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/module.h>
  23 #include <linux/slab.h>
  24 #include <linux/spinlock.h>
  25 #include <linux/raid/multipath.h>
  26 #include <linux/buffer_head.h>
  27 #include <asm/atomic.h>
  28
  29 #define MAJOR_NR MD_MAJOR
  30 #define MD_DRIVER
  31 #define MD_PERSONALITY
  32 #define DEVICE_NR(device) (minor(device))
  33
  34 #define MAX_WORK_PER_DISK 128
  35
  36 #define NR_RESERVED_BUFS        32
  37
  38
  39 static mdk_personality_t multipath_personality;
  40 static spinlock_t retry_list_lock = SPIN_LOCK_UNLOCKED;
  41 struct multipath_bh *multipath_retry_list = NULL, **multipath_retry_tail;
  42
  43
  44 static void *mp_pool_alloc(int gfp_flags, void *data)
  45 {
  46         struct multipath_bh *mpb;
  47         mpb = kmalloc(sizeof(*mpb), gfp_flags);
  48         if (mpb)
  49                 memset(mpb, 0, sizeof(*mpb));
  50         return mpb;
  51 }
  52
  53 static void mp_pool_free(void *mpb, void *data)
  54 {
  55         kfree(mpb);
  56 }
  57
  58 static int multipath_map (mddev_t *mddev, mdk_rdev_t **rdevp)
  59 {
  60         multipath_conf_t *conf = mddev_to_conf(mddev);
  61         int i, disks = conf->raid_disks;
  62
  63         /*
  64          * Later we do read balancing on the read side
  65          * now we use the first available disk.
  66          */
  67
  68         spin_lock_irq(&conf->device_lock);
  69         for (i = 0; i < disks; i++) {
  70                 mdk_rdev_t *rdev = conf->multipaths[i].rdev;
  71                 if (rdev && rdev->in_sync) {
  72                         *rdevp = rdev;
  73                         atomic_inc(&rdev->nr_pending);
  74                         spin_unlock_irq(&conf->device_lock);
  75                         return 0;
  76                 }
  77         }
  78         spin_unlock_irq(&conf->device_lock);
  79
  80         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
  81         return (-1);
  82 }
  83
  84 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
  85 {
  86         unsigned long flags;
  87         mddev_t *mddev = mp_bh->mddev;
  88
  89         spin_lock_irqsave(&retry_list_lock, flags);
  90         if (multipath_retry_list == NULL)
  91                 multipath_retry_tail = &multipath_retry_list;
  92         *multipath_retry_tail = mp_bh;
  93         multipath_retry_tail = &mp_bh->next_mp;
  94         mp_bh->next_mp = NULL;
  95         spin_unlock_irqrestore(&retry_list_lock, flags);
  96         md_wakeup_thread(mddev->thread);
  97 }
  98
  99
 100 /*
 101  * multipath_end_bh_io() is called when we have finished servicing a multipathed
 102  * operation and are ready to return a success/failure code to the buffer
 103  * cache layer.
 104  */
 105 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int uptodate)
 106 {
 107         struct bio *bio = mp_bh->master_bio;
 108         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
 109
 110         bio_endio(bio, bio->bi_size, uptodate ? 0 : -EIO);
 111         mempool_free(mp_bh, conf->pool);
 112 }
 113
 114 int multipath_end_request(struct bio *bio, unsigned int bytes_done, int error)
 115 {
 116         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 117         struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
 118         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
 119         mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
 120
 121         if (bio->bi_size)
 122                 return 1;
 123
 124         if (uptodate)
 125                 multipath_end_bh_io(mp_bh, uptodate);
 126         else {
 127                 /*
 128                  * oops, IO error:
 129                  */
 130                 char b[BDEVNAME_SIZE];
 131                 md_error (mp_bh->mddev, rdev);
 132                 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
 133                        bdevname(rdev->bdev,b),
 134                        (unsigned long long)bio->bi_sector);
 135                 multipath_reschedule_retry(mp_bh);
 136         }
 137         atomic_dec(&rdev->nr_pending);
 138         return 0;
 139 }
 140
 141 /*
 142  * This routine returns the disk from which the requested read should
 143  * be done.
 144  */
 145
 146 static int multipath_read_balance (multipath_conf_t *conf)
 147 {
 148         int disk;
 149
 150         for (disk = 0; disk < conf->raid_disks; disk++) {
 151                 mdk_rdev_t *rdev = conf->multipaths[disk].rdev;
 152                 if (rdev && rdev->in_sync)
 153                         return disk;
 154         }
 155         BUG();
 156         return 0;
 157 }
 158
 159 static int multipath_make_request (request_queue_t *q, struct bio * bio)
 160 {
 161         mddev_t *mddev = q->queuedata;
 162         multipath_conf_t *conf = mddev_to_conf(mddev);
 163         struct multipath_bh * mp_bh;
 164         struct multipath_info *multipath;
 165
 166         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
 167
 168         mp_bh->master_bio = bio;
 169         mp_bh->mddev = mddev;
 170
 171         /*
 172          * read balancing logic:
 173          */
 174         spin_lock_irq(&conf->device_lock);
 175         mp_bh->path = multipath_read_balance(conf);
 176         multipath = conf->multipaths + mp_bh->path;
 177         atomic_inc(&multipath->rdev->nr_pending);
 178         spin_unlock_irq(&conf->device_lock);
 179
 180         mp_bh->bio = *bio;
 181         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
 182         mp_bh->bio.bi_end_io = multipath_end_request;
 183         mp_bh->bio.bi_private = mp_bh;
 184         generic_make_request(&mp_bh->bio);
 185         return 0;
 186 }
 187
 188 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
 189 {
 190         multipath_conf_t *conf = mddev_to_conf(mddev);
 191         int i;
 192
 193         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
 194                                                  conf->working_disks);
 195         for (i = 0; i < conf->raid_disks; i++)
 196                 seq_printf (seq, "%s",
 197                                conf->multipaths[i].rdev &&
 198                                conf->multipaths[i].rdev->in_sync ? "U" : "_");
 199         seq_printf (seq, "]");
 200 }
 201
 202
 203 /*
 204  * Careful, this can execute in IRQ contexts as well!
 205  */
 206 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
 207 {
 208         multipath_conf_t *conf = mddev_to_conf(mddev);
 209
 210         if (conf->working_disks <= 1) {
 211                 /*
 212                  * Uh oh, we can do nothing if this is our last path, but
 213                  * first check if this is a queued request for a device
 214                  * which has just failed.
 215                  */
 216                 printk(KERN_ALERT
 217                         "multipath: only one IO path left and IO error.\n");
 218                 /* leave it active... it's all we have */
 219         } else {
 220                 /*
 221                  * Mark disk as unusable
 222                  */
 223                 if (!rdev->faulty) {
 224                         char b[BDEVNAME_SIZE];
 225                         rdev->in_sync = 0;
 226                         rdev->faulty = 1;
 227                         mddev->sb_dirty = 1;
 228                         conf->working_disks--;
 229                         printk(KERN_ALERT "multipath: IO failure on %s,"
 230                                 " disabling IO path. \n Operation continuing"
 231                                 " on %d IO paths.\n",
 232                                 bdevname (rdev->bdev,b),
 233                                 conf->working_disks);
 234                 }
 235         }
 236 }
 237
 238 static void print_multipath_conf (multipath_conf_t *conf)
 239 {
 240         int i;
 241         struct multipath_info *tmp;
 242
 243         printk("MULTIPATH conf printout:\n");
 244         if (!conf) {
 245                 printk("(conf==NULL)\n");
 246                 return;
 247         }
 248         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 249                          conf->raid_disks);
 250
 251         for (i = 0; i < conf->raid_disks; i++) {
 252                 char b[BDEVNAME_SIZE];
 253                 tmp = conf->multipaths + i;
 254                 if (tmp->rdev)
 255                         printk(" disk%d, o:%d, dev:%s\n",
 256                                 i,!tmp->rdev->faulty,
 257                                bdevname(tmp->rdev->bdev,b));
 258         }
 259 }
 260
 261
 262 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 263 {
 264         multipath_conf_t *conf = mddev->private;
 265         int found = 0;
 266         int path;
 267         struct multipath_info *p;
 268
 269         print_multipath_conf(conf);
 270         spin_lock_irq(&conf->device_lock);
 271         for (path=0; path<mddev->raid_disks; path++)
 272                 if ((p=conf->multipaths+path)->rdev == NULL) {
 273                         p->rdev = rdev;
 274                         conf->working_disks++;
 275                         rdev->raid_disk = path;
 276                         rdev->in_sync = 1;
 277                         found = 1;
 278                 }
 279         spin_unlock_irq(&conf->device_lock);
 280
 281         print_multipath_conf(conf);
 282         return found;
 283 }
 284
 285 static int multipath_remove_disk(mddev_t *mddev, int number)
 286 {
 287         multipath_conf_t *conf = mddev->private;
 288         int err = 1;
 289         struct multipath_info *p = conf->multipaths + number;
 290
 291         print_multipath_conf(conf);
 292         spin_lock_irq(&conf->device_lock);
 293
 294         if (p->rdev) {
 295                 if (p->rdev->in_sync ||
 296                     atomic_read(&p->rdev->nr_pending)) {
 297                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"                                " but is still operational!\n", number);
 298                         err = -EBUSY;
 299                         goto abort;
 300                 }
 301                 p->rdev = NULL;
 302                 err = 0;
 303         }
 304         if (err)
 305                 MD_BUG();
 306 abort:
 307         spin_unlock_irq(&conf->device_lock);
 308
 309         print_multipath_conf(conf);
 310         return err;
 311 }
 312
 313
 314
 315 /*
 316  * This is a kernel thread which:
 317  *
 318  *      1.      Retries failed read operations on working multipaths.
 319  *      2.      Updates the raid superblock when problems encounter.
 320  *      3.      Performs writes following reads for array syncronising.
 321  */
 322
 323 static void multipathd (mddev_t *mddev)
 324 {
 325         struct multipath_bh *mp_bh;
 326         struct bio *bio;
 327         unsigned long flags;
 328         mdk_rdev_t *rdev;
 329
 330         md_check_recovery(mddev);
 331         for (;;) {
 332                 char b[BDEVNAME_SIZE];
 333                 spin_lock_irqsave(&retry_list_lock, flags);
 334                 mp_bh = multipath_retry_list;
 335                 if (!mp_bh)
 336                         break;
 337                 multipath_retry_list = mp_bh->next_mp;
 338                 spin_unlock_irqrestore(&retry_list_lock, flags);
 339
 340                 mddev = mp_bh->mddev;
 341                 bio = &mp_bh->bio;
 342                 bio->bi_sector = mp_bh->master_bio->bi_sector;
 343
 344                 rdev = NULL;
 345                 if (multipath_map (mddev, &rdev)<0) {
 346                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
 347                                 " error for block %llu\n",
 348                                 bdevname(bio->bi_bdev,b),
 349                                 (unsigned long long)bio->bi_sector);
 350                         multipath_end_bh_io(mp_bh, 0);
 351                 } else {
 352                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
 353                                 " to another IO path\n",
 354                                 bdevname(bio->bi_bdev,b),
 355                                 (unsigned long long)bio->bi_sector);
 356                         bio->bi_bdev = rdev->bdev;
 357                         generic_make_request(bio);
 358                 }
 359         }
 360         spin_unlock_irqrestore(&retry_list_lock, flags);
 361 }
 362
 363 static int multipath_run (mddev_t *mddev)
 364 {
 365         multipath_conf_t *conf;
 366         int disk_idx;
 367         struct multipath_info *disk;
 368         mdk_rdev_t *rdev;
 369         struct list_head *tmp;
 370
 371         if (mddev->level != LEVEL_MULTIPATH) {
 372                 printk("multipath: md%d: raid level not set to multipath IO (%d)\n",
 373                        mdidx(mddev), mddev->level);
 374                 goto out;
 375         }
 376         /*
 377          * copy the already verified devices into our private MULTIPATH
 378          * bookkeeping area. [whatever we allocate in multipath_run(),
 379          * should be freed in multipath_stop()]
 380          */
 381
 382         conf = kmalloc(sizeof(multipath_conf_t), GFP_KERNEL);
 383         mddev->private = conf;
 384         if (!conf) {
 385                 printk(KERN_ERR
 386                         "multipath: couldn't allocate memory for md%d\n",
 387                         mdidx(mddev));
 388                 goto out;
 389         }
 390         memset(conf, 0, sizeof(*conf));
 391
 392         conf->multipaths = kmalloc(sizeof(struct multipath_info)*mddev->raid_disks,
 393                                    GFP_KERNEL);
 394         if (!conf->multipaths) {
 395                 printk(KERN_ERR
 396                         "multipath: couldn't allocate memory for md%d\n",
 397                         mdidx(mddev));
 398                 goto out_free_conf;
 399         }
 400         memset(conf->multipaths, 0, sizeof(struct multipath_info)*mddev->raid_disks);
 401
 402         conf->working_disks = 0;
 403         ITERATE_RDEV(mddev,rdev,tmp) {
 404                 disk_idx = rdev->raid_disk;
 405                 if (disk_idx < 0 ||
 406                     disk_idx >= mddev->raid_disks)
 407                         continue;
 408
 409                 disk = conf->multipaths + disk_idx;
 410                 disk->rdev = rdev;
 411                 if (!rdev->faulty)
 412                         conf->working_disks++;
 413         }
 414
 415         conf->raid_disks = mddev->raid_disks;
 416         mddev->sb_dirty = 1;
 417         conf->mddev = mddev;
 418         conf->device_lock = SPIN_LOCK_UNLOCKED;
 419
 420         if (!conf->working_disks) {
 421                 printk(KERN_ERR "multipath: no operational IO paths for md%d\n",
 422                         mdidx(mddev));
 423                 goto out_free_conf;
 424         }
 425         mddev->degraded = conf->raid_disks = conf->working_disks;
 426
 427         conf->pool = mempool_create(NR_RESERVED_BUFS,
 428                                     mp_pool_alloc, mp_pool_free,
 429                                     NULL);
 430         if (conf->pool == NULL) {
 431                 printk(KERN_ERR
 432                         "multipath: couldn't allocate memory for md%d\n",
 433                         mdidx(mddev));
 434                 goto out_free_conf;
 435         }
 436
 437         {
 438                 const char * name = "md%d_multipath";
 439
 440                 mddev->thread = md_register_thread(multipathd, mddev, name);
 441                 if (!mddev->thread) {
 442                         printk(KERN_ERR "multipath: couldn't allocate thread"
 443                                 " for md%d\n", mdidx(mddev));
 444                         goto out_free_conf;
 445                 }
 446         }
 447
 448         printk(KERN_INFO
 449                 "multipath: array md%d active with %d out of %d IO paths\n",
 450                 mdidx(mddev), conf->working_disks, mddev->raid_disks);
 451         /*
 452          * Ok, everything is just fine now
 453          */
 454         mddev->array_size = mddev->size;
 455         return 0;
 456
 457 out_free_conf:
 458         if (conf->pool)
 459                 mempool_destroy(conf->pool);
 460         if (conf->multipaths)
 461                 kfree(conf->multipaths);
 462         kfree(conf);
 463         mddev->private = NULL;
 464 out:
 465         return -EIO;
 466 }
 467
 468
 469 static int multipath_stop (mddev_t *mddev)
 470 {
 471         multipath_conf_t *conf = mddev_to_conf(mddev);
 472
 473         md_unregister_thread(mddev->thread);
 474         mempool_destroy(conf->pool);
 475         kfree(conf->multipaths);
 476         kfree(conf);
 477         mddev->private = NULL;
 478         return 0;
 479 }
 480
 481 static mdk_personality_t multipath_personality=
 482 {
 483         .name           = "multipath",
 484         .owner          = THIS_MODULE,
 485         .make_request   = multipath_make_request,
 486         .run            = multipath_run,
 487         .stop           = multipath_stop,
 488         .status         = multipath_status,
 489         .error_handler  = multipath_error,
 490         .hot_add_disk   = multipath_add_disk,
 491         .hot_remove_disk= multipath_remove_disk,
 492 };
 493
 494 static int __init multipath_init (void)
 495 {
 496         return register_md_personality (MULTIPATH, &multipath_personality);
 497 }
 498
 499 static void __exit multipath_exit (void)
 500 {
 501         unregister_md_personality (MULTIPATH);
 502 }
 503
 504 module_init(multipath_init);
 505 module_exit(multipath_exit);
 506 MODULE_LICENSE("GPL");