2 * multipath.c : Multiple Devices driver for Linux
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 * MULTIPATH management functions.
10 * derived from raid1.c.
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)
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.
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>
29 #define MAJOR_NR MD_MAJOR
31 #define MD_PERSONALITY
33 #define MAX_WORK_PER_DISK 128
35 #define NR_RESERVED_BUFS 32
38 static mdk_personality_t multipath_personality;
41 static void *mp_pool_alloc(gfp_t gfp_flags, void *data)
43 struct multipath_bh *mpb;
44 mpb = kzalloc(sizeof(*mpb), gfp_flags);
48 static void mp_pool_free(void *mpb, void *data)
53 static int multipath_map (multipath_conf_t *conf)
55 int i, disks = conf->raid_disks;
58 * Later we do read balancing on the read side
59 * now we use the first available disk.
63 for (i = 0; i < disks; i++) {
64 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
65 if (rdev && test_bit(In_sync, &rdev->flags)) {
66 atomic_inc(&rdev->nr_pending);
73 printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
77 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
80 mddev_t *mddev = mp_bh->mddev;
81 multipath_conf_t *conf = mddev_to_conf(mddev);
83 spin_lock_irqsave(&conf->device_lock, flags);
84 list_add(&mp_bh->retry_list, &conf->retry_list);
85 spin_unlock_irqrestore(&conf->device_lock, flags);
86 md_wakeup_thread(mddev->thread);
91 * multipath_end_bh_io() is called when we have finished servicing a multipathed
92 * operation and are ready to return a success/failure code to the buffer
95 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
97 struct bio *bio = mp_bh->master_bio;
98 multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
100 bio_endio(bio, bio->bi_size, err);
101 mempool_free(mp_bh, conf->pool);
104 static int multipath_end_request(struct bio *bio, unsigned int bytes_done,
107 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
108 struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
109 multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
110 mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
116 multipath_end_bh_io(mp_bh, 0);
117 else if (!bio_rw_ahead(bio)) {
121 char b[BDEVNAME_SIZE];
122 md_error (mp_bh->mddev, rdev);
123 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
124 bdevname(rdev->bdev,b),
125 (unsigned long long)bio->bi_sector);
126 multipath_reschedule_retry(mp_bh);
128 multipath_end_bh_io(mp_bh, error);
129 rdev_dec_pending(rdev, conf->mddev);
133 static void unplug_slaves(mddev_t *mddev)
135 multipath_conf_t *conf = mddev_to_conf(mddev);
139 for (i=0; i<mddev->raid_disks; i++) {
140 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
141 if (rdev && !test_bit(Faulty, &rdev->flags)
142 && atomic_read(&rdev->nr_pending)) {
143 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
145 atomic_inc(&rdev->nr_pending);
148 if (r_queue->unplug_fn)
149 r_queue->unplug_fn(r_queue);
151 rdev_dec_pending(rdev, mddev);
158 static void multipath_unplug(request_queue_t *q)
160 unplug_slaves(q->queuedata);
164 static int multipath_make_request (request_queue_t *q, struct bio * bio)
166 mddev_t *mddev = q->queuedata;
167 multipath_conf_t *conf = mddev_to_conf(mddev);
168 struct multipath_bh * mp_bh;
169 struct multipath_info *multipath;
170 const int rw = bio_data_dir(bio);
172 if (unlikely(bio_barrier(bio))) {
173 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
177 mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
179 mp_bh->master_bio = bio;
180 mp_bh->mddev = mddev;
182 disk_stat_inc(mddev->gendisk, ios[rw]);
183 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
185 mp_bh->path = multipath_map(conf);
186 if (mp_bh->path < 0) {
187 bio_endio(bio, bio->bi_size, -EIO);
188 mempool_free(mp_bh, conf->pool);
191 multipath = conf->multipaths + mp_bh->path;
194 mp_bh->bio.bi_sector += multipath->rdev->data_offset;
195 mp_bh->bio.bi_bdev = multipath->rdev->bdev;
196 mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
197 mp_bh->bio.bi_end_io = multipath_end_request;
198 mp_bh->bio.bi_private = mp_bh;
199 generic_make_request(&mp_bh->bio);
203 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
205 multipath_conf_t *conf = mddev_to_conf(mddev);
208 seq_printf (seq, " [%d/%d] [", conf->raid_disks,
209 conf->working_disks);
210 for (i = 0; i < conf->raid_disks; i++)
211 seq_printf (seq, "%s",
212 conf->multipaths[i].rdev &&
213 test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
214 seq_printf (seq, "]");
217 static int multipath_issue_flush(request_queue_t *q, struct gendisk *disk,
218 sector_t *error_sector)
220 mddev_t *mddev = q->queuedata;
221 multipath_conf_t *conf = mddev_to_conf(mddev);
225 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
226 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
227 if (rdev && !test_bit(Faulty, &rdev->flags)) {
228 struct block_device *bdev = rdev->bdev;
229 request_queue_t *r_queue = bdev_get_queue(bdev);
231 if (!r_queue->issue_flush_fn)
234 atomic_inc(&rdev->nr_pending);
236 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
238 rdev_dec_pending(rdev, mddev);
248 * Careful, this can execute in IRQ contexts as well!
250 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
252 multipath_conf_t *conf = mddev_to_conf(mddev);
254 if (conf->working_disks <= 1) {
256 * Uh oh, we can do nothing if this is our last path, but
257 * first check if this is a queued request for a device
258 * which has just failed.
261 "multipath: only one IO path left and IO error.\n");
262 /* leave it active... it's all we have */
265 * Mark disk as unusable
267 if (!test_bit(Faulty, &rdev->flags)) {
268 char b[BDEVNAME_SIZE];
269 clear_bit(In_sync, &rdev->flags);
270 set_bit(Faulty, &rdev->flags);
272 conf->working_disks--;
273 printk(KERN_ALERT "multipath: IO failure on %s,"
274 " disabling IO path. \n Operation continuing"
275 " on %d IO paths.\n",
276 bdevname (rdev->bdev,b),
277 conf->working_disks);
282 static void print_multipath_conf (multipath_conf_t *conf)
285 struct multipath_info *tmp;
287 printk("MULTIPATH conf printout:\n");
289 printk("(conf==NULL)\n");
292 printk(" --- wd:%d rd:%d\n", conf->working_disks,
295 for (i = 0; i < conf->raid_disks; i++) {
296 char b[BDEVNAME_SIZE];
297 tmp = conf->multipaths + i;
299 printk(" disk%d, o:%d, dev:%s\n",
300 i,!test_bit(Faulty, &tmp->rdev->flags),
301 bdevname(tmp->rdev->bdev,b));
306 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
308 multipath_conf_t *conf = mddev->private;
311 struct multipath_info *p;
313 print_multipath_conf(conf);
315 for (path=0; path<mddev->raid_disks; path++)
316 if ((p=conf->multipaths+path)->rdev == NULL) {
317 blk_queue_stack_limits(mddev->queue,
318 rdev->bdev->bd_disk->queue);
320 /* as we don't honour merge_bvec_fn, we must never risk
321 * violating it, so limit ->max_sector to one PAGE, as
322 * a one page request is never in violation.
323 * (Note: it is very unlikely that a device with
324 * merge_bvec_fn will be involved in multipath.)
326 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
327 mddev->queue->max_sectors > (PAGE_SIZE>>9))
328 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
330 conf->working_disks++;
331 rdev->raid_disk = path;
332 set_bit(In_sync, &rdev->flags);
333 rcu_assign_pointer(p->rdev, rdev);
337 print_multipath_conf(conf);
341 static int multipath_remove_disk(mddev_t *mddev, int number)
343 multipath_conf_t *conf = mddev->private;
346 struct multipath_info *p = conf->multipaths + number;
348 print_multipath_conf(conf);
352 if (test_bit(In_sync, &rdev->flags) ||
353 atomic_read(&rdev->nr_pending)) {
354 printk(KERN_ERR "hot-remove-disk, slot %d is identified" " but is still operational!\n", number);
360 if (atomic_read(&rdev->nr_pending)) {
361 /* lost the race, try later */
368 print_multipath_conf(conf);
375 * This is a kernel thread which:
377 * 1. Retries failed read operations on working multipaths.
378 * 2. Updates the raid superblock when problems encounter.
379 * 3. Performs writes following reads for array syncronising.
382 static void multipathd (mddev_t *mddev)
384 struct multipath_bh *mp_bh;
387 multipath_conf_t *conf = mddev_to_conf(mddev);
388 struct list_head *head = &conf->retry_list;
390 md_check_recovery(mddev);
392 char b[BDEVNAME_SIZE];
393 spin_lock_irqsave(&conf->device_lock, flags);
394 if (list_empty(head))
396 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
397 list_del(head->prev);
398 spin_unlock_irqrestore(&conf->device_lock, flags);
401 bio->bi_sector = mp_bh->master_bio->bi_sector;
403 if ((mp_bh->path = multipath_map (conf))<0) {
404 printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
405 " error for block %llu\n",
406 bdevname(bio->bi_bdev,b),
407 (unsigned long long)bio->bi_sector);
408 multipath_end_bh_io(mp_bh, -EIO);
410 printk(KERN_ERR "multipath: %s: redirecting sector %llu"
411 " to another IO path\n",
412 bdevname(bio->bi_bdev,b),
413 (unsigned long long)bio->bi_sector);
414 *bio = *(mp_bh->master_bio);
415 bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
416 bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
417 bio->bi_rw |= (1 << BIO_RW_FAILFAST);
418 bio->bi_end_io = multipath_end_request;
419 bio->bi_private = mp_bh;
420 generic_make_request(bio);
423 spin_unlock_irqrestore(&conf->device_lock, flags);
426 static int multipath_run (mddev_t *mddev)
428 multipath_conf_t *conf;
430 struct multipath_info *disk;
432 struct list_head *tmp;
434 if (mddev->level != LEVEL_MULTIPATH) {
435 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
436 mdname(mddev), mddev->level);
440 * copy the already verified devices into our private MULTIPATH
441 * bookkeeping area. [whatever we allocate in multipath_run(),
442 * should be freed in multipath_stop()]
445 conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
446 mddev->private = conf;
449 "multipath: couldn't allocate memory for %s\n",
454 conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
456 if (!conf->multipaths) {
458 "multipath: couldn't allocate memory for %s\n",
463 conf->working_disks = 0;
464 ITERATE_RDEV(mddev,rdev,tmp) {
465 disk_idx = rdev->raid_disk;
467 disk_idx >= mddev->raid_disks)
470 disk = conf->multipaths + disk_idx;
473 blk_queue_stack_limits(mddev->queue,
474 rdev->bdev->bd_disk->queue);
475 /* as we don't honour merge_bvec_fn, we must never risk
476 * violating it, not that we ever expect a device with
477 * a merge_bvec_fn to be involved in multipath */
478 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
479 mddev->queue->max_sectors > (PAGE_SIZE>>9))
480 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
482 if (!test_bit(Faulty, &rdev->flags))
483 conf->working_disks++;
486 conf->raid_disks = mddev->raid_disks;
489 spin_lock_init(&conf->device_lock);
490 INIT_LIST_HEAD(&conf->retry_list);
492 if (!conf->working_disks) {
493 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
497 mddev->degraded = conf->raid_disks = conf->working_disks;
499 conf->pool = mempool_create(NR_RESERVED_BUFS,
500 mp_pool_alloc, mp_pool_free,
502 if (conf->pool == NULL) {
504 "multipath: couldn't allocate memory for %s\n",
510 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
511 if (!mddev->thread) {
512 printk(KERN_ERR "multipath: couldn't allocate thread"
513 " for %s\n", mdname(mddev));
519 "multipath: array %s active with %d out of %d IO paths\n",
520 mdname(mddev), conf->working_disks, mddev->raid_disks);
522 * Ok, everything is just fine now
524 mddev->array_size = mddev->size;
526 mddev->queue->unplug_fn = multipath_unplug;
527 mddev->queue->issue_flush_fn = multipath_issue_flush;
533 mempool_destroy(conf->pool);
534 kfree(conf->multipaths);
536 mddev->private = NULL;
542 static int multipath_stop (mddev_t *mddev)
544 multipath_conf_t *conf = mddev_to_conf(mddev);
546 md_unregister_thread(mddev->thread);
547 mddev->thread = NULL;
548 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
549 mempool_destroy(conf->pool);
550 kfree(conf->multipaths);
552 mddev->private = NULL;
556 static mdk_personality_t multipath_personality=
559 .owner = THIS_MODULE,
560 .make_request = multipath_make_request,
561 .run = multipath_run,
562 .stop = multipath_stop,
563 .status = multipath_status,
564 .error_handler = multipath_error,
565 .hot_add_disk = multipath_add_disk,
566 .hot_remove_disk= multipath_remove_disk,
569 static int __init multipath_init (void)
571 return register_md_personality (MULTIPATH, &multipath_personality);
574 static void __exit multipath_exit (void)
576 unregister_md_personality (MULTIPATH);
579 module_init(multipath_init);
580 module_exit(multipath_exit);
581 MODULE_LICENSE("GPL");
582 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */