drivers/md/multipath.c

   1 /*
   2  * multipath.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  * MULTIPATH management functions.
   9  *
  10  * derived from raid1.c.
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU General Public License as published by
  14  * the Free Software Foundation; either version 2, or (at your option)
  15  * any later version.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * (for example /usr/src/linux/COPYING); if not, write to the Free
  19  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #include <linux/module.h>
  23 #include <linux/slab.h>
  24 #include <linux/spinlock.h>
  25 #include <linux/raid/multipath.h>
  26 #include <linux/buffer_head.h>
  27 #include <asm/atomic.h>
  28
  29 #define MAJOR_NR MD_MAJOR
  30 #define MD_DRIVER
  31 #define MD_PERSONALITY
  32
  33 #define MAX_WORK_PER_DISK 128
  34
  35 #define NR_RESERVED_BUFS        32
  36
  37
  38 static void *mp_pool_alloc(gfp_t gfp_flags, void *data)
  39 {
  40         struct multipath_bh *mpb;
  41         mpb = kzalloc(sizeof(*mpb), gfp_flags);
  42         return mpb;
  43 }
  44
  45 static void mp_pool_free(void *mpb, void *data)
  46 {
  47         kfree(mpb);
  48 }
  49
  50 static int multipath_map (multipath_conf_t *conf)
  51 {
  52         int i, disks = conf->raid_disks;
  53
  54         /*
  55          * Later we do read balancing on the read side
  56          * now we use the first available disk.
  57          */
  58
  59         rcu_read_lock();
  60         for (i = 0; i < disks; i++) {
  61                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
  62                 if (rdev && test_bit(In_sync, &rdev->flags)) {
  63                         atomic_inc(&rdev->nr_pending);
  64                         rcu_read_unlock();
  65                         return i;
  66                 }
  67         }
  68         rcu_read_unlock();
  69
  70         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
  71         return (-1);
  72 }
  73
  74 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
  75 {
  76         unsigned long flags;
  77         mddev_t *mddev = mp_bh->mddev;
  78         multipath_conf_t *conf = mddev_to_conf(mddev);
  79
  80         spin_lock_irqsave(&conf->device_lock, flags);
  81         list_add(&mp_bh->retry_list, &conf->retry_list);
  82         spin_unlock_irqrestore(&conf->device_lock, flags);
  83         md_wakeup_thread(mddev->thread);
  84 }
  85
  86
  87 /*
  88  * multipath_end_bh_io() is called when we have finished servicing a multipathed
  89  * operation and are ready to return a success/failure code to the buffer
  90  * cache layer.
  91  */
  92 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
  93 {
  94         struct bio *bio = mp_bh->master_bio;
  95         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
  96
  97         bio_endio(bio, bio->bi_size, err);
  98         mempool_free(mp_bh, conf->pool);
  99 }
 100
 101 static int multipath_end_request(struct bio *bio, unsigned int bytes_done,
 102                                  int error)
 103 {
 104         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 105         struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
 106         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
 107         mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
 108
 109         if (bio->bi_size)
 110                 return 1;
 111
 112         if (uptodate)
 113                 multipath_end_bh_io(mp_bh, 0);
 114         else if (!bio_rw_ahead(bio)) {
 115                 /*
 116                  * oops, IO error:
 117                  */
 118                 char b[BDEVNAME_SIZE];
 119                 md_error (mp_bh->mddev, rdev);
 120                 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
 121                        bdevname(rdev->bdev,b),
 122                        (unsigned long long)bio->bi_sector);
 123                 multipath_reschedule_retry(mp_bh);
 124         } else
 125                 multipath_end_bh_io(mp_bh, error);
 126         rdev_dec_pending(rdev, conf->mddev);
 127         return 0;
 128 }
 129
 130 static void unplug_slaves(mddev_t *mddev)
 131 {
 132         multipath_conf_t *conf = mddev_to_conf(mddev);
 133         int i;
 134
 135         rcu_read_lock();
 136         for (i=0; i<mddev->raid_disks; i++) {
 137                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 138                 if (rdev && !test_bit(Faulty, &rdev->flags)
 139                     && atomic_read(&rdev->nr_pending)) {
 140                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 141
 142                         atomic_inc(&rdev->nr_pending);
 143                         rcu_read_unlock();
 144
 145                         if (r_queue->unplug_fn)
 146                                 r_queue->unplug_fn(r_queue);
 147
 148                         rdev_dec_pending(rdev, mddev);
 149                         rcu_read_lock();
 150                 }
 151         }
 152         rcu_read_unlock();
 153 }
 154
 155 static void multipath_unplug(request_queue_t *q)
 156 {
 157         unplug_slaves(q->queuedata);
 158 }
 159
 160
 161 static int multipath_make_request (request_queue_t *q, struct bio * bio)
 162 {
 163         mddev_t *mddev = q->queuedata;
 164         multipath_conf_t *conf = mddev_to_conf(mddev);
 165         struct multipath_bh * mp_bh;
 166         struct multipath_info *multipath;
 167         const int rw = bio_data_dir(bio);
 168
 169         if (unlikely(bio_barrier(bio))) {
 170                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 171                 return 0;
 172         }
 173
 174         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
 175
 176         mp_bh->master_bio = bio;
 177         mp_bh->mddev = mddev;
 178
 179         disk_stat_inc(mddev->gendisk, ios[rw]);
 180         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 181
 182         mp_bh->path = multipath_map(conf);
 183         if (mp_bh->path < 0) {
 184                 bio_endio(bio, bio->bi_size, -EIO);
 185                 mempool_free(mp_bh, conf->pool);
 186                 return 0;
 187         }
 188         multipath = conf->multipaths + mp_bh->path;
 189
 190         mp_bh->bio = *bio;
 191         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
 192         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
 193         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
 194         mp_bh->bio.bi_end_io = multipath_end_request;
 195         mp_bh->bio.bi_private = mp_bh;
 196         generic_make_request(&mp_bh->bio);
 197         return 0;
 198 }
 199
 200 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
 201 {
 202         multipath_conf_t *conf = mddev_to_conf(mddev);
 203         int i;
 204
 205         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
 206                                                  conf->working_disks);
 207         for (i = 0; i < conf->raid_disks; i++)
 208                 seq_printf (seq, "%s",
 209                                conf->multipaths[i].rdev &&
 210                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
 211         seq_printf (seq, "]");
 212 }
 213
 214 static int multipath_issue_flush(request_queue_t *q, struct gendisk *disk,
 215                                  sector_t *error_sector)
 216 {
 217         mddev_t *mddev = q->queuedata;
 218         multipath_conf_t *conf = mddev_to_conf(mddev);
 219         int i, ret = 0;
 220
 221         rcu_read_lock();
 222         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 223                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 224                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 225                         struct block_device *bdev = rdev->bdev;
 226                         request_queue_t *r_queue = bdev_get_queue(bdev);
 227
 228                         if (!r_queue->issue_flush_fn)
 229                                 ret = -EOPNOTSUPP;
 230                         else {
 231                                 atomic_inc(&rdev->nr_pending);
 232                                 rcu_read_unlock();
 233                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 234                                                               error_sector);
 235                                 rdev_dec_pending(rdev, mddev);
 236                                 rcu_read_lock();
 237                         }
 238                 }
 239         }
 240         rcu_read_unlock();
 241         return ret;
 242 }
 243
 244 /*
 245  * Careful, this can execute in IRQ contexts as well!
 246  */
 247 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
 248 {
 249         multipath_conf_t *conf = mddev_to_conf(mddev);
 250
 251         if (conf->working_disks <= 1) {
 252                 /*
 253                  * Uh oh, we can do nothing if this is our last path, but
 254                  * first check if this is a queued request for a device
 255                  * which has just failed.
 256                  */
 257                 printk(KERN_ALERT
 258                         "multipath: only one IO path left and IO error.\n");
 259                 /* leave it active... it's all we have */
 260         } else {
 261                 /*
 262                  * Mark disk as unusable
 263                  */
 264                 if (!test_bit(Faulty, &rdev->flags)) {
 265                         char b[BDEVNAME_SIZE];
 266                         clear_bit(In_sync, &rdev->flags);
 267                         set_bit(Faulty, &rdev->flags);
 268                         mddev->sb_dirty = 1;
 269                         conf->working_disks--;
 270                         printk(KERN_ALERT "multipath: IO failure on %s,"
 271                                 " disabling IO path. \n Operation continuing"
 272                                 " on %d IO paths.\n",
 273                                 bdevname (rdev->bdev,b),
 274                                 conf->working_disks);
 275                 }
 276         }
 277 }
 278
 279 static void print_multipath_conf (multipath_conf_t *conf)
 280 {
 281         int i;
 282         struct multipath_info *tmp;
 283
 284         printk("MULTIPATH conf printout:\n");
 285         if (!conf) {
 286                 printk("(conf==NULL)\n");
 287                 return;
 288         }
 289         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 290                          conf->raid_disks);
 291
 292         for (i = 0; i < conf->raid_disks; i++) {
 293                 char b[BDEVNAME_SIZE];
 294                 tmp = conf->multipaths + i;
 295                 if (tmp->rdev)
 296                         printk(" disk%d, o:%d, dev:%s\n",
 297                                 i,!test_bit(Faulty, &tmp->rdev->flags),
 298                                bdevname(tmp->rdev->bdev,b));
 299         }
 300 }
 301
 302
 303 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 304 {
 305         multipath_conf_t *conf = mddev->private;
 306         int found = 0;
 307         int path;
 308         struct multipath_info *p;
 309
 310         print_multipath_conf(conf);
 311
 312         for (path=0; path<mddev->raid_disks; path++)
 313                 if ((p=conf->multipaths+path)->rdev == NULL) {
 314                         blk_queue_stack_limits(mddev->queue,
 315                                                rdev->bdev->bd_disk->queue);
 316
 317                 /* as we don't honour merge_bvec_fn, we must never risk
 318                  * violating it, so limit ->max_sector to one PAGE, as
 319                  * a one page request is never in violation.
 320                  * (Note: it is very unlikely that a device with
 321                  * merge_bvec_fn will be involved in multipath.)
 322                  */
 323                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 324                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 325                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 326
 327                         conf->working_disks++;
 328                         rdev->raid_disk = path;
 329                         set_bit(In_sync, &rdev->flags);
 330                         rcu_assign_pointer(p->rdev, rdev);
 331                         found = 1;
 332                 }
 333
 334         print_multipath_conf(conf);
 335         return found;
 336 }
 337
 338 static int multipath_remove_disk(mddev_t *mddev, int number)
 339 {
 340         multipath_conf_t *conf = mddev->private;
 341         int err = 0;
 342         mdk_rdev_t *rdev;
 343         struct multipath_info *p = conf->multipaths + number;
 344
 345         print_multipath_conf(conf);
 346
 347         rdev = p->rdev;
 348         if (rdev) {
 349                 if (test_bit(In_sync, &rdev->flags) ||
 350                     atomic_read(&rdev->nr_pending)) {
 351                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"                                " but is still operational!\n", number);
 352                         err = -EBUSY;
 353                         goto abort;
 354                 }
 355                 p->rdev = NULL;
 356                 synchronize_rcu();
 357                 if (atomic_read(&rdev->nr_pending)) {
 358                         /* lost the race, try later */
 359                         err = -EBUSY;
 360                         p->rdev = rdev;
 361                 }
 362         }
 363 abort:
 364
 365         print_multipath_conf(conf);
 366         return err;
 367 }
 368
 369
 370
 371 /*
 372  * This is a kernel thread which:
 373  *
 374  *      1.      Retries failed read operations on working multipaths.
 375  *      2.      Updates the raid superblock when problems encounter.
 376  *      3.      Performs writes following reads for array syncronising.
 377  */
 378
 379 static void multipathd (mddev_t *mddev)
 380 {
 381         struct multipath_bh *mp_bh;
 382         struct bio *bio;
 383         unsigned long flags;
 384         multipath_conf_t *conf = mddev_to_conf(mddev);
 385         struct list_head *head = &conf->retry_list;
 386
 387         md_check_recovery(mddev);
 388         for (;;) {
 389                 char b[BDEVNAME_SIZE];
 390                 spin_lock_irqsave(&conf->device_lock, flags);
 391                 if (list_empty(head))
 392                         break;
 393                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
 394                 list_del(head->prev);
 395                 spin_unlock_irqrestore(&conf->device_lock, flags);
 396
 397                 bio = &mp_bh->bio;
 398                 bio->bi_sector = mp_bh->master_bio->bi_sector;
 399
 400                 if ((mp_bh->path = multipath_map (conf))<0) {
 401                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
 402                                 " error for block %llu\n",
 403                                 bdevname(bio->bi_bdev,b),
 404                                 (unsigned long long)bio->bi_sector);
 405                         multipath_end_bh_io(mp_bh, -EIO);
 406                 } else {
 407                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
 408                                 " to another IO path\n",
 409                                 bdevname(bio->bi_bdev,b),
 410                                 (unsigned long long)bio->bi_sector);
 411                         *bio = *(mp_bh->master_bio);
 412                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
 413                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
 414                         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 415                         bio->bi_end_io = multipath_end_request;
 416                         bio->bi_private = mp_bh;
 417                         generic_make_request(bio);
 418                 }
 419         }
 420         spin_unlock_irqrestore(&conf->device_lock, flags);
 421 }
 422
 423 static int multipath_run (mddev_t *mddev)
 424 {
 425         multipath_conf_t *conf;
 426         int disk_idx;
 427         struct multipath_info *disk;
 428         mdk_rdev_t *rdev;
 429         struct list_head *tmp;
 430
 431         if (mddev->level != LEVEL_MULTIPATH) {
 432                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
 433                        mdname(mddev), mddev->level);
 434                 goto out;
 435         }
 436         /*
 437          * copy the already verified devices into our private MULTIPATH
 438          * bookkeeping area. [whatever we allocate in multipath_run(),
 439          * should be freed in multipath_stop()]
 440          */
 441
 442         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
 443         mddev->private = conf;
 444         if (!conf) {
 445                 printk(KERN_ERR
 446                         "multipath: couldn't allocate memory for %s\n",
 447                         mdname(mddev));
 448                 goto out;
 449         }
 450
 451         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
 452                                    GFP_KERNEL);
 453         if (!conf->multipaths) {
 454                 printk(KERN_ERR
 455                         "multipath: couldn't allocate memory for %s\n",
 456                         mdname(mddev));
 457                 goto out_free_conf;
 458         }
 459
 460         conf->working_disks = 0;
 461         ITERATE_RDEV(mddev,rdev,tmp) {
 462                 disk_idx = rdev->raid_disk;
 463                 if (disk_idx < 0 ||
 464                     disk_idx >= mddev->raid_disks)
 465                         continue;
 466
 467                 disk = conf->multipaths + disk_idx;
 468                 disk->rdev = rdev;
 469
 470                 blk_queue_stack_limits(mddev->queue,
 471                                        rdev->bdev->bd_disk->queue);
 472                 /* as we don't honour merge_bvec_fn, we must never risk
 473                  * violating it, not that we ever expect a device with
 474                  * a merge_bvec_fn to be involved in multipath */
 475                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 476                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
 477                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 478
 479                 if (!test_bit(Faulty, &rdev->flags))
 480                         conf->working_disks++;
 481         }
 482
 483         conf->raid_disks = mddev->raid_disks;
 484         mddev->sb_dirty = 1;
 485         conf->mddev = mddev;
 486         spin_lock_init(&conf->device_lock);
 487         INIT_LIST_HEAD(&conf->retry_list);
 488
 489         if (!conf->working_disks) {
 490                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
 491                         mdname(mddev));
 492                 goto out_free_conf;
 493         }
 494         mddev->degraded = conf->raid_disks = conf->working_disks;
 495
 496         conf->pool = mempool_create(NR_RESERVED_BUFS,
 497                                     mp_pool_alloc, mp_pool_free,
 498                                     NULL);
 499         if (conf->pool == NULL) {
 500                 printk(KERN_ERR
 501                         "multipath: couldn't allocate memory for %s\n",
 502                         mdname(mddev));
 503                 goto out_free_conf;
 504         }
 505
 506         {
 507                 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
 508                 if (!mddev->thread) {
 509                         printk(KERN_ERR "multipath: couldn't allocate thread"
 510                                 " for %s\n", mdname(mddev));
 511                         goto out_free_conf;
 512                 }
 513         }
 514
 515         printk(KERN_INFO
 516                 "multipath: array %s active with %d out of %d IO paths\n",
 517                 mdname(mddev), conf->working_disks, mddev->raid_disks);
 518         /*
 519          * Ok, everything is just fine now
 520          */
 521         mddev->array_size = mddev->size;
 522
 523         mddev->queue->unplug_fn = multipath_unplug;
 524         mddev->queue->issue_flush_fn = multipath_issue_flush;
 525
 526         return 0;
 527
 528 out_free_conf:
 529         if (conf->pool)
 530                 mempool_destroy(conf->pool);
 531         kfree(conf->multipaths);
 532         kfree(conf);
 533         mddev->private = NULL;
 534 out:
 535         return -EIO;
 536 }
 537
 538
 539 static int multipath_stop (mddev_t *mddev)
 540 {
 541         multipath_conf_t *conf = mddev_to_conf(mddev);
 542
 543         md_unregister_thread(mddev->thread);
 544         mddev->thread = NULL;
 545         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
 546         mempool_destroy(conf->pool);
 547         kfree(conf->multipaths);
 548         kfree(conf);
 549         mddev->private = NULL;
 550         return 0;
 551 }
 552
 553 static struct mdk_personality multipath_personality =
 554 {
 555         .name           = "multipath",
 556         .level          = LEVEL_MULTIPATH,
 557         .owner          = THIS_MODULE,
 558         .make_request   = multipath_make_request,
 559         .run            = multipath_run,
 560         .stop           = multipath_stop,
 561         .status         = multipath_status,
 562         .error_handler  = multipath_error,
 563         .hot_add_disk   = multipath_add_disk,
 564         .hot_remove_disk= multipath_remove_disk,
 565 };
 566
 567 static int __init multipath_init (void)
 568 {
 569         return register_md_personality (&multipath_personality);
 570 }
 571
 572 static void __exit multipath_exit (void)
 573 {
 574         unregister_md_personality (&multipath_personality);
 575 }
 576
 577 module_init(multipath_init);
 578 module_exit(multipath_exit);
 579 MODULE_LICENSE("GPL");
 580 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
 581 MODULE_ALIAS("md-level--4");