git.oblomov.eu Git - linux-2.6/blob - 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  *
   6  * RAID-5 management functions.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2, or (at your option)
  11  * any later version.
  12  *
  13  * You should have received a copy of the GNU General Public License
  14  * (for example /usr/src/linux/COPYING); if not, write to the Free
  15  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  16  */
  17
  18
  19 #include <linux/config.h>
  20 #include <linux/module.h>
  21 #include <linux/slab.h>
  22 #include <linux/raid/raid5.h>
  23 #include <linux/highmem.h>
  24 #include <linux/bitops.h>
  25 #include <asm/atomic.h>
  26
  27 #include <linux/raid/bitmap.h>
  28
  29 /*
  30  * Stripe cache
  31  */
  32
  33 #define NR_STRIPES              256
  34 #define STRIPE_SIZE             PAGE_SIZE
  35 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  36 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  37 #define IO_THRESHOLD            1
  38 #define HASH_PAGES              1
  39 #define HASH_PAGES_ORDER        0
  40 #define NR_HASH                 (HASH_PAGES * PAGE_SIZE / sizeof(struct stripe_head *))
  41 #define HASH_MASK               (NR_HASH - 1)
  42
  43 #define stripe_hash(conf, sect) ((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK])
  44
  45 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  46  * order without overlap.  There may be several bio's per stripe+device, and
  47  * a bio could span several devices.
  48  * When walking this list for a particular stripe+device, we must never proceed
  49  * beyond a bio that extends past this device, as the next bio might no longer
  50  * be valid.
  51  * This macro is used to determine the 'next' bio in the list, given the sector
  52  * of the current stripe+device
  53  */
  54 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
  55 /*
  56  * The following can be used to debug the driver
  57  */
  58 #define RAID5_DEBUG     0
  59 #define RAID5_PARANOIA  1
  60 #if RAID5_PARANOIA && defined(CONFIG_SMP)
  61 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
  62 #else
  63 # define CHECK_DEVLOCK()
  64 #endif
  65
  66 #define PRINTK(x...) ((void)(RAID5_DEBUG && printk(x)))
  67 #if RAID5_DEBUG
  68 #define inline
  69 #define __inline__
  70 #endif
  71
  72 static void print_raid5_conf (raid5_conf_t *conf);
  73
  74 static inline void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
  75 {
  76         if (atomic_dec_and_test(&sh->count)) {
  77                 if (!list_empty(&sh->lru))
  78                         BUG();
  79                 if (atomic_read(&conf->active_stripes)==0)
  80                         BUG();
  81                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
  82                         if (test_bit(STRIPE_DELAYED, &sh->state))
  83                                 list_add_tail(&sh->lru, &conf->delayed_list);
  84                         else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
  85                                  conf->seq_write == sh->bm_seq)
  86                                 list_add_tail(&sh->lru, &conf->bitmap_list);
  87                         else {
  88                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
  89                                 list_add_tail(&sh->lru, &conf->handle_list);
  90                         }
  91                         md_wakeup_thread(conf->mddev->thread);
  92                 } else {
  93                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
  94                                 atomic_dec(&conf->preread_active_stripes);
  95                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
  96                                         md_wakeup_thread(conf->mddev->thread);
  97                         }
  98                         list_add_tail(&sh->lru, &conf->inactive_list);
  99                         atomic_dec(&conf->active_stripes);
 100                         if (!conf->inactive_blocked ||
 101                             atomic_read(&conf->active_stripes) < (NR_STRIPES*3/4))
 102                                 wake_up(&conf->wait_for_stripe);
 103                 }
 104         }
 105 }
 106 static void release_stripe(struct stripe_head *sh)
 107 {
 108         raid5_conf_t *conf = sh->raid_conf;
 109         unsigned long flags;
 110
 111         spin_lock_irqsave(&conf->device_lock, flags);
 112         __release_stripe(conf, sh);
 113         spin_unlock_irqrestore(&conf->device_lock, flags);
 114 }
 115
 116 static void remove_hash(struct stripe_head *sh)
 117 {
 118         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 119
 120         if (sh->hash_pprev) {
 121                 if (sh->hash_next)
 122                         sh->hash_next->hash_pprev = sh->hash_pprev;
 123                 *sh->hash_pprev = sh->hash_next;
 124                 sh->hash_pprev = NULL;
 125         }
 126 }
 127
 128 static __inline__ void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
 129 {
 130         struct stripe_head **shp = &stripe_hash(conf, sh->sector);
 131
 132         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 133
 134         CHECK_DEVLOCK();
 135         if ((sh->hash_next = *shp) != NULL)
 136                 (*shp)->hash_pprev = &sh->hash_next;
 137         *shp = sh;
 138         sh->hash_pprev = shp;
 139 }
 140
 141
 142 /* find an idle stripe, make sure it is unhashed, and return it. */
 143 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
 144 {
 145         struct stripe_head *sh = NULL;
 146         struct list_head *first;
 147
 148         CHECK_DEVLOCK();
 149         if (list_empty(&conf->inactive_list))
 150                 goto out;
 151         first = conf->inactive_list.next;
 152         sh = list_entry(first, struct stripe_head, lru);
 153         list_del_init(first);
 154         remove_hash(sh);
 155         atomic_inc(&conf->active_stripes);
 156 out:
 157         return sh;
 158 }
 159
 160 static void shrink_buffers(struct stripe_head *sh, int num)
 161 {
 162         struct page *p;
 163         int i;
 164
 165         for (i=0; i<num ; i++) {
 166                 p = sh->dev[i].page;
 167                 if (!p)
 168                         continue;
 169                 sh->dev[i].page = NULL;
 170                 page_cache_release(p);
 171         }
 172 }
 173
 174 static int grow_buffers(struct stripe_head *sh, int num)
 175 {
 176         int i;
 177
 178         for (i=0; i<num; i++) {
 179                 struct page *page;
 180
 181                 if (!(page = alloc_page(GFP_KERNEL))) {
 182                         return 1;
 183                 }
 184                 sh->dev[i].page = page;
 185         }
 186         return 0;
 187 }
 188
 189 static void raid5_build_block (struct stripe_head *sh, int i);
 190
 191 static inline void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx)
 192 {
 193         raid5_conf_t *conf = sh->raid_conf;
 194         int disks = conf->raid_disks, i;
 195
 196         if (atomic_read(&sh->count) != 0)
 197                 BUG();
 198         if (test_bit(STRIPE_HANDLE, &sh->state))
 199                 BUG();
 200
 201         CHECK_DEVLOCK();
 202         PRINTK("init_stripe called, stripe %llu\n",
 203                 (unsigned long long)sh->sector);
 204
 205         remove_hash(sh);
 206
 207         sh->sector = sector;
 208         sh->pd_idx = pd_idx;
 209         sh->state = 0;
 210
 211         for (i=disks; i--; ) {
 212                 struct r5dev *dev = &sh->dev[i];
 213
 214                 if (dev->toread || dev->towrite || dev->written ||
 215                     test_bit(R5_LOCKED, &dev->flags)) {
 216                         printk("sector=%llx i=%d %p %p %p %d\n",
 217                                (unsigned long long)sh->sector, i, dev->toread,
 218                                dev->towrite, dev->written,
 219                                test_bit(R5_LOCKED, &dev->flags));
 220                         BUG();
 221                 }
 222                 dev->flags = 0;
 223                 raid5_build_block(sh, i);
 224         }
 225         insert_hash(conf, sh);
 226 }
 227
 228 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector)
 229 {
 230         struct stripe_head *sh;
 231
 232         CHECK_DEVLOCK();
 233         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
 234         for (sh = stripe_hash(conf, sector); sh; sh = sh->hash_next)
 235                 if (sh->sector == sector)
 236                         return sh;
 237         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
 238         return NULL;
 239 }
 240
 241 static void unplug_slaves(mddev_t *mddev);
 242 static void raid5_unplug_device(request_queue_t *q);
 243
 244 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector,
 245                                              int pd_idx, int noblock)
 246 {
 247         struct stripe_head *sh;
 248
 249         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
 250
 251         spin_lock_irq(&conf->device_lock);
 252
 253         do {
 254                 wait_event_lock_irq(conf->wait_for_stripe,
 255                                     conf->quiesce == 0,
 256                                     conf->device_lock, /* nothing */);
 257                 sh = __find_stripe(conf, sector);
 258                 if (!sh) {
 259                         if (!conf->inactive_blocked)
 260                                 sh = get_free_stripe(conf);
 261                         if (noblock && sh == NULL)
 262                                 break;
 263                         if (!sh) {
 264                                 conf->inactive_blocked = 1;
 265                                 wait_event_lock_irq(conf->wait_for_stripe,
 266                                                     !list_empty(&conf->inactive_list) &&
 267                                                     (atomic_read(&conf->active_stripes) < (NR_STRIPES *3/4)
 268                                                      || !conf->inactive_blocked),
 269                                                     conf->device_lock,
 270                                                     unplug_slaves(conf->mddev);
 271                                         );
 272                                 conf->inactive_blocked = 0;
 273                         } else
 274                                 init_stripe(sh, sector, pd_idx);
 275                 } else {
 276                         if (atomic_read(&sh->count)) {
 277                                 if (!list_empty(&sh->lru))
 278                                         BUG();
 279                         } else {
 280                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 281                                         atomic_inc(&conf->active_stripes);
 282                                 if (list_empty(&sh->lru))
 283                                         BUG();
 284                                 list_del_init(&sh->lru);
 285                         }
 286                 }
 287         } while (sh == NULL);
 288
 289         if (sh)
 290                 atomic_inc(&sh->count);
 291
 292         spin_unlock_irq(&conf->device_lock);
 293         return sh;
 294 }
 295
 296 static int grow_stripes(raid5_conf_t *conf, int num)
 297 {
 298         struct stripe_head *sh;
 299         kmem_cache_t *sc;
 300         int devs = conf->raid_disks;
 301
 302         sprintf(conf->cache_name, "raid5/%s", mdname(conf->mddev));
 303
 304         sc = kmem_cache_create(conf->cache_name,
 305                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
 306                                0, 0, NULL, NULL);
 307         if (!sc)
 308                 return 1;
 309         conf->slab_cache = sc;
 310         while (num--) {
 311                 sh = kmem_cache_alloc(sc, GFP_KERNEL);
 312                 if (!sh)
 313                         return 1;
 314                 memset(sh, 0, sizeof(*sh) + (devs-1)*sizeof(struct r5dev));
 315                 sh->raid_conf = conf;
 316                 spin_lock_init(&sh->lock);
 317
 318                 if (grow_buffers(sh, conf->raid_disks)) {
 319                         shrink_buffers(sh, conf->raid_disks);
 320                         kmem_cache_free(sc, sh);
 321                         return 1;
 322                 }
 323                 /* we just created an active stripe so... */
 324                 atomic_set(&sh->count, 1);
 325                 atomic_inc(&conf->active_stripes);
 326                 INIT_LIST_HEAD(&sh->lru);
 327                 release_stripe(sh);
 328         }
 329         return 0;
 330 }
 331
 332 static void shrink_stripes(raid5_conf_t *conf)
 333 {
 334         struct stripe_head *sh;
 335
 336         while (1) {
 337                 spin_lock_irq(&conf->device_lock);
 338                 sh = get_free_stripe(conf);
 339                 spin_unlock_irq(&conf->device_lock);
 340                 if (!sh)
 341                         break;
 342                 if (atomic_read(&sh->count))
 343                         BUG();
 344                 shrink_buffers(sh, conf->raid_disks);
 345                 kmem_cache_free(conf->slab_cache, sh);
 346                 atomic_dec(&conf->active_stripes);
 347         }
 348         kmem_cache_destroy(conf->slab_cache);
 349         conf->slab_cache = NULL;
 350 }
 351
 352 static int raid5_end_read_request (struct bio * bi, unsigned int bytes_done,
 353                                    int error)
 354 {
 355         struct stripe_head *sh = bi->bi_private;
 356         raid5_conf_t *conf = sh->raid_conf;
 357         int disks = conf->raid_disks, i;
 358         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 359
 360         if (bi->bi_size)
 361                 return 1;
 362
 363         for (i=0 ; i<disks; i++)
 364                 if (bi == &sh->dev[i].req)
 365                         break;
 366
 367         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
 368                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 369                 uptodate);
 370         if (i == disks) {
 371                 BUG();
 372                 return 0;
 373         }
 374
 375         if (uptodate) {
 376 #if 0
 377                 struct bio *bio;
 378                 unsigned long flags;
 379                 spin_lock_irqsave(&conf->device_lock, flags);
 380                 /* we can return a buffer if we bypassed the cache or
 381                  * if the top buffer is not in highmem.  If there are
 382                  * multiple buffers, leave the extra work to
 383                  * handle_stripe
 384                  */
 385                 buffer = sh->bh_read[i];
 386                 if (buffer &&
 387                     (!PageHighMem(buffer->b_page)
 388                      || buffer->b_page == bh->b_page )
 389                         ) {
 390                         sh->bh_read[i] = buffer->b_reqnext;
 391                         buffer->b_reqnext = NULL;
 392                 } else
 393                         buffer = NULL;
 394                 spin_unlock_irqrestore(&conf->device_lock, flags);
 395                 if (sh->bh_page[i]==bh->b_page)
 396                         set_buffer_uptodate(bh);
 397                 if (buffer) {
 398                         if (buffer->b_page != bh->b_page)
 399                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
 400                         buffer->b_end_io(buffer, 1);
 401                 }
 402 #else
 403                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
 404 #endif
 405         } else {
 406                 md_error(conf->mddev, conf->disks[i].rdev);
 407                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
 408         }
 409         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 410 #if 0
 411         /* must restore b_page before unlocking buffer... */
 412         if (sh->bh_page[i] != bh->b_page) {
 413                 bh->b_page = sh->bh_page[i];
 414                 bh->b_data = page_address(bh->b_page);
 415                 clear_buffer_uptodate(bh);
 416         }
 417 #endif
 418         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 419         set_bit(STRIPE_HANDLE, &sh->state);
 420         release_stripe(sh);
 421         return 0;
 422 }
 423
 424 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
 425                                     int error)
 426 {
 427         struct stripe_head *sh = bi->bi_private;
 428         raid5_conf_t *conf = sh->raid_conf;
 429         int disks = conf->raid_disks, i;
 430         unsigned long flags;
 431         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 432
 433         if (bi->bi_size)
 434                 return 1;
 435
 436         for (i=0 ; i<disks; i++)
 437                 if (bi == &sh->dev[i].req)
 438                         break;
 439
 440         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
 441                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 442                 uptodate);
 443         if (i == disks) {
 444                 BUG();
 445                 return 0;
 446         }
 447
 448         spin_lock_irqsave(&conf->device_lock, flags);
 449         if (!uptodate)
 450                 md_error(conf->mddev, conf->disks[i].rdev);
 451
 452         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 453
 454         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 455         set_bit(STRIPE_HANDLE, &sh->state);
 456         __release_stripe(conf, sh);
 457         spin_unlock_irqrestore(&conf->device_lock, flags);
 458         return 0;
 459 }
 460
 461
 462 static sector_t compute_blocknr(struct stripe_head *sh, int i);
 463
 464 static void raid5_build_block (struct stripe_head *sh, int i)
 465 {
 466         struct r5dev *dev = &sh->dev[i];
 467
 468         bio_init(&dev->req);
 469         dev->req.bi_io_vec = &dev->vec;
 470         dev->req.bi_vcnt++;
 471         dev->req.bi_max_vecs++;
 472         dev->vec.bv_page = dev->page;
 473         dev->vec.bv_len = STRIPE_SIZE;
 474         dev->vec.bv_offset = 0;
 475
 476         dev->req.bi_sector = sh->sector;
 477         dev->req.bi_private = sh;
 478
 479         dev->flags = 0;
 480         if (i != sh->pd_idx)
 481                 dev->sector = compute_blocknr(sh, i);
 482 }
 483
 484 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 485 {
 486         char b[BDEVNAME_SIZE];
 487         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
 488         PRINTK("raid5: error called\n");
 489
 490         if (!rdev->faulty) {
 491                 mddev->sb_dirty = 1;
 492                 if (rdev->in_sync) {
 493                         conf->working_disks--;
 494                         mddev->degraded++;
 495                         conf->failed_disks++;
 496                         rdev->in_sync = 0;
 497                         /*
 498                          * if recovery was running, make sure it aborts.
 499                          */
 500                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 501                 }
 502                 rdev->faulty = 1;
 503                 printk (KERN_ALERT
 504                         "raid5: Disk failure on %s, disabling device."
 505                         " Operation continuing on %d devices\n",
 506                         bdevname(rdev->bdev,b), conf->working_disks);
 507         }
 508 }
 509
 510 /*
 511  * Input: a 'big' sector number,
 512  * Output: index of the data and parity disk, and the sector # in them.
 513  */
 514 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
 515                         unsigned int data_disks, unsigned int * dd_idx,
 516                         unsigned int * pd_idx, raid5_conf_t *conf)
 517 {
 518         long stripe;
 519         unsigned long chunk_number;
 520         unsigned int chunk_offset;
 521         sector_t new_sector;
 522         int sectors_per_chunk = conf->chunk_size >> 9;
 523
 524         /* First compute the information on this sector */
 525
 526         /*
 527          * Compute the chunk number and the sector offset inside the chunk
 528          */
 529         chunk_offset = sector_div(r_sector, sectors_per_chunk);
 530         chunk_number = r_sector;
 531         BUG_ON(r_sector != chunk_number);
 532
 533         /*
 534          * Compute the stripe number
 535          */
 536         stripe = chunk_number / data_disks;
 537
 538         /*
 539          * Compute the data disk and parity disk indexes inside the stripe
 540          */
 541         *dd_idx = chunk_number % data_disks;
 542
 543         /*
 544          * Select the parity disk based on the user selected algorithm.
 545          */
 546         if (conf->level == 4)
 547                 *pd_idx = data_disks;
 548         else switch (conf->algorithm) {
 549                 case ALGORITHM_LEFT_ASYMMETRIC:
 550                         *pd_idx = data_disks - stripe % raid_disks;
 551                         if (*dd_idx >= *pd_idx)
 552                                 (*dd_idx)++;
 553                         break;
 554                 case ALGORITHM_RIGHT_ASYMMETRIC:
 555                         *pd_idx = stripe % raid_disks;
 556                         if (*dd_idx >= *pd_idx)
 557                                 (*dd_idx)++;
 558                         break;
 559                 case ALGORITHM_LEFT_SYMMETRIC:
 560                         *pd_idx = data_disks - stripe % raid_disks;
 561                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 562                         break;
 563                 case ALGORITHM_RIGHT_SYMMETRIC:
 564                         *pd_idx = stripe % raid_disks;
 565                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 566                         break;
 567                 default:
 568                         printk("raid5: unsupported algorithm %d\n",
 569                                 conf->algorithm);
 570         }
 571
 572         /*
 573          * Finally, compute the new sector number
 574          */
 575         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
 576         return new_sector;
 577 }
 578
 579
 580 static sector_t compute_blocknr(struct stripe_head *sh, int i)
 581 {
 582         raid5_conf_t *conf = sh->raid_conf;
 583         int raid_disks = conf->raid_disks, data_disks = raid_disks - 1;
 584         sector_t new_sector = sh->sector, check;
 585         int sectors_per_chunk = conf->chunk_size >> 9;
 586         sector_t stripe;
 587         int chunk_offset;
 588         int chunk_number, dummy1, dummy2, dd_idx = i;
 589         sector_t r_sector;
 590
 591         chunk_offset = sector_div(new_sector, sectors_per_chunk);
 592         stripe = new_sector;
 593         BUG_ON(new_sector != stripe);
 594
 595
 596         switch (conf->algorithm) {
 597                 case ALGORITHM_LEFT_ASYMMETRIC:
 598                 case ALGORITHM_RIGHT_ASYMMETRIC:
 599                         if (i > sh->pd_idx)
 600                                 i--;
 601                         break;
 602                 case ALGORITHM_LEFT_SYMMETRIC:
 603                 case ALGORITHM_RIGHT_SYMMETRIC:
 604                         if (i < sh->pd_idx)
 605                                 i += raid_disks;
 606                         i -= (sh->pd_idx + 1);
 607                         break;
 608                 default:
 609                         printk("raid5: unsupported algorithm %d\n",
 610                                 conf->algorithm);
 611         }
 612
 613         chunk_number = stripe * data_disks + i;
 614         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
 615
 616         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
 617         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
 618                 printk("compute_blocknr: map not correct\n");
 619                 return 0;
 620         }
 621         return r_sector;
 622 }
 623
 624
 625
 626 /*
 627  * Copy data between a page in the stripe cache, and a bio.
 628  * There are no alignment or size guarantees between the page or the
 629  * bio except that there is some overlap.
 630  * All iovecs in the bio must be considered.
 631  */
 632 static void copy_data(int frombio, struct bio *bio,
 633                      struct page *page,
 634                      sector_t sector)
 635 {
 636         char *pa = page_address(page);
 637         struct bio_vec *bvl;
 638         int i;
 639         int page_offset;
 640
 641         if (bio->bi_sector >= sector)
 642                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 643         else
 644                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 645         bio_for_each_segment(bvl, bio, i) {
 646                 int len = bio_iovec_idx(bio,i)->bv_len;
 647                 int clen;
 648                 int b_offset = 0;
 649
 650                 if (page_offset < 0) {
 651                         b_offset = -page_offset;
 652                         page_offset += b_offset;
 653                         len -= b_offset;
 654                 }
 655
 656                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 657                         clen = STRIPE_SIZE - page_offset;
 658                 else clen = len;
 659
 660                 if (clen > 0) {
 661                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
 662                         if (frombio)
 663                                 memcpy(pa+page_offset, ba+b_offset, clen);
 664                         else
 665                                 memcpy(ba+b_offset, pa+page_offset, clen);
 666                         __bio_kunmap_atomic(ba, KM_USER0);
 667                 }
 668                 if (clen < len) /* hit end of page */
 669                         break;
 670                 page_offset +=  len;
 671         }
 672 }
 673
 674 #define check_xor()     do {                                            \
 675                            if (count == MAX_XOR_BLOCKS) {               \
 676                                 xor_block(count, STRIPE_SIZE, ptr);     \
 677                                 count = 1;                              \
 678                            }                                            \
 679                         } while(0)
 680
 681
 682 static void compute_block(struct stripe_head *sh, int dd_idx)
 683 {
 684         raid5_conf_t *conf = sh->raid_conf;
 685         int i, count, disks = conf->raid_disks;
 686         void *ptr[MAX_XOR_BLOCKS], *p;
 687
 688         PRINTK("compute_block, stripe %llu, idx %d\n",
 689                 (unsigned long long)sh->sector, dd_idx);
 690
 691         ptr[0] = page_address(sh->dev[dd_idx].page);
 692         memset(ptr[0], 0, STRIPE_SIZE);
 693         count = 1;
 694         for (i = disks ; i--; ) {
 695                 if (i == dd_idx)
 696                         continue;
 697                 p = page_address(sh->dev[i].page);
 698                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
 699                         ptr[count++] = p;
 700                 else
 701                         printk("compute_block() %d, stripe %llu, %d"
 702                                 " not present\n", dd_idx,
 703                                 (unsigned long long)sh->sector, i);
 704
 705                 check_xor();
 706         }
 707         if (count != 1)
 708                 xor_block(count, STRIPE_SIZE, ptr);
 709         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
 710 }
 711
 712 static void compute_parity(struct stripe_head *sh, int method)
 713 {
 714         raid5_conf_t *conf = sh->raid_conf;
 715         int i, pd_idx = sh->pd_idx, disks = conf->raid_disks, count;
 716         void *ptr[MAX_XOR_BLOCKS];
 717         struct bio *chosen;
 718
 719         PRINTK("compute_parity, stripe %llu, method %d\n",
 720                 (unsigned long long)sh->sector, method);
 721
 722         count = 1;
 723         ptr[0] = page_address(sh->dev[pd_idx].page);
 724         switch(method) {
 725         case READ_MODIFY_WRITE:
 726                 if (!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags))
 727                         BUG();
 728                 for (i=disks ; i-- ;) {
 729                         if (i==pd_idx)
 730                                 continue;
 731                         if (sh->dev[i].towrite &&
 732                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
 733                                 ptr[count++] = page_address(sh->dev[i].page);
 734                                 chosen = sh->dev[i].towrite;
 735                                 sh->dev[i].towrite = NULL;
 736
 737                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 738                                         wake_up(&conf->wait_for_overlap);
 739
 740                                 if (sh->dev[i].written) BUG();
 741                                 sh->dev[i].written = chosen;
 742                                 check_xor();
 743                         }
 744                 }
 745                 break;
 746         case RECONSTRUCT_WRITE:
 747                 memset(ptr[0], 0, STRIPE_SIZE);
 748                 for (i= disks; i-- ;)
 749                         if (i!=pd_idx && sh->dev[i].towrite) {
 750                                 chosen = sh->dev[i].towrite;
 751                                 sh->dev[i].towrite = NULL;
 752
 753                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 754                                         wake_up(&conf->wait_for_overlap);
 755
 756                                 if (sh->dev[i].written) BUG();
 757                                 sh->dev[i].written = chosen;
 758                         }
 759                 break;
 760         case CHECK_PARITY:
 761                 break;
 762         }
 763         if (count>1) {
 764                 xor_block(count, STRIPE_SIZE, ptr);
 765                 count = 1;
 766         }
 767
 768         for (i = disks; i--;)
 769                 if (sh->dev[i].written) {
 770                         sector_t sector = sh->dev[i].sector;
 771                         struct bio *wbi = sh->dev[i].written;
 772                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
 773                                 copy_data(1, wbi, sh->dev[i].page, sector);
 774                                 wbi = r5_next_bio(wbi, sector);
 775                         }
 776
 777                         set_bit(R5_LOCKED, &sh->dev[i].flags);
 778                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
 779                 }
 780
 781         switch(method) {
 782         case RECONSTRUCT_WRITE:
 783         case CHECK_PARITY:
 784                 for (i=disks; i--;)
 785                         if (i != pd_idx) {
 786                                 ptr[count++] = page_address(sh->dev[i].page);
 787                                 check_xor();
 788                         }
 789                 break;
 790         case READ_MODIFY_WRITE:
 791                 for (i = disks; i--;)
 792                         if (sh->dev[i].written) {
 793                                 ptr[count++] = page_address(sh->dev[i].page);
 794                                 check_xor();
 795                         }
 796         }
 797         if (count != 1)
 798                 xor_block(count, STRIPE_SIZE, ptr);
 799
 800         if (method != CHECK_PARITY) {
 801                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 802                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
 803         } else
 804                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 805 }
 806
 807 /*
 808  * Each stripe/dev can have one or more bion attached.
 809  * toread/towrite point to the first in a chain.
 810  * The bi_next chain must be in order.
 811  */
 812 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
 813 {
 814         struct bio **bip;
 815         raid5_conf_t *conf = sh->raid_conf;
 816         int firstwrite=0;
 817
 818         PRINTK("adding bh b#%llu to stripe s#%llu\n",
 819                 (unsigned long long)bi->bi_sector,
 820                 (unsigned long long)sh->sector);
 821
 822
 823         spin_lock(&sh->lock);
 824         spin_lock_irq(&conf->device_lock);
 825         if (forwrite) {
 826                 bip = &sh->dev[dd_idx].towrite;
 827                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
 828                         firstwrite = 1;
 829         } else
 830                 bip = &sh->dev[dd_idx].toread;
 831         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
 832                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
 833                         goto overlap;
 834                 bip = & (*bip)->bi_next;
 835         }
 836         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
 837                 goto overlap;
 838
 839         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
 840                 BUG();
 841         if (*bip)
 842                 bi->bi_next = *bip;
 843         *bip = bi;
 844         bi->bi_phys_segments ++;
 845         spin_unlock_irq(&conf->device_lock);
 846         spin_unlock(&sh->lock);
 847
 848         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
 849                 (unsigned long long)bi->bi_sector,
 850                 (unsigned long long)sh->sector, dd_idx);
 851
 852         if (conf->mddev->bitmap && firstwrite) {
 853                 sh->bm_seq = conf->seq_write;
 854                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
 855                                   STRIPE_SECTORS, 0);
 856                 set_bit(STRIPE_BIT_DELAY, &sh->state);
 857         }
 858
 859         if (forwrite) {
 860                 /* check if page is covered */
 861                 sector_t sector = sh->dev[dd_idx].sector;
 862                 for (bi=sh->dev[dd_idx].towrite;
 863                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
 864                              bi && bi->bi_sector <= sector;
 865                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
 866                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
 867                                 sector = bi->bi_sector + (bi->bi_size>>9);
 868                 }
 869                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
 870                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
 871         }
 872         return 1;
 873
 874  overlap:
 875         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
 876         spin_unlock_irq(&conf->device_lock);
 877         spin_unlock(&sh->lock);
 878         return 0;
 879 }
 880
 881
 882 /*
 883  * handle_stripe - do things to a stripe.
 884  *
 885  * We lock the stripe and then examine the state of various bits
 886  * to see what needs to be done.
 887  * Possible results:
 888  *    return some read request which now have data
 889  *    return some write requests which are safely on disc
 890  *    schedule a read on some buffers
 891  *    schedule a write of some buffers
 892  *    return confirmation of parity correctness
 893  *
 894  * Parity calculations are done inside the stripe lock
 895  * buffers are taken off read_list or write_list, and bh_cache buffers
 896  * get BH_Lock set before the stripe lock is released.
 897  *
 898  */
 899
 900 static void handle_stripe(struct stripe_head *sh)
 901 {
 902         raid5_conf_t *conf = sh->raid_conf;
 903         int disks = conf->raid_disks;
 904         struct bio *return_bi= NULL;
 905         struct bio *bi;
 906         int i;
 907         int syncing;
 908         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
 909         int non_overwrite = 0;
 910         int failed_num=0;
 911         struct r5dev *dev;
 912
 913         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
 914                 (unsigned long long)sh->sector, atomic_read(&sh->count),
 915                 sh->pd_idx);
 916
 917         spin_lock(&sh->lock);
 918         clear_bit(STRIPE_HANDLE, &sh->state);
 919         clear_bit(STRIPE_DELAYED, &sh->state);
 920
 921         syncing = test_bit(STRIPE_SYNCING, &sh->state);
 922         /* Now to look around and see what can be done */
 923
 924         for (i=disks; i--; ) {
 925                 mdk_rdev_t *rdev;
 926                 dev = &sh->dev[i];
 927                 clear_bit(R5_Insync, &dev->flags);
 928                 clear_bit(R5_Syncio, &dev->flags);
 929
 930                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
 931                         i, dev->flags, dev->toread, dev->towrite, dev->written);
 932                 /* maybe we can reply to a read */
 933                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
 934                         struct bio *rbi, *rbi2;
 935                         PRINTK("Return read for disc %d\n", i);
 936                         spin_lock_irq(&conf->device_lock);
 937                         rbi = dev->toread;
 938                         dev->toread = NULL;
 939                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
 940                                 wake_up(&conf->wait_for_overlap);
 941                         spin_unlock_irq(&conf->device_lock);
 942                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
 943                                 copy_data(0, rbi, dev->page, dev->sector);
 944                                 rbi2 = r5_next_bio(rbi, dev->sector);
 945                                 spin_lock_irq(&conf->device_lock);
 946                                 if (--rbi->bi_phys_segments == 0) {
 947                                         rbi->bi_next = return_bi;
 948                                         return_bi = rbi;
 949                                 }
 950                                 spin_unlock_irq(&conf->device_lock);
 951                                 rbi = rbi2;
 952                         }
 953                 }
 954
 955                 /* now count some things */
 956                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
 957                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
 958
 959
 960                 if (dev->toread) to_read++;
 961                 if (dev->towrite) {
 962                         to_write++;
 963                         if (!test_bit(R5_OVERWRITE, &dev->flags))
 964                                 non_overwrite++;
 965                 }
 966                 if (dev->written) written++;
 967                 rdev = conf->disks[i].rdev; /* FIXME, should I be looking rdev */
 968                 if (!rdev || !rdev->in_sync) {
 969                         failed++;
 970                         failed_num = i;
 971                 } else
 972                         set_bit(R5_Insync, &dev->flags);
 973         }
 974         PRINTK("locked=%d uptodate=%d to_read=%d"
 975                 " to_write=%d failed=%d failed_num=%d\n",
 976                 locked, uptodate, to_read, to_write, failed, failed_num);
 977         /* check if the array has lost two devices and, if so, some requests might
 978          * need to be failed
 979          */
 980         if (failed > 1 && to_read+to_write+written) {
 981                 for (i=disks; i--; ) {
 982                         int bitmap_end = 0;
 983                         spin_lock_irq(&conf->device_lock);
 984                         /* fail all writes first */
 985                         bi = sh->dev[i].towrite;
 986                         sh->dev[i].towrite = NULL;
 987                         if (bi) { to_write--; bitmap_end = 1; }
 988
 989                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 990                                 wake_up(&conf->wait_for_overlap);
 991
 992                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
 993                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
 994                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
 995                                 if (--bi->bi_phys_segments == 0) {
 996                                         md_write_end(conf->mddev);
 997                                         bi->bi_next = return_bi;
 998                                         return_bi = bi;
 999                                 }
1000                                 bi = nextbi;
1001                         }
1002                         /* and fail all 'written' */
1003                         bi = sh->dev[i].written;
1004                         sh->dev[i].written = NULL;
1005                         if (bi) bitmap_end = 1;
1006                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1007                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1008                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1009                                 if (--bi->bi_phys_segments == 0) {
1010                                         md_write_end(conf->mddev);
1011                                         bi->bi_next = return_bi;
1012                                         return_bi = bi;
1013                                 }
1014                                 bi = bi2;
1015                         }
1016
1017                         /* fail any reads if this device is non-operational */
1018                         if (!test_bit(R5_Insync, &sh->dev[i].flags)) {
1019                                 bi = sh->dev[i].toread;
1020                                 sh->dev[i].toread = NULL;
1021                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1022                                         wake_up(&conf->wait_for_overlap);
1023                                 if (bi) to_read--;
1024                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1025                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1026                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1027                                         if (--bi->bi_phys_segments == 0) {
1028                                                 bi->bi_next = return_bi;
1029                                                 return_bi = bi;
1030                                         }
1031                                         bi = nextbi;
1032                                 }
1033                         }
1034                         spin_unlock_irq(&conf->device_lock);
1035                         if (bitmap_end)
1036                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1037                                                 STRIPE_SECTORS, 0, 0);
1038                 }
1039         }
1040         if (failed > 1 && syncing) {
1041                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1042                 clear_bit(STRIPE_SYNCING, &sh->state);
1043                 syncing = 0;
1044         }
1045
1046         /* might be able to return some write requests if the parity block
1047          * is safe, or on a failed drive
1048          */
1049         dev = &sh->dev[sh->pd_idx];
1050         if ( written &&
1051              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1052                 test_bit(R5_UPTODATE, &dev->flags))
1053                || (failed == 1 && failed_num == sh->pd_idx))
1054             ) {
1055             /* any written block on an uptodate or failed drive can be returned.
1056              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
1057              * never LOCKED, so we don't need to test 'failed' directly.
1058              */
1059             for (i=disks; i--; )
1060                 if (sh->dev[i].written) {
1061                     dev = &sh->dev[i];
1062                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1063                          test_bit(R5_UPTODATE, &dev->flags) ) {
1064                         /* We can return any write requests */
1065                             struct bio *wbi, *wbi2;
1066                             int bitmap_end = 0;
1067                             PRINTK("Return write for disc %d\n", i);
1068                             spin_lock_irq(&conf->device_lock);
1069                             wbi = dev->written;
1070                             dev->written = NULL;
1071                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1072                                     wbi2 = r5_next_bio(wbi, dev->sector);
1073                                     if (--wbi->bi_phys_segments == 0) {
1074                                             md_write_end(conf->mddev);
1075                                             wbi->bi_next = return_bi;
1076                                             return_bi = wbi;
1077                                     }
1078                                     wbi = wbi2;
1079                             }
1080                             if (dev->towrite == NULL)
1081                                     bitmap_end = 1;
1082                             spin_unlock_irq(&conf->device_lock);
1083                             if (bitmap_end)
1084                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1085                                                     STRIPE_SECTORS,
1086                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1087                     }
1088                 }
1089         }
1090
1091         /* Now we might consider reading some blocks, either to check/generate
1092          * parity, or to satisfy requests
1093          * or to load a block that is being partially written.
1094          */
1095         if (to_read || non_overwrite || (syncing && (uptodate < disks))) {
1096                 for (i=disks; i--;) {
1097                         dev = &sh->dev[i];
1098                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1099                             (dev->toread ||
1100                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1101                              syncing ||
1102                              (failed && (sh->dev[failed_num].toread ||
1103                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1104                                     )
1105                                 ) {
1106                                 /* we would like to get this block, possibly
1107                                  * by computing it, but we might not be able to
1108                                  */
1109                                 if (uptodate == disks-1) {
1110                                         PRINTK("Computing block %d\n", i);
1111                                         compute_block(sh, i);
1112                                         uptodate++;
1113                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1114                                         set_bit(R5_LOCKED, &dev->flags);
1115                                         set_bit(R5_Wantread, &dev->flags);
1116 #if 0
1117                                         /* if I am just reading this block and we don't have
1118                                            a failed drive, or any pending writes then sidestep the cache */
1119                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1120                                             ! syncing && !failed && !to_write) {
1121                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1122                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1123                                         }
1124 #endif
1125                                         locked++;
1126                                         PRINTK("Reading block %d (sync=%d)\n",
1127                                                 i, syncing);
1128                                         if (syncing)
1129                                                 md_sync_acct(conf->disks[i].rdev->bdev,
1130                                                              STRIPE_SECTORS);
1131                                 }
1132                         }
1133                 }
1134                 set_bit(STRIPE_HANDLE, &sh->state);
1135         }
1136
1137         /* now to consider writing and what else, if anything should be read */
1138         if (to_write) {
1139                 int rmw=0, rcw=0;
1140                 for (i=disks ; i--;) {
1141                         /* would I have to read this buffer for read_modify_write */
1142                         dev = &sh->dev[i];
1143                         if ((dev->towrite || i == sh->pd_idx) &&
1144                             (!test_bit(R5_LOCKED, &dev->flags)
1145 #if 0
1146 || sh->bh_page[i]!=bh->b_page
1147 #endif
1148                                     ) &&
1149                             !test_bit(R5_UPTODATE, &dev->flags)) {
1150                                 if (test_bit(R5_Insync, &dev->flags)
1151 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1152                                         )
1153                                         rmw++;
1154                                 else rmw += 2*disks;  /* cannot read it */
1155                         }
1156                         /* Would I have to read this buffer for reconstruct_write */
1157                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1158                             (!test_bit(R5_LOCKED, &dev->flags)
1159 #if 0
1160 || sh->bh_page[i] != bh->b_page
1161 #endif
1162                                     ) &&
1163                             !test_bit(R5_UPTODATE, &dev->flags)) {
1164                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1165                                 else rcw += 2*disks;
1166                         }
1167                 }
1168                 PRINTK("for sector %llu, rmw=%d rcw=%d\n",
1169                         (unsigned long long)sh->sector, rmw, rcw);
1170                 set_bit(STRIPE_HANDLE, &sh->state);
1171                 if (rmw < rcw && rmw > 0)
1172                         /* prefer read-modify-write, but need to get some data */
1173                         for (i=disks; i--;) {
1174                                 dev = &sh->dev[i];
1175                                 if ((dev->towrite || i == sh->pd_idx) &&
1176                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1177                                     test_bit(R5_Insync, &dev->flags)) {
1178                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1179                                         {
1180                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1181                                                 set_bit(R5_LOCKED, &dev->flags);
1182                                                 set_bit(R5_Wantread, &dev->flags);
1183                                                 locked++;
1184                                         } else {
1185                                                 set_bit(STRIPE_DELAYED, &sh->state);
1186                                                 set_bit(STRIPE_HANDLE, &sh->state);
1187                                         }
1188                                 }
1189                         }
1190                 if (rcw <= rmw && rcw > 0)
1191                         /* want reconstruct write, but need to get some data */
1192                         for (i=disks; i--;) {
1193                                 dev = &sh->dev[i];
1194                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1195                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1196                                     test_bit(R5_Insync, &dev->flags)) {
1197                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1198                                         {
1199                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1200                                                 set_bit(R5_LOCKED, &dev->flags);
1201                                                 set_bit(R5_Wantread, &dev->flags);
1202                                                 locked++;
1203                                         } else {
1204                                                 set_bit(STRIPE_DELAYED, &sh->state);
1205                                                 set_bit(STRIPE_HANDLE, &sh->state);
1206                                         }
1207                                 }
1208                         }
1209                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1210                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1211                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1212                         PRINTK("Computing parity...\n");
1213                         compute_parity(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1214                         /* now every locked buffer is ready to be written */
1215                         for (i=disks; i--;)
1216                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1217                                         PRINTK("Writing block %d\n", i);
1218                                         locked++;
1219                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1220                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1221                                             || (i==sh->pd_idx && failed == 0))
1222                                                 set_bit(STRIPE_INSYNC, &sh->state);
1223                                 }
1224                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1225                                 atomic_dec(&conf->preread_active_stripes);
1226                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1227                                         md_wakeup_thread(conf->mddev->thread);
1228                         }
1229                 }
1230         }
1231
1232         /* maybe we need to check and possibly fix the parity for this stripe
1233          * Any reads will already have been scheduled, so we just see if enough data
1234          * is available
1235          */
1236         if (syncing && locked == 0 &&
1237             !test_bit(STRIPE_INSYNC, &sh->state) && failed <= 1) {
1238                 set_bit(STRIPE_HANDLE, &sh->state);
1239                 if (failed == 0) {
1240                         char *pagea;
1241                         if (uptodate != disks)
1242                                 BUG();
1243                         compute_parity(sh, CHECK_PARITY);
1244                         uptodate--;
1245                         pagea = page_address(sh->dev[sh->pd_idx].page);
1246                         if ((*(u32*)pagea) == 0 &&
1247                             !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1248                                 /* parity is correct (on disc, not in buffer any more) */
1249                                 set_bit(STRIPE_INSYNC, &sh->state);
1250                         }
1251                 }
1252                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1253                         if (failed==0)
1254                                 failed_num = sh->pd_idx;
1255                         /* should be able to compute the missing block and write it to spare */
1256                         if (!test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)) {
1257                                 if (uptodate+1 != disks)
1258                                         BUG();
1259                                 compute_block(sh, failed_num);
1260                                 uptodate++;
1261                         }
1262                         if (uptodate != disks)
1263                                 BUG();
1264                         dev = &sh->dev[failed_num];
1265                         set_bit(R5_LOCKED, &dev->flags);
1266                         set_bit(R5_Wantwrite, &dev->flags);
1267                         clear_bit(STRIPE_DEGRADED, &sh->state);
1268                         locked++;
1269                         set_bit(STRIPE_INSYNC, &sh->state);
1270                         set_bit(R5_Syncio, &dev->flags);
1271                 }
1272         }
1273         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1274                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1275                 clear_bit(STRIPE_SYNCING, &sh->state);
1276         }
1277
1278         spin_unlock(&sh->lock);
1279
1280         while ((bi=return_bi)) {
1281                 int bytes = bi->bi_size;
1282
1283                 return_bi = bi->bi_next;
1284                 bi->bi_next = NULL;
1285                 bi->bi_size = 0;
1286                 bi->bi_end_io(bi, bytes, 0);
1287         }
1288         for (i=disks; i-- ;) {
1289                 int rw;
1290                 struct bio *bi;
1291                 mdk_rdev_t *rdev;
1292                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1293                         rw = 1;
1294                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1295                         rw = 0;
1296                 else
1297                         continue;
1298
1299                 bi = &sh->dev[i].req;
1300
1301                 bi->bi_rw = rw;
1302                 if (rw)
1303                         bi->bi_end_io = raid5_end_write_request;
1304                 else
1305                         bi->bi_end_io = raid5_end_read_request;
1306
1307                 rcu_read_lock();
1308                 rdev = conf->disks[i].rdev;
1309                 if (rdev && rdev->faulty)
1310                         rdev = NULL;
1311                 if (rdev)
1312                         atomic_inc(&rdev->nr_pending);
1313                 rcu_read_unlock();
1314
1315                 if (rdev) {
1316                         if (test_bit(R5_Syncio, &sh->dev[i].flags))
1317                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1318
1319                         bi->bi_bdev = rdev->bdev;
1320                         PRINTK("for %llu schedule op %ld on disc %d\n",
1321                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1322                         atomic_inc(&sh->count);
1323                         bi->bi_sector = sh->sector + rdev->data_offset;
1324                         bi->bi_flags = 1 << BIO_UPTODATE;
1325                         bi->bi_vcnt = 1;
1326                         bi->bi_max_vecs = 1;
1327                         bi->bi_idx = 0;
1328                         bi->bi_io_vec = &sh->dev[i].vec;
1329                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1330                         bi->bi_io_vec[0].bv_offset = 0;
1331                         bi->bi_size = STRIPE_SIZE;
1332                         bi->bi_next = NULL;
1333                         generic_make_request(bi);
1334                 } else {
1335                         if (rw == 1)
1336                                 set_bit(STRIPE_DEGRADED, &sh->state);
1337                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1338                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1339                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1340                         set_bit(STRIPE_HANDLE, &sh->state);
1341                 }
1342         }
1343 }
1344
1345 static inline void raid5_activate_delayed(raid5_conf_t *conf)
1346 {
1347         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1348                 while (!list_empty(&conf->delayed_list)) {
1349                         struct list_head *l = conf->delayed_list.next;
1350                         struct stripe_head *sh;
1351                         sh = list_entry(l, struct stripe_head, lru);
1352                         list_del_init(l);
1353                         clear_bit(STRIPE_DELAYED, &sh->state);
1354                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1355                                 atomic_inc(&conf->preread_active_stripes);
1356                         list_add_tail(&sh->lru, &conf->handle_list);
1357                 }
1358         }
1359 }
1360
1361 static inline void activate_bit_delay(raid5_conf_t *conf)
1362 {
1363         /* device_lock is held */
1364         struct list_head head;
1365         list_add(&head, &conf->bitmap_list);
1366         list_del_init(&conf->bitmap_list);
1367         while (!list_empty(&head)) {
1368                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1369                 list_del_init(&sh->lru);
1370                 atomic_inc(&sh->count);
1371                 __release_stripe(conf, sh);
1372         }
1373 }
1374
1375 static void unplug_slaves(mddev_t *mddev)
1376 {
1377         raid5_conf_t *conf = mddev_to_conf(mddev);
1378         int i;
1379
1380         rcu_read_lock();
1381         for (i=0; i<mddev->raid_disks; i++) {
1382                 mdk_rdev_t *rdev = conf->disks[i].rdev;
1383                 if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
1384                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1385
1386                         atomic_inc(&rdev->nr_pending);
1387                         rcu_read_unlock();
1388
1389                         if (r_queue->unplug_fn)
1390                                 r_queue->unplug_fn(r_queue);
1391
1392                         rdev_dec_pending(rdev, mddev);
1393                         rcu_read_lock();
1394                 }
1395         }
1396         rcu_read_unlock();
1397 }
1398
1399 static void raid5_unplug_device(request_queue_t *q)
1400 {
1401         mddev_t *mddev = q->queuedata;
1402         raid5_conf_t *conf = mddev_to_conf(mddev);
1403         unsigned long flags;
1404
1405         spin_lock_irqsave(&conf->device_lock, flags);
1406
1407         if (blk_remove_plug(q)) {
1408                 conf->seq_flush++;
1409                 raid5_activate_delayed(conf);
1410         }
1411         md_wakeup_thread(mddev->thread);
1412
1413         spin_unlock_irqrestore(&conf->device_lock, flags);
1414
1415         unplug_slaves(mddev);
1416 }
1417
1418 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
1419                              sector_t *error_sector)
1420 {
1421         mddev_t *mddev = q->queuedata;
1422         raid5_conf_t *conf = mddev_to_conf(mddev);
1423         int i, ret = 0;
1424
1425         rcu_read_lock();
1426         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1427                 mdk_rdev_t *rdev = conf->disks[i].rdev;
1428                 if (rdev && !rdev->faulty) {
1429                         struct block_device *bdev = rdev->bdev;
1430                         request_queue_t *r_queue = bdev_get_queue(bdev);
1431
1432                         if (!r_queue->issue_flush_fn)
1433                                 ret = -EOPNOTSUPP;
1434                         else {
1435                                 atomic_inc(&rdev->nr_pending);
1436                                 rcu_read_unlock();
1437                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1438                                                               error_sector);
1439                                 rdev_dec_pending(rdev, mddev);
1440                                 rcu_read_lock();
1441                         }
1442                 }
1443         }
1444         rcu_read_unlock();
1445         return ret;
1446 }
1447
1448 static inline void raid5_plug_device(raid5_conf_t *conf)
1449 {
1450         spin_lock_irq(&conf->device_lock);
1451         blk_plug_device(conf->mddev->queue);
1452         spin_unlock_irq(&conf->device_lock);
1453 }
1454
1455 static int make_request (request_queue_t *q, struct bio * bi)
1456 {
1457         mddev_t *mddev = q->queuedata;
1458         raid5_conf_t *conf = mddev_to_conf(mddev);
1459         const unsigned int raid_disks = conf->raid_disks;
1460         const unsigned int data_disks = raid_disks - 1;
1461         unsigned int dd_idx, pd_idx;
1462         sector_t new_sector;
1463         sector_t logical_sector, last_sector;
1464         struct stripe_head *sh;
1465
1466         if (unlikely(bio_barrier(bi))) {
1467                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1468                 return 0;
1469         }
1470
1471         md_write_start(mddev, bi);
1472
1473         if (bio_data_dir(bi)==WRITE) {
1474                 disk_stat_inc(mddev->gendisk, writes);
1475                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bi));
1476         } else {
1477                 disk_stat_inc(mddev->gendisk, reads);
1478                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bi));
1479         }
1480
1481         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1482         last_sector = bi->bi_sector + (bi->bi_size>>9);
1483         bi->bi_next = NULL;
1484         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1485
1486         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1487                 DEFINE_WAIT(w);
1488
1489                 new_sector = raid5_compute_sector(logical_sector,
1490                                                   raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1491
1492                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
1493                         (unsigned long long)new_sector,
1494                         (unsigned long long)logical_sector);
1495
1496         retry:
1497                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1498                 sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
1499                 if (sh) {
1500                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1501                                 /* Add failed due to overlap.  Flush everything
1502                                  * and wait a while
1503                                  */
1504                                 raid5_unplug_device(mddev->queue);
1505                                 release_stripe(sh);
1506                                 schedule();
1507                                 goto retry;
1508                         }
1509                         finish_wait(&conf->wait_for_overlap, &w);
1510                         raid5_plug_device(conf);
1511                         handle_stripe(sh);
1512                         release_stripe(sh);
1513
1514                 } else {
1515                         /* cannot get stripe for read-ahead, just give-up */
1516                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1517                         finish_wait(&conf->wait_for_overlap, &w);
1518                         break;
1519                 }
1520
1521         }
1522         spin_lock_irq(&conf->device_lock);
1523         if (--bi->bi_phys_segments == 0) {
1524                 int bytes = bi->bi_size;
1525
1526                 if ( bio_data_dir(bi) == WRITE )
1527                         md_write_end(mddev);
1528                 bi->bi_size = 0;
1529                 bi->bi_end_io(bi, bytes, 0);
1530         }
1531         spin_unlock_irq(&conf->device_lock);
1532         return 0;
1533 }
1534
1535 /* FIXME go_faster isn't used */
1536 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1537 {
1538         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1539         struct stripe_head *sh;
1540         int sectors_per_chunk = conf->chunk_size >> 9;
1541         sector_t x;
1542         unsigned long stripe;
1543         int chunk_offset;
1544         int dd_idx, pd_idx;
1545         sector_t first_sector;
1546         int raid_disks = conf->raid_disks;
1547         int data_disks = raid_disks-1;
1548         sector_t max_sector = mddev->size << 1;
1549         int sync_blocks;
1550
1551         if (sector_nr >= max_sector) {
1552                 /* just being told to finish up .. nothing much to do */
1553                 unplug_slaves(mddev);
1554
1555                 if (mddev->curr_resync < max_sector) /* aborted */
1556                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1557                                         &sync_blocks, 1);
1558                 else /* compelted sync */
1559                         conf->fullsync = 0;
1560                 bitmap_close_sync(mddev->bitmap);
1561
1562                 return 0;
1563         }
1564         /* if there is 1 or more failed drives and we are trying
1565          * to resync, then assert that we are finished, because there is
1566          * nothing we can do.
1567          */
1568         if (mddev->degraded >= 1 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1569                 sector_t rv = (mddev->size << 1) - sector_nr;
1570                 *skipped = 1;
1571                 return rv;
1572         }
1573         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1574             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
1575                 /* we can skip this block, and probably more */
1576                 sync_blocks /= STRIPE_SECTORS;
1577                 *skipped = 1;
1578                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
1579         }
1580
1581         x = sector_nr;
1582         chunk_offset = sector_div(x, sectors_per_chunk);
1583         stripe = x;
1584         BUG_ON(x != stripe);
1585
1586         first_sector = raid5_compute_sector((sector_t)stripe*data_disks*sectors_per_chunk
1587                 + chunk_offset, raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1588         sh = get_active_stripe(conf, sector_nr, pd_idx, 1);
1589         if (sh == NULL) {
1590                 sh = get_active_stripe(conf, sector_nr, pd_idx, 0);
1591                 /* make sure we don't swamp the stripe cache if someone else
1592                  * is trying to get access
1593                  */
1594                 set_current_state(TASK_UNINTERRUPTIBLE);
1595                 schedule_timeout(1);
1596         }
1597         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 0);
1598         spin_lock(&sh->lock);
1599         set_bit(STRIPE_SYNCING, &sh->state);
1600         clear_bit(STRIPE_INSYNC, &sh->state);
1601         spin_unlock(&sh->lock);
1602
1603         handle_stripe(sh);
1604         release_stripe(sh);
1605
1606         return STRIPE_SECTORS;
1607 }
1608
1609 /*
1610  * This is our raid5 kernel thread.
1611  *
1612  * We scan the hash table for stripes which can be handled now.
1613  * During the scan, completed stripes are saved for us by the interrupt
1614  * handler, so that they will not have to wait for our next wakeup.
1615  */
1616 static void raid5d (mddev_t *mddev)
1617 {
1618         struct stripe_head *sh;
1619         raid5_conf_t *conf = mddev_to_conf(mddev);
1620         int handled;
1621
1622         PRINTK("+++ raid5d active\n");
1623
1624         md_check_recovery(mddev);
1625
1626         handled = 0;
1627         spin_lock_irq(&conf->device_lock);
1628         while (1) {
1629                 struct list_head *first;
1630
1631                 if (conf->seq_flush - conf->seq_write > 0) {
1632                         int seq = conf->seq_flush;
1633                         bitmap_unplug(mddev->bitmap);
1634                         conf->seq_write = seq;
1635                         activate_bit_delay(conf);
1636                 }
1637
1638                 if (list_empty(&conf->handle_list) &&
1639                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
1640                     !blk_queue_plugged(mddev->queue) &&
1641                     !list_empty(&conf->delayed_list))
1642                         raid5_activate_delayed(conf);
1643
1644                 if (list_empty(&conf->handle_list))
1645                         break;
1646
1647                 first = conf->handle_list.next;
1648                 sh = list_entry(first, struct stripe_head, lru);
1649
1650                 list_del_init(first);
1651                 atomic_inc(&sh->count);
1652                 if (atomic_read(&sh->count)!= 1)
1653                         BUG();
1654                 spin_unlock_irq(&conf->device_lock);
1655
1656                 handled++;
1657                 handle_stripe(sh);
1658                 release_stripe(sh);
1659
1660                 spin_lock_irq(&conf->device_lock);
1661         }
1662         PRINTK("%d stripes handled\n", handled);
1663
1664         spin_unlock_irq(&conf->device_lock);
1665
1666         unplug_slaves(mddev);
1667
1668         PRINTK("--- raid5d inactive\n");
1669 }
1670
1671 static int run(mddev_t *mddev)
1672 {
1673         raid5_conf_t *conf;
1674         int raid_disk, memory;
1675         mdk_rdev_t *rdev;
1676         struct disk_info *disk;
1677         struct list_head *tmp;
1678
1679         if (mddev->level != 5 && mddev->level != 4) {
1680                 printk("raid5: %s: raid level not set to 4/5 (%d)\n", mdname(mddev), mddev->level);
1681                 return -EIO;
1682         }
1683
1684         mddev->private = kmalloc (sizeof (raid5_conf_t)
1685                                   + mddev->raid_disks * sizeof(struct disk_info),
1686                                   GFP_KERNEL);
1687         if ((conf = mddev->private) == NULL)
1688                 goto abort;
1689         memset (conf, 0, sizeof (*conf) + mddev->raid_disks * sizeof(struct disk_info) );
1690         conf->mddev = mddev;
1691
1692         if ((conf->stripe_hashtbl = (struct stripe_head **) __get_free_pages(GFP_ATOMIC, HASH_PAGES_ORDER)) == NULL)
1693                 goto abort;
1694         memset(conf->stripe_hashtbl, 0, HASH_PAGES * PAGE_SIZE);
1695
1696         spin_lock_init(&conf->device_lock);
1697         init_waitqueue_head(&conf->wait_for_stripe);
1698         init_waitqueue_head(&conf->wait_for_overlap);
1699         INIT_LIST_HEAD(&conf->handle_list);
1700         INIT_LIST_HEAD(&conf->delayed_list);
1701         INIT_LIST_HEAD(&conf->bitmap_list);
1702         INIT_LIST_HEAD(&conf->inactive_list);
1703         atomic_set(&conf->active_stripes, 0);
1704         atomic_set(&conf->preread_active_stripes, 0);
1705
1706         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
1707
1708         ITERATE_RDEV(mddev,rdev,tmp) {
1709                 raid_disk = rdev->raid_disk;
1710                 if (raid_disk >= mddev->raid_disks
1711                     || raid_disk < 0)
1712                         continue;
1713                 disk = conf->disks + raid_disk;
1714
1715                 disk->rdev = rdev;
1716
1717                 if (rdev->in_sync) {
1718                         char b[BDEVNAME_SIZE];
1719                         printk(KERN_INFO "raid5: device %s operational as raid"
1720                                 " disk %d\n", bdevname(rdev->bdev,b),
1721                                 raid_disk);
1722                         conf->working_disks++;
1723                 }
1724         }
1725
1726         conf->raid_disks = mddev->raid_disks;
1727         /*
1728          * 0 for a fully functional array, 1 for a degraded array.
1729          */
1730         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
1731         conf->mddev = mddev;
1732         conf->chunk_size = mddev->chunk_size;
1733         conf->level = mddev->level;
1734         conf->algorithm = mddev->layout;
1735         conf->max_nr_stripes = NR_STRIPES;
1736
1737         /* device size must be a multiple of chunk size */
1738         mddev->size &= ~(mddev->chunk_size/1024 -1);
1739         mddev->resync_max_sectors = mddev->size << 1;
1740
1741         if (!conf->chunk_size || conf->chunk_size % 4) {
1742                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
1743                         conf->chunk_size, mdname(mddev));
1744                 goto abort;
1745         }
1746         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
1747                 printk(KERN_ERR
1748                         "raid5: unsupported parity algorithm %d for %s\n",
1749                         conf->algorithm, mdname(mddev));
1750                 goto abort;
1751         }
1752         if (mddev->degraded > 1) {
1753                 printk(KERN_ERR "raid5: not enough operational devices for %s"
1754                         " (%d/%d failed)\n",
1755                         mdname(mddev), conf->failed_disks, conf->raid_disks);
1756                 goto abort;
1757         }
1758
1759         if (mddev->degraded == 1 &&
1760             mddev->recovery_cp != MaxSector) {
1761                 printk(KERN_ERR
1762                         "raid5: cannot start dirty degraded array for %s\n",
1763                         mdname(mddev));
1764                 goto abort;
1765         }
1766
1767         {
1768                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
1769                 if (!mddev->thread) {
1770                         printk(KERN_ERR
1771                                 "raid5: couldn't allocate thread for %s\n",
1772                                 mdname(mddev));
1773                         goto abort;
1774                 }
1775         }
1776 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
1777                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
1778         if (grow_stripes(conf, conf->max_nr_stripes)) {
1779                 printk(KERN_ERR
1780                         "raid5: couldn't allocate %dkB for buffers\n", memory);
1781                 shrink_stripes(conf);
1782                 md_unregister_thread(mddev->thread);
1783                 goto abort;
1784         } else
1785                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
1786                         memory, mdname(mddev));
1787
1788         if (mddev->degraded == 0)
1789                 printk("raid5: raid level %d set %s active with %d out of %d"
1790                         " devices, algorithm %d\n", conf->level, mdname(mddev),
1791                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
1792                         conf->algorithm);
1793         else
1794                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
1795                         " out of %d devices, algorithm %d\n", conf->level,
1796                         mdname(mddev), mddev->raid_disks - mddev->degraded,
1797                         mddev->raid_disks, conf->algorithm);
1798
1799         print_raid5_conf(conf);
1800
1801         /* read-ahead size must cover two whole stripes, which is
1802          * 2 * (n-1) * chunksize where 'n' is the number of raid devices
1803          */
1804         {
1805                 int stripe = (mddev->raid_disks-1) * mddev->chunk_size
1806                         / PAGE_CACHE_SIZE;
1807                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
1808                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
1809         }
1810
1811         /* Ok, everything is just fine now */
1812
1813         if (mddev->bitmap)
1814                 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
1815
1816         mddev->queue->unplug_fn = raid5_unplug_device;
1817         mddev->queue->issue_flush_fn = raid5_issue_flush;
1818
1819         mddev->array_size =  mddev->size * (mddev->raid_disks - 1);
1820         return 0;
1821 abort:
1822         if (conf) {
1823                 print_raid5_conf(conf);
1824                 if (conf->stripe_hashtbl)
1825                         free_pages((unsigned long) conf->stripe_hashtbl,
1826                                                         HASH_PAGES_ORDER);
1827                 kfree(conf);
1828         }
1829         mddev->private = NULL;
1830         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
1831         return -EIO;
1832 }
1833
1834
1835
1836 static int stop (mddev_t *mddev)
1837 {
1838         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1839
1840         md_unregister_thread(mddev->thread);
1841         mddev->thread = NULL;
1842         shrink_stripes(conf);
1843         free_pages((unsigned long) conf->stripe_hashtbl, HASH_PAGES_ORDER);
1844         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1845         kfree(conf);
1846         mddev->private = NULL;
1847         return 0;
1848 }
1849
1850 #if RAID5_DEBUG
1851 static void print_sh (struct stripe_head *sh)
1852 {
1853         int i;
1854
1855         printk("sh %llu, pd_idx %d, state %ld.\n",
1856                 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
1857         printk("sh %llu,  count %d.\n",
1858                 (unsigned long long)sh->sector, atomic_read(&sh->count));
1859         printk("sh %llu, ", (unsigned long long)sh->sector);
1860         for (i = 0; i < sh->raid_conf->raid_disks; i++) {
1861                 printk("(cache%d: %p %ld) ",
1862                         i, sh->dev[i].page, sh->dev[i].flags);
1863         }
1864         printk("\n");
1865 }
1866
1867 static void printall (raid5_conf_t *conf)
1868 {
1869         struct stripe_head *sh;
1870         int i;
1871
1872         spin_lock_irq(&conf->device_lock);
1873         for (i = 0; i < NR_HASH; i++) {
1874                 sh = conf->stripe_hashtbl[i];
1875                 for (; sh; sh = sh->hash_next) {
1876                         if (sh->raid_conf != conf)
1877                                 continue;
1878                         print_sh(sh);
1879                 }
1880         }
1881         spin_unlock_irq(&conf->device_lock);
1882 }
1883 #endif
1884
1885 static void status (struct seq_file *seq, mddev_t *mddev)
1886 {
1887         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1888         int i;
1889
1890         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
1891         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
1892         for (i = 0; i < conf->raid_disks; i++)
1893                 seq_printf (seq, "%s",
1894                                conf->disks[i].rdev &&
1895                                conf->disks[i].rdev->in_sync ? "U" : "_");
1896         seq_printf (seq, "]");
1897 #if RAID5_DEBUG
1898 #define D(x) \
1899         seq_printf (seq, "<"#x":%d>", atomic_read(&conf->x))
1900         printall(conf);
1901 #endif
1902 }
1903
1904 static void print_raid5_conf (raid5_conf_t *conf)
1905 {
1906         int i;
1907         struct disk_info *tmp;
1908
1909         printk("RAID5 conf printout:\n");
1910         if (!conf) {
1911                 printk("(conf==NULL)\n");
1912                 return;
1913         }
1914         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
1915                  conf->working_disks, conf->failed_disks);
1916
1917         for (i = 0; i < conf->raid_disks; i++) {
1918                 char b[BDEVNAME_SIZE];
1919                 tmp = conf->disks + i;
1920                 if (tmp->rdev)
1921                 printk(" disk %d, o:%d, dev:%s\n",
1922                         i, !tmp->rdev->faulty,
1923                         bdevname(tmp->rdev->bdev,b));
1924         }
1925 }
1926
1927 static int raid5_spare_active(mddev_t *mddev)
1928 {
1929         int i;
1930         raid5_conf_t *conf = mddev->private;
1931         struct disk_info *tmp;
1932
1933         for (i = 0; i < conf->raid_disks; i++) {
1934                 tmp = conf->disks + i;
1935                 if (tmp->rdev
1936                     && !tmp->rdev->faulty
1937                     && !tmp->rdev->in_sync) {
1938                         mddev->degraded--;
1939                         conf->failed_disks--;
1940                         conf->working_disks++;
1941                         tmp->rdev->in_sync = 1;
1942                 }
1943         }
1944         print_raid5_conf(conf);
1945         return 0;
1946 }
1947
1948 static int raid5_remove_disk(mddev_t *mddev, int number)
1949 {
1950         raid5_conf_t *conf = mddev->private;
1951         int err = 0;
1952         mdk_rdev_t *rdev;
1953         struct disk_info *p = conf->disks + number;
1954
1955         print_raid5_conf(conf);
1956         rdev = p->rdev;
1957         if (rdev) {
1958                 if (rdev->in_sync ||
1959                     atomic_read(&rdev->nr_pending)) {
1960                         err = -EBUSY;
1961                         goto abort;
1962                 }
1963                 p->rdev = NULL;
1964                 synchronize_rcu();
1965                 if (atomic_read(&rdev->nr_pending)) {
1966                         /* lost the race, try later */
1967                         err = -EBUSY;
1968                         p->rdev = rdev;
1969                 }
1970         }
1971 abort:
1972
1973         print_raid5_conf(conf);
1974         return err;
1975 }
1976
1977 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1978 {
1979         raid5_conf_t *conf = mddev->private;
1980         int found = 0;
1981         int disk;
1982         struct disk_info *p;
1983
1984         if (mddev->degraded > 1)
1985                 /* no point adding a device */
1986                 return 0;
1987
1988         /*
1989          * find the disk ...
1990          */
1991         for (disk=0; disk < mddev->raid_disks; disk++)
1992                 if ((p=conf->disks + disk)->rdev == NULL) {
1993                         rdev->in_sync = 0;
1994                         rdev->raid_disk = disk;
1995                         found = 1;
1996                         if (rdev->saved_raid_disk != disk)
1997                                 conf->fullsync = 1;
1998                         p->rdev = rdev;
1999                         break;
2000                 }
2001         print_raid5_conf(conf);
2002         return found;
2003 }
2004
2005 static int raid5_resize(mddev_t *mddev, sector_t sectors)
2006 {
2007         /* no resync is happening, and there is enough space
2008          * on all devices, so we can resize.
2009          * We need to make sure resync covers any new space.
2010          * If the array is shrinking we should possibly wait until
2011          * any io in the removed space completes, but it hardly seems
2012          * worth it.
2013          */
2014         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2015         mddev->array_size = (sectors * (mddev->raid_disks-1))>>1;
2016         set_capacity(mddev->gendisk, mddev->array_size << 1);
2017         mddev->changed = 1;
2018         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2019                 mddev->recovery_cp = mddev->size << 1;
2020                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2021         }
2022         mddev->size = sectors /2;
2023         mddev->resync_max_sectors = sectors;
2024         return 0;
2025 }
2026
2027 static void raid5_quiesce(mddev_t *mddev, int state)
2028 {
2029         raid5_conf_t *conf = mddev_to_conf(mddev);
2030
2031         switch(state) {
2032         case 1: /* stop all writes */
2033                 spin_lock_irq(&conf->device_lock);
2034                 conf->quiesce = 1;
2035                 wait_event_lock_irq(conf->wait_for_stripe,
2036                                     atomic_read(&conf->active_stripes) == 0,
2037                                     conf->device_lock, /* nothing */);
2038                 spin_unlock_irq(&conf->device_lock);
2039                 break;
2040
2041         case 0: /* re-enable writes */
2042                 spin_lock_irq(&conf->device_lock);
2043                 conf->quiesce = 0;
2044                 wake_up(&conf->wait_for_stripe);
2045                 spin_unlock_irq(&conf->device_lock);
2046                 break;
2047         }
2048         if (mddev->thread) {
2049                 if (mddev->bitmap)
2050                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2051                 else
2052                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2053                 md_wakeup_thread(mddev->thread);
2054         }
2055 }
2056 static mdk_personality_t raid5_personality=
2057 {
2058         .name           = "raid5",
2059         .owner          = THIS_MODULE,
2060         .make_request   = make_request,
2061         .run            = run,
2062         .stop           = stop,
2063         .status         = status,
2064         .error_handler  = error,
2065         .hot_add_disk   = raid5_add_disk,
2066         .hot_remove_disk= raid5_remove_disk,
2067         .spare_active   = raid5_spare_active,
2068         .sync_request   = sync_request,
2069         .resize         = raid5_resize,
2070         .quiesce        = raid5_quiesce,
2071 };
2072
2073 static int __init raid5_init (void)
2074 {
2075         return register_md_personality (RAID5, &raid5_personality);
2076 }
2077
2078 static void raid5_exit (void)
2079 {
2080         unregister_md_personality (RAID5);
2081 }
2082
2083 module_init(raid5_init);
2084 module_exit(raid5_exit);
2085 MODULE_LICENSE("GPL");
2086 MODULE_ALIAS("md-personality-4"); /* RAID5 */