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