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 "md.h"
  26 #include "multipath.h"
  27
  28 #define MAX_WORK_PER_DISK 128
  29
  30 #define NR_RESERVED_BUFS        32
  31
  32
  33 static int multipath_map (multipath_conf_t *conf)
  34 {
  35         int i, disks = conf->raid_disks;
  36
  37         /*
  38          * Later we do read balancing on the read side
  39          * now we use the first available disk.
  40          */
  41
  42         rcu_read_lock();
  43         for (i = 0; i < disks; i++) {
  44                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
  45                 if (rdev && test_bit(In_sync, &rdev->flags)) {
  46                         atomic_inc(&rdev->nr_pending);
  47                         rcu_read_unlock();
  48                         return i;
  49                 }
  50         }
  51         rcu_read_unlock();
  52
  53         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
  54         return (-1);
  55 }
  56
  57 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
  58 {
  59         unsigned long flags;
  60         mddev_t *mddev = mp_bh->mddev;
  61         multipath_conf_t *conf = mddev_to_conf(mddev);
  62
  63         spin_lock_irqsave(&conf->device_lock, flags);
  64         list_add(&mp_bh->retry_list, &conf->retry_list);
  65         spin_unlock_irqrestore(&conf->device_lock, flags);
  66         md_wakeup_thread(mddev->thread);
  67 }
  68
  69
  70 /*
  71  * multipath_end_bh_io() is called when we have finished servicing a multipathed
  72  * operation and are ready to return a success/failure code to the buffer
  73  * cache layer.
  74  */
  75 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
  76 {
  77         struct bio *bio = mp_bh->master_bio;
  78         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
  79
  80         bio_endio(bio, err);
  81         mempool_free(mp_bh, conf->pool);
  82 }
  83
  84 static void multipath_end_request(struct bio *bio, int error)
  85 {
  86         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
  87         struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
  88         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
  89         mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
  90
  91         if (uptodate)
  92                 multipath_end_bh_io(mp_bh, 0);
  93         else if (!bio_rw_ahead(bio)) {
  94                 /*
  95                  * oops, IO error:
  96                  */
  97                 char b[BDEVNAME_SIZE];
  98                 md_error (mp_bh->mddev, rdev);
  99                 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
 100                        bdevname(rdev->bdev,b),
 101                        (unsigned long long)bio->bi_sector);
 102                 multipath_reschedule_retry(mp_bh);
 103         } else
 104                 multipath_end_bh_io(mp_bh, error);
 105         rdev_dec_pending(rdev, conf->mddev);
 106 }
 107
 108 static void unplug_slaves(mddev_t *mddev)
 109 {
 110         multipath_conf_t *conf = mddev_to_conf(mddev);
 111         int i;
 112
 113         rcu_read_lock();
 114         for (i=0; i<mddev->raid_disks; i++) {
 115                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 116                 if (rdev && !test_bit(Faulty, &rdev->flags)
 117                     && atomic_read(&rdev->nr_pending)) {
 118                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 119
 120                         atomic_inc(&rdev->nr_pending);
 121                         rcu_read_unlock();
 122
 123                         blk_unplug(r_queue);
 124
 125                         rdev_dec_pending(rdev, mddev);
 126                         rcu_read_lock();
 127                 }
 128         }
 129         rcu_read_unlock();
 130 }
 131
 132 static void multipath_unplug(struct request_queue *q)
 133 {
 134         unplug_slaves(q->queuedata);
 135 }
 136
 137
 138 static int multipath_make_request (struct request_queue *q, struct bio * bio)
 139 {
 140         mddev_t *mddev = q->queuedata;
 141         multipath_conf_t *conf = mddev_to_conf(mddev);
 142         struct multipath_bh * mp_bh;
 143         struct multipath_info *multipath;
 144         const int rw = bio_data_dir(bio);
 145         int cpu;
 146
 147         if (unlikely(bio_barrier(bio))) {
 148                 bio_endio(bio, -EOPNOTSUPP);
 149                 return 0;
 150         }
 151
 152         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
 153
 154         mp_bh->master_bio = bio;
 155         mp_bh->mddev = mddev;
 156
 157         cpu = part_stat_lock();
 158         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 159         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 160                       bio_sectors(bio));
 161         part_stat_unlock();
 162
 163         mp_bh->path = multipath_map(conf);
 164         if (mp_bh->path < 0) {
 165                 bio_endio(bio, -EIO);
 166                 mempool_free(mp_bh, conf->pool);
 167                 return 0;
 168         }
 169         multipath = conf->multipaths + mp_bh->path;
 170
 171         mp_bh->bio = *bio;
 172         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
 173         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
 174         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
 175         mp_bh->bio.bi_end_io = multipath_end_request;
 176         mp_bh->bio.bi_private = mp_bh;
 177         generic_make_request(&mp_bh->bio);
 178         return 0;
 179 }
 180
 181 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
 182 {
 183         multipath_conf_t *conf = mddev_to_conf(mddev);
 184         int i;
 185
 186         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
 187                                                  conf->working_disks);
 188         for (i = 0; i < conf->raid_disks; i++)
 189                 seq_printf (seq, "%s",
 190                                conf->multipaths[i].rdev &&
 191                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
 192         seq_printf (seq, "]");
 193 }
 194
 195 static int multipath_congested(void *data, int bits)
 196 {
 197         mddev_t *mddev = data;
 198         multipath_conf_t *conf = mddev_to_conf(mddev);
 199         int i, ret = 0;
 200
 201         rcu_read_lock();
 202         for (i = 0; i < mddev->raid_disks ; i++) {
 203                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 204                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 205                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 206
 207                         ret |= bdi_congested(&q->backing_dev_info, bits);
 208                         /* Just like multipath_map, we just check the
 209                          * first available device
 210                          */
 211                         break;
 212                 }
 213         }
 214         rcu_read_unlock();
 215         return ret;
 216 }
 217
 218 /*
 219  * Careful, this can execute in IRQ contexts as well!
 220  */
 221 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
 222 {
 223         multipath_conf_t *conf = mddev_to_conf(mddev);
 224
 225         if (conf->working_disks <= 1) {
 226                 /*
 227                  * Uh oh, we can do nothing if this is our last path, but
 228                  * first check if this is a queued request for a device
 229                  * which has just failed.
 230                  */
 231                 printk(KERN_ALERT
 232                         "multipath: only one IO path left and IO error.\n");
 233                 /* leave it active... it's all we have */
 234         } else {
 235                 /*
 236                  * Mark disk as unusable
 237                  */
 238                 if (!test_bit(Faulty, &rdev->flags)) {
 239                         char b[BDEVNAME_SIZE];
 240                         clear_bit(In_sync, &rdev->flags);
 241                         set_bit(Faulty, &rdev->flags);
 242                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 243                         conf->working_disks--;
 244                         mddev->degraded++;
 245                         printk(KERN_ALERT "multipath: IO failure on %s,"
 246                                 " disabling IO path.\n"
 247                                 "multipath: Operation continuing"
 248                                 " on %d IO paths.\n",
 249                                 bdevname (rdev->bdev,b),
 250                                 conf->working_disks);
 251                 }
 252         }
 253 }
 254
 255 static void print_multipath_conf (multipath_conf_t *conf)
 256 {
 257         int i;
 258         struct multipath_info *tmp;
 259
 260         printk("MULTIPATH conf printout:\n");
 261         if (!conf) {
 262                 printk("(conf==NULL)\n");
 263                 return;
 264         }
 265         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 266                          conf->raid_disks);
 267
 268         for (i = 0; i < conf->raid_disks; i++) {
 269                 char b[BDEVNAME_SIZE];
 270                 tmp = conf->multipaths + i;
 271                 if (tmp->rdev)
 272                         printk(" disk%d, o:%d, dev:%s\n",
 273                                 i,!test_bit(Faulty, &tmp->rdev->flags),
 274                                bdevname(tmp->rdev->bdev,b));
 275         }
 276 }
 277
 278
 279 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 280 {
 281         multipath_conf_t *conf = mddev->private;
 282         struct request_queue *q;
 283         int err = -EEXIST;
 284         int path;
 285         struct multipath_info *p;
 286         int first = 0;
 287         int last = mddev->raid_disks - 1;
 288
 289         if (rdev->raid_disk >= 0)
 290                 first = last = rdev->raid_disk;
 291
 292         print_multipath_conf(conf);
 293
 294         for (path = first; path <= last; path++)
 295                 if ((p=conf->multipaths+path)->rdev == NULL) {
 296                         q = rdev->bdev->bd_disk->queue;
 297                         blk_queue_stack_limits(mddev->queue, q);
 298
 299                 /* as we don't honour merge_bvec_fn, we must never risk
 300                  * violating it, so limit ->max_sector to one PAGE, as
 301                  * a one page request is never in violation.
 302                  * (Note: it is very unlikely that a device with
 303                  * merge_bvec_fn will be involved in multipath.)
 304                  */
 305                         if (q->merge_bvec_fn &&
 306                             queue_max_sectors(q) > (PAGE_SIZE>>9))
 307                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 308
 309                         conf->working_disks++;
 310                         mddev->degraded--;
 311                         rdev->raid_disk = path;
 312                         set_bit(In_sync, &rdev->flags);
 313                         rcu_assign_pointer(p->rdev, rdev);
 314                         err = 0;
 315                         break;
 316                 }
 317
 318         print_multipath_conf(conf);
 319
 320         return err;
 321 }
 322
 323 static int multipath_remove_disk(mddev_t *mddev, int number)
 324 {
 325         multipath_conf_t *conf = mddev->private;
 326         int err = 0;
 327         mdk_rdev_t *rdev;
 328         struct multipath_info *p = conf->multipaths + number;
 329
 330         print_multipath_conf(conf);
 331
 332         rdev = p->rdev;
 333         if (rdev) {
 334                 if (test_bit(In_sync, &rdev->flags) ||
 335                     atomic_read(&rdev->nr_pending)) {
 336                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"
 337                                " but is still operational!\n", number);
 338                         err = -EBUSY;
 339                         goto abort;
 340                 }
 341                 p->rdev = NULL;
 342                 synchronize_rcu();
 343                 if (atomic_read(&rdev->nr_pending)) {
 344                         /* lost the race, try later */
 345                         err = -EBUSY;
 346                         p->rdev = rdev;
 347                 }
 348         }
 349 abort:
 350
 351         print_multipath_conf(conf);
 352         return err;
 353 }
 354
 355
 356
 357 /*
 358  * This is a kernel thread which:
 359  *
 360  *      1.      Retries failed read operations on working multipaths.
 361  *      2.      Updates the raid superblock when problems encounter.
 362  *      3.      Performs writes following reads for array syncronising.
 363  */
 364
 365 static void multipathd (mddev_t *mddev)
 366 {
 367         struct multipath_bh *mp_bh;
 368         struct bio *bio;
 369         unsigned long flags;
 370         multipath_conf_t *conf = mddev_to_conf(mddev);
 371         struct list_head *head = &conf->retry_list;
 372
 373         md_check_recovery(mddev);
 374         for (;;) {
 375                 char b[BDEVNAME_SIZE];
 376                 spin_lock_irqsave(&conf->device_lock, flags);
 377                 if (list_empty(head))
 378                         break;
 379                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
 380                 list_del(head->prev);
 381                 spin_unlock_irqrestore(&conf->device_lock, flags);
 382
 383                 bio = &mp_bh->bio;
 384                 bio->bi_sector = mp_bh->master_bio->bi_sector;
 385
 386                 if ((mp_bh->path = multipath_map (conf))<0) {
 387                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
 388                                 " error for block %llu\n",
 389                                 bdevname(bio->bi_bdev,b),
 390                                 (unsigned long long)bio->bi_sector);
 391                         multipath_end_bh_io(mp_bh, -EIO);
 392                 } else {
 393                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
 394                                 " to another IO path\n",
 395                                 bdevname(bio->bi_bdev,b),
 396                                 (unsigned long long)bio->bi_sector);
 397                         *bio = *(mp_bh->master_bio);
 398                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
 399                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
 400                         bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
 401                         bio->bi_end_io = multipath_end_request;
 402                         bio->bi_private = mp_bh;
 403                         generic_make_request(bio);
 404                 }
 405         }
 406         spin_unlock_irqrestore(&conf->device_lock, flags);
 407 }
 408
 409 static sector_t multipath_size(mddev_t *mddev, sector_t sectors, int raid_disks)
 410 {
 411         WARN_ONCE(sectors || raid_disks,
 412                   "%s does not support generic reshape\n", __func__);
 413
 414         return mddev->dev_sectors;
 415 }
 416
 417 static int multipath_run (mddev_t *mddev)
 418 {
 419         multipath_conf_t *conf;
 420         int disk_idx;
 421         struct multipath_info *disk;
 422         mdk_rdev_t *rdev;
 423
 424         if (mddev->level != LEVEL_MULTIPATH) {
 425                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
 426                        mdname(mddev), mddev->level);
 427                 goto out;
 428         }
 429         /*
 430          * copy the already verified devices into our private MULTIPATH
 431          * bookkeeping area. [whatever we allocate in multipath_run(),
 432          * should be freed in multipath_stop()]
 433          */
 434         mddev->queue->queue_lock = &mddev->queue->__queue_lock;
 435
 436         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
 437         mddev->private = conf;
 438         if (!conf) {
 439                 printk(KERN_ERR
 440                         "multipath: couldn't allocate memory for %s\n",
 441                         mdname(mddev));
 442                 goto out;
 443         }
 444
 445         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
 446                                    GFP_KERNEL);
 447         if (!conf->multipaths) {
 448                 printk(KERN_ERR
 449                         "multipath: couldn't allocate memory for %s\n",
 450                         mdname(mddev));
 451                 goto out_free_conf;
 452         }
 453
 454         conf->working_disks = 0;
 455         list_for_each_entry(rdev, &mddev->disks, same_set) {
 456                 disk_idx = rdev->raid_disk;
 457                 if (disk_idx < 0 ||
 458                     disk_idx >= mddev->raid_disks)
 459                         continue;
 460
 461                 disk = conf->multipaths + disk_idx;
 462                 disk->rdev = rdev;
 463
 464                 blk_queue_stack_limits(mddev->queue,
 465                                        rdev->bdev->bd_disk->queue);
 466                 /* as we don't honour merge_bvec_fn, we must never risk
 467                  * violating it, not that we ever expect a device with
 468                  * a merge_bvec_fn to be involved in multipath */
 469                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 470                     queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
 471                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 472
 473                 if (!test_bit(Faulty, &rdev->flags))
 474                         conf->working_disks++;
 475         }
 476
 477         conf->raid_disks = mddev->raid_disks;
 478         conf->mddev = mddev;
 479         spin_lock_init(&conf->device_lock);
 480         INIT_LIST_HEAD(&conf->retry_list);
 481
 482         if (!conf->working_disks) {
 483                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
 484                         mdname(mddev));
 485                 goto out_free_conf;
 486         }
 487         mddev->degraded = conf->raid_disks - conf->working_disks;
 488
 489         conf->pool = mempool_create_kzalloc_pool(NR_RESERVED_BUFS,
 490                                                  sizeof(struct multipath_bh));
 491         if (conf->pool == NULL) {
 492                 printk(KERN_ERR
 493                         "multipath: couldn't allocate memory for %s\n",
 494                         mdname(mddev));
 495                 goto out_free_conf;
 496         }
 497
 498         {
 499                 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
 500                 if (!mddev->thread) {
 501                         printk(KERN_ERR "multipath: couldn't allocate thread"
 502                                 " for %s\n", mdname(mddev));
 503                         goto out_free_conf;
 504                 }
 505         }
 506
 507         printk(KERN_INFO
 508                 "multipath: array %s active with %d out of %d IO paths\n",
 509                 mdname(mddev), conf->working_disks, mddev->raid_disks);
 510         /*
 511          * Ok, everything is just fine now
 512          */
 513         md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
 514
 515         mddev->queue->unplug_fn = multipath_unplug;
 516         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
 517         mddev->queue->backing_dev_info.congested_data = mddev;
 518
 519         return 0;
 520
 521 out_free_conf:
 522         if (conf->pool)
 523                 mempool_destroy(conf->pool);
 524         kfree(conf->multipaths);
 525         kfree(conf);
 526         mddev->private = NULL;
 527 out:
 528         return -EIO;
 529 }
 530
 531
 532 static int multipath_stop (mddev_t *mddev)
 533 {
 534         multipath_conf_t *conf = mddev_to_conf(mddev);
 535
 536         md_unregister_thread(mddev->thread);
 537         mddev->thread = NULL;
 538         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
 539         mempool_destroy(conf->pool);
 540         kfree(conf->multipaths);
 541         kfree(conf);
 542         mddev->private = NULL;
 543         return 0;
 544 }
 545
 546 static struct mdk_personality multipath_personality =
 547 {
 548         .name           = "multipath",
 549         .level          = LEVEL_MULTIPATH,
 550         .owner          = THIS_MODULE,
 551         .make_request   = multipath_make_request,
 552         .run            = multipath_run,
 553         .stop           = multipath_stop,
 554         .status         = multipath_status,
 555         .error_handler  = multipath_error,
 556         .hot_add_disk   = multipath_add_disk,
 557         .hot_remove_disk= multipath_remove_disk,
 558         .size           = multipath_size,
 559 };
 560
 561 static int __init multipath_init (void)
 562 {
 563         return register_md_personality (&multipath_personality);
 564 }
 565
 566 static void __exit multipath_exit (void)
 567 {
 568         unregister_md_personality (&multipath_personality);
 569 }
 570
 571 module_init(multipath_init);
 572 module_exit(multipath_exit);
 573 MODULE_LICENSE("GPL");
 574 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
 575 MODULE_ALIAS("md-multipath");
 576 MODULE_ALIAS("md-level--4");