SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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->private;
  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 = mp_bh->mddev->private;
  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 = bio->bi_private;
  88         multipath_conf_t *conf = mp_bh->mddev->private;
  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_flagged(bio, BIO_RW_AHEAD)) {
  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->private;
 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(mddev_t *mddev, struct bio * bio)
 139 {
 140         multipath_conf_t *conf = mddev->private;
 141         struct multipath_bh * mp_bh;
 142         struct multipath_info *multipath;
 143
 144         if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 145                 md_barrier_request(mddev, bio);
 146                 return 0;
 147         }
 148
 149         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
 150
 151         mp_bh->master_bio = bio;
 152         mp_bh->mddev = mddev;
 153
 154         mp_bh->path = multipath_map(conf);
 155         if (mp_bh->path < 0) {
 156                 bio_endio(bio, -EIO);
 157                 mempool_free(mp_bh, conf->pool);
 158                 return 0;
 159         }
 160         multipath = conf->multipaths + mp_bh->path;
 161
 162         mp_bh->bio = *bio;
 163         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
 164         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
 165         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
 166         mp_bh->bio.bi_end_io = multipath_end_request;
 167         mp_bh->bio.bi_private = mp_bh;
 168         generic_make_request(&mp_bh->bio);
 169         return 0;
 170 }
 171
 172 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
 173 {
 174         multipath_conf_t *conf = mddev->private;
 175         int i;
 176
 177         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
 178                                                  conf->working_disks);
 179         for (i = 0; i < conf->raid_disks; i++)
 180                 seq_printf (seq, "%s",
 181                                conf->multipaths[i].rdev &&
 182                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
 183         seq_printf (seq, "]");
 184 }
 185
 186 static int multipath_congested(void *data, int bits)
 187 {
 188         mddev_t *mddev = data;
 189         multipath_conf_t *conf = mddev->private;
 190         int i, ret = 0;
 191
 192         if (mddev_congested(mddev, bits))
 193                 return 1;
 194
 195         rcu_read_lock();
 196         for (i = 0; i < mddev->raid_disks ; i++) {
 197                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 198                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 199                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 200
 201                         ret |= bdi_congested(&q->backing_dev_info, bits);
 202                         /* Just like multipath_map, we just check the
 203                          * first available device
 204                          */
 205                         break;
 206                 }
 207         }
 208         rcu_read_unlock();
 209         return ret;
 210 }
 211
 212 /*
 213  * Careful, this can execute in IRQ contexts as well!
 214  */
 215 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
 216 {
 217         multipath_conf_t *conf = mddev->private;
 218
 219         if (conf->working_disks <= 1) {
 220                 /*
 221                  * Uh oh, we can do nothing if this is our last path, but
 222                  * first check if this is a queued request for a device
 223                  * which has just failed.
 224                  */
 225                 printk(KERN_ALERT
 226                         "multipath: only one IO path left and IO error.\n");
 227                 /* leave it active... it's all we have */
 228         } else {
 229                 /*
 230                  * Mark disk as unusable
 231                  */
 232                 if (!test_bit(Faulty, &rdev->flags)) {
 233                         char b[BDEVNAME_SIZE];
 234                         clear_bit(In_sync, &rdev->flags);
 235                         set_bit(Faulty, &rdev->flags);
 236                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 237                         conf->working_disks--;
 238                         mddev->degraded++;
 239                         printk(KERN_ALERT "multipath: IO failure on %s,"
 240                                 " disabling IO path.\n"
 241                                 "multipath: Operation continuing"
 242                                 " on %d IO paths.\n",
 243                                 bdevname (rdev->bdev,b),
 244                                 conf->working_disks);
 245                 }
 246         }
 247 }
 248
 249 static void print_multipath_conf (multipath_conf_t *conf)
 250 {
 251         int i;
 252         struct multipath_info *tmp;
 253
 254         printk("MULTIPATH conf printout:\n");
 255         if (!conf) {
 256                 printk("(conf==NULL)\n");
 257                 return;
 258         }
 259         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 260                          conf->raid_disks);
 261
 262         for (i = 0; i < conf->raid_disks; i++) {
 263                 char b[BDEVNAME_SIZE];
 264                 tmp = conf->multipaths + i;
 265                 if (tmp->rdev)
 266                         printk(" disk%d, o:%d, dev:%s\n",
 267                                 i,!test_bit(Faulty, &tmp->rdev->flags),
 268                                bdevname(tmp->rdev->bdev,b));
 269         }
 270 }
 271
 272
 273 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 274 {
 275         multipath_conf_t *conf = mddev->private;
 276         struct request_queue *q;
 277         int err = -EEXIST;
 278         int path;
 279         struct multipath_info *p;
 280         int first = 0;
 281         int last = mddev->raid_disks - 1;
 282
 283         if (rdev->raid_disk >= 0)
 284                 first = last = rdev->raid_disk;
 285
 286         print_multipath_conf(conf);
 287
 288         for (path = first; path <= last; path++)
 289                 if ((p=conf->multipaths+path)->rdev == NULL) {
 290                         q = rdev->bdev->bd_disk->queue;
 291                         disk_stack_limits(mddev->gendisk, rdev->bdev,
 292                                           rdev->data_offset << 9);
 293
 294                 /* as we don't honour merge_bvec_fn, we must never risk
 295                  * violating it, so limit ->max_segments to one, lying
 296                  * within a single page.
 297                  * (Note: it is very unlikely that a device with
 298                  * merge_bvec_fn will be involved in multipath.)
 299                  */
 300                         if (q->merge_bvec_fn) {
 301                                 blk_queue_max_segments(mddev->queue, 1);
 302                                 blk_queue_segment_boundary(mddev->queue,
 303                                                            PAGE_CACHE_SIZE - 1);
 304                         }
 305
 306                         conf->working_disks++;
 307                         mddev->degraded--;
 308                         rdev->raid_disk = path;
 309                         set_bit(In_sync, &rdev->flags);
 310                         rcu_assign_pointer(p->rdev, rdev);
 311                         err = 0;
 312                         md_integrity_add_rdev(rdev, mddev);
 313                         break;
 314                 }
 315
 316         print_multipath_conf(conf);
 317
 318         return err;
 319 }
 320
 321 static int multipath_remove_disk(mddev_t *mddev, int number)
 322 {
 323         multipath_conf_t *conf = mddev->private;
 324         int err = 0;
 325         mdk_rdev_t *rdev;
 326         struct multipath_info *p = conf->multipaths + number;
 327
 328         print_multipath_conf(conf);
 329
 330         rdev = p->rdev;
 331         if (rdev) {
 332                 if (test_bit(In_sync, &rdev->flags) ||
 333                     atomic_read(&rdev->nr_pending)) {
 334                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"
 335                                " but is still operational!\n", number);
 336                         err = -EBUSY;
 337                         goto abort;
 338                 }
 339                 p->rdev = NULL;
 340                 synchronize_rcu();
 341                 if (atomic_read(&rdev->nr_pending)) {
 342                         /* lost the race, try later */
 343                         err = -EBUSY;
 344                         p->rdev = rdev;
 345                         goto abort;
 346                 }
 347                 md_integrity_register(mddev);
 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->private;
 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 (md_check_no_bitmap(mddev))
 425                 return -EINVAL;
 426
 427         if (mddev->level != LEVEL_MULTIPATH) {
 428                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
 429                        mdname(mddev), mddev->level);
 430                 goto out;
 431         }
 432         /*
 433          * copy the already verified devices into our private MULTIPATH
 434          * bookkeeping area. [whatever we allocate in multipath_run(),
 435          * should be freed in multipath_stop()]
 436          */
 437         mddev->queue->queue_lock = &mddev->queue->__queue_lock;
 438
 439         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
 440         mddev->private = conf;
 441         if (!conf) {
 442                 printk(KERN_ERR
 443                         "multipath: couldn't allocate memory for %s\n",
 444                         mdname(mddev));
 445                 goto out;
 446         }
 447
 448         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
 449                                    GFP_KERNEL);
 450         if (!conf->multipaths) {
 451                 printk(KERN_ERR
 452                         "multipath: couldn't allocate memory for %s\n",
 453                         mdname(mddev));
 454                 goto out_free_conf;
 455         }
 456
 457         conf->working_disks = 0;
 458         list_for_each_entry(rdev, &mddev->disks, same_set) {
 459                 disk_idx = rdev->raid_disk;
 460                 if (disk_idx < 0 ||
 461                     disk_idx >= mddev->raid_disks)
 462                         continue;
 463
 464                 disk = conf->multipaths + disk_idx;
 465                 disk->rdev = rdev;
 466                 disk_stack_limits(mddev->gendisk, rdev->bdev,
 467                                   rdev->data_offset << 9);
 468
 469                 /* as we don't honour merge_bvec_fn, we must never risk
 470                  * violating it, not that we ever expect a device with
 471                  * a merge_bvec_fn to be involved in multipath */
 472                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
 473                         blk_queue_max_segments(mddev->queue, 1);
 474                         blk_queue_segment_boundary(mddev->queue,
 475                                                    PAGE_CACHE_SIZE - 1);
 476                 }
 477
 478                 if (!test_bit(Faulty, &rdev->flags))
 479                         conf->working_disks++;
 480         }
 481
 482         conf->raid_disks = mddev->raid_disks;
 483         conf->mddev = mddev;
 484         spin_lock_init(&conf->device_lock);
 485         INIT_LIST_HEAD(&conf->retry_list);
 486
 487         if (!conf->working_disks) {
 488                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
 489                         mdname(mddev));
 490                 goto out_free_conf;
 491         }
 492         mddev->degraded = conf->raid_disks - conf->working_disks;
 493
 494         conf->pool = mempool_create_kmalloc_pool(NR_RESERVED_BUFS,
 495                                                  sizeof(struct multipath_bh));
 496         if (conf->pool == NULL) {
 497                 printk(KERN_ERR
 498                         "multipath: couldn't allocate memory for %s\n",
 499                         mdname(mddev));
 500                 goto out_free_conf;
 501         }
 502
 503         {
 504                 mddev->thread = md_register_thread(multipathd, mddev, NULL);
 505                 if (!mddev->thread) {
 506                         printk(KERN_ERR "multipath: couldn't allocate thread"
 507                                 " for %s\n", mdname(mddev));
 508                         goto out_free_conf;
 509                 }
 510         }
 511
 512         printk(KERN_INFO
 513                 "multipath: array %s active with %d out of %d IO paths\n",
 514                 mdname(mddev), conf->working_disks, mddev->raid_disks);
 515         /*
 516          * Ok, everything is just fine now
 517          */
 518         md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
 519
 520         mddev->queue->unplug_fn = multipath_unplug;
 521         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
 522         mddev->queue->backing_dev_info.congested_data = mddev;
 523         md_integrity_register(mddev);
 524         return 0;
 525
 526 out_free_conf:
 527         if (conf->pool)
 528                 mempool_destroy(conf->pool);
 529         kfree(conf->multipaths);
 530         kfree(conf);
 531         mddev->private = NULL;
 532 out:
 533         return -EIO;
 534 }
 535
 536
 537 static int multipath_stop (mddev_t *mddev)
 538 {
 539         multipath_conf_t *conf = mddev->private;
 540
 541         md_unregister_thread(mddev->thread);
 542         mddev->thread = NULL;
 543         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
 544         mempool_destroy(conf->pool);
 545         kfree(conf->multipaths);
 546         kfree(conf);
 547         mddev->private = NULL;
 548         return 0;
 549 }
 550
 551 static struct mdk_personality multipath_personality =
 552 {
 553         .name           = "multipath",
 554         .level          = LEVEL_MULTIPATH,
 555         .owner          = THIS_MODULE,
 556         .make_request   = multipath_make_request,
 557         .run            = multipath_run,
 558         .stop           = multipath_stop,
 559         .status         = multipath_status,
 560         .error_handler  = multipath_error,
 561         .hot_add_disk   = multipath_add_disk,
 562         .hot_remove_disk= multipath_remove_disk,
 563         .size           = multipath_size,
 564 };
 565
 566 static int __init multipath_init (void)
 567 {
 568         return register_md_personality (&multipath_personality);
 569 }
 570
 571 static void __exit multipath_exit (void)
 572 {
 573         unregister_md_personality (&multipath_personality);
 574 }
 575
 576 module_init(multipath_init);
 577 module_exit(multipath_exit);
 578 MODULE_LICENSE("GPL");
 579 MODULE_DESCRIPTION("simple multi-path personality for MD");
 580 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
 581 MODULE_ALIAS("md-multipath");
 582 MODULE_ALIAS("md-level--4");