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 = (struct multipath_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_rw_flagged(bio, BIO_RW_AHEAD)) {
  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 void unplug_slaves(mddev_t *mddev)
 110 {
 111         multipath_conf_t *conf = mddev->private;
 112         int i;
 113
 114         rcu_read_lock();
 115         for (i=0; i<mddev->raid_disks; i++) {
 116                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 117                 if (rdev && !test_bit(Faulty, &rdev->flags)
 118                     && atomic_read(&rdev->nr_pending)) {
 119                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 120
 121                         atomic_inc(&rdev->nr_pending);
 122                         rcu_read_unlock();
 123
 124                         blk_unplug(r_queue);
 125
 126                         rdev_dec_pending(rdev, mddev);
 127                         rcu_read_lock();
 128                 }
 129         }
 130         rcu_read_unlock();
 131 }
 132
 133 static void multipath_unplug(struct request_queue *q)
 134 {
 135         unplug_slaves(q->queuedata);
 136 }
 137
 138
 139 static int multipath_make_request (struct request_queue *q, struct bio * bio)
 140 {
 141         mddev_t *mddev = q->queuedata;
 142         multipath_conf_t *conf = mddev->private;
 143         struct multipath_bh * mp_bh;
 144         struct multipath_info *multipath;
 145         const int rw = bio_data_dir(bio);
 146         int cpu;
 147
 148         if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 149                 md_barrier_request(mddev, bio);
 150                 return 0;
 151         }
 152
 153         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
 154
 155         mp_bh->master_bio = bio;
 156         mp_bh->mddev = mddev;
 157
 158         cpu = part_stat_lock();
 159         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 160         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 161                       bio_sectors(bio));
 162         part_stat_unlock();
 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_TRANSPORT);
 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->private;
 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->private;
 200         int i, ret = 0;
 201
 202         if (mddev_congested(mddev, bits))
 203                 return 1;
 204
 205         rcu_read_lock();
 206         for (i = 0; i < mddev->raid_disks ; i++) {
 207                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 208                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 209                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 210
 211                         ret |= bdi_congested(&q->backing_dev_info, bits);
 212                         /* Just like multipath_map, we just check the
 213                          * first available device
 214                          */
 215                         break;
 216                 }
 217         }
 218         rcu_read_unlock();
 219         return ret;
 220 }
 221
 222 /*
 223  * Careful, this can execute in IRQ contexts as well!
 224  */
 225 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
 226 {
 227         multipath_conf_t *conf = mddev->private;
 228
 229         if (conf->working_disks <= 1) {
 230                 /*
 231                  * Uh oh, we can do nothing if this is our last path, but
 232                  * first check if this is a queued request for a device
 233                  * which has just failed.
 234                  */
 235                 printk(KERN_ALERT
 236                         "multipath: only one IO path left and IO error.\n");
 237                 /* leave it active... it's all we have */
 238         } else {
 239                 /*
 240                  * Mark disk as unusable
 241                  */
 242                 if (!test_bit(Faulty, &rdev->flags)) {
 243                         char b[BDEVNAME_SIZE];
 244                         clear_bit(In_sync, &rdev->flags);
 245                         set_bit(Faulty, &rdev->flags);
 246                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 247                         conf->working_disks--;
 248                         mddev->degraded++;
 249                         printk(KERN_ALERT "multipath: IO failure on %s,"
 250                                 " disabling IO path.\n"
 251                                 "multipath: Operation continuing"
 252                                 " on %d IO paths.\n",
 253                                 bdevname (rdev->bdev,b),
 254                                 conf->working_disks);
 255                 }
 256         }
 257 }
 258
 259 static void print_multipath_conf (multipath_conf_t *conf)
 260 {
 261         int i;
 262         struct multipath_info *tmp;
 263
 264         printk("MULTIPATH conf printout:\n");
 265         if (!conf) {
 266                 printk("(conf==NULL)\n");
 267                 return;
 268         }
 269         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 270                          conf->raid_disks);
 271
 272         for (i = 0; i < conf->raid_disks; i++) {
 273                 char b[BDEVNAME_SIZE];
 274                 tmp = conf->multipaths + i;
 275                 if (tmp->rdev)
 276                         printk(" disk%d, o:%d, dev:%s\n",
 277                                 i,!test_bit(Faulty, &tmp->rdev->flags),
 278                                bdevname(tmp->rdev->bdev,b));
 279         }
 280 }
 281
 282
 283 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 284 {
 285         multipath_conf_t *conf = mddev->private;
 286         struct request_queue *q;
 287         int err = -EEXIST;
 288         int path;
 289         struct multipath_info *p;
 290         int first = 0;
 291         int last = mddev->raid_disks - 1;
 292
 293         if (rdev->raid_disk >= 0)
 294                 first = last = rdev->raid_disk;
 295
 296         print_multipath_conf(conf);
 297
 298         for (path = first; path <= last; path++)
 299                 if ((p=conf->multipaths+path)->rdev == NULL) {
 300                         q = rdev->bdev->bd_disk->queue;
 301                         disk_stack_limits(mddev->gendisk, rdev->bdev,
 302                                           rdev->data_offset << 9);
 303
 304                 /* as we don't honour merge_bvec_fn, we must never risk
 305                  * violating it, so limit ->max_segments to one, lying
 306                  * within a single page.
 307                  * (Note: it is very unlikely that a device with
 308                  * merge_bvec_fn will be involved in multipath.)
 309                  */
 310                         if (q->merge_bvec_fn) {
 311                                 blk_queue_max_segments(mddev->queue, 1);
 312                                 blk_queue_segment_boundary(mddev->queue,
 313                                                            PAGE_CACHE_SIZE - 1);
 314                         }
 315
 316                         conf->working_disks++;
 317                         mddev->degraded--;
 318                         rdev->raid_disk = path;
 319                         set_bit(In_sync, &rdev->flags);
 320                         rcu_assign_pointer(p->rdev, rdev);
 321                         err = 0;
 322                         md_integrity_add_rdev(rdev, mddev);
 323                         break;
 324                 }
 325
 326         print_multipath_conf(conf);
 327
 328         return err;
 329 }
 330
 331 static int multipath_remove_disk(mddev_t *mddev, int number)
 332 {
 333         multipath_conf_t *conf = mddev->private;
 334         int err = 0;
 335         mdk_rdev_t *rdev;
 336         struct multipath_info *p = conf->multipaths + number;
 337
 338         print_multipath_conf(conf);
 339
 340         rdev = p->rdev;
 341         if (rdev) {
 342                 if (test_bit(In_sync, &rdev->flags) ||
 343                     atomic_read(&rdev->nr_pending)) {
 344                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"
 345                                " but is still operational!\n", number);
 346                         err = -EBUSY;
 347                         goto abort;
 348                 }
 349                 p->rdev = NULL;
 350                 synchronize_rcu();
 351                 if (atomic_read(&rdev->nr_pending)) {
 352                         /* lost the race, try later */
 353                         err = -EBUSY;
 354                         p->rdev = rdev;
 355                         goto abort;
 356                 }
 357                 md_integrity_register(mddev);
 358         }
 359 abort:
 360
 361         print_multipath_conf(conf);
 362         return err;
 363 }
 364
 365
 366
 367 /*
 368  * This is a kernel thread which:
 369  *
 370  *      1.      Retries failed read operations on working multipaths.
 371  *      2.      Updates the raid superblock when problems encounter.
 372  *      3.      Performs writes following reads for array syncronising.
 373  */
 374
 375 static void multipathd (mddev_t *mddev)
 376 {
 377         struct multipath_bh *mp_bh;
 378         struct bio *bio;
 379         unsigned long flags;
 380         multipath_conf_t *conf = mddev->private;
 381         struct list_head *head = &conf->retry_list;
 382
 383         md_check_recovery(mddev);
 384         for (;;) {
 385                 char b[BDEVNAME_SIZE];
 386                 spin_lock_irqsave(&conf->device_lock, flags);
 387                 if (list_empty(head))
 388                         break;
 389                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
 390                 list_del(head->prev);
 391                 spin_unlock_irqrestore(&conf->device_lock, flags);
 392
 393                 bio = &mp_bh->bio;
 394                 bio->bi_sector = mp_bh->master_bio->bi_sector;
 395
 396                 if ((mp_bh->path = multipath_map (conf))<0) {
 397                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
 398                                 " error for block %llu\n",
 399                                 bdevname(bio->bi_bdev,b),
 400                                 (unsigned long long)bio->bi_sector);
 401                         multipath_end_bh_io(mp_bh, -EIO);
 402                 } else {
 403                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
 404                                 " to another IO path\n",
 405                                 bdevname(bio->bi_bdev,b),
 406                                 (unsigned long long)bio->bi_sector);
 407                         *bio = *(mp_bh->master_bio);
 408                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
 409                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
 410                         bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
 411                         bio->bi_end_io = multipath_end_request;
 412                         bio->bi_private = mp_bh;
 413                         generic_make_request(bio);
 414                 }
 415         }
 416         spin_unlock_irqrestore(&conf->device_lock, flags);
 417 }
 418
 419 static sector_t multipath_size(mddev_t *mddev, sector_t sectors, int raid_disks)
 420 {
 421         WARN_ONCE(sectors || raid_disks,
 422                   "%s does not support generic reshape\n", __func__);
 423
 424         return mddev->dev_sectors;
 425 }
 426
 427 static int multipath_run (mddev_t *mddev)
 428 {
 429         multipath_conf_t *conf;
 430         int disk_idx;
 431         struct multipath_info *disk;
 432         mdk_rdev_t *rdev;
 433
 434         if (md_check_no_bitmap(mddev))
 435                 return -EINVAL;
 436
 437         if (mddev->level != LEVEL_MULTIPATH) {
 438                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
 439                        mdname(mddev), mddev->level);
 440                 goto out;
 441         }
 442         /*
 443          * copy the already verified devices into our private MULTIPATH
 444          * bookkeeping area. [whatever we allocate in multipath_run(),
 445          * should be freed in multipath_stop()]
 446          */
 447         mddev->queue->queue_lock = &mddev->queue->__queue_lock;
 448
 449         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
 450         mddev->private = conf;
 451         if (!conf) {
 452                 printk(KERN_ERR
 453                         "multipath: couldn't allocate memory for %s\n",
 454                         mdname(mddev));
 455                 goto out;
 456         }
 457
 458         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
 459                                    GFP_KERNEL);
 460         if (!conf->multipaths) {
 461                 printk(KERN_ERR
 462                         "multipath: couldn't allocate memory for %s\n",
 463                         mdname(mddev));
 464                 goto out_free_conf;
 465         }
 466
 467         conf->working_disks = 0;
 468         list_for_each_entry(rdev, &mddev->disks, same_set) {
 469                 disk_idx = rdev->raid_disk;
 470                 if (disk_idx < 0 ||
 471                     disk_idx >= mddev->raid_disks)
 472                         continue;
 473
 474                 disk = conf->multipaths + disk_idx;
 475                 disk->rdev = rdev;
 476                 disk_stack_limits(mddev->gendisk, rdev->bdev,
 477                                   rdev->data_offset << 9);
 478
 479                 /* as we don't honour merge_bvec_fn, we must never risk
 480                  * violating it, not that we ever expect a device with
 481                  * a merge_bvec_fn to be involved in multipath */
 482                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
 483                         blk_queue_max_segments(mddev->queue, 1);
 484                         blk_queue_segment_boundary(mddev->queue,
 485                                                    PAGE_CACHE_SIZE - 1);
 486                 }
 487
 488                 if (!test_bit(Faulty, &rdev->flags))
 489                         conf->working_disks++;
 490         }
 491
 492         conf->raid_disks = mddev->raid_disks;
 493         conf->mddev = mddev;
 494         spin_lock_init(&conf->device_lock);
 495         INIT_LIST_HEAD(&conf->retry_list);
 496
 497         if (!conf->working_disks) {
 498                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
 499                         mdname(mddev));
 500                 goto out_free_conf;
 501         }
 502         mddev->degraded = conf->raid_disks - conf->working_disks;
 503
 504         conf->pool = mempool_create_kmalloc_pool(NR_RESERVED_BUFS,
 505                                                  sizeof(struct multipath_bh));
 506         if (conf->pool == NULL) {
 507                 printk(KERN_ERR
 508                         "multipath: couldn't allocate memory for %s\n",
 509                         mdname(mddev));
 510                 goto out_free_conf;
 511         }
 512
 513         {
 514                 mddev->thread = md_register_thread(multipathd, mddev, NULL);
 515                 if (!mddev->thread) {
 516                         printk(KERN_ERR "multipath: couldn't allocate thread"
 517                                 " for %s\n", mdname(mddev));
 518                         goto out_free_conf;
 519                 }
 520         }
 521
 522         printk(KERN_INFO
 523                 "multipath: array %s active with %d out of %d IO paths\n",
 524                 mdname(mddev), conf->working_disks, mddev->raid_disks);
 525         /*
 526          * Ok, everything is just fine now
 527          */
 528         md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
 529
 530         mddev->queue->unplug_fn = multipath_unplug;
 531         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
 532         mddev->queue->backing_dev_info.congested_data = mddev;
 533         md_integrity_register(mddev);
 534         return 0;
 535
 536 out_free_conf:
 537         if (conf->pool)
 538                 mempool_destroy(conf->pool);
 539         kfree(conf->multipaths);
 540         kfree(conf);
 541         mddev->private = NULL;
 542 out:
 543         return -EIO;
 544 }
 545
 546
 547 static int multipath_stop (mddev_t *mddev)
 548 {
 549         multipath_conf_t *conf = mddev->private;
 550
 551         md_unregister_thread(mddev->thread);
 552         mddev->thread = NULL;
 553         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
 554         mempool_destroy(conf->pool);
 555         kfree(conf->multipaths);
 556         kfree(conf);
 557         mddev->private = NULL;
 558         return 0;
 559 }
 560
 561 static struct mdk_personality multipath_personality =
 562 {
 563         .name           = "multipath",
 564         .level          = LEVEL_MULTIPATH,
 565         .owner          = THIS_MODULE,
 566         .make_request   = multipath_make_request,
 567         .run            = multipath_run,
 568         .stop           = multipath_stop,
 569         .status         = multipath_status,
 570         .error_handler  = multipath_error,
 571         .hot_add_disk   = multipath_add_disk,
 572         .hot_remove_disk= multipath_remove_disk,
 573         .size           = multipath_size,
 574 };
 575
 576 static int __init multipath_init (void)
 577 {
 578         return register_md_personality (&multipath_personality);
 579 }
 580
 581 static void __exit multipath_exit (void)
 582 {
 583         unregister_md_personality (&multipath_personality);
 584 }
 585
 586 module_init(multipath_init);
 587 module_exit(multipath_exit);
 588 MODULE_LICENSE("GPL");
 589 MODULE_DESCRIPTION("simple multi-path personality for MD");
 590 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
 591 MODULE_ALIAS("md-multipath");
 592 MODULE_ALIAS("md-level--4");