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_one_stripe(raid5_conf_t *conf)
 297 {
 298         struct stripe_head *sh;
 299         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
 300         if (!sh)
 301                 return 0;
 302         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
 303         sh->raid_conf = conf;
 304         spin_lock_init(&sh->lock);
 305
 306         if (grow_buffers(sh, conf->raid_disks)) {
 307                 shrink_buffers(sh, conf->raid_disks);
 308                 kmem_cache_free(conf->slab_cache, sh);
 309                 return 0;
 310         }
 311         /* we just created an active stripe so... */
 312         atomic_set(&sh->count, 1);
 313         atomic_inc(&conf->active_stripes);
 314         INIT_LIST_HEAD(&sh->lru);
 315         release_stripe(sh);
 316         return 1;
 317 }
 318
 319 static int grow_stripes(raid5_conf_t *conf, int num)
 320 {
 321         kmem_cache_t *sc;
 322         int devs = conf->raid_disks;
 323
 324         sprintf(conf->cache_name, "raid5/%s", mdname(conf->mddev));
 325
 326         sc = kmem_cache_create(conf->cache_name,
 327                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
 328                                0, 0, NULL, NULL);
 329         if (!sc)
 330                 return 1;
 331         conf->slab_cache = sc;
 332         while (num--) {
 333                 if (!grow_one_stripe(conf))
 334                         return 1;
 335         }
 336         return 0;
 337 }
 338
 339 static int drop_one_stripe(raid5_conf_t *conf)
 340 {
 341         struct stripe_head *sh;
 342
 343         spin_lock_irq(&conf->device_lock);
 344         sh = get_free_stripe(conf);
 345         spin_unlock_irq(&conf->device_lock);
 346         if (!sh)
 347                 return 0;
 348         if (atomic_read(&sh->count))
 349                 BUG();
 350         shrink_buffers(sh, conf->raid_disks);
 351         kmem_cache_free(conf->slab_cache, sh);
 352         atomic_dec(&conf->active_stripes);
 353         return 1;
 354 }
 355
 356 static void shrink_stripes(raid5_conf_t *conf)
 357 {
 358         while (drop_one_stripe(conf))
 359                 ;
 360
 361         kmem_cache_destroy(conf->slab_cache);
 362         conf->slab_cache = NULL;
 363 }
 364
 365 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
 366                                    int error)
 367 {
 368         struct stripe_head *sh = bi->bi_private;
 369         raid5_conf_t *conf = sh->raid_conf;
 370         int disks = conf->raid_disks, i;
 371         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 372
 373         if (bi->bi_size)
 374                 return 1;
 375
 376         for (i=0 ; i<disks; i++)
 377                 if (bi == &sh->dev[i].req)
 378                         break;
 379
 380         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
 381                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 382                 uptodate);
 383         if (i == disks) {
 384                 BUG();
 385                 return 0;
 386         }
 387
 388         if (uptodate) {
 389 #if 0
 390                 struct bio *bio;
 391                 unsigned long flags;
 392                 spin_lock_irqsave(&conf->device_lock, flags);
 393                 /* we can return a buffer if we bypassed the cache or
 394                  * if the top buffer is not in highmem.  If there are
 395                  * multiple buffers, leave the extra work to
 396                  * handle_stripe
 397                  */
 398                 buffer = sh->bh_read[i];
 399                 if (buffer &&
 400                     (!PageHighMem(buffer->b_page)
 401                      || buffer->b_page == bh->b_page )
 402                         ) {
 403                         sh->bh_read[i] = buffer->b_reqnext;
 404                         buffer->b_reqnext = NULL;
 405                 } else
 406                         buffer = NULL;
 407                 spin_unlock_irqrestore(&conf->device_lock, flags);
 408                 if (sh->bh_page[i]==bh->b_page)
 409                         set_buffer_uptodate(bh);
 410                 if (buffer) {
 411                         if (buffer->b_page != bh->b_page)
 412                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
 413                         buffer->b_end_io(buffer, 1);
 414                 }
 415 #else
 416                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
 417 #endif
 418                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
 419                         printk("R5: read error corrected!!\n");
 420                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 421                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 422                 }
 423         } else {
 424                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
 425                 if (conf->mddev->degraded) {
 426                         printk("R5: read error not correctable.\n");
 427                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 428                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 429                         md_error(conf->mddev, conf->disks[i].rdev);
 430                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
 431                         /* Oh, no!!! */
 432                         printk("R5: read error NOT corrected!!\n");
 433                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 434                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 435                         md_error(conf->mddev, conf->disks[i].rdev);
 436                 } else
 437                         set_bit(R5_ReadError, &sh->dev[i].flags);
 438         }
 439         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 440 #if 0
 441         /* must restore b_page before unlocking buffer... */
 442         if (sh->bh_page[i] != bh->b_page) {
 443                 bh->b_page = sh->bh_page[i];
 444                 bh->b_data = page_address(bh->b_page);
 445                 clear_buffer_uptodate(bh);
 446         }
 447 #endif
 448         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 449         set_bit(STRIPE_HANDLE, &sh->state);
 450         release_stripe(sh);
 451         return 0;
 452 }
 453
 454 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
 455                                     int error)
 456 {
 457         struct stripe_head *sh = bi->bi_private;
 458         raid5_conf_t *conf = sh->raid_conf;
 459         int disks = conf->raid_disks, i;
 460         unsigned long flags;
 461         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 462
 463         if (bi->bi_size)
 464                 return 1;
 465
 466         for (i=0 ; i<disks; i++)
 467                 if (bi == &sh->dev[i].req)
 468                         break;
 469
 470         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
 471                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 472                 uptodate);
 473         if (i == disks) {
 474                 BUG();
 475                 return 0;
 476         }
 477
 478         spin_lock_irqsave(&conf->device_lock, flags);
 479         if (!uptodate)
 480                 md_error(conf->mddev, conf->disks[i].rdev);
 481
 482         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 483
 484         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 485         set_bit(STRIPE_HANDLE, &sh->state);
 486         __release_stripe(conf, sh);
 487         spin_unlock_irqrestore(&conf->device_lock, flags);
 488         return 0;
 489 }
 490
 491
 492 static sector_t compute_blocknr(struct stripe_head *sh, int i);
 493
 494 static void raid5_build_block (struct stripe_head *sh, int i)
 495 {
 496         struct r5dev *dev = &sh->dev[i];
 497
 498         bio_init(&dev->req);
 499         dev->req.bi_io_vec = &dev->vec;
 500         dev->req.bi_vcnt++;
 501         dev->req.bi_max_vecs++;
 502         dev->vec.bv_page = dev->page;
 503         dev->vec.bv_len = STRIPE_SIZE;
 504         dev->vec.bv_offset = 0;
 505
 506         dev->req.bi_sector = sh->sector;
 507         dev->req.bi_private = sh;
 508
 509         dev->flags = 0;
 510         if (i != sh->pd_idx)
 511                 dev->sector = compute_blocknr(sh, i);
 512 }
 513
 514 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 515 {
 516         char b[BDEVNAME_SIZE];
 517         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
 518         PRINTK("raid5: error called\n");
 519
 520         if (!rdev->faulty) {
 521                 mddev->sb_dirty = 1;
 522                 if (rdev->in_sync) {
 523                         conf->working_disks--;
 524                         mddev->degraded++;
 525                         conf->failed_disks++;
 526                         rdev->in_sync = 0;
 527                         /*
 528                          * if recovery was running, make sure it aborts.
 529                          */
 530                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 531                 }
 532                 rdev->faulty = 1;
 533                 printk (KERN_ALERT
 534                         "raid5: Disk failure on %s, disabling device."
 535                         " Operation continuing on %d devices\n",
 536                         bdevname(rdev->bdev,b), conf->working_disks);
 537         }
 538 }
 539
 540 /*
 541  * Input: a 'big' sector number,
 542  * Output: index of the data and parity disk, and the sector # in them.
 543  */
 544 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
 545                         unsigned int data_disks, unsigned int * dd_idx,
 546                         unsigned int * pd_idx, raid5_conf_t *conf)
 547 {
 548         long stripe;
 549         unsigned long chunk_number;
 550         unsigned int chunk_offset;
 551         sector_t new_sector;
 552         int sectors_per_chunk = conf->chunk_size >> 9;
 553
 554         /* First compute the information on this sector */
 555
 556         /*
 557          * Compute the chunk number and the sector offset inside the chunk
 558          */
 559         chunk_offset = sector_div(r_sector, sectors_per_chunk);
 560         chunk_number = r_sector;
 561         BUG_ON(r_sector != chunk_number);
 562
 563         /*
 564          * Compute the stripe number
 565          */
 566         stripe = chunk_number / data_disks;
 567
 568         /*
 569          * Compute the data disk and parity disk indexes inside the stripe
 570          */
 571         *dd_idx = chunk_number % data_disks;
 572
 573         /*
 574          * Select the parity disk based on the user selected algorithm.
 575          */
 576         if (conf->level == 4)
 577                 *pd_idx = data_disks;
 578         else switch (conf->algorithm) {
 579                 case ALGORITHM_LEFT_ASYMMETRIC:
 580                         *pd_idx = data_disks - stripe % raid_disks;
 581                         if (*dd_idx >= *pd_idx)
 582                                 (*dd_idx)++;
 583                         break;
 584                 case ALGORITHM_RIGHT_ASYMMETRIC:
 585                         *pd_idx = stripe % raid_disks;
 586                         if (*dd_idx >= *pd_idx)
 587                                 (*dd_idx)++;
 588                         break;
 589                 case ALGORITHM_LEFT_SYMMETRIC:
 590                         *pd_idx = data_disks - stripe % raid_disks;
 591                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 592                         break;
 593                 case ALGORITHM_RIGHT_SYMMETRIC:
 594                         *pd_idx = stripe % raid_disks;
 595                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 596                         break;
 597                 default:
 598                         printk("raid5: unsupported algorithm %d\n",
 599                                 conf->algorithm);
 600         }
 601
 602         /*
 603          * Finally, compute the new sector number
 604          */
 605         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
 606         return new_sector;
 607 }
 608
 609
 610 static sector_t compute_blocknr(struct stripe_head *sh, int i)
 611 {
 612         raid5_conf_t *conf = sh->raid_conf;
 613         int raid_disks = conf->raid_disks, data_disks = raid_disks - 1;
 614         sector_t new_sector = sh->sector, check;
 615         int sectors_per_chunk = conf->chunk_size >> 9;
 616         sector_t stripe;
 617         int chunk_offset;
 618         int chunk_number, dummy1, dummy2, dd_idx = i;
 619         sector_t r_sector;
 620
 621         chunk_offset = sector_div(new_sector, sectors_per_chunk);
 622         stripe = new_sector;
 623         BUG_ON(new_sector != stripe);
 624
 625
 626         switch (conf->algorithm) {
 627                 case ALGORITHM_LEFT_ASYMMETRIC:
 628                 case ALGORITHM_RIGHT_ASYMMETRIC:
 629                         if (i > sh->pd_idx)
 630                                 i--;
 631                         break;
 632                 case ALGORITHM_LEFT_SYMMETRIC:
 633                 case ALGORITHM_RIGHT_SYMMETRIC:
 634                         if (i < sh->pd_idx)
 635                                 i += raid_disks;
 636                         i -= (sh->pd_idx + 1);
 637                         break;
 638                 default:
 639                         printk("raid5: unsupported algorithm %d\n",
 640                                 conf->algorithm);
 641         }
 642
 643         chunk_number = stripe * data_disks + i;
 644         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
 645
 646         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
 647         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
 648                 printk("compute_blocknr: map not correct\n");
 649                 return 0;
 650         }
 651         return r_sector;
 652 }
 653
 654
 655
 656 /*
 657  * Copy data between a page in the stripe cache, and a bio.
 658  * There are no alignment or size guarantees between the page or the
 659  * bio except that there is some overlap.
 660  * All iovecs in the bio must be considered.
 661  */
 662 static void copy_data(int frombio, struct bio *bio,
 663                      struct page *page,
 664                      sector_t sector)
 665 {
 666         char *pa = page_address(page);
 667         struct bio_vec *bvl;
 668         int i;
 669         int page_offset;
 670
 671         if (bio->bi_sector >= sector)
 672                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 673         else
 674                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 675         bio_for_each_segment(bvl, bio, i) {
 676                 int len = bio_iovec_idx(bio,i)->bv_len;
 677                 int clen;
 678                 int b_offset = 0;
 679
 680                 if (page_offset < 0) {
 681                         b_offset = -page_offset;
 682                         page_offset += b_offset;
 683                         len -= b_offset;
 684                 }
 685
 686                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 687                         clen = STRIPE_SIZE - page_offset;
 688                 else clen = len;
 689
 690                 if (clen > 0) {
 691                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
 692                         if (frombio)
 693                                 memcpy(pa+page_offset, ba+b_offset, clen);
 694                         else
 695                                 memcpy(ba+b_offset, pa+page_offset, clen);
 696                         __bio_kunmap_atomic(ba, KM_USER0);
 697                 }
 698                 if (clen < len) /* hit end of page */
 699                         break;
 700                 page_offset +=  len;
 701         }
 702 }
 703
 704 #define check_xor()     do {                                            \
 705                            if (count == MAX_XOR_BLOCKS) {               \
 706                                 xor_block(count, STRIPE_SIZE, ptr);     \
 707                                 count = 1;                              \
 708                            }                                            \
 709                         } while(0)
 710
 711
 712 static void compute_block(struct stripe_head *sh, int dd_idx)
 713 {
 714         raid5_conf_t *conf = sh->raid_conf;
 715         int i, count, disks = conf->raid_disks;
 716         void *ptr[MAX_XOR_BLOCKS], *p;
 717
 718         PRINTK("compute_block, stripe %llu, idx %d\n",
 719                 (unsigned long long)sh->sector, dd_idx);
 720
 721         ptr[0] = page_address(sh->dev[dd_idx].page);
 722         memset(ptr[0], 0, STRIPE_SIZE);
 723         count = 1;
 724         for (i = disks ; i--; ) {
 725                 if (i == dd_idx)
 726                         continue;
 727                 p = page_address(sh->dev[i].page);
 728                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
 729                         ptr[count++] = p;
 730                 else
 731                         printk("compute_block() %d, stripe %llu, %d"
 732                                 " not present\n", dd_idx,
 733                                 (unsigned long long)sh->sector, i);
 734
 735                 check_xor();
 736         }
 737         if (count != 1)
 738                 xor_block(count, STRIPE_SIZE, ptr);
 739         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
 740 }
 741
 742 static void compute_parity(struct stripe_head *sh, int method)
 743 {
 744         raid5_conf_t *conf = sh->raid_conf;
 745         int i, pd_idx = sh->pd_idx, disks = conf->raid_disks, count;
 746         void *ptr[MAX_XOR_BLOCKS];
 747         struct bio *chosen;
 748
 749         PRINTK("compute_parity, stripe %llu, method %d\n",
 750                 (unsigned long long)sh->sector, method);
 751
 752         count = 1;
 753         ptr[0] = page_address(sh->dev[pd_idx].page);
 754         switch(method) {
 755         case READ_MODIFY_WRITE:
 756                 if (!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags))
 757                         BUG();
 758                 for (i=disks ; i-- ;) {
 759                         if (i==pd_idx)
 760                                 continue;
 761                         if (sh->dev[i].towrite &&
 762                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
 763                                 ptr[count++] = page_address(sh->dev[i].page);
 764                                 chosen = sh->dev[i].towrite;
 765                                 sh->dev[i].towrite = NULL;
 766
 767                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 768                                         wake_up(&conf->wait_for_overlap);
 769
 770                                 if (sh->dev[i].written) BUG();
 771                                 sh->dev[i].written = chosen;
 772                                 check_xor();
 773                         }
 774                 }
 775                 break;
 776         case RECONSTRUCT_WRITE:
 777                 memset(ptr[0], 0, STRIPE_SIZE);
 778                 for (i= disks; i-- ;)
 779                         if (i!=pd_idx && sh->dev[i].towrite) {
 780                                 chosen = sh->dev[i].towrite;
 781                                 sh->dev[i].towrite = NULL;
 782
 783                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 784                                         wake_up(&conf->wait_for_overlap);
 785
 786                                 if (sh->dev[i].written) BUG();
 787                                 sh->dev[i].written = chosen;
 788                         }
 789                 break;
 790         case CHECK_PARITY:
 791                 break;
 792         }
 793         if (count>1) {
 794                 xor_block(count, STRIPE_SIZE, ptr);
 795                 count = 1;
 796         }
 797
 798         for (i = disks; i--;)
 799                 if (sh->dev[i].written) {
 800                         sector_t sector = sh->dev[i].sector;
 801                         struct bio *wbi = sh->dev[i].written;
 802                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
 803                                 copy_data(1, wbi, sh->dev[i].page, sector);
 804                                 wbi = r5_next_bio(wbi, sector);
 805                         }
 806
 807                         set_bit(R5_LOCKED, &sh->dev[i].flags);
 808                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
 809                 }
 810
 811         switch(method) {
 812         case RECONSTRUCT_WRITE:
 813         case CHECK_PARITY:
 814                 for (i=disks; i--;)
 815                         if (i != pd_idx) {
 816                                 ptr[count++] = page_address(sh->dev[i].page);
 817                                 check_xor();
 818                         }
 819                 break;
 820         case READ_MODIFY_WRITE:
 821                 for (i = disks; i--;)
 822                         if (sh->dev[i].written) {
 823                                 ptr[count++] = page_address(sh->dev[i].page);
 824                                 check_xor();
 825                         }
 826         }
 827         if (count != 1)
 828                 xor_block(count, STRIPE_SIZE, ptr);
 829
 830         if (method != CHECK_PARITY) {
 831                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 832                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
 833         } else
 834                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 835 }
 836
 837 /*
 838  * Each stripe/dev can have one or more bion attached.
 839  * toread/towrite point to the first in a chain.
 840  * The bi_next chain must be in order.
 841  */
 842 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
 843 {
 844         struct bio **bip;
 845         raid5_conf_t *conf = sh->raid_conf;
 846         int firstwrite=0;
 847
 848         PRINTK("adding bh b#%llu to stripe s#%llu\n",
 849                 (unsigned long long)bi->bi_sector,
 850                 (unsigned long long)sh->sector);
 851
 852
 853         spin_lock(&sh->lock);
 854         spin_lock_irq(&conf->device_lock);
 855         if (forwrite) {
 856                 bip = &sh->dev[dd_idx].towrite;
 857                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
 858                         firstwrite = 1;
 859         } else
 860                 bip = &sh->dev[dd_idx].toread;
 861         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
 862                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
 863                         goto overlap;
 864                 bip = & (*bip)->bi_next;
 865         }
 866         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
 867                 goto overlap;
 868
 869         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
 870                 BUG();
 871         if (*bip)
 872                 bi->bi_next = *bip;
 873         *bip = bi;
 874         bi->bi_phys_segments ++;
 875         spin_unlock_irq(&conf->device_lock);
 876         spin_unlock(&sh->lock);
 877
 878         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
 879                 (unsigned long long)bi->bi_sector,
 880                 (unsigned long long)sh->sector, dd_idx);
 881
 882         if (conf->mddev->bitmap && firstwrite) {
 883                 sh->bm_seq = conf->seq_write;
 884                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
 885                                   STRIPE_SECTORS, 0);
 886                 set_bit(STRIPE_BIT_DELAY, &sh->state);
 887         }
 888
 889         if (forwrite) {
 890                 /* check if page is covered */
 891                 sector_t sector = sh->dev[dd_idx].sector;
 892                 for (bi=sh->dev[dd_idx].towrite;
 893                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
 894                              bi && bi->bi_sector <= sector;
 895                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
 896                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
 897                                 sector = bi->bi_sector + (bi->bi_size>>9);
 898                 }
 899                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
 900                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
 901         }
 902         return 1;
 903
 904  overlap:
 905         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
 906         spin_unlock_irq(&conf->device_lock);
 907         spin_unlock(&sh->lock);
 908         return 0;
 909 }
 910
 911
 912 /*
 913  * handle_stripe - do things to a stripe.
 914  *
 915  * We lock the stripe and then examine the state of various bits
 916  * to see what needs to be done.
 917  * Possible results:
 918  *    return some read request which now have data
 919  *    return some write requests which are safely on disc
 920  *    schedule a read on some buffers
 921  *    schedule a write of some buffers
 922  *    return confirmation of parity correctness
 923  *
 924  * Parity calculations are done inside the stripe lock
 925  * buffers are taken off read_list or write_list, and bh_cache buffers
 926  * get BH_Lock set before the stripe lock is released.
 927  *
 928  */
 929
 930 static void handle_stripe(struct stripe_head *sh)
 931 {
 932         raid5_conf_t *conf = sh->raid_conf;
 933         int disks = conf->raid_disks;
 934         struct bio *return_bi= NULL;
 935         struct bio *bi;
 936         int i;
 937         int syncing;
 938         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
 939         int non_overwrite = 0;
 940         int failed_num=0;
 941         struct r5dev *dev;
 942
 943         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
 944                 (unsigned long long)sh->sector, atomic_read(&sh->count),
 945                 sh->pd_idx);
 946
 947         spin_lock(&sh->lock);
 948         clear_bit(STRIPE_HANDLE, &sh->state);
 949         clear_bit(STRIPE_DELAYED, &sh->state);
 950
 951         syncing = test_bit(STRIPE_SYNCING, &sh->state);
 952         /* Now to look around and see what can be done */
 953
 954         for (i=disks; i--; ) {
 955                 mdk_rdev_t *rdev;
 956                 dev = &sh->dev[i];
 957                 clear_bit(R5_Insync, &dev->flags);
 958                 clear_bit(R5_Syncio, &dev->flags);
 959
 960                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
 961                         i, dev->flags, dev->toread, dev->towrite, dev->written);
 962                 /* maybe we can reply to a read */
 963                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
 964                         struct bio *rbi, *rbi2;
 965                         PRINTK("Return read for disc %d\n", i);
 966                         spin_lock_irq(&conf->device_lock);
 967                         rbi = dev->toread;
 968                         dev->toread = NULL;
 969                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
 970                                 wake_up(&conf->wait_for_overlap);
 971                         spin_unlock_irq(&conf->device_lock);
 972                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
 973                                 copy_data(0, rbi, dev->page, dev->sector);
 974                                 rbi2 = r5_next_bio(rbi, dev->sector);
 975                                 spin_lock_irq(&conf->device_lock);
 976                                 if (--rbi->bi_phys_segments == 0) {
 977                                         rbi->bi_next = return_bi;
 978                                         return_bi = rbi;
 979                                 }
 980                                 spin_unlock_irq(&conf->device_lock);
 981                                 rbi = rbi2;
 982                         }
 983                 }
 984
 985                 /* now count some things */
 986                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
 987                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
 988
 989
 990                 if (dev->toread) to_read++;
 991                 if (dev->towrite) {
 992                         to_write++;
 993                         if (!test_bit(R5_OVERWRITE, &dev->flags))
 994                                 non_overwrite++;
 995                 }
 996                 if (dev->written) written++;
 997                 rdev = conf->disks[i].rdev; /* FIXME, should I be looking rdev */
 998                 if (!rdev || !rdev->in_sync) {
 999                         /* The ReadError flag wil just be confusing now */
1000                         clear_bit(R5_ReadError, &dev->flags);
1001                         clear_bit(R5_ReWrite, &dev->flags);
1002                 }
1003                 if (!rdev || !rdev->in_sync
1004                     || test_bit(R5_ReadError, &dev->flags)) {
1005                         failed++;
1006                         failed_num = i;
1007                 } else
1008                         set_bit(R5_Insync, &dev->flags);
1009         }
1010         PRINTK("locked=%d uptodate=%d to_read=%d"
1011                 " to_write=%d failed=%d failed_num=%d\n",
1012                 locked, uptodate, to_read, to_write, failed, failed_num);
1013         /* check if the array has lost two devices and, if so, some requests might
1014          * need to be failed
1015          */
1016         if (failed > 1 && to_read+to_write+written) {
1017                 for (i=disks; i--; ) {
1018                         int bitmap_end = 0;
1019
1020                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1021                                 mdk_rdev_t *rdev = conf->disks[i].rdev;
1022                                 if (rdev && rdev->in_sync)
1023                                         /* multiple read failures in one stripe */
1024                                         md_error(conf->mddev, rdev);
1025                         }
1026
1027                         spin_lock_irq(&conf->device_lock);
1028                         /* fail all writes first */
1029                         bi = sh->dev[i].towrite;
1030                         sh->dev[i].towrite = NULL;
1031                         if (bi) { to_write--; bitmap_end = 1; }
1032
1033                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1034                                 wake_up(&conf->wait_for_overlap);
1035
1036                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1037                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1038                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1039                                 if (--bi->bi_phys_segments == 0) {
1040                                         md_write_end(conf->mddev);
1041                                         bi->bi_next = return_bi;
1042                                         return_bi = bi;
1043                                 }
1044                                 bi = nextbi;
1045                         }
1046                         /* and fail all 'written' */
1047                         bi = sh->dev[i].written;
1048                         sh->dev[i].written = NULL;
1049                         if (bi) bitmap_end = 1;
1050                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1051                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1052                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1053                                 if (--bi->bi_phys_segments == 0) {
1054                                         md_write_end(conf->mddev);
1055                                         bi->bi_next = return_bi;
1056                                         return_bi = bi;
1057                                 }
1058                                 bi = bi2;
1059                         }
1060
1061                         /* fail any reads if this device is non-operational */
1062                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1063                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1064                                 bi = sh->dev[i].toread;
1065                                 sh->dev[i].toread = NULL;
1066                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1067                                         wake_up(&conf->wait_for_overlap);
1068                                 if (bi) to_read--;
1069                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1070                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1071                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1072                                         if (--bi->bi_phys_segments == 0) {
1073                                                 bi->bi_next = return_bi;
1074                                                 return_bi = bi;
1075                                         }
1076                                         bi = nextbi;
1077                                 }
1078                         }
1079                         spin_unlock_irq(&conf->device_lock);
1080                         if (bitmap_end)
1081                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1082                                                 STRIPE_SECTORS, 0, 0);
1083                 }
1084         }
1085         if (failed > 1 && syncing) {
1086                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1087                 clear_bit(STRIPE_SYNCING, &sh->state);
1088                 syncing = 0;
1089         }
1090
1091         /* might be able to return some write requests if the parity block
1092          * is safe, or on a failed drive
1093          */
1094         dev = &sh->dev[sh->pd_idx];
1095         if ( written &&
1096              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1097                 test_bit(R5_UPTODATE, &dev->flags))
1098                || (failed == 1 && failed_num == sh->pd_idx))
1099             ) {
1100             /* any written block on an uptodate or failed drive can be returned.
1101              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
1102              * never LOCKED, so we don't need to test 'failed' directly.
1103              */
1104             for (i=disks; i--; )
1105                 if (sh->dev[i].written) {
1106                     dev = &sh->dev[i];
1107                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1108                          test_bit(R5_UPTODATE, &dev->flags) ) {
1109                         /* We can return any write requests */
1110                             struct bio *wbi, *wbi2;
1111                             int bitmap_end = 0;
1112                             PRINTK("Return write for disc %d\n", i);
1113                             spin_lock_irq(&conf->device_lock);
1114                             wbi = dev->written;
1115                             dev->written = NULL;
1116                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1117                                     wbi2 = r5_next_bio(wbi, dev->sector);
1118                                     if (--wbi->bi_phys_segments == 0) {
1119                                             md_write_end(conf->mddev);
1120                                             wbi->bi_next = return_bi;
1121                                             return_bi = wbi;
1122                                     }
1123                                     wbi = wbi2;
1124                             }
1125                             if (dev->towrite == NULL)
1126                                     bitmap_end = 1;
1127                             spin_unlock_irq(&conf->device_lock);
1128                             if (bitmap_end)
1129                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1130                                                     STRIPE_SECTORS,
1131                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1132                     }
1133                 }
1134         }
1135
1136         /* Now we might consider reading some blocks, either to check/generate
1137          * parity, or to satisfy requests
1138          * or to load a block that is being partially written.
1139          */
1140         if (to_read || non_overwrite || (syncing && (uptodate < disks))) {
1141                 for (i=disks; i--;) {
1142                         dev = &sh->dev[i];
1143                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1144                             (dev->toread ||
1145                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1146                              syncing ||
1147                              (failed && (sh->dev[failed_num].toread ||
1148                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1149                                     )
1150                                 ) {
1151                                 /* we would like to get this block, possibly
1152                                  * by computing it, but we might not be able to
1153                                  */
1154                                 if (uptodate == disks-1) {
1155                                         PRINTK("Computing block %d\n", i);
1156                                         compute_block(sh, i);
1157                                         uptodate++;
1158                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1159                                         set_bit(R5_LOCKED, &dev->flags);
1160                                         set_bit(R5_Wantread, &dev->flags);
1161 #if 0
1162                                         /* if I am just reading this block and we don't have
1163                                            a failed drive, or any pending writes then sidestep the cache */
1164                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1165                                             ! syncing && !failed && !to_write) {
1166                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1167                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1168                                         }
1169 #endif
1170                                         locked++;
1171                                         PRINTK("Reading block %d (sync=%d)\n",
1172                                                 i, syncing);
1173                                         if (syncing)
1174                                                 md_sync_acct(conf->disks[i].rdev->bdev,
1175                                                              STRIPE_SECTORS);
1176                                 }
1177                         }
1178                 }
1179                 set_bit(STRIPE_HANDLE, &sh->state);
1180         }
1181
1182         /* now to consider writing and what else, if anything should be read */
1183         if (to_write) {
1184                 int rmw=0, rcw=0;
1185                 for (i=disks ; i--;) {
1186                         /* would I have to read this buffer for read_modify_write */
1187                         dev = &sh->dev[i];
1188                         if ((dev->towrite || i == sh->pd_idx) &&
1189                             (!test_bit(R5_LOCKED, &dev->flags)
1190 #if 0
1191 || sh->bh_page[i]!=bh->b_page
1192 #endif
1193                                     ) &&
1194                             !test_bit(R5_UPTODATE, &dev->flags)) {
1195                                 if (test_bit(R5_Insync, &dev->flags)
1196 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1197                                         )
1198                                         rmw++;
1199                                 else rmw += 2*disks;  /* cannot read it */
1200                         }
1201                         /* Would I have to read this buffer for reconstruct_write */
1202                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1203                             (!test_bit(R5_LOCKED, &dev->flags)
1204 #if 0
1205 || sh->bh_page[i] != bh->b_page
1206 #endif
1207                                     ) &&
1208                             !test_bit(R5_UPTODATE, &dev->flags)) {
1209                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1210                                 else rcw += 2*disks;
1211                         }
1212                 }
1213                 PRINTK("for sector %llu, rmw=%d rcw=%d\n",
1214                         (unsigned long long)sh->sector, rmw, rcw);
1215                 set_bit(STRIPE_HANDLE, &sh->state);
1216                 if (rmw < rcw && rmw > 0)
1217                         /* prefer read-modify-write, but need to get some data */
1218                         for (i=disks; i--;) {
1219                                 dev = &sh->dev[i];
1220                                 if ((dev->towrite || i == sh->pd_idx) &&
1221                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1222                                     test_bit(R5_Insync, &dev->flags)) {
1223                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1224                                         {
1225                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1226                                                 set_bit(R5_LOCKED, &dev->flags);
1227                                                 set_bit(R5_Wantread, &dev->flags);
1228                                                 locked++;
1229                                         } else {
1230                                                 set_bit(STRIPE_DELAYED, &sh->state);
1231                                                 set_bit(STRIPE_HANDLE, &sh->state);
1232                                         }
1233                                 }
1234                         }
1235                 if (rcw <= rmw && rcw > 0)
1236                         /* want reconstruct write, but need to get some data */
1237                         for (i=disks; i--;) {
1238                                 dev = &sh->dev[i];
1239                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1240                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1241                                     test_bit(R5_Insync, &dev->flags)) {
1242                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1243                                         {
1244                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1245                                                 set_bit(R5_LOCKED, &dev->flags);
1246                                                 set_bit(R5_Wantread, &dev->flags);
1247                                                 locked++;
1248                                         } else {
1249                                                 set_bit(STRIPE_DELAYED, &sh->state);
1250                                                 set_bit(STRIPE_HANDLE, &sh->state);
1251                                         }
1252                                 }
1253                         }
1254                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1255                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1256                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1257                         PRINTK("Computing parity...\n");
1258                         compute_parity(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1259                         /* now every locked buffer is ready to be written */
1260                         for (i=disks; i--;)
1261                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1262                                         PRINTK("Writing block %d\n", i);
1263                                         locked++;
1264                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1265                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1266                                             || (i==sh->pd_idx && failed == 0))
1267                                                 set_bit(STRIPE_INSYNC, &sh->state);
1268                                 }
1269                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1270                                 atomic_dec(&conf->preread_active_stripes);
1271                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1272                                         md_wakeup_thread(conf->mddev->thread);
1273                         }
1274                 }
1275         }
1276
1277         /* maybe we need to check and possibly fix the parity for this stripe
1278          * Any reads will already have been scheduled, so we just see if enough data
1279          * is available
1280          */
1281         if (syncing && locked == 0 &&
1282             !test_bit(STRIPE_INSYNC, &sh->state) && failed <= 1) {
1283                 set_bit(STRIPE_HANDLE, &sh->state);
1284                 if (failed == 0) {
1285                         char *pagea;
1286                         if (uptodate != disks)
1287                                 BUG();
1288                         compute_parity(sh, CHECK_PARITY);
1289                         uptodate--;
1290                         pagea = page_address(sh->dev[sh->pd_idx].page);
1291                         if ((*(u32*)pagea) == 0 &&
1292                             !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1293                                 /* parity is correct (on disc, not in buffer any more) */
1294                                 set_bit(STRIPE_INSYNC, &sh->state);
1295                         } else {
1296                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1297                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1298                                         /* don't try to repair!! */
1299                                         set_bit(STRIPE_INSYNC, &sh->state);
1300                         }
1301                 }
1302                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1303                         if (failed==0)
1304                                 failed_num = sh->pd_idx;
1305                         /* should be able to compute the missing block and write it to spare */
1306                         if (!test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)) {
1307                                 if (uptodate+1 != disks)
1308                                         BUG();
1309                                 compute_block(sh, failed_num);
1310                                 uptodate++;
1311                         }
1312                         if (uptodate != disks)
1313                                 BUG();
1314                         dev = &sh->dev[failed_num];
1315                         set_bit(R5_LOCKED, &dev->flags);
1316                         set_bit(R5_Wantwrite, &dev->flags);
1317                         clear_bit(STRIPE_DEGRADED, &sh->state);
1318                         locked++;
1319                         set_bit(STRIPE_INSYNC, &sh->state);
1320                         set_bit(R5_Syncio, &dev->flags);
1321                 }
1322         }
1323         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1324                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1325                 clear_bit(STRIPE_SYNCING, &sh->state);
1326         }
1327
1328         /* If the failed drive is just a ReadError, then we might need to progress
1329          * the repair/check process
1330          */
1331         if (failed == 1 && test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1332             && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1333             && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1334                 ) {
1335                 dev = &sh->dev[failed_num];
1336                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1337                         set_bit(R5_Wantwrite, &dev->flags);
1338                         set_bit(R5_ReWrite, &dev->flags);
1339                         set_bit(R5_LOCKED, &dev->flags);
1340                 } else {
1341                         /* let's read it back */
1342                         set_bit(R5_Wantread, &dev->flags);
1343                         set_bit(R5_LOCKED, &dev->flags);
1344                 }
1345         }
1346
1347         spin_unlock(&sh->lock);
1348
1349         while ((bi=return_bi)) {
1350                 int bytes = bi->bi_size;
1351
1352                 return_bi = bi->bi_next;
1353                 bi->bi_next = NULL;
1354                 bi->bi_size = 0;
1355                 bi->bi_end_io(bi, bytes, 0);
1356         }
1357         for (i=disks; i-- ;) {
1358                 int rw;
1359                 struct bio *bi;
1360                 mdk_rdev_t *rdev;
1361                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1362                         rw = 1;
1363                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1364                         rw = 0;
1365                 else
1366                         continue;
1367
1368                 bi = &sh->dev[i].req;
1369
1370                 bi->bi_rw = rw;
1371                 if (rw)
1372                         bi->bi_end_io = raid5_end_write_request;
1373                 else
1374                         bi->bi_end_io = raid5_end_read_request;
1375
1376                 rcu_read_lock();
1377                 rdev = rcu_dereference(conf->disks[i].rdev);
1378                 if (rdev && rdev->faulty)
1379                         rdev = NULL;
1380                 if (rdev)
1381                         atomic_inc(&rdev->nr_pending);
1382                 rcu_read_unlock();
1383
1384                 if (rdev) {
1385                         if (test_bit(R5_Syncio, &sh->dev[i].flags))
1386                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1387
1388                         bi->bi_bdev = rdev->bdev;
1389                         PRINTK("for %llu schedule op %ld on disc %d\n",
1390                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1391                         atomic_inc(&sh->count);
1392                         bi->bi_sector = sh->sector + rdev->data_offset;
1393                         bi->bi_flags = 1 << BIO_UPTODATE;
1394                         bi->bi_vcnt = 1;
1395                         bi->bi_max_vecs = 1;
1396                         bi->bi_idx = 0;
1397                         bi->bi_io_vec = &sh->dev[i].vec;
1398                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1399                         bi->bi_io_vec[0].bv_offset = 0;
1400                         bi->bi_size = STRIPE_SIZE;
1401                         bi->bi_next = NULL;
1402                         generic_make_request(bi);
1403                 } else {
1404                         if (rw == 1)
1405                                 set_bit(STRIPE_DEGRADED, &sh->state);
1406                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1407                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1408                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1409                         set_bit(STRIPE_HANDLE, &sh->state);
1410                 }
1411         }
1412 }
1413
1414 static inline void raid5_activate_delayed(raid5_conf_t *conf)
1415 {
1416         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1417                 while (!list_empty(&conf->delayed_list)) {
1418                         struct list_head *l = conf->delayed_list.next;
1419                         struct stripe_head *sh;
1420                         sh = list_entry(l, struct stripe_head, lru);
1421                         list_del_init(l);
1422                         clear_bit(STRIPE_DELAYED, &sh->state);
1423                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1424                                 atomic_inc(&conf->preread_active_stripes);
1425                         list_add_tail(&sh->lru, &conf->handle_list);
1426                 }
1427         }
1428 }
1429
1430 static inline void activate_bit_delay(raid5_conf_t *conf)
1431 {
1432         /* device_lock is held */
1433         struct list_head head;
1434         list_add(&head, &conf->bitmap_list);
1435         list_del_init(&conf->bitmap_list);
1436         while (!list_empty(&head)) {
1437                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1438                 list_del_init(&sh->lru);
1439                 atomic_inc(&sh->count);
1440                 __release_stripe(conf, sh);
1441         }
1442 }
1443
1444 static void unplug_slaves(mddev_t *mddev)
1445 {
1446         raid5_conf_t *conf = mddev_to_conf(mddev);
1447         int i;
1448
1449         rcu_read_lock();
1450         for (i=0; i<mddev->raid_disks; i++) {
1451                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1452                 if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
1453                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1454
1455                         atomic_inc(&rdev->nr_pending);
1456                         rcu_read_unlock();
1457
1458                         if (r_queue->unplug_fn)
1459                                 r_queue->unplug_fn(r_queue);
1460
1461                         rdev_dec_pending(rdev, mddev);
1462                         rcu_read_lock();
1463                 }
1464         }
1465         rcu_read_unlock();
1466 }
1467
1468 static void raid5_unplug_device(request_queue_t *q)
1469 {
1470         mddev_t *mddev = q->queuedata;
1471         raid5_conf_t *conf = mddev_to_conf(mddev);
1472         unsigned long flags;
1473
1474         spin_lock_irqsave(&conf->device_lock, flags);
1475
1476         if (blk_remove_plug(q)) {
1477                 conf->seq_flush++;
1478                 raid5_activate_delayed(conf);
1479         }
1480         md_wakeup_thread(mddev->thread);
1481
1482         spin_unlock_irqrestore(&conf->device_lock, flags);
1483
1484         unplug_slaves(mddev);
1485 }
1486
1487 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
1488                              sector_t *error_sector)
1489 {
1490         mddev_t *mddev = q->queuedata;
1491         raid5_conf_t *conf = mddev_to_conf(mddev);
1492         int i, ret = 0;
1493
1494         rcu_read_lock();
1495         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1496                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1497                 if (rdev && !rdev->faulty) {
1498                         struct block_device *bdev = rdev->bdev;
1499                         request_queue_t *r_queue = bdev_get_queue(bdev);
1500
1501                         if (!r_queue->issue_flush_fn)
1502                                 ret = -EOPNOTSUPP;
1503                         else {
1504                                 atomic_inc(&rdev->nr_pending);
1505                                 rcu_read_unlock();
1506                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1507                                                               error_sector);
1508                                 rdev_dec_pending(rdev, mddev);
1509                                 rcu_read_lock();
1510                         }
1511                 }
1512         }
1513         rcu_read_unlock();
1514         return ret;
1515 }
1516
1517 static inline void raid5_plug_device(raid5_conf_t *conf)
1518 {
1519         spin_lock_irq(&conf->device_lock);
1520         blk_plug_device(conf->mddev->queue);
1521         spin_unlock_irq(&conf->device_lock);
1522 }
1523
1524 static int make_request (request_queue_t *q, struct bio * bi)
1525 {
1526         mddev_t *mddev = q->queuedata;
1527         raid5_conf_t *conf = mddev_to_conf(mddev);
1528         const unsigned int raid_disks = conf->raid_disks;
1529         const unsigned int data_disks = raid_disks - 1;
1530         unsigned int dd_idx, pd_idx;
1531         sector_t new_sector;
1532         sector_t logical_sector, last_sector;
1533         struct stripe_head *sh;
1534         const int rw = bio_data_dir(bi);
1535
1536         if (unlikely(bio_barrier(bi))) {
1537                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1538                 return 0;
1539         }
1540
1541         md_write_start(mddev, bi);
1542
1543         disk_stat_inc(mddev->gendisk, ios[rw]);
1544         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
1545
1546         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1547         last_sector = bi->bi_sector + (bi->bi_size>>9);
1548         bi->bi_next = NULL;
1549         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1550
1551         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1552                 DEFINE_WAIT(w);
1553
1554                 new_sector = raid5_compute_sector(logical_sector,
1555                                                   raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1556
1557                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
1558                         (unsigned long long)new_sector,
1559                         (unsigned long long)logical_sector);
1560
1561         retry:
1562                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1563                 sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
1564                 if (sh) {
1565                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1566                                 /* Add failed due to overlap.  Flush everything
1567                                  * and wait a while
1568                                  */
1569                                 raid5_unplug_device(mddev->queue);
1570                                 release_stripe(sh);
1571                                 schedule();
1572                                 goto retry;
1573                         }
1574                         finish_wait(&conf->wait_for_overlap, &w);
1575                         raid5_plug_device(conf);
1576                         handle_stripe(sh);
1577                         release_stripe(sh);
1578
1579                 } else {
1580                         /* cannot get stripe for read-ahead, just give-up */
1581                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1582                         finish_wait(&conf->wait_for_overlap, &w);
1583                         break;
1584                 }
1585
1586         }
1587         spin_lock_irq(&conf->device_lock);
1588         if (--bi->bi_phys_segments == 0) {
1589                 int bytes = bi->bi_size;
1590
1591                 if ( bio_data_dir(bi) == WRITE )
1592                         md_write_end(mddev);
1593                 bi->bi_size = 0;
1594                 bi->bi_end_io(bi, bytes, 0);
1595         }
1596         spin_unlock_irq(&conf->device_lock);
1597         return 0;
1598 }
1599
1600 /* FIXME go_faster isn't used */
1601 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1602 {
1603         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1604         struct stripe_head *sh;
1605         int sectors_per_chunk = conf->chunk_size >> 9;
1606         sector_t x;
1607         unsigned long stripe;
1608         int chunk_offset;
1609         int dd_idx, pd_idx;
1610         sector_t first_sector;
1611         int raid_disks = conf->raid_disks;
1612         int data_disks = raid_disks-1;
1613         sector_t max_sector = mddev->size << 1;
1614         int sync_blocks;
1615
1616         if (sector_nr >= max_sector) {
1617                 /* just being told to finish up .. nothing much to do */
1618                 unplug_slaves(mddev);
1619
1620                 if (mddev->curr_resync < max_sector) /* aborted */
1621                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1622                                         &sync_blocks, 1);
1623                 else /* compelted sync */
1624                         conf->fullsync = 0;
1625                 bitmap_close_sync(mddev->bitmap);
1626
1627                 return 0;
1628         }
1629         /* if there is 1 or more failed drives and we are trying
1630          * to resync, then assert that we are finished, because there is
1631          * nothing we can do.
1632          */
1633         if (mddev->degraded >= 1 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1634                 sector_t rv = (mddev->size << 1) - sector_nr;
1635                 *skipped = 1;
1636                 return rv;
1637         }
1638         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1639             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
1640                 /* we can skip this block, and probably more */
1641                 sync_blocks /= STRIPE_SECTORS;
1642                 *skipped = 1;
1643                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
1644         }
1645
1646         x = sector_nr;
1647         chunk_offset = sector_div(x, sectors_per_chunk);
1648         stripe = x;
1649         BUG_ON(x != stripe);
1650
1651         first_sector = raid5_compute_sector((sector_t)stripe*data_disks*sectors_per_chunk
1652                 + chunk_offset, raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1653         sh = get_active_stripe(conf, sector_nr, pd_idx, 1);
1654         if (sh == NULL) {
1655                 sh = get_active_stripe(conf, sector_nr, pd_idx, 0);
1656                 /* make sure we don't swamp the stripe cache if someone else
1657                  * is trying to get access
1658                  */
1659                 schedule_timeout_uninterruptible(1);
1660         }
1661         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 0);
1662         spin_lock(&sh->lock);
1663         set_bit(STRIPE_SYNCING, &sh->state);
1664         clear_bit(STRIPE_INSYNC, &sh->state);
1665         spin_unlock(&sh->lock);
1666
1667         handle_stripe(sh);
1668         release_stripe(sh);
1669
1670         return STRIPE_SECTORS;
1671 }
1672
1673 /*
1674  * This is our raid5 kernel thread.
1675  *
1676  * We scan the hash table for stripes which can be handled now.
1677  * During the scan, completed stripes are saved for us by the interrupt
1678  * handler, so that they will not have to wait for our next wakeup.
1679  */
1680 static void raid5d (mddev_t *mddev)
1681 {
1682         struct stripe_head *sh;
1683         raid5_conf_t *conf = mddev_to_conf(mddev);
1684         int handled;
1685
1686         PRINTK("+++ raid5d active\n");
1687
1688         md_check_recovery(mddev);
1689
1690         handled = 0;
1691         spin_lock_irq(&conf->device_lock);
1692         while (1) {
1693                 struct list_head *first;
1694
1695                 if (conf->seq_flush - conf->seq_write > 0) {
1696                         int seq = conf->seq_flush;
1697                         bitmap_unplug(mddev->bitmap);
1698                         conf->seq_write = seq;
1699                         activate_bit_delay(conf);
1700                 }
1701
1702                 if (list_empty(&conf->handle_list) &&
1703                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
1704                     !blk_queue_plugged(mddev->queue) &&
1705                     !list_empty(&conf->delayed_list))
1706                         raid5_activate_delayed(conf);
1707
1708                 if (list_empty(&conf->handle_list))
1709                         break;
1710
1711                 first = conf->handle_list.next;
1712                 sh = list_entry(first, struct stripe_head, lru);
1713
1714                 list_del_init(first);
1715                 atomic_inc(&sh->count);
1716                 if (atomic_read(&sh->count)!= 1)
1717                         BUG();
1718                 spin_unlock_irq(&conf->device_lock);
1719
1720                 handled++;
1721                 handle_stripe(sh);
1722                 release_stripe(sh);
1723
1724                 spin_lock_irq(&conf->device_lock);
1725         }
1726         PRINTK("%d stripes handled\n", handled);
1727
1728         spin_unlock_irq(&conf->device_lock);
1729
1730         unplug_slaves(mddev);
1731
1732         PRINTK("--- raid5d inactive\n");
1733 }
1734
1735 struct raid5_sysfs_entry {
1736         struct attribute attr;
1737         ssize_t (*show)(raid5_conf_t *, char *);
1738         ssize_t (*store)(raid5_conf_t *, const char *, ssize_t);
1739 };
1740
1741 static ssize_t
1742 raid5_show_stripe_cache_size(raid5_conf_t *conf, char *page)
1743 {
1744         return sprintf(page, "%d\n", conf->max_nr_stripes);
1745 }
1746
1747 static ssize_t
1748 raid5_store_stripe_cache_size(raid5_conf_t *conf, const char *page, ssize_t len)
1749 {
1750         char *end;
1751         int new;
1752         if (len >= PAGE_SIZE)
1753                 return -EINVAL;
1754
1755         new = simple_strtoul(page, &end, 10);
1756         if (!*page || (*end && *end != '\n') )
1757                 return -EINVAL;
1758         if (new <= 16 || new > 32768)
1759                 return -EINVAL;
1760         while (new < conf->max_nr_stripes) {
1761                 if (drop_one_stripe(conf))
1762                         conf->max_nr_stripes--;
1763                 else
1764                         break;
1765         }
1766         while (new > conf->max_nr_stripes) {
1767                 if (grow_one_stripe(conf))
1768                         conf->max_nr_stripes++;
1769                 else break;
1770         }
1771         return len;
1772 }
1773 static struct raid5_sysfs_entry raid5_stripecache_size = {
1774         .attr = {.name = "stripe_cache_size", .mode = S_IRUGO | S_IWUSR },
1775         .show = raid5_show_stripe_cache_size,
1776         .store = raid5_store_stripe_cache_size,
1777 };
1778
1779 static ssize_t
1780 raid5_show_stripe_cache_active(raid5_conf_t *conf, char *page)
1781 {
1782         return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
1783 }
1784
1785 static struct raid5_sysfs_entry raid5_stripecache_active = {
1786         .attr = {.name = "stripe_cache_active", .mode = S_IRUGO},
1787         .show = raid5_show_stripe_cache_active,
1788 };
1789
1790 static struct attribute *raid5_default_attrs[] = {
1791         &raid5_stripecache_size.attr,
1792         &raid5_stripecache_active.attr,
1793         NULL,
1794 };
1795
1796 static ssize_t
1797 raid5_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
1798 {
1799         struct raid5_sysfs_entry *entry = container_of(attr, struct raid5_sysfs_entry, attr);
1800         raid5_conf_t *conf = container_of(kobj, raid5_conf_t, kobj);
1801
1802         if (!entry->show)
1803                 return -EIO;
1804         return entry->show(conf, page);
1805 }
1806
1807 static ssize_t
1808 raid5_attr_store(struct kobject *kobj, struct attribute *attr,
1809               const char *page, size_t length)
1810 {
1811         struct raid5_sysfs_entry *entry = container_of(attr, struct raid5_sysfs_entry, attr);
1812         raid5_conf_t *conf = container_of(kobj, raid5_conf_t, kobj);
1813
1814         if (!entry->store)
1815                 return -EIO;
1816         return entry->store(conf, page, length);
1817 }
1818
1819 static void raid5_free(struct kobject *ko)
1820 {
1821         raid5_conf_t *conf = container_of(ko, raid5_conf_t, kobj);
1822         kfree(conf);
1823 }
1824
1825
1826 static struct sysfs_ops raid5_sysfs_ops = {
1827         .show           = raid5_attr_show,
1828         .store          = raid5_attr_store,
1829 };
1830
1831 static struct kobj_type raid5_ktype = {
1832         .release        = raid5_free,
1833         .sysfs_ops      = &raid5_sysfs_ops,
1834         .default_attrs  = raid5_default_attrs,
1835 };
1836
1837 static int run(mddev_t *mddev)
1838 {
1839         raid5_conf_t *conf;
1840         int raid_disk, memory;
1841         mdk_rdev_t *rdev;
1842         struct disk_info *disk;
1843         struct list_head *tmp;
1844
1845         if (mddev->level != 5 && mddev->level != 4) {
1846                 printk("raid5: %s: raid level not set to 4/5 (%d)\n", mdname(mddev), mddev->level);
1847                 return -EIO;
1848         }
1849
1850         mddev->private = kmalloc (sizeof (raid5_conf_t)
1851                                   + mddev->raid_disks * sizeof(struct disk_info),
1852                                   GFP_KERNEL);
1853         if ((conf = mddev->private) == NULL)
1854                 goto abort;
1855         memset (conf, 0, sizeof (*conf) + mddev->raid_disks * sizeof(struct disk_info) );
1856         conf->mddev = mddev;
1857
1858         if ((conf->stripe_hashtbl = (struct stripe_head **) __get_free_pages(GFP_ATOMIC, HASH_PAGES_ORDER)) == NULL)
1859                 goto abort;
1860         memset(conf->stripe_hashtbl, 0, HASH_PAGES * PAGE_SIZE);
1861
1862         spin_lock_init(&conf->device_lock);
1863         init_waitqueue_head(&conf->wait_for_stripe);
1864         init_waitqueue_head(&conf->wait_for_overlap);
1865         INIT_LIST_HEAD(&conf->handle_list);
1866         INIT_LIST_HEAD(&conf->delayed_list);
1867         INIT_LIST_HEAD(&conf->bitmap_list);
1868         INIT_LIST_HEAD(&conf->inactive_list);
1869         atomic_set(&conf->active_stripes, 0);
1870         atomic_set(&conf->preread_active_stripes, 0);
1871
1872         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
1873
1874         ITERATE_RDEV(mddev,rdev,tmp) {
1875                 raid_disk = rdev->raid_disk;
1876                 if (raid_disk >= mddev->raid_disks
1877                     || raid_disk < 0)
1878                         continue;
1879                 disk = conf->disks + raid_disk;
1880
1881                 disk->rdev = rdev;
1882
1883                 if (rdev->in_sync) {
1884                         char b[BDEVNAME_SIZE];
1885                         printk(KERN_INFO "raid5: device %s operational as raid"
1886                                 " disk %d\n", bdevname(rdev->bdev,b),
1887                                 raid_disk);
1888                         conf->working_disks++;
1889                 }
1890         }
1891
1892         conf->raid_disks = mddev->raid_disks;
1893         /*
1894          * 0 for a fully functional array, 1 for a degraded array.
1895          */
1896         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
1897         conf->mddev = mddev;
1898         conf->chunk_size = mddev->chunk_size;
1899         conf->level = mddev->level;
1900         conf->algorithm = mddev->layout;
1901         conf->max_nr_stripes = NR_STRIPES;
1902
1903         /* device size must be a multiple of chunk size */
1904         mddev->size &= ~(mddev->chunk_size/1024 -1);
1905         mddev->resync_max_sectors = mddev->size << 1;
1906
1907         if (!conf->chunk_size || conf->chunk_size % 4) {
1908                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
1909                         conf->chunk_size, mdname(mddev));
1910                 goto abort;
1911         }
1912         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
1913                 printk(KERN_ERR
1914                         "raid5: unsupported parity algorithm %d for %s\n",
1915                         conf->algorithm, mdname(mddev));
1916                 goto abort;
1917         }
1918         if (mddev->degraded > 1) {
1919                 printk(KERN_ERR "raid5: not enough operational devices for %s"
1920                         " (%d/%d failed)\n",
1921                         mdname(mddev), conf->failed_disks, conf->raid_disks);
1922                 goto abort;
1923         }
1924
1925         if (mddev->degraded == 1 &&
1926             mddev->recovery_cp != MaxSector) {
1927                 printk(KERN_ERR
1928                         "raid5: cannot start dirty degraded array for %s\n",
1929                         mdname(mddev));
1930                 goto abort;
1931         }
1932
1933         {
1934                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
1935                 if (!mddev->thread) {
1936                         printk(KERN_ERR
1937                                 "raid5: couldn't allocate thread for %s\n",
1938                                 mdname(mddev));
1939                         goto abort;
1940                 }
1941         }
1942 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
1943                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
1944         if (grow_stripes(conf, conf->max_nr_stripes)) {
1945                 printk(KERN_ERR
1946                         "raid5: couldn't allocate %dkB for buffers\n", memory);
1947                 shrink_stripes(conf);
1948                 md_unregister_thread(mddev->thread);
1949                 goto abort;
1950         } else
1951                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
1952                         memory, mdname(mddev));
1953
1954         if (mddev->degraded == 0)
1955                 printk("raid5: raid level %d set %s active with %d out of %d"
1956                         " devices, algorithm %d\n", conf->level, mdname(mddev),
1957                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
1958                         conf->algorithm);
1959         else
1960                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
1961                         " out of %d devices, algorithm %d\n", conf->level,
1962                         mdname(mddev), mddev->raid_disks - mddev->degraded,
1963                         mddev->raid_disks, conf->algorithm);
1964
1965         print_raid5_conf(conf);
1966
1967         /* read-ahead size must cover two whole stripes, which is
1968          * 2 * (n-1) * chunksize where 'n' is the number of raid devices
1969          */
1970         {
1971                 int stripe = (mddev->raid_disks-1) * mddev->chunk_size
1972                         / PAGE_CACHE_SIZE;
1973                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
1974                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
1975         }
1976
1977         /* Ok, everything is just fine now */
1978         conf->kobj.parent = kobject_get(&mddev->kobj);
1979         strcpy(conf->kobj.name, "raid5");
1980         conf->kobj.ktype = &raid5_ktype;
1981         kobject_register(&conf->kobj);
1982
1983         if (mddev->bitmap)
1984                 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
1985
1986         mddev->queue->unplug_fn = raid5_unplug_device;
1987         mddev->queue->issue_flush_fn = raid5_issue_flush;
1988
1989         mddev->array_size =  mddev->size * (mddev->raid_disks - 1);
1990         return 0;
1991 abort:
1992         if (conf) {
1993                 print_raid5_conf(conf);
1994                 if (conf->stripe_hashtbl)
1995                         free_pages((unsigned long) conf->stripe_hashtbl,
1996                                                         HASH_PAGES_ORDER);
1997                 kfree(conf);
1998         }
1999         mddev->private = NULL;
2000         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
2001         return -EIO;
2002 }
2003
2004
2005
2006 static int stop(mddev_t *mddev)
2007 {
2008         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2009
2010         md_unregister_thread(mddev->thread);
2011         mddev->thread = NULL;
2012         shrink_stripes(conf);
2013         free_pages((unsigned long) conf->stripe_hashtbl, HASH_PAGES_ORDER);
2014         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2015         kobject_unregister(&conf->kobj);
2016         mddev->private = NULL;
2017         return 0;
2018 }
2019
2020 #if RAID5_DEBUG
2021 static void print_sh (struct stripe_head *sh)
2022 {
2023         int i;
2024
2025         printk("sh %llu, pd_idx %d, state %ld.\n",
2026                 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
2027         printk("sh %llu,  count %d.\n",
2028                 (unsigned long long)sh->sector, atomic_read(&sh->count));
2029         printk("sh %llu, ", (unsigned long long)sh->sector);
2030         for (i = 0; i < sh->raid_conf->raid_disks; i++) {
2031                 printk("(cache%d: %p %ld) ",
2032                         i, sh->dev[i].page, sh->dev[i].flags);
2033         }
2034         printk("\n");
2035 }
2036
2037 static void printall (raid5_conf_t *conf)
2038 {
2039         struct stripe_head *sh;
2040         int i;
2041
2042         spin_lock_irq(&conf->device_lock);
2043         for (i = 0; i < NR_HASH; i++) {
2044                 sh = conf->stripe_hashtbl[i];
2045                 for (; sh; sh = sh->hash_next) {
2046                         if (sh->raid_conf != conf)
2047                                 continue;
2048                         print_sh(sh);
2049                 }
2050         }
2051         spin_unlock_irq(&conf->device_lock);
2052 }
2053 #endif
2054
2055 static void status (struct seq_file *seq, mddev_t *mddev)
2056 {
2057         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2058         int i;
2059
2060         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2061         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2062         for (i = 0; i < conf->raid_disks; i++)
2063                 seq_printf (seq, "%s",
2064                                conf->disks[i].rdev &&
2065                                conf->disks[i].rdev->in_sync ? "U" : "_");
2066         seq_printf (seq, "]");
2067 #if RAID5_DEBUG
2068 #define D(x) \
2069         seq_printf (seq, "<"#x":%d>", atomic_read(&conf->x))
2070         printall(conf);
2071 #endif
2072 }
2073
2074 static void print_raid5_conf (raid5_conf_t *conf)
2075 {
2076         int i;
2077         struct disk_info *tmp;
2078
2079         printk("RAID5 conf printout:\n");
2080         if (!conf) {
2081                 printk("(conf==NULL)\n");
2082                 return;
2083         }
2084         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2085                  conf->working_disks, conf->failed_disks);
2086
2087         for (i = 0; i < conf->raid_disks; i++) {
2088                 char b[BDEVNAME_SIZE];
2089                 tmp = conf->disks + i;
2090                 if (tmp->rdev)
2091                 printk(" disk %d, o:%d, dev:%s\n",
2092                         i, !tmp->rdev->faulty,
2093                         bdevname(tmp->rdev->bdev,b));
2094         }
2095 }
2096
2097 static int raid5_spare_active(mddev_t *mddev)
2098 {
2099         int i;
2100         raid5_conf_t *conf = mddev->private;
2101         struct disk_info *tmp;
2102
2103         for (i = 0; i < conf->raid_disks; i++) {
2104                 tmp = conf->disks + i;
2105                 if (tmp->rdev
2106                     && !tmp->rdev->faulty
2107                     && !tmp->rdev->in_sync) {
2108                         mddev->degraded--;
2109                         conf->failed_disks--;
2110                         conf->working_disks++;
2111                         tmp->rdev->in_sync = 1;
2112                 }
2113         }
2114         print_raid5_conf(conf);
2115         return 0;
2116 }
2117
2118 static int raid5_remove_disk(mddev_t *mddev, int number)
2119 {
2120         raid5_conf_t *conf = mddev->private;
2121         int err = 0;
2122         mdk_rdev_t *rdev;
2123         struct disk_info *p = conf->disks + number;
2124
2125         print_raid5_conf(conf);
2126         rdev = p->rdev;
2127         if (rdev) {
2128                 if (rdev->in_sync ||
2129                     atomic_read(&rdev->nr_pending)) {
2130                         err = -EBUSY;
2131                         goto abort;
2132                 }
2133                 p->rdev = NULL;
2134                 synchronize_rcu();
2135                 if (atomic_read(&rdev->nr_pending)) {
2136                         /* lost the race, try later */
2137                         err = -EBUSY;
2138                         p->rdev = rdev;
2139                 }
2140         }
2141 abort:
2142
2143         print_raid5_conf(conf);
2144         return err;
2145 }
2146
2147 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2148 {
2149         raid5_conf_t *conf = mddev->private;
2150         int found = 0;
2151         int disk;
2152         struct disk_info *p;
2153
2154         if (mddev->degraded > 1)
2155                 /* no point adding a device */
2156                 return 0;
2157
2158         /*
2159          * find the disk ...
2160          */
2161         for (disk=0; disk < mddev->raid_disks; disk++)
2162                 if ((p=conf->disks + disk)->rdev == NULL) {
2163                         rdev->in_sync = 0;
2164                         rdev->raid_disk = disk;
2165                         found = 1;
2166                         if (rdev->saved_raid_disk != disk)
2167                                 conf->fullsync = 1;
2168                         rcu_assign_pointer(p->rdev, rdev);
2169                         break;
2170                 }
2171         print_raid5_conf(conf);
2172         return found;
2173 }
2174
2175 static int raid5_resize(mddev_t *mddev, sector_t sectors)
2176 {
2177         /* no resync is happening, and there is enough space
2178          * on all devices, so we can resize.
2179          * We need to make sure resync covers any new space.
2180          * If the array is shrinking we should possibly wait until
2181          * any io in the removed space completes, but it hardly seems
2182          * worth it.
2183          */
2184         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2185         mddev->array_size = (sectors * (mddev->raid_disks-1))>>1;
2186         set_capacity(mddev->gendisk, mddev->array_size << 1);
2187         mddev->changed = 1;
2188         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2189                 mddev->recovery_cp = mddev->size << 1;
2190                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2191         }
2192         mddev->size = sectors /2;
2193         mddev->resync_max_sectors = sectors;
2194         return 0;
2195 }
2196
2197 static void raid5_quiesce(mddev_t *mddev, int state)
2198 {
2199         raid5_conf_t *conf = mddev_to_conf(mddev);
2200
2201         switch(state) {
2202         case 1: /* stop all writes */
2203                 spin_lock_irq(&conf->device_lock);
2204                 conf->quiesce = 1;
2205                 wait_event_lock_irq(conf->wait_for_stripe,
2206                                     atomic_read(&conf->active_stripes) == 0,
2207                                     conf->device_lock, /* nothing */);
2208                 spin_unlock_irq(&conf->device_lock);
2209                 break;
2210
2211         case 0: /* re-enable writes */
2212                 spin_lock_irq(&conf->device_lock);
2213                 conf->quiesce = 0;
2214                 wake_up(&conf->wait_for_stripe);
2215                 spin_unlock_irq(&conf->device_lock);
2216                 break;
2217         }
2218         if (mddev->thread) {
2219                 if (mddev->bitmap)
2220                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2221                 else
2222                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2223                 md_wakeup_thread(mddev->thread);
2224         }
2225 }
2226 static mdk_personality_t raid5_personality=
2227 {
2228         .name           = "raid5",
2229         .owner          = THIS_MODULE,
2230         .make_request   = make_request,
2231         .run            = run,
2232         .stop           = stop,
2233         .status         = status,
2234         .error_handler  = error,
2235         .hot_add_disk   = raid5_add_disk,
2236         .hot_remove_disk= raid5_remove_disk,
2237         .spare_active   = raid5_spare_active,
2238         .sync_request   = sync_request,
2239         .resize         = raid5_resize,
2240         .quiesce        = raid5_quiesce,
2241 };
2242
2243 static int __init raid5_init (void)
2244 {
2245         return register_md_personality (RAID5, &raid5_personality);
2246 }
2247
2248 static void raid5_exit (void)
2249 {
2250         unregister_md_personality (RAID5);
2251 }
2252
2253 module_init(raid5_init);
2254 module_exit(raid5_exit);
2255 MODULE_LICENSE("GPL");
2256 MODULE_ALIAS("md-personality-4"); /* RAID5 */