git.oblomov.eu Git - linux-2.6/blob - drivers/md/raid5.c

   1 /*
   2  * raid5.c : Multiple Devices driver for Linux
   3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   4  *         Copyright (C) 1999, 2000 Ingo Molnar
   5  *
   6  * RAID-5 management functions.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2, or (at your option)
  11  * any later version.
  12  *
  13  * You should have received a copy of the GNU General Public License
  14  * (for example /usr/src/linux/COPYING); if not, write to the Free
  15  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  16  */
  17
  18
  19 #include <linux/config.h>
  20 #include <linux/module.h>
  21 #include <linux/slab.h>
  22 #include <linux/raid/raid5.h>
  23 #include <linux/highmem.h>
  24 #include <linux/bitops.h>
  25 #include <asm/atomic.h>
  26
  27 #include <linux/raid/bitmap.h>
  28
  29 /*
  30  * Stripe cache
  31  */
  32
  33 #define NR_STRIPES              256
  34 #define STRIPE_SIZE             PAGE_SIZE
  35 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  36 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  37 #define IO_THRESHOLD            1
  38 #define HASH_PAGES              1
  39 #define HASH_PAGES_ORDER        0
  40 #define NR_HASH                 (HASH_PAGES * PAGE_SIZE / sizeof(struct stripe_head *))
  41 #define HASH_MASK               (NR_HASH - 1)
  42
  43 #define stripe_hash(conf, sect) ((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK])
  44
  45 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  46  * order without overlap.  There may be several bio's per stripe+device, and
  47  * a bio could span several devices.
  48  * When walking this list for a particular stripe+device, we must never proceed
  49  * beyond a bio that extends past this device, as the next bio might no longer
  50  * be valid.
  51  * This macro is used to determine the 'next' bio in the list, given the sector
  52  * of the current stripe+device
  53  */
  54 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
  55 /*
  56  * The following can be used to debug the driver
  57  */
  58 #define RAID5_DEBUG     0
  59 #define RAID5_PARANOIA  1
  60 #if RAID5_PARANOIA && defined(CONFIG_SMP)
  61 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
  62 #else
  63 # define CHECK_DEVLOCK()
  64 #endif
  65
  66 #define PRINTK(x...) ((void)(RAID5_DEBUG && printk(x)))
  67 #if RAID5_DEBUG
  68 #define inline
  69 #define __inline__
  70 #endif
  71
  72 static void print_raid5_conf (raid5_conf_t *conf);
  73
  74 static inline void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
  75 {
  76         if (atomic_dec_and_test(&sh->count)) {
  77                 if (!list_empty(&sh->lru))
  78                         BUG();
  79                 if (atomic_read(&conf->active_stripes)==0)
  80                         BUG();
  81                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
  82                         if (test_bit(STRIPE_DELAYED, &sh->state))
  83                                 list_add_tail(&sh->lru, &conf->delayed_list);
  84                         else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
  85                                  conf->seq_write == sh->bm_seq)
  86                                 list_add_tail(&sh->lru, &conf->bitmap_list);
  87                         else {
  88                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
  89                                 list_add_tail(&sh->lru, &conf->handle_list);
  90                         }
  91                         md_wakeup_thread(conf->mddev->thread);
  92                 } else {
  93                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
  94                                 atomic_dec(&conf->preread_active_stripes);
  95                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
  96                                         md_wakeup_thread(conf->mddev->thread);
  97                         }
  98                         list_add_tail(&sh->lru, &conf->inactive_list);
  99                         atomic_dec(&conf->active_stripes);
 100                         if (!conf->inactive_blocked ||
 101                             atomic_read(&conf->active_stripes) < (NR_STRIPES*3/4))
 102                                 wake_up(&conf->wait_for_stripe);
 103                 }
 104         }
 105 }
 106 static void release_stripe(struct stripe_head *sh)
 107 {
 108         raid5_conf_t *conf = sh->raid_conf;
 109         unsigned long flags;
 110
 111         spin_lock_irqsave(&conf->device_lock, flags);
 112         __release_stripe(conf, sh);
 113         spin_unlock_irqrestore(&conf->device_lock, flags);
 114 }
 115
 116 static void remove_hash(struct stripe_head *sh)
 117 {
 118         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 119
 120         if (sh->hash_pprev) {
 121                 if (sh->hash_next)
 122                         sh->hash_next->hash_pprev = sh->hash_pprev;
 123                 *sh->hash_pprev = sh->hash_next;
 124                 sh->hash_pprev = NULL;
 125         }
 126 }
 127
 128 static __inline__ void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
 129 {
 130         struct stripe_head **shp = &stripe_hash(conf, sh->sector);
 131
 132         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 133
 134         CHECK_DEVLOCK();
 135         if ((sh->hash_next = *shp) != NULL)
 136                 (*shp)->hash_pprev = &sh->hash_next;
 137         *shp = sh;
 138         sh->hash_pprev = shp;
 139 }
 140
 141
 142 /* find an idle stripe, make sure it is unhashed, and return it. */
 143 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
 144 {
 145         struct stripe_head *sh = NULL;
 146         struct list_head *first;
 147
 148         CHECK_DEVLOCK();
 149         if (list_empty(&conf->inactive_list))
 150                 goto out;
 151         first = conf->inactive_list.next;
 152         sh = list_entry(first, struct stripe_head, lru);
 153         list_del_init(first);
 154         remove_hash(sh);
 155         atomic_inc(&conf->active_stripes);
 156 out:
 157         return sh;
 158 }
 159
 160 static void shrink_buffers(struct stripe_head *sh, int num)
 161 {
 162         struct page *p;
 163         int i;
 164
 165         for (i=0; i<num ; i++) {
 166                 p = sh->dev[i].page;
 167                 if (!p)
 168                         continue;
 169                 sh->dev[i].page = NULL;
 170                 page_cache_release(p);
 171         }
 172 }
 173
 174 static int grow_buffers(struct stripe_head *sh, int num)
 175 {
 176         int i;
 177
 178         for (i=0; i<num; i++) {
 179                 struct page *page;
 180
 181                 if (!(page = alloc_page(GFP_KERNEL))) {
 182                         return 1;
 183                 }
 184                 sh->dev[i].page = page;
 185         }
 186         return 0;
 187 }
 188
 189 static void raid5_build_block (struct stripe_head *sh, int i);
 190
 191 static inline void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx)
 192 {
 193         raid5_conf_t *conf = sh->raid_conf;
 194         int disks = conf->raid_disks, i;
 195
 196         if (atomic_read(&sh->count) != 0)
 197                 BUG();
 198         if (test_bit(STRIPE_HANDLE, &sh->state))
 199                 BUG();
 200
 201         CHECK_DEVLOCK();
 202         PRINTK("init_stripe called, stripe %llu\n",
 203                 (unsigned long long)sh->sector);
 204
 205         remove_hash(sh);
 206
 207         sh->sector = sector;
 208         sh->pd_idx = pd_idx;
 209         sh->state = 0;
 210
 211         for (i=disks; i--; ) {
 212                 struct r5dev *dev = &sh->dev[i];
 213
 214                 if (dev->toread || dev->towrite || dev->written ||
 215                     test_bit(R5_LOCKED, &dev->flags)) {
 216                         printk("sector=%llx i=%d %p %p %p %d\n",
 217                                (unsigned long long)sh->sector, i, dev->toread,
 218                                dev->towrite, dev->written,
 219                                test_bit(R5_LOCKED, &dev->flags));
 220                         BUG();
 221                 }
 222                 dev->flags = 0;
 223                 raid5_build_block(sh, i);
 224         }
 225         insert_hash(conf, sh);
 226 }
 227
 228 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector)
 229 {
 230         struct stripe_head *sh;
 231
 232         CHECK_DEVLOCK();
 233         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
 234         for (sh = stripe_hash(conf, sector); sh; sh = sh->hash_next)
 235                 if (sh->sector == sector)
 236                         return sh;
 237         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
 238         return NULL;
 239 }
 240
 241 static void unplug_slaves(mddev_t *mddev);
 242 static void raid5_unplug_device(request_queue_t *q);
 243
 244 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector,
 245                                              int pd_idx, int noblock)
 246 {
 247         struct stripe_head *sh;
 248
 249         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
 250
 251         spin_lock_irq(&conf->device_lock);
 252
 253         do {
 254                 wait_event_lock_irq(conf->wait_for_stripe,
 255                                     conf->quiesce == 0,
 256                                     conf->device_lock, /* nothing */);
 257                 sh = __find_stripe(conf, sector);
 258                 if (!sh) {
 259                         if (!conf->inactive_blocked)
 260                                 sh = get_free_stripe(conf);
 261                         if (noblock && sh == NULL)
 262                                 break;
 263                         if (!sh) {
 264                                 conf->inactive_blocked = 1;
 265                                 wait_event_lock_irq(conf->wait_for_stripe,
 266                                                     !list_empty(&conf->inactive_list) &&
 267                                                     (atomic_read(&conf->active_stripes) < (NR_STRIPES *3/4)
 268                                                      || !conf->inactive_blocked),
 269                                                     conf->device_lock,
 270                                                     unplug_slaves(conf->mddev);
 271                                         );
 272                                 conf->inactive_blocked = 0;
 273                         } else
 274                                 init_stripe(sh, sector, pd_idx);
 275                 } else {
 276                         if (atomic_read(&sh->count)) {
 277                                 if (!list_empty(&sh->lru))
 278                                         BUG();
 279                         } else {
 280                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 281                                         atomic_inc(&conf->active_stripes);
 282                                 if (list_empty(&sh->lru))
 283                                         BUG();
 284                                 list_del_init(&sh->lru);
 285                         }
 286                 }
 287         } while (sh == NULL);
 288
 289         if (sh)
 290                 atomic_inc(&sh->count);
 291
 292         spin_unlock_irq(&conf->device_lock);
 293         return sh;
 294 }
 295
 296 static int grow_one_stripe(raid5_conf_t *conf)
 297 {
 298         struct stripe_head *sh;
 299         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
 300         if (!sh)
 301                 return 0;
 302         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
 303         sh->raid_conf = conf;
 304         spin_lock_init(&sh->lock);
 305
 306         if (grow_buffers(sh, conf->raid_disks)) {
 307                 shrink_buffers(sh, conf->raid_disks);
 308                 kmem_cache_free(conf->slab_cache, sh);
 309                 return 0;
 310         }
 311         /* we just created an active stripe so... */
 312         atomic_set(&sh->count, 1);
 313         atomic_inc(&conf->active_stripes);
 314         INIT_LIST_HEAD(&sh->lru);
 315         release_stripe(sh);
 316         return 1;
 317 }
 318
 319 static int grow_stripes(raid5_conf_t *conf, int num)
 320 {
 321         kmem_cache_t *sc;
 322         int devs = conf->raid_disks;
 323
 324         sprintf(conf->cache_name, "raid5/%s", mdname(conf->mddev));
 325
 326         sc = kmem_cache_create(conf->cache_name,
 327                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
 328                                0, 0, NULL, NULL);
 329         if (!sc)
 330                 return 1;
 331         conf->slab_cache = sc;
 332         while (num--) {
 333                 if (!grow_one_stripe(conf))
 334                         return 1;
 335         }
 336         return 0;
 337 }
 338
 339 static int drop_one_stripe(raid5_conf_t *conf)
 340 {
 341         struct stripe_head *sh;
 342
 343         spin_lock_irq(&conf->device_lock);
 344         sh = get_free_stripe(conf);
 345         spin_unlock_irq(&conf->device_lock);
 346         if (!sh)
 347                 return 0;
 348         if (atomic_read(&sh->count))
 349                 BUG();
 350         shrink_buffers(sh, conf->raid_disks);
 351         kmem_cache_free(conf->slab_cache, sh);
 352         atomic_dec(&conf->active_stripes);
 353         return 1;
 354 }
 355
 356 static void shrink_stripes(raid5_conf_t *conf)
 357 {
 358         while (drop_one_stripe(conf))
 359                 ;
 360
 361         kmem_cache_destroy(conf->slab_cache);
 362         conf->slab_cache = NULL;
 363 }
 364
 365 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
 366                                    int error)
 367 {
 368         struct stripe_head *sh = bi->bi_private;
 369         raid5_conf_t *conf = sh->raid_conf;
 370         int disks = conf->raid_disks, i;
 371         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 372
 373         if (bi->bi_size)
 374                 return 1;
 375
 376         for (i=0 ; i<disks; i++)
 377                 if (bi == &sh->dev[i].req)
 378                         break;
 379
 380         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
 381                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 382                 uptodate);
 383         if (i == disks) {
 384                 BUG();
 385                 return 0;
 386         }
 387
 388         if (uptodate) {
 389 #if 0
 390                 struct bio *bio;
 391                 unsigned long flags;
 392                 spin_lock_irqsave(&conf->device_lock, flags);
 393                 /* we can return a buffer if we bypassed the cache or
 394                  * if the top buffer is not in highmem.  If there are
 395                  * multiple buffers, leave the extra work to
 396                  * handle_stripe
 397                  */
 398                 buffer = sh->bh_read[i];
 399                 if (buffer &&
 400                     (!PageHighMem(buffer->b_page)
 401                      || buffer->b_page == bh->b_page )
 402                         ) {
 403                         sh->bh_read[i] = buffer->b_reqnext;
 404                         buffer->b_reqnext = NULL;
 405                 } else
 406                         buffer = NULL;
 407                 spin_unlock_irqrestore(&conf->device_lock, flags);
 408                 if (sh->bh_page[i]==bh->b_page)
 409                         set_buffer_uptodate(bh);
 410                 if (buffer) {
 411                         if (buffer->b_page != bh->b_page)
 412                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
 413                         buffer->b_end_io(buffer, 1);
 414                 }
 415 #else
 416                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
 417 #endif
 418                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
 419                         printk("R5: read error corrected!!\n");
 420                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 421                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 422                 }
 423                 if (atomic_read(&conf->disks[i].rdev->read_errors))
 424                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
 425         } else {
 426                 int retry = 0;
 427                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
 428                 atomic_inc(&conf->disks[i].rdev->read_errors);
 429                 if (conf->mddev->degraded)
 430                         printk("R5: read error not correctable.\n");
 431                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
 432                         /* Oh, no!!! */
 433                         printk("R5: read error NOT corrected!!\n");
 434                 else if (atomic_read(&conf->disks[i].rdev->read_errors)
 435                          > conf->max_nr_stripes)
 436                         printk("raid5: Too many read errors, failing device.\n");
 437                 else
 438                         retry = 1;
 439                 if (retry)
 440                         set_bit(R5_ReadError, &sh->dev[i].flags);
 441                 else {
 442                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 443                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 444                         md_error(conf->mddev, conf->disks[i].rdev);
 445                 }
 446         }
 447         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 448 #if 0
 449         /* must restore b_page before unlocking buffer... */
 450         if (sh->bh_page[i] != bh->b_page) {
 451                 bh->b_page = sh->bh_page[i];
 452                 bh->b_data = page_address(bh->b_page);
 453                 clear_buffer_uptodate(bh);
 454         }
 455 #endif
 456         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 457         set_bit(STRIPE_HANDLE, &sh->state);
 458         release_stripe(sh);
 459         return 0;
 460 }
 461
 462 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
 463                                     int error)
 464 {
 465         struct stripe_head *sh = bi->bi_private;
 466         raid5_conf_t *conf = sh->raid_conf;
 467         int disks = conf->raid_disks, i;
 468         unsigned long flags;
 469         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 470
 471         if (bi->bi_size)
 472                 return 1;
 473
 474         for (i=0 ; i<disks; i++)
 475                 if (bi == &sh->dev[i].req)
 476                         break;
 477
 478         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
 479                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 480                 uptodate);
 481         if (i == disks) {
 482                 BUG();
 483                 return 0;
 484         }
 485
 486         spin_lock_irqsave(&conf->device_lock, flags);
 487         if (!uptodate)
 488                 md_error(conf->mddev, conf->disks[i].rdev);
 489
 490         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 491
 492         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 493         set_bit(STRIPE_HANDLE, &sh->state);
 494         __release_stripe(conf, sh);
 495         spin_unlock_irqrestore(&conf->device_lock, flags);
 496         return 0;
 497 }
 498
 499
 500 static sector_t compute_blocknr(struct stripe_head *sh, int i);
 501
 502 static void raid5_build_block (struct stripe_head *sh, int i)
 503 {
 504         struct r5dev *dev = &sh->dev[i];
 505
 506         bio_init(&dev->req);
 507         dev->req.bi_io_vec = &dev->vec;
 508         dev->req.bi_vcnt++;
 509         dev->req.bi_max_vecs++;
 510         dev->vec.bv_page = dev->page;
 511         dev->vec.bv_len = STRIPE_SIZE;
 512         dev->vec.bv_offset = 0;
 513
 514         dev->req.bi_sector = sh->sector;
 515         dev->req.bi_private = sh;
 516
 517         dev->flags = 0;
 518         if (i != sh->pd_idx)
 519                 dev->sector = compute_blocknr(sh, i);
 520 }
 521
 522 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 523 {
 524         char b[BDEVNAME_SIZE];
 525         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
 526         PRINTK("raid5: error called\n");
 527
 528         if (!test_bit(Faulty, &rdev->flags)) {
 529                 mddev->sb_dirty = 1;
 530                 if (test_bit(In_sync, &rdev->flags)) {
 531                         conf->working_disks--;
 532                         mddev->degraded++;
 533                         conf->failed_disks++;
 534                         clear_bit(In_sync, &rdev->flags);
 535                         /*
 536                          * if recovery was running, make sure it aborts.
 537                          */
 538                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 539                 }
 540                 set_bit(Faulty, &rdev->flags);
 541                 printk (KERN_ALERT
 542                         "raid5: Disk failure on %s, disabling device."
 543                         " Operation continuing on %d devices\n",
 544                         bdevname(rdev->bdev,b), conf->working_disks);
 545         }
 546 }
 547
 548 /*
 549  * Input: a 'big' sector number,
 550  * Output: index of the data and parity disk, and the sector # in them.
 551  */
 552 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
 553                         unsigned int data_disks, unsigned int * dd_idx,
 554                         unsigned int * pd_idx, raid5_conf_t *conf)
 555 {
 556         long stripe;
 557         unsigned long chunk_number;
 558         unsigned int chunk_offset;
 559         sector_t new_sector;
 560         int sectors_per_chunk = conf->chunk_size >> 9;
 561
 562         /* First compute the information on this sector */
 563
 564         /*
 565          * Compute the chunk number and the sector offset inside the chunk
 566          */
 567         chunk_offset = sector_div(r_sector, sectors_per_chunk);
 568         chunk_number = r_sector;
 569         BUG_ON(r_sector != chunk_number);
 570
 571         /*
 572          * Compute the stripe number
 573          */
 574         stripe = chunk_number / data_disks;
 575
 576         /*
 577          * Compute the data disk and parity disk indexes inside the stripe
 578          */
 579         *dd_idx = chunk_number % data_disks;
 580
 581         /*
 582          * Select the parity disk based on the user selected algorithm.
 583          */
 584         if (conf->level == 4)
 585                 *pd_idx = data_disks;
 586         else switch (conf->algorithm) {
 587                 case ALGORITHM_LEFT_ASYMMETRIC:
 588                         *pd_idx = data_disks - stripe % raid_disks;
 589                         if (*dd_idx >= *pd_idx)
 590                                 (*dd_idx)++;
 591                         break;
 592                 case ALGORITHM_RIGHT_ASYMMETRIC:
 593                         *pd_idx = stripe % raid_disks;
 594                         if (*dd_idx >= *pd_idx)
 595                                 (*dd_idx)++;
 596                         break;
 597                 case ALGORITHM_LEFT_SYMMETRIC:
 598                         *pd_idx = data_disks - stripe % raid_disks;
 599                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 600                         break;
 601                 case ALGORITHM_RIGHT_SYMMETRIC:
 602                         *pd_idx = stripe % raid_disks;
 603                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 604                         break;
 605                 default:
 606                         printk("raid5: unsupported algorithm %d\n",
 607                                 conf->algorithm);
 608         }
 609
 610         /*
 611          * Finally, compute the new sector number
 612          */
 613         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
 614         return new_sector;
 615 }
 616
 617
 618 static sector_t compute_blocknr(struct stripe_head *sh, int i)
 619 {
 620         raid5_conf_t *conf = sh->raid_conf;
 621         int raid_disks = conf->raid_disks, data_disks = raid_disks - 1;
 622         sector_t new_sector = sh->sector, check;
 623         int sectors_per_chunk = conf->chunk_size >> 9;
 624         sector_t stripe;
 625         int chunk_offset;
 626         int chunk_number, dummy1, dummy2, dd_idx = i;
 627         sector_t r_sector;
 628
 629         chunk_offset = sector_div(new_sector, sectors_per_chunk);
 630         stripe = new_sector;
 631         BUG_ON(new_sector != stripe);
 632
 633
 634         switch (conf->algorithm) {
 635                 case ALGORITHM_LEFT_ASYMMETRIC:
 636                 case ALGORITHM_RIGHT_ASYMMETRIC:
 637                         if (i > sh->pd_idx)
 638                                 i--;
 639                         break;
 640                 case ALGORITHM_LEFT_SYMMETRIC:
 641                 case ALGORITHM_RIGHT_SYMMETRIC:
 642                         if (i < sh->pd_idx)
 643                                 i += raid_disks;
 644                         i -= (sh->pd_idx + 1);
 645                         break;
 646                 default:
 647                         printk("raid5: unsupported algorithm %d\n",
 648                                 conf->algorithm);
 649         }
 650
 651         chunk_number = stripe * data_disks + i;
 652         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
 653
 654         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
 655         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
 656                 printk("compute_blocknr: map not correct\n");
 657                 return 0;
 658         }
 659         return r_sector;
 660 }
 661
 662
 663
 664 /*
 665  * Copy data between a page in the stripe cache, and a bio.
 666  * There are no alignment or size guarantees between the page or the
 667  * bio except that there is some overlap.
 668  * All iovecs in the bio must be considered.
 669  */
 670 static void copy_data(int frombio, struct bio *bio,
 671                      struct page *page,
 672                      sector_t sector)
 673 {
 674         char *pa = page_address(page);
 675         struct bio_vec *bvl;
 676         int i;
 677         int page_offset;
 678
 679         if (bio->bi_sector >= sector)
 680                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 681         else
 682                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 683         bio_for_each_segment(bvl, bio, i) {
 684                 int len = bio_iovec_idx(bio,i)->bv_len;
 685                 int clen;
 686                 int b_offset = 0;
 687
 688                 if (page_offset < 0) {
 689                         b_offset = -page_offset;
 690                         page_offset += b_offset;
 691                         len -= b_offset;
 692                 }
 693
 694                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 695                         clen = STRIPE_SIZE - page_offset;
 696                 else clen = len;
 697
 698                 if (clen > 0) {
 699                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
 700                         if (frombio)
 701                                 memcpy(pa+page_offset, ba+b_offset, clen);
 702                         else
 703                                 memcpy(ba+b_offset, pa+page_offset, clen);
 704                         __bio_kunmap_atomic(ba, KM_USER0);
 705                 }
 706                 if (clen < len) /* hit end of page */
 707                         break;
 708                 page_offset +=  len;
 709         }
 710 }
 711
 712 #define check_xor()     do {                                            \
 713                            if (count == MAX_XOR_BLOCKS) {               \
 714                                 xor_block(count, STRIPE_SIZE, ptr);     \
 715                                 count = 1;                              \
 716                            }                                            \
 717                         } while(0)
 718
 719
 720 static void compute_block(struct stripe_head *sh, int dd_idx)
 721 {
 722         raid5_conf_t *conf = sh->raid_conf;
 723         int i, count, disks = conf->raid_disks;
 724         void *ptr[MAX_XOR_BLOCKS], *p;
 725
 726         PRINTK("compute_block, stripe %llu, idx %d\n",
 727                 (unsigned long long)sh->sector, dd_idx);
 728
 729         ptr[0] = page_address(sh->dev[dd_idx].page);
 730         memset(ptr[0], 0, STRIPE_SIZE);
 731         count = 1;
 732         for (i = disks ; i--; ) {
 733                 if (i == dd_idx)
 734                         continue;
 735                 p = page_address(sh->dev[i].page);
 736                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
 737                         ptr[count++] = p;
 738                 else
 739                         printk("compute_block() %d, stripe %llu, %d"
 740                                 " not present\n", dd_idx,
 741                                 (unsigned long long)sh->sector, i);
 742
 743                 check_xor();
 744         }
 745         if (count != 1)
 746                 xor_block(count, STRIPE_SIZE, ptr);
 747         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
 748 }
 749
 750 static void compute_parity(struct stripe_head *sh, int method)
 751 {
 752         raid5_conf_t *conf = sh->raid_conf;
 753         int i, pd_idx = sh->pd_idx, disks = conf->raid_disks, count;
 754         void *ptr[MAX_XOR_BLOCKS];
 755         struct bio *chosen;
 756
 757         PRINTK("compute_parity, stripe %llu, method %d\n",
 758                 (unsigned long long)sh->sector, method);
 759
 760         count = 1;
 761         ptr[0] = page_address(sh->dev[pd_idx].page);
 762         switch(method) {
 763         case READ_MODIFY_WRITE:
 764                 if (!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags))
 765                         BUG();
 766                 for (i=disks ; i-- ;) {
 767                         if (i==pd_idx)
 768                                 continue;
 769                         if (sh->dev[i].towrite &&
 770                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
 771                                 ptr[count++] = page_address(sh->dev[i].page);
 772                                 chosen = sh->dev[i].towrite;
 773                                 sh->dev[i].towrite = NULL;
 774
 775                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 776                                         wake_up(&conf->wait_for_overlap);
 777
 778                                 if (sh->dev[i].written) BUG();
 779                                 sh->dev[i].written = chosen;
 780                                 check_xor();
 781                         }
 782                 }
 783                 break;
 784         case RECONSTRUCT_WRITE:
 785                 memset(ptr[0], 0, STRIPE_SIZE);
 786                 for (i= disks; i-- ;)
 787                         if (i!=pd_idx && sh->dev[i].towrite) {
 788                                 chosen = sh->dev[i].towrite;
 789                                 sh->dev[i].towrite = NULL;
 790
 791                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 792                                         wake_up(&conf->wait_for_overlap);
 793
 794                                 if (sh->dev[i].written) BUG();
 795                                 sh->dev[i].written = chosen;
 796                         }
 797                 break;
 798         case CHECK_PARITY:
 799                 break;
 800         }
 801         if (count>1) {
 802                 xor_block(count, STRIPE_SIZE, ptr);
 803                 count = 1;
 804         }
 805
 806         for (i = disks; i--;)
 807                 if (sh->dev[i].written) {
 808                         sector_t sector = sh->dev[i].sector;
 809                         struct bio *wbi = sh->dev[i].written;
 810                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
 811                                 copy_data(1, wbi, sh->dev[i].page, sector);
 812                                 wbi = r5_next_bio(wbi, sector);
 813                         }
 814
 815                         set_bit(R5_LOCKED, &sh->dev[i].flags);
 816                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
 817                 }
 818
 819         switch(method) {
 820         case RECONSTRUCT_WRITE:
 821         case CHECK_PARITY:
 822                 for (i=disks; i--;)
 823                         if (i != pd_idx) {
 824                                 ptr[count++] = page_address(sh->dev[i].page);
 825                                 check_xor();
 826                         }
 827                 break;
 828         case READ_MODIFY_WRITE:
 829                 for (i = disks; i--;)
 830                         if (sh->dev[i].written) {
 831                                 ptr[count++] = page_address(sh->dev[i].page);
 832                                 check_xor();
 833                         }
 834         }
 835         if (count != 1)
 836                 xor_block(count, STRIPE_SIZE, ptr);
 837
 838         if (method != CHECK_PARITY) {
 839                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 840                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
 841         } else
 842                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 843 }
 844
 845 /*
 846  * Each stripe/dev can have one or more bion attached.
 847  * toread/towrite point to the first in a chain.
 848  * The bi_next chain must be in order.
 849  */
 850 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
 851 {
 852         struct bio **bip;
 853         raid5_conf_t *conf = sh->raid_conf;
 854         int firstwrite=0;
 855
 856         PRINTK("adding bh b#%llu to stripe s#%llu\n",
 857                 (unsigned long long)bi->bi_sector,
 858                 (unsigned long long)sh->sector);
 859
 860
 861         spin_lock(&sh->lock);
 862         spin_lock_irq(&conf->device_lock);
 863         if (forwrite) {
 864                 bip = &sh->dev[dd_idx].towrite;
 865                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
 866                         firstwrite = 1;
 867         } else
 868                 bip = &sh->dev[dd_idx].toread;
 869         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
 870                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
 871                         goto overlap;
 872                 bip = & (*bip)->bi_next;
 873         }
 874         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
 875                 goto overlap;
 876
 877         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
 878                 BUG();
 879         if (*bip)
 880                 bi->bi_next = *bip;
 881         *bip = bi;
 882         bi->bi_phys_segments ++;
 883         spin_unlock_irq(&conf->device_lock);
 884         spin_unlock(&sh->lock);
 885
 886         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
 887                 (unsigned long long)bi->bi_sector,
 888                 (unsigned long long)sh->sector, dd_idx);
 889
 890         if (conf->mddev->bitmap && firstwrite) {
 891                 sh->bm_seq = conf->seq_write;
 892                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
 893                                   STRIPE_SECTORS, 0);
 894                 set_bit(STRIPE_BIT_DELAY, &sh->state);
 895         }
 896
 897         if (forwrite) {
 898                 /* check if page is covered */
 899                 sector_t sector = sh->dev[dd_idx].sector;
 900                 for (bi=sh->dev[dd_idx].towrite;
 901                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
 902                              bi && bi->bi_sector <= sector;
 903                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
 904                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
 905                                 sector = bi->bi_sector + (bi->bi_size>>9);
 906                 }
 907                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
 908                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
 909         }
 910         return 1;
 911
 912  overlap:
 913         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
 914         spin_unlock_irq(&conf->device_lock);
 915         spin_unlock(&sh->lock);
 916         return 0;
 917 }
 918
 919
 920 /*
 921  * handle_stripe - do things to a stripe.
 922  *
 923  * We lock the stripe and then examine the state of various bits
 924  * to see what needs to be done.
 925  * Possible results:
 926  *    return some read request which now have data
 927  *    return some write requests which are safely on disc
 928  *    schedule a read on some buffers
 929  *    schedule a write of some buffers
 930  *    return confirmation of parity correctness
 931  *
 932  * Parity calculations are done inside the stripe lock
 933  * buffers are taken off read_list or write_list, and bh_cache buffers
 934  * get BH_Lock set before the stripe lock is released.
 935  *
 936  */
 937
 938 static void handle_stripe(struct stripe_head *sh)
 939 {
 940         raid5_conf_t *conf = sh->raid_conf;
 941         int disks = conf->raid_disks;
 942         struct bio *return_bi= NULL;
 943         struct bio *bi;
 944         int i;
 945         int syncing;
 946         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
 947         int non_overwrite = 0;
 948         int failed_num=0;
 949         struct r5dev *dev;
 950
 951         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
 952                 (unsigned long long)sh->sector, atomic_read(&sh->count),
 953                 sh->pd_idx);
 954
 955         spin_lock(&sh->lock);
 956         clear_bit(STRIPE_HANDLE, &sh->state);
 957         clear_bit(STRIPE_DELAYED, &sh->state);
 958
 959         syncing = test_bit(STRIPE_SYNCING, &sh->state);
 960         /* Now to look around and see what can be done */
 961
 962         for (i=disks; i--; ) {
 963                 mdk_rdev_t *rdev;
 964                 dev = &sh->dev[i];
 965                 clear_bit(R5_Insync, &dev->flags);
 966                 clear_bit(R5_Syncio, &dev->flags);
 967
 968                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
 969                         i, dev->flags, dev->toread, dev->towrite, dev->written);
 970                 /* maybe we can reply to a read */
 971                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
 972                         struct bio *rbi, *rbi2;
 973                         PRINTK("Return read for disc %d\n", i);
 974                         spin_lock_irq(&conf->device_lock);
 975                         rbi = dev->toread;
 976                         dev->toread = NULL;
 977                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
 978                                 wake_up(&conf->wait_for_overlap);
 979                         spin_unlock_irq(&conf->device_lock);
 980                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
 981                                 copy_data(0, rbi, dev->page, dev->sector);
 982                                 rbi2 = r5_next_bio(rbi, dev->sector);
 983                                 spin_lock_irq(&conf->device_lock);
 984                                 if (--rbi->bi_phys_segments == 0) {
 985                                         rbi->bi_next = return_bi;
 986                                         return_bi = rbi;
 987                                 }
 988                                 spin_unlock_irq(&conf->device_lock);
 989                                 rbi = rbi2;
 990                         }
 991                 }
 992
 993                 /* now count some things */
 994                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
 995                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
 996
 997
 998                 if (dev->toread) to_read++;
 999                 if (dev->towrite) {
1000                         to_write++;
1001                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1002                                 non_overwrite++;
1003                 }
1004                 if (dev->written) written++;
1005                 rdev = conf->disks[i].rdev; /* FIXME, should I be looking rdev */
1006                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1007                         /* The ReadError flag wil just be confusing now */
1008                         clear_bit(R5_ReadError, &dev->flags);
1009                         clear_bit(R5_ReWrite, &dev->flags);
1010                 }
1011                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1012                     || test_bit(R5_ReadError, &dev->flags)) {
1013                         failed++;
1014                         failed_num = i;
1015                 } else
1016                         set_bit(R5_Insync, &dev->flags);
1017         }
1018         PRINTK("locked=%d uptodate=%d to_read=%d"
1019                 " to_write=%d failed=%d failed_num=%d\n",
1020                 locked, uptodate, to_read, to_write, failed, failed_num);
1021         /* check if the array has lost two devices and, if so, some requests might
1022          * need to be failed
1023          */
1024         if (failed > 1 && to_read+to_write+written) {
1025                 for (i=disks; i--; ) {
1026                         int bitmap_end = 0;
1027
1028                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1029                                 mdk_rdev_t *rdev = conf->disks[i].rdev;
1030                                 if (rdev && test_bit(In_sync, &rdev->flags))
1031                                         /* multiple read failures in one stripe */
1032                                         md_error(conf->mddev, rdev);
1033                         }
1034
1035                         spin_lock_irq(&conf->device_lock);
1036                         /* fail all writes first */
1037                         bi = sh->dev[i].towrite;
1038                         sh->dev[i].towrite = NULL;
1039                         if (bi) { to_write--; bitmap_end = 1; }
1040
1041                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1042                                 wake_up(&conf->wait_for_overlap);
1043
1044                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1045                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1046                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1047                                 if (--bi->bi_phys_segments == 0) {
1048                                         md_write_end(conf->mddev);
1049                                         bi->bi_next = return_bi;
1050                                         return_bi = bi;
1051                                 }
1052                                 bi = nextbi;
1053                         }
1054                         /* and fail all 'written' */
1055                         bi = sh->dev[i].written;
1056                         sh->dev[i].written = NULL;
1057                         if (bi) bitmap_end = 1;
1058                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1059                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1060                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1061                                 if (--bi->bi_phys_segments == 0) {
1062                                         md_write_end(conf->mddev);
1063                                         bi->bi_next = return_bi;
1064                                         return_bi = bi;
1065                                 }
1066                                 bi = bi2;
1067                         }
1068
1069                         /* fail any reads if this device is non-operational */
1070                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1071                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1072                                 bi = sh->dev[i].toread;
1073                                 sh->dev[i].toread = NULL;
1074                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1075                                         wake_up(&conf->wait_for_overlap);
1076                                 if (bi) to_read--;
1077                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1078                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1079                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1080                                         if (--bi->bi_phys_segments == 0) {
1081                                                 bi->bi_next = return_bi;
1082                                                 return_bi = bi;
1083                                         }
1084                                         bi = nextbi;
1085                                 }
1086                         }
1087                         spin_unlock_irq(&conf->device_lock);
1088                         if (bitmap_end)
1089                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1090                                                 STRIPE_SECTORS, 0, 0);
1091                 }
1092         }
1093         if (failed > 1 && syncing) {
1094                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1095                 clear_bit(STRIPE_SYNCING, &sh->state);
1096                 syncing = 0;
1097         }
1098
1099         /* might be able to return some write requests if the parity block
1100          * is safe, or on a failed drive
1101          */
1102         dev = &sh->dev[sh->pd_idx];
1103         if ( written &&
1104              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1105                 test_bit(R5_UPTODATE, &dev->flags))
1106                || (failed == 1 && failed_num == sh->pd_idx))
1107             ) {
1108             /* any written block on an uptodate or failed drive can be returned.
1109              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
1110              * never LOCKED, so we don't need to test 'failed' directly.
1111              */
1112             for (i=disks; i--; )
1113                 if (sh->dev[i].written) {
1114                     dev = &sh->dev[i];
1115                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1116                          test_bit(R5_UPTODATE, &dev->flags) ) {
1117                         /* We can return any write requests */
1118                             struct bio *wbi, *wbi2;
1119                             int bitmap_end = 0;
1120                             PRINTK("Return write for disc %d\n", i);
1121                             spin_lock_irq(&conf->device_lock);
1122                             wbi = dev->written;
1123                             dev->written = NULL;
1124                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1125                                     wbi2 = r5_next_bio(wbi, dev->sector);
1126                                     if (--wbi->bi_phys_segments == 0) {
1127                                             md_write_end(conf->mddev);
1128                                             wbi->bi_next = return_bi;
1129                                             return_bi = wbi;
1130                                     }
1131                                     wbi = wbi2;
1132                             }
1133                             if (dev->towrite == NULL)
1134                                     bitmap_end = 1;
1135                             spin_unlock_irq(&conf->device_lock);
1136                             if (bitmap_end)
1137                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1138                                                     STRIPE_SECTORS,
1139                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1140                     }
1141                 }
1142         }
1143
1144         /* Now we might consider reading some blocks, either to check/generate
1145          * parity, or to satisfy requests
1146          * or to load a block that is being partially written.
1147          */
1148         if (to_read || non_overwrite || (syncing && (uptodate < disks))) {
1149                 for (i=disks; i--;) {
1150                         dev = &sh->dev[i];
1151                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1152                             (dev->toread ||
1153                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1154                              syncing ||
1155                              (failed && (sh->dev[failed_num].toread ||
1156                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1157                                     )
1158                                 ) {
1159                                 /* we would like to get this block, possibly
1160                                  * by computing it, but we might not be able to
1161                                  */
1162                                 if (uptodate == disks-1) {
1163                                         PRINTK("Computing block %d\n", i);
1164                                         compute_block(sh, i);
1165                                         uptodate++;
1166                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1167                                         set_bit(R5_LOCKED, &dev->flags);
1168                                         set_bit(R5_Wantread, &dev->flags);
1169 #if 0
1170                                         /* if I am just reading this block and we don't have
1171                                            a failed drive, or any pending writes then sidestep the cache */
1172                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1173                                             ! syncing && !failed && !to_write) {
1174                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1175                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1176                                         }
1177 #endif
1178                                         locked++;
1179                                         PRINTK("Reading block %d (sync=%d)\n",
1180                                                 i, syncing);
1181                                         if (syncing)
1182                                                 md_sync_acct(conf->disks[i].rdev->bdev,
1183                                                              STRIPE_SECTORS);
1184                                 }
1185                         }
1186                 }
1187                 set_bit(STRIPE_HANDLE, &sh->state);
1188         }
1189
1190         /* now to consider writing and what else, if anything should be read */
1191         if (to_write) {
1192                 int rmw=0, rcw=0;
1193                 for (i=disks ; i--;) {
1194                         /* would I have to read this buffer for read_modify_write */
1195                         dev = &sh->dev[i];
1196                         if ((dev->towrite || i == sh->pd_idx) &&
1197                             (!test_bit(R5_LOCKED, &dev->flags)
1198 #if 0
1199 || sh->bh_page[i]!=bh->b_page
1200 #endif
1201                                     ) &&
1202                             !test_bit(R5_UPTODATE, &dev->flags)) {
1203                                 if (test_bit(R5_Insync, &dev->flags)
1204 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1205                                         )
1206                                         rmw++;
1207                                 else rmw += 2*disks;  /* cannot read it */
1208                         }
1209                         /* Would I have to read this buffer for reconstruct_write */
1210                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1211                             (!test_bit(R5_LOCKED, &dev->flags)
1212 #if 0
1213 || sh->bh_page[i] != bh->b_page
1214 #endif
1215                                     ) &&
1216                             !test_bit(R5_UPTODATE, &dev->flags)) {
1217                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1218                                 else rcw += 2*disks;
1219                         }
1220                 }
1221                 PRINTK("for sector %llu, rmw=%d rcw=%d\n",
1222                         (unsigned long long)sh->sector, rmw, rcw);
1223                 set_bit(STRIPE_HANDLE, &sh->state);
1224                 if (rmw < rcw && rmw > 0)
1225                         /* prefer read-modify-write, but need to get some data */
1226                         for (i=disks; i--;) {
1227                                 dev = &sh->dev[i];
1228                                 if ((dev->towrite || i == sh->pd_idx) &&
1229                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1230                                     test_bit(R5_Insync, &dev->flags)) {
1231                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1232                                         {
1233                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1234                                                 set_bit(R5_LOCKED, &dev->flags);
1235                                                 set_bit(R5_Wantread, &dev->flags);
1236                                                 locked++;
1237                                         } else {
1238                                                 set_bit(STRIPE_DELAYED, &sh->state);
1239                                                 set_bit(STRIPE_HANDLE, &sh->state);
1240                                         }
1241                                 }
1242                         }
1243                 if (rcw <= rmw && rcw > 0)
1244                         /* want reconstruct write, but need to get some data */
1245                         for (i=disks; i--;) {
1246                                 dev = &sh->dev[i];
1247                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1248                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1249                                     test_bit(R5_Insync, &dev->flags)) {
1250                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1251                                         {
1252                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1253                                                 set_bit(R5_LOCKED, &dev->flags);
1254                                                 set_bit(R5_Wantread, &dev->flags);
1255                                                 locked++;
1256                                         } else {
1257                                                 set_bit(STRIPE_DELAYED, &sh->state);
1258                                                 set_bit(STRIPE_HANDLE, &sh->state);
1259                                         }
1260                                 }
1261                         }
1262                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1263                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1264                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1265                         PRINTK("Computing parity...\n");
1266                         compute_parity(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1267                         /* now every locked buffer is ready to be written */
1268                         for (i=disks; i--;)
1269                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1270                                         PRINTK("Writing block %d\n", i);
1271                                         locked++;
1272                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1273                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1274                                             || (i==sh->pd_idx && failed == 0))
1275                                                 set_bit(STRIPE_INSYNC, &sh->state);
1276                                 }
1277                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1278                                 atomic_dec(&conf->preread_active_stripes);
1279                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1280                                         md_wakeup_thread(conf->mddev->thread);
1281                         }
1282                 }
1283         }
1284
1285         /* maybe we need to check and possibly fix the parity for this stripe
1286          * Any reads will already have been scheduled, so we just see if enough data
1287          * is available
1288          */
1289         if (syncing && locked == 0 &&
1290             !test_bit(STRIPE_INSYNC, &sh->state) && failed <= 1) {
1291                 set_bit(STRIPE_HANDLE, &sh->state);
1292                 if (failed == 0) {
1293                         char *pagea;
1294                         if (uptodate != disks)
1295                                 BUG();
1296                         compute_parity(sh, CHECK_PARITY);
1297                         uptodate--;
1298                         pagea = page_address(sh->dev[sh->pd_idx].page);
1299                         if ((*(u32*)pagea) == 0 &&
1300                             !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1301                                 /* parity is correct (on disc, not in buffer any more) */
1302                                 set_bit(STRIPE_INSYNC, &sh->state);
1303                         } else {
1304                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1305                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1306                                         /* don't try to repair!! */
1307                                         set_bit(STRIPE_INSYNC, &sh->state);
1308                         }
1309                 }
1310                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1311                         if (failed==0)
1312                                 failed_num = sh->pd_idx;
1313                         /* should be able to compute the missing block and write it to spare */
1314                         if (!test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)) {
1315                                 if (uptodate+1 != disks)
1316                                         BUG();
1317                                 compute_block(sh, failed_num);
1318                                 uptodate++;
1319                         }
1320                         if (uptodate != disks)
1321                                 BUG();
1322                         dev = &sh->dev[failed_num];
1323                         set_bit(R5_LOCKED, &dev->flags);
1324                         set_bit(R5_Wantwrite, &dev->flags);
1325                         clear_bit(STRIPE_DEGRADED, &sh->state);
1326                         locked++;
1327                         set_bit(STRIPE_INSYNC, &sh->state);
1328                         set_bit(R5_Syncio, &dev->flags);
1329                 }
1330         }
1331         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1332                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1333                 clear_bit(STRIPE_SYNCING, &sh->state);
1334         }
1335
1336         /* If the failed drive is just a ReadError, then we might need to progress
1337          * the repair/check process
1338          */
1339         if (failed == 1 && ! conf->mddev->ro &&
1340             test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1341             && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1342             && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1343                 ) {
1344                 dev = &sh->dev[failed_num];
1345                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1346                         set_bit(R5_Wantwrite, &dev->flags);
1347                         set_bit(R5_ReWrite, &dev->flags);
1348                         set_bit(R5_LOCKED, &dev->flags);
1349                 } else {
1350                         /* let's read it back */
1351                         set_bit(R5_Wantread, &dev->flags);
1352                         set_bit(R5_LOCKED, &dev->flags);
1353                 }
1354         }
1355
1356         spin_unlock(&sh->lock);
1357
1358         while ((bi=return_bi)) {
1359                 int bytes = bi->bi_size;
1360
1361                 return_bi = bi->bi_next;
1362                 bi->bi_next = NULL;
1363                 bi->bi_size = 0;
1364                 bi->bi_end_io(bi, bytes, 0);
1365         }
1366         for (i=disks; i-- ;) {
1367                 int rw;
1368                 struct bio *bi;
1369                 mdk_rdev_t *rdev;
1370                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1371                         rw = 1;
1372                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1373                         rw = 0;
1374                 else
1375                         continue;
1376
1377                 bi = &sh->dev[i].req;
1378
1379                 bi->bi_rw = rw;
1380                 if (rw)
1381                         bi->bi_end_io = raid5_end_write_request;
1382                 else
1383                         bi->bi_end_io = raid5_end_read_request;
1384
1385                 rcu_read_lock();
1386                 rdev = rcu_dereference(conf->disks[i].rdev);
1387                 if (rdev && test_bit(Faulty, &rdev->flags))
1388                         rdev = NULL;
1389                 if (rdev)
1390                         atomic_inc(&rdev->nr_pending);
1391                 rcu_read_unlock();
1392
1393                 if (rdev) {
1394                         if (test_bit(R5_Syncio, &sh->dev[i].flags))
1395                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1396
1397                         bi->bi_bdev = rdev->bdev;
1398                         PRINTK("for %llu schedule op %ld on disc %d\n",
1399                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1400                         atomic_inc(&sh->count);
1401                         bi->bi_sector = sh->sector + rdev->data_offset;
1402                         bi->bi_flags = 1 << BIO_UPTODATE;
1403                         bi->bi_vcnt = 1;
1404                         bi->bi_max_vecs = 1;
1405                         bi->bi_idx = 0;
1406                         bi->bi_io_vec = &sh->dev[i].vec;
1407                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1408                         bi->bi_io_vec[0].bv_offset = 0;
1409                         bi->bi_size = STRIPE_SIZE;
1410                         bi->bi_next = NULL;
1411                         generic_make_request(bi);
1412                 } else {
1413                         if (rw == 1)
1414                                 set_bit(STRIPE_DEGRADED, &sh->state);
1415                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1416                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1417                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1418                         set_bit(STRIPE_HANDLE, &sh->state);
1419                 }
1420         }
1421 }
1422
1423 static inline void raid5_activate_delayed(raid5_conf_t *conf)
1424 {
1425         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1426                 while (!list_empty(&conf->delayed_list)) {
1427                         struct list_head *l = conf->delayed_list.next;
1428                         struct stripe_head *sh;
1429                         sh = list_entry(l, struct stripe_head, lru);
1430                         list_del_init(l);
1431                         clear_bit(STRIPE_DELAYED, &sh->state);
1432                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1433                                 atomic_inc(&conf->preread_active_stripes);
1434                         list_add_tail(&sh->lru, &conf->handle_list);
1435                 }
1436         }
1437 }
1438
1439 static inline void activate_bit_delay(raid5_conf_t *conf)
1440 {
1441         /* device_lock is held */
1442         struct list_head head;
1443         list_add(&head, &conf->bitmap_list);
1444         list_del_init(&conf->bitmap_list);
1445         while (!list_empty(&head)) {
1446                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1447                 list_del_init(&sh->lru);
1448                 atomic_inc(&sh->count);
1449                 __release_stripe(conf, sh);
1450         }
1451 }
1452
1453 static void unplug_slaves(mddev_t *mddev)
1454 {
1455         raid5_conf_t *conf = mddev_to_conf(mddev);
1456         int i;
1457
1458         rcu_read_lock();
1459         for (i=0; i<mddev->raid_disks; i++) {
1460                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1461                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
1462                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1463
1464                         atomic_inc(&rdev->nr_pending);
1465                         rcu_read_unlock();
1466
1467                         if (r_queue->unplug_fn)
1468                                 r_queue->unplug_fn(r_queue);
1469
1470                         rdev_dec_pending(rdev, mddev);
1471                         rcu_read_lock();
1472                 }
1473         }
1474         rcu_read_unlock();
1475 }
1476
1477 static void raid5_unplug_device(request_queue_t *q)
1478 {
1479         mddev_t *mddev = q->queuedata;
1480         raid5_conf_t *conf = mddev_to_conf(mddev);
1481         unsigned long flags;
1482
1483         spin_lock_irqsave(&conf->device_lock, flags);
1484
1485         if (blk_remove_plug(q)) {
1486                 conf->seq_flush++;
1487                 raid5_activate_delayed(conf);
1488         }
1489         md_wakeup_thread(mddev->thread);
1490
1491         spin_unlock_irqrestore(&conf->device_lock, flags);
1492
1493         unplug_slaves(mddev);
1494 }
1495
1496 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
1497                              sector_t *error_sector)
1498 {
1499         mddev_t *mddev = q->queuedata;
1500         raid5_conf_t *conf = mddev_to_conf(mddev);
1501         int i, ret = 0;
1502
1503         rcu_read_lock();
1504         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1505                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1506                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
1507                         struct block_device *bdev = rdev->bdev;
1508                         request_queue_t *r_queue = bdev_get_queue(bdev);
1509
1510                         if (!r_queue->issue_flush_fn)
1511                                 ret = -EOPNOTSUPP;
1512                         else {
1513                                 atomic_inc(&rdev->nr_pending);
1514                                 rcu_read_unlock();
1515                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1516                                                               error_sector);
1517                                 rdev_dec_pending(rdev, mddev);
1518                                 rcu_read_lock();
1519                         }
1520                 }
1521         }
1522         rcu_read_unlock();
1523         return ret;
1524 }
1525
1526 static inline void raid5_plug_device(raid5_conf_t *conf)
1527 {
1528         spin_lock_irq(&conf->device_lock);
1529         blk_plug_device(conf->mddev->queue);
1530         spin_unlock_irq(&conf->device_lock);
1531 }
1532
1533 static int make_request (request_queue_t *q, struct bio * bi)
1534 {
1535         mddev_t *mddev = q->queuedata;
1536         raid5_conf_t *conf = mddev_to_conf(mddev);
1537         const unsigned int raid_disks = conf->raid_disks;
1538         const unsigned int data_disks = raid_disks - 1;
1539         unsigned int dd_idx, pd_idx;
1540         sector_t new_sector;
1541         sector_t logical_sector, last_sector;
1542         struct stripe_head *sh;
1543         const int rw = bio_data_dir(bi);
1544
1545         if (unlikely(bio_barrier(bi))) {
1546                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1547                 return 0;
1548         }
1549
1550         md_write_start(mddev, bi);
1551
1552         disk_stat_inc(mddev->gendisk, ios[rw]);
1553         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
1554
1555         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1556         last_sector = bi->bi_sector + (bi->bi_size>>9);
1557         bi->bi_next = NULL;
1558         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1559
1560         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1561                 DEFINE_WAIT(w);
1562
1563                 new_sector = raid5_compute_sector(logical_sector,
1564                                                   raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1565
1566                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
1567                         (unsigned long long)new_sector,
1568                         (unsigned long long)logical_sector);
1569
1570         retry:
1571                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1572                 sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
1573                 if (sh) {
1574                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1575                                 /* Add failed due to overlap.  Flush everything
1576                                  * and wait a while
1577                                  */
1578                                 raid5_unplug_device(mddev->queue);
1579                                 release_stripe(sh);
1580                                 schedule();
1581                                 goto retry;
1582                         }
1583                         finish_wait(&conf->wait_for_overlap, &w);
1584                         raid5_plug_device(conf);
1585                         handle_stripe(sh);
1586                         release_stripe(sh);
1587
1588                 } else {
1589                         /* cannot get stripe for read-ahead, just give-up */
1590                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1591                         finish_wait(&conf->wait_for_overlap, &w);
1592                         break;
1593                 }
1594
1595         }
1596         spin_lock_irq(&conf->device_lock);
1597         if (--bi->bi_phys_segments == 0) {
1598                 int bytes = bi->bi_size;
1599
1600                 if ( bio_data_dir(bi) == WRITE )
1601                         md_write_end(mddev);
1602                 bi->bi_size = 0;
1603                 bi->bi_end_io(bi, bytes, 0);
1604         }
1605         spin_unlock_irq(&conf->device_lock);
1606         return 0;
1607 }
1608
1609 /* FIXME go_faster isn't used */
1610 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1611 {
1612         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1613         struct stripe_head *sh;
1614         int sectors_per_chunk = conf->chunk_size >> 9;
1615         sector_t x;
1616         unsigned long stripe;
1617         int chunk_offset;
1618         int dd_idx, pd_idx;
1619         sector_t first_sector;
1620         int raid_disks = conf->raid_disks;
1621         int data_disks = raid_disks-1;
1622         sector_t max_sector = mddev->size << 1;
1623         int sync_blocks;
1624
1625         if (sector_nr >= max_sector) {
1626                 /* just being told to finish up .. nothing much to do */
1627                 unplug_slaves(mddev);
1628
1629                 if (mddev->curr_resync < max_sector) /* aborted */
1630                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1631                                         &sync_blocks, 1);
1632                 else /* compelted sync */
1633                         conf->fullsync = 0;
1634                 bitmap_close_sync(mddev->bitmap);
1635
1636                 return 0;
1637         }
1638         /* if there is 1 or more failed drives and we are trying
1639          * to resync, then assert that we are finished, because there is
1640          * nothing we can do.
1641          */
1642         if (mddev->degraded >= 1 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1643                 sector_t rv = (mddev->size << 1) - sector_nr;
1644                 *skipped = 1;
1645                 return rv;
1646         }
1647         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1648             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1649             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
1650                 /* we can skip this block, and probably more */
1651                 sync_blocks /= STRIPE_SECTORS;
1652                 *skipped = 1;
1653                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
1654         }
1655
1656         x = sector_nr;
1657         chunk_offset = sector_div(x, sectors_per_chunk);
1658         stripe = x;
1659         BUG_ON(x != stripe);
1660
1661         first_sector = raid5_compute_sector((sector_t)stripe*data_disks*sectors_per_chunk
1662                 + chunk_offset, raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1663         sh = get_active_stripe(conf, sector_nr, pd_idx, 1);
1664         if (sh == NULL) {
1665                 sh = get_active_stripe(conf, sector_nr, pd_idx, 0);
1666                 /* make sure we don't swamp the stripe cache if someone else
1667                  * is trying to get access
1668                  */
1669                 schedule_timeout_uninterruptible(1);
1670         }
1671         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 0);
1672         spin_lock(&sh->lock);
1673         set_bit(STRIPE_SYNCING, &sh->state);
1674         clear_bit(STRIPE_INSYNC, &sh->state);
1675         spin_unlock(&sh->lock);
1676
1677         handle_stripe(sh);
1678         release_stripe(sh);
1679
1680         return STRIPE_SECTORS;
1681 }
1682
1683 /*
1684  * This is our raid5 kernel thread.
1685  *
1686  * We scan the hash table for stripes which can be handled now.
1687  * During the scan, completed stripes are saved for us by the interrupt
1688  * handler, so that they will not have to wait for our next wakeup.
1689  */
1690 static void raid5d (mddev_t *mddev)
1691 {
1692         struct stripe_head *sh;
1693         raid5_conf_t *conf = mddev_to_conf(mddev);
1694         int handled;
1695
1696         PRINTK("+++ raid5d active\n");
1697
1698         md_check_recovery(mddev);
1699
1700         handled = 0;
1701         spin_lock_irq(&conf->device_lock);
1702         while (1) {
1703                 struct list_head *first;
1704
1705                 if (conf->seq_flush - conf->seq_write > 0) {
1706                         int seq = conf->seq_flush;
1707                         bitmap_unplug(mddev->bitmap);
1708                         conf->seq_write = seq;
1709                         activate_bit_delay(conf);
1710                 }
1711
1712                 if (list_empty(&conf->handle_list) &&
1713                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
1714                     !blk_queue_plugged(mddev->queue) &&
1715                     !list_empty(&conf->delayed_list))
1716                         raid5_activate_delayed(conf);
1717
1718                 if (list_empty(&conf->handle_list))
1719                         break;
1720
1721                 first = conf->handle_list.next;
1722                 sh = list_entry(first, struct stripe_head, lru);
1723
1724                 list_del_init(first);
1725                 atomic_inc(&sh->count);
1726                 if (atomic_read(&sh->count)!= 1)
1727                         BUG();
1728                 spin_unlock_irq(&conf->device_lock);
1729
1730                 handled++;
1731                 handle_stripe(sh);
1732                 release_stripe(sh);
1733
1734                 spin_lock_irq(&conf->device_lock);
1735         }
1736         PRINTK("%d stripes handled\n", handled);
1737
1738         spin_unlock_irq(&conf->device_lock);
1739
1740         unplug_slaves(mddev);
1741
1742         PRINTK("--- raid5d inactive\n");
1743 }
1744
1745 static ssize_t
1746 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
1747 {
1748         raid5_conf_t *conf = mddev_to_conf(mddev);
1749         if (conf)
1750                 return sprintf(page, "%d\n", conf->max_nr_stripes);
1751         else
1752                 return 0;
1753 }
1754
1755 static ssize_t
1756 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
1757 {
1758         raid5_conf_t *conf = mddev_to_conf(mddev);
1759         char *end;
1760         int new;
1761         if (len >= PAGE_SIZE)
1762                 return -EINVAL;
1763         if (!conf)
1764                 return -ENODEV;
1765
1766         new = simple_strtoul(page, &end, 10);
1767         if (!*page || (*end && *end != '\n') )
1768                 return -EINVAL;
1769         if (new <= 16 || new > 32768)
1770                 return -EINVAL;
1771         while (new < conf->max_nr_stripes) {
1772                 if (drop_one_stripe(conf))
1773                         conf->max_nr_stripes--;
1774                 else
1775                         break;
1776         }
1777         while (new > conf->max_nr_stripes) {
1778                 if (grow_one_stripe(conf))
1779                         conf->max_nr_stripes++;
1780                 else break;
1781         }
1782         return len;
1783 }
1784
1785 static struct md_sysfs_entry
1786 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
1787                                 raid5_show_stripe_cache_size,
1788                                 raid5_store_stripe_cache_size);
1789
1790 static ssize_t
1791 stripe_cache_active_show(mddev_t *mddev, char *page)
1792 {
1793         raid5_conf_t *conf = mddev_to_conf(mddev);
1794         if (conf)
1795                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
1796         else
1797                 return 0;
1798 }
1799
1800 static struct md_sysfs_entry
1801 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
1802
1803 static struct attribute *raid5_attrs[] =  {
1804         &raid5_stripecache_size.attr,
1805         &raid5_stripecache_active.attr,
1806         NULL,
1807 };
1808 static struct attribute_group raid5_attrs_group = {
1809         .name = NULL,
1810         .attrs = raid5_attrs,
1811 };
1812
1813 static int run(mddev_t *mddev)
1814 {
1815         raid5_conf_t *conf;
1816         int raid_disk, memory;
1817         mdk_rdev_t *rdev;
1818         struct disk_info *disk;
1819         struct list_head *tmp;
1820
1821         if (mddev->level != 5 && mddev->level != 4) {
1822                 printk("raid5: %s: raid level not set to 4/5 (%d)\n", mdname(mddev), mddev->level);
1823                 return -EIO;
1824         }
1825
1826         mddev->private = kmalloc (sizeof (raid5_conf_t)
1827                                   + mddev->raid_disks * sizeof(struct disk_info),
1828                                   GFP_KERNEL);
1829         if ((conf = mddev->private) == NULL)
1830                 goto abort;
1831         memset (conf, 0, sizeof (*conf) + mddev->raid_disks * sizeof(struct disk_info) );
1832         conf->mddev = mddev;
1833
1834         if ((conf->stripe_hashtbl = (struct stripe_head **) __get_free_pages(GFP_ATOMIC, HASH_PAGES_ORDER)) == NULL)
1835                 goto abort;
1836         memset(conf->stripe_hashtbl, 0, HASH_PAGES * PAGE_SIZE);
1837
1838         spin_lock_init(&conf->device_lock);
1839         init_waitqueue_head(&conf->wait_for_stripe);
1840         init_waitqueue_head(&conf->wait_for_overlap);
1841         INIT_LIST_HEAD(&conf->handle_list);
1842         INIT_LIST_HEAD(&conf->delayed_list);
1843         INIT_LIST_HEAD(&conf->bitmap_list);
1844         INIT_LIST_HEAD(&conf->inactive_list);
1845         atomic_set(&conf->active_stripes, 0);
1846         atomic_set(&conf->preread_active_stripes, 0);
1847
1848         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
1849
1850         ITERATE_RDEV(mddev,rdev,tmp) {
1851                 raid_disk = rdev->raid_disk;
1852                 if (raid_disk >= mddev->raid_disks
1853                     || raid_disk < 0)
1854                         continue;
1855                 disk = conf->disks + raid_disk;
1856
1857                 disk->rdev = rdev;
1858
1859                 if (test_bit(In_sync, &rdev->flags)) {
1860                         char b[BDEVNAME_SIZE];
1861                         printk(KERN_INFO "raid5: device %s operational as raid"
1862                                 " disk %d\n", bdevname(rdev->bdev,b),
1863                                 raid_disk);
1864                         conf->working_disks++;
1865                 }
1866         }
1867
1868         conf->raid_disks = mddev->raid_disks;
1869         /*
1870          * 0 for a fully functional array, 1 for a degraded array.
1871          */
1872         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
1873         conf->mddev = mddev;
1874         conf->chunk_size = mddev->chunk_size;
1875         conf->level = mddev->level;
1876         conf->algorithm = mddev->layout;
1877         conf->max_nr_stripes = NR_STRIPES;
1878
1879         /* device size must be a multiple of chunk size */
1880         mddev->size &= ~(mddev->chunk_size/1024 -1);
1881         mddev->resync_max_sectors = mddev->size << 1;
1882
1883         if (!conf->chunk_size || conf->chunk_size % 4) {
1884                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
1885                         conf->chunk_size, mdname(mddev));
1886                 goto abort;
1887         }
1888         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
1889                 printk(KERN_ERR
1890                         "raid5: unsupported parity algorithm %d for %s\n",
1891                         conf->algorithm, mdname(mddev));
1892                 goto abort;
1893         }
1894         if (mddev->degraded > 1) {
1895                 printk(KERN_ERR "raid5: not enough operational devices for %s"
1896                         " (%d/%d failed)\n",
1897                         mdname(mddev), conf->failed_disks, conf->raid_disks);
1898                 goto abort;
1899         }
1900
1901         if (mddev->degraded == 1 &&
1902             mddev->recovery_cp != MaxSector) {
1903                 printk(KERN_ERR
1904                         "raid5: cannot start dirty degraded array for %s\n",
1905                         mdname(mddev));
1906                 goto abort;
1907         }
1908
1909         {
1910                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
1911                 if (!mddev->thread) {
1912                         printk(KERN_ERR
1913                                 "raid5: couldn't allocate thread for %s\n",
1914                                 mdname(mddev));
1915                         goto abort;
1916                 }
1917         }
1918 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
1919                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
1920         if (grow_stripes(conf, conf->max_nr_stripes)) {
1921                 printk(KERN_ERR
1922                         "raid5: couldn't allocate %dkB for buffers\n", memory);
1923                 shrink_stripes(conf);
1924                 md_unregister_thread(mddev->thread);
1925                 goto abort;
1926         } else
1927                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
1928                         memory, mdname(mddev));
1929
1930         if (mddev->degraded == 0)
1931                 printk("raid5: raid level %d set %s active with %d out of %d"
1932                         " devices, algorithm %d\n", conf->level, mdname(mddev),
1933                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
1934                         conf->algorithm);
1935         else
1936                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
1937                         " out of %d devices, algorithm %d\n", conf->level,
1938                         mdname(mddev), mddev->raid_disks - mddev->degraded,
1939                         mddev->raid_disks, conf->algorithm);
1940
1941         print_raid5_conf(conf);
1942
1943         /* read-ahead size must cover two whole stripes, which is
1944          * 2 * (n-1) * chunksize where 'n' is the number of raid devices
1945          */
1946         {
1947                 int stripe = (mddev->raid_disks-1) * mddev->chunk_size
1948                         / PAGE_CACHE_SIZE;
1949                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
1950                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
1951         }
1952
1953         /* Ok, everything is just fine now */
1954         sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
1955
1956         if (mddev->bitmap)
1957                 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
1958
1959         mddev->queue->unplug_fn = raid5_unplug_device;
1960         mddev->queue->issue_flush_fn = raid5_issue_flush;
1961
1962         mddev->array_size =  mddev->size * (mddev->raid_disks - 1);
1963         return 0;
1964 abort:
1965         if (conf) {
1966                 print_raid5_conf(conf);
1967                 if (conf->stripe_hashtbl)
1968                         free_pages((unsigned long) conf->stripe_hashtbl,
1969                                                         HASH_PAGES_ORDER);
1970                 kfree(conf);
1971         }
1972         mddev->private = NULL;
1973         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
1974         return -EIO;
1975 }
1976
1977
1978
1979 static int stop(mddev_t *mddev)
1980 {
1981         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1982
1983         md_unregister_thread(mddev->thread);
1984         mddev->thread = NULL;
1985         shrink_stripes(conf);
1986         free_pages((unsigned long) conf->stripe_hashtbl, HASH_PAGES_ORDER);
1987         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1988         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
1989         kfree(conf);
1990         mddev->private = NULL;
1991         return 0;
1992 }
1993
1994 #if RAID5_DEBUG
1995 static void print_sh (struct stripe_head *sh)
1996 {
1997         int i;
1998
1999         printk("sh %llu, pd_idx %d, state %ld.\n",
2000                 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
2001         printk("sh %llu,  count %d.\n",
2002                 (unsigned long long)sh->sector, atomic_read(&sh->count));
2003         printk("sh %llu, ", (unsigned long long)sh->sector);
2004         for (i = 0; i < sh->raid_conf->raid_disks; i++) {
2005                 printk("(cache%d: %p %ld) ",
2006                         i, sh->dev[i].page, sh->dev[i].flags);
2007         }
2008         printk("\n");
2009 }
2010
2011 static void printall (raid5_conf_t *conf)
2012 {
2013         struct stripe_head *sh;
2014         int i;
2015
2016         spin_lock_irq(&conf->device_lock);
2017         for (i = 0; i < NR_HASH; i++) {
2018                 sh = conf->stripe_hashtbl[i];
2019                 for (; sh; sh = sh->hash_next) {
2020                         if (sh->raid_conf != conf)
2021                                 continue;
2022                         print_sh(sh);
2023                 }
2024         }
2025         spin_unlock_irq(&conf->device_lock);
2026 }
2027 #endif
2028
2029 static void status (struct seq_file *seq, mddev_t *mddev)
2030 {
2031         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2032         int i;
2033
2034         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2035         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2036         for (i = 0; i < conf->raid_disks; i++)
2037                 seq_printf (seq, "%s",
2038                                conf->disks[i].rdev &&
2039                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
2040         seq_printf (seq, "]");
2041 #if RAID5_DEBUG
2042 #define D(x) \
2043         seq_printf (seq, "<"#x":%d>", atomic_read(&conf->x))
2044         printall(conf);
2045 #endif
2046 }
2047
2048 static void print_raid5_conf (raid5_conf_t *conf)
2049 {
2050         int i;
2051         struct disk_info *tmp;
2052
2053         printk("RAID5 conf printout:\n");
2054         if (!conf) {
2055                 printk("(conf==NULL)\n");
2056                 return;
2057         }
2058         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2059                  conf->working_disks, conf->failed_disks);
2060
2061         for (i = 0; i < conf->raid_disks; i++) {
2062                 char b[BDEVNAME_SIZE];
2063                 tmp = conf->disks + i;
2064                 if (tmp->rdev)
2065                 printk(" disk %d, o:%d, dev:%s\n",
2066                         i, !test_bit(Faulty, &tmp->rdev->flags),
2067                         bdevname(tmp->rdev->bdev,b));
2068         }
2069 }
2070
2071 static int raid5_spare_active(mddev_t *mddev)
2072 {
2073         int i;
2074         raid5_conf_t *conf = mddev->private;
2075         struct disk_info *tmp;
2076
2077         for (i = 0; i < conf->raid_disks; i++) {
2078                 tmp = conf->disks + i;
2079                 if (tmp->rdev
2080                     && !test_bit(Faulty, &tmp->rdev->flags)
2081                     && !test_bit(In_sync, &tmp->rdev->flags)) {
2082                         mddev->degraded--;
2083                         conf->failed_disks--;
2084                         conf->working_disks++;
2085                         set_bit(In_sync, &tmp->rdev->flags);
2086                 }
2087         }
2088         print_raid5_conf(conf);
2089         return 0;
2090 }
2091
2092 static int raid5_remove_disk(mddev_t *mddev, int number)
2093 {
2094         raid5_conf_t *conf = mddev->private;
2095         int err = 0;
2096         mdk_rdev_t *rdev;
2097         struct disk_info *p = conf->disks + number;
2098
2099         print_raid5_conf(conf);
2100         rdev = p->rdev;
2101         if (rdev) {
2102                 if (test_bit(In_sync, &rdev->flags) ||
2103                     atomic_read(&rdev->nr_pending)) {
2104                         err = -EBUSY;
2105                         goto abort;
2106                 }
2107                 p->rdev = NULL;
2108                 synchronize_rcu();
2109                 if (atomic_read(&rdev->nr_pending)) {
2110                         /* lost the race, try later */
2111                         err = -EBUSY;
2112                         p->rdev = rdev;
2113                 }
2114         }
2115 abort:
2116
2117         print_raid5_conf(conf);
2118         return err;
2119 }
2120
2121 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2122 {
2123         raid5_conf_t *conf = mddev->private;
2124         int found = 0;
2125         int disk;
2126         struct disk_info *p;
2127
2128         if (mddev->degraded > 1)
2129                 /* no point adding a device */
2130                 return 0;
2131
2132         /*
2133          * find the disk ...
2134          */
2135         for (disk=0; disk < mddev->raid_disks; disk++)
2136                 if ((p=conf->disks + disk)->rdev == NULL) {
2137                         clear_bit(In_sync, &rdev->flags);
2138                         rdev->raid_disk = disk;
2139                         found = 1;
2140                         if (rdev->saved_raid_disk != disk)
2141                                 conf->fullsync = 1;
2142                         rcu_assign_pointer(p->rdev, rdev);
2143                         break;
2144                 }
2145         print_raid5_conf(conf);
2146         return found;
2147 }
2148
2149 static int raid5_resize(mddev_t *mddev, sector_t sectors)
2150 {
2151         /* no resync is happening, and there is enough space
2152          * on all devices, so we can resize.
2153          * We need to make sure resync covers any new space.
2154          * If the array is shrinking we should possibly wait until
2155          * any io in the removed space completes, but it hardly seems
2156          * worth it.
2157          */
2158         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2159         mddev->array_size = (sectors * (mddev->raid_disks-1))>>1;
2160         set_capacity(mddev->gendisk, mddev->array_size << 1);
2161         mddev->changed = 1;
2162         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2163                 mddev->recovery_cp = mddev->size << 1;
2164                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2165         }
2166         mddev->size = sectors /2;
2167         mddev->resync_max_sectors = sectors;
2168         return 0;
2169 }
2170
2171 static void raid5_quiesce(mddev_t *mddev, int state)
2172 {
2173         raid5_conf_t *conf = mddev_to_conf(mddev);
2174
2175         switch(state) {
2176         case 1: /* stop all writes */
2177                 spin_lock_irq(&conf->device_lock);
2178                 conf->quiesce = 1;
2179                 wait_event_lock_irq(conf->wait_for_stripe,
2180                                     atomic_read(&conf->active_stripes) == 0,
2181                                     conf->device_lock, /* nothing */);
2182                 spin_unlock_irq(&conf->device_lock);
2183                 break;
2184
2185         case 0: /* re-enable writes */
2186                 spin_lock_irq(&conf->device_lock);
2187                 conf->quiesce = 0;
2188                 wake_up(&conf->wait_for_stripe);
2189                 spin_unlock_irq(&conf->device_lock);
2190                 break;
2191         }
2192         if (mddev->thread) {
2193                 if (mddev->bitmap)
2194                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2195                 else
2196                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2197                 md_wakeup_thread(mddev->thread);
2198         }
2199 }
2200 static mdk_personality_t raid5_personality=
2201 {
2202         .name           = "raid5",
2203         .owner          = THIS_MODULE,
2204         .make_request   = make_request,
2205         .run            = run,
2206         .stop           = stop,
2207         .status         = status,
2208         .error_handler  = error,
2209         .hot_add_disk   = raid5_add_disk,
2210         .hot_remove_disk= raid5_remove_disk,
2211         .spare_active   = raid5_spare_active,
2212         .sync_request   = sync_request,
2213         .resize         = raid5_resize,
2214         .quiesce        = raid5_quiesce,
2215 };
2216
2217 static int __init raid5_init (void)
2218 {
2219         return register_md_personality (RAID5, &raid5_personality);
2220 }
2221
2222 static void raid5_exit (void)
2223 {
2224         unregister_md_personality (RAID5);
2225 }
2226
2227 module_init(raid5_init);
2228 module_exit(raid5_exit);
2229 MODULE_LICENSE("GPL");
2230 MODULE_ALIAS("md-personality-4"); /* RAID5 */