kernel/drivers/md/raid5.c

   1 /*
   2  * raid5.c : Multiple Devices driver for Linux
   3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   4  *         Copyright (C) 1999, 2000 Ingo Molnar
   5  *         Copyright (C) 2002, 2003 H. Peter Anvin
   6  *
   7  * RAID-4/5/6 management functions.
   8  * Thanks to Penguin Computing for making the RAID-6 development possible
   9  * by donating a test server!
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 /*
  22  * BITMAP UNPLUGGING:
  23  *
  24  * The sequencing for updating the bitmap reliably is a little
  25  * subtle (and I got it wrong the first time) so it deserves some
  26  * explanation.
  27  *
  28  * We group bitmap updates into batches.  Each batch has a number.
  29  * We may write out several batches at once, but that isn't very important.
  30  * conf->seq_write is the number of the last batch successfully written.
  31  * conf->seq_flush is the number of the last batch that was closed to
  32  *    new additions.
  33  * When we discover that we will need to write to any block in a stripe
  34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  35  * the number of the batch it will be in. This is seq_flush+1.
  36  * When we are ready to do a write, if that batch hasn't been written yet,
  37  *   we plug the array and queue the stripe for later.
  38  * When an unplug happens, we increment bm_flush, thus closing the current
  39  *   batch.
  40  * When we notice that bm_flush > bm_write, we write out all pending updates
  41  * to the bitmap, and advance bm_write to where bm_flush was.
  42  * This may occasionally write a bit out twice, but is sure never to
  43  * miss any bits.
  44  */
  45
  46 #include <linux/blkdev.h>
  47 #include <linux/kthread.h>
  48 #include <linux/raid/pq.h>
  49 #include <linux/async_tx.h>
  50 #include <linux/module.h>
  51 #include <linux/async.h>
  52 #include <linux/seq_file.h>
  53 #include <linux/cpu.h>
  54 #include <linux/slab.h>
  55 #include <linux/ratelimit.h>
  56 #include <linux/nodemask.h>
  57 #include <linux/flex_array.h>
  58 #include <trace/events/block.h>
  59
  60 #include "md.h"
  61 #include "raid5.h"
  62 #include "raid0.h"
  63 #include "bitmap.h"
  64
  65 #define cpu_to_group(cpu) cpu_to_node(cpu)
  66 #define ANY_GROUP NUMA_NO_NODE
  67
  68 static bool devices_handle_discard_safely = false;
  69 module_param(devices_handle_discard_safely, bool, 0644);
  70 MODULE_PARM_DESC(devices_handle_discard_safely,
  71                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  72 static struct workqueue_struct *raid5_wq;
  73 /*
  74  * Stripe cache
  75  */
  76
  77 #define NR_STRIPES              256
  78 #define STRIPE_SIZE             PAGE_SIZE
  79 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  80 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  81 #define IO_THRESHOLD            1
  82 #define BYPASS_THRESHOLD        1
  83 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  84 #define HASH_MASK               (NR_HASH - 1)
  85 #define MAX_STRIPE_BATCH        8
  86
  87 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  88 {
  89         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  90         return &conf->stripe_hashtbl[hash];
  91 }
  92
  93 static inline int stripe_hash_locks_hash(sector_t sect)
  94 {
  95         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  96 }
  97
  98 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  99 {
 100         spin_lock_irq(conf->hash_locks + hash);
 101         spin_lock(&conf->device_lock);
 102 }
 103
 104 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
 105 {
 106         spin_unlock(&conf->device_lock);
 107         spin_unlock_irq(conf->hash_locks + hash);
 108 }
 109
 110 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 111 {
 112         int i;
 113         local_irq_disable();
 114         spin_lock(conf->hash_locks);
 115         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 116                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 117         spin_lock(&conf->device_lock);
 118 }
 119
 120 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 121 {
 122         int i;
 123         spin_unlock(&conf->device_lock);
 124         for (i = NR_STRIPE_HASH_LOCKS; i; i--)
 125                 spin_unlock(conf->hash_locks + i - 1);
 126         local_irq_enable();
 127 }
 128
 129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
 130  * order without overlap.  There may be several bio's per stripe+device, and
 131  * a bio could span several devices.
 132  * When walking this list for a particular stripe+device, we must never proceed
 133  * beyond a bio that extends past this device, as the next bio might no longer
 134  * be valid.
 135  * This function is used to determine the 'next' bio in the list, given the sector
 136  * of the current stripe+device
 137  */
 138 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
 139 {
 140         int sectors = bio_sectors(bio);
 141         if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
 142                 return bio->bi_next;
 143         else
 144                 return NULL;
 145 }
 146
 147 /*
 148  * We maintain a biased count of active stripes in the bottom 16 bits of
 149  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
 150  */
 151 static inline int raid5_bi_processed_stripes(struct bio *bio)
 152 {
 153         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 154         return (atomic_read(segments) >> 16) & 0xffff;
 155 }
 156
 157 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
 158 {
 159         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 160         return atomic_sub_return(1, segments) & 0xffff;
 161 }
 162
 163 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
 164 {
 165         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 166         atomic_inc(segments);
 167 }
 168
 169 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
 170         unsigned int cnt)
 171 {
 172         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 173         int old, new;
 174
 175         do {
 176                 old = atomic_read(segments);
 177                 new = (old & 0xffff) | (cnt << 16);
 178         } while (atomic_cmpxchg(segments, old, new) != old);
 179 }
 180
 181 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
 182 {
 183         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 184         atomic_set(segments, cnt);
 185 }
 186
 187 /* Find first data disk in a raid6 stripe */
 188 static inline int raid6_d0(struct stripe_head *sh)
 189 {
 190         if (sh->ddf_layout)
 191                 /* ddf always start from first device */
 192                 return 0;
 193         /* md starts just after Q block */
 194         if (sh->qd_idx == sh->disks - 1)
 195                 return 0;
 196         else
 197                 return sh->qd_idx + 1;
 198 }
 199 static inline int raid6_next_disk(int disk, int raid_disks)
 200 {
 201         disk++;
 202         return (disk < raid_disks) ? disk : 0;
 203 }
 204
 205 /* When walking through the disks in a raid5, starting at raid6_d0,
 206  * We need to map each disk to a 'slot', where the data disks are slot
 207  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 208  * is raid_disks-1.  This help does that mapping.
 209  */
 210 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 211                              int *count, int syndrome_disks)
 212 {
 213         int slot = *count;
 214
 215         if (sh->ddf_layout)
 216                 (*count)++;
 217         if (idx == sh->pd_idx)
 218                 return syndrome_disks;
 219         if (idx == sh->qd_idx)
 220                 return syndrome_disks + 1;
 221         if (!sh->ddf_layout)
 222                 (*count)++;
 223         return slot;
 224 }
 225
 226 static void return_io(struct bio_list *return_bi)
 227 {
 228         struct bio *bi;
 229         while ((bi = bio_list_pop(return_bi)) != NULL) {
 230                 bi->bi_iter.bi_size = 0;
 231                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 232                                          bi, 0);
 233                 bio_endio(bi);
 234         }
 235 }
 236
 237 static void print_raid5_conf (struct r5conf *conf);
 238
 239 static int stripe_operations_active(struct stripe_head *sh)
 240 {
 241         return sh->check_state || sh->reconstruct_state ||
 242                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 243                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 244 }
 245
 246 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 247 {
 248         struct r5conf *conf = sh->raid_conf;
 249         struct r5worker_group *group;
 250         int thread_cnt;
 251         int i, cpu = sh->cpu;
 252
 253         if (!cpu_online(cpu)) {
 254                 cpu = cpumask_any(cpu_online_mask);
 255                 sh->cpu = cpu;
 256         }
 257
 258         if (list_empty(&sh->lru)) {
 259                 struct r5worker_group *group;
 260                 group = conf->worker_groups + cpu_to_group(cpu);
 261                 list_add_tail(&sh->lru, &group->handle_list);
 262                 group->stripes_cnt++;
 263                 sh->group = group;
 264         }
 265
 266         if (conf->worker_cnt_per_group == 0) {
 267                 md_wakeup_thread(conf->mddev->thread);
 268                 return;
 269         }
 270
 271         group = conf->worker_groups + cpu_to_group(sh->cpu);
 272
 273         group->workers[0].working = true;
 274         /* at least one worker should run to avoid race */
 275         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 276
 277         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 278         /* wakeup more workers */
 279         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 280                 if (group->workers[i].working == false) {
 281                         group->workers[i].working = true;
 282                         queue_work_on(sh->cpu, raid5_wq,
 283                                       &group->workers[i].work);
 284                         thread_cnt--;
 285                 }
 286         }
 287 }
 288
 289 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 290                               struct list_head *temp_inactive_list)
 291 {
 292         BUG_ON(!list_empty(&sh->lru));
 293         BUG_ON(atomic_read(&conf->active_stripes)==0);
 294         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 295                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 296                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 297                         list_add_tail(&sh->lru, &conf->delayed_list);
 298                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 299                            sh->bm_seq - conf->seq_write > 0)
 300                         list_add_tail(&sh->lru, &conf->bitmap_list);
 301                 else {
 302                         clear_bit(STRIPE_DELAYED, &sh->state);
 303                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 304                         if (conf->worker_cnt_per_group == 0) {
 305                                 list_add_tail(&sh->lru, &conf->handle_list);
 306                         } else {
 307                                 raid5_wakeup_stripe_thread(sh);
 308                                 return;
 309                         }
 310                 }
 311                 md_wakeup_thread(conf->mddev->thread);
 312         } else {
 313                 BUG_ON(stripe_operations_active(sh));
 314                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 315                         if (atomic_dec_return(&conf->preread_active_stripes)
 316                             < IO_THRESHOLD)
 317                                 md_wakeup_thread(conf->mddev->thread);
 318                 atomic_dec(&conf->active_stripes);
 319                 if (!test_bit(STRIPE_EXPANDING, &sh->state))
 320                         list_add_tail(&sh->lru, temp_inactive_list);
 321         }
 322 }
 323
 324 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 325                              struct list_head *temp_inactive_list)
 326 {
 327         if (atomic_dec_and_test(&sh->count))
 328                 do_release_stripe(conf, sh, temp_inactive_list);
 329 }
 330
 331 /*
 332  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 333  *
 334  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 335  * given time. Adding stripes only takes device lock, while deleting stripes
 336  * only takes hash lock.
 337  */
 338 static void release_inactive_stripe_list(struct r5conf *conf,
 339                                          struct list_head *temp_inactive_list,
 340                                          int hash)
 341 {
 342         int size;
 343         unsigned long do_wakeup = 0;
 344         int i = 0;
 345         unsigned long flags;
 346
 347         if (hash == NR_STRIPE_HASH_LOCKS) {
 348                 size = NR_STRIPE_HASH_LOCKS;
 349                 hash = NR_STRIPE_HASH_LOCKS - 1;
 350         } else
 351                 size = 1;
 352         while (size) {
 353                 struct list_head *list = &temp_inactive_list[size - 1];
 354
 355                 /*
 356                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 357                  * remove stripes from the list
 358                  */
 359                 if (!list_empty_careful(list)) {
 360                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 361                         if (list_empty(conf->inactive_list + hash) &&
 362                             !list_empty(list))
 363                                 atomic_dec(&conf->empty_inactive_list_nr);
 364                         list_splice_tail_init(list, conf->inactive_list + hash);
 365                         do_wakeup |= 1 << hash;
 366                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 367                 }
 368                 size--;
 369                 hash--;
 370         }
 371
 372         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
 373                 if (do_wakeup & (1 << i))
 374                         wake_up(&conf->wait_for_stripe[i]);
 375         }
 376
 377         if (do_wakeup) {
 378                 if (atomic_read(&conf->active_stripes) == 0)
 379                         wake_up(&conf->wait_for_quiescent);
 380                 if (conf->retry_read_aligned)
 381                         md_wakeup_thread(conf->mddev->thread);
 382         }
 383 }
 384
 385 /* should hold conf->device_lock already */
 386 static int release_stripe_list(struct r5conf *conf,
 387                                struct list_head *temp_inactive_list)
 388 {
 389         struct stripe_head *sh;
 390         int count = 0;
 391         struct llist_node *head;
 392
 393         head = llist_del_all(&conf->released_stripes);
 394         head = llist_reverse_order(head);
 395         while (head) {
 396                 int hash;
 397
 398                 sh = llist_entry(head, struct stripe_head, release_list);
 399                 head = llist_next(head);
 400                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 401                 smp_mb();
 402                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 403                 /*
 404                  * Don't worry the bit is set here, because if the bit is set
 405                  * again, the count is always > 1. This is true for
 406                  * STRIPE_ON_UNPLUG_LIST bit too.
 407                  */
 408                 hash = sh->hash_lock_index;
 409                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 410                 count++;
 411         }
 412
 413         return count;
 414 }
 415
 416 void raid5_release_stripe(struct stripe_head *sh)
 417 {
 418         struct r5conf *conf = sh->raid_conf;
 419         unsigned long flags;
 420         struct list_head list;
 421         int hash;
 422         bool wakeup;
 423
 424         /* Avoid release_list until the last reference.
 425          */
 426         if (atomic_add_unless(&sh->count, -1, 1))
 427                 return;
 428
 429         if (unlikely(!conf->mddev->thread) ||
 430                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 431                 goto slow_path;
 432         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 433         if (wakeup)
 434                 md_wakeup_thread(conf->mddev->thread);
 435         return;
 436 slow_path:
 437         local_irq_save(flags);
 438         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 439         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 440                 INIT_LIST_HEAD(&list);
 441                 hash = sh->hash_lock_index;
 442                 do_release_stripe(conf, sh, &list);
 443                 spin_unlock(&conf->device_lock);
 444                 release_inactive_stripe_list(conf, &list, hash);
 445         }
 446         local_irq_restore(flags);
 447 }
 448
 449 static inline void remove_hash(struct stripe_head *sh)
 450 {
 451         pr_debug("remove_hash(), stripe %llu\n",
 452                 (unsigned long long)sh->sector);
 453
 454         hlist_del_init(&sh->hash);
 455 }
 456
 457 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 458 {
 459         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 460
 461         pr_debug("insert_hash(), stripe %llu\n",
 462                 (unsigned long long)sh->sector);
 463
 464         hlist_add_head(&sh->hash, hp);
 465 }
 466
 467 /* find an idle stripe, make sure it is unhashed, and return it. */
 468 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 469 {
 470         struct stripe_head *sh = NULL;
 471         struct list_head *first;
 472
 473         if (list_empty(conf->inactive_list + hash))
 474                 goto out;
 475         first = (conf->inactive_list + hash)->next;
 476         sh = list_entry(first, struct stripe_head, lru);
 477         list_del_init(first);
 478         remove_hash(sh);
 479         atomic_inc(&conf->active_stripes);
 480         BUG_ON(hash != sh->hash_lock_index);
 481         if (list_empty(conf->inactive_list + hash))
 482                 atomic_inc(&conf->empty_inactive_list_nr);
 483 out:
 484         return sh;
 485 }
 486
 487 static void shrink_buffers(struct stripe_head *sh)
 488 {
 489         struct page *p;
 490         int i;
 491         int num = sh->raid_conf->pool_size;
 492
 493         for (i = 0; i < num ; i++) {
 494                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 495                 p = sh->dev[i].page;
 496                 if (!p)
 497                         continue;
 498                 sh->dev[i].page = NULL;
 499                 put_page(p);
 500         }
 501 }
 502
 503 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 504 {
 505         int i;
 506         int num = sh->raid_conf->pool_size;
 507
 508         for (i = 0; i < num; i++) {
 509                 struct page *page;
 510
 511                 if (!(page = alloc_page(gfp))) {
 512                         return 1;
 513                 }
 514                 sh->dev[i].page = page;
 515                 sh->dev[i].orig_page = page;
 516         }
 517         return 0;
 518 }
 519
 520 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 521 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 522                             struct stripe_head *sh);
 523
 524 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 525 {
 526         struct r5conf *conf = sh->raid_conf;
 527         int i, seq;
 528
 529         BUG_ON(atomic_read(&sh->count) != 0);
 530         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 531         BUG_ON(stripe_operations_active(sh));
 532         BUG_ON(sh->batch_head);
 533
 534         pr_debug("init_stripe called, stripe %llu\n",
 535                 (unsigned long long)sector);
 536 retry:
 537         seq = read_seqcount_begin(&conf->gen_lock);
 538         sh->generation = conf->generation - previous;
 539         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 540         sh->sector = sector;
 541         stripe_set_idx(sector, conf, previous, sh);
 542         sh->state = 0;
 543
 544         for (i = sh->disks; i--; ) {
 545                 struct r5dev *dev = &sh->dev[i];
 546
 547                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 548                     test_bit(R5_LOCKED, &dev->flags)) {
 549                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 550                                (unsigned long long)sh->sector, i, dev->toread,
 551                                dev->read, dev->towrite, dev->written,
 552                                test_bit(R5_LOCKED, &dev->flags));
 553                         WARN_ON(1);
 554                 }
 555                 dev->flags = 0;
 556                 raid5_build_block(sh, i, previous);
 557         }
 558         if (read_seqcount_retry(&conf->gen_lock, seq))
 559                 goto retry;
 560         sh->overwrite_disks = 0;
 561         insert_hash(conf, sh);
 562         sh->cpu = smp_processor_id();
 563         set_bit(STRIPE_BATCH_READY, &sh->state);
 564 }
 565
 566 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 567                                          short generation)
 568 {
 569         struct stripe_head *sh;
 570
 571         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 572         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 573                 if (sh->sector == sector && sh->generation == generation)
 574                         return sh;
 575         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 576         return NULL;
 577 }
 578
 579 /*
 580  * Need to check if array has failed when deciding whether to:
 581  *  - start an array
 582  *  - remove non-faulty devices
 583  *  - add a spare
 584  *  - allow a reshape
 585  * This determination is simple when no reshape is happening.
 586  * However if there is a reshape, we need to carefully check
 587  * both the before and after sections.
 588  * This is because some failed devices may only affect one
 589  * of the two sections, and some non-in_sync devices may
 590  * be insync in the section most affected by failed devices.
 591  */
 592 static int calc_degraded(struct r5conf *conf)
 593 {
 594         int degraded, degraded2;
 595         int i;
 596
 597         rcu_read_lock();
 598         degraded = 0;
 599         for (i = 0; i < conf->previous_raid_disks; i++) {
 600                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 601                 if (rdev && test_bit(Faulty, &rdev->flags))
 602                         rdev = rcu_dereference(conf->disks[i].replacement);
 603                 if (!rdev || test_bit(Faulty, &rdev->flags))
 604                         degraded++;
 605                 else if (test_bit(In_sync, &rdev->flags))
 606                         ;
 607                 else
 608                         /* not in-sync or faulty.
 609                          * If the reshape increases the number of devices,
 610                          * this is being recovered by the reshape, so
 611                          * this 'previous' section is not in_sync.
 612                          * If the number of devices is being reduced however,
 613                          * the device can only be part of the array if
 614                          * we are reverting a reshape, so this section will
 615                          * be in-sync.
 616                          */
 617                         if (conf->raid_disks >= conf->previous_raid_disks)
 618                                 degraded++;
 619         }
 620         rcu_read_unlock();
 621         if (conf->raid_disks == conf->previous_raid_disks)
 622                 return degraded;
 623         rcu_read_lock();
 624         degraded2 = 0;
 625         for (i = 0; i < conf->raid_disks; i++) {
 626                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 627                 if (rdev && test_bit(Faulty, &rdev->flags))
 628                         rdev = rcu_dereference(conf->disks[i].replacement);
 629                 if (!rdev || test_bit(Faulty, &rdev->flags))
 630                         degraded2++;
 631                 else if (test_bit(In_sync, &rdev->flags))
 632                         ;
 633                 else
 634                         /* not in-sync or faulty.
 635                          * If reshape increases the number of devices, this
 636                          * section has already been recovered, else it
 637                          * almost certainly hasn't.
 638                          */
 639                         if (conf->raid_disks <= conf->previous_raid_disks)
 640                                 degraded2++;
 641         }
 642         rcu_read_unlock();
 643         if (degraded2 > degraded)
 644                 return degraded2;
 645         return degraded;
 646 }
 647
 648 static int has_failed(struct r5conf *conf)
 649 {
 650         int degraded;
 651
 652         if (conf->mddev->reshape_position == MaxSector)
 653                 return conf->mddev->degraded > conf->max_degraded;
 654
 655         degraded = calc_degraded(conf);
 656         if (degraded > conf->max_degraded)
 657                 return 1;
 658         return 0;
 659 }
 660
 661 struct stripe_head *
 662 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 663                         int previous, int noblock, int noquiesce)
 664 {
 665         struct stripe_head *sh;
 666         int hash = stripe_hash_locks_hash(sector);
 667
 668         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 669
 670         spin_lock_irq(conf->hash_locks + hash);
 671
 672         do {
 673                 wait_event_lock_irq(conf->wait_for_quiescent,
 674                                     conf->quiesce == 0 || noquiesce,
 675                                     *(conf->hash_locks + hash));
 676                 sh = __find_stripe(conf, sector, conf->generation - previous);
 677                 if (!sh) {
 678                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 679                                 sh = get_free_stripe(conf, hash);
 680                                 if (!sh && !test_bit(R5_DID_ALLOC,
 681                                                      &conf->cache_state))
 682                                         set_bit(R5_ALLOC_MORE,
 683                                                 &conf->cache_state);
 684                         }
 685                         if (noblock && sh == NULL)
 686                                 break;
 687                         if (!sh) {
 688                                 set_bit(R5_INACTIVE_BLOCKED,
 689                                         &conf->cache_state);
 690                                 wait_event_exclusive_cmd(
 691                                         conf->wait_for_stripe[hash],
 692                                         !list_empty(conf->inactive_list + hash) &&
 693                                         (atomic_read(&conf->active_stripes)
 694                                          < (conf->max_nr_stripes * 3 / 4)
 695                                          || !test_bit(R5_INACTIVE_BLOCKED,
 696                                                       &conf->cache_state)),
 697                                         spin_unlock_irq(conf->hash_locks + hash),
 698                                         spin_lock_irq(conf->hash_locks + hash));
 699                                 clear_bit(R5_INACTIVE_BLOCKED,
 700                                           &conf->cache_state);
 701                         } else {
 702                                 init_stripe(sh, sector, previous);
 703                                 atomic_inc(&sh->count);
 704                         }
 705                 } else if (!atomic_inc_not_zero(&sh->count)) {
 706                         spin_lock(&conf->device_lock);
 707                         if (!atomic_read(&sh->count)) {
 708                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 709                                         atomic_inc(&conf->active_stripes);
 710                                 BUG_ON(list_empty(&sh->lru) &&
 711                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 712                                 list_del_init(&sh->lru);
 713                                 if (sh->group) {
 714                                         sh->group->stripes_cnt--;
 715                                         sh->group = NULL;
 716                                 }
 717                         }
 718                         atomic_inc(&sh->count);
 719                         spin_unlock(&conf->device_lock);
 720                 }
 721         } while (sh == NULL);
 722
 723         if (!list_empty(conf->inactive_list + hash))
 724                 wake_up(&conf->wait_for_stripe[hash]);
 725
 726         spin_unlock_irq(conf->hash_locks + hash);
 727         return sh;
 728 }
 729
 730 static bool is_full_stripe_write(struct stripe_head *sh)
 731 {
 732         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 733         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 734 }
 735
 736 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 737 {
 738         local_irq_disable();
 739         if (sh1 > sh2) {
 740                 spin_lock(&sh2->stripe_lock);
 741                 spin_lock_nested(&sh1->stripe_lock, 1);
 742         } else {
 743                 spin_lock(&sh1->stripe_lock);
 744                 spin_lock_nested(&sh2->stripe_lock, 1);
 745         }
 746 }
 747
 748 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 749 {
 750         spin_unlock(&sh1->stripe_lock);
 751         spin_unlock(&sh2->stripe_lock);
 752         local_irq_enable();
 753 }
 754
 755 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 756 static bool stripe_can_batch(struct stripe_head *sh)
 757 {
 758         struct r5conf *conf = sh->raid_conf;
 759
 760         if (conf->log)
 761                 return false;
 762         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 763                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 764                 is_full_stripe_write(sh);
 765 }
 766
 767 /* we only do back search */
 768 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 769 {
 770         struct stripe_head *head;
 771         sector_t head_sector, tmp_sec;
 772         int hash;
 773         int dd_idx;
 774
 775         if (!stripe_can_batch(sh))
 776                 return;
 777         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 778         tmp_sec = sh->sector;
 779         if (!sector_div(tmp_sec, conf->chunk_sectors))
 780                 return;
 781         head_sector = sh->sector - STRIPE_SECTORS;
 782
 783         hash = stripe_hash_locks_hash(head_sector);
 784         spin_lock_irq(conf->hash_locks + hash);
 785         head = __find_stripe(conf, head_sector, conf->generation);
 786         if (head && !atomic_inc_not_zero(&head->count)) {
 787                 spin_lock(&conf->device_lock);
 788                 if (!atomic_read(&head->count)) {
 789                         if (!test_bit(STRIPE_HANDLE, &head->state))
 790                                 atomic_inc(&conf->active_stripes);
 791                         BUG_ON(list_empty(&head->lru) &&
 792                                !test_bit(STRIPE_EXPANDING, &head->state));
 793                         list_del_init(&head->lru);
 794                         if (head->group) {
 795                                 head->group->stripes_cnt--;
 796                                 head->group = NULL;
 797                         }
 798                 }
 799                 atomic_inc(&head->count);
 800                 spin_unlock(&conf->device_lock);
 801         }
 802         spin_unlock_irq(conf->hash_locks + hash);
 803
 804         if (!head)
 805                 return;
 806         if (!stripe_can_batch(head))
 807                 goto out;
 808
 809         lock_two_stripes(head, sh);
 810         /* clear_batch_ready clear the flag */
 811         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 812                 goto unlock_out;
 813
 814         if (sh->batch_head)
 815                 goto unlock_out;
 816
 817         dd_idx = 0;
 818         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 819                 dd_idx++;
 820         if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
 821                 goto unlock_out;
 822
 823         if (head->batch_head) {
 824                 spin_lock(&head->batch_head->batch_lock);
 825                 /* This batch list is already running */
 826                 if (!stripe_can_batch(head)) {
 827                         spin_unlock(&head->batch_head->batch_lock);
 828                         goto unlock_out;
 829                 }
 830
 831                 /*
 832                  * at this point, head's BATCH_READY could be cleared, but we
 833                  * can still add the stripe to batch list
 834                  */
 835                 list_add(&sh->batch_list, &head->batch_list);
 836                 spin_unlock(&head->batch_head->batch_lock);
 837
 838                 sh->batch_head = head->batch_head;
 839         } else {
 840                 head->batch_head = head;
 841                 sh->batch_head = head->batch_head;
 842                 spin_lock(&head->batch_lock);
 843                 list_add_tail(&sh->batch_list, &head->batch_list);
 844                 spin_unlock(&head->batch_lock);
 845         }
 846
 847         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 848                 if (atomic_dec_return(&conf->preread_active_stripes)
 849                     < IO_THRESHOLD)
 850                         md_wakeup_thread(conf->mddev->thread);
 851
 852         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 853                 int seq = sh->bm_seq;
 854                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 855                     sh->batch_head->bm_seq > seq)
 856                         seq = sh->batch_head->bm_seq;
 857                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 858                 sh->batch_head->bm_seq = seq;
 859         }
 860
 861         atomic_inc(&sh->count);
 862 unlock_out:
 863         unlock_two_stripes(head, sh);
 864 out:
 865         raid5_release_stripe(head);
 866 }
 867
 868 /* Determine if 'data_offset' or 'new_data_offset' should be used
 869  * in this stripe_head.
 870  */
 871 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 872 {
 873         sector_t progress = conf->reshape_progress;
 874         /* Need a memory barrier to make sure we see the value
 875          * of conf->generation, or ->data_offset that was set before
 876          * reshape_progress was updated.
 877          */
 878         smp_rmb();
 879         if (progress == MaxSector)
 880                 return 0;
 881         if (sh->generation == conf->generation - 1)
 882                 return 0;
 883         /* We are in a reshape, and this is a new-generation stripe,
 884          * so use new_data_offset.
 885          */
 886         return 1;
 887 }
 888
 889 static void
 890 raid5_end_read_request(struct bio *bi);
 891 static void
 892 raid5_end_write_request(struct bio *bi);
 893
 894 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 895 {
 896         struct r5conf *conf = sh->raid_conf;
 897         int i, disks = sh->disks;
 898         struct stripe_head *head_sh = sh;
 899
 900         might_sleep();
 901
 902         if (r5l_write_stripe(conf->log, sh) == 0)
 903                 return;
 904         for (i = disks; i--; ) {
 905                 int rw;
 906                 int replace_only = 0;
 907                 struct bio *bi, *rbi;
 908                 struct md_rdev *rdev, *rrdev = NULL;
 909
 910                 sh = head_sh;
 911                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 912                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 913                                 rw = WRITE_FUA;
 914                         else
 915                                 rw = WRITE;
 916                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 917                                 rw |= REQ_DISCARD;
 918                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 919                         rw = READ;
 920                 else if (test_and_clear_bit(R5_WantReplace,
 921                                             &sh->dev[i].flags)) {
 922                         rw = WRITE;
 923                         replace_only = 1;
 924                 } else
 925                         continue;
 926                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 927                         rw |= REQ_SYNC;
 928
 929 again:
 930                 bi = &sh->dev[i].req;
 931                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 932
 933                 rcu_read_lock();
 934                 rrdev = rcu_dereference(conf->disks[i].replacement);
 935                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 936                 rdev = rcu_dereference(conf->disks[i].rdev);
 937                 if (!rdev) {
 938                         rdev = rrdev;
 939                         rrdev = NULL;
 940                 }
 941                 if (rw & WRITE) {
 942                         if (replace_only)
 943                                 rdev = NULL;
 944                         if (rdev == rrdev)
 945                                 /* We raced and saw duplicates */
 946                                 rrdev = NULL;
 947                 } else {
 948                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
 949                                 rdev = rrdev;
 950                         rrdev = NULL;
 951                 }
 952
 953                 if (rdev && test_bit(Faulty, &rdev->flags))
 954                         rdev = NULL;
 955                 if (rdev)
 956                         atomic_inc(&rdev->nr_pending);
 957                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 958                         rrdev = NULL;
 959                 if (rrdev)
 960                         atomic_inc(&rrdev->nr_pending);
 961                 rcu_read_unlock();
 962
 963                 /* We have already checked bad blocks for reads.  Now
 964                  * need to check for writes.  We never accept write errors
 965                  * on the replacement, so we don't to check rrdev.
 966                  */
 967                 while ((rw & WRITE) && rdev &&
 968                        test_bit(WriteErrorSeen, &rdev->flags)) {
 969                         sector_t first_bad;
 970                         int bad_sectors;
 971                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 972                                               &first_bad, &bad_sectors);
 973                         if (!bad)
 974                                 break;
 975
 976                         if (bad < 0) {
 977                                 set_bit(BlockedBadBlocks, &rdev->flags);
 978                                 if (!conf->mddev->external &&
 979                                     conf->mddev->flags) {
 980                                         /* It is very unlikely, but we might
 981                                          * still need to write out the
 982                                          * bad block log - better give it
 983                                          * a chance*/
 984                                         md_check_recovery(conf->mddev);
 985                                 }
 986                                 /*
 987                                  * Because md_wait_for_blocked_rdev
 988                                  * will dec nr_pending, we must
 989                                  * increment it first.
 990                                  */
 991                                 atomic_inc(&rdev->nr_pending);
 992                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 993                         } else {
 994                                 /* Acknowledged bad block - skip the write */
 995                                 rdev_dec_pending(rdev, conf->mddev);
 996                                 rdev = NULL;
 997                         }
 998                 }
 999
1000                 if (rdev) {
1001                         if (s->syncing || s->expanding || s->expanded
1002                             || s->replacing)
1003                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1004
1005                         set_bit(STRIPE_IO_STARTED, &sh->state);
1006
1007                         bio_reset(bi);
1008                         bi->bi_bdev = rdev->bdev;
1009                         bi->bi_rw = rw;
1010                         bi->bi_end_io = (rw & WRITE)
1011                                 ? raid5_end_write_request
1012                                 : raid5_end_read_request;
1013                         bi->bi_private = sh;
1014
1015                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1016                                 __func__, (unsigned long long)sh->sector,
1017                                 bi->bi_rw, i);
1018                         atomic_inc(&sh->count);
1019                         if (sh != head_sh)
1020                                 atomic_inc(&head_sh->count);
1021                         if (use_new_offset(conf, sh))
1022                                 bi->bi_iter.bi_sector = (sh->sector
1023                                                  + rdev->new_data_offset);
1024                         else
1025                                 bi->bi_iter.bi_sector = (sh->sector
1026                                                  + rdev->data_offset);
1027                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1028                                 bi->bi_rw |= REQ_NOMERGE;
1029
1030                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1031                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1032                         sh->dev[i].vec.bv_page = sh->dev[i].page;
1033                         bi->bi_vcnt = 1;
1034                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1035                         bi->bi_io_vec[0].bv_offset = 0;
1036                         bi->bi_iter.bi_size = STRIPE_SIZE;
1037                         /*
1038                          * If this is discard request, set bi_vcnt 0. We don't
1039                          * want to confuse SCSI because SCSI will replace payload
1040                          */
1041                         if (rw & REQ_DISCARD)
1042                                 bi->bi_vcnt = 0;
1043                         if (rrdev)
1044                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1045
1046                         if (conf->mddev->gendisk)
1047                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1048                                                       bi, disk_devt(conf->mddev->gendisk),
1049                                                       sh->dev[i].sector);
1050                         generic_make_request(bi);
1051                 }
1052                 if (rrdev) {
1053                         if (s->syncing || s->expanding || s->expanded
1054                             || s->replacing)
1055                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1056
1057                         set_bit(STRIPE_IO_STARTED, &sh->state);
1058
1059                         bio_reset(rbi);
1060                         rbi->bi_bdev = rrdev->bdev;
1061                         rbi->bi_rw = rw;
1062                         BUG_ON(!(rw & WRITE));
1063                         rbi->bi_end_io = raid5_end_write_request;
1064                         rbi->bi_private = sh;
1065
1066                         pr_debug("%s: for %llu schedule op %ld on "
1067                                  "replacement disc %d\n",
1068                                 __func__, (unsigned long long)sh->sector,
1069                                 rbi->bi_rw, i);
1070                         atomic_inc(&sh->count);
1071                         if (sh != head_sh)
1072                                 atomic_inc(&head_sh->count);
1073                         if (use_new_offset(conf, sh))
1074                                 rbi->bi_iter.bi_sector = (sh->sector
1075                                                   + rrdev->new_data_offset);
1076                         else
1077                                 rbi->bi_iter.bi_sector = (sh->sector
1078                                                   + rrdev->data_offset);
1079                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1080                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1081                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1082                         rbi->bi_vcnt = 1;
1083                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1084                         rbi->bi_io_vec[0].bv_offset = 0;
1085                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1086                         /*
1087                          * If this is discard request, set bi_vcnt 0. We don't
1088                          * want to confuse SCSI because SCSI will replace payload
1089                          */
1090                         if (rw & REQ_DISCARD)
1091                                 rbi->bi_vcnt = 0;
1092                         if (conf->mddev->gendisk)
1093                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1094                                                       rbi, disk_devt(conf->mddev->gendisk),
1095                                                       sh->dev[i].sector);
1096                         generic_make_request(rbi);
1097                 }
1098                 if (!rdev && !rrdev) {
1099                         if (rw & WRITE)
1100                                 set_bit(STRIPE_DEGRADED, &sh->state);
1101                         pr_debug("skip op %ld on disc %d for sector %llu\n",
1102                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1103                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1104                         set_bit(STRIPE_HANDLE, &sh->state);
1105                 }
1106
1107                 if (!head_sh->batch_head)
1108                         continue;
1109                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1110                                       batch_list);
1111                 if (sh != head_sh)
1112                         goto again;
1113         }
1114 }
1115
1116 static struct dma_async_tx_descriptor *
1117 async_copy_data(int frombio, struct bio *bio, struct page **page,
1118         sector_t sector, struct dma_async_tx_descriptor *tx,
1119         struct stripe_head *sh)
1120 {
1121         struct bio_vec bvl;
1122         struct bvec_iter iter;
1123         struct page *bio_page;
1124         int page_offset;
1125         struct async_submit_ctl submit;
1126         enum async_tx_flags flags = 0;
1127
1128         if (bio->bi_iter.bi_sector >= sector)
1129                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1130         else
1131                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1132
1133         if (frombio)
1134                 flags |= ASYNC_TX_FENCE;
1135         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1136
1137         bio_for_each_segment(bvl, bio, iter) {
1138                 int len = bvl.bv_len;
1139                 int clen;
1140                 int b_offset = 0;
1141
1142                 if (page_offset < 0) {
1143                         b_offset = -page_offset;
1144                         page_offset += b_offset;
1145                         len -= b_offset;
1146                 }
1147
1148                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1149                         clen = STRIPE_SIZE - page_offset;
1150                 else
1151                         clen = len;
1152
1153                 if (clen > 0) {
1154                         b_offset += bvl.bv_offset;
1155                         bio_page = bvl.bv_page;
1156                         if (frombio) {
1157                                 if (sh->raid_conf->skip_copy &&
1158                                     b_offset == 0 && page_offset == 0 &&
1159                                     clen == STRIPE_SIZE)
1160                                         *page = bio_page;
1161                                 else
1162                                         tx = async_memcpy(*page, bio_page, page_offset,
1163                                                   b_offset, clen, &submit);
1164                         } else
1165                                 tx = async_memcpy(bio_page, *page, b_offset,
1166                                                   page_offset, clen, &submit);
1167                 }
1168                 /* chain the operations */
1169                 submit.depend_tx = tx;
1170
1171                 if (clen < len) /* hit end of page */
1172                         break;
1173                 page_offset +=  len;
1174         }
1175
1176         return tx;
1177 }
1178
1179 static void ops_complete_biofill(void *stripe_head_ref)
1180 {
1181         struct stripe_head *sh = stripe_head_ref;
1182         struct bio_list return_bi = BIO_EMPTY_LIST;
1183         int i;
1184
1185         pr_debug("%s: stripe %llu\n", __func__,
1186                 (unsigned long long)sh->sector);
1187
1188         /* clear completed biofills */
1189         for (i = sh->disks; i--; ) {
1190                 struct r5dev *dev = &sh->dev[i];
1191
1192                 /* acknowledge completion of a biofill operation */
1193                 /* and check if we need to reply to a read request,
1194                  * new R5_Wantfill requests are held off until
1195                  * !STRIPE_BIOFILL_RUN
1196                  */
1197                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1198                         struct bio *rbi, *rbi2;
1199
1200                         BUG_ON(!dev->read);
1201                         rbi = dev->read;
1202                         dev->read = NULL;
1203                         while (rbi && rbi->bi_iter.bi_sector <
1204                                 dev->sector + STRIPE_SECTORS) {
1205                                 rbi2 = r5_next_bio(rbi, dev->sector);
1206                                 if (!raid5_dec_bi_active_stripes(rbi))
1207                                         bio_list_add(&return_bi, rbi);
1208                                 rbi = rbi2;
1209                         }
1210                 }
1211         }
1212         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1213
1214         return_io(&return_bi);
1215
1216         set_bit(STRIPE_HANDLE, &sh->state);
1217         raid5_release_stripe(sh);
1218 }
1219
1220 static void ops_run_biofill(struct stripe_head *sh)
1221 {
1222         struct dma_async_tx_descriptor *tx = NULL;
1223         struct async_submit_ctl submit;
1224         int i;
1225
1226         BUG_ON(sh->batch_head);
1227         pr_debug("%s: stripe %llu\n", __func__,
1228                 (unsigned long long)sh->sector);
1229
1230         for (i = sh->disks; i--; ) {
1231                 struct r5dev *dev = &sh->dev[i];
1232                 if (test_bit(R5_Wantfill, &dev->flags)) {
1233                         struct bio *rbi;
1234                         spin_lock_irq(&sh->stripe_lock);
1235                         dev->read = rbi = dev->toread;
1236                         dev->toread = NULL;
1237                         spin_unlock_irq(&sh->stripe_lock);
1238                         while (rbi && rbi->bi_iter.bi_sector <
1239                                 dev->sector + STRIPE_SECTORS) {
1240                                 tx = async_copy_data(0, rbi, &dev->page,
1241                                         dev->sector, tx, sh);
1242                                 rbi = r5_next_bio(rbi, dev->sector);
1243                         }
1244                 }
1245         }
1246
1247         atomic_inc(&sh->count);
1248         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1249         async_trigger_callback(&submit);
1250 }
1251
1252 static void mark_target_uptodate(struct stripe_head *sh, int target)
1253 {
1254         struct r5dev *tgt;
1255
1256         if (target < 0)
1257                 return;
1258
1259         tgt = &sh->dev[target];
1260         set_bit(R5_UPTODATE, &tgt->flags);
1261         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1262         clear_bit(R5_Wantcompute, &tgt->flags);
1263 }
1264
1265 static void ops_complete_compute(void *stripe_head_ref)
1266 {
1267         struct stripe_head *sh = stripe_head_ref;
1268
1269         pr_debug("%s: stripe %llu\n", __func__,
1270                 (unsigned long long)sh->sector);
1271
1272         /* mark the computed target(s) as uptodate */
1273         mark_target_uptodate(sh, sh->ops.target);
1274         mark_target_uptodate(sh, sh->ops.target2);
1275
1276         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1277         if (sh->check_state == check_state_compute_run)
1278                 sh->check_state = check_state_compute_result;
1279         set_bit(STRIPE_HANDLE, &sh->state);
1280         raid5_release_stripe(sh);
1281 }
1282
1283 /* return a pointer to the address conversion region of the scribble buffer */
1284 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1285                                  struct raid5_percpu *percpu, int i)
1286 {
1287         void *addr;
1288
1289         addr = flex_array_get(percpu->scribble, i);
1290         return addr + sizeof(struct page *) * (sh->disks + 2);
1291 }
1292
1293 /* return a pointer to the address conversion region of the scribble buffer */
1294 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1295 {
1296         void *addr;
1297
1298         addr = flex_array_get(percpu->scribble, i);
1299         return addr;
1300 }
1301
1302 static struct dma_async_tx_descriptor *
1303 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1304 {
1305         int disks = sh->disks;
1306         struct page **xor_srcs = to_addr_page(percpu, 0);
1307         int target = sh->ops.target;
1308         struct r5dev *tgt = &sh->dev[target];
1309         struct page *xor_dest = tgt->page;
1310         int count = 0;
1311         struct dma_async_tx_descriptor *tx;
1312         struct async_submit_ctl submit;
1313         int i;
1314
1315         BUG_ON(sh->batch_head);
1316
1317         pr_debug("%s: stripe %llu block: %d\n",
1318                 __func__, (unsigned long long)sh->sector, target);
1319         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1320
1321         for (i = disks; i--; )
1322                 if (i != target)
1323                         xor_srcs[count++] = sh->dev[i].page;
1324
1325         atomic_inc(&sh->count);
1326
1327         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1328                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1329         if (unlikely(count == 1))
1330                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1331         else
1332                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1333
1334         return tx;
1335 }
1336
1337 /* set_syndrome_sources - populate source buffers for gen_syndrome
1338  * @srcs - (struct page *) array of size sh->disks
1339  * @sh - stripe_head to parse
1340  *
1341  * Populates srcs in proper layout order for the stripe and returns the
1342  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1343  * destination buffer is recorded in srcs[count] and the Q destination
1344  * is recorded in srcs[count+1]].
1345  */
1346 static int set_syndrome_sources(struct page **srcs,
1347                                 struct stripe_head *sh,
1348                                 int srctype)
1349 {
1350         int disks = sh->disks;
1351         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1352         int d0_idx = raid6_d0(sh);
1353         int count;
1354         int i;
1355
1356         for (i = 0; i < disks; i++)
1357                 srcs[i] = NULL;
1358
1359         count = 0;
1360         i = d0_idx;
1361         do {
1362                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1363                 struct r5dev *dev = &sh->dev[i];
1364
1365                 if (i == sh->qd_idx || i == sh->pd_idx ||
1366                     (srctype == SYNDROME_SRC_ALL) ||
1367                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1368                      test_bit(R5_Wantdrain, &dev->flags)) ||
1369                     (srctype == SYNDROME_SRC_WRITTEN &&
1370                      dev->written))
1371                         srcs[slot] = sh->dev[i].page;
1372                 i = raid6_next_disk(i, disks);
1373         } while (i != d0_idx);
1374
1375         return syndrome_disks;
1376 }
1377
1378 static struct dma_async_tx_descriptor *
1379 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1380 {
1381         int disks = sh->disks;
1382         struct page **blocks = to_addr_page(percpu, 0);
1383         int target;
1384         int qd_idx = sh->qd_idx;
1385         struct dma_async_tx_descriptor *tx;
1386         struct async_submit_ctl submit;
1387         struct r5dev *tgt;
1388         struct page *dest;
1389         int i;
1390         int count;
1391
1392         BUG_ON(sh->batch_head);
1393         if (sh->ops.target < 0)
1394                 target = sh->ops.target2;
1395         else if (sh->ops.target2 < 0)
1396                 target = sh->ops.target;
1397         else
1398                 /* we should only have one valid target */
1399                 BUG();
1400         BUG_ON(target < 0);
1401         pr_debug("%s: stripe %llu block: %d\n",
1402                 __func__, (unsigned long long)sh->sector, target);
1403
1404         tgt = &sh->dev[target];
1405         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1406         dest = tgt->page;
1407
1408         atomic_inc(&sh->count);
1409
1410         if (target == qd_idx) {
1411                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1412                 blocks[count] = NULL; /* regenerating p is not necessary */
1413                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1414                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1415                                   ops_complete_compute, sh,
1416                                   to_addr_conv(sh, percpu, 0));
1417                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1418         } else {
1419                 /* Compute any data- or p-drive using XOR */
1420                 count = 0;
1421                 for (i = disks; i-- ; ) {
1422                         if (i == target || i == qd_idx)
1423                                 continue;
1424                         blocks[count++] = sh->dev[i].page;
1425                 }
1426
1427                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1428                                   NULL, ops_complete_compute, sh,
1429                                   to_addr_conv(sh, percpu, 0));
1430                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1431         }
1432
1433         return tx;
1434 }
1435
1436 static struct dma_async_tx_descriptor *
1437 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1438 {
1439         int i, count, disks = sh->disks;
1440         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1441         int d0_idx = raid6_d0(sh);
1442         int faila = -1, failb = -1;
1443         int target = sh->ops.target;
1444         int target2 = sh->ops.target2;
1445         struct r5dev *tgt = &sh->dev[target];
1446         struct r5dev *tgt2 = &sh->dev[target2];
1447         struct dma_async_tx_descriptor *tx;
1448         struct page **blocks = to_addr_page(percpu, 0);
1449         struct async_submit_ctl submit;
1450
1451         BUG_ON(sh->batch_head);
1452         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1453                  __func__, (unsigned long long)sh->sector, target, target2);
1454         BUG_ON(target < 0 || target2 < 0);
1455         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1456         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1457
1458         /* we need to open-code set_syndrome_sources to handle the
1459          * slot number conversion for 'faila' and 'failb'
1460          */
1461         for (i = 0; i < disks ; i++)
1462                 blocks[i] = NULL;
1463         count = 0;
1464         i = d0_idx;
1465         do {
1466                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1467
1468                 blocks[slot] = sh->dev[i].page;
1469
1470                 if (i == target)
1471                         faila = slot;
1472                 if (i == target2)
1473                         failb = slot;
1474                 i = raid6_next_disk(i, disks);
1475         } while (i != d0_idx);
1476
1477         BUG_ON(faila == failb);
1478         if (failb < faila)
1479                 swap(faila, failb);
1480         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1481                  __func__, (unsigned long long)sh->sector, faila, failb);
1482
1483         atomic_inc(&sh->count);
1484
1485         if (failb == syndrome_disks+1) {
1486                 /* Q disk is one of the missing disks */
1487                 if (faila == syndrome_disks) {
1488                         /* Missing P+Q, just recompute */
1489                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1490                                           ops_complete_compute, sh,
1491                                           to_addr_conv(sh, percpu, 0));
1492                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1493                                                   STRIPE_SIZE, &submit);
1494                 } else {
1495                         struct page *dest;
1496                         int data_target;
1497                         int qd_idx = sh->qd_idx;
1498
1499                         /* Missing D+Q: recompute D from P, then recompute Q */
1500                         if (target == qd_idx)
1501                                 data_target = target2;
1502                         else
1503                                 data_target = target;
1504
1505                         count = 0;
1506                         for (i = disks; i-- ; ) {
1507                                 if (i == data_target || i == qd_idx)
1508                                         continue;
1509                                 blocks[count++] = sh->dev[i].page;
1510                         }
1511                         dest = sh->dev[data_target].page;
1512                         init_async_submit(&submit,
1513                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1514                                           NULL, NULL, NULL,
1515                                           to_addr_conv(sh, percpu, 0));
1516                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1517                                        &submit);
1518
1519                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1520                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1521                                           ops_complete_compute, sh,
1522                                           to_addr_conv(sh, percpu, 0));
1523                         return async_gen_syndrome(blocks, 0, count+2,
1524                                                   STRIPE_SIZE, &submit);
1525                 }
1526         } else {
1527                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1528                                   ops_complete_compute, sh,
1529                                   to_addr_conv(sh, percpu, 0));
1530                 if (failb == syndrome_disks) {
1531                         /* We're missing D+P. */
1532                         return async_raid6_datap_recov(syndrome_disks+2,
1533                                                        STRIPE_SIZE, faila,
1534                                                        blocks, &submit);
1535                 } else {
1536                         /* We're missing D+D. */
1537                         return async_raid6_2data_recov(syndrome_disks+2,
1538                                                        STRIPE_SIZE, faila, failb,
1539                                                        blocks, &submit);
1540                 }
1541         }
1542 }
1543
1544 static void ops_complete_prexor(void *stripe_head_ref)
1545 {
1546         struct stripe_head *sh = stripe_head_ref;
1547
1548         pr_debug("%s: stripe %llu\n", __func__,
1549                 (unsigned long long)sh->sector);
1550 }
1551
1552 static struct dma_async_tx_descriptor *
1553 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1554                 struct dma_async_tx_descriptor *tx)
1555 {
1556         int disks = sh->disks;
1557         struct page **xor_srcs = to_addr_page(percpu, 0);
1558         int count = 0, pd_idx = sh->pd_idx, i;
1559         struct async_submit_ctl submit;
1560
1561         /* existing parity data subtracted */
1562         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1563
1564         BUG_ON(sh->batch_head);
1565         pr_debug("%s: stripe %llu\n", __func__,
1566                 (unsigned long long)sh->sector);
1567
1568         for (i = disks; i--; ) {
1569                 struct r5dev *dev = &sh->dev[i];
1570                 /* Only process blocks that are known to be uptodate */
1571                 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         int disks = sh->disks;
1606         int i;
1607         struct stripe_head *head_sh = sh;
1608
1609         pr_debug("%s: stripe %llu\n", __func__,
1610                 (unsigned long long)sh->sector);
1611
1612         for (i = disks; i--; ) {
1613                 struct r5dev *dev;
1614                 struct bio *chosen;
1615
1616                 sh = head_sh;
1617                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1618                         struct bio *wbi;
1619
1620 again:
1621                         dev = &sh->dev[i];
1622                         spin_lock_irq(&sh->stripe_lock);
1623                         chosen = dev->towrite;
1624                         dev->towrite = NULL;
1625                         sh->overwrite_disks = 0;
1626                         BUG_ON(dev->written);
1627                         wbi = dev->written = chosen;
1628                         spin_unlock_irq(&sh->stripe_lock);
1629                         WARN_ON(dev->page != dev->orig_page);
1630
1631                         while (wbi && wbi->bi_iter.bi_sector <
1632                                 dev->sector + STRIPE_SECTORS) {
1633                                 if (wbi->bi_rw & REQ_FUA)
1634                                         set_bit(R5_WantFUA, &dev->flags);
1635                                 if (wbi->bi_rw & REQ_SYNC)
1636                                         set_bit(R5_SyncIO, &dev->flags);
1637                                 if (wbi->bi_rw & REQ_DISCARD)
1638                                         set_bit(R5_Discard, &dev->flags);
1639                                 else {
1640                                         tx = async_copy_data(1, wbi, &dev->page,
1641                                                 dev->sector, tx, sh);
1642                                         if (dev->page != dev->orig_page) {
1643                                                 set_bit(R5_SkipCopy, &dev->flags);
1644                                                 clear_bit(R5_UPTODATE, &dev->flags);
1645                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1646                                         }
1647                                 }
1648                                 wbi = r5_next_bio(wbi, dev->sector);
1649                         }
1650
1651                         if (head_sh->batch_head) {
1652                                 sh = list_first_entry(&sh->batch_list,
1653                                                       struct stripe_head,
1654                                                       batch_list);
1655                                 if (sh == head_sh)
1656                                         continue;
1657                                 goto again;
1658                         }
1659                 }
1660         }
1661
1662         return tx;
1663 }
1664
1665 static void ops_complete_reconstruct(void *stripe_head_ref)
1666 {
1667         struct stripe_head *sh = stripe_head_ref;
1668         int disks = sh->disks;
1669         int pd_idx = sh->pd_idx;
1670         int qd_idx = sh->qd_idx;
1671         int i;
1672         bool fua = false, sync = false, discard = false;
1673
1674         pr_debug("%s: stripe %llu\n", __func__,
1675                 (unsigned long long)sh->sector);
1676
1677         for (i = disks; i--; ) {
1678                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1679                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1680                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1681         }
1682
1683         for (i = disks; i--; ) {
1684                 struct r5dev *dev = &sh->dev[i];
1685
1686                 if (dev->written || i == pd_idx || i == qd_idx) {
1687                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1688                                 set_bit(R5_UPTODATE, &dev->flags);
1689                         if (fua)
1690                                 set_bit(R5_WantFUA, &dev->flags);
1691                         if (sync)
1692                                 set_bit(R5_SyncIO, &dev->flags);
1693                 }
1694         }
1695
1696         if (sh->reconstruct_state == reconstruct_state_drain_run)
1697                 sh->reconstruct_state = reconstruct_state_drain_result;
1698         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1699                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1700         else {
1701                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1702                 sh->reconstruct_state = reconstruct_state_result;
1703         }
1704
1705         set_bit(STRIPE_HANDLE, &sh->state);
1706         raid5_release_stripe(sh);
1707 }
1708
1709 static void
1710 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1711                      struct dma_async_tx_descriptor *tx)
1712 {
1713         int disks = sh->disks;
1714         struct page **xor_srcs;
1715         struct async_submit_ctl submit;
1716         int count, pd_idx = sh->pd_idx, i;
1717         struct page *xor_dest;
1718         int prexor = 0;
1719         unsigned long flags;
1720         int j = 0;
1721         struct stripe_head *head_sh = sh;
1722         int last_stripe;
1723
1724         pr_debug("%s: stripe %llu\n", __func__,
1725                 (unsigned long long)sh->sector);
1726
1727         for (i = 0; i < sh->disks; i++) {
1728                 if (pd_idx == i)
1729                         continue;
1730                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1731                         break;
1732         }
1733         if (i >= sh->disks) {
1734                 atomic_inc(&sh->count);
1735                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1736                 ops_complete_reconstruct(sh);
1737                 return;
1738         }
1739 again:
1740         count = 0;
1741         xor_srcs = to_addr_page(percpu, j);
1742         /* check if prexor is active which means only process blocks
1743          * that are part of a read-modify-write (written)
1744          */
1745         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1746                 prexor = 1;
1747                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1748                 for (i = disks; i--; ) {
1749                         struct r5dev *dev = &sh->dev[i];
1750                         if (head_sh->dev[i].written)
1751                                 xor_srcs[count++] = dev->page;
1752                 }
1753         } else {
1754                 xor_dest = sh->dev[pd_idx].page;
1755                 for (i = disks; i--; ) {
1756                         struct r5dev *dev = &sh->dev[i];
1757                         if (i != pd_idx)
1758                                 xor_srcs[count++] = dev->page;
1759                 }
1760         }
1761
1762         /* 1/ if we prexor'd then the dest is reused as a source
1763          * 2/ if we did not prexor then we are redoing the parity
1764          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1765          * for the synchronous xor case
1766          */
1767         last_stripe = !head_sh->batch_head ||
1768                 list_first_entry(&sh->batch_list,
1769                                  struct stripe_head, batch_list) == head_sh;
1770         if (last_stripe) {
1771                 flags = ASYNC_TX_ACK |
1772                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1773
1774                 atomic_inc(&head_sh->count);
1775                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1776                                   to_addr_conv(sh, percpu, j));
1777         } else {
1778                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1779                 init_async_submit(&submit, flags, tx, NULL, NULL,
1780                                   to_addr_conv(sh, percpu, j));
1781         }
1782
1783         if (unlikely(count == 1))
1784                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1785         else
1786                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1787         if (!last_stripe) {
1788                 j++;
1789                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1790                                       batch_list);
1791                 goto again;
1792         }
1793 }
1794
1795 static void
1796 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1797                      struct dma_async_tx_descriptor *tx)
1798 {
1799         struct async_submit_ctl submit;
1800         struct page **blocks;
1801         int count, i, j = 0;
1802         struct stripe_head *head_sh = sh;
1803         int last_stripe;
1804         int synflags;
1805         unsigned long txflags;
1806
1807         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1808
1809         for (i = 0; i < sh->disks; i++) {
1810                 if (sh->pd_idx == i || sh->qd_idx == i)
1811                         continue;
1812                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1813                         break;
1814         }
1815         if (i >= sh->disks) {
1816                 atomic_inc(&sh->count);
1817                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1818                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1819                 ops_complete_reconstruct(sh);
1820                 return;
1821         }
1822
1823 again:
1824         blocks = to_addr_page(percpu, j);
1825
1826         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1827                 synflags = SYNDROME_SRC_WRITTEN;
1828                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1829         } else {
1830                 synflags = SYNDROME_SRC_ALL;
1831                 txflags = ASYNC_TX_ACK;
1832         }
1833
1834         count = set_syndrome_sources(blocks, sh, synflags);
1835         last_stripe = !head_sh->batch_head ||
1836                 list_first_entry(&sh->batch_list,
1837                                  struct stripe_head, batch_list) == head_sh;
1838
1839         if (last_stripe) {
1840                 atomic_inc(&head_sh->count);
1841                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1842                                   head_sh, to_addr_conv(sh, percpu, j));
1843         } else
1844                 init_async_submit(&submit, 0, tx, NULL, NULL,
1845                                   to_addr_conv(sh, percpu, j));
1846         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1847         if (!last_stripe) {
1848                 j++;
1849                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1850                                       batch_list);
1851                 goto again;
1852         }
1853 }
1854
1855 static void ops_complete_check(void *stripe_head_ref)
1856 {
1857         struct stripe_head *sh = stripe_head_ref;
1858
1859         pr_debug("%s: stripe %llu\n", __func__,
1860                 (unsigned long long)sh->sector);
1861
1862         sh->check_state = check_state_check_result;
1863         set_bit(STRIPE_HANDLE, &sh->state);
1864         raid5_release_stripe(sh);
1865 }
1866
1867 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1868 {
1869         int disks = sh->disks;
1870         int pd_idx = sh->pd_idx;
1871         int qd_idx = sh->qd_idx;
1872         struct page *xor_dest;
1873         struct page **xor_srcs = to_addr_page(percpu, 0);
1874         struct dma_async_tx_descriptor *tx;
1875         struct async_submit_ctl submit;
1876         int count;
1877         int i;
1878
1879         pr_debug("%s: stripe %llu\n", __func__,
1880                 (unsigned long long)sh->sector);
1881
1882         BUG_ON(sh->batch_head);
1883         count = 0;
1884         xor_dest = sh->dev[pd_idx].page;
1885         xor_srcs[count++] = xor_dest;
1886         for (i = disks; i--; ) {
1887                 if (i == pd_idx || i == qd_idx)
1888                         continue;
1889                 xor_srcs[count++] = sh->dev[i].page;
1890         }
1891
1892         init_async_submit(&submit, 0, NULL, NULL, NULL,
1893                           to_addr_conv(sh, percpu, 0));
1894         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1895                            &sh->ops.zero_sum_result, &submit);
1896
1897         atomic_inc(&sh->count);
1898         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1899         tx = async_trigger_callback(&submit);
1900 }
1901
1902 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1903 {
1904         struct page **srcs = to_addr_page(percpu, 0);
1905         struct async_submit_ctl submit;
1906         int count;
1907
1908         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1909                 (unsigned long long)sh->sector, checkp);
1910
1911         BUG_ON(sh->batch_head);
1912         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1913         if (!checkp)
1914                 srcs[count] = NULL;
1915
1916         atomic_inc(&sh->count);
1917         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1918                           sh, to_addr_conv(sh, percpu, 0));
1919         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1920                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1921 }
1922
1923 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1924 {
1925         int overlap_clear = 0, i, disks = sh->disks;
1926         struct dma_async_tx_descriptor *tx = NULL;
1927         struct r5conf *conf = sh->raid_conf;
1928         int level = conf->level;
1929         struct raid5_percpu *percpu;
1930         unsigned long cpu;
1931
1932         cpu = get_cpu_light();
1933         percpu = per_cpu_ptr(conf->percpu, cpu);
1934         spin_lock(&percpu->lock);
1935         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1936                 ops_run_biofill(sh);
1937                 overlap_clear++;
1938         }
1939
1940         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1941                 if (level < 6)
1942                         tx = ops_run_compute5(sh, percpu);
1943                 else {
1944                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1945                                 tx = ops_run_compute6_1(sh, percpu);
1946                         else
1947                                 tx = ops_run_compute6_2(sh, percpu);
1948                 }
1949                 /* terminate the chain if reconstruct is not set to be run */
1950                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1951                         async_tx_ack(tx);
1952         }
1953
1954         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1955                 if (level < 6)
1956                         tx = ops_run_prexor5(sh, percpu, tx);
1957                 else
1958                         tx = ops_run_prexor6(sh, percpu, tx);
1959         }
1960
1961         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1962                 tx = ops_run_biodrain(sh, tx);
1963                 overlap_clear++;
1964         }
1965
1966         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1967                 if (level < 6)
1968                         ops_run_reconstruct5(sh, percpu, tx);
1969                 else
1970                         ops_run_reconstruct6(sh, percpu, tx);
1971         }
1972
1973         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1974                 if (sh->check_state == check_state_run)
1975                         ops_run_check_p(sh, percpu);
1976                 else if (sh->check_state == check_state_run_q)
1977                         ops_run_check_pq(sh, percpu, 0);
1978                 else if (sh->check_state == check_state_run_pq)
1979                         ops_run_check_pq(sh, percpu, 1);
1980                 else
1981                         BUG();
1982         }
1983
1984         if (overlap_clear && !sh->batch_head)
1985                 for (i = disks; i--; ) {
1986                         struct r5dev *dev = &sh->dev[i];
1987                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1988                                 wake_up(&sh->raid_conf->wait_for_overlap);
1989                 }
1990         spin_unlock(&percpu->lock);
1991         put_cpu_light();
1992 }
1993
1994 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1995 {
1996         struct stripe_head *sh;
1997
1998         sh = kmem_cache_zalloc(sc, gfp);
1999         if (sh) {
2000                 spin_lock_init(&sh->stripe_lock);
2001                 spin_lock_init(&sh->batch_lock);
2002                 INIT_LIST_HEAD(&sh->batch_list);
2003                 INIT_LIST_HEAD(&sh->lru);
2004                 atomic_set(&sh->count, 1);
2005         }
2006         return sh;
2007 }
2008 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2009 {
2010         struct stripe_head *sh;
2011
2012         sh = alloc_stripe(conf->slab_cache, gfp);
2013         if (!sh)
2014                 return 0;
2015
2016         sh->raid_conf = conf;
2017
2018         if (grow_buffers(sh, gfp)) {
2019                 shrink_buffers(sh);
2020                 kmem_cache_free(conf->slab_cache, sh);
2021                 return 0;
2022         }
2023         sh->hash_lock_index =
2024                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2025         /* we just created an active stripe so... */
2026         atomic_inc(&conf->active_stripes);
2027
2028         raid5_release_stripe(sh);
2029         conf->max_nr_stripes++;
2030         return 1;
2031 }
2032
2033 static int grow_stripes(struct r5conf *conf, int num)
2034 {
2035         struct kmem_cache *sc;
2036         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2037
2038         if (conf->mddev->gendisk)
2039                 sprintf(conf->cache_name[0],
2040                         "raid%d-%s", conf->level, mdname(conf->mddev));
2041         else
2042                 sprintf(conf->cache_name[0],
2043                         "raid%d-%p", conf->level, conf->mddev);
2044         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2045
2046         conf->active_name = 0;
2047         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2048                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2049                                0, 0, NULL);
2050         if (!sc)
2051                 return 1;
2052         conf->slab_cache = sc;
2053         conf->pool_size = devs;
2054         while (num--)
2055                 if (!grow_one_stripe(conf, GFP_KERNEL))
2056                         return 1;
2057
2058         return 0;
2059 }
2060
2061 /**
2062  * scribble_len - return the required size of the scribble region
2063  * @num - total number of disks in the array
2064  *
2065  * The size must be enough to contain:
2066  * 1/ a struct page pointer for each device in the array +2
2067  * 2/ room to convert each entry in (1) to its corresponding dma
2068  *    (dma_map_page()) or page (page_address()) address.
2069  *
2070  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2071  * calculate over all devices (not just the data blocks), using zeros in place
2072  * of the P and Q blocks.
2073  */
2074 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2075 {
2076         struct flex_array *ret;
2077         size_t len;
2078
2079         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2080         ret = flex_array_alloc(len, cnt, flags);
2081         if (!ret)
2082                 return NULL;
2083         /* always prealloc all elements, so no locking is required */
2084         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2085                 flex_array_free(ret);
2086                 return NULL;
2087         }
2088         return ret;
2089 }
2090
2091 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2092 {
2093         unsigned long cpu;
2094         int err = 0;
2095
2096         mddev_suspend(conf->mddev);
2097         get_online_cpus();
2098         for_each_present_cpu(cpu) {
2099                 struct raid5_percpu *percpu;
2100                 struct flex_array *scribble;
2101
2102                 percpu = per_cpu_ptr(conf->percpu, cpu);
2103                 scribble = scribble_alloc(new_disks,
2104                                           new_sectors / STRIPE_SECTORS,
2105                                           GFP_NOIO);
2106
2107                 if (scribble) {
2108                         flex_array_free(percpu->scribble);
2109                         percpu->scribble = scribble;
2110                 } else {
2111                         err = -ENOMEM;
2112                         break;
2113                 }
2114         }
2115         put_online_cpus();
2116         mddev_resume(conf->mddev);
2117         return err;
2118 }
2119
2120 static int resize_stripes(struct r5conf *conf, int newsize)
2121 {
2122         /* Make all the stripes able to hold 'newsize' devices.
2123          * New slots in each stripe get 'page' set to a new page.
2124          *
2125          * This happens in stages:
2126          * 1/ create a new kmem_cache and allocate the required number of
2127          *    stripe_heads.
2128          * 2/ gather all the old stripe_heads and transfer the pages across
2129          *    to the new stripe_heads.  This will have the side effect of
2130          *    freezing the array as once all stripe_heads have been collected,
2131          *    no IO will be possible.  Old stripe heads are freed once their
2132          *    pages have been transferred over, and the old kmem_cache is
2133          *    freed when all stripes are done.
2134          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2135          *    we simple return a failre status - no need to clean anything up.
2136          * 4/ allocate new pages for the new slots in the new stripe_heads.
2137          *    If this fails, we don't bother trying the shrink the
2138          *    stripe_heads down again, we just leave them as they are.
2139          *    As each stripe_head is processed the new one is released into
2140          *    active service.
2141          *
2142          * Once step2 is started, we cannot afford to wait for a write,
2143          * so we use GFP_NOIO allocations.
2144          */
2145         struct stripe_head *osh, *nsh;
2146         LIST_HEAD(newstripes);
2147         struct disk_info *ndisks;
2148         int err;
2149         struct kmem_cache *sc;
2150         int i;
2151         int hash, cnt;
2152
2153         if (newsize <= conf->pool_size)
2154                 return 0; /* never bother to shrink */
2155
2156         err = md_allow_write(conf->mddev);
2157         if (err)
2158                 return err;
2159
2160         /* Step 1 */
2161         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2162                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2163                                0, 0, NULL);
2164         if (!sc)
2165                 return -ENOMEM;
2166
2167         /* Need to ensure auto-resizing doesn't interfere */
2168         mutex_lock(&conf->cache_size_mutex);
2169
2170         for (i = conf->max_nr_stripes; i; i--) {
2171                 nsh = alloc_stripe(sc, GFP_KERNEL);
2172                 if (!nsh)
2173                         break;
2174
2175                 nsh->raid_conf = conf;
2176                 list_add(&nsh->lru, &newstripes);
2177         }
2178         if (i) {
2179                 /* didn't get enough, give up */
2180                 while (!list_empty(&newstripes)) {
2181                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2182                         list_del(&nsh->lru);
2183                         kmem_cache_free(sc, nsh);
2184                 }
2185                 kmem_cache_destroy(sc);
2186                 mutex_unlock(&conf->cache_size_mutex);
2187                 return -ENOMEM;
2188         }
2189         /* Step 2 - Must use GFP_NOIO now.
2190          * OK, we have enough stripes, start collecting inactive
2191          * stripes and copying them over
2192          */
2193         hash = 0;
2194         cnt = 0;
2195         list_for_each_entry(nsh, &newstripes, lru) {
2196                 lock_device_hash_lock(conf, hash);
2197                 wait_event_exclusive_cmd(conf->wait_for_stripe[hash],
2198                                     !list_empty(conf->inactive_list + hash),
2199                                     unlock_device_hash_lock(conf, hash),
2200                                     lock_device_hash_lock(conf, hash));
2201                 osh = get_free_stripe(conf, hash);
2202                 unlock_device_hash_lock(conf, hash);
2203
2204                 for(i=0; i<conf->pool_size; i++) {
2205                         nsh->dev[i].page = osh->dev[i].page;
2206                         nsh->dev[i].orig_page = osh->dev[i].page;
2207                 }
2208                 nsh->hash_lock_index = hash;
2209                 kmem_cache_free(conf->slab_cache, osh);
2210                 cnt++;
2211                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2212                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2213                         hash++;
2214                         cnt = 0;
2215                 }
2216         }
2217         kmem_cache_destroy(conf->slab_cache);
2218
2219         /* Step 3.
2220          * At this point, we are holding all the stripes so the array
2221          * is completely stalled, so now is a good time to resize
2222          * conf->disks and the scribble region
2223          */
2224         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2225         if (ndisks) {
2226                 for (i=0; i<conf->raid_disks; i++)
2227                         ndisks[i] = conf->disks[i];
2228                 kfree(conf->disks);
2229                 conf->disks = ndisks;
2230         } else
2231                 err = -ENOMEM;
2232
2233         mutex_unlock(&conf->cache_size_mutex);
2234         /* Step 4, return new stripes to service */
2235         while(!list_empty(&newstripes)) {
2236                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2237                 list_del_init(&nsh->lru);
2238
2239                 for (i=conf->raid_disks; i < newsize; i++)
2240                         if (nsh->dev[i].page == NULL) {
2241                                 struct page *p = alloc_page(GFP_NOIO);
2242                                 nsh->dev[i].page = p;
2243                                 nsh->dev[i].orig_page = p;
2244                                 if (!p)
2245                                         err = -ENOMEM;
2246                         }
2247                 raid5_release_stripe(nsh);
2248         }
2249         /* critical section pass, GFP_NOIO no longer needed */
2250
2251         conf->slab_cache = sc;
2252         conf->active_name = 1-conf->active_name;
2253         if (!err)
2254                 conf->pool_size = newsize;
2255         return err;
2256 }
2257
2258 static int drop_one_stripe(struct r5conf *conf)
2259 {
2260         struct stripe_head *sh;
2261         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2262
2263         spin_lock_irq(conf->hash_locks + hash);
2264         sh = get_free_stripe(conf, hash);
2265         spin_unlock_irq(conf->hash_locks + hash);
2266         if (!sh)
2267                 return 0;
2268         BUG_ON(atomic_read(&sh->count));
2269         shrink_buffers(sh);
2270         kmem_cache_free(conf->slab_cache, sh);
2271         atomic_dec(&conf->active_stripes);
2272         conf->max_nr_stripes--;
2273         return 1;
2274 }
2275
2276 static void shrink_stripes(struct r5conf *conf)
2277 {
2278         while (conf->max_nr_stripes &&
2279                drop_one_stripe(conf))
2280                 ;
2281
2282         kmem_cache_destroy(conf->slab_cache);
2283         conf->slab_cache = NULL;
2284 }
2285
2286 static void raid5_end_read_request(struct bio * bi)
2287 {
2288         struct stripe_head *sh = bi->bi_private;
2289         struct r5conf *conf = sh->raid_conf;
2290         int disks = sh->disks, i;
2291         char b[BDEVNAME_SIZE];
2292         struct md_rdev *rdev = NULL;
2293         sector_t s;
2294
2295         for (i=0 ; i<disks; i++)
2296                 if (bi == &sh->dev[i].req)
2297                         break;
2298
2299         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2300                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2301                 bi->bi_error);
2302         if (i == disks) {
2303                 BUG();
2304                 return;
2305         }
2306         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2307                 /* If replacement finished while this request was outstanding,
2308                  * 'replacement' might be NULL already.
2309                  * In that case it moved down to 'rdev'.
2310                  * rdev is not removed until all requests are finished.
2311                  */
2312                 rdev = conf->disks[i].replacement;
2313         if (!rdev)
2314                 rdev = conf->disks[i].rdev;
2315
2316         if (use_new_offset(conf, sh))
2317                 s = sh->sector + rdev->new_data_offset;
2318         else
2319                 s = sh->sector + rdev->data_offset;
2320         if (!bi->bi_error) {
2321                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2322                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2323                         /* Note that this cannot happen on a
2324                          * replacement device.  We just fail those on
2325                          * any error
2326                          */
2327                         printk_ratelimited(
2328                                 KERN_INFO
2329                                 "md/raid:%s: read error corrected"
2330                                 " (%lu sectors at %llu on %s)\n",
2331                                 mdname(conf->mddev), STRIPE_SECTORS,
2332                                 (unsigned long long)s,
2333                                 bdevname(rdev->bdev, b));
2334                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2335                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2336                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2337                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2338                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2339
2340                 if (atomic_read(&rdev->read_errors))
2341                         atomic_set(&rdev->read_errors, 0);
2342         } else {
2343                 const char *bdn = bdevname(rdev->bdev, b);
2344                 int retry = 0;
2345                 int set_bad = 0;
2346
2347                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2348                 atomic_inc(&rdev->read_errors);
2349                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2350                         printk_ratelimited(
2351                                 KERN_WARNING
2352                                 "md/raid:%s: read error on replacement device "
2353                                 "(sector %llu on %s).\n",
2354                                 mdname(conf->mddev),
2355                                 (unsigned long long)s,
2356                                 bdn);
2357                 else if (conf->mddev->degraded >= conf->max_degraded) {
2358                         set_bad = 1;
2359                         printk_ratelimited(
2360                                 KERN_WARNING
2361                                 "md/raid:%s: read error not correctable "
2362                                 "(sector %llu on %s).\n",
2363                                 mdname(conf->mddev),
2364                                 (unsigned long long)s,
2365                                 bdn);
2366                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2367                         /* Oh, no!!! */
2368                         set_bad = 1;
2369                         printk_ratelimited(
2370                                 KERN_WARNING
2371                                 "md/raid:%s: read error NOT corrected!! "
2372                                 "(sector %llu on %s).\n",
2373                                 mdname(conf->mddev),
2374                                 (unsigned long long)s,
2375                                 bdn);
2376                 } else if (atomic_read(&rdev->read_errors)
2377                          > conf->max_nr_stripes)
2378                         printk(KERN_WARNING
2379                                "md/raid:%s: Too many read errors, failing device %s.\n",
2380                                mdname(conf->mddev), bdn);
2381                 else
2382                         retry = 1;
2383                 if (set_bad && test_bit(In_sync, &rdev->flags)
2384                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2385                         retry = 1;
2386                 if (retry)
2387                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2388                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2389                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2390                         } else
2391                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2392                 else {
2393                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2394                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2395                         if (!(set_bad
2396                               && test_bit(In_sync, &rdev->flags)
2397                               && rdev_set_badblocks(
2398                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2399                                 md_error(conf->mddev, rdev);
2400                 }
2401         }
2402         rdev_dec_pending(rdev, conf->mddev);
2403         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2404         set_bit(STRIPE_HANDLE, &sh->state);
2405         raid5_release_stripe(sh);
2406 }
2407
2408 static void raid5_end_write_request(struct bio *bi)
2409 {
2410         struct stripe_head *sh = bi->bi_private;
2411         struct r5conf *conf = sh->raid_conf;
2412         int disks = sh->disks, i;
2413         struct md_rdev *uninitialized_var(rdev);
2414         sector_t first_bad;
2415         int bad_sectors;
2416         int replacement = 0;
2417
2418         for (i = 0 ; i < disks; i++) {
2419                 if (bi == &sh->dev[i].req) {
2420                         rdev = conf->disks[i].rdev;
2421                         break;
2422                 }
2423                 if (bi == &sh->dev[i].rreq) {
2424                         rdev = conf->disks[i].replacement;
2425                         if (rdev)
2426                                 replacement = 1;
2427                         else
2428                                 /* rdev was removed and 'replacement'
2429                                  * replaced it.  rdev is not removed
2430                                  * until all requests are finished.
2431                                  */
2432                                 rdev = conf->disks[i].rdev;
2433                         break;
2434                 }
2435         }
2436         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2437                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2438                 bi->bi_error);
2439         if (i == disks) {
2440                 BUG();
2441                 return;
2442         }
2443
2444         if (replacement) {
2445                 if (bi->bi_error)
2446                         md_error(conf->mddev, rdev);
2447                 else if (is_badblock(rdev, sh->sector,
2448                                      STRIPE_SECTORS,
2449                                      &first_bad, &bad_sectors))
2450                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2451         } else {
2452                 if (bi->bi_error) {
2453                         set_bit(STRIPE_DEGRADED, &sh->state);
2454                         set_bit(WriteErrorSeen, &rdev->flags);
2455                         set_bit(R5_WriteError, &sh->dev[i].flags);
2456                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2457                                 set_bit(MD_RECOVERY_NEEDED,
2458                                         &rdev->mddev->recovery);
2459                 } else if (is_badblock(rdev, sh->sector,
2460                                        STRIPE_SECTORS,
2461                                        &first_bad, &bad_sectors)) {
2462                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2463                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2464                                 /* That was a successful write so make
2465                                  * sure it looks like we already did
2466                                  * a re-write.
2467                                  */
2468                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2469                 }
2470         }
2471         rdev_dec_pending(rdev, conf->mddev);
2472
2473         if (sh->batch_head && bi->bi_error && !replacement)
2474                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2475
2476         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2477                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2478         set_bit(STRIPE_HANDLE, &sh->state);
2479         raid5_release_stripe(sh);
2480
2481         if (sh->batch_head && sh != sh->batch_head)
2482                 raid5_release_stripe(sh->batch_head);
2483 }
2484
2485 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2486 {
2487         struct r5dev *dev = &sh->dev[i];
2488
2489         bio_init(&dev->req);
2490         dev->req.bi_io_vec = &dev->vec;
2491         dev->req.bi_max_vecs = 1;
2492         dev->req.bi_private = sh;
2493
2494         bio_init(&dev->rreq);
2495         dev->rreq.bi_io_vec = &dev->rvec;
2496         dev->rreq.bi_max_vecs = 1;
2497         dev->rreq.bi_private = sh;
2498
2499         dev->flags = 0;
2500         dev->sector = raid5_compute_blocknr(sh, i, previous);
2501 }
2502
2503 static void error(struct mddev *mddev, struct md_rdev *rdev)
2504 {
2505         char b[BDEVNAME_SIZE];
2506         struct r5conf *conf = mddev->private;
2507         unsigned long flags;
2508         pr_debug("raid456: error called\n");
2509
2510         spin_lock_irqsave(&conf->device_lock, flags);
2511         clear_bit(In_sync, &rdev->flags);
2512         mddev->degraded = calc_degraded(conf);
2513         spin_unlock_irqrestore(&conf->device_lock, flags);
2514         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2515
2516         set_bit(Blocked, &rdev->flags);
2517         set_bit(Faulty, &rdev->flags);
2518         set_bit(MD_CHANGE_DEVS, &mddev->flags);
2519         set_bit(MD_CHANGE_PENDING, &mddev->flags);
2520         printk(KERN_ALERT
2521                "md/raid:%s: Disk failure on %s, disabling device.\n"
2522                "md/raid:%s: Operation continuing on %d devices.\n",
2523                mdname(mddev),
2524                bdevname(rdev->bdev, b),
2525                mdname(mddev),
2526                conf->raid_disks - mddev->degraded);
2527 }
2528
2529 /*
2530  * Input: a 'big' sector number,
2531  * Output: index of the data and parity disk, and the sector # in them.
2532  */
2533 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2534                               int previous, int *dd_idx,
2535                               struct stripe_head *sh)
2536 {
2537         sector_t stripe, stripe2;
2538         sector_t chunk_number;
2539         unsigned int chunk_offset;
2540         int pd_idx, qd_idx;
2541         int ddf_layout = 0;
2542         sector_t new_sector;
2543         int algorithm = previous ? conf->prev_algo
2544                                  : conf->algorithm;
2545         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2546                                          : conf->chunk_sectors;
2547         int raid_disks = previous ? conf->previous_raid_disks
2548                                   : conf->raid_disks;
2549         int data_disks = raid_disks - conf->max_degraded;
2550
2551         /* First compute the information on this sector */
2552
2553         /*
2554          * Compute the chunk number and the sector offset inside the chunk
2555          */
2556         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2557         chunk_number = r_sector;
2558
2559         /*
2560          * Compute the stripe number
2561          */
2562         stripe = chunk_number;
2563         *dd_idx = sector_div(stripe, data_disks);
2564         stripe2 = stripe;
2565         /*
2566          * Select the parity disk based on the user selected algorithm.
2567          */
2568         pd_idx = qd_idx = -1;
2569         switch(conf->level) {
2570         case 4:
2571                 pd_idx = data_disks;
2572                 break;
2573         case 5:
2574                 switch (algorithm) {
2575                 case ALGORITHM_LEFT_ASYMMETRIC:
2576                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2577                         if (*dd_idx >= pd_idx)
2578                                 (*dd_idx)++;
2579                         break;
2580                 case ALGORITHM_RIGHT_ASYMMETRIC:
2581                         pd_idx = sector_div(stripe2, raid_disks);
2582                         if (*dd_idx >= pd_idx)
2583                                 (*dd_idx)++;
2584                         break;
2585                 case ALGORITHM_LEFT_SYMMETRIC:
2586                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2587                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2588                         break;
2589                 case ALGORITHM_RIGHT_SYMMETRIC:
2590                         pd_idx = sector_div(stripe2, raid_disks);
2591                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2592                         break;
2593                 case ALGORITHM_PARITY_0:
2594                         pd_idx = 0;
2595                         (*dd_idx)++;
2596                         break;
2597                 case ALGORITHM_PARITY_N:
2598                         pd_idx = data_disks;
2599                         break;
2600                 default:
2601                         BUG();
2602                 }
2603                 break;
2604         case 6:
2605
2606                 switch (algorithm) {
2607                 case ALGORITHM_LEFT_ASYMMETRIC:
2608                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2609                         qd_idx = pd_idx + 1;
2610                         if (pd_idx == raid_disks-1) {
2611                                 (*dd_idx)++;    /* Q D D D P */
2612                                 qd_idx = 0;
2613                         } else if (*dd_idx >= pd_idx)
2614                                 (*dd_idx) += 2; /* D D P Q D */
2615                         break;
2616                 case ALGORITHM_RIGHT_ASYMMETRIC:
2617                         pd_idx = sector_div(stripe2, raid_disks);
2618                         qd_idx = pd_idx + 1;
2619                         if (pd_idx == raid_disks-1) {
2620                                 (*dd_idx)++;    /* Q D D D P */
2621                                 qd_idx = 0;
2622                         } else if (*dd_idx >= pd_idx)
2623                                 (*dd_idx) += 2; /* D D P Q D */
2624                         break;
2625                 case ALGORITHM_LEFT_SYMMETRIC:
2626                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2627                         qd_idx = (pd_idx + 1) % raid_disks;
2628                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2629                         break;
2630                 case ALGORITHM_RIGHT_SYMMETRIC:
2631                         pd_idx = sector_div(stripe2, raid_disks);
2632                         qd_idx = (pd_idx + 1) % raid_disks;
2633                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2634                         break;
2635
2636                 case ALGORITHM_PARITY_0:
2637                         pd_idx = 0;
2638                         qd_idx = 1;
2639                         (*dd_idx) += 2;
2640                         break;
2641                 case ALGORITHM_PARITY_N:
2642                         pd_idx = data_disks;
2643                         qd_idx = data_disks + 1;
2644                         break;
2645
2646                 case ALGORITHM_ROTATING_ZERO_RESTART:
2647                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2648                          * of blocks for computing Q is different.
2649                          */
2650                         pd_idx = sector_div(stripe2, raid_disks);
2651                         qd_idx = pd_idx + 1;
2652                         if (pd_idx == raid_disks-1) {
2653                                 (*dd_idx)++;    /* Q D D D P */
2654                                 qd_idx = 0;
2655                         } else if (*dd_idx >= pd_idx)
2656                                 (*dd_idx) += 2; /* D D P Q D */
2657                         ddf_layout = 1;
2658                         break;
2659
2660                 case ALGORITHM_ROTATING_N_RESTART:
2661                         /* Same a left_asymmetric, by first stripe is
2662                          * D D D P Q  rather than
2663                          * Q D D D P
2664                          */
2665                         stripe2 += 1;
2666                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2667                         qd_idx = pd_idx + 1;
2668                         if (pd_idx == raid_disks-1) {
2669                                 (*dd_idx)++;    /* Q D D D P */
2670                                 qd_idx = 0;
2671                         } else if (*dd_idx >= pd_idx)
2672                                 (*dd_idx) += 2; /* D D P Q D */
2673                         ddf_layout = 1;
2674                         break;
2675
2676                 case ALGORITHM_ROTATING_N_CONTINUE:
2677                         /* Same as left_symmetric but Q is before P */
2678                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2679                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2680                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2681                         ddf_layout = 1;
2682                         break;
2683
2684                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2685                         /* RAID5 left_asymmetric, with Q on last device */
2686                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2687                         if (*dd_idx >= pd_idx)
2688                                 (*dd_idx)++;
2689                         qd_idx = raid_disks - 1;
2690                         break;
2691
2692                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2693                         pd_idx = sector_div(stripe2, raid_disks-1);
2694                         if (*dd_idx >= pd_idx)
2695                                 (*dd_idx)++;
2696                         qd_idx = raid_disks - 1;
2697                         break;
2698
2699                 case ALGORITHM_LEFT_SYMMETRIC_6:
2700                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2701                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2702                         qd_idx = raid_disks - 1;
2703                         break;
2704
2705                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2706                         pd_idx = sector_div(stripe2, raid_disks-1);
2707                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2708                         qd_idx = raid_disks - 1;
2709                         break;
2710
2711                 case ALGORITHM_PARITY_0_6:
2712                         pd_idx = 0;
2713                         (*dd_idx)++;
2714                         qd_idx = raid_disks - 1;
2715                         break;
2716
2717                 default:
2718                         BUG();
2719                 }
2720                 break;
2721         }
2722
2723         if (sh) {
2724                 sh->pd_idx = pd_idx;
2725                 sh->qd_idx = qd_idx;
2726                 sh->ddf_layout = ddf_layout;
2727         }
2728         /*
2729          * Finally, compute the new sector number
2730          */
2731         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2732         return new_sector;
2733 }
2734
2735 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2736 {
2737         struct r5conf *conf = sh->raid_conf;
2738         int raid_disks = sh->disks;
2739         int data_disks = raid_disks - conf->max_degraded;
2740         sector_t new_sector = sh->sector, check;
2741         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2742                                          : conf->chunk_sectors;
2743         int algorithm = previous ? conf->prev_algo
2744                                  : conf->algorithm;
2745         sector_t stripe;
2746         int chunk_offset;
2747         sector_t chunk_number;
2748         int dummy1, dd_idx = i;
2749         sector_t r_sector;
2750         struct stripe_head sh2;
2751
2752         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2753         stripe = new_sector;
2754
2755         if (i == sh->pd_idx)
2756                 return 0;
2757         switch(conf->level) {
2758         case 4: break;
2759         case 5:
2760                 switch (algorithm) {
2761                 case ALGORITHM_LEFT_ASYMMETRIC:
2762                 case ALGORITHM_RIGHT_ASYMMETRIC:
2763                         if (i > sh->pd_idx)
2764                                 i--;
2765                         break;
2766                 case ALGORITHM_LEFT_SYMMETRIC:
2767                 case ALGORITHM_RIGHT_SYMMETRIC:
2768                         if (i < sh->pd_idx)
2769                                 i += raid_disks;
2770                         i -= (sh->pd_idx + 1);
2771                         break;
2772                 case ALGORITHM_PARITY_0:
2773                         i -= 1;
2774                         break;
2775                 case ALGORITHM_PARITY_N:
2776                         break;
2777                 default:
2778                         BUG();
2779                 }
2780                 break;
2781         case 6:
2782                 if (i == sh->qd_idx)
2783                         return 0; /* It is the Q disk */
2784                 switch (algorithm) {
2785                 case ALGORITHM_LEFT_ASYMMETRIC:
2786                 case ALGORITHM_RIGHT_ASYMMETRIC:
2787                 case ALGORITHM_ROTATING_ZERO_RESTART:
2788                 case ALGORITHM_ROTATING_N_RESTART:
2789                         if (sh->pd_idx == raid_disks-1)
2790                                 i--;    /* Q D D D P */
2791                         else if (i > sh->pd_idx)
2792                                 i -= 2; /* D D P Q D */
2793                         break;
2794                 case ALGORITHM_LEFT_SYMMETRIC:
2795                 case ALGORITHM_RIGHT_SYMMETRIC:
2796                         if (sh->pd_idx == raid_disks-1)
2797                                 i--; /* Q D D D P */
2798                         else {
2799                                 /* D D P Q D */
2800                                 if (i < sh->pd_idx)
2801                                         i += raid_disks;
2802                                 i -= (sh->pd_idx + 2);
2803                         }
2804                         break;
2805                 case ALGORITHM_PARITY_0:
2806                         i -= 2;
2807                         break;
2808                 case ALGORITHM_PARITY_N:
2809                         break;
2810                 case ALGORITHM_ROTATING_N_CONTINUE:
2811                         /* Like left_symmetric, but P is before Q */
2812                         if (sh->pd_idx == 0)
2813                                 i--;    /* P D D D Q */
2814                         else {
2815                                 /* D D Q P D */
2816                                 if (i < sh->pd_idx)
2817                                         i += raid_disks;
2818                                 i -= (sh->pd_idx + 1);
2819                         }
2820                         break;
2821                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2822                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2823                         if (i > sh->pd_idx)
2824                                 i--;
2825                         break;
2826                 case ALGORITHM_LEFT_SYMMETRIC_6:
2827                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2828                         if (i < sh->pd_idx)
2829                                 i += data_disks + 1;
2830                         i -= (sh->pd_idx + 1);
2831                         break;
2832                 case ALGORITHM_PARITY_0_6:
2833                         i -= 1;
2834                         break;
2835                 default:
2836                         BUG();
2837                 }
2838                 break;
2839         }
2840
2841         chunk_number = stripe * data_disks + i;
2842         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2843
2844         check = raid5_compute_sector(conf, r_sector,
2845                                      previous, &dummy1, &sh2);
2846         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2847                 || sh2.qd_idx != sh->qd_idx) {
2848                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2849                        mdname(conf->mddev));
2850                 return 0;
2851         }
2852         return r_sector;
2853 }
2854
2855 static void
2856 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2857                          int rcw, int expand)
2858 {
2859         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2860         struct r5conf *conf = sh->raid_conf;
2861         int level = conf->level;
2862
2863         if (rcw) {
2864
2865                 for (i = disks; i--; ) {
2866                         struct r5dev *dev = &sh->dev[i];
2867
2868                         if (dev->towrite) {
2869                                 set_bit(R5_LOCKED, &dev->flags);
2870                                 set_bit(R5_Wantdrain, &dev->flags);
2871                                 if (!expand)
2872                                         clear_bit(R5_UPTODATE, &dev->flags);
2873                                 s->locked++;
2874                         }
2875                 }
2876                 /* if we are not expanding this is a proper write request, and
2877                  * there will be bios with new data to be drained into the
2878                  * stripe cache
2879                  */
2880                 if (!expand) {
2881                         if (!s->locked)
2882                                 /* False alarm, nothing to do */
2883                                 return;
2884                         sh->reconstruct_state = reconstruct_state_drain_run;
2885                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2886                 } else
2887                         sh->reconstruct_state = reconstruct_state_run;
2888
2889                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2890
2891                 if (s->locked + conf->max_degraded == disks)
2892                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2893                                 atomic_inc(&conf->pending_full_writes);
2894         } else {
2895                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2896                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2897                 BUG_ON(level == 6 &&
2898                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2899                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2900
2901                 for (i = disks; i--; ) {
2902                         struct r5dev *dev = &sh->dev[i];
2903                         if (i == pd_idx || i == qd_idx)
2904                                 continue;
2905
2906                         if (dev->towrite &&
2907                             (test_bit(R5_UPTODATE, &dev->flags) ||
2908                              test_bit(R5_Wantcompute, &dev->flags))) {
2909                                 set_bit(R5_Wantdrain, &dev->flags);
2910                                 set_bit(R5_LOCKED, &dev->flags);
2911                                 clear_bit(R5_UPTODATE, &dev->flags);
2912                                 s->locked++;
2913                         }
2914                 }
2915                 if (!s->locked)
2916                         /* False alarm - nothing to do */
2917                         return;
2918                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2919                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2920                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2921                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2922         }
2923
2924         /* keep the parity disk(s) locked while asynchronous operations
2925          * are in flight
2926          */
2927         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2928         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2929         s->locked++;
2930
2931         if (level == 6) {
2932                 int qd_idx = sh->qd_idx;
2933                 struct r5dev *dev = &sh->dev[qd_idx];
2934
2935                 set_bit(R5_LOCKED, &dev->flags);
2936                 clear_bit(R5_UPTODATE, &dev->flags);
2937                 s->locked++;
2938         }
2939
2940         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2941                 __func__, (unsigned long long)sh->sector,
2942                 s->locked, s->ops_request);
2943 }
2944
2945 /*
2946  * Each stripe/dev can have one or more bion attached.
2947  * toread/towrite point to the first in a chain.
2948  * The bi_next chain must be in order.
2949  */
2950 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2951                           int forwrite, int previous)
2952 {
2953         struct bio **bip;
2954         struct r5conf *conf = sh->raid_conf;
2955         int firstwrite=0;
2956
2957         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2958                 (unsigned long long)bi->bi_iter.bi_sector,
2959                 (unsigned long long)sh->sector);
2960
2961         /*
2962          * If several bio share a stripe. The bio bi_phys_segments acts as a
2963          * reference count to avoid race. The reference count should already be
2964          * increased before this function is called (for example, in
2965          * make_request()), so other bio sharing this stripe will not free the
2966          * stripe. If a stripe is owned by one stripe, the stripe lock will
2967          * protect it.
2968          */
2969         spin_lock_irq(&sh->stripe_lock);
2970         /* Don't allow new IO added to stripes in batch list */
2971         if (sh->batch_head)
2972                 goto overlap;
2973         if (forwrite) {
2974                 bip = &sh->dev[dd_idx].towrite;
2975                 if (*bip == NULL)
2976                         firstwrite = 1;
2977         } else
2978                 bip = &sh->dev[dd_idx].toread;
2979         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2980                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2981                         goto overlap;
2982                 bip = & (*bip)->bi_next;
2983         }
2984         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2985                 goto overlap;
2986
2987         if (!forwrite || previous)
2988                 clear_bit(STRIPE_BATCH_READY, &sh->state);
2989
2990         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2991         if (*bip)
2992                 bi->bi_next = *bip;
2993         *bip = bi;
2994         raid5_inc_bi_active_stripes(bi);
2995
2996         if (forwrite) {
2997                 /* check if page is covered */
2998                 sector_t sector = sh->dev[dd_idx].sector;
2999                 for (bi=sh->dev[dd_idx].towrite;
3000                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3001                              bi && bi->bi_iter.bi_sector <= sector;
3002                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3003                         if (bio_end_sector(bi) >= sector)
3004                                 sector = bio_end_sector(bi);
3005                 }
3006                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3007                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3008                                 sh->overwrite_disks++;
3009         }
3010
3011         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3012                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3013                 (unsigned long long)sh->sector, dd_idx);
3014
3015         if (conf->mddev->bitmap && firstwrite) {
3016                 /* Cannot hold spinlock over bitmap_startwrite,
3017                  * but must ensure this isn't added to a batch until
3018                  * we have added to the bitmap and set bm_seq.
3019                  * So set STRIPE_BITMAP_PENDING to prevent
3020                  * batching.
3021                  * If multiple add_stripe_bio() calls race here they
3022                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3023                  * to complete "bitmap_startwrite" gets to set
3024                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3025                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3026                  * any more.
3027                  */
3028                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3029                 spin_unlock_irq(&sh->stripe_lock);
3030                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3031                                   STRIPE_SECTORS, 0);
3032                 spin_lock_irq(&sh->stripe_lock);
3033                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3034                 if (!sh->batch_head) {
3035                         sh->bm_seq = conf->seq_flush+1;
3036                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3037                 }
3038         }
3039         spin_unlock_irq(&sh->stripe_lock);
3040
3041         if (stripe_can_batch(sh))
3042                 stripe_add_to_batch_list(conf, sh);
3043         return 1;
3044
3045  overlap:
3046         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3047         spin_unlock_irq(&sh->stripe_lock);
3048         return 0;
3049 }
3050
3051 static void end_reshape(struct r5conf *conf);
3052
3053 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3054                             struct stripe_head *sh)
3055 {
3056         int sectors_per_chunk =
3057                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3058         int dd_idx;
3059         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3060         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3061
3062         raid5_compute_sector(conf,
3063                              stripe * (disks - conf->max_degraded)
3064                              *sectors_per_chunk + chunk_offset,
3065                              previous,
3066                              &dd_idx, sh);
3067 }
3068
3069 static void
3070 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3071                                 struct stripe_head_state *s, int disks,
3072                                 struct bio_list *return_bi)
3073 {
3074         int i;
3075         BUG_ON(sh->batch_head);
3076         for (i = disks; i--; ) {
3077                 struct bio *bi;
3078                 int bitmap_end = 0;
3079
3080                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3081                         struct md_rdev *rdev;
3082                         rcu_read_lock();
3083                         rdev = rcu_dereference(conf->disks[i].rdev);
3084                         if (rdev && test_bit(In_sync, &rdev->flags))
3085                                 atomic_inc(&rdev->nr_pending);
3086                         else
3087                                 rdev = NULL;
3088                         rcu_read_unlock();
3089                         if (rdev) {
3090                                 if (!rdev_set_badblocks(
3091                                             rdev,
3092                                             sh->sector,
3093                                             STRIPE_SECTORS, 0))
3094                                         md_error(conf->mddev, rdev);
3095                                 rdev_dec_pending(rdev, conf->mddev);
3096                         }
3097                 }
3098                 spin_lock_irq(&sh->stripe_lock);
3099                 /* fail all writes first */
3100                 bi = sh->dev[i].towrite;
3101                 sh->dev[i].towrite = NULL;
3102                 sh->overwrite_disks = 0;
3103                 spin_unlock_irq(&sh->stripe_lock);
3104                 if (bi)
3105                         bitmap_end = 1;
3106
3107                 r5l_stripe_write_finished(sh);
3108
3109                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3110                         wake_up(&conf->wait_for_overlap);
3111
3112                 while (bi && bi->bi_iter.bi_sector <
3113                         sh->dev[i].sector + STRIPE_SECTORS) {
3114                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3115
3116                         bi->bi_error = -EIO;
3117                         if (!raid5_dec_bi_active_stripes(bi)) {
3118                                 md_write_end(conf->mddev);
3119                                 bio_list_add(return_bi, bi);
3120                         }
3121                         bi = nextbi;
3122                 }
3123                 if (bitmap_end)
3124                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3125                                 STRIPE_SECTORS, 0, 0);
3126                 bitmap_end = 0;
3127                 /* and fail all 'written' */
3128                 bi = sh->dev[i].written;
3129                 sh->dev[i].written = NULL;
3130                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3131                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3132                         sh->dev[i].page = sh->dev[i].orig_page;
3133                 }
3134
3135                 if (bi) bitmap_end = 1;
3136                 while (bi && bi->bi_iter.bi_sector <
3137                        sh->dev[i].sector + STRIPE_SECTORS) {
3138                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3139
3140                         bi->bi_error = -EIO;
3141                         if (!raid5_dec_bi_active_stripes(bi)) {
3142                                 md_write_end(conf->mddev);
3143                                 bio_list_add(return_bi, bi);
3144                         }
3145                         bi = bi2;
3146                 }
3147
3148                 /* fail any reads if this device is non-operational and
3149                  * the data has not reached the cache yet.
3150                  */
3151                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3152                     s->failed > conf->max_degraded &&
3153                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3154                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3155                         spin_lock_irq(&sh->stripe_lock);
3156                         bi = sh->dev[i].toread;
3157                         sh->dev[i].toread = NULL;
3158                         spin_unlock_irq(&sh->stripe_lock);
3159                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3160                                 wake_up(&conf->wait_for_overlap);
3161                         if (bi)
3162                                 s->to_read--;
3163                         while (bi && bi->bi_iter.bi_sector <
3164                                sh->dev[i].sector + STRIPE_SECTORS) {
3165                                 struct bio *nextbi =
3166                                         r5_next_bio(bi, sh->dev[i].sector);
3167
3168                                 bi->bi_error = -EIO;
3169                                 if (!raid5_dec_bi_active_stripes(bi))
3170                                         bio_list_add(return_bi, bi);
3171                                 bi = nextbi;
3172                         }
3173                 }
3174                 if (bitmap_end)
3175                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3176                                         STRIPE_SECTORS, 0, 0);
3177                 /* If we were in the middle of a write the parity block might
3178                  * still be locked - so just clear all R5_LOCKED flags
3179                  */
3180                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3181         }
3182         s->to_write = 0;
3183         s->written = 0;
3184
3185         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3186                 if (atomic_dec_and_test(&conf->pending_full_writes))
3187                         md_wakeup_thread(conf->mddev->thread);
3188 }
3189
3190 static void
3191 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3192                    struct stripe_head_state *s)
3193 {
3194         int abort = 0;
3195         int i;
3196
3197         BUG_ON(sh->batch_head);
3198         clear_bit(STRIPE_SYNCING, &sh->state);
3199         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3200                 wake_up(&conf->wait_for_overlap);
3201         s->syncing = 0;
3202         s->replacing = 0;
3203         /* There is nothing more to do for sync/check/repair.
3204          * Don't even need to abort as that is handled elsewhere
3205          * if needed, and not always wanted e.g. if there is a known
3206          * bad block here.
3207          * For recover/replace we need to record a bad block on all
3208          * non-sync devices, or abort the recovery
3209          */
3210         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3211                 /* During recovery devices cannot be removed, so
3212                  * locking and refcounting of rdevs is not needed
3213                  */
3214                 for (i = 0; i < conf->raid_disks; i++) {
3215                         struct md_rdev *rdev = conf->disks[i].rdev;
3216                         if (rdev
3217                             && !test_bit(Faulty, &rdev->flags)
3218                             && !test_bit(In_sync, &rdev->flags)
3219                             && !rdev_set_badblocks(rdev, sh->sector,
3220                                                    STRIPE_SECTORS, 0))
3221                                 abort = 1;
3222                         rdev = conf->disks[i].replacement;
3223                         if (rdev
3224                             && !test_bit(Faulty, &rdev->flags)
3225                             && !test_bit(In_sync, &rdev->flags)
3226                             && !rdev_set_badblocks(rdev, sh->sector,
3227                                                    STRIPE_SECTORS, 0))
3228                                 abort = 1;
3229                 }
3230                 if (abort)
3231                         conf->recovery_disabled =
3232                                 conf->mddev->recovery_disabled;
3233         }
3234         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3235 }
3236
3237 static int want_replace(struct stripe_head *sh, int disk_idx)
3238 {
3239         struct md_rdev *rdev;
3240         int rv = 0;
3241         /* Doing recovery so rcu locking not required */
3242         rdev = sh->raid_conf->disks[disk_idx].replacement;
3243         if (rdev
3244             && !test_bit(Faulty, &rdev->flags)
3245             && !test_bit(In_sync, &rdev->flags)
3246             && (rdev->recovery_offset <= sh->sector
3247                 || rdev->mddev->recovery_cp <= sh->sector))
3248                 rv = 1;
3249
3250         return rv;
3251 }
3252
3253 /* fetch_block - checks the given member device to see if its data needs
3254  * to be read or computed to satisfy a request.
3255  *
3256  * Returns 1 when no more member devices need to be checked, otherwise returns
3257  * 0 to tell the loop in handle_stripe_fill to continue
3258  */
3259
3260 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3261                            int disk_idx, int disks)
3262 {
3263         struct r5dev *dev = &sh->dev[disk_idx];
3264         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3265                                   &sh->dev[s->failed_num[1]] };
3266         int i;
3267
3268
3269         if (test_bit(R5_LOCKED, &dev->flags) ||
3270             test_bit(R5_UPTODATE, &dev->flags))
3271                 /* No point reading this as we already have it or have
3272                  * decided to get it.
3273                  */
3274                 return 0;
3275
3276         if (dev->toread ||
3277             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3278                 /* We need this block to directly satisfy a request */
3279                 return 1;
3280
3281         if (s->syncing || s->expanding ||
3282             (s->replacing && want_replace(sh, disk_idx)))
3283                 /* When syncing, or expanding we read everything.
3284                  * When replacing, we need the replaced block.
3285                  */
3286                 return 1;
3287
3288         if ((s->failed >= 1 && fdev[0]->toread) ||
3289             (s->failed >= 2 && fdev[1]->toread))
3290                 /* If we want to read from a failed device, then
3291                  * we need to actually read every other device.
3292                  */
3293                 return 1;
3294
3295         /* Sometimes neither read-modify-write nor reconstruct-write
3296          * cycles can work.  In those cases we read every block we
3297          * can.  Then the parity-update is certain to have enough to
3298          * work with.
3299          * This can only be a problem when we need to write something,
3300          * and some device has failed.  If either of those tests
3301          * fail we need look no further.
3302          */
3303         if (!s->failed || !s->to_write)
3304                 return 0;
3305
3306         if (test_bit(R5_Insync, &dev->flags) &&
3307             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3308                 /* Pre-reads at not permitted until after short delay
3309                  * to gather multiple requests.  However if this
3310                  * device is no Insync, the block could only be be computed
3311                  * and there is no need to delay that.
3312                  */
3313                 return 0;
3314
3315         for (i = 0; i < s->failed && i < 2; i++) {
3316                 if (fdev[i]->towrite &&
3317                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3318                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3319                         /* If we have a partial write to a failed
3320                          * device, then we will need to reconstruct
3321                          * the content of that device, so all other
3322                          * devices must be read.
3323                          */
3324                         return 1;
3325         }
3326
3327         /* If we are forced to do a reconstruct-write, either because
3328          * the current RAID6 implementation only supports that, or
3329          * or because parity cannot be trusted and we are currently
3330          * recovering it, there is extra need to be careful.
3331          * If one of the devices that we would need to read, because
3332          * it is not being overwritten (and maybe not written at all)
3333          * is missing/faulty, then we need to read everything we can.
3334          */
3335         if (sh->raid_conf->level != 6 &&
3336             sh->sector < sh->raid_conf->mddev->recovery_cp)
3337                 /* reconstruct-write isn't being forced */
3338                 return 0;
3339         for (i = 0; i < s->failed && i < 2; i++) {
3340                 if (s->failed_num[i] != sh->pd_idx &&
3341                     s->failed_num[i] != sh->qd_idx &&
3342                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3343                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3344                         return 1;
3345         }
3346
3347         return 0;
3348 }
3349
3350 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3351                        int disk_idx, int disks)
3352 {
3353         struct r5dev *dev = &sh->dev[disk_idx];
3354
3355         /* is the data in this block needed, and can we get it? */
3356         if (need_this_block(sh, s, disk_idx, disks)) {
3357                 /* we would like to get this block, possibly by computing it,
3358                  * otherwise read it if the backing disk is insync
3359                  */
3360                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3361                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3362                 BUG_ON(sh->batch_head);
3363                 if ((s->uptodate == disks - 1) &&
3364                     (s->failed && (disk_idx == s->failed_num[0] ||
3365                                    disk_idx == s->failed_num[1]))) {
3366                         /* have disk failed, and we're requested to fetch it;
3367                          * do compute it
3368                          */
3369                         pr_debug("Computing stripe %llu block %d\n",
3370                                (unsigned long long)sh->sector, disk_idx);
3371                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3372                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3373                         set_bit(R5_Wantcompute, &dev->flags);
3374                         sh->ops.target = disk_idx;
3375                         sh->ops.target2 = -1; /* no 2nd target */
3376                         s->req_compute = 1;
3377                         /* Careful: from this point on 'uptodate' is in the eye
3378                          * of raid_run_ops which services 'compute' operations
3379                          * before writes. R5_Wantcompute flags a block that will
3380                          * be R5_UPTODATE by the time it is needed for a
3381                          * subsequent operation.
3382                          */
3383                         s->uptodate++;
3384                         return 1;
3385                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3386                         /* Computing 2-failure is *very* expensive; only
3387                          * do it if failed >= 2
3388                          */
3389                         int other;
3390                         for (other = disks; other--; ) {
3391                                 if (other == disk_idx)
3392                                         continue;
3393                                 if (!test_bit(R5_UPTODATE,
3394                                       &sh->dev[other].flags))
3395                                         break;
3396                         }
3397                         BUG_ON(other < 0);
3398                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3399                                (unsigned long long)sh->sector,
3400                                disk_idx, other);
3401                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3402                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3403                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3404                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3405                         sh->ops.target = disk_idx;
3406                         sh->ops.target2 = other;
3407                         s->uptodate += 2;
3408                         s->req_compute = 1;
3409                         return 1;
3410                 } else if (test_bit(R5_Insync, &dev->flags)) {
3411                         set_bit(R5_LOCKED, &dev->flags);
3412                         set_bit(R5_Wantread, &dev->flags);
3413                         s->locked++;
3414                         pr_debug("Reading block %d (sync=%d)\n",
3415                                 disk_idx, s->syncing);
3416                 }
3417         }
3418
3419         return 0;
3420 }
3421
3422 /**
3423  * handle_stripe_fill - read or compute data to satisfy pending requests.
3424  */
3425 static void handle_stripe_fill(struct stripe_head *sh,
3426                                struct stripe_head_state *s,
3427                                int disks)
3428 {
3429         int i;
3430
3431         /* look for blocks to read/compute, skip this if a compute
3432          * is already in flight, or if the stripe contents are in the
3433          * midst of changing due to a write
3434          */
3435         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3436             !sh->reconstruct_state)
3437                 for (i = disks; i--; )
3438                         if (fetch_block(sh, s, i, disks))
3439                                 break;
3440         set_bit(STRIPE_HANDLE, &sh->state);
3441 }
3442
3443 static void break_stripe_batch_list(struct stripe_head *head_sh,
3444                                     unsigned long handle_flags);
3445 /* handle_stripe_clean_event
3446  * any written block on an uptodate or failed drive can be returned.
3447  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3448  * never LOCKED, so we don't need to test 'failed' directly.
3449  */
3450 static void handle_stripe_clean_event(struct r5conf *conf,
3451         struct stripe_head *sh, int disks, struct bio_list *return_bi)
3452 {
3453         int i;
3454         struct r5dev *dev;
3455         int discard_pending = 0;
3456         struct stripe_head *head_sh = sh;
3457         bool do_endio = false;
3458
3459         for (i = disks; i--; )
3460                 if (sh->dev[i].written) {
3461                         dev = &sh->dev[i];
3462                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3463                             (test_bit(R5_UPTODATE, &dev->flags) ||
3464                              test_bit(R5_Discard, &dev->flags) ||
3465                              test_bit(R5_SkipCopy, &dev->flags))) {
3466                                 /* We can return any write requests */
3467                                 struct bio *wbi, *wbi2;
3468                                 pr_debug("Return write for disc %d\n", i);
3469                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3470                                         clear_bit(R5_UPTODATE, &dev->flags);
3471                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3472                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3473                                 }
3474                                 do_endio = true;
3475
3476 returnbi:
3477                                 dev->page = dev->orig_page;
3478                                 wbi = dev->written;
3479                                 dev->written = NULL;
3480                                 while (wbi && wbi->bi_iter.bi_sector <
3481                                         dev->sector + STRIPE_SECTORS) {
3482                                         wbi2 = r5_next_bio(wbi, dev->sector);
3483                                         if (!raid5_dec_bi_active_stripes(wbi)) {
3484                                                 md_write_end(conf->mddev);
3485                                                 bio_list_add(return_bi, wbi);
3486                                         }
3487                                         wbi = wbi2;
3488                                 }
3489                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3490                                                 STRIPE_SECTORS,
3491                                          !test_bit(STRIPE_DEGRADED, &sh->state),
3492                                                 0);
3493                                 if (head_sh->batch_head) {
3494                                         sh = list_first_entry(&sh->batch_list,
3495                                                               struct stripe_head,
3496                                                               batch_list);
3497                                         if (sh != head_sh) {
3498                                                 dev = &sh->dev[i];
3499                                                 goto returnbi;
3500                                         }
3501                                 }
3502                                 sh = head_sh;
3503                                 dev = &sh->dev[i];
3504                         } else if (test_bit(R5_Discard, &dev->flags))
3505                                 discard_pending = 1;
3506                         WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3507                         WARN_ON(dev->page != dev->orig_page);
3508                 }
3509
3510         r5l_stripe_write_finished(sh);
3511
3512         if (!discard_pending &&
3513             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3514                 int hash;
3515                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3516                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3517                 if (sh->qd_idx >= 0) {
3518                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3519                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3520                 }
3521                 /* now that discard is done we can proceed with any sync */
3522                 clear_bit(STRIPE_DISCARD, &sh->state);
3523                 /*
3524                  * SCSI discard will change some bio fields and the stripe has
3525                  * no updated data, so remove it from hash list and the stripe
3526                  * will be reinitialized
3527                  */
3528 unhash:
3529                 hash = sh->hash_lock_index;
3530                 spin_lock_irq(conf->hash_locks + hash);
3531                 remove_hash(sh);
3532                 spin_unlock_irq(conf->hash_locks + hash);
3533                 if (head_sh->batch_head) {
3534                         sh = list_first_entry(&sh->batch_list,
3535                                               struct stripe_head, batch_list);
3536                         if (sh != head_sh)
3537                                         goto unhash;
3538                 }
3539                 sh = head_sh;
3540
3541                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3542                         set_bit(STRIPE_HANDLE, &sh->state);
3543
3544         }
3545
3546         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3547                 if (atomic_dec_and_test(&conf->pending_full_writes))
3548                         md_wakeup_thread(conf->mddev->thread);
3549
3550         if (head_sh->batch_head && do_endio)
3551                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3552 }
3553
3554 static void handle_stripe_dirtying(struct r5conf *conf,
3555                                    struct stripe_head *sh,
3556                                    struct stripe_head_state *s,
3557                                    int disks)
3558 {
3559         int rmw = 0, rcw = 0, i;
3560         sector_t recovery_cp = conf->mddev->recovery_cp;
3561
3562         /* Check whether resync is now happening or should start.
3563          * If yes, then the array is dirty (after unclean shutdown or
3564          * initial creation), so parity in some stripes might be inconsistent.
3565          * In this case, we need to always do reconstruct-write, to ensure
3566          * that in case of drive failure or read-error correction, we
3567          * generate correct data from the parity.
3568          */
3569         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3570             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3571              s->failed == 0)) {
3572                 /* Calculate the real rcw later - for now make it
3573                  * look like rcw is cheaper
3574                  */
3575                 rcw = 1; rmw = 2;
3576                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3577                          conf->rmw_level, (unsigned long long)recovery_cp,
3578                          (unsigned long long)sh->sector);
3579         } else for (i = disks; i--; ) {
3580                 /* would I have to read this buffer for read_modify_write */
3581                 struct r5dev *dev = &sh->dev[i];
3582                 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3583                     !test_bit(R5_LOCKED, &dev->flags) &&
3584                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3585                       test_bit(R5_Wantcompute, &dev->flags))) {
3586                         if (test_bit(R5_Insync, &dev->flags))
3587                                 rmw++;
3588                         else
3589                                 rmw += 2*disks;  /* cannot read it */
3590                 }
3591                 /* Would I have to read this buffer for reconstruct_write */
3592                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3593                     i != sh->pd_idx && i != sh->qd_idx &&
3594                     !test_bit(R5_LOCKED, &dev->flags) &&
3595                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3596                     test_bit(R5_Wantcompute, &dev->flags))) {
3597                         if (test_bit(R5_Insync, &dev->flags))
3598                                 rcw++;
3599                         else
3600                                 rcw += 2*disks;
3601                 }
3602         }
3603         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3604                 (unsigned long long)sh->sector, rmw, rcw);
3605         set_bit(STRIPE_HANDLE, &sh->state);
3606         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3607                 /* prefer read-modify-write, but need to get some data */
3608                 if (conf->mddev->queue)
3609                         blk_add_trace_msg(conf->mddev->queue,
3610                                           "raid5 rmw %llu %d",
3611                                           (unsigned long long)sh->sector, rmw);
3612                 for (i = disks; i--; ) {
3613                         struct r5dev *dev = &sh->dev[i];
3614                         if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3615                             !test_bit(R5_LOCKED, &dev->flags) &&
3616                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3617                             test_bit(R5_Wantcompute, &dev->flags)) &&
3618                             test_bit(R5_Insync, &dev->flags)) {
3619                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3620                                              &sh->state)) {
3621                                         pr_debug("Read_old block %d for r-m-w\n",
3622                                                  i);
3623                                         set_bit(R5_LOCKED, &dev->flags);
3624                                         set_bit(R5_Wantread, &dev->flags);
3625                                         s->locked++;
3626                                 } else {
3627                                         set_bit(STRIPE_DELAYED, &sh->state);
3628                                         set_bit(STRIPE_HANDLE, &sh->state);
3629                                 }
3630                         }
3631                 }
3632         }
3633         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3634                 /* want reconstruct write, but need to get some data */
3635                 int qread =0;
3636                 rcw = 0;
3637                 for (i = disks; i--; ) {
3638                         struct r5dev *dev = &sh->dev[i];
3639                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3640                             i != sh->pd_idx && i != sh->qd_idx &&
3641                             !test_bit(R5_LOCKED, &dev->flags) &&
3642                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3643                               test_bit(R5_Wantcompute, &dev->flags))) {
3644                                 rcw++;
3645                                 if (test_bit(R5_Insync, &dev->flags) &&
3646                                     test_bit(STRIPE_PREREAD_ACTIVE,
3647                                              &sh->state)) {
3648                                         pr_debug("Read_old block "
3649                                                 "%d for Reconstruct\n", i);
3650                                         set_bit(R5_LOCKED, &dev->flags);
3651                                         set_bit(R5_Wantread, &dev->flags);
3652                                         s->locked++;
3653                                         qread++;
3654                                 } else {
3655                                         set_bit(STRIPE_DELAYED, &sh->state);
3656                                         set_bit(STRIPE_HANDLE, &sh->state);
3657                                 }
3658                         }
3659                 }
3660                 if (rcw && conf->mddev->queue)
3661                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3662                                           (unsigned long long)sh->sector,
3663                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3664         }
3665
3666         if (rcw > disks && rmw > disks &&
3667             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3668                 set_bit(STRIPE_DELAYED, &sh->state);
3669
3670         /* now if nothing is locked, and if we have enough data,
3671          * we can start a write request
3672          */
3673         /* since handle_stripe can be called at any time we need to handle the
3674          * case where a compute block operation has been submitted and then a
3675          * subsequent call wants to start a write request.  raid_run_ops only
3676          * handles the case where compute block and reconstruct are requested
3677          * simultaneously.  If this is not the case then new writes need to be
3678          * held off until the compute completes.
3679          */
3680         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3681             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3682             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3683                 schedule_reconstruction(sh, s, rcw == 0, 0);
3684 }
3685
3686 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3687                                 struct stripe_head_state *s, int disks)
3688 {
3689         struct r5dev *dev = NULL;
3690
3691         BUG_ON(sh->batch_head);
3692         set_bit(STRIPE_HANDLE, &sh->state);
3693
3694         switch (sh->check_state) {
3695         case check_state_idle:
3696                 /* start a new check operation if there are no failures */
3697                 if (s->failed == 0) {
3698                         BUG_ON(s->uptodate != disks);
3699                         sh->check_state = check_state_run;
3700                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3701                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3702                         s->uptodate--;
3703                         break;
3704                 }
3705                 dev = &sh->dev[s->failed_num[0]];
3706                 /* fall through */
3707         case check_state_compute_result:
3708                 sh->check_state = check_state_idle;
3709                 if (!dev)
3710                         dev = &sh->dev[sh->pd_idx];
3711
3712                 /* check that a write has not made the stripe insync */
3713                 if (test_bit(STRIPE_INSYNC, &sh->state))
3714                         break;
3715
3716                 /* either failed parity check, or recovery is happening */
3717                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3718                 BUG_ON(s->uptodate != disks);
3719
3720                 set_bit(R5_LOCKED, &dev->flags);
3721                 s->locked++;
3722                 set_bit(R5_Wantwrite, &dev->flags);
3723
3724                 clear_bit(STRIPE_DEGRADED, &sh->state);
3725                 set_bit(STRIPE_INSYNC, &sh->state);
3726                 break;
3727         case check_state_run:
3728                 break; /* we will be called again upon completion */
3729         case check_state_check_result:
3730                 sh->check_state = check_state_idle;
3731
3732                 /* if a failure occurred during the check operation, leave
3733                  * STRIPE_INSYNC not set and let the stripe be handled again
3734                  */
3735                 if (s->failed)
3736                         break;
3737
3738                 /* handle a successful check operation, if parity is correct
3739                  * we are done.  Otherwise update the mismatch count and repair
3740                  * parity if !MD_RECOVERY_CHECK
3741                  */
3742                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3743                         /* parity is correct (on disc,
3744                          * not in buffer any more)
3745                          */
3746                         set_bit(STRIPE_INSYNC, &sh->state);
3747                 else {
3748                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3749                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3750                                 /* don't try to repair!! */
3751                                 set_bit(STRIPE_INSYNC, &sh->state);
3752                         else {
3753                                 sh->check_state = check_state_compute_run;
3754                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3755                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3756                                 set_bit(R5_Wantcompute,
3757                                         &sh->dev[sh->pd_idx].flags);
3758                                 sh->ops.target = sh->pd_idx;
3759                                 sh->ops.target2 = -1;
3760                                 s->uptodate++;
3761                         }
3762                 }
3763                 break;
3764         case check_state_compute_run:
3765                 break;
3766         default:
3767                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3768                        __func__, sh->check_state,
3769                        (unsigned long long) sh->sector);
3770                 BUG();
3771         }
3772 }
3773
3774 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3775                                   struct stripe_head_state *s,
3776                                   int disks)
3777 {
3778         int pd_idx = sh->pd_idx;
3779         int qd_idx = sh->qd_idx;
3780         struct r5dev *dev;
3781
3782         BUG_ON(sh->batch_head);
3783         set_bit(STRIPE_HANDLE, &sh->state);
3784
3785         BUG_ON(s->failed > 2);
3786
3787         /* Want to check and possibly repair P and Q.
3788          * However there could be one 'failed' device, in which
3789          * case we can only check one of them, possibly using the
3790          * other to generate missing data
3791          */
3792
3793         switch (sh->check_state) {
3794         case check_state_idle:
3795                 /* start a new check operation if there are < 2 failures */
3796                 if (s->failed == s->q_failed) {
3797                         /* The only possible failed device holds Q, so it
3798                          * makes sense to check P (If anything else were failed,
3799                          * we would have used P to recreate it).
3800                          */
3801                         sh->check_state = check_state_run;
3802                 }
3803                 if (!s->q_failed && s->failed < 2) {
3804                         /* Q is not failed, and we didn't use it to generate
3805                          * anything, so it makes sense to check it
3806                          */
3807                         if (sh->check_state == check_state_run)
3808                                 sh->check_state = check_state_run_pq;
3809                         else
3810                                 sh->check_state = check_state_run_q;
3811                 }
3812
3813                 /* discard potentially stale zero_sum_result */
3814                 sh->ops.zero_sum_result = 0;
3815
3816                 if (sh->check_state == check_state_run) {
3817                         /* async_xor_zero_sum destroys the contents of P */
3818                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3819                         s->uptodate--;
3820                 }
3821                 if (sh->check_state >= check_state_run &&
3822                     sh->check_state <= check_state_run_pq) {
3823                         /* async_syndrome_zero_sum preserves P and Q, so
3824                          * no need to mark them !uptodate here
3825                          */
3826                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3827                         break;
3828                 }
3829
3830                 /* we have 2-disk failure */
3831                 BUG_ON(s->failed != 2);
3832                 /* fall through */
3833         case check_state_compute_result:
3834                 sh->check_state = check_state_idle;
3835
3836                 /* check that a write has not made the stripe insync */
3837                 if (test_bit(STRIPE_INSYNC, &sh->state))
3838                         break;
3839
3840                 /* now write out any block on a failed drive,
3841                  * or P or Q if they were recomputed
3842                  */
3843                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3844                 if (s->failed == 2) {
3845                         dev = &sh->dev[s->failed_num[1]];
3846                         s->locked++;
3847                         set_bit(R5_LOCKED, &dev->flags);
3848                         set_bit(R5_Wantwrite, &dev->flags);
3849                 }
3850                 if (s->failed >= 1) {
3851                         dev = &sh->dev[s->failed_num[0]];
3852                         s->locked++;
3853                         set_bit(R5_LOCKED, &dev->flags);
3854                         set_bit(R5_Wantwrite, &dev->flags);
3855                 }
3856                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3857                         dev = &sh->dev[pd_idx];
3858                         s->locked++;
3859                         set_bit(R5_LOCKED, &dev->flags);
3860                         set_bit(R5_Wantwrite, &dev->flags);
3861                 }
3862                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3863                         dev = &sh->dev[qd_idx];
3864                         s->locked++;
3865                         set_bit(R5_LOCKED, &dev->flags);
3866                         set_bit(R5_Wantwrite, &dev->flags);
3867                 }
3868                 clear_bit(STRIPE_DEGRADED, &sh->state);
3869
3870                 set_bit(STRIPE_INSYNC, &sh->state);
3871                 break;
3872         case check_state_run:
3873         case check_state_run_q:
3874         case check_state_run_pq:
3875                 break; /* we will be called again upon completion */
3876         case check_state_check_result:
3877                 sh->check_state = check_state_idle;
3878
3879                 /* handle a successful check operation, if parity is correct
3880                  * we are done.  Otherwise update the mismatch count and repair
3881                  * parity if !MD_RECOVERY_CHECK
3882                  */
3883                 if (sh->ops.zero_sum_result == 0) {
3884                         /* both parities are correct */
3885                         if (!s->failed)
3886                                 set_bit(STRIPE_INSYNC, &sh->state);
3887                         else {
3888                                 /* in contrast to the raid5 case we can validate
3889                                  * parity, but still have a failure to write
3890                                  * back
3891                                  */
3892                                 sh->check_state = check_state_compute_result;
3893                                 /* Returning at this point means that we may go
3894                                  * off and bring p and/or q uptodate again so
3895                                  * we make sure to check zero_sum_result again
3896                                  * to verify if p or q need writeback
3897                                  */
3898                         }
3899                 } else {
3900                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3901                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3902                                 /* don't try to repair!! */
3903                                 set_bit(STRIPE_INSYNC, &sh->state);
3904                         else {
3905                                 int *target = &sh->ops.target;
3906
3907                                 sh->ops.target = -1;
3908                                 sh->ops.target2 = -1;
3909                                 sh->check_state = check_state_compute_run;
3910                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3911                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3912                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3913                                         set_bit(R5_Wantcompute,
3914                                                 &sh->dev[pd_idx].flags);
3915                                         *target = pd_idx;
3916                                         target = &sh->ops.target2;
3917                                         s->uptodate++;
3918                                 }
3919                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3920                                         set_bit(R5_Wantcompute,
3921                                                 &sh->dev[qd_idx].flags);
3922                                         *target = qd_idx;
3923                                         s->uptodate++;
3924                                 }
3925                         }
3926                 }
3927                 break;
3928         case check_state_compute_run:
3929                 break;
3930         default:
3931                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3932                        __func__, sh->check_state,
3933                        (unsigned long long) sh->sector);
3934                 BUG();
3935         }
3936 }
3937
3938 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3939 {
3940         int i;
3941
3942         /* We have read all the blocks in this stripe and now we need to
3943          * copy some of them into a target stripe for expand.
3944          */
3945         struct dma_async_tx_descriptor *tx = NULL;
3946         BUG_ON(sh->batch_head);
3947         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3948         for (i = 0; i < sh->disks; i++)
3949                 if (i != sh->pd_idx && i != sh->qd_idx) {
3950                         int dd_idx, j;
3951                         struct stripe_head *sh2;
3952                         struct async_submit_ctl submit;
3953
3954                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
3955                         sector_t s = raid5_compute_sector(conf, bn, 0,
3956                                                           &dd_idx, NULL);
3957                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3958                         if (sh2 == NULL)
3959                                 /* so far only the early blocks of this stripe
3960                                  * have been requested.  When later blocks
3961                                  * get requested, we will try again
3962                                  */
3963                                 continue;
3964                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3965                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3966                                 /* must have already done this block */
3967                                 raid5_release_stripe(sh2);
3968                                 continue;
3969                         }
3970
3971                         /* place all the copies on one channel */
3972                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3973                         tx = async_memcpy(sh2->dev[dd_idx].page,
3974                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
3975                                           &submit);
3976
3977                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3978                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3979                         for (j = 0; j < conf->raid_disks; j++)
3980                                 if (j != sh2->pd_idx &&
3981                                     j != sh2->qd_idx &&
3982                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
3983                                         break;
3984                         if (j == conf->raid_disks) {
3985                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3986                                 set_bit(STRIPE_HANDLE, &sh2->state);
3987                         }
3988                         raid5_release_stripe(sh2);
3989
3990                 }
3991         /* done submitting copies, wait for them to complete */
3992         async_tx_quiesce(&tx);
3993 }
3994
3995 /*
3996  * handle_stripe - do things to a stripe.
3997  *
3998  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3999  * state of various bits to see what needs to be done.
4000  * Possible results:
4001  *    return some read requests which now have data
4002  *    return some write requests which are safely on storage
4003  *    schedule a read on some buffers
4004  *    schedule a write of some buffers
4005  *    return confirmation of parity correctness
4006  *
4007  */
4008
4009 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4010 {
4011         struct r5conf *conf = sh->raid_conf;
4012         int disks = sh->disks;
4013         struct r5dev *dev;
4014         int i;
4015         int do_recovery = 0;
4016
4017         memset(s, 0, sizeof(*s));
4018
4019         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4020         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4021         s->failed_num[0] = -1;
4022         s->failed_num[1] = -1;
4023         s->log_failed = r5l_log_disk_error(conf);
4024
4025         /* Now to look around and see what can be done */
4026         rcu_read_lock();
4027         for (i=disks; i--; ) {
4028                 struct md_rdev *rdev;
4029                 sector_t first_bad;
4030                 int bad_sectors;
4031                 int is_bad = 0;
4032
4033                 dev = &sh->dev[i];
4034
4035                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4036                          i, dev->flags,
4037                          dev->toread, dev->towrite, dev->written);
4038                 /* maybe we can reply to a read
4039                  *
4040                  * new wantfill requests are only permitted while
4041                  * ops_complete_biofill is guaranteed to be inactive
4042                  */
4043                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4044                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4045                         set_bit(R5_Wantfill, &dev->flags);
4046
4047                 /* now count some things */
4048                 if (test_bit(R5_LOCKED, &dev->flags))
4049                         s->locked++;
4050                 if (test_bit(R5_UPTODATE, &dev->flags))
4051                         s->uptodate++;
4052                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4053                         s->compute++;
4054                         BUG_ON(s->compute > 2);
4055                 }
4056
4057                 if (test_bit(R5_Wantfill, &dev->flags))
4058                         s->to_fill++;
4059                 else if (dev->toread)
4060                         s->to_read++;
4061                 if (dev->towrite) {
4062                         s->to_write++;
4063                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4064                                 s->non_overwrite++;
4065                 }
4066                 if (dev->written)
4067                         s->written++;
4068                 /* Prefer to use the replacement for reads, but only
4069                  * if it is recovered enough and has no bad blocks.
4070                  */
4071                 rdev = rcu_dereference(conf->disks[i].replacement);
4072                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4073                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4074                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4075                                  &first_bad, &bad_sectors))
4076                         set_bit(R5_ReadRepl, &dev->flags);
4077                 else {
4078                         if (rdev && !test_bit(Faulty, &rdev->flags))
4079                                 set_bit(R5_NeedReplace, &dev->flags);
4080                         else
4081                                 clear_bit(R5_NeedReplace, &dev->flags);
4082                         rdev = rcu_dereference(conf->disks[i].rdev);
4083                         clear_bit(R5_ReadRepl, &dev->flags);
4084                 }
4085                 if (rdev && test_bit(Faulty, &rdev->flags))
4086                         rdev = NULL;
4087                 if (rdev) {
4088                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4089                                              &first_bad, &bad_sectors);
4090                         if (s->blocked_rdev == NULL
4091                             && (test_bit(Blocked, &rdev->flags)
4092                                 || is_bad < 0)) {
4093                                 if (is_bad < 0)
4094                                         set_bit(BlockedBadBlocks,
4095                                                 &rdev->flags);
4096                                 s->blocked_rdev = rdev;
4097                                 atomic_inc(&rdev->nr_pending);
4098                         }
4099                 }
4100                 clear_bit(R5_Insync, &dev->flags);
4101                 if (!rdev)
4102                         /* Not in-sync */;
4103                 else if (is_bad) {
4104                         /* also not in-sync */
4105                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4106                             test_bit(R5_UPTODATE, &dev->flags)) {
4107                                 /* treat as in-sync, but with a read error
4108                                  * which we can now try to correct
4109                                  */
4110                                 set_bit(R5_Insync, &dev->flags);
4111                                 set_bit(R5_ReadError, &dev->flags);
4112                         }
4113                 } else if (test_bit(In_sync, &rdev->flags))
4114                         set_bit(R5_Insync, &dev->flags);
4115                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4116                         /* in sync if before recovery_offset */
4117                         set_bit(R5_Insync, &dev->flags);
4118                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4119                          test_bit(R5_Expanded, &dev->flags))
4120                         /* If we've reshaped into here, we assume it is Insync.
4121                          * We will shortly update recovery_offset to make
4122                          * it official.
4123                          */
4124                         set_bit(R5_Insync, &dev->flags);
4125
4126                 if (test_bit(R5_WriteError, &dev->flags)) {
4127                         /* This flag does not apply to '.replacement'
4128                          * only to .rdev, so make sure to check that*/
4129                         struct md_rdev *rdev2 = rcu_dereference(
4130                                 conf->disks[i].rdev);
4131                         if (rdev2 == rdev)
4132                                 clear_bit(R5_Insync, &dev->flags);
4133                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4134                                 s->handle_bad_blocks = 1;
4135                                 atomic_inc(&rdev2->nr_pending);
4136                         } else
4137                                 clear_bit(R5_WriteError, &dev->flags);
4138                 }
4139                 if (test_bit(R5_MadeGood, &dev->flags)) {
4140                         /* This flag does not apply to '.replacement'
4141                          * only to .rdev, so make sure to check that*/
4142                         struct md_rdev *rdev2 = rcu_dereference(
4143                                 conf->disks[i].rdev);
4144                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4145                                 s->handle_bad_blocks = 1;
4146                                 atomic_inc(&rdev2->nr_pending);
4147                         } else
4148                                 clear_bit(R5_MadeGood, &dev->flags);
4149                 }
4150                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4151                         struct md_rdev *rdev2 = rcu_dereference(
4152                                 conf->disks[i].replacement);
4153                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4154                                 s->handle_bad_blocks = 1;
4155                                 atomic_inc(&rdev2->nr_pending);
4156                         } else
4157                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4158                 }
4159                 if (!test_bit(R5_Insync, &dev->flags)) {
4160                         /* The ReadError flag will just be confusing now */
4161                         clear_bit(R5_ReadError, &dev->flags);
4162                         clear_bit(R5_ReWrite, &dev->flags);
4163                 }
4164                 if (test_bit(R5_ReadError, &dev->flags))
4165                         clear_bit(R5_Insync, &dev->flags);
4166                 if (!test_bit(R5_Insync, &dev->flags)) {
4167                         if (s->failed < 2)
4168                                 s->failed_num[s->failed] = i;
4169                         s->failed++;
4170                         if (rdev && !test_bit(Faulty, &rdev->flags))
4171                                 do_recovery = 1;
4172                 }
4173         }
4174         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4175                 /* If there is a failed device being replaced,
4176                  *     we must be recovering.
4177                  * else if we are after recovery_cp, we must be syncing
4178                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4179                  * else we can only be replacing
4180                  * sync and recovery both need to read all devices, and so
4181                  * use the same flag.
4182                  */
4183                 if (do_recovery ||
4184                     sh->sector >= conf->mddev->recovery_cp ||
4185                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4186                         s->syncing = 1;
4187                 else
4188                         s->replacing = 1;
4189         }
4190         rcu_read_unlock();
4191 }
4192
4193 static int clear_batch_ready(struct stripe_head *sh)
4194 {
4195         /* Return '1' if this is a member of batch, or
4196          * '0' if it is a lone stripe or a head which can now be
4197          * handled.
4198          */
4199         struct stripe_head *tmp;
4200         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4201                 return (sh->batch_head && sh->batch_head != sh);
4202         spin_lock(&sh->stripe_lock);
4203         if (!sh->batch_head) {
4204                 spin_unlock(&sh->stripe_lock);
4205                 return 0;
4206         }
4207
4208         /*
4209          * this stripe could be added to a batch list before we check
4210          * BATCH_READY, skips it
4211          */
4212         if (sh->batch_head != sh) {
4213                 spin_unlock(&sh->stripe_lock);
4214                 return 1;
4215         }
4216         spin_lock(&sh->batch_lock);
4217         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4218                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4219         spin_unlock(&sh->batch_lock);
4220         spin_unlock(&sh->stripe_lock);
4221
4222         /*
4223          * BATCH_READY is cleared, no new stripes can be added.
4224          * batch_list can be accessed without lock
4225          */
4226         return 0;
4227 }
4228
4229 static void break_stripe_batch_list(struct stripe_head *head_sh,
4230                                     unsigned long handle_flags)
4231 {
4232         struct stripe_head *sh, *next;
4233         int i;
4234         int do_wakeup = 0;
4235
4236         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4237
4238                 list_del_init(&sh->batch_list);
4239
4240                 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4241                                           (1 << STRIPE_SYNCING) |
4242                                           (1 << STRIPE_REPLACED) |
4243                                           (1 << STRIPE_PREREAD_ACTIVE) |
4244                                           (1 << STRIPE_DELAYED) |
4245                                           (1 << STRIPE_BIT_DELAY) |
4246                                           (1 << STRIPE_FULL_WRITE) |
4247                                           (1 << STRIPE_BIOFILL_RUN) |
4248                                           (1 << STRIPE_COMPUTE_RUN)  |
4249                                           (1 << STRIPE_OPS_REQ_PENDING) |
4250                                           (1 << STRIPE_DISCARD) |
4251                                           (1 << STRIPE_BATCH_READY) |
4252                                           (1 << STRIPE_BATCH_ERR) |
4253                                           (1 << STRIPE_BITMAP_PENDING)));
4254                 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4255                                               (1 << STRIPE_REPLACED)));
4256
4257                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4258                                             (1 << STRIPE_DEGRADED)),
4259                               head_sh->state & (1 << STRIPE_INSYNC));
4260
4261                 sh->check_state = head_sh->check_state;
4262                 sh->reconstruct_state = head_sh->reconstruct_state;
4263                 for (i = 0; i < sh->disks; i++) {
4264                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4265                                 do_wakeup = 1;
4266                         sh->dev[i].flags = head_sh->dev[i].flags &
4267                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4268                 }
4269                 spin_lock_irq(&sh->stripe_lock);
4270                 sh->batch_head = NULL;
4271                 spin_unlock_irq(&sh->stripe_lock);
4272                 if (handle_flags == 0 ||
4273                     sh->state & handle_flags)
4274                         set_bit(STRIPE_HANDLE, &sh->state);
4275                 raid5_release_stripe(sh);
4276         }
4277         spin_lock_irq(&head_sh->stripe_lock);
4278         head_sh->batch_head = NULL;
4279         spin_unlock_irq(&head_sh->stripe_lock);
4280         for (i = 0; i < head_sh->disks; i++)
4281                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4282                         do_wakeup = 1;
4283         if (head_sh->state & handle_flags)
4284                 set_bit(STRIPE_HANDLE, &head_sh->state);
4285
4286         if (do_wakeup)
4287                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4288 }
4289
4290 static void handle_stripe(struct stripe_head *sh)
4291 {
4292         struct stripe_head_state s;
4293         struct r5conf *conf = sh->raid_conf;
4294         int i;
4295         int prexor;
4296         int disks = sh->disks;
4297         struct r5dev *pdev, *qdev;
4298
4299         clear_bit(STRIPE_HANDLE, &sh->state);
4300         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4301                 /* already being handled, ensure it gets handled
4302                  * again when current action finishes */
4303                 set_bit(STRIPE_HANDLE, &sh->state);
4304                 return;
4305         }
4306
4307         if (clear_batch_ready(sh) ) {
4308                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4309                 return;
4310         }
4311
4312         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4313                 break_stripe_batch_list(sh, 0);
4314
4315         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4316                 spin_lock(&sh->stripe_lock);
4317                 /* Cannot process 'sync' concurrently with 'discard' */
4318                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4319                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4320                         set_bit(STRIPE_SYNCING, &sh->state);
4321                         clear_bit(STRIPE_INSYNC, &sh->state);
4322                         clear_bit(STRIPE_REPLACED, &sh->state);
4323                 }
4324                 spin_unlock(&sh->stripe_lock);
4325         }
4326         clear_bit(STRIPE_DELAYED, &sh->state);
4327
4328         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4329                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4330                (unsigned long long)sh->sector, sh->state,
4331                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4332                sh->check_state, sh->reconstruct_state);
4333
4334         analyse_stripe(sh, &s);
4335
4336         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4337                 goto finish;
4338
4339         if (s.handle_bad_blocks) {
4340                 set_bit(STRIPE_HANDLE, &sh->state);
4341                 goto finish;
4342         }
4343
4344         if (unlikely(s.blocked_rdev)) {
4345                 if (s.syncing || s.expanding || s.expanded ||
4346                     s.replacing || s.to_write || s.written) {
4347                         set_bit(STRIPE_HANDLE, &sh->state);
4348                         goto finish;
4349                 }
4350                 /* There is nothing for the blocked_rdev to block */
4351                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4352                 s.blocked_rdev = NULL;
4353         }
4354
4355         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4356                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4357                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4358         }
4359
4360         pr_debug("locked=%d uptodate=%d to_read=%d"
4361                " to_write=%d failed=%d failed_num=%d,%d\n",
4362                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4363                s.failed_num[0], s.failed_num[1]);
4364         /* check if the array has lost more than max_degraded devices and,
4365          * if so, some requests might need to be failed.
4366          */
4367         if (s.failed > conf->max_degraded || s.log_failed) {
4368                 sh->check_state = 0;
4369                 sh->reconstruct_state = 0;
4370                 break_stripe_batch_list(sh, 0);
4371                 if (s.to_read+s.to_write+s.written)
4372                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4373                 if (s.syncing + s.replacing)
4374                         handle_failed_sync(conf, sh, &s);
4375         }
4376
4377         /* Now we check to see if any write operations have recently
4378          * completed
4379          */
4380         prexor = 0;
4381         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4382                 prexor = 1;
4383         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4384             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4385                 sh->reconstruct_state = reconstruct_state_idle;
4386
4387                 /* All the 'written' buffers and the parity block are ready to
4388                  * be written back to disk
4389                  */
4390                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4391                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4392                 BUG_ON(sh->qd_idx >= 0 &&
4393                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4394                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4395                 for (i = disks; i--; ) {
4396                         struct r5dev *dev = &sh->dev[i];
4397                         if (test_bit(R5_LOCKED, &dev->flags) &&
4398                                 (i == sh->pd_idx || i == sh->qd_idx ||
4399                                  dev->written)) {
4400                                 pr_debug("Writing block %d\n", i);
4401                                 set_bit(R5_Wantwrite, &dev->flags);
4402                                 if (prexor)
4403                                         continue;
4404                                 if (s.failed > 1)
4405                                         continue;
4406                                 if (!test_bit(R5_Insync, &dev->flags) ||
4407                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4408                                      s.failed == 0))
4409                                         set_bit(STRIPE_INSYNC, &sh->state);
4410                         }
4411                 }
4412                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4413                         s.dec_preread_active = 1;
4414         }
4415
4416         /*
4417          * might be able to return some write requests if the parity blocks
4418          * are safe, or on a failed drive
4419          */
4420         pdev = &sh->dev[sh->pd_idx];
4421         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4422                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4423         qdev = &sh->dev[sh->qd_idx];
4424         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4425                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4426                 || conf->level < 6;
4427
4428         if (s.written &&
4429             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4430                              && !test_bit(R5_LOCKED, &pdev->flags)
4431                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4432                                  test_bit(R5_Discard, &pdev->flags))))) &&
4433             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4434                              && !test_bit(R5_LOCKED, &qdev->flags)
4435                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4436                                  test_bit(R5_Discard, &qdev->flags))))))
4437                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4438
4439         /* Now we might consider reading some blocks, either to check/generate
4440          * parity, or to satisfy requests
4441          * or to load a block that is being partially written.
4442          */
4443         if (s.to_read || s.non_overwrite
4444             || (conf->level == 6 && s.to_write && s.failed)
4445             || (s.syncing && (s.uptodate + s.compute < disks))
4446             || s.replacing
4447             || s.expanding)
4448                 handle_stripe_fill(sh, &s, disks);
4449
4450         /* Now to consider new write requests and what else, if anything
4451          * should be read.  We do not handle new writes when:
4452          * 1/ A 'write' operation (copy+xor) is already in flight.
4453          * 2/ A 'check' operation is in flight, as it may clobber the parity
4454          *    block.
4455          */
4456         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4457                 handle_stripe_dirtying(conf, sh, &s, disks);
4458
4459         /* maybe we need to check and possibly fix the parity for this stripe
4460          * Any reads will already have been scheduled, so we just see if enough
4461          * data is available.  The parity check is held off while parity
4462          * dependent operations are in flight.
4463          */
4464         if (sh->check_state ||
4465             (s.syncing && s.locked == 0 &&
4466              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4467              !test_bit(STRIPE_INSYNC, &sh->state))) {
4468                 if (conf->level == 6)
4469                         handle_parity_checks6(conf, sh, &s, disks);
4470                 else
4471                         handle_parity_checks5(conf, sh, &s, disks);
4472         }
4473
4474         if ((s.replacing || s.syncing) && s.locked == 0
4475             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4476             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4477                 /* Write out to replacement devices where possible */
4478                 for (i = 0; i < conf->raid_disks; i++)
4479                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4480                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4481                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4482                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4483                                 s.locked++;
4484                         }
4485                 if (s.replacing)
4486                         set_bit(STRIPE_INSYNC, &sh->state);
4487                 set_bit(STRIPE_REPLACED, &sh->state);
4488         }
4489         if ((s.syncing || s.replacing) && s.locked == 0 &&
4490             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4491             test_bit(STRIPE_INSYNC, &sh->state)) {
4492                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4493                 clear_bit(STRIPE_SYNCING, &sh->state);
4494                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4495                         wake_up(&conf->wait_for_overlap);
4496         }
4497
4498         /* If the failed drives are just a ReadError, then we might need
4499          * to progress the repair/check process
4500          */
4501         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4502                 for (i = 0; i < s.failed; i++) {
4503                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4504                         if (test_bit(R5_ReadError, &dev->flags)
4505                             && !test_bit(R5_LOCKED, &dev->flags)
4506                             && test_bit(R5_UPTODATE, &dev->flags)
4507                                 ) {
4508                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4509                                         set_bit(R5_Wantwrite, &dev->flags);
4510                                         set_bit(R5_ReWrite, &dev->flags);
4511                                         set_bit(R5_LOCKED, &dev->flags);
4512                                         s.locked++;
4513                                 } else {
4514                                         /* let's read it back */
4515                                         set_bit(R5_Wantread, &dev->flags);
4516                                         set_bit(R5_LOCKED, &dev->flags);
4517                                         s.locked++;
4518                                 }
4519                         }
4520                 }
4521
4522         /* Finish reconstruct operations initiated by the expansion process */
4523         if (sh->reconstruct_state == reconstruct_state_result) {
4524                 struct stripe_head *sh_src
4525                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4526                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4527                         /* sh cannot be written until sh_src has been read.
4528                          * so arrange for sh to be delayed a little
4529                          */
4530                         set_bit(STRIPE_DELAYED, &sh->state);
4531                         set_bit(STRIPE_HANDLE, &sh->state);
4532                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4533                                               &sh_src->state))
4534                                 atomic_inc(&conf->preread_active_stripes);
4535                         raid5_release_stripe(sh_src);
4536                         goto finish;
4537                 }
4538                 if (sh_src)
4539                         raid5_release_stripe(sh_src);
4540
4541                 sh->reconstruct_state = reconstruct_state_idle;
4542                 clear_bit(STRIPE_EXPANDING, &sh->state);
4543                 for (i = conf->raid_disks; i--; ) {
4544                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4545                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4546                         s.locked++;
4547                 }
4548         }
4549
4550         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4551             !sh->reconstruct_state) {
4552                 /* Need to write out all blocks after computing parity */
4553                 sh->disks = conf->raid_disks;
4554                 stripe_set_idx(sh->sector, conf, 0, sh);
4555                 schedule_reconstruction(sh, &s, 1, 1);
4556         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4557                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4558                 atomic_dec(&conf->reshape_stripes);
4559                 wake_up(&conf->wait_for_overlap);
4560                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4561         }
4562
4563         if (s.expanding && s.locked == 0 &&
4564             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4565                 handle_stripe_expansion(conf, sh);
4566
4567 finish:
4568         /* wait for this device to become unblocked */
4569         if (unlikely(s.blocked_rdev)) {
4570                 if (conf->mddev->external)
4571                         md_wait_for_blocked_rdev(s.blocked_rdev,
4572                                                  conf->mddev);
4573                 else
4574                         /* Internal metadata will immediately
4575                          * be written by raid5d, so we don't
4576                          * need to wait here.
4577                          */
4578                         rdev_dec_pending(s.blocked_rdev,
4579                                          conf->mddev);
4580         }
4581
4582         if (s.handle_bad_blocks)
4583                 for (i = disks; i--; ) {
4584                         struct md_rdev *rdev;
4585                         struct r5dev *dev = &sh->dev[i];
4586                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4587                                 /* We own a safe reference to the rdev */
4588                                 rdev = conf->disks[i].rdev;
4589                                 if (!rdev_set_badblocks(rdev, sh->sector,
4590                                                         STRIPE_SECTORS, 0))
4591                                         md_error(conf->mddev, rdev);
4592                                 rdev_dec_pending(rdev, conf->mddev);
4593                         }
4594                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4595                                 rdev = conf->disks[i].rdev;
4596                                 rdev_clear_badblocks(rdev, sh->sector,
4597                                                      STRIPE_SECTORS, 0);
4598                                 rdev_dec_pending(rdev, conf->mddev);
4599                         }
4600                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4601                                 rdev = conf->disks[i].replacement;
4602                                 if (!rdev)
4603                                         /* rdev have been moved down */
4604                                         rdev = conf->disks[i].rdev;
4605                                 rdev_clear_badblocks(rdev, sh->sector,
4606                                                      STRIPE_SECTORS, 0);
4607                                 rdev_dec_pending(rdev, conf->mddev);
4608                         }
4609                 }
4610
4611         if (s.ops_request)
4612                 raid_run_ops(sh, s.ops_request);
4613
4614         ops_run_io(sh, &s);
4615
4616         if (s.dec_preread_active) {
4617                 /* We delay this until after ops_run_io so that if make_request
4618                  * is waiting on a flush, it won't continue until the writes
4619                  * have actually been submitted.
4620                  */
4621                 atomic_dec(&conf->preread_active_stripes);
4622                 if (atomic_read(&conf->preread_active_stripes) <
4623                     IO_THRESHOLD)
4624                         md_wakeup_thread(conf->mddev->thread);
4625         }
4626
4627         if (!bio_list_empty(&s.return_bi)) {
4628                 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags)) {
4629                         spin_lock_irq(&conf->device_lock);
4630                         bio_list_merge(&conf->return_bi, &s.return_bi);
4631                         spin_unlock_irq(&conf->device_lock);
4632                         md_wakeup_thread(conf->mddev->thread);
4633                 } else
4634                         return_io(&s.return_bi);
4635         }
4636
4637         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4638 }
4639
4640 static void raid5_activate_delayed(struct r5conf *conf)
4641 {
4642         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4643                 while (!list_empty(&conf->delayed_list)) {
4644                         struct list_head *l = conf->delayed_list.next;
4645                         struct stripe_head *sh;
4646                         sh = list_entry(l, struct stripe_head, lru);
4647                         list_del_init(l);
4648                         clear_bit(STRIPE_DELAYED, &sh->state);
4649                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4650                                 atomic_inc(&conf->preread_active_stripes);
4651                         list_add_tail(&sh->lru, &conf->hold_list);
4652                         raid5_wakeup_stripe_thread(sh);
4653                 }
4654         }
4655 }
4656
4657 static void activate_bit_delay(struct r5conf *conf,
4658         struct list_head *temp_inactive_list)
4659 {
4660         /* device_lock is held */
4661         struct list_head head;
4662         list_add(&head, &conf->bitmap_list);
4663         list_del_init(&conf->bitmap_list);
4664         while (!list_empty(&head)) {
4665                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4666                 int hash;
4667                 list_del_init(&sh->lru);
4668                 atomic_inc(&sh->count);
4669                 hash = sh->hash_lock_index;
4670                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4671         }
4672 }
4673
4674 static int raid5_congested(struct mddev *mddev, int bits)
4675 {
4676         struct r5conf *conf = mddev->private;
4677
4678         /* No difference between reads and writes.  Just check
4679          * how busy the stripe_cache is
4680          */
4681
4682         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4683                 return 1;
4684         if (conf->quiesce)
4685                 return 1;
4686         if (atomic_read(&conf->empty_inactive_list_nr))
4687                 return 1;
4688
4689         return 0;
4690 }
4691
4692 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4693 {
4694         struct r5conf *conf = mddev->private;
4695         sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4696         unsigned int chunk_sectors;
4697         unsigned int bio_sectors = bio_sectors(bio);
4698
4699         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4700         return  chunk_sectors >=
4701                 ((sector & (chunk_sectors - 1)) + bio_sectors);
4702 }
4703
4704 /*
4705  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
4706  *  later sampled by raid5d.
4707  */
4708 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4709 {
4710         unsigned long flags;
4711
4712         spin_lock_irqsave(&conf->device_lock, flags);
4713
4714         bi->bi_next = conf->retry_read_aligned_list;
4715         conf->retry_read_aligned_list = bi;
4716
4717         spin_unlock_irqrestore(&conf->device_lock, flags);
4718         md_wakeup_thread(conf->mddev->thread);
4719 }
4720
4721 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4722 {
4723         struct bio *bi;
4724
4725         bi = conf->retry_read_aligned;
4726         if (bi) {
4727                 conf->retry_read_aligned = NULL;
4728                 return bi;
4729         }
4730         bi = conf->retry_read_aligned_list;
4731         if(bi) {
4732                 conf->retry_read_aligned_list = bi->bi_next;
4733                 bi->bi_next = NULL;
4734                 /*
4735                  * this sets the active strip count to 1 and the processed
4736                  * strip count to zero (upper 8 bits)
4737                  */
4738                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4739         }
4740
4741         return bi;
4742 }
4743
4744 /*
4745  *  The "raid5_align_endio" should check if the read succeeded and if it
4746  *  did, call bio_endio on the original bio (having bio_put the new bio
4747  *  first).
4748  *  If the read failed..
4749  */
4750 static void raid5_align_endio(struct bio *bi)
4751 {
4752         struct bio* raid_bi  = bi->bi_private;
4753         struct mddev *mddev;
4754         struct r5conf *conf;
4755         struct md_rdev *rdev;
4756         int error = bi->bi_error;
4757
4758         bio_put(bi);
4759
4760         rdev = (void*)raid_bi->bi_next;
4761         raid_bi->bi_next = NULL;
4762         mddev = rdev->mddev;
4763         conf = mddev->private;
4764
4765         rdev_dec_pending(rdev, conf->mddev);
4766
4767         if (!error) {
4768                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4769                                          raid_bi, 0);
4770                 bio_endio(raid_bi);
4771                 if (atomic_dec_and_test(&conf->active_aligned_reads))
4772                         wake_up(&conf->wait_for_quiescent);
4773                 return;
4774         }
4775
4776         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4777
4778         add_bio_to_retry(raid_bi, conf);
4779 }
4780
4781 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4782 {
4783         struct r5conf *conf = mddev->private;
4784         int dd_idx;
4785         struct bio* align_bi;
4786         struct md_rdev *rdev;
4787         sector_t end_sector;
4788
4789         if (!in_chunk_boundary(mddev, raid_bio)) {
4790                 pr_debug("%s: non aligned\n", __func__);
4791                 return 0;
4792         }
4793         /*
4794          * use bio_clone_mddev to make a copy of the bio
4795          */
4796         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4797         if (!align_bi)
4798                 return 0;
4799         /*
4800          *   set bi_end_io to a new function, and set bi_private to the
4801          *     original bio.
4802          */
4803         align_bi->bi_end_io  = raid5_align_endio;
4804         align_bi->bi_private = raid_bio;
4805         /*
4806          *      compute position
4807          */
4808         align_bi->bi_iter.bi_sector =
4809                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4810                                      0, &dd_idx, NULL);
4811
4812         end_sector = bio_end_sector(align_bi);
4813         rcu_read_lock();
4814         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4815         if (!rdev || test_bit(Faulty, &rdev->flags) ||
4816             rdev->recovery_offset < end_sector) {
4817                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4818                 if (rdev &&
4819                     (test_bit(Faulty, &rdev->flags) ||
4820                     !(test_bit(In_sync, &rdev->flags) ||
4821                       rdev->recovery_offset >= end_sector)))
4822                         rdev = NULL;
4823         }
4824         if (rdev) {
4825                 sector_t first_bad;
4826                 int bad_sectors;
4827
4828                 atomic_inc(&rdev->nr_pending);
4829                 rcu_read_unlock();
4830                 raid_bio->bi_next = (void*)rdev;
4831                 align_bi->bi_bdev =  rdev->bdev;
4832                 bio_clear_flag(align_bi, BIO_SEG_VALID);
4833
4834                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4835                                 bio_sectors(align_bi),
4836                                 &first_bad, &bad_sectors)) {
4837                         bio_put(align_bi);
4838                         rdev_dec_pending(rdev, mddev);
4839                         return 0;
4840                 }
4841
4842                 /* No reshape active, so we can trust rdev->data_offset */
4843                 align_bi->bi_iter.bi_sector += rdev->data_offset;
4844
4845                 spin_lock_irq(&conf->device_lock);
4846                 wait_event_lock_irq(conf->wait_for_quiescent,
4847                                     conf->quiesce == 0,
4848                                     conf->device_lock);
4849                 atomic_inc(&conf->active_aligned_reads);
4850                 spin_unlock_irq(&conf->device_lock);
4851
4852                 if (mddev->gendisk)
4853                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4854                                               align_bi, disk_devt(mddev->gendisk),
4855                                               raid_bio->bi_iter.bi_sector);
4856                 generic_make_request(align_bi);
4857                 return 1;
4858         } else {
4859                 rcu_read_unlock();
4860                 bio_put(align_bi);
4861                 return 0;
4862         }
4863 }
4864
4865 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4866 {
4867         struct bio *split;
4868
4869         do {
4870                 sector_t sector = raid_bio->bi_iter.bi_sector;
4871                 unsigned chunk_sects = mddev->chunk_sectors;
4872                 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4873
4874                 if (sectors < bio_sectors(raid_bio)) {
4875                         split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4876                         bio_chain(split, raid_bio);
4877                 } else
4878                         split = raid_bio;
4879
4880                 if (!raid5_read_one_chunk(mddev, split)) {
4881                         if (split != raid_bio)
4882                                 generic_make_request(raid_bio);
4883                         return split;
4884                 }
4885         } while (split != raid_bio);
4886
4887         return NULL;
4888 }
4889
4890 /* __get_priority_stripe - get the next stripe to process
4891  *
4892  * Full stripe writes are allowed to pass preread active stripes up until
4893  * the bypass_threshold is exceeded.  In general the bypass_count
4894  * increments when the handle_list is handled before the hold_list; however, it
4895  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4896  * stripe with in flight i/o.  The bypass_count will be reset when the
4897  * head of the hold_list has changed, i.e. the head was promoted to the
4898  * handle_list.
4899  */
4900 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4901 {
4902         struct stripe_head *sh = NULL, *tmp;
4903         struct list_head *handle_list = NULL;
4904         struct r5worker_group *wg = NULL;
4905
4906         if (conf->worker_cnt_per_group == 0) {
4907                 handle_list = &conf->handle_list;
4908         } else if (group != ANY_GROUP) {
4909                 handle_list = &conf->worker_groups[group].handle_list;
4910                 wg = &conf->worker_groups[group];
4911         } else {
4912                 int i;
4913                 for (i = 0; i < conf->group_cnt; i++) {
4914                         handle_list = &conf->worker_groups[i].handle_list;
4915                         wg = &conf->worker_groups[i];
4916                         if (!list_empty(handle_list))
4917                                 break;
4918                 }
4919         }
4920
4921         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4922                   __func__,
4923                   list_empty(handle_list) ? "empty" : "busy",
4924                   list_empty(&conf->hold_list) ? "empty" : "busy",
4925                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
4926
4927         if (!list_empty(handle_list)) {
4928                 sh = list_entry(handle_list->next, typeof(*sh), lru);
4929
4930                 if (list_empty(&conf->hold_list))
4931                         conf->bypass_count = 0;
4932                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4933                         if (conf->hold_list.next == conf->last_hold)
4934                                 conf->bypass_count++;
4935                         else {
4936                                 conf->last_hold = conf->hold_list.next;
4937                                 conf->bypass_count -= conf->bypass_threshold;
4938                                 if (conf->bypass_count < 0)
4939                                         conf->bypass_count = 0;
4940                         }
4941                 }
4942         } else if (!list_empty(&conf->hold_list) &&
4943                    ((conf->bypass_threshold &&
4944                      conf->bypass_count > conf->bypass_threshold) ||
4945                     atomic_read(&conf->pending_full_writes) == 0)) {
4946
4947                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
4948                         if (conf->worker_cnt_per_group == 0 ||
4949                             group == ANY_GROUP ||
4950                             !cpu_online(tmp->cpu) ||
4951                             cpu_to_group(tmp->cpu) == group) {
4952                                 sh = tmp;
4953                                 break;
4954                         }
4955                 }
4956
4957                 if (sh) {
4958                         conf->bypass_count -= conf->bypass_threshold;
4959                         if (conf->bypass_count < 0)
4960                                 conf->bypass_count = 0;
4961                 }
4962                 wg = NULL;
4963         }
4964
4965         if (!sh)
4966                 return NULL;
4967
4968         if (wg) {
4969                 wg->stripes_cnt--;
4970                 sh->group = NULL;
4971         }
4972         list_del_init(&sh->lru);
4973         BUG_ON(atomic_inc_return(&sh->count) != 1);
4974         return sh;
4975 }
4976
4977 struct raid5_plug_cb {
4978         struct blk_plug_cb      cb;
4979         struct list_head        list;
4980         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4981 };
4982
4983 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4984 {
4985         struct raid5_plug_cb *cb = container_of(
4986                 blk_cb, struct raid5_plug_cb, cb);
4987         struct stripe_head *sh;
4988         struct mddev *mddev = cb->cb.data;
4989         struct r5conf *conf = mddev->private;
4990         int cnt = 0;
4991         int hash;
4992
4993         if (cb->list.next && !list_empty(&cb->list)) {
4994                 spin_lock_irq(&conf->device_lock);
4995                 while (!list_empty(&cb->list)) {
4996                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
4997                         list_del_init(&sh->lru);
4998                         /*
4999                          * avoid race release_stripe_plug() sees
5000                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5001                          * is still in our list
5002                          */
5003                         smp_mb__before_atomic();
5004                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5005                         /*
5006                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5007                          * case, the count is always > 1 here
5008                          */
5009                         hash = sh->hash_lock_index;
5010                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5011                         cnt++;
5012                 }
5013                 spin_unlock_irq(&conf->device_lock);
5014         }
5015         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5016                                      NR_STRIPE_HASH_LOCKS);
5017         if (mddev->queue)
5018                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5019         kfree(cb);
5020 }
5021
5022 static void release_stripe_plug(struct mddev *mddev,
5023                                 struct stripe_head *sh)
5024 {
5025         struct blk_plug_cb *blk_cb = blk_check_plugged(
5026                 raid5_unplug, mddev,
5027                 sizeof(struct raid5_plug_cb));
5028         struct raid5_plug_cb *cb;
5029
5030         if (!blk_cb) {
5031                 raid5_release_stripe(sh);
5032                 return;
5033         }
5034
5035         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5036
5037         if (cb->list.next == NULL) {
5038                 int i;
5039                 INIT_LIST_HEAD(&cb->list);
5040                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5041                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5042         }
5043
5044         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5045                 list_add_tail(&sh->lru, &cb->list);
5046         else
5047                 raid5_release_stripe(sh);
5048 }
5049
5050 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5051 {
5052         struct r5conf *conf = mddev->private;
5053         sector_t logical_sector, last_sector;
5054         struct stripe_head *sh;
5055         int remaining;
5056         int stripe_sectors;
5057
5058         if (mddev->reshape_position != MaxSector)
5059                 /* Skip discard while reshape is happening */
5060                 return;
5061
5062         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5063         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5064
5065         bi->bi_next = NULL;
5066         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5067
5068         stripe_sectors = conf->chunk_sectors *
5069                 (conf->raid_disks - conf->max_degraded);
5070         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5071                                                stripe_sectors);
5072         sector_div(last_sector, stripe_sectors);
5073
5074         logical_sector *= conf->chunk_sectors;
5075         last_sector *= conf->chunk_sectors;
5076
5077         for (; logical_sector < last_sector;
5078              logical_sector += STRIPE_SECTORS) {
5079                 DEFINE_WAIT(w);
5080                 int d;
5081         again:
5082                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5083                 prepare_to_wait(&conf->wait_for_overlap, &w,
5084                                 TASK_UNINTERRUPTIBLE);
5085                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5086                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5087                         raid5_release_stripe(sh);
5088                         schedule();
5089                         goto again;
5090                 }
5091                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5092                 spin_lock_irq(&sh->stripe_lock);
5093                 for (d = 0; d < conf->raid_disks; d++) {
5094                         if (d == sh->pd_idx || d == sh->qd_idx)
5095                                 continue;
5096                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5097                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5098                                 spin_unlock_irq(&sh->stripe_lock);
5099                                 raid5_release_stripe(sh);
5100                                 schedule();
5101                                 goto again;
5102                         }
5103                 }
5104                 set_bit(STRIPE_DISCARD, &sh->state);
5105                 finish_wait(&conf->wait_for_overlap, &w);
5106                 sh->overwrite_disks = 0;
5107                 for (d = 0; d < conf->raid_disks; d++) {
5108                         if (d == sh->pd_idx || d == sh->qd_idx)
5109                                 continue;
5110                         sh->dev[d].towrite = bi;
5111                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5112                         raid5_inc_bi_active_stripes(bi);
5113                         sh->overwrite_disks++;
5114                 }
5115                 spin_unlock_irq(&sh->stripe_lock);
5116                 if (conf->mddev->bitmap) {
5117                         for (d = 0;
5118                              d < conf->raid_disks - conf->max_degraded;
5119                              d++)
5120                                 bitmap_startwrite(mddev->bitmap,
5121                                                   sh->sector,
5122                                                   STRIPE_SECTORS,
5123                                                   0);
5124                         sh->bm_seq = conf->seq_flush + 1;
5125                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5126                 }
5127
5128                 set_bit(STRIPE_HANDLE, &sh->state);
5129                 clear_bit(STRIPE_DELAYED, &sh->state);
5130                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5131                         atomic_inc(&conf->preread_active_stripes);
5132                 release_stripe_plug(mddev, sh);
5133         }
5134
5135         remaining = raid5_dec_bi_active_stripes(bi);
5136         if (remaining == 0) {
5137                 md_write_end(mddev);
5138                 bio_endio(bi);
5139         }
5140 }
5141
5142 static void make_request(struct mddev *mddev, struct bio * bi)
5143 {
5144         struct r5conf *conf = mddev->private;
5145         int dd_idx;
5146         sector_t new_sector;
5147         sector_t logical_sector, last_sector;
5148         struct stripe_head *sh;
5149         const int rw = bio_data_dir(bi);
5150         int remaining;
5151         DEFINE_WAIT(w);
5152         bool do_prepare;
5153
5154         if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5155                 int ret = r5l_handle_flush_request(conf->log, bi);
5156
5157                 if (ret == 0)
5158                         return;
5159                 if (ret == -ENODEV) {
5160                         md_flush_request(mddev, bi);
5161                         return;
5162                 }
5163                 /* ret == -EAGAIN, fallback */
5164         }
5165
5166         md_write_start(mddev, bi);
5167
5168         /*
5169          * If array is degraded, better not do chunk aligned read because
5170          * later we might have to read it again in order to reconstruct
5171          * data on failed drives.
5172          */
5173         if (rw == READ && mddev->degraded == 0 &&
5174             mddev->reshape_position == MaxSector) {
5175                 bi = chunk_aligned_read(mddev, bi);
5176                 if (!bi)
5177                         return;
5178         }
5179
5180         if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5181                 make_discard_request(mddev, bi);
5182                 return;
5183         }
5184
5185         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5186         last_sector = bio_end_sector(bi);
5187         bi->bi_next = NULL;
5188         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
5189
5190         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5191         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5192                 int previous;
5193                 int seq;
5194
5195                 do_prepare = false;
5196         retry:
5197                 seq = read_seqcount_begin(&conf->gen_lock);
5198                 previous = 0;
5199                 if (do_prepare)
5200                         prepare_to_wait(&conf->wait_for_overlap, &w,
5201                                 TASK_UNINTERRUPTIBLE);
5202                 if (unlikely(conf->reshape_progress != MaxSector)) {
5203                         /* spinlock is needed as reshape_progress may be
5204                          * 64bit on a 32bit platform, and so it might be
5205                          * possible to see a half-updated value
5206                          * Of course reshape_progress could change after
5207                          * the lock is dropped, so once we get a reference
5208                          * to the stripe that we think it is, we will have
5209                          * to check again.
5210                          */
5211                         spin_lock_irq(&conf->device_lock);
5212                         if (mddev->reshape_backwards
5213                             ? logical_sector < conf->reshape_progress
5214                             : logical_sector >= conf->reshape_progress) {
5215                                 previous = 1;
5216                         } else {
5217                                 if (mddev->reshape_backwards
5218                                     ? logical_sector < conf->reshape_safe
5219                                     : logical_sector >= conf->reshape_safe) {
5220                                         spin_unlock_irq(&conf->device_lock);
5221                                         schedule();
5222                                         do_prepare = true;
5223                                         goto retry;
5224                                 }
5225                         }
5226                         spin_unlock_irq(&conf->device_lock);
5227                 }
5228
5229                 new_sector = raid5_compute_sector(conf, logical_sector,
5230                                                   previous,
5231                                                   &dd_idx, NULL);
5232                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5233                         (unsigned long long)new_sector,
5234                         (unsigned long long)logical_sector);
5235
5236                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5237                                        (bi->bi_rw&RWA_MASK), 0);
5238                 if (sh) {
5239                         if (unlikely(previous)) {
5240                                 /* expansion might have moved on while waiting for a
5241                                  * stripe, so we must do the range check again.
5242                                  * Expansion could still move past after this
5243                                  * test, but as we are holding a reference to
5244                                  * 'sh', we know that if that happens,
5245                                  *  STRIPE_EXPANDING will get set and the expansion
5246                                  * won't proceed until we finish with the stripe.
5247                                  */
5248                                 int must_retry = 0;
5249                                 spin_lock_irq(&conf->device_lock);
5250                                 if (mddev->reshape_backwards
5251                                     ? logical_sector >= conf->reshape_progress
5252                                     : logical_sector < conf->reshape_progress)
5253                                         /* mismatch, need to try again */
5254                                         must_retry = 1;
5255                                 spin_unlock_irq(&conf->device_lock);
5256                                 if (must_retry) {
5257                                         raid5_release_stripe(sh);
5258                                         schedule();
5259                                         do_prepare = true;
5260                                         goto retry;
5261                                 }
5262                         }
5263                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5264                                 /* Might have got the wrong stripe_head
5265                                  * by accident
5266                                  */
5267                                 raid5_release_stripe(sh);
5268                                 goto retry;
5269                         }
5270
5271                         if (rw == WRITE &&
5272                             logical_sector >= mddev->suspend_lo &&
5273                             logical_sector < mddev->suspend_hi) {
5274                                 raid5_release_stripe(sh);
5275                                 /* As the suspend_* range is controlled by
5276                                  * userspace, we want an interruptible
5277                                  * wait.
5278                                  */
5279                                 flush_signals(current);
5280                                 prepare_to_wait(&conf->wait_for_overlap,
5281                                                 &w, TASK_INTERRUPTIBLE);
5282                                 if (logical_sector >= mddev->suspend_lo &&
5283                                     logical_sector < mddev->suspend_hi) {
5284                                         schedule();
5285                                         do_prepare = true;
5286                                 }
5287                                 goto retry;
5288                         }
5289
5290                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5291                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5292                                 /* Stripe is busy expanding or
5293                                  * add failed due to overlap.  Flush everything
5294                                  * and wait a while
5295                                  */
5296                                 md_wakeup_thread(mddev->thread);
5297                                 raid5_release_stripe(sh);
5298                                 schedule();
5299                                 do_prepare = true;
5300                                 goto retry;
5301                         }
5302                         set_bit(STRIPE_HANDLE, &sh->state);
5303                         clear_bit(STRIPE_DELAYED, &sh->state);
5304                         if ((!sh->batch_head || sh == sh->batch_head) &&
5305                             (bi->bi_rw & REQ_SYNC) &&
5306                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5307                                 atomic_inc(&conf->preread_active_stripes);
5308                         release_stripe_plug(mddev, sh);
5309                 } else {
5310                         /* cannot get stripe for read-ahead, just give-up */
5311                         bi->bi_error = -EIO;
5312                         break;
5313                 }
5314         }
5315         finish_wait(&conf->wait_for_overlap, &w);
5316
5317         remaining = raid5_dec_bi_active_stripes(bi);
5318         if (remaining == 0) {
5319
5320                 if ( rw == WRITE )
5321                         md_write_end(mddev);
5322
5323                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5324                                          bi, 0);
5325                 bio_endio(bi);
5326         }
5327 }
5328
5329 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5330
5331 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5332 {
5333         /* reshaping is quite different to recovery/resync so it is
5334          * handled quite separately ... here.
5335          *
5336          * On each call to sync_request, we gather one chunk worth of
5337          * destination stripes and flag them as expanding.
5338          * Then we find all the source stripes and request reads.
5339          * As the reads complete, handle_stripe will copy the data
5340          * into the destination stripe and release that stripe.
5341          */
5342         struct r5conf *conf = mddev->private;
5343         struct stripe_head *sh;
5344         sector_t first_sector, last_sector;
5345         int raid_disks = conf->previous_raid_disks;
5346         int data_disks = raid_disks - conf->max_degraded;
5347         int new_data_disks = conf->raid_disks - conf->max_degraded;
5348         int i;
5349         int dd_idx;
5350         sector_t writepos, readpos, safepos;
5351         sector_t stripe_addr;
5352         int reshape_sectors;
5353         struct list_head stripes;
5354         sector_t retn;
5355
5356         if (sector_nr == 0) {
5357                 /* If restarting in the middle, skip the initial sectors */
5358                 if (mddev->reshape_backwards &&
5359                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5360                         sector_nr = raid5_size(mddev, 0, 0)
5361                                 - conf->reshape_progress;
5362                 } else if (mddev->reshape_backwards &&
5363                            conf->reshape_progress == MaxSector) {
5364                         /* shouldn't happen, but just in case, finish up.*/
5365                         sector_nr = MaxSector;
5366                 } else if (!mddev->reshape_backwards &&
5367                            conf->reshape_progress > 0)
5368                         sector_nr = conf->reshape_progress;
5369                 sector_div(sector_nr, new_data_disks);
5370                 if (sector_nr) {
5371                         mddev->curr_resync_completed = sector_nr;
5372                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5373                         *skipped = 1;
5374                         retn = sector_nr;
5375                         goto finish;
5376                 }
5377         }
5378
5379         /* We need to process a full chunk at a time.
5380          * If old and new chunk sizes differ, we need to process the
5381          * largest of these
5382          */
5383
5384         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5385
5386         /* We update the metadata at least every 10 seconds, or when
5387          * the data about to be copied would over-write the source of
5388          * the data at the front of the range.  i.e. one new_stripe
5389          * along from reshape_progress new_maps to after where
5390          * reshape_safe old_maps to
5391          */
5392         writepos = conf->reshape_progress;
5393         sector_div(writepos, new_data_disks);
5394         readpos = conf->reshape_progress;
5395         sector_div(readpos, data_disks);
5396         safepos = conf->reshape_safe;
5397         sector_div(safepos, data_disks);
5398         if (mddev->reshape_backwards) {
5399                 BUG_ON(writepos < reshape_sectors);
5400                 writepos -= reshape_sectors;
5401                 readpos += reshape_sectors;
5402                 safepos += reshape_sectors;
5403         } else {
5404                 writepos += reshape_sectors;
5405                 /* readpos and safepos are worst-case calculations.
5406                  * A negative number is overly pessimistic, and causes
5407                  * obvious problems for unsigned storage.  So clip to 0.
5408                  */
5409                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5410                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5411         }
5412
5413         /* Having calculated the 'writepos' possibly use it
5414          * to set 'stripe_addr' which is where we will write to.
5415          */
5416         if (mddev->reshape_backwards) {
5417                 BUG_ON(conf->reshape_progress == 0);
5418                 stripe_addr = writepos;
5419                 BUG_ON((mddev->dev_sectors &
5420                         ~((sector_t)reshape_sectors - 1))
5421                        - reshape_sectors - stripe_addr
5422                        != sector_nr);
5423         } else {
5424                 BUG_ON(writepos != sector_nr + reshape_sectors);
5425                 stripe_addr = sector_nr;
5426         }
5427
5428         /* 'writepos' is the most advanced device address we might write.
5429          * 'readpos' is the least advanced device address we might read.
5430          * 'safepos' is the least address recorded in the metadata as having
5431          *     been reshaped.
5432          * If there is a min_offset_diff, these are adjusted either by
5433          * increasing the safepos/readpos if diff is negative, or
5434          * increasing writepos if diff is positive.
5435          * If 'readpos' is then behind 'writepos', there is no way that we can
5436          * ensure safety in the face of a crash - that must be done by userspace
5437          * making a backup of the data.  So in that case there is no particular
5438          * rush to update metadata.
5439          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5440          * update the metadata to advance 'safepos' to match 'readpos' so that
5441          * we can be safe in the event of a crash.
5442          * So we insist on updating metadata if safepos is behind writepos and
5443          * readpos is beyond writepos.
5444          * In any case, update the metadata every 10 seconds.
5445          * Maybe that number should be configurable, but I'm not sure it is
5446          * worth it.... maybe it could be a multiple of safemode_delay???
5447          */
5448         if (conf->min_offset_diff < 0) {
5449                 safepos += -conf->min_offset_diff;
5450                 readpos += -conf->min_offset_diff;
5451         } else
5452                 writepos += conf->min_offset_diff;
5453
5454         if ((mddev->reshape_backwards
5455              ? (safepos > writepos && readpos < writepos)
5456              : (safepos < writepos && readpos > writepos)) ||
5457             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5458                 /* Cannot proceed until we've updated the superblock... */
5459                 wait_event(conf->wait_for_overlap,
5460                            atomic_read(&conf->reshape_stripes)==0
5461                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5462                 if (atomic_read(&conf->reshape_stripes) != 0)
5463                         return 0;
5464                 mddev->reshape_position = conf->reshape_progress;
5465                 mddev->curr_resync_completed = sector_nr;
5466                 conf->reshape_checkpoint = jiffies;
5467                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5468                 md_wakeup_thread(mddev->thread);
5469                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5470                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5471                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5472                         return 0;
5473                 spin_lock_irq(&conf->device_lock);
5474                 conf->reshape_safe = mddev->reshape_position;
5475                 spin_unlock_irq(&conf->device_lock);
5476                 wake_up(&conf->wait_for_overlap);
5477                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5478         }
5479
5480         INIT_LIST_HEAD(&stripes);
5481         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5482                 int j;
5483                 int skipped_disk = 0;
5484                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5485                 set_bit(STRIPE_EXPANDING, &sh->state);
5486                 atomic_inc(&conf->reshape_stripes);
5487                 /* If any of this stripe is beyond the end of the old
5488                  * array, then we need to zero those blocks
5489                  */
5490                 for (j=sh->disks; j--;) {
5491                         sector_t s;
5492                         if (j == sh->pd_idx)
5493                                 continue;
5494                         if (conf->level == 6 &&
5495                             j == sh->qd_idx)
5496                                 continue;
5497                         s = raid5_compute_blocknr(sh, j, 0);
5498                         if (s < raid5_size(mddev, 0, 0)) {
5499                                 skipped_disk = 1;
5500                                 continue;
5501                         }
5502                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5503                         set_bit(R5_Expanded, &sh->dev[j].flags);
5504                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5505                 }
5506                 if (!skipped_disk) {
5507                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5508                         set_bit(STRIPE_HANDLE, &sh->state);
5509                 }
5510                 list_add(&sh->lru, &stripes);
5511         }
5512         spin_lock_irq(&conf->device_lock);
5513         if (mddev->reshape_backwards)
5514                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5515         else
5516                 conf->reshape_progress += reshape_sectors * new_data_disks;
5517         spin_unlock_irq(&conf->device_lock);
5518         /* Ok, those stripe are ready. We can start scheduling
5519          * reads on the source stripes.
5520          * The source stripes are determined by mapping the first and last
5521          * block on the destination stripes.
5522          */
5523         first_sector =
5524                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5525                                      1, &dd_idx, NULL);
5526         last_sector =
5527                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5528                                             * new_data_disks - 1),
5529                                      1, &dd_idx, NULL);
5530         if (last_sector >= mddev->dev_sectors)
5531                 last_sector = mddev->dev_sectors - 1;
5532         while (first_sector <= last_sector) {
5533                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5534                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5535                 set_bit(STRIPE_HANDLE, &sh->state);
5536                 raid5_release_stripe(sh);
5537                 first_sector += STRIPE_SECTORS;
5538         }
5539         /* Now that the sources are clearly marked, we can release
5540          * the destination stripes
5541          */
5542         while (!list_empty(&stripes)) {
5543                 sh = list_entry(stripes.next, struct stripe_head, lru);
5544                 list_del_init(&sh->lru);
5545                 raid5_release_stripe(sh);
5546         }
5547         /* If this takes us to the resync_max point where we have to pause,
5548          * then we need to write out the superblock.
5549          */
5550         sector_nr += reshape_sectors;
5551         retn = reshape_sectors;
5552 finish:
5553         if (mddev->curr_resync_completed > mddev->resync_max ||
5554             (sector_nr - mddev->curr_resync_completed) * 2
5555             >= mddev->resync_max - mddev->curr_resync_completed) {
5556                 /* Cannot proceed until we've updated the superblock... */
5557                 wait_event(conf->wait_for_overlap,
5558                            atomic_read(&conf->reshape_stripes) == 0
5559                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5560                 if (atomic_read(&conf->reshape_stripes) != 0)
5561                         goto ret;
5562                 mddev->reshape_position = conf->reshape_progress;
5563                 mddev->curr_resync_completed = sector_nr;
5564                 conf->reshape_checkpoint = jiffies;
5565                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5566                 md_wakeup_thread(mddev->thread);
5567                 wait_event(mddev->sb_wait,
5568                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5569                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5570                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5571                         goto ret;
5572                 spin_lock_irq(&conf->device_lock);
5573                 conf->reshape_safe = mddev->reshape_position;
5574                 spin_unlock_irq(&conf->device_lock);
5575                 wake_up(&conf->wait_for_overlap);
5576                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5577         }
5578 ret:
5579         return retn;
5580 }
5581
5582 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5583 {
5584         struct r5conf *conf = mddev->private;
5585         struct stripe_head *sh;
5586         sector_t max_sector = mddev->dev_sectors;
5587         sector_t sync_blocks;
5588         int still_degraded = 0;
5589         int i;
5590
5591         if (sector_nr >= max_sector) {
5592                 /* just being told to finish up .. nothing much to do */
5593
5594                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5595                         end_reshape(conf);
5596                         return 0;
5597                 }
5598
5599                 if (mddev->curr_resync < max_sector) /* aborted */
5600                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5601                                         &sync_blocks, 1);
5602                 else /* completed sync */
5603                         conf->fullsync = 0;
5604                 bitmap_close_sync(mddev->bitmap);
5605
5606                 return 0;
5607         }
5608
5609         /* Allow raid5_quiesce to complete */
5610         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5611
5612         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5613                 return reshape_request(mddev, sector_nr, skipped);
5614
5615         /* No need to check resync_max as we never do more than one
5616          * stripe, and as resync_max will always be on a chunk boundary,
5617          * if the check in md_do_sync didn't fire, there is no chance
5618          * of overstepping resync_max here
5619          */
5620
5621         /* if there is too many failed drives and we are trying
5622          * to resync, then assert that we are finished, because there is
5623          * nothing we can do.
5624          */
5625         if (mddev->degraded >= conf->max_degraded &&
5626             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5627                 sector_t rv = mddev->dev_sectors - sector_nr;
5628                 *skipped = 1;
5629                 return rv;
5630         }
5631         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5632             !conf->fullsync &&
5633             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5634             sync_blocks >= STRIPE_SECTORS) {
5635                 /* we can skip this block, and probably more */
5636                 sync_blocks /= STRIPE_SECTORS;
5637                 *skipped = 1;
5638                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5639         }
5640
5641         bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5642
5643         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5644         if (sh == NULL) {
5645                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5646                 /* make sure we don't swamp the stripe cache if someone else
5647                  * is trying to get access
5648                  */
5649                 schedule_timeout_uninterruptible(1);
5650         }
5651         /* Need to check if array will still be degraded after recovery/resync
5652          * Note in case of > 1 drive failures it's possible we're rebuilding
5653          * one drive while leaving another faulty drive in array.
5654          */
5655         rcu_read_lock();
5656         for (i = 0; i < conf->raid_disks; i++) {
5657                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5658
5659                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5660                         still_degraded = 1;
5661         }
5662         rcu_read_unlock();
5663
5664         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5665
5666         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5667         set_bit(STRIPE_HANDLE, &sh->state);
5668
5669         raid5_release_stripe(sh);
5670
5671         return STRIPE_SECTORS;
5672 }
5673
5674 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5675 {
5676         /* We may not be able to submit a whole bio at once as there
5677          * may not be enough stripe_heads available.
5678          * We cannot pre-allocate enough stripe_heads as we may need
5679          * more than exist in the cache (if we allow ever large chunks).
5680          * So we do one stripe head at a time and record in
5681          * ->bi_hw_segments how many have been done.
5682          *
5683          * We *know* that this entire raid_bio is in one chunk, so
5684          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5685          */
5686         struct stripe_head *sh;
5687         int dd_idx;
5688         sector_t sector, logical_sector, last_sector;
5689         int scnt = 0;
5690         int remaining;
5691         int handled = 0;
5692
5693         logical_sector = raid_bio->bi_iter.bi_sector &
5694                 ~((sector_t)STRIPE_SECTORS-1);
5695         sector = raid5_compute_sector(conf, logical_sector,
5696                                       0, &dd_idx, NULL);
5697         last_sector = bio_end_sector(raid_bio);
5698
5699         for (; logical_sector < last_sector;
5700              logical_sector += STRIPE_SECTORS,
5701                      sector += STRIPE_SECTORS,
5702                      scnt++) {
5703
5704                 if (scnt < raid5_bi_processed_stripes(raid_bio))
5705                         /* already done this stripe */
5706                         continue;
5707
5708                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5709
5710                 if (!sh) {
5711                         /* failed to get a stripe - must wait */
5712                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5713                         conf->retry_read_aligned = raid_bio;
5714                         return handled;
5715                 }
5716
5717                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5718                         raid5_release_stripe(sh);
5719                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5720                         conf->retry_read_aligned = raid_bio;
5721                         return handled;
5722                 }
5723
5724                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5725                 handle_stripe(sh);
5726                 raid5_release_stripe(sh);
5727                 handled++;
5728         }
5729         remaining = raid5_dec_bi_active_stripes(raid_bio);
5730         if (remaining == 0) {
5731                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5732                                          raid_bio, 0);
5733                 bio_endio(raid_bio);
5734         }
5735         if (atomic_dec_and_test(&conf->active_aligned_reads))
5736                 wake_up(&conf->wait_for_quiescent);
5737         return handled;
5738 }
5739
5740 static int handle_active_stripes(struct r5conf *conf, int group,
5741                                  struct r5worker *worker,
5742                                  struct list_head *temp_inactive_list)
5743 {
5744         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5745         int i, batch_size = 0, hash;
5746         bool release_inactive = false;
5747
5748         while (batch_size < MAX_STRIPE_BATCH &&
5749                         (sh = __get_priority_stripe(conf, group)) != NULL)
5750                 batch[batch_size++] = sh;
5751
5752         if (batch_size == 0) {
5753                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5754                         if (!list_empty(temp_inactive_list + i))
5755                                 break;
5756                 if (i == NR_STRIPE_HASH_LOCKS) {
5757                         spin_unlock_irq(&conf->device_lock);
5758                         r5l_flush_stripe_to_raid(conf->log);
5759                         spin_lock_irq(&conf->device_lock);
5760                         return batch_size;
5761                 }
5762                 release_inactive = true;
5763         }
5764         spin_unlock_irq(&conf->device_lock);
5765
5766         release_inactive_stripe_list(conf, temp_inactive_list,
5767                                      NR_STRIPE_HASH_LOCKS);
5768
5769         r5l_flush_stripe_to_raid(conf->log);
5770         if (release_inactive) {
5771                 spin_lock_irq(&conf->device_lock);
5772                 return 0;
5773         }
5774
5775         for (i = 0; i < batch_size; i++)
5776                 handle_stripe(batch[i]);
5777         r5l_write_stripe_run(conf->log);
5778
5779         cond_resched();
5780
5781         spin_lock_irq(&conf->device_lock);
5782         for (i = 0; i < batch_size; i++) {
5783                 hash = batch[i]->hash_lock_index;
5784                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5785         }
5786         return batch_size;
5787 }
5788
5789 static void raid5_do_work(struct work_struct *work)
5790 {
5791         struct r5worker *worker = container_of(work, struct r5worker, work);
5792         struct r5worker_group *group = worker->group;
5793         struct r5conf *conf = group->conf;
5794         int group_id = group - conf->worker_groups;
5795         int handled;
5796         struct blk_plug plug;
5797
5798         pr_debug("+++ raid5worker active\n");
5799
5800         blk_start_plug(&plug);
5801         handled = 0;
5802         spin_lock_irq(&conf->device_lock);
5803         while (1) {
5804                 int batch_size, released;
5805
5806                 released = release_stripe_list(conf, worker->temp_inactive_list);
5807
5808                 batch_size = handle_active_stripes(conf, group_id, worker,
5809                                                    worker->temp_inactive_list);
5810                 worker->working = false;
5811                 if (!batch_size && !released)
5812                         break;
5813                 handled += batch_size;
5814         }
5815         pr_debug("%d stripes handled\n", handled);
5816
5817         spin_unlock_irq(&conf->device_lock);
5818         blk_finish_plug(&plug);
5819
5820         pr_debug("--- raid5worker inactive\n");
5821 }
5822
5823 /*
5824  * This is our raid5 kernel thread.
5825  *
5826  * We scan the hash table for stripes which can be handled now.
5827  * During the scan, completed stripes are saved for us by the interrupt
5828  * handler, so that they will not have to wait for our next wakeup.
5829  */
5830 static void raid5d(struct md_thread *thread)
5831 {
5832         struct mddev *mddev = thread->mddev;
5833         struct r5conf *conf = mddev->private;
5834         int handled;
5835         struct blk_plug plug;
5836
5837         pr_debug("+++ raid5d active\n");
5838
5839         md_check_recovery(mddev);
5840
5841         if (!bio_list_empty(&conf->return_bi) &&
5842             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5843                 struct bio_list tmp = BIO_EMPTY_LIST;
5844                 spin_lock_irq(&conf->device_lock);
5845                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5846                         bio_list_merge(&tmp, &conf->return_bi);
5847                         bio_list_init(&conf->return_bi);
5848                 }
5849                 spin_unlock_irq(&conf->device_lock);
5850                 return_io(&tmp);
5851         }
5852
5853         blk_start_plug(&plug);
5854         handled = 0;
5855         spin_lock_irq(&conf->device_lock);
5856         while (1) {
5857                 struct bio *bio;
5858                 int batch_size, released;
5859
5860                 released = release_stripe_list(conf, conf->temp_inactive_list);
5861                 if (released)
5862                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
5863
5864                 if (
5865                     !list_empty(&conf->bitmap_list)) {
5866                         /* Now is a good time to flush some bitmap updates */
5867                         conf->seq_flush++;
5868                         spin_unlock_irq(&conf->device_lock);
5869                         bitmap_unplug(mddev->bitmap);
5870                         spin_lock_irq(&conf->device_lock);
5871                         conf->seq_write = conf->seq_flush;
5872                         activate_bit_delay(conf, conf->temp_inactive_list);
5873                 }
5874                 raid5_activate_delayed(conf);
5875
5876                 while ((bio = remove_bio_from_retry(conf))) {
5877                         int ok;
5878                         spin_unlock_irq(&conf->device_lock);
5879                         ok = retry_aligned_read(conf, bio);
5880                         spin_lock_irq(&conf->device_lock);
5881                         if (!ok)
5882                                 break;
5883                         handled++;
5884                 }
5885
5886                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5887                                                    conf->temp_inactive_list);
5888                 if (!batch_size && !released)
5889                         break;
5890                 handled += batch_size;
5891
5892                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5893                         spin_unlock_irq(&conf->device_lock);
5894                         md_check_recovery(mddev);
5895                         spin_lock_irq(&conf->device_lock);
5896                 }
5897         }
5898         pr_debug("%d stripes handled\n", handled);
5899
5900         spin_unlock_irq(&conf->device_lock);
5901         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5902             mutex_trylock(&conf->cache_size_mutex)) {
5903                 grow_one_stripe(conf, __GFP_NOWARN);
5904                 /* Set flag even if allocation failed.  This helps
5905                  * slow down allocation requests when mem is short
5906                  */
5907                 set_bit(R5_DID_ALLOC, &conf->cache_state);
5908                 mutex_unlock(&conf->cache_size_mutex);
5909         }
5910
5911         r5l_flush_stripe_to_raid(conf->log);
5912
5913         async_tx_issue_pending_all();
5914         blk_finish_plug(&plug);
5915
5916         pr_debug("--- raid5d inactive\n");
5917 }
5918
5919 static ssize_t
5920 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5921 {
5922         struct r5conf *conf;
5923         int ret = 0;
5924         spin_lock(&mddev->lock);
5925         conf = mddev->private;
5926         if (conf)
5927                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5928         spin_unlock(&mddev->lock);
5929         return ret;
5930 }
5931
5932 int
5933 raid5_set_cache_size(struct mddev *mddev, int size)
5934 {
5935         struct r5conf *conf = mddev->private;
5936         int err;
5937
5938         if (size <= 16 || size > 32768)
5939                 return -EINVAL;
5940
5941         conf->min_nr_stripes = size;
5942         mutex_lock(&conf->cache_size_mutex);
5943         while (size < conf->max_nr_stripes &&
5944                drop_one_stripe(conf))
5945                 ;
5946         mutex_unlock(&conf->cache_size_mutex);
5947
5948
5949         err = md_allow_write(mddev);
5950         if (err)
5951                 return err;
5952
5953         mutex_lock(&conf->cache_size_mutex);
5954         while (size > conf->max_nr_stripes)
5955                 if (!grow_one_stripe(conf, GFP_KERNEL))
5956                         break;
5957         mutex_unlock(&conf->cache_size_mutex);
5958
5959         return 0;
5960 }
5961 EXPORT_SYMBOL(raid5_set_cache_size);
5962
5963 static ssize_t
5964 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5965 {
5966         struct r5conf *conf;
5967         unsigned long new;
5968         int err;
5969
5970         if (len >= PAGE_SIZE)
5971                 return -EINVAL;
5972         if (kstrtoul(page, 10, &new))
5973                 return -EINVAL;
5974         err = mddev_lock(mddev);
5975         if (err)
5976                 return err;
5977         conf = mddev->private;
5978         if (!conf)
5979                 err = -ENODEV;
5980         else
5981                 err = raid5_set_cache_size(mddev, new);
5982         mddev_unlock(mddev);
5983
5984         return err ?: len;
5985 }
5986
5987 static struct md_sysfs_entry
5988 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5989                                 raid5_show_stripe_cache_size,
5990                                 raid5_store_stripe_cache_size);
5991
5992 static ssize_t
5993 raid5_show_rmw_level(struct mddev  *mddev, char *page)
5994 {
5995         struct r5conf *conf = mddev->private;
5996         if (conf)
5997                 return sprintf(page, "%d\n", conf->rmw_level);
5998         else
5999                 return 0;
6000 }
6001
6002 static ssize_t
6003 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6004 {
6005         struct r5conf *conf = mddev->private;
6006         unsigned long new;
6007
6008         if (!conf)
6009                 return -ENODEV;
6010
6011         if (len >= PAGE_SIZE)
6012                 return -EINVAL;
6013
6014         if (kstrtoul(page, 10, &new))
6015                 return -EINVAL;
6016
6017         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6018                 return -EINVAL;
6019
6020         if (new != PARITY_DISABLE_RMW &&
6021             new != PARITY_ENABLE_RMW &&
6022             new != PARITY_PREFER_RMW)
6023                 return -EINVAL;
6024
6025         conf->rmw_level = new;
6026         return len;
6027 }
6028
6029 static struct md_sysfs_entry
6030 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6031                          raid5_show_rmw_level,
6032                          raid5_store_rmw_level);
6033
6034
6035 static ssize_t
6036 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6037 {
6038         struct r5conf *conf;
6039         int ret = 0;
6040         spin_lock(&mddev->lock);
6041         conf = mddev->private;
6042         if (conf)
6043                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6044         spin_unlock(&mddev->lock);
6045         return ret;
6046 }
6047
6048 static ssize_t
6049 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6050 {
6051         struct r5conf *conf;
6052         unsigned long new;
6053         int err;
6054
6055         if (len >= PAGE_SIZE)
6056                 return -EINVAL;
6057         if (kstrtoul(page, 10, &new))
6058                 return -EINVAL;
6059
6060         err = mddev_lock(mddev);
6061         if (err)
6062                 return err;
6063         conf = mddev->private;
6064         if (!conf)
6065                 err = -ENODEV;
6066         else if (new > conf->min_nr_stripes)
6067                 err = -EINVAL;
6068         else
6069                 conf->bypass_threshold = new;
6070         mddev_unlock(mddev);
6071         return err ?: len;
6072 }
6073
6074 static struct md_sysfs_entry
6075 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6076                                         S_IRUGO | S_IWUSR,
6077                                         raid5_show_preread_threshold,
6078                                         raid5_store_preread_threshold);
6079
6080 static ssize_t
6081 raid5_show_skip_copy(struct mddev *mddev, char *page)
6082 {
6083         struct r5conf *conf;
6084         int ret = 0;
6085         spin_lock(&mddev->lock);
6086         conf = mddev->private;
6087         if (conf)
6088                 ret = sprintf(page, "%d\n", conf->skip_copy);
6089         spin_unlock(&mddev->lock);
6090         return ret;
6091 }
6092
6093 static ssize_t
6094 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6095 {
6096         struct r5conf *conf;
6097         unsigned long new;
6098         int err;
6099
6100         if (len >= PAGE_SIZE)
6101                 return -EINVAL;
6102         if (kstrtoul(page, 10, &new))
6103                 return -EINVAL;
6104         new = !!new;
6105
6106         err = mddev_lock(mddev);
6107         if (err)
6108                 return err;
6109         conf = mddev->private;
6110         if (!conf)
6111                 err = -ENODEV;
6112         else if (new != conf->skip_copy) {
6113                 mddev_suspend(mddev);
6114                 conf->skip_copy = new;
6115                 if (new)
6116                         mddev->queue->backing_dev_info.capabilities |=
6117                                 BDI_CAP_STABLE_WRITES;
6118                 else
6119                         mddev->queue->backing_dev_info.capabilities &=
6120                                 ~BDI_CAP_STABLE_WRITES;
6121                 mddev_resume(mddev);
6122         }
6123         mddev_unlock(mddev);
6124         return err ?: len;
6125 }
6126
6127 static struct md_sysfs_entry
6128 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6129                                         raid5_show_skip_copy,
6130                                         raid5_store_skip_copy);
6131
6132 static ssize_t
6133 stripe_cache_active_show(struct mddev *mddev, char *page)
6134 {
6135         struct r5conf *conf = mddev->private;
6136         if (conf)
6137                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6138         else
6139                 return 0;
6140 }
6141
6142 static struct md_sysfs_entry
6143 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6144
6145 static ssize_t
6146 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6147 {
6148         struct r5conf *conf;
6149         int ret = 0;
6150         spin_lock(&mddev->lock);
6151         conf = mddev->private;
6152         if (conf)
6153                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6154         spin_unlock(&mddev->lock);
6155         return ret;
6156 }
6157
6158 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6159                                int *group_cnt,
6160                                int *worker_cnt_per_group,
6161                                struct r5worker_group **worker_groups);
6162 static ssize_t
6163 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6164 {
6165         struct r5conf *conf;
6166         unsigned long new;
6167         int err;
6168         struct r5worker_group *new_groups, *old_groups;
6169         int group_cnt, worker_cnt_per_group;
6170
6171         if (len >= PAGE_SIZE)
6172                 return -EINVAL;
6173         if (kstrtoul(page, 10, &new))
6174                 return -EINVAL;
6175
6176         err = mddev_lock(mddev);
6177         if (err)
6178                 return err;
6179         conf = mddev->private;
6180         if (!conf)
6181                 err = -ENODEV;
6182         else if (new != conf->worker_cnt_per_group) {
6183                 mddev_suspend(mddev);
6184
6185                 old_groups = conf->worker_groups;
6186                 if (old_groups)
6187                         flush_workqueue(raid5_wq);
6188
6189                 err = alloc_thread_groups(conf, new,
6190                                           &group_cnt, &worker_cnt_per_group,
6191                                           &new_groups);
6192                 if (!err) {
6193                         spin_lock_irq(&conf->device_lock);
6194                         conf->group_cnt = group_cnt;
6195                         conf->worker_cnt_per_group = worker_cnt_per_group;
6196                         conf->worker_groups = new_groups;
6197                         spin_unlock_irq(&conf->device_lock);
6198
6199                         if (old_groups)
6200                                 kfree(old_groups[0].workers);
6201                         kfree(old_groups);
6202                 }
6203                 mddev_resume(mddev);
6204         }
6205         mddev_unlock(mddev);
6206
6207         return err ?: len;
6208 }
6209
6210 static struct md_sysfs_entry
6211 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6212                                 raid5_show_group_thread_cnt,
6213                                 raid5_store_group_thread_cnt);
6214
6215 static struct attribute *raid5_attrs[] =  {
6216         &raid5_stripecache_size.attr,
6217         &raid5_stripecache_active.attr,
6218         &raid5_preread_bypass_threshold.attr,
6219         &raid5_group_thread_cnt.attr,
6220         &raid5_skip_copy.attr,
6221         &raid5_rmw_level.attr,
6222         NULL,
6223 };
6224 static struct attribute_group raid5_attrs_group = {
6225         .name = NULL,
6226         .attrs = raid5_attrs,
6227 };
6228
6229 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6230                                int *group_cnt,
6231                                int *worker_cnt_per_group,
6232                                struct r5worker_group **worker_groups)
6233 {
6234         int i, j, k;
6235         ssize_t size;
6236         struct r5worker *workers;
6237
6238         *worker_cnt_per_group = cnt;
6239         if (cnt == 0) {
6240                 *group_cnt = 0;
6241                 *worker_groups = NULL;
6242                 return 0;
6243         }
6244         *group_cnt = num_possible_nodes();
6245         size = sizeof(struct r5worker) * cnt;
6246         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6247         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6248                                 *group_cnt, GFP_NOIO);
6249         if (!*worker_groups || !workers) {
6250                 kfree(workers);
6251                 kfree(*worker_groups);
6252                 return -ENOMEM;
6253         }
6254
6255         for (i = 0; i < *group_cnt; i++) {
6256                 struct r5worker_group *group;
6257
6258                 group = &(*worker_groups)[i];
6259                 INIT_LIST_HEAD(&group->handle_list);
6260                 group->conf = conf;
6261                 group->workers = workers + i * cnt;
6262
6263                 for (j = 0; j < cnt; j++) {
6264                         struct r5worker *worker = group->workers + j;
6265                         worker->group = group;
6266                         INIT_WORK(&worker->work, raid5_do_work);
6267
6268                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6269                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6270                 }
6271         }
6272
6273         return 0;
6274 }
6275
6276 static void free_thread_groups(struct r5conf *conf)
6277 {
6278         if (conf->worker_groups)
6279                 kfree(conf->worker_groups[0].workers);
6280         kfree(conf->worker_groups);
6281         conf->worker_groups = NULL;
6282 }
6283
6284 static sector_t
6285 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6286 {
6287         struct r5conf *conf = mddev->private;
6288
6289         if (!sectors)
6290                 sectors = mddev->dev_sectors;
6291         if (!raid_disks)
6292                 /* size is defined by the smallest of previous and new size */
6293                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6294
6295         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6296         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6297         return sectors * (raid_disks - conf->max_degraded);
6298 }
6299
6300 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6301 {
6302         safe_put_page(percpu->spare_page);
6303         if (percpu->scribble)
6304                 flex_array_free(percpu->scribble);
6305         percpu->spare_page = NULL;
6306         percpu->scribble = NULL;
6307 }
6308
6309 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6310 {
6311         if (conf->level == 6 && !percpu->spare_page)
6312                 percpu->spare_page = alloc_page(GFP_KERNEL);
6313         if (!percpu->scribble)
6314                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6315                                                       conf->previous_raid_disks),
6316                                                   max(conf->chunk_sectors,
6317                                                       conf->prev_chunk_sectors)
6318                                                    / STRIPE_SECTORS,
6319                                                   GFP_KERNEL);
6320
6321         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6322                 free_scratch_buffer(conf, percpu);
6323                 return -ENOMEM;
6324         }
6325
6326         return 0;
6327 }
6328
6329 static void raid5_free_percpu(struct r5conf *conf)
6330 {
6331         unsigned long cpu;
6332
6333         if (!conf->percpu)
6334                 return;
6335
6336 #ifdef CONFIG_HOTPLUG_CPU
6337         unregister_cpu_notifier(&conf->cpu_notify);
6338 #endif
6339
6340         get_online_cpus();
6341         for_each_possible_cpu(cpu)
6342                 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6343         put_online_cpus();
6344
6345         free_percpu(conf->percpu);
6346 }
6347
6348 static void free_conf(struct r5conf *conf)
6349 {
6350         if (conf->log)
6351                 r5l_exit_log(conf->log);
6352         if (conf->shrinker.seeks)
6353                 unregister_shrinker(&conf->shrinker);
6354
6355         free_thread_groups(conf);
6356         shrink_stripes(conf);
6357         raid5_free_percpu(conf);
6358         kfree(conf->disks);
6359         kfree(conf->stripe_hashtbl);
6360         kfree(conf);
6361 }
6362
6363 #ifdef CONFIG_HOTPLUG_CPU
6364 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6365                               void *hcpu)
6366 {
6367         struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6368         long cpu = (long)hcpu;
6369         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6370
6371         switch (action) {
6372         case CPU_UP_PREPARE:
6373         case CPU_UP_PREPARE_FROZEN:
6374                 if (alloc_scratch_buffer(conf, percpu)) {
6375                         pr_err("%s: failed memory allocation for cpu%ld\n",
6376                                __func__, cpu);
6377                         return notifier_from_errno(-ENOMEM);
6378                 }
6379                 break;
6380         case CPU_DEAD:
6381         case CPU_DEAD_FROZEN:
6382                 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6383                 break;
6384         default:
6385                 break;
6386         }
6387         return NOTIFY_OK;
6388 }
6389 #endif
6390
6391 static int raid5_alloc_percpu(struct r5conf *conf)
6392 {
6393         unsigned long cpu;
6394         int err = 0;
6395
6396         conf->percpu = alloc_percpu(struct raid5_percpu);
6397         if (!conf->percpu)
6398                 return -ENOMEM;
6399
6400 #ifdef CONFIG_HOTPLUG_CPU
6401         conf->cpu_notify.notifier_call = raid456_cpu_notify;
6402         conf->cpu_notify.priority = 0;
6403         err = register_cpu_notifier(&conf->cpu_notify);
6404         if (err)
6405                 return err;
6406 #endif
6407
6408         get_online_cpus();
6409         for_each_present_cpu(cpu) {
6410                 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6411                 if (err) {
6412                         pr_err("%s: failed memory allocation for cpu%ld\n",
6413                                __func__, cpu);
6414                         break;
6415                 }
6416                 spin_lock_init(&per_cpu_ptr(conf->percpu, cpu)->lock);
6417         }
6418         put_online_cpus();
6419
6420         return err;
6421 }
6422
6423 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6424                                       struct shrink_control *sc)
6425 {
6426         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6427         unsigned long ret = SHRINK_STOP;
6428
6429         if (mutex_trylock(&conf->cache_size_mutex)) {
6430                 ret= 0;
6431                 while (ret < sc->nr_to_scan &&
6432                        conf->max_nr_stripes > conf->min_nr_stripes) {
6433                         if (drop_one_stripe(conf) == 0) {
6434                                 ret = SHRINK_STOP;
6435                                 break;
6436                         }
6437                         ret++;
6438                 }
6439                 mutex_unlock(&conf->cache_size_mutex);
6440         }
6441         return ret;
6442 }
6443
6444 static unsigned long raid5_cache_count(struct shrinker *shrink,
6445                                        struct shrink_control *sc)
6446 {
6447         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6448
6449         if (conf->max_nr_stripes < conf->min_nr_stripes)
6450                 /* unlikely, but not impossible */
6451                 return 0;
6452         return conf->max_nr_stripes - conf->min_nr_stripes;
6453 }
6454
6455 static struct r5conf *setup_conf(struct mddev *mddev)
6456 {
6457         struct r5conf *conf;
6458         int raid_disk, memory, max_disks;
6459         struct md_rdev *rdev;
6460         struct disk_info *disk;
6461         char pers_name[6];
6462         int i;
6463         int group_cnt, worker_cnt_per_group;
6464         struct r5worker_group *new_group;
6465
6466         if (mddev->new_level != 5
6467             && mddev->new_level != 4
6468             && mddev->new_level != 6) {
6469                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6470                        mdname(mddev), mddev->new_level);
6471                 return ERR_PTR(-EIO);
6472         }
6473         if ((mddev->new_level == 5
6474              && !algorithm_valid_raid5(mddev->new_layout)) ||
6475             (mddev->new_level == 6
6476              && !algorithm_valid_raid6(mddev->new_layout))) {
6477                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6478                        mdname(mddev), mddev->new_layout);
6479                 return ERR_PTR(-EIO);
6480         }
6481         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6482                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6483                        mdname(mddev), mddev->raid_disks);
6484                 return ERR_PTR(-EINVAL);
6485         }
6486
6487         if (!mddev->new_chunk_sectors ||
6488             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6489             !is_power_of_2(mddev->new_chunk_sectors)) {
6490                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6491                        mdname(mddev), mddev->new_chunk_sectors << 9);
6492                 return ERR_PTR(-EINVAL);
6493         }
6494
6495         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6496         if (conf == NULL)
6497                 goto abort;
6498         /* Don't enable multi-threading by default*/
6499         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6500                                  &new_group)) {
6501                 conf->group_cnt = group_cnt;
6502                 conf->worker_cnt_per_group = worker_cnt_per_group;
6503                 conf->worker_groups = new_group;
6504         } else
6505                 goto abort;
6506         spin_lock_init(&conf->device_lock);
6507         seqcount_init(&conf->gen_lock);
6508         mutex_init(&conf->cache_size_mutex);
6509         init_waitqueue_head(&conf->wait_for_quiescent);
6510         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
6511                 init_waitqueue_head(&conf->wait_for_stripe[i]);
6512         }
6513         init_waitqueue_head(&conf->wait_for_overlap);
6514         INIT_LIST_HEAD(&conf->handle_list);
6515         INIT_LIST_HEAD(&conf->hold_list);
6516         INIT_LIST_HEAD(&conf->delayed_list);
6517         INIT_LIST_HEAD(&conf->bitmap_list);
6518         bio_list_init(&conf->return_bi);
6519         init_llist_head(&conf->released_stripes);
6520         atomic_set(&conf->active_stripes, 0);
6521         atomic_set(&conf->preread_active_stripes, 0);
6522         atomic_set(&conf->active_aligned_reads, 0);
6523         conf->bypass_threshold = BYPASS_THRESHOLD;
6524         conf->recovery_disabled = mddev->recovery_disabled - 1;
6525
6526         conf->raid_disks = mddev->raid_disks;
6527         if (mddev->reshape_position == MaxSector)
6528                 conf->previous_raid_disks = mddev->raid_disks;
6529         else
6530                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6531         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6532
6533         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6534                               GFP_KERNEL);
6535         if (!conf->disks)
6536                 goto abort;
6537
6538         conf->mddev = mddev;
6539
6540         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6541                 goto abort;
6542
6543         /* We init hash_locks[0] separately to that it can be used
6544          * as the reference lock in the spin_lock_nest_lock() call
6545          * in lock_all_device_hash_locks_irq in order to convince
6546          * lockdep that we know what we are doing.
6547          */
6548         spin_lock_init(conf->hash_locks);
6549         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6550                 spin_lock_init(conf->hash_locks + i);
6551
6552         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6553                 INIT_LIST_HEAD(conf->inactive_list + i);
6554
6555         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6556                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6557
6558         conf->level = mddev->new_level;
6559         conf->chunk_sectors = mddev->new_chunk_sectors;
6560         if (raid5_alloc_percpu(conf) != 0)
6561                 goto abort;
6562
6563         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6564
6565         rdev_for_each(rdev, mddev) {
6566                 raid_disk = rdev->raid_disk;
6567                 if (raid_disk >= max_disks
6568                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6569                         continue;
6570                 disk = conf->disks + raid_disk;
6571
6572                 if (test_bit(Replacement, &rdev->flags)) {
6573                         if (disk->replacement)
6574                                 goto abort;
6575                         disk->replacement = rdev;
6576                 } else {
6577                         if (disk->rdev)
6578                                 goto abort;
6579                         disk->rdev = rdev;
6580                 }
6581
6582                 if (test_bit(In_sync, &rdev->flags)) {
6583                         char b[BDEVNAME_SIZE];
6584                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6585                                " disk %d\n",
6586                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6587                 } else if (rdev->saved_raid_disk != raid_disk)
6588                         /* Cannot rely on bitmap to complete recovery */
6589                         conf->fullsync = 1;
6590         }
6591
6592         conf->level = mddev->new_level;
6593         if (conf->level == 6) {
6594                 conf->max_degraded = 2;
6595                 if (raid6_call.xor_syndrome)
6596                         conf->rmw_level = PARITY_ENABLE_RMW;
6597                 else
6598                         conf->rmw_level = PARITY_DISABLE_RMW;
6599         } else {
6600                 conf->max_degraded = 1;
6601                 conf->rmw_level = PARITY_ENABLE_RMW;
6602         }
6603         conf->algorithm = mddev->new_layout;
6604         conf->reshape_progress = mddev->reshape_position;
6605         if (conf->reshape_progress != MaxSector) {
6606                 conf->prev_chunk_sectors = mddev->chunk_sectors;
6607                 conf->prev_algo = mddev->layout;
6608         } else {
6609                 conf->prev_chunk_sectors = conf->chunk_sectors;
6610                 conf->prev_algo = conf->algorithm;
6611         }
6612
6613         conf->min_nr_stripes = NR_STRIPES;
6614         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6615                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6616         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6617         if (grow_stripes(conf, conf->min_nr_stripes)) {
6618                 printk(KERN_ERR
6619                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
6620                        mdname(mddev), memory);
6621                 goto abort;
6622         } else
6623                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6624                        mdname(mddev), memory);
6625         /*
6626          * Losing a stripe head costs more than the time to refill it,
6627          * it reduces the queue depth and so can hurt throughput.
6628          * So set it rather large, scaled by number of devices.
6629          */
6630         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6631         conf->shrinker.scan_objects = raid5_cache_scan;
6632         conf->shrinker.count_objects = raid5_cache_count;
6633         conf->shrinker.batch = 128;
6634         conf->shrinker.flags = 0;
6635         register_shrinker(&conf->shrinker);
6636
6637         sprintf(pers_name, "raid%d", mddev->new_level);
6638         conf->thread = md_register_thread(raid5d, mddev, pers_name);
6639         if (!conf->thread) {
6640                 printk(KERN_ERR
6641                        "md/raid:%s: couldn't allocate thread.\n",
6642                        mdname(mddev));
6643                 goto abort;
6644         }
6645
6646         return conf;
6647
6648  abort:
6649         if (conf) {
6650                 free_conf(conf);
6651                 return ERR_PTR(-EIO);
6652         } else
6653                 return ERR_PTR(-ENOMEM);
6654 }
6655
6656 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6657 {
6658         switch (algo) {
6659         case ALGORITHM_PARITY_0:
6660                 if (raid_disk < max_degraded)
6661                         return 1;
6662                 break;
6663         case ALGORITHM_PARITY_N:
6664                 if (raid_disk >= raid_disks - max_degraded)
6665                         return 1;
6666                 break;
6667         case ALGORITHM_PARITY_0_6:
6668                 if (raid_disk == 0 ||
6669                     raid_disk == raid_disks - 1)
6670                         return 1;
6671                 break;
6672         case ALGORITHM_LEFT_ASYMMETRIC_6:
6673         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6674         case ALGORITHM_LEFT_SYMMETRIC_6:
6675         case ALGORITHM_RIGHT_SYMMETRIC_6:
6676                 if (raid_disk == raid_disks - 1)
6677                         return 1;
6678         }
6679         return 0;
6680 }
6681
6682 static int run(struct mddev *mddev)
6683 {
6684         struct r5conf *conf;
6685         int working_disks = 0;
6686         int dirty_parity_disks = 0;
6687         struct md_rdev *rdev;
6688         struct md_rdev *journal_dev = NULL;
6689         sector_t reshape_offset = 0;
6690         int i;
6691         long long min_offset_diff = 0;
6692         int first = 1;
6693
6694         if (mddev->recovery_cp != MaxSector)
6695                 printk(KERN_NOTICE "md/raid:%s: not clean"
6696                        " -- starting background reconstruction\n",
6697                        mdname(mddev));
6698
6699         rdev_for_each(rdev, mddev) {
6700                 long long diff;
6701
6702                 if (test_bit(Journal, &rdev->flags)) {
6703                         journal_dev = rdev;
6704                         continue;
6705                 }
6706                 if (rdev->raid_disk < 0)
6707                         continue;
6708                 diff = (rdev->new_data_offset - rdev->data_offset);
6709                 if (first) {
6710                         min_offset_diff = diff;
6711                         first = 0;
6712                 } else if (mddev->reshape_backwards &&
6713                          diff < min_offset_diff)
6714                         min_offset_diff = diff;
6715                 else if (!mddev->reshape_backwards &&
6716                          diff > min_offset_diff)
6717                         min_offset_diff = diff;
6718         }
6719
6720         if (mddev->reshape_position != MaxSector) {
6721                 /* Check that we can continue the reshape.
6722                  * Difficulties arise if the stripe we would write to
6723                  * next is at or after the stripe we would read from next.
6724                  * For a reshape that changes the number of devices, this
6725                  * is only possible for a very short time, and mdadm makes
6726                  * sure that time appears to have past before assembling
6727                  * the array.  So we fail if that time hasn't passed.
6728                  * For a reshape that keeps the number of devices the same
6729                  * mdadm must be monitoring the reshape can keeping the
6730                  * critical areas read-only and backed up.  It will start
6731                  * the array in read-only mode, so we check for that.
6732                  */
6733                 sector_t here_new, here_old;
6734                 int old_disks;
6735                 int max_degraded = (mddev->level == 6 ? 2 : 1);
6736                 int chunk_sectors;
6737                 int new_data_disks;
6738
6739                 if (journal_dev) {
6740                         printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6741                                mdname(mddev));
6742                         return -EINVAL;
6743                 }
6744
6745                 if (mddev->new_level != mddev->level) {
6746                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
6747                                "required - aborting.\n",
6748                                mdname(mddev));
6749                         return -EINVAL;
6750                 }
6751                 old_disks = mddev->raid_disks - mddev->delta_disks;
6752                 /* reshape_position must be on a new-stripe boundary, and one
6753                  * further up in new geometry must map after here in old
6754                  * geometry.
6755                  * If the chunk sizes are different, then as we perform reshape
6756                  * in units of the largest of the two, reshape_position needs
6757                  * be a multiple of the largest chunk size times new data disks.
6758                  */
6759                 here_new = mddev->reshape_position;
6760                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6761                 new_data_disks = mddev->raid_disks - max_degraded;
6762                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6763                         printk(KERN_ERR "md/raid:%s: reshape_position not "
6764                                "on a stripe boundary\n", mdname(mddev));
6765                         return -EINVAL;
6766                 }
6767                 reshape_offset = here_new * chunk_sectors;
6768                 /* here_new is the stripe we will write to */
6769                 here_old = mddev->reshape_position;
6770                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6771                 /* here_old is the first stripe that we might need to read
6772                  * from */
6773                 if (mddev->delta_disks == 0) {
6774                         /* We cannot be sure it is safe to start an in-place
6775                          * reshape.  It is only safe if user-space is monitoring
6776                          * and taking constant backups.
6777                          * mdadm always starts a situation like this in
6778                          * readonly mode so it can take control before
6779                          * allowing any writes.  So just check for that.
6780                          */
6781                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6782                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
6783                                 /* not really in-place - so OK */;
6784                         else if (mddev->ro == 0) {
6785                                 printk(KERN_ERR "md/raid:%s: in-place reshape "
6786                                        "must be started in read-only mode "
6787                                        "- aborting\n",
6788                                        mdname(mddev));
6789                                 return -EINVAL;
6790                         }
6791                 } else if (mddev->reshape_backwards
6792                     ? (here_new * chunk_sectors + min_offset_diff <=
6793                        here_old * chunk_sectors)
6794                     : (here_new * chunk_sectors >=
6795                        here_old * chunk_sectors + (-min_offset_diff))) {
6796                         /* Reading from the same stripe as writing to - bad */
6797                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6798                                "auto-recovery - aborting.\n",
6799                                mdname(mddev));
6800                         return -EINVAL;
6801                 }
6802                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6803                        mdname(mddev));
6804                 /* OK, we should be able to continue; */
6805         } else {
6806                 BUG_ON(mddev->level != mddev->new_level);
6807                 BUG_ON(mddev->layout != mddev->new_layout);
6808                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6809                 BUG_ON(mddev->delta_disks != 0);
6810         }
6811
6812         if (mddev->private == NULL)
6813                 conf = setup_conf(mddev);
6814         else
6815                 conf = mddev->private;
6816
6817         if (IS_ERR(conf))
6818                 return PTR_ERR(conf);
6819
6820         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) && !journal_dev) {
6821                 printk(KERN_ERR "md/raid:%s: journal disk is missing, force array readonly\n",
6822                        mdname(mddev));
6823                 mddev->ro = 1;
6824                 set_disk_ro(mddev->gendisk, 1);
6825         }
6826
6827         conf->min_offset_diff = min_offset_diff;
6828         mddev->thread = conf->thread;
6829         conf->thread = NULL;
6830         mddev->private = conf;
6831
6832         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6833              i++) {
6834                 rdev = conf->disks[i].rdev;
6835                 if (!rdev && conf->disks[i].replacement) {
6836                         /* The replacement is all we have yet */
6837                         rdev = conf->disks[i].replacement;
6838                         conf->disks[i].replacement = NULL;
6839                         clear_bit(Replacement, &rdev->flags);
6840                         conf->disks[i].rdev = rdev;
6841                 }
6842                 if (!rdev)
6843                         continue;
6844                 if (conf->disks[i].replacement &&
6845                     conf->reshape_progress != MaxSector) {
6846                         /* replacements and reshape simply do not mix. */
6847                         printk(KERN_ERR "md: cannot handle concurrent "
6848                                "replacement and reshape.\n");
6849                         goto abort;
6850                 }
6851                 if (test_bit(In_sync, &rdev->flags)) {
6852                         working_disks++;
6853                         continue;
6854                 }
6855                 /* This disc is not fully in-sync.  However if it
6856                  * just stored parity (beyond the recovery_offset),
6857                  * when we don't need to be concerned about the
6858                  * array being dirty.
6859                  * When reshape goes 'backwards', we never have
6860                  * partially completed devices, so we only need
6861                  * to worry about reshape going forwards.
6862                  */
6863                 /* Hack because v0.91 doesn't store recovery_offset properly. */
6864                 if (mddev->major_version == 0 &&
6865                     mddev->minor_version > 90)
6866                         rdev->recovery_offset = reshape_offset;
6867
6868                 if (rdev->recovery_offset < reshape_offset) {
6869                         /* We need to check old and new layout */
6870                         if (!only_parity(rdev->raid_disk,
6871                                          conf->algorithm,
6872                                          conf->raid_disks,
6873                                          conf->max_degraded))
6874                                 continue;
6875                 }
6876                 if (!only_parity(rdev->raid_disk,
6877                                  conf->prev_algo,
6878                                  conf->previous_raid_disks,
6879                                  conf->max_degraded))
6880                         continue;
6881                 dirty_parity_disks++;
6882         }
6883
6884         /*
6885          * 0 for a fully functional array, 1 or 2 for a degraded array.
6886          */
6887         mddev->degraded = calc_degraded(conf);
6888
6889         if (has_failed(conf)) {
6890                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6891                         " (%d/%d failed)\n",
6892                         mdname(mddev), mddev->degraded, conf->raid_disks);
6893                 goto abort;
6894         }
6895
6896         /* device size must be a multiple of chunk size */
6897         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6898         mddev->resync_max_sectors = mddev->dev_sectors;
6899
6900         if (mddev->degraded > dirty_parity_disks &&
6901             mddev->recovery_cp != MaxSector) {
6902                 if (mddev->ok_start_degraded)
6903                         printk(KERN_WARNING
6904                                "md/raid:%s: starting dirty degraded array"
6905                                " - data corruption possible.\n",
6906                                mdname(mddev));
6907                 else {
6908                         printk(KERN_ERR
6909                                "md/raid:%s: cannot start dirty degraded array.\n",
6910                                mdname(mddev));
6911                         goto abort;
6912                 }
6913         }
6914
6915         if (mddev->degraded == 0)
6916                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6917                        " devices, algorithm %d\n", mdname(mddev), conf->level,
6918                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6919                        mddev->new_layout);
6920         else
6921                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6922                        " out of %d devices, algorithm %d\n",
6923                        mdname(mddev), conf->level,
6924                        mddev->raid_disks - mddev->degraded,
6925                        mddev->raid_disks, mddev->new_layout);
6926
6927         print_raid5_conf(conf);
6928
6929         if (conf->reshape_progress != MaxSector) {
6930                 conf->reshape_safe = conf->reshape_progress;
6931                 atomic_set(&conf->reshape_stripes, 0);
6932                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6933                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6934                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6935                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6936                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6937                                                         "reshape");
6938         }
6939
6940         /* Ok, everything is just fine now */
6941         if (mddev->to_remove == &raid5_attrs_group)
6942                 mddev->to_remove = NULL;
6943         else if (mddev->kobj.sd &&
6944             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6945                 printk(KERN_WARNING
6946                        "raid5: failed to create sysfs attributes for %s\n",
6947                        mdname(mddev));
6948         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6949
6950         if (mddev->queue) {
6951                 int chunk_size;
6952                 bool discard_supported = true;
6953                 /* read-ahead size must cover two whole stripes, which
6954                  * is 2 * (datadisks) * chunksize where 'n' is the
6955                  * number of raid devices
6956                  */
6957                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6958                 int stripe = data_disks *
6959                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6960                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6961                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6962
6963                 chunk_size = mddev->chunk_sectors << 9;
6964                 blk_queue_io_min(mddev->queue, chunk_size);
6965                 blk_queue_io_opt(mddev->queue, chunk_size *
6966                                  (conf->raid_disks - conf->max_degraded));
6967                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6968                 /*
6969                  * We can only discard a whole stripe. It doesn't make sense to
6970                  * discard data disk but write parity disk
6971                  */
6972                 stripe = stripe * PAGE_SIZE;
6973                 /* Round up to power of 2, as discard handling
6974                  * currently assumes that */
6975                 while ((stripe-1) & stripe)
6976                         stripe = (stripe | (stripe-1)) + 1;
6977                 mddev->queue->limits.discard_alignment = stripe;
6978                 mddev->queue->limits.discard_granularity = stripe;
6979                 /*
6980                  * unaligned part of discard request will be ignored, so can't
6981                  * guarantee discard_zeroes_data
6982                  */
6983                 mddev->queue->limits.discard_zeroes_data = 0;
6984
6985                 blk_queue_max_write_same_sectors(mddev->queue, 0);
6986
6987                 rdev_for_each(rdev, mddev) {
6988                         disk_stack_limits(mddev->gendisk, rdev->bdev,
6989                                           rdev->data_offset << 9);
6990                         disk_stack_limits(mddev->gendisk, rdev->bdev,
6991                                           rdev->new_data_offset << 9);
6992                         /*
6993                          * discard_zeroes_data is required, otherwise data
6994                          * could be lost. Consider a scenario: discard a stripe
6995                          * (the stripe could be inconsistent if
6996                          * discard_zeroes_data is 0); write one disk of the
6997                          * stripe (the stripe could be inconsistent again
6998                          * depending on which disks are used to calculate
6999                          * parity); the disk is broken; The stripe data of this
7000                          * disk is lost.
7001                          */
7002                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7003                             !bdev_get_queue(rdev->bdev)->
7004                                                 limits.discard_zeroes_data)
7005                                 discard_supported = false;
7006                         /* Unfortunately, discard_zeroes_data is not currently
7007                          * a guarantee - just a hint.  So we only allow DISCARD
7008                          * if the sysadmin has confirmed that only safe devices
7009                          * are in use by setting a module parameter.
7010                          */
7011                         if (!devices_handle_discard_safely) {
7012                                 if (discard_supported) {
7013                                         pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7014                                         pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7015                                 }
7016                                 discard_supported = false;
7017                         }
7018                 }
7019
7020                 if (discard_supported &&
7021                    mddev->queue->limits.max_discard_sectors >= stripe &&
7022                    mddev->queue->limits.discard_granularity >= stripe)
7023                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7024                                                 mddev->queue);
7025                 else
7026                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7027                                                 mddev->queue);
7028         }
7029
7030         if (journal_dev) {
7031                 char b[BDEVNAME_SIZE];
7032
7033                 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7034                        mdname(mddev), bdevname(journal_dev->bdev, b));
7035                 r5l_init_log(conf, journal_dev);
7036         }
7037
7038         return 0;
7039 abort:
7040         md_unregister_thread(&mddev->thread);
7041         print_raid5_conf(conf);
7042         free_conf(conf);
7043         mddev->private = NULL;
7044         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7045         return -EIO;
7046 }
7047
7048 static void raid5_free(struct mddev *mddev, void *priv)
7049 {
7050         struct r5conf *conf = priv;
7051
7052         free_conf(conf);
7053         mddev->to_remove = &raid5_attrs_group;
7054 }
7055
7056 static void status(struct seq_file *seq, struct mddev *mddev)
7057 {
7058         struct r5conf *conf = mddev->private;
7059         int i;
7060
7061         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7062                 conf->chunk_sectors / 2, mddev->layout);
7063         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7064         for (i = 0; i < conf->raid_disks; i++)
7065                 seq_printf (seq, "%s",
7066                                conf->disks[i].rdev &&
7067                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
7068         seq_printf (seq, "]");
7069 }
7070
7071 static void print_raid5_conf (struct r5conf *conf)
7072 {
7073         int i;
7074         struct disk_info *tmp;
7075
7076         printk(KERN_DEBUG "RAID conf printout:\n");
7077         if (!conf) {
7078                 printk("(conf==NULL)\n");
7079                 return;
7080         }
7081         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7082                conf->raid_disks,
7083                conf->raid_disks - conf->mddev->degraded);
7084
7085         for (i = 0; i < conf->raid_disks; i++) {
7086                 char b[BDEVNAME_SIZE];
7087                 tmp = conf->disks + i;
7088                 if (tmp->rdev)
7089                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7090                                i, !test_bit(Faulty, &tmp->rdev->flags),
7091                                bdevname(tmp->rdev->bdev, b));
7092         }
7093 }
7094
7095 static int raid5_spare_active(struct mddev *mddev)
7096 {
7097         int i;
7098         struct r5conf *conf = mddev->private;
7099         struct disk_info *tmp;
7100         int count = 0;
7101         unsigned long flags;
7102
7103         for (i = 0; i < conf->raid_disks; i++) {
7104                 tmp = conf->disks + i;
7105                 if (tmp->replacement
7106                     && tmp->replacement->recovery_offset == MaxSector
7107                     && !test_bit(Faulty, &tmp->replacement->flags)
7108                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7109                         /* Replacement has just become active. */
7110                         if (!tmp->rdev
7111                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7112                                 count++;
7113                         if (tmp->rdev) {
7114                                 /* Replaced device not technically faulty,
7115                                  * but we need to be sure it gets removed
7116                                  * and never re-added.
7117                                  */
7118                                 set_bit(Faulty, &tmp->rdev->flags);
7119                                 sysfs_notify_dirent_safe(
7120                                         tmp->rdev->sysfs_state);
7121                         }
7122                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7123                 } else if (tmp->rdev
7124                     && tmp->rdev->recovery_offset == MaxSector
7125                     && !test_bit(Faulty, &tmp->rdev->flags)
7126                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7127                         count++;
7128                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7129                 }
7130         }
7131         spin_lock_irqsave(&conf->device_lock, flags);
7132         mddev->degraded = calc_degraded(conf);
7133         spin_unlock_irqrestore(&conf->device_lock, flags);
7134         print_raid5_conf(conf);
7135         return count;
7136 }
7137
7138 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7139 {
7140         struct r5conf *conf = mddev->private;
7141         int err = 0;
7142         int number = rdev->raid_disk;
7143         struct md_rdev **rdevp;
7144         struct disk_info *p = conf->disks + number;
7145
7146         print_raid5_conf(conf);
7147         if (test_bit(Journal, &rdev->flags)) {
7148                 /*
7149                  * journal disk is not removable, but we need give a chance to
7150                  * update superblock of other disks. Otherwise journal disk
7151                  * will be considered as 'fresh'
7152                  */
7153                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7154                 return -EINVAL;
7155         }
7156         if (rdev == p->rdev)
7157                 rdevp = &p->rdev;
7158         else if (rdev == p->replacement)
7159                 rdevp = &p->replacement;
7160         else
7161                 return 0;
7162
7163         if (number >= conf->raid_disks &&
7164             conf->reshape_progress == MaxSector)
7165                 clear_bit(In_sync, &rdev->flags);
7166
7167         if (test_bit(In_sync, &rdev->flags) ||
7168             atomic_read(&rdev->nr_pending)) {
7169                 err = -EBUSY;
7170                 goto abort;
7171         }
7172         /* Only remove non-faulty devices if recovery
7173          * isn't possible.
7174          */
7175         if (!test_bit(Faulty, &rdev->flags) &&
7176             mddev->recovery_disabled != conf->recovery_disabled &&
7177             !has_failed(conf) &&
7178             (!p->replacement || p->replacement == rdev) &&
7179             number < conf->raid_disks) {
7180                 err = -EBUSY;
7181                 goto abort;
7182         }
7183         *rdevp = NULL;
7184         synchronize_rcu();
7185         if (atomic_read(&rdev->nr_pending)) {
7186                 /* lost the race, try later */
7187                 err = -EBUSY;
7188                 *rdevp = rdev;
7189         } else if (p->replacement) {
7190                 /* We must have just cleared 'rdev' */
7191                 p->rdev = p->replacement;
7192                 clear_bit(Replacement, &p->replacement->flags);
7193                 smp_mb(); /* Make sure other CPUs may see both as identical
7194                            * but will never see neither - if they are careful
7195                            */
7196                 p->replacement = NULL;
7197                 clear_bit(WantReplacement, &rdev->flags);
7198         } else
7199                 /* We might have just removed the Replacement as faulty-
7200                  * clear the bit just in case
7201                  */
7202                 clear_bit(WantReplacement, &rdev->flags);
7203 abort:
7204
7205         print_raid5_conf(conf);
7206         return err;
7207 }
7208
7209 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7210 {
7211         struct r5conf *conf = mddev->private;
7212         int err = -EEXIST;
7213         int disk;
7214         struct disk_info *p;
7215         int first = 0;
7216         int last = conf->raid_disks - 1;
7217
7218         if (test_bit(Journal, &rdev->flags))
7219                 return -EINVAL;
7220         if (mddev->recovery_disabled == conf->recovery_disabled)
7221                 return -EBUSY;
7222
7223         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7224                 /* no point adding a device */
7225                 return -EINVAL;
7226
7227         if (rdev->raid_disk >= 0)
7228                 first = last = rdev->raid_disk;
7229
7230         /*
7231          * find the disk ... but prefer rdev->saved_raid_disk
7232          * if possible.
7233          */
7234         if (rdev->saved_raid_disk >= 0 &&
7235             rdev->saved_raid_disk >= first &&
7236             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7237                 first = rdev->saved_raid_disk;
7238
7239         for (disk = first; disk <= last; disk++) {
7240                 p = conf->disks + disk;
7241                 if (p->rdev == NULL) {
7242                         clear_bit(In_sync, &rdev->flags);
7243                         rdev->raid_disk = disk;
7244                         err = 0;
7245                         if (rdev->saved_raid_disk != disk)
7246                                 conf->fullsync = 1;
7247                         rcu_assign_pointer(p->rdev, rdev);
7248                         goto out;
7249                 }
7250         }
7251         for (disk = first; disk <= last; disk++) {
7252                 p = conf->disks + disk;
7253                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7254                     p->replacement == NULL) {
7255                         clear_bit(In_sync, &rdev->flags);
7256                         set_bit(Replacement, &rdev->flags);
7257                         rdev->raid_disk = disk;
7258                         err = 0;
7259                         conf->fullsync = 1;
7260                         rcu_assign_pointer(p->replacement, rdev);
7261                         break;
7262                 }
7263         }
7264 out:
7265         print_raid5_conf(conf);
7266         return err;
7267 }
7268
7269 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7270 {
7271         /* no resync is happening, and there is enough space
7272          * on all devices, so we can resize.
7273          * We need to make sure resync covers any new space.
7274          * If the array is shrinking we should possibly wait until
7275          * any io in the removed space completes, but it hardly seems
7276          * worth it.
7277          */
7278         sector_t newsize;
7279         struct r5conf *conf = mddev->private;
7280
7281         if (conf->log)
7282                 return -EINVAL;
7283         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7284         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7285         if (mddev->external_size &&
7286             mddev->array_sectors > newsize)
7287                 return -EINVAL;
7288         if (mddev->bitmap) {
7289                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7290                 if (ret)
7291                         return ret;
7292         }
7293         md_set_array_sectors(mddev, newsize);
7294         set_capacity(mddev->gendisk, mddev->array_sectors);
7295         revalidate_disk(mddev->gendisk);
7296         if (sectors > mddev->dev_sectors &&
7297             mddev->recovery_cp > mddev->dev_sectors) {
7298                 mddev->recovery_cp = mddev->dev_sectors;
7299                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7300         }
7301         mddev->dev_sectors = sectors;
7302         mddev->resync_max_sectors = sectors;
7303         return 0;
7304 }
7305
7306 static int check_stripe_cache(struct mddev *mddev)
7307 {
7308         /* Can only proceed if there are plenty of stripe_heads.
7309          * We need a minimum of one full stripe,, and for sensible progress
7310          * it is best to have about 4 times that.
7311          * If we require 4 times, then the default 256 4K stripe_heads will
7312          * allow for chunk sizes up to 256K, which is probably OK.
7313          * If the chunk size is greater, user-space should request more
7314          * stripe_heads first.
7315          */
7316         struct r5conf *conf = mddev->private;
7317         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7318             > conf->min_nr_stripes ||
7319             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7320             > conf->min_nr_stripes) {
7321                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7322                        mdname(mddev),
7323                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7324                         / STRIPE_SIZE)*4);
7325                 return 0;
7326         }
7327         return 1;
7328 }
7329
7330 static int check_reshape(struct mddev *mddev)
7331 {
7332         struct r5conf *conf = mddev->private;
7333
7334         if (conf->log)
7335                 return -EINVAL;
7336         if (mddev->delta_disks == 0 &&
7337             mddev->new_layout == mddev->layout &&
7338             mddev->new_chunk_sectors == mddev->chunk_sectors)
7339                 return 0; /* nothing to do */
7340         if (has_failed(conf))
7341                 return -EINVAL;
7342         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7343                 /* We might be able to shrink, but the devices must
7344                  * be made bigger first.
7345                  * For raid6, 4 is the minimum size.
7346                  * Otherwise 2 is the minimum
7347                  */
7348                 int min = 2;
7349                 if (mddev->level == 6)
7350                         min = 4;
7351                 if (mddev->raid_disks + mddev->delta_disks < min)
7352                         return -EINVAL;
7353         }
7354
7355         if (!check_stripe_cache(mddev))
7356                 return -ENOSPC;
7357
7358         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7359             mddev->delta_disks > 0)
7360                 if (resize_chunks(conf,
7361                                   conf->previous_raid_disks
7362                                   + max(0, mddev->delta_disks),
7363                                   max(mddev->new_chunk_sectors,
7364                                       mddev->chunk_sectors)
7365                             ) < 0)
7366                         return -ENOMEM;
7367         return resize_stripes(conf, (conf->previous_raid_disks
7368                                      + mddev->delta_disks));
7369 }
7370
7371 static int raid5_start_reshape(struct mddev *mddev)
7372 {
7373         struct r5conf *conf = mddev->private;
7374         struct md_rdev *rdev;
7375         int spares = 0;
7376         unsigned long flags;
7377
7378         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7379                 return -EBUSY;
7380
7381         if (!check_stripe_cache(mddev))
7382                 return -ENOSPC;
7383
7384         if (has_failed(conf))
7385                 return -EINVAL;
7386
7387         rdev_for_each(rdev, mddev) {
7388                 if (!test_bit(In_sync, &rdev->flags)
7389                     && !test_bit(Faulty, &rdev->flags))
7390                         spares++;
7391         }
7392
7393         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7394                 /* Not enough devices even to make a degraded array
7395                  * of that size
7396                  */
7397                 return -EINVAL;
7398
7399         /* Refuse to reduce size of the array.  Any reductions in
7400          * array size must be through explicit setting of array_size
7401          * attribute.
7402          */
7403         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7404             < mddev->array_sectors) {
7405                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7406                        "before number of disks\n", mdname(mddev));
7407                 return -EINVAL;
7408         }
7409
7410         atomic_set(&conf->reshape_stripes, 0);
7411         spin_lock_irq(&conf->device_lock);
7412         write_seqcount_begin(&conf->gen_lock);
7413         conf->previous_raid_disks = conf->raid_disks;
7414         conf->raid_disks += mddev->delta_disks;
7415         conf->prev_chunk_sectors = conf->chunk_sectors;
7416         conf->chunk_sectors = mddev->new_chunk_sectors;
7417         conf->prev_algo = conf->algorithm;
7418         conf->algorithm = mddev->new_layout;
7419         conf->generation++;
7420         /* Code that selects data_offset needs to see the generation update
7421          * if reshape_progress has been set - so a memory barrier needed.
7422          */
7423         smp_mb();
7424         if (mddev->reshape_backwards)
7425                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7426         else
7427                 conf->reshape_progress = 0;
7428         conf->reshape_safe = conf->reshape_progress;
7429         write_seqcount_end(&conf->gen_lock);
7430         spin_unlock_irq(&conf->device_lock);
7431
7432         /* Now make sure any requests that proceeded on the assumption
7433          * the reshape wasn't running - like Discard or Read - have
7434          * completed.
7435          */
7436         mddev_suspend(mddev);
7437         mddev_resume(mddev);
7438
7439         /* Add some new drives, as many as will fit.
7440          * We know there are enough to make the newly sized array work.
7441          * Don't add devices if we are reducing the number of
7442          * devices in the array.  This is because it is not possible
7443          * to correctly record the "partially reconstructed" state of
7444          * such devices during the reshape and confusion could result.
7445          */
7446         if (mddev->delta_disks >= 0) {
7447                 rdev_for_each(rdev, mddev)
7448                         if (rdev->raid_disk < 0 &&
7449                             !test_bit(Faulty, &rdev->flags)) {
7450                                 if (raid5_add_disk(mddev, rdev) == 0) {
7451                                         if (rdev->raid_disk
7452                                             >= conf->previous_raid_disks)
7453                                                 set_bit(In_sync, &rdev->flags);
7454                                         else
7455                                                 rdev->recovery_offset = 0;
7456
7457                                         if (sysfs_link_rdev(mddev, rdev))
7458                                                 /* Failure here is OK */;
7459                                 }
7460                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7461                                    && !test_bit(Faulty, &rdev->flags)) {
7462                                 /* This is a spare that was manually added */
7463                                 set_bit(In_sync, &rdev->flags);
7464                         }
7465
7466                 /* When a reshape changes the number of devices,
7467                  * ->degraded is measured against the larger of the
7468                  * pre and post number of devices.
7469                  */
7470                 spin_lock_irqsave(&conf->device_lock, flags);
7471                 mddev->degraded = calc_degraded(conf);
7472                 spin_unlock_irqrestore(&conf->device_lock, flags);
7473         }
7474         mddev->raid_disks = conf->raid_disks;
7475         mddev->reshape_position = conf->reshape_progress;
7476         set_bit(MD_CHANGE_DEVS, &mddev->flags);
7477
7478         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7479         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7480         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7481         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7482         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7483         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7484                                                 "reshape");
7485         if (!mddev->sync_thread) {
7486                 mddev->recovery = 0;
7487                 spin_lock_irq(&conf->device_lock);
7488                 write_seqcount_begin(&conf->gen_lock);
7489                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7490                 mddev->new_chunk_sectors =
7491                         conf->chunk_sectors = conf->prev_chunk_sectors;
7492                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7493                 rdev_for_each(rdev, mddev)
7494                         rdev->new_data_offset = rdev->data_offset;
7495                 smp_wmb();
7496                 conf->generation --;
7497                 conf->reshape_progress = MaxSector;
7498                 mddev->reshape_position = MaxSector;
7499                 write_seqcount_end(&conf->gen_lock);
7500                 spin_unlock_irq(&conf->device_lock);
7501                 return -EAGAIN;
7502         }
7503         conf->reshape_checkpoint = jiffies;
7504         md_wakeup_thread(mddev->sync_thread);
7505         md_new_event(mddev);
7506         return 0;
7507 }
7508
7509 /* This is called from the reshape thread and should make any
7510  * changes needed in 'conf'
7511  */
7512 static void end_reshape(struct r5conf *conf)
7513 {
7514
7515         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7516                 struct md_rdev *rdev;
7517
7518                 spin_lock_irq(&conf->device_lock);
7519                 conf->previous_raid_disks = conf->raid_disks;
7520                 rdev_for_each(rdev, conf->mddev)
7521                         rdev->data_offset = rdev->new_data_offset;
7522                 smp_wmb();
7523                 conf->reshape_progress = MaxSector;
7524                 conf->mddev->reshape_position = MaxSector;
7525                 spin_unlock_irq(&conf->device_lock);
7526                 wake_up(&conf->wait_for_overlap);
7527
7528                 /* read-ahead size must cover two whole stripes, which is
7529                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7530                  */
7531                 if (conf->mddev->queue) {
7532                         int data_disks = conf->raid_disks - conf->max_degraded;
7533                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7534                                                    / PAGE_SIZE);
7535                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7536                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7537                 }
7538         }
7539 }
7540
7541 /* This is called from the raid5d thread with mddev_lock held.
7542  * It makes config changes to the device.
7543  */
7544 static void raid5_finish_reshape(struct mddev *mddev)
7545 {
7546         struct r5conf *conf = mddev->private;
7547
7548         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7549
7550                 if (mddev->delta_disks > 0) {
7551                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7552                         set_capacity(mddev->gendisk, mddev->array_sectors);
7553                         revalidate_disk(mddev->gendisk);
7554                 } else {
7555                         int d;
7556                         spin_lock_irq(&conf->device_lock);
7557                         mddev->degraded = calc_degraded(conf);
7558                         spin_unlock_irq(&conf->device_lock);
7559                         for (d = conf->raid_disks ;
7560                              d < conf->raid_disks - mddev->delta_disks;
7561                              d++) {
7562                                 struct md_rdev *rdev = conf->disks[d].rdev;
7563                                 if (rdev)
7564                                         clear_bit(In_sync, &rdev->flags);
7565                                 rdev = conf->disks[d].replacement;
7566                                 if (rdev)
7567                                         clear_bit(In_sync, &rdev->flags);
7568                         }
7569                 }
7570                 mddev->layout = conf->algorithm;
7571                 mddev->chunk_sectors = conf->chunk_sectors;
7572                 mddev->reshape_position = MaxSector;
7573                 mddev->delta_disks = 0;
7574                 mddev->reshape_backwards = 0;
7575         }
7576 }
7577
7578 static void raid5_quiesce(struct mddev *mddev, int state)
7579 {
7580         struct r5conf *conf = mddev->private;
7581
7582         switch(state) {
7583         case 2: /* resume for a suspend */
7584                 wake_up(&conf->wait_for_overlap);
7585                 break;
7586
7587         case 1: /* stop all writes */
7588                 lock_all_device_hash_locks_irq(conf);
7589                 /* '2' tells resync/reshape to pause so that all
7590                  * active stripes can drain
7591                  */
7592                 conf->quiesce = 2;
7593                 wait_event_cmd(conf->wait_for_quiescent,
7594                                     atomic_read(&conf->active_stripes) == 0 &&
7595                                     atomic_read(&conf->active_aligned_reads) == 0,
7596                                     unlock_all_device_hash_locks_irq(conf),
7597                                     lock_all_device_hash_locks_irq(conf));
7598                 conf->quiesce = 1;
7599                 unlock_all_device_hash_locks_irq(conf);
7600                 /* allow reshape to continue */
7601                 wake_up(&conf->wait_for_overlap);
7602                 break;
7603
7604         case 0: /* re-enable writes */
7605                 lock_all_device_hash_locks_irq(conf);
7606                 conf->quiesce = 0;
7607                 wake_up(&conf->wait_for_quiescent);
7608                 wake_up(&conf->wait_for_overlap);
7609                 unlock_all_device_hash_locks_irq(conf);
7610                 break;
7611         }
7612         r5l_quiesce(conf->log, state);
7613 }
7614
7615 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7616 {
7617         struct r0conf *raid0_conf = mddev->private;
7618         sector_t sectors;
7619
7620         /* for raid0 takeover only one zone is supported */
7621         if (raid0_conf->nr_strip_zones > 1) {
7622                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7623                        mdname(mddev));
7624                 return ERR_PTR(-EINVAL);
7625         }
7626
7627         sectors = raid0_conf->strip_zone[0].zone_end;
7628         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7629         mddev->dev_sectors = sectors;
7630         mddev->new_level = level;
7631         mddev->new_layout = ALGORITHM_PARITY_N;
7632         mddev->new_chunk_sectors = mddev->chunk_sectors;
7633         mddev->raid_disks += 1;
7634         mddev->delta_disks = 1;
7635         /* make sure it will be not marked as dirty */
7636         mddev->recovery_cp = MaxSector;
7637
7638         return setup_conf(mddev);
7639 }
7640
7641 static void *raid5_takeover_raid1(struct mddev *mddev)
7642 {
7643         int chunksect;
7644
7645         if (mddev->raid_disks != 2 ||
7646             mddev->degraded > 1)
7647                 return ERR_PTR(-EINVAL);
7648
7649         /* Should check if there are write-behind devices? */
7650
7651         chunksect = 64*2; /* 64K by default */
7652
7653         /* The array must be an exact multiple of chunksize */
7654         while (chunksect && (mddev->array_sectors & (chunksect-1)))
7655                 chunksect >>= 1;
7656
7657         if ((chunksect<<9) < STRIPE_SIZE)
7658                 /* array size does not allow a suitable chunk size */
7659                 return ERR_PTR(-EINVAL);
7660
7661         mddev->new_level = 5;
7662         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7663         mddev->new_chunk_sectors = chunksect;
7664
7665         return setup_conf(mddev);
7666 }
7667
7668 static void *raid5_takeover_raid6(struct mddev *mddev)
7669 {
7670         int new_layout;
7671
7672         switch (mddev->layout) {
7673         case ALGORITHM_LEFT_ASYMMETRIC_6:
7674                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7675                 break;
7676         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7677                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7678                 break;
7679         case ALGORITHM_LEFT_SYMMETRIC_6:
7680                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7681                 break;
7682         case ALGORITHM_RIGHT_SYMMETRIC_6:
7683                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7684                 break;
7685         case ALGORITHM_PARITY_0_6:
7686                 new_layout = ALGORITHM_PARITY_0;
7687                 break;
7688         case ALGORITHM_PARITY_N:
7689                 new_layout = ALGORITHM_PARITY_N;
7690                 break;
7691         default:
7692                 return ERR_PTR(-EINVAL);
7693         }
7694         mddev->new_level = 5;
7695         mddev->new_layout = new_layout;
7696         mddev->delta_disks = -1;
7697         mddev->raid_disks -= 1;
7698         return setup_conf(mddev);
7699 }
7700
7701 static int raid5_check_reshape(struct mddev *mddev)
7702 {
7703         /* For a 2-drive array, the layout and chunk size can be changed
7704          * immediately as not restriping is needed.
7705          * For larger arrays we record the new value - after validation
7706          * to be used by a reshape pass.
7707          */
7708         struct r5conf *conf = mddev->private;
7709         int new_chunk = mddev->new_chunk_sectors;
7710
7711         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7712                 return -EINVAL;
7713         if (new_chunk > 0) {
7714                 if (!is_power_of_2(new_chunk))
7715                         return -EINVAL;
7716                 if (new_chunk < (PAGE_SIZE>>9))
7717                         return -EINVAL;
7718                 if (mddev->array_sectors & (new_chunk-1))
7719                         /* not factor of array size */
7720                         return -EINVAL;
7721         }
7722
7723         /* They look valid */
7724
7725         if (mddev->raid_disks == 2) {
7726                 /* can make the change immediately */
7727                 if (mddev->new_layout >= 0) {
7728                         conf->algorithm = mddev->new_layout;
7729                         mddev->layout = mddev->new_layout;
7730                 }
7731                 if (new_chunk > 0) {
7732                         conf->chunk_sectors = new_chunk ;
7733                         mddev->chunk_sectors = new_chunk;
7734                 }
7735                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7736                 md_wakeup_thread(mddev->thread);
7737         }
7738         return check_reshape(mddev);
7739 }
7740
7741 static int raid6_check_reshape(struct mddev *mddev)
7742 {
7743         int new_chunk = mddev->new_chunk_sectors;
7744
7745         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7746                 return -EINVAL;
7747         if (new_chunk > 0) {
7748                 if (!is_power_of_2(new_chunk))
7749                         return -EINVAL;
7750                 if (new_chunk < (PAGE_SIZE >> 9))
7751                         return -EINVAL;
7752                 if (mddev->array_sectors & (new_chunk-1))
7753                         /* not factor of array size */
7754                         return -EINVAL;
7755         }
7756
7757         /* They look valid */
7758         return check_reshape(mddev);
7759 }
7760
7761 static void *raid5_takeover(struct mddev *mddev)
7762 {
7763         /* raid5 can take over:
7764          *  raid0 - if there is only one strip zone - make it a raid4 layout
7765          *  raid1 - if there are two drives.  We need to know the chunk size
7766          *  raid4 - trivial - just use a raid4 layout.
7767          *  raid6 - Providing it is a *_6 layout
7768          */
7769         if (mddev->level == 0)
7770                 return raid45_takeover_raid0(mddev, 5);
7771         if (mddev->level == 1)
7772                 return raid5_takeover_raid1(mddev);
7773         if (mddev->level == 4) {
7774                 mddev->new_layout = ALGORITHM_PARITY_N;
7775                 mddev->new_level = 5;
7776                 return setup_conf(mddev);
7777         }
7778         if (mddev->level == 6)
7779                 return raid5_takeover_raid6(mddev);
7780
7781         return ERR_PTR(-EINVAL);
7782 }
7783
7784 static void *raid4_takeover(struct mddev *mddev)
7785 {
7786         /* raid4 can take over:
7787          *  raid0 - if there is only one strip zone
7788          *  raid5 - if layout is right
7789          */
7790         if (mddev->level == 0)
7791                 return raid45_takeover_raid0(mddev, 4);
7792         if (mddev->level == 5 &&
7793             mddev->layout == ALGORITHM_PARITY_N) {
7794                 mddev->new_layout = 0;
7795                 mddev->new_level = 4;
7796                 return setup_conf(mddev);
7797         }
7798         return ERR_PTR(-EINVAL);
7799 }
7800
7801 static struct md_personality raid5_personality;
7802
7803 static void *raid6_takeover(struct mddev *mddev)
7804 {
7805         /* Currently can only take over a raid5.  We map the
7806          * personality to an equivalent raid6 personality
7807          * with the Q block at the end.
7808          */
7809         int new_layout;
7810
7811         if (mddev->pers != &raid5_personality)
7812                 return ERR_PTR(-EINVAL);
7813         if (mddev->degraded > 1)
7814                 return ERR_PTR(-EINVAL);
7815         if (mddev->raid_disks > 253)
7816                 return ERR_PTR(-EINVAL);
7817         if (mddev->raid_disks < 3)
7818                 return ERR_PTR(-EINVAL);
7819
7820         switch (mddev->layout) {
7821         case ALGORITHM_LEFT_ASYMMETRIC:
7822                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7823                 break;
7824         case ALGORITHM_RIGHT_ASYMMETRIC:
7825                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7826                 break;
7827         case ALGORITHM_LEFT_SYMMETRIC:
7828                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7829                 break;
7830         case ALGORITHM_RIGHT_SYMMETRIC:
7831                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7832                 break;
7833         case ALGORITHM_PARITY_0:
7834                 new_layout = ALGORITHM_PARITY_0_6;
7835                 break;
7836         case ALGORITHM_PARITY_N:
7837                 new_layout = ALGORITHM_PARITY_N;
7838                 break;
7839         default:
7840                 return ERR_PTR(-EINVAL);
7841         }
7842         mddev->new_level = 6;
7843         mddev->new_layout = new_layout;
7844         mddev->delta_disks = 1;
7845         mddev->raid_disks += 1;
7846         return setup_conf(mddev);
7847 }
7848
7849 static struct md_personality raid6_personality =
7850 {
7851         .name           = "raid6",
7852         .level          = 6,
7853         .owner          = THIS_MODULE,
7854         .make_request   = make_request,
7855         .run            = run,
7856         .free           = raid5_free,
7857         .status         = status,
7858         .error_handler  = error,
7859         .hot_add_disk   = raid5_add_disk,
7860         .hot_remove_disk= raid5_remove_disk,
7861         .spare_active   = raid5_spare_active,
7862         .sync_request   = sync_request,
7863         .resize         = raid5_resize,
7864         .size           = raid5_size,
7865         .check_reshape  = raid6_check_reshape,
7866         .start_reshape  = raid5_start_reshape,
7867         .finish_reshape = raid5_finish_reshape,
7868         .quiesce        = raid5_quiesce,
7869         .takeover       = raid6_takeover,
7870         .congested      = raid5_congested,
7871 };
7872 static struct md_personality raid5_personality =
7873 {
7874         .name           = "raid5",
7875         .level          = 5,
7876         .owner          = THIS_MODULE,
7877         .make_request   = make_request,
7878         .run            = run,
7879         .free           = raid5_free,
7880         .status         = status,
7881         .error_handler  = error,
7882         .hot_add_disk   = raid5_add_disk,
7883         .hot_remove_disk= raid5_remove_disk,
7884         .spare_active   = raid5_spare_active,
7885         .sync_request   = sync_request,
7886         .resize         = raid5_resize,
7887         .size           = raid5_size,
7888         .check_reshape  = raid5_check_reshape,
7889         .start_reshape  = raid5_start_reshape,
7890         .finish_reshape = raid5_finish_reshape,
7891         .quiesce        = raid5_quiesce,
7892         .takeover       = raid5_takeover,
7893         .congested      = raid5_congested,
7894 };
7895
7896 static struct md_personality raid4_personality =
7897 {
7898         .name           = "raid4",
7899         .level          = 4,
7900         .owner          = THIS_MODULE,
7901         .make_request   = make_request,
7902         .run            = run,
7903         .free           = raid5_free,
7904         .status         = status,
7905         .error_handler  = error,
7906         .hot_add_disk   = raid5_add_disk,
7907         .hot_remove_disk= raid5_remove_disk,
7908         .spare_active   = raid5_spare_active,
7909         .sync_request   = sync_request,
7910         .resize         = raid5_resize,
7911         .size           = raid5_size,
7912         .check_reshape  = raid5_check_reshape,
7913         .start_reshape  = raid5_start_reshape,
7914         .finish_reshape = raid5_finish_reshape,
7915         .quiesce        = raid5_quiesce,
7916         .takeover       = raid4_takeover,
7917         .congested      = raid5_congested,
7918 };
7919
7920 static int __init raid5_init(void)
7921 {
7922         raid5_wq = alloc_workqueue("raid5wq",
7923                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7924         if (!raid5_wq)
7925                 return -ENOMEM;
7926         register_md_personality(&raid6_personality);
7927         register_md_personality(&raid5_personality);
7928         register_md_personality(&raid4_personality);
7929         return 0;
7930 }
7931
7932 static void raid5_exit(void)
7933 {
7934         unregister_md_personality(&raid6_personality);
7935         unregister_md_personality(&raid5_personality);
7936         unregister_md_personality(&raid4_personality);
7937         destroy_workqueue(raid5_wq);
7938 }
7939
7940 module_init(raid5_init);
7941 module_exit(raid5_exit);
7942 MODULE_LICENSE("GPL");
7943 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7944 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7945 MODULE_ALIAS("md-raid5");
7946 MODULE_ALIAS("md-raid4");
7947 MODULE_ALIAS("md-level-5");
7948 MODULE_ALIAS("md-level-4");
7949 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7950 MODULE_ALIAS("md-raid6");
7951 MODULE_ALIAS("md-level-6");
7952
7953 /* This used to be two separate modules, they were: */
7954 MODULE_ALIAS("raid5");
7955 MODULE_ALIAS("raid6");