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 int multipath_map (multipath_conf_t *conf)
40 int i, disks = conf->raid_disks;
43 * Later we do read balancing on the read side
44 * now we use the first available disk.
48 for (i = 0; i < disks; i++) {
49 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
50 if (rdev && test_bit(In_sync, &rdev->flags)) {
51 atomic_inc(&rdev->nr_pending);
58 printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
62 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
65 mddev_t *mddev = mp_bh->mddev;
66 multipath_conf_t *conf = mddev_to_conf(mddev);
68 spin_lock_irqsave(&conf->device_lock, flags);
69 list_add(&mp_bh->retry_list, &conf->retry_list);
70 spin_unlock_irqrestore(&conf->device_lock, flags);
71 md_wakeup_thread(mddev->thread);
76 * multipath_end_bh_io() is called when we have finished servicing a multipathed
77 * operation and are ready to return a success/failure code to the buffer
80 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
82 struct bio *bio = mp_bh->master_bio;
83 multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
86 mempool_free(mp_bh, conf->pool);
89 static void multipath_end_request(struct bio *bio, int error)
91 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
92 struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
93 multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
94 mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
97 multipath_end_bh_io(mp_bh, 0);
98 else if (!bio_rw_ahead(bio)) {
102 char b[BDEVNAME_SIZE];
103 md_error (mp_bh->mddev, rdev);
104 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
105 bdevname(rdev->bdev,b),
106 (unsigned long long)bio->bi_sector);
107 multipath_reschedule_retry(mp_bh);
109 multipath_end_bh_io(mp_bh, error);
110 rdev_dec_pending(rdev, conf->mddev);
113 static void unplug_slaves(mddev_t *mddev)
115 multipath_conf_t *conf = mddev_to_conf(mddev);
119 for (i=0; i<mddev->raid_disks; i++) {
120 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
121 if (rdev && !test_bit(Faulty, &rdev->flags)
122 && atomic_read(&rdev->nr_pending)) {
123 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
125 atomic_inc(&rdev->nr_pending);
130 rdev_dec_pending(rdev, mddev);
137 static void multipath_unplug(struct request_queue *q)
139 unplug_slaves(q->queuedata);
143 static int multipath_make_request (struct request_queue *q, struct bio * bio)
145 mddev_t *mddev = q->queuedata;
146 multipath_conf_t *conf = mddev_to_conf(mddev);
147 struct multipath_bh * mp_bh;
148 struct multipath_info *multipath;
149 const int rw = bio_data_dir(bio);
152 if (unlikely(bio_barrier(bio))) {
153 bio_endio(bio, -EOPNOTSUPP);
157 mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
159 mp_bh->master_bio = bio;
160 mp_bh->mddev = mddev;
162 cpu = disk_stat_lock();
163 disk_stat_inc(cpu, mddev->gendisk, ios[rw]);
164 disk_stat_add(cpu, mddev->gendisk, sectors[rw], bio_sectors(bio));
167 mp_bh->path = multipath_map(conf);
168 if (mp_bh->path < 0) {
169 bio_endio(bio, -EIO);
170 mempool_free(mp_bh, conf->pool);
173 multipath = conf->multipaths + mp_bh->path;
176 mp_bh->bio.bi_sector += multipath->rdev->data_offset;
177 mp_bh->bio.bi_bdev = multipath->rdev->bdev;
178 mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
179 mp_bh->bio.bi_end_io = multipath_end_request;
180 mp_bh->bio.bi_private = mp_bh;
181 generic_make_request(&mp_bh->bio);
185 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
187 multipath_conf_t *conf = mddev_to_conf(mddev);
190 seq_printf (seq, " [%d/%d] [", conf->raid_disks,
191 conf->working_disks);
192 for (i = 0; i < conf->raid_disks; i++)
193 seq_printf (seq, "%s",
194 conf->multipaths[i].rdev &&
195 test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
196 seq_printf (seq, "]");
199 static int multipath_congested(void *data, int bits)
201 mddev_t *mddev = data;
202 multipath_conf_t *conf = mddev_to_conf(mddev);
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);
211 ret |= bdi_congested(&q->backing_dev_info, bits);
212 /* Just like multipath_map, we just check the
213 * first available device
223 * Careful, this can execute in IRQ contexts as well!
225 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
227 multipath_conf_t *conf = mddev_to_conf(mddev);
229 if (conf->working_disks <= 1) {
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.
236 "multipath: only one IO path left and IO error.\n");
237 /* leave it active... it's all we have */
240 * Mark disk as unusable
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--;
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);
259 static void print_multipath_conf (multipath_conf_t *conf)
262 struct multipath_info *tmp;
264 printk("MULTIPATH conf printout:\n");
266 printk("(conf==NULL)\n");
269 printk(" --- wd:%d rd:%d\n", conf->working_disks,
272 for (i = 0; i < conf->raid_disks; i++) {
273 char b[BDEVNAME_SIZE];
274 tmp = conf->multipaths + i;
276 printk(" disk%d, o:%d, dev:%s\n",
277 i,!test_bit(Faulty, &tmp->rdev->flags),
278 bdevname(tmp->rdev->bdev,b));
283 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
285 multipath_conf_t *conf = mddev->private;
286 struct request_queue *q;
289 struct multipath_info *p;
291 int last = mddev->raid_disks - 1;
293 if (rdev->raid_disk >= 0)
294 first = last = rdev->raid_disk;
296 print_multipath_conf(conf);
298 for (path = first; path <= last; path++)
299 if ((p=conf->multipaths+path)->rdev == NULL) {
300 q = rdev->bdev->bd_disk->queue;
301 blk_queue_stack_limits(mddev->queue, q);
303 /* as we don't honour merge_bvec_fn, we must never risk
304 * violating it, so limit ->max_sector to one PAGE, as
305 * a one page request is never in violation.
306 * (Note: it is very unlikely that a device with
307 * merge_bvec_fn will be involved in multipath.)
309 if (q->merge_bvec_fn &&
310 mddev->queue->max_sectors > (PAGE_SIZE>>9))
311 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
313 conf->working_disks++;
315 rdev->raid_disk = path;
316 set_bit(In_sync, &rdev->flags);
317 rcu_assign_pointer(p->rdev, rdev);
322 print_multipath_conf(conf);
327 static int multipath_remove_disk(mddev_t *mddev, int number)
329 multipath_conf_t *conf = mddev->private;
332 struct multipath_info *p = conf->multipaths + number;
334 print_multipath_conf(conf);
338 if (test_bit(In_sync, &rdev->flags) ||
339 atomic_read(&rdev->nr_pending)) {
340 printk(KERN_ERR "hot-remove-disk, slot %d is identified"
341 " but is still operational!\n", number);
347 if (atomic_read(&rdev->nr_pending)) {
348 /* lost the race, try later */
355 print_multipath_conf(conf);
362 * This is a kernel thread which:
364 * 1. Retries failed read operations on working multipaths.
365 * 2. Updates the raid superblock when problems encounter.
366 * 3. Performs writes following reads for array syncronising.
369 static void multipathd (mddev_t *mddev)
371 struct multipath_bh *mp_bh;
374 multipath_conf_t *conf = mddev_to_conf(mddev);
375 struct list_head *head = &conf->retry_list;
377 md_check_recovery(mddev);
379 char b[BDEVNAME_SIZE];
380 spin_lock_irqsave(&conf->device_lock, flags);
381 if (list_empty(head))
383 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
384 list_del(head->prev);
385 spin_unlock_irqrestore(&conf->device_lock, flags);
388 bio->bi_sector = mp_bh->master_bio->bi_sector;
390 if ((mp_bh->path = multipath_map (conf))<0) {
391 printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
392 " error for block %llu\n",
393 bdevname(bio->bi_bdev,b),
394 (unsigned long long)bio->bi_sector);
395 multipath_end_bh_io(mp_bh, -EIO);
397 printk(KERN_ERR "multipath: %s: redirecting sector %llu"
398 " to another IO path\n",
399 bdevname(bio->bi_bdev,b),
400 (unsigned long long)bio->bi_sector);
401 *bio = *(mp_bh->master_bio);
402 bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
403 bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
404 bio->bi_rw |= (1 << BIO_RW_FAILFAST);
405 bio->bi_end_io = multipath_end_request;
406 bio->bi_private = mp_bh;
407 generic_make_request(bio);
410 spin_unlock_irqrestore(&conf->device_lock, flags);
413 static int multipath_run (mddev_t *mddev)
415 multipath_conf_t *conf;
417 struct multipath_info *disk;
419 struct list_head *tmp;
421 if (mddev->level != LEVEL_MULTIPATH) {
422 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
423 mdname(mddev), mddev->level);
427 * copy the already verified devices into our private MULTIPATH
428 * bookkeeping area. [whatever we allocate in multipath_run(),
429 * should be freed in multipath_stop()]
431 mddev->queue->queue_lock = &mddev->queue->__queue_lock;
433 conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
434 mddev->private = conf;
437 "multipath: couldn't allocate memory for %s\n",
442 conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
444 if (!conf->multipaths) {
446 "multipath: couldn't allocate memory for %s\n",
451 conf->working_disks = 0;
452 rdev_for_each(rdev, tmp, mddev) {
453 disk_idx = rdev->raid_disk;
455 disk_idx >= mddev->raid_disks)
458 disk = conf->multipaths + disk_idx;
461 blk_queue_stack_limits(mddev->queue,
462 rdev->bdev->bd_disk->queue);
463 /* as we don't honour merge_bvec_fn, we must never risk
464 * violating it, not that we ever expect a device with
465 * a merge_bvec_fn to be involved in multipath */
466 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
467 mddev->queue->max_sectors > (PAGE_SIZE>>9))
468 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
470 if (!test_bit(Faulty, &rdev->flags))
471 conf->working_disks++;
474 conf->raid_disks = mddev->raid_disks;
476 spin_lock_init(&conf->device_lock);
477 INIT_LIST_HEAD(&conf->retry_list);
479 if (!conf->working_disks) {
480 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
484 mddev->degraded = conf->raid_disks - conf->working_disks;
486 conf->pool = mempool_create_kzalloc_pool(NR_RESERVED_BUFS,
487 sizeof(struct multipath_bh));
488 if (conf->pool == NULL) {
490 "multipath: couldn't allocate memory for %s\n",
496 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
497 if (!mddev->thread) {
498 printk(KERN_ERR "multipath: couldn't allocate thread"
499 " for %s\n", mdname(mddev));
505 "multipath: array %s active with %d out of %d IO paths\n",
506 mdname(mddev), conf->working_disks, mddev->raid_disks);
508 * Ok, everything is just fine now
510 mddev->array_sectors = mddev->size * 2;
512 mddev->queue->unplug_fn = multipath_unplug;
513 mddev->queue->backing_dev_info.congested_fn = multipath_congested;
514 mddev->queue->backing_dev_info.congested_data = mddev;
520 mempool_destroy(conf->pool);
521 kfree(conf->multipaths);
523 mddev->private = NULL;
529 static int multipath_stop (mddev_t *mddev)
531 multipath_conf_t *conf = mddev_to_conf(mddev);
533 md_unregister_thread(mddev->thread);
534 mddev->thread = NULL;
535 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
536 mempool_destroy(conf->pool);
537 kfree(conf->multipaths);
539 mddev->private = NULL;
543 static struct mdk_personality multipath_personality =
546 .level = LEVEL_MULTIPATH,
547 .owner = THIS_MODULE,
548 .make_request = multipath_make_request,
549 .run = multipath_run,
550 .stop = multipath_stop,
551 .status = multipath_status,
552 .error_handler = multipath_error,
553 .hot_add_disk = multipath_add_disk,
554 .hot_remove_disk= multipath_remove_disk,
557 static int __init multipath_init (void)
559 return register_md_personality (&multipath_personality);
562 static void __exit multipath_exit (void)
564 unregister_md_personality (&multipath_personality);
567 module_init(multipath_init);
568 module_exit(multipath_exit);
569 MODULE_LICENSE("GPL");
570 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
571 MODULE_ALIAS("md-multipath");
572 MODULE_ALIAS("md-level--4");