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