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