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 <trace/events/block.h>
  57
  58 #include "md.h"
  59 #include "raid5.h"
  60 #include "raid0.h"
  61 #include "bitmap.h"
  62
  63 /*
  64  * Stripe cache
  65  */
  66
  67 #define NR_STRIPES              256
  68 #define STRIPE_SIZE             PAGE_SIZE
  69 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  70 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  71 #define IO_THRESHOLD            1
  72 #define BYPASS_THRESHOLD        1
  73 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  74 #define HASH_MASK               (NR_HASH - 1)
  75
  76 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  77 {
  78         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  79         return &conf->stripe_hashtbl[hash];
  80 }
  81
  82 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  83  * order without overlap.  There may be several bio's per stripe+device, and
  84  * a bio could span several devices.
  85  * When walking this list for a particular stripe+device, we must never proceed
  86  * beyond a bio that extends past this device, as the next bio might no longer
  87  * be valid.
  88  * This function is used to determine the 'next' bio in the list, given the sector
  89  * of the current stripe+device
  90  */
  91 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
  92 {
  93         int sectors = bio->bi_size >> 9;
  94         if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
  95                 return bio->bi_next;
  96         else
  97                 return NULL;
  98 }
  99
 100 /*
 101  * We maintain a biased count of active stripes in the bottom 16 bits of
 102  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
 103  */
 104 static inline int raid5_bi_processed_stripes(struct bio *bio)
 105 {
 106         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 107         return (atomic_read(segments) >> 16) & 0xffff;
 108 }
 109
 110 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
 111 {
 112         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 113         return atomic_sub_return(1, segments) & 0xffff;
 114 }
 115
 116 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
 117 {
 118         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 119         atomic_inc(segments);
 120 }
 121
 122 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
 123         unsigned int cnt)
 124 {
 125         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 126         int old, new;
 127
 128         do {
 129                 old = atomic_read(segments);
 130                 new = (old & 0xffff) | (cnt << 16);
 131         } while (atomic_cmpxchg(segments, old, new) != old);
 132 }
 133
 134 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
 135 {
 136         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 137         atomic_set(segments, cnt);
 138 }
 139
 140 /* Find first data disk in a raid6 stripe */
 141 static inline int raid6_d0(struct stripe_head *sh)
 142 {
 143         if (sh->ddf_layout)
 144                 /* ddf always start from first device */
 145                 return 0;
 146         /* md starts just after Q block */
 147         if (sh->qd_idx == sh->disks - 1)
 148                 return 0;
 149         else
 150                 return sh->qd_idx + 1;
 151 }
 152 static inline int raid6_next_disk(int disk, int raid_disks)
 153 {
 154         disk++;
 155         return (disk < raid_disks) ? disk : 0;
 156 }
 157
 158 /* When walking through the disks in a raid5, starting at raid6_d0,
 159  * We need to map each disk to a 'slot', where the data disks are slot
 160  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 161  * is raid_disks-1.  This help does that mapping.
 162  */
 163 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 164                              int *count, int syndrome_disks)
 165 {
 166         int slot = *count;
 167
 168         if (sh->ddf_layout)
 169                 (*count)++;
 170         if (idx == sh->pd_idx)
 171                 return syndrome_disks;
 172         if (idx == sh->qd_idx)
 173                 return syndrome_disks + 1;
 174         if (!sh->ddf_layout)
 175                 (*count)++;
 176         return slot;
 177 }
 178
 179 static void return_io(struct bio *return_bi)
 180 {
 181         struct bio *bi = return_bi;
 182         while (bi) {
 183
 184                 return_bi = bi->bi_next;
 185                 bi->bi_next = NULL;
 186                 bi->bi_size = 0;
 187                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 188                                          bi, 0);
 189                 bio_endio(bi, 0);
 190                 bi = return_bi;
 191         }
 192 }
 193
 194 static void print_raid5_conf (struct r5conf *conf);
 195
 196 static int stripe_operations_active(struct stripe_head *sh)
 197 {
 198         return sh->check_state || sh->reconstruct_state ||
 199                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 200                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 201 }
 202
 203 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
 204 {
 205         BUG_ON(!list_empty(&sh->lru));
 206         BUG_ON(atomic_read(&conf->active_stripes)==0);
 207         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 208                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 209                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 210                         list_add_tail(&sh->lru, &conf->delayed_list);
 211                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 212                            sh->bm_seq - conf->seq_write > 0)
 213                         list_add_tail(&sh->lru, &conf->bitmap_list);
 214                 else {
 215                         clear_bit(STRIPE_DELAYED, &sh->state);
 216                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 217                         list_add_tail(&sh->lru, &conf->handle_list);
 218                 }
 219                 md_wakeup_thread(conf->mddev->thread);
 220         } else {
 221                 BUG_ON(stripe_operations_active(sh));
 222                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 223                         if (atomic_dec_return(&conf->preread_active_stripes)
 224                             < IO_THRESHOLD)
 225                                 md_wakeup_thread(conf->mddev->thread);
 226                 atomic_dec(&conf->active_stripes);
 227                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 228                         list_add_tail(&sh->lru, &conf->inactive_list);
 229                         wake_up(&conf->wait_for_stripe);
 230                         if (conf->retry_read_aligned)
 231                                 md_wakeup_thread(conf->mddev->thread);
 232                 }
 233         }
 234 }
 235
 236 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
 237 {
 238         if (atomic_dec_and_test(&sh->count))
 239                 do_release_stripe(conf, sh);
 240 }
 241
 242 static void release_stripe(struct stripe_head *sh)
 243 {
 244         struct r5conf *conf = sh->raid_conf;
 245         unsigned long flags;
 246
 247         local_irq_save(flags);
 248         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 249                 do_release_stripe(conf, sh);
 250                 spin_unlock(&conf->device_lock);
 251         }
 252         local_irq_restore(flags);
 253 }
 254
 255 static inline void remove_hash(struct stripe_head *sh)
 256 {
 257         pr_debug("remove_hash(), stripe %llu\n",
 258                 (unsigned long long)sh->sector);
 259
 260         hlist_del_init(&sh->hash);
 261 }
 262
 263 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 264 {
 265         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 266
 267         pr_debug("insert_hash(), stripe %llu\n",
 268                 (unsigned long long)sh->sector);
 269
 270         hlist_add_head(&sh->hash, hp);
 271 }
 272
 273
 274 /* find an idle stripe, make sure it is unhashed, and return it. */
 275 static struct stripe_head *get_free_stripe(struct r5conf *conf)
 276 {
 277         struct stripe_head *sh = NULL;
 278         struct list_head *first;
 279
 280         if (list_empty(&conf->inactive_list))
 281                 goto out;
 282         first = conf->inactive_list.next;
 283         sh = list_entry(first, struct stripe_head, lru);
 284         list_del_init(first);
 285         remove_hash(sh);
 286         atomic_inc(&conf->active_stripes);
 287 out:
 288         return sh;
 289 }
 290
 291 static void shrink_buffers(struct stripe_head *sh)
 292 {
 293         struct page *p;
 294         int i;
 295         int num = sh->raid_conf->pool_size;
 296
 297         for (i = 0; i < num ; i++) {
 298                 p = sh->dev[i].page;
 299                 if (!p)
 300                         continue;
 301                 sh->dev[i].page = NULL;
 302                 put_page(p);
 303         }
 304 }
 305
 306 static int grow_buffers(struct stripe_head *sh)
 307 {
 308         int i;
 309         int num = sh->raid_conf->pool_size;
 310
 311         for (i = 0; i < num; i++) {
 312                 struct page *page;
 313
 314                 if (!(page = alloc_page(GFP_KERNEL))) {
 315                         return 1;
 316                 }
 317                 sh->dev[i].page = page;
 318         }
 319         return 0;
 320 }
 321
 322 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 323 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 324                             struct stripe_head *sh);
 325
 326 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 327 {
 328         struct r5conf *conf = sh->raid_conf;
 329         int i;
 330
 331         BUG_ON(atomic_read(&sh->count) != 0);
 332         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 333         BUG_ON(stripe_operations_active(sh));
 334
 335         pr_debug("init_stripe called, stripe %llu\n",
 336                 (unsigned long long)sh->sector);
 337
 338         remove_hash(sh);
 339
 340         sh->generation = conf->generation - previous;
 341         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 342         sh->sector = sector;
 343         stripe_set_idx(sector, conf, previous, sh);
 344         sh->state = 0;
 345
 346
 347         for (i = sh->disks; i--; ) {
 348                 struct r5dev *dev = &sh->dev[i];
 349
 350                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 351                     test_bit(R5_LOCKED, &dev->flags)) {
 352                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 353                                (unsigned long long)sh->sector, i, dev->toread,
 354                                dev->read, dev->towrite, dev->written,
 355                                test_bit(R5_LOCKED, &dev->flags));
 356                         WARN_ON(1);
 357                 }
 358                 dev->flags = 0;
 359                 raid5_build_block(sh, i, previous);
 360         }
 361         insert_hash(conf, sh);
 362 }
 363
 364 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 365                                          short generation)
 366 {
 367         struct stripe_head *sh;
 368
 369         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 370         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 371                 if (sh->sector == sector && sh->generation == generation)
 372                         return sh;
 373         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 374         return NULL;
 375 }
 376
 377 /*
 378  * Need to check if array has failed when deciding whether to:
 379  *  - start an array
 380  *  - remove non-faulty devices
 381  *  - add a spare
 382  *  - allow a reshape
 383  * This determination is simple when no reshape is happening.
 384  * However if there is a reshape, we need to carefully check
 385  * both the before and after sections.
 386  * This is because some failed devices may only affect one
 387  * of the two sections, and some non-in_sync devices may
 388  * be insync in the section most affected by failed devices.
 389  */
 390 static int calc_degraded(struct r5conf *conf)
 391 {
 392         int degraded, degraded2;
 393         int i;
 394
 395         rcu_read_lock();
 396         degraded = 0;
 397         for (i = 0; i < conf->previous_raid_disks; i++) {
 398                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 399                 if (rdev && test_bit(Faulty, &rdev->flags))
 400                         rdev = rcu_dereference(conf->disks[i].replacement);
 401                 if (!rdev || test_bit(Faulty, &rdev->flags))
 402                         degraded++;
 403                 else if (test_bit(In_sync, &rdev->flags))
 404                         ;
 405                 else
 406                         /* not in-sync or faulty.
 407                          * If the reshape increases the number of devices,
 408                          * this is being recovered by the reshape, so
 409                          * this 'previous' section is not in_sync.
 410                          * If the number of devices is being reduced however,
 411                          * the device can only be part of the array if
 412                          * we are reverting a reshape, so this section will
 413                          * be in-sync.
 414                          */
 415                         if (conf->raid_disks >= conf->previous_raid_disks)
 416                                 degraded++;
 417         }
 418         rcu_read_unlock();
 419         if (conf->raid_disks == conf->previous_raid_disks)
 420                 return degraded;
 421         rcu_read_lock();
 422         degraded2 = 0;
 423         for (i = 0; i < conf->raid_disks; i++) {
 424                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 425                 if (rdev && test_bit(Faulty, &rdev->flags))
 426                         rdev = rcu_dereference(conf->disks[i].replacement);
 427                 if (!rdev || test_bit(Faulty, &rdev->flags))
 428                         degraded2++;
 429                 else if (test_bit(In_sync, &rdev->flags))
 430                         ;
 431                 else
 432                         /* not in-sync or faulty.
 433                          * If reshape increases the number of devices, this
 434                          * section has already been recovered, else it
 435                          * almost certainly hasn't.
 436                          */
 437                         if (conf->raid_disks <= conf->previous_raid_disks)
 438                                 degraded2++;
 439         }
 440         rcu_read_unlock();
 441         if (degraded2 > degraded)
 442                 return degraded2;
 443         return degraded;
 444 }
 445
 446 static int has_failed(struct r5conf *conf)
 447 {
 448         int degraded;
 449
 450         if (conf->mddev->reshape_position == MaxSector)
 451                 return conf->mddev->degraded > conf->max_degraded;
 452
 453         degraded = calc_degraded(conf);
 454         if (degraded > conf->max_degraded)
 455                 return 1;
 456         return 0;
 457 }
 458
 459 static struct stripe_head *
 460 get_active_stripe(struct r5conf *conf, sector_t sector,
 461                   int previous, int noblock, int noquiesce)
 462 {
 463         struct stripe_head *sh;
 464
 465         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 466
 467         spin_lock_irq(&conf->device_lock);
 468
 469         do {
 470                 wait_event_lock_irq(conf->wait_for_stripe,
 471                                     conf->quiesce == 0 || noquiesce,
 472                                     conf->device_lock);
 473                 sh = __find_stripe(conf, sector, conf->generation - previous);
 474                 if (!sh) {
 475                         if (!conf->inactive_blocked)
 476                                 sh = get_free_stripe(conf);
 477                         if (noblock && sh == NULL)
 478                                 break;
 479                         if (!sh) {
 480                                 conf->inactive_blocked = 1;
 481                                 wait_event_lock_irq(conf->wait_for_stripe,
 482                                                     !list_empty(&conf->inactive_list) &&
 483                                                     (atomic_read(&conf->active_stripes)
 484                                                      < (conf->max_nr_stripes *3/4)
 485                                                      || !conf->inactive_blocked),
 486                                                     conf->device_lock);
 487                                 conf->inactive_blocked = 0;
 488                         } else
 489                                 init_stripe(sh, sector, previous);
 490                 } else {
 491                         if (atomic_read(&sh->count)) {
 492                                 BUG_ON(!list_empty(&sh->lru)
 493                                     && !test_bit(STRIPE_EXPANDING, &sh->state)
 494                                     && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state));
 495                         } else {
 496                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 497                                         atomic_inc(&conf->active_stripes);
 498                                 if (list_empty(&sh->lru) &&
 499                                     !test_bit(STRIPE_EXPANDING, &sh->state))
 500                                         BUG();
 501                                 list_del_init(&sh->lru);
 502                         }
 503                 }
 504         } while (sh == NULL);
 505
 506         if (sh)
 507                 atomic_inc(&sh->count);
 508
 509         spin_unlock_irq(&conf->device_lock);
 510         return sh;
 511 }
 512
 513 /* Determine if 'data_offset' or 'new_data_offset' should be used
 514  * in this stripe_head.
 515  */
 516 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 517 {
 518         sector_t progress = conf->reshape_progress;
 519         /* Need a memory barrier to make sure we see the value
 520          * of conf->generation, or ->data_offset that was set before
 521          * reshape_progress was updated.
 522          */
 523         smp_rmb();
 524         if (progress == MaxSector)
 525                 return 0;
 526         if (sh->generation == conf->generation - 1)
 527                 return 0;
 528         /* We are in a reshape, and this is a new-generation stripe,
 529          * so use new_data_offset.
 530          */
 531         return 1;
 532 }
 533
 534 static void
 535 raid5_end_read_request(struct bio *bi, int error);
 536 static void
 537 raid5_end_write_request(struct bio *bi, int error);
 538
 539 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 540 {
 541         struct r5conf *conf = sh->raid_conf;
 542         int i, disks = sh->disks;
 543
 544         might_sleep();
 545
 546         for (i = disks; i--; ) {
 547                 int rw;
 548                 int replace_only = 0;
 549                 struct bio *bi, *rbi;
 550                 struct md_rdev *rdev, *rrdev = NULL;
 551                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 552                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 553                                 rw = WRITE_FUA;
 554                         else
 555                                 rw = WRITE;
 556                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 557                                 rw |= REQ_DISCARD;
 558                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 559                         rw = READ;
 560                 else if (test_and_clear_bit(R5_WantReplace,
 561                                             &sh->dev[i].flags)) {
 562                         rw = WRITE;
 563                         replace_only = 1;
 564                 } else
 565                         continue;
 566                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 567                         rw |= REQ_SYNC;
 568
 569                 bi = &sh->dev[i].req;
 570                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 571
 572                 bi->bi_rw = rw;
 573                 rbi->bi_rw = rw;
 574                 if (rw & WRITE) {
 575                         bi->bi_end_io = raid5_end_write_request;
 576                         rbi->bi_end_io = raid5_end_write_request;
 577                 } else
 578                         bi->bi_end_io = raid5_end_read_request;
 579
 580                 rcu_read_lock();
 581                 rrdev = rcu_dereference(conf->disks[i].replacement);
 582                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 583                 rdev = rcu_dereference(conf->disks[i].rdev);
 584                 if (!rdev) {
 585                         rdev = rrdev;
 586                         rrdev = NULL;
 587                 }
 588                 if (rw & WRITE) {
 589                         if (replace_only)
 590                                 rdev = NULL;
 591                         if (rdev == rrdev)
 592                                 /* We raced and saw duplicates */
 593                                 rrdev = NULL;
 594                 } else {
 595                         if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
 596                                 rdev = rrdev;
 597                         rrdev = NULL;
 598                 }
 599
 600                 if (rdev && test_bit(Faulty, &rdev->flags))
 601                         rdev = NULL;
 602                 if (rdev)
 603                         atomic_inc(&rdev->nr_pending);
 604                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 605                         rrdev = NULL;
 606                 if (rrdev)
 607                         atomic_inc(&rrdev->nr_pending);
 608                 rcu_read_unlock();
 609
 610                 /* We have already checked bad blocks for reads.  Now
 611                  * need to check for writes.  We never accept write errors
 612                  * on the replacement, so we don't to check rrdev.
 613                  */
 614                 while ((rw & WRITE) && rdev &&
 615                        test_bit(WriteErrorSeen, &rdev->flags)) {
 616                         sector_t first_bad;
 617                         int bad_sectors;
 618                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 619                                               &first_bad, &bad_sectors);
 620                         if (!bad)
 621                                 break;
 622
 623                         if (bad < 0) {
 624                                 set_bit(BlockedBadBlocks, &rdev->flags);
 625                                 if (!conf->mddev->external &&
 626                                     conf->mddev->flags) {
 627                                         /* It is very unlikely, but we might
 628                                          * still need to write out the
 629                                          * bad block log - better give it
 630                                          * a chance*/
 631                                         md_check_recovery(conf->mddev);
 632                                 }
 633                                 /*
 634                                  * Because md_wait_for_blocked_rdev
 635                                  * will dec nr_pending, we must
 636                                  * increment it first.
 637                                  */
 638                                 atomic_inc(&rdev->nr_pending);
 639                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 640                         } else {
 641                                 /* Acknowledged bad block - skip the write */
 642                                 rdev_dec_pending(rdev, conf->mddev);
 643                                 rdev = NULL;
 644                         }
 645                 }
 646
 647                 if (rdev) {
 648                         if (s->syncing || s->expanding || s->expanded
 649                             || s->replacing)
 650                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 651
 652                         set_bit(STRIPE_IO_STARTED, &sh->state);
 653
 654                         bi->bi_bdev = rdev->bdev;
 655                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
 656                                 __func__, (unsigned long long)sh->sector,
 657                                 bi->bi_rw, i);
 658                         atomic_inc(&sh->count);
 659                         if (use_new_offset(conf, sh))
 660                                 bi->bi_sector = (sh->sector
 661                                                  + rdev->new_data_offset);
 662                         else
 663                                 bi->bi_sector = (sh->sector
 664                                                  + rdev->data_offset);
 665                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
 666                                 bi->bi_rw |= REQ_FLUSH;
 667
 668                         bi->bi_flags = 1 << BIO_UPTODATE;
 669                         bi->bi_idx = 0;
 670                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
 671                         bi->bi_io_vec[0].bv_offset = 0;
 672                         bi->bi_size = STRIPE_SIZE;
 673                         bi->bi_next = NULL;
 674                         if (rrdev)
 675                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
 676
 677                         if (conf->mddev->gendisk)
 678                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
 679                                                       bi, disk_devt(conf->mddev->gendisk),
 680                                                       sh->dev[i].sector);
 681                         generic_make_request(bi);
 682                 }
 683                 if (rrdev) {
 684                         if (s->syncing || s->expanding || s->expanded
 685                             || s->replacing)
 686                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
 687
 688                         set_bit(STRIPE_IO_STARTED, &sh->state);
 689
 690                         rbi->bi_bdev = rrdev->bdev;
 691                         pr_debug("%s: for %llu schedule op %ld on "
 692                                  "replacement disc %d\n",
 693                                 __func__, (unsigned long long)sh->sector,
 694                                 rbi->bi_rw, i);
 695                         atomic_inc(&sh->count);
 696                         if (use_new_offset(conf, sh))
 697                                 rbi->bi_sector = (sh->sector
 698                                                   + rrdev->new_data_offset);
 699                         else
 700                                 rbi->bi_sector = (sh->sector
 701                                                   + rrdev->data_offset);
 702                         rbi->bi_flags = 1 << BIO_UPTODATE;
 703                         rbi->bi_idx = 0;
 704                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
 705                         rbi->bi_io_vec[0].bv_offset = 0;
 706                         rbi->bi_size = STRIPE_SIZE;
 707                         rbi->bi_next = NULL;
 708                         if (conf->mddev->gendisk)
 709                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
 710                                                       rbi, disk_devt(conf->mddev->gendisk),
 711                                                       sh->dev[i].sector);
 712                         generic_make_request(rbi);
 713                 }
 714                 if (!rdev && !rrdev) {
 715                         if (rw & WRITE)
 716                                 set_bit(STRIPE_DEGRADED, &sh->state);
 717                         pr_debug("skip op %ld on disc %d for sector %llu\n",
 718                                 bi->bi_rw, i, (unsigned long long)sh->sector);
 719                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 720                         set_bit(STRIPE_HANDLE, &sh->state);
 721                 }
 722         }
 723 }
 724
 725 static struct dma_async_tx_descriptor *
 726 async_copy_data(int frombio, struct bio *bio, struct page *page,
 727         sector_t sector, struct dma_async_tx_descriptor *tx)
 728 {
 729         struct bio_vec *bvl;
 730         struct page *bio_page;
 731         int i;
 732         int page_offset;
 733         struct async_submit_ctl submit;
 734         enum async_tx_flags flags = 0;
 735
 736         if (bio->bi_sector >= sector)
 737                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 738         else
 739                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 740
 741         if (frombio)
 742                 flags |= ASYNC_TX_FENCE;
 743         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
 744
 745         bio_for_each_segment(bvl, bio, i) {
 746                 int len = bvl->bv_len;
 747                 int clen;
 748                 int b_offset = 0;
 749
 750                 if (page_offset < 0) {
 751                         b_offset = -page_offset;
 752                         page_offset += b_offset;
 753                         len -= b_offset;
 754                 }
 755
 756                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 757                         clen = STRIPE_SIZE - page_offset;
 758                 else
 759                         clen = len;
 760
 761                 if (clen > 0) {
 762                         b_offset += bvl->bv_offset;
 763                         bio_page = bvl->bv_page;
 764                         if (frombio)
 765                                 tx = async_memcpy(page, bio_page, page_offset,
 766                                                   b_offset, clen, &submit);
 767                         else
 768                                 tx = async_memcpy(bio_page, page, b_offset,
 769                                                   page_offset, clen, &submit);
 770                 }
 771                 /* chain the operations */
 772                 submit.depend_tx = tx;
 773
 774                 if (clen < len) /* hit end of page */
 775                         break;
 776                 page_offset +=  len;
 777         }
 778
 779         return tx;
 780 }
 781
 782 static void ops_complete_biofill(void *stripe_head_ref)
 783 {
 784         struct stripe_head *sh = stripe_head_ref;
 785         struct bio *return_bi = NULL;
 786         int i;
 787
 788         pr_debug("%s: stripe %llu\n", __func__,
 789                 (unsigned long long)sh->sector);
 790
 791         /* clear completed biofills */
 792         for (i = sh->disks; i--; ) {
 793                 struct r5dev *dev = &sh->dev[i];
 794
 795                 /* acknowledge completion of a biofill operation */
 796                 /* and check if we need to reply to a read request,
 797                  * new R5_Wantfill requests are held off until
 798                  * !STRIPE_BIOFILL_RUN
 799                  */
 800                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
 801                         struct bio *rbi, *rbi2;
 802
 803                         BUG_ON(!dev->read);
 804                         rbi = dev->read;
 805                         dev->read = NULL;
 806                         while (rbi && rbi->bi_sector <
 807                                 dev->sector + STRIPE_SECTORS) {
 808                                 rbi2 = r5_next_bio(rbi, dev->sector);
 809                                 if (!raid5_dec_bi_active_stripes(rbi)) {
 810                                         rbi->bi_next = return_bi;
 811                                         return_bi = rbi;
 812                                 }
 813                                 rbi = rbi2;
 814                         }
 815                 }
 816         }
 817         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
 818
 819         return_io(return_bi);
 820
 821         set_bit(STRIPE_HANDLE, &sh->state);
 822         release_stripe(sh);
 823 }
 824
 825 static void ops_run_biofill(struct stripe_head *sh)
 826 {
 827         struct dma_async_tx_descriptor *tx = NULL;
 828         struct async_submit_ctl submit;
 829         int i;
 830
 831         pr_debug("%s: stripe %llu\n", __func__,
 832                 (unsigned long long)sh->sector);
 833
 834         for (i = sh->disks; i--; ) {
 835                 struct r5dev *dev = &sh->dev[i];
 836                 if (test_bit(R5_Wantfill, &dev->flags)) {
 837                         struct bio *rbi;
 838                         spin_lock_irq(&sh->stripe_lock);
 839                         dev->read = rbi = dev->toread;
 840                         dev->toread = NULL;
 841                         spin_unlock_irq(&sh->stripe_lock);
 842                         while (rbi && rbi->bi_sector <
 843                                 dev->sector + STRIPE_SECTORS) {
 844                                 tx = async_copy_data(0, rbi, dev->page,
 845                                         dev->sector, tx);
 846                                 rbi = r5_next_bio(rbi, dev->sector);
 847                         }
 848                 }
 849         }
 850
 851         atomic_inc(&sh->count);
 852         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
 853         async_trigger_callback(&submit);
 854 }
 855
 856 static void mark_target_uptodate(struct stripe_head *sh, int target)
 857 {
 858         struct r5dev *tgt;
 859
 860         if (target < 0)
 861                 return;
 862
 863         tgt = &sh->dev[target];
 864         set_bit(R5_UPTODATE, &tgt->flags);
 865         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 866         clear_bit(R5_Wantcompute, &tgt->flags);
 867 }
 868
 869 static void ops_complete_compute(void *stripe_head_ref)
 870 {
 871         struct stripe_head *sh = stripe_head_ref;
 872
 873         pr_debug("%s: stripe %llu\n", __func__,
 874                 (unsigned long long)sh->sector);
 875
 876         /* mark the computed target(s) as uptodate */
 877         mark_target_uptodate(sh, sh->ops.target);
 878         mark_target_uptodate(sh, sh->ops.target2);
 879
 880         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
 881         if (sh->check_state == check_state_compute_run)
 882                 sh->check_state = check_state_compute_result;
 883         set_bit(STRIPE_HANDLE, &sh->state);
 884         release_stripe(sh);
 885 }
 886
 887 /* return a pointer to the address conversion region of the scribble buffer */
 888 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
 889                                  struct raid5_percpu *percpu)
 890 {
 891         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
 892 }
 893
 894 static struct dma_async_tx_descriptor *
 895 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
 896 {
 897         int disks = sh->disks;
 898         struct page **xor_srcs = percpu->scribble;
 899         int target = sh->ops.target;
 900         struct r5dev *tgt = &sh->dev[target];
 901         struct page *xor_dest = tgt->page;
 902         int count = 0;
 903         struct dma_async_tx_descriptor *tx;
 904         struct async_submit_ctl submit;
 905         int i;
 906
 907         pr_debug("%s: stripe %llu block: %d\n",
 908                 __func__, (unsigned long long)sh->sector, target);
 909         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 910
 911         for (i = disks; i--; )
 912                 if (i != target)
 913                         xor_srcs[count++] = sh->dev[i].page;
 914
 915         atomic_inc(&sh->count);
 916
 917         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
 918                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
 919         if (unlikely(count == 1))
 920                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
 921         else
 922                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
 923
 924         return tx;
 925 }
 926
 927 /* set_syndrome_sources - populate source buffers for gen_syndrome
 928  * @srcs - (struct page *) array of size sh->disks
 929  * @sh - stripe_head to parse
 930  *
 931  * Populates srcs in proper layout order for the stripe and returns the
 932  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
 933  * destination buffer is recorded in srcs[count] and the Q destination
 934  * is recorded in srcs[count+1]].
 935  */
 936 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
 937 {
 938         int disks = sh->disks;
 939         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
 940         int d0_idx = raid6_d0(sh);
 941         int count;
 942         int i;
 943
 944         for (i = 0; i < disks; i++)
 945                 srcs[i] = NULL;
 946
 947         count = 0;
 948         i = d0_idx;
 949         do {
 950                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
 951
 952                 srcs[slot] = sh->dev[i].page;
 953                 i = raid6_next_disk(i, disks);
 954         } while (i != d0_idx);
 955
 956         return syndrome_disks;
 957 }
 958
 959 static struct dma_async_tx_descriptor *
 960 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
 961 {
 962         int disks = sh->disks;
 963         struct page **blocks = percpu->scribble;
 964         int target;
 965         int qd_idx = sh->qd_idx;
 966         struct dma_async_tx_descriptor *tx;
 967         struct async_submit_ctl submit;
 968         struct r5dev *tgt;
 969         struct page *dest;
 970         int i;
 971         int count;
 972
 973         if (sh->ops.target < 0)
 974                 target = sh->ops.target2;
 975         else if (sh->ops.target2 < 0)
 976                 target = sh->ops.target;
 977         else
 978                 /* we should only have one valid target */
 979                 BUG();
 980         BUG_ON(target < 0);
 981         pr_debug("%s: stripe %llu block: %d\n",
 982                 __func__, (unsigned long long)sh->sector, target);
 983
 984         tgt = &sh->dev[target];
 985         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 986         dest = tgt->page;
 987
 988         atomic_inc(&sh->count);
 989
 990         if (target == qd_idx) {
 991                 count = set_syndrome_sources(blocks, sh);
 992                 blocks[count] = NULL; /* regenerating p is not necessary */
 993                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
 994                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
 995                                   ops_complete_compute, sh,
 996                                   to_addr_conv(sh, percpu));
 997                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
 998         } else {
 999                 /* Compute any data- or p-drive using XOR */
1000                 count = 0;
1001                 for (i = disks; i-- ; ) {
1002                         if (i == target || i == qd_idx)
1003                                 continue;
1004                         blocks[count++] = sh->dev[i].page;
1005                 }
1006
1007                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1008                                   NULL, ops_complete_compute, sh,
1009                                   to_addr_conv(sh, percpu));
1010                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1011         }
1012
1013         return tx;
1014 }
1015
1016 static struct dma_async_tx_descriptor *
1017 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1018 {
1019         int i, count, disks = sh->disks;
1020         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1021         int d0_idx = raid6_d0(sh);
1022         int faila = -1, failb = -1;
1023         int target = sh->ops.target;
1024         int target2 = sh->ops.target2;
1025         struct r5dev *tgt = &sh->dev[target];
1026         struct r5dev *tgt2 = &sh->dev[target2];
1027         struct dma_async_tx_descriptor *tx;
1028         struct page **blocks = percpu->scribble;
1029         struct async_submit_ctl submit;
1030
1031         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1032                  __func__, (unsigned long long)sh->sector, target, target2);
1033         BUG_ON(target < 0 || target2 < 0);
1034         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1035         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1036
1037         /* we need to open-code set_syndrome_sources to handle the
1038          * slot number conversion for 'faila' and 'failb'
1039          */
1040         for (i = 0; i < disks ; i++)
1041                 blocks[i] = NULL;
1042         count = 0;
1043         i = d0_idx;
1044         do {
1045                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1046
1047                 blocks[slot] = sh->dev[i].page;
1048
1049                 if (i == target)
1050                         faila = slot;
1051                 if (i == target2)
1052                         failb = slot;
1053                 i = raid6_next_disk(i, disks);
1054         } while (i != d0_idx);
1055
1056         BUG_ON(faila == failb);
1057         if (failb < faila)
1058                 swap(faila, failb);
1059         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1060                  __func__, (unsigned long long)sh->sector, faila, failb);
1061
1062         atomic_inc(&sh->count);
1063
1064         if (failb == syndrome_disks+1) {
1065                 /* Q disk is one of the missing disks */
1066                 if (faila == syndrome_disks) {
1067                         /* Missing P+Q, just recompute */
1068                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1069                                           ops_complete_compute, sh,
1070                                           to_addr_conv(sh, percpu));
1071                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1072                                                   STRIPE_SIZE, &submit);
1073                 } else {
1074                         struct page *dest;
1075                         int data_target;
1076                         int qd_idx = sh->qd_idx;
1077
1078                         /* Missing D+Q: recompute D from P, then recompute Q */
1079                         if (target == qd_idx)
1080                                 data_target = target2;
1081                         else
1082                                 data_target = target;
1083
1084                         count = 0;
1085                         for (i = disks; i-- ; ) {
1086                                 if (i == data_target || i == qd_idx)
1087                                         continue;
1088                                 blocks[count++] = sh->dev[i].page;
1089                         }
1090                         dest = sh->dev[data_target].page;
1091                         init_async_submit(&submit,
1092                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1093                                           NULL, NULL, NULL,
1094                                           to_addr_conv(sh, percpu));
1095                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1096                                        &submit);
1097
1098                         count = set_syndrome_sources(blocks, sh);
1099                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1100                                           ops_complete_compute, sh,
1101                                           to_addr_conv(sh, percpu));
1102                         return async_gen_syndrome(blocks, 0, count+2,
1103                                                   STRIPE_SIZE, &submit);
1104                 }
1105         } else {
1106                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1107                                   ops_complete_compute, sh,
1108                                   to_addr_conv(sh, percpu));
1109                 if (failb == syndrome_disks) {
1110                         /* We're missing D+P. */
1111                         return async_raid6_datap_recov(syndrome_disks+2,
1112                                                        STRIPE_SIZE, faila,
1113                                                        blocks, &submit);
1114                 } else {
1115                         /* We're missing D+D. */
1116                         return async_raid6_2data_recov(syndrome_disks+2,
1117                                                        STRIPE_SIZE, faila, failb,
1118                                                        blocks, &submit);
1119                 }
1120         }
1121 }
1122
1123
1124 static void ops_complete_prexor(void *stripe_head_ref)
1125 {
1126         struct stripe_head *sh = stripe_head_ref;
1127
1128         pr_debug("%s: stripe %llu\n", __func__,
1129                 (unsigned long long)sh->sector);
1130 }
1131
1132 static struct dma_async_tx_descriptor *
1133 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1134                struct dma_async_tx_descriptor *tx)
1135 {
1136         int disks = sh->disks;
1137         struct page **xor_srcs = percpu->scribble;
1138         int count = 0, pd_idx = sh->pd_idx, i;
1139         struct async_submit_ctl submit;
1140
1141         /* existing parity data subtracted */
1142         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1143
1144         pr_debug("%s: stripe %llu\n", __func__,
1145                 (unsigned long long)sh->sector);
1146
1147         for (i = disks; i--; ) {
1148                 struct r5dev *dev = &sh->dev[i];
1149                 /* Only process blocks that are known to be uptodate */
1150                 if (test_bit(R5_Wantdrain, &dev->flags))
1151                         xor_srcs[count++] = dev->page;
1152         }
1153
1154         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1155                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1156         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1157
1158         return tx;
1159 }
1160
1161 static struct dma_async_tx_descriptor *
1162 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1163 {
1164         int disks = sh->disks;
1165         int i;
1166
1167         pr_debug("%s: stripe %llu\n", __func__,
1168                 (unsigned long long)sh->sector);
1169
1170         for (i = disks; i--; ) {
1171                 struct r5dev *dev = &sh->dev[i];
1172                 struct bio *chosen;
1173
1174                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1175                         struct bio *wbi;
1176
1177                         spin_lock_irq(&sh->stripe_lock);
1178                         chosen = dev->towrite;
1179                         dev->towrite = NULL;
1180                         BUG_ON(dev->written);
1181                         wbi = dev->written = chosen;
1182                         spin_unlock_irq(&sh->stripe_lock);
1183
1184                         while (wbi && wbi->bi_sector <
1185                                 dev->sector + STRIPE_SECTORS) {
1186                                 if (wbi->bi_rw & REQ_FUA)
1187                                         set_bit(R5_WantFUA, &dev->flags);
1188                                 if (wbi->bi_rw & REQ_SYNC)
1189                                         set_bit(R5_SyncIO, &dev->flags);
1190                                 if (wbi->bi_rw & REQ_DISCARD)
1191                                         set_bit(R5_Discard, &dev->flags);
1192                                 else
1193                                         tx = async_copy_data(1, wbi, dev->page,
1194                                                 dev->sector, tx);
1195                                 wbi = r5_next_bio(wbi, dev->sector);
1196                         }
1197                 }
1198         }
1199
1200         return tx;
1201 }
1202
1203 static void ops_complete_reconstruct(void *stripe_head_ref)
1204 {
1205         struct stripe_head *sh = stripe_head_ref;
1206         int disks = sh->disks;
1207         int pd_idx = sh->pd_idx;
1208         int qd_idx = sh->qd_idx;
1209         int i;
1210         bool fua = false, sync = false, discard = false;
1211
1212         pr_debug("%s: stripe %llu\n", __func__,
1213                 (unsigned long long)sh->sector);
1214
1215         for (i = disks; i--; ) {
1216                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1217                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1218                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1219         }
1220
1221         for (i = disks; i--; ) {
1222                 struct r5dev *dev = &sh->dev[i];
1223
1224                 if (dev->written || i == pd_idx || i == qd_idx) {
1225                         if (!discard)
1226                                 set_bit(R5_UPTODATE, &dev->flags);
1227                         if (fua)
1228                                 set_bit(R5_WantFUA, &dev->flags);
1229                         if (sync)
1230                                 set_bit(R5_SyncIO, &dev->flags);
1231                 }
1232         }
1233
1234         if (sh->reconstruct_state == reconstruct_state_drain_run)
1235                 sh->reconstruct_state = reconstruct_state_drain_result;
1236         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1237                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1238         else {
1239                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1240                 sh->reconstruct_state = reconstruct_state_result;
1241         }
1242
1243         set_bit(STRIPE_HANDLE, &sh->state);
1244         release_stripe(sh);
1245 }
1246
1247 static void
1248 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1249                      struct dma_async_tx_descriptor *tx)
1250 {
1251         int disks = sh->disks;
1252         struct page **xor_srcs = percpu->scribble;
1253         struct async_submit_ctl submit;
1254         int count = 0, pd_idx = sh->pd_idx, i;
1255         struct page *xor_dest;
1256         int prexor = 0;
1257         unsigned long flags;
1258
1259         pr_debug("%s: stripe %llu\n", __func__,
1260                 (unsigned long long)sh->sector);
1261
1262         for (i = 0; i < sh->disks; i++) {
1263                 if (pd_idx == i)
1264                         continue;
1265                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1266                         break;
1267         }
1268         if (i >= sh->disks) {
1269                 atomic_inc(&sh->count);
1270                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1271                 ops_complete_reconstruct(sh);
1272                 return;
1273         }
1274         /* check if prexor is active which means only process blocks
1275          * that are part of a read-modify-write (written)
1276          */
1277         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1278                 prexor = 1;
1279                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1280                 for (i = disks; i--; ) {
1281                         struct r5dev *dev = &sh->dev[i];
1282                         if (dev->written)
1283                                 xor_srcs[count++] = dev->page;
1284                 }
1285         } else {
1286                 xor_dest = sh->dev[pd_idx].page;
1287                 for (i = disks; i--; ) {
1288                         struct r5dev *dev = &sh->dev[i];
1289                         if (i != pd_idx)
1290                                 xor_srcs[count++] = dev->page;
1291                 }
1292         }
1293
1294         /* 1/ if we prexor'd then the dest is reused as a source
1295          * 2/ if we did not prexor then we are redoing the parity
1296          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1297          * for the synchronous xor case
1298          */
1299         flags = ASYNC_TX_ACK |
1300                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1301
1302         atomic_inc(&sh->count);
1303
1304         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1305                           to_addr_conv(sh, percpu));
1306         if (unlikely(count == 1))
1307                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1308         else
1309                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1310 }
1311
1312 static void
1313 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1314                      struct dma_async_tx_descriptor *tx)
1315 {
1316         struct async_submit_ctl submit;
1317         struct page **blocks = percpu->scribble;
1318         int count, i;
1319
1320         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1321
1322         for (i = 0; i < sh->disks; i++) {
1323                 if (sh->pd_idx == i || sh->qd_idx == i)
1324                         continue;
1325                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1326                         break;
1327         }
1328         if (i >= sh->disks) {
1329                 atomic_inc(&sh->count);
1330                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1331                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1332                 ops_complete_reconstruct(sh);
1333                 return;
1334         }
1335
1336         count = set_syndrome_sources(blocks, sh);
1337
1338         atomic_inc(&sh->count);
1339
1340         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1341                           sh, to_addr_conv(sh, percpu));
1342         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1343 }
1344
1345 static void ops_complete_check(void *stripe_head_ref)
1346 {
1347         struct stripe_head *sh = stripe_head_ref;
1348
1349         pr_debug("%s: stripe %llu\n", __func__,
1350                 (unsigned long long)sh->sector);
1351
1352         sh->check_state = check_state_check_result;
1353         set_bit(STRIPE_HANDLE, &sh->state);
1354         release_stripe(sh);
1355 }
1356
1357 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1358 {
1359         int disks = sh->disks;
1360         int pd_idx = sh->pd_idx;
1361         int qd_idx = sh->qd_idx;
1362         struct page *xor_dest;
1363         struct page **xor_srcs = percpu->scribble;
1364         struct dma_async_tx_descriptor *tx;
1365         struct async_submit_ctl submit;
1366         int count;
1367         int i;
1368
1369         pr_debug("%s: stripe %llu\n", __func__,
1370                 (unsigned long long)sh->sector);
1371
1372         count = 0;
1373         xor_dest = sh->dev[pd_idx].page;
1374         xor_srcs[count++] = xor_dest;
1375         for (i = disks; i--; ) {
1376                 if (i == pd_idx || i == qd_idx)
1377                         continue;
1378                 xor_srcs[count++] = sh->dev[i].page;
1379         }
1380
1381         init_async_submit(&submit, 0, NULL, NULL, NULL,
1382                           to_addr_conv(sh, percpu));
1383         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1384                            &sh->ops.zero_sum_result, &submit);
1385
1386         atomic_inc(&sh->count);
1387         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1388         tx = async_trigger_callback(&submit);
1389 }
1390
1391 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1392 {
1393         struct page **srcs = percpu->scribble;
1394         struct async_submit_ctl submit;
1395         int count;
1396
1397         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1398                 (unsigned long long)sh->sector, checkp);
1399
1400         count = set_syndrome_sources(srcs, sh);
1401         if (!checkp)
1402                 srcs[count] = NULL;
1403
1404         atomic_inc(&sh->count);
1405         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1406                           sh, to_addr_conv(sh, percpu));
1407         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1408                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1409 }
1410
1411 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1412 {
1413         int overlap_clear = 0, i, disks = sh->disks;
1414         struct dma_async_tx_descriptor *tx = NULL;
1415         struct r5conf *conf = sh->raid_conf;
1416         int level = conf->level;
1417         struct raid5_percpu *percpu;
1418         unsigned long cpu;
1419
1420         cpu = get_cpu();
1421         percpu = per_cpu_ptr(conf->percpu, cpu);
1422         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1423                 ops_run_biofill(sh);
1424                 overlap_clear++;
1425         }
1426
1427         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1428                 if (level < 6)
1429                         tx = ops_run_compute5(sh, percpu);
1430                 else {
1431                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1432                                 tx = ops_run_compute6_1(sh, percpu);
1433                         else
1434                                 tx = ops_run_compute6_2(sh, percpu);
1435                 }
1436                 /* terminate the chain if reconstruct is not set to be run */
1437                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1438                         async_tx_ack(tx);
1439         }
1440
1441         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1442                 tx = ops_run_prexor(sh, percpu, tx);
1443
1444         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1445                 tx = ops_run_biodrain(sh, tx);
1446                 overlap_clear++;
1447         }
1448
1449         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1450                 if (level < 6)
1451                         ops_run_reconstruct5(sh, percpu, tx);
1452                 else
1453                         ops_run_reconstruct6(sh, percpu, tx);
1454         }
1455
1456         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1457                 if (sh->check_state == check_state_run)
1458                         ops_run_check_p(sh, percpu);
1459                 else if (sh->check_state == check_state_run_q)
1460                         ops_run_check_pq(sh, percpu, 0);
1461                 else if (sh->check_state == check_state_run_pq)
1462                         ops_run_check_pq(sh, percpu, 1);
1463                 else
1464                         BUG();
1465         }
1466
1467         if (overlap_clear)
1468                 for (i = disks; i--; ) {
1469                         struct r5dev *dev = &sh->dev[i];
1470                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1471                                 wake_up(&sh->raid_conf->wait_for_overlap);
1472                 }
1473         put_cpu();
1474 }
1475
1476 static int grow_one_stripe(struct r5conf *conf)
1477 {
1478         struct stripe_head *sh;
1479         sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1480         if (!sh)
1481                 return 0;
1482
1483         sh->raid_conf = conf;
1484
1485         spin_lock_init(&sh->stripe_lock);
1486
1487         if (grow_buffers(sh)) {
1488                 shrink_buffers(sh);
1489                 kmem_cache_free(conf->slab_cache, sh);
1490                 return 0;
1491         }
1492         /* we just created an active stripe so... */
1493         atomic_set(&sh->count, 1);
1494         atomic_inc(&conf->active_stripes);
1495         INIT_LIST_HEAD(&sh->lru);
1496         release_stripe(sh);
1497         return 1;
1498 }
1499
1500 static int grow_stripes(struct r5conf *conf, int num)
1501 {
1502         struct kmem_cache *sc;
1503         int devs = max(conf->raid_disks, conf->previous_raid_disks);
1504
1505         if (conf->mddev->gendisk)
1506                 sprintf(conf->cache_name[0],
1507                         "raid%d-%s", conf->level, mdname(conf->mddev));
1508         else
1509                 sprintf(conf->cache_name[0],
1510                         "raid%d-%p", conf->level, conf->mddev);
1511         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1512
1513         conf->active_name = 0;
1514         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1515                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1516                                0, 0, NULL);
1517         if (!sc)
1518                 return 1;
1519         conf->slab_cache = sc;
1520         conf->pool_size = devs;
1521         while (num--)
1522                 if (!grow_one_stripe(conf))
1523                         return 1;
1524         return 0;
1525 }
1526
1527 /**
1528  * scribble_len - return the required size of the scribble region
1529  * @num - total number of disks in the array
1530  *
1531  * The size must be enough to contain:
1532  * 1/ a struct page pointer for each device in the array +2
1533  * 2/ room to convert each entry in (1) to its corresponding dma
1534  *    (dma_map_page()) or page (page_address()) address.
1535  *
1536  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1537  * calculate over all devices (not just the data blocks), using zeros in place
1538  * of the P and Q blocks.
1539  */
1540 static size_t scribble_len(int num)
1541 {
1542         size_t len;
1543
1544         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1545
1546         return len;
1547 }
1548
1549 static int resize_stripes(struct r5conf *conf, int newsize)
1550 {
1551         /* Make all the stripes able to hold 'newsize' devices.
1552          * New slots in each stripe get 'page' set to a new page.
1553          *
1554          * This happens in stages:
1555          * 1/ create a new kmem_cache and allocate the required number of
1556          *    stripe_heads.
1557          * 2/ gather all the old stripe_heads and transfer the pages across
1558          *    to the new stripe_heads.  This will have the side effect of
1559          *    freezing the array as once all stripe_heads have been collected,
1560          *    no IO will be possible.  Old stripe heads are freed once their
1561          *    pages have been transferred over, and the old kmem_cache is
1562          *    freed when all stripes are done.
1563          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1564          *    we simple return a failre status - no need to clean anything up.
1565          * 4/ allocate new pages for the new slots in the new stripe_heads.
1566          *    If this fails, we don't bother trying the shrink the
1567          *    stripe_heads down again, we just leave them as they are.
1568          *    As each stripe_head is processed the new one is released into
1569          *    active service.
1570          *
1571          * Once step2 is started, we cannot afford to wait for a write,
1572          * so we use GFP_NOIO allocations.
1573          */
1574         struct stripe_head *osh, *nsh;
1575         LIST_HEAD(newstripes);
1576         struct disk_info *ndisks;
1577         unsigned long cpu;
1578         int err;
1579         struct kmem_cache *sc;
1580         int i;
1581
1582         if (newsize <= conf->pool_size)
1583                 return 0; /* never bother to shrink */
1584
1585         err = md_allow_write(conf->mddev);
1586         if (err)
1587                 return err;
1588
1589         /* Step 1 */
1590         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1591                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1592                                0, 0, NULL);
1593         if (!sc)
1594                 return -ENOMEM;
1595
1596         for (i = conf->max_nr_stripes; i; i--) {
1597                 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1598                 if (!nsh)
1599                         break;
1600
1601                 nsh->raid_conf = conf;
1602                 spin_lock_init(&nsh->stripe_lock);
1603
1604                 list_add(&nsh->lru, &newstripes);
1605         }
1606         if (i) {
1607                 /* didn't get enough, give up */
1608                 while (!list_empty(&newstripes)) {
1609                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1610                         list_del(&nsh->lru);
1611                         kmem_cache_free(sc, nsh);
1612                 }
1613                 kmem_cache_destroy(sc);
1614                 return -ENOMEM;
1615         }
1616         /* Step 2 - Must use GFP_NOIO now.
1617          * OK, we have enough stripes, start collecting inactive
1618          * stripes and copying them over
1619          */
1620         list_for_each_entry(nsh, &newstripes, lru) {
1621                 spin_lock_irq(&conf->device_lock);
1622                 wait_event_lock_irq(conf->wait_for_stripe,
1623                                     !list_empty(&conf->inactive_list),
1624                                     conf->device_lock);
1625                 osh = get_free_stripe(conf);
1626                 spin_unlock_irq(&conf->device_lock);
1627                 atomic_set(&nsh->count, 1);
1628                 for(i=0; i<conf->pool_size; i++)
1629                         nsh->dev[i].page = osh->dev[i].page;
1630                 for( ; i<newsize; i++)
1631                         nsh->dev[i].page = NULL;
1632                 kmem_cache_free(conf->slab_cache, osh);
1633         }
1634         kmem_cache_destroy(conf->slab_cache);
1635
1636         /* Step 3.
1637          * At this point, we are holding all the stripes so the array
1638          * is completely stalled, so now is a good time to resize
1639          * conf->disks and the scribble region
1640          */
1641         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1642         if (ndisks) {
1643                 for (i=0; i<conf->raid_disks; i++)
1644                         ndisks[i] = conf->disks[i];
1645                 kfree(conf->disks);
1646                 conf->disks = ndisks;
1647         } else
1648                 err = -ENOMEM;
1649
1650         get_online_cpus();
1651         conf->scribble_len = scribble_len(newsize);
1652         for_each_present_cpu(cpu) {
1653                 struct raid5_percpu *percpu;
1654                 void *scribble;
1655
1656                 percpu = per_cpu_ptr(conf->percpu, cpu);
1657                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1658
1659                 if (scribble) {
1660                         kfree(percpu->scribble);
1661                         percpu->scribble = scribble;
1662                 } else {
1663                         err = -ENOMEM;
1664                         break;
1665                 }
1666         }
1667         put_online_cpus();
1668
1669         /* Step 4, return new stripes to service */
1670         while(!list_empty(&newstripes)) {
1671                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1672                 list_del_init(&nsh->lru);
1673
1674                 for (i=conf->raid_disks; i < newsize; i++)
1675                         if (nsh->dev[i].page == NULL) {
1676                                 struct page *p = alloc_page(GFP_NOIO);
1677                                 nsh->dev[i].page = p;
1678                                 if (!p)
1679                                         err = -ENOMEM;
1680                         }
1681                 release_stripe(nsh);
1682         }
1683         /* critical section pass, GFP_NOIO no longer needed */
1684
1685         conf->slab_cache = sc;
1686         conf->active_name = 1-conf->active_name;
1687         conf->pool_size = newsize;
1688         return err;
1689 }
1690
1691 static int drop_one_stripe(struct r5conf *conf)
1692 {
1693         struct stripe_head *sh;
1694
1695         spin_lock_irq(&conf->device_lock);
1696         sh = get_free_stripe(conf);
1697         spin_unlock_irq(&conf->device_lock);
1698         if (!sh)
1699                 return 0;
1700         BUG_ON(atomic_read(&sh->count));
1701         shrink_buffers(sh);
1702         kmem_cache_free(conf->slab_cache, sh);
1703         atomic_dec(&conf->active_stripes);
1704         return 1;
1705 }
1706
1707 static void shrink_stripes(struct r5conf *conf)
1708 {
1709         while (drop_one_stripe(conf))
1710                 ;
1711
1712         if (conf->slab_cache)
1713                 kmem_cache_destroy(conf->slab_cache);
1714         conf->slab_cache = NULL;
1715 }
1716
1717 static void raid5_end_read_request(struct bio * bi, int error)
1718 {
1719         struct stripe_head *sh = bi->bi_private;
1720         struct r5conf *conf = sh->raid_conf;
1721         int disks = sh->disks, i;
1722         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1723         char b[BDEVNAME_SIZE];
1724         struct md_rdev *rdev = NULL;
1725         sector_t s;
1726
1727         for (i=0 ; i<disks; i++)
1728                 if (bi == &sh->dev[i].req)
1729                         break;
1730
1731         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1732                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1733                 uptodate);
1734         if (i == disks) {
1735                 BUG();
1736                 return;
1737         }
1738         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1739                 /* If replacement finished while this request was outstanding,
1740                  * 'replacement' might be NULL already.
1741                  * In that case it moved down to 'rdev'.
1742                  * rdev is not removed until all requests are finished.
1743                  */
1744                 rdev = conf->disks[i].replacement;
1745         if (!rdev)
1746                 rdev = conf->disks[i].rdev;
1747
1748         if (use_new_offset(conf, sh))
1749                 s = sh->sector + rdev->new_data_offset;
1750         else
1751                 s = sh->sector + rdev->data_offset;
1752         if (uptodate) {
1753                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1754                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1755                         /* Note that this cannot happen on a
1756                          * replacement device.  We just fail those on
1757                          * any error
1758                          */
1759                         printk_ratelimited(
1760                                 KERN_INFO
1761                                 "md/raid:%s: read error corrected"
1762                                 " (%lu sectors at %llu on %s)\n",
1763                                 mdname(conf->mddev), STRIPE_SECTORS,
1764                                 (unsigned long long)s,
1765                                 bdevname(rdev->bdev, b));
1766                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1767                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1768                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1769                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1770                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1771
1772                 if (atomic_read(&rdev->read_errors))
1773                         atomic_set(&rdev->read_errors, 0);
1774         } else {
1775                 const char *bdn = bdevname(rdev->bdev, b);
1776                 int retry = 0;
1777                 int set_bad = 0;
1778
1779                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1780                 atomic_inc(&rdev->read_errors);
1781                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1782                         printk_ratelimited(
1783                                 KERN_WARNING
1784                                 "md/raid:%s: read error on replacement device "
1785                                 "(sector %llu on %s).\n",
1786                                 mdname(conf->mddev),
1787                                 (unsigned long long)s,
1788                                 bdn);
1789                 else if (conf->mddev->degraded >= conf->max_degraded) {
1790                         set_bad = 1;
1791                         printk_ratelimited(
1792                                 KERN_WARNING
1793                                 "md/raid:%s: read error not correctable "
1794                                 "(sector %llu on %s).\n",
1795                                 mdname(conf->mddev),
1796                                 (unsigned long long)s,
1797                                 bdn);
1798                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1799                         /* Oh, no!!! */
1800                         set_bad = 1;
1801                         printk_ratelimited(
1802                                 KERN_WARNING
1803                                 "md/raid:%s: read error NOT corrected!! "
1804                                 "(sector %llu on %s).\n",
1805                                 mdname(conf->mddev),
1806                                 (unsigned long long)s,
1807                                 bdn);
1808                 } else if (atomic_read(&rdev->read_errors)
1809                          > conf->max_nr_stripes)
1810                         printk(KERN_WARNING
1811                                "md/raid:%s: Too many read errors, failing device %s.\n",
1812                                mdname(conf->mddev), bdn);
1813                 else
1814                         retry = 1;
1815                 if (retry)
1816                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1817                                 set_bit(R5_ReadError, &sh->dev[i].flags);
1818                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1819                         } else
1820                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1821                 else {
1822                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1823                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1824                         if (!(set_bad
1825                               && test_bit(In_sync, &rdev->flags)
1826                               && rdev_set_badblocks(
1827                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
1828                                 md_error(conf->mddev, rdev);
1829                 }
1830         }
1831         rdev_dec_pending(rdev, conf->mddev);
1832         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1833         set_bit(STRIPE_HANDLE, &sh->state);
1834         release_stripe(sh);
1835 }
1836
1837 static void raid5_end_write_request(struct bio *bi, int error)
1838 {
1839         struct stripe_head *sh = bi->bi_private;
1840         struct r5conf *conf = sh->raid_conf;
1841         int disks = sh->disks, i;
1842         struct md_rdev *uninitialized_var(rdev);
1843         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1844         sector_t first_bad;
1845         int bad_sectors;
1846         int replacement = 0;
1847
1848         for (i = 0 ; i < disks; i++) {
1849                 if (bi == &sh->dev[i].req) {
1850                         rdev = conf->disks[i].rdev;
1851                         break;
1852                 }
1853                 if (bi == &sh->dev[i].rreq) {
1854                         rdev = conf->disks[i].replacement;
1855                         if (rdev)
1856                                 replacement = 1;
1857                         else
1858                                 /* rdev was removed and 'replacement'
1859                                  * replaced it.  rdev is not removed
1860                                  * until all requests are finished.
1861                                  */
1862                                 rdev = conf->disks[i].rdev;
1863                         break;
1864                 }
1865         }
1866         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1867                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1868                 uptodate);
1869         if (i == disks) {
1870                 BUG();
1871                 return;
1872         }
1873
1874         if (replacement) {
1875                 if (!uptodate)
1876                         md_error(conf->mddev, rdev);
1877                 else if (is_badblock(rdev, sh->sector,
1878                                      STRIPE_SECTORS,
1879                                      &first_bad, &bad_sectors))
1880                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1881         } else {
1882                 if (!uptodate) {
1883                         set_bit(WriteErrorSeen, &rdev->flags);
1884                         set_bit(R5_WriteError, &sh->dev[i].flags);
1885                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1886                                 set_bit(MD_RECOVERY_NEEDED,
1887                                         &rdev->mddev->recovery);
1888                 } else if (is_badblock(rdev, sh->sector,
1889                                        STRIPE_SECTORS,
1890                                        &first_bad, &bad_sectors)) {
1891                         set_bit(R5_MadeGood, &sh->dev[i].flags);
1892                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
1893                                 /* That was a successful write so make
1894                                  * sure it looks like we already did
1895                                  * a re-write.
1896                                  */
1897                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
1898                 }
1899         }
1900         rdev_dec_pending(rdev, conf->mddev);
1901
1902         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1903                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1904         set_bit(STRIPE_HANDLE, &sh->state);
1905         release_stripe(sh);
1906 }
1907
1908 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1909
1910 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1911 {
1912         struct r5dev *dev = &sh->dev[i];
1913
1914         bio_init(&dev->req);
1915         dev->req.bi_io_vec = &dev->vec;
1916         dev->req.bi_vcnt++;
1917         dev->req.bi_max_vecs++;
1918         dev->req.bi_private = sh;
1919         dev->vec.bv_page = dev->page;
1920
1921         bio_init(&dev->rreq);
1922         dev->rreq.bi_io_vec = &dev->rvec;
1923         dev->rreq.bi_vcnt++;
1924         dev->rreq.bi_max_vecs++;
1925         dev->rreq.bi_private = sh;
1926         dev->rvec.bv_page = dev->page;
1927
1928         dev->flags = 0;
1929         dev->sector = compute_blocknr(sh, i, previous);
1930 }
1931
1932 static void error(struct mddev *mddev, struct md_rdev *rdev)
1933 {
1934         char b[BDEVNAME_SIZE];
1935         struct r5conf *conf = mddev->private;
1936         unsigned long flags;
1937         pr_debug("raid456: error called\n");
1938
1939         spin_lock_irqsave(&conf->device_lock, flags);
1940         clear_bit(In_sync, &rdev->flags);
1941         mddev->degraded = calc_degraded(conf);
1942         spin_unlock_irqrestore(&conf->device_lock, flags);
1943         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1944
1945         set_bit(Blocked, &rdev->flags);
1946         set_bit(Faulty, &rdev->flags);
1947         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1948         printk(KERN_ALERT
1949                "md/raid:%s: Disk failure on %s, disabling device.\n"
1950                "md/raid:%s: Operation continuing on %d devices.\n",
1951                mdname(mddev),
1952                bdevname(rdev->bdev, b),
1953                mdname(mddev),
1954                conf->raid_disks - mddev->degraded);
1955 }
1956
1957 /*
1958  * Input: a 'big' sector number,
1959  * Output: index of the data and parity disk, and the sector # in them.
1960  */
1961 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1962                                      int previous, int *dd_idx,
1963                                      struct stripe_head *sh)
1964 {
1965         sector_t stripe, stripe2;
1966         sector_t chunk_number;
1967         unsigned int chunk_offset;
1968         int pd_idx, qd_idx;
1969         int ddf_layout = 0;
1970         sector_t new_sector;
1971         int algorithm = previous ? conf->prev_algo
1972                                  : conf->algorithm;
1973         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1974                                          : conf->chunk_sectors;
1975         int raid_disks = previous ? conf->previous_raid_disks
1976                                   : conf->raid_disks;
1977         int data_disks = raid_disks - conf->max_degraded;
1978
1979         /* First compute the information on this sector */
1980
1981         /*
1982          * Compute the chunk number and the sector offset inside the chunk
1983          */
1984         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1985         chunk_number = r_sector;
1986
1987         /*
1988          * Compute the stripe number
1989          */
1990         stripe = chunk_number;
1991         *dd_idx = sector_div(stripe, data_disks);
1992         stripe2 = stripe;
1993         /*
1994          * Select the parity disk based on the user selected algorithm.
1995          */
1996         pd_idx = qd_idx = -1;
1997         switch(conf->level) {
1998         case 4:
1999                 pd_idx = data_disks;
2000                 break;
2001         case 5:
2002                 switch (algorithm) {
2003                 case ALGORITHM_LEFT_ASYMMETRIC:
2004                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2005                         if (*dd_idx >= pd_idx)
2006                                 (*dd_idx)++;
2007                         break;
2008                 case ALGORITHM_RIGHT_ASYMMETRIC:
2009                         pd_idx = sector_div(stripe2, raid_disks);
2010                         if (*dd_idx >= pd_idx)
2011                                 (*dd_idx)++;
2012                         break;
2013                 case ALGORITHM_LEFT_SYMMETRIC:
2014                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2015                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2016                         break;
2017                 case ALGORITHM_RIGHT_SYMMETRIC:
2018                         pd_idx = sector_div(stripe2, raid_disks);
2019                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2020                         break;
2021                 case ALGORITHM_PARITY_0:
2022                         pd_idx = 0;
2023                         (*dd_idx)++;
2024                         break;
2025                 case ALGORITHM_PARITY_N:
2026                         pd_idx = data_disks;
2027                         break;
2028                 default:
2029                         BUG();
2030                 }
2031                 break;
2032         case 6:
2033
2034                 switch (algorithm) {
2035                 case ALGORITHM_LEFT_ASYMMETRIC:
2036                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2037                         qd_idx = pd_idx + 1;
2038                         if (pd_idx == raid_disks-1) {
2039                                 (*dd_idx)++;    /* Q D D D P */
2040                                 qd_idx = 0;
2041                         } else if (*dd_idx >= pd_idx)
2042                                 (*dd_idx) += 2; /* D D P Q D */
2043                         break;
2044                 case ALGORITHM_RIGHT_ASYMMETRIC:
2045                         pd_idx = sector_div(stripe2, raid_disks);
2046                         qd_idx = pd_idx + 1;
2047                         if (pd_idx == raid_disks-1) {
2048                                 (*dd_idx)++;    /* Q D D D P */
2049                                 qd_idx = 0;
2050                         } else if (*dd_idx >= pd_idx)
2051                                 (*dd_idx) += 2; /* D D P Q D */
2052                         break;
2053                 case ALGORITHM_LEFT_SYMMETRIC:
2054                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2055                         qd_idx = (pd_idx + 1) % raid_disks;
2056                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2057                         break;
2058                 case ALGORITHM_RIGHT_SYMMETRIC:
2059                         pd_idx = sector_div(stripe2, raid_disks);
2060                         qd_idx = (pd_idx + 1) % raid_disks;
2061                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2062                         break;
2063
2064                 case ALGORITHM_PARITY_0:
2065                         pd_idx = 0;
2066                         qd_idx = 1;
2067                         (*dd_idx) += 2;
2068                         break;
2069                 case ALGORITHM_PARITY_N:
2070                         pd_idx = data_disks;
2071                         qd_idx = data_disks + 1;
2072                         break;
2073
2074                 case ALGORITHM_ROTATING_ZERO_RESTART:
2075                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2076                          * of blocks for computing Q is different.
2077                          */
2078                         pd_idx = sector_div(stripe2, raid_disks);
2079                         qd_idx = pd_idx + 1;
2080                         if (pd_idx == raid_disks-1) {
2081                                 (*dd_idx)++;    /* Q D D D P */
2082                                 qd_idx = 0;
2083                         } else if (*dd_idx >= pd_idx)
2084                                 (*dd_idx) += 2; /* D D P Q D */
2085                         ddf_layout = 1;
2086                         break;
2087
2088                 case ALGORITHM_ROTATING_N_RESTART:
2089                         /* Same a left_asymmetric, by first stripe is
2090                          * D D D P Q  rather than
2091                          * Q D D D P
2092                          */
2093                         stripe2 += 1;
2094                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2095                         qd_idx = pd_idx + 1;
2096                         if (pd_idx == raid_disks-1) {
2097                                 (*dd_idx)++;    /* Q D D D P */
2098                                 qd_idx = 0;
2099                         } else if (*dd_idx >= pd_idx)
2100                                 (*dd_idx) += 2; /* D D P Q D */
2101                         ddf_layout = 1;
2102                         break;
2103
2104                 case ALGORITHM_ROTATING_N_CONTINUE:
2105                         /* Same as left_symmetric but Q is before P */
2106                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2107                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2108                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2109                         ddf_layout = 1;
2110                         break;
2111
2112                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2113                         /* RAID5 left_asymmetric, with Q on last device */
2114                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2115                         if (*dd_idx >= pd_idx)
2116                                 (*dd_idx)++;
2117                         qd_idx = raid_disks - 1;
2118                         break;
2119
2120                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2121                         pd_idx = sector_div(stripe2, raid_disks-1);
2122                         if (*dd_idx >= pd_idx)
2123                                 (*dd_idx)++;
2124                         qd_idx = raid_disks - 1;
2125                         break;
2126
2127                 case ALGORITHM_LEFT_SYMMETRIC_6:
2128                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2129                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2130                         qd_idx = raid_disks - 1;
2131                         break;
2132
2133                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2134                         pd_idx = sector_div(stripe2, raid_disks-1);
2135                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2136                         qd_idx = raid_disks - 1;
2137                         break;
2138
2139                 case ALGORITHM_PARITY_0_6:
2140                         pd_idx = 0;
2141                         (*dd_idx)++;
2142                         qd_idx = raid_disks - 1;
2143                         break;
2144
2145                 default:
2146                         BUG();
2147                 }
2148                 break;
2149         }
2150
2151         if (sh) {
2152                 sh->pd_idx = pd_idx;
2153                 sh->qd_idx = qd_idx;
2154                 sh->ddf_layout = ddf_layout;
2155         }
2156         /*
2157          * Finally, compute the new sector number
2158          */
2159         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2160         return new_sector;
2161 }
2162
2163
2164 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2165 {
2166         struct r5conf *conf = sh->raid_conf;
2167         int raid_disks = sh->disks;
2168         int data_disks = raid_disks - conf->max_degraded;
2169         sector_t new_sector = sh->sector, check;
2170         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2171                                          : conf->chunk_sectors;
2172         int algorithm = previous ? conf->prev_algo
2173                                  : conf->algorithm;
2174         sector_t stripe;
2175         int chunk_offset;
2176         sector_t chunk_number;
2177         int dummy1, dd_idx = i;
2178         sector_t r_sector;
2179         struct stripe_head sh2;
2180
2181
2182         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2183         stripe = new_sector;
2184
2185         if (i == sh->pd_idx)
2186                 return 0;
2187         switch(conf->level) {
2188         case 4: break;
2189         case 5:
2190                 switch (algorithm) {
2191                 case ALGORITHM_LEFT_ASYMMETRIC:
2192                 case ALGORITHM_RIGHT_ASYMMETRIC:
2193                         if (i > sh->pd_idx)
2194                                 i--;
2195                         break;
2196                 case ALGORITHM_LEFT_SYMMETRIC:
2197                 case ALGORITHM_RIGHT_SYMMETRIC:
2198                         if (i < sh->pd_idx)
2199                                 i += raid_disks;
2200                         i -= (sh->pd_idx + 1);
2201                         break;
2202                 case ALGORITHM_PARITY_0:
2203                         i -= 1;
2204                         break;
2205                 case ALGORITHM_PARITY_N:
2206                         break;
2207                 default:
2208                         BUG();
2209                 }
2210                 break;
2211         case 6:
2212                 if (i == sh->qd_idx)
2213                         return 0; /* It is the Q disk */
2214                 switch (algorithm) {
2215                 case ALGORITHM_LEFT_ASYMMETRIC:
2216                 case ALGORITHM_RIGHT_ASYMMETRIC:
2217                 case ALGORITHM_ROTATING_ZERO_RESTART:
2218                 case ALGORITHM_ROTATING_N_RESTART:
2219                         if (sh->pd_idx == raid_disks-1)
2220                                 i--;    /* Q D D D P */
2221                         else if (i > sh->pd_idx)
2222                                 i -= 2; /* D D P Q D */
2223                         break;
2224                 case ALGORITHM_LEFT_SYMMETRIC:
2225                 case ALGORITHM_RIGHT_SYMMETRIC:
2226                         if (sh->pd_idx == raid_disks-1)
2227                                 i--; /* Q D D D P */
2228                         else {
2229                                 /* D D P Q D */
2230                                 if (i < sh->pd_idx)
2231                                         i += raid_disks;
2232                                 i -= (sh->pd_idx + 2);
2233                         }
2234                         break;
2235                 case ALGORITHM_PARITY_0:
2236                         i -= 2;
2237                         break;
2238                 case ALGORITHM_PARITY_N:
2239                         break;
2240                 case ALGORITHM_ROTATING_N_CONTINUE:
2241                         /* Like left_symmetric, but P is before Q */
2242                         if (sh->pd_idx == 0)
2243                                 i--;    /* P D D D Q */
2244                         else {
2245                                 /* D D Q P D */
2246                                 if (i < sh->pd_idx)
2247                                         i += raid_disks;
2248                                 i -= (sh->pd_idx + 1);
2249                         }
2250                         break;
2251                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2252                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2253                         if (i > sh->pd_idx)
2254                                 i--;
2255                         break;
2256                 case ALGORITHM_LEFT_SYMMETRIC_6:
2257                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2258                         if (i < sh->pd_idx)
2259                                 i += data_disks + 1;
2260                         i -= (sh->pd_idx + 1);
2261                         break;
2262                 case ALGORITHM_PARITY_0_6:
2263                         i -= 1;
2264                         break;
2265                 default:
2266                         BUG();
2267                 }
2268                 break;
2269         }
2270
2271         chunk_number = stripe * data_disks + i;
2272         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2273
2274         check = raid5_compute_sector(conf, r_sector,
2275                                      previous, &dummy1, &sh2);
2276         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2277                 || sh2.qd_idx != sh->qd_idx) {
2278                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2279                        mdname(conf->mddev));
2280                 return 0;
2281         }
2282         return r_sector;
2283 }
2284
2285
2286 static void
2287 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2288                          int rcw, int expand)
2289 {
2290         int i, pd_idx = sh->pd_idx, disks = sh->disks;
2291         struct r5conf *conf = sh->raid_conf;
2292         int level = conf->level;
2293
2294         if (rcw) {
2295
2296                 for (i = disks; i--; ) {
2297                         struct r5dev *dev = &sh->dev[i];
2298
2299                         if (dev->towrite) {
2300                                 set_bit(R5_LOCKED, &dev->flags);
2301                                 set_bit(R5_Wantdrain, &dev->flags);
2302                                 if (!expand)
2303                                         clear_bit(R5_UPTODATE, &dev->flags);
2304                                 s->locked++;
2305                         }
2306                 }
2307                 /* if we are not expanding this is a proper write request, and
2308                  * there will be bios with new data to be drained into the
2309                  * stripe cache
2310                  */
2311                 if (!expand) {
2312                         if (!s->locked)
2313                                 /* False alarm, nothing to do */
2314                                 return;
2315                         sh->reconstruct_state = reconstruct_state_drain_run;
2316                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2317                 } else
2318                         sh->reconstruct_state = reconstruct_state_run;
2319
2320                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2321
2322                 if (s->locked + conf->max_degraded == disks)
2323                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2324                                 atomic_inc(&conf->pending_full_writes);
2325         } else {
2326                 BUG_ON(level == 6);
2327                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2328                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2329
2330                 for (i = disks; i--; ) {
2331                         struct r5dev *dev = &sh->dev[i];
2332                         if (i == pd_idx)
2333                                 continue;
2334
2335                         if (dev->towrite &&
2336                             (test_bit(R5_UPTODATE, &dev->flags) ||
2337                              test_bit(R5_Wantcompute, &dev->flags))) {
2338                                 set_bit(R5_Wantdrain, &dev->flags);
2339                                 set_bit(R5_LOCKED, &dev->flags);
2340                                 clear_bit(R5_UPTODATE, &dev->flags);
2341                                 s->locked++;
2342                         }
2343                 }
2344                 if (!s->locked)
2345                         /* False alarm - nothing to do */
2346                         return;
2347                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2348                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2349                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2350                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2351         }
2352
2353         /* keep the parity disk(s) locked while asynchronous operations
2354          * are in flight
2355          */
2356         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2357         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2358         s->locked++;
2359
2360         if (level == 6) {
2361                 int qd_idx = sh->qd_idx;
2362                 struct r5dev *dev = &sh->dev[qd_idx];
2363
2364                 set_bit(R5_LOCKED, &dev->flags);
2365                 clear_bit(R5_UPTODATE, &dev->flags);
2366                 s->locked++;
2367         }
2368
2369         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2370                 __func__, (unsigned long long)sh->sector,
2371                 s->locked, s->ops_request);
2372 }
2373
2374 /*
2375  * Each stripe/dev can have one or more bion attached.
2376  * toread/towrite point to the first in a chain.
2377  * The bi_next chain must be in order.
2378  */
2379 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2380 {
2381         struct bio **bip;
2382         struct r5conf *conf = sh->raid_conf;
2383         int firstwrite=0;
2384
2385         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2386                 (unsigned long long)bi->bi_sector,
2387                 (unsigned long long)sh->sector);
2388
2389         /*
2390          * If several bio share a stripe. The bio bi_phys_segments acts as a
2391          * reference count to avoid race. The reference count should already be
2392          * increased before this function is called (for example, in
2393          * make_request()), so other bio sharing this stripe will not free the
2394          * stripe. If a stripe is owned by one stripe, the stripe lock will
2395          * protect it.
2396          */
2397         spin_lock_irq(&sh->stripe_lock);
2398         if (forwrite) {
2399                 bip = &sh->dev[dd_idx].towrite;
2400                 if (*bip == NULL)
2401                         firstwrite = 1;
2402         } else
2403                 bip = &sh->dev[dd_idx].toread;
2404         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2405                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2406                         goto overlap;
2407                 bip = & (*bip)->bi_next;
2408         }
2409         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2410                 goto overlap;
2411
2412         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2413         if (*bip)
2414                 bi->bi_next = *bip;
2415         *bip = bi;
2416         raid5_inc_bi_active_stripes(bi);
2417
2418         if (forwrite) {
2419                 /* check if page is covered */
2420                 sector_t sector = sh->dev[dd_idx].sector;
2421                 for (bi=sh->dev[dd_idx].towrite;
2422                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2423                              bi && bi->bi_sector <= sector;
2424                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2425                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2426                                 sector = bi->bi_sector + (bi->bi_size>>9);
2427                 }
2428                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2429                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2430         }
2431
2432         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2433                 (unsigned long long)(*bip)->bi_sector,
2434                 (unsigned long long)sh->sector, dd_idx);
2435         spin_unlock_irq(&sh->stripe_lock);
2436
2437         if (conf->mddev->bitmap && firstwrite) {
2438                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2439                                   STRIPE_SECTORS, 0);
2440                 sh->bm_seq = conf->seq_flush+1;
2441                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2442         }
2443         return 1;
2444
2445  overlap:
2446         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2447         spin_unlock_irq(&sh->stripe_lock);
2448         return 0;
2449 }
2450
2451 static void end_reshape(struct r5conf *conf);
2452
2453 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2454                             struct stripe_head *sh)
2455 {
2456         int sectors_per_chunk =
2457                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2458         int dd_idx;
2459         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2460         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2461
2462         raid5_compute_sector(conf,
2463                              stripe * (disks - conf->max_degraded)
2464                              *sectors_per_chunk + chunk_offset,
2465                              previous,
2466                              &dd_idx, sh);
2467 }
2468
2469 static void
2470 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2471                                 struct stripe_head_state *s, int disks,
2472                                 struct bio **return_bi)
2473 {
2474         int i;
2475         for (i = disks; i--; ) {
2476                 struct bio *bi;
2477                 int bitmap_end = 0;
2478
2479                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2480                         struct md_rdev *rdev;
2481                         rcu_read_lock();
2482                         rdev = rcu_dereference(conf->disks[i].rdev);
2483                         if (rdev && test_bit(In_sync, &rdev->flags))
2484                                 atomic_inc(&rdev->nr_pending);
2485                         else
2486                                 rdev = NULL;
2487                         rcu_read_unlock();
2488                         if (rdev) {
2489                                 if (!rdev_set_badblocks(
2490                                             rdev,
2491                                             sh->sector,
2492                                             STRIPE_SECTORS, 0))
2493                                         md_error(conf->mddev, rdev);
2494                                 rdev_dec_pending(rdev, conf->mddev);
2495                         }
2496                 }
2497                 spin_lock_irq(&sh->stripe_lock);
2498                 /* fail all writes first */
2499                 bi = sh->dev[i].towrite;
2500                 sh->dev[i].towrite = NULL;
2501                 spin_unlock_irq(&sh->stripe_lock);
2502                 if (bi)
2503                         bitmap_end = 1;
2504
2505                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2506                         wake_up(&conf->wait_for_overlap);
2507
2508                 while (bi && bi->bi_sector <
2509                         sh->dev[i].sector + STRIPE_SECTORS) {
2510                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2511                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2512                         if (!raid5_dec_bi_active_stripes(bi)) {
2513                                 md_write_end(conf->mddev);
2514                                 bi->bi_next = *return_bi;
2515                                 *return_bi = bi;
2516                         }
2517                         bi = nextbi;
2518                 }
2519                 if (bitmap_end)
2520                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2521                                 STRIPE_SECTORS, 0, 0);
2522                 bitmap_end = 0;
2523                 /* and fail all 'written' */
2524                 bi = sh->dev[i].written;
2525                 sh->dev[i].written = NULL;
2526                 if (bi) bitmap_end = 1;
2527                 while (bi && bi->bi_sector <
2528                        sh->dev[i].sector + STRIPE_SECTORS) {
2529                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2530                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2531                         if (!raid5_dec_bi_active_stripes(bi)) {
2532                                 md_write_end(conf->mddev);
2533                                 bi->bi_next = *return_bi;
2534                                 *return_bi = bi;
2535                         }
2536                         bi = bi2;
2537                 }
2538
2539                 /* fail any reads if this device is non-operational and
2540                  * the data has not reached the cache yet.
2541                  */
2542                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2543                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2544                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2545                         spin_lock_irq(&sh->stripe_lock);
2546                         bi = sh->dev[i].toread;
2547                         sh->dev[i].toread = NULL;
2548                         spin_unlock_irq(&sh->stripe_lock);
2549                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2550                                 wake_up(&conf->wait_for_overlap);
2551                         while (bi && bi->bi_sector <
2552                                sh->dev[i].sector + STRIPE_SECTORS) {
2553                                 struct bio *nextbi =
2554                                         r5_next_bio(bi, sh->dev[i].sector);
2555                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2556                                 if (!raid5_dec_bi_active_stripes(bi)) {
2557                                         bi->bi_next = *return_bi;
2558                                         *return_bi = bi;
2559                                 }
2560                                 bi = nextbi;
2561                         }
2562                 }
2563                 if (bitmap_end)
2564                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2565                                         STRIPE_SECTORS, 0, 0);
2566                 /* If we were in the middle of a write the parity block might
2567                  * still be locked - so just clear all R5_LOCKED flags
2568                  */
2569                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2570         }
2571
2572         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2573                 if (atomic_dec_and_test(&conf->pending_full_writes))
2574                         md_wakeup_thread(conf->mddev->thread);
2575 }
2576
2577 static void
2578 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2579                    struct stripe_head_state *s)
2580 {
2581         int abort = 0;
2582         int i;
2583
2584         clear_bit(STRIPE_SYNCING, &sh->state);
2585         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2586                 wake_up(&conf->wait_for_overlap);
2587         s->syncing = 0;
2588         s->replacing = 0;
2589         /* There is nothing more to do for sync/check/repair.
2590          * Don't even need to abort as that is handled elsewhere
2591          * if needed, and not always wanted e.g. if there is a known
2592          * bad block here.
2593          * For recover/replace we need to record a bad block on all
2594          * non-sync devices, or abort the recovery
2595          */
2596         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2597                 /* During recovery devices cannot be removed, so
2598                  * locking and refcounting of rdevs is not needed
2599                  */
2600                 for (i = 0; i < conf->raid_disks; i++) {
2601                         struct md_rdev *rdev = conf->disks[i].rdev;
2602                         if (rdev
2603                             && !test_bit(Faulty, &rdev->flags)
2604                             && !test_bit(In_sync, &rdev->flags)
2605                             && !rdev_set_badblocks(rdev, sh->sector,
2606                                                    STRIPE_SECTORS, 0))
2607                                 abort = 1;
2608                         rdev = conf->disks[i].replacement;
2609                         if (rdev
2610                             && !test_bit(Faulty, &rdev->flags)
2611                             && !test_bit(In_sync, &rdev->flags)
2612                             && !rdev_set_badblocks(rdev, sh->sector,
2613                                                    STRIPE_SECTORS, 0))
2614                                 abort = 1;
2615                 }
2616                 if (abort)
2617                         conf->recovery_disabled =
2618                                 conf->mddev->recovery_disabled;
2619         }
2620         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2621 }
2622
2623 static int want_replace(struct stripe_head *sh, int disk_idx)
2624 {
2625         struct md_rdev *rdev;
2626         int rv = 0;
2627         /* Doing recovery so rcu locking not required */
2628         rdev = sh->raid_conf->disks[disk_idx].replacement;
2629         if (rdev
2630             && !test_bit(Faulty, &rdev->flags)
2631             && !test_bit(In_sync, &rdev->flags)
2632             && (rdev->recovery_offset <= sh->sector
2633                 || rdev->mddev->recovery_cp <= sh->sector))
2634                 rv = 1;
2635
2636         return rv;
2637 }
2638
2639 /* fetch_block - checks the given member device to see if its data needs
2640  * to be read or computed to satisfy a request.
2641  *
2642  * Returns 1 when no more member devices need to be checked, otherwise returns
2643  * 0 to tell the loop in handle_stripe_fill to continue
2644  */
2645 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2646                        int disk_idx, int disks)
2647 {
2648         struct r5dev *dev = &sh->dev[disk_idx];
2649         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2650                                   &sh->dev[s->failed_num[1]] };
2651
2652         /* is the data in this block needed, and can we get it? */
2653         if (!test_bit(R5_LOCKED, &dev->flags) &&
2654             !test_bit(R5_UPTODATE, &dev->flags) &&
2655             (dev->toread ||
2656              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2657              s->syncing || s->expanding ||
2658              (s->replacing && want_replace(sh, disk_idx)) ||
2659              (s->failed >= 1 && fdev[0]->toread) ||
2660              (s->failed >= 2 && fdev[1]->toread) ||
2661              (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2662               !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2663              (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2664                 /* we would like to get this block, possibly by computing it,
2665                  * otherwise read it if the backing disk is insync
2666                  */
2667                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2668                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2669                 if ((s->uptodate == disks - 1) &&
2670                     (s->failed && (disk_idx == s->failed_num[0] ||
2671                                    disk_idx == s->failed_num[1]))) {
2672                         /* have disk failed, and we're requested to fetch it;
2673                          * do compute it
2674                          */
2675                         pr_debug("Computing stripe %llu block %d\n",
2676                                (unsigned long long)sh->sector, disk_idx);
2677                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2678                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2679                         set_bit(R5_Wantcompute, &dev->flags);
2680                         sh->ops.target = disk_idx;
2681                         sh->ops.target2 = -1; /* no 2nd target */
2682                         s->req_compute = 1;
2683                         /* Careful: from this point on 'uptodate' is in the eye
2684                          * of raid_run_ops which services 'compute' operations
2685                          * before writes. R5_Wantcompute flags a block that will
2686                          * be R5_UPTODATE by the time it is needed for a
2687                          * subsequent operation.
2688                          */
2689                         s->uptodate++;
2690                         return 1;
2691                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2692                         /* Computing 2-failure is *very* expensive; only
2693                          * do it if failed >= 2
2694                          */
2695                         int other;
2696                         for (other = disks; other--; ) {
2697                                 if (other == disk_idx)
2698                                         continue;
2699                                 if (!test_bit(R5_UPTODATE,
2700                                       &sh->dev[other].flags))
2701                                         break;
2702                         }
2703                         BUG_ON(other < 0);
2704                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2705                                (unsigned long long)sh->sector,
2706                                disk_idx, other);
2707                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2708                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2709                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2710                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2711                         sh->ops.target = disk_idx;
2712                         sh->ops.target2 = other;
2713                         s->uptodate += 2;
2714                         s->req_compute = 1;
2715                         return 1;
2716                 } else if (test_bit(R5_Insync, &dev->flags)) {
2717                         set_bit(R5_LOCKED, &dev->flags);
2718                         set_bit(R5_Wantread, &dev->flags);
2719                         s->locked++;
2720                         pr_debug("Reading block %d (sync=%d)\n",
2721                                 disk_idx, s->syncing);
2722                 }
2723         }
2724
2725         return 0;
2726 }
2727
2728 /**
2729  * handle_stripe_fill - read or compute data to satisfy pending requests.
2730  */
2731 static void handle_stripe_fill(struct stripe_head *sh,
2732                                struct stripe_head_state *s,
2733                                int disks)
2734 {
2735         int i;
2736
2737         /* look for blocks to read/compute, skip this if a compute
2738          * is already in flight, or if the stripe contents are in the
2739          * midst of changing due to a write
2740          */
2741         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2742             !sh->reconstruct_state)
2743                 for (i = disks; i--; )
2744                         if (fetch_block(sh, s, i, disks))
2745                                 break;
2746         set_bit(STRIPE_HANDLE, &sh->state);
2747 }
2748
2749
2750 /* handle_stripe_clean_event
2751  * any written block on an uptodate or failed drive can be returned.
2752  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2753  * never LOCKED, so we don't need to test 'failed' directly.
2754  */
2755 static void handle_stripe_clean_event(struct r5conf *conf,
2756         struct stripe_head *sh, int disks, struct bio **return_bi)
2757 {
2758         int i;
2759         struct r5dev *dev;
2760         int discard_pending = 0;
2761
2762         for (i = disks; i--; )
2763                 if (sh->dev[i].written) {
2764                         dev = &sh->dev[i];
2765                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2766                             (test_bit(R5_UPTODATE, &dev->flags) ||
2767                              test_bit(R5_Discard, &dev->flags))) {
2768                                 /* We can return any write requests */
2769                                 struct bio *wbi, *wbi2;
2770                                 pr_debug("Return write for disc %d\n", i);
2771                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
2772                                         clear_bit(R5_UPTODATE, &dev->flags);
2773                                 wbi = dev->written;
2774                                 dev->written = NULL;
2775                                 while (wbi && wbi->bi_sector <
2776                                         dev->sector + STRIPE_SECTORS) {
2777                                         wbi2 = r5_next_bio(wbi, dev->sector);
2778                                         if (!raid5_dec_bi_active_stripes(wbi)) {
2779                                                 md_write_end(conf->mddev);
2780                                                 wbi->bi_next = *return_bi;
2781                                                 *return_bi = wbi;
2782                                         }
2783                                         wbi = wbi2;
2784                                 }
2785                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2786                                                 STRIPE_SECTORS,
2787                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2788                                                 0);
2789                         } else if (test_bit(R5_Discard, &dev->flags))
2790                                 discard_pending = 1;
2791                 }
2792         if (!discard_pending &&
2793             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
2794                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2795                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2796                 if (sh->qd_idx >= 0) {
2797                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2798                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
2799                 }
2800                 /* now that discard is done we can proceed with any sync */
2801                 clear_bit(STRIPE_DISCARD, &sh->state);
2802                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2803                         set_bit(STRIPE_HANDLE, &sh->state);
2804
2805         }
2806
2807         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2808                 if (atomic_dec_and_test(&conf->pending_full_writes))
2809                         md_wakeup_thread(conf->mddev->thread);
2810 }
2811
2812 static void handle_stripe_dirtying(struct r5conf *conf,
2813                                    struct stripe_head *sh,
2814                                    struct stripe_head_state *s,
2815                                    int disks)
2816 {
2817         int rmw = 0, rcw = 0, i;
2818         sector_t recovery_cp = conf->mddev->recovery_cp;
2819
2820         /* RAID6 requires 'rcw' in current implementation.
2821          * Otherwise, check whether resync is now happening or should start.
2822          * If yes, then the array is dirty (after unclean shutdown or
2823          * initial creation), so parity in some stripes might be inconsistent.
2824          * In this case, we need to always do reconstruct-write, to ensure
2825          * that in case of drive failure or read-error correction, we
2826          * generate correct data from the parity.
2827          */
2828         if (conf->max_degraded == 2 ||
2829             (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2830                 /* Calculate the real rcw later - for now make it
2831                  * look like rcw is cheaper
2832                  */
2833                 rcw = 1; rmw = 2;
2834                 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2835                          conf->max_degraded, (unsigned long long)recovery_cp,
2836                          (unsigned long long)sh->sector);
2837         } else for (i = disks; i--; ) {
2838                 /* would I have to read this buffer for read_modify_write */
2839                 struct r5dev *dev = &sh->dev[i];
2840                 if ((dev->towrite || i == sh->pd_idx) &&
2841                     !test_bit(R5_LOCKED, &dev->flags) &&
2842                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2843                       test_bit(R5_Wantcompute, &dev->flags))) {
2844                         if (test_bit(R5_Insync, &dev->flags))
2845                                 rmw++;
2846                         else
2847                                 rmw += 2*disks;  /* cannot read it */
2848                 }
2849                 /* Would I have to read this buffer for reconstruct_write */
2850                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2851                     !test_bit(R5_LOCKED, &dev->flags) &&
2852                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2853                     test_bit(R5_Wantcompute, &dev->flags))) {
2854                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2855                         else
2856                                 rcw += 2*disks;
2857                 }
2858         }
2859         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2860                 (unsigned long long)sh->sector, rmw, rcw);
2861         set_bit(STRIPE_HANDLE, &sh->state);
2862         if (rmw < rcw && rmw > 0) {
2863                 /* prefer read-modify-write, but need to get some data */
2864                 if (conf->mddev->queue)
2865                         blk_add_trace_msg(conf->mddev->queue,
2866                                           "raid5 rmw %llu %d",
2867                                           (unsigned long long)sh->sector, rmw);
2868                 for (i = disks; i--; ) {
2869                         struct r5dev *dev = &sh->dev[i];
2870                         if ((dev->towrite || i == sh->pd_idx) &&
2871                             !test_bit(R5_LOCKED, &dev->flags) &&
2872                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2873                             test_bit(R5_Wantcompute, &dev->flags)) &&
2874                             test_bit(R5_Insync, &dev->flags)) {
2875                                 if (
2876                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2877                                         pr_debug("Read_old block "
2878                                                  "%d for r-m-w\n", i);
2879                                         set_bit(R5_LOCKED, &dev->flags);
2880                                         set_bit(R5_Wantread, &dev->flags);
2881                                         s->locked++;
2882                                 } else {
2883                                         set_bit(STRIPE_DELAYED, &sh->state);
2884                                         set_bit(STRIPE_HANDLE, &sh->state);
2885                                 }
2886                         }
2887                 }
2888         }
2889         if (rcw <= rmw && rcw > 0) {
2890                 /* want reconstruct write, but need to get some data */
2891                 int qread =0;
2892                 rcw = 0;
2893                 for (i = disks; i--; ) {
2894                         struct r5dev *dev = &sh->dev[i];
2895                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2896                             i != sh->pd_idx && i != sh->qd_idx &&
2897                             !test_bit(R5_LOCKED, &dev->flags) &&
2898                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2899                               test_bit(R5_Wantcompute, &dev->flags))) {
2900                                 rcw++;
2901                                 if (!test_bit(R5_Insync, &dev->flags))
2902                                         continue; /* it's a failed drive */
2903                                 if (
2904                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2905                                         pr_debug("Read_old block "
2906                                                 "%d for Reconstruct\n", i);
2907                                         set_bit(R5_LOCKED, &dev->flags);
2908                                         set_bit(R5_Wantread, &dev->flags);
2909                                         s->locked++;
2910                                         qread++;
2911                                 } else {
2912                                         set_bit(STRIPE_DELAYED, &sh->state);
2913                                         set_bit(STRIPE_HANDLE, &sh->state);
2914                                 }
2915                         }
2916                 }
2917                 if (rcw && conf->mddev->queue)
2918                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
2919                                           (unsigned long long)sh->sector,
2920                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
2921         }
2922         /* now if nothing is locked, and if we have enough data,
2923          * we can start a write request
2924          */
2925         /* since handle_stripe can be called at any time we need to handle the
2926          * case where a compute block operation has been submitted and then a
2927          * subsequent call wants to start a write request.  raid_run_ops only
2928          * handles the case where compute block and reconstruct are requested
2929          * simultaneously.  If this is not the case then new writes need to be
2930          * held off until the compute completes.
2931          */
2932         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2933             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2934             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2935                 schedule_reconstruction(sh, s, rcw == 0, 0);
2936 }
2937
2938 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2939                                 struct stripe_head_state *s, int disks)
2940 {
2941         struct r5dev *dev = NULL;
2942
2943         set_bit(STRIPE_HANDLE, &sh->state);
2944
2945         switch (sh->check_state) {
2946         case check_state_idle:
2947                 /* start a new check operation if there are no failures */
2948                 if (s->failed == 0) {
2949                         BUG_ON(s->uptodate != disks);
2950                         sh->check_state = check_state_run;
2951                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2952                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2953                         s->uptodate--;
2954                         break;
2955                 }
2956                 dev = &sh->dev[s->failed_num[0]];
2957                 /* fall through */
2958         case check_state_compute_result:
2959                 sh->check_state = check_state_idle;
2960                 if (!dev)
2961                         dev = &sh->dev[sh->pd_idx];
2962
2963                 /* check that a write has not made the stripe insync */
2964                 if (test_bit(STRIPE_INSYNC, &sh->state))
2965                         break;
2966
2967                 /* either failed parity check, or recovery is happening */
2968                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2969                 BUG_ON(s->uptodate != disks);
2970
2971                 set_bit(R5_LOCKED, &dev->flags);
2972                 s->locked++;
2973                 set_bit(R5_Wantwrite, &dev->flags);
2974
2975                 clear_bit(STRIPE_DEGRADED, &sh->state);
2976                 set_bit(STRIPE_INSYNC, &sh->state);
2977                 break;
2978         case check_state_run:
2979                 break; /* we will be called again upon completion */
2980         case check_state_check_result:
2981                 sh->check_state = check_state_idle;
2982
2983                 /* if a failure occurred during the check operation, leave
2984                  * STRIPE_INSYNC not set and let the stripe be handled again
2985                  */
2986                 if (s->failed)
2987                         break;
2988
2989                 /* handle a successful check operation, if parity is correct
2990                  * we are done.  Otherwise update the mismatch count and repair
2991                  * parity if !MD_RECOVERY_CHECK
2992                  */
2993                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2994                         /* parity is correct (on disc,
2995                          * not in buffer any more)
2996                          */
2997                         set_bit(STRIPE_INSYNC, &sh->state);
2998                 else {
2999                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3000                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3001                                 /* don't try to repair!! */
3002                                 set_bit(STRIPE_INSYNC, &sh->state);
3003                         else {
3004                                 sh->check_state = check_state_compute_run;
3005                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3006                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3007                                 set_bit(R5_Wantcompute,
3008                                         &sh->dev[sh->pd_idx].flags);
3009                                 sh->ops.target = sh->pd_idx;
3010                                 sh->ops.target2 = -1;
3011                                 s->uptodate++;
3012                         }
3013                 }
3014                 break;
3015         case check_state_compute_run:
3016                 break;
3017         default:
3018                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3019                        __func__, sh->check_state,
3020                        (unsigned long long) sh->sector);
3021                 BUG();
3022         }
3023 }
3024
3025
3026 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3027                                   struct stripe_head_state *s,
3028                                   int disks)
3029 {
3030         int pd_idx = sh->pd_idx;
3031         int qd_idx = sh->qd_idx;
3032         struct r5dev *dev;
3033
3034         set_bit(STRIPE_HANDLE, &sh->state);
3035
3036         BUG_ON(s->failed > 2);
3037
3038         /* Want to check and possibly repair P and Q.
3039          * However there could be one 'failed' device, in which
3040          * case we can only check one of them, possibly using the
3041          * other to generate missing data
3042          */
3043
3044         switch (sh->check_state) {
3045         case check_state_idle:
3046                 /* start a new check operation if there are < 2 failures */
3047                 if (s->failed == s->q_failed) {
3048                         /* The only possible failed device holds Q, so it
3049                          * makes sense to check P (If anything else were failed,
3050                          * we would have used P to recreate it).
3051                          */
3052                         sh->check_state = check_state_run;
3053                 }
3054                 if (!s->q_failed && s->failed < 2) {
3055                         /* Q is not failed, and we didn't use it to generate
3056                          * anything, so it makes sense to check it
3057                          */
3058                         if (sh->check_state == check_state_run)
3059                                 sh->check_state = check_state_run_pq;
3060                         else
3061                                 sh->check_state = check_state_run_q;
3062                 }
3063
3064                 /* discard potentially stale zero_sum_result */
3065                 sh->ops.zero_sum_result = 0;
3066
3067                 if (sh->check_state == check_state_run) {
3068                         /* async_xor_zero_sum destroys the contents of P */
3069                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3070                         s->uptodate--;
3071                 }
3072                 if (sh->check_state >= check_state_run &&
3073                     sh->check_state <= check_state_run_pq) {
3074                         /* async_syndrome_zero_sum preserves P and Q, so
3075                          * no need to mark them !uptodate here
3076                          */
3077                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3078                         break;
3079                 }
3080
3081                 /* we have 2-disk failure */
3082                 BUG_ON(s->failed != 2);
3083                 /* fall through */
3084         case check_state_compute_result:
3085                 sh->check_state = check_state_idle;
3086
3087                 /* check that a write has not made the stripe insync */
3088                 if (test_bit(STRIPE_INSYNC, &sh->state))
3089                         break;
3090
3091                 /* now write out any block on a failed drive,
3092                  * or P or Q if they were recomputed
3093                  */
3094                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3095                 if (s->failed == 2) {
3096                         dev = &sh->dev[s->failed_num[1]];
3097                         s->locked++;
3098                         set_bit(R5_LOCKED, &dev->flags);
3099                         set_bit(R5_Wantwrite, &dev->flags);
3100                 }
3101                 if (s->failed >= 1) {
3102                         dev = &sh->dev[s->failed_num[0]];
3103                         s->locked++;
3104                         set_bit(R5_LOCKED, &dev->flags);
3105                         set_bit(R5_Wantwrite, &dev->flags);
3106                 }
3107                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3108                         dev = &sh->dev[pd_idx];
3109                         s->locked++;
3110                         set_bit(R5_LOCKED, &dev->flags);
3111                         set_bit(R5_Wantwrite, &dev->flags);
3112                 }
3113                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3114                         dev = &sh->dev[qd_idx];
3115                         s->locked++;
3116                         set_bit(R5_LOCKED, &dev->flags);
3117                         set_bit(R5_Wantwrite, &dev->flags);
3118                 }
3119                 clear_bit(STRIPE_DEGRADED, &sh->state);
3120
3121                 set_bit(STRIPE_INSYNC, &sh->state);
3122                 break;
3123         case check_state_run:
3124         case check_state_run_q:
3125         case check_state_run_pq:
3126                 break; /* we will be called again upon completion */
3127         case check_state_check_result:
3128                 sh->check_state = check_state_idle;
3129
3130                 /* handle a successful check operation, if parity is correct
3131                  * we are done.  Otherwise update the mismatch count and repair
3132                  * parity if !MD_RECOVERY_CHECK
3133                  */
3134                 if (sh->ops.zero_sum_result == 0) {
3135                         /* both parities are correct */
3136                         if (!s->failed)
3137                                 set_bit(STRIPE_INSYNC, &sh->state);
3138                         else {
3139                                 /* in contrast to the raid5 case we can validate
3140                                  * parity, but still have a failure to write
3141                                  * back
3142                                  */
3143                                 sh->check_state = check_state_compute_result;
3144                                 /* Returning at this point means that we may go
3145                                  * off and bring p and/or q uptodate again so
3146                                  * we make sure to check zero_sum_result again
3147                                  * to verify if p or q need writeback
3148                                  */
3149                         }
3150                 } else {
3151                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3152                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3153                                 /* don't try to repair!! */
3154                                 set_bit(STRIPE_INSYNC, &sh->state);
3155                         else {
3156                                 int *target = &sh->ops.target;
3157
3158                                 sh->ops.target = -1;
3159                                 sh->ops.target2 = -1;
3160                                 sh->check_state = check_state_compute_run;
3161                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3162                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3163                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3164                                         set_bit(R5_Wantcompute,
3165                                                 &sh->dev[pd_idx].flags);
3166                                         *target = pd_idx;
3167                                         target = &sh->ops.target2;
3168                                         s->uptodate++;
3169                                 }
3170                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3171                                         set_bit(R5_Wantcompute,
3172                                                 &sh->dev[qd_idx].flags);
3173                                         *target = qd_idx;
3174                                         s->uptodate++;
3175                                 }
3176                         }
3177                 }
3178                 break;
3179         case check_state_compute_run:
3180                 break;
3181         default:
3182                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3183                        __func__, sh->check_state,
3184                        (unsigned long long) sh->sector);
3185                 BUG();
3186         }
3187 }
3188
3189 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3190 {
3191         int i;
3192
3193         /* We have read all the blocks in this stripe and now we need to
3194          * copy some of them into a target stripe for expand.
3195          */
3196         struct dma_async_tx_descriptor *tx = NULL;
3197         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3198         for (i = 0; i < sh->disks; i++)
3199                 if (i != sh->pd_idx && i != sh->qd_idx) {
3200                         int dd_idx, j;
3201                         struct stripe_head *sh2;
3202                         struct async_submit_ctl submit;
3203
3204                         sector_t bn = compute_blocknr(sh, i, 1);
3205                         sector_t s = raid5_compute_sector(conf, bn, 0,
3206                                                           &dd_idx, NULL);
3207                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
3208                         if (sh2 == NULL)
3209                                 /* so far only the early blocks of this stripe
3210                                  * have been requested.  When later blocks
3211                                  * get requested, we will try again
3212                                  */
3213                                 continue;
3214                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3215                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3216                                 /* must have already done this block */
3217                                 release_stripe(sh2);
3218                                 continue;
3219                         }
3220
3221                         /* place all the copies on one channel */
3222                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3223                         tx = async_memcpy(sh2->dev[dd_idx].page,
3224                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
3225                                           &submit);
3226
3227                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3228                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3229                         for (j = 0; j < conf->raid_disks; j++)
3230                                 if (j != sh2->pd_idx &&
3231                                     j != sh2->qd_idx &&
3232                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
3233                                         break;
3234                         if (j == conf->raid_disks) {
3235                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3236                                 set_bit(STRIPE_HANDLE, &sh2->state);
3237                         }
3238                         release_stripe(sh2);
3239
3240                 }
3241         /* done submitting copies, wait for them to complete */
3242         async_tx_quiesce(&tx);
3243 }
3244
3245 /*
3246  * handle_stripe - do things to a stripe.
3247  *
3248  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3249  * state of various bits to see what needs to be done.
3250  * Possible results:
3251  *    return some read requests which now have data
3252  *    return some write requests which are safely on storage
3253  *    schedule a read on some buffers
3254  *    schedule a write of some buffers
3255  *    return confirmation of parity correctness
3256  *
3257  */
3258
3259 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3260 {
3261         struct r5conf *conf = sh->raid_conf;
3262         int disks = sh->disks;
3263         struct r5dev *dev;
3264         int i;
3265         int do_recovery = 0;
3266
3267         memset(s, 0, sizeof(*s));
3268
3269         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3270         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3271         s->failed_num[0] = -1;
3272         s->failed_num[1] = -1;
3273
3274         /* Now to look around and see what can be done */
3275         rcu_read_lock();
3276         for (i=disks; i--; ) {
3277                 struct md_rdev *rdev;
3278                 sector_t first_bad;
3279                 int bad_sectors;
3280                 int is_bad = 0;
3281
3282                 dev = &sh->dev[i];
3283
3284                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3285                          i, dev->flags,
3286                          dev->toread, dev->towrite, dev->written);
3287                 /* maybe we can reply to a read
3288                  *
3289                  * new wantfill requests are only permitted while
3290                  * ops_complete_biofill is guaranteed to be inactive
3291                  */
3292                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3293                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3294                         set_bit(R5_Wantfill, &dev->flags);
3295
3296                 /* now count some things */
3297                 if (test_bit(R5_LOCKED, &dev->flags))
3298                         s->locked++;
3299                 if (test_bit(R5_UPTODATE, &dev->flags))
3300                         s->uptodate++;
3301                 if (test_bit(R5_Wantcompute, &dev->flags)) {
3302                         s->compute++;
3303                         BUG_ON(s->compute > 2);
3304                 }
3305
3306                 if (test_bit(R5_Wantfill, &dev->flags))
3307                         s->to_fill++;
3308                 else if (dev->toread)
3309                         s->to_read++;
3310                 if (dev->towrite) {
3311                         s->to_write++;
3312                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3313                                 s->non_overwrite++;
3314                 }
3315                 if (dev->written)
3316                         s->written++;
3317                 /* Prefer to use the replacement for reads, but only
3318                  * if it is recovered enough and has no bad blocks.
3319                  */
3320                 rdev = rcu_dereference(conf->disks[i].replacement);
3321                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3322                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3323                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3324                                  &first_bad, &bad_sectors))
3325                         set_bit(R5_ReadRepl, &dev->flags);
3326                 else {
3327                         if (rdev)
3328                                 set_bit(R5_NeedReplace, &dev->flags);
3329                         rdev = rcu_dereference(conf->disks[i].rdev);
3330                         clear_bit(R5_ReadRepl, &dev->flags);
3331                 }
3332                 if (rdev && test_bit(Faulty, &rdev->flags))
3333                         rdev = NULL;
3334                 if (rdev) {
3335                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3336                                              &first_bad, &bad_sectors);
3337                         if (s->blocked_rdev == NULL
3338                             && (test_bit(Blocked, &rdev->flags)
3339                                 || is_bad < 0)) {
3340                                 if (is_bad < 0)
3341                                         set_bit(BlockedBadBlocks,
3342                                                 &rdev->flags);
3343                                 s->blocked_rdev = rdev;
3344                                 atomic_inc(&rdev->nr_pending);
3345                         }
3346                 }
3347                 clear_bit(R5_Insync, &dev->flags);
3348                 if (!rdev)
3349                         /* Not in-sync */;
3350                 else if (is_bad) {
3351                         /* also not in-sync */
3352                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3353                             test_bit(R5_UPTODATE, &dev->flags)) {
3354                                 /* treat as in-sync, but with a read error
3355                                  * which we can now try to correct
3356                                  */
3357                                 set_bit(R5_Insync, &dev->flags);
3358                                 set_bit(R5_ReadError, &dev->flags);
3359                         }
3360                 } else if (test_bit(In_sync, &rdev->flags))
3361                         set_bit(R5_Insync, &dev->flags);
3362                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3363                         /* in sync if before recovery_offset */
3364                         set_bit(R5_Insync, &dev->flags);
3365                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3366                          test_bit(R5_Expanded, &dev->flags))
3367                         /* If we've reshaped into here, we assume it is Insync.
3368                          * We will shortly update recovery_offset to make
3369                          * it official.
3370                          */
3371                         set_bit(R5_Insync, &dev->flags);
3372
3373                 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3374                         /* This flag does not apply to '.replacement'
3375                          * only to .rdev, so make sure to check that*/
3376                         struct md_rdev *rdev2 = rcu_dereference(
3377                                 conf->disks[i].rdev);
3378                         if (rdev2 == rdev)
3379                                 clear_bit(R5_Insync, &dev->flags);
3380                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3381                                 s->handle_bad_blocks = 1;
3382                                 atomic_inc(&rdev2->nr_pending);
3383                         } else
3384                                 clear_bit(R5_WriteError, &dev->flags);
3385                 }
3386                 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3387                         /* This flag does not apply to '.replacement'
3388                          * only to .rdev, so make sure to check that*/
3389                         struct md_rdev *rdev2 = rcu_dereference(
3390                                 conf->disks[i].rdev);
3391                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3392                                 s->handle_bad_blocks = 1;
3393                                 atomic_inc(&rdev2->nr_pending);
3394                         } else
3395                                 clear_bit(R5_MadeGood, &dev->flags);
3396                 }
3397                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3398                         struct md_rdev *rdev2 = rcu_dereference(
3399                                 conf->disks[i].replacement);
3400                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3401                                 s->handle_bad_blocks = 1;
3402                                 atomic_inc(&rdev2->nr_pending);
3403                         } else
3404                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
3405                 }
3406                 if (!test_bit(R5_Insync, &dev->flags)) {
3407                         /* The ReadError flag will just be confusing now */
3408                         clear_bit(R5_ReadError, &dev->flags);
3409                         clear_bit(R5_ReWrite, &dev->flags);
3410                 }
3411                 if (test_bit(R5_ReadError, &dev->flags))
3412                         clear_bit(R5_Insync, &dev->flags);
3413                 if (!test_bit(R5_Insync, &dev->flags)) {
3414                         if (s->failed < 2)
3415                                 s->failed_num[s->failed] = i;
3416                         s->failed++;
3417                         if (rdev && !test_bit(Faulty, &rdev->flags))
3418                                 do_recovery = 1;
3419                 }
3420         }
3421         if (test_bit(STRIPE_SYNCING, &sh->state)) {
3422                 /* If there is a failed device being replaced,
3423                  *     we must be recovering.
3424                  * else if we are after recovery_cp, we must be syncing
3425                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3426                  * else we can only be replacing
3427                  * sync and recovery both need to read all devices, and so
3428                  * use the same flag.
3429                  */
3430                 if (do_recovery ||
3431                     sh->sector >= conf->mddev->recovery_cp ||
3432                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3433                         s->syncing = 1;
3434                 else
3435                         s->replacing = 1;
3436         }
3437         rcu_read_unlock();
3438 }
3439
3440 static void handle_stripe(struct stripe_head *sh)
3441 {
3442         struct stripe_head_state s;
3443         struct r5conf *conf = sh->raid_conf;
3444         int i;
3445         int prexor;
3446         int disks = sh->disks;
3447         struct r5dev *pdev, *qdev;
3448
3449         clear_bit(STRIPE_HANDLE, &sh->state);
3450         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3451                 /* already being handled, ensure it gets handled
3452                  * again when current action finishes */
3453                 set_bit(STRIPE_HANDLE, &sh->state);
3454                 return;
3455         }
3456
3457         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3458                 spin_lock(&sh->stripe_lock);
3459                 /* Cannot process 'sync' concurrently with 'discard' */
3460                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3461                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3462                         set_bit(STRIPE_SYNCING, &sh->state);
3463                         clear_bit(STRIPE_INSYNC, &sh->state);
3464                 }
3465                 spin_unlock(&sh->stripe_lock);
3466         }
3467         clear_bit(STRIPE_DELAYED, &sh->state);
3468
3469         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3470                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3471                (unsigned long long)sh->sector, sh->state,
3472                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3473                sh->check_state, sh->reconstruct_state);
3474
3475         analyse_stripe(sh, &s);
3476
3477         if (s.handle_bad_blocks) {
3478                 set_bit(STRIPE_HANDLE, &sh->state);
3479                 goto finish;
3480         }
3481
3482         if (unlikely(s.blocked_rdev)) {
3483                 if (s.syncing || s.expanding || s.expanded ||
3484                     s.replacing || s.to_write || s.written) {
3485                         set_bit(STRIPE_HANDLE, &sh->state);
3486                         goto finish;
3487                 }
3488                 /* There is nothing for the blocked_rdev to block */
3489                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3490                 s.blocked_rdev = NULL;
3491         }
3492
3493         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3494                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3495                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3496         }
3497
3498         pr_debug("locked=%d uptodate=%d to_read=%d"
3499                " to_write=%d failed=%d failed_num=%d,%d\n",
3500                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3501                s.failed_num[0], s.failed_num[1]);
3502         /* check if the array has lost more than max_degraded devices and,
3503          * if so, some requests might need to be failed.
3504          */
3505         if (s.failed > conf->max_degraded) {
3506                 sh->check_state = 0;
3507                 sh->reconstruct_state = 0;
3508                 if (s.to_read+s.to_write+s.written)
3509                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3510                 if (s.syncing + s.replacing)
3511                         handle_failed_sync(conf, sh, &s);
3512         }
3513
3514         /* Now we check to see if any write operations have recently
3515          * completed
3516          */
3517         prexor = 0;
3518         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3519                 prexor = 1;
3520         if (sh->reconstruct_state == reconstruct_state_drain_result ||
3521             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3522                 sh->reconstruct_state = reconstruct_state_idle;
3523
3524                 /* All the 'written' buffers and the parity block are ready to
3525                  * be written back to disk
3526                  */
3527                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3528                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3529                 BUG_ON(sh->qd_idx >= 0 &&
3530                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3531                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3532                 for (i = disks; i--; ) {
3533                         struct r5dev *dev = &sh->dev[i];
3534                         if (test_bit(R5_LOCKED, &dev->flags) &&
3535                                 (i == sh->pd_idx || i == sh->qd_idx ||
3536                                  dev->written)) {
3537                                 pr_debug("Writing block %d\n", i);
3538                                 set_bit(R5_Wantwrite, &dev->flags);
3539                                 if (prexor)
3540                                         continue;
3541                                 if (!test_bit(R5_Insync, &dev->flags) ||
3542                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
3543                                      s.failed == 0))
3544                                         set_bit(STRIPE_INSYNC, &sh->state);
3545                         }
3546                 }
3547                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3548                         s.dec_preread_active = 1;
3549         }
3550
3551         /*
3552          * might be able to return some write requests if the parity blocks
3553          * are safe, or on a failed drive
3554          */
3555         pdev = &sh->dev[sh->pd_idx];
3556         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3557                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3558         qdev = &sh->dev[sh->qd_idx];
3559         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3560                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3561                 || conf->level < 6;
3562
3563         if (s.written &&
3564             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3565                              && !test_bit(R5_LOCKED, &pdev->flags)
3566                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
3567                                  test_bit(R5_Discard, &pdev->flags))))) &&
3568             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3569                              && !test_bit(R5_LOCKED, &qdev->flags)
3570                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
3571                                  test_bit(R5_Discard, &qdev->flags))))))
3572                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3573
3574         /* Now we might consider reading some blocks, either to check/generate
3575          * parity, or to satisfy requests
3576          * or to load a block that is being partially written.
3577          */
3578         if (s.to_read || s.non_overwrite
3579             || (conf->level == 6 && s.to_write && s.failed)
3580             || (s.syncing && (s.uptodate + s.compute < disks))
3581             || s.replacing
3582             || s.expanding)
3583                 handle_stripe_fill(sh, &s, disks);
3584
3585         /* Now to consider new write requests and what else, if anything
3586          * should be read.  We do not handle new writes when:
3587          * 1/ A 'write' operation (copy+xor) is already in flight.
3588          * 2/ A 'check' operation is in flight, as it may clobber the parity
3589          *    block.
3590          */
3591         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3592                 handle_stripe_dirtying(conf, sh, &s, disks);
3593
3594         /* maybe we need to check and possibly fix the parity for this stripe
3595          * Any reads will already have been scheduled, so we just see if enough
3596          * data is available.  The parity check is held off while parity
3597          * dependent operations are in flight.
3598          */
3599         if (sh->check_state ||
3600             (s.syncing && s.locked == 0 &&
3601              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3602              !test_bit(STRIPE_INSYNC, &sh->state))) {
3603                 if (conf->level == 6)
3604                         handle_parity_checks6(conf, sh, &s, disks);
3605                 else
3606                         handle_parity_checks5(conf, sh, &s, disks);
3607         }
3608
3609         if (s.replacing && s.locked == 0
3610             && !test_bit(STRIPE_INSYNC, &sh->state)) {
3611                 /* Write out to replacement devices where possible */
3612                 for (i = 0; i < conf->raid_disks; i++)
3613                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3614                             test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3615                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
3616                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
3617                                 s.locked++;
3618                         }
3619                 set_bit(STRIPE_INSYNC, &sh->state);
3620         }
3621         if ((s.syncing || s.replacing) && s.locked == 0 &&
3622             test_bit(STRIPE_INSYNC, &sh->state)) {
3623                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3624                 clear_bit(STRIPE_SYNCING, &sh->state);
3625                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3626                         wake_up(&conf->wait_for_overlap);
3627         }
3628
3629         /* If the failed drives are just a ReadError, then we might need
3630          * to progress the repair/check process
3631          */
3632         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3633                 for (i = 0; i < s.failed; i++) {
3634                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
3635                         if (test_bit(R5_ReadError, &dev->flags)
3636                             && !test_bit(R5_LOCKED, &dev->flags)
3637                             && test_bit(R5_UPTODATE, &dev->flags)
3638                                 ) {
3639                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3640                                         set_bit(R5_Wantwrite, &dev->flags);
3641                                         set_bit(R5_ReWrite, &dev->flags);
3642                                         set_bit(R5_LOCKED, &dev->flags);
3643                                         s.locked++;
3644                                 } else {
3645                                         /* let's read it back */
3646                                         set_bit(R5_Wantread, &dev->flags);
3647                                         set_bit(R5_LOCKED, &dev->flags);
3648                                         s.locked++;
3649                                 }
3650                         }
3651                 }
3652
3653
3654         /* Finish reconstruct operations initiated by the expansion process */
3655         if (sh->reconstruct_state == reconstruct_state_result) {
3656                 struct stripe_head *sh_src
3657                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3658                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3659                         /* sh cannot be written until sh_src has been read.
3660                          * so arrange for sh to be delayed a little
3661                          */
3662                         set_bit(STRIPE_DELAYED, &sh->state);
3663                         set_bit(STRIPE_HANDLE, &sh->state);
3664                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3665                                               &sh_src->state))
3666                                 atomic_inc(&conf->preread_active_stripes);
3667                         release_stripe(sh_src);
3668                         goto finish;
3669                 }
3670                 if (sh_src)
3671                         release_stripe(sh_src);
3672
3673                 sh->reconstruct_state = reconstruct_state_idle;
3674                 clear_bit(STRIPE_EXPANDING, &sh->state);
3675                 for (i = conf->raid_disks; i--; ) {
3676                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3677                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3678                         s.locked++;
3679                 }
3680         }
3681
3682         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3683             !sh->reconstruct_state) {
3684                 /* Need to write out all blocks after computing parity */
3685                 sh->disks = conf->raid_disks;
3686                 stripe_set_idx(sh->sector, conf, 0, sh);
3687                 schedule_reconstruction(sh, &s, 1, 1);
3688         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3689                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3690                 atomic_dec(&conf->reshape_stripes);
3691                 wake_up(&conf->wait_for_overlap);
3692                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3693         }
3694
3695         if (s.expanding && s.locked == 0 &&
3696             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3697                 handle_stripe_expansion(conf, sh);
3698
3699 finish:
3700         /* wait for this device to become unblocked */
3701         if (unlikely(s.blocked_rdev)) {
3702                 if (conf->mddev->external)
3703                         md_wait_for_blocked_rdev(s.blocked_rdev,
3704                                                  conf->mddev);
3705                 else
3706                         /* Internal metadata will immediately
3707                          * be written by raid5d, so we don't
3708                          * need to wait here.
3709                          */
3710                         rdev_dec_pending(s.blocked_rdev,
3711                                          conf->mddev);
3712         }
3713
3714         if (s.handle_bad_blocks)
3715                 for (i = disks; i--; ) {
3716                         struct md_rdev *rdev;
3717                         struct r5dev *dev = &sh->dev[i];
3718                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3719                                 /* We own a safe reference to the rdev */
3720                                 rdev = conf->disks[i].rdev;
3721                                 if (!rdev_set_badblocks(rdev, sh->sector,
3722                                                         STRIPE_SECTORS, 0))
3723                                         md_error(conf->mddev, rdev);
3724                                 rdev_dec_pending(rdev, conf->mddev);
3725                         }
3726                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3727                                 rdev = conf->disks[i].rdev;
3728                                 rdev_clear_badblocks(rdev, sh->sector,
3729                                                      STRIPE_SECTORS, 0);
3730                                 rdev_dec_pending(rdev, conf->mddev);
3731                         }
3732                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3733                                 rdev = conf->disks[i].replacement;
3734                                 if (!rdev)
3735                                         /* rdev have been moved down */
3736                                         rdev = conf->disks[i].rdev;
3737                                 rdev_clear_badblocks(rdev, sh->sector,
3738                                                      STRIPE_SECTORS, 0);
3739                                 rdev_dec_pending(rdev, conf->mddev);
3740                         }
3741                 }
3742
3743         if (s.ops_request)
3744                 raid_run_ops(sh, s.ops_request);
3745
3746         ops_run_io(sh, &s);
3747
3748         if (s.dec_preread_active) {
3749                 /* We delay this until after ops_run_io so that if make_request
3750                  * is waiting on a flush, it won't continue until the writes
3751                  * have actually been submitted.
3752                  */
3753                 atomic_dec(&conf->preread_active_stripes);
3754                 if (atomic_read(&conf->preread_active_stripes) <
3755                     IO_THRESHOLD)
3756                         md_wakeup_thread(conf->mddev->thread);
3757         }
3758
3759         return_io(s.return_bi);
3760
3761         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3762 }
3763
3764 static void raid5_activate_delayed(struct r5conf *conf)
3765 {
3766         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3767                 while (!list_empty(&conf->delayed_list)) {
3768                         struct list_head *l = conf->delayed_list.next;
3769                         struct stripe_head *sh;
3770                         sh = list_entry(l, struct stripe_head, lru);
3771                         list_del_init(l);
3772                         clear_bit(STRIPE_DELAYED, &sh->state);
3773                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3774                                 atomic_inc(&conf->preread_active_stripes);
3775                         list_add_tail(&sh->lru, &conf->hold_list);
3776                 }
3777         }
3778 }
3779
3780 static void activate_bit_delay(struct r5conf *conf)
3781 {
3782         /* device_lock is held */
3783         struct list_head head;
3784         list_add(&head, &conf->bitmap_list);
3785         list_del_init(&conf->bitmap_list);
3786         while (!list_empty(&head)) {
3787                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3788                 list_del_init(&sh->lru);
3789                 atomic_inc(&sh->count);
3790                 __release_stripe(conf, sh);
3791         }
3792 }
3793
3794 int md_raid5_congested(struct mddev *mddev, int bits)
3795 {
3796         struct r5conf *conf = mddev->private;
3797
3798         /* No difference between reads and writes.  Just check
3799          * how busy the stripe_cache is
3800          */
3801
3802         if (conf->inactive_blocked)
3803                 return 1;
3804         if (conf->quiesce)
3805                 return 1;
3806         if (list_empty_careful(&conf->inactive_list))
3807                 return 1;
3808
3809         return 0;
3810 }
3811 EXPORT_SYMBOL_GPL(md_raid5_congested);
3812
3813 static int raid5_congested(void *data, int bits)
3814 {
3815         struct mddev *mddev = data;
3816
3817         return mddev_congested(mddev, bits) ||
3818                 md_raid5_congested(mddev, bits);
3819 }
3820
3821 /* We want read requests to align with chunks where possible,
3822  * but write requests don't need to.
3823  */
3824 static int raid5_mergeable_bvec(struct request_queue *q,
3825                                 struct bvec_merge_data *bvm,
3826                                 struct bio_vec *biovec)
3827 {
3828         struct mddev *mddev = q->queuedata;
3829         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3830         int max;
3831         unsigned int chunk_sectors = mddev->chunk_sectors;
3832         unsigned int bio_sectors = bvm->bi_size >> 9;
3833
3834         if ((bvm->bi_rw & 1) == WRITE)
3835                 return biovec->bv_len; /* always allow writes to be mergeable */
3836
3837         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3838                 chunk_sectors = mddev->new_chunk_sectors;
3839         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3840         if (max < 0) max = 0;
3841         if (max <= biovec->bv_len && bio_sectors == 0)
3842                 return biovec->bv_len;
3843         else
3844                 return max;
3845 }
3846
3847
3848 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3849 {
3850         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3851         unsigned int chunk_sectors = mddev->chunk_sectors;
3852         unsigned int bio_sectors = bio->bi_size >> 9;
3853
3854         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3855                 chunk_sectors = mddev->new_chunk_sectors;
3856         return  chunk_sectors >=
3857                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3858 }
3859
3860 /*
3861  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3862  *  later sampled by raid5d.
3863  */
3864 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3865 {
3866         unsigned long flags;
3867
3868         spin_lock_irqsave(&conf->device_lock, flags);
3869
3870         bi->bi_next = conf->retry_read_aligned_list;
3871         conf->retry_read_aligned_list = bi;
3872
3873         spin_unlock_irqrestore(&conf->device_lock, flags);
3874         md_wakeup_thread(conf->mddev->thread);
3875 }
3876
3877
3878 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3879 {
3880         struct bio *bi;
3881
3882         bi = conf->retry_read_aligned;
3883         if (bi) {
3884                 conf->retry_read_aligned = NULL;
3885                 return bi;
3886         }
3887         bi = conf->retry_read_aligned_list;
3888         if(bi) {
3889                 conf->retry_read_aligned_list = bi->bi_next;
3890                 bi->bi_next = NULL;
3891                 /*
3892                  * this sets the active strip count to 1 and the processed
3893                  * strip count to zero (upper 8 bits)
3894                  */
3895                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3896         }
3897
3898         return bi;
3899 }
3900
3901
3902 /*
3903  *  The "raid5_align_endio" should check if the read succeeded and if it
3904  *  did, call bio_endio on the original bio (having bio_put the new bio
3905  *  first).
3906  *  If the read failed..
3907  */
3908 static void raid5_align_endio(struct bio *bi, int error)
3909 {
3910         struct bio* raid_bi  = bi->bi_private;
3911         struct mddev *mddev;
3912         struct r5conf *conf;
3913         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3914         struct md_rdev *rdev;
3915
3916         bio_put(bi);
3917
3918         rdev = (void*)raid_bi->bi_next;
3919         raid_bi->bi_next = NULL;
3920         mddev = rdev->mddev;
3921         conf = mddev->private;
3922
3923         rdev_dec_pending(rdev, conf->mddev);
3924
3925         if (!error && uptodate) {
3926                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
3927                                          raid_bi, 0);
3928                 bio_endio(raid_bi, 0);
3929                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3930                         wake_up(&conf->wait_for_stripe);
3931                 return;
3932         }
3933
3934
3935         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3936
3937         add_bio_to_retry(raid_bi, conf);
3938 }
3939
3940 static int bio_fits_rdev(struct bio *bi)
3941 {
3942         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3943
3944         if ((bi->bi_size>>9) > queue_max_sectors(q))
3945                 return 0;
3946         blk_recount_segments(q, bi);
3947         if (bi->bi_phys_segments > queue_max_segments(q))
3948                 return 0;
3949
3950         if (q->merge_bvec_fn)
3951                 /* it's too hard to apply the merge_bvec_fn at this stage,
3952                  * just just give up
3953                  */
3954                 return 0;
3955
3956         return 1;
3957 }
3958
3959
3960 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3961 {
3962         struct r5conf *conf = mddev->private;
3963         int dd_idx;
3964         struct bio* align_bi;
3965         struct md_rdev *rdev;
3966         sector_t end_sector;
3967
3968         if (!in_chunk_boundary(mddev, raid_bio)) {
3969                 pr_debug("chunk_aligned_read : non aligned\n");
3970                 return 0;
3971         }
3972         /*
3973          * use bio_clone_mddev to make a copy of the bio
3974          */
3975         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3976         if (!align_bi)
3977                 return 0;
3978         /*
3979          *   set bi_end_io to a new function, and set bi_private to the
3980          *     original bio.
3981          */
3982         align_bi->bi_end_io  = raid5_align_endio;
3983         align_bi->bi_private = raid_bio;
3984         /*
3985          *      compute position
3986          */
3987         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3988                                                     0,
3989                                                     &dd_idx, NULL);
3990
3991         end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3992         rcu_read_lock();
3993         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3994         if (!rdev || test_bit(Faulty, &rdev->flags) ||
3995             rdev->recovery_offset < end_sector) {
3996                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3997                 if (rdev &&
3998                     (test_bit(Faulty, &rdev->flags) ||
3999                     !(test_bit(In_sync, &rdev->flags) ||
4000                       rdev->recovery_offset >= end_sector)))
4001                         rdev = NULL;
4002         }
4003         if (rdev) {
4004                 sector_t first_bad;
4005                 int bad_sectors;
4006
4007                 atomic_inc(&rdev->nr_pending);
4008                 rcu_read_unlock();
4009                 raid_bio->bi_next = (void*)rdev;
4010                 align_bi->bi_bdev =  rdev->bdev;
4011                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4012
4013                 if (!bio_fits_rdev(align_bi) ||
4014                     is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
4015                                 &first_bad, &bad_sectors)) {
4016                         /* too big in some way, or has a known bad block */
4017                         bio_put(align_bi);
4018                         rdev_dec_pending(rdev, mddev);
4019                         return 0;
4020                 }
4021
4022                 /* No reshape active, so we can trust rdev->data_offset */
4023                 align_bi->bi_sector += rdev->data_offset;
4024
4025                 spin_lock_irq(&conf->device_lock);
4026                 wait_event_lock_irq(conf->wait_for_stripe,
4027                                     conf->quiesce == 0,
4028                                     conf->device_lock);
4029                 atomic_inc(&conf->active_aligned_reads);
4030                 spin_unlock_irq(&conf->device_lock);
4031
4032                 if (mddev->gendisk)
4033                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4034                                               align_bi, disk_devt(mddev->gendisk),
4035                                               raid_bio->bi_sector);
4036                 generic_make_request(align_bi);
4037                 return 1;
4038         } else {
4039                 rcu_read_unlock();
4040                 bio_put(align_bi);
4041                 return 0;
4042         }
4043 }
4044
4045 /* __get_priority_stripe - get the next stripe to process
4046  *
4047  * Full stripe writes are allowed to pass preread active stripes up until
4048  * the bypass_threshold is exceeded.  In general the bypass_count
4049  * increments when the handle_list is handled before the hold_list; however, it
4050  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4051  * stripe with in flight i/o.  The bypass_count will be reset when the
4052  * head of the hold_list has changed, i.e. the head was promoted to the
4053  * handle_list.
4054  */
4055 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
4056 {
4057         struct stripe_head *sh;
4058
4059         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4060                   __func__,
4061                   list_empty(&conf->handle_list) ? "empty" : "busy",
4062                   list_empty(&conf->hold_list) ? "empty" : "busy",
4063                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
4064
4065         if (!list_empty(&conf->handle_list)) {
4066                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4067
4068                 if (list_empty(&conf->hold_list))
4069                         conf->bypass_count = 0;
4070                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4071                         if (conf->hold_list.next == conf->last_hold)
4072                                 conf->bypass_count++;
4073                         else {
4074                                 conf->last_hold = conf->hold_list.next;
4075                                 conf->bypass_count -= conf->bypass_threshold;
4076                                 if (conf->bypass_count < 0)
4077                                         conf->bypass_count = 0;
4078                         }
4079                 }
4080         } else if (!list_empty(&conf->hold_list) &&
4081                    ((conf->bypass_threshold &&
4082                      conf->bypass_count > conf->bypass_threshold) ||
4083                     atomic_read(&conf->pending_full_writes) == 0)) {
4084                 sh = list_entry(conf->hold_list.next,
4085                                 typeof(*sh), lru);
4086                 conf->bypass_count -= conf->bypass_threshold;
4087                 if (conf->bypass_count < 0)
4088                         conf->bypass_count = 0;
4089         } else
4090                 return NULL;
4091
4092         list_del_init(&sh->lru);
4093         atomic_inc(&sh->count);
4094         BUG_ON(atomic_read(&sh->count) != 1);
4095         return sh;
4096 }
4097
4098 struct raid5_plug_cb {
4099         struct blk_plug_cb      cb;
4100         struct list_head        list;
4101 };
4102
4103 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4104 {
4105         struct raid5_plug_cb *cb = container_of(
4106                 blk_cb, struct raid5_plug_cb, cb);
4107         struct stripe_head *sh;
4108         struct mddev *mddev = cb->cb.data;
4109         struct r5conf *conf = mddev->private;
4110         int cnt = 0;
4111
4112         if (cb->list.next && !list_empty(&cb->list)) {
4113                 spin_lock_irq(&conf->device_lock);
4114                 while (!list_empty(&cb->list)) {
4115                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
4116                         list_del_init(&sh->lru);
4117                         /*
4118                          * avoid race release_stripe_plug() sees
4119                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
4120                          * is still in our list
4121                          */
4122                         smp_mb__before_clear_bit();
4123                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4124                         __release_stripe(conf, sh);
4125                         cnt++;
4126                 }
4127                 spin_unlock_irq(&conf->device_lock);
4128         }
4129         if (mddev->queue)
4130                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4131         kfree(cb);
4132 }
4133
4134 static void release_stripe_plug(struct mddev *mddev,
4135                                 struct stripe_head *sh)
4136 {
4137         struct blk_plug_cb *blk_cb = blk_check_plugged(
4138                 raid5_unplug, mddev,
4139                 sizeof(struct raid5_plug_cb));
4140         struct raid5_plug_cb *cb;
4141
4142         if (!blk_cb) {
4143                 release_stripe(sh);
4144                 return;
4145         }
4146
4147         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4148
4149         if (cb->list.next == NULL)
4150                 INIT_LIST_HEAD(&cb->list);
4151
4152         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4153                 list_add_tail(&sh->lru, &cb->list);
4154         else
4155                 release_stripe(sh);
4156 }
4157
4158 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4159 {
4160         struct r5conf *conf = mddev->private;
4161         sector_t logical_sector, last_sector;
4162         struct stripe_head *sh;
4163         int remaining;
4164         int stripe_sectors;
4165
4166         if (mddev->reshape_position != MaxSector)
4167                 /* Skip discard while reshape is happening */
4168                 return;
4169
4170         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4171         last_sector = bi->bi_sector + (bi->bi_size>>9);
4172
4173         bi->bi_next = NULL;
4174         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4175
4176         stripe_sectors = conf->chunk_sectors *
4177                 (conf->raid_disks - conf->max_degraded);
4178         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4179                                                stripe_sectors);
4180         sector_div(last_sector, stripe_sectors);
4181
4182         logical_sector *= conf->chunk_sectors;
4183         last_sector *= conf->chunk_sectors;
4184
4185         for (; logical_sector < last_sector;
4186              logical_sector += STRIPE_SECTORS) {
4187                 DEFINE_WAIT(w);
4188                 int d;
4189         again:
4190                 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4191                 prepare_to_wait(&conf->wait_for_overlap, &w,
4192                                 TASK_UNINTERRUPTIBLE);
4193                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4194                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4195                         release_stripe(sh);
4196                         schedule();
4197                         goto again;
4198                 }
4199                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4200                 spin_lock_irq(&sh->stripe_lock);
4201                 for (d = 0; d < conf->raid_disks; d++) {
4202                         if (d == sh->pd_idx || d == sh->qd_idx)
4203                                 continue;
4204                         if (sh->dev[d].towrite || sh->dev[d].toread) {
4205                                 set_bit(R5_Overlap, &sh->dev[d].flags);
4206                                 spin_unlock_irq(&sh->stripe_lock);
4207                                 release_stripe(sh);
4208                                 schedule();
4209                                 goto again;
4210                         }
4211                 }
4212                 set_bit(STRIPE_DISCARD, &sh->state);
4213                 finish_wait(&conf->wait_for_overlap, &w);
4214                 for (d = 0; d < conf->raid_disks; d++) {
4215                         if (d == sh->pd_idx || d == sh->qd_idx)
4216                                 continue;
4217                         sh->dev[d].towrite = bi;
4218                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4219                         raid5_inc_bi_active_stripes(bi);
4220                 }
4221                 spin_unlock_irq(&sh->stripe_lock);
4222                 if (conf->mddev->bitmap) {
4223                         for (d = 0;
4224                              d < conf->raid_disks - conf->max_degraded;
4225                              d++)
4226                                 bitmap_startwrite(mddev->bitmap,
4227                                                   sh->sector,
4228                                                   STRIPE_SECTORS,
4229                                                   0);
4230                         sh->bm_seq = conf->seq_flush + 1;
4231                         set_bit(STRIPE_BIT_DELAY, &sh->state);
4232                 }
4233
4234                 set_bit(STRIPE_HANDLE, &sh->state);
4235                 clear_bit(STRIPE_DELAYED, &sh->state);
4236                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4237                         atomic_inc(&conf->preread_active_stripes);
4238                 release_stripe_plug(mddev, sh);
4239         }
4240
4241         remaining = raid5_dec_bi_active_stripes(bi);
4242         if (remaining == 0) {
4243                 md_write_end(mddev);
4244                 bio_endio(bi, 0);
4245         }
4246 }
4247
4248 static void make_request(struct mddev *mddev, struct bio * bi)
4249 {
4250         struct r5conf *conf = mddev->private;
4251         int dd_idx;
4252         sector_t new_sector;
4253         sector_t logical_sector, last_sector;
4254         struct stripe_head *sh;
4255         const int rw = bio_data_dir(bi);
4256         int remaining;
4257
4258         if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4259                 md_flush_request(mddev, bi);
4260                 return;
4261         }
4262
4263         md_write_start(mddev, bi);
4264
4265         if (rw == READ &&
4266              mddev->reshape_position == MaxSector &&
4267              chunk_aligned_read(mddev,bi))
4268                 return;
4269
4270         if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4271                 make_discard_request(mddev, bi);
4272                 return;
4273         }
4274
4275         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4276         last_sector = bi->bi_sector + (bi->bi_size>>9);
4277         bi->bi_next = NULL;
4278         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
4279
4280         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4281                 DEFINE_WAIT(w);
4282                 int previous;
4283
4284         retry:
4285                 previous = 0;
4286                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4287                 if (unlikely(conf->reshape_progress != MaxSector)) {
4288                         /* spinlock is needed as reshape_progress may be
4289                          * 64bit on a 32bit platform, and so it might be
4290                          * possible to see a half-updated value
4291                          * Of course reshape_progress could change after
4292                          * the lock is dropped, so once we get a reference
4293                          * to the stripe that we think it is, we will have
4294                          * to check again.
4295                          */
4296                         spin_lock_irq(&conf->device_lock);
4297                         if (mddev->reshape_backwards
4298                             ? logical_sector < conf->reshape_progress
4299                             : logical_sector >= conf->reshape_progress) {
4300                                 previous = 1;
4301                         } else {
4302                                 if (mddev->reshape_backwards
4303                                     ? logical_sector < conf->reshape_safe
4304                                     : logical_sector >= conf->reshape_safe) {
4305                                         spin_unlock_irq(&conf->device_lock);
4306                                         schedule();
4307                                         goto retry;
4308                                 }
4309                         }
4310                         spin_unlock_irq(&conf->device_lock);
4311                 }
4312
4313                 new_sector = raid5_compute_sector(conf, logical_sector,
4314                                                   previous,
4315                                                   &dd_idx, NULL);
4316                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4317                         (unsigned long long)new_sector,
4318                         (unsigned long long)logical_sector);
4319
4320                 sh = get_active_stripe(conf, new_sector, previous,
4321                                        (bi->bi_rw&RWA_MASK), 0);
4322                 if (sh) {
4323                         if (unlikely(previous)) {
4324                                 /* expansion might have moved on while waiting for a
4325                                  * stripe, so we must do the range check again.
4326                                  * Expansion could still move past after this
4327                                  * test, but as we are holding a reference to
4328                                  * 'sh', we know that if that happens,
4329                                  *  STRIPE_EXPANDING will get set and the expansion
4330                                  * won't proceed until we finish with the stripe.
4331                                  */
4332                                 int must_retry = 0;
4333                                 spin_lock_irq(&conf->device_lock);
4334                                 if (mddev->reshape_backwards
4335                                     ? logical_sector >= conf->reshape_progress
4336                                     : logical_sector < conf->reshape_progress)
4337                                         /* mismatch, need to try again */
4338                                         must_retry = 1;
4339                                 spin_unlock_irq(&conf->device_lock);
4340                                 if (must_retry) {
4341                                         release_stripe(sh);
4342                                         schedule();
4343                                         goto retry;
4344                                 }
4345                         }
4346
4347                         if (rw == WRITE &&
4348                             logical_sector >= mddev->suspend_lo &&
4349                             logical_sector < mddev->suspend_hi) {
4350                                 release_stripe(sh);
4351                                 /* As the suspend_* range is controlled by
4352                                  * userspace, we want an interruptible
4353                                  * wait.
4354                                  */
4355                                 flush_signals(current);
4356                                 prepare_to_wait(&conf->wait_for_overlap,
4357                                                 &w, TASK_INTERRUPTIBLE);
4358                                 if (logical_sector >= mddev->suspend_lo &&
4359                                     logical_sector < mddev->suspend_hi)
4360                                         schedule();
4361                                 goto retry;
4362                         }
4363
4364                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4365                             !add_stripe_bio(sh, bi, dd_idx, rw)) {
4366                                 /* Stripe is busy expanding or
4367                                  * add failed due to overlap.  Flush everything
4368                                  * and wait a while
4369                                  */
4370                                 md_wakeup_thread(mddev->thread);
4371                                 release_stripe(sh);
4372                                 schedule();
4373                                 goto retry;
4374                         }
4375                         finish_wait(&conf->wait_for_overlap, &w);
4376                         set_bit(STRIPE_HANDLE, &sh->state);
4377                         clear_bit(STRIPE_DELAYED, &sh->state);
4378                         if ((bi->bi_rw & REQ_SYNC) &&
4379                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4380                                 atomic_inc(&conf->preread_active_stripes);
4381                         release_stripe_plug(mddev, sh);
4382                 } else {
4383                         /* cannot get stripe for read-ahead, just give-up */
4384                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
4385                         finish_wait(&conf->wait_for_overlap, &w);
4386                         break;
4387                 }
4388         }
4389
4390         remaining = raid5_dec_bi_active_stripes(bi);
4391         if (remaining == 0) {
4392
4393                 if ( rw == WRITE )
4394                         md_write_end(mddev);
4395
4396                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4397                                          bi, 0);
4398                 bio_endio(bi, 0);
4399         }
4400 }
4401
4402 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4403
4404 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4405 {
4406         /* reshaping is quite different to recovery/resync so it is
4407          * handled quite separately ... here.
4408          *
4409          * On each call to sync_request, we gather one chunk worth of
4410          * destination stripes and flag them as expanding.
4411          * Then we find all the source stripes and request reads.
4412          * As the reads complete, handle_stripe will copy the data
4413          * into the destination stripe and release that stripe.
4414          */
4415         struct r5conf *conf = mddev->private;
4416         struct stripe_head *sh;
4417         sector_t first_sector, last_sector;
4418         int raid_disks = conf->previous_raid_disks;
4419         int data_disks = raid_disks - conf->max_degraded;
4420         int new_data_disks = conf->raid_disks - conf->max_degraded;
4421         int i;
4422         int dd_idx;
4423         sector_t writepos, readpos, safepos;
4424         sector_t stripe_addr;
4425         int reshape_sectors;
4426         struct list_head stripes;
4427
4428         if (sector_nr == 0) {
4429                 /* If restarting in the middle, skip the initial sectors */
4430                 if (mddev->reshape_backwards &&
4431                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4432                         sector_nr = raid5_size(mddev, 0, 0)
4433                                 - conf->reshape_progress;
4434                 } else if (!mddev->reshape_backwards &&
4435                            conf->reshape_progress > 0)
4436                         sector_nr = conf->reshape_progress;
4437                 sector_div(sector_nr, new_data_disks);
4438                 if (sector_nr) {
4439                         mddev->curr_resync_completed = sector_nr;
4440                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4441                         *skipped = 1;
4442                         return sector_nr;
4443                 }
4444         }
4445
4446         /* We need to process a full chunk at a time.
4447          * If old and new chunk sizes differ, we need to process the
4448          * largest of these
4449          */
4450         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4451                 reshape_sectors = mddev->new_chunk_sectors;
4452         else
4453                 reshape_sectors = mddev->chunk_sectors;
4454
4455         /* We update the metadata at least every 10 seconds, or when
4456          * the data about to be copied would over-write the source of
4457          * the data at the front of the range.  i.e. one new_stripe
4458          * along from reshape_progress new_maps to after where
4459          * reshape_safe old_maps to
4460          */
4461         writepos = conf->reshape_progress;
4462         sector_div(writepos, new_data_disks);
4463         readpos = conf->reshape_progress;
4464         sector_div(readpos, data_disks);
4465         safepos = conf->reshape_safe;
4466         sector_div(safepos, data_disks);
4467         if (mddev->reshape_backwards) {
4468                 writepos -= min_t(sector_t, reshape_sectors, writepos);
4469                 readpos += reshape_sectors;
4470                 safepos += reshape_sectors;
4471         } else {
4472                 writepos += reshape_sectors;
4473                 readpos -= min_t(sector_t, reshape_sectors, readpos);
4474                 safepos -= min_t(sector_t, reshape_sectors, safepos);
4475         }
4476
4477         /* Having calculated the 'writepos' possibly use it
4478          * to set 'stripe_addr' which is where we will write to.
4479          */
4480         if (mddev->reshape_backwards) {
4481                 BUG_ON(conf->reshape_progress == 0);
4482                 stripe_addr = writepos;
4483                 BUG_ON((mddev->dev_sectors &
4484                         ~((sector_t)reshape_sectors - 1))
4485                        - reshape_sectors - stripe_addr
4486                        != sector_nr);
4487         } else {
4488                 BUG_ON(writepos != sector_nr + reshape_sectors);
4489                 stripe_addr = sector_nr;
4490         }
4491
4492         /* 'writepos' is the most advanced device address we might write.
4493          * 'readpos' is the least advanced device address we might read.
4494          * 'safepos' is the least address recorded in the metadata as having
4495          *     been reshaped.
4496          * If there is a min_offset_diff, these are adjusted either by
4497          * increasing the safepos/readpos if diff is negative, or
4498          * increasing writepos if diff is positive.
4499          * If 'readpos' is then behind 'writepos', there is no way that we can
4500          * ensure safety in the face of a crash - that must be done by userspace
4501          * making a backup of the data.  So in that case there is no particular
4502          * rush to update metadata.
4503          * Otherwise if 'safepos' is behind 'writepos', then we really need to
4504          * update the metadata to advance 'safepos' to match 'readpos' so that
4505          * we can be safe in the event of a crash.
4506          * So we insist on updating metadata if safepos is behind writepos and
4507          * readpos is beyond writepos.
4508          * In any case, update the metadata every 10 seconds.
4509          * Maybe that number should be configurable, but I'm not sure it is
4510          * worth it.... maybe it could be a multiple of safemode_delay???
4511          */
4512         if (conf->min_offset_diff < 0) {
4513                 safepos += -conf->min_offset_diff;
4514                 readpos += -conf->min_offset_diff;
4515         } else
4516                 writepos += conf->min_offset_diff;
4517
4518         if ((mddev->reshape_backwards
4519              ? (safepos > writepos && readpos < writepos)
4520              : (safepos < writepos && readpos > writepos)) ||
4521             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4522                 /* Cannot proceed until we've updated the superblock... */
4523                 wait_event(conf->wait_for_overlap,
4524                            atomic_read(&conf->reshape_stripes)==0);
4525                 mddev->reshape_position = conf->reshape_progress;
4526                 mddev->curr_resync_completed = sector_nr;
4527                 conf->reshape_checkpoint = jiffies;
4528                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4529                 md_wakeup_thread(mddev->thread);
4530                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4531                            kthread_should_stop());
4532                 spin_lock_irq(&conf->device_lock);
4533                 conf->reshape_safe = mddev->reshape_position;
4534                 spin_unlock_irq(&conf->device_lock);
4535                 wake_up(&conf->wait_for_overlap);
4536                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4537         }
4538
4539         INIT_LIST_HEAD(&stripes);
4540         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4541                 int j;
4542                 int skipped_disk = 0;
4543                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4544                 set_bit(STRIPE_EXPANDING, &sh->state);
4545                 atomic_inc(&conf->reshape_stripes);
4546                 /* If any of this stripe is beyond the end of the old
4547                  * array, then we need to zero those blocks
4548                  */
4549                 for (j=sh->disks; j--;) {
4550                         sector_t s;
4551                         if (j == sh->pd_idx)
4552                                 continue;
4553                         if (conf->level == 6 &&
4554                             j == sh->qd_idx)
4555                                 continue;
4556                         s = compute_blocknr(sh, j, 0);
4557                         if (s < raid5_size(mddev, 0, 0)) {
4558                                 skipped_disk = 1;
4559                                 continue;
4560                         }
4561                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4562                         set_bit(R5_Expanded, &sh->dev[j].flags);
4563                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
4564                 }
4565                 if (!skipped_disk) {
4566                         set_bit(STRIPE_EXPAND_READY, &sh->state);
4567                         set_bit(STRIPE_HANDLE, &sh->state);
4568                 }
4569                 list_add(&sh->lru, &stripes);
4570         }
4571         spin_lock_irq(&conf->device_lock);
4572         if (mddev->reshape_backwards)
4573                 conf->reshape_progress -= reshape_sectors * new_data_disks;
4574         else
4575                 conf->reshape_progress += reshape_sectors * new_data_disks;
4576         spin_unlock_irq(&conf->device_lock);
4577         /* Ok, those stripe are ready. We can start scheduling
4578          * reads on the source stripes.
4579          * The source stripes are determined by mapping the first and last
4580          * block on the destination stripes.
4581          */
4582         first_sector =
4583                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4584                                      1, &dd_idx, NULL);
4585         last_sector =
4586                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4587                                             * new_data_disks - 1),
4588                                      1, &dd_idx, NULL);
4589         if (last_sector >= mddev->dev_sectors)
4590                 last_sector = mddev->dev_sectors - 1;
4591         while (first_sector <= last_sector) {
4592                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4593                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4594                 set_bit(STRIPE_HANDLE, &sh->state);
4595                 release_stripe(sh);
4596                 first_sector += STRIPE_SECTORS;
4597         }
4598         /* Now that the sources are clearly marked, we can release
4599          * the destination stripes
4600          */
4601         while (!list_empty(&stripes)) {
4602                 sh = list_entry(stripes.next, struct stripe_head, lru);
4603                 list_del_init(&sh->lru);
4604                 release_stripe(sh);
4605         }
4606         /* If this takes us to the resync_max point where we have to pause,
4607          * then we need to write out the superblock.
4608          */
4609         sector_nr += reshape_sectors;
4610         if ((sector_nr - mddev->curr_resync_completed) * 2
4611             >= mddev->resync_max - mddev->curr_resync_completed) {
4612                 /* Cannot proceed until we've updated the superblock... */
4613                 wait_event(conf->wait_for_overlap,
4614                            atomic_read(&conf->reshape_stripes) == 0);
4615                 mddev->reshape_position = conf->reshape_progress;
4616                 mddev->curr_resync_completed = sector_nr;
4617                 conf->reshape_checkpoint = jiffies;
4618                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4619                 md_wakeup_thread(mddev->thread);
4620                 wait_event(mddev->sb_wait,
4621                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4622                            || kthread_should_stop());
4623                 spin_lock_irq(&conf->device_lock);
4624                 conf->reshape_safe = mddev->reshape_position;
4625                 spin_unlock_irq(&conf->device_lock);
4626                 wake_up(&conf->wait_for_overlap);
4627                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4628         }
4629         return reshape_sectors;
4630 }
4631
4632 /* FIXME go_faster isn't used */
4633 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4634 {
4635         struct r5conf *conf = mddev->private;
4636         struct stripe_head *sh;
4637         sector_t max_sector = mddev->dev_sectors;
4638         sector_t sync_blocks;
4639         int still_degraded = 0;
4640         int i;
4641
4642         if (sector_nr >= max_sector) {
4643                 /* just being told to finish up .. nothing much to do */
4644
4645                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4646                         end_reshape(conf);
4647                         return 0;
4648                 }
4649
4650                 if (mddev->curr_resync < max_sector) /* aborted */
4651                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4652                                         &sync_blocks, 1);
4653                 else /* completed sync */
4654                         conf->fullsync = 0;
4655                 bitmap_close_sync(mddev->bitmap);
4656
4657                 return 0;
4658         }
4659
4660         /* Allow raid5_quiesce to complete */
4661         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4662
4663         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4664                 return reshape_request(mddev, sector_nr, skipped);
4665
4666         /* No need to check resync_max as we never do more than one
4667          * stripe, and as resync_max will always be on a chunk boundary,
4668          * if the check in md_do_sync didn't fire, there is no chance
4669          * of overstepping resync_max here
4670          */
4671
4672         /* if there is too many failed drives and we are trying
4673          * to resync, then assert that we are finished, because there is
4674          * nothing we can do.
4675          */
4676         if (mddev->degraded >= conf->max_degraded &&
4677             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4678                 sector_t rv = mddev->dev_sectors - sector_nr;
4679                 *skipped = 1;
4680                 return rv;
4681         }
4682         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4683             !conf->fullsync &&
4684             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4685             sync_blocks >= STRIPE_SECTORS) {
4686                 /* we can skip this block, and probably more */
4687                 sync_blocks /= STRIPE_SECTORS;
4688                 *skipped = 1;
4689                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4690         }
4691
4692         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4693
4694         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4695         if (sh == NULL) {
4696                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4697                 /* make sure we don't swamp the stripe cache if someone else
4698                  * is trying to get access
4699                  */
4700                 schedule_timeout_uninterruptible(1);
4701         }
4702         /* Need to check if array will still be degraded after recovery/resync
4703          * We don't need to check the 'failed' flag as when that gets set,
4704          * recovery aborts.
4705          */
4706         for (i = 0; i < conf->raid_disks; i++)
4707                 if (conf->disks[i].rdev == NULL)
4708                         still_degraded = 1;
4709
4710         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4711
4712         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4713
4714         handle_stripe(sh);
4715         release_stripe(sh);
4716
4717         return STRIPE_SECTORS;
4718 }
4719
4720 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4721 {
4722         /* We may not be able to submit a whole bio at once as there
4723          * may not be enough stripe_heads available.
4724          * We cannot pre-allocate enough stripe_heads as we may need
4725          * more than exist in the cache (if we allow ever large chunks).
4726          * So we do one stripe head at a time and record in
4727          * ->bi_hw_segments how many have been done.
4728          *
4729          * We *know* that this entire raid_bio is in one chunk, so
4730          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4731          */
4732         struct stripe_head *sh;
4733         int dd_idx;
4734         sector_t sector, logical_sector, last_sector;
4735         int scnt = 0;
4736         int remaining;
4737         int handled = 0;
4738
4739         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4740         sector = raid5_compute_sector(conf, logical_sector,
4741                                       0, &dd_idx, NULL);
4742         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4743
4744         for (; logical_sector < last_sector;
4745              logical_sector += STRIPE_SECTORS,
4746                      sector += STRIPE_SECTORS,
4747                      scnt++) {
4748
4749                 if (scnt < raid5_bi_processed_stripes(raid_bio))
4750                         /* already done this stripe */
4751                         continue;
4752
4753                 sh = get_active_stripe(conf, sector, 0, 1, 0);
4754
4755                 if (!sh) {
4756                         /* failed to get a stripe - must wait */
4757                         raid5_set_bi_processed_stripes(raid_bio, scnt);
4758                         conf->retry_read_aligned = raid_bio;
4759                         return handled;
4760                 }
4761
4762                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4763                         release_stripe(sh);
4764                         raid5_set_bi_processed_stripes(raid_bio, scnt);
4765                         conf->retry_read_aligned = raid_bio;
4766                         return handled;
4767                 }
4768
4769                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4770                 handle_stripe(sh);
4771                 release_stripe(sh);
4772                 handled++;
4773         }
4774         remaining = raid5_dec_bi_active_stripes(raid_bio);
4775         if (remaining == 0) {
4776                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4777                                          raid_bio, 0);
4778                 bio_endio(raid_bio, 0);
4779         }
4780         if (atomic_dec_and_test(&conf->active_aligned_reads))
4781                 wake_up(&conf->wait_for_stripe);
4782         return handled;
4783 }
4784
4785 #define MAX_STRIPE_BATCH 8
4786 static int handle_active_stripes(struct r5conf *conf)
4787 {
4788         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4789         int i, batch_size = 0;
4790
4791         while (batch_size < MAX_STRIPE_BATCH &&
4792                         (sh = __get_priority_stripe(conf)) != NULL)
4793                 batch[batch_size++] = sh;
4794
4795         if (batch_size == 0)
4796                 return batch_size;
4797         spin_unlock_irq(&conf->device_lock);
4798
4799         for (i = 0; i < batch_size; i++)
4800                 handle_stripe(batch[i]);
4801
4802         cond_resched();
4803
4804         spin_lock_irq(&conf->device_lock);
4805         for (i = 0; i < batch_size; i++)
4806                 __release_stripe(conf, batch[i]);
4807         return batch_size;
4808 }
4809
4810 /*
4811  * This is our raid5 kernel thread.
4812  *
4813  * We scan the hash table for stripes which can be handled now.
4814  * During the scan, completed stripes are saved for us by the interrupt
4815  * handler, so that they will not have to wait for our next wakeup.
4816  */
4817 static void raid5d(struct md_thread *thread)
4818 {
4819         struct mddev *mddev = thread->mddev;
4820         struct r5conf *conf = mddev->private;
4821         int handled;
4822         struct blk_plug plug;
4823
4824         pr_debug("+++ raid5d active\n");
4825
4826         md_check_recovery(mddev);
4827
4828         blk_start_plug(&plug);
4829         handled = 0;
4830         spin_lock_irq(&conf->device_lock);
4831         while (1) {
4832                 struct bio *bio;
4833                 int batch_size;
4834
4835                 if (
4836                     !list_empty(&conf->bitmap_list)) {
4837                         /* Now is a good time to flush some bitmap updates */
4838                         conf->seq_flush++;
4839                         spin_unlock_irq(&conf->device_lock);
4840                         bitmap_unplug(mddev->bitmap);
4841                         spin_lock_irq(&conf->device_lock);
4842                         conf->seq_write = conf->seq_flush;
4843                         activate_bit_delay(conf);
4844                 }
4845                 raid5_activate_delayed(conf);
4846
4847                 while ((bio = remove_bio_from_retry(conf))) {
4848                         int ok;
4849                         spin_unlock_irq(&conf->device_lock);
4850                         ok = retry_aligned_read(conf, bio);
4851                         spin_lock_irq(&conf->device_lock);
4852                         if (!ok)
4853                                 break;
4854                         handled++;
4855                 }
4856
4857                 batch_size = handle_active_stripes(conf);
4858                 if (!batch_size)
4859                         break;
4860                 handled += batch_size;
4861
4862                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4863                         spin_unlock_irq(&conf->device_lock);
4864                         md_check_recovery(mddev);
4865                         spin_lock_irq(&conf->device_lock);
4866                 }
4867         }
4868         pr_debug("%d stripes handled\n", handled);
4869
4870         spin_unlock_irq(&conf->device_lock);
4871
4872         async_tx_issue_pending_all();
4873         blk_finish_plug(&plug);
4874
4875         pr_debug("--- raid5d inactive\n");
4876 }
4877
4878 static ssize_t
4879 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4880 {
4881         struct r5conf *conf = mddev->private;
4882         if (conf)
4883                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4884         else
4885                 return 0;
4886 }
4887
4888 int
4889 raid5_set_cache_size(struct mddev *mddev, int size)
4890 {
4891         struct r5conf *conf = mddev->private;
4892         int err;
4893
4894         if (size <= 16 || size > 32768)
4895                 return -EINVAL;
4896         while (size < conf->max_nr_stripes) {
4897                 if (drop_one_stripe(conf))
4898                         conf->max_nr_stripes--;
4899                 else
4900                         break;
4901         }
4902         err = md_allow_write(mddev);
4903         if (err)
4904                 return err;
4905         while (size > conf->max_nr_stripes) {
4906                 if (grow_one_stripe(conf))
4907                         conf->max_nr_stripes++;
4908                 else break;
4909         }
4910         return 0;
4911 }
4912 EXPORT_SYMBOL(raid5_set_cache_size);
4913
4914 static ssize_t
4915 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4916 {
4917         struct r5conf *conf = mddev->private;
4918         unsigned long new;
4919         int err;
4920
4921         if (len >= PAGE_SIZE)
4922                 return -EINVAL;
4923         if (!conf)
4924                 return -ENODEV;
4925
4926         if (strict_strtoul(page, 10, &new))
4927                 return -EINVAL;
4928         err = raid5_set_cache_size(mddev, new);
4929         if (err)
4930                 return err;
4931         return len;
4932 }
4933
4934 static struct md_sysfs_entry
4935 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4936                                 raid5_show_stripe_cache_size,
4937                                 raid5_store_stripe_cache_size);
4938
4939 static ssize_t
4940 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4941 {
4942         struct r5conf *conf = mddev->private;
4943         if (conf)
4944                 return sprintf(page, "%d\n", conf->bypass_threshold);
4945         else
4946                 return 0;
4947 }
4948
4949 static ssize_t
4950 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4951 {
4952         struct r5conf *conf = mddev->private;
4953         unsigned long new;
4954         if (len >= PAGE_SIZE)
4955                 return -EINVAL;
4956         if (!conf)
4957                 return -ENODEV;
4958
4959         if (strict_strtoul(page, 10, &new))
4960                 return -EINVAL;
4961         if (new > conf->max_nr_stripes)
4962                 return -EINVAL;
4963         conf->bypass_threshold = new;
4964         return len;
4965 }
4966
4967 static struct md_sysfs_entry
4968 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4969                                         S_IRUGO | S_IWUSR,
4970                                         raid5_show_preread_threshold,
4971                                         raid5_store_preread_threshold);
4972
4973 static ssize_t
4974 stripe_cache_active_show(struct mddev *mddev, char *page)
4975 {
4976         struct r5conf *conf = mddev->private;
4977         if (conf)
4978                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4979         else
4980                 return 0;
4981 }
4982
4983 static struct md_sysfs_entry
4984 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4985
4986 static struct attribute *raid5_attrs[] =  {
4987         &raid5_stripecache_size.attr,
4988         &raid5_stripecache_active.attr,
4989         &raid5_preread_bypass_threshold.attr,
4990         NULL,
4991 };
4992 static struct attribute_group raid5_attrs_group = {
4993         .name = NULL,
4994         .attrs = raid5_attrs,
4995 };
4996
4997 static sector_t
4998 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4999 {
5000         struct r5conf *conf = mddev->private;
5001
5002         if (!sectors)
5003                 sectors = mddev->dev_sectors;
5004         if (!raid_disks)
5005                 /* size is defined by the smallest of previous and new size */
5006                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
5007
5008         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5009         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5010         return sectors * (raid_disks - conf->max_degraded);
5011 }
5012
5013 static void raid5_free_percpu(struct r5conf *conf)
5014 {
5015         struct raid5_percpu *percpu;
5016         unsigned long cpu;
5017
5018         if (!conf->percpu)
5019                 return;
5020
5021         get_online_cpus();
5022         for_each_possible_cpu(cpu) {
5023                 percpu = per_cpu_ptr(conf->percpu, cpu);
5024                 safe_put_page(percpu->spare_page);
5025                 kfree(percpu->scribble);
5026         }
5027 #ifdef CONFIG_HOTPLUG_CPU
5028         unregister_cpu_notifier(&conf->cpu_notify);
5029 #endif
5030         put_online_cpus();
5031
5032         free_percpu(conf->percpu);
5033 }
5034
5035 static void free_conf(struct r5conf *conf)
5036 {
5037         shrink_stripes(conf);
5038         raid5_free_percpu(conf);
5039         kfree(conf->disks);
5040         kfree(conf->stripe_hashtbl);
5041         kfree(conf);
5042 }
5043
5044 #ifdef CONFIG_HOTPLUG_CPU
5045 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5046                               void *hcpu)
5047 {
5048         struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5049         long cpu = (long)hcpu;
5050         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5051
5052         switch (action) {
5053         case CPU_UP_PREPARE:
5054         case CPU_UP_PREPARE_FROZEN:
5055                 if (conf->level == 6 && !percpu->spare_page)
5056                         percpu->spare_page = alloc_page(GFP_KERNEL);
5057                 if (!percpu->scribble)
5058                         percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5059
5060                 if (!percpu->scribble ||
5061                     (conf->level == 6 && !percpu->spare_page)) {
5062                         safe_put_page(percpu->spare_page);
5063                         kfree(percpu->scribble);
5064                         pr_err("%s: failed memory allocation for cpu%ld\n",
5065                                __func__, cpu);
5066                         return notifier_from_errno(-ENOMEM);
5067                 }
5068                 break;
5069         case CPU_DEAD:
5070         case CPU_DEAD_FROZEN:
5071                 safe_put_page(percpu->spare_page);
5072                 kfree(percpu->scribble);
5073                 percpu->spare_page = NULL;
5074                 percpu->scribble = NULL;
5075                 break;
5076         default:
5077                 break;
5078         }
5079         return NOTIFY_OK;
5080 }
5081 #endif
5082
5083 static int raid5_alloc_percpu(struct r5conf *conf)
5084 {
5085         unsigned long cpu;
5086         struct page *spare_page;
5087         struct raid5_percpu __percpu *allcpus;
5088         void *scribble;
5089         int err;
5090
5091         allcpus = alloc_percpu(struct raid5_percpu);
5092         if (!allcpus)
5093                 return -ENOMEM;
5094         conf->percpu = allcpus;
5095
5096         get_online_cpus();
5097         err = 0;
5098         for_each_present_cpu(cpu) {
5099                 if (conf->level == 6) {
5100                         spare_page = alloc_page(GFP_KERNEL);
5101                         if (!spare_page) {
5102                                 err = -ENOMEM;
5103                                 break;
5104                         }
5105                         per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5106                 }
5107                 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5108                 if (!scribble) {
5109                         err = -ENOMEM;
5110                         break;
5111                 }
5112                 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5113         }
5114 #ifdef CONFIG_HOTPLUG_CPU
5115         conf->cpu_notify.notifier_call = raid456_cpu_notify;
5116         conf->cpu_notify.priority = 0;
5117         if (err == 0)
5118                 err = register_cpu_notifier(&conf->cpu_notify);
5119 #endif
5120         put_online_cpus();
5121
5122         return err;
5123 }
5124
5125 static struct r5conf *setup_conf(struct mddev *mddev)
5126 {
5127         struct r5conf *conf;
5128         int raid_disk, memory, max_disks;
5129         struct md_rdev *rdev;
5130         struct disk_info *disk;
5131         char pers_name[6];
5132
5133         if (mddev->new_level != 5
5134             && mddev->new_level != 4
5135             && mddev->new_level != 6) {
5136                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5137                        mdname(mddev), mddev->new_level);
5138                 return ERR_PTR(-EIO);
5139         }
5140         if ((mddev->new_level == 5
5141              && !algorithm_valid_raid5(mddev->new_layout)) ||
5142             (mddev->new_level == 6
5143              && !algorithm_valid_raid6(mddev->new_layout))) {
5144                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5145                        mdname(mddev), mddev->new_layout);
5146                 return ERR_PTR(-EIO);
5147         }
5148         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5149                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5150                        mdname(mddev), mddev->raid_disks);
5151                 return ERR_PTR(-EINVAL);
5152         }
5153
5154         if (!mddev->new_chunk_sectors ||
5155             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5156             !is_power_of_2(mddev->new_chunk_sectors)) {
5157                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5158                        mdname(mddev), mddev->new_chunk_sectors << 9);
5159                 return ERR_PTR(-EINVAL);
5160         }
5161
5162         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5163         if (conf == NULL)
5164                 goto abort;
5165         spin_lock_init(&conf->device_lock);
5166         init_waitqueue_head(&conf->wait_for_stripe);
5167         init_waitqueue_head(&conf->wait_for_overlap);
5168         INIT_LIST_HEAD(&conf->handle_list);
5169         INIT_LIST_HEAD(&conf->hold_list);
5170         INIT_LIST_HEAD(&conf->delayed_list);
5171         INIT_LIST_HEAD(&conf->bitmap_list);
5172         INIT_LIST_HEAD(&conf->inactive_list);
5173         atomic_set(&conf->active_stripes, 0);
5174         atomic_set(&conf->preread_active_stripes, 0);
5175         atomic_set(&conf->active_aligned_reads, 0);
5176         conf->bypass_threshold = BYPASS_THRESHOLD;
5177         conf->recovery_disabled = mddev->recovery_disabled - 1;
5178
5179         conf->raid_disks = mddev->raid_disks;
5180         if (mddev->reshape_position == MaxSector)
5181                 conf->previous_raid_disks = mddev->raid_disks;
5182         else
5183                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5184         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5185         conf->scribble_len = scribble_len(max_disks);
5186
5187         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5188                               GFP_KERNEL);
5189         if (!conf->disks)
5190                 goto abort;
5191
5192         conf->mddev = mddev;
5193
5194         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5195                 goto abort;
5196
5197         conf->level = mddev->new_level;
5198         if (raid5_alloc_percpu(conf) != 0)
5199                 goto abort;
5200
5201         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5202
5203         rdev_for_each(rdev, mddev) {
5204                 raid_disk = rdev->raid_disk;
5205                 if (raid_disk >= max_disks
5206                     || raid_disk < 0)
5207                         continue;
5208                 disk = conf->disks + raid_disk;
5209
5210                 if (test_bit(Replacement, &rdev->flags)) {
5211                         if (disk->replacement)
5212                                 goto abort;
5213                         disk->replacement = rdev;
5214                 } else {
5215                         if (disk->rdev)
5216                                 goto abort;
5217                         disk->rdev = rdev;
5218                 }
5219
5220                 if (test_bit(In_sync, &rdev->flags)) {
5221                         char b[BDEVNAME_SIZE];
5222                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5223                                " disk %d\n",
5224                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5225                 } else if (rdev->saved_raid_disk != raid_disk)
5226                         /* Cannot rely on bitmap to complete recovery */
5227                         conf->fullsync = 1;
5228         }
5229
5230         conf->chunk_sectors = mddev->new_chunk_sectors;
5231         conf->level = mddev->new_level;
5232         if (conf->level == 6)
5233                 conf->max_degraded = 2;
5234         else
5235                 conf->max_degraded = 1;
5236         conf->algorithm = mddev->new_layout;
5237         conf->max_nr_stripes = NR_STRIPES;
5238         conf->reshape_progress = mddev->reshape_position;
5239         if (conf->reshape_progress != MaxSector) {
5240                 conf->prev_chunk_sectors = mddev->chunk_sectors;
5241                 conf->prev_algo = mddev->layout;
5242         }
5243
5244         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5245                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5246         if (grow_stripes(conf, conf->max_nr_stripes)) {
5247                 printk(KERN_ERR
5248                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
5249                        mdname(mddev), memory);
5250                 goto abort;
5251         } else
5252                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5253                        mdname(mddev), memory);
5254
5255         sprintf(pers_name, "raid%d", mddev->new_level);
5256         conf->thread = md_register_thread(raid5d, mddev, pers_name);
5257         if (!conf->thread) {
5258                 printk(KERN_ERR
5259                        "md/raid:%s: couldn't allocate thread.\n",
5260                        mdname(mddev));
5261                 goto abort;
5262         }
5263
5264         return conf;
5265
5266  abort:
5267         if (conf) {
5268                 free_conf(conf);
5269                 return ERR_PTR(-EIO);
5270         } else
5271                 return ERR_PTR(-ENOMEM);
5272 }
5273
5274
5275 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5276 {
5277         switch (algo) {
5278         case ALGORITHM_PARITY_0:
5279                 if (raid_disk < max_degraded)
5280                         return 1;
5281                 break;
5282         case ALGORITHM_PARITY_N:
5283                 if (raid_disk >= raid_disks - max_degraded)
5284                         return 1;
5285                 break;
5286         case ALGORITHM_PARITY_0_6:
5287                 if (raid_disk == 0 ||
5288                     raid_disk == raid_disks - 1)
5289                         return 1;
5290                 break;
5291         case ALGORITHM_LEFT_ASYMMETRIC_6:
5292         case ALGORITHM_RIGHT_ASYMMETRIC_6:
5293         case ALGORITHM_LEFT_SYMMETRIC_6:
5294         case ALGORITHM_RIGHT_SYMMETRIC_6:
5295                 if (raid_disk == raid_disks - 1)
5296                         return 1;
5297         }
5298         return 0;
5299 }
5300
5301 static int run(struct mddev *mddev)
5302 {
5303         struct r5conf *conf;
5304         int working_disks = 0;
5305         int dirty_parity_disks = 0;
5306         struct md_rdev *rdev;
5307         sector_t reshape_offset = 0;
5308         int i;
5309         long long min_offset_diff = 0;
5310         int first = 1;
5311
5312         if (mddev->recovery_cp != MaxSector)
5313                 printk(KERN_NOTICE "md/raid:%s: not clean"
5314                        " -- starting background reconstruction\n",
5315                        mdname(mddev));
5316
5317         rdev_for_each(rdev, mddev) {
5318                 long long diff;
5319                 if (rdev->raid_disk < 0)
5320                         continue;
5321                 diff = (rdev->new_data_offset - rdev->data_offset);
5322                 if (first) {
5323                         min_offset_diff = diff;
5324                         first = 0;
5325                 } else if (mddev->reshape_backwards &&
5326                          diff < min_offset_diff)
5327                         min_offset_diff = diff;
5328                 else if (!mddev->reshape_backwards &&
5329                          diff > min_offset_diff)
5330                         min_offset_diff = diff;
5331         }
5332
5333         if (mddev->reshape_position != MaxSector) {
5334                 /* Check that we can continue the reshape.
5335                  * Difficulties arise if the stripe we would write to
5336                  * next is at or after the stripe we would read from next.
5337                  * For a reshape that changes the number of devices, this
5338                  * is only possible for a very short time, and mdadm makes
5339                  * sure that time appears to have past before assembling
5340                  * the array.  So we fail if that time hasn't passed.
5341                  * For a reshape that keeps the number of devices the same
5342                  * mdadm must be monitoring the reshape can keeping the
5343                  * critical areas read-only and backed up.  It will start
5344                  * the array in read-only mode, so we check for that.
5345                  */
5346                 sector_t here_new, here_old;
5347                 int old_disks;
5348                 int max_degraded = (mddev->level == 6 ? 2 : 1);
5349
5350                 if (mddev->new_level != mddev->level) {
5351                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
5352                                "required - aborting.\n",
5353                                mdname(mddev));
5354                         return -EINVAL;
5355                 }
5356                 old_disks = mddev->raid_disks - mddev->delta_disks;
5357                 /* reshape_position must be on a new-stripe boundary, and one
5358                  * further up in new geometry must map after here in old
5359                  * geometry.
5360                  */
5361                 here_new = mddev->reshape_position;
5362                 if (sector_div(here_new, mddev->new_chunk_sectors *
5363                                (mddev->raid_disks - max_degraded))) {
5364                         printk(KERN_ERR "md/raid:%s: reshape_position not "
5365                                "on a stripe boundary\n", mdname(mddev));
5366                         return -EINVAL;
5367                 }
5368                 reshape_offset = here_new * mddev->new_chunk_sectors;
5369                 /* here_new is the stripe we will write to */
5370                 here_old = mddev->reshape_position;
5371                 sector_div(here_old, mddev->chunk_sectors *
5372                            (old_disks-max_degraded));
5373                 /* here_old is the first stripe that we might need to read
5374                  * from */
5375                 if (mddev->delta_disks == 0) {
5376                         if ((here_new * mddev->new_chunk_sectors !=
5377                              here_old * mddev->chunk_sectors)) {
5378                                 printk(KERN_ERR "md/raid:%s: reshape position is"
5379                                        " confused - aborting\n", mdname(mddev));
5380                                 return -EINVAL;
5381                         }
5382                         /* We cannot be sure it is safe to start an in-place
5383                          * reshape.  It is only safe if user-space is monitoring
5384                          * and taking constant backups.
5385                          * mdadm always starts a situation like this in
5386                          * readonly mode so it can take control before
5387                          * allowing any writes.  So just check for that.
5388                          */
5389                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5390                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
5391                                 /* not really in-place - so OK */;
5392                         else if (mddev->ro == 0) {
5393                                 printk(KERN_ERR "md/raid:%s: in-place reshape "
5394                                        "must be started in read-only mode "
5395                                        "- aborting\n",
5396                                        mdname(mddev));
5397                                 return -EINVAL;
5398                         }
5399                 } else if (mddev->reshape_backwards
5400                     ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5401                        here_old * mddev->chunk_sectors)
5402                     : (here_new * mddev->new_chunk_sectors >=
5403                        here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5404                         /* Reading from the same stripe as writing to - bad */
5405                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5406                                "auto-recovery - aborting.\n",
5407                                mdname(mddev));
5408                         return -EINVAL;
5409                 }
5410                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5411                        mdname(mddev));
5412                 /* OK, we should be able to continue; */
5413         } else {
5414                 BUG_ON(mddev->level != mddev->new_level);
5415                 BUG_ON(mddev->layout != mddev->new_layout);
5416                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5417                 BUG_ON(mddev->delta_disks != 0);
5418         }
5419
5420         if (mddev->private == NULL)
5421                 conf = setup_conf(mddev);
5422         else
5423                 conf = mddev->private;
5424
5425         if (IS_ERR(conf))
5426                 return PTR_ERR(conf);
5427
5428         conf->min_offset_diff = min_offset_diff;
5429         mddev->thread = conf->thread;
5430         conf->thread = NULL;
5431         mddev->private = conf;
5432
5433         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5434              i++) {
5435                 rdev = conf->disks[i].rdev;
5436                 if (!rdev && conf->disks[i].replacement) {
5437                         /* The replacement is all we have yet */
5438                         rdev = conf->disks[i].replacement;
5439                         conf->disks[i].replacement = NULL;
5440                         clear_bit(Replacement, &rdev->flags);
5441                         conf->disks[i].rdev = rdev;
5442                 }
5443                 if (!rdev)
5444                         continue;
5445                 if (conf->disks[i].replacement &&
5446                     conf->reshape_progress != MaxSector) {
5447                         /* replacements and reshape simply do not mix. */
5448                         printk(KERN_ERR "md: cannot handle concurrent "
5449                                "replacement and reshape.\n");
5450                         goto abort;
5451                 }
5452                 if (test_bit(In_sync, &rdev->flags)) {
5453                         working_disks++;
5454                         continue;
5455                 }
5456                 /* This disc is not fully in-sync.  However if it
5457                  * just stored parity (beyond the recovery_offset),
5458                  * when we don't need to be concerned about the
5459                  * array being dirty.
5460                  * When reshape goes 'backwards', we never have
5461                  * partially completed devices, so we only need
5462                  * to worry about reshape going forwards.
5463                  */
5464                 /* Hack because v0.91 doesn't store recovery_offset properly. */
5465                 if (mddev->major_version == 0 &&
5466                     mddev->minor_version > 90)
5467                         rdev->recovery_offset = reshape_offset;
5468
5469                 if (rdev->recovery_offset < reshape_offset) {
5470                         /* We need to check old and new layout */
5471                         if (!only_parity(rdev->raid_disk,
5472                                          conf->algorithm,
5473                                          conf->raid_disks,
5474                                          conf->max_degraded))
5475                                 continue;
5476                 }
5477                 if (!only_parity(rdev->raid_disk,
5478                                  conf->prev_algo,
5479                                  conf->previous_raid_disks,
5480                                  conf->max_degraded))
5481                         continue;
5482                 dirty_parity_disks++;
5483         }
5484
5485         /*
5486          * 0 for a fully functional array, 1 or 2 for a degraded array.
5487          */
5488         mddev->degraded = calc_degraded(conf);
5489
5490         if (has_failed(conf)) {
5491                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5492                         " (%d/%d failed)\n",
5493                         mdname(mddev), mddev->degraded, conf->raid_disks);
5494                 goto abort;
5495         }
5496
5497         /* device size must be a multiple of chunk size */
5498         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5499         mddev->resync_max_sectors = mddev->dev_sectors;
5500
5501         if (mddev->degraded > dirty_parity_disks &&
5502             mddev->recovery_cp != MaxSector) {
5503                 if (mddev->ok_start_degraded)
5504                         printk(KERN_WARNING
5505                                "md/raid:%s: starting dirty degraded array"
5506                                " - data corruption possible.\n",
5507                                mdname(mddev));
5508                 else {
5509                         printk(KERN_ERR
5510                                "md/raid:%s: cannot start dirty degraded array.\n",
5511                                mdname(mddev));
5512                         goto abort;
5513                 }
5514         }
5515
5516         if (mddev->degraded == 0)
5517                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5518                        " devices, algorithm %d\n", mdname(mddev), conf->level,
5519                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5520                        mddev->new_layout);
5521         else
5522                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5523                        " out of %d devices, algorithm %d\n",
5524                        mdname(mddev), conf->level,
5525                        mddev->raid_disks - mddev->degraded,
5526                        mddev->raid_disks, mddev->new_layout);
5527
5528         print_raid5_conf(conf);
5529
5530         if (conf->reshape_progress != MaxSector) {
5531                 conf->reshape_safe = conf->reshape_progress;
5532                 atomic_set(&conf->reshape_stripes, 0);
5533                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5534                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5535                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5536                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5537                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5538                                                         "reshape");
5539         }
5540
5541
5542         /* Ok, everything is just fine now */
5543         if (mddev->to_remove == &raid5_attrs_group)
5544                 mddev->to_remove = NULL;
5545         else if (mddev->kobj.sd &&
5546             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5547                 printk(KERN_WARNING
5548                        "raid5: failed to create sysfs attributes for %s\n",
5549                        mdname(mddev));
5550         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5551
5552         if (mddev->queue) {
5553                 int chunk_size;
5554                 bool discard_supported = true;
5555                 /* read-ahead size must cover two whole stripes, which
5556                  * is 2 * (datadisks) * chunksize where 'n' is the
5557                  * number of raid devices
5558                  */
5559                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5560                 int stripe = data_disks *
5561                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5562                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5563                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5564
5565                 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5566
5567                 mddev->queue->backing_dev_info.congested_data = mddev;
5568                 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5569
5570                 chunk_size = mddev->chunk_sectors << 9;
5571                 blk_queue_io_min(mddev->queue, chunk_size);
5572                 blk_queue_io_opt(mddev->queue, chunk_size *
5573                                  (conf->raid_disks - conf->max_degraded));
5574                 /*
5575                  * We can only discard a whole stripe. It doesn't make sense to
5576                  * discard data disk but write parity disk
5577                  */
5578                 stripe = stripe * PAGE_SIZE;
5579                 /* Round up to power of 2, as discard handling
5580                  * currently assumes that */
5581                 while ((stripe-1) & stripe)
5582                         stripe = (stripe | (stripe-1)) + 1;
5583                 mddev->queue->limits.discard_alignment = stripe;
5584                 mddev->queue->limits.discard_granularity = stripe;
5585                 /*
5586                  * unaligned part of discard request will be ignored, so can't
5587                  * guarantee discard_zerors_data
5588                  */
5589                 mddev->queue->limits.discard_zeroes_data = 0;
5590
5591                 rdev_for_each(rdev, mddev) {
5592                         disk_stack_limits(mddev->gendisk, rdev->bdev,
5593                                           rdev->data_offset << 9);
5594                         disk_stack_limits(mddev->gendisk, rdev->bdev,
5595                                           rdev->new_data_offset << 9);
5596                         /*
5597                          * discard_zeroes_data is required, otherwise data
5598                          * could be lost. Consider a scenario: discard a stripe
5599                          * (the stripe could be inconsistent if
5600                          * discard_zeroes_data is 0); write one disk of the
5601                          * stripe (the stripe could be inconsistent again
5602                          * depending on which disks are used to calculate
5603                          * parity); the disk is broken; The stripe data of this
5604                          * disk is lost.
5605                          */
5606                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5607                             !bdev_get_queue(rdev->bdev)->
5608                                                 limits.discard_zeroes_data)
5609                                 discard_supported = false;
5610                 }
5611
5612                 if (discard_supported &&
5613                    mddev->queue->limits.max_discard_sectors >= stripe &&
5614                    mddev->queue->limits.discard_granularity >= stripe)
5615                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5616                                                 mddev->queue);
5617                 else
5618                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5619                                                 mddev->queue);
5620         }
5621
5622         return 0;
5623 abort:
5624         md_unregister_thread(&mddev->thread);
5625         print_raid5_conf(conf);
5626         free_conf(conf);
5627         mddev->private = NULL;
5628         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5629         return -EIO;
5630 }
5631
5632 static int stop(struct mddev *mddev)
5633 {
5634         struct r5conf *conf = mddev->private;
5635
5636         md_unregister_thread(&mddev->thread);
5637         if (mddev->queue)
5638                 mddev->queue->backing_dev_info.congested_fn = NULL;
5639         free_conf(conf);
5640         mddev->private = NULL;
5641         mddev->to_remove = &raid5_attrs_group;
5642         return 0;
5643 }
5644
5645 static void status(struct seq_file *seq, struct mddev *mddev)
5646 {
5647         struct r5conf *conf = mddev->private;
5648         int i;
5649
5650         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5651                 mddev->chunk_sectors / 2, mddev->layout);
5652         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5653         for (i = 0; i < conf->raid_disks; i++)
5654                 seq_printf (seq, "%s",
5655                                conf->disks[i].rdev &&
5656                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5657         seq_printf (seq, "]");
5658 }
5659
5660 static void print_raid5_conf (struct r5conf *conf)
5661 {
5662         int i;
5663         struct disk_info *tmp;
5664
5665         printk(KERN_DEBUG "RAID conf printout:\n");
5666         if (!conf) {
5667                 printk("(conf==NULL)\n");
5668                 return;
5669         }
5670         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5671                conf->raid_disks,
5672                conf->raid_disks - conf->mddev->degraded);
5673
5674         for (i = 0; i < conf->raid_disks; i++) {
5675                 char b[BDEVNAME_SIZE];
5676                 tmp = conf->disks + i;
5677                 if (tmp->rdev)
5678                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5679                                i, !test_bit(Faulty, &tmp->rdev->flags),
5680                                bdevname(tmp->rdev->bdev, b));
5681         }
5682 }
5683
5684 static int raid5_spare_active(struct mddev *mddev)
5685 {
5686         int i;
5687         struct r5conf *conf = mddev->private;
5688         struct disk_info *tmp;
5689         int count = 0;
5690         unsigned long flags;
5691
5692         for (i = 0; i < conf->raid_disks; i++) {
5693                 tmp = conf->disks + i;
5694                 if (tmp->replacement
5695                     && tmp->replacement->recovery_offset == MaxSector
5696                     && !test_bit(Faulty, &tmp->replacement->flags)
5697                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5698                         /* Replacement has just become active. */
5699                         if (!tmp->rdev
5700                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5701                                 count++;
5702                         if (tmp->rdev) {
5703                                 /* Replaced device not technically faulty,
5704                                  * but we need to be sure it gets removed
5705                                  * and never re-added.
5706                                  */
5707                                 set_bit(Faulty, &tmp->rdev->flags);
5708                                 sysfs_notify_dirent_safe(
5709                                         tmp->rdev->sysfs_state);
5710                         }
5711                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5712                 } else if (tmp->rdev
5713                     && tmp->rdev->recovery_offset == MaxSector
5714                     && !test_bit(Faulty, &tmp->rdev->flags)
5715                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5716                         count++;
5717                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5718                 }
5719         }
5720         spin_lock_irqsave(&conf->device_lock, flags);
5721         mddev->degraded = calc_degraded(conf);
5722         spin_unlock_irqrestore(&conf->device_lock, flags);
5723         print_raid5_conf(conf);
5724         return count;
5725 }
5726
5727 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5728 {
5729         struct r5conf *conf = mddev->private;
5730         int err = 0;
5731         int number = rdev->raid_disk;
5732         struct md_rdev **rdevp;
5733         struct disk_info *p = conf->disks + number;
5734
5735         print_raid5_conf(conf);
5736         if (rdev == p->rdev)
5737                 rdevp = &p->rdev;
5738         else if (rdev == p->replacement)
5739                 rdevp = &p->replacement;
5740         else
5741                 return 0;
5742
5743         if (number >= conf->raid_disks &&
5744             conf->reshape_progress == MaxSector)
5745                 clear_bit(In_sync, &rdev->flags);
5746
5747         if (test_bit(In_sync, &rdev->flags) ||
5748             atomic_read(&rdev->nr_pending)) {
5749                 err = -EBUSY;
5750                 goto abort;
5751         }
5752         /* Only remove non-faulty devices if recovery
5753          * isn't possible.
5754          */
5755         if (!test_bit(Faulty, &rdev->flags) &&
5756             mddev->recovery_disabled != conf->recovery_disabled &&
5757             !has_failed(conf) &&
5758             (!p->replacement || p->replacement == rdev) &&
5759             number < conf->raid_disks) {
5760                 err = -EBUSY;
5761                 goto abort;
5762         }
5763         *rdevp = NULL;
5764         synchronize_rcu();
5765         if (atomic_read(&rdev->nr_pending)) {
5766                 /* lost the race, try later */
5767                 err = -EBUSY;
5768                 *rdevp = rdev;
5769         } else if (p->replacement) {
5770                 /* We must have just cleared 'rdev' */
5771                 p->rdev = p->replacement;
5772                 clear_bit(Replacement, &p->replacement->flags);
5773                 smp_mb(); /* Make sure other CPUs may see both as identical
5774                            * but will never see neither - if they are careful
5775                            */
5776                 p->replacement = NULL;
5777                 clear_bit(WantReplacement, &rdev->flags);
5778         } else
5779                 /* We might have just removed the Replacement as faulty-
5780                  * clear the bit just in case
5781                  */
5782                 clear_bit(WantReplacement, &rdev->flags);
5783 abort:
5784
5785         print_raid5_conf(conf);
5786         return err;
5787 }
5788
5789 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5790 {
5791         struct r5conf *conf = mddev->private;
5792         int err = -EEXIST;
5793         int disk;
5794         struct disk_info *p;
5795         int first = 0;
5796         int last = conf->raid_disks - 1;
5797
5798         if (mddev->recovery_disabled == conf->recovery_disabled)
5799                 return -EBUSY;
5800
5801         if (rdev->saved_raid_disk < 0 && has_failed(conf))
5802                 /* no point adding a device */
5803                 return -EINVAL;
5804
5805         if (rdev->raid_disk >= 0)
5806                 first = last = rdev->raid_disk;
5807
5808         /*
5809          * find the disk ... but prefer rdev->saved_raid_disk
5810          * if possible.
5811          */
5812         if (rdev->saved_raid_disk >= 0 &&
5813             rdev->saved_raid_disk >= first &&
5814             conf->disks[rdev->saved_raid_disk].rdev == NULL)
5815                 first = rdev->saved_raid_disk;
5816
5817         for (disk = first; disk <= last; disk++) {
5818                 p = conf->disks + disk;
5819                 if (p->rdev == NULL) {
5820                         clear_bit(In_sync, &rdev->flags);
5821                         rdev->raid_disk = disk;
5822                         err = 0;
5823                         if (rdev->saved_raid_disk != disk)
5824                                 conf->fullsync = 1;
5825                         rcu_assign_pointer(p->rdev, rdev);
5826                         goto out;
5827                 }
5828         }
5829         for (disk = first; disk <= last; disk++) {
5830                 p = conf->disks + disk;
5831                 if (test_bit(WantReplacement, &p->rdev->flags) &&
5832                     p->replacement == NULL) {
5833                         clear_bit(In_sync, &rdev->flags);
5834                         set_bit(Replacement, &rdev->flags);
5835                         rdev->raid_disk = disk;
5836                         err = 0;
5837                         conf->fullsync = 1;
5838                         rcu_assign_pointer(p->replacement, rdev);
5839                         break;
5840                 }
5841         }
5842 out:
5843         print_raid5_conf(conf);
5844         return err;
5845 }
5846
5847 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5848 {
5849         /* no resync is happening, and there is enough space
5850          * on all devices, so we can resize.
5851          * We need to make sure resync covers any new space.
5852          * If the array is shrinking we should possibly wait until
5853          * any io in the removed space completes, but it hardly seems
5854          * worth it.
5855          */
5856         sector_t newsize;
5857         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5858         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5859         if (mddev->external_size &&
5860             mddev->array_sectors > newsize)
5861                 return -EINVAL;
5862         if (mddev->bitmap) {
5863                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5864                 if (ret)
5865                         return ret;
5866         }
5867         md_set_array_sectors(mddev, newsize);
5868         set_capacity(mddev->gendisk, mddev->array_sectors);
5869         revalidate_disk(mddev->gendisk);
5870         if (sectors > mddev->dev_sectors &&
5871             mddev->recovery_cp > mddev->dev_sectors) {
5872                 mddev->recovery_cp = mddev->dev_sectors;
5873                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5874         }
5875         mddev->dev_sectors = sectors;
5876         mddev->resync_max_sectors = sectors;
5877         return 0;
5878 }
5879
5880 static int check_stripe_cache(struct mddev *mddev)
5881 {
5882         /* Can only proceed if there are plenty of stripe_heads.
5883          * We need a minimum of one full stripe,, and for sensible progress
5884          * it is best to have about 4 times that.
5885          * If we require 4 times, then the default 256 4K stripe_heads will
5886          * allow for chunk sizes up to 256K, which is probably OK.
5887          * If the chunk size is greater, user-space should request more
5888          * stripe_heads first.
5889          */
5890         struct r5conf *conf = mddev->private;
5891         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5892             > conf->max_nr_stripes ||
5893             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5894             > conf->max_nr_stripes) {
5895                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
5896                        mdname(mddev),
5897                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5898                         / STRIPE_SIZE)*4);
5899                 return 0;
5900         }
5901         return 1;
5902 }
5903
5904 static int check_reshape(struct mddev *mddev)
5905 {
5906         struct r5conf *conf = mddev->private;
5907
5908         if (mddev->delta_disks == 0 &&
5909             mddev->new_layout == mddev->layout &&
5910             mddev->new_chunk_sectors == mddev->chunk_sectors)
5911                 return 0; /* nothing to do */
5912         if (has_failed(conf))
5913                 return -EINVAL;
5914         if (mddev->delta_disks < 0) {
5915                 /* We might be able to shrink, but the devices must
5916                  * be made bigger first.
5917                  * For raid6, 4 is the minimum size.
5918                  * Otherwise 2 is the minimum
5919                  */
5920                 int min = 2;
5921                 if (mddev->level == 6)
5922                         min = 4;
5923                 if (mddev->raid_disks + mddev->delta_disks < min)
5924                         return -EINVAL;
5925         }
5926
5927         if (!check_stripe_cache(mddev))
5928                 return -ENOSPC;
5929
5930         return resize_stripes(conf, (conf->previous_raid_disks
5931                                      + mddev->delta_disks));
5932 }
5933
5934 static int raid5_start_reshape(struct mddev *mddev)
5935 {
5936         struct r5conf *conf = mddev->private;
5937         struct md_rdev *rdev;
5938         int spares = 0;
5939         unsigned long flags;
5940
5941         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5942                 return -EBUSY;
5943
5944         if (!check_stripe_cache(mddev))
5945                 return -ENOSPC;
5946
5947         if (has_failed(conf))
5948                 return -EINVAL;
5949
5950         rdev_for_each(rdev, mddev) {
5951                 if (!test_bit(In_sync, &rdev->flags)
5952                     && !test_bit(Faulty, &rdev->flags))
5953                         spares++;
5954         }
5955
5956         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5957                 /* Not enough devices even to make a degraded array
5958                  * of that size
5959                  */
5960                 return -EINVAL;
5961
5962         /* Refuse to reduce size of the array.  Any reductions in
5963          * array size must be through explicit setting of array_size
5964          * attribute.
5965          */
5966         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5967             < mddev->array_sectors) {
5968                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5969                        "before number of disks\n", mdname(mddev));
5970                 return -EINVAL;
5971         }
5972
5973         atomic_set(&conf->reshape_stripes, 0);
5974         spin_lock_irq(&conf->device_lock);
5975         conf->previous_raid_disks = conf->raid_disks;
5976         conf->raid_disks += mddev->delta_disks;
5977         conf->prev_chunk_sectors = conf->chunk_sectors;
5978         conf->chunk_sectors = mddev->new_chunk_sectors;
5979         conf->prev_algo = conf->algorithm;
5980         conf->algorithm = mddev->new_layout;
5981         conf->generation++;
5982         /* Code that selects data_offset needs to see the generation update
5983          * if reshape_progress has been set - so a memory barrier needed.
5984          */
5985         smp_mb();
5986         if (mddev->reshape_backwards)
5987                 conf->reshape_progress = raid5_size(mddev, 0, 0);
5988         else
5989                 conf->reshape_progress = 0;
5990         conf->reshape_safe = conf->reshape_progress;
5991         spin_unlock_irq(&conf->device_lock);
5992
5993         /* Add some new drives, as many as will fit.
5994          * We know there are enough to make the newly sized array work.
5995          * Don't add devices if we are reducing the number of
5996          * devices in the array.  This is because it is not possible
5997          * to correctly record the "partially reconstructed" state of
5998          * such devices during the reshape and confusion could result.
5999          */
6000         if (mddev->delta_disks >= 0) {
6001                 rdev_for_each(rdev, mddev)
6002                         if (rdev->raid_disk < 0 &&
6003                             !test_bit(Faulty, &rdev->flags)) {
6004                                 if (raid5_add_disk(mddev, rdev) == 0) {
6005                                         if (rdev->raid_disk
6006                                             >= conf->previous_raid_disks)
6007                                                 set_bit(In_sync, &rdev->flags);
6008                                         else
6009                                                 rdev->recovery_offset = 0;
6010
6011                                         if (sysfs_link_rdev(mddev, rdev))
6012                                                 /* Failure here is OK */;
6013                                 }
6014                         } else if (rdev->raid_disk >= conf->previous_raid_disks
6015                                    && !test_bit(Faulty, &rdev->flags)) {
6016                                 /* This is a spare that was manually added */
6017                                 set_bit(In_sync, &rdev->flags);
6018                         }
6019
6020                 /* When a reshape changes the number of devices,
6021                  * ->degraded is measured against the larger of the
6022                  * pre and post number of devices.
6023                  */
6024                 spin_lock_irqsave(&conf->device_lock, flags);
6025                 mddev->degraded = calc_degraded(conf);
6026                 spin_unlock_irqrestore(&conf->device_lock, flags);
6027         }
6028         mddev->raid_disks = conf->raid_disks;
6029         mddev->reshape_position = conf->reshape_progress;
6030         set_bit(MD_CHANGE_DEVS, &mddev->flags);
6031
6032         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6033         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6034         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6035         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6036         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6037                                                 "reshape");
6038         if (!mddev->sync_thread) {
6039                 mddev->recovery = 0;
6040                 spin_lock_irq(&conf->device_lock);
6041                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6042                 rdev_for_each(rdev, mddev)
6043                         rdev->new_data_offset = rdev->data_offset;
6044                 smp_wmb();
6045                 conf->reshape_progress = MaxSector;
6046                 mddev->reshape_position = MaxSector;
6047                 spin_unlock_irq(&conf->device_lock);
6048                 return -EAGAIN;
6049         }
6050         conf->reshape_checkpoint = jiffies;
6051         md_wakeup_thread(mddev->sync_thread);
6052         md_new_event(mddev);
6053         return 0;
6054 }
6055
6056 /* This is called from the reshape thread and should make any
6057  * changes needed in 'conf'
6058  */
6059 static void end_reshape(struct r5conf *conf)
6060 {
6061
6062         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6063                 struct md_rdev *rdev;
6064
6065                 spin_lock_irq(&conf->device_lock);
6066                 conf->previous_raid_disks = conf->raid_disks;
6067                 rdev_for_each(rdev, conf->mddev)
6068                         rdev->data_offset = rdev->new_data_offset;
6069                 smp_wmb();
6070                 conf->reshape_progress = MaxSector;
6071                 spin_unlock_irq(&conf->device_lock);
6072                 wake_up(&conf->wait_for_overlap);
6073
6074                 /* read-ahead size must cover two whole stripes, which is
6075                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6076                  */
6077                 if (conf->mddev->queue) {
6078                         int data_disks = conf->raid_disks - conf->max_degraded;
6079                         int stripe = data_disks * ((conf->chunk_sectors << 9)
6080                                                    / PAGE_SIZE);
6081                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6082                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6083                 }
6084         }
6085 }
6086
6087 /* This is called from the raid5d thread with mddev_lock held.
6088  * It makes config changes to the device.
6089  */
6090 static void raid5_finish_reshape(struct mddev *mddev)
6091 {
6092         struct r5conf *conf = mddev->private;
6093
6094         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6095
6096                 if (mddev->delta_disks > 0) {
6097                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6098                         set_capacity(mddev->gendisk, mddev->array_sectors);
6099                         revalidate_disk(mddev->gendisk);
6100                 } else {
6101                         int d;
6102                         spin_lock_irq(&conf->device_lock);
6103                         mddev->degraded = calc_degraded(conf);
6104                         spin_unlock_irq(&conf->device_lock);
6105                         for (d = conf->raid_disks ;
6106                              d < conf->raid_disks - mddev->delta_disks;
6107                              d++) {
6108                                 struct md_rdev *rdev = conf->disks[d].rdev;
6109                                 if (rdev)
6110                                         clear_bit(In_sync, &rdev->flags);
6111                                 rdev = conf->disks[d].replacement;
6112                                 if (rdev)
6113                                         clear_bit(In_sync, &rdev->flags);
6114                         }
6115                 }
6116                 mddev->layout = conf->algorithm;
6117                 mddev->chunk_sectors = conf->chunk_sectors;
6118                 mddev->reshape_position = MaxSector;
6119                 mddev->delta_disks = 0;
6120                 mddev->reshape_backwards = 0;
6121         }
6122 }
6123
6124 static void raid5_quiesce(struct mddev *mddev, int state)
6125 {
6126         struct r5conf *conf = mddev->private;
6127
6128         switch(state) {
6129         case 2: /* resume for a suspend */
6130                 wake_up(&conf->wait_for_overlap);
6131                 break;
6132
6133         case 1: /* stop all writes */
6134                 spin_lock_irq(&conf->device_lock);
6135                 /* '2' tells resync/reshape to pause so that all
6136                  * active stripes can drain
6137                  */
6138                 conf->quiesce = 2;
6139                 wait_event_lock_irq(conf->wait_for_stripe,
6140                                     atomic_read(&conf->active_stripes) == 0 &&
6141                                     atomic_read(&conf->active_aligned_reads) == 0,
6142                                     conf->device_lock);
6143                 conf->quiesce = 1;
6144                 spin_unlock_irq(&conf->device_lock);
6145                 /* allow reshape to continue */
6146                 wake_up(&conf->wait_for_overlap);
6147                 break;
6148
6149         case 0: /* re-enable writes */
6150                 spin_lock_irq(&conf->device_lock);
6151                 conf->quiesce = 0;
6152                 wake_up(&conf->wait_for_stripe);
6153                 wake_up(&conf->wait_for_overlap);
6154                 spin_unlock_irq(&conf->device_lock);
6155                 break;
6156         }
6157 }
6158
6159
6160 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6161 {
6162         struct r0conf *raid0_conf = mddev->private;
6163         sector_t sectors;
6164
6165         /* for raid0 takeover only one zone is supported */
6166         if (raid0_conf->nr_strip_zones > 1) {
6167                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6168                        mdname(mddev));
6169                 return ERR_PTR(-EINVAL);
6170         }
6171
6172         sectors = raid0_conf->strip_zone[0].zone_end;
6173         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6174         mddev->dev_sectors = sectors;
6175         mddev->new_level = level;
6176         mddev->new_layout = ALGORITHM_PARITY_N;
6177         mddev->new_chunk_sectors = mddev->chunk_sectors;
6178         mddev->raid_disks += 1;
6179         mddev->delta_disks = 1;
6180         /* make sure it will be not marked as dirty */
6181         mddev->recovery_cp = MaxSector;
6182
6183         return setup_conf(mddev);
6184 }
6185
6186
6187 static void *raid5_takeover_raid1(struct mddev *mddev)
6188 {
6189         int chunksect;
6190
6191         if (mddev->raid_disks != 2 ||
6192             mddev->degraded > 1)
6193                 return ERR_PTR(-EINVAL);
6194
6195         /* Should check if there are write-behind devices? */
6196
6197         chunksect = 64*2; /* 64K by default */
6198
6199         /* The array must be an exact multiple of chunksize */
6200         while (chunksect && (mddev->array_sectors & (chunksect-1)))
6201                 chunksect >>= 1;
6202
6203         if ((chunksect<<9) < STRIPE_SIZE)
6204                 /* array size does not allow a suitable chunk size */
6205                 return ERR_PTR(-EINVAL);
6206
6207         mddev->new_level = 5;
6208         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6209         mddev->new_chunk_sectors = chunksect;
6210
6211         return setup_conf(mddev);
6212 }
6213
6214 static void *raid5_takeover_raid6(struct mddev *mddev)
6215 {
6216         int new_layout;
6217
6218         switch (mddev->layout) {
6219         case ALGORITHM_LEFT_ASYMMETRIC_6:
6220                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6221                 break;
6222         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6223                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6224                 break;
6225         case ALGORITHM_LEFT_SYMMETRIC_6:
6226                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6227                 break;
6228         case ALGORITHM_RIGHT_SYMMETRIC_6:
6229                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6230                 break;
6231         case ALGORITHM_PARITY_0_6:
6232                 new_layout = ALGORITHM_PARITY_0;
6233                 break;
6234         case ALGORITHM_PARITY_N:
6235                 new_layout = ALGORITHM_PARITY_N;
6236                 break;
6237         default:
6238                 return ERR_PTR(-EINVAL);
6239         }
6240         mddev->new_level = 5;
6241         mddev->new_layout = new_layout;
6242         mddev->delta_disks = -1;
6243         mddev->raid_disks -= 1;
6244         return setup_conf(mddev);
6245 }
6246
6247
6248 static int raid5_check_reshape(struct mddev *mddev)
6249 {
6250         /* For a 2-drive array, the layout and chunk size can be changed
6251          * immediately as not restriping is needed.
6252          * For larger arrays we record the new value - after validation
6253          * to be used by a reshape pass.
6254          */
6255         struct r5conf *conf = mddev->private;
6256         int new_chunk = mddev->new_chunk_sectors;
6257
6258         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6259                 return -EINVAL;
6260         if (new_chunk > 0) {
6261                 if (!is_power_of_2(new_chunk))
6262                         return -EINVAL;
6263                 if (new_chunk < (PAGE_SIZE>>9))
6264                         return -EINVAL;
6265                 if (mddev->array_sectors & (new_chunk-1))
6266                         /* not factor of array size */
6267                         return -EINVAL;
6268         }
6269
6270         /* They look valid */
6271
6272         if (mddev->raid_disks == 2) {
6273                 /* can make the change immediately */
6274                 if (mddev->new_layout >= 0) {
6275                         conf->algorithm = mddev->new_layout;
6276                         mddev->layout = mddev->new_layout;
6277                 }
6278                 if (new_chunk > 0) {
6279                         conf->chunk_sectors = new_chunk ;
6280                         mddev->chunk_sectors = new_chunk;
6281                 }
6282                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6283                 md_wakeup_thread(mddev->thread);
6284         }
6285         return check_reshape(mddev);
6286 }
6287
6288 static int raid6_check_reshape(struct mddev *mddev)
6289 {
6290         int new_chunk = mddev->new_chunk_sectors;
6291
6292         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6293                 return -EINVAL;
6294         if (new_chunk > 0) {
6295                 if (!is_power_of_2(new_chunk))
6296                         return -EINVAL;
6297                 if (new_chunk < (PAGE_SIZE >> 9))
6298                         return -EINVAL;
6299                 if (mddev->array_sectors & (new_chunk-1))
6300                         /* not factor of array size */
6301                         return -EINVAL;
6302         }
6303
6304         /* They look valid */
6305         return check_reshape(mddev);
6306 }
6307
6308 static void *raid5_takeover(struct mddev *mddev)
6309 {
6310         /* raid5 can take over:
6311          *  raid0 - if there is only one strip zone - make it a raid4 layout
6312          *  raid1 - if there are two drives.  We need to know the chunk size
6313          *  raid4 - trivial - just use a raid4 layout.
6314          *  raid6 - Providing it is a *_6 layout
6315          */
6316         if (mddev->level == 0)
6317                 return raid45_takeover_raid0(mddev, 5);
6318         if (mddev->level == 1)
6319                 return raid5_takeover_raid1(mddev);
6320         if (mddev->level == 4) {
6321                 mddev->new_layout = ALGORITHM_PARITY_N;
6322                 mddev->new_level = 5;
6323                 return setup_conf(mddev);
6324         }
6325         if (mddev->level == 6)
6326                 return raid5_takeover_raid6(mddev);
6327
6328         return ERR_PTR(-EINVAL);
6329 }
6330
6331 static void *raid4_takeover(struct mddev *mddev)
6332 {
6333         /* raid4 can take over:
6334          *  raid0 - if there is only one strip zone
6335          *  raid5 - if layout is right
6336          */
6337         if (mddev->level == 0)
6338                 return raid45_takeover_raid0(mddev, 4);
6339         if (mddev->level == 5 &&
6340             mddev->layout == ALGORITHM_PARITY_N) {
6341                 mddev->new_layout = 0;
6342                 mddev->new_level = 4;
6343                 return setup_conf(mddev);
6344         }
6345         return ERR_PTR(-EINVAL);
6346 }
6347
6348 static struct md_personality raid5_personality;
6349
6350 static void *raid6_takeover(struct mddev *mddev)
6351 {
6352         /* Currently can only take over a raid5.  We map the
6353          * personality to an equivalent raid6 personality
6354          * with the Q block at the end.
6355          */
6356         int new_layout;
6357
6358         if (mddev->pers != &raid5_personality)
6359                 return ERR_PTR(-EINVAL);
6360         if (mddev->degraded > 1)
6361                 return ERR_PTR(-EINVAL);
6362         if (mddev->raid_disks > 253)
6363                 return ERR_PTR(-EINVAL);
6364         if (mddev->raid_disks < 3)
6365                 return ERR_PTR(-EINVAL);
6366
6367         switch (mddev->layout) {
6368         case ALGORITHM_LEFT_ASYMMETRIC:
6369                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6370                 break;
6371         case ALGORITHM_RIGHT_ASYMMETRIC:
6372                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6373                 break;
6374         case ALGORITHM_LEFT_SYMMETRIC:
6375                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6376                 break;
6377         case ALGORITHM_RIGHT_SYMMETRIC:
6378                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6379                 break;
6380         case ALGORITHM_PARITY_0:
6381                 new_layout = ALGORITHM_PARITY_0_6;
6382                 break;
6383         case ALGORITHM_PARITY_N:
6384                 new_layout = ALGORITHM_PARITY_N;
6385                 break;
6386         default:
6387                 return ERR_PTR(-EINVAL);
6388         }
6389         mddev->new_level = 6;
6390         mddev->new_layout = new_layout;
6391         mddev->delta_disks = 1;
6392         mddev->raid_disks += 1;
6393         return setup_conf(mddev);
6394 }
6395
6396
6397 static struct md_personality raid6_personality =
6398 {
6399         .name           = "raid6",
6400         .level          = 6,
6401         .owner          = THIS_MODULE,
6402         .make_request   = make_request,
6403         .run            = run,
6404         .stop           = stop,
6405         .status         = status,
6406         .error_handler  = error,
6407         .hot_add_disk   = raid5_add_disk,
6408         .hot_remove_disk= raid5_remove_disk,
6409         .spare_active   = raid5_spare_active,
6410         .sync_request   = sync_request,
6411         .resize         = raid5_resize,
6412         .size           = raid5_size,
6413         .check_reshape  = raid6_check_reshape,
6414         .start_reshape  = raid5_start_reshape,
6415         .finish_reshape = raid5_finish_reshape,
6416         .quiesce        = raid5_quiesce,
6417         .takeover       = raid6_takeover,
6418 };
6419 static struct md_personality raid5_personality =
6420 {
6421         .name           = "raid5",
6422         .level          = 5,
6423         .owner          = THIS_MODULE,
6424         .make_request   = make_request,
6425         .run            = run,
6426         .stop           = stop,
6427         .status         = status,
6428         .error_handler  = error,
6429         .hot_add_disk   = raid5_add_disk,
6430         .hot_remove_disk= raid5_remove_disk,
6431         .spare_active   = raid5_spare_active,
6432         .sync_request   = sync_request,
6433         .resize         = raid5_resize,
6434         .size           = raid5_size,
6435         .check_reshape  = raid5_check_reshape,
6436         .start_reshape  = raid5_start_reshape,
6437         .finish_reshape = raid5_finish_reshape,
6438         .quiesce        = raid5_quiesce,
6439         .takeover       = raid5_takeover,
6440 };
6441
6442 static struct md_personality raid4_personality =
6443 {
6444         .name           = "raid4",
6445         .level          = 4,
6446         .owner          = THIS_MODULE,
6447         .make_request   = make_request,
6448         .run            = run,
6449         .stop           = stop,
6450         .status         = status,
6451         .error_handler  = error,
6452         .hot_add_disk   = raid5_add_disk,
6453         .hot_remove_disk= raid5_remove_disk,
6454         .spare_active   = raid5_spare_active,
6455         .sync_request   = sync_request,
6456         .resize         = raid5_resize,
6457         .size           = raid5_size,
6458         .check_reshape  = raid5_check_reshape,
6459         .start_reshape  = raid5_start_reshape,
6460         .finish_reshape = raid5_finish_reshape,
6461         .quiesce        = raid5_quiesce,
6462         .takeover       = raid4_takeover,
6463 };
6464
6465 static int __init raid5_init(void)
6466 {
6467         register_md_personality(&raid6_personality);
6468         register_md_personality(&raid5_personality);
6469         register_md_personality(&raid4_personality);
6470         return 0;
6471 }
6472
6473 static void raid5_exit(void)
6474 {
6475         unregister_md_personality(&raid6_personality);
6476         unregister_md_personality(&raid5_personality);
6477         unregister_md_personality(&raid4_personality);
6478 }
6479
6480 module_init(raid5_init);
6481 module_exit(raid5_exit);
6482 MODULE_LICENSE("GPL");
6483 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6484 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6485 MODULE_ALIAS("md-raid5");
6486 MODULE_ALIAS("md-raid4");
6487 MODULE_ALIAS("md-level-5");
6488 MODULE_ALIAS("md-level-4");
6489 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6490 MODULE_ALIAS("md-raid6");
6491 MODULE_ALIAS("md-level-6");
6492
6493 /* This used to be two separate modules, they were: */
6494 MODULE_ALIAS("raid5");
6495 MODULE_ALIAS("raid6");