drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * This program is free software; you can redistribute it and/or modify
  16  * it under the terms of the GNU General Public License as published by
  17  * the Free Software Foundation; either version 2, or (at your option)
  18  * any later version.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * (for example /usr/src/linux/COPYING); if not, write to the Free
  22  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  23  */
  24
  25 #include <linux/raid/raid1.h>
  26
  27 #define MAJOR_NR MD_MAJOR
  28 #define MD_DRIVER
  29 #define MD_PERSONALITY
  30
  31 /*
  32  * Number of guaranteed r1bios in case of extreme VM load:
  33  */
  34 #define NR_RAID1_BIOS 256
  35
  36 static mdk_personality_t raid1_personality;
  37 static spinlock_t retry_list_lock = SPIN_LOCK_UNLOCKED;
  38 static LIST_HEAD(retry_list_head);
  39
  40 static void unplug_slaves(mddev_t *mddev);
  41
  42
  43 static void * r1bio_pool_alloc(int gfp_flags, void *data)
  44 {
  45         struct pool_info *pi = data;
  46         r1bio_t *r1_bio;
  47
  48         /* allocate a r1bio with room for raid_disks entries in the bios array */
  49         r1_bio = kmalloc(sizeof(r1bio_t) + sizeof(struct bio*)*pi->raid_disks,
  50                          gfp_flags);
  51         if (r1_bio)
  52                 memset(r1_bio, 0, sizeof(*r1_bio) +
  53                                sizeof(struct bio*) * pi->raid_disks);
  54         else
  55                 unplug_slaves(pi->mddev);
  56
  57         return r1_bio;
  58 }
  59
  60 static void r1bio_pool_free(void *r1_bio, void *data)
  61 {
  62         kfree(r1_bio);
  63 }
  64
  65 #define RESYNC_BLOCK_SIZE (64*1024)
  66 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  67 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  68 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  69 #define RESYNC_WINDOW (2048*1024)
  70
  71 static void * r1buf_pool_alloc(int gfp_flags, void *data)
  72 {
  73         struct pool_info *pi = data;
  74         struct page *page;
  75         r1bio_t *r1_bio;
  76         struct bio *bio;
  77         int i, j;
  78
  79         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  80         if (!r1_bio) {
  81                 unplug_slaves(pi->mddev);
  82                 return NULL;
  83         }
  84
  85         /*
  86          * Allocate bios : 1 for reading, n-1 for writing
  87          */
  88         for (j = pi->raid_disks ; j-- ; ) {
  89                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
  90                 if (!bio)
  91                         goto out_free_bio;
  92                 r1_bio->bios[j] = bio;
  93         }
  94         /*
  95          * Allocate RESYNC_PAGES data pages and attach them to
  96          * the first bio;
  97          */
  98         bio = r1_bio->bios[0];
  99         for (i = 0; i < RESYNC_PAGES; i++) {
 100                 page = alloc_page(gfp_flags);
 101                 if (unlikely(!page))
 102                         goto out_free_pages;
 103
 104                 bio->bi_io_vec[i].bv_page = page;
 105         }
 106
 107         r1_bio->master_bio = bio;
 108
 109         return r1_bio;
 110
 111 out_free_pages:
 112         for ( ; i > 0 ; i--)
 113                 __free_page(bio->bi_io_vec[i-1].bv_page);
 114 out_free_bio:
 115         while ( ++j < pi->raid_disks )
 116                 bio_put(r1_bio->bios[j]);
 117         r1bio_pool_free(r1_bio, data);
 118         return NULL;
 119 }
 120
 121 static void r1buf_pool_free(void *__r1_bio, void *data)
 122 {
 123         struct pool_info *pi = data;
 124         int i;
 125         r1bio_t *r1bio = __r1_bio;
 126         struct bio *bio = r1bio->bios[0];
 127
 128         for (i = 0; i < RESYNC_PAGES; i++) {
 129                 __free_page(bio->bi_io_vec[i].bv_page);
 130                 bio->bi_io_vec[i].bv_page = NULL;
 131         }
 132         for (i=0 ; i < pi->raid_disks; i++)
 133                 bio_put(r1bio->bios[i]);
 134
 135         r1bio_pool_free(r1bio, data);
 136 }
 137
 138 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 139 {
 140         int i;
 141
 142         for (i = 0; i < conf->raid_disks; i++) {
 143                 struct bio **bio = r1_bio->bios + i;
 144                 if (*bio)
 145                         bio_put(*bio);
 146                 *bio = NULL;
 147         }
 148 }
 149
 150 static inline void free_r1bio(r1bio_t *r1_bio)
 151 {
 152         unsigned long flags;
 153
 154         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 155
 156         /*
 157          * Wake up any possible resync thread that waits for the device
 158          * to go idle.
 159          */
 160         spin_lock_irqsave(&conf->resync_lock, flags);
 161         if (!--conf->nr_pending) {
 162                 wake_up(&conf->wait_idle);
 163                 wake_up(&conf->wait_resume);
 164         }
 165         spin_unlock_irqrestore(&conf->resync_lock, flags);
 166
 167         put_all_bios(conf, r1_bio);
 168         mempool_free(r1_bio, conf->r1bio_pool);
 169 }
 170
 171 static inline void put_buf(r1bio_t *r1_bio)
 172 {
 173         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 174         unsigned long flags;
 175
 176         mempool_free(r1_bio, conf->r1buf_pool);
 177
 178         spin_lock_irqsave(&conf->resync_lock, flags);
 179         if (!conf->barrier)
 180                 BUG();
 181         --conf->barrier;
 182         wake_up(&conf->wait_resume);
 183         wake_up(&conf->wait_idle);
 184
 185         if (!--conf->nr_pending) {
 186                 wake_up(&conf->wait_idle);
 187                 wake_up(&conf->wait_resume);
 188         }
 189         spin_unlock_irqrestore(&conf->resync_lock, flags);
 190 }
 191
 192 static int map(mddev_t *mddev, mdk_rdev_t **rdevp)
 193 {
 194         conf_t *conf = mddev_to_conf(mddev);
 195         int i, disks = conf->raid_disks;
 196
 197         /*
 198          * Later we do read balancing on the read side
 199          * now we use the first available disk.
 200          */
 201
 202         spin_lock_irq(&conf->device_lock);
 203         for (i = 0; i < disks; i++) {
 204                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 205                 if (rdev && rdev->in_sync) {
 206                         *rdevp = rdev;
 207                         atomic_inc(&rdev->nr_pending);
 208                         spin_unlock_irq(&conf->device_lock);
 209                         return i;
 210                 }
 211         }
 212         spin_unlock_irq(&conf->device_lock);
 213
 214         printk(KERN_ERR "raid1_map(): huh, no more operational devices?\n");
 215         return -1;
 216 }
 217
 218 static void reschedule_retry(r1bio_t *r1_bio)
 219 {
 220         unsigned long flags;
 221         mddev_t *mddev = r1_bio->mddev;
 222
 223         spin_lock_irqsave(&retry_list_lock, flags);
 224         list_add(&r1_bio->retry_list, &retry_list_head);
 225         spin_unlock_irqrestore(&retry_list_lock, flags);
 226
 227         md_wakeup_thread(mddev->thread);
 228 }
 229
 230 /*
 231  * raid_end_bio_io() is called when we have finished servicing a mirrored
 232  * operation and are ready to return a success/failure code to the buffer
 233  * cache layer.
 234  */
 235 static void raid_end_bio_io(r1bio_t *r1_bio)
 236 {
 237         struct bio *bio = r1_bio->master_bio;
 238
 239         bio_endio(bio, bio->bi_size,
 240                 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 241         free_r1bio(r1_bio);
 242 }
 243
 244 /*
 245  * Update disk head position estimator based on IRQ completion info.
 246  */
 247 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 248 {
 249         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 250
 251         conf->mirrors[disk].head_position =
 252                 r1_bio->sector + (r1_bio->sectors);
 253 }
 254
 255 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
 256 {
 257         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 258         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 259         int mirror;
 260         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 261
 262         if (bio->bi_size)
 263                 return 1;
 264
 265         mirror = r1_bio->read_disk;
 266         /*
 267          * this branch is our 'one mirror IO has finished' event handler:
 268          */
 269         if (!uptodate)
 270                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 271         else
 272                 /*
 273                  * Set R1BIO_Uptodate in our master bio, so that
 274                  * we will return a good error code for to the higher
 275                  * levels even if IO on some other mirrored buffer fails.
 276                  *
 277                  * The 'master' represents the composite IO operation to
 278                  * user-side. So if something waits for IO, then it will
 279                  * wait for the 'master' bio.
 280                  */
 281                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 282
 283         update_head_pos(mirror, r1_bio);
 284
 285         /*
 286          * we have only one bio on the read side
 287          */
 288         if (uptodate)
 289                 raid_end_bio_io(r1_bio);
 290         else {
 291                 /*
 292                  * oops, read error:
 293                  */
 294                 char b[BDEVNAME_SIZE];
 295                 printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 296                        bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 297                 reschedule_retry(r1_bio);
 298         }
 299
 300         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 301         return 0;
 302 }
 303
 304 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 305 {
 306         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 307         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 308         int mirror;
 309         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 310
 311         if (bio->bi_size)
 312                 return 1;
 313
 314         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 315                 if (r1_bio->bios[mirror] == bio)
 316                         break;
 317
 318         /*
 319          * this branch is our 'one mirror IO has finished' event handler:
 320          */
 321         if (!uptodate)
 322                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 323         else
 324                 /*
 325                  * Set R1BIO_Uptodate in our master bio, so that
 326                  * we will return a good error code for to the higher
 327                  * levels even if IO on some other mirrored buffer fails.
 328                  *
 329                  * The 'master' represents the composite IO operation to
 330                  * user-side. So if something waits for IO, then it will
 331                  * wait for the 'master' bio.
 332                  */
 333                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 334
 335         update_head_pos(mirror, r1_bio);
 336
 337         /*
 338          *
 339          * Let's see if all mirrored write operations have finished
 340          * already.
 341          */
 342         if (atomic_dec_and_test(&r1_bio->remaining)) {
 343                 md_write_end(r1_bio->mddev);
 344                 raid_end_bio_io(r1_bio);
 345         }
 346
 347         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 348         return 0;
 349 }
 350
 351
 352 /*
 353  * This routine returns the disk from which the requested read should
 354  * be done. There is a per-array 'next expected sequential IO' sector
 355  * number - if this matches on the next IO then we use the last disk.
 356  * There is also a per-disk 'last know head position' sector that is
 357  * maintained from IRQ contexts, both the normal and the resync IO
 358  * completion handlers update this position correctly. If there is no
 359  * perfect sequential match then we pick the disk whose head is closest.
 360  *
 361  * If there are 2 mirrors in the same 2 devices, performance degrades
 362  * because position is mirror, not device based.
 363  *
 364  * The rdev for the device selected will have nr_pending incremented.
 365  */
 366 static int read_balance(conf_t *conf, struct bio *bio, r1bio_t *r1_bio)
 367 {
 368         const unsigned long this_sector = r1_bio->sector;
 369         int new_disk = conf->last_used, disk = new_disk;
 370         const int sectors = bio->bi_size >> 9;
 371         sector_t new_distance, current_distance;
 372
 373         spin_lock_irq(&conf->device_lock);
 374         /*
 375          * Check if it if we can balance. We can balance on the whole
 376          * device if no resync is going on, or below the resync window.
 377          * We take the first readable disk when above the resync window.
 378          */
 379         if (conf->mddev->recovery_cp < MaxSector &&
 380             (this_sector + sectors >= conf->next_resync)) {
 381                 /* make sure that disk is operational */
 382                 new_disk = 0;
 383
 384                 while (!conf->mirrors[new_disk].rdev ||
 385                        !conf->mirrors[new_disk].rdev->in_sync) {
 386                         new_disk++;
 387                         if (new_disk == conf->raid_disks) {
 388                                 new_disk = 0;
 389                                 break;
 390                         }
 391                 }
 392                 goto rb_out;
 393         }
 394
 395
 396         /* make sure the disk is operational */
 397         while (!conf->mirrors[new_disk].rdev ||
 398                !conf->mirrors[new_disk].rdev->in_sync) {
 399                 if (new_disk <= 0)
 400                         new_disk = conf->raid_disks;
 401                 new_disk--;
 402                 if (new_disk == disk) {
 403                         new_disk = conf->last_used;
 404                         goto rb_out;
 405                 }
 406         }
 407         disk = new_disk;
 408         /* now disk == new_disk == starting point for search */
 409
 410         /*
 411          * Don't change to another disk for sequential reads:
 412          */
 413         if (conf->next_seq_sect == this_sector)
 414                 goto rb_out;
 415         if (this_sector == conf->mirrors[new_disk].head_position)
 416                 goto rb_out;
 417
 418         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 419
 420         /* Find the disk whose head is closest */
 421
 422         do {
 423                 if (disk <= 0)
 424                         disk = conf->raid_disks;
 425                 disk--;
 426
 427                 if (!conf->mirrors[disk].rdev ||
 428                     !conf->mirrors[disk].rdev->in_sync)
 429                         continue;
 430
 431                 if (!atomic_read(&conf->mirrors[disk].rdev->nr_pending)) {
 432                         new_disk = disk;
 433                         break;
 434                 }
 435                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 436                 if (new_distance < current_distance) {
 437                         current_distance = new_distance;
 438                         new_disk = disk;
 439                 }
 440         } while (disk != conf->last_used);
 441
 442 rb_out:
 443         r1_bio->read_disk = new_disk;
 444         conf->next_seq_sect = this_sector + sectors;
 445
 446         conf->last_used = new_disk;
 447
 448         if (conf->mirrors[new_disk].rdev)
 449                 atomic_inc(&conf->mirrors[new_disk].rdev->nr_pending);
 450         spin_unlock_irq(&conf->device_lock);
 451
 452         return new_disk;
 453 }
 454
 455 static void unplug_slaves(mddev_t *mddev)
 456 {
 457         conf_t *conf = mddev_to_conf(mddev);
 458         int i;
 459         unsigned long flags;
 460
 461         spin_lock_irqsave(&conf->device_lock, flags);
 462         for (i=0; i<mddev->raid_disks; i++) {
 463                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 464                 if (rdev && atomic_read(&rdev->nr_pending)) {
 465                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 466
 467                         atomic_inc(&rdev->nr_pending);
 468                         spin_unlock_irqrestore(&conf->device_lock, flags);
 469
 470                         if (r_queue->unplug_fn)
 471                                 r_queue->unplug_fn(r_queue);
 472
 473                         spin_lock_irqsave(&conf->device_lock, flags);
 474                         atomic_dec(&rdev->nr_pending);
 475                 }
 476         }
 477         spin_unlock_irqrestore(&conf->device_lock, flags);
 478 }
 479 static void raid1_unplug(request_queue_t *q)
 480 {
 481         unplug_slaves(q->queuedata);
 482 }
 483
 484 /*
 485  * Throttle resync depth, so that we can both get proper overlapping of
 486  * requests, but are still able to handle normal requests quickly.
 487  */
 488 #define RESYNC_DEPTH 32
 489
 490 static void device_barrier(conf_t *conf, sector_t sect)
 491 {
 492         spin_lock_irq(&conf->resync_lock);
 493         wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
 494                             conf->resync_lock, unplug_slaves(conf->mddev));
 495
 496         if (!conf->barrier++) {
 497                 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
 498                                     conf->resync_lock, unplug_slaves(conf->mddev));
 499                 if (conf->nr_pending)
 500                         BUG();
 501         }
 502         wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
 503                             conf->resync_lock, unplug_slaves(conf->mddev));
 504         conf->next_resync = sect;
 505         spin_unlock_irq(&conf->resync_lock);
 506 }
 507
 508 static int make_request(request_queue_t *q, struct bio * bio)
 509 {
 510         mddev_t *mddev = q->queuedata;
 511         conf_t *conf = mddev_to_conf(mddev);
 512         mirror_info_t *mirror;
 513         r1bio_t *r1_bio;
 514         struct bio *read_bio;
 515         int i, disks;
 516
 517         /*
 518          * Register the new request and wait if the reconstruction
 519          * thread has put up a bar for new requests.
 520          * Continue immediately if no resync is active currently.
 521          */
 522         spin_lock_irq(&conf->resync_lock);
 523         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
 524         conf->nr_pending++;
 525         spin_unlock_irq(&conf->resync_lock);
 526
 527         if (bio_data_dir(bio)==WRITE) {
 528                 disk_stat_inc(mddev->gendisk, writes);
 529                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
 530         } else {
 531                 disk_stat_inc(mddev->gendisk, reads);
 532                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
 533         }
 534
 535         /*
 536          * make_request() can abort the operation when READA is being
 537          * used and no empty request is available.
 538          *
 539          */
 540         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 541
 542         r1_bio->master_bio = bio;
 543         r1_bio->sectors = bio->bi_size >> 9;
 544
 545         r1_bio->mddev = mddev;
 546         r1_bio->sector = bio->bi_sector;
 547
 548         if (bio_data_dir(bio) == READ) {
 549                 /*
 550                  * read balancing logic:
 551                  */
 552                 mirror = conf->mirrors + read_balance(conf, bio, r1_bio);
 553
 554                 read_bio = bio_clone(bio, GFP_NOIO);
 555
 556                 r1_bio->bios[r1_bio->read_disk] = read_bio;
 557
 558                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 559                 read_bio->bi_bdev = mirror->rdev->bdev;
 560                 read_bio->bi_end_io = raid1_end_read_request;
 561                 read_bio->bi_rw = READ;
 562                 read_bio->bi_private = r1_bio;
 563
 564                 generic_make_request(read_bio);
 565                 return 0;
 566         }
 567
 568         /*
 569          * WRITE:
 570          */
 571         /* first select target devices under spinlock and
 572          * inc refcount on their rdev.  Record them by setting
 573          * bios[x] to bio
 574          */
 575         disks = conf->raid_disks;
 576         spin_lock_irq(&conf->device_lock);
 577         for (i = 0;  i < disks; i++) {
 578                 if (conf->mirrors[i].rdev &&
 579                     !conf->mirrors[i].rdev->faulty) {
 580                         atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 581                         r1_bio->bios[i] = bio;
 582                 } else
 583                         r1_bio->bios[i] = NULL;
 584         }
 585         spin_unlock_irq(&conf->device_lock);
 586
 587         atomic_set(&r1_bio->remaining, 1);
 588         md_write_start(mddev);
 589         for (i = 0; i < disks; i++) {
 590                 struct bio *mbio;
 591                 if (!r1_bio->bios[i])
 592                         continue;
 593
 594                 mbio = bio_clone(bio, GFP_NOIO);
 595                 r1_bio->bios[i] = mbio;
 596
 597                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 598                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 599                 mbio->bi_end_io = raid1_end_write_request;
 600                 mbio->bi_rw = WRITE;
 601                 mbio->bi_private = r1_bio;
 602
 603                 atomic_inc(&r1_bio->remaining);
 604                 generic_make_request(mbio);
 605         }
 606
 607         if (atomic_dec_and_test(&r1_bio->remaining)) {
 608                 md_write_end(mddev);
 609                 raid_end_bio_io(r1_bio);
 610         }
 611
 612         return 0;
 613 }
 614
 615 static void status(struct seq_file *seq, mddev_t *mddev)
 616 {
 617         conf_t *conf = mddev_to_conf(mddev);
 618         int i;
 619
 620         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 621                                                 conf->working_disks);
 622         for (i = 0; i < conf->raid_disks; i++)
 623                 seq_printf(seq, "%s",
 624                               conf->mirrors[i].rdev &&
 625                               conf->mirrors[i].rdev->in_sync ? "U" : "_");
 626         seq_printf(seq, "]");
 627 }
 628
 629
 630 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 631 {
 632         char b[BDEVNAME_SIZE];
 633         conf_t *conf = mddev_to_conf(mddev);
 634
 635         /*
 636          * If it is not operational, then we have already marked it as dead
 637          * else if it is the last working disks, ignore the error, let the
 638          * next level up know.
 639          * else mark the drive as failed
 640          */
 641         if (rdev->in_sync
 642             && conf->working_disks == 1)
 643                 /*
 644                  * Don't fail the drive, act as though we were just a
 645                  * normal single drive
 646                  */
 647                 return;
 648         if (rdev->in_sync) {
 649                 mddev->degraded++;
 650                 conf->working_disks--;
 651                 /*
 652                  * if recovery is running, make sure it aborts.
 653                  */
 654                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 655         }
 656         rdev->in_sync = 0;
 657         rdev->faulty = 1;
 658         mddev->sb_dirty = 1;
 659         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 660                 "       Operation continuing on %d devices\n",
 661                 bdevname(rdev->bdev,b), conf->working_disks);
 662 }
 663
 664 static void print_conf(conf_t *conf)
 665 {
 666         int i;
 667         mirror_info_t *tmp;
 668
 669         printk("RAID1 conf printout:\n");
 670         if (!conf) {
 671                 printk("(!conf)\n");
 672                 return;
 673         }
 674         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 675                 conf->raid_disks);
 676
 677         for (i = 0; i < conf->raid_disks; i++) {
 678                 char b[BDEVNAME_SIZE];
 679                 tmp = conf->mirrors + i;
 680                 if (tmp->rdev)
 681                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 682                                 i, !tmp->rdev->in_sync, !tmp->rdev->faulty,
 683                                 bdevname(tmp->rdev->bdev,b));
 684         }
 685 }
 686
 687 static void close_sync(conf_t *conf)
 688 {
 689         spin_lock_irq(&conf->resync_lock);
 690         wait_event_lock_irq(conf->wait_resume, !conf->barrier,
 691                             conf->resync_lock,  unplug_slaves(conf->mddev));
 692         spin_unlock_irq(&conf->resync_lock);
 693
 694         if (conf->barrier) BUG();
 695         if (waitqueue_active(&conf->wait_idle)) BUG();
 696
 697         mempool_destroy(conf->r1buf_pool);
 698         conf->r1buf_pool = NULL;
 699 }
 700
 701 static int raid1_spare_active(mddev_t *mddev)
 702 {
 703         int i;
 704         conf_t *conf = mddev->private;
 705         mirror_info_t *tmp;
 706
 707         spin_lock_irq(&conf->device_lock);
 708         /*
 709          * Find all failed disks within the RAID1 configuration
 710          * and mark them readable
 711          */
 712         for (i = 0; i < conf->raid_disks; i++) {
 713                 tmp = conf->mirrors + i;
 714                 if (tmp->rdev
 715                     && !tmp->rdev->faulty
 716                     && !tmp->rdev->in_sync) {
 717                         conf->working_disks++;
 718                         mddev->degraded--;
 719                         tmp->rdev->in_sync = 1;
 720                 }
 721         }
 722         spin_unlock_irq(&conf->device_lock);
 723
 724         print_conf(conf);
 725         return 0;
 726 }
 727
 728
 729 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 730 {
 731         conf_t *conf = mddev->private;
 732         int found = 0;
 733         int mirror;
 734         mirror_info_t *p;
 735
 736         spin_lock_irq(&conf->device_lock);
 737         for (mirror=0; mirror < mddev->raid_disks; mirror++)
 738                 if ( !(p=conf->mirrors+mirror)->rdev) {
 739                         p->rdev = rdev;
 740
 741                         blk_queue_stack_limits(mddev->queue,
 742                                                rdev->bdev->bd_disk->queue);
 743                         /* as we don't honour merge_bvec_fn, we must never risk
 744                          * violating it, so limit ->max_sector to one PAGE, as
 745                          * a one page request is never in violation.
 746                          */
 747                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 748                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 749                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
 750
 751                         p->head_position = 0;
 752                         rdev->raid_disk = mirror;
 753                         found = 1;
 754                         break;
 755                 }
 756         spin_unlock_irq(&conf->device_lock);
 757
 758         print_conf(conf);
 759         return found;
 760 }
 761
 762 static int raid1_remove_disk(mddev_t *mddev, int number)
 763 {
 764         conf_t *conf = mddev->private;
 765         int err = 1;
 766         mirror_info_t *p = conf->mirrors+ number;
 767
 768         print_conf(conf);
 769         spin_lock_irq(&conf->device_lock);
 770         if (p->rdev) {
 771                 if (p->rdev->in_sync ||
 772                     atomic_read(&p->rdev->nr_pending)) {
 773                         err = -EBUSY;
 774                         goto abort;
 775                 }
 776                 p->rdev = NULL;
 777                 err = 0;
 778         }
 779         if (err)
 780                 MD_BUG();
 781 abort:
 782         spin_unlock_irq(&conf->device_lock);
 783
 784         print_conf(conf);
 785         return err;
 786 }
 787
 788
 789 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
 790 {
 791         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 792         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 793         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 794
 795         if (bio->bi_size)
 796                 return 1;
 797
 798         if (r1_bio->bios[r1_bio->read_disk] != bio)
 799                 BUG();
 800         update_head_pos(r1_bio->read_disk, r1_bio);
 801         /*
 802          * we have read a block, now it needs to be re-written,
 803          * or re-read if the read failed.
 804          * We don't do much here, just schedule handling by raid1d
 805          */
 806         if (!uptodate)
 807                 md_error(r1_bio->mddev,
 808                          conf->mirrors[r1_bio->read_disk].rdev);
 809         else
 810                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 811         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
 812         reschedule_retry(r1_bio);
 813         return 0;
 814 }
 815
 816 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
 817 {
 818         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 819         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 820         mddev_t *mddev = r1_bio->mddev;
 821         conf_t *conf = mddev_to_conf(mddev);
 822         int i;
 823         int mirror=0;
 824
 825         if (bio->bi_size)
 826                 return 1;
 827
 828         for (i = 0; i < conf->raid_disks; i++)
 829                 if (r1_bio->bios[i] == bio) {
 830                         mirror = i;
 831                         break;
 832                 }
 833         if (!uptodate)
 834                 md_error(mddev, conf->mirrors[mirror].rdev);
 835         update_head_pos(mirror, r1_bio);
 836
 837         if (atomic_dec_and_test(&r1_bio->remaining)) {
 838                 md_done_sync(mddev, r1_bio->sectors, uptodate);
 839                 put_buf(r1_bio);
 840         }
 841         rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
 842         return 0;
 843 }
 844
 845 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
 846 {
 847         conf_t *conf = mddev_to_conf(mddev);
 848         int i;
 849         int disks = conf->raid_disks;
 850         struct bio *bio, *wbio;
 851
 852         bio = r1_bio->bios[r1_bio->read_disk];
 853
 854         /*
 855          * schedule writes
 856          */
 857         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 858                 /*
 859                  * There is no point trying a read-for-reconstruct as
 860                  * reconstruct is about to be aborted
 861                  */
 862                 char b[BDEVNAME_SIZE];
 863                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
 864                         " for block %llu\n",
 865                         bdevname(bio->bi_bdev,b),
 866                         (unsigned long long)r1_bio->sector);
 867                 md_done_sync(mddev, r1_bio->sectors, 0);
 868                 put_buf(r1_bio);
 869                 return;
 870         }
 871
 872         atomic_set(&r1_bio->remaining, 1);
 873         for (i = 0; i < disks ; i++) {
 874                 wbio = r1_bio->bios[i];
 875                 if (wbio->bi_end_io != end_sync_write)
 876                         continue;
 877
 878                 atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 879                 atomic_inc(&r1_bio->remaining);
 880                 md_sync_acct(conf->mirrors[i].rdev, wbio->bi_size >> 9);
 881                 generic_make_request(wbio);
 882         }
 883
 884         if (atomic_dec_and_test(&r1_bio->remaining)) {
 885                 md_done_sync(mddev, r1_bio->sectors, 1);
 886                 put_buf(r1_bio);
 887         }
 888 }
 889
 890 /*
 891  * This is a kernel thread which:
 892  *
 893  *      1.      Retries failed read operations on working mirrors.
 894  *      2.      Updates the raid superblock when problems encounter.
 895  *      3.      Performs writes following reads for array syncronising.
 896  */
 897
 898 static void raid1d(mddev_t *mddev)
 899 {
 900         struct list_head *head = &retry_list_head;
 901         r1bio_t *r1_bio;
 902         struct bio *bio;
 903         unsigned long flags;
 904         conf_t *conf = mddev_to_conf(mddev);
 905         int unplug=0;
 906         mdk_rdev_t *rdev;
 907
 908         md_check_recovery(mddev);
 909         md_handle_safemode(mddev);
 910
 911         for (;;) {
 912                 char b[BDEVNAME_SIZE];
 913                 spin_lock_irqsave(&retry_list_lock, flags);
 914                 if (list_empty(head))
 915                         break;
 916                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
 917                 list_del(head->prev);
 918                 spin_unlock_irqrestore(&retry_list_lock, flags);
 919
 920                 mddev = r1_bio->mddev;
 921                 conf = mddev_to_conf(mddev);
 922                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
 923                         sync_request_write(mddev, r1_bio);
 924                         unplug = 1;
 925                 } else {
 926                         int disk;
 927                         bio = r1_bio->bios[r1_bio->read_disk];
 928                         if ((disk=map(mddev, &rdev)) == -1) {
 929                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
 930                                        " read error for block %llu\n",
 931                                        bdevname(bio->bi_bdev,b),
 932                                        (unsigned long long)r1_bio->sector);
 933                                 raid_end_bio_io(r1_bio);
 934                         } else {
 935                                 r1_bio->bios[r1_bio->read_disk] = NULL;
 936                                 r1_bio->read_disk = disk;
 937                                 r1_bio->bios[r1_bio->read_disk] = bio;
 938                                 printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
 939                                        " another mirror\n",
 940                                        bdevname(rdev->bdev,b),
 941                                        (unsigned long long)r1_bio->sector);
 942                                 bio->bi_bdev = rdev->bdev;
 943                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
 944                                 bio->bi_rw = READ;
 945                                 unplug = 1;
 946                                 generic_make_request(bio);
 947                         }
 948                 }
 949         }
 950         spin_unlock_irqrestore(&retry_list_lock, flags);
 951         if (unplug)
 952                 unplug_slaves(mddev);
 953 }
 954
 955
 956 static int init_resync(conf_t *conf)
 957 {
 958         int buffs;
 959
 960         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
 961         if (conf->r1buf_pool)
 962                 BUG();
 963         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
 964                                           conf->poolinfo);
 965         if (!conf->r1buf_pool)
 966                 return -ENOMEM;
 967         conf->next_resync = 0;
 968         return 0;
 969 }
 970
 971 /*
 972  * perform a "sync" on one "block"
 973  *
 974  * We need to make sure that no normal I/O request - particularly write
 975  * requests - conflict with active sync requests.
 976  *
 977  * This is achieved by tracking pending requests and a 'barrier' concept
 978  * that can be installed to exclude normal IO requests.
 979  */
 980
 981 static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
 982 {
 983         conf_t *conf = mddev_to_conf(mddev);
 984         mirror_info_t *mirror;
 985         r1bio_t *r1_bio;
 986         struct bio *bio;
 987         sector_t max_sector, nr_sectors;
 988         int disk;
 989         int i;
 990         int write_targets = 0;
 991
 992         if (!conf->r1buf_pool)
 993                 if (init_resync(conf))
 994                         return -ENOMEM;
 995
 996         max_sector = mddev->size << 1;
 997         if (sector_nr >= max_sector) {
 998                 close_sync(conf);
 999                 return 0;
1000         }
1001
1002         /*
1003          * If there is non-resync activity waiting for us then
1004          * put in a delay to throttle resync.
1005          */
1006         if (!go_faster && waitqueue_active(&conf->wait_resume))
1007                 schedule_timeout(HZ);
1008         device_barrier(conf, sector_nr + RESYNC_SECTORS);
1009
1010         /*
1011          * If reconstructing, and >1 working disc,
1012          * could dedicate one to rebuild and others to
1013          * service read requests ..
1014          */
1015         disk = conf->last_used;
1016         /* make sure disk is operational */
1017         spin_lock_irq(&conf->device_lock);
1018         while (conf->mirrors[disk].rdev == NULL ||
1019                !conf->mirrors[disk].rdev->in_sync) {
1020                 if (disk <= 0)
1021                         disk = conf->raid_disks;
1022                 disk--;
1023                 if (disk == conf->last_used)
1024                         break;
1025         }
1026         conf->last_used = disk;
1027         atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
1028         spin_unlock_irq(&conf->device_lock);
1029
1030         mirror = conf->mirrors + disk;
1031
1032         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1033
1034         spin_lock_irq(&conf->resync_lock);
1035         conf->nr_pending++;
1036         spin_unlock_irq(&conf->resync_lock);
1037
1038         r1_bio->mddev = mddev;
1039         r1_bio->sector = sector_nr;
1040         set_bit(R1BIO_IsSync, &r1_bio->state);
1041         r1_bio->read_disk = disk;
1042
1043         for (i=0; i < conf->raid_disks; i++) {
1044                 bio = r1_bio->bios[i];
1045
1046                 /* take from bio_init */
1047                 bio->bi_next = NULL;
1048                 bio->bi_flags |= 1 << BIO_UPTODATE;
1049                 bio->bi_rw = 0;
1050                 bio->bi_vcnt = 0;
1051                 bio->bi_idx = 0;
1052                 bio->bi_phys_segments = 0;
1053                 bio->bi_hw_segments = 0;
1054                 bio->bi_size = 0;
1055                 bio->bi_end_io = NULL;
1056                 bio->bi_private = NULL;
1057
1058                 if (i == disk) {
1059                         bio->bi_rw = READ;
1060                         bio->bi_end_io = end_sync_read;
1061                 } else if (conf->mirrors[i].rdev &&
1062                            !conf->mirrors[i].rdev->faulty &&
1063                            (!conf->mirrors[i].rdev->in_sync ||
1064                             sector_nr + RESYNC_SECTORS > mddev->recovery_cp)) {
1065                         bio->bi_rw = WRITE;
1066                         bio->bi_end_io = end_sync_write;
1067                         write_targets ++;
1068                 } else
1069                         continue;
1070                 bio->bi_sector = sector_nr + conf->mirrors[i].rdev->data_offset;
1071                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1072                 bio->bi_private = r1_bio;
1073         }
1074         if (write_targets == 0) {
1075                 /* There is nowhere to write, so all non-sync
1076                  * drives must be failed - so we are finished
1077                  */
1078                 int rv = max_sector - sector_nr;
1079                 md_done_sync(mddev, rv, 1);
1080                 put_buf(r1_bio);
1081                 atomic_dec(&conf->mirrors[disk].rdev->nr_pending);
1082                 return rv;
1083         }
1084
1085         nr_sectors = 0;
1086         do {
1087                 struct page *page;
1088                 int len = PAGE_SIZE;
1089                 if (sector_nr + (len>>9) > max_sector)
1090                         len = (max_sector - sector_nr) << 9;
1091                 if (len == 0)
1092                         break;
1093                 for (i=0 ; i < conf->raid_disks; i++) {
1094                         bio = r1_bio->bios[i];
1095                         if (bio->bi_end_io) {
1096                                 page = r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page;
1097                                 if (bio_add_page(bio, page, len, 0) == 0) {
1098                                         /* stop here */
1099                                         r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page = page;
1100                                         while (i > 0) {
1101                                                 i--;
1102                                                 bio = r1_bio->bios[i];
1103                                                 if (bio->bi_end_io==NULL) continue;
1104                                                 /* remove last page from this bio */
1105                                                 bio->bi_vcnt--;
1106                                                 bio->bi_size -= len;
1107                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1108                                         }
1109                                         goto bio_full;
1110                                 }
1111                         }
1112                 }
1113                 nr_sectors += len>>9;
1114                 sector_nr += len>>9;
1115         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1116  bio_full:
1117         bio = r1_bio->bios[disk];
1118         r1_bio->sectors = nr_sectors;
1119
1120         md_sync_acct(mirror->rdev, nr_sectors);
1121
1122         generic_make_request(bio);
1123
1124         return nr_sectors;
1125 }
1126
1127 static int run(mddev_t *mddev)
1128 {
1129         conf_t *conf;
1130         int i, j, disk_idx;
1131         mirror_info_t *disk;
1132         mdk_rdev_t *rdev;
1133         struct list_head *tmp;
1134
1135         if (mddev->level != 1) {
1136                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1137                        mdname(mddev), mddev->level);
1138                 goto out;
1139         }
1140         /*
1141          * copy the already verified devices into our private RAID1
1142          * bookkeeping area. [whatever we allocate in run(),
1143          * should be freed in stop()]
1144          */
1145         conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1146         mddev->private = conf;
1147         if (!conf)
1148                 goto out_no_mem;
1149
1150         memset(conf, 0, sizeof(*conf));
1151         conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1152                                  GFP_KERNEL);
1153         if (!conf->mirrors)
1154                 goto out_no_mem;
1155
1156         memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1157
1158         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1159         if (!conf->poolinfo)
1160                 goto out_no_mem;
1161         conf->poolinfo->mddev = mddev;
1162         conf->poolinfo->raid_disks = mddev->raid_disks;
1163         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1164                                           r1bio_pool_free,
1165                                           conf->poolinfo);
1166         if (!conf->r1bio_pool)
1167                 goto out_no_mem;
1168
1169         mddev->queue->unplug_fn = raid1_unplug;
1170
1171
1172         ITERATE_RDEV(mddev, rdev, tmp) {
1173                 disk_idx = rdev->raid_disk;
1174                 if (disk_idx >= mddev->raid_disks
1175                     || disk_idx < 0)
1176                         continue;
1177                 disk = conf->mirrors + disk_idx;
1178
1179                 disk->rdev = rdev;
1180
1181                 blk_queue_stack_limits(mddev->queue,
1182                                        rdev->bdev->bd_disk->queue);
1183                 /* as we don't honour merge_bvec_fn, we must never risk
1184                  * violating it, so limit ->max_sector to one PAGE, as
1185                  * a one page request is never in violation.
1186                  */
1187                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1188                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1189                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
1190
1191                 disk->head_position = 0;
1192                 if (!rdev->faulty && rdev->in_sync)
1193                         conf->working_disks++;
1194         }
1195         conf->raid_disks = mddev->raid_disks;
1196         conf->mddev = mddev;
1197         conf->device_lock = SPIN_LOCK_UNLOCKED;
1198         if (conf->working_disks == 1)
1199                 mddev->recovery_cp = MaxSector;
1200
1201         conf->resync_lock = SPIN_LOCK_UNLOCKED;
1202         init_waitqueue_head(&conf->wait_idle);
1203         init_waitqueue_head(&conf->wait_resume);
1204
1205         if (!conf->working_disks) {
1206                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1207                         mdname(mddev));
1208                 goto out_free_conf;
1209         }
1210
1211         mddev->degraded = 0;
1212         for (i = 0; i < conf->raid_disks; i++) {
1213
1214                 disk = conf->mirrors + i;
1215
1216                 if (!disk->rdev) {
1217                         disk->head_position = 0;
1218                         mddev->degraded++;
1219                 }
1220         }
1221
1222         /*
1223          * find the first working one and use it as a starting point
1224          * to read balancing.
1225          */
1226         for (j = 0; j < conf->raid_disks &&
1227                      (!conf->mirrors[j].rdev ||
1228                       !conf->mirrors[j].rdev->in_sync) ; j++)
1229                 /* nothing */;
1230         conf->last_used = j;
1231
1232
1233
1234         {
1235                 mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1236                 if (!mddev->thread) {
1237                         printk(KERN_ERR
1238                                 "raid1: couldn't allocate thread for %s\n",
1239                                 mdname(mddev));
1240                         goto out_free_conf;
1241                 }
1242         }
1243         printk(KERN_INFO
1244                 "raid1: raid set %s active with %d out of %d mirrors\n",
1245                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1246                 mddev->raid_disks);
1247         /*
1248          * Ok, everything is just fine now
1249          */
1250         mddev->array_size = mddev->size;
1251
1252         return 0;
1253
1254 out_no_mem:
1255         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1256                mdname(mddev));
1257
1258 out_free_conf:
1259         if (conf) {
1260                 if (conf->r1bio_pool)
1261                         mempool_destroy(conf->r1bio_pool);
1262                 if (conf->mirrors)
1263                         kfree(conf->mirrors);
1264                 if (conf->poolinfo)
1265                         kfree(conf->poolinfo);
1266                 kfree(conf);
1267                 mddev->private = NULL;
1268         }
1269 out:
1270         return -EIO;
1271 }
1272
1273 static int stop(mddev_t *mddev)
1274 {
1275         conf_t *conf = mddev_to_conf(mddev);
1276
1277         md_unregister_thread(mddev->thread);
1278         mddev->thread = NULL;
1279         if (conf->r1bio_pool)
1280                 mempool_destroy(conf->r1bio_pool);
1281         if (conf->mirrors)
1282                 kfree(conf->mirrors);
1283         if (conf->poolinfo)
1284                 kfree(conf->poolinfo);
1285         kfree(conf);
1286         mddev->private = NULL;
1287         return 0;
1288 }
1289
1290 static int raid1_resize(mddev_t *mddev, sector_t sectors)
1291 {
1292         /* no resync is happening, and there is enough space
1293          * on all devices, so we can resize.
1294          * We need to make sure resync covers any new space.
1295          * If the array is shrinking we should possibly wait until
1296          * any io in the removed space completes, but it hardly seems
1297          * worth it.
1298          */
1299         mddev->array_size = sectors>>1;
1300         set_capacity(mddev->gendisk, mddev->array_size << 1);
1301         mddev->changed = 1;
1302         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
1303                 mddev->recovery_cp = mddev->size << 1;
1304                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1305         }
1306         mddev->size = mddev->array_size;
1307         return 0;
1308 }
1309
1310 static int raid1_reshape(mddev_t *mddev, int raid_disks)
1311 {
1312         /* We need to:
1313          * 1/ resize the r1bio_pool
1314          * 2/ resize conf->mirrors
1315          *
1316          * We allocate a new r1bio_pool if we can.
1317          * Then raise a device barrier and wait until all IO stops.
1318          * Then resize conf->mirrors and swap in the new r1bio pool.
1319          */
1320         mempool_t *newpool, *oldpool;
1321         struct pool_info *newpoolinfo;
1322         mirror_info_t *newmirrors;
1323         conf_t *conf = mddev_to_conf(mddev);
1324
1325         int d;
1326
1327         for (d= raid_disks; d < conf->raid_disks; d++)
1328                 if (conf->mirrors[d].rdev)
1329                         return -EBUSY;
1330
1331         newpoolinfo = kmalloc(sizeof(newpoolinfo), GFP_KERNEL);
1332         if (!newpoolinfo)
1333                 return -ENOMEM;
1334         newpoolinfo->mddev = mddev;
1335         newpoolinfo->raid_disks = raid_disks;
1336
1337         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1338                                  r1bio_pool_free, newpoolinfo);
1339         if (!newpool) {
1340                 kfree(newpoolinfo);
1341                 return -ENOMEM;
1342         }
1343         newmirrors = kmalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
1344         if (!newmirrors) {
1345                 kfree(newpoolinfo);
1346                 mempool_destroy(newpool);
1347                 return -ENOMEM;
1348         }
1349         memset(newmirrors, 0, sizeof(struct mirror_info)*raid_disks);
1350
1351         spin_lock_irq(&conf->resync_lock);
1352         conf->barrier++;
1353         wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
1354                             conf->resync_lock, unplug_slaves(mddev));
1355         spin_unlock_irq(&conf->resync_lock);
1356
1357         /* ok, everything is stopped */
1358         oldpool = conf->r1bio_pool;
1359         conf->r1bio_pool = newpool;
1360         for (d=0; d < raid_disks && d < conf->raid_disks; d++)
1361                 newmirrors[d] = conf->mirrors[d];
1362         kfree(conf->mirrors);
1363         conf->mirrors = newmirrors;
1364         kfree(conf->poolinfo);
1365         conf->poolinfo = newpoolinfo;
1366
1367         mddev->degraded += (raid_disks - conf->raid_disks);
1368         conf->raid_disks = mddev->raid_disks = raid_disks;
1369
1370         spin_lock_irq(&conf->resync_lock);
1371         conf->barrier--;
1372         spin_unlock_irq(&conf->resync_lock);
1373         wake_up(&conf->wait_resume);
1374         wake_up(&conf->wait_idle);
1375
1376
1377         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1378         md_wakeup_thread(mddev->thread);
1379
1380         mempool_destroy(oldpool);
1381         return 0;
1382 }
1383
1384
1385 static mdk_personality_t raid1_personality =
1386 {
1387         .name           = "raid1",
1388         .owner          = THIS_MODULE,
1389         .make_request   = make_request,
1390         .run            = run,
1391         .stop           = stop,
1392         .status         = status,
1393         .error_handler  = error,
1394         .hot_add_disk   = raid1_add_disk,
1395         .hot_remove_disk= raid1_remove_disk,
1396         .spare_active   = raid1_spare_active,
1397         .sync_request   = sync_request,
1398         .resize         = raid1_resize,
1399         .reshape        = raid1_reshape,
1400 };
1401
1402 static int __init raid_init(void)
1403 {
1404         return register_md_personality(RAID1, &raid1_personality);
1405 }
1406
1407 static void raid_exit(void)
1408 {
1409         unregister_md_personality(RAID1);
1410 }
1411
1412 module_init(raid_init);
1413 module_exit(raid_exit);
1414 MODULE_LICENSE("GPL");
1415 MODULE_ALIAS("md-personality-3"); /* RAID1 */