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 <linux/kthread.h>
  26 #include <asm/atomic.h>
  27
  28 #include <linux/raid/bitmap.h>
  29
  30 /*
  31  * Stripe cache
  32  */
  33
  34 #define NR_STRIPES              256
  35 #define STRIPE_SIZE             PAGE_SIZE
  36 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  37 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  38 #define IO_THRESHOLD            1
  39 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  40 #define HASH_MASK               (NR_HASH - 1)
  41
  42 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
  43
  44 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  45  * order without overlap.  There may be several bio's per stripe+device, and
  46  * a bio could span several devices.
  47  * When walking this list for a particular stripe+device, we must never proceed
  48  * beyond a bio that extends past this device, as the next bio might no longer
  49  * be valid.
  50  * This macro is used to determine the 'next' bio in the list, given the sector
  51  * of the current stripe+device
  52  */
  53 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
  54 /*
  55  * The following can be used to debug the driver
  56  */
  57 #define RAID5_DEBUG     0
  58 #define RAID5_PARANOIA  1
  59 #if RAID5_PARANOIA && defined(CONFIG_SMP)
  60 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
  61 #else
  62 # define CHECK_DEVLOCK()
  63 #endif
  64
  65 #define PRINTK(x...) ((void)(RAID5_DEBUG && printk(x)))
  66 #if RAID5_DEBUG
  67 #define inline
  68 #define __inline__
  69 #endif
  70
  71 static void print_raid5_conf (raid5_conf_t *conf);
  72
  73 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
  74 {
  75         if (atomic_dec_and_test(&sh->count)) {
  76                 BUG_ON(!list_empty(&sh->lru));
  77                 BUG_ON(atomic_read(&conf->active_stripes)==0);
  78                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
  79                         if (test_bit(STRIPE_DELAYED, &sh->state))
  80                                 list_add_tail(&sh->lru, &conf->delayed_list);
  81                         else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
  82                                  conf->seq_write == sh->bm_seq)
  83                                 list_add_tail(&sh->lru, &conf->bitmap_list);
  84                         else {
  85                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
  86                                 list_add_tail(&sh->lru, &conf->handle_list);
  87                         }
  88                         md_wakeup_thread(conf->mddev->thread);
  89                 } else {
  90                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
  91                                 atomic_dec(&conf->preread_active_stripes);
  92                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
  93                                         md_wakeup_thread(conf->mddev->thread);
  94                         }
  95                         atomic_dec(&conf->active_stripes);
  96                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
  97                                 list_add_tail(&sh->lru, &conf->inactive_list);
  98                                 wake_up(&conf->wait_for_stripe);
  99                         }
 100                 }
 101         }
 102 }
 103 static void release_stripe(struct stripe_head *sh)
 104 {
 105         raid5_conf_t *conf = sh->raid_conf;
 106         unsigned long flags;
 107
 108         spin_lock_irqsave(&conf->device_lock, flags);
 109         __release_stripe(conf, sh);
 110         spin_unlock_irqrestore(&conf->device_lock, flags);
 111 }
 112
 113 static inline void remove_hash(struct stripe_head *sh)
 114 {
 115         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 116
 117         hlist_del_init(&sh->hash);
 118 }
 119
 120 static void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
 121 {
 122         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 123
 124         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 125
 126         CHECK_DEVLOCK();
 127         hlist_add_head(&sh->hash, hp);
 128 }
 129
 130
 131 /* find an idle stripe, make sure it is unhashed, and return it. */
 132 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
 133 {
 134         struct stripe_head *sh = NULL;
 135         struct list_head *first;
 136
 137         CHECK_DEVLOCK();
 138         if (list_empty(&conf->inactive_list))
 139                 goto out;
 140         first = conf->inactive_list.next;
 141         sh = list_entry(first, struct stripe_head, lru);
 142         list_del_init(first);
 143         remove_hash(sh);
 144         atomic_inc(&conf->active_stripes);
 145 out:
 146         return sh;
 147 }
 148
 149 static void shrink_buffers(struct stripe_head *sh, int num)
 150 {
 151         struct page *p;
 152         int i;
 153
 154         for (i=0; i<num ; i++) {
 155                 p = sh->dev[i].page;
 156                 if (!p)
 157                         continue;
 158                 sh->dev[i].page = NULL;
 159                 put_page(p);
 160         }
 161 }
 162
 163 static int grow_buffers(struct stripe_head *sh, int num)
 164 {
 165         int i;
 166
 167         for (i=0; i<num; i++) {
 168                 struct page *page;
 169
 170                 if (!(page = alloc_page(GFP_KERNEL))) {
 171                         return 1;
 172                 }
 173                 sh->dev[i].page = page;
 174         }
 175         return 0;
 176 }
 177
 178 static void raid5_build_block (struct stripe_head *sh, int i);
 179
 180 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
 181 {
 182         raid5_conf_t *conf = sh->raid_conf;
 183         int i;
 184
 185         BUG_ON(atomic_read(&sh->count) != 0);
 186         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 187
 188         CHECK_DEVLOCK();
 189         PRINTK("init_stripe called, stripe %llu\n",
 190                 (unsigned long long)sh->sector);
 191
 192         remove_hash(sh);
 193
 194         sh->sector = sector;
 195         sh->pd_idx = pd_idx;
 196         sh->state = 0;
 197
 198         sh->disks = disks;
 199
 200         for (i = sh->disks; i--; ) {
 201                 struct r5dev *dev = &sh->dev[i];
 202
 203                 if (dev->toread || dev->towrite || dev->written ||
 204                     test_bit(R5_LOCKED, &dev->flags)) {
 205                         printk("sector=%llx i=%d %p %p %p %d\n",
 206                                (unsigned long long)sh->sector, i, dev->toread,
 207                                dev->towrite, dev->written,
 208                                test_bit(R5_LOCKED, &dev->flags));
 209                         BUG();
 210                 }
 211                 dev->flags = 0;
 212                 raid5_build_block(sh, i);
 213         }
 214         insert_hash(conf, sh);
 215 }
 216
 217 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
 218 {
 219         struct stripe_head *sh;
 220         struct hlist_node *hn;
 221
 222         CHECK_DEVLOCK();
 223         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
 224         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
 225                 if (sh->sector == sector && sh->disks == disks)
 226                         return sh;
 227         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
 228         return NULL;
 229 }
 230
 231 static void unplug_slaves(mddev_t *mddev);
 232 static void raid5_unplug_device(request_queue_t *q);
 233
 234 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
 235                                              int pd_idx, int noblock)
 236 {
 237         struct stripe_head *sh;
 238
 239         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
 240
 241         spin_lock_irq(&conf->device_lock);
 242
 243         do {
 244                 wait_event_lock_irq(conf->wait_for_stripe,
 245                                     conf->quiesce == 0,
 246                                     conf->device_lock, /* nothing */);
 247                 sh = __find_stripe(conf, sector, disks);
 248                 if (!sh) {
 249                         if (!conf->inactive_blocked)
 250                                 sh = get_free_stripe(conf);
 251                         if (noblock && sh == NULL)
 252                                 break;
 253                         if (!sh) {
 254                                 conf->inactive_blocked = 1;
 255                                 wait_event_lock_irq(conf->wait_for_stripe,
 256                                                     !list_empty(&conf->inactive_list) &&
 257                                                     (atomic_read(&conf->active_stripes)
 258                                                      < (conf->max_nr_stripes *3/4)
 259                                                      || !conf->inactive_blocked),
 260                                                     conf->device_lock,
 261                                                     unplug_slaves(conf->mddev)
 262                                         );
 263                                 conf->inactive_blocked = 0;
 264                         } else
 265                                 init_stripe(sh, sector, pd_idx, disks);
 266                 } else {
 267                         if (atomic_read(&sh->count)) {
 268                           BUG_ON(!list_empty(&sh->lru));
 269                         } else {
 270                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 271                                         atomic_inc(&conf->active_stripes);
 272                                 if (!list_empty(&sh->lru))
 273                                         list_del_init(&sh->lru);
 274                         }
 275                 }
 276         } while (sh == NULL);
 277
 278         if (sh)
 279                 atomic_inc(&sh->count);
 280
 281         spin_unlock_irq(&conf->device_lock);
 282         return sh;
 283 }
 284
 285 static int grow_one_stripe(raid5_conf_t *conf)
 286 {
 287         struct stripe_head *sh;
 288         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
 289         if (!sh)
 290                 return 0;
 291         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
 292         sh->raid_conf = conf;
 293         spin_lock_init(&sh->lock);
 294
 295         if (grow_buffers(sh, conf->raid_disks)) {
 296                 shrink_buffers(sh, conf->raid_disks);
 297                 kmem_cache_free(conf->slab_cache, sh);
 298                 return 0;
 299         }
 300         sh->disks = conf->raid_disks;
 301         /* we just created an active stripe so... */
 302         atomic_set(&sh->count, 1);
 303         atomic_inc(&conf->active_stripes);
 304         INIT_LIST_HEAD(&sh->lru);
 305         release_stripe(sh);
 306         return 1;
 307 }
 308
 309 static int grow_stripes(raid5_conf_t *conf, int num)
 310 {
 311         kmem_cache_t *sc;
 312         int devs = conf->raid_disks;
 313
 314         sprintf(conf->cache_name[0], "raid5/%s", mdname(conf->mddev));
 315         sprintf(conf->cache_name[1], "raid5/%s-alt", mdname(conf->mddev));
 316         conf->active_name = 0;
 317         sc = kmem_cache_create(conf->cache_name[conf->active_name],
 318                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
 319                                0, 0, NULL, NULL);
 320         if (!sc)
 321                 return 1;
 322         conf->slab_cache = sc;
 323         conf->pool_size = devs;
 324         while (num--) {
 325                 if (!grow_one_stripe(conf))
 326                         return 1;
 327         }
 328         return 0;
 329 }
 330
 331 #ifdef CONFIG_MD_RAID5_RESHAPE
 332 static int resize_stripes(raid5_conf_t *conf, int newsize)
 333 {
 334         /* Make all the stripes able to hold 'newsize' devices.
 335          * New slots in each stripe get 'page' set to a new page.
 336          *
 337          * This happens in stages:
 338          * 1/ create a new kmem_cache and allocate the required number of
 339          *    stripe_heads.
 340          * 2/ gather all the old stripe_heads and tranfer the pages across
 341          *    to the new stripe_heads.  This will have the side effect of
 342          *    freezing the array as once all stripe_heads have been collected,
 343          *    no IO will be possible.  Old stripe heads are freed once their
 344          *    pages have been transferred over, and the old kmem_cache is
 345          *    freed when all stripes are done.
 346          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
 347          *    we simple return a failre status - no need to clean anything up.
 348          * 4/ allocate new pages for the new slots in the new stripe_heads.
 349          *    If this fails, we don't bother trying the shrink the
 350          *    stripe_heads down again, we just leave them as they are.
 351          *    As each stripe_head is processed the new one is released into
 352          *    active service.
 353          *
 354          * Once step2 is started, we cannot afford to wait for a write,
 355          * so we use GFP_NOIO allocations.
 356          */
 357         struct stripe_head *osh, *nsh;
 358         LIST_HEAD(newstripes);
 359         struct disk_info *ndisks;
 360         int err = 0;
 361         kmem_cache_t *sc;
 362         int i;
 363
 364         if (newsize <= conf->pool_size)
 365                 return 0; /* never bother to shrink */
 366
 367         /* Step 1 */
 368         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
 369                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
 370                                0, 0, NULL, NULL);
 371         if (!sc)
 372                 return -ENOMEM;
 373
 374         for (i = conf->max_nr_stripes; i; i--) {
 375                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
 376                 if (!nsh)
 377                         break;
 378
 379                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
 380
 381                 nsh->raid_conf = conf;
 382                 spin_lock_init(&nsh->lock);
 383
 384                 list_add(&nsh->lru, &newstripes);
 385         }
 386         if (i) {
 387                 /* didn't get enough, give up */
 388                 while (!list_empty(&newstripes)) {
 389                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
 390                         list_del(&nsh->lru);
 391                         kmem_cache_free(sc, nsh);
 392                 }
 393                 kmem_cache_destroy(sc);
 394                 return -ENOMEM;
 395         }
 396         /* Step 2 - Must use GFP_NOIO now.
 397          * OK, we have enough stripes, start collecting inactive
 398          * stripes and copying them over
 399          */
 400         list_for_each_entry(nsh, &newstripes, lru) {
 401                 spin_lock_irq(&conf->device_lock);
 402                 wait_event_lock_irq(conf->wait_for_stripe,
 403                                     !list_empty(&conf->inactive_list),
 404                                     conf->device_lock,
 405                                     unplug_slaves(conf->mddev)
 406                         );
 407                 osh = get_free_stripe(conf);
 408                 spin_unlock_irq(&conf->device_lock);
 409                 atomic_set(&nsh->count, 1);
 410                 for(i=0; i<conf->pool_size; i++)
 411                         nsh->dev[i].page = osh->dev[i].page;
 412                 for( ; i<newsize; i++)
 413                         nsh->dev[i].page = NULL;
 414                 kmem_cache_free(conf->slab_cache, osh);
 415         }
 416         kmem_cache_destroy(conf->slab_cache);
 417
 418         /* Step 3.
 419          * At this point, we are holding all the stripes so the array
 420          * is completely stalled, so now is a good time to resize
 421          * conf->disks.
 422          */
 423         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
 424         if (ndisks) {
 425                 for (i=0; i<conf->raid_disks; i++)
 426                         ndisks[i] = conf->disks[i];
 427                 kfree(conf->disks);
 428                 conf->disks = ndisks;
 429         } else
 430                 err = -ENOMEM;
 431
 432         /* Step 4, return new stripes to service */
 433         while(!list_empty(&newstripes)) {
 434                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
 435                 list_del_init(&nsh->lru);
 436                 for (i=conf->raid_disks; i < newsize; i++)
 437                         if (nsh->dev[i].page == NULL) {
 438                                 struct page *p = alloc_page(GFP_NOIO);
 439                                 nsh->dev[i].page = p;
 440                                 if (!p)
 441                                         err = -ENOMEM;
 442                         }
 443                 release_stripe(nsh);
 444         }
 445         /* critical section pass, GFP_NOIO no longer needed */
 446
 447         conf->slab_cache = sc;
 448         conf->active_name = 1-conf->active_name;
 449         conf->pool_size = newsize;
 450         return err;
 451 }
 452 #endif
 453
 454 static int drop_one_stripe(raid5_conf_t *conf)
 455 {
 456         struct stripe_head *sh;
 457
 458         spin_lock_irq(&conf->device_lock);
 459         sh = get_free_stripe(conf);
 460         spin_unlock_irq(&conf->device_lock);
 461         if (!sh)
 462                 return 0;
 463         BUG_ON(atomic_read(&sh->count));
 464         shrink_buffers(sh, conf->pool_size);
 465         kmem_cache_free(conf->slab_cache, sh);
 466         atomic_dec(&conf->active_stripes);
 467         return 1;
 468 }
 469
 470 static void shrink_stripes(raid5_conf_t *conf)
 471 {
 472         while (drop_one_stripe(conf))
 473                 ;
 474
 475         if (conf->slab_cache)
 476                 kmem_cache_destroy(conf->slab_cache);
 477         conf->slab_cache = NULL;
 478 }
 479
 480 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
 481                                    int error)
 482 {
 483         struct stripe_head *sh = bi->bi_private;
 484         raid5_conf_t *conf = sh->raid_conf;
 485         int disks = sh->disks, i;
 486         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 487
 488         if (bi->bi_size)
 489                 return 1;
 490
 491         for (i=0 ; i<disks; i++)
 492                 if (bi == &sh->dev[i].req)
 493                         break;
 494
 495         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
 496                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 497                 uptodate);
 498         if (i == disks) {
 499                 BUG();
 500                 return 0;
 501         }
 502
 503         if (uptodate) {
 504 #if 0
 505                 struct bio *bio;
 506                 unsigned long flags;
 507                 spin_lock_irqsave(&conf->device_lock, flags);
 508                 /* we can return a buffer if we bypassed the cache or
 509                  * if the top buffer is not in highmem.  If there are
 510                  * multiple buffers, leave the extra work to
 511                  * handle_stripe
 512                  */
 513                 buffer = sh->bh_read[i];
 514                 if (buffer &&
 515                     (!PageHighMem(buffer->b_page)
 516                      || buffer->b_page == bh->b_page )
 517                         ) {
 518                         sh->bh_read[i] = buffer->b_reqnext;
 519                         buffer->b_reqnext = NULL;
 520                 } else
 521                         buffer = NULL;
 522                 spin_unlock_irqrestore(&conf->device_lock, flags);
 523                 if (sh->bh_page[i]==bh->b_page)
 524                         set_buffer_uptodate(bh);
 525                 if (buffer) {
 526                         if (buffer->b_page != bh->b_page)
 527                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
 528                         buffer->b_end_io(buffer, 1);
 529                 }
 530 #else
 531                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
 532 #endif
 533                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
 534                         printk(KERN_INFO "raid5: read error corrected!!\n");
 535                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 536                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 537                 }
 538                 if (atomic_read(&conf->disks[i].rdev->read_errors))
 539                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
 540         } else {
 541                 int retry = 0;
 542                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
 543                 atomic_inc(&conf->disks[i].rdev->read_errors);
 544                 if (conf->mddev->degraded)
 545                         printk(KERN_WARNING "raid5: read error not correctable.\n");
 546                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
 547                         /* Oh, no!!! */
 548                         printk(KERN_WARNING "raid5: read error NOT corrected!!\n");
 549                 else if (atomic_read(&conf->disks[i].rdev->read_errors)
 550                          > conf->max_nr_stripes)
 551                         printk(KERN_WARNING
 552                                "raid5: Too many read errors, failing device.\n");
 553                 else
 554                         retry = 1;
 555                 if (retry)
 556                         set_bit(R5_ReadError, &sh->dev[i].flags);
 557                 else {
 558                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 559                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 560                         md_error(conf->mddev, conf->disks[i].rdev);
 561                 }
 562         }
 563         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 564 #if 0
 565         /* must restore b_page before unlocking buffer... */
 566         if (sh->bh_page[i] != bh->b_page) {
 567                 bh->b_page = sh->bh_page[i];
 568                 bh->b_data = page_address(bh->b_page);
 569                 clear_buffer_uptodate(bh);
 570         }
 571 #endif
 572         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 573         set_bit(STRIPE_HANDLE, &sh->state);
 574         release_stripe(sh);
 575         return 0;
 576 }
 577
 578 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
 579                                     int error)
 580 {
 581         struct stripe_head *sh = bi->bi_private;
 582         raid5_conf_t *conf = sh->raid_conf;
 583         int disks = sh->disks, i;
 584         unsigned long flags;
 585         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 586
 587         if (bi->bi_size)
 588                 return 1;
 589
 590         for (i=0 ; i<disks; i++)
 591                 if (bi == &sh->dev[i].req)
 592                         break;
 593
 594         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
 595                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 596                 uptodate);
 597         if (i == disks) {
 598                 BUG();
 599                 return 0;
 600         }
 601
 602         spin_lock_irqsave(&conf->device_lock, flags);
 603         if (!uptodate)
 604                 md_error(conf->mddev, conf->disks[i].rdev);
 605
 606         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 607
 608         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 609         set_bit(STRIPE_HANDLE, &sh->state);
 610         __release_stripe(conf, sh);
 611         spin_unlock_irqrestore(&conf->device_lock, flags);
 612         return 0;
 613 }
 614
 615
 616 static sector_t compute_blocknr(struct stripe_head *sh, int i);
 617
 618 static void raid5_build_block (struct stripe_head *sh, int i)
 619 {
 620         struct r5dev *dev = &sh->dev[i];
 621
 622         bio_init(&dev->req);
 623         dev->req.bi_io_vec = &dev->vec;
 624         dev->req.bi_vcnt++;
 625         dev->req.bi_max_vecs++;
 626         dev->vec.bv_page = dev->page;
 627         dev->vec.bv_len = STRIPE_SIZE;
 628         dev->vec.bv_offset = 0;
 629
 630         dev->req.bi_sector = sh->sector;
 631         dev->req.bi_private = sh;
 632
 633         dev->flags = 0;
 634         if (i != sh->pd_idx)
 635                 dev->sector = compute_blocknr(sh, i);
 636 }
 637
 638 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 639 {
 640         char b[BDEVNAME_SIZE];
 641         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
 642         PRINTK("raid5: error called\n");
 643
 644         if (!test_bit(Faulty, &rdev->flags)) {
 645                 mddev->sb_dirty = 1;
 646                 if (test_bit(In_sync, &rdev->flags)) {
 647                         conf->working_disks--;
 648                         mddev->degraded++;
 649                         conf->failed_disks++;
 650                         clear_bit(In_sync, &rdev->flags);
 651                         /*
 652                          * if recovery was running, make sure it aborts.
 653                          */
 654                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 655                 }
 656                 set_bit(Faulty, &rdev->flags);
 657                 printk (KERN_ALERT
 658                         "raid5: Disk failure on %s, disabling device."
 659                         " Operation continuing on %d devices\n",
 660                         bdevname(rdev->bdev,b), conf->working_disks);
 661         }
 662 }
 663
 664 /*
 665  * Input: a 'big' sector number,
 666  * Output: index of the data and parity disk, and the sector # in them.
 667  */
 668 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
 669                         unsigned int data_disks, unsigned int * dd_idx,
 670                         unsigned int * pd_idx, raid5_conf_t *conf)
 671 {
 672         long stripe;
 673         unsigned long chunk_number;
 674         unsigned int chunk_offset;
 675         sector_t new_sector;
 676         int sectors_per_chunk = conf->chunk_size >> 9;
 677
 678         /* First compute the information on this sector */
 679
 680         /*
 681          * Compute the chunk number and the sector offset inside the chunk
 682          */
 683         chunk_offset = sector_div(r_sector, sectors_per_chunk);
 684         chunk_number = r_sector;
 685         BUG_ON(r_sector != chunk_number);
 686
 687         /*
 688          * Compute the stripe number
 689          */
 690         stripe = chunk_number / data_disks;
 691
 692         /*
 693          * Compute the data disk and parity disk indexes inside the stripe
 694          */
 695         *dd_idx = chunk_number % data_disks;
 696
 697         /*
 698          * Select the parity disk based on the user selected algorithm.
 699          */
 700         if (conf->level == 4)
 701                 *pd_idx = data_disks;
 702         else switch (conf->algorithm) {
 703                 case ALGORITHM_LEFT_ASYMMETRIC:
 704                         *pd_idx = data_disks - stripe % raid_disks;
 705                         if (*dd_idx >= *pd_idx)
 706                                 (*dd_idx)++;
 707                         break;
 708                 case ALGORITHM_RIGHT_ASYMMETRIC:
 709                         *pd_idx = stripe % raid_disks;
 710                         if (*dd_idx >= *pd_idx)
 711                                 (*dd_idx)++;
 712                         break;
 713                 case ALGORITHM_LEFT_SYMMETRIC:
 714                         *pd_idx = data_disks - stripe % raid_disks;
 715                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 716                         break;
 717                 case ALGORITHM_RIGHT_SYMMETRIC:
 718                         *pd_idx = stripe % raid_disks;
 719                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 720                         break;
 721                 default:
 722                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
 723                                 conf->algorithm);
 724         }
 725
 726         /*
 727          * Finally, compute the new sector number
 728          */
 729         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
 730         return new_sector;
 731 }
 732
 733
 734 static sector_t compute_blocknr(struct stripe_head *sh, int i)
 735 {
 736         raid5_conf_t *conf = sh->raid_conf;
 737         int raid_disks = sh->disks, data_disks = raid_disks - 1;
 738         sector_t new_sector = sh->sector, check;
 739         int sectors_per_chunk = conf->chunk_size >> 9;
 740         sector_t stripe;
 741         int chunk_offset;
 742         int chunk_number, dummy1, dummy2, dd_idx = i;
 743         sector_t r_sector;
 744
 745         chunk_offset = sector_div(new_sector, sectors_per_chunk);
 746         stripe = new_sector;
 747         BUG_ON(new_sector != stripe);
 748
 749
 750         switch (conf->algorithm) {
 751                 case ALGORITHM_LEFT_ASYMMETRIC:
 752                 case ALGORITHM_RIGHT_ASYMMETRIC:
 753                         if (i > sh->pd_idx)
 754                                 i--;
 755                         break;
 756                 case ALGORITHM_LEFT_SYMMETRIC:
 757                 case ALGORITHM_RIGHT_SYMMETRIC:
 758                         if (i < sh->pd_idx)
 759                                 i += raid_disks;
 760                         i -= (sh->pd_idx + 1);
 761                         break;
 762                 default:
 763                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
 764                                 conf->algorithm);
 765         }
 766
 767         chunk_number = stripe * data_disks + i;
 768         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
 769
 770         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
 771         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
 772                 printk(KERN_ERR "compute_blocknr: map not correct\n");
 773                 return 0;
 774         }
 775         return r_sector;
 776 }
 777
 778
 779
 780 /*
 781  * Copy data between a page in the stripe cache, and a bio.
 782  * There are no alignment or size guarantees between the page or the
 783  * bio except that there is some overlap.
 784  * All iovecs in the bio must be considered.
 785  */
 786 static void copy_data(int frombio, struct bio *bio,
 787                      struct page *page,
 788                      sector_t sector)
 789 {
 790         char *pa = page_address(page);
 791         struct bio_vec *bvl;
 792         int i;
 793         int page_offset;
 794
 795         if (bio->bi_sector >= sector)
 796                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 797         else
 798                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 799         bio_for_each_segment(bvl, bio, i) {
 800                 int len = bio_iovec_idx(bio,i)->bv_len;
 801                 int clen;
 802                 int b_offset = 0;
 803
 804                 if (page_offset < 0) {
 805                         b_offset = -page_offset;
 806                         page_offset += b_offset;
 807                         len -= b_offset;
 808                 }
 809
 810                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 811                         clen = STRIPE_SIZE - page_offset;
 812                 else clen = len;
 813
 814                 if (clen > 0) {
 815                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
 816                         if (frombio)
 817                                 memcpy(pa+page_offset, ba+b_offset, clen);
 818                         else
 819                                 memcpy(ba+b_offset, pa+page_offset, clen);
 820                         __bio_kunmap_atomic(ba, KM_USER0);
 821                 }
 822                 if (clen < len) /* hit end of page */
 823                         break;
 824                 page_offset +=  len;
 825         }
 826 }
 827
 828 #define check_xor()     do {                                            \
 829                            if (count == MAX_XOR_BLOCKS) {               \
 830                                 xor_block(count, STRIPE_SIZE, ptr);     \
 831                                 count = 1;                              \
 832                            }                                            \
 833                         } while(0)
 834
 835
 836 static void compute_block(struct stripe_head *sh, int dd_idx)
 837 {
 838         int i, count, disks = sh->disks;
 839         void *ptr[MAX_XOR_BLOCKS], *p;
 840
 841         PRINTK("compute_block, stripe %llu, idx %d\n",
 842                 (unsigned long long)sh->sector, dd_idx);
 843
 844         ptr[0] = page_address(sh->dev[dd_idx].page);
 845         memset(ptr[0], 0, STRIPE_SIZE);
 846         count = 1;
 847         for (i = disks ; i--; ) {
 848                 if (i == dd_idx)
 849                         continue;
 850                 p = page_address(sh->dev[i].page);
 851                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
 852                         ptr[count++] = p;
 853                 else
 854                         printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
 855                                 " not present\n", dd_idx,
 856                                 (unsigned long long)sh->sector, i);
 857
 858                 check_xor();
 859         }
 860         if (count != 1)
 861                 xor_block(count, STRIPE_SIZE, ptr);
 862         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
 863 }
 864
 865 static void compute_parity(struct stripe_head *sh, int method)
 866 {
 867         raid5_conf_t *conf = sh->raid_conf;
 868         int i, pd_idx = sh->pd_idx, disks = sh->disks, count;
 869         void *ptr[MAX_XOR_BLOCKS];
 870         struct bio *chosen;
 871
 872         PRINTK("compute_parity, stripe %llu, method %d\n",
 873                 (unsigned long long)sh->sector, method);
 874
 875         count = 1;
 876         ptr[0] = page_address(sh->dev[pd_idx].page);
 877         switch(method) {
 878         case READ_MODIFY_WRITE:
 879                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags));
 880                 for (i=disks ; i-- ;) {
 881                         if (i==pd_idx)
 882                                 continue;
 883                         if (sh->dev[i].towrite &&
 884                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
 885                                 ptr[count++] = page_address(sh->dev[i].page);
 886                                 chosen = sh->dev[i].towrite;
 887                                 sh->dev[i].towrite = NULL;
 888
 889                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 890                                         wake_up(&conf->wait_for_overlap);
 891
 892                                 BUG_ON(sh->dev[i].written);
 893                                 sh->dev[i].written = chosen;
 894                                 check_xor();
 895                         }
 896                 }
 897                 break;
 898         case RECONSTRUCT_WRITE:
 899                 memset(ptr[0], 0, STRIPE_SIZE);
 900                 for (i= disks; i-- ;)
 901                         if (i!=pd_idx && sh->dev[i].towrite) {
 902                                 chosen = sh->dev[i].towrite;
 903                                 sh->dev[i].towrite = NULL;
 904
 905                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 906                                         wake_up(&conf->wait_for_overlap);
 907
 908                                 BUG_ON(sh->dev[i].written);
 909                                 sh->dev[i].written = chosen;
 910                         }
 911                 break;
 912         case CHECK_PARITY:
 913                 break;
 914         }
 915         if (count>1) {
 916                 xor_block(count, STRIPE_SIZE, ptr);
 917                 count = 1;
 918         }
 919
 920         for (i = disks; i--;)
 921                 if (sh->dev[i].written) {
 922                         sector_t sector = sh->dev[i].sector;
 923                         struct bio *wbi = sh->dev[i].written;
 924                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
 925                                 copy_data(1, wbi, sh->dev[i].page, sector);
 926                                 wbi = r5_next_bio(wbi, sector);
 927                         }
 928
 929                         set_bit(R5_LOCKED, &sh->dev[i].flags);
 930                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
 931                 }
 932
 933         switch(method) {
 934         case RECONSTRUCT_WRITE:
 935         case CHECK_PARITY:
 936                 for (i=disks; i--;)
 937                         if (i != pd_idx) {
 938                                 ptr[count++] = page_address(sh->dev[i].page);
 939                                 check_xor();
 940                         }
 941                 break;
 942         case READ_MODIFY_WRITE:
 943                 for (i = disks; i--;)
 944                         if (sh->dev[i].written) {
 945                                 ptr[count++] = page_address(sh->dev[i].page);
 946                                 check_xor();
 947                         }
 948         }
 949         if (count != 1)
 950                 xor_block(count, STRIPE_SIZE, ptr);
 951
 952         if (method != CHECK_PARITY) {
 953                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 954                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
 955         } else
 956                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 957 }
 958
 959 /*
 960  * Each stripe/dev can have one or more bion attached.
 961  * toread/towrite point to the first in a chain.
 962  * The bi_next chain must be in order.
 963  */
 964 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
 965 {
 966         struct bio **bip;
 967         raid5_conf_t *conf = sh->raid_conf;
 968         int firstwrite=0;
 969
 970         PRINTK("adding bh b#%llu to stripe s#%llu\n",
 971                 (unsigned long long)bi->bi_sector,
 972                 (unsigned long long)sh->sector);
 973
 974
 975         spin_lock(&sh->lock);
 976         spin_lock_irq(&conf->device_lock);
 977         if (forwrite) {
 978                 bip = &sh->dev[dd_idx].towrite;
 979                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
 980                         firstwrite = 1;
 981         } else
 982                 bip = &sh->dev[dd_idx].toread;
 983         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
 984                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
 985                         goto overlap;
 986                 bip = & (*bip)->bi_next;
 987         }
 988         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
 989                 goto overlap;
 990
 991         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
 992         if (*bip)
 993                 bi->bi_next = *bip;
 994         *bip = bi;
 995         bi->bi_phys_segments ++;
 996         spin_unlock_irq(&conf->device_lock);
 997         spin_unlock(&sh->lock);
 998
 999         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
1000                 (unsigned long long)bi->bi_sector,
1001                 (unsigned long long)sh->sector, dd_idx);
1002
1003         if (conf->mddev->bitmap && firstwrite) {
1004                 sh->bm_seq = conf->seq_write;
1005                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1006                                   STRIPE_SECTORS, 0);
1007                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1008         }
1009
1010         if (forwrite) {
1011                 /* check if page is covered */
1012                 sector_t sector = sh->dev[dd_idx].sector;
1013                 for (bi=sh->dev[dd_idx].towrite;
1014                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1015                              bi && bi->bi_sector <= sector;
1016                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1017                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1018                                 sector = bi->bi_sector + (bi->bi_size>>9);
1019                 }
1020                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1021                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1022         }
1023         return 1;
1024
1025  overlap:
1026         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1027         spin_unlock_irq(&conf->device_lock);
1028         spin_unlock(&sh->lock);
1029         return 0;
1030 }
1031
1032 static void end_reshape(raid5_conf_t *conf);
1033
1034 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1035 {
1036         int sectors_per_chunk = conf->chunk_size >> 9;
1037         sector_t x = stripe;
1038         int pd_idx, dd_idx;
1039         int chunk_offset = sector_div(x, sectors_per_chunk);
1040         stripe = x;
1041         raid5_compute_sector(stripe*(disks-1)*sectors_per_chunk
1042                              + chunk_offset, disks, disks-1, &dd_idx, &pd_idx, conf);
1043         return pd_idx;
1044 }
1045
1046
1047 /*
1048  * handle_stripe - do things to a stripe.
1049  *
1050  * We lock the stripe and then examine the state of various bits
1051  * to see what needs to be done.
1052  * Possible results:
1053  *    return some read request which now have data
1054  *    return some write requests which are safely on disc
1055  *    schedule a read on some buffers
1056  *    schedule a write of some buffers
1057  *    return confirmation of parity correctness
1058  *
1059  * Parity calculations are done inside the stripe lock
1060  * buffers are taken off read_list or write_list, and bh_cache buffers
1061  * get BH_Lock set before the stripe lock is released.
1062  *
1063  */
1064
1065 static void handle_stripe(struct stripe_head *sh)
1066 {
1067         raid5_conf_t *conf = sh->raid_conf;
1068         int disks = sh->disks;
1069         struct bio *return_bi= NULL;
1070         struct bio *bi;
1071         int i;
1072         int syncing, expanding, expanded;
1073         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1074         int non_overwrite = 0;
1075         int failed_num=0;
1076         struct r5dev *dev;
1077
1078         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
1079                 (unsigned long long)sh->sector, atomic_read(&sh->count),
1080                 sh->pd_idx);
1081
1082         spin_lock(&sh->lock);
1083         clear_bit(STRIPE_HANDLE, &sh->state);
1084         clear_bit(STRIPE_DELAYED, &sh->state);
1085
1086         syncing = test_bit(STRIPE_SYNCING, &sh->state);
1087         expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1088         expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
1089         /* Now to look around and see what can be done */
1090
1091         rcu_read_lock();
1092         for (i=disks; i--; ) {
1093                 mdk_rdev_t *rdev;
1094                 dev = &sh->dev[i];
1095                 clear_bit(R5_Insync, &dev->flags);
1096
1097                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1098                         i, dev->flags, dev->toread, dev->towrite, dev->written);
1099                 /* maybe we can reply to a read */
1100                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1101                         struct bio *rbi, *rbi2;
1102                         PRINTK("Return read for disc %d\n", i);
1103                         spin_lock_irq(&conf->device_lock);
1104                         rbi = dev->toread;
1105                         dev->toread = NULL;
1106                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1107                                 wake_up(&conf->wait_for_overlap);
1108                         spin_unlock_irq(&conf->device_lock);
1109                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1110                                 copy_data(0, rbi, dev->page, dev->sector);
1111                                 rbi2 = r5_next_bio(rbi, dev->sector);
1112                                 spin_lock_irq(&conf->device_lock);
1113                                 if (--rbi->bi_phys_segments == 0) {
1114                                         rbi->bi_next = return_bi;
1115                                         return_bi = rbi;
1116                                 }
1117                                 spin_unlock_irq(&conf->device_lock);
1118                                 rbi = rbi2;
1119                         }
1120                 }
1121
1122                 /* now count some things */
1123                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1124                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1125
1126
1127                 if (dev->toread) to_read++;
1128                 if (dev->towrite) {
1129                         to_write++;
1130                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1131                                 non_overwrite++;
1132                 }
1133                 if (dev->written) written++;
1134                 rdev = rcu_dereference(conf->disks[i].rdev);
1135                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1136                         /* The ReadError flag will just be confusing now */
1137                         clear_bit(R5_ReadError, &dev->flags);
1138                         clear_bit(R5_ReWrite, &dev->flags);
1139                 }
1140                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1141                     || test_bit(R5_ReadError, &dev->flags)) {
1142                         failed++;
1143                         failed_num = i;
1144                 } else
1145                         set_bit(R5_Insync, &dev->flags);
1146         }
1147         rcu_read_unlock();
1148         PRINTK("locked=%d uptodate=%d to_read=%d"
1149                 " to_write=%d failed=%d failed_num=%d\n",
1150                 locked, uptodate, to_read, to_write, failed, failed_num);
1151         /* check if the array has lost two devices and, if so, some requests might
1152          * need to be failed
1153          */
1154         if (failed > 1 && to_read+to_write+written) {
1155                 for (i=disks; i--; ) {
1156                         int bitmap_end = 0;
1157
1158                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1159                                 mdk_rdev_t *rdev;
1160                                 rcu_read_lock();
1161                                 rdev = rcu_dereference(conf->disks[i].rdev);
1162                                 if (rdev && test_bit(In_sync, &rdev->flags))
1163                                         /* multiple read failures in one stripe */
1164                                         md_error(conf->mddev, rdev);
1165                                 rcu_read_unlock();
1166                         }
1167
1168                         spin_lock_irq(&conf->device_lock);
1169                         /* fail all writes first */
1170                         bi = sh->dev[i].towrite;
1171                         sh->dev[i].towrite = NULL;
1172                         if (bi) { to_write--; bitmap_end = 1; }
1173
1174                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1175                                 wake_up(&conf->wait_for_overlap);
1176
1177                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1178                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1179                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1180                                 if (--bi->bi_phys_segments == 0) {
1181                                         md_write_end(conf->mddev);
1182                                         bi->bi_next = return_bi;
1183                                         return_bi = bi;
1184                                 }
1185                                 bi = nextbi;
1186                         }
1187                         /* and fail all 'written' */
1188                         bi = sh->dev[i].written;
1189                         sh->dev[i].written = NULL;
1190                         if (bi) bitmap_end = 1;
1191                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1192                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1193                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1194                                 if (--bi->bi_phys_segments == 0) {
1195                                         md_write_end(conf->mddev);
1196                                         bi->bi_next = return_bi;
1197                                         return_bi = bi;
1198                                 }
1199                                 bi = bi2;
1200                         }
1201
1202                         /* fail any reads if this device is non-operational */
1203                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1204                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1205                                 bi = sh->dev[i].toread;
1206                                 sh->dev[i].toread = NULL;
1207                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1208                                         wake_up(&conf->wait_for_overlap);
1209                                 if (bi) to_read--;
1210                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1211                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1212                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1213                                         if (--bi->bi_phys_segments == 0) {
1214                                                 bi->bi_next = return_bi;
1215                                                 return_bi = bi;
1216                                         }
1217                                         bi = nextbi;
1218                                 }
1219                         }
1220                         spin_unlock_irq(&conf->device_lock);
1221                         if (bitmap_end)
1222                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1223                                                 STRIPE_SECTORS, 0, 0);
1224                 }
1225         }
1226         if (failed > 1 && syncing) {
1227                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1228                 clear_bit(STRIPE_SYNCING, &sh->state);
1229                 syncing = 0;
1230         }
1231
1232         /* might be able to return some write requests if the parity block
1233          * is safe, or on a failed drive
1234          */
1235         dev = &sh->dev[sh->pd_idx];
1236         if ( written &&
1237              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1238                 test_bit(R5_UPTODATE, &dev->flags))
1239                || (failed == 1 && failed_num == sh->pd_idx))
1240             ) {
1241             /* any written block on an uptodate or failed drive can be returned.
1242              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
1243              * never LOCKED, so we don't need to test 'failed' directly.
1244              */
1245             for (i=disks; i--; )
1246                 if (sh->dev[i].written) {
1247                     dev = &sh->dev[i];
1248                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1249                          test_bit(R5_UPTODATE, &dev->flags) ) {
1250                         /* We can return any write requests */
1251                             struct bio *wbi, *wbi2;
1252                             int bitmap_end = 0;
1253                             PRINTK("Return write for disc %d\n", i);
1254                             spin_lock_irq(&conf->device_lock);
1255                             wbi = dev->written;
1256                             dev->written = NULL;
1257                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1258                                     wbi2 = r5_next_bio(wbi, dev->sector);
1259                                     if (--wbi->bi_phys_segments == 0) {
1260                                             md_write_end(conf->mddev);
1261                                             wbi->bi_next = return_bi;
1262                                             return_bi = wbi;
1263                                     }
1264                                     wbi = wbi2;
1265                             }
1266                             if (dev->towrite == NULL)
1267                                     bitmap_end = 1;
1268                             spin_unlock_irq(&conf->device_lock);
1269                             if (bitmap_end)
1270                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1271                                                     STRIPE_SECTORS,
1272                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1273                     }
1274                 }
1275         }
1276
1277         /* Now we might consider reading some blocks, either to check/generate
1278          * parity, or to satisfy requests
1279          * or to load a block that is being partially written.
1280          */
1281         if (to_read || non_overwrite || (syncing && (uptodate < disks)) || expanding) {
1282                 for (i=disks; i--;) {
1283                         dev = &sh->dev[i];
1284                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1285                             (dev->toread ||
1286                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1287                              syncing ||
1288                              expanding ||
1289                              (failed && (sh->dev[failed_num].toread ||
1290                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1291                                     )
1292                                 ) {
1293                                 /* we would like to get this block, possibly
1294                                  * by computing it, but we might not be able to
1295                                  */
1296                                 if (uptodate == disks-1) {
1297                                         PRINTK("Computing block %d\n", i);
1298                                         compute_block(sh, i);
1299                                         uptodate++;
1300                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1301                                         set_bit(R5_LOCKED, &dev->flags);
1302                                         set_bit(R5_Wantread, &dev->flags);
1303 #if 0
1304                                         /* if I am just reading this block and we don't have
1305                                            a failed drive, or any pending writes then sidestep the cache */
1306                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1307                                             ! syncing && !failed && !to_write) {
1308                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1309                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1310                                         }
1311 #endif
1312                                         locked++;
1313                                         PRINTK("Reading block %d (sync=%d)\n",
1314                                                 i, syncing);
1315                                 }
1316                         }
1317                 }
1318                 set_bit(STRIPE_HANDLE, &sh->state);
1319         }
1320
1321         /* now to consider writing and what else, if anything should be read */
1322         if (to_write) {
1323                 int rmw=0, rcw=0;
1324                 for (i=disks ; i--;) {
1325                         /* would I have to read this buffer for read_modify_write */
1326                         dev = &sh->dev[i];
1327                         if ((dev->towrite || i == sh->pd_idx) &&
1328                             (!test_bit(R5_LOCKED, &dev->flags)
1329 #if 0
1330 || sh->bh_page[i]!=bh->b_page
1331 #endif
1332                                     ) &&
1333                             !test_bit(R5_UPTODATE, &dev->flags)) {
1334                                 if (test_bit(R5_Insync, &dev->flags)
1335 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1336                                         )
1337                                         rmw++;
1338                                 else rmw += 2*disks;  /* cannot read it */
1339                         }
1340                         /* Would I have to read this buffer for reconstruct_write */
1341                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1342                             (!test_bit(R5_LOCKED, &dev->flags)
1343 #if 0
1344 || sh->bh_page[i] != bh->b_page
1345 #endif
1346                                     ) &&
1347                             !test_bit(R5_UPTODATE, &dev->flags)) {
1348                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1349                                 else rcw += 2*disks;
1350                         }
1351                 }
1352                 PRINTK("for sector %llu, rmw=%d rcw=%d\n",
1353                         (unsigned long long)sh->sector, rmw, rcw);
1354                 set_bit(STRIPE_HANDLE, &sh->state);
1355                 if (rmw < rcw && rmw > 0)
1356                         /* prefer read-modify-write, but need to get some data */
1357                         for (i=disks; i--;) {
1358                                 dev = &sh->dev[i];
1359                                 if ((dev->towrite || i == sh->pd_idx) &&
1360                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1361                                     test_bit(R5_Insync, &dev->flags)) {
1362                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1363                                         {
1364                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1365                                                 set_bit(R5_LOCKED, &dev->flags);
1366                                                 set_bit(R5_Wantread, &dev->flags);
1367                                                 locked++;
1368                                         } else {
1369                                                 set_bit(STRIPE_DELAYED, &sh->state);
1370                                                 set_bit(STRIPE_HANDLE, &sh->state);
1371                                         }
1372                                 }
1373                         }
1374                 if (rcw <= rmw && rcw > 0)
1375                         /* want reconstruct write, but need to get some data */
1376                         for (i=disks; i--;) {
1377                                 dev = &sh->dev[i];
1378                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1379                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1380                                     test_bit(R5_Insync, &dev->flags)) {
1381                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1382                                         {
1383                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1384                                                 set_bit(R5_LOCKED, &dev->flags);
1385                                                 set_bit(R5_Wantread, &dev->flags);
1386                                                 locked++;
1387                                         } else {
1388                                                 set_bit(STRIPE_DELAYED, &sh->state);
1389                                                 set_bit(STRIPE_HANDLE, &sh->state);
1390                                         }
1391                                 }
1392                         }
1393                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1394                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1395                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1396                         PRINTK("Computing parity...\n");
1397                         compute_parity(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1398                         /* now every locked buffer is ready to be written */
1399                         for (i=disks; i--;)
1400                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1401                                         PRINTK("Writing block %d\n", i);
1402                                         locked++;
1403                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1404                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1405                                             || (i==sh->pd_idx && failed == 0))
1406                                                 set_bit(STRIPE_INSYNC, &sh->state);
1407                                 }
1408                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1409                                 atomic_dec(&conf->preread_active_stripes);
1410                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1411                                         md_wakeup_thread(conf->mddev->thread);
1412                         }
1413                 }
1414         }
1415
1416         /* maybe we need to check and possibly fix the parity for this stripe
1417          * Any reads will already have been scheduled, so we just see if enough data
1418          * is available
1419          */
1420         if (syncing && locked == 0 &&
1421             !test_bit(STRIPE_INSYNC, &sh->state)) {
1422                 set_bit(STRIPE_HANDLE, &sh->state);
1423                 if (failed == 0) {
1424                         char *pagea;
1425                         BUG_ON(uptodate != disks);
1426                         compute_parity(sh, CHECK_PARITY);
1427                         uptodate--;
1428                         pagea = page_address(sh->dev[sh->pd_idx].page);
1429                         if ((*(u32*)pagea) == 0 &&
1430                             !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1431                                 /* parity is correct (on disc, not in buffer any more) */
1432                                 set_bit(STRIPE_INSYNC, &sh->state);
1433                         } else {
1434                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1435                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1436                                         /* don't try to repair!! */
1437                                         set_bit(STRIPE_INSYNC, &sh->state);
1438                                 else {
1439                                         compute_block(sh, sh->pd_idx);
1440                                         uptodate++;
1441                                 }
1442                         }
1443                 }
1444                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1445                         /* either failed parity check, or recovery is happening */
1446                         if (failed==0)
1447                                 failed_num = sh->pd_idx;
1448                         dev = &sh->dev[failed_num];
1449                         BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
1450                         BUG_ON(uptodate != disks);
1451
1452                         set_bit(R5_LOCKED, &dev->flags);
1453                         set_bit(R5_Wantwrite, &dev->flags);
1454                         clear_bit(STRIPE_DEGRADED, &sh->state);
1455                         locked++;
1456                         set_bit(STRIPE_INSYNC, &sh->state);
1457                 }
1458         }
1459         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1460                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1461                 clear_bit(STRIPE_SYNCING, &sh->state);
1462         }
1463
1464         /* If the failed drive is just a ReadError, then we might need to progress
1465          * the repair/check process
1466          */
1467         if (failed == 1 && ! conf->mddev->ro &&
1468             test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1469             && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1470             && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1471                 ) {
1472                 dev = &sh->dev[failed_num];
1473                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1474                         set_bit(R5_Wantwrite, &dev->flags);
1475                         set_bit(R5_ReWrite, &dev->flags);
1476                         set_bit(R5_LOCKED, &dev->flags);
1477                         locked++;
1478                 } else {
1479                         /* let's read it back */
1480                         set_bit(R5_Wantread, &dev->flags);
1481                         set_bit(R5_LOCKED, &dev->flags);
1482                         locked++;
1483                 }
1484         }
1485
1486         if (expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
1487                 /* Need to write out all blocks after computing parity */
1488                 sh->disks = conf->raid_disks;
1489                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, conf->raid_disks);
1490                 compute_parity(sh, RECONSTRUCT_WRITE);
1491                 for (i= conf->raid_disks; i--;) {
1492                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1493                         locked++;
1494                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1495                 }
1496                 clear_bit(STRIPE_EXPANDING, &sh->state);
1497         } else if (expanded) {
1498                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
1499                 atomic_dec(&conf->reshape_stripes);
1500                 wake_up(&conf->wait_for_overlap);
1501                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
1502         }
1503
1504         if (expanding && locked == 0) {
1505                 /* We have read all the blocks in this stripe and now we need to
1506                  * copy some of them into a target stripe for expand.
1507                  */
1508                 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1509                 for (i=0; i< sh->disks; i++)
1510                         if (i != sh->pd_idx) {
1511                                 int dd_idx, pd_idx, j;
1512                                 struct stripe_head *sh2;
1513
1514                                 sector_t bn = compute_blocknr(sh, i);
1515                                 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
1516                                                                   conf->raid_disks-1,
1517                                                                   &dd_idx, &pd_idx, conf);
1518                                 sh2 = get_active_stripe(conf, s, conf->raid_disks, pd_idx, 1);
1519                                 if (sh2 == NULL)
1520                                         /* so far only the early blocks of this stripe
1521                                          * have been requested.  When later blocks
1522                                          * get requested, we will try again
1523                                          */
1524                                         continue;
1525                                 if(!test_bit(STRIPE_EXPANDING, &sh2->state) ||
1526                                    test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
1527                                         /* must have already done this block */
1528                                         release_stripe(sh2);
1529                                         continue;
1530                                 }
1531                                 memcpy(page_address(sh2->dev[dd_idx].page),
1532                                        page_address(sh->dev[i].page),
1533                                        STRIPE_SIZE);
1534                                 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
1535                                 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
1536                                 for (j=0; j<conf->raid_disks; j++)
1537                                         if (j != sh2->pd_idx &&
1538                                             !test_bit(R5_Expanded, &sh2->dev[j].flags))
1539                                                 break;
1540                                 if (j == conf->raid_disks) {
1541                                         set_bit(STRIPE_EXPAND_READY, &sh2->state);
1542                                         set_bit(STRIPE_HANDLE, &sh2->state);
1543                                 }
1544                                 release_stripe(sh2);
1545                         }
1546         }
1547
1548         spin_unlock(&sh->lock);
1549
1550         while ((bi=return_bi)) {
1551                 int bytes = bi->bi_size;
1552
1553                 return_bi = bi->bi_next;
1554                 bi->bi_next = NULL;
1555                 bi->bi_size = 0;
1556                 bi->bi_end_io(bi, bytes, 0);
1557         }
1558         for (i=disks; i-- ;) {
1559                 int rw;
1560                 struct bio *bi;
1561                 mdk_rdev_t *rdev;
1562                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1563                         rw = 1;
1564                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1565                         rw = 0;
1566                 else
1567                         continue;
1568
1569                 bi = &sh->dev[i].req;
1570
1571                 bi->bi_rw = rw;
1572                 if (rw)
1573                         bi->bi_end_io = raid5_end_write_request;
1574                 else
1575                         bi->bi_end_io = raid5_end_read_request;
1576
1577                 rcu_read_lock();
1578                 rdev = rcu_dereference(conf->disks[i].rdev);
1579                 if (rdev && test_bit(Faulty, &rdev->flags))
1580                         rdev = NULL;
1581                 if (rdev)
1582                         atomic_inc(&rdev->nr_pending);
1583                 rcu_read_unlock();
1584
1585                 if (rdev) {
1586                         if (syncing || expanding || expanded)
1587                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1588
1589                         bi->bi_bdev = rdev->bdev;
1590                         PRINTK("for %llu schedule op %ld on disc %d\n",
1591                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1592                         atomic_inc(&sh->count);
1593                         bi->bi_sector = sh->sector + rdev->data_offset;
1594                         bi->bi_flags = 1 << BIO_UPTODATE;
1595                         bi->bi_vcnt = 1;
1596                         bi->bi_max_vecs = 1;
1597                         bi->bi_idx = 0;
1598                         bi->bi_io_vec = &sh->dev[i].vec;
1599                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1600                         bi->bi_io_vec[0].bv_offset = 0;
1601                         bi->bi_size = STRIPE_SIZE;
1602                         bi->bi_next = NULL;
1603                         if (rw == WRITE &&
1604                             test_bit(R5_ReWrite, &sh->dev[i].flags))
1605                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1606                         generic_make_request(bi);
1607                 } else {
1608                         if (rw == 1)
1609                                 set_bit(STRIPE_DEGRADED, &sh->state);
1610                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1611                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1612                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1613                         set_bit(STRIPE_HANDLE, &sh->state);
1614                 }
1615         }
1616 }
1617
1618 static void raid5_activate_delayed(raid5_conf_t *conf)
1619 {
1620         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1621                 while (!list_empty(&conf->delayed_list)) {
1622                         struct list_head *l = conf->delayed_list.next;
1623                         struct stripe_head *sh;
1624                         sh = list_entry(l, struct stripe_head, lru);
1625                         list_del_init(l);
1626                         clear_bit(STRIPE_DELAYED, &sh->state);
1627                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1628                                 atomic_inc(&conf->preread_active_stripes);
1629                         list_add_tail(&sh->lru, &conf->handle_list);
1630                 }
1631         }
1632 }
1633
1634 static void activate_bit_delay(raid5_conf_t *conf)
1635 {
1636         /* device_lock is held */
1637         struct list_head head;
1638         list_add(&head, &conf->bitmap_list);
1639         list_del_init(&conf->bitmap_list);
1640         while (!list_empty(&head)) {
1641                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1642                 list_del_init(&sh->lru);
1643                 atomic_inc(&sh->count);
1644                 __release_stripe(conf, sh);
1645         }
1646 }
1647
1648 static void unplug_slaves(mddev_t *mddev)
1649 {
1650         raid5_conf_t *conf = mddev_to_conf(mddev);
1651         int i;
1652
1653         rcu_read_lock();
1654         for (i=0; i<mddev->raid_disks; i++) {
1655                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1656                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
1657                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1658
1659                         atomic_inc(&rdev->nr_pending);
1660                         rcu_read_unlock();
1661
1662                         if (r_queue->unplug_fn)
1663                                 r_queue->unplug_fn(r_queue);
1664
1665                         rdev_dec_pending(rdev, mddev);
1666                         rcu_read_lock();
1667                 }
1668         }
1669         rcu_read_unlock();
1670 }
1671
1672 static void raid5_unplug_device(request_queue_t *q)
1673 {
1674         mddev_t *mddev = q->queuedata;
1675         raid5_conf_t *conf = mddev_to_conf(mddev);
1676         unsigned long flags;
1677
1678         spin_lock_irqsave(&conf->device_lock, flags);
1679
1680         if (blk_remove_plug(q)) {
1681                 conf->seq_flush++;
1682                 raid5_activate_delayed(conf);
1683         }
1684         md_wakeup_thread(mddev->thread);
1685
1686         spin_unlock_irqrestore(&conf->device_lock, flags);
1687
1688         unplug_slaves(mddev);
1689 }
1690
1691 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
1692                              sector_t *error_sector)
1693 {
1694         mddev_t *mddev = q->queuedata;
1695         raid5_conf_t *conf = mddev_to_conf(mddev);
1696         int i, ret = 0;
1697
1698         rcu_read_lock();
1699         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1700                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1701                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
1702                         struct block_device *bdev = rdev->bdev;
1703                         request_queue_t *r_queue = bdev_get_queue(bdev);
1704
1705                         if (!r_queue->issue_flush_fn)
1706                                 ret = -EOPNOTSUPP;
1707                         else {
1708                                 atomic_inc(&rdev->nr_pending);
1709                                 rcu_read_unlock();
1710                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1711                                                               error_sector);
1712                                 rdev_dec_pending(rdev, mddev);
1713                                 rcu_read_lock();
1714                         }
1715                 }
1716         }
1717         rcu_read_unlock();
1718         return ret;
1719 }
1720
1721 static inline void raid5_plug_device(raid5_conf_t *conf)
1722 {
1723         spin_lock_irq(&conf->device_lock);
1724         blk_plug_device(conf->mddev->queue);
1725         spin_unlock_irq(&conf->device_lock);
1726 }
1727
1728 static int make_request(request_queue_t *q, struct bio * bi)
1729 {
1730         mddev_t *mddev = q->queuedata;
1731         raid5_conf_t *conf = mddev_to_conf(mddev);
1732         unsigned int dd_idx, pd_idx;
1733         sector_t new_sector;
1734         sector_t logical_sector, last_sector;
1735         struct stripe_head *sh;
1736         const int rw = bio_data_dir(bi);
1737         int remaining;
1738
1739         if (unlikely(bio_barrier(bi))) {
1740                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1741                 return 0;
1742         }
1743
1744         md_write_start(mddev, bi);
1745
1746         disk_stat_inc(mddev->gendisk, ios[rw]);
1747         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
1748
1749         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1750         last_sector = bi->bi_sector + (bi->bi_size>>9);
1751         bi->bi_next = NULL;
1752         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1753
1754         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1755                 DEFINE_WAIT(w);
1756                 int disks;
1757
1758         retry:
1759                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1760                 if (likely(conf->expand_progress == MaxSector))
1761                         disks = conf->raid_disks;
1762                 else {
1763                         /* spinlock is needed as expand_progress may be
1764                          * 64bit on a 32bit platform, and so it might be
1765                          * possible to see a half-updated value
1766                          * Ofcourse expand_progress could change after
1767                          * the lock is dropped, so once we get a reference
1768                          * to the stripe that we think it is, we will have
1769                          * to check again.
1770                          */
1771                         spin_lock_irq(&conf->device_lock);
1772                         disks = conf->raid_disks;
1773                         if (logical_sector >= conf->expand_progress)
1774                                 disks = conf->previous_raid_disks;
1775                         else {
1776                                 if (logical_sector >= conf->expand_lo) {
1777                                         spin_unlock_irq(&conf->device_lock);
1778                                         schedule();
1779                                         goto retry;
1780                                 }
1781                         }
1782                         spin_unlock_irq(&conf->device_lock);
1783                 }
1784                 new_sector = raid5_compute_sector(logical_sector, disks, disks - 1,
1785                                                   &dd_idx, &pd_idx, conf);
1786                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
1787                         (unsigned long long)new_sector,
1788                         (unsigned long long)logical_sector);
1789
1790                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
1791                 if (sh) {
1792                         if (unlikely(conf->expand_progress != MaxSector)) {
1793                                 /* expansion might have moved on while waiting for a
1794                                  * stripe, so we must do the range check again.
1795                                  * Expansion could still move past after this
1796                                  * test, but as we are holding a reference to
1797                                  * 'sh', we know that if that happens,
1798                                  *  STRIPE_EXPANDING will get set and the expansion
1799                                  * won't proceed until we finish with the stripe.
1800                                  */
1801                                 int must_retry = 0;
1802                                 spin_lock_irq(&conf->device_lock);
1803                                 if (logical_sector <  conf->expand_progress &&
1804                                     disks == conf->previous_raid_disks)
1805                                         /* mismatch, need to try again */
1806                                         must_retry = 1;
1807                                 spin_unlock_irq(&conf->device_lock);
1808                                 if (must_retry) {
1809                                         release_stripe(sh);
1810                                         goto retry;
1811                                 }
1812                         }
1813                         /* FIXME what if we get a false positive because these
1814                          * are being updated.
1815                          */
1816                         if (logical_sector >= mddev->suspend_lo &&
1817                             logical_sector < mddev->suspend_hi) {
1818                                 release_stripe(sh);
1819                                 schedule();
1820                                 goto retry;
1821                         }
1822
1823                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
1824                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1825                                 /* Stripe is busy expanding or
1826                                  * add failed due to overlap.  Flush everything
1827                                  * and wait a while
1828                                  */
1829                                 raid5_unplug_device(mddev->queue);
1830                                 release_stripe(sh);
1831                                 schedule();
1832                                 goto retry;
1833                         }
1834                         finish_wait(&conf->wait_for_overlap, &w);
1835                         raid5_plug_device(conf);
1836                         handle_stripe(sh);
1837                         release_stripe(sh);
1838                 } else {
1839                         /* cannot get stripe for read-ahead, just give-up */
1840                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1841                         finish_wait(&conf->wait_for_overlap, &w);
1842                         break;
1843                 }
1844
1845         }
1846         spin_lock_irq(&conf->device_lock);
1847         remaining = --bi->bi_phys_segments;
1848         spin_unlock_irq(&conf->device_lock);
1849         if (remaining == 0) {
1850                 int bytes = bi->bi_size;
1851
1852                 if ( bio_data_dir(bi) == WRITE )
1853                         md_write_end(mddev);
1854                 bi->bi_size = 0;
1855                 bi->bi_end_io(bi, bytes, 0);
1856         }
1857         return 0;
1858 }
1859
1860 /* FIXME go_faster isn't used */
1861 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1862 {
1863         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1864         struct stripe_head *sh;
1865         int pd_idx;
1866         sector_t first_sector, last_sector;
1867         int raid_disks = conf->raid_disks;
1868         int data_disks = raid_disks-1;
1869         sector_t max_sector = mddev->size << 1;
1870         int sync_blocks;
1871
1872         if (sector_nr >= max_sector) {
1873                 /* just being told to finish up .. nothing much to do */
1874                 unplug_slaves(mddev);
1875                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
1876                         end_reshape(conf);
1877                         return 0;
1878                 }
1879
1880                 if (mddev->curr_resync < max_sector) /* aborted */
1881                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1882                                         &sync_blocks, 1);
1883                 else /* compelted sync */
1884                         conf->fullsync = 0;
1885                 bitmap_close_sync(mddev->bitmap);
1886
1887                 return 0;
1888         }
1889
1890         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
1891                 /* reshaping is quite different to recovery/resync so it is
1892                  * handled quite separately ... here.
1893                  *
1894                  * On each call to sync_request, we gather one chunk worth of
1895                  * destination stripes and flag them as expanding.
1896                  * Then we find all the source stripes and request reads.
1897                  * As the reads complete, handle_stripe will copy the data
1898                  * into the destination stripe and release that stripe.
1899                  */
1900                 int i;
1901                 int dd_idx;
1902                 sector_t writepos, safepos, gap;
1903
1904                 if (sector_nr == 0 &&
1905                     conf->expand_progress != 0) {
1906                         /* restarting in the middle, skip the initial sectors */
1907                         sector_nr = conf->expand_progress;
1908                         sector_div(sector_nr, conf->raid_disks-1);
1909                         *skipped = 1;
1910                         return sector_nr;
1911                 }
1912
1913                 /* we update the metadata when there is more than 3Meg
1914                  * in the block range (that is rather arbitrary, should
1915                  * probably be time based) or when the data about to be
1916                  * copied would over-write the source of the data at
1917                  * the front of the range.
1918                  * i.e. one new_stripe forward from expand_progress new_maps
1919                  * to after where expand_lo old_maps to
1920                  */
1921                 writepos = conf->expand_progress +
1922                         conf->chunk_size/512*(conf->raid_disks-1);
1923                 sector_div(writepos, conf->raid_disks-1);
1924                 safepos = conf->expand_lo;
1925                 sector_div(safepos, conf->previous_raid_disks-1);
1926                 gap = conf->expand_progress - conf->expand_lo;
1927
1928                 if (writepos >= safepos ||
1929                     gap > (conf->raid_disks-1)*3000*2 /*3Meg*/) {
1930                         /* Cannot proceed until we've updated the superblock... */
1931                         wait_event(conf->wait_for_overlap,
1932                                    atomic_read(&conf->reshape_stripes)==0);
1933                         mddev->reshape_position = conf->expand_progress;
1934                         mddev->sb_dirty = 1;
1935                         md_wakeup_thread(mddev->thread);
1936                         wait_event(mddev->sb_wait, mddev->sb_dirty == 0 ||
1937                                    kthread_should_stop());
1938                         spin_lock_irq(&conf->device_lock);
1939                         conf->expand_lo = mddev->reshape_position;
1940                         spin_unlock_irq(&conf->device_lock);
1941                         wake_up(&conf->wait_for_overlap);
1942                 }
1943
1944                 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
1945                         int j;
1946                         int skipped = 0;
1947                         pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
1948                         sh = get_active_stripe(conf, sector_nr+i,
1949                                                conf->raid_disks, pd_idx, 0);
1950                         set_bit(STRIPE_EXPANDING, &sh->state);
1951                         atomic_inc(&conf->reshape_stripes);
1952                         /* If any of this stripe is beyond the end of the old
1953                          * array, then we need to zero those blocks
1954                          */
1955                         for (j=sh->disks; j--;) {
1956                                 sector_t s;
1957                                 if (j == sh->pd_idx)
1958                                         continue;
1959                                 s = compute_blocknr(sh, j);
1960                                 if (s < (mddev->array_size<<1)) {
1961                                         skipped = 1;
1962                                         continue;
1963                                 }
1964                                 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
1965                                 set_bit(R5_Expanded, &sh->dev[j].flags);
1966                                 set_bit(R5_UPTODATE, &sh->dev[j].flags);
1967                         }
1968                         if (!skipped) {
1969                                 set_bit(STRIPE_EXPAND_READY, &sh->state);
1970                                 set_bit(STRIPE_HANDLE, &sh->state);
1971                         }
1972                         release_stripe(sh);
1973                 }
1974                 spin_lock_irq(&conf->device_lock);
1975                 conf->expand_progress = (sector_nr + i)*(conf->raid_disks-1);
1976                 spin_unlock_irq(&conf->device_lock);
1977                 /* Ok, those stripe are ready. We can start scheduling
1978                  * reads on the source stripes.
1979                  * The source stripes are determined by mapping the first and last
1980                  * block on the destination stripes.
1981                  */
1982                 raid_disks = conf->previous_raid_disks;
1983                 data_disks = raid_disks - 1;
1984                 first_sector =
1985                         raid5_compute_sector(sector_nr*(conf->raid_disks-1),
1986                                              raid_disks, data_disks,
1987                                              &dd_idx, &pd_idx, conf);
1988                 last_sector =
1989                         raid5_compute_sector((sector_nr+conf->chunk_size/512)
1990                                                *(conf->raid_disks-1) -1,
1991                                              raid_disks, data_disks,
1992                                              &dd_idx, &pd_idx, conf);
1993                 if (last_sector >= (mddev->size<<1))
1994                         last_sector = (mddev->size<<1)-1;
1995                 while (first_sector <= last_sector) {
1996                         pd_idx = stripe_to_pdidx(first_sector, conf, conf->previous_raid_disks);
1997                         sh = get_active_stripe(conf, first_sector,
1998                                                conf->previous_raid_disks, pd_idx, 0);
1999                         set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2000                         set_bit(STRIPE_HANDLE, &sh->state);
2001                         release_stripe(sh);
2002                         first_sector += STRIPE_SECTORS;
2003                 }
2004                 return conf->chunk_size>>9;
2005         }
2006         /* if there is 1 or more failed drives and we are trying
2007          * to resync, then assert that we are finished, because there is
2008          * nothing we can do.
2009          */
2010         if (mddev->degraded >= 1 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2011                 sector_t rv = (mddev->size << 1) - sector_nr;
2012                 *skipped = 1;
2013                 return rv;
2014         }
2015         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2016             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2017             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
2018                 /* we can skip this block, and probably more */
2019                 sync_blocks /= STRIPE_SECTORS;
2020                 *skipped = 1;
2021                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
2022         }
2023
2024         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
2025         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
2026         if (sh == NULL) {
2027                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
2028                 /* make sure we don't swamp the stripe cache if someone else
2029                  * is trying to get access
2030                  */
2031                 schedule_timeout_uninterruptible(1);
2032         }
2033         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 0);
2034         spin_lock(&sh->lock);
2035         set_bit(STRIPE_SYNCING, &sh->state);
2036         clear_bit(STRIPE_INSYNC, &sh->state);
2037         spin_unlock(&sh->lock);
2038
2039         handle_stripe(sh);
2040         release_stripe(sh);
2041
2042         return STRIPE_SECTORS;
2043 }
2044
2045 /*
2046  * This is our raid5 kernel thread.
2047  *
2048  * We scan the hash table for stripes which can be handled now.
2049  * During the scan, completed stripes are saved for us by the interrupt
2050  * handler, so that they will not have to wait for our next wakeup.
2051  */
2052 static void raid5d (mddev_t *mddev)
2053 {
2054         struct stripe_head *sh;
2055         raid5_conf_t *conf = mddev_to_conf(mddev);
2056         int handled;
2057
2058         PRINTK("+++ raid5d active\n");
2059
2060         md_check_recovery(mddev);
2061
2062         handled = 0;
2063         spin_lock_irq(&conf->device_lock);
2064         while (1) {
2065                 struct list_head *first;
2066
2067                 if (conf->seq_flush - conf->seq_write > 0) {
2068                         int seq = conf->seq_flush;
2069                         spin_unlock_irq(&conf->device_lock);
2070                         bitmap_unplug(mddev->bitmap);
2071                         spin_lock_irq(&conf->device_lock);
2072                         conf->seq_write = seq;
2073                         activate_bit_delay(conf);
2074                 }
2075
2076                 if (list_empty(&conf->handle_list) &&
2077                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
2078                     !blk_queue_plugged(mddev->queue) &&
2079                     !list_empty(&conf->delayed_list))
2080                         raid5_activate_delayed(conf);
2081
2082                 if (list_empty(&conf->handle_list))
2083                         break;
2084
2085                 first = conf->handle_list.next;
2086                 sh = list_entry(first, struct stripe_head, lru);
2087
2088                 list_del_init(first);
2089                 atomic_inc(&sh->count);
2090                 BUG_ON(atomic_read(&sh->count)!= 1);
2091                 spin_unlock_irq(&conf->device_lock);
2092
2093                 handled++;
2094                 handle_stripe(sh);
2095                 release_stripe(sh);
2096
2097                 spin_lock_irq(&conf->device_lock);
2098         }
2099         PRINTK("%d stripes handled\n", handled);
2100
2101         spin_unlock_irq(&conf->device_lock);
2102
2103         unplug_slaves(mddev);
2104
2105         PRINTK("--- raid5d inactive\n");
2106 }
2107
2108 static ssize_t
2109 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
2110 {
2111         raid5_conf_t *conf = mddev_to_conf(mddev);
2112         if (conf)
2113                 return sprintf(page, "%d\n", conf->max_nr_stripes);
2114         else
2115                 return 0;
2116 }
2117
2118 static ssize_t
2119 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
2120 {
2121         raid5_conf_t *conf = mddev_to_conf(mddev);
2122         char *end;
2123         int new;
2124         if (len >= PAGE_SIZE)
2125                 return -EINVAL;
2126         if (!conf)
2127                 return -ENODEV;
2128
2129         new = simple_strtoul(page, &end, 10);
2130         if (!*page || (*end && *end != '\n') )
2131                 return -EINVAL;
2132         if (new <= 16 || new > 32768)
2133                 return -EINVAL;
2134         while (new < conf->max_nr_stripes) {
2135                 if (drop_one_stripe(conf))
2136                         conf->max_nr_stripes--;
2137                 else
2138                         break;
2139         }
2140         while (new > conf->max_nr_stripes) {
2141                 if (grow_one_stripe(conf))
2142                         conf->max_nr_stripes++;
2143                 else break;
2144         }
2145         return len;
2146 }
2147
2148 static struct md_sysfs_entry
2149 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
2150                                 raid5_show_stripe_cache_size,
2151                                 raid5_store_stripe_cache_size);
2152
2153 static ssize_t
2154 stripe_cache_active_show(mddev_t *mddev, char *page)
2155 {
2156         raid5_conf_t *conf = mddev_to_conf(mddev);
2157         if (conf)
2158                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
2159         else
2160                 return 0;
2161 }
2162
2163 static struct md_sysfs_entry
2164 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
2165
2166 static struct attribute *raid5_attrs[] =  {
2167         &raid5_stripecache_size.attr,
2168         &raid5_stripecache_active.attr,
2169         NULL,
2170 };
2171 static struct attribute_group raid5_attrs_group = {
2172         .name = NULL,
2173         .attrs = raid5_attrs,
2174 };
2175
2176 static int run(mddev_t *mddev)
2177 {
2178         raid5_conf_t *conf;
2179         int raid_disk, memory;
2180         mdk_rdev_t *rdev;
2181         struct disk_info *disk;
2182         struct list_head *tmp;
2183
2184         if (mddev->level != 5 && mddev->level != 4) {
2185                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5 (%d)\n",
2186                        mdname(mddev), mddev->level);
2187                 return -EIO;
2188         }
2189
2190         if (mddev->reshape_position != MaxSector) {
2191                 /* Check that we can continue the reshape.
2192                  * Currently only disks can change, it must
2193                  * increase, and we must be past the point where
2194                  * a stripe over-writes itself
2195                  */
2196                 sector_t here_new, here_old;
2197                 int old_disks;
2198
2199                 if (mddev->new_level != mddev->level ||
2200                     mddev->new_layout != mddev->layout ||
2201                     mddev->new_chunk != mddev->chunk_size) {
2202                         printk(KERN_ERR "raid5: %s: unsupported reshape required - aborting.\n",
2203                                mdname(mddev));
2204                         return -EINVAL;
2205                 }
2206                 if (mddev->delta_disks <= 0) {
2207                         printk(KERN_ERR "raid5: %s: unsupported reshape (reduce disks) required - aborting.\n",
2208                                mdname(mddev));
2209                         return -EINVAL;
2210                 }
2211                 old_disks = mddev->raid_disks - mddev->delta_disks;
2212                 /* reshape_position must be on a new-stripe boundary, and one
2213                  * further up in new geometry must map after here in old geometry.
2214                  */
2215                 here_new = mddev->reshape_position;
2216                 if (sector_div(here_new, (mddev->chunk_size>>9)*(mddev->raid_disks-1))) {
2217                         printk(KERN_ERR "raid5: reshape_position not on a stripe boundary\n");
2218                         return -EINVAL;
2219                 }
2220                 /* here_new is the stripe we will write to */
2221                 here_old = mddev->reshape_position;
2222                 sector_div(here_old, (mddev->chunk_size>>9)*(old_disks-1));
2223                 /* here_old is the first stripe that we might need to read from */
2224                 if (here_new >= here_old) {
2225                         /* Reading from the same stripe as writing to - bad */
2226                         printk(KERN_ERR "raid5: reshape_position too early for auto-recovery - aborting.\n");
2227                         return -EINVAL;
2228                 }
2229                 printk(KERN_INFO "raid5: reshape will continue\n");
2230                 /* OK, we should be able to continue; */
2231         }
2232
2233
2234         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
2235         if ((conf = mddev->private) == NULL)
2236                 goto abort;
2237         if (mddev->reshape_position == MaxSector) {
2238                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
2239         } else {
2240                 conf->raid_disks = mddev->raid_disks;
2241                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
2242         }
2243
2244         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
2245                               GFP_KERNEL);
2246         if (!conf->disks)
2247                 goto abort;
2248
2249         conf->mddev = mddev;
2250
2251         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
2252                 goto abort;
2253
2254         spin_lock_init(&conf->device_lock);
2255         init_waitqueue_head(&conf->wait_for_stripe);
2256         init_waitqueue_head(&conf->wait_for_overlap);
2257         INIT_LIST_HEAD(&conf->handle_list);
2258         INIT_LIST_HEAD(&conf->delayed_list);
2259         INIT_LIST_HEAD(&conf->bitmap_list);
2260         INIT_LIST_HEAD(&conf->inactive_list);
2261         atomic_set(&conf->active_stripes, 0);
2262         atomic_set(&conf->preread_active_stripes, 0);
2263
2264         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
2265
2266         ITERATE_RDEV(mddev,rdev,tmp) {
2267                 raid_disk = rdev->raid_disk;
2268                 if (raid_disk >= conf->raid_disks
2269                     || raid_disk < 0)
2270                         continue;
2271                 disk = conf->disks + raid_disk;
2272
2273                 disk->rdev = rdev;
2274
2275                 if (test_bit(In_sync, &rdev->flags)) {
2276                         char b[BDEVNAME_SIZE];
2277                         printk(KERN_INFO "raid5: device %s operational as raid"
2278                                 " disk %d\n", bdevname(rdev->bdev,b),
2279                                 raid_disk);
2280                         conf->working_disks++;
2281                 }
2282         }
2283
2284         /*
2285          * 0 for a fully functional array, 1 for a degraded array.
2286          */
2287         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
2288         conf->mddev = mddev;
2289         conf->chunk_size = mddev->chunk_size;
2290         conf->level = mddev->level;
2291         conf->algorithm = mddev->layout;
2292         conf->max_nr_stripes = NR_STRIPES;
2293         conf->expand_progress = mddev->reshape_position;
2294
2295         /* device size must be a multiple of chunk size */
2296         mddev->size &= ~(mddev->chunk_size/1024 -1);
2297         mddev->resync_max_sectors = mddev->size << 1;
2298
2299         if (!conf->chunk_size || conf->chunk_size % 4) {
2300                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
2301                         conf->chunk_size, mdname(mddev));
2302                 goto abort;
2303         }
2304         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
2305                 printk(KERN_ERR
2306                         "raid5: unsupported parity algorithm %d for %s\n",
2307                         conf->algorithm, mdname(mddev));
2308                 goto abort;
2309         }
2310         if (mddev->degraded > 1) {
2311                 printk(KERN_ERR "raid5: not enough operational devices for %s"
2312                         " (%d/%d failed)\n",
2313                         mdname(mddev), conf->failed_disks, conf->raid_disks);
2314                 goto abort;
2315         }
2316
2317         if (mddev->degraded == 1 &&
2318             mddev->recovery_cp != MaxSector) {
2319                 if (mddev->ok_start_degraded)
2320                         printk(KERN_WARNING
2321                                "raid5: starting dirty degraded array: %s"
2322                                "- data corruption possible.\n",
2323                                mdname(mddev));
2324                 else {
2325                         printk(KERN_ERR
2326                                "raid5: cannot start dirty degraded array for %s\n",
2327                                mdname(mddev));
2328                         goto abort;
2329                 }
2330         }
2331
2332         {
2333                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
2334                 if (!mddev->thread) {
2335                         printk(KERN_ERR
2336                                 "raid5: couldn't allocate thread for %s\n",
2337                                 mdname(mddev));
2338                         goto abort;
2339                 }
2340         }
2341         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
2342                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
2343         if (grow_stripes(conf, conf->max_nr_stripes)) {
2344                 printk(KERN_ERR
2345                         "raid5: couldn't allocate %dkB for buffers\n", memory);
2346                 shrink_stripes(conf);
2347                 md_unregister_thread(mddev->thread);
2348                 goto abort;
2349         } else
2350                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
2351                         memory, mdname(mddev));
2352
2353         if (mddev->degraded == 0)
2354                 printk("raid5: raid level %d set %s active with %d out of %d"
2355                         " devices, algorithm %d\n", conf->level, mdname(mddev),
2356                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
2357                         conf->algorithm);
2358         else
2359                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
2360                         " out of %d devices, algorithm %d\n", conf->level,
2361                         mdname(mddev), mddev->raid_disks - mddev->degraded,
2362                         mddev->raid_disks, conf->algorithm);
2363
2364         print_raid5_conf(conf);
2365
2366         if (conf->expand_progress != MaxSector) {
2367                 printk("...ok start reshape thread\n");
2368                 conf->expand_lo = conf->expand_progress;
2369                 atomic_set(&conf->reshape_stripes, 0);
2370                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
2371                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
2372                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
2373                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
2374                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
2375                                                         "%s_reshape");
2376                 /* FIXME if md_register_thread fails?? */
2377                 md_wakeup_thread(mddev->sync_thread);
2378
2379         }
2380
2381         /* read-ahead size must cover two whole stripes, which is
2382          * 2 * (n-1) * chunksize where 'n' is the number of raid devices
2383          */
2384         {
2385                 int stripe = (mddev->raid_disks-1) * mddev->chunk_size
2386                         / PAGE_SIZE;
2387                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
2388                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
2389         }
2390
2391         /* Ok, everything is just fine now */
2392         sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
2393
2394         mddev->queue->unplug_fn = raid5_unplug_device;
2395         mddev->queue->issue_flush_fn = raid5_issue_flush;
2396         mddev->array_size =  mddev->size * (conf->previous_raid_disks - 1);
2397
2398         return 0;
2399 abort:
2400         if (conf) {
2401                 print_raid5_conf(conf);
2402                 kfree(conf->disks);
2403                 kfree(conf->stripe_hashtbl);
2404                 kfree(conf);
2405         }
2406         mddev->private = NULL;
2407         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
2408         return -EIO;
2409 }
2410
2411
2412
2413 static int stop(mddev_t *mddev)
2414 {
2415         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2416
2417         md_unregister_thread(mddev->thread);
2418         mddev->thread = NULL;
2419         shrink_stripes(conf);
2420         kfree(conf->stripe_hashtbl);
2421         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2422         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
2423         kfree(conf->disks);
2424         kfree(conf);
2425         mddev->private = NULL;
2426         return 0;
2427 }
2428
2429 #if RAID5_DEBUG
2430 static void print_sh (struct stripe_head *sh)
2431 {
2432         int i;
2433
2434         printk("sh %llu, pd_idx %d, state %ld.\n",
2435                 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
2436         printk("sh %llu,  count %d.\n",
2437                 (unsigned long long)sh->sector, atomic_read(&sh->count));
2438         printk("sh %llu, ", (unsigned long long)sh->sector);
2439         for (i = 0; i < sh->disks; i++) {
2440                 printk("(cache%d: %p %ld) ",
2441                         i, sh->dev[i].page, sh->dev[i].flags);
2442         }
2443         printk("\n");
2444 }
2445
2446 static void printall (raid5_conf_t *conf)
2447 {
2448         struct stripe_head *sh;
2449         struct hlist_node *hn;
2450         int i;
2451
2452         spin_lock_irq(&conf->device_lock);
2453         for (i = 0; i < NR_HASH; i++) {
2454                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
2455                         if (sh->raid_conf != conf)
2456                                 continue;
2457                         print_sh(sh);
2458                 }
2459         }
2460         spin_unlock_irq(&conf->device_lock);
2461 }
2462 #endif
2463
2464 static void status (struct seq_file *seq, mddev_t *mddev)
2465 {
2466         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2467         int i;
2468
2469         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2470         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2471         for (i = 0; i < conf->raid_disks; i++)
2472                 seq_printf (seq, "%s",
2473                                conf->disks[i].rdev &&
2474                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
2475         seq_printf (seq, "]");
2476 #if RAID5_DEBUG
2477 #define D(x) \
2478         seq_printf (seq, "<"#x":%d>", atomic_read(&conf->x))
2479         printall(conf);
2480 #endif
2481 }
2482
2483 static void print_raid5_conf (raid5_conf_t *conf)
2484 {
2485         int i;
2486         struct disk_info *tmp;
2487
2488         printk("RAID5 conf printout:\n");
2489         if (!conf) {
2490                 printk("(conf==NULL)\n");
2491                 return;
2492         }
2493         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2494                  conf->working_disks, conf->failed_disks);
2495
2496         for (i = 0; i < conf->raid_disks; i++) {
2497                 char b[BDEVNAME_SIZE];
2498                 tmp = conf->disks + i;
2499                 if (tmp->rdev)
2500                 printk(" disk %d, o:%d, dev:%s\n",
2501                         i, !test_bit(Faulty, &tmp->rdev->flags),
2502                         bdevname(tmp->rdev->bdev,b));
2503         }
2504 }
2505
2506 static int raid5_spare_active(mddev_t *mddev)
2507 {
2508         int i;
2509         raid5_conf_t *conf = mddev->private;
2510         struct disk_info *tmp;
2511
2512         for (i = 0; i < conf->raid_disks; i++) {
2513                 tmp = conf->disks + i;
2514                 if (tmp->rdev
2515                     && !test_bit(Faulty, &tmp->rdev->flags)
2516                     && !test_bit(In_sync, &tmp->rdev->flags)) {
2517                         mddev->degraded--;
2518                         conf->failed_disks--;
2519                         conf->working_disks++;
2520                         set_bit(In_sync, &tmp->rdev->flags);
2521                 }
2522         }
2523         print_raid5_conf(conf);
2524         return 0;
2525 }
2526
2527 static int raid5_remove_disk(mddev_t *mddev, int number)
2528 {
2529         raid5_conf_t *conf = mddev->private;
2530         int err = 0;
2531         mdk_rdev_t *rdev;
2532         struct disk_info *p = conf->disks + number;
2533
2534         print_raid5_conf(conf);
2535         rdev = p->rdev;
2536         if (rdev) {
2537                 if (test_bit(In_sync, &rdev->flags) ||
2538                     atomic_read(&rdev->nr_pending)) {
2539                         err = -EBUSY;
2540                         goto abort;
2541                 }
2542                 p->rdev = NULL;
2543                 synchronize_rcu();
2544                 if (atomic_read(&rdev->nr_pending)) {
2545                         /* lost the race, try later */
2546                         err = -EBUSY;
2547                         p->rdev = rdev;
2548                 }
2549         }
2550 abort:
2551
2552         print_raid5_conf(conf);
2553         return err;
2554 }
2555
2556 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2557 {
2558         raid5_conf_t *conf = mddev->private;
2559         int found = 0;
2560         int disk;
2561         struct disk_info *p;
2562
2563         if (mddev->degraded > 1)
2564                 /* no point adding a device */
2565                 return 0;
2566
2567         /*
2568          * find the disk ...
2569          */
2570         for (disk=0; disk < conf->raid_disks; disk++)
2571                 if ((p=conf->disks + disk)->rdev == NULL) {
2572                         clear_bit(In_sync, &rdev->flags);
2573                         rdev->raid_disk = disk;
2574                         found = 1;
2575                         if (rdev->saved_raid_disk != disk)
2576                                 conf->fullsync = 1;
2577                         rcu_assign_pointer(p->rdev, rdev);
2578                         break;
2579                 }
2580         print_raid5_conf(conf);
2581         return found;
2582 }
2583
2584 static int raid5_resize(mddev_t *mddev, sector_t sectors)
2585 {
2586         /* no resync is happening, and there is enough space
2587          * on all devices, so we can resize.
2588          * We need to make sure resync covers any new space.
2589          * If the array is shrinking we should possibly wait until
2590          * any io in the removed space completes, but it hardly seems
2591          * worth it.
2592          */
2593         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2594         mddev->array_size = (sectors * (mddev->raid_disks-1))>>1;
2595         set_capacity(mddev->gendisk, mddev->array_size << 1);
2596         mddev->changed = 1;
2597         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2598                 mddev->recovery_cp = mddev->size << 1;
2599                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2600         }
2601         mddev->size = sectors /2;
2602         mddev->resync_max_sectors = sectors;
2603         return 0;
2604 }
2605
2606 #ifdef CONFIG_MD_RAID5_RESHAPE
2607 static int raid5_check_reshape(mddev_t *mddev)
2608 {
2609         raid5_conf_t *conf = mddev_to_conf(mddev);
2610         int err;
2611
2612         if (mddev->delta_disks < 0 ||
2613             mddev->new_level != mddev->level)
2614                 return -EINVAL; /* Cannot shrink array or change level yet */
2615         if (mddev->delta_disks == 0)
2616                 return 0; /* nothing to do */
2617
2618         /* Can only proceed if there are plenty of stripe_heads.
2619          * We need a minimum of one full stripe,, and for sensible progress
2620          * it is best to have about 4 times that.
2621          * If we require 4 times, then the default 256 4K stripe_heads will
2622          * allow for chunk sizes up to 256K, which is probably OK.
2623          * If the chunk size is greater, user-space should request more
2624          * stripe_heads first.
2625          */
2626         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
2627             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
2628                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
2629                        (mddev->chunk_size / STRIPE_SIZE)*4);
2630                 return -ENOSPC;
2631         }
2632
2633         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
2634         if (err)
2635                 return err;
2636
2637         /* looks like we might be able to manage this */
2638         return 0;
2639 }
2640
2641 static int raid5_start_reshape(mddev_t *mddev)
2642 {
2643         raid5_conf_t *conf = mddev_to_conf(mddev);
2644         mdk_rdev_t *rdev;
2645         struct list_head *rtmp;
2646         int spares = 0;
2647         int added_devices = 0;
2648
2649         if (mddev->degraded ||
2650             test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
2651                 return -EBUSY;
2652
2653         ITERATE_RDEV(mddev, rdev, rtmp)
2654                 if (rdev->raid_disk < 0 &&
2655                     !test_bit(Faulty, &rdev->flags))
2656                         spares++;
2657
2658         if (spares < mddev->delta_disks-1)
2659                 /* Not enough devices even to make a degraded array
2660                  * of that size
2661                  */
2662                 return -EINVAL;
2663
2664         atomic_set(&conf->reshape_stripes, 0);
2665         spin_lock_irq(&conf->device_lock);
2666         conf->previous_raid_disks = conf->raid_disks;
2667         conf->raid_disks += mddev->delta_disks;
2668         conf->expand_progress = 0;
2669         conf->expand_lo = 0;
2670         spin_unlock_irq(&conf->device_lock);
2671
2672         /* Add some new drives, as many as will fit.
2673          * We know there are enough to make the newly sized array work.
2674          */
2675         ITERATE_RDEV(mddev, rdev, rtmp)
2676                 if (rdev->raid_disk < 0 &&
2677                     !test_bit(Faulty, &rdev->flags)) {
2678                         if (raid5_add_disk(mddev, rdev)) {
2679                                 char nm[20];
2680                                 set_bit(In_sync, &rdev->flags);
2681                                 conf->working_disks++;
2682                                 added_devices++;
2683                                 sprintf(nm, "rd%d", rdev->raid_disk);
2684                                 sysfs_create_link(&mddev->kobj, &rdev->kobj, nm);
2685                         } else
2686                                 break;
2687                 }
2688
2689         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
2690         mddev->raid_disks = conf->raid_disks;
2691         mddev->reshape_position = 0;
2692         mddev->sb_dirty = 1;
2693
2694         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
2695         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
2696         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
2697         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
2698         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
2699                                                 "%s_reshape");
2700         if (!mddev->sync_thread) {
2701                 mddev->recovery = 0;
2702                 spin_lock_irq(&conf->device_lock);
2703                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
2704                 conf->expand_progress = MaxSector;
2705                 spin_unlock_irq(&conf->device_lock);
2706                 return -EAGAIN;
2707         }
2708         md_wakeup_thread(mddev->sync_thread);
2709         md_new_event(mddev);
2710         return 0;
2711 }
2712 #endif
2713
2714 static void end_reshape(raid5_conf_t *conf)
2715 {
2716         struct block_device *bdev;
2717
2718         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
2719                 conf->mddev->array_size = conf->mddev->size * (conf->raid_disks-1);
2720                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
2721                 conf->mddev->changed = 1;
2722
2723                 bdev = bdget_disk(conf->mddev->gendisk, 0);
2724                 if (bdev) {
2725                         mutex_lock(&bdev->bd_inode->i_mutex);
2726                         i_size_write(bdev->bd_inode, conf->mddev->array_size << 10);
2727                         mutex_unlock(&bdev->bd_inode->i_mutex);
2728                         bdput(bdev);
2729                 }
2730                 spin_lock_irq(&conf->device_lock);
2731                 conf->expand_progress = MaxSector;
2732                 spin_unlock_irq(&conf->device_lock);
2733                 conf->mddev->reshape_position = MaxSector;
2734         }
2735 }
2736
2737 static void raid5_quiesce(mddev_t *mddev, int state)
2738 {
2739         raid5_conf_t *conf = mddev_to_conf(mddev);
2740
2741         switch(state) {
2742         case 2: /* resume for a suspend */
2743                 wake_up(&conf->wait_for_overlap);
2744                 break;
2745
2746         case 1: /* stop all writes */
2747                 spin_lock_irq(&conf->device_lock);
2748                 conf->quiesce = 1;
2749                 wait_event_lock_irq(conf->wait_for_stripe,
2750                                     atomic_read(&conf->active_stripes) == 0,
2751                                     conf->device_lock, /* nothing */);
2752                 spin_unlock_irq(&conf->device_lock);
2753                 break;
2754
2755         case 0: /* re-enable writes */
2756                 spin_lock_irq(&conf->device_lock);
2757                 conf->quiesce = 0;
2758                 wake_up(&conf->wait_for_stripe);
2759                 wake_up(&conf->wait_for_overlap);
2760                 spin_unlock_irq(&conf->device_lock);
2761                 break;
2762         }
2763 }
2764
2765 static struct mdk_personality raid5_personality =
2766 {
2767         .name           = "raid5",
2768         .level          = 5,
2769         .owner          = THIS_MODULE,
2770         .make_request   = make_request,
2771         .run            = run,
2772         .stop           = stop,
2773         .status         = status,
2774         .error_handler  = error,
2775         .hot_add_disk   = raid5_add_disk,
2776         .hot_remove_disk= raid5_remove_disk,
2777         .spare_active   = raid5_spare_active,
2778         .sync_request   = sync_request,
2779         .resize         = raid5_resize,
2780 #ifdef CONFIG_MD_RAID5_RESHAPE
2781         .check_reshape  = raid5_check_reshape,
2782         .start_reshape  = raid5_start_reshape,
2783 #endif
2784         .quiesce        = raid5_quiesce,
2785 };
2786
2787 static struct mdk_personality raid4_personality =
2788 {
2789         .name           = "raid4",
2790         .level          = 4,
2791         .owner          = THIS_MODULE,
2792         .make_request   = make_request,
2793         .run            = run,
2794         .stop           = stop,
2795         .status         = status,
2796         .error_handler  = error,
2797         .hot_add_disk   = raid5_add_disk,
2798         .hot_remove_disk= raid5_remove_disk,
2799         .spare_active   = raid5_spare_active,
2800         .sync_request   = sync_request,
2801         .resize         = raid5_resize,
2802         .quiesce        = raid5_quiesce,
2803 };
2804
2805 static int __init raid5_init(void)
2806 {
2807         register_md_personality(&raid5_personality);
2808         register_md_personality(&raid4_personality);
2809         return 0;
2810 }
2811
2812 static void raid5_exit(void)
2813 {
2814         unregister_md_personality(&raid5_personality);
2815         unregister_md_personality(&raid4_personality);
2816 }
2817
2818 module_init(raid5_init);
2819 module_exit(raid5_exit);
2820 MODULE_LICENSE("GPL");
2821 MODULE_ALIAS("md-personality-4"); /* RAID5 */
2822 MODULE_ALIAS("md-raid5");
2823 MODULE_ALIAS("md-raid4");
2824 MODULE_ALIAS("md-level-5");
2825 MODULE_ALIAS("md-level-4");