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 int multipath_map (multipath_conf_t *conf)
  39 {
  40         int i, disks = conf->raid_disks;
  41
  42         /*
  43          * Later we do read balancing on the read side
  44          * now we use the first available disk.
  45          */
  46
  47         rcu_read_lock();
  48         for (i = 0; i < disks; i++) {
  49                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
  50                 if (rdev && test_bit(In_sync, &rdev->flags)) {
  51                         atomic_inc(&rdev->nr_pending);
  52                         rcu_read_unlock();
  53                         return i;
  54                 }
  55         }
  56         rcu_read_unlock();
  57
  58         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
  59         return (-1);
  60 }
  61
  62 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
  63 {
  64         unsigned long flags;
  65         mddev_t *mddev = mp_bh->mddev;
  66         multipath_conf_t *conf = mddev_to_conf(mddev);
  67
  68         spin_lock_irqsave(&conf->device_lock, flags);
  69         list_add(&mp_bh->retry_list, &conf->retry_list);
  70         spin_unlock_irqrestore(&conf->device_lock, flags);
  71         md_wakeup_thread(mddev->thread);
  72 }
  73
  74
  75 /*
  76  * multipath_end_bh_io() is called when we have finished servicing a multipathed
  77  * operation and are ready to return a success/failure code to the buffer
  78  * cache layer.
  79  */
  80 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
  81 {
  82         struct bio *bio = mp_bh->master_bio;
  83         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
  84
  85         bio_endio(bio, bio->bi_size, err);
  86         mempool_free(mp_bh, conf->pool);
  87 }
  88
  89 static int multipath_end_request(struct bio *bio, unsigned int bytes_done,
  90                                  int error)
  91 {
  92         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
  93         struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
  94         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
  95         mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
  96
  97         if (bio->bi_size)
  98                 return 1;
  99
 100         if (uptodate)
 101                 multipath_end_bh_io(mp_bh, 0);
 102         else if (!bio_rw_ahead(bio)) {
 103                 /*
 104                  * oops, IO error:
 105                  */
 106                 char b[BDEVNAME_SIZE];
 107                 md_error (mp_bh->mddev, rdev);
 108                 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
 109                        bdevname(rdev->bdev,b),
 110                        (unsigned long long)bio->bi_sector);
 111                 multipath_reschedule_retry(mp_bh);
 112         } else
 113                 multipath_end_bh_io(mp_bh, error);
 114         rdev_dec_pending(rdev, conf->mddev);
 115         return 0;
 116 }
 117
 118 static void unplug_slaves(mddev_t *mddev)
 119 {
 120         multipath_conf_t *conf = mddev_to_conf(mddev);
 121         int i;
 122
 123         rcu_read_lock();
 124         for (i=0; i<mddev->raid_disks; i++) {
 125                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 126                 if (rdev && !test_bit(Faulty, &rdev->flags)
 127                     && atomic_read(&rdev->nr_pending)) {
 128                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 129
 130                         atomic_inc(&rdev->nr_pending);
 131                         rcu_read_unlock();
 132
 133                         if (r_queue->unplug_fn)
 134                                 r_queue->unplug_fn(r_queue);
 135
 136                         rdev_dec_pending(rdev, mddev);
 137                         rcu_read_lock();
 138                 }
 139         }
 140         rcu_read_unlock();
 141 }
 142
 143 static void multipath_unplug(request_queue_t *q)
 144 {
 145         unplug_slaves(q->queuedata);
 146 }
 147
 148
 149 static int multipath_make_request (request_queue_t *q, struct bio * bio)
 150 {
 151         mddev_t *mddev = q->queuedata;
 152         multipath_conf_t *conf = mddev_to_conf(mddev);
 153         struct multipath_bh * mp_bh;
 154         struct multipath_info *multipath;
 155         const int rw = bio_data_dir(bio);
 156
 157         if (unlikely(bio_barrier(bio))) {
 158                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 159                 return 0;
 160         }
 161
 162         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
 163
 164         mp_bh->master_bio = bio;
 165         mp_bh->mddev = mddev;
 166
 167         disk_stat_inc(mddev->gendisk, ios[rw]);
 168         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 169
 170         mp_bh->path = multipath_map(conf);
 171         if (mp_bh->path < 0) {
 172                 bio_endio(bio, bio->bi_size, -EIO);
 173                 mempool_free(mp_bh, conf->pool);
 174                 return 0;
 175         }
 176         multipath = conf->multipaths + mp_bh->path;
 177
 178         mp_bh->bio = *bio;
 179         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
 180         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
 181         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
 182         mp_bh->bio.bi_end_io = multipath_end_request;
 183         mp_bh->bio.bi_private = mp_bh;
 184         generic_make_request(&mp_bh->bio);
 185         return 0;
 186 }
 187
 188 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
 189 {
 190         multipath_conf_t *conf = mddev_to_conf(mddev);
 191         int i;
 192
 193         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
 194                                                  conf->working_disks);
 195         for (i = 0; i < conf->raid_disks; i++)
 196                 seq_printf (seq, "%s",
 197                                conf->multipaths[i].rdev &&
 198                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
 199         seq_printf (seq, "]");
 200 }
 201
 202 static int multipath_issue_flush(request_queue_t *q, struct gendisk *disk,
 203                                  sector_t *error_sector)
 204 {
 205         mddev_t *mddev = q->queuedata;
 206         multipath_conf_t *conf = mddev_to_conf(mddev);
 207         int i, ret = 0;
 208
 209         rcu_read_lock();
 210         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 211                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 212                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 213                         struct block_device *bdev = rdev->bdev;
 214                         request_queue_t *r_queue = bdev_get_queue(bdev);
 215
 216                         if (!r_queue->issue_flush_fn)
 217                                 ret = -EOPNOTSUPP;
 218                         else {
 219                                 atomic_inc(&rdev->nr_pending);
 220                                 rcu_read_unlock();
 221                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 222                                                               error_sector);
 223                                 rdev_dec_pending(rdev, mddev);
 224                                 rcu_read_lock();
 225                         }
 226                 }
 227         }
 228         rcu_read_unlock();
 229         return ret;
 230 }
 231 static int multipath_congested(void *data, int bits)
 232 {
 233         mddev_t *mddev = data;
 234         multipath_conf_t *conf = mddev_to_conf(mddev);
 235         int i, ret = 0;
 236
 237         rcu_read_lock();
 238         for (i = 0; i < mddev->raid_disks ; i++) {
 239                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 240                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 241                         request_queue_t *q = bdev_get_queue(rdev->bdev);
 242
 243                         ret |= bdi_congested(&q->backing_dev_info, bits);
 244                         /* Just like multipath_map, we just check the
 245                          * first available device
 246                          */
 247                         break;
 248                 }
 249         }
 250         rcu_read_unlock();
 251         return ret;
 252 }
 253
 254 /*
 255  * Careful, this can execute in IRQ contexts as well!
 256  */
 257 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
 258 {
 259         multipath_conf_t *conf = mddev_to_conf(mddev);
 260
 261         if (conf->working_disks <= 1) {
 262                 /*
 263                  * Uh oh, we can do nothing if this is our last path, but
 264                  * first check if this is a queued request for a device
 265                  * which has just failed.
 266                  */
 267                 printk(KERN_ALERT
 268                         "multipath: only one IO path left and IO error.\n");
 269                 /* leave it active... it's all we have */
 270         } else {
 271                 /*
 272                  * Mark disk as unusable
 273                  */
 274                 if (!test_bit(Faulty, &rdev->flags)) {
 275                         char b[BDEVNAME_SIZE];
 276                         clear_bit(In_sync, &rdev->flags);
 277                         set_bit(Faulty, &rdev->flags);
 278                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 279                         conf->working_disks--;
 280                         printk(KERN_ALERT "multipath: IO failure on %s,"
 281                                 " disabling IO path. \n Operation continuing"
 282                                 " on %d IO paths.\n",
 283                                 bdevname (rdev->bdev,b),
 284                                 conf->working_disks);
 285                 }
 286         }
 287 }
 288
 289 static void print_multipath_conf (multipath_conf_t *conf)
 290 {
 291         int i;
 292         struct multipath_info *tmp;
 293
 294         printk("MULTIPATH conf printout:\n");
 295         if (!conf) {
 296                 printk("(conf==NULL)\n");
 297                 return;
 298         }
 299         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 300                          conf->raid_disks);
 301
 302         for (i = 0; i < conf->raid_disks; i++) {
 303                 char b[BDEVNAME_SIZE];
 304                 tmp = conf->multipaths + i;
 305                 if (tmp->rdev)
 306                         printk(" disk%d, o:%d, dev:%s\n",
 307                                 i,!test_bit(Faulty, &tmp->rdev->flags),
 308                                bdevname(tmp->rdev->bdev,b));
 309         }
 310 }
 311
 312
 313 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 314 {
 315         multipath_conf_t *conf = mddev->private;
 316         struct request_queue *q;
 317         int found = 0;
 318         int path;
 319         struct multipath_info *p;
 320
 321         print_multipath_conf(conf);
 322
 323         for (path=0; path<mddev->raid_disks; path++)
 324                 if ((p=conf->multipaths+path)->rdev == NULL) {
 325                         q = rdev->bdev->bd_disk->queue;
 326                         blk_queue_stack_limits(mddev->queue, q);
 327
 328                 /* as we don't honour merge_bvec_fn, we must never risk
 329                  * violating it, so limit ->max_sector to one PAGE, as
 330                  * a one page request is never in violation.
 331                  * (Note: it is very unlikely that a device with
 332                  * merge_bvec_fn will be involved in multipath.)
 333                  */
 334                         if (q->merge_bvec_fn &&
 335                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 336                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 337
 338                         conf->working_disks++;
 339                         rdev->raid_disk = path;
 340                         set_bit(In_sync, &rdev->flags);
 341                         rcu_assign_pointer(p->rdev, rdev);
 342                         found = 1;
 343                 }
 344
 345         print_multipath_conf(conf);
 346         return found;
 347 }
 348
 349 static int multipath_remove_disk(mddev_t *mddev, int number)
 350 {
 351         multipath_conf_t *conf = mddev->private;
 352         int err = 0;
 353         mdk_rdev_t *rdev;
 354         struct multipath_info *p = conf->multipaths + number;
 355
 356         print_multipath_conf(conf);
 357
 358         rdev = p->rdev;
 359         if (rdev) {
 360                 if (test_bit(In_sync, &rdev->flags) ||
 361                     atomic_read(&rdev->nr_pending)) {
 362                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"                                " but is still operational!\n", number);
 363                         err = -EBUSY;
 364                         goto abort;
 365                 }
 366                 p->rdev = NULL;
 367                 synchronize_rcu();
 368                 if (atomic_read(&rdev->nr_pending)) {
 369                         /* lost the race, try later */
 370                         err = -EBUSY;
 371                         p->rdev = rdev;
 372                 }
 373         }
 374 abort:
 375
 376         print_multipath_conf(conf);
 377         return err;
 378 }
 379
 380
 381
 382 /*
 383  * This is a kernel thread which:
 384  *
 385  *      1.      Retries failed read operations on working multipaths.
 386  *      2.      Updates the raid superblock when problems encounter.
 387  *      3.      Performs writes following reads for array syncronising.
 388  */
 389
 390 static void multipathd (mddev_t *mddev)
 391 {
 392         struct multipath_bh *mp_bh;
 393         struct bio *bio;
 394         unsigned long flags;
 395         multipath_conf_t *conf = mddev_to_conf(mddev);
 396         struct list_head *head = &conf->retry_list;
 397
 398         md_check_recovery(mddev);
 399         for (;;) {
 400                 char b[BDEVNAME_SIZE];
 401                 spin_lock_irqsave(&conf->device_lock, flags);
 402                 if (list_empty(head))
 403                         break;
 404                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
 405                 list_del(head->prev);
 406                 spin_unlock_irqrestore(&conf->device_lock, flags);
 407
 408                 bio = &mp_bh->bio;
 409                 bio->bi_sector = mp_bh->master_bio->bi_sector;
 410
 411                 if ((mp_bh->path = multipath_map (conf))<0) {
 412                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
 413                                 " error for block %llu\n",
 414                                 bdevname(bio->bi_bdev,b),
 415                                 (unsigned long long)bio->bi_sector);
 416                         multipath_end_bh_io(mp_bh, -EIO);
 417                 } else {
 418                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
 419                                 " to another IO path\n",
 420                                 bdevname(bio->bi_bdev,b),
 421                                 (unsigned long long)bio->bi_sector);
 422                         *bio = *(mp_bh->master_bio);
 423                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
 424                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
 425                         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 426                         bio->bi_end_io = multipath_end_request;
 427                         bio->bi_private = mp_bh;
 428                         generic_make_request(bio);
 429                 }
 430         }
 431         spin_unlock_irqrestore(&conf->device_lock, flags);
 432 }
 433
 434 static int multipath_run (mddev_t *mddev)
 435 {
 436         multipath_conf_t *conf;
 437         int disk_idx;
 438         struct multipath_info *disk;
 439         mdk_rdev_t *rdev;
 440         struct list_head *tmp;
 441
 442         if (mddev->level != LEVEL_MULTIPATH) {
 443                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
 444                        mdname(mddev), mddev->level);
 445                 goto out;
 446         }
 447         /*
 448          * copy the already verified devices into our private MULTIPATH
 449          * bookkeeping area. [whatever we allocate in multipath_run(),
 450          * should be freed in multipath_stop()]
 451          */
 452
 453         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
 454         mddev->private = conf;
 455         if (!conf) {
 456                 printk(KERN_ERR
 457                         "multipath: couldn't allocate memory for %s\n",
 458                         mdname(mddev));
 459                 goto out;
 460         }
 461
 462         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
 463                                    GFP_KERNEL);
 464         if (!conf->multipaths) {
 465                 printk(KERN_ERR
 466                         "multipath: couldn't allocate memory for %s\n",
 467                         mdname(mddev));
 468                 goto out_free_conf;
 469         }
 470
 471         conf->working_disks = 0;
 472         ITERATE_RDEV(mddev,rdev,tmp) {
 473                 disk_idx = rdev->raid_disk;
 474                 if (disk_idx < 0 ||
 475                     disk_idx >= mddev->raid_disks)
 476                         continue;
 477
 478                 disk = conf->multipaths + disk_idx;
 479                 disk->rdev = rdev;
 480
 481                 blk_queue_stack_limits(mddev->queue,
 482                                        rdev->bdev->bd_disk->queue);
 483                 /* as we don't honour merge_bvec_fn, we must never risk
 484                  * violating it, not that we ever expect a device with
 485                  * a merge_bvec_fn to be involved in multipath */
 486                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 487                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
 488                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 489
 490                 if (!test_bit(Faulty, &rdev->flags))
 491                         conf->working_disks++;
 492         }
 493
 494         conf->raid_disks = mddev->raid_disks;
 495         conf->mddev = mddev;
 496         spin_lock_init(&conf->device_lock);
 497         INIT_LIST_HEAD(&conf->retry_list);
 498
 499         if (!conf->working_disks) {
 500                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
 501                         mdname(mddev));
 502                 goto out_free_conf;
 503         }
 504         mddev->degraded = conf->raid_disks = conf->working_disks;
 505
 506         conf->pool = mempool_create_kzalloc_pool(NR_RESERVED_BUFS,
 507                                                  sizeof(struct multipath_bh));
 508         if (conf->pool == NULL) {
 509                 printk(KERN_ERR
 510                         "multipath: couldn't allocate memory for %s\n",
 511                         mdname(mddev));
 512                 goto out_free_conf;
 513         }
 514
 515         {
 516                 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
 517                 if (!mddev->thread) {
 518                         printk(KERN_ERR "multipath: couldn't allocate thread"
 519                                 " for %s\n", mdname(mddev));
 520                         goto out_free_conf;
 521                 }
 522         }
 523
 524         printk(KERN_INFO
 525                 "multipath: array %s active with %d out of %d IO paths\n",
 526                 mdname(mddev), conf->working_disks, mddev->raid_disks);
 527         /*
 528          * Ok, everything is just fine now
 529          */
 530         mddev->array_size = mddev->size;
 531
 532         mddev->queue->unplug_fn = multipath_unplug;
 533         mddev->queue->issue_flush_fn = multipath_issue_flush;
 534         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
 535         mddev->queue->backing_dev_info.congested_data = mddev;
 536
 537         return 0;
 538
 539 out_free_conf:
 540         if (conf->pool)
 541                 mempool_destroy(conf->pool);
 542         kfree(conf->multipaths);
 543         kfree(conf);
 544         mddev->private = NULL;
 545 out:
 546         return -EIO;
 547 }
 548
 549
 550 static int multipath_stop (mddev_t *mddev)
 551 {
 552         multipath_conf_t *conf = mddev_to_conf(mddev);
 553
 554         md_unregister_thread(mddev->thread);
 555         mddev->thread = NULL;
 556         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
 557         mempool_destroy(conf->pool);
 558         kfree(conf->multipaths);
 559         kfree(conf);
 560         mddev->private = NULL;
 561         return 0;
 562 }
 563
 564 static struct mdk_personality multipath_personality =
 565 {
 566         .name           = "multipath",
 567         .level          = LEVEL_MULTIPATH,
 568         .owner          = THIS_MODULE,
 569         .make_request   = multipath_make_request,
 570         .run            = multipath_run,
 571         .stop           = multipath_stop,
 572         .status         = multipath_status,
 573         .error_handler  = multipath_error,
 574         .hot_add_disk   = multipath_add_disk,
 575         .hot_remove_disk= multipath_remove_disk,
 576 };
 577
 578 static int __init multipath_init (void)
 579 {
 580         return register_md_personality (&multipath_personality);
 581 }
 582
 583 static void __exit multipath_exit (void)
 584 {
 585         unregister_md_personality (&multipath_personality);
 586 }
 587
 588 module_init(multipath_init);
 589 module_exit(multipath_exit);
 590 MODULE_LICENSE("GPL");
 591 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
 592 MODULE_ALIAS("md-multipath");
 593 MODULE_ALIAS("md-level--4");