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 0;
 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->in_sync && (this_sector + sectors >= conf->next_resync)) {
 380                 /* make sure that disk is operational */
 381                 new_disk = 0;
 382
 383                 while (!conf->mirrors[new_disk].rdev ||
 384                        !conf->mirrors[new_disk].rdev->in_sync) {
 385                         new_disk++;
 386                         if (new_disk == conf->raid_disks) {
 387                                 new_disk = 0;
 388                                 break;
 389                         }
 390                 }
 391                 goto rb_out;
 392         }
 393
 394
 395         /* make sure the disk is operational */
 396         while (!conf->mirrors[new_disk].rdev ||
 397                !conf->mirrors[new_disk].rdev->in_sync) {
 398                 if (new_disk <= 0)
 399                         new_disk = conf->raid_disks;
 400                 new_disk--;
 401                 if (new_disk == disk) {
 402                         new_disk = conf->last_used;
 403                         goto rb_out;
 404                 }
 405         }
 406         disk = new_disk;
 407         /* now disk == new_disk == starting point for search */
 408
 409         /*
 410          * Don't change to another disk for sequential reads:
 411          */
 412         if (conf->next_seq_sect == this_sector)
 413                 goto rb_out;
 414         if (this_sector == conf->mirrors[new_disk].head_position)
 415                 goto rb_out;
 416
 417         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 418
 419         /* Find the disk whose head is closest */
 420
 421         do {
 422                 if (disk <= 0)
 423                         disk = conf->raid_disks;
 424                 disk--;
 425
 426                 if (!conf->mirrors[disk].rdev ||
 427                     !conf->mirrors[disk].rdev->in_sync)
 428                         continue;
 429
 430                 if (!atomic_read(&conf->mirrors[disk].rdev->nr_pending)) {
 431                         new_disk = disk;
 432                         break;
 433                 }
 434                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 435                 if (new_distance < current_distance) {
 436                         current_distance = new_distance;
 437                         new_disk = disk;
 438                 }
 439         } while (disk != conf->last_used);
 440
 441 rb_out:
 442         r1_bio->read_disk = new_disk;
 443         conf->next_seq_sect = this_sector + sectors;
 444
 445         conf->last_used = new_disk;
 446
 447         if (conf->mirrors[new_disk].rdev)
 448                 atomic_inc(&conf->mirrors[new_disk].rdev->nr_pending);
 449         spin_unlock_irq(&conf->device_lock);
 450
 451         return new_disk;
 452 }
 453
 454 static void unplug_slaves(mddev_t *mddev)
 455 {
 456         conf_t *conf = mddev_to_conf(mddev);
 457         int i;
 458         unsigned long flags;
 459
 460         spin_lock_irqsave(&conf->device_lock, flags);
 461         for (i=0; i<mddev->raid_disks; i++) {
 462                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 463                 if (rdev && atomic_read(&rdev->nr_pending)) {
 464                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 465
 466                         atomic_inc(&rdev->nr_pending);
 467                         spin_unlock_irqrestore(&conf->device_lock, flags);
 468
 469                         if (r_queue->unplug_fn)
 470                                 r_queue->unplug_fn(r_queue);
 471
 472                         spin_lock_irqsave(&conf->device_lock, flags);
 473                         atomic_dec(&rdev->nr_pending);
 474                 }
 475         }
 476         spin_unlock_irqrestore(&conf->device_lock, flags);
 477 }
 478 static void raid1_unplug(request_queue_t *q)
 479 {
 480         unplug_slaves(q->queuedata);
 481 }
 482
 483 /*
 484  * Throttle resync depth, so that we can both get proper overlapping of
 485  * requests, but are still able to handle normal requests quickly.
 486  */
 487 #define RESYNC_DEPTH 32
 488
 489 static void device_barrier(conf_t *conf, sector_t sect)
 490 {
 491         spin_lock_irq(&conf->resync_lock);
 492         wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
 493                             conf->resync_lock, unplug_slaves(conf->mddev));
 494
 495         if (!conf->barrier++) {
 496                 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
 497                                     conf->resync_lock, unplug_slaves(conf->mddev));
 498                 if (conf->nr_pending)
 499                         BUG();
 500         }
 501         wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
 502                             conf->resync_lock, unplug_slaves(conf->mddev));
 503         conf->next_resync = sect;
 504         spin_unlock_irq(&conf->resync_lock);
 505 }
 506
 507 static int make_request(request_queue_t *q, struct bio * bio)
 508 {
 509         mddev_t *mddev = q->queuedata;
 510         conf_t *conf = mddev_to_conf(mddev);
 511         mirror_info_t *mirror;
 512         r1bio_t *r1_bio;
 513         struct bio *read_bio;
 514         int i, disks;
 515
 516         /*
 517          * Register the new request and wait if the reconstruction
 518          * thread has put up a bar for new requests.
 519          * Continue immediately if no resync is active currently.
 520          */
 521         spin_lock_irq(&conf->resync_lock);
 522         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
 523         conf->nr_pending++;
 524         spin_unlock_irq(&conf->resync_lock);
 525
 526         if (bio_data_dir(bio)==WRITE) {
 527                 disk_stat_inc(mddev->gendisk, writes);
 528                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
 529         } else {
 530                 disk_stat_inc(mddev->gendisk, reads);
 531                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
 532         }
 533
 534         /*
 535          * make_request() can abort the operation when READA is being
 536          * used and no empty request is available.
 537          *
 538          */
 539         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 540
 541         r1_bio->master_bio = bio;
 542         r1_bio->sectors = bio->bi_size >> 9;
 543
 544         r1_bio->mddev = mddev;
 545         r1_bio->sector = bio->bi_sector;
 546
 547         if (bio_data_dir(bio) == READ) {
 548                 /*
 549                  * read balancing logic:
 550                  */
 551                 mirror = conf->mirrors + read_balance(conf, bio, r1_bio);
 552
 553                 read_bio = bio_clone(bio, GFP_NOIO);
 554
 555                 r1_bio->bios[r1_bio->read_disk] = read_bio;
 556
 557                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 558                 read_bio->bi_bdev = mirror->rdev->bdev;
 559                 read_bio->bi_end_io = raid1_end_read_request;
 560                 read_bio->bi_rw = READ;
 561                 read_bio->bi_private = r1_bio;
 562
 563                 generic_make_request(read_bio);
 564                 return 0;
 565         }
 566
 567         /*
 568          * WRITE:
 569          */
 570         /* first select target devices under spinlock and
 571          * inc refcount on their rdev.  Record them by setting
 572          * bios[x] to bio
 573          */
 574         disks = conf->raid_disks;
 575         spin_lock_irq(&conf->device_lock);
 576         for (i = 0;  i < disks; i++) {
 577                 if (conf->mirrors[i].rdev &&
 578                     !conf->mirrors[i].rdev->faulty) {
 579                         atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 580                         r1_bio->bios[i] = bio;
 581                 } else
 582                         r1_bio->bios[i] = NULL;
 583         }
 584         spin_unlock_irq(&conf->device_lock);
 585
 586         atomic_set(&r1_bio->remaining, 1);
 587         md_write_start(mddev);
 588         for (i = 0; i < disks; i++) {
 589                 struct bio *mbio;
 590                 if (!r1_bio->bios[i])
 591                         continue;
 592
 593                 mbio = bio_clone(bio, GFP_NOIO);
 594                 r1_bio->bios[i] = mbio;
 595
 596                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 597                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 598                 mbio->bi_end_io = raid1_end_write_request;
 599                 mbio->bi_rw = WRITE;
 600                 mbio->bi_private = r1_bio;
 601
 602                 atomic_inc(&r1_bio->remaining);
 603                 generic_make_request(mbio);
 604         }
 605
 606         if (atomic_dec_and_test(&r1_bio->remaining)) {
 607                 md_write_end(mddev);
 608                 raid_end_bio_io(r1_bio);
 609         }
 610
 611         return 0;
 612 }
 613
 614 static void status(struct seq_file *seq, mddev_t *mddev)
 615 {
 616         conf_t *conf = mddev_to_conf(mddev);
 617         int i;
 618
 619         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 620                                                 conf->working_disks);
 621         for (i = 0; i < conf->raid_disks; i++)
 622                 seq_printf(seq, "%s",
 623                               conf->mirrors[i].rdev &&
 624                               conf->mirrors[i].rdev->in_sync ? "U" : "_");
 625         seq_printf(seq, "]");
 626 }
 627
 628
 629 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 630 {
 631         char b[BDEVNAME_SIZE];
 632         conf_t *conf = mddev_to_conf(mddev);
 633
 634         /*
 635          * If it is not operational, then we have already marked it as dead
 636          * else if it is the last working disks, ignore the error, let the
 637          * next level up know.
 638          * else mark the drive as failed
 639          */
 640         if (rdev->in_sync
 641             && conf->working_disks == 1)
 642                 /*
 643                  * Don't fail the drive, act as though we were just a
 644                  * normal single drive
 645                  */
 646                 return;
 647         if (rdev->in_sync) {
 648                 mddev->degraded++;
 649                 conf->working_disks--;
 650                 /*
 651                  * if recovery is running, make sure it aborts.
 652                  */
 653                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 654         }
 655         rdev->in_sync = 0;
 656         rdev->faulty = 1;
 657         mddev->sb_dirty = 1;
 658         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 659                 "       Operation continuing on %d devices\n",
 660                 bdevname(rdev->bdev,b), conf->working_disks);
 661 }
 662
 663 static void print_conf(conf_t *conf)
 664 {
 665         int i;
 666         mirror_info_t *tmp;
 667
 668         printk("RAID1 conf printout:\n");
 669         if (!conf) {
 670                 printk("(!conf)\n");
 671                 return;
 672         }
 673         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 674                 conf->raid_disks);
 675
 676         for (i = 0; i < conf->raid_disks; i++) {
 677                 char b[BDEVNAME_SIZE];
 678                 tmp = conf->mirrors + i;
 679                 if (tmp->rdev)
 680                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 681                                 i, !tmp->rdev->in_sync, !tmp->rdev->faulty,
 682                                 bdevname(tmp->rdev->bdev,b));
 683         }
 684 }
 685
 686 static void close_sync(conf_t *conf)
 687 {
 688         spin_lock_irq(&conf->resync_lock);
 689         wait_event_lock_irq(conf->wait_resume, !conf->barrier,
 690                             conf->resync_lock,  unplug_slaves(conf->mddev));
 691         spin_unlock_irq(&conf->resync_lock);
 692
 693         if (conf->barrier) BUG();
 694         if (waitqueue_active(&conf->wait_idle)) BUG();
 695
 696         mempool_destroy(conf->r1buf_pool);
 697         conf->r1buf_pool = NULL;
 698 }
 699
 700 static int raid1_spare_active(mddev_t *mddev)
 701 {
 702         int i;
 703         conf_t *conf = mddev->private;
 704         mirror_info_t *tmp;
 705
 706         spin_lock_irq(&conf->device_lock);
 707         /*
 708          * Find all failed disks within the RAID1 configuration
 709          * and mark them readable
 710          */
 711         for (i = 0; i < conf->raid_disks; i++) {
 712                 tmp = conf->mirrors + i;
 713                 if (tmp->rdev
 714                     && !tmp->rdev->faulty
 715                     && !tmp->rdev->in_sync) {
 716                         conf->working_disks++;
 717                         mddev->degraded--;
 718                         tmp->rdev->in_sync = 1;
 719                 }
 720         }
 721         spin_unlock_irq(&conf->device_lock);
 722
 723         print_conf(conf);
 724         return 0;
 725 }
 726
 727
 728 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 729 {
 730         conf_t *conf = mddev->private;
 731         int found = 0;
 732         int mirror;
 733         mirror_info_t *p;
 734
 735         spin_lock_irq(&conf->device_lock);
 736         for (mirror=0; mirror < mddev->raid_disks; mirror++)
 737                 if ( !(p=conf->mirrors+mirror)->rdev) {
 738                         p->rdev = rdev;
 739
 740                         blk_queue_stack_limits(mddev->queue,
 741                                                rdev->bdev->bd_disk->queue);
 742                         /* as we don't honour merge_bvec_fn, we must never risk
 743                          * violating it, so limit ->max_sector to one PAGE, as
 744                          * a one page request is never in violation.
 745                          */
 746                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 747                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 748                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
 749
 750                         p->head_position = 0;
 751                         rdev->raid_disk = mirror;
 752                         found = 1;
 753                         break;
 754                 }
 755         spin_unlock_irq(&conf->device_lock);
 756
 757         print_conf(conf);
 758         return found;
 759 }
 760
 761 static int raid1_remove_disk(mddev_t *mddev, int number)
 762 {
 763         conf_t *conf = mddev->private;
 764         int err = 1;
 765         mirror_info_t *p = conf->mirrors+ number;
 766
 767         print_conf(conf);
 768         spin_lock_irq(&conf->device_lock);
 769         if (p->rdev) {
 770                 if (p->rdev->in_sync ||
 771                     atomic_read(&p->rdev->nr_pending)) {
 772                         err = -EBUSY;
 773                         goto abort;
 774                 }
 775                 p->rdev = NULL;
 776                 err = 0;
 777         }
 778         if (err)
 779                 MD_BUG();
 780 abort:
 781         spin_unlock_irq(&conf->device_lock);
 782
 783         print_conf(conf);
 784         return err;
 785 }
 786
 787
 788 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
 789 {
 790         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 791         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 792         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 793
 794         if (bio->bi_size)
 795                 return 1;
 796
 797         if (r1_bio->bios[r1_bio->read_disk] != bio)
 798                 BUG();
 799         update_head_pos(r1_bio->read_disk, r1_bio);
 800         /*
 801          * we have read a block, now it needs to be re-written,
 802          * or re-read if the read failed.
 803          * We don't do much here, just schedule handling by raid1d
 804          */
 805         if (!uptodate)
 806                 md_error(r1_bio->mddev,
 807                          conf->mirrors[r1_bio->read_disk].rdev);
 808         else
 809                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 810         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
 811         reschedule_retry(r1_bio);
 812         return 0;
 813 }
 814
 815 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
 816 {
 817         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 818         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 819         mddev_t *mddev = r1_bio->mddev;
 820         conf_t *conf = mddev_to_conf(mddev);
 821         int i;
 822         int mirror=0;
 823
 824         if (bio->bi_size)
 825                 return 1;
 826
 827         for (i = 0; i < conf->raid_disks; i++)
 828                 if (r1_bio->bios[i] == bio) {
 829                         mirror = i;
 830                         break;
 831                 }
 832         if (!uptodate)
 833                 md_error(mddev, conf->mirrors[mirror].rdev);
 834         update_head_pos(mirror, r1_bio);
 835
 836         if (atomic_dec_and_test(&r1_bio->remaining)) {
 837                 md_done_sync(mddev, r1_bio->sectors, uptodate);
 838                 put_buf(r1_bio);
 839         }
 840         rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
 841         return 0;
 842 }
 843
 844 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
 845 {
 846         conf_t *conf = mddev_to_conf(mddev);
 847         int i;
 848         int disks = conf->raid_disks;
 849         struct bio *bio, *wbio;
 850
 851         bio = r1_bio->bios[r1_bio->read_disk];
 852
 853         /*
 854          * schedule writes
 855          */
 856         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 857                 /*
 858                  * There is no point trying a read-for-reconstruct as
 859                  * reconstruct is about to be aborted
 860                  */
 861                 char b[BDEVNAME_SIZE];
 862                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
 863                         " for block %llu\n",
 864                         bdevname(bio->bi_bdev,b),
 865                         (unsigned long long)r1_bio->sector);
 866                 md_done_sync(mddev, r1_bio->sectors, 0);
 867                 put_buf(r1_bio);
 868                 return;
 869         }
 870
 871         atomic_set(&r1_bio->remaining, 1);
 872         for (i = 0; i < disks ; i++) {
 873                 wbio = r1_bio->bios[i];
 874                 if (wbio->bi_end_io != end_sync_write)
 875                         continue;
 876
 877                 atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 878                 atomic_inc(&r1_bio->remaining);
 879                 md_sync_acct(conf->mirrors[i].rdev, wbio->bi_size >> 9);
 880                 generic_make_request(wbio);
 881         }
 882
 883         if (atomic_dec_and_test(&r1_bio->remaining)) {
 884                 md_done_sync(mddev, r1_bio->sectors, 1);
 885                 put_buf(r1_bio);
 886         }
 887 }
 888
 889 /*
 890  * This is a kernel thread which:
 891  *
 892  *      1.      Retries failed read operations on working mirrors.
 893  *      2.      Updates the raid superblock when problems encounter.
 894  *      3.      Performs writes following reads for array syncronising.
 895  */
 896
 897 static void raid1d(mddev_t *mddev)
 898 {
 899         struct list_head *head = &retry_list_head;
 900         r1bio_t *r1_bio;
 901         struct bio *bio;
 902         unsigned long flags;
 903         conf_t *conf = mddev_to_conf(mddev);
 904         int unplug=0;
 905         mdk_rdev_t *rdev;
 906
 907         md_check_recovery(mddev);
 908         md_handle_safemode(mddev);
 909
 910         for (;;) {
 911                 char b[BDEVNAME_SIZE];
 912                 spin_lock_irqsave(&retry_list_lock, flags);
 913                 if (list_empty(head))
 914                         break;
 915                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
 916                 list_del(head->prev);
 917                 spin_unlock_irqrestore(&retry_list_lock, flags);
 918
 919                 mddev = r1_bio->mddev;
 920                 conf = mddev_to_conf(mddev);
 921                 bio = r1_bio->master_bio;
 922                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
 923                         sync_request_write(mddev, r1_bio);
 924                         unplug = 1;
 925                 } else {
 926                         if (map(mddev, &rdev) == -1) {
 927                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
 928                                        " read error for block %llu\n",
 929                                        bdevname(bio->bi_bdev,b),
 930                                        (unsigned long long)r1_bio->sector);
 931                                 raid_end_bio_io(r1_bio);
 932                         } else {
 933                                 printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
 934                                        " another mirror\n",
 935                                        bdevname(rdev->bdev,b),
 936                                        (unsigned long long)r1_bio->sector);
 937                                 bio->bi_bdev = rdev->bdev;
 938                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
 939                                 bio->bi_rw = READ;
 940                                 unplug = 1;
 941                                 generic_make_request(bio);
 942                         }
 943                 }
 944         }
 945         spin_unlock_irqrestore(&retry_list_lock, flags);
 946         if (unplug)
 947                 unplug_slaves(mddev);
 948 }
 949
 950
 951 static int init_resync(conf_t *conf)
 952 {
 953         int buffs;
 954
 955         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
 956         if (conf->r1buf_pool)
 957                 BUG();
 958         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
 959                                           conf->poolinfo);
 960         if (!conf->r1buf_pool)
 961                 return -ENOMEM;
 962         conf->next_resync = 0;
 963         return 0;
 964 }
 965
 966 /*
 967  * perform a "sync" on one "block"
 968  *
 969  * We need to make sure that no normal I/O request - particularly write
 970  * requests - conflict with active sync requests.
 971  *
 972  * This is achieved by tracking pending requests and a 'barrier' concept
 973  * that can be installed to exclude normal IO requests.
 974  */
 975
 976 static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
 977 {
 978         conf_t *conf = mddev_to_conf(mddev);
 979         mirror_info_t *mirror;
 980         r1bio_t *r1_bio;
 981         struct bio *bio;
 982         sector_t max_sector, nr_sectors;
 983         int disk;
 984         int i;
 985         int write_targets = 0;
 986
 987         if (!conf->r1buf_pool)
 988                 if (init_resync(conf))
 989                         return -ENOMEM;
 990
 991         max_sector = mddev->size << 1;
 992         if (sector_nr >= max_sector) {
 993                 close_sync(conf);
 994                 return 0;
 995         }
 996
 997         /*
 998          * If there is non-resync activity waiting for us then
 999          * put in a delay to throttle resync.
1000          */
1001         if (!go_faster && waitqueue_active(&conf->wait_resume))
1002                 schedule_timeout(HZ);
1003         device_barrier(conf, sector_nr + RESYNC_SECTORS);
1004
1005         /*
1006          * If reconstructing, and >1 working disc,
1007          * could dedicate one to rebuild and others to
1008          * service read requests ..
1009          */
1010         disk = conf->last_used;
1011         /* make sure disk is operational */
1012         spin_lock_irq(&conf->device_lock);
1013         while (conf->mirrors[disk].rdev == NULL ||
1014                !conf->mirrors[disk].rdev->in_sync) {
1015                 if (disk <= 0)
1016                         disk = conf->raid_disks;
1017                 disk--;
1018                 if (disk == conf->last_used)
1019                         break;
1020         }
1021         conf->last_used = disk;
1022         atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
1023         spin_unlock_irq(&conf->device_lock);
1024
1025         mirror = conf->mirrors + disk;
1026
1027         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1028
1029         spin_lock_irq(&conf->resync_lock);
1030         conf->nr_pending++;
1031         spin_unlock_irq(&conf->resync_lock);
1032
1033         r1_bio->mddev = mddev;
1034         r1_bio->sector = sector_nr;
1035         set_bit(R1BIO_IsSync, &r1_bio->state);
1036         r1_bio->read_disk = disk;
1037
1038         for (i=0; i < conf->raid_disks; i++) {
1039                 bio = r1_bio->bios[i];
1040
1041                 /* take from bio_init */
1042                 bio->bi_next = NULL;
1043                 bio->bi_flags |= 1 << BIO_UPTODATE;
1044                 bio->bi_rw = 0;
1045                 bio->bi_vcnt = 0;
1046                 bio->bi_idx = 0;
1047                 bio->bi_phys_segments = 0;
1048                 bio->bi_hw_segments = 0;
1049                 bio->bi_size = 0;
1050                 bio->bi_end_io = NULL;
1051                 bio->bi_private = NULL;
1052
1053                 if (i == disk) {
1054                         bio->bi_rw = READ;
1055                         bio->bi_end_io = end_sync_read;
1056                 } else if (conf->mirrors[i].rdev &&
1057                            !conf->mirrors[i].rdev->faulty &&
1058                            (!conf->mirrors[i].rdev->in_sync ||
1059                             sector_nr + RESYNC_SECTORS > mddev->recovery_cp)) {
1060                         bio->bi_rw = WRITE;
1061                         bio->bi_end_io = end_sync_write;
1062                         write_targets ++;
1063                 } else
1064                         continue;
1065                 bio->bi_sector = sector_nr + conf->mirrors[i].rdev->data_offset;
1066                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1067                 bio->bi_private = r1_bio;
1068         }
1069         if (write_targets == 0) {
1070                 /* There is nowhere to write, so all non-sync
1071                  * drives must be failed - so we are finished
1072                  */
1073                 int rv = max_sector - sector_nr;
1074                 md_done_sync(mddev, rv, 1);
1075                 put_buf(r1_bio);
1076                 atomic_dec(&conf->mirrors[disk].rdev->nr_pending);
1077                 return rv;
1078         }
1079
1080         nr_sectors = 0;
1081         do {
1082                 struct page *page;
1083                 int len = PAGE_SIZE;
1084                 if (sector_nr + (len>>9) > max_sector)
1085                         len = (max_sector - sector_nr) << 9;
1086                 if (len == 0)
1087                         break;
1088                 for (i=0 ; i < conf->raid_disks; i++) {
1089                         bio = r1_bio->bios[i];
1090                         if (bio->bi_end_io) {
1091                                 page = r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page;
1092                                 if (bio_add_page(bio, page, len, 0) == 0) {
1093                                         /* stop here */
1094                                         r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page = page;
1095                                         while (i > 0) {
1096                                                 i--;
1097                                                 bio = r1_bio->bios[i];
1098                                                 if (bio->bi_end_io==NULL) continue;
1099                                                 /* remove last page from this bio */
1100                                                 bio->bi_vcnt--;
1101                                                 bio->bi_size -= len;
1102                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1103                                         }
1104                                         goto bio_full;
1105                                 }
1106                         }
1107                 }
1108                 nr_sectors += len>>9;
1109                 sector_nr += len>>9;
1110         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1111  bio_full:
1112         bio = r1_bio->bios[disk];
1113         r1_bio->sectors = nr_sectors;
1114
1115         md_sync_acct(mirror->rdev, nr_sectors);
1116
1117         generic_make_request(bio);
1118
1119         return nr_sectors;
1120 }
1121
1122 static int run(mddev_t *mddev)
1123 {
1124         conf_t *conf;
1125         int i, j, disk_idx;
1126         mirror_info_t *disk;
1127         mdk_rdev_t *rdev;
1128         struct list_head *tmp;
1129
1130         if (mddev->level != 1) {
1131                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1132                        mdname(mddev), mddev->level);
1133                 goto out;
1134         }
1135         /*
1136          * copy the already verified devices into our private RAID1
1137          * bookkeeping area. [whatever we allocate in run(),
1138          * should be freed in stop()]
1139          */
1140         conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1141         mddev->private = conf;
1142         if (!conf)
1143                 goto out_no_mem;
1144
1145         memset(conf, 0, sizeof(*conf));
1146         conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1147                                  GFP_KERNEL);
1148         if (!conf->mirrors)
1149                 goto out_no_mem;
1150
1151         memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1152
1153         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1154         if (!conf->poolinfo)
1155                 goto out_no_mem;
1156         conf->poolinfo->mddev = mddev;
1157         conf->poolinfo->raid_disks = mddev->raid_disks;
1158         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1159                                           r1bio_pool_free,
1160                                           conf->poolinfo);
1161         if (!conf->r1bio_pool)
1162                 goto out_no_mem;
1163
1164         mddev->queue->unplug_fn = raid1_unplug;
1165
1166
1167         ITERATE_RDEV(mddev, rdev, tmp) {
1168                 disk_idx = rdev->raid_disk;
1169                 if (disk_idx >= mddev->raid_disks
1170                     || disk_idx < 0)
1171                         continue;
1172                 disk = conf->mirrors + disk_idx;
1173
1174                 disk->rdev = rdev;
1175
1176                 blk_queue_stack_limits(mddev->queue,
1177                                        rdev->bdev->bd_disk->queue);
1178                 /* as we don't honour merge_bvec_fn, we must never risk
1179                  * violating it, so limit ->max_sector to one PAGE, as
1180                  * a one page request is never in violation.
1181                  */
1182                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1183                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1184                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
1185
1186                 disk->head_position = 0;
1187                 if (!rdev->faulty && rdev->in_sync)
1188                         conf->working_disks++;
1189         }
1190         conf->raid_disks = mddev->raid_disks;
1191         conf->mddev = mddev;
1192         conf->device_lock = SPIN_LOCK_UNLOCKED;
1193         if (conf->working_disks == 1)
1194                 mddev->recovery_cp = MaxSector;
1195
1196         conf->resync_lock = SPIN_LOCK_UNLOCKED;
1197         init_waitqueue_head(&conf->wait_idle);
1198         init_waitqueue_head(&conf->wait_resume);
1199
1200         if (!conf->working_disks) {
1201                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1202                         mdname(mddev));
1203                 goto out_free_conf;
1204         }
1205
1206         mddev->degraded = 0;
1207         for (i = 0; i < conf->raid_disks; i++) {
1208
1209                 disk = conf->mirrors + i;
1210
1211                 if (!disk->rdev) {
1212                         disk->head_position = 0;
1213                         mddev->degraded++;
1214                 }
1215         }
1216
1217         /*
1218          * find the first working one and use it as a starting point
1219          * to read balancing.
1220          */
1221         for (j = 0; j < conf->raid_disks &&
1222                      (!conf->mirrors[j].rdev ||
1223                       !conf->mirrors[j].rdev->in_sync) ; j++)
1224                 /* nothing */;
1225         conf->last_used = j;
1226
1227
1228
1229         {
1230                 mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1231                 if (!mddev->thread) {
1232                         printk(KERN_ERR
1233                                 "raid1: couldn't allocate thread for %s\n",
1234                                 mdname(mddev));
1235                         goto out_free_conf;
1236                 }
1237         }
1238         printk(KERN_INFO
1239                 "raid1: raid set %s active with %d out of %d mirrors\n",
1240                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1241                 mddev->raid_disks);
1242         /*
1243          * Ok, everything is just fine now
1244          */
1245         mddev->array_size = mddev->size;
1246
1247         return 0;
1248
1249 out_no_mem:
1250         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1251                mdname(mddev));
1252
1253 out_free_conf:
1254         if (conf) {
1255                 if (conf->r1bio_pool)
1256                         mempool_destroy(conf->r1bio_pool);
1257                 if (conf->mirrors)
1258                         kfree(conf->mirrors);
1259                 if (conf->poolinfo)
1260                         kfree(conf->poolinfo);
1261                 kfree(conf);
1262                 mddev->private = NULL;
1263         }
1264 out:
1265         return -EIO;
1266 }
1267
1268 static int stop(mddev_t *mddev)
1269 {
1270         conf_t *conf = mddev_to_conf(mddev);
1271
1272         md_unregister_thread(mddev->thread);
1273         mddev->thread = NULL;
1274         if (conf->r1bio_pool)
1275                 mempool_destroy(conf->r1bio_pool);
1276         if (conf->mirrors)
1277                 kfree(conf->mirrors);
1278         if (conf->poolinfo)
1279                 kfree(conf->poolinfo);
1280         kfree(conf);
1281         mddev->private = NULL;
1282         return 0;
1283 }
1284
1285 static int raid1_resize(mddev_t *mddev, sector_t sectors)
1286 {
1287         /* no resync is happening, and there is enough space
1288          * on all devices, so we can resize.
1289          * We need to make sure resync covers any new space.
1290          * If the array is shrinking we should possibly wait until
1291          * any io in the removed space completes, but it hardly seems
1292          * worth it.
1293          */
1294         mddev->array_size = sectors>>1;
1295         set_capacity(mddev->gendisk, mddev->array_size << 1);
1296         mddev->changed = 1;
1297         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
1298                 mddev->recovery_cp = mddev->size << 1;
1299                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1300         }
1301         mddev->size = mddev->array_size;
1302         return 0;
1303 }
1304
1305 static int raid1_reshape(mddev_t *mddev, int raid_disks)
1306 {
1307         /* We need to:
1308          * 1/ resize the r1bio_pool
1309          * 2/ resize conf->mirrors
1310          *
1311          * We allocate a new r1bio_pool if we can.
1312          * Then raise a device barrier and wait until all IO stops.
1313          * Then resize conf->mirrors and swap in the new r1bio pool.
1314          */
1315         mempool_t *newpool, *oldpool;
1316         struct pool_info *newpoolinfo;
1317         mirror_info_t *newmirrors;
1318         conf_t *conf = mddev_to_conf(mddev);
1319
1320         int d;
1321
1322         for (d= raid_disks; d < conf->raid_disks; d++)
1323                 if (conf->mirrors[d].rdev)
1324                         return -EBUSY;
1325
1326         newpoolinfo = kmalloc(sizeof(newpoolinfo), GFP_KERNEL);
1327         if (!newpoolinfo)
1328                 return -ENOMEM;
1329         newpoolinfo->mddev = mddev;
1330         newpoolinfo->raid_disks = raid_disks;
1331
1332         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1333                                  r1bio_pool_free, newpoolinfo);
1334         if (!newpool) {
1335                 kfree(newpoolinfo);
1336                 return -ENOMEM;
1337         }
1338         newmirrors = kmalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
1339         if (!newmirrors) {
1340                 kfree(newpoolinfo);
1341                 mempool_destroy(newpool);
1342                 return -ENOMEM;
1343         }
1344         memset(newmirrors, 0, sizeof(struct mirror_info)*raid_disks);
1345
1346         spin_lock_irq(&conf->resync_lock);
1347         conf->barrier++;
1348         wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
1349                             conf->resync_lock, unplug_slaves(mddev));
1350         spin_unlock_irq(&conf->resync_lock);
1351
1352         /* ok, everything is stopped */
1353         oldpool = conf->r1bio_pool;
1354         conf->r1bio_pool = newpool;
1355         for (d=0; d < raid_disks && d < conf->raid_disks; d++)
1356                 newmirrors[d] = conf->mirrors[d];
1357         kfree(conf->mirrors);
1358         conf->mirrors = newmirrors;
1359         kfree(conf->poolinfo);
1360         conf->poolinfo = newpoolinfo;
1361
1362         mddev->degraded += (raid_disks - conf->raid_disks);
1363         conf->raid_disks = mddev->raid_disks = raid_disks;
1364
1365         spin_lock_irq(&conf->resync_lock);
1366         conf->barrier--;
1367         spin_unlock_irq(&conf->resync_lock);
1368         wake_up(&conf->wait_resume);
1369         wake_up(&conf->wait_idle);
1370
1371
1372         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1373         md_wakeup_thread(mddev->thread);
1374
1375         mempool_destroy(oldpool);
1376         return 0;
1377 }
1378
1379
1380 static mdk_personality_t raid1_personality =
1381 {
1382         .name           = "raid1",
1383         .owner          = THIS_MODULE,
1384         .make_request   = make_request,
1385         .run            = run,
1386         .stop           = stop,
1387         .status         = status,
1388         .error_handler  = error,
1389         .hot_add_disk   = raid1_add_disk,
1390         .hot_remove_disk= raid1_remove_disk,
1391         .spare_active   = raid1_spare_active,
1392         .sync_request   = sync_request,
1393         .resize         = raid1_resize,
1394         .reshape        = raid1_reshape,
1395 };
1396
1397 static int __init raid_init(void)
1398 {
1399         return register_md_personality(RAID1, &raid1_personality);
1400 }
1401
1402 static void raid_exit(void)
1403 {
1404         unregister_md_personality(RAID1);
1405 }
1406
1407 module_init(raid_init);
1408 module_exit(raid_exit);
1409 MODULE_LICENSE("GPL");
1410 MODULE_ALIAS("md-personality-3"); /* RAID1 */