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) < (conf->max_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)
 268                                                      < (conf->max_nr_stripes *3/4)
 269                                                      || !conf->inactive_blocked),
 270                                                     conf->device_lock,
 271                                                     unplug_slaves(conf->mddev);
 272                                         );
 273                                 conf->inactive_blocked = 0;
 274                         } else
 275                                 init_stripe(sh, sector, pd_idx);
 276                 } else {
 277                         if (atomic_read(&sh->count)) {
 278                                 if (!list_empty(&sh->lru))
 279                                         BUG();
 280                         } else {
 281                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 282                                         atomic_inc(&conf->active_stripes);
 283                                 if (list_empty(&sh->lru))
 284                                         BUG();
 285                                 list_del_init(&sh->lru);
 286                         }
 287                 }
 288         } while (sh == NULL);
 289
 290         if (sh)
 291                 atomic_inc(&sh->count);
 292
 293         spin_unlock_irq(&conf->device_lock);
 294         return sh;
 295 }
 296
 297 static int grow_one_stripe(raid5_conf_t *conf)
 298 {
 299         struct stripe_head *sh;
 300         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
 301         if (!sh)
 302                 return 0;
 303         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
 304         sh->raid_conf = conf;
 305         spin_lock_init(&sh->lock);
 306
 307         if (grow_buffers(sh, conf->raid_disks)) {
 308                 shrink_buffers(sh, conf->raid_disks);
 309                 kmem_cache_free(conf->slab_cache, sh);
 310                 return 0;
 311         }
 312         /* we just created an active stripe so... */
 313         atomic_set(&sh->count, 1);
 314         atomic_inc(&conf->active_stripes);
 315         INIT_LIST_HEAD(&sh->lru);
 316         release_stripe(sh);
 317         return 1;
 318 }
 319
 320 static int grow_stripes(raid5_conf_t *conf, int num)
 321 {
 322         kmem_cache_t *sc;
 323         int devs = conf->raid_disks;
 324
 325         sprintf(conf->cache_name, "raid5/%s", mdname(conf->mddev));
 326
 327         sc = kmem_cache_create(conf->cache_name,
 328                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
 329                                0, 0, NULL, NULL);
 330         if (!sc)
 331                 return 1;
 332         conf->slab_cache = sc;
 333         while (num--) {
 334                 if (!grow_one_stripe(conf))
 335                         return 1;
 336         }
 337         return 0;
 338 }
 339
 340 static int drop_one_stripe(raid5_conf_t *conf)
 341 {
 342         struct stripe_head *sh;
 343
 344         spin_lock_irq(&conf->device_lock);
 345         sh = get_free_stripe(conf);
 346         spin_unlock_irq(&conf->device_lock);
 347         if (!sh)
 348                 return 0;
 349         if (atomic_read(&sh->count))
 350                 BUG();
 351         shrink_buffers(sh, conf->raid_disks);
 352         kmem_cache_free(conf->slab_cache, sh);
 353         atomic_dec(&conf->active_stripes);
 354         return 1;
 355 }
 356
 357 static void shrink_stripes(raid5_conf_t *conf)
 358 {
 359         while (drop_one_stripe(conf))
 360                 ;
 361
 362         kmem_cache_destroy(conf->slab_cache);
 363         conf->slab_cache = NULL;
 364 }
 365
 366 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
 367                                    int error)
 368 {
 369         struct stripe_head *sh = bi->bi_private;
 370         raid5_conf_t *conf = sh->raid_conf;
 371         int disks = conf->raid_disks, i;
 372         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 373
 374         if (bi->bi_size)
 375                 return 1;
 376
 377         for (i=0 ; i<disks; i++)
 378                 if (bi == &sh->dev[i].req)
 379                         break;
 380
 381         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
 382                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 383                 uptodate);
 384         if (i == disks) {
 385                 BUG();
 386                 return 0;
 387         }
 388
 389         if (uptodate) {
 390 #if 0
 391                 struct bio *bio;
 392                 unsigned long flags;
 393                 spin_lock_irqsave(&conf->device_lock, flags);
 394                 /* we can return a buffer if we bypassed the cache or
 395                  * if the top buffer is not in highmem.  If there are
 396                  * multiple buffers, leave the extra work to
 397                  * handle_stripe
 398                  */
 399                 buffer = sh->bh_read[i];
 400                 if (buffer &&
 401                     (!PageHighMem(buffer->b_page)
 402                      || buffer->b_page == bh->b_page )
 403                         ) {
 404                         sh->bh_read[i] = buffer->b_reqnext;
 405                         buffer->b_reqnext = NULL;
 406                 } else
 407                         buffer = NULL;
 408                 spin_unlock_irqrestore(&conf->device_lock, flags);
 409                 if (sh->bh_page[i]==bh->b_page)
 410                         set_buffer_uptodate(bh);
 411                 if (buffer) {
 412                         if (buffer->b_page != bh->b_page)
 413                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
 414                         buffer->b_end_io(buffer, 1);
 415                 }
 416 #else
 417                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
 418 #endif
 419                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
 420                         printk(KERN_INFO "raid5: read error corrected!!\n");
 421                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 422                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 423                 }
 424                 if (atomic_read(&conf->disks[i].rdev->read_errors))
 425                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
 426         } else {
 427                 int retry = 0;
 428                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
 429                 atomic_inc(&conf->disks[i].rdev->read_errors);
 430                 if (conf->mddev->degraded)
 431                         printk(KERN_WARNING "raid5: read error not correctable.\n");
 432                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
 433                         /* Oh, no!!! */
 434                         printk(KERN_WARNING "raid5: read error NOT corrected!!\n");
 435                 else if (atomic_read(&conf->disks[i].rdev->read_errors)
 436                          > conf->max_nr_stripes)
 437                         printk(KERN_WARNING
 438                                "raid5: Too many read errors, failing device.\n");
 439                 else
 440                         retry = 1;
 441                 if (retry)
 442                         set_bit(R5_ReadError, &sh->dev[i].flags);
 443                 else {
 444                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 445                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 446                         md_error(conf->mddev, conf->disks[i].rdev);
 447                 }
 448         }
 449         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 450 #if 0
 451         /* must restore b_page before unlocking buffer... */
 452         if (sh->bh_page[i] != bh->b_page) {
 453                 bh->b_page = sh->bh_page[i];
 454                 bh->b_data = page_address(bh->b_page);
 455                 clear_buffer_uptodate(bh);
 456         }
 457 #endif
 458         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 459         set_bit(STRIPE_HANDLE, &sh->state);
 460         release_stripe(sh);
 461         return 0;
 462 }
 463
 464 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
 465                                     int error)
 466 {
 467         struct stripe_head *sh = bi->bi_private;
 468         raid5_conf_t *conf = sh->raid_conf;
 469         int disks = conf->raid_disks, i;
 470         unsigned long flags;
 471         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 472
 473         if (bi->bi_size)
 474                 return 1;
 475
 476         for (i=0 ; i<disks; i++)
 477                 if (bi == &sh->dev[i].req)
 478                         break;
 479
 480         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
 481                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 482                 uptodate);
 483         if (i == disks) {
 484                 BUG();
 485                 return 0;
 486         }
 487
 488         spin_lock_irqsave(&conf->device_lock, flags);
 489         if (!uptodate)
 490                 md_error(conf->mddev, conf->disks[i].rdev);
 491
 492         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 493
 494         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 495         set_bit(STRIPE_HANDLE, &sh->state);
 496         __release_stripe(conf, sh);
 497         spin_unlock_irqrestore(&conf->device_lock, flags);
 498         return 0;
 499 }
 500
 501
 502 static sector_t compute_blocknr(struct stripe_head *sh, int i);
 503
 504 static void raid5_build_block (struct stripe_head *sh, int i)
 505 {
 506         struct r5dev *dev = &sh->dev[i];
 507
 508         bio_init(&dev->req);
 509         dev->req.bi_io_vec = &dev->vec;
 510         dev->req.bi_vcnt++;
 511         dev->req.bi_max_vecs++;
 512         dev->vec.bv_page = dev->page;
 513         dev->vec.bv_len = STRIPE_SIZE;
 514         dev->vec.bv_offset = 0;
 515
 516         dev->req.bi_sector = sh->sector;
 517         dev->req.bi_private = sh;
 518
 519         dev->flags = 0;
 520         if (i != sh->pd_idx)
 521                 dev->sector = compute_blocknr(sh, i);
 522 }
 523
 524 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 525 {
 526         char b[BDEVNAME_SIZE];
 527         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
 528         PRINTK("raid5: error called\n");
 529
 530         if (!test_bit(Faulty, &rdev->flags)) {
 531                 mddev->sb_dirty = 1;
 532                 if (test_bit(In_sync, &rdev->flags)) {
 533                         conf->working_disks--;
 534                         mddev->degraded++;
 535                         conf->failed_disks++;
 536                         clear_bit(In_sync, &rdev->flags);
 537                         /*
 538                          * if recovery was running, make sure it aborts.
 539                          */
 540                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 541                 }
 542                 set_bit(Faulty, &rdev->flags);
 543                 printk (KERN_ALERT
 544                         "raid5: Disk failure on %s, disabling device."
 545                         " Operation continuing on %d devices\n",
 546                         bdevname(rdev->bdev,b), conf->working_disks);
 547         }
 548 }
 549
 550 /*
 551  * Input: a 'big' sector number,
 552  * Output: index of the data and parity disk, and the sector # in them.
 553  */
 554 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
 555                         unsigned int data_disks, unsigned int * dd_idx,
 556                         unsigned int * pd_idx, raid5_conf_t *conf)
 557 {
 558         long stripe;
 559         unsigned long chunk_number;
 560         unsigned int chunk_offset;
 561         sector_t new_sector;
 562         int sectors_per_chunk = conf->chunk_size >> 9;
 563
 564         /* First compute the information on this sector */
 565
 566         /*
 567          * Compute the chunk number and the sector offset inside the chunk
 568          */
 569         chunk_offset = sector_div(r_sector, sectors_per_chunk);
 570         chunk_number = r_sector;
 571         BUG_ON(r_sector != chunk_number);
 572
 573         /*
 574          * Compute the stripe number
 575          */
 576         stripe = chunk_number / data_disks;
 577
 578         /*
 579          * Compute the data disk and parity disk indexes inside the stripe
 580          */
 581         *dd_idx = chunk_number % data_disks;
 582
 583         /*
 584          * Select the parity disk based on the user selected algorithm.
 585          */
 586         if (conf->level == 4)
 587                 *pd_idx = data_disks;
 588         else switch (conf->algorithm) {
 589                 case ALGORITHM_LEFT_ASYMMETRIC:
 590                         *pd_idx = data_disks - stripe % raid_disks;
 591                         if (*dd_idx >= *pd_idx)
 592                                 (*dd_idx)++;
 593                         break;
 594                 case ALGORITHM_RIGHT_ASYMMETRIC:
 595                         *pd_idx = stripe % raid_disks;
 596                         if (*dd_idx >= *pd_idx)
 597                                 (*dd_idx)++;
 598                         break;
 599                 case ALGORITHM_LEFT_SYMMETRIC:
 600                         *pd_idx = data_disks - stripe % raid_disks;
 601                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 602                         break;
 603                 case ALGORITHM_RIGHT_SYMMETRIC:
 604                         *pd_idx = stripe % raid_disks;
 605                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 606                         break;
 607                 default:
 608                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
 609                                 conf->algorithm);
 610         }
 611
 612         /*
 613          * Finally, compute the new sector number
 614          */
 615         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
 616         return new_sector;
 617 }
 618
 619
 620 static sector_t compute_blocknr(struct stripe_head *sh, int i)
 621 {
 622         raid5_conf_t *conf = sh->raid_conf;
 623         int raid_disks = conf->raid_disks, data_disks = raid_disks - 1;
 624         sector_t new_sector = sh->sector, check;
 625         int sectors_per_chunk = conf->chunk_size >> 9;
 626         sector_t stripe;
 627         int chunk_offset;
 628         int chunk_number, dummy1, dummy2, dd_idx = i;
 629         sector_t r_sector;
 630
 631         chunk_offset = sector_div(new_sector, sectors_per_chunk);
 632         stripe = new_sector;
 633         BUG_ON(new_sector != stripe);
 634
 635
 636         switch (conf->algorithm) {
 637                 case ALGORITHM_LEFT_ASYMMETRIC:
 638                 case ALGORITHM_RIGHT_ASYMMETRIC:
 639                         if (i > sh->pd_idx)
 640                                 i--;
 641                         break;
 642                 case ALGORITHM_LEFT_SYMMETRIC:
 643                 case ALGORITHM_RIGHT_SYMMETRIC:
 644                         if (i < sh->pd_idx)
 645                                 i += raid_disks;
 646                         i -= (sh->pd_idx + 1);
 647                         break;
 648                 default:
 649                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
 650                                 conf->algorithm);
 651         }
 652
 653         chunk_number = stripe * data_disks + i;
 654         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
 655
 656         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
 657         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
 658                 printk(KERN_ERR "compute_blocknr: map not correct\n");
 659                 return 0;
 660         }
 661         return r_sector;
 662 }
 663
 664
 665
 666 /*
 667  * Copy data between a page in the stripe cache, and a bio.
 668  * There are no alignment or size guarantees between the page or the
 669  * bio except that there is some overlap.
 670  * All iovecs in the bio must be considered.
 671  */
 672 static void copy_data(int frombio, struct bio *bio,
 673                      struct page *page,
 674                      sector_t sector)
 675 {
 676         char *pa = page_address(page);
 677         struct bio_vec *bvl;
 678         int i;
 679         int page_offset;
 680
 681         if (bio->bi_sector >= sector)
 682                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 683         else
 684                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 685         bio_for_each_segment(bvl, bio, i) {
 686                 int len = bio_iovec_idx(bio,i)->bv_len;
 687                 int clen;
 688                 int b_offset = 0;
 689
 690                 if (page_offset < 0) {
 691                         b_offset = -page_offset;
 692                         page_offset += b_offset;
 693                         len -= b_offset;
 694                 }
 695
 696                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 697                         clen = STRIPE_SIZE - page_offset;
 698                 else clen = len;
 699
 700                 if (clen > 0) {
 701                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
 702                         if (frombio)
 703                                 memcpy(pa+page_offset, ba+b_offset, clen);
 704                         else
 705                                 memcpy(ba+b_offset, pa+page_offset, clen);
 706                         __bio_kunmap_atomic(ba, KM_USER0);
 707                 }
 708                 if (clen < len) /* hit end of page */
 709                         break;
 710                 page_offset +=  len;
 711         }
 712 }
 713
 714 #define check_xor()     do {                                            \
 715                            if (count == MAX_XOR_BLOCKS) {               \
 716                                 xor_block(count, STRIPE_SIZE, ptr);     \
 717                                 count = 1;                              \
 718                            }                                            \
 719                         } while(0)
 720
 721
 722 static void compute_block(struct stripe_head *sh, int dd_idx)
 723 {
 724         raid5_conf_t *conf = sh->raid_conf;
 725         int i, count, disks = conf->raid_disks;
 726         void *ptr[MAX_XOR_BLOCKS], *p;
 727
 728         PRINTK("compute_block, stripe %llu, idx %d\n",
 729                 (unsigned long long)sh->sector, dd_idx);
 730
 731         ptr[0] = page_address(sh->dev[dd_idx].page);
 732         memset(ptr[0], 0, STRIPE_SIZE);
 733         count = 1;
 734         for (i = disks ; i--; ) {
 735                 if (i == dd_idx)
 736                         continue;
 737                 p = page_address(sh->dev[i].page);
 738                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
 739                         ptr[count++] = p;
 740                 else
 741                         printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
 742                                 " not present\n", dd_idx,
 743                                 (unsigned long long)sh->sector, i);
 744
 745                 check_xor();
 746         }
 747         if (count != 1)
 748                 xor_block(count, STRIPE_SIZE, ptr);
 749         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
 750 }
 751
 752 static void compute_parity(struct stripe_head *sh, int method)
 753 {
 754         raid5_conf_t *conf = sh->raid_conf;
 755         int i, pd_idx = sh->pd_idx, disks = conf->raid_disks, count;
 756         void *ptr[MAX_XOR_BLOCKS];
 757         struct bio *chosen;
 758
 759         PRINTK("compute_parity, stripe %llu, method %d\n",
 760                 (unsigned long long)sh->sector, method);
 761
 762         count = 1;
 763         ptr[0] = page_address(sh->dev[pd_idx].page);
 764         switch(method) {
 765         case READ_MODIFY_WRITE:
 766                 if (!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags))
 767                         BUG();
 768                 for (i=disks ; i-- ;) {
 769                         if (i==pd_idx)
 770                                 continue;
 771                         if (sh->dev[i].towrite &&
 772                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
 773                                 ptr[count++] = page_address(sh->dev[i].page);
 774                                 chosen = sh->dev[i].towrite;
 775                                 sh->dev[i].towrite = NULL;
 776
 777                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 778                                         wake_up(&conf->wait_for_overlap);
 779
 780                                 if (sh->dev[i].written) BUG();
 781                                 sh->dev[i].written = chosen;
 782                                 check_xor();
 783                         }
 784                 }
 785                 break;
 786         case RECONSTRUCT_WRITE:
 787                 memset(ptr[0], 0, STRIPE_SIZE);
 788                 for (i= disks; i-- ;)
 789                         if (i!=pd_idx && sh->dev[i].towrite) {
 790                                 chosen = sh->dev[i].towrite;
 791                                 sh->dev[i].towrite = NULL;
 792
 793                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 794                                         wake_up(&conf->wait_for_overlap);
 795
 796                                 if (sh->dev[i].written) BUG();
 797                                 sh->dev[i].written = chosen;
 798                         }
 799                 break;
 800         case CHECK_PARITY:
 801                 break;
 802         }
 803         if (count>1) {
 804                 xor_block(count, STRIPE_SIZE, ptr);
 805                 count = 1;
 806         }
 807
 808         for (i = disks; i--;)
 809                 if (sh->dev[i].written) {
 810                         sector_t sector = sh->dev[i].sector;
 811                         struct bio *wbi = sh->dev[i].written;
 812                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
 813                                 copy_data(1, wbi, sh->dev[i].page, sector);
 814                                 wbi = r5_next_bio(wbi, sector);
 815                         }
 816
 817                         set_bit(R5_LOCKED, &sh->dev[i].flags);
 818                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
 819                 }
 820
 821         switch(method) {
 822         case RECONSTRUCT_WRITE:
 823         case CHECK_PARITY:
 824                 for (i=disks; i--;)
 825                         if (i != pd_idx) {
 826                                 ptr[count++] = page_address(sh->dev[i].page);
 827                                 check_xor();
 828                         }
 829                 break;
 830         case READ_MODIFY_WRITE:
 831                 for (i = disks; i--;)
 832                         if (sh->dev[i].written) {
 833                                 ptr[count++] = page_address(sh->dev[i].page);
 834                                 check_xor();
 835                         }
 836         }
 837         if (count != 1)
 838                 xor_block(count, STRIPE_SIZE, ptr);
 839
 840         if (method != CHECK_PARITY) {
 841                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 842                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
 843         } else
 844                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 845 }
 846
 847 /*
 848  * Each stripe/dev can have one or more bion attached.
 849  * toread/towrite point to the first in a chain.
 850  * The bi_next chain must be in order.
 851  */
 852 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
 853 {
 854         struct bio **bip;
 855         raid5_conf_t *conf = sh->raid_conf;
 856         int firstwrite=0;
 857
 858         PRINTK("adding bh b#%llu to stripe s#%llu\n",
 859                 (unsigned long long)bi->bi_sector,
 860                 (unsigned long long)sh->sector);
 861
 862
 863         spin_lock(&sh->lock);
 864         spin_lock_irq(&conf->device_lock);
 865         if (forwrite) {
 866                 bip = &sh->dev[dd_idx].towrite;
 867                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
 868                         firstwrite = 1;
 869         } else
 870                 bip = &sh->dev[dd_idx].toread;
 871         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
 872                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
 873                         goto overlap;
 874                 bip = & (*bip)->bi_next;
 875         }
 876         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
 877                 goto overlap;
 878
 879         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
 880                 BUG();
 881         if (*bip)
 882                 bi->bi_next = *bip;
 883         *bip = bi;
 884         bi->bi_phys_segments ++;
 885         spin_unlock_irq(&conf->device_lock);
 886         spin_unlock(&sh->lock);
 887
 888         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
 889                 (unsigned long long)bi->bi_sector,
 890                 (unsigned long long)sh->sector, dd_idx);
 891
 892         if (conf->mddev->bitmap && firstwrite) {
 893                 sh->bm_seq = conf->seq_write;
 894                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
 895                                   STRIPE_SECTORS, 0);
 896                 set_bit(STRIPE_BIT_DELAY, &sh->state);
 897         }
 898
 899         if (forwrite) {
 900                 /* check if page is covered */
 901                 sector_t sector = sh->dev[dd_idx].sector;
 902                 for (bi=sh->dev[dd_idx].towrite;
 903                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
 904                              bi && bi->bi_sector <= sector;
 905                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
 906                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
 907                                 sector = bi->bi_sector + (bi->bi_size>>9);
 908                 }
 909                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
 910                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
 911         }
 912         return 1;
 913
 914  overlap:
 915         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
 916         spin_unlock_irq(&conf->device_lock);
 917         spin_unlock(&sh->lock);
 918         return 0;
 919 }
 920
 921
 922 /*
 923  * handle_stripe - do things to a stripe.
 924  *
 925  * We lock the stripe and then examine the state of various bits
 926  * to see what needs to be done.
 927  * Possible results:
 928  *    return some read request which now have data
 929  *    return some write requests which are safely on disc
 930  *    schedule a read on some buffers
 931  *    schedule a write of some buffers
 932  *    return confirmation of parity correctness
 933  *
 934  * Parity calculations are done inside the stripe lock
 935  * buffers are taken off read_list or write_list, and bh_cache buffers
 936  * get BH_Lock set before the stripe lock is released.
 937  *
 938  */
 939
 940 static void handle_stripe(struct stripe_head *sh)
 941 {
 942         raid5_conf_t *conf = sh->raid_conf;
 943         int disks = conf->raid_disks;
 944         struct bio *return_bi= NULL;
 945         struct bio *bi;
 946         int i;
 947         int syncing;
 948         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
 949         int non_overwrite = 0;
 950         int failed_num=0;
 951         struct r5dev *dev;
 952
 953         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
 954                 (unsigned long long)sh->sector, atomic_read(&sh->count),
 955                 sh->pd_idx);
 956
 957         spin_lock(&sh->lock);
 958         clear_bit(STRIPE_HANDLE, &sh->state);
 959         clear_bit(STRIPE_DELAYED, &sh->state);
 960
 961         syncing = test_bit(STRIPE_SYNCING, &sh->state);
 962         /* Now to look around and see what can be done */
 963
 964         rcu_read_lock();
 965         for (i=disks; i--; ) {
 966                 mdk_rdev_t *rdev;
 967                 dev = &sh->dev[i];
 968                 clear_bit(R5_Insync, &dev->flags);
 969
 970                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
 971                         i, dev->flags, dev->toread, dev->towrite, dev->written);
 972                 /* maybe we can reply to a read */
 973                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
 974                         struct bio *rbi, *rbi2;
 975                         PRINTK("Return read for disc %d\n", i);
 976                         spin_lock_irq(&conf->device_lock);
 977                         rbi = dev->toread;
 978                         dev->toread = NULL;
 979                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
 980                                 wake_up(&conf->wait_for_overlap);
 981                         spin_unlock_irq(&conf->device_lock);
 982                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
 983                                 copy_data(0, rbi, dev->page, dev->sector);
 984                                 rbi2 = r5_next_bio(rbi, dev->sector);
 985                                 spin_lock_irq(&conf->device_lock);
 986                                 if (--rbi->bi_phys_segments == 0) {
 987                                         rbi->bi_next = return_bi;
 988                                         return_bi = rbi;
 989                                 }
 990                                 spin_unlock_irq(&conf->device_lock);
 991                                 rbi = rbi2;
 992                         }
 993                 }
 994
 995                 /* now count some things */
 996                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
 997                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
 998
 999
1000                 if (dev->toread) to_read++;
1001                 if (dev->towrite) {
1002                         to_write++;
1003                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1004                                 non_overwrite++;
1005                 }
1006                 if (dev->written) written++;
1007                 rdev = rcu_dereference(conf->disks[i].rdev);
1008                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1009                         /* The ReadError flag will just be confusing now */
1010                         clear_bit(R5_ReadError, &dev->flags);
1011                         clear_bit(R5_ReWrite, &dev->flags);
1012                 }
1013                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1014                     || test_bit(R5_ReadError, &dev->flags)) {
1015                         failed++;
1016                         failed_num = i;
1017                 } else
1018                         set_bit(R5_Insync, &dev->flags);
1019         }
1020         rcu_read_unlock();
1021         PRINTK("locked=%d uptodate=%d to_read=%d"
1022                 " to_write=%d failed=%d failed_num=%d\n",
1023                 locked, uptodate, to_read, to_write, failed, failed_num);
1024         /* check if the array has lost two devices and, if so, some requests might
1025          * need to be failed
1026          */
1027         if (failed > 1 && to_read+to_write+written) {
1028                 for (i=disks; i--; ) {
1029                         int bitmap_end = 0;
1030
1031                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1032                                 mdk_rdev_t *rdev;
1033                                 rcu_read_lock();
1034                                 rdev = rcu_dereference(conf->disks[i].rdev);
1035                                 if (rdev && test_bit(In_sync, &rdev->flags))
1036                                         /* multiple read failures in one stripe */
1037                                         md_error(conf->mddev, rdev);
1038                                 rcu_read_unlock();
1039                         }
1040
1041                         spin_lock_irq(&conf->device_lock);
1042                         /* fail all writes first */
1043                         bi = sh->dev[i].towrite;
1044                         sh->dev[i].towrite = NULL;
1045                         if (bi) { to_write--; bitmap_end = 1; }
1046
1047                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1048                                 wake_up(&conf->wait_for_overlap);
1049
1050                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1051                                 struct bio *nextbi = 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 = nextbi;
1059                         }
1060                         /* and fail all 'written' */
1061                         bi = sh->dev[i].written;
1062                         sh->dev[i].written = NULL;
1063                         if (bi) bitmap_end = 1;
1064                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1065                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1066                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1067                                 if (--bi->bi_phys_segments == 0) {
1068                                         md_write_end(conf->mddev);
1069                                         bi->bi_next = return_bi;
1070                                         return_bi = bi;
1071                                 }
1072                                 bi = bi2;
1073                         }
1074
1075                         /* fail any reads if this device is non-operational */
1076                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1077                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1078                                 bi = sh->dev[i].toread;
1079                                 sh->dev[i].toread = NULL;
1080                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1081                                         wake_up(&conf->wait_for_overlap);
1082                                 if (bi) to_read--;
1083                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1084                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1085                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1086                                         if (--bi->bi_phys_segments == 0) {
1087                                                 bi->bi_next = return_bi;
1088                                                 return_bi = bi;
1089                                         }
1090                                         bi = nextbi;
1091                                 }
1092                         }
1093                         spin_unlock_irq(&conf->device_lock);
1094                         if (bitmap_end)
1095                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1096                                                 STRIPE_SECTORS, 0, 0);
1097                 }
1098         }
1099         if (failed > 1 && syncing) {
1100                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1101                 clear_bit(STRIPE_SYNCING, &sh->state);
1102                 syncing = 0;
1103         }
1104
1105         /* might be able to return some write requests if the parity block
1106          * is safe, or on a failed drive
1107          */
1108         dev = &sh->dev[sh->pd_idx];
1109         if ( written &&
1110              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1111                 test_bit(R5_UPTODATE, &dev->flags))
1112                || (failed == 1 && failed_num == sh->pd_idx))
1113             ) {
1114             /* any written block on an uptodate or failed drive can be returned.
1115              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
1116              * never LOCKED, so we don't need to test 'failed' directly.
1117              */
1118             for (i=disks; i--; )
1119                 if (sh->dev[i].written) {
1120                     dev = &sh->dev[i];
1121                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1122                          test_bit(R5_UPTODATE, &dev->flags) ) {
1123                         /* We can return any write requests */
1124                             struct bio *wbi, *wbi2;
1125                             int bitmap_end = 0;
1126                             PRINTK("Return write for disc %d\n", i);
1127                             spin_lock_irq(&conf->device_lock);
1128                             wbi = dev->written;
1129                             dev->written = NULL;
1130                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1131                                     wbi2 = r5_next_bio(wbi, dev->sector);
1132                                     if (--wbi->bi_phys_segments == 0) {
1133                                             md_write_end(conf->mddev);
1134                                             wbi->bi_next = return_bi;
1135                                             return_bi = wbi;
1136                                     }
1137                                     wbi = wbi2;
1138                             }
1139                             if (dev->towrite == NULL)
1140                                     bitmap_end = 1;
1141                             spin_unlock_irq(&conf->device_lock);
1142                             if (bitmap_end)
1143                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1144                                                     STRIPE_SECTORS,
1145                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1146                     }
1147                 }
1148         }
1149
1150         /* Now we might consider reading some blocks, either to check/generate
1151          * parity, or to satisfy requests
1152          * or to load a block that is being partially written.
1153          */
1154         if (to_read || non_overwrite || (syncing && (uptodate < disks))) {
1155                 for (i=disks; i--;) {
1156                         dev = &sh->dev[i];
1157                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1158                             (dev->toread ||
1159                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1160                              syncing ||
1161                              (failed && (sh->dev[failed_num].toread ||
1162                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1163                                     )
1164                                 ) {
1165                                 /* we would like to get this block, possibly
1166                                  * by computing it, but we might not be able to
1167                                  */
1168                                 if (uptodate == disks-1) {
1169                                         PRINTK("Computing block %d\n", i);
1170                                         compute_block(sh, i);
1171                                         uptodate++;
1172                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1173                                         set_bit(R5_LOCKED, &dev->flags);
1174                                         set_bit(R5_Wantread, &dev->flags);
1175 #if 0
1176                                         /* if I am just reading this block and we don't have
1177                                            a failed drive, or any pending writes then sidestep the cache */
1178                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1179                                             ! syncing && !failed && !to_write) {
1180                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1181                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1182                                         }
1183 #endif
1184                                         locked++;
1185                                         PRINTK("Reading block %d (sync=%d)\n",
1186                                                 i, syncing);
1187                                 }
1188                         }
1189                 }
1190                 set_bit(STRIPE_HANDLE, &sh->state);
1191         }
1192
1193         /* now to consider writing and what else, if anything should be read */
1194         if (to_write) {
1195                 int rmw=0, rcw=0;
1196                 for (i=disks ; i--;) {
1197                         /* would I have to read this buffer for read_modify_write */
1198                         dev = &sh->dev[i];
1199                         if ((dev->towrite || i == sh->pd_idx) &&
1200                             (!test_bit(R5_LOCKED, &dev->flags)
1201 #if 0
1202 || sh->bh_page[i]!=bh->b_page
1203 #endif
1204                                     ) &&
1205                             !test_bit(R5_UPTODATE, &dev->flags)) {
1206                                 if (test_bit(R5_Insync, &dev->flags)
1207 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1208                                         )
1209                                         rmw++;
1210                                 else rmw += 2*disks;  /* cannot read it */
1211                         }
1212                         /* Would I have to read this buffer for reconstruct_write */
1213                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1214                             (!test_bit(R5_LOCKED, &dev->flags)
1215 #if 0
1216 || sh->bh_page[i] != bh->b_page
1217 #endif
1218                                     ) &&
1219                             !test_bit(R5_UPTODATE, &dev->flags)) {
1220                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1221                                 else rcw += 2*disks;
1222                         }
1223                 }
1224                 PRINTK("for sector %llu, rmw=%d rcw=%d\n",
1225                         (unsigned long long)sh->sector, rmw, rcw);
1226                 set_bit(STRIPE_HANDLE, &sh->state);
1227                 if (rmw < rcw && rmw > 0)
1228                         /* prefer read-modify-write, but need to get some data */
1229                         for (i=disks; i--;) {
1230                                 dev = &sh->dev[i];
1231                                 if ((dev->towrite || i == sh->pd_idx) &&
1232                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1233                                     test_bit(R5_Insync, &dev->flags)) {
1234                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1235                                         {
1236                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1237                                                 set_bit(R5_LOCKED, &dev->flags);
1238                                                 set_bit(R5_Wantread, &dev->flags);
1239                                                 locked++;
1240                                         } else {
1241                                                 set_bit(STRIPE_DELAYED, &sh->state);
1242                                                 set_bit(STRIPE_HANDLE, &sh->state);
1243                                         }
1244                                 }
1245                         }
1246                 if (rcw <= rmw && rcw > 0)
1247                         /* want reconstruct write, but need to get some data */
1248                         for (i=disks; i--;) {
1249                                 dev = &sh->dev[i];
1250                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1251                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1252                                     test_bit(R5_Insync, &dev->flags)) {
1253                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1254                                         {
1255                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1256                                                 set_bit(R5_LOCKED, &dev->flags);
1257                                                 set_bit(R5_Wantread, &dev->flags);
1258                                                 locked++;
1259                                         } else {
1260                                                 set_bit(STRIPE_DELAYED, &sh->state);
1261                                                 set_bit(STRIPE_HANDLE, &sh->state);
1262                                         }
1263                                 }
1264                         }
1265                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1266                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1267                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1268                         PRINTK("Computing parity...\n");
1269                         compute_parity(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1270                         /* now every locked buffer is ready to be written */
1271                         for (i=disks; i--;)
1272                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1273                                         PRINTK("Writing block %d\n", i);
1274                                         locked++;
1275                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1276                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1277                                             || (i==sh->pd_idx && failed == 0))
1278                                                 set_bit(STRIPE_INSYNC, &sh->state);
1279                                 }
1280                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1281                                 atomic_dec(&conf->preread_active_stripes);
1282                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1283                                         md_wakeup_thread(conf->mddev->thread);
1284                         }
1285                 }
1286         }
1287
1288         /* maybe we need to check and possibly fix the parity for this stripe
1289          * Any reads will already have been scheduled, so we just see if enough data
1290          * is available
1291          */
1292         if (syncing && locked == 0 &&
1293             !test_bit(STRIPE_INSYNC, &sh->state)) {
1294                 set_bit(STRIPE_HANDLE, &sh->state);
1295                 if (failed == 0) {
1296                         char *pagea;
1297                         if (uptodate != disks)
1298                                 BUG();
1299                         compute_parity(sh, CHECK_PARITY);
1300                         uptodate--;
1301                         pagea = page_address(sh->dev[sh->pd_idx].page);
1302                         if ((*(u32*)pagea) == 0 &&
1303                             !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1304                                 /* parity is correct (on disc, not in buffer any more) */
1305                                 set_bit(STRIPE_INSYNC, &sh->state);
1306                         } else {
1307                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1308                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1309                                         /* don't try to repair!! */
1310                                         set_bit(STRIPE_INSYNC, &sh->state);
1311                                 else {
1312                                         compute_block(sh, sh->pd_idx);
1313                                         uptodate++;
1314                                 }
1315                         }
1316                 }
1317                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1318                         /* either failed parity check, or recovery is happening */
1319                         if (failed==0)
1320                                 failed_num = sh->pd_idx;
1321                         dev = &sh->dev[failed_num];
1322                         BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
1323                         BUG_ON(uptodate != disks);
1324
1325                         set_bit(R5_LOCKED, &dev->flags);
1326                         set_bit(R5_Wantwrite, &dev->flags);
1327                         clear_bit(STRIPE_DEGRADED, &sh->state);
1328                         locked++;
1329                         set_bit(STRIPE_INSYNC, &sh->state);
1330                 }
1331         }
1332         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1333                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1334                 clear_bit(STRIPE_SYNCING, &sh->state);
1335         }
1336
1337         /* If the failed drive is just a ReadError, then we might need to progress
1338          * the repair/check process
1339          */
1340         if (failed == 1 && ! conf->mddev->ro &&
1341             test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1342             && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1343             && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1344                 ) {
1345                 dev = &sh->dev[failed_num];
1346                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1347                         set_bit(R5_Wantwrite, &dev->flags);
1348                         set_bit(R5_ReWrite, &dev->flags);
1349                         set_bit(R5_LOCKED, &dev->flags);
1350                 } else {
1351                         /* let's read it back */
1352                         set_bit(R5_Wantread, &dev->flags);
1353                         set_bit(R5_LOCKED, &dev->flags);
1354                 }
1355         }
1356
1357         spin_unlock(&sh->lock);
1358
1359         while ((bi=return_bi)) {
1360                 int bytes = bi->bi_size;
1361
1362                 return_bi = bi->bi_next;
1363                 bi->bi_next = NULL;
1364                 bi->bi_size = 0;
1365                 bi->bi_end_io(bi, bytes, 0);
1366         }
1367         for (i=disks; i-- ;) {
1368                 int rw;
1369                 struct bio *bi;
1370                 mdk_rdev_t *rdev;
1371                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1372                         rw = 1;
1373                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1374                         rw = 0;
1375                 else
1376                         continue;
1377
1378                 bi = &sh->dev[i].req;
1379
1380                 bi->bi_rw = rw;
1381                 if (rw)
1382                         bi->bi_end_io = raid5_end_write_request;
1383                 else
1384                         bi->bi_end_io = raid5_end_read_request;
1385
1386                 rcu_read_lock();
1387                 rdev = rcu_dereference(conf->disks[i].rdev);
1388                 if (rdev && test_bit(Faulty, &rdev->flags))
1389                         rdev = NULL;
1390                 if (rdev)
1391                         atomic_inc(&rdev->nr_pending);
1392                 rcu_read_unlock();
1393
1394                 if (rdev) {
1395                         if (syncing)
1396                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1397
1398                         bi->bi_bdev = rdev->bdev;
1399                         PRINTK("for %llu schedule op %ld on disc %d\n",
1400                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1401                         atomic_inc(&sh->count);
1402                         bi->bi_sector = sh->sector + rdev->data_offset;
1403                         bi->bi_flags = 1 << BIO_UPTODATE;
1404                         bi->bi_vcnt = 1;
1405                         bi->bi_max_vecs = 1;
1406                         bi->bi_idx = 0;
1407                         bi->bi_io_vec = &sh->dev[i].vec;
1408                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1409                         bi->bi_io_vec[0].bv_offset = 0;
1410                         bi->bi_size = STRIPE_SIZE;
1411                         bi->bi_next = NULL;
1412                         generic_make_request(bi);
1413                 } else {
1414                         if (rw == 1)
1415                                 set_bit(STRIPE_DEGRADED, &sh->state);
1416                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1417                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1418                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1419                         set_bit(STRIPE_HANDLE, &sh->state);
1420                 }
1421         }
1422 }
1423
1424 static inline void raid5_activate_delayed(raid5_conf_t *conf)
1425 {
1426         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1427                 while (!list_empty(&conf->delayed_list)) {
1428                         struct list_head *l = conf->delayed_list.next;
1429                         struct stripe_head *sh;
1430                         sh = list_entry(l, struct stripe_head, lru);
1431                         list_del_init(l);
1432                         clear_bit(STRIPE_DELAYED, &sh->state);
1433                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1434                                 atomic_inc(&conf->preread_active_stripes);
1435                         list_add_tail(&sh->lru, &conf->handle_list);
1436                 }
1437         }
1438 }
1439
1440 static inline void activate_bit_delay(raid5_conf_t *conf)
1441 {
1442         /* device_lock is held */
1443         struct list_head head;
1444         list_add(&head, &conf->bitmap_list);
1445         list_del_init(&conf->bitmap_list);
1446         while (!list_empty(&head)) {
1447                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1448                 list_del_init(&sh->lru);
1449                 atomic_inc(&sh->count);
1450                 __release_stripe(conf, sh);
1451         }
1452 }
1453
1454 static void unplug_slaves(mddev_t *mddev)
1455 {
1456         raid5_conf_t *conf = mddev_to_conf(mddev);
1457         int i;
1458
1459         rcu_read_lock();
1460         for (i=0; i<mddev->raid_disks; i++) {
1461                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1462                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
1463                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1464
1465                         atomic_inc(&rdev->nr_pending);
1466                         rcu_read_unlock();
1467
1468                         if (r_queue->unplug_fn)
1469                                 r_queue->unplug_fn(r_queue);
1470
1471                         rdev_dec_pending(rdev, mddev);
1472                         rcu_read_lock();
1473                 }
1474         }
1475         rcu_read_unlock();
1476 }
1477
1478 static void raid5_unplug_device(request_queue_t *q)
1479 {
1480         mddev_t *mddev = q->queuedata;
1481         raid5_conf_t *conf = mddev_to_conf(mddev);
1482         unsigned long flags;
1483
1484         spin_lock_irqsave(&conf->device_lock, flags);
1485
1486         if (blk_remove_plug(q)) {
1487                 conf->seq_flush++;
1488                 raid5_activate_delayed(conf);
1489         }
1490         md_wakeup_thread(mddev->thread);
1491
1492         spin_unlock_irqrestore(&conf->device_lock, flags);
1493
1494         unplug_slaves(mddev);
1495 }
1496
1497 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
1498                              sector_t *error_sector)
1499 {
1500         mddev_t *mddev = q->queuedata;
1501         raid5_conf_t *conf = mddev_to_conf(mddev);
1502         int i, ret = 0;
1503
1504         rcu_read_lock();
1505         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1506                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1507                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
1508                         struct block_device *bdev = rdev->bdev;
1509                         request_queue_t *r_queue = bdev_get_queue(bdev);
1510
1511                         if (!r_queue->issue_flush_fn)
1512                                 ret = -EOPNOTSUPP;
1513                         else {
1514                                 atomic_inc(&rdev->nr_pending);
1515                                 rcu_read_unlock();
1516                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1517                                                               error_sector);
1518                                 rdev_dec_pending(rdev, mddev);
1519                                 rcu_read_lock();
1520                         }
1521                 }
1522         }
1523         rcu_read_unlock();
1524         return ret;
1525 }
1526
1527 static inline void raid5_plug_device(raid5_conf_t *conf)
1528 {
1529         spin_lock_irq(&conf->device_lock);
1530         blk_plug_device(conf->mddev->queue);
1531         spin_unlock_irq(&conf->device_lock);
1532 }
1533
1534 static int make_request (request_queue_t *q, struct bio * bi)
1535 {
1536         mddev_t *mddev = q->queuedata;
1537         raid5_conf_t *conf = mddev_to_conf(mddev);
1538         const unsigned int raid_disks = conf->raid_disks;
1539         const unsigned int data_disks = raid_disks - 1;
1540         unsigned int dd_idx, pd_idx;
1541         sector_t new_sector;
1542         sector_t logical_sector, last_sector;
1543         struct stripe_head *sh;
1544         const int rw = bio_data_dir(bi);
1545
1546         if (unlikely(bio_barrier(bi))) {
1547                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1548                 return 0;
1549         }
1550
1551         md_write_start(mddev, bi);
1552
1553         disk_stat_inc(mddev->gendisk, ios[rw]);
1554         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
1555
1556         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1557         last_sector = bi->bi_sector + (bi->bi_size>>9);
1558         bi->bi_next = NULL;
1559         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1560
1561         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1562                 DEFINE_WAIT(w);
1563
1564                 new_sector = raid5_compute_sector(logical_sector,
1565                                                   raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1566
1567                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
1568                         (unsigned long long)new_sector,
1569                         (unsigned long long)logical_sector);
1570
1571         retry:
1572                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1573                 sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
1574                 if (sh) {
1575                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1576                                 /* Add failed due to overlap.  Flush everything
1577                                  * and wait a while
1578                                  */
1579                                 raid5_unplug_device(mddev->queue);
1580                                 release_stripe(sh);
1581                                 schedule();
1582                                 goto retry;
1583                         }
1584                         finish_wait(&conf->wait_for_overlap, &w);
1585                         raid5_plug_device(conf);
1586                         handle_stripe(sh);
1587                         release_stripe(sh);
1588
1589                 } else {
1590                         /* cannot get stripe for read-ahead, just give-up */
1591                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1592                         finish_wait(&conf->wait_for_overlap, &w);
1593                         break;
1594                 }
1595
1596         }
1597         spin_lock_irq(&conf->device_lock);
1598         if (--bi->bi_phys_segments == 0) {
1599                 int bytes = bi->bi_size;
1600
1601                 if ( bio_data_dir(bi) == WRITE )
1602                         md_write_end(mddev);
1603                 bi->bi_size = 0;
1604                 bi->bi_end_io(bi, bytes, 0);
1605         }
1606         spin_unlock_irq(&conf->device_lock);
1607         return 0;
1608 }
1609
1610 /* FIXME go_faster isn't used */
1611 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1612 {
1613         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1614         struct stripe_head *sh;
1615         int sectors_per_chunk = conf->chunk_size >> 9;
1616         sector_t x;
1617         unsigned long stripe;
1618         int chunk_offset;
1619         int dd_idx, pd_idx;
1620         sector_t first_sector;
1621         int raid_disks = conf->raid_disks;
1622         int data_disks = raid_disks-1;
1623         sector_t max_sector = mddev->size << 1;
1624         int sync_blocks;
1625
1626         if (sector_nr >= max_sector) {
1627                 /* just being told to finish up .. nothing much to do */
1628                 unplug_slaves(mddev);
1629
1630                 if (mddev->curr_resync < max_sector) /* aborted */
1631                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1632                                         &sync_blocks, 1);
1633                 else /* compelted sync */
1634                         conf->fullsync = 0;
1635                 bitmap_close_sync(mddev->bitmap);
1636
1637                 return 0;
1638         }
1639         /* if there is 1 or more failed drives and we are trying
1640          * to resync, then assert that we are finished, because there is
1641          * nothing we can do.
1642          */
1643         if (mddev->degraded >= 1 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1644                 sector_t rv = (mddev->size << 1) - sector_nr;
1645                 *skipped = 1;
1646                 return rv;
1647         }
1648         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1649             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1650             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
1651                 /* we can skip this block, and probably more */
1652                 sync_blocks /= STRIPE_SECTORS;
1653                 *skipped = 1;
1654                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
1655         }
1656
1657         x = sector_nr;
1658         chunk_offset = sector_div(x, sectors_per_chunk);
1659         stripe = x;
1660         BUG_ON(x != stripe);
1661
1662         first_sector = raid5_compute_sector((sector_t)stripe*data_disks*sectors_per_chunk
1663                 + chunk_offset, raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1664         sh = get_active_stripe(conf, sector_nr, pd_idx, 1);
1665         if (sh == NULL) {
1666                 sh = get_active_stripe(conf, sector_nr, pd_idx, 0);
1667                 /* make sure we don't swamp the stripe cache if someone else
1668                  * is trying to get access
1669                  */
1670                 schedule_timeout_uninterruptible(1);
1671         }
1672         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 0);
1673         spin_lock(&sh->lock);
1674         set_bit(STRIPE_SYNCING, &sh->state);
1675         clear_bit(STRIPE_INSYNC, &sh->state);
1676         spin_unlock(&sh->lock);
1677
1678         handle_stripe(sh);
1679         release_stripe(sh);
1680
1681         return STRIPE_SECTORS;
1682 }
1683
1684 /*
1685  * This is our raid5 kernel thread.
1686  *
1687  * We scan the hash table for stripes which can be handled now.
1688  * During the scan, completed stripes are saved for us by the interrupt
1689  * handler, so that they will not have to wait for our next wakeup.
1690  */
1691 static void raid5d (mddev_t *mddev)
1692 {
1693         struct stripe_head *sh;
1694         raid5_conf_t *conf = mddev_to_conf(mddev);
1695         int handled;
1696
1697         PRINTK("+++ raid5d active\n");
1698
1699         md_check_recovery(mddev);
1700
1701         handled = 0;
1702         spin_lock_irq(&conf->device_lock);
1703         while (1) {
1704                 struct list_head *first;
1705
1706                 if (conf->seq_flush - conf->seq_write > 0) {
1707                         int seq = conf->seq_flush;
1708                         spin_unlock_irq(&conf->device_lock);
1709                         bitmap_unplug(mddev->bitmap);
1710                         spin_lock_irq(&conf->device_lock);
1711                         conf->seq_write = seq;
1712                         activate_bit_delay(conf);
1713                 }
1714
1715                 if (list_empty(&conf->handle_list) &&
1716                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
1717                     !blk_queue_plugged(mddev->queue) &&
1718                     !list_empty(&conf->delayed_list))
1719                         raid5_activate_delayed(conf);
1720
1721                 if (list_empty(&conf->handle_list))
1722                         break;
1723
1724                 first = conf->handle_list.next;
1725                 sh = list_entry(first, struct stripe_head, lru);
1726
1727                 list_del_init(first);
1728                 atomic_inc(&sh->count);
1729                 if (atomic_read(&sh->count)!= 1)
1730                         BUG();
1731                 spin_unlock_irq(&conf->device_lock);
1732
1733                 handled++;
1734                 handle_stripe(sh);
1735                 release_stripe(sh);
1736
1737                 spin_lock_irq(&conf->device_lock);
1738         }
1739         PRINTK("%d stripes handled\n", handled);
1740
1741         spin_unlock_irq(&conf->device_lock);
1742
1743         unplug_slaves(mddev);
1744
1745         PRINTK("--- raid5d inactive\n");
1746 }
1747
1748 static ssize_t
1749 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
1750 {
1751         raid5_conf_t *conf = mddev_to_conf(mddev);
1752         if (conf)
1753                 return sprintf(page, "%d\n", conf->max_nr_stripes);
1754         else
1755                 return 0;
1756 }
1757
1758 static ssize_t
1759 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
1760 {
1761         raid5_conf_t *conf = mddev_to_conf(mddev);
1762         char *end;
1763         int new;
1764         if (len >= PAGE_SIZE)
1765                 return -EINVAL;
1766         if (!conf)
1767                 return -ENODEV;
1768
1769         new = simple_strtoul(page, &end, 10);
1770         if (!*page || (*end && *end != '\n') )
1771                 return -EINVAL;
1772         if (new <= 16 || new > 32768)
1773                 return -EINVAL;
1774         while (new < conf->max_nr_stripes) {
1775                 if (drop_one_stripe(conf))
1776                         conf->max_nr_stripes--;
1777                 else
1778                         break;
1779         }
1780         while (new > conf->max_nr_stripes) {
1781                 if (grow_one_stripe(conf))
1782                         conf->max_nr_stripes++;
1783                 else break;
1784         }
1785         return len;
1786 }
1787
1788 static struct md_sysfs_entry
1789 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
1790                                 raid5_show_stripe_cache_size,
1791                                 raid5_store_stripe_cache_size);
1792
1793 static ssize_t
1794 stripe_cache_active_show(mddev_t *mddev, char *page)
1795 {
1796         raid5_conf_t *conf = mddev_to_conf(mddev);
1797         if (conf)
1798                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
1799         else
1800                 return 0;
1801 }
1802
1803 static struct md_sysfs_entry
1804 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
1805
1806 static struct attribute *raid5_attrs[] =  {
1807         &raid5_stripecache_size.attr,
1808         &raid5_stripecache_active.attr,
1809         NULL,
1810 };
1811 static struct attribute_group raid5_attrs_group = {
1812         .name = NULL,
1813         .attrs = raid5_attrs,
1814 };
1815
1816 static int run(mddev_t *mddev)
1817 {
1818         raid5_conf_t *conf;
1819         int raid_disk, memory;
1820         mdk_rdev_t *rdev;
1821         struct disk_info *disk;
1822         struct list_head *tmp;
1823
1824         if (mddev->level != 5 && mddev->level != 4) {
1825                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5 (%d)\n",
1826                        mdname(mddev), mddev->level);
1827                 return -EIO;
1828         }
1829
1830         mddev->private = kmalloc (sizeof (raid5_conf_t)
1831                                   + mddev->raid_disks * sizeof(struct disk_info),
1832                                   GFP_KERNEL);
1833         if ((conf = mddev->private) == NULL)
1834                 goto abort;
1835         memset (conf, 0, sizeof (*conf) + mddev->raid_disks * sizeof(struct disk_info) );
1836         conf->mddev = mddev;
1837
1838         if ((conf->stripe_hashtbl = (struct stripe_head **) __get_free_pages(GFP_ATOMIC, HASH_PAGES_ORDER)) == NULL)
1839                 goto abort;
1840         memset(conf->stripe_hashtbl, 0, HASH_PAGES * PAGE_SIZE);
1841
1842         spin_lock_init(&conf->device_lock);
1843         init_waitqueue_head(&conf->wait_for_stripe);
1844         init_waitqueue_head(&conf->wait_for_overlap);
1845         INIT_LIST_HEAD(&conf->handle_list);
1846         INIT_LIST_HEAD(&conf->delayed_list);
1847         INIT_LIST_HEAD(&conf->bitmap_list);
1848         INIT_LIST_HEAD(&conf->inactive_list);
1849         atomic_set(&conf->active_stripes, 0);
1850         atomic_set(&conf->preread_active_stripes, 0);
1851
1852         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
1853
1854         ITERATE_RDEV(mddev,rdev,tmp) {
1855                 raid_disk = rdev->raid_disk;
1856                 if (raid_disk >= mddev->raid_disks
1857                     || raid_disk < 0)
1858                         continue;
1859                 disk = conf->disks + raid_disk;
1860
1861                 disk->rdev = rdev;
1862
1863                 if (test_bit(In_sync, &rdev->flags)) {
1864                         char b[BDEVNAME_SIZE];
1865                         printk(KERN_INFO "raid5: device %s operational as raid"
1866                                 " disk %d\n", bdevname(rdev->bdev,b),
1867                                 raid_disk);
1868                         conf->working_disks++;
1869                 }
1870         }
1871
1872         conf->raid_disks = mddev->raid_disks;
1873         /*
1874          * 0 for a fully functional array, 1 for a degraded array.
1875          */
1876         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
1877         conf->mddev = mddev;
1878         conf->chunk_size = mddev->chunk_size;
1879         conf->level = mddev->level;
1880         conf->algorithm = mddev->layout;
1881         conf->max_nr_stripes = NR_STRIPES;
1882
1883         /* device size must be a multiple of chunk size */
1884         mddev->size &= ~(mddev->chunk_size/1024 -1);
1885         mddev->resync_max_sectors = mddev->size << 1;
1886
1887         if (!conf->chunk_size || conf->chunk_size % 4) {
1888                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
1889                         conf->chunk_size, mdname(mddev));
1890                 goto abort;
1891         }
1892         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
1893                 printk(KERN_ERR
1894                         "raid5: unsupported parity algorithm %d for %s\n",
1895                         conf->algorithm, mdname(mddev));
1896                 goto abort;
1897         }
1898         if (mddev->degraded > 1) {
1899                 printk(KERN_ERR "raid5: not enough operational devices for %s"
1900                         " (%d/%d failed)\n",
1901                         mdname(mddev), conf->failed_disks, conf->raid_disks);
1902                 goto abort;
1903         }
1904
1905         if (mddev->degraded == 1 &&
1906             mddev->recovery_cp != MaxSector) {
1907                 if (mddev->ok_start_degraded)
1908                         printk(KERN_WARNING
1909                                "raid5: starting dirty degraded array: %s"
1910                                "- data corruption possible.\n",
1911                                mdname(mddev));
1912                 else {
1913                         printk(KERN_ERR
1914                                "raid5: cannot start dirty degraded array for %s\n",
1915                                mdname(mddev));
1916                         goto abort;
1917                 }
1918         }
1919
1920         {
1921                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
1922                 if (!mddev->thread) {
1923                         printk(KERN_ERR
1924                                 "raid5: couldn't allocate thread for %s\n",
1925                                 mdname(mddev));
1926                         goto abort;
1927                 }
1928         }
1929         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
1930                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
1931         if (grow_stripes(conf, conf->max_nr_stripes)) {
1932                 printk(KERN_ERR
1933                         "raid5: couldn't allocate %dkB for buffers\n", memory);
1934                 shrink_stripes(conf);
1935                 md_unregister_thread(mddev->thread);
1936                 goto abort;
1937         } else
1938                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
1939                         memory, mdname(mddev));
1940
1941         if (mddev->degraded == 0)
1942                 printk("raid5: raid level %d set %s active with %d out of %d"
1943                         " devices, algorithm %d\n", conf->level, mdname(mddev),
1944                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
1945                         conf->algorithm);
1946         else
1947                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
1948                         " out of %d devices, algorithm %d\n", conf->level,
1949                         mdname(mddev), mddev->raid_disks - mddev->degraded,
1950                         mddev->raid_disks, conf->algorithm);
1951
1952         print_raid5_conf(conf);
1953
1954         /* read-ahead size must cover two whole stripes, which is
1955          * 2 * (n-1) * chunksize where 'n' is the number of raid devices
1956          */
1957         {
1958                 int stripe = (mddev->raid_disks-1) * mddev->chunk_size
1959                         / PAGE_CACHE_SIZE;
1960                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
1961                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
1962         }
1963
1964         /* Ok, everything is just fine now */
1965         sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
1966
1967         mddev->queue->unplug_fn = raid5_unplug_device;
1968         mddev->queue->issue_flush_fn = raid5_issue_flush;
1969
1970         mddev->array_size =  mddev->size * (mddev->raid_disks - 1);
1971         return 0;
1972 abort:
1973         if (conf) {
1974                 print_raid5_conf(conf);
1975                 if (conf->stripe_hashtbl)
1976                         free_pages((unsigned long) conf->stripe_hashtbl,
1977                                                         HASH_PAGES_ORDER);
1978                 kfree(conf);
1979         }
1980         mddev->private = NULL;
1981         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
1982         return -EIO;
1983 }
1984
1985
1986
1987 static int stop(mddev_t *mddev)
1988 {
1989         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1990
1991         md_unregister_thread(mddev->thread);
1992         mddev->thread = NULL;
1993         shrink_stripes(conf);
1994         free_pages((unsigned long) conf->stripe_hashtbl, HASH_PAGES_ORDER);
1995         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1996         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
1997         kfree(conf);
1998         mddev->private = NULL;
1999         return 0;
2000 }
2001
2002 #if RAID5_DEBUG
2003 static void print_sh (struct stripe_head *sh)
2004 {
2005         int i;
2006
2007         printk("sh %llu, pd_idx %d, state %ld.\n",
2008                 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
2009         printk("sh %llu,  count %d.\n",
2010                 (unsigned long long)sh->sector, atomic_read(&sh->count));
2011         printk("sh %llu, ", (unsigned long long)sh->sector);
2012         for (i = 0; i < sh->raid_conf->raid_disks; i++) {
2013                 printk("(cache%d: %p %ld) ",
2014                         i, sh->dev[i].page, sh->dev[i].flags);
2015         }
2016         printk("\n");
2017 }
2018
2019 static void printall (raid5_conf_t *conf)
2020 {
2021         struct stripe_head *sh;
2022         int i;
2023
2024         spin_lock_irq(&conf->device_lock);
2025         for (i = 0; i < NR_HASH; i++) {
2026                 sh = conf->stripe_hashtbl[i];
2027                 for (; sh; sh = sh->hash_next) {
2028                         if (sh->raid_conf != conf)
2029                                 continue;
2030                         print_sh(sh);
2031                 }
2032         }
2033         spin_unlock_irq(&conf->device_lock);
2034 }
2035 #endif
2036
2037 static void status (struct seq_file *seq, mddev_t *mddev)
2038 {
2039         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2040         int i;
2041
2042         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2043         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2044         for (i = 0; i < conf->raid_disks; i++)
2045                 seq_printf (seq, "%s",
2046                                conf->disks[i].rdev &&
2047                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
2048         seq_printf (seq, "]");
2049 #if RAID5_DEBUG
2050 #define D(x) \
2051         seq_printf (seq, "<"#x":%d>", atomic_read(&conf->x))
2052         printall(conf);
2053 #endif
2054 }
2055
2056 static void print_raid5_conf (raid5_conf_t *conf)
2057 {
2058         int i;
2059         struct disk_info *tmp;
2060
2061         printk("RAID5 conf printout:\n");
2062         if (!conf) {
2063                 printk("(conf==NULL)\n");
2064                 return;
2065         }
2066         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2067                  conf->working_disks, conf->failed_disks);
2068
2069         for (i = 0; i < conf->raid_disks; i++) {
2070                 char b[BDEVNAME_SIZE];
2071                 tmp = conf->disks + i;
2072                 if (tmp->rdev)
2073                 printk(" disk %d, o:%d, dev:%s\n",
2074                         i, !test_bit(Faulty, &tmp->rdev->flags),
2075                         bdevname(tmp->rdev->bdev,b));
2076         }
2077 }
2078
2079 static int raid5_spare_active(mddev_t *mddev)
2080 {
2081         int i;
2082         raid5_conf_t *conf = mddev->private;
2083         struct disk_info *tmp;
2084
2085         for (i = 0; i < conf->raid_disks; i++) {
2086                 tmp = conf->disks + i;
2087                 if (tmp->rdev
2088                     && !test_bit(Faulty, &tmp->rdev->flags)
2089                     && !test_bit(In_sync, &tmp->rdev->flags)) {
2090                         mddev->degraded--;
2091                         conf->failed_disks--;
2092                         conf->working_disks++;
2093                         set_bit(In_sync, &tmp->rdev->flags);
2094                 }
2095         }
2096         print_raid5_conf(conf);
2097         return 0;
2098 }
2099
2100 static int raid5_remove_disk(mddev_t *mddev, int number)
2101 {
2102         raid5_conf_t *conf = mddev->private;
2103         int err = 0;
2104         mdk_rdev_t *rdev;
2105         struct disk_info *p = conf->disks + number;
2106
2107         print_raid5_conf(conf);
2108         rdev = p->rdev;
2109         if (rdev) {
2110                 if (test_bit(In_sync, &rdev->flags) ||
2111                     atomic_read(&rdev->nr_pending)) {
2112                         err = -EBUSY;
2113                         goto abort;
2114                 }
2115                 p->rdev = NULL;
2116                 synchronize_rcu();
2117                 if (atomic_read(&rdev->nr_pending)) {
2118                         /* lost the race, try later */
2119                         err = -EBUSY;
2120                         p->rdev = rdev;
2121                 }
2122         }
2123 abort:
2124
2125         print_raid5_conf(conf);
2126         return err;
2127 }
2128
2129 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2130 {
2131         raid5_conf_t *conf = mddev->private;
2132         int found = 0;
2133         int disk;
2134         struct disk_info *p;
2135
2136         if (mddev->degraded > 1)
2137                 /* no point adding a device */
2138                 return 0;
2139
2140         /*
2141          * find the disk ...
2142          */
2143         for (disk=0; disk < mddev->raid_disks; disk++)
2144                 if ((p=conf->disks + disk)->rdev == NULL) {
2145                         clear_bit(In_sync, &rdev->flags);
2146                         rdev->raid_disk = disk;
2147                         found = 1;
2148                         if (rdev->saved_raid_disk != disk)
2149                                 conf->fullsync = 1;
2150                         rcu_assign_pointer(p->rdev, rdev);
2151                         break;
2152                 }
2153         print_raid5_conf(conf);
2154         return found;
2155 }
2156
2157 static int raid5_resize(mddev_t *mddev, sector_t sectors)
2158 {
2159         /* no resync is happening, and there is enough space
2160          * on all devices, so we can resize.
2161          * We need to make sure resync covers any new space.
2162          * If the array is shrinking we should possibly wait until
2163          * any io in the removed space completes, but it hardly seems
2164          * worth it.
2165          */
2166         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2167         mddev->array_size = (sectors * (mddev->raid_disks-1))>>1;
2168         set_capacity(mddev->gendisk, mddev->array_size << 1);
2169         mddev->changed = 1;
2170         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2171                 mddev->recovery_cp = mddev->size << 1;
2172                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2173         }
2174         mddev->size = sectors /2;
2175         mddev->resync_max_sectors = sectors;
2176         return 0;
2177 }
2178
2179 static void raid5_quiesce(mddev_t *mddev, int state)
2180 {
2181         raid5_conf_t *conf = mddev_to_conf(mddev);
2182
2183         switch(state) {
2184         case 1: /* stop all writes */
2185                 spin_lock_irq(&conf->device_lock);
2186                 conf->quiesce = 1;
2187                 wait_event_lock_irq(conf->wait_for_stripe,
2188                                     atomic_read(&conf->active_stripes) == 0,
2189                                     conf->device_lock, /* nothing */);
2190                 spin_unlock_irq(&conf->device_lock);
2191                 break;
2192
2193         case 0: /* re-enable writes */
2194                 spin_lock_irq(&conf->device_lock);
2195                 conf->quiesce = 0;
2196                 wake_up(&conf->wait_for_stripe);
2197                 spin_unlock_irq(&conf->device_lock);
2198                 break;
2199         }
2200 }
2201
2202 static mdk_personality_t raid5_personality=
2203 {
2204         .name           = "raid5",
2205         .owner          = THIS_MODULE,
2206         .make_request   = make_request,
2207         .run            = run,
2208         .stop           = stop,
2209         .status         = status,
2210         .error_handler  = error,
2211         .hot_add_disk   = raid5_add_disk,
2212         .hot_remove_disk= raid5_remove_disk,
2213         .spare_active   = raid5_spare_active,
2214         .sync_request   = sync_request,
2215         .resize         = raid5_resize,
2216         .quiesce        = raid5_quiesce,
2217 };
2218
2219 static int __init raid5_init (void)
2220 {
2221         return register_md_personality (RAID5, &raid5_personality);
2222 }
2223
2224 static void raid5_exit (void)
2225 {
2226         unregister_md_personality (RAID5);
2227 }
2228
2229 module_init(raid5_init);
2230 module_exit(raid5_exit);
2231 MODULE_LICENSE("GPL");
2232 MODULE_ALIAS("md-personality-4"); /* RAID5 */