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