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                 if (!list_empty(&sh->lru))
  77                         BUG();
  78                 if (atomic_read(&conf->active_stripes)==0)
  79                         BUG();
  80                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
  81                         if (test_bit(STRIPE_DELAYED, &sh->state))
  82                                 list_add_tail(&sh->lru, &conf->delayed_list);
  83                         else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
  84                                  conf->seq_write == sh->bm_seq)
  85                                 list_add_tail(&sh->lru, &conf->bitmap_list);
  86                         else {
  87                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
  88                                 list_add_tail(&sh->lru, &conf->handle_list);
  89                         }
  90                         md_wakeup_thread(conf->mddev->thread);
  91                 } else {
  92                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
  93                                 atomic_dec(&conf->preread_active_stripes);
  94                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
  95                                         md_wakeup_thread(conf->mddev->thread);
  96                         }
  97                         atomic_dec(&conf->active_stripes);
  98                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
  99                                 list_add_tail(&sh->lru, &conf->inactive_list);
 100                                 wake_up(&conf->wait_for_stripe);
 101                         }
 102                 }
 103         }
 104 }
 105 static void release_stripe(struct stripe_head *sh)
 106 {
 107         raid5_conf_t *conf = sh->raid_conf;
 108         unsigned long flags;
 109
 110         spin_lock_irqsave(&conf->device_lock, flags);
 111         __release_stripe(conf, sh);
 112         spin_unlock_irqrestore(&conf->device_lock, flags);
 113 }
 114
 115 static inline void remove_hash(struct stripe_head *sh)
 116 {
 117         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 118
 119         hlist_del_init(&sh->hash);
 120 }
 121
 122 static void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
 123 {
 124         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 125
 126         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 127
 128         CHECK_DEVLOCK();
 129         hlist_add_head(&sh->hash, hp);
 130 }
 131
 132
 133 /* find an idle stripe, make sure it is unhashed, and return it. */
 134 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
 135 {
 136         struct stripe_head *sh = NULL;
 137         struct list_head *first;
 138
 139         CHECK_DEVLOCK();
 140         if (list_empty(&conf->inactive_list))
 141                 goto out;
 142         first = conf->inactive_list.next;
 143         sh = list_entry(first, struct stripe_head, lru);
 144         list_del_init(first);
 145         remove_hash(sh);
 146         atomic_inc(&conf->active_stripes);
 147 out:
 148         return sh;
 149 }
 150
 151 static void shrink_buffers(struct stripe_head *sh, int num)
 152 {
 153         struct page *p;
 154         int i;
 155
 156         for (i=0; i<num ; i++) {
 157                 p = sh->dev[i].page;
 158                 if (!p)
 159                         continue;
 160                 sh->dev[i].page = NULL;
 161                 put_page(p);
 162         }
 163 }
 164
 165 static int grow_buffers(struct stripe_head *sh, int num)
 166 {
 167         int i;
 168
 169         for (i=0; i<num; i++) {
 170                 struct page *page;
 171
 172                 if (!(page = alloc_page(GFP_KERNEL))) {
 173                         return 1;
 174                 }
 175                 sh->dev[i].page = page;
 176         }
 177         return 0;
 178 }
 179
 180 static void raid5_build_block (struct stripe_head *sh, int i);
 181
 182 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
 183 {
 184         raid5_conf_t *conf = sh->raid_conf;
 185         int i;
 186
 187         if (atomic_read(&sh->count) != 0)
 188                 BUG();
 189         if (test_bit(STRIPE_HANDLE, &sh->state))
 190                 BUG();
 191
 192         CHECK_DEVLOCK();
 193         PRINTK("init_stripe called, stripe %llu\n",
 194                 (unsigned long long)sh->sector);
 195
 196         remove_hash(sh);
 197
 198         sh->sector = sector;
 199         sh->pd_idx = pd_idx;
 200         sh->state = 0;
 201
 202         sh->disks = disks;
 203
 204         for (i = sh->disks; i--; ) {
 205                 struct r5dev *dev = &sh->dev[i];
 206
 207                 if (dev->toread || dev->towrite || dev->written ||
 208                     test_bit(R5_LOCKED, &dev->flags)) {
 209                         printk("sector=%llx i=%d %p %p %p %d\n",
 210                                (unsigned long long)sh->sector, i, dev->toread,
 211                                dev->towrite, dev->written,
 212                                test_bit(R5_LOCKED, &dev->flags));
 213                         BUG();
 214                 }
 215                 dev->flags = 0;
 216                 raid5_build_block(sh, i);
 217         }
 218         insert_hash(conf, sh);
 219 }
 220
 221 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
 222 {
 223         struct stripe_head *sh;
 224         struct hlist_node *hn;
 225
 226         CHECK_DEVLOCK();
 227         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
 228         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
 229                 if (sh->sector == sector && sh->disks == disks)
 230                         return sh;
 231         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
 232         return NULL;
 233 }
 234
 235 static void unplug_slaves(mddev_t *mddev);
 236 static void raid5_unplug_device(request_queue_t *q);
 237
 238 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
 239                                              int pd_idx, int noblock)
 240 {
 241         struct stripe_head *sh;
 242
 243         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
 244
 245         spin_lock_irq(&conf->device_lock);
 246
 247         do {
 248                 wait_event_lock_irq(conf->wait_for_stripe,
 249                                     conf->quiesce == 0,
 250                                     conf->device_lock, /* nothing */);
 251                 sh = __find_stripe(conf, sector, disks);
 252                 if (!sh) {
 253                         if (!conf->inactive_blocked)
 254                                 sh = get_free_stripe(conf);
 255                         if (noblock && sh == NULL)
 256                                 break;
 257                         if (!sh) {
 258                                 conf->inactive_blocked = 1;
 259                                 wait_event_lock_irq(conf->wait_for_stripe,
 260                                                     !list_empty(&conf->inactive_list) &&
 261                                                     (atomic_read(&conf->active_stripes)
 262                                                      < (conf->max_nr_stripes *3/4)
 263                                                      || !conf->inactive_blocked),
 264                                                     conf->device_lock,
 265                                                     unplug_slaves(conf->mddev)
 266                                         );
 267                                 conf->inactive_blocked = 0;
 268                         } else
 269                                 init_stripe(sh, sector, pd_idx, disks);
 270                 } else {
 271                         if (atomic_read(&sh->count)) {
 272                                 if (!list_empty(&sh->lru))
 273                                         BUG();
 274                         } else {
 275                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 276                                         atomic_inc(&conf->active_stripes);
 277                                 if (!list_empty(&sh->lru))
 278                                         list_del_init(&sh->lru);
 279                         }
 280                 }
 281         } while (sh == NULL);
 282
 283         if (sh)
 284                 atomic_inc(&sh->count);
 285
 286         spin_unlock_irq(&conf->device_lock);
 287         return sh;
 288 }
 289
 290 static int grow_one_stripe(raid5_conf_t *conf)
 291 {
 292         struct stripe_head *sh;
 293         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
 294         if (!sh)
 295                 return 0;
 296         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
 297         sh->raid_conf = conf;
 298         spin_lock_init(&sh->lock);
 299
 300         if (grow_buffers(sh, conf->raid_disks)) {
 301                 shrink_buffers(sh, conf->raid_disks);
 302                 kmem_cache_free(conf->slab_cache, sh);
 303                 return 0;
 304         }
 305         sh->disks = conf->raid_disks;
 306         /* we just created an active stripe so... */
 307         atomic_set(&sh->count, 1);
 308         atomic_inc(&conf->active_stripes);
 309         INIT_LIST_HEAD(&sh->lru);
 310         release_stripe(sh);
 311         return 1;
 312 }
 313
 314 static int grow_stripes(raid5_conf_t *conf, int num)
 315 {
 316         kmem_cache_t *sc;
 317         int devs = conf->raid_disks;
 318
 319         sprintf(conf->cache_name[0], "raid5/%s", mdname(conf->mddev));
 320         sprintf(conf->cache_name[1], "raid5/%s-alt", mdname(conf->mddev));
 321         conf->active_name = 0;
 322         sc = kmem_cache_create(conf->cache_name[conf->active_name],
 323                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
 324                                0, 0, NULL, NULL);
 325         if (!sc)
 326                 return 1;
 327         conf->slab_cache = sc;
 328         conf->pool_size = devs;
 329         while (num--) {
 330                 if (!grow_one_stripe(conf))
 331                         return 1;
 332         }
 333         return 0;
 334 }
 335
 336 #ifdef CONFIG_MD_RAID5_RESHAPE
 337 static int resize_stripes(raid5_conf_t *conf, int newsize)
 338 {
 339         /* Make all the stripes able to hold 'newsize' devices.
 340          * New slots in each stripe get 'page' set to a new page.
 341          *
 342          * This happens in stages:
 343          * 1/ create a new kmem_cache and allocate the required number of
 344          *    stripe_heads.
 345          * 2/ gather all the old stripe_heads and tranfer the pages across
 346          *    to the new stripe_heads.  This will have the side effect of
 347          *    freezing the array as once all stripe_heads have been collected,
 348          *    no IO will be possible.  Old stripe heads are freed once their
 349          *    pages have been transferred over, and the old kmem_cache is
 350          *    freed when all stripes are done.
 351          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
 352          *    we simple return a failre status - no need to clean anything up.
 353          * 4/ allocate new pages for the new slots in the new stripe_heads.
 354          *    If this fails, we don't bother trying the shrink the
 355          *    stripe_heads down again, we just leave them as they are.
 356          *    As each stripe_head is processed the new one is released into
 357          *    active service.
 358          *
 359          * Once step2 is started, we cannot afford to wait for a write,
 360          * so we use GFP_NOIO allocations.
 361          */
 362         struct stripe_head *osh, *nsh;
 363         LIST_HEAD(newstripes);
 364         struct disk_info *ndisks;
 365         int err = 0;
 366         kmem_cache_t *sc;
 367         int i;
 368
 369         if (newsize <= conf->pool_size)
 370                 return 0; /* never bother to shrink */
 371
 372         /* Step 1 */
 373         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
 374                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
 375                                0, 0, NULL, NULL);
 376         if (!sc)
 377                 return -ENOMEM;
 378
 379         for (i = conf->max_nr_stripes; i; i--) {
 380                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
 381                 if (!nsh)
 382                         break;
 383
 384                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
 385
 386                 nsh->raid_conf = conf;
 387                 spin_lock_init(&nsh->lock);
 388
 389                 list_add(&nsh->lru, &newstripes);
 390         }
 391         if (i) {
 392                 /* didn't get enough, give up */
 393                 while (!list_empty(&newstripes)) {
 394                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
 395                         list_del(&nsh->lru);
 396                         kmem_cache_free(sc, nsh);
 397                 }
 398                 kmem_cache_destroy(sc);
 399                 return -ENOMEM;
 400         }
 401         /* Step 2 - Must use GFP_NOIO now.
 402          * OK, we have enough stripes, start collecting inactive
 403          * stripes and copying them over
 404          */
 405         list_for_each_entry(nsh, &newstripes, lru) {
 406                 spin_lock_irq(&conf->device_lock);
 407                 wait_event_lock_irq(conf->wait_for_stripe,
 408                                     !list_empty(&conf->inactive_list),
 409                                     conf->device_lock,
 410                                     unplug_slaves(conf->mddev)
 411                         );
 412                 osh = get_free_stripe(conf);
 413                 spin_unlock_irq(&conf->device_lock);
 414                 atomic_set(&nsh->count, 1);
 415                 for(i=0; i<conf->pool_size; i++)
 416                         nsh->dev[i].page = osh->dev[i].page;
 417                 for( ; i<newsize; i++)
 418                         nsh->dev[i].page = NULL;
 419                 kmem_cache_free(conf->slab_cache, osh);
 420         }
 421         kmem_cache_destroy(conf->slab_cache);
 422
 423         /* Step 3.
 424          * At this point, we are holding all the stripes so the array
 425          * is completely stalled, so now is a good time to resize
 426          * conf->disks.
 427          */
 428         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
 429         if (ndisks) {
 430                 for (i=0; i<conf->raid_disks; i++)
 431                         ndisks[i] = conf->disks[i];
 432                 kfree(conf->disks);
 433                 conf->disks = ndisks;
 434         } else
 435                 err = -ENOMEM;
 436
 437         /* Step 4, return new stripes to service */
 438         while(!list_empty(&newstripes)) {
 439                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
 440                 list_del_init(&nsh->lru);
 441                 for (i=conf->raid_disks; i < newsize; i++)
 442                         if (nsh->dev[i].page == NULL) {
 443                                 struct page *p = alloc_page(GFP_NOIO);
 444                                 nsh->dev[i].page = p;
 445                                 if (!p)
 446                                         err = -ENOMEM;
 447                         }
 448                 release_stripe(nsh);
 449         }
 450         /* critical section pass, GFP_NOIO no longer needed */
 451
 452         conf->slab_cache = sc;
 453         conf->active_name = 1-conf->active_name;
 454         conf->pool_size = newsize;
 455         return err;
 456 }
 457 #endif
 458
 459 static int drop_one_stripe(raid5_conf_t *conf)
 460 {
 461         struct stripe_head *sh;
 462
 463         spin_lock_irq(&conf->device_lock);
 464         sh = get_free_stripe(conf);
 465         spin_unlock_irq(&conf->device_lock);
 466         if (!sh)
 467                 return 0;
 468         if (atomic_read(&sh->count))
 469                 BUG();
 470         shrink_buffers(sh, conf->pool_size);
 471         kmem_cache_free(conf->slab_cache, sh);
 472         atomic_dec(&conf->active_stripes);
 473         return 1;
 474 }
 475
 476 static void shrink_stripes(raid5_conf_t *conf)
 477 {
 478         while (drop_one_stripe(conf))
 479                 ;
 480
 481         if (conf->slab_cache)
 482                 kmem_cache_destroy(conf->slab_cache);
 483         conf->slab_cache = NULL;
 484 }
 485
 486 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
 487                                    int error)
 488 {
 489         struct stripe_head *sh = bi->bi_private;
 490         raid5_conf_t *conf = sh->raid_conf;
 491         int disks = sh->disks, i;
 492         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 493
 494         if (bi->bi_size)
 495                 return 1;
 496
 497         for (i=0 ; i<disks; i++)
 498                 if (bi == &sh->dev[i].req)
 499                         break;
 500
 501         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
 502                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 503                 uptodate);
 504         if (i == disks) {
 505                 BUG();
 506                 return 0;
 507         }
 508
 509         if (uptodate) {
 510 #if 0
 511                 struct bio *bio;
 512                 unsigned long flags;
 513                 spin_lock_irqsave(&conf->device_lock, flags);
 514                 /* we can return a buffer if we bypassed the cache or
 515                  * if the top buffer is not in highmem.  If there are
 516                  * multiple buffers, leave the extra work to
 517                  * handle_stripe
 518                  */
 519                 buffer = sh->bh_read[i];
 520                 if (buffer &&
 521                     (!PageHighMem(buffer->b_page)
 522                      || buffer->b_page == bh->b_page )
 523                         ) {
 524                         sh->bh_read[i] = buffer->b_reqnext;
 525                         buffer->b_reqnext = NULL;
 526                 } else
 527                         buffer = NULL;
 528                 spin_unlock_irqrestore(&conf->device_lock, flags);
 529                 if (sh->bh_page[i]==bh->b_page)
 530                         set_buffer_uptodate(bh);
 531                 if (buffer) {
 532                         if (buffer->b_page != bh->b_page)
 533                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
 534                         buffer->b_end_io(buffer, 1);
 535                 }
 536 #else
 537                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
 538 #endif
 539                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
 540                         printk(KERN_INFO "raid5: read error corrected!!\n");
 541                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 542                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 543                 }
 544                 if (atomic_read(&conf->disks[i].rdev->read_errors))
 545                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
 546         } else {
 547                 int retry = 0;
 548                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
 549                 atomic_inc(&conf->disks[i].rdev->read_errors);
 550                 if (conf->mddev->degraded)
 551                         printk(KERN_WARNING "raid5: read error not correctable.\n");
 552                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
 553                         /* Oh, no!!! */
 554                         printk(KERN_WARNING "raid5: read error NOT corrected!!\n");
 555                 else if (atomic_read(&conf->disks[i].rdev->read_errors)
 556                          > conf->max_nr_stripes)
 557                         printk(KERN_WARNING
 558                                "raid5: Too many read errors, failing device.\n");
 559                 else
 560                         retry = 1;
 561                 if (retry)
 562                         set_bit(R5_ReadError, &sh->dev[i].flags);
 563                 else {
 564                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 565                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 566                         md_error(conf->mddev, conf->disks[i].rdev);
 567                 }
 568         }
 569         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 570 #if 0
 571         /* must restore b_page before unlocking buffer... */
 572         if (sh->bh_page[i] != bh->b_page) {
 573                 bh->b_page = sh->bh_page[i];
 574                 bh->b_data = page_address(bh->b_page);
 575                 clear_buffer_uptodate(bh);
 576         }
 577 #endif
 578         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 579         set_bit(STRIPE_HANDLE, &sh->state);
 580         release_stripe(sh);
 581         return 0;
 582 }
 583
 584 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
 585                                     int error)
 586 {
 587         struct stripe_head *sh = bi->bi_private;
 588         raid5_conf_t *conf = sh->raid_conf;
 589         int disks = sh->disks, i;
 590         unsigned long flags;
 591         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 592
 593         if (bi->bi_size)
 594                 return 1;
 595
 596         for (i=0 ; i<disks; i++)
 597                 if (bi == &sh->dev[i].req)
 598                         break;
 599
 600         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
 601                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 602                 uptodate);
 603         if (i == disks) {
 604                 BUG();
 605                 return 0;
 606         }
 607
 608         spin_lock_irqsave(&conf->device_lock, flags);
 609         if (!uptodate)
 610                 md_error(conf->mddev, conf->disks[i].rdev);
 611
 612         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 613
 614         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 615         set_bit(STRIPE_HANDLE, &sh->state);
 616         __release_stripe(conf, sh);
 617         spin_unlock_irqrestore(&conf->device_lock, flags);
 618         return 0;
 619 }
 620
 621
 622 static sector_t compute_blocknr(struct stripe_head *sh, int i);
 623
 624 static void raid5_build_block (struct stripe_head *sh, int i)
 625 {
 626         struct r5dev *dev = &sh->dev[i];
 627
 628         bio_init(&dev->req);
 629         dev->req.bi_io_vec = &dev->vec;
 630         dev->req.bi_vcnt++;
 631         dev->req.bi_max_vecs++;
 632         dev->vec.bv_page = dev->page;
 633         dev->vec.bv_len = STRIPE_SIZE;
 634         dev->vec.bv_offset = 0;
 635
 636         dev->req.bi_sector = sh->sector;
 637         dev->req.bi_private = sh;
 638
 639         dev->flags = 0;
 640         if (i != sh->pd_idx)
 641                 dev->sector = compute_blocknr(sh, i);
 642 }
 643
 644 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 645 {
 646         char b[BDEVNAME_SIZE];
 647         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
 648         PRINTK("raid5: error called\n");
 649
 650         if (!test_bit(Faulty, &rdev->flags)) {
 651                 mddev->sb_dirty = 1;
 652                 if (test_bit(In_sync, &rdev->flags)) {
 653                         conf->working_disks--;
 654                         mddev->degraded++;
 655                         conf->failed_disks++;
 656                         clear_bit(In_sync, &rdev->flags);
 657                         /*
 658                          * if recovery was running, make sure it aborts.
 659                          */
 660                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 661                 }
 662                 set_bit(Faulty, &rdev->flags);
 663                 printk (KERN_ALERT
 664                         "raid5: Disk failure on %s, disabling device."
 665                         " Operation continuing on %d devices\n",
 666                         bdevname(rdev->bdev,b), conf->working_disks);
 667         }
 668 }
 669
 670 /*
 671  * Input: a 'big' sector number,
 672  * Output: index of the data and parity disk, and the sector # in them.
 673  */
 674 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
 675                         unsigned int data_disks, unsigned int * dd_idx,
 676                         unsigned int * pd_idx, raid5_conf_t *conf)
 677 {
 678         long stripe;
 679         unsigned long chunk_number;
 680         unsigned int chunk_offset;
 681         sector_t new_sector;
 682         int sectors_per_chunk = conf->chunk_size >> 9;
 683
 684         /* First compute the information on this sector */
 685
 686         /*
 687          * Compute the chunk number and the sector offset inside the chunk
 688          */
 689         chunk_offset = sector_div(r_sector, sectors_per_chunk);
 690         chunk_number = r_sector;
 691         BUG_ON(r_sector != chunk_number);
 692
 693         /*
 694          * Compute the stripe number
 695          */
 696         stripe = chunk_number / data_disks;
 697
 698         /*
 699          * Compute the data disk and parity disk indexes inside the stripe
 700          */
 701         *dd_idx = chunk_number % data_disks;
 702
 703         /*
 704          * Select the parity disk based on the user selected algorithm.
 705          */
 706         if (conf->level == 4)
 707                 *pd_idx = data_disks;
 708         else switch (conf->algorithm) {
 709                 case ALGORITHM_LEFT_ASYMMETRIC:
 710                         *pd_idx = data_disks - stripe % raid_disks;
 711                         if (*dd_idx >= *pd_idx)
 712                                 (*dd_idx)++;
 713                         break;
 714                 case ALGORITHM_RIGHT_ASYMMETRIC:
 715                         *pd_idx = stripe % raid_disks;
 716                         if (*dd_idx >= *pd_idx)
 717                                 (*dd_idx)++;
 718                         break;
 719                 case ALGORITHM_LEFT_SYMMETRIC:
 720                         *pd_idx = data_disks - stripe % raid_disks;
 721                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 722                         break;
 723                 case ALGORITHM_RIGHT_SYMMETRIC:
 724                         *pd_idx = stripe % raid_disks;
 725                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
 726                         break;
 727                 default:
 728                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
 729                                 conf->algorithm);
 730         }
 731
 732         /*
 733          * Finally, compute the new sector number
 734          */
 735         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
 736         return new_sector;
 737 }
 738
 739
 740 static sector_t compute_blocknr(struct stripe_head *sh, int i)
 741 {
 742         raid5_conf_t *conf = sh->raid_conf;
 743         int raid_disks = sh->disks, data_disks = raid_disks - 1;
 744         sector_t new_sector = sh->sector, check;
 745         int sectors_per_chunk = conf->chunk_size >> 9;
 746         sector_t stripe;
 747         int chunk_offset;
 748         int chunk_number, dummy1, dummy2, dd_idx = i;
 749         sector_t r_sector;
 750
 751         chunk_offset = sector_div(new_sector, sectors_per_chunk);
 752         stripe = new_sector;
 753         BUG_ON(new_sector != stripe);
 754
 755
 756         switch (conf->algorithm) {
 757                 case ALGORITHM_LEFT_ASYMMETRIC:
 758                 case ALGORITHM_RIGHT_ASYMMETRIC:
 759                         if (i > sh->pd_idx)
 760                                 i--;
 761                         break;
 762                 case ALGORITHM_LEFT_SYMMETRIC:
 763                 case ALGORITHM_RIGHT_SYMMETRIC:
 764                         if (i < sh->pd_idx)
 765                                 i += raid_disks;
 766                         i -= (sh->pd_idx + 1);
 767                         break;
 768                 default:
 769                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
 770                                 conf->algorithm);
 771         }
 772
 773         chunk_number = stripe * data_disks + i;
 774         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
 775
 776         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
 777         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
 778                 printk(KERN_ERR "compute_blocknr: map not correct\n");
 779                 return 0;
 780         }
 781         return r_sector;
 782 }
 783
 784
 785
 786 /*
 787  * Copy data between a page in the stripe cache, and a bio.
 788  * There are no alignment or size guarantees between the page or the
 789  * bio except that there is some overlap.
 790  * All iovecs in the bio must be considered.
 791  */
 792 static void copy_data(int frombio, struct bio *bio,
 793                      struct page *page,
 794                      sector_t sector)
 795 {
 796         char *pa = page_address(page);
 797         struct bio_vec *bvl;
 798         int i;
 799         int page_offset;
 800
 801         if (bio->bi_sector >= sector)
 802                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 803         else
 804                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 805         bio_for_each_segment(bvl, bio, i) {
 806                 int len = bio_iovec_idx(bio,i)->bv_len;
 807                 int clen;
 808                 int b_offset = 0;
 809
 810                 if (page_offset < 0) {
 811                         b_offset = -page_offset;
 812                         page_offset += b_offset;
 813                         len -= b_offset;
 814                 }
 815
 816                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 817                         clen = STRIPE_SIZE - page_offset;
 818                 else clen = len;
 819
 820                 if (clen > 0) {
 821                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
 822                         if (frombio)
 823                                 memcpy(pa+page_offset, ba+b_offset, clen);
 824                         else
 825                                 memcpy(ba+b_offset, pa+page_offset, clen);
 826                         __bio_kunmap_atomic(ba, KM_USER0);
 827                 }
 828                 if (clen < len) /* hit end of page */
 829                         break;
 830                 page_offset +=  len;
 831         }
 832 }
 833
 834 #define check_xor()     do {                                            \
 835                            if (count == MAX_XOR_BLOCKS) {               \
 836                                 xor_block(count, STRIPE_SIZE, ptr);     \
 837                                 count = 1;                              \
 838                            }                                            \
 839                         } while(0)
 840
 841
 842 static void compute_block(struct stripe_head *sh, int dd_idx)
 843 {
 844         int i, count, disks = sh->disks;
 845         void *ptr[MAX_XOR_BLOCKS], *p;
 846
 847         PRINTK("compute_block, stripe %llu, idx %d\n",
 848                 (unsigned long long)sh->sector, dd_idx);
 849
 850         ptr[0] = page_address(sh->dev[dd_idx].page);
 851         memset(ptr[0], 0, STRIPE_SIZE);
 852         count = 1;
 853         for (i = disks ; i--; ) {
 854                 if (i == dd_idx)
 855                         continue;
 856                 p = page_address(sh->dev[i].page);
 857                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
 858                         ptr[count++] = p;
 859                 else
 860                         printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
 861                                 " not present\n", dd_idx,
 862                                 (unsigned long long)sh->sector, i);
 863
 864                 check_xor();
 865         }
 866         if (count != 1)
 867                 xor_block(count, STRIPE_SIZE, ptr);
 868         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
 869 }
 870
 871 static void compute_parity(struct stripe_head *sh, int method)
 872 {
 873         raid5_conf_t *conf = sh->raid_conf;
 874         int i, pd_idx = sh->pd_idx, disks = sh->disks, count;
 875         void *ptr[MAX_XOR_BLOCKS];
 876         struct bio *chosen;
 877
 878         PRINTK("compute_parity, stripe %llu, method %d\n",
 879                 (unsigned long long)sh->sector, method);
 880
 881         count = 1;
 882         ptr[0] = page_address(sh->dev[pd_idx].page);
 883         switch(method) {
 884         case READ_MODIFY_WRITE:
 885                 if (!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags))
 886                         BUG();
 887                 for (i=disks ; i-- ;) {
 888                         if (i==pd_idx)
 889                                 continue;
 890                         if (sh->dev[i].towrite &&
 891                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
 892                                 ptr[count++] = page_address(sh->dev[i].page);
 893                                 chosen = sh->dev[i].towrite;
 894                                 sh->dev[i].towrite = NULL;
 895
 896                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 897                                         wake_up(&conf->wait_for_overlap);
 898
 899                                 if (sh->dev[i].written) BUG();
 900                                 sh->dev[i].written = chosen;
 901                                 check_xor();
 902                         }
 903                 }
 904                 break;
 905         case RECONSTRUCT_WRITE:
 906                 memset(ptr[0], 0, STRIPE_SIZE);
 907                 for (i= disks; i-- ;)
 908                         if (i!=pd_idx && sh->dev[i].towrite) {
 909                                 chosen = sh->dev[i].towrite;
 910                                 sh->dev[i].towrite = NULL;
 911
 912                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 913                                         wake_up(&conf->wait_for_overlap);
 914
 915                                 if (sh->dev[i].written) BUG();
 916                                 sh->dev[i].written = chosen;
 917                         }
 918                 break;
 919         case CHECK_PARITY:
 920                 break;
 921         }
 922         if (count>1) {
 923                 xor_block(count, STRIPE_SIZE, ptr);
 924                 count = 1;
 925         }
 926
 927         for (i = disks; i--;)
 928                 if (sh->dev[i].written) {
 929                         sector_t sector = sh->dev[i].sector;
 930                         struct bio *wbi = sh->dev[i].written;
 931                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
 932                                 copy_data(1, wbi, sh->dev[i].page, sector);
 933                                 wbi = r5_next_bio(wbi, sector);
 934                         }
 935
 936                         set_bit(R5_LOCKED, &sh->dev[i].flags);
 937                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
 938                 }
 939
 940         switch(method) {
 941         case RECONSTRUCT_WRITE:
 942         case CHECK_PARITY:
 943                 for (i=disks; i--;)
 944                         if (i != pd_idx) {
 945                                 ptr[count++] = page_address(sh->dev[i].page);
 946                                 check_xor();
 947                         }
 948                 break;
 949         case READ_MODIFY_WRITE:
 950                 for (i = disks; i--;)
 951                         if (sh->dev[i].written) {
 952                                 ptr[count++] = page_address(sh->dev[i].page);
 953                                 check_xor();
 954                         }
 955         }
 956         if (count != 1)
 957                 xor_block(count, STRIPE_SIZE, ptr);
 958
 959         if (method != CHECK_PARITY) {
 960                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 961                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
 962         } else
 963                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 964 }
 965
 966 /*
 967  * Each stripe/dev can have one or more bion attached.
 968  * toread/towrite point to the first in a chain.
 969  * The bi_next chain must be in order.
 970  */
 971 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
 972 {
 973         struct bio **bip;
 974         raid5_conf_t *conf = sh->raid_conf;
 975         int firstwrite=0;
 976
 977         PRINTK("adding bh b#%llu to stripe s#%llu\n",
 978                 (unsigned long long)bi->bi_sector,
 979                 (unsigned long long)sh->sector);
 980
 981
 982         spin_lock(&sh->lock);
 983         spin_lock_irq(&conf->device_lock);
 984         if (forwrite) {
 985                 bip = &sh->dev[dd_idx].towrite;
 986                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
 987                         firstwrite = 1;
 988         } else
 989                 bip = &sh->dev[dd_idx].toread;
 990         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
 991                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
 992                         goto overlap;
 993                 bip = & (*bip)->bi_next;
 994         }
 995         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
 996                 goto overlap;
 997
 998         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
 999                 BUG();
1000         if (*bip)
1001                 bi->bi_next = *bip;
1002         *bip = bi;
1003         bi->bi_phys_segments ++;
1004         spin_unlock_irq(&conf->device_lock);
1005         spin_unlock(&sh->lock);
1006
1007         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
1008                 (unsigned long long)bi->bi_sector,
1009                 (unsigned long long)sh->sector, dd_idx);
1010
1011         if (conf->mddev->bitmap && firstwrite) {
1012                 sh->bm_seq = conf->seq_write;
1013                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1014                                   STRIPE_SECTORS, 0);
1015                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1016         }
1017
1018         if (forwrite) {
1019                 /* check if page is covered */
1020                 sector_t sector = sh->dev[dd_idx].sector;
1021                 for (bi=sh->dev[dd_idx].towrite;
1022                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1023                              bi && bi->bi_sector <= sector;
1024                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1025                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1026                                 sector = bi->bi_sector + (bi->bi_size>>9);
1027                 }
1028                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1029                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1030         }
1031         return 1;
1032
1033  overlap:
1034         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1035         spin_unlock_irq(&conf->device_lock);
1036         spin_unlock(&sh->lock);
1037         return 0;
1038 }
1039
1040 static void end_reshape(raid5_conf_t *conf);
1041
1042 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1043 {
1044         int sectors_per_chunk = conf->chunk_size >> 9;
1045         sector_t x = stripe;
1046         int pd_idx, dd_idx;
1047         int chunk_offset = sector_div(x, sectors_per_chunk);
1048         stripe = x;
1049         raid5_compute_sector(stripe*(disks-1)*sectors_per_chunk
1050                              + chunk_offset, disks, disks-1, &dd_idx, &pd_idx, conf);
1051         return pd_idx;
1052 }
1053
1054
1055 /*
1056  * handle_stripe - do things to a stripe.
1057  *
1058  * We lock the stripe and then examine the state of various bits
1059  * to see what needs to be done.
1060  * Possible results:
1061  *    return some read request which now have data
1062  *    return some write requests which are safely on disc
1063  *    schedule a read on some buffers
1064  *    schedule a write of some buffers
1065  *    return confirmation of parity correctness
1066  *
1067  * Parity calculations are done inside the stripe lock
1068  * buffers are taken off read_list or write_list, and bh_cache buffers
1069  * get BH_Lock set before the stripe lock is released.
1070  *
1071  */
1072
1073 static void handle_stripe(struct stripe_head *sh)
1074 {
1075         raid5_conf_t *conf = sh->raid_conf;
1076         int disks = sh->disks;
1077         struct bio *return_bi= NULL;
1078         struct bio *bi;
1079         int i;
1080         int syncing, expanding, expanded;
1081         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1082         int non_overwrite = 0;
1083         int failed_num=0;
1084         struct r5dev *dev;
1085
1086         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
1087                 (unsigned long long)sh->sector, atomic_read(&sh->count),
1088                 sh->pd_idx);
1089
1090         spin_lock(&sh->lock);
1091         clear_bit(STRIPE_HANDLE, &sh->state);
1092         clear_bit(STRIPE_DELAYED, &sh->state);
1093
1094         syncing = test_bit(STRIPE_SYNCING, &sh->state);
1095         expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1096         expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
1097         /* Now to look around and see what can be done */
1098
1099         rcu_read_lock();
1100         for (i=disks; i--; ) {
1101                 mdk_rdev_t *rdev;
1102                 dev = &sh->dev[i];
1103                 clear_bit(R5_Insync, &dev->flags);
1104
1105                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1106                         i, dev->flags, dev->toread, dev->towrite, dev->written);
1107                 /* maybe we can reply to a read */
1108                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1109                         struct bio *rbi, *rbi2;
1110                         PRINTK("Return read for disc %d\n", i);
1111                         spin_lock_irq(&conf->device_lock);
1112                         rbi = dev->toread;
1113                         dev->toread = NULL;
1114                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1115                                 wake_up(&conf->wait_for_overlap);
1116                         spin_unlock_irq(&conf->device_lock);
1117                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1118                                 copy_data(0, rbi, dev->page, dev->sector);
1119                                 rbi2 = r5_next_bio(rbi, dev->sector);
1120                                 spin_lock_irq(&conf->device_lock);
1121                                 if (--rbi->bi_phys_segments == 0) {
1122                                         rbi->bi_next = return_bi;
1123                                         return_bi = rbi;
1124                                 }
1125                                 spin_unlock_irq(&conf->device_lock);
1126                                 rbi = rbi2;
1127                         }
1128                 }
1129
1130                 /* now count some things */
1131                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1132                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1133
1134
1135                 if (dev->toread) to_read++;
1136                 if (dev->towrite) {
1137                         to_write++;
1138                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1139                                 non_overwrite++;
1140                 }
1141                 if (dev->written) written++;
1142                 rdev = rcu_dereference(conf->disks[i].rdev);
1143                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1144                         /* The ReadError flag will just be confusing now */
1145                         clear_bit(R5_ReadError, &dev->flags);
1146                         clear_bit(R5_ReWrite, &dev->flags);
1147                 }
1148                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1149                     || test_bit(R5_ReadError, &dev->flags)) {
1150                         failed++;
1151                         failed_num = i;
1152                 } else
1153                         set_bit(R5_Insync, &dev->flags);
1154         }
1155         rcu_read_unlock();
1156         PRINTK("locked=%d uptodate=%d to_read=%d"
1157                 " to_write=%d failed=%d failed_num=%d\n",
1158                 locked, uptodate, to_read, to_write, failed, failed_num);
1159         /* check if the array has lost two devices and, if so, some requests might
1160          * need to be failed
1161          */
1162         if (failed > 1 && to_read+to_write+written) {
1163                 for (i=disks; i--; ) {
1164                         int bitmap_end = 0;
1165
1166                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1167                                 mdk_rdev_t *rdev;
1168                                 rcu_read_lock();
1169                                 rdev = rcu_dereference(conf->disks[i].rdev);
1170                                 if (rdev && test_bit(In_sync, &rdev->flags))
1171                                         /* multiple read failures in one stripe */
1172                                         md_error(conf->mddev, rdev);
1173                                 rcu_read_unlock();
1174                         }
1175
1176                         spin_lock_irq(&conf->device_lock);
1177                         /* fail all writes first */
1178                         bi = sh->dev[i].towrite;
1179                         sh->dev[i].towrite = NULL;
1180                         if (bi) { to_write--; bitmap_end = 1; }
1181
1182                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1183                                 wake_up(&conf->wait_for_overlap);
1184
1185                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1186                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1187                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1188                                 if (--bi->bi_phys_segments == 0) {
1189                                         md_write_end(conf->mddev);
1190                                         bi->bi_next = return_bi;
1191                                         return_bi = bi;
1192                                 }
1193                                 bi = nextbi;
1194                         }
1195                         /* and fail all 'written' */
1196                         bi = sh->dev[i].written;
1197                         sh->dev[i].written = NULL;
1198                         if (bi) bitmap_end = 1;
1199                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1200                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1201                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1202                                 if (--bi->bi_phys_segments == 0) {
1203                                         md_write_end(conf->mddev);
1204                                         bi->bi_next = return_bi;
1205                                         return_bi = bi;
1206                                 }
1207                                 bi = bi2;
1208                         }
1209
1210                         /* fail any reads if this device is non-operational */
1211                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1212                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1213                                 bi = sh->dev[i].toread;
1214                                 sh->dev[i].toread = NULL;
1215                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1216                                         wake_up(&conf->wait_for_overlap);
1217                                 if (bi) to_read--;
1218                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1219                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1220                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1221                                         if (--bi->bi_phys_segments == 0) {
1222                                                 bi->bi_next = return_bi;
1223                                                 return_bi = bi;
1224                                         }
1225                                         bi = nextbi;
1226                                 }
1227                         }
1228                         spin_unlock_irq(&conf->device_lock);
1229                         if (bitmap_end)
1230                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1231                                                 STRIPE_SECTORS, 0, 0);
1232                 }
1233         }
1234         if (failed > 1 && syncing) {
1235                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1236                 clear_bit(STRIPE_SYNCING, &sh->state);
1237                 syncing = 0;
1238         }
1239
1240         /* might be able to return some write requests if the parity block
1241          * is safe, or on a failed drive
1242          */
1243         dev = &sh->dev[sh->pd_idx];
1244         if ( written &&
1245              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1246                 test_bit(R5_UPTODATE, &dev->flags))
1247                || (failed == 1 && failed_num == sh->pd_idx))
1248             ) {
1249             /* any written block on an uptodate or failed drive can be returned.
1250              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
1251              * never LOCKED, so we don't need to test 'failed' directly.
1252              */
1253             for (i=disks; i--; )
1254                 if (sh->dev[i].written) {
1255                     dev = &sh->dev[i];
1256                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1257                          test_bit(R5_UPTODATE, &dev->flags) ) {
1258                         /* We can return any write requests */
1259                             struct bio *wbi, *wbi2;
1260                             int bitmap_end = 0;
1261                             PRINTK("Return write for disc %d\n", i);
1262                             spin_lock_irq(&conf->device_lock);
1263                             wbi = dev->written;
1264                             dev->written = NULL;
1265                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1266                                     wbi2 = r5_next_bio(wbi, dev->sector);
1267                                     if (--wbi->bi_phys_segments == 0) {
1268                                             md_write_end(conf->mddev);
1269                                             wbi->bi_next = return_bi;
1270                                             return_bi = wbi;
1271                                     }
1272                                     wbi = wbi2;
1273                             }
1274                             if (dev->towrite == NULL)
1275                                     bitmap_end = 1;
1276                             spin_unlock_irq(&conf->device_lock);
1277                             if (bitmap_end)
1278                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1279                                                     STRIPE_SECTORS,
1280                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1281                     }
1282                 }
1283         }
1284
1285         /* Now we might consider reading some blocks, either to check/generate
1286          * parity, or to satisfy requests
1287          * or to load a block that is being partially written.
1288          */
1289         if (to_read || non_overwrite || (syncing && (uptodate < disks)) || expanding) {
1290                 for (i=disks; i--;) {
1291                         dev = &sh->dev[i];
1292                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1293                             (dev->toread ||
1294                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1295                              syncing ||
1296                              expanding ||
1297                              (failed && (sh->dev[failed_num].toread ||
1298                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1299                                     )
1300                                 ) {
1301                                 /* we would like to get this block, possibly
1302                                  * by computing it, but we might not be able to
1303                                  */
1304                                 if (uptodate == disks-1) {
1305                                         PRINTK("Computing block %d\n", i);
1306                                         compute_block(sh, i);
1307                                         uptodate++;
1308                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1309                                         set_bit(R5_LOCKED, &dev->flags);
1310                                         set_bit(R5_Wantread, &dev->flags);
1311 #if 0
1312                                         /* if I am just reading this block and we don't have
1313                                            a failed drive, or any pending writes then sidestep the cache */
1314                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1315                                             ! syncing && !failed && !to_write) {
1316                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1317                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1318                                         }
1319 #endif
1320                                         locked++;
1321                                         PRINTK("Reading block %d (sync=%d)\n",
1322                                                 i, syncing);
1323                                 }
1324                         }
1325                 }
1326                 set_bit(STRIPE_HANDLE, &sh->state);
1327         }
1328
1329         /* now to consider writing and what else, if anything should be read */
1330         if (to_write) {
1331                 int rmw=0, rcw=0;
1332                 for (i=disks ; i--;) {
1333                         /* would I have to read this buffer for read_modify_write */
1334                         dev = &sh->dev[i];
1335                         if ((dev->towrite || i == sh->pd_idx) &&
1336                             (!test_bit(R5_LOCKED, &dev->flags)
1337 #if 0
1338 || sh->bh_page[i]!=bh->b_page
1339 #endif
1340                                     ) &&
1341                             !test_bit(R5_UPTODATE, &dev->flags)) {
1342                                 if (test_bit(R5_Insync, &dev->flags)
1343 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1344                                         )
1345                                         rmw++;
1346                                 else rmw += 2*disks;  /* cannot read it */
1347                         }
1348                         /* Would I have to read this buffer for reconstruct_write */
1349                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1350                             (!test_bit(R5_LOCKED, &dev->flags)
1351 #if 0
1352 || sh->bh_page[i] != bh->b_page
1353 #endif
1354                                     ) &&
1355                             !test_bit(R5_UPTODATE, &dev->flags)) {
1356                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1357                                 else rcw += 2*disks;
1358                         }
1359                 }
1360                 PRINTK("for sector %llu, rmw=%d rcw=%d\n",
1361                         (unsigned long long)sh->sector, rmw, rcw);
1362                 set_bit(STRIPE_HANDLE, &sh->state);
1363                 if (rmw < rcw && rmw > 0)
1364                         /* prefer read-modify-write, but need to get some data */
1365                         for (i=disks; i--;) {
1366                                 dev = &sh->dev[i];
1367                                 if ((dev->towrite || i == sh->pd_idx) &&
1368                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1369                                     test_bit(R5_Insync, &dev->flags)) {
1370                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1371                                         {
1372                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1373                                                 set_bit(R5_LOCKED, &dev->flags);
1374                                                 set_bit(R5_Wantread, &dev->flags);
1375                                                 locked++;
1376                                         } else {
1377                                                 set_bit(STRIPE_DELAYED, &sh->state);
1378                                                 set_bit(STRIPE_HANDLE, &sh->state);
1379                                         }
1380                                 }
1381                         }
1382                 if (rcw <= rmw && rcw > 0)
1383                         /* want reconstruct write, but need to get some data */
1384                         for (i=disks; i--;) {
1385                                 dev = &sh->dev[i];
1386                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1387                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1388                                     test_bit(R5_Insync, &dev->flags)) {
1389                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1390                                         {
1391                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1392                                                 set_bit(R5_LOCKED, &dev->flags);
1393                                                 set_bit(R5_Wantread, &dev->flags);
1394                                                 locked++;
1395                                         } else {
1396                                                 set_bit(STRIPE_DELAYED, &sh->state);
1397                                                 set_bit(STRIPE_HANDLE, &sh->state);
1398                                         }
1399                                 }
1400                         }
1401                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1402                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1403                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1404                         PRINTK("Computing parity...\n");
1405                         compute_parity(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1406                         /* now every locked buffer is ready to be written */
1407                         for (i=disks; i--;)
1408                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1409                                         PRINTK("Writing block %d\n", i);
1410                                         locked++;
1411                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1412                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1413                                             || (i==sh->pd_idx && failed == 0))
1414                                                 set_bit(STRIPE_INSYNC, &sh->state);
1415                                 }
1416                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1417                                 atomic_dec(&conf->preread_active_stripes);
1418                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1419                                         md_wakeup_thread(conf->mddev->thread);
1420                         }
1421                 }
1422         }
1423
1424         /* maybe we need to check and possibly fix the parity for this stripe
1425          * Any reads will already have been scheduled, so we just see if enough data
1426          * is available
1427          */
1428         if (syncing && locked == 0 &&
1429             !test_bit(STRIPE_INSYNC, &sh->state)) {
1430                 set_bit(STRIPE_HANDLE, &sh->state);
1431                 if (failed == 0) {
1432                         char *pagea;
1433                         if (uptodate != disks)
1434                                 BUG();
1435                         compute_parity(sh, CHECK_PARITY);
1436                         uptodate--;
1437                         pagea = page_address(sh->dev[sh->pd_idx].page);
1438                         if ((*(u32*)pagea) == 0 &&
1439                             !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1440                                 /* parity is correct (on disc, not in buffer any more) */
1441                                 set_bit(STRIPE_INSYNC, &sh->state);
1442                         } else {
1443                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1444                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1445                                         /* don't try to repair!! */
1446                                         set_bit(STRIPE_INSYNC, &sh->state);
1447                                 else {
1448                                         compute_block(sh, sh->pd_idx);
1449                                         uptodate++;
1450                                 }
1451                         }
1452                 }
1453                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1454                         /* either failed parity check, or recovery is happening */
1455                         if (failed==0)
1456                                 failed_num = sh->pd_idx;
1457                         dev = &sh->dev[failed_num];
1458                         BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
1459                         BUG_ON(uptodate != disks);
1460
1461                         set_bit(R5_LOCKED, &dev->flags);
1462                         set_bit(R5_Wantwrite, &dev->flags);
1463                         clear_bit(STRIPE_DEGRADED, &sh->state);
1464                         locked++;
1465                         set_bit(STRIPE_INSYNC, &sh->state);
1466                 }
1467         }
1468         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1469                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1470                 clear_bit(STRIPE_SYNCING, &sh->state);
1471         }
1472
1473         /* If the failed drive is just a ReadError, then we might need to progress
1474          * the repair/check process
1475          */
1476         if (failed == 1 && ! conf->mddev->ro &&
1477             test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1478             && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1479             && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1480                 ) {
1481                 dev = &sh->dev[failed_num];
1482                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1483                         set_bit(R5_Wantwrite, &dev->flags);
1484                         set_bit(R5_ReWrite, &dev->flags);
1485                         set_bit(R5_LOCKED, &dev->flags);
1486                         locked++;
1487                 } else {
1488                         /* let's read it back */
1489                         set_bit(R5_Wantread, &dev->flags);
1490                         set_bit(R5_LOCKED, &dev->flags);
1491                         locked++;
1492                 }
1493         }
1494
1495         if (expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
1496                 /* Need to write out all blocks after computing parity */
1497                 sh->disks = conf->raid_disks;
1498                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, conf->raid_disks);
1499                 compute_parity(sh, RECONSTRUCT_WRITE);
1500                 for (i= conf->raid_disks; i--;) {
1501                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1502                         locked++;
1503                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1504                 }
1505                 clear_bit(STRIPE_EXPANDING, &sh->state);
1506         } else if (expanded) {
1507                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
1508                 atomic_dec(&conf->reshape_stripes);
1509                 wake_up(&conf->wait_for_overlap);
1510                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
1511         }
1512
1513         if (expanding && locked == 0) {
1514                 /* We have read all the blocks in this stripe and now we need to
1515                  * copy some of them into a target stripe for expand.
1516                  */
1517                 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1518                 for (i=0; i< sh->disks; i++)
1519                         if (i != sh->pd_idx) {
1520                                 int dd_idx, pd_idx, j;
1521                                 struct stripe_head *sh2;
1522
1523                                 sector_t bn = compute_blocknr(sh, i);
1524                                 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
1525                                                                   conf->raid_disks-1,
1526                                                                   &dd_idx, &pd_idx, conf);
1527                                 sh2 = get_active_stripe(conf, s, conf->raid_disks, pd_idx, 1);
1528                                 if (sh2 == NULL)
1529                                         /* so far only the early blocks of this stripe
1530                                          * have been requested.  When later blocks
1531                                          * get requested, we will try again
1532                                          */
1533                                         continue;
1534                                 if(!test_bit(STRIPE_EXPANDING, &sh2->state) ||
1535                                    test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
1536                                         /* must have already done this block */
1537                                         release_stripe(sh2);
1538                                         continue;
1539                                 }
1540                                 memcpy(page_address(sh2->dev[dd_idx].page),
1541                                        page_address(sh->dev[i].page),
1542                                        STRIPE_SIZE);
1543                                 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
1544                                 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
1545                                 for (j=0; j<conf->raid_disks; j++)
1546                                         if (j != sh2->pd_idx &&
1547                                             !test_bit(R5_Expanded, &sh2->dev[j].flags))
1548                                                 break;
1549                                 if (j == conf->raid_disks) {
1550                                         set_bit(STRIPE_EXPAND_READY, &sh2->state);
1551                                         set_bit(STRIPE_HANDLE, &sh2->state);
1552                                 }
1553                                 release_stripe(sh2);
1554                         }
1555         }
1556
1557         spin_unlock(&sh->lock);
1558
1559         while ((bi=return_bi)) {
1560                 int bytes = bi->bi_size;
1561
1562                 return_bi = bi->bi_next;
1563                 bi->bi_next = NULL;
1564                 bi->bi_size = 0;
1565                 bi->bi_end_io(bi, bytes, 0);
1566         }
1567         for (i=disks; i-- ;) {
1568                 int rw;
1569                 struct bio *bi;
1570                 mdk_rdev_t *rdev;
1571                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1572                         rw = 1;
1573                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1574                         rw = 0;
1575                 else
1576                         continue;
1577
1578                 bi = &sh->dev[i].req;
1579
1580                 bi->bi_rw = rw;
1581                 if (rw)
1582                         bi->bi_end_io = raid5_end_write_request;
1583                 else
1584                         bi->bi_end_io = raid5_end_read_request;
1585
1586                 rcu_read_lock();
1587                 rdev = rcu_dereference(conf->disks[i].rdev);
1588                 if (rdev && test_bit(Faulty, &rdev->flags))
1589                         rdev = NULL;
1590                 if (rdev)
1591                         atomic_inc(&rdev->nr_pending);
1592                 rcu_read_unlock();
1593
1594                 if (rdev) {
1595                         if (syncing || expanding || expanded)
1596                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1597
1598                         bi->bi_bdev = rdev->bdev;
1599                         PRINTK("for %llu schedule op %ld on disc %d\n",
1600                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1601                         atomic_inc(&sh->count);
1602                         bi->bi_sector = sh->sector + rdev->data_offset;
1603                         bi->bi_flags = 1 << BIO_UPTODATE;
1604                         bi->bi_vcnt = 1;
1605                         bi->bi_max_vecs = 1;
1606                         bi->bi_idx = 0;
1607                         bi->bi_io_vec = &sh->dev[i].vec;
1608                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1609                         bi->bi_io_vec[0].bv_offset = 0;
1610                         bi->bi_size = STRIPE_SIZE;
1611                         bi->bi_next = NULL;
1612                         if (rw == WRITE &&
1613                             test_bit(R5_ReWrite, &sh->dev[i].flags))
1614                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1615                         generic_make_request(bi);
1616                 } else {
1617                         if (rw == 1)
1618                                 set_bit(STRIPE_DEGRADED, &sh->state);
1619                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1620                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1621                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1622                         set_bit(STRIPE_HANDLE, &sh->state);
1623                 }
1624         }
1625 }
1626
1627 static void raid5_activate_delayed(raid5_conf_t *conf)
1628 {
1629         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1630                 while (!list_empty(&conf->delayed_list)) {
1631                         struct list_head *l = conf->delayed_list.next;
1632                         struct stripe_head *sh;
1633                         sh = list_entry(l, struct stripe_head, lru);
1634                         list_del_init(l);
1635                         clear_bit(STRIPE_DELAYED, &sh->state);
1636                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1637                                 atomic_inc(&conf->preread_active_stripes);
1638                         list_add_tail(&sh->lru, &conf->handle_list);
1639                 }
1640         }
1641 }
1642
1643 static void activate_bit_delay(raid5_conf_t *conf)
1644 {
1645         /* device_lock is held */
1646         struct list_head head;
1647         list_add(&head, &conf->bitmap_list);
1648         list_del_init(&conf->bitmap_list);
1649         while (!list_empty(&head)) {
1650                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1651                 list_del_init(&sh->lru);
1652                 atomic_inc(&sh->count);
1653                 __release_stripe(conf, sh);
1654         }
1655 }
1656
1657 static void unplug_slaves(mddev_t *mddev)
1658 {
1659         raid5_conf_t *conf = mddev_to_conf(mddev);
1660         int i;
1661
1662         rcu_read_lock();
1663         for (i=0; i<mddev->raid_disks; i++) {
1664                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1665                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
1666                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1667
1668                         atomic_inc(&rdev->nr_pending);
1669                         rcu_read_unlock();
1670
1671                         if (r_queue->unplug_fn)
1672                                 r_queue->unplug_fn(r_queue);
1673
1674                         rdev_dec_pending(rdev, mddev);
1675                         rcu_read_lock();
1676                 }
1677         }
1678         rcu_read_unlock();
1679 }
1680
1681 static void raid5_unplug_device(request_queue_t *q)
1682 {
1683         mddev_t *mddev = q->queuedata;
1684         raid5_conf_t *conf = mddev_to_conf(mddev);
1685         unsigned long flags;
1686
1687         spin_lock_irqsave(&conf->device_lock, flags);
1688
1689         if (blk_remove_plug(q)) {
1690                 conf->seq_flush++;
1691                 raid5_activate_delayed(conf);
1692         }
1693         md_wakeup_thread(mddev->thread);
1694
1695         spin_unlock_irqrestore(&conf->device_lock, flags);
1696
1697         unplug_slaves(mddev);
1698 }
1699
1700 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
1701                              sector_t *error_sector)
1702 {
1703         mddev_t *mddev = q->queuedata;
1704         raid5_conf_t *conf = mddev_to_conf(mddev);
1705         int i, ret = 0;
1706
1707         rcu_read_lock();
1708         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1709                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1710                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
1711                         struct block_device *bdev = rdev->bdev;
1712                         request_queue_t *r_queue = bdev_get_queue(bdev);
1713
1714                         if (!r_queue->issue_flush_fn)
1715                                 ret = -EOPNOTSUPP;
1716                         else {
1717                                 atomic_inc(&rdev->nr_pending);
1718                                 rcu_read_unlock();
1719                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1720                                                               error_sector);
1721                                 rdev_dec_pending(rdev, mddev);
1722                                 rcu_read_lock();
1723                         }
1724                 }
1725         }
1726         rcu_read_unlock();
1727         return ret;
1728 }
1729
1730 static inline void raid5_plug_device(raid5_conf_t *conf)
1731 {
1732         spin_lock_irq(&conf->device_lock);
1733         blk_plug_device(conf->mddev->queue);
1734         spin_unlock_irq(&conf->device_lock);
1735 }
1736
1737 static int make_request(request_queue_t *q, struct bio * bi)
1738 {
1739         mddev_t *mddev = q->queuedata;
1740         raid5_conf_t *conf = mddev_to_conf(mddev);
1741         unsigned int dd_idx, pd_idx;
1742         sector_t new_sector;
1743         sector_t logical_sector, last_sector;
1744         struct stripe_head *sh;
1745         const int rw = bio_data_dir(bi);
1746         int remaining;
1747
1748         if (unlikely(bio_barrier(bi))) {
1749                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1750                 return 0;
1751         }
1752
1753         md_write_start(mddev, bi);
1754
1755         disk_stat_inc(mddev->gendisk, ios[rw]);
1756         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
1757
1758         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1759         last_sector = bi->bi_sector + (bi->bi_size>>9);
1760         bi->bi_next = NULL;
1761         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1762
1763         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1764                 DEFINE_WAIT(w);
1765                 int disks;
1766
1767         retry:
1768                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1769                 if (likely(conf->expand_progress == MaxSector))
1770                         disks = conf->raid_disks;
1771                 else {
1772                         /* spinlock is needed as expand_progress may be
1773                          * 64bit on a 32bit platform, and so it might be
1774                          * possible to see a half-updated value
1775                          * Ofcourse expand_progress could change after
1776                          * the lock is dropped, so once we get a reference
1777                          * to the stripe that we think it is, we will have
1778                          * to check again.
1779                          */
1780                         spin_lock_irq(&conf->device_lock);
1781                         disks = conf->raid_disks;
1782                         if (logical_sector >= conf->expand_progress)
1783                                 disks = conf->previous_raid_disks;
1784                         else {
1785                                 if (logical_sector >= conf->expand_lo) {
1786                                         spin_unlock_irq(&conf->device_lock);
1787                                         schedule();
1788                                         goto retry;
1789                                 }
1790                         }
1791                         spin_unlock_irq(&conf->device_lock);
1792                 }
1793                 new_sector = raid5_compute_sector(logical_sector, disks, disks - 1,
1794                                                   &dd_idx, &pd_idx, conf);
1795                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
1796                         (unsigned long long)new_sector,
1797                         (unsigned long long)logical_sector);
1798
1799                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
1800                 if (sh) {
1801                         if (unlikely(conf->expand_progress != MaxSector)) {
1802                                 /* expansion might have moved on while waiting for a
1803                                  * stripe, so we must do the range check again.
1804                                  * Expansion could still move past after this
1805                                  * test, but as we are holding a reference to
1806                                  * 'sh', we know that if that happens,
1807                                  *  STRIPE_EXPANDING will get set and the expansion
1808                                  * won't proceed until we finish with the stripe.
1809                                  */
1810                                 int must_retry = 0;
1811                                 spin_lock_irq(&conf->device_lock);
1812                                 if (logical_sector <  conf->expand_progress &&
1813                                     disks == conf->previous_raid_disks)
1814                                         /* mismatch, need to try again */
1815                                         must_retry = 1;
1816                                 spin_unlock_irq(&conf->device_lock);
1817                                 if (must_retry) {
1818                                         release_stripe(sh);
1819                                         goto retry;
1820                                 }
1821                         }
1822                         /* FIXME what if we get a false positive because these
1823                          * are being updated.
1824                          */
1825                         if (logical_sector >= mddev->suspend_lo &&
1826                             logical_sector < mddev->suspend_hi) {
1827                                 release_stripe(sh);
1828                                 schedule();
1829                                 goto retry;
1830                         }
1831
1832                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
1833                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1834                                 /* Stripe is busy expanding or
1835                                  * add failed due to overlap.  Flush everything
1836                                  * and wait a while
1837                                  */
1838                                 raid5_unplug_device(mddev->queue);
1839                                 release_stripe(sh);
1840                                 schedule();
1841                                 goto retry;
1842                         }
1843                         finish_wait(&conf->wait_for_overlap, &w);
1844                         raid5_plug_device(conf);
1845                         handle_stripe(sh);
1846                         release_stripe(sh);
1847                 } else {
1848                         /* cannot get stripe for read-ahead, just give-up */
1849                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1850                         finish_wait(&conf->wait_for_overlap, &w);
1851                         break;
1852                 }
1853
1854         }
1855         spin_lock_irq(&conf->device_lock);
1856         remaining = --bi->bi_phys_segments;
1857         spin_unlock_irq(&conf->device_lock);
1858         if (remaining == 0) {
1859                 int bytes = bi->bi_size;
1860
1861                 if ( bio_data_dir(bi) == WRITE )
1862                         md_write_end(mddev);
1863                 bi->bi_size = 0;
1864                 bi->bi_end_io(bi, bytes, 0);
1865         }
1866         return 0;
1867 }
1868
1869 /* FIXME go_faster isn't used */
1870 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1871 {
1872         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1873         struct stripe_head *sh;
1874         int pd_idx;
1875         sector_t first_sector, last_sector;
1876         int raid_disks = conf->raid_disks;
1877         int data_disks = raid_disks-1;
1878         sector_t max_sector = mddev->size << 1;
1879         int sync_blocks;
1880
1881         if (sector_nr >= max_sector) {
1882                 /* just being told to finish up .. nothing much to do */
1883                 unplug_slaves(mddev);
1884                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
1885                         end_reshape(conf);
1886                         return 0;
1887                 }
1888
1889                 if (mddev->curr_resync < max_sector) /* aborted */
1890                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1891                                         &sync_blocks, 1);
1892                 else /* compelted sync */
1893                         conf->fullsync = 0;
1894                 bitmap_close_sync(mddev->bitmap);
1895
1896                 return 0;
1897         }
1898
1899         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
1900                 /* reshaping is quite different to recovery/resync so it is
1901                  * handled quite separately ... here.
1902                  *
1903                  * On each call to sync_request, we gather one chunk worth of
1904                  * destination stripes and flag them as expanding.
1905                  * Then we find all the source stripes and request reads.
1906                  * As the reads complete, handle_stripe will copy the data
1907                  * into the destination stripe and release that stripe.
1908                  */
1909                 int i;
1910                 int dd_idx;
1911                 sector_t writepos, safepos, gap;
1912
1913                 if (sector_nr == 0 &&
1914                     conf->expand_progress != 0) {
1915                         /* restarting in the middle, skip the initial sectors */
1916                         sector_nr = conf->expand_progress;
1917                         sector_div(sector_nr, conf->raid_disks-1);
1918                         *skipped = 1;
1919                         return sector_nr;
1920                 }
1921
1922                 /* we update the metadata when there is more than 3Meg
1923                  * in the block range (that is rather arbitrary, should
1924                  * probably be time based) or when the data about to be
1925                  * copied would over-write the source of the data at
1926                  * the front of the range.
1927                  * i.e. one new_stripe forward from expand_progress new_maps
1928                  * to after where expand_lo old_maps to
1929                  */
1930                 writepos = conf->expand_progress +
1931                         conf->chunk_size/512*(conf->raid_disks-1);
1932                 sector_div(writepos, conf->raid_disks-1);
1933                 safepos = conf->expand_lo;
1934                 sector_div(safepos, conf->previous_raid_disks-1);
1935                 gap = conf->expand_progress - conf->expand_lo;
1936
1937                 if (writepos >= safepos ||
1938                     gap > (conf->raid_disks-1)*3000*2 /*3Meg*/) {
1939                         /* Cannot proceed until we've updated the superblock... */
1940                         wait_event(conf->wait_for_overlap,
1941                                    atomic_read(&conf->reshape_stripes)==0);
1942                         mddev->reshape_position = conf->expand_progress;
1943                         mddev->sb_dirty = 1;
1944                         md_wakeup_thread(mddev->thread);
1945                         wait_event(mddev->sb_wait, mddev->sb_dirty == 0 ||
1946                                    kthread_should_stop());
1947                         spin_lock_irq(&conf->device_lock);
1948                         conf->expand_lo = mddev->reshape_position;
1949                         spin_unlock_irq(&conf->device_lock);
1950                         wake_up(&conf->wait_for_overlap);
1951                 }
1952
1953                 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
1954                         int j;
1955                         int skipped = 0;
1956                         pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
1957                         sh = get_active_stripe(conf, sector_nr+i,
1958                                                conf->raid_disks, pd_idx, 0);
1959                         set_bit(STRIPE_EXPANDING, &sh->state);
1960                         atomic_inc(&conf->reshape_stripes);
1961                         /* If any of this stripe is beyond the end of the old
1962                          * array, then we need to zero those blocks
1963                          */
1964                         for (j=sh->disks; j--;) {
1965                                 sector_t s;
1966                                 if (j == sh->pd_idx)
1967                                         continue;
1968                                 s = compute_blocknr(sh, j);
1969                                 if (s < (mddev->array_size<<1)) {
1970                                         skipped = 1;
1971                                         continue;
1972                                 }
1973                                 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
1974                                 set_bit(R5_Expanded, &sh->dev[j].flags);
1975                                 set_bit(R5_UPTODATE, &sh->dev[j].flags);
1976                         }
1977                         if (!skipped) {
1978                                 set_bit(STRIPE_EXPAND_READY, &sh->state);
1979                                 set_bit(STRIPE_HANDLE, &sh->state);
1980                         }
1981                         release_stripe(sh);
1982                 }
1983                 spin_lock_irq(&conf->device_lock);
1984                 conf->expand_progress = (sector_nr + i)*(conf->raid_disks-1);
1985                 spin_unlock_irq(&conf->device_lock);
1986                 /* Ok, those stripe are ready. We can start scheduling
1987                  * reads on the source stripes.
1988                  * The source stripes are determined by mapping the first and last
1989                  * block on the destination stripes.
1990                  */
1991                 raid_disks = conf->previous_raid_disks;
1992                 data_disks = raid_disks - 1;
1993                 first_sector =
1994                         raid5_compute_sector(sector_nr*(conf->raid_disks-1),
1995                                              raid_disks, data_disks,
1996                                              &dd_idx, &pd_idx, conf);
1997                 last_sector =
1998                         raid5_compute_sector((sector_nr+conf->chunk_size/512)
1999                                                *(conf->raid_disks-1) -1,
2000                                              raid_disks, data_disks,
2001                                              &dd_idx, &pd_idx, conf);
2002                 if (last_sector >= (mddev->size<<1))
2003                         last_sector = (mddev->size<<1)-1;
2004                 while (first_sector <= last_sector) {
2005                         pd_idx = stripe_to_pdidx(first_sector, conf, conf->previous_raid_disks);
2006                         sh = get_active_stripe(conf, first_sector,
2007                                                conf->previous_raid_disks, pd_idx, 0);
2008                         set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2009                         set_bit(STRIPE_HANDLE, &sh->state);
2010                         release_stripe(sh);
2011                         first_sector += STRIPE_SECTORS;
2012                 }
2013                 return conf->chunk_size>>9;
2014         }
2015         /* if there is 1 or more failed drives and we are trying
2016          * to resync, then assert that we are finished, because there is
2017          * nothing we can do.
2018          */
2019         if (mddev->degraded >= 1 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2020                 sector_t rv = (mddev->size << 1) - sector_nr;
2021                 *skipped = 1;
2022                 return rv;
2023         }
2024         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2025             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2026             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
2027                 /* we can skip this block, and probably more */
2028                 sync_blocks /= STRIPE_SECTORS;
2029                 *skipped = 1;
2030                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
2031         }
2032
2033         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
2034         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
2035         if (sh == NULL) {
2036                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
2037                 /* make sure we don't swamp the stripe cache if someone else
2038                  * is trying to get access
2039                  */
2040                 schedule_timeout_uninterruptible(1);
2041         }
2042         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 0);
2043         spin_lock(&sh->lock);
2044         set_bit(STRIPE_SYNCING, &sh->state);
2045         clear_bit(STRIPE_INSYNC, &sh->state);
2046         spin_unlock(&sh->lock);
2047
2048         handle_stripe(sh);
2049         release_stripe(sh);
2050
2051         return STRIPE_SECTORS;
2052 }
2053
2054 /*
2055  * This is our raid5 kernel thread.
2056  *
2057  * We scan the hash table for stripes which can be handled now.
2058  * During the scan, completed stripes are saved for us by the interrupt
2059  * handler, so that they will not have to wait for our next wakeup.
2060  */
2061 static void raid5d (mddev_t *mddev)
2062 {
2063         struct stripe_head *sh;
2064         raid5_conf_t *conf = mddev_to_conf(mddev);
2065         int handled;
2066
2067         PRINTK("+++ raid5d active\n");
2068
2069         md_check_recovery(mddev);
2070
2071         handled = 0;
2072         spin_lock_irq(&conf->device_lock);
2073         while (1) {
2074                 struct list_head *first;
2075
2076                 if (conf->seq_flush - conf->seq_write > 0) {
2077                         int seq = conf->seq_flush;
2078                         spin_unlock_irq(&conf->device_lock);
2079                         bitmap_unplug(mddev->bitmap);
2080                         spin_lock_irq(&conf->device_lock);
2081                         conf->seq_write = seq;
2082                         activate_bit_delay(conf);
2083                 }
2084
2085                 if (list_empty(&conf->handle_list) &&
2086                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
2087                     !blk_queue_plugged(mddev->queue) &&
2088                     !list_empty(&conf->delayed_list))
2089                         raid5_activate_delayed(conf);
2090
2091                 if (list_empty(&conf->handle_list))
2092                         break;
2093
2094                 first = conf->handle_list.next;
2095                 sh = list_entry(first, struct stripe_head, lru);
2096
2097                 list_del_init(first);
2098                 atomic_inc(&sh->count);
2099                 if (atomic_read(&sh->count)!= 1)
2100                         BUG();
2101                 spin_unlock_irq(&conf->device_lock);
2102
2103                 handled++;
2104                 handle_stripe(sh);
2105                 release_stripe(sh);
2106
2107                 spin_lock_irq(&conf->device_lock);
2108         }
2109         PRINTK("%d stripes handled\n", handled);
2110
2111         spin_unlock_irq(&conf->device_lock);
2112
2113         unplug_slaves(mddev);
2114
2115         PRINTK("--- raid5d inactive\n");
2116 }
2117
2118 static ssize_t
2119 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
2120 {
2121         raid5_conf_t *conf = mddev_to_conf(mddev);
2122         if (conf)
2123                 return sprintf(page, "%d\n", conf->max_nr_stripes);
2124         else
2125                 return 0;
2126 }
2127
2128 static ssize_t
2129 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
2130 {
2131         raid5_conf_t *conf = mddev_to_conf(mddev);
2132         char *end;
2133         int new;
2134         if (len >= PAGE_SIZE)
2135                 return -EINVAL;
2136         if (!conf)
2137                 return -ENODEV;
2138
2139         new = simple_strtoul(page, &end, 10);
2140         if (!*page || (*end && *end != '\n') )
2141                 return -EINVAL;
2142         if (new <= 16 || new > 32768)
2143                 return -EINVAL;
2144         while (new < conf->max_nr_stripes) {
2145                 if (drop_one_stripe(conf))
2146                         conf->max_nr_stripes--;
2147                 else
2148                         break;
2149         }
2150         while (new > conf->max_nr_stripes) {
2151                 if (grow_one_stripe(conf))
2152                         conf->max_nr_stripes++;
2153                 else break;
2154         }
2155         return len;
2156 }
2157
2158 static struct md_sysfs_entry
2159 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
2160                                 raid5_show_stripe_cache_size,
2161                                 raid5_store_stripe_cache_size);
2162
2163 static ssize_t
2164 stripe_cache_active_show(mddev_t *mddev, char *page)
2165 {
2166         raid5_conf_t *conf = mddev_to_conf(mddev);
2167         if (conf)
2168                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
2169         else
2170                 return 0;
2171 }
2172
2173 static struct md_sysfs_entry
2174 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
2175
2176 static struct attribute *raid5_attrs[] =  {
2177         &raid5_stripecache_size.attr,
2178         &raid5_stripecache_active.attr,
2179         NULL,
2180 };
2181 static struct attribute_group raid5_attrs_group = {
2182         .name = NULL,
2183         .attrs = raid5_attrs,
2184 };
2185
2186 static int run(mddev_t *mddev)
2187 {
2188         raid5_conf_t *conf;
2189         int raid_disk, memory;
2190         mdk_rdev_t *rdev;
2191         struct disk_info *disk;
2192         struct list_head *tmp;
2193
2194         if (mddev->level != 5 && mddev->level != 4) {
2195                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5 (%d)\n",
2196                        mdname(mddev), mddev->level);
2197                 return -EIO;
2198         }
2199
2200         if (mddev->reshape_position != MaxSector) {
2201                 /* Check that we can continue the reshape.
2202                  * Currently only disks can change, it must
2203                  * increase, and we must be past the point where
2204                  * a stripe over-writes itself
2205                  */
2206                 sector_t here_new, here_old;
2207                 int old_disks;
2208
2209                 if (mddev->new_level != mddev->level ||
2210                     mddev->new_layout != mddev->layout ||
2211                     mddev->new_chunk != mddev->chunk_size) {
2212                         printk(KERN_ERR "raid5: %s: unsupported reshape required - aborting.\n",
2213                                mdname(mddev));
2214                         return -EINVAL;
2215                 }
2216                 if (mddev->delta_disks <= 0) {
2217                         printk(KERN_ERR "raid5: %s: unsupported reshape (reduce disks) required - aborting.\n",
2218                                mdname(mddev));
2219                         return -EINVAL;
2220                 }
2221                 old_disks = mddev->raid_disks - mddev->delta_disks;
2222                 /* reshape_position must be on a new-stripe boundary, and one
2223                  * further up in new geometry must map after here in old geometry.
2224                  */
2225                 here_new = mddev->reshape_position;
2226                 if (sector_div(here_new, (mddev->chunk_size>>9)*(mddev->raid_disks-1))) {
2227                         printk(KERN_ERR "raid5: reshape_position not on a stripe boundary\n");
2228                         return -EINVAL;
2229                 }
2230                 /* here_new is the stripe we will write to */
2231                 here_old = mddev->reshape_position;
2232                 sector_div(here_old, (mddev->chunk_size>>9)*(old_disks-1));
2233                 /* here_old is the first stripe that we might need to read from */
2234                 if (here_new >= here_old) {
2235                         /* Reading from the same stripe as writing to - bad */
2236                         printk(KERN_ERR "raid5: reshape_position too early for auto-recovery - aborting.\n");
2237                         return -EINVAL;
2238                 }
2239                 printk(KERN_INFO "raid5: reshape will continue\n");
2240                 /* OK, we should be able to continue; */
2241         }
2242
2243
2244         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
2245         if ((conf = mddev->private) == NULL)
2246                 goto abort;
2247         if (mddev->reshape_position == MaxSector) {
2248                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
2249         } else {
2250                 conf->raid_disks = mddev->raid_disks;
2251                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
2252         }
2253
2254         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
2255                               GFP_KERNEL);
2256         if (!conf->disks)
2257                 goto abort;
2258
2259         conf->mddev = mddev;
2260
2261         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
2262                 goto abort;
2263
2264         spin_lock_init(&conf->device_lock);
2265         init_waitqueue_head(&conf->wait_for_stripe);
2266         init_waitqueue_head(&conf->wait_for_overlap);
2267         INIT_LIST_HEAD(&conf->handle_list);
2268         INIT_LIST_HEAD(&conf->delayed_list);
2269         INIT_LIST_HEAD(&conf->bitmap_list);
2270         INIT_LIST_HEAD(&conf->inactive_list);
2271         atomic_set(&conf->active_stripes, 0);
2272         atomic_set(&conf->preread_active_stripes, 0);
2273
2274         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
2275
2276         ITERATE_RDEV(mddev,rdev,tmp) {
2277                 raid_disk = rdev->raid_disk;
2278                 if (raid_disk >= conf->raid_disks
2279                     || raid_disk < 0)
2280                         continue;
2281                 disk = conf->disks + raid_disk;
2282
2283                 disk->rdev = rdev;
2284
2285                 if (test_bit(In_sync, &rdev->flags)) {
2286                         char b[BDEVNAME_SIZE];
2287                         printk(KERN_INFO "raid5: device %s operational as raid"
2288                                 " disk %d\n", bdevname(rdev->bdev,b),
2289                                 raid_disk);
2290                         conf->working_disks++;
2291                 }
2292         }
2293
2294         /*
2295          * 0 for a fully functional array, 1 for a degraded array.
2296          */
2297         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
2298         conf->mddev = mddev;
2299         conf->chunk_size = mddev->chunk_size;
2300         conf->level = mddev->level;
2301         conf->algorithm = mddev->layout;
2302         conf->max_nr_stripes = NR_STRIPES;
2303         conf->expand_progress = mddev->reshape_position;
2304
2305         /* device size must be a multiple of chunk size */
2306         mddev->size &= ~(mddev->chunk_size/1024 -1);
2307         mddev->resync_max_sectors = mddev->size << 1;
2308
2309         if (!conf->chunk_size || conf->chunk_size % 4) {
2310                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
2311                         conf->chunk_size, mdname(mddev));
2312                 goto abort;
2313         }
2314         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
2315                 printk(KERN_ERR
2316                         "raid5: unsupported parity algorithm %d for %s\n",
2317                         conf->algorithm, mdname(mddev));
2318                 goto abort;
2319         }
2320         if (mddev->degraded > 1) {
2321                 printk(KERN_ERR "raid5: not enough operational devices for %s"
2322                         " (%d/%d failed)\n",
2323                         mdname(mddev), conf->failed_disks, conf->raid_disks);
2324                 goto abort;
2325         }
2326
2327         if (mddev->degraded == 1 &&
2328             mddev->recovery_cp != MaxSector) {
2329                 if (mddev->ok_start_degraded)
2330                         printk(KERN_WARNING
2331                                "raid5: starting dirty degraded array: %s"
2332                                "- data corruption possible.\n",
2333                                mdname(mddev));
2334                 else {
2335                         printk(KERN_ERR
2336                                "raid5: cannot start dirty degraded array for %s\n",
2337                                mdname(mddev));
2338                         goto abort;
2339                 }
2340         }
2341
2342         {
2343                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
2344                 if (!mddev->thread) {
2345                         printk(KERN_ERR
2346                                 "raid5: couldn't allocate thread for %s\n",
2347                                 mdname(mddev));
2348                         goto abort;
2349                 }
2350         }
2351         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
2352                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
2353         if (grow_stripes(conf, conf->max_nr_stripes)) {
2354                 printk(KERN_ERR
2355                         "raid5: couldn't allocate %dkB for buffers\n", memory);
2356                 shrink_stripes(conf);
2357                 md_unregister_thread(mddev->thread);
2358                 goto abort;
2359         } else
2360                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
2361                         memory, mdname(mddev));
2362
2363         if (mddev->degraded == 0)
2364                 printk("raid5: raid level %d set %s active with %d out of %d"
2365                         " devices, algorithm %d\n", conf->level, mdname(mddev),
2366                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
2367                         conf->algorithm);
2368         else
2369                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
2370                         " out of %d devices, algorithm %d\n", conf->level,
2371                         mdname(mddev), mddev->raid_disks - mddev->degraded,
2372                         mddev->raid_disks, conf->algorithm);
2373
2374         print_raid5_conf(conf);
2375
2376         if (conf->expand_progress != MaxSector) {
2377                 printk("...ok start reshape thread\n");
2378                 conf->expand_lo = conf->expand_progress;
2379                 atomic_set(&conf->reshape_stripes, 0);
2380                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
2381                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
2382                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
2383                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
2384                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
2385                                                         "%s_reshape");
2386                 /* FIXME if md_register_thread fails?? */
2387                 md_wakeup_thread(mddev->sync_thread);
2388
2389         }
2390
2391         /* read-ahead size must cover two whole stripes, which is
2392          * 2 * (n-1) * chunksize where 'n' is the number of raid devices
2393          */
2394         {
2395                 int stripe = (mddev->raid_disks-1) * mddev->chunk_size
2396                         / PAGE_SIZE;
2397                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
2398                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
2399         }
2400
2401         /* Ok, everything is just fine now */
2402         sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
2403
2404         mddev->queue->unplug_fn = raid5_unplug_device;
2405         mddev->queue->issue_flush_fn = raid5_issue_flush;
2406         mddev->array_size =  mddev->size * (conf->previous_raid_disks - 1);
2407
2408         return 0;
2409 abort:
2410         if (conf) {
2411                 print_raid5_conf(conf);
2412                 kfree(conf->disks);
2413                 kfree(conf->stripe_hashtbl);
2414                 kfree(conf);
2415         }
2416         mddev->private = NULL;
2417         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
2418         return -EIO;
2419 }
2420
2421
2422
2423 static int stop(mddev_t *mddev)
2424 {
2425         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2426
2427         md_unregister_thread(mddev->thread);
2428         mddev->thread = NULL;
2429         shrink_stripes(conf);
2430         kfree(conf->stripe_hashtbl);
2431         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2432         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
2433         kfree(conf->disks);
2434         kfree(conf);
2435         mddev->private = NULL;
2436         return 0;
2437 }
2438
2439 #if RAID5_DEBUG
2440 static void print_sh (struct stripe_head *sh)
2441 {
2442         int i;
2443
2444         printk("sh %llu, pd_idx %d, state %ld.\n",
2445                 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
2446         printk("sh %llu,  count %d.\n",
2447                 (unsigned long long)sh->sector, atomic_read(&sh->count));
2448         printk("sh %llu, ", (unsigned long long)sh->sector);
2449         for (i = 0; i < sh->disks; i++) {
2450                 printk("(cache%d: %p %ld) ",
2451                         i, sh->dev[i].page, sh->dev[i].flags);
2452         }
2453         printk("\n");
2454 }
2455
2456 static void printall (raid5_conf_t *conf)
2457 {
2458         struct stripe_head *sh;
2459         struct hlist_node *hn;
2460         int i;
2461
2462         spin_lock_irq(&conf->device_lock);
2463         for (i = 0; i < NR_HASH; i++) {
2464                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
2465                         if (sh->raid_conf != conf)
2466                                 continue;
2467                         print_sh(sh);
2468                 }
2469         }
2470         spin_unlock_irq(&conf->device_lock);
2471 }
2472 #endif
2473
2474 static void status (struct seq_file *seq, mddev_t *mddev)
2475 {
2476         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2477         int i;
2478
2479         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2480         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2481         for (i = 0; i < conf->raid_disks; i++)
2482                 seq_printf (seq, "%s",
2483                                conf->disks[i].rdev &&
2484                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
2485         seq_printf (seq, "]");
2486 #if RAID5_DEBUG
2487 #define D(x) \
2488         seq_printf (seq, "<"#x":%d>", atomic_read(&conf->x))
2489         printall(conf);
2490 #endif
2491 }
2492
2493 static void print_raid5_conf (raid5_conf_t *conf)
2494 {
2495         int i;
2496         struct disk_info *tmp;
2497
2498         printk("RAID5 conf printout:\n");
2499         if (!conf) {
2500                 printk("(conf==NULL)\n");
2501                 return;
2502         }
2503         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2504                  conf->working_disks, conf->failed_disks);
2505
2506         for (i = 0; i < conf->raid_disks; i++) {
2507                 char b[BDEVNAME_SIZE];
2508                 tmp = conf->disks + i;
2509                 if (tmp->rdev)
2510                 printk(" disk %d, o:%d, dev:%s\n",
2511                         i, !test_bit(Faulty, &tmp->rdev->flags),
2512                         bdevname(tmp->rdev->bdev,b));
2513         }
2514 }
2515
2516 static int raid5_spare_active(mddev_t *mddev)
2517 {
2518         int i;
2519         raid5_conf_t *conf = mddev->private;
2520         struct disk_info *tmp;
2521
2522         for (i = 0; i < conf->raid_disks; i++) {
2523                 tmp = conf->disks + i;
2524                 if (tmp->rdev
2525                     && !test_bit(Faulty, &tmp->rdev->flags)
2526                     && !test_bit(In_sync, &tmp->rdev->flags)) {
2527                         mddev->degraded--;
2528                         conf->failed_disks--;
2529                         conf->working_disks++;
2530                         set_bit(In_sync, &tmp->rdev->flags);
2531                 }
2532         }
2533         print_raid5_conf(conf);
2534         return 0;
2535 }
2536
2537 static int raid5_remove_disk(mddev_t *mddev, int number)
2538 {
2539         raid5_conf_t *conf = mddev->private;
2540         int err = 0;
2541         mdk_rdev_t *rdev;
2542         struct disk_info *p = conf->disks + number;
2543
2544         print_raid5_conf(conf);
2545         rdev = p->rdev;
2546         if (rdev) {
2547                 if (test_bit(In_sync, &rdev->flags) ||
2548                     atomic_read(&rdev->nr_pending)) {
2549                         err = -EBUSY;
2550                         goto abort;
2551                 }
2552                 p->rdev = NULL;
2553                 synchronize_rcu();
2554                 if (atomic_read(&rdev->nr_pending)) {
2555                         /* lost the race, try later */
2556                         err = -EBUSY;
2557                         p->rdev = rdev;
2558                 }
2559         }
2560 abort:
2561
2562         print_raid5_conf(conf);
2563         return err;
2564 }
2565
2566 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2567 {
2568         raid5_conf_t *conf = mddev->private;
2569         int found = 0;
2570         int disk;
2571         struct disk_info *p;
2572
2573         if (mddev->degraded > 1)
2574                 /* no point adding a device */
2575                 return 0;
2576
2577         /*
2578          * find the disk ...
2579          */
2580         for (disk=0; disk < conf->raid_disks; disk++)
2581                 if ((p=conf->disks + disk)->rdev == NULL) {
2582                         clear_bit(In_sync, &rdev->flags);
2583                         rdev->raid_disk = disk;
2584                         found = 1;
2585                         if (rdev->saved_raid_disk != disk)
2586                                 conf->fullsync = 1;
2587                         rcu_assign_pointer(p->rdev, rdev);
2588                         break;
2589                 }
2590         print_raid5_conf(conf);
2591         return found;
2592 }
2593
2594 static int raid5_resize(mddev_t *mddev, sector_t sectors)
2595 {
2596         /* no resync is happening, and there is enough space
2597          * on all devices, so we can resize.
2598          * We need to make sure resync covers any new space.
2599          * If the array is shrinking we should possibly wait until
2600          * any io in the removed space completes, but it hardly seems
2601          * worth it.
2602          */
2603         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2604         mddev->array_size = (sectors * (mddev->raid_disks-1))>>1;
2605         set_capacity(mddev->gendisk, mddev->array_size << 1);
2606         mddev->changed = 1;
2607         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2608                 mddev->recovery_cp = mddev->size << 1;
2609                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2610         }
2611         mddev->size = sectors /2;
2612         mddev->resync_max_sectors = sectors;
2613         return 0;
2614 }
2615
2616 #ifdef CONFIG_MD_RAID5_RESHAPE
2617 static int raid5_check_reshape(mddev_t *mddev)
2618 {
2619         raid5_conf_t *conf = mddev_to_conf(mddev);
2620         int err;
2621
2622         if (mddev->delta_disks < 0 ||
2623             mddev->new_level != mddev->level)
2624                 return -EINVAL; /* Cannot shrink array or change level yet */
2625         if (mddev->delta_disks == 0)
2626                 return 0; /* nothing to do */
2627
2628         /* Can only proceed if there are plenty of stripe_heads.
2629          * We need a minimum of one full stripe,, and for sensible progress
2630          * it is best to have about 4 times that.
2631          * If we require 4 times, then the default 256 4K stripe_heads will
2632          * allow for chunk sizes up to 256K, which is probably OK.
2633          * If the chunk size is greater, user-space should request more
2634          * stripe_heads first.
2635          */
2636         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
2637             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
2638                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
2639                        (mddev->chunk_size / STRIPE_SIZE)*4);
2640                 return -ENOSPC;
2641         }
2642
2643         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
2644         if (err)
2645                 return err;
2646
2647         /* looks like we might be able to manage this */
2648         return 0;
2649 }
2650
2651 static int raid5_start_reshape(mddev_t *mddev)
2652 {
2653         raid5_conf_t *conf = mddev_to_conf(mddev);
2654         mdk_rdev_t *rdev;
2655         struct list_head *rtmp;
2656         int spares = 0;
2657         int added_devices = 0;
2658
2659         if (mddev->degraded ||
2660             test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
2661                 return -EBUSY;
2662
2663         ITERATE_RDEV(mddev, rdev, rtmp)
2664                 if (rdev->raid_disk < 0 &&
2665                     !test_bit(Faulty, &rdev->flags))
2666                         spares++;
2667
2668         if (spares < mddev->delta_disks-1)
2669                 /* Not enough devices even to make a degraded array
2670                  * of that size
2671                  */
2672                 return -EINVAL;
2673
2674         atomic_set(&conf->reshape_stripes, 0);
2675         spin_lock_irq(&conf->device_lock);
2676         conf->previous_raid_disks = conf->raid_disks;
2677         conf->raid_disks += mddev->delta_disks;
2678         conf->expand_progress = 0;
2679         conf->expand_lo = 0;
2680         spin_unlock_irq(&conf->device_lock);
2681
2682         /* Add some new drives, as many as will fit.
2683          * We know there are enough to make the newly sized array work.
2684          */
2685         ITERATE_RDEV(mddev, rdev, rtmp)
2686                 if (rdev->raid_disk < 0 &&
2687                     !test_bit(Faulty, &rdev->flags)) {
2688                         if (raid5_add_disk(mddev, rdev)) {
2689                                 char nm[20];
2690                                 set_bit(In_sync, &rdev->flags);
2691                                 conf->working_disks++;
2692                                 added_devices++;
2693                                 sprintf(nm, "rd%d", rdev->raid_disk);
2694                                 sysfs_create_link(&mddev->kobj, &rdev->kobj, nm);
2695                         } else
2696                                 break;
2697                 }
2698
2699         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
2700         mddev->raid_disks = conf->raid_disks;
2701         mddev->reshape_position = 0;
2702         mddev->sb_dirty = 1;
2703
2704         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
2705         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
2706         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
2707         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
2708         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
2709                                                 "%s_reshape");
2710         if (!mddev->sync_thread) {
2711                 mddev->recovery = 0;
2712                 spin_lock_irq(&conf->device_lock);
2713                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
2714                 conf->expand_progress = MaxSector;
2715                 spin_unlock_irq(&conf->device_lock);
2716                 return -EAGAIN;
2717         }
2718         md_wakeup_thread(mddev->sync_thread);
2719         md_new_event(mddev);
2720         return 0;
2721 }
2722 #endif
2723
2724 static void end_reshape(raid5_conf_t *conf)
2725 {
2726         struct block_device *bdev;
2727
2728         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
2729                 conf->mddev->array_size = conf->mddev->size * (conf->raid_disks-1);
2730                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
2731                 conf->mddev->changed = 1;
2732
2733                 bdev = bdget_disk(conf->mddev->gendisk, 0);
2734                 if (bdev) {
2735                         mutex_lock(&bdev->bd_inode->i_mutex);
2736                         i_size_write(bdev->bd_inode, conf->mddev->array_size << 10);
2737                         mutex_unlock(&bdev->bd_inode->i_mutex);
2738                         bdput(bdev);
2739                 }
2740                 spin_lock_irq(&conf->device_lock);
2741                 conf->expand_progress = MaxSector;
2742                 spin_unlock_irq(&conf->device_lock);
2743                 conf->mddev->reshape_position = MaxSector;
2744         }
2745 }
2746
2747 static void raid5_quiesce(mddev_t *mddev, int state)
2748 {
2749         raid5_conf_t *conf = mddev_to_conf(mddev);
2750
2751         switch(state) {
2752         case 2: /* resume for a suspend */
2753                 wake_up(&conf->wait_for_overlap);
2754                 break;
2755
2756         case 1: /* stop all writes */
2757                 spin_lock_irq(&conf->device_lock);
2758                 conf->quiesce = 1;
2759                 wait_event_lock_irq(conf->wait_for_stripe,
2760                                     atomic_read(&conf->active_stripes) == 0,
2761                                     conf->device_lock, /* nothing */);
2762                 spin_unlock_irq(&conf->device_lock);
2763                 break;
2764
2765         case 0: /* re-enable writes */
2766                 spin_lock_irq(&conf->device_lock);
2767                 conf->quiesce = 0;
2768                 wake_up(&conf->wait_for_stripe);
2769                 wake_up(&conf->wait_for_overlap);
2770                 spin_unlock_irq(&conf->device_lock);
2771                 break;
2772         }
2773 }
2774
2775 static struct mdk_personality raid5_personality =
2776 {
2777         .name           = "raid5",
2778         .level          = 5,
2779         .owner          = THIS_MODULE,
2780         .make_request   = make_request,
2781         .run            = run,
2782         .stop           = stop,
2783         .status         = status,
2784         .error_handler  = error,
2785         .hot_add_disk   = raid5_add_disk,
2786         .hot_remove_disk= raid5_remove_disk,
2787         .spare_active   = raid5_spare_active,
2788         .sync_request   = sync_request,
2789         .resize         = raid5_resize,
2790 #ifdef CONFIG_MD_RAID5_RESHAPE
2791         .check_reshape  = raid5_check_reshape,
2792         .start_reshape  = raid5_start_reshape,
2793 #endif
2794         .quiesce        = raid5_quiesce,
2795 };
2796
2797 static struct mdk_personality raid4_personality =
2798 {
2799         .name           = "raid4",
2800         .level          = 4,
2801         .owner          = THIS_MODULE,
2802         .make_request   = make_request,
2803         .run            = run,
2804         .stop           = stop,
2805         .status         = status,
2806         .error_handler  = error,
2807         .hot_add_disk   = raid5_add_disk,
2808         .hot_remove_disk= raid5_remove_disk,
2809         .spare_active   = raid5_spare_active,
2810         .sync_request   = sync_request,
2811         .resize         = raid5_resize,
2812         .quiesce        = raid5_quiesce,
2813 };
2814
2815 static int __init raid5_init(void)
2816 {
2817         register_md_personality(&raid5_personality);
2818         register_md_personality(&raid4_personality);
2819         return 0;
2820 }
2821
2822 static void raid5_exit(void)
2823 {
2824         unregister_md_personality(&raid5_personality);
2825         unregister_md_personality(&raid4_personality);
2826 }
2827
2828 module_init(raid5_init);
2829 module_exit(raid5_exit);
2830 MODULE_LICENSE("GPL");
2831 MODULE_ALIAS("md-personality-4"); /* RAID5 */
2832 MODULE_ALIAS("md-raid5");
2833 MODULE_ALIAS("md-raid4");
2834 MODULE_ALIAS("md-level-5");
2835 MODULE_ALIAS("md-level-4");