIPE_DELAYED, &sh->state);
4782                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4783                                 atomic_inc(&conf->preread_active_stripes);
4784                         list_add_tail(&sh->lru, &conf->hold_list);
4785                         raid5_wakeup_stripe_thread(sh);
4786                 }
4787         }
4788 }
4789
4790 static void activate_bit_delay(struct r5conf *conf,
4791         struct list_head *temp_inactive_list)
4792 {
4793         /* device_lock is held */
4794         struct list_head head;
4795         list_add(&head, &conf->bitmap_list);
4796         list_del_init(&conf->bitmap_list);
4797         while (!list_empty(&head)) {
4798                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4799                 int hash;
4800                 list_del_init(&sh->lru);
4801                 atomic_inc(&sh->count);
4802                 hash = sh->hash_lock_index;
4803                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4804         }
4805 }
4806
4807 static int raid5_congested(struct mddev *mddev, int bits)
4808 {
4809         struct r5conf *conf = mddev->private;
4810
4811         /* No difference between reads and writes.  Just check
4812          * how busy the stripe_cache is
4813          */
4814
4815         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4816                 return 1;
4817
4818         /* Also checks whether there is pressure on r5cache log space */
4819         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
4820                 return 1;
4821         if (conf->quiesce)
4822                 return 1;
4823         if (atomic_read(&conf->empty_inactive_list_nr))
4824                 return 1;
4825
4826         return 0;
4827 }
4828
4829 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4830 {
4831         struct r5conf *conf = mddev->private;
4832         sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4833         unsigned int chunk_sectors;
4834         unsigned int bio_sectors = bio_sectors(bio);
4835
4836         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4837         return  chunk_sectors >=
4838                 ((sector & (chunk_sectors - 1)) + bio_sectors);
4839 }
4840
4841 /*
4842  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
4843  *  later sampled by raid5d.
4844  */
4845 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4846 {
4847         unsigned long flags;
4848
4849         spin_lock_irqsave(&conf->device_lock, flags);
4850
4851         bi->bi_next = conf->retry_read_aligned_list;
4852         conf->retry_read_aligned_list = bi;
4853
4854         spin_unlock_irqrestore(&conf->device_lock, flags);
4855         md_wakeup_thread(conf->mddev->thread);
4856 }
4857
4858 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4859 {
4860         struct bio *bi;
4861
4862         bi = conf->retry_read_aligned;
4863         if (bi) {
4864                 conf->retry_read_aligned = NULL;
4865                 return bi;
4866         }
4867         bi = conf->retry_read_aligned_list;
4868         if(bi) {
4869                 conf->retry_read_aligned_list = bi->bi_next;
4870                 bi->bi_next = NULL;
4871                 /*
4872                  * this sets the active strip count to 1 and the processed
4873                  * strip count to zero (upper 8 bits)
4874                  */
4875                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4876         }
4877
4878         return bi;
4879 }
4880
4881 /*
4882  *  The "raid5_align_endio" should check if the read succeeded and if it
4883  *  did, call bio_endio on the original bio (having bio_put the new bio
4884  *  first).
4885  *  If the read failed..
4886  */
4887 static void raid5_align_endio(struct bio *bi)
4888 {
4889         struct bio* raid_bi  = bi->bi_private;
4890         struct mddev *mddev;
4891         struct r5conf *conf;
4892         struct md_rdev *rdev;
4893         int error = bi->bi_error;
4894
4895         bio_put(bi);
4896
4897         rdev = (void*)raid_bi->bi_next;
4898         raid_bi->bi_next = NULL;
4899         mddev = rdev->mddev;
4900         conf = mddev->private;
4901
4902         rdev_dec_pending(rdev, conf->mddev);
4903
4904         if (!error) {
4905                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4906                                          raid_bi, 0);
4907                 bio_endio(raid_bi);
4908                 if (atomic_dec_and_test(&conf->active_aligned_reads))
4909                         wake_up(&conf->wait_for_quiescent);
4910                 return;
4911         }
4912
4913         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4914
4915         add_bio_to_retry(raid_bi, conf);
4916 }
4917
4918 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4919 {
4920         struct r5conf *conf = mddev->private;
4921         int dd_idx;
4922         struct bio* align_bi;
4923         struct md_rdev *rdev;
4924         sector_t end_sector;
4925
4926         if (!in_chunk_boundary(mddev, raid_bio)) {
4927                 pr_debug("%s: non aligned\n", __func__);
4928                 return 0;
4929         }
4930         /*
4931          * use bio_clone_mddev to make a copy of the bio
4932          */
4933         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4934         if (!align_bi)
4935                 return 0;
4936         /*
4937          *   set bi_end_io to a new function, and set bi_private to the
4938          *     original bio.
4939          */
4940         align_bi->bi_end_io  = raid5_align_endio;
4941         align_bi->bi_private = raid_bio;
4942         /*
4943          *      compute position
4944          */
4945         align_bi->bi_iter.bi_sector =
4946                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4947                                      0, &dd_idx, NULL);
4948
4949         end_sector = bio_end_sector(align_bi);
4950         rcu_read_lock();
4951         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4952         if (!rdev || test_bit(Faulty, &rdev->flags) ||
4953             rdev->recovery_offset < end_sector) {
4954                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4955                 if (rdev &&
4956                     (test_bit(Faulty, &rdev->flags) ||
4957                     !(test_bit(In_sync, &rdev->flags) ||
4958                       rdev->recovery_offset >= end_sector)))
4959                         rdev = NULL;
4960         }
4961         if (rdev) {
4962                 sector_t first_bad;
4963                 int bad_sectors;
4964
4965                 atomic_inc(&rdev->nr_pending);
4966                 rcu_read_unlock();
4967                 raid_bio->bi_next = (void*)rdev;
4968                 align_bi->bi_bdev =  rdev->bdev;
4969                 bio_clear_flag(align_bi, BIO_SEG_VALID);
4970
4971                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4972                                 bio_sectors(align_bi),
4973                                 &first_bad, &bad_sectors)) {
4974                         bio_put(align_bi);
4975                         rdev_dec_pending(rdev, mddev);
4976                         return 0;
4977                 }
4978
4979                 /* No reshape active, so we can trust rdev->data_offset */
4980                 align_bi->bi_iter.bi_sector += rdev->data_offset;
4981
4982                 spin_lock_irq(&conf->device_lock);
4983                 wait_event_lock_irq(conf->wait_for_quiescent,
4984                                     conf->quiesce == 0,
4985                                     conf->device_lock);
4986                 atomic_inc(&conf->active_aligned_reads);
4987                 spin_unlock_irq(&conf->device_lock);
4988
4989                 if (mddev->gendisk)
4990                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4991                                               align_bi, disk_devt(mddev->gendisk),
4992                                               raid_bio->bi_iter.bi_sector);
4993                 generic_make_request(align_bi);
4994                 return 1;
4995         } else {
4996                 rcu_read_unlock();
4997                 bio_put(align_bi);
4998                 return 0;
4999         }
5000 }
5001
5002 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5003 {
5004         struct bio *split;
5005
5006         do {
5007                 sector_t sector = raid_bio->bi_iter.bi_sector;
5008                 unsigned chunk_sects = mddev->chunk_sectors;
5009                 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5010
5011                 if (sectors < bio_sectors(raid_bio)) {
5012                         split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
5013                         bio_chain(split, raid_bio);
5014                 } else
5015                         split = raid_bio;
5016
5017                 if (!raid5_read_one_chunk(mddev, split)) {
5018                         if (split != raid_bio)
5019                                 generic_make_request(raid_bio);
5020                         return split;
5021                 }
5022         } while (split != raid_bio);
5023
5024         return NULL;
5025 }
5026
5027 /* __get_priority_stripe - get the next stripe to process
5028  *
5029  * Full stripe writes are allowed to pass preread active stripes up until
5030  * the bypass_threshold is exceeded.  In general the bypass_count
5031  * increments when the handle_list is handled before the hold_list; however, it
5032  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5033  * stripe with in flight i/o.  The bypass_count will be reset when the
5034  * head of the hold_list has changed, i.e. the head was promoted to the
5035  * handle_list.
5036  */
5037 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5038 {
5039         struct stripe_head *sh = NULL, *tmp;
5040         struct list_head *handle_list = NULL;
5041         struct r5worker_group *wg = NULL;
5042
5043         if (conf->worker_cnt_per_group == 0) {
5044                 handle_list = &conf->handle_list;
5045         } else if (group != ANY_GROUP) {
5046                 handle_list = &conf->worker_groups[group].handle_list;
5047                 wg = &conf->worker_groups[group];
5048         } else {
5049                 int i;
5050                 for (i = 0; i < conf->group_cnt; i++) {
5051                         handle_list = &conf->worker_groups[i].handle_list;
5052                         wg = &conf->worker_groups[i];
5053                         if (!list_empty(handle_list))
5054                                 break;
5055                 }
5056         }
5057
5058         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5059                   __func__,
5060                   list_empty(handle_list) ? "empty" : "busy",
5061                   list_empty(&conf->hold_list) ? "empty" : "busy",
5062                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5063
5064         if (!list_empty(handle_list)) {
5065                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5066
5067                 if (list_empty(&conf->hold_list))
5068                         conf->bypass_count = 0;
5069                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5070                         if (conf->hold_list.next == conf->last_hold)
5071                                 conf->bypass_count++;
5072                         else {
5073                                 conf->last_hold = conf->hold_list.next;
5074                                 conf->bypass_count -= conf->bypass_threshold;
5075                                 if (conf->bypass_count < 0)
5076                                         conf->bypass_count = 0;
5077                         }
5078                 }
5079         } else if (!list_empty(&conf->hold_list) &&
5080                    ((conf->bypass_threshold &&
5081                      conf->bypass_count > conf->bypass_threshold) ||
5082                     atomic_read(&conf->pending_full_writes) == 0)) {
5083
5084                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5085                         if (conf->worker_cnt_per_group == 0 ||
5086                             group == ANY_GROUP ||
5087                             !cpu_online(tmp->cpu) ||
5088                             cpu_to_group(tmp->cpu) == group) {
5089                                 sh = tmp;
5090                                 break;
5091                         }
5092                 }
5093
5094                 if (sh) {
5095                         conf->bypass_count -= conf->bypass_threshold;
5096                         if (conf->bypass_count < 0)
5097                                 conf->bypass_count = 0;
5098                 }
5099                 wg = NULL;
5100         }
5101
5102         if (!sh)
5103                 return NULL;
5104
5105         if (wg) {
5106                 wg->stripes_cnt--;
5107                 sh->group = NULL;
5108         }
5109         list_del_init(&sh->lru);
5110         BUG_ON(atomic_inc_return(&sh->count) != 1);
5111         return sh;
5112 }
5113
5114 struct raid5_plug_cb {
5115         struct blk_plug_cb      cb;
5116         struct list_head        list;
5117         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5118 };
5119
5120 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5121 {
5122         struct raid5_plug_cb *cb = container_of(
5123                 blk_cb, struct raid5_plug_cb, cb);
5124         struct stripe_head *sh;
5125         struct mddev *mddev = cb->cb.data;
5126         struct r5conf *conf = mddev->private;
5127         int cnt = 0;
5128         int hash;
5129
5130         if (cb->list.next && !list_empty(&cb->list)) {
5131                 spin_lock_irq(&conf->device_lock);
5132                 while (!list_empty(&cb->list)) {
5133                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5134                         list_del_init(&sh->lru);
5135                         /*
5136                          * avoid race release_stripe_plug() sees
5137                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5138                          * is still in our list
5139                          */
5140                         smp_mb__before_atomic();
5141                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5142                         /*
5143                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5144                          * case, the count is always > 1 here
5145                          */
5146                         hash = sh->hash_lock_index;
5147                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5148                         cnt++;
5149                 }
5150                 spin_unlock_irq(&conf->device_lock);
5151         }
5152         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5153                                      NR_STRIPE_HASH_LOCKS);
5154         if (mddev->queue)
5155                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5156         kfree(cb);
5157 }
5158
5159 static void release_stripe_plug(struct mddev *mddev,
5160                                 struct stripe_head *sh)
5161 {
5162         struct blk_plug_cb *blk_cb = blk_check_plugged(
5163                 raid5_unplug, mddev,
5164                 sizeof(struct raid5_plug_cb));
5165         struct raid5_plug_cb *cb;
5166
5167         if (!blk_cb) {
5168                 raid5_release_stripe(sh);
5169                 return;
5170         }
5171
5172         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5173
5174         if (cb->list.next == NULL) {
5175                 int i;
5176                 INIT_LIST_HEAD(&cb->list);
5177                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5178                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5179         }
5180
5181         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5182                 list_add_tail(&sh->lru, &cb->list);
5183         else
5184                 raid5_release_stripe(sh);
5185 }
5186
5187 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5188 {
5189         struct r5conf *conf = mddev->private;
5190         sector_t logical_sector, last_sector;
5191         struct stripe_head *sh;
5192         int remaining;
5193         int stripe_sectors;
5194
5195         if (mddev->reshape_position != MaxSector)
5196                 /* Skip discard while reshape is happening */
5197                 return;
5198
5199         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5200         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5201
5202         bi->bi_next = NULL;
5203         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5204
5205         stripe_sectors = conf->chunk_sectors *
5206                 (conf->raid_disks - conf->max_degraded);
5207         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5208                                                stripe_sectors);
5209         sector_div(last_sector, stripe_sectors);
5210
5211         logical_sector *= conf->chunk_sectors;
5212         last_sector *= conf->chunk_sectors;
5213
5214         for (; logical_sector < last_sector;
5215              logical_sector += STRIPE_SECTORS) {
5216                 DEFINE_WAIT(w);
5217                 int d;
5218         again:
5219                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5220                 prepare_to_wait(&conf->wait_for_overlap, &w,
5221                                 TASK_UNINTERRUPTIBLE);
5222                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5223                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5224                         raid5_release_stripe(sh);
5225                         schedule();
5226                         goto again;
5227                 }
5228                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5229                 spin_lock_irq(&sh->stripe_lock);
5230                 for (d = 0; d < conf->raid_disks; d++) {
5231                         if (d == sh->pd_idx || d == sh->qd_idx)
5232                                 continue;
5233                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5234                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5235                                 spin_unlock_irq(&sh->stripe_lock);
5236                                 raid5_release_stripe(sh);
5237                                 schedule();
5238                                 goto again;
5239                         }
5240                 }
5241                 set_bit(STRIPE_DISCARD, &sh->state);
5242                 finish_wait(&conf->wait_for_overlap, &w);
5243                 sh->overwrite_disks = 0;
5244                 for (d = 0; d < conf->raid_disks; d++) {
5245                         if (d == sh->pd_idx || d == sh->qd_idx)
5246                                 continue;
5247                         sh->dev[d].towrite = bi;
5248                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5249                         raid5_inc_bi_active_stripes(bi);
5250                         sh->overwrite_disks++;
5251                 }
5252                 spin_unlock_irq(&sh->stripe_lock);
5253                 if (conf->mddev->bitmap) {
5254                         for (d = 0;
5255                              d < conf->raid_disks - conf->max_degraded;
5256                              d++)
5257                                 bitmap_startwrite(mddev->bitmap,
5258                                                   sh->sector,
5259                                                   STRIPE_SECTORS,
5260                                                   0);
5261                         sh->bm_seq = conf->seq_flush + 1;
5262                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5263                 }
5264
5265                 set_bit(STRIPE_HANDLE, &sh->state);
5266                 clear_bit(STRIPE_DELAYED, &sh->state);
5267                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5268                         atomic_inc(&conf->preread_active_stripes);
5269                 release_stripe_plug(mddev, sh);
5270         }
5271
5272         remaining = raid5_dec_bi_active_stripes(bi);
5273         if (remaining == 0) {
5274                 md_write_end(mddev);
5275                 bio_endio(bi);
5276         }
5277 }
5278
5279 static void raid5_make_request(struct mddev *mddev, struct bio * bi)
5280 {
5281         struct r5conf *conf = mddev->private;
5282         int dd_idx;
5283         sector_t new_sector;
5284         sector_t logical_sector, last_sector;
5285         struct stripe_head *sh;
5286         const int rw = bio_data_dir(bi);
5287         int remaining;
5288         DEFINE_WAIT(w);
5289         bool do_prepare;
5290         bool do_flush = false;
5291
5292         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5293                 int ret = r5l_handle_flush_request(conf->log, bi);
5294
5295                 if (ret == 0)
5296                         return;
5297                 if (ret == -ENODEV) {
5298                         md_flush_request(mddev, bi);
5299                         return;
5300                 }
5301                 /* ret == -EAGAIN, fallback */
5302                 /*
5303                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5304                  * we need to flush journal device
5305                  */
5306                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5307         }
5308
5309         md_write_start(mddev, bi);
5310
5311         /*
5312          * If array is degraded, better not do chunk aligned read because
5313          * later we might have to read it again in order to reconstruct
5314          * data on failed drives.
5315          */
5316         if (rw == READ && mddev->degraded == 0 &&
5317             !r5c_is_writeback(conf->log) &&
5318             mddev->reshape_position == MaxSector) {
5319                 bi = chunk_aligned_read(mddev, bi);
5320                 if (!bi)
5321                         return;
5322         }
5323
5324         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5325                 make_discard_request(mddev, bi);
5326                 return;
5327         }
5328
5329         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5330         last_sector = bio_end_sector(bi);
5331         bi->bi_next = NULL;
5332         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
5333
5334         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5335         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5336                 int previous;
5337                 int seq;
5338
5339                 do_prepare = false;
5340         retry:
5341                 seq = read_seqcount_begin(&conf->gen_lock);
5342                 previous = 0;
5343                 if (do_prepare)
5344                         prepare_to_wait(&conf->wait_for_overlap, &w,
5345                                 TASK_UNINTERRUPTIBLE);
5346                 if (unlikely(conf->reshape_progress != MaxSector)) {
5347                         /* spinlock is needed as reshape_progress may be
5348                          * 64bit on a 32bit platform, and so it might be
5349                          * possible to see a half-updated value
5350                          * Of course reshape_progress could change after
5351                          * the lock is dropped, so once we get a reference
5352                          * to the stripe that we think it is, we will have
5353                          * to check again.
5354                          */
5355                         spin_lock_irq(&conf->device_lock);
5356                         if (mddev->reshape_backwards
5357                             ? logical_sector < conf->reshape_progress
5358                             : logical_sector >= conf->reshape_progress) {
5359                                 previous = 1;
5360                         } else {
5361                                 if (mddev->reshape_backwards
5362                                     ? logical_sector < conf->reshape_safe
5363                                     : logical_sector >= conf->reshape_safe) {
5364                                         spin_unlock_irq(&conf->device_lock);
5365                                         schedule();
5366                                         do_prepare = true;
5367                                         goto retry;
5368                                 }
5369                         }
5370                         spin_unlock_irq(&conf->device_lock);
5371                 }
5372
5373                 new_sector = raid5_compute_sector(conf, logical_sector,
5374                                                   previous,
5375                                                   &dd_idx, NULL);
5376                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5377                         (unsigned long long)new_sector,
5378                         (unsigned long long)logical_sector);
5379
5380                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5381                                        (bi->bi_opf & REQ_RAHEAD), 0);
5382                 if (sh) {
5383                         if (unlikely(previous)) {
5384                                 /* expansion might have moved on while waiting for a
5385                                  * stripe, so we must do the range check again.
5386                                  * Expansion could still move past after this
5387                                  * test, but as we are holding a reference to
5388                                  * 'sh', we know that if that happens,
5389                                  *  STRIPE_EXPANDING will get set and the expansion
5390                                  * won't proceed until we finish with the stripe.
5391                                  */
5392                                 int must_retry = 0;
5393                                 spin_lock_irq(&conf->device_lock);
5394                                 if (mddev->reshape_backwards
5395                                     ? logical_sector >= conf->reshape_progress
5396                                     : logical_sector < conf->reshape_progress)
5397                                         /* mismatch, need to try again */
5398                                         must_retry = 1;
5399                                 spin_unlock_irq(&conf->device_lock);
5400                                 if (must_retry) {
5401                                         raid5_release_stripe(sh);
5402                                         schedule();
5403                                         do_prepare = true;
5404                                         goto retry;
5405                                 }
5406                         }
5407                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5408                                 /* Might have got the wrong stripe_head
5409                                  * by accident
5410                                  */
5411                                 raid5_release_stripe(sh);
5412                                 goto retry;
5413                         }
5414
5415                         if (rw == WRITE &&
5416                             logical_sector >= mddev->suspend_lo &&
5417                             logical_sector < mddev->suspend_hi) {
5418                                 raid5_release_stripe(sh);
5419                                 /* As the suspend_* range is controlled by
5420                                  * userspace, we want an interruptible
5421                                  * wait.
5422                                  */
5423                                 flush_signals(current);
5424                                 prepare_to_wait(&conf->wait_for_overlap,
5425                                                 &w, TASK_INTERRUPTIBLE);
5426                                 if (logical_sector >= mddev->suspend_lo &&
5427                                     logical_sector < mddev->suspend_hi) {
5428                                         schedule();
5429                                         do_prepare = true;
5430                                 }
5431                                 goto retry;
5432                         }
5433
5434                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5435                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5436                                 /* Stripe is busy expanding or
5437                                  * add failed due to overlap.  Flush everything
5438                                  * and wait a while
5439                                  */
5440                                 md_wakeup_thread(mddev->thread);
5441                                 raid5_release_stripe(sh);
5442                                 schedule();
5443                                 do_prepare = true;
5444                                 goto retry;
5445                         }
5446                         if (do_flush) {
5447                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5448                                 /* we only need flush for one stripe */
5449                                 do_flush = false;
5450                         }
5451
5452                         set_bit(STRIPE_HANDLE, &sh->state);
5453                         clear_bit(STRIPE_DELAYED, &sh->state);
5454                         if ((!sh->batch_head || sh == sh->batch_head) &&
5455                             (bi->bi_opf & REQ_SYNC) &&
5456                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5457                                 atomic_inc(&conf->preread_active_stripes);
5458                         release_stripe_plug(mddev, sh);
5459                 } else {
5460                         /* cannot get stripe for read-ahead, just give-up */
5461                         bi->bi_error = -EIO;
5462                         break;
5463                 }
5464         }
5465         finish_wait(&conf->wait_for_overlap, &w);
5466
5467         remaining = raid5_dec_bi_active_stripes(bi);
5468         if (remaining == 0) {
5469
5470                 if ( rw == WRITE )
5471                         md_write_end(mddev);
5472
5473                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5474                                          bi, 0);
5475                 bio_endio(bi);
5476         }
5477 }
5478
5479 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5480
5481 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5482 {
5483         /* reshaping is quite different to recovery/resync so it is
5484          * handled quite separately ... here.
5485          *
5486          * On each call to sync_request, we gather one chunk worth of
5487          * destination stripes and flag them as expanding.
5488          * Then we find all the source stripes and request reads.
5489          * As the reads complete, handle_stripe will copy the data
5490          * into the destination stripe and release that stripe.
5491          */
5492         struct r5conf *conf = mddev->private;
5493         struct stripe_head *sh;
5494         sector_t first_sector, last_sector;
5495         int raid_disks = conf->previous_raid_disks;
5496         int data_disks = raid_disks - conf->max_degraded;
5497         int new_data_disks = conf->raid_disks - conf->max_degraded;
5498         int i;
5499         int dd_idx;
5500         sector_t writepos, readpos, safepos;
5501         sector_t stripe_addr;
5502         int reshape_sectors;
5503         struct list_head stripes;
5504         sector_t retn;
5505
5506         if (sector_nr == 0) {
5507                 /* If restarting in the middle, skip the initial sectors */
5508                 if (mddev->reshape_backwards &&
5509                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5510                         sector_nr = raid5_size(mddev, 0, 0)
5511                                 - conf->reshape_progress;
5512                 } else if (mddev->reshape_backwards &&
5513                            conf->reshape_progress == MaxSector) {
5514                         /* shouldn't happen, but just in case, finish up.*/
5515                         sector_nr = MaxSector;
5516                 } else if (!mddev->reshape_backwards &&
5517                            conf->reshape_progress > 0)
5518                         sector_nr = conf->reshape_progress;
5519                 sector_div(sector_nr, new_data_disks);
5520                 if (sector_nr) {
5521                         mddev->curr_resync_completed = sector_nr;
5522                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5523                         *skipped = 1;
5524                         retn = sector_nr;
5525                         goto finish;
5526                 }
5527         }
5528
5529         /* We need to process a full chunk at a time.
5530          * If old and new chunk sizes differ, we need to process the
5531          * largest of these
5532          */
5533
5534         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5535
5536         /* We update the metadata at least every 10 seconds, or when
5537          * the data about to be copied would over-write the source of
5538          * the data at the front of the range.  i.e. one new_stripe
5539          * along from reshape_progress new_maps to after where
5540          * reshape_safe old_maps to
5541          */
5542         writepos = conf->reshape_progress;
5543         sector_div(writepos, new_data_disks);
5544         readpos = conf->reshape_progress;
5545         sector_div(readpos, data_disks);
5546         safepos = conf->reshape_safe;
5547         sector_div(safepos, data_disks);
5548         if (mddev->reshape_backwards) {
5549                 BUG_ON(writepos < reshape_sectors);
5550                 writepos -= reshape_sectors;
5551                 readpos += reshape_sectors;
5552                 safepos += reshape_sectors;
5553         } else {
5554                 writepos += reshape_sectors;
5555                 /* readpos and safepos are worst-case calculations.
5556                  * A negative number is overly pessimistic, and causes
5557                  * obvious problems for unsigned storage.  So clip to 0.
5558                  */
5559                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5560                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5561         }
5562
5563         /* Having calculated the 'writepos' possibly use it
5564          * to set 'stripe_addr' which is where we will write to.
5565          */
5566         if (mddev->reshape_backwards) {
5567                 BUG_ON(conf->reshape_progress == 0);
5568                 stripe_addr = writepos;
5569                 BUG_ON((mddev->dev_sectors &
5570                         ~((sector_t)reshape_sectors - 1))
5571                        - reshape_sectors - stripe_addr
5572                        != sector_nr);
5573         } else {
5574                 BUG_ON(writepos != sector_nr + reshape_sectors);
5575                 stripe_addr = sector_nr;
5576         }
5577
5578         /* 'writepos' is the most advanced device address we might write.
5579          * 'readpos' is the least advanced device address we might read.
5580          * 'safepos' is the least address recorded in the metadata as having
5581          *     been reshaped.
5582          * If there is a min_offset_diff, these are adjusted either by
5583          * increasing the safepos/readpos if diff is negative, or
5584          * increasing writepos if diff is positive.
5585          * If 'readpos' is then behind 'writepos', there is no way that we can
5586          * ensure safety in the face of a crash - that must be done by userspace
5587          * making a backup of the data.  So in that case there is no particular
5588          * rush to update metadata.
5589          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5590          * update the metadata to advance 'safepos' to match 'readpos' so that
5591          * we can be safe in the event of a crash.
5592          * So we insist on updating metadata if safepos is behind writepos and
5593          * readpos is beyond writepos.
5594          * In any case, update the metadata every 10 seconds.
5595          * Maybe that number should be configurable, but I'm not sure it is
5596          * worth it.... maybe it could be a multiple of safemode_delay???
5597          */
5598         if (conf->min_offset_diff < 0) {
5599                 safepos += -conf->min_offset_diff;
5600                 readpos += -conf->min_offset_diff;
5601         } else
5602                 writepos += conf->min_offset_diff;
5603
5604         if ((mddev->reshape_backwards
5605              ? (safepos > writepos && readpos < writepos)
5606              : (safepos < writepos && readpos > writepos)) ||
5607             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5608                 /* Cannot proceed until we've updated the superblock... */
5609                 wait_event(conf->wait_for_overlap,
5610                            atomic_read(&conf->reshape_stripes)==0
5611                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5612                 if (atomic_read(&conf->reshape_stripes) != 0)
5613                         return 0;
5614                 mddev->reshape_position = conf->reshape_progress;
5615                 mddev->curr_resync_completed = sector_nr;
5616                 conf->reshape_checkpoint = jiffies;
5617                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5618                 md_wakeup_thread(mddev->thread);
5619                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5620                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5621                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5622                         return 0;
5623                 spin_lock_irq(&conf->device_lock);
5624                 conf->reshape_safe = mddev->reshape_position;
5625                 spin_unlock_irq(&conf->device_lock);
5626                 wake_up(&conf->wait_for_overlap);
5627                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5628         }
5629
5630         INIT_LIST_HEAD(&stripes);
5631         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5632                 int j;
5633                 int skipped_disk = 0;
5634                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5635                 set_bit(STRIPE_EXPANDING, &sh->state);
5636                 atomic_inc(&conf->reshape_stripes);
5637                 /* If any of this stripe is beyond the end of the old
5638                  * array, then we need to zero those blocks
5639                  */
5640                 for (j=sh->disks; j--;) {
5641                         sector_t s;
5642                         if (j == sh->pd_idx)
5643                                 continue;
5644                         if (conf->level == 6 &&
5645                             j == sh->qd_idx)
5646                                 continue;
5647                         s = raid5_compute_blocknr(sh, j, 0);
5648                         if (s < raid5_size(mddev, 0, 0)) {
5649                                 skipped_disk = 1;
5650                                 continue;
5651                         }
5652                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5653                         set_bit(R5_Expanded, &sh->dev[j].flags);
5654                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5655                 }
5656                 if (!skipped_disk) {
5657                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5658                         set_bit(STRIPE_HANDLE, &sh->state);
5659                 }
5660                 list_add(&sh->lru, &stripes);
5661         }
5662         spin_lock_irq(&conf->device_lock);
5663         if (mddev->reshape_backwards)
5664                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5665         else
5666                 conf->reshape_progress += reshape_sectors * new_data_disks;
5667         spin_unlock_irq(&conf->device_lock);
5668         /* Ok, those stripe are ready. We can start scheduling
5669          * reads on the source stripes.
5670          * The source stripes are determined by mapping the first and last
5671          * block on the destination stripes.
5672          */
5673         first_sector =
5674                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5675                                      1, &dd_idx, NULL);
5676         last_sector =
5677                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5678                                             * new_data_disks - 1),
5679                                      1, &dd_idx, NULL);
5680         if (last_sector >= mddev->dev_sectors)
5681                 last_sector = mddev->dev_sectors - 1;
5682         while (first_sector <= last_sector) {
5683                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5684                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5685                 set_bit(STRIPE_HANDLE, &sh->state);
5686                 raid5_release_stripe(sh);
5687                 first_sector += STRIPE_SECTORS;
5688         }
5689         /* Now that the sources are clearly marked, we can release
5690          * the destination stripes
5691          */
5692         while (!list_empty(&stripes)) {
5693                 sh = list_entry(stripes.next, struct stripe_head, lru);
5694                 list_del_init(&sh->lru);
5695                 raid5_release_stripe(sh);
5696         }
5697         /* If this takes us to the resync_max point where we have to pause,
5698          * then we need to write out the superblock.
5699          */
5700         sector_nr += reshape_sectors;
5701         retn = reshape_sectors;
5702 finish:
5703         if (mddev->curr_resync_completed > mddev->resync_max ||
5704             (sector_nr - mddev->curr_resync_completed) * 2
5705             >= mddev->resync_max - mddev->curr_resync_completed) {
5706                 /* Cannot proceed until we've updated the superblock... */
5707                 wait_event(conf->wait_for_overlap,
5708                            atomic_read(&conf->reshape_stripes) == 0
5709                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5710                 if (atomic_read(&conf->reshape_stripes) != 0)
5711                         goto ret;
5712                 mddev->reshape_position = conf->reshape_progress;
5713                 mddev->curr_resync_completed = sector_nr;
5714                 conf->reshape_checkpoint = jiffies;
5715                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5716                 md_wakeup_thread(mddev->thread);
5717                 wait_event(mddev->sb_wait,
5718                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5719                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5720                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5721                         goto ret;
5722                 spin_lock_irq(&conf->device_lock);
5723                 conf->reshape_safe = mddev->reshape_position;
5724                 spin_unlock_irq(&conf->device_lock);
5725                 wake_up(&conf->wait_for_overlap);
5726                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5727         }
5728 ret:
5729         return retn;
5730 }
5731
5732 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
5733                                           int *skipped)
5734 {
5735         struct r5conf *conf = mddev->private;
5736         struct stripe_head *sh;
5737         sector_t max_sector = mddev->dev_sectors;
5738         sector_t sync_blocks;
5739         int still_degraded = 0;
5740         int i;
5741
5742         if (sector_nr >= max_sector) {
5743                 /* just being told to finish up .. nothing much to do */
5744
5745                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5746                         end_reshape(conf);
5747                         return 0;
5748                 }
5749
5750                 if (mddev->curr_resync < max_sector) /* aborted */
5751                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5752                                         &sync_blocks, 1);
5753                 else /* completed sync */
5754                         conf->fullsync = 0;
5755                 bitmap_close_sync(mddev->bitmap);
5756
5757                 return 0;
5758         }
5759
5760         /* Allow raid5_quiesce to complete */
5761         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5762
5763         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5764                 return reshape_request(mddev, sector_nr, skipped);
5765
5766         /* No need to check resync_max as we never do more than one
5767          * stripe, and as resync_max will always be on a chunk boundary,
5768          * if the check in md_do_sync didn't fire, there is no chance
5769          * of overstepping resync_max here
5770          */
5771
5772         /* if there is too many failed drives and we are trying
5773          * to resync, then assert that we are finished, because there is
5774          * nothing we can do.
5775          */
5776         if (mddev->degraded >= conf->max_degraded &&
5777             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5778                 sector_t rv = mddev->dev_sectors - sector_nr;
5779                 *skipped = 1;
5780                 return rv;
5781         }
5782         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5783             !conf->fullsync &&
5784             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5785             sync_blocks >= STRIPE_SECTORS) {
5786                 /* we can skip this block, and probably more */
5787                 sync_blocks /= STRIPE_SECTORS;
5788                 *skipped = 1;
5789                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5790         }
5791
5792         bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5793
5794         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5795         if (sh == NULL) {
5796                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5797                 /* make sure we don't swamp the stripe cache if someone else
5798                  * is trying to get access
5799                  */
5800                 schedule_timeout_uninterruptible(1);
5801         }
5802         /* Need to check if array will still be degraded after recovery/resync
5803          * Note in case of > 1 drive failures it's possible we're rebuilding
5804          * one drive while leaving another faulty drive in array.
5805          */
5806         rcu_read_lock();
5807         for (i = 0; i < conf->raid_disks; i++) {
5808                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5809
5810                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5811                         still_degraded = 1;
5812         }
5813         rcu_read_unlock();
5814
5815         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5816
5817         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5818         set_bit(STRIPE_HANDLE, &sh->state);
5819
5820         raid5_release_stripe(sh);
5821
5822         return STRIPE_SECTORS;
5823 }
5824
5825 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5826 {
5827         /* We may not be able to submit a whole bio at once as there
5828          * may not be enough stripe_heads available.
5829          * We cannot pre-allocate enough stripe_heads as we may need
5830          * more than exist in the cache (if we allow ever large chunks).
5831          * So we do one stripe head at a time and record in
5832          * ->bi_hw_segments how many have been done.
5833          *
5834          * We *know* that this entire raid_bio is in one chunk, so
5835          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5836          */
5837         struct stripe_head *sh;
5838         int dd_idx;
5839         sector_t sector, logical_sector, last_sector;
5840         int scnt = 0;
5841         int remaining;
5842         int handled = 0;
5843
5844         logical_sector = raid_bio->bi_iter.bi_sector &
5845                 ~((sector_t)STRIPE_SECTORS-1);
5846         sector = raid5_compute_sector(conf, logical_sector,
5847                                       0, &dd_idx, NULL);
5848         last_sector = bio_end_sector(raid_bio);
5849
5850         for (; logical_sector < last_sector;
5851              logical_sector += STRIPE_SECTORS,
5852                      sector += STRIPE_SECTORS,
5853                      scnt++) {
5854
5855                 if (scnt < raid5_bi_processed_stripes(raid_bio))
5856                         /* already done this stripe */
5857                         continue;
5858
5859                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5860
5861                 if (!sh) {
5862                         /* failed to get a stripe - must wait */
5863                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5864                         conf->retry_read_aligned = raid_bio;
5865                         return handled;
5866                 }
5867
5868                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5869                         raid5_release_stripe(sh);
5870                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5871                         conf->retry_read_aligned = raid_bio;
5872                         return handled;
5873                 }
5874
5875                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5876                 handle_stripe(sh);
5877                 raid5_release_stripe(sh);
5878                 handled++;
5879         }
5880         remaining = raid5_dec_bi_active_stripes(raid_bio);
5881         if (remaining == 0) {
5882                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5883                                          raid_bio, 0);
5884                 bio_endio(raid_bio);
5885         }
5886         if (atomic_dec_and_test(&conf->active_aligned_reads))
5887                 wake_up(&conf->wait_for_quiescent);
5888         return handled;
5889 }
5890
5891 static int handle_active_stripes(struct r5conf *conf, int group,
5892                                  struct r5worker *worker,
5893                                  struct list_head *temp_inactive_list)
5894 {
5895         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5896         int i, batch_size = 0, hash;
5897         bool release_inactive = false;
5898
5899         while (batch_size < MAX_STRIPE_BATCH &&
5900                         (sh = __get_priority_stripe(conf, group)) != NULL)
5901                 batch[batch_size++] = sh;
5902
5903         if (batch_size == 0) {
5904                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5905                         if (!list_empty(temp_inactive_list + i))
5906                                 break;
5907                 if (i == NR_STRIPE_HASH_LOCKS) {
5908                         spin_unlock_irq(&conf->device_lock);
5909                         r5l_flush_stripe_to_raid(conf->log);
5910                         spin_lock_irq(&conf->device_lock);
5911                         return batch_size;
5912                 }
5913                 release_inactive = true;
5914         }
5915         spin_unlock_irq(&conf->device_lock);
5916
5917         release_inactive_stripe_list(conf, temp_inactive_list,
5918                                      NR_STRIPE_HASH_LOCKS);
5919
5920         r5l_flush_stripe_to_raid(conf->log);
5921         if (release_inactive) {
5922                 spin_lock_irq(&conf->device_lock);
5923                 return 0;
5924         }
5925
5926         for (i = 0; i < batch_size; i++)
5927                 handle_stripe(batch[i]);
5928         r5l_write_stripe_run(conf->log);
5929
5930         cond_resched();
5931
5932         spin_lock_irq(&conf->device_lock);
5933         for (i = 0; i < batch_size; i++) {
5934                 hash = batch[i]->hash_lock_index;
5935                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5936         }
5937         return batch_size;
5938 }
5939
5940 static void raid5_do_work(struct work_struct *work)
5941 {
5942         struct r5worker *worker = container_of(work, struct r5worker, work);
5943         struct r5worker_group *group = worker->group;
5944         struct r5conf *conf = group->conf;
5945         int group_id = group - conf->worker_groups;
5946         int handled;
5947         struct blk_plug plug;
5948
5949         pr_debug("+++ raid5worker active\n");
5950
5951         blk_start_plug(&plug);
5952         handled = 0;
5953         spin_lock_irq(&conf->device_lock);
5954         while (1) {
5955                 int batch_size, released;
5956
5957                 released = release_stripe_list(conf, worker->temp_inactive_list);
5958
5959                 batch_size = handle_active_stripes(conf, group_id, worker,
5960                                                    worker->temp_inactive_list);
5961                 worker->working = false;
5962                 if (!batch_size && !released)
5963                         break;
5964                 handled += batch_size;
5965         }
5966         pr_debug("%d stripes handled\n", handled);
5967
5968         spin_unlock_irq(&conf->device_lock);
5969         blk_finish_plug(&plug);
5970
5971         pr_debug("--- raid5worker inactive\n");
5972 }
5973
5974 /*
5975  * This is our raid5 kernel thread.
5976  *
5977  * We scan the hash table for stripes which can be handled now.
5978  * During the scan, completed stripes are saved for us by the interrupt
5979  * handler, so that they will not have to wait for our next wakeup.
5980  */
5981 static void raid5d(struct md_thread *thread)
5982 {
5983         struct mddev *mddev = thread->mddev;
5984         struct r5conf *conf = mddev->private;
5985         int handled;
5986         struct blk_plug plug;
5987
5988         pr_debug("+++ raid5d active\n");
5989
5990         md_check_recovery(mddev);
5991
5992         if (!bio_list_empty(&conf->return_bi) &&
5993             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
5994                 struct bio_list tmp = BIO_EMPTY_LIST;
5995                 spin_lock_irq(&conf->device_lock);
5996                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
5997                         bio_list_merge(&tmp, &conf->return_bi);
5998                         bio_list_init(&conf->return_bi);
5999                 }
6000                 spin_unlock_irq(&conf->device_lock);
6001                 return_io(&tmp);
6002         }
6003
6004         blk_start_plug(&plug);
6005         handled = 0;
6006         spin_lock_irq(&conf->device_lock);
6007         while (1) {
6008                 struct bio *bio;
6009                 int batch_size, released;
6010
6011                 released = release_stripe_list(conf, conf->temp_inactive_list);
6012                 if (released)
6013                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6014
6015                 if (
6016                     !list_empty(&conf->bitmap_list)) {
6017                         /* Now is a good time to flush some bitmap updates */
6018                         conf->seq_flush++;
6019                         spin_unlock_irq(&conf->device_lock);
6020                         bitmap_unplug(mddev->bitmap);
6021                         spin_lock_irq(&conf->device_lock);
6022                         conf->seq_write = conf->seq_flush;
6023                         activate_bit_delay(conf, conf->temp_inactive_list);
6024                 }
6025                 raid5_activate_delayed(conf);
6026
6027                 while ((bio = remove_bio_from_retry(conf))) {
6028                         int ok;
6029                         spin_unlock_irq(&conf->device_lock);
6030                         ok = retry_aligned_read(conf, bio);
6031                         spin_lock_irq(&conf->device_lock);
6032                         if (!ok)
6033                                 break;
6034                         handled++;
6035                 }
6036
6037                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6038                                                    conf->temp_inactive_list);
6039                 if (!batch_size && !released)
6040                         break;
6041                 handled += batch_size;
6042
6043                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6044                         spin_unlock_irq(&conf->device_lock);
6045                         md_check_recovery(mddev);
6046                         spin_lock_irq(&conf->device_lock);
6047                 }
6048         }
6049         pr_debug("%d stripes handled\n", handled);
6050
6051         spin_unlock_irq(&conf->device_lock);
6052         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6053             mutex_trylock(&conf->cache_size_mutex)) {
6054                 grow_one_stripe(conf, __GFP_NOWARN);
6055                 /* Set flag even if allocation failed.  This helps
6056                  * slow down allocation requests when mem is short
6057                  */
6058                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6059                 mutex_unlock(&conf->cache_size_mutex);
6060         }
6061
6062         r5l_flush_stripe_to_raid(conf->log);
6063
6064         async_tx_issue_pending_all();
6065         blk_finish_plug(&plug);
6066
6067         pr_debug("--- raid5d inactive\n");
6068 }
6069
6070 static ssize_t
6071 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6072 {
6073         struct r5conf *conf;
6074         int ret = 0;
6075         spin_lock(&mddev->lock);
6076         conf = mddev->private;
6077         if (conf)
6078                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6079         spin_unlock(&mddev->lock);
6080         return ret;
6081 }
6082
6083 int
6084 raid5_set_cache_size(struct mddev *mddev, int size)
6085 {
6086         struct r5conf *conf = mddev->private;
6087         int err;
6088
6089         if (size <= 16 || size > 32768)
6090                 return -EINVAL;
6091
6092         conf->min_nr_stripes = size;
6093         mutex_lock(&conf->cache_size_mutex);
6094         while (size < conf->max_nr_stripes &&
6095                drop_one_stripe(conf))
6096                 ;
6097         mutex_unlock(&conf->cache_size_mutex);
6098
6099
6100         err = md_allow_write(mddev);
6101         if (err)
6102                 return err;
6103
6104         mutex_lock(&conf->cache_size_mutex);
6105         while (size > conf->max_nr_stripes)
6106                 if (!grow_one_stripe(conf, GFP_KERNEL))
6107                         break;
6108         mutex_unlock(&conf->cache_size_mutex);
6109
6110         return 0;
6111 }
6112 EXPORT_SYMBOL(raid5_set_cache_size);
6113
6114 static ssize_t
6115 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6116 {
6117         struct r5conf *conf;
6118         unsigned long new;
6119         int err;
6120
6121         if (len >= PAGE_SIZE)
6122                 return -EINVAL;
6123         if (kstrtoul(page, 10, &new))
6124                 return -EINVAL;
6125         err = mddev_lock(mddev);
6126         if (err)
6127                 return err;
6128         conf = mddev->private;
6129         if (!conf)
6130                 err = -ENODEV;
6131         else
6132                 err = raid5_set_cache_size(mddev, new);
6133         mddev_unlock(mddev);
6134
6135         return err ?: len;
6136 }
6137
6138 static struct md_sysfs_entry
6139 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6140                                 raid5_show_stripe_cache_size,
6141                                 raid5_store_stripe_cache_size);
6142
6143 static ssize_t
6144 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6145 {
6146         struct r5conf *conf = mddev->private;
6147         if (conf)
6148                 return sprintf(page, "%d\n", conf->rmw_level);
6149         else
6150                 return 0;
6151 }
6152
6153 static ssize_t
6154 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6155 {
6156         struct r5conf *conf = mddev->private;
6157         unsigned long new;
6158
6159         if (!conf)
6160                 return -ENODEV;
6161
6162         if (len >= PAGE_SIZE)
6163                 return -EINVAL;
6164
6165         if (kstrtoul(page, 10, &new))
6166                 return -EINVAL;
6167
6168         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6169                 return -EINVAL;
6170
6171         if (new != PARITY_DISABLE_RMW &&
6172             new != PARITY_ENABLE_RMW &&
6173             new != PARITY_PREFER_RMW)
6174                 return -EINVAL;
6175
6176         conf->rmw_level = new;
6177         return len;
6178 }
6179
6180 static struct md_sysfs_entry
6181 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6182                          raid5_show_rmw_level,
6183                          raid5_store_rmw_level);
6184
6185
6186 static ssize_t
6187 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6188 {
6189         struct r5conf *conf;
6190         int ret = 0;
6191         spin_lock(&mddev->lock);
6192         conf = mddev->private;
6193         if (conf)
6194                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6195         spin_unlock(&mddev->lock);
6196         return ret;
6197 }
6198
6199 static ssize_t
6200 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6201 {
6202         struct r5conf *conf;
6203         unsigned long new;
6204         int err;
6205
6206         if (len >= PAGE_SIZE)
6207                 return -EINVAL;
6208         if (kstrtoul(page, 10, &new))
6209                 return -EINVAL;
6210
6211         err = mddev_lock(mddev);
6212         if (err)
6213                 return err;
6214         conf = mddev->private;
6215         if (!conf)
6216                 err = -ENODEV;
6217         else if (new > conf->min_nr_stripes)
6218                 err = -EINVAL;
6219         else
6220                 conf->bypass_threshold = new;
6221         mddev_unlock(mddev);
6222         return err ?: len;
6223 }
6224
6225 static struct md_sysfs_entry
6226 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6227                                         S_IRUGO | S_IWUSR,
6228                                         raid5_show_preread_threshold,
6229                                         raid5_store_preread_threshold);
6230
6231 static ssize_t
6232 raid5_show_skip_copy(struct mddev *mddev, char *page)
6233 {
6234         struct r5conf *conf;
6235         int ret = 0;
6236         spin_lock(&mddev->lock);
6237         conf = mddev->private;
6238         if (conf)
6239                 ret = sprintf(page, "%d\n", conf->skip_copy);
6240         spin_unlock(&mddev->lock);
6241         return ret;
6242 }
6243
6244 static ssize_t
6245 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6246 {
6247         struct r5conf *conf;
6248         unsigned long new;
6249         int err;
6250
6251         if (len >= PAGE_SIZE)
6252                 return -EINVAL;
6253         if (kstrtoul(page, 10, &new))
6254                 return -EINVAL;
6255         new = !!new;
6256
6257         err = mddev_lock(mddev);
6258         if (err)
6259                 return err;
6260         conf = mddev->private;
6261         if (!conf)
6262                 err = -ENODEV;
6263         else if (new != conf->skip_copy) {
6264                 mddev_suspend(mddev);
6265                 conf->skip_copy = new;
6266                 if (new)
6267                         mddev->queue->backing_dev_info.capabilities |=
6268                                 BDI_CAP_STABLE_WRITES;
6269                 else
6270                         mddev->queue->backing_dev_info.capabilities &=
6271                                 ~BDI_CAP_STABLE_WRITES;
6272                 mddev_resume(mddev);
6273         }
6274         mddev_unlock(mddev);
6275         return err ?: len;
6276 }
6277
6278 static struct md_sysfs_entry
6279 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6280                                         raid5_show_skip_copy,
6281                                         raid5_store_skip_copy);
6282
6283 static ssize_t
6284 stripe_cache_active_show(struct mddev *mddev, char *page)
6285 {
6286         struct r5conf *conf = mddev->private;
6287         if (conf)
6288                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6289         else
6290                 return 0;
6291 }
6292
6293 static struct md_sysfs_entry
6294 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6295
6296 static ssize_t
6297 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6298 {
6299         struct r5conf *conf;
6300         int ret = 0;
6301         spin_lock(&mddev->lock);
6302         conf = mddev->private;
6303         if (conf)
6304                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6305         spin_unlock(&mddev->lock);
6306         return ret;
6307 }
6308
6309 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6310                                int *group_cnt,
6311                                int *worker_cnt_per_group,
6312                                struct r5worker_group **worker_groups);
6313 static ssize_t
6314 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6315 {
6316         struct r5conf *conf;
6317         unsigned long new;
6318         int err;
6319         struct r5worker_group *new_groups, *old_groups;
6320         int group_cnt, worker_cnt_per_group;
6321
6322         if (len >= PAGE_SIZE)
6323                 return -EINVAL;
6324         if (kstrtoul(page, 10, &new))
6325                 return -EINVAL;
6326
6327         err = mddev_lock(mddev);
6328         if (err)
6329                 return err;
6330         conf = mddev->private;
6331         if (!conf)
6332                 err = -ENODEV;
6333         else if (new != conf->worker_cnt_per_group) {
6334                 mddev_suspend(mddev);
6335
6336                 old_groups = conf->worker_groups;
6337                 if (old_groups)
6338                         flush_workqueue(raid5_wq);
6339
6340                 err = alloc_thread_groups(conf, new,
6341                                           &group_cnt, &worker_cnt_per_group,
6342                                           &new_groups);
6343                 if (!err) {
6344                         spin_lock_irq(&conf->device_lock);
6345                         conf->group_cnt = group_cnt;
6346                         conf->worker_cnt_per_group = worker_cnt_per_group;
6347                         conf->worker_groups = new_groups;
6348                         spin_unlock_irq(&conf->device_lock);
6349
6350                         if (old_groups)
6351                                 kfree(old_groups[0].workers);
6352                         kfree(old_groups);
6353                 }
6354                 mddev_resume(mddev);
6355         }
6356         mddev_unlock(mddev);
6357
6358         return err ?: len;
6359 }
6360
6361 static struct md_sysfs_entry
6362 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6363                                 raid5_show_group_thread_cnt,
6364                                 raid5_store_group_thread_cnt);
6365
6366 static struct attribute *raid5_attrs[] =  {
6367         &raid5_stripecache_size.attr,
6368         &raid5_stripecache_active.attr,
6369         &raid5_preread_bypass_threshold.attr,
6370         &raid5_group_thread_cnt.attr,
6371         &raid5_skip_copy.attr,
6372         &raid5_rmw_level.attr,
6373         &r5c_journal_mode.attr,
6374         NULL,
6375 };
6376 static struct attribute_group raid5_attrs_group = {
6377         .name = NULL,
6378         .attrs = raid5_attrs,
6379 };
6380
6381 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6382                                int *group_cnt,
6383                                int *worker_cnt_per_group,
6384                                struct r5worker_group **worker_groups)
6385 {
6386         int i, j, k;
6387         ssize_t size;
6388         struct r5worker *workers;
6389
6390         *worker_cnt_per_group = cnt;
6391         if (cnt == 0) {
6392                 *group_cnt = 0;
6393                 *worker_groups = NULL;
6394                 return 0;
6395         }
6396         *group_cnt = num_possible_nodes();
6397         size = sizeof(struct r5worker) * cnt;
6398         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6399         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6400                                 *group_cnt, GFP_NOIO);
6401         if (!*worker_groups || !workers) {
6402                 kfree(workers);
6403                 kfree(*worker_groups);
6404                 return -ENOMEM;
6405         }
6406
6407         for (i = 0; i < *group_cnt; i++) {
6408                 struct r5worker_group *group;
6409
6410                 group = &(*worker_groups)[i];
6411                 INIT_LIST_HEAD(&group->handle_list);
6412                 group->conf = conf;
6413                 group->workers = workers + i * cnt;
6414
6415                 for (j = 0; j < cnt; j++) {
6416                         struct r5worker *worker = group->workers + j;
6417                         worker->group = group;
6418                         INIT_WORK(&worker->work, raid5_do_work);
6419
6420                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6421                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6422                 }
6423         }
6424
6425         return 0;
6426 }
6427
6428 static void free_thread_groups(struct r5conf *conf)
6429 {
6430         if (conf->worker_groups)
6431                 kfree(conf->worker_groups[0].workers);
6432         kfree(conf->worker_groups);
6433         conf->worker_groups = NULL;
6434 }
6435
6436 static sector_t
6437 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6438 {
6439         struct r5conf *conf = mddev->private;
6440
6441         if (!sectors)
6442                 sectors = mddev->dev_sectors;
6443         if (!raid_disks)
6444                 /* size is defined by the smallest of previous and new size */
6445                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6446
6447         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6448         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6449         return sectors * (raid_disks - conf->max_degraded);
6450 }
6451
6452 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6453 {
6454         safe_put_page(percpu->spare_page);
6455         if (percpu->scribble)
6456                 flex_array_free(percpu->scribble);
6457         percpu->spare_page = NULL;
6458         percpu->scribble = NULL;
6459 }
6460
6461 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6462 {
6463         if (conf->level == 6 && !percpu->spare_page)
6464                 percpu->spare_page = alloc_page(GFP_KERNEL);
6465         if (!percpu->scribble)
6466                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6467                                                       conf->previous_raid_disks),
6468                                                   max(conf->chunk_sectors,
6469                                                       conf->prev_chunk_sectors)
6470                                                    / STRIPE_SECTORS,
6471                                                   GFP_KERNEL);
6472
6473         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6474                 free_scratch_buffer(conf, percpu);
6475                 return -ENOMEM;
6476         }
6477
6478         return 0;
6479 }
6480
6481 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6482 {
6483         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6484
6485         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6486         return 0;
6487 }
6488
6489 static void raid5_free_percpu(struct r5conf *conf)
6490 {
6491         if (!conf->percpu)
6492                 return;
6493
6494         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6495         free_percpu(conf->percpu);
6496 }
6497
6498 static void free_conf(struct r5conf *conf)
6499 {
6500         int i;
6501
6502         if (conf->log)
6503                 r5l_exit_log(conf->log);
6504         if (conf->shrinker.nr_deferred)
6505                 unregister_shrinker(&conf->shrinker);
6506
6507         free_thread_groups(conf);
6508         shrink_stripes(conf);
6509         raid5_free_percpu(conf);
6510         for (i = 0; i < conf->pool_size; i++)
6511                 if (conf->disks[i].extra_page)
6512                         put_page(conf->disks[i].extra_page);
6513         kfree(conf->disks);
6514         kfree(conf->stripe_hashtbl);
6515         kfree(conf);
6516 }
6517
6518 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6519 {
6520         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6521         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6522
6523         if (alloc_scratch_buffer(conf, percpu)) {
6524                 pr_warn("%s: failed memory allocation for cpu%u\n",
6525                         __func__, cpu);
6526                 return -ENOMEM;
6527         }
6528         return 0;
6529 }
6530
6531 static int raid5_alloc_percpu(struct r5conf *conf)
6532 {
6533         int err = 0;
6534
6535         conf->percpu = alloc_percpu(struct raid5_percpu);
6536         if (!conf->percpu)
6537                 return -ENOMEM;
6538
6539         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6540         if (!err) {
6541                 conf->scribble_disks = max(conf->raid_disks,
6542                         conf->previous_raid_disks);
6543                 conf->scribble_sectors = max(conf->chunk_sectors,
6544                         conf->prev_chunk_sectors);
6545         }
6546         return err;
6547 }
6548
6549 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6550                                       struct shrink_control *sc)
6551 {
6552         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6553         unsigned long ret = SHRINK_STOP;
6554
6555         if (mutex_trylock(&conf->cache_size_mutex)) {
6556                 ret= 0;
6557                 while (ret < sc->nr_to_scan &&
6558                        conf->max_nr_stripes > conf->min_nr_stripes) {
6559                         if (drop_one_stripe(conf) == 0) {
6560                                 ret = SHRINK_STOP;
6561                                 break;
6562                         }
6563                         ret++;
6564                 }
6565                 mutex_unlock(&conf->cache_size_mutex);
6566         }
6567         return ret;
6568 }
6569
6570 static unsigned long raid5_cache_count(struct shrinker *shrink,
6571                                        struct shrink_control *sc)
6572 {
6573         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6574
6575         if (conf->max_nr_stripes < conf->min_nr_stripes)
6576                 /* unlikely, but not impossible */
6577                 return 0;
6578         return conf->max_nr_stripes - conf->min_nr_stripes;
6579 }
6580
6581 static struct r5conf *setup_conf(struct mddev *mddev)
6582 {
6583         struct r5conf *conf;
6584         int raid_disk, memory, max_disks;
6585         struct md_rdev *rdev;
6586         struct disk_info *disk;
6587         char pers_name[6];
6588         int i;
6589         int group_cnt, worker_cnt_per_group;
6590         struct r5worker_group *new_group;
6591
6592         if (mddev->new_level != 5
6593             && mddev->new_level != 4
6594             && mddev->new_level != 6) {
6595                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6596                         mdname(mddev), mddev->new_level);
6597                 return ERR_PTR(-EIO);
6598         }
6599         if ((mddev->new_level == 5
6600              && !algorithm_valid_raid5(mddev->new_layout)) ||
6601             (mddev->new_level == 6
6602              && !algorithm_valid_raid6(mddev->new_layout))) {
6603                 pr_warn("md/raid:%s: layout %d not supported\n",
6604                         mdname(mddev), mddev->new_layout);
6605                 return ERR_PTR(-EIO);
6606         }
6607         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6608                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6609                         mdname(mddev), mddev->raid_disks);
6610                 return ERR_PTR(-EINVAL);
6611         }
6612
6613         if (!mddev->new_chunk_sectors ||
6614             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6615             !is_power_of_2(mddev->new_chunk_sectors)) {
6616                 pr_warn("md/raid:%s: invalid chunk size %d\n",
6617                         mdname(mddev), mddev->new_chunk_sectors << 9);
6618                 return ERR_PTR(-EINVAL);
6619         }
6620
6621         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6622         if (conf == NULL)
6623                 goto abort;
6624         /* Don't enable multi-threading by default*/
6625         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6626                                  &new_group)) {
6627                 conf->group_cnt = group_cnt;
6628                 conf->worker_cnt_per_group = worker_cnt_per_group;
6629                 conf->worker_groups = new_group;
6630         } else
6631                 goto abort;
6632         spin_lock_init(&conf->device_lock);
6633         seqcount_init(&conf->gen_lock);
6634         mutex_init(&conf->cache_size_mutex);
6635         init_waitqueue_head(&conf->wait_for_quiescent);
6636         init_waitqueue_head(&conf->wait_for_stripe);
6637         init_waitqueue_head(&conf->wait_for_overlap);
6638         INIT_LIST_HEAD(&conf->handle_list);
6639         INIT_LIST_HEAD(&conf->hold_list);
6640         INIT_LIST_HEAD(&conf->delayed_list);
6641         INIT_LIST_HEAD(&conf->bitmap_list);
6642         bio_list_init(&conf->return_bi);
6643         init_llist_head(&conf->released_stripes);
6644         atomic_set(&conf->active_stripes, 0);
6645         atomic_set(&conf->preread_active_stripes, 0);
6646         atomic_set(&conf->active_aligned_reads, 0);
6647         conf->bypass_threshold = BYPASS_THRESHOLD;
6648         conf->recovery_disabled = mddev->recovery_disabled - 1;
6649
6650         conf->raid_disks = mddev->raid_disks;
6651         if (mddev->reshape_position == MaxSector)
6652                 conf->previous_raid_disks = mddev->raid_disks;
6653         else
6654                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6655         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6656
6657         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6658                               GFP_KERNEL);
6659
6660         if (!conf->disks)
6661                 goto abort;
6662
6663         for (i = 0; i < max_disks; i++) {
6664                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6665                 if (!conf->disks[i].extra_page)
6666                         goto abort;
6667         }
6668
6669         conf->mddev = mddev;
6670
6671         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6672                 goto abort;
6673
6674         /* We init hash_locks[0] separately to that it can be used
6675          * as the reference lock in the spin_lock_nest_lock() call
6676          * in lock_all_device_hash_locks_irq in order to convince
6677          * lockdep that we know what we are doing.
6678          */
6679         spin_lock_init(conf->hash_locks);
6680         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6681                 spin_lock_init(conf->hash_locks + i);
6682
6683         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6684                 INIT_LIST_HEAD(conf->inactive_list + i);
6685
6686         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6687                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6688
6689         atomic_set(&conf->r5c_cached_full_stripes, 0);
6690         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6691         atomic_set(&conf->r5c_cached_partial_stripes, 0);
6692         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6693
6694         conf->level = mddev->new_level;
6695         conf->chunk_sectors = mddev->new_chunk_sectors;
6696         if (raid5_alloc_percpu(conf) != 0)
6697                 goto abort;
6698
6699         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6700
6701         rdev_for_each(rdev, mddev) {
6702                 raid_disk = rdev->raid_disk;
6703                 if (raid_disk >= max_disks
6704                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6705                         continue;
6706                 disk = conf->disks + raid_disk;
6707
6708                 if (test_bit(Replacement, &rdev->flags)) {
6709                         if (disk->replacement)
6710                                 goto abort;
6711                         disk->replacement = rdev;
6712                 } else {
6713                         if (disk->rdev)
6714                                 goto abort;
6715                         disk->rdev = rdev;
6716                 }
6717
6718                 if (test_bit(In_sync, &rdev->flags)) {
6719                         char b[BDEVNAME_SIZE];
6720                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
6721                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6722                 } else if (rdev->saved_raid_disk != raid_disk)
6723                         /* Cannot rely on bitmap to complete recovery */
6724                         conf->fullsync = 1;
6725         }
6726
6727         conf->level = mddev->new_level;
6728         if (conf->level == 6) {
6729                 conf->max_degraded = 2;
6730                 if (raid6_call.xor_syndrome)
6731                         conf->rmw_level = PARITY_ENABLE_RMW;
6732                 else
6733                         conf->rmw_level = PARITY_DISABLE_RMW;
6734         } else {
6735                 conf->max_degraded = 1;
6736                 conf->rmw_level = PARITY_ENABLE_RMW;
6737         }
6738         conf->algorithm = mddev->new_layout;
6739         conf->reshape_progress = mddev->reshape_position;
6740         if (conf->reshape_progress != MaxSector) {
6741                 conf->prev_chunk_sectors = mddev->chunk_sectors;
6742                 conf->prev_algo = mddev->layout;
6743         } else {
6744                 conf->prev_chunk_sectors = conf->chunk_sectors;
6745                 conf->prev_algo = conf->algorithm;
6746         }
6747
6748         conf->min_nr_stripes = NR_STRIPES;
6749         if (mddev->reshape_position != MaxSector) {
6750                 int stripes = max_t(int,
6751                         ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
6752                         ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
6753                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
6754                 if (conf->min_nr_stripes != NR_STRIPES)
6755                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
6756                                 mdname(mddev), conf->min_nr_stripes);
6757         }
6758         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6759                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6760         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6761         if (grow_stripes(conf, conf->min_nr_stripes)) {
6762                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
6763                         mdname(mddev), memory);
6764                 goto abort;
6765         } else
6766                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
6767         /*
6768          * Losing a stripe head costs more than the time to refill it,
6769          * it reduces the queue depth and so can hurt throughput.
6770          * So set it rather large, scaled by number of devices.
6771          */
6772         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6773         conf->shrinker.scan_objects = raid5_cache_scan;
6774         conf->shrinker.count_objects = raid5_cache_count;
6775         conf->shrinker.batch = 128;
6776         conf->shrinker.flags = 0;
6777         if (register_shrinker(&conf->shrinker)) {
6778                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
6779                         mdname(mddev));
6780                 goto abort;
6781         }
6782
6783         sprintf(pers_name, "raid%d", mddev->new_level);
6784         conf->thread = md_register_thread(raid5d, mddev, pers_name);
6785         if (!conf->thread) {
6786                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
6787                         mdname(mddev));
6788                 goto abort;
6789         }
6790
6791         return conf;
6792
6793  abort:
6794         if (conf) {
6795                 free_conf(conf);
6796                 return ERR_PTR(-EIO);
6797         } else
6798                 return ERR_PTR(-ENOMEM);
6799 }
6800
6801 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6802 {
6803         switch (algo) {
6804         case ALGORITHM_PARITY_0:
6805                 if (raid_disk < max_degraded)
6806                         return 1;
6807                 break;
6808         case ALGORITHM_PARITY_N:
6809                 if (raid_disk >= raid_disks - max_degraded)
6810                         return 1;
6811                 break;
6812         case ALGORITHM_PARITY_0_6:
6813                 if (raid_disk == 0 ||
6814                     raid_disk == raid_disks - 1)
6815                         return 1;
6816                 break;
6817         case ALGORITHM_LEFT_ASYMMETRIC_6:
6818         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6819         case ALGORITHM_LEFT_SYMMETRIC_6:
6820         case ALGORITHM_RIGHT_SYMMETRIC_6:
6821                 if (raid_disk == raid_disks - 1)
6822                         return 1;
6823         }
6824         return 0;
6825 }
6826
6827 static int raid5_run(struct mddev *mddev)
6828 {
6829         struct r5conf *conf;
6830         int working_disks = 0;
6831         int dirty_parity_disks = 0;
6832         struct md_rdev *rdev;
6833         struct md_rdev *journal_dev = NULL;
6834         sector_t reshape_offset = 0;
6835         int i;
6836         long long min_offset_diff = 0;
6837         int first = 1;
6838
6839         if (mddev->recovery_cp != MaxSector)
6840                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
6841                           mdname(mddev));
6842
6843         rdev_for_each(rdev, mddev) {
6844                 long long diff;
6845
6846                 if (test_bit(Journal, &rdev->flags)) {
6847                         journal_dev = rdev;
6848                         continue;
6849                 }
6850                 if (rdev->raid_disk < 0)
6851                         continue;
6852                 diff = (rdev->new_data_offset - rdev->data_offset);
6853                 if (first) {
6854                         min_offset_diff = diff;
6855                         first = 0;
6856                 } else if (mddev->reshape_backwards &&
6857                          diff < min_offset_diff)
6858                         min_offset_diff = diff;
6859                 else if (!mddev->reshape_backwards &&
6860                          diff > min_offset_diff)
6861                         min_offset_diff = diff;
6862         }
6863
6864         if (mddev->reshape_position != MaxSector) {
6865                 /* Check that we can continue the reshape.
6866                  * Difficulties arise if the stripe we would write to
6867                  * next is at or after the stripe we would read from next.
6868                  * For a reshape that changes the number of devices, this
6869                  * is only possible for a very short time, and mdadm makes
6870                  * sure that time appears to have past before assembling
6871                  * the array.  So we fail if that time hasn't passed.
6872                  * For a reshape that keeps the number of devices the same
6873                  * mdadm must be monitoring the reshape can keeping the
6874                  * critical areas read-only and backed up.  It will start
6875                  * the array in read-only mode, so we check for that.
6876                  */
6877                 sector_t here_new, here_old;
6878                 int old_disks;
6879                 int max_degraded = (mddev->level == 6 ? 2 : 1);
6880                 int chunk_sectors;
6881                 int new_data_disks;
6882
6883                 if (journal_dev) {
6884                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
6885                                 mdname(mddev));
6886                         return -EINVAL;
6887                 }
6888
6889                 if (mddev->new_level != mddev->level) {
6890                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
6891                                 mdname(mddev));
6892                         return -EINVAL;
6893                 }
6894                 old_disks = mddev->raid_disks - mddev->delta_disks;
6895                 /* reshape_position must be on a new-stripe boundary, and one
6896                  * further up in new geometry must map after here in old
6897                  * geometry.
6898                  * If the chunk sizes are different, then as we perform reshape
6899                  * in units of the largest of the two, reshape_position needs
6900                  * be a multiple of the largest chunk size times new data disks.
6901                  */
6902                 here_new = mddev->reshape_position;
6903                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6904                 new_data_disks = mddev->raid_disks - max_degraded;
6905                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6906                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
6907                                 mdname(mddev));
6908                         return -EINVAL;
6909                 }
6910                 reshape_offset = here_new * chunk_sectors;
6911                 /* here_new is the stripe we will write to */
6912                 here_old = mddev->reshape_position;
6913                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6914                 /* here_old is the first stripe that we might need to read
6915                  * from */
6916                 if (mddev->delta_disks == 0) {
6917                         /* We cannot be sure it is safe to start an in-place
6918                          * reshape.  It is only safe if user-space is monitoring
6919                          * and taking constant backups.
6920                          * mdadm always starts a situation like this in
6921                          * readonly mode so it can take control before
6922                          * allowing any writes.  So just check for that.
6923                          */
6924                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6925                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
6926                                 /* not really in-place - so OK */;
6927                         else if (mddev->ro == 0) {
6928                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
6929                                         mdname(mddev));
6930                                 return -EINVAL;
6931                         }
6932                 } else if (mddev->reshape_backwards
6933                     ? (here_new * chunk_sectors + min_offset_diff <=
6934                        here_old * chunk_sectors)
6935                     : (here_new * chunk_sectors >=
6936                        here_old * chunk_sectors + (-min_offset_diff))) {
6937                         /* Reading from the same stripe as writing to - bad */
6938                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
6939                                 mdname(mddev));
6940                         return -EINVAL;
6941                 }
6942                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
6943                 /* OK, we should be able to continue; */
6944         } else {
6945                 BUG_ON(mddev->level != mddev->new_level);
6946                 BUG_ON(mddev->layout != mddev->new_layout);
6947                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6948                 BUG_ON(mddev->delta_disks != 0);
6949         }
6950
6951         if (mddev->private == NULL)
6952                 conf = setup_conf(mddev);
6953         else
6954                 conf = mddev->private;
6955
6956         if (IS_ERR(conf))
6957                 return PTR_ERR(conf);
6958
6959         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
6960                 if (!journal_dev) {
6961                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
6962                                 mdname(mddev));
6963                         mddev->ro = 1;
6964                         set_disk_ro(mddev->gendisk, 1);
6965                 } else if (mddev->recovery_cp == MaxSector)
6966                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
6967         }
6968
6969         conf->min_offset_diff = min_offset_diff;
6970         mddev->thread = conf->thread;
6971         conf->thread = NULL;
6972         mddev->private = conf;
6973
6974         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6975              i++) {
6976                 rdev = conf->disks[i].rdev;
6977                 if (!rdev && conf->disks[i].replacement) {
6978                         /* The replacement is all we have yet */
6979                         rdev = conf->disks[i].replacement;
6980                         conf->disks[i].replacement = NULL;
6981                         clear_bit(Replacement, &rdev->flags);
6982                         conf->disks[i].rdev = rdev;
6983                 }
6984                 if (!rdev)
6985                         continue;
6986                 if (conf->disks[i].replacement &&
6987                     conf->reshape_progress != MaxSector) {
6988                         /* replacements and reshape simply do not mix. */
6989                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
6990                         goto abort;
6991                 }
6992                 if (test_bit(In_sync, &rdev->flags)) {
6993                         working_disks++;
6994                         continue;
6995                 }
6996                 /* This disc is not fully in-sync.  However if it
6997                  * just stored parity (beyond the recovery_offset),
6998                  * when we don't need to be concerned about the
6999                  * array being dirty.
7000                  * When reshape goes 'backwards', we never have
7001                  * partially completed devices, so we only need
7002                  * to worry about reshape going forwards.
7003                  */
7004                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7005                 if (mddev->major_version == 0 &&
7006                     mddev->minor_version > 90)
7007                         rdev->recovery_offset = reshape_offset;
7008
7009                 if (rdev->recovery_offset < reshape_offset) {
7010                         /* We need to check old and new layout */
7011                         if (!only_parity(rdev->raid_disk,
7012                                          conf->algorithm,
7013                                          conf->raid_disks,
7014                                          conf->max_degraded))
7015                                 continue;
7016                 }
7017                 if (!only_parity(rdev->raid_disk,
7018                                  conf->prev_algo,
7019                                  conf->previous_raid_disks,
7020                                  conf->max_degraded))
7021                         continue;
7022                 dirty_parity_disks++;
7023         }
7024
7025         /*
7026          * 0 for a fully functional array, 1 or 2 for a degraded array.
7027          */
7028         mddev->degraded = calc_degraded(conf);
7029
7030         if (has_failed(conf)) {
7031                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7032                         mdname(mddev), mddev->degraded, conf->raid_disks);
7033                 goto abort;
7034         }
7035
7036         /* device size must be a multiple of chunk size */
7037         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7038         mddev->resync_max_sectors = mddev->dev_sectors;
7039
7040         if (mddev->degraded > dirty_parity_disks &&
7041             mddev->recovery_cp != MaxSector) {
7042                 if (mddev->ok_start_degraded)
7043                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7044                                 mdname(mddev));
7045                 else {
7046                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7047                                 mdname(mddev));
7048                         goto abort;
7049                 }
7050         }
7051
7052         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7053                 mdname(mddev), conf->level,
7054                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7055                 mddev->new_layout);
7056
7057         print_raid5_conf(conf);
7058
7059         if (conf->reshape_progress != MaxSector) {
7060                 conf->reshape_safe = conf->reshape_progress;
7061                 atomic_set(&conf->reshape_stripes, 0);
7062                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7063                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7064                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7065                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7066                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7067                                                         "reshape");
7068         }
7069
7070         /* Ok, everything is just fine now */
7071         if (mddev->to_remove == &raid5_attrs_group)
7072                 mddev->to_remove = NULL;
7073         else if (mddev->kobj.sd &&
7074             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7075                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7076                         mdname(mddev));
7077         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7078
7079         if (mddev->queue) {
7080                 int chunk_size;
7081                 bool discard_supported = true;
7082                 /* read-ahead size must cover two whole stripes, which
7083                  * is 2 * (datadisks) * chunksize where 'n' is the
7084                  * number of raid devices
7085                  */
7086                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7087                 int stripe = data_disks *
7088                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7089                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7090                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7091
7092                 chunk_size = mddev->chunk_sectors << 9;
7093                 blk_queue_io_min(mddev->queue, chunk_size);
7094                 blk_queue_io_opt(mddev->queue, chunk_size *
7095                                  (conf->raid_disks - conf->max_degraded));
7096                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7097                 /*
7098                  * We can only discard a whole stripe. It doesn't make sense to
7099                  * discard data disk but write parity disk
7100                  */
7101                 stripe = stripe * PAGE_SIZE;
7102                 /* Round up to power of 2, as discard handling
7103                  * currently assumes that */
7104                 while ((stripe-1) & stripe)
7105                         stripe = (stripe | (stripe-1)) + 1;
7106                 mddev->queue->limits.discard_alignment = stripe;
7107                 mddev->queue->limits.discard_granularity = stripe;
7108
7109                 /*
7110                  * We use 16-bit counter of active stripes in bi_phys_segments
7111                  * (minus one for over-loaded initialization)
7112                  */
7113                 blk_queue_max_hw_sectors(mddev->queue, 0xfffe * STRIPE_SECTORS);
7114                 blk_queue_max_discard_sectors(mddev->queue,
7115                                               0xfffe * STRIPE_SECTORS);
7116
7117                 /*
7118                  * unaligned part of discard request will be ignored, so can't
7119                  * guarantee discard_zeroes_data
7120                  */
7121                 mddev->queue->limits.discard_zeroes_data = 0;
7122
7123                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7124
7125                 rdev_for_each(rdev, mddev) {
7126                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7127                                           rdev->data_offset << 9);
7128                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7129                                           rdev->new_data_offset << 9);
7130                         /*
7131                          * discard_zeroes_data is required, otherwise data
7132                          * could be lost. Consider a scenario: discard a stripe
7133                          * (the stripe could be inconsistent if
7134                          * discard_zeroes_data is 0); write one disk of the
7135                          * stripe (the stripe could be inconsistent again
7136                          * depending on which disks are used to calculate
7137                          * parity); the disk is broken; The stripe data of this
7138                          * disk is lost.
7139                          */
7140                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7141                             !bdev_get_queue(rdev->bdev)->
7142                                                 limits.discard_zeroes_data)
7143                                 discard_supported = false;
7144                         /* Unfortunately, discard_zeroes_data is not currently
7145                          * a guarantee - just a hint.  So we only allow DISCARD
7146                          * if the sysadmin has confirmed that only safe devices
7147                          * are in use by setting a module parameter.
7148                          */
7149                         if (!devices_handle_discard_safely) {
7150                                 if (discard_supported) {
7151                                         pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7152                                         pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7153                                 }
7154                                 discard_supported = false;
7155                         }
7156                 }
7157
7158                 if (discard_supported &&
7159                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7160                     mddev->queue->limits.discard_granularity >= stripe)
7161                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7162                                                 mddev->queue);
7163                 else
7164                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7165                                                 mddev->queue);
7166
7167                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7168         }
7169
7170         if (journal_dev) {
7171                 char b[BDEVNAME_SIZE];
7172
7173                 pr_debug("md/raid:%s: using device %s as journal\n",
7174                          mdname(mddev), bdevname(journal_dev->bdev, b));
7175                 if (r5l_init_log(conf, journal_dev))
7176                         goto abort;
7177         }
7178
7179         return 0;
7180 abort:
7181         md_unregister_thread(&mddev->thread);
7182         print_raid5_conf(conf);
7183         free_conf(conf);
7184         mddev->private = NULL;
7185         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7186         return -EIO;
7187 }
7188
7189 static void raid5_free(struct mddev *mddev, void *priv)
7190 {
7191         struct r5conf *conf = priv;
7192
7193         free_conf(conf);
7194         mddev->to_remove = &raid5_attrs_group;
7195 }
7196
7197 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7198 {
7199         struct r5conf *conf = mddev->private;
7200         int i;
7201
7202         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7203                 conf->chunk_sectors / 2, mddev->layout);
7204         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7205         rcu_read_lock();
7206         for (i = 0; i < conf->raid_disks; i++) {
7207                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7208                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7209         }
7210         rcu_read_unlock();
7211         seq_printf (seq, "]");
7212 }
7213
7214 static void print_raid5_conf (struct r5conf *conf)
7215 {
7216         int i;
7217         struct disk_info *tmp;
7218
7219         pr_debug("RAID conf printout:\n");
7220         if (!conf) {
7221                 pr_debug("(conf==NULL)\n");
7222                 return;
7223         }
7224         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7225                conf->raid_disks,
7226                conf->raid_disks - conf->mddev->degraded);
7227
7228         for (i = 0; i < conf->raid_disks; i++) {
7229                 char b[BDEVNAME_SIZE];
7230                 tmp = conf->disks + i;
7231                 if (tmp->rdev)
7232                         pr_debug(" disk %d, o:%d, dev:%s\n",
7233                                i, !test_bit(Faulty, &tmp->rdev->flags),
7234                                bdevname(tmp->rdev->bdev, b));
7235         }
7236 }
7237
7238 static int raid5_spare_active(struct mddev *mddev)
7239 {
7240         int i;
7241         struct r5conf *conf = mddev->private;
7242         struct disk_info *tmp;
7243         int count = 0;
7244         unsigned long flags;
7245
7246         for (i = 0; i < conf->raid_disks; i++) {
7247                 tmp = conf->disks + i;
7248                 if (tmp->replacement
7249                     && tmp->replacement->recovery_offset == MaxSector
7250                     && !test_bit(Faulty, &tmp->replacement->flags)
7251                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7252                         /* Replacement has just become active. */
7253                         if (!tmp->rdev
7254                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7255                                 count++;
7256                         if (tmp->rdev) {
7257                                 /* Replaced device not technically faulty,
7258                                  * but we need to be sure it gets removed
7259                                  * and never re-added.
7260                                  */
7261                                 set_bit(Faulty, &tmp->rdev->flags);
7262                                 sysfs_notify_dirent_safe(
7263                                         tmp->rdev->sysfs_state);
7264                         }
7265                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7266                 } else if (tmp->rdev
7267                     && tmp->rdev->recovery_offset == MaxSector
7268                     && !test_bit(Faulty, &tmp->rdev->flags)
7269                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7270                         count++;
7271                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7272                 }
7273         }
7274         spin_lock_irqsave(&conf->device_lock, flags);
7275         mddev->degraded = calc_degraded(conf);
7276         spin_unlock_irqrestore(&conf->device_lock, flags);
7277         print_raid5_conf(conf);
7278         return count;
7279 }
7280
7281 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7282 {
7283         struct r5conf *conf = mddev->private;
7284         int err = 0;
7285         int number = rdev->raid_disk;
7286         struct md_rdev **rdevp;
7287         struct disk_info *p = conf->disks + number;
7288
7289         print_raid5_conf(conf);
7290         if (test_bit(Journal, &rdev->flags) && conf->log) {
7291                 struct r5l_log *log;
7292                 /*
7293                  * we can't wait pending write here, as this is called in
7294                  * raid5d, wait will deadlock.
7295                  */
7296                 if (atomic_read(&mddev->writes_pending))
7297                         return -EBUSY;
7298                 log = conf->log;
7299                 conf->log = NULL;
7300                 synchronize_rcu();
7301                 r5l_exit_log(log);
7302                 return 0;
7303         }
7304         if (rdev == p->rdev)
7305                 rdevp = &p->rdev;
7306         else if (rdev == p->replacement)
7307                 rdevp = &p->replacement;
7308         else
7309                 return 0;
7310
7311         if (number >= conf->raid_disks &&
7312             conf->reshape_progress == MaxSector)
7313                 clear_bit(In_sync, &rdev->flags);
7314
7315         if (test_bit(In_sync, &rdev->flags) ||
7316             atomic_read(&rdev->nr_pending)) {
7317                 err = -EBUSY;
7318                 goto abort;
7319         }
7320         /* Only remove non-faulty devices if recovery
7321          * isn't possible.
7322          */
7323         if (!test_bit(Faulty, &rdev->flags) &&
7324             mddev->recovery_disabled != conf->recovery_disabled &&
7325             !has_failed(conf) &&
7326             (!p->replacement || p->replacement == rdev) &&
7327             number < conf->raid_disks) {
7328                 err = -EBUSY;
7329                 goto abort;
7330         }
7331         *rdevp = NULL;
7332         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7333                 synchronize_rcu();
7334                 if (atomic_read(&rdev->nr_pending)) {
7335                         /* lost the race, try later */
7336                         err = -EBUSY;
7337                         *rdevp = rdev;
7338                 }
7339         }
7340         if (p->replacement) {
7341                 /* We must have just cleared 'rdev' */
7342                 p->rdev = p->replacement;
7343                 clear_bit(Replacement, &p->replacement->flags);
7344                 smp_mb(); /* Make sure other CPUs may see both as identical
7345                            * but will never see neither - if they are careful
7346                            */
7347                 p->replacement = NULL;
7348                 clear_bit(WantReplacement, &rdev->flags);
7349         } else
7350                 /* We might have just removed the Replacement as faulty-
7351                  * clear the bit just in case
7352                  */
7353                 clear_bit(WantReplacement, &rdev->flags);
7354 abort:
7355
7356         print_raid5_conf(conf);
7357         return err;
7358 }
7359
7360 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7361 {
7362         struct r5conf *conf = mddev->private;
7363         int err = -EEXIST;
7364         int disk;
7365         struct disk_info *p;
7366         int first = 0;
7367         int last = conf->raid_disks - 1;
7368
7369         if (test_bit(Journal, &rdev->flags)) {
7370                 char b[BDEVNAME_SIZE];
7371                 if (conf->log)
7372                         return -EBUSY;
7373
7374                 rdev->raid_disk = 0;
7375                 /*
7376                  * The array is in readonly mode if journal is missing, so no
7377                  * write requests running. We should be safe
7378                  */
7379                 r5l_init_log(conf, rdev);
7380                 pr_debug("md/raid:%s: using device %s as journal\n",
7381                          mdname(mddev), bdevname(rdev->bdev, b));
7382                 return 0;
7383         }
7384         if (mddev->recovery_disabled == conf->recovery_disabled)
7385                 return -EBUSY;
7386
7387         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7388                 /* no point adding a device */
7389                 return -EINVAL;
7390
7391         if (rdev->raid_disk >= 0)
7392                 first = last = rdev->raid_disk;
7393
7394         /*
7395          * find the disk ... but prefer rdev->saved_raid_disk
7396          * if possible.
7397          */
7398         if (rdev->saved_raid_disk >= 0 &&
7399             rdev->saved_raid_disk >= first &&
7400             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7401                 first = rdev->saved_raid_disk;
7402
7403         for (disk = first; disk <= last; disk++) {
7404                 p = conf->disks + disk;
7405                 if (p->rdev == NULL) {
7406                         clear_bit(In_sync, &rdev->flags);
7407                         rdev->raid_disk = disk;
7408                         err = 0;
7409                         if (rdev->saved_raid_disk != disk)
7410                                 conf->fullsync = 1;
7411                         rcu_assign_pointer(p->rdev, rdev);
7412                         goto out;
7413                 }
7414         }
7415         for (disk = first; disk <= last; disk++) {
7416                 p = conf->disks + disk;
7417                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7418                     p->replacement == NULL) {
7419                         clear_bit(In_sync, &rdev->flags);
7420                         set_bit(Replacement, &rdev->flags);
7421                         rdev->raid_disk = disk;
7422                         err = 0;
7423                         conf->fullsync = 1;
7424                         rcu_assign_pointer(p->replacement, rdev);
7425                         break;
7426                 }
7427         }
7428 out:
7429         print_raid5_conf(conf);
7430         return err;
7431 }
7432
7433 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7434 {
7435         /* no resync is happening, and there is enough space
7436          * on all devices, so we can resize.
7437          * We need to make sure resync covers any new space.
7438          * If the array is shrinking we should possibly wait until
7439          * any io in the removed space completes, but it hardly seems
7440          * worth it.
7441          */
7442         sector_t newsize;
7443         struct r5conf *conf = mddev->private;
7444
7445         if (conf->log)
7446                 return -EINVAL;
7447         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7448         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7449         if (mddev->external_size &&
7450             mddev->array_sectors > newsize)
7451                 return -EINVAL;
7452         if (mddev->bitmap) {
7453                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7454                 if (ret)
7455                         return ret;
7456         }
7457         md_set_array_sectors(mddev, newsize);
7458         set_capacity(mddev->gendisk, mddev->array_sectors);
7459         revalidate_disk(mddev->gendisk);
7460         if (sectors > mddev->dev_sectors &&
7461             mddev->recovery_cp > mddev->dev_sectors) {
7462                 mddev->recovery_cp = mddev->dev_sectors;
7463                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7464         }
7465         mddev->dev_sectors = sectors;
7466         mddev->resync_max_sectors = sectors;
7467         return 0;
7468 }
7469
7470 static int check_stripe_cache(struct mddev *mddev)
7471 {
7472         /* Can only proceed if there are plenty of stripe_heads.
7473          * We need a minimum of one full stripe,, and for sensible progress
7474          * it is best to have about 4 times that.
7475          * If we require 4 times, then the default 256 4K stripe_heads will
7476          * allow for chunk sizes up to 256K, which is probably OK.
7477          * If the chunk size is greater, user-space should request more
7478          * stripe_heads first.
7479          */
7480         struct r5conf *conf = mddev->private;
7481         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7482             > conf->min_nr_stripes ||
7483             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7484             > conf->min_nr_stripes) {
7485                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7486                         mdname(mddev),
7487                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7488                          / STRIPE_SIZE)*4);
7489                 return 0;
7490         }
7491         return 1;
7492 }
7493
7494 static int check_reshape(struct mddev *mddev)
7495 {
7496         struct r5conf *conf = mddev->private;
7497
7498         if (conf->log)
7499                 return -EINVAL;
7500         if (mddev->delta_disks == 0 &&
7501             mddev->new_layout == mddev->layout &&
7502             mddev->new_chunk_sectors == mddev->chunk_sectors)
7503                 return 0; /* nothing to do */
7504         if (has_failed(conf))
7505                 return -EINVAL;
7506         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7507                 /* We might be able to shrink, but the devices must
7508                  * be made bigger first.
7509                  * For raid6, 4 is the minimum size.
7510                  * Otherwise 2 is the minimum
7511                  */
7512                 int min = 2;
7513                 if (mddev->level == 6)
7514                         min = 4;
7515                 if (mddev->raid_disks + mddev->delta_disks < min)
7516                         return -EINVAL;
7517         }
7518
7519         if (!check_stripe_cache(mddev))
7520                 return -ENOSPC;
7521
7522         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7523             mddev->delta_disks > 0)
7524                 if (resize_chunks(conf,
7525                                   conf->previous_raid_disks
7526                                   + max(0, mddev->delta_disks),
7527                                   max(mddev->new_chunk_sectors,
7528                                       mddev->chunk_sectors)
7529                             ) < 0)
7530                         return -ENOMEM;
7531         return resize_stripes(conf, (conf->previous_raid_disks
7532                                      + mddev->delta_disks));
7533 }
7534
7535 static int raid5_start_reshape(struct mddev *mddev)
7536 {
7537         struct r5conf *conf = mddev->private;
7538         struct md_rdev *rdev;
7539         int spares = 0;
7540         unsigned long flags;
7541
7542         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7543                 return -EBUSY;
7544
7545         if (!check_stripe_cache(mddev))
7546                 return -ENOSPC;
7547
7548         if (has_failed(conf))
7549                 return -EINVAL;
7550
7551         rdev_for_each(rdev, mddev) {
7552                 if (!test_bit(In_sync, &rdev->flags)
7553                     && !test_bit(Faulty, &rdev->flags))
7554                         spares++;
7555         }
7556
7557         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7558                 /* Not enough devices even to make a degraded array
7559                  * of that size
7560                  */
7561                 return -EINVAL;
7562
7563         /* Refuse to reduce size of the array.  Any reductions in
7564          * array size must be through explicit setting of array_size
7565          * attribute.
7566          */
7567         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7568             < mddev->array_sectors) {
7569                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7570                         mdname(mddev));
7571                 return -EINVAL;
7572         }
7573
7574         atomic_set(&conf->reshape_stripes, 0);
7575         spin_lock_irq(&conf->device_lock);
7576         write_seqcount_begin(&conf->gen_lock);
7577         conf->previous_raid_disks = conf->raid_disks;
7578         conf->raid_disks += mddev->delta_disks;
7579         conf->prev_chunk_sectors = conf->chunk_sectors;
7580         conf->chunk_sectors = mddev->new_chunk_sectors;
7581         conf->prev_algo = conf->algorithm;
7582         conf->algorithm = mddev->new_layout;
7583         conf->generation++;
7584         /* Code that selects data_offset needs to see the generation update
7585          * if reshape_progress has been set - so a memory barrier needed.
7586          */
7587         smp_mb();
7588         if (mddev->reshape_backwards)
7589                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7590         else
7591                 conf->reshape_progress = 0;
7592         conf->reshape_safe = conf->reshape_progress;
7593         write_seqcount_end(&conf->gen_lock);
7594         spin_unlock_irq(&conf->device_lock);
7595
7596         /* Now make sure any requests that proceeded on the assumption
7597          * the reshape wasn't running - like Discard or Read - have
7598          * completed.
7599          */
7600         mddev_suspend(mddev);
7601         mddev_resume(mddev);
7602
7603         /* Add some new drives, as many as will fit.
7604          * We know there are enough to make the newly sized array work.
7605          * Don't add devices if we are reducing the number of
7606          * devices in the array.  This is because it is not possible
7607          * to correctly record the "partially reconstructed" state of
7608          * such devices during the reshape and confusion could result.
7609          */
7610         if (mddev->delta_disks >= 0) {
7611                 rdev_for_each(rdev, mddev)
7612                         if (rdev->raid_disk < 0 &&
7613                             !test_bit(Faulty, &rdev->flags)) {
7614                                 if (raid5_add_disk(mddev, rdev) == 0) {
7615                                         if (rdev->raid_disk
7616                                             >= conf->previous_raid_disks)
7617                                                 set_bit(In_sync, &rdev->flags);
7618                                         else
7619                                                 rdev->recovery_offset = 0;
7620
7621                                         if (sysfs_link_rdev(mddev, rdev))
7622                                                 /* Failure here is OK */;
7623                                 }
7624                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7625                                    && !test_bit(Faulty, &rdev->flags)) {
7626                                 /* This is a spare that was manually added */
7627                                 set_bit(In_sync, &rdev->flags);
7628                         }
7629
7630                 /* When a reshape changes the number of devices,
7631                  * ->degraded is measured against the larger of the
7632                  * pre and post number of devices.
7633                  */
7634                 spin_lock_irqsave(&conf->device_lock, flags);
7635                 mddev->degraded = calc_degraded(conf);
7636                 spin_unlock_irqrestore(&conf->device_lock, flags);
7637         }
7638         mddev->raid_disks = conf->raid_disks;
7639         mddev->reshape_position = conf->reshape_progress;
7640         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7641
7642         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7643         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7644         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7645         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7646         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7647         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7648                                                 "reshape");
7649         if (!mddev->sync_thread) {
7650                 mddev->recovery = 0;
7651                 spin_lock_irq(&conf->device_lock);
7652                 write_seqcount_begin(&conf->gen_lock);
7653                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7654                 mddev->new_chunk_sectors =
7655                         conf->chunk_sectors = conf->prev_chunk_sectors;
7656                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7657                 rdev_for_each(rdev, mddev)
7658                         rdev->new_data_offset = rdev->data_offset;
7659                 smp_wmb();
7660                 conf->generation --;
7661                 conf->reshape_progress = MaxSector;
7662                 mddev->reshape_position = MaxSector;
7663                 write_seqcount_end(&conf->gen_lock);
7664                 spin_unlock_irq(&conf->device_lock);
7665                 return -EAGAIN;
7666         }
7667         conf->reshape_checkpoint = jiffies;
7668         md_wakeup_thread(mddev->sync_thread);
7669         md_new_event(mddev);
7670         return 0;
7671 }
7672
7673 /* This is called from the reshape thread and should make any
7674  * changes needed in 'conf'
7675  */
7676 static void end_reshape(struct r5conf *conf)
7677 {
7678
7679         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7680                 struct md_rdev *rdev;
7681
7682                 spin_lock_irq(&conf->device_lock);
7683                 conf->previous_raid_disks = conf->raid_disks;
7684                 rdev_for_each(rdev, conf->mddev)
7685                         rdev->data_offset = rdev->new_data_offset;
7686                 smp_wmb();
7687                 conf->reshape_progress = MaxSector;
7688                 conf->mddev->reshape_position = MaxSector;
7689                 spin_unlock_irq(&conf->device_lock);
7690                 wake_up(&conf->wait_for_overlap);
7691
7692                 /* read-ahead size must cover two whole stripes, which is
7693                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7694                  */
7695                 if (conf->mddev->queue) {
7696                         int data_disks = conf->raid_disks - conf->max_degraded;
7697                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7698                                                    / PAGE_SIZE);
7699                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7700                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7701                 }
7702         }
7703 }
7704
7705 /* This is called from the raid5d thread with mddev_lock held.
7706  * It makes config changes to the device.
7707  */
7708 static void raid5_finish_reshape(struct mddev *mddev)
7709 {
7710         struct r5conf *conf = mddev->private;
7711
7712         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7713
7714                 if (mddev->delta_disks > 0) {
7715                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7716                         if (mddev->queue) {
7717                                 set_capacity(mddev->gendisk, mddev->array_sectors);
7718                                 revalidate_disk(mddev->gendisk);
7719                         }
7720                 } else {
7721                         int d;
7722                         spin_lock_irq(&conf->device_lock);
7723                         mddev->degraded = calc_degraded(conf);
7724                         spin_unlock_irq(&conf->device_lock);
7725                         for (d = conf->raid_disks ;
7726                              d < conf->raid_disks - mddev->delta_disks;
7727                              d++) {
7728                                 struct md_rdev *rdev = conf->disks[d].rdev;
7729                                 if (rdev)
7730                                         clear_bit(In_sync, &rdev->flags);
7731                                 rdev = conf->disks[d].replacement;
7732                                 if (rdev)
7733                                         clear_bit(In_sync, &rdev->flags);
7734                         }
7735                 }
7736                 mddev->layout = conf->algorithm;
7737                 mddev->chunk_sectors = conf->chunk_sectors;
7738                 mddev->reshape_position = MaxSector;
7739                 mddev->delta_disks = 0;
7740                 mddev->reshape_backwards = 0;
7741         }
7742 }
7743
7744 static void raid5_quiesce(struct mddev *mddev, int state)
7745 {
7746         struct r5conf *conf = mddev->private;
7747
7748         switch(state) {
7749         case 2: /* resume for a suspend */
7750                 wake_up(&conf->wait_for_overlap);
7751                 break;
7752
7753         case 1: /* stop all writes */
7754                 lock_all_device_hash_locks_irq(conf);
7755                 /* '2' tells resync/reshape to pause so that all
7756                  * active stripes can drain
7757                  */
7758                 r5c_flush_cache(conf, INT_MAX);
7759                 conf->quiesce = 2;
7760                 wait_event_cmd(conf->wait_for_quiescent,
7761                                     atomic_read(&conf->active_stripes) == 0 &&
7762                                     atomic_read(&conf->active_aligned_reads) == 0,
7763                                     unlock_all_device_hash_locks_irq(conf),
7764                                     lock_all_device_hash_locks_irq(conf));
7765                 conf->quiesce = 1;
7766                 unlock_all_device_hash_locks_irq(conf);
7767                 /* allow reshape to continue */
7768                 wake_up(&conf->wait_for_overlap);
7769                 break;
7770
7771         case 0: /* re-enable writes */
7772                 lock_all_device_hash_locks_irq(conf);
7773                 conf->quiesce = 0;
7774                 wake_up(&conf->wait_for_quiescent);
7775                 wake_up(&conf->wait_for_overlap);
7776                 unlock_all_device_hash_locks_irq(conf);
7777                 break;
7778         }
7779         r5l_quiesce(conf->log, state);
7780 }
7781
7782 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7783 {
7784         struct r0conf *raid0_conf = mddev->private;
7785         sector_t sectors;
7786
7787         /* for raid0 takeover only one zone is supported */
7788         if (raid0_conf->nr_strip_zones > 1) {
7789                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7790                         mdname(mddev));
7791                 return ERR_PTR(-EINVAL);
7792         }
7793
7794         sectors = raid0_conf->strip_zone[0].zone_end;
7795         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7796         mddev->dev_sectors = sectors;
7797         mddev->new_level = level;
7798         mddev->new_layout = ALGORITHM_PARITY_N;
7799         mddev->new_chunk_sectors = mddev->chunk_sectors;
7800         mddev->raid_disks += 1;
7801         mddev->delta_disks = 1;
7802         /* make sure it will be not marked as dirty */
7803         mddev->recovery_cp = MaxSector;
7804
7805         return setup_conf(mddev);
7806 }
7807
7808 static void *raid5_takeover_raid1(struct mddev *mddev)
7809 {
7810         int chunksect;
7811         void *ret;
7812
7813         if (mddev->raid_disks != 2 ||
7814             mddev->degraded > 1)
7815                 return ERR_PTR(-EINVAL);
7816
7817         /* Should check if there are write-behind devices? */
7818
7819         chunksect = 64*2; /* 64K by default */
7820
7821         /* The array must be an exact multiple of chunksize */
7822         while (chunksect && (mddev->array_sectors & (chunksect-1)))
7823                 chunksect >>= 1;
7824
7825         if ((chunksect<<9) < STRIPE_SIZE)
7826                 /* array size does not allow a suitable chunk size */
7827                 return ERR_PTR(-EINVAL);
7828
7829         mddev->new_level = 5;
7830         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7831         mddev->new_chunk_sectors = chunksect;
7832
7833         ret = setup_conf(mddev);
7834         if (!IS_ERR(ret))
7835                 mddev_clear_unsupported_flags(mddev,
7836                         UNSUPPORTED_MDDEV_FLAGS);
7837         return ret;
7838 }
7839
7840 static void *raid5_takeover_raid6(struct mddev *mddev)
7841 {
7842         int new_layout;
7843
7844         switch (mddev->layout) {
7845         case ALGORITHM_LEFT_ASYMMETRIC_6:
7846                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7847                 break;
7848         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7849                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7850                 break;
7851         case ALGORITHM_LEFT_SYMMETRIC_6:
7852                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7853                 break;
7854         case ALGORITHM_RIGHT_SYMMETRIC_6:
7855                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7856                 break;
7857         case ALGORITHM_PARITY_0_6:
7858                 new_layout = ALGORITHM_PARITY_0;
7859                 break;
7860         case ALGORITHM_PARITY_N:
7861                 new_layout = ALGORITHM_PARITY_N;
7862                 break;
7863         default:
7864                 return ERR_PTR(-EINVAL);
7865         }
7866         mddev->new_level = 5;
7867         mddev->new_layout = new_layout;
7868         mddev->delta_disks = -1;
7869         mddev->raid_disks -= 1;
7870         return setup_conf(mddev);
7871 }
7872
7873 static int raid5_check_reshape(struct mddev *mddev)
7874 {
7875         /* For a 2-drive array, the layout and chunk size can be changed
7876          * immediately as not restriping is needed.
7877          * For larger arrays we record the new value - after validation
7878          * to be used by a reshape pass.
7879          */
7880         struct r5conf *conf = mddev->private;
7881         int new_chunk = mddev->new_chunk_sectors;
7882
7883         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7884                 return -EINVAL;
7885         if (new_chunk > 0) {
7886                 if (!is_power_of_2(new_chunk))
7887                         return -EINVAL;
7888                 if (new_chunk < (PAGE_SIZE>>9))
7889                         return -EINVAL;
7890                 if (mddev->array_sectors & (new_chunk-1))
7891                         /* not factor of array size */
7892                         return -EINVAL;
7893         }
7894
7895         /* They look valid */
7896
7897         if (mddev->raid_disks == 2) {
7898                 /* can make the change immediately */
7899                 if (mddev->new_layout >= 0) {
7900                         conf->algorithm = mddev->new_layout;
7901                         mddev->layout = mddev->new_layout;
7902                 }
7903                 if (new_chunk > 0) {
7904                         conf->chunk_sectors = new_chunk ;
7905                         mddev->chunk_sectors = new_chunk;
7906                 }
7907                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7908                 md_wakeup_thread(mddev->thread);
7909         }
7910         return check_reshape(mddev);
7911 }
7912
7913 static int raid6_check_reshape(struct mddev *mddev)
7914 {
7915         int new_chunk = mddev->new_chunk_sectors;
7916
7917         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7918                 return -EINVAL;
7919         if (new_chunk > 0) {
7920                 if (!is_power_of_2(new_chunk))
7921                         return -EINVAL;
7922                 if (new_chunk < (PAGE_SIZE >> 9))
7923                         return -EINVAL;
7924                 if (mddev->array_sectors & (new_chunk-1))
7925                         /* not factor of array size */
7926                         return -EINVAL;
7927         }
7928
7929         /* They look valid */
7930         return check_reshape(mddev);
7931 }
7932
7933 static void *raid5_takeover(struct mddev *mddev)
7934 {
7935         /* raid5 can take over:
7936          *  raid0 - if there is only one strip zone - make it a raid4 layout
7937          *  raid1 - if there are two drives.  We need to know the chunk size
7938          *  raid4 - trivial - just use a raid4 layout.
7939          *  raid6 - Providing it is a *_6 layout
7940          */
7941         if (mddev->level == 0)
7942                 return raid45_takeover_raid0(mddev, 5);
7943         if (mddev->level == 1)
7944                 return raid5_takeover_raid1(mddev);
7945         if (mddev->level == 4) {
7946                 mddev->new_layout = ALGORITHM_PARITY_N;
7947                 mddev->new_level = 5;
7948                 return setup_conf(mddev);
7949         }
7950         if (mddev->level == 6)
7951                 return raid5_takeover_raid6(mddev);
7952
7953         return ERR_PTR(-EINVAL);
7954 }
7955
7956 static void *raid4_takeover(struct mddev *mddev)
7957 {
7958         /* raid4 can take over:
7959          *  raid0 - if there is only one strip zone
7960          *  raid5 - if layout is right
7961          */
7962         if (mddev->level == 0)
7963                 return raid45_takeover_raid0(mddev, 4);
7964         if (mddev->level == 5 &&
7965             mddev->layout == ALGORITHM_PARITY_N) {
7966                 mddev->new_layout = 0;
7967                 mddev->new_level = 4;
7968                 return setup_conf(mddev);
7969         }
7970         return ERR_PTR(-EINVAL);
7971 }
7972
7973 static struct md_personality raid5_personality;
7974
7975 static void *raid6_takeover(struct mddev *mddev)
7976 {
7977         /* Currently can only take over a raid5.  We map the
7978          * personality to an equivalent raid6 personality
7979          * with the Q block at the end.
7980          */
7981         int new_layout;
7982
7983         if (mddev->pers != &raid5_personality)
7984                 return ERR_PTR(-EINVAL);
7985         if (mddev->degraded > 1)
7986                 return ERR_PTR(-EINVAL);
7987         if (mddev->raid_disks > 253)
7988                 return ERR_PTR(-EINVAL);
7989         if (mddev->raid_disks < 3)
7990                 return ERR_PTR(-EINVAL);
7991
7992         switch (mddev->layout) {
7993         case ALGORITHM_LEFT_ASYMMETRIC:
7994                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7995                 break;
7996         case ALGORITHM_RIGHT_ASYMMETRIC:
7997                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7998                 break;
7999         case ALGORITHM_LEFT_SYMMETRIC:
8000                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8001                 break;
8002         case ALGORITHM_RIGHT_SYMMETRIC:
8003                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8004                 break;
8005         case ALGORITHM_PARITY_0:
8006                 new_layout = ALGORITHM_PARITY_0_6;
8007                 break;
8008         case ALGORITHM_PARITY_N:
8009                 new_layout = ALGORITHM_PARITY_N;
8010                 break;
8011         default:
8012                 return ERR_PTR(-EINVAL);
8013         }
8014         mddev->new_level = 6;
8015         mddev->new_layout = new_layout;
8016         mddev->delta_disks = 1;
8017         mddev->raid_disks += 1;
8018         return setup_conf(mddev);
8019 }
8020
8021 static struct md_personality raid6_personality =
8022 {
8023         .name           = "raid6",
8024         .level          = 6,
8025         .owner          = THIS_MODULE,
8026         .make_request   = raid5_make_request,
8027         .run            = raid5_run,
8028         .free           = raid5_free,
8029         .status         = raid5_status,
8030         .error_handler  = raid5_error,
8031         .hot_add_disk   = raid5_add_disk,
8032         .hot_remove_disk= raid5_remove_disk,
8033         .spare_active   = raid5_spare_active,
8034         .sync_request   = raid5_sync_request,
8035         .resize         = raid5_resize,
8036         .size           = raid5_size,
8037         .check_reshape  = raid6_check_reshape,
8038         .start_reshape  = raid5_start_reshape,
8039         .finish_reshape = raid5_finish_reshape,
8040         .quiesce        = raid5_quiesce,
8041         .takeover       = raid6_takeover,
8042         .congested      = raid5_congested,
8043 };
8044 static struct md_personality raid5_personality =
8045 {
8046         .name           = "raid5",
8047         .level          = 5,
8048         .owner          = THIS_MODULE,
8049         .make_request   = raid5_make_request,
8050         .run            = raid5_run,
8051         .free           = raid5_free,
8052         .status         = raid5_status,
8053         .error_handler  = raid5_error,
8054         .hot_add_disk   = raid5_add_disk,
8055         .hot_remove_disk= raid5_remove_disk,
8056         .spare_active   = raid5_spare_active,
8057         .sync_request   = raid5_sync_request,
8058         .resize         = raid5_resize,
8059         .size           = raid5_size,
8060         .check_reshape  = raid5_check_reshape,
8061         .start_reshape  = raid5_start_reshape,
8062         .finish_reshape = raid5_finish_reshape,
8063         .quiesce        = raid5_quiesce,
8064         .takeover       = raid5_takeover,
8065         .congested      = raid5_congested,
8066 };
8067
8068 static struct md_personality raid4_personality =
8069 {
8070         .name           = "raid4",
8071         .level          = 4,
8072         .owner          = THIS_MODULE,
8073         .make_request   = raid5_make_request,
8074         .run            = raid5_run,
8075         .free           = raid5_free,
8076         .status         = raid5_status,
8077         .error_handler  = raid5_error,
8078         .hot_add_disk   = raid5_add_disk,
8079         .hot_remove_disk= raid5_remove_disk,
8080         .spare_active   = raid5_spare_active,
8081         .sync_request   = raid5_sync_request,
8082         .resize         = raid5_resize,
8083         .size           = raid5_size,
8084         .check_reshape  = raid5_check_reshape,
8085         .start_reshape  = raid5_start_reshape,
8086         .finish_reshape = raid5_finish_reshape,
8087         .quiesce        = raid5_quiesce,
8088         .takeover       = raid4_takeover,
8089         .congested      = raid5_congested,
8090 };
8091
8092 static int __init raid5_init(void)
8093 {
8094         int ret;
8095
8096         raid5_wq = alloc_workqueue("raid5wq",
8097                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8098         if (!raid5_wq)
8099                 return -ENOMEM;
8100
8101         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8102                                       "md/raid5:prepare",
8103                                       raid456_cpu_up_prepare,
8104                                       raid456_cpu_dead);
8105         if (ret) {
8106                 destroy_workqueue(raid5_wq);
8107                 return ret;
8108         }
8109         register_md_personality(&raid6_personality);
8110         register_md_personality(&raid5_personality);
8111         register_md_personality(&raid4_personality);
8112         return 0;
8113 }
8114
8115 static void raid5_exit(void)
8116 {
8117         unregister_md_personality(&raid6_personality);
8118         unregister_md_personality(&raid5_personality);
8119         unregister_md_personality(&raid4_personality);
8120         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8121         destroy_workqueue(raid5_wq);
8122 }
8123
8124 module_init(raid5_init);
8125 module_exit(raid5_exit);
8126 MODULE_LICENSE("GPL");
8127 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8128 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8129 MODULE_ALIAS("md-raid5");
8130 MODULE_ALIAS("md-raid4");
8131 MODULE_ALIAS("md-level-5");
8132 MODULE_ALIAS("md-level-4");
8133 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8134 MODULE_ALIAS("md-raid6");
8135 MODULE_ALIAS("md-level-6");
8136
8137 /* This used to be two separate modules, they were: */
8138 MODULE_ALIAS("raid5");
8139 MODULE_ALIAS("raid6");