blob: ae2c5fd6105d5daadfa9db3a020f338ef6d0ae63 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
33#define MAX_WORK_PER_DISK 128
34
35#define NR_RESERVED_BUFS 32
36
37
38static mdk_personality_t multipath_personality;
39
40
Al Virodd0fc662005-10-07 07:46:04 +010041static void *mp_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070042{
43 struct multipath_bh *mpb;
44 mpb = kmalloc(sizeof(*mpb), gfp_flags);
45 if (mpb)
46 memset(mpb, 0, sizeof(*mpb));
47 return mpb;
48}
49
50static void mp_pool_free(void *mpb, void *data)
51{
52 kfree(mpb);
53}
54
55static int multipath_map (multipath_conf_t *conf)
56{
57 int i, disks = conf->raid_disks;
58
59 /*
60 * Later we do read balancing on the read side
61 * now we use the first available disk.
62 */
63
64 rcu_read_lock();
65 for (i = 0; i < disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -080066 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -070067 if (rdev && rdev->in_sync) {
68 atomic_inc(&rdev->nr_pending);
69 rcu_read_unlock();
70 return i;
71 }
72 }
73 rcu_read_unlock();
74
75 printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
76 return (-1);
77}
78
79static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
80{
81 unsigned long flags;
82 mddev_t *mddev = mp_bh->mddev;
83 multipath_conf_t *conf = mddev_to_conf(mddev);
84
85 spin_lock_irqsave(&conf->device_lock, flags);
86 list_add(&mp_bh->retry_list, &conf->retry_list);
87 spin_unlock_irqrestore(&conf->device_lock, flags);
88 md_wakeup_thread(mddev->thread);
89}
90
91
92/*
93 * multipath_end_bh_io() is called when we have finished servicing a multipathed
94 * operation and are ready to return a success/failure code to the buffer
95 * cache layer.
96 */
97static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
98{
99 struct bio *bio = mp_bh->master_bio;
100 multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
101
102 bio_endio(bio, bio->bi_size, err);
103 mempool_free(mp_bh, conf->pool);
104}
105
Adrian Bunk75c96f82005-05-05 16:16:09 -0700106static int multipath_end_request(struct bio *bio, unsigned int bytes_done,
107 int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108{
109 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
110 struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
111 multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
112 mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
113
114 if (bio->bi_size)
115 return 1;
116
117 if (uptodate)
118 multipath_end_bh_io(mp_bh, 0);
119 else if (!bio_rw_ahead(bio)) {
120 /*
121 * oops, IO error:
122 */
123 char b[BDEVNAME_SIZE];
124 md_error (mp_bh->mddev, rdev);
125 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
126 bdevname(rdev->bdev,b),
127 (unsigned long long)bio->bi_sector);
128 multipath_reschedule_retry(mp_bh);
129 } else
130 multipath_end_bh_io(mp_bh, error);
131 rdev_dec_pending(rdev, conf->mddev);
132 return 0;
133}
134
135static void unplug_slaves(mddev_t *mddev)
136{
137 multipath_conf_t *conf = mddev_to_conf(mddev);
138 int i;
139
140 rcu_read_lock();
141 for (i=0; i<mddev->raid_disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800142 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
144 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
145
146 atomic_inc(&rdev->nr_pending);
147 rcu_read_unlock();
148
149 if (r_queue->unplug_fn)
150 r_queue->unplug_fn(r_queue);
151
152 rdev_dec_pending(rdev, mddev);
153 rcu_read_lock();
154 }
155 }
156 rcu_read_unlock();
157}
158
159static void multipath_unplug(request_queue_t *q)
160{
161 unplug_slaves(q->queuedata);
162}
163
164
165static int multipath_make_request (request_queue_t *q, struct bio * bio)
166{
167 mddev_t *mddev = q->queuedata;
168 multipath_conf_t *conf = mddev_to_conf(mddev);
169 struct multipath_bh * mp_bh;
170 struct multipath_info *multipath;
Jens Axboea3623572005-11-01 09:26:16 +0100171 const int rw = bio_data_dir(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172
NeilBrowne5dcdd82005-09-09 16:23:41 -0700173 if (unlikely(bio_barrier(bio))) {
174 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
175 return 0;
176 }
177
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
179
180 mp_bh->master_bio = bio;
181 mp_bh->mddev = mddev;
182
Jens Axboea3623572005-11-01 09:26:16 +0100183 disk_stat_inc(mddev->gendisk, ios[rw]);
184 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185
186 mp_bh->path = multipath_map(conf);
187 if (mp_bh->path < 0) {
188 bio_endio(bio, bio->bi_size, -EIO);
189 mempool_free(mp_bh, conf->pool);
190 return 0;
191 }
192 multipath = conf->multipaths + mp_bh->path;
193
194 mp_bh->bio = *bio;
195 mp_bh->bio.bi_sector += multipath->rdev->data_offset;
196 mp_bh->bio.bi_bdev = multipath->rdev->bdev;
197 mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
198 mp_bh->bio.bi_end_io = multipath_end_request;
199 mp_bh->bio.bi_private = mp_bh;
200 generic_make_request(&mp_bh->bio);
201 return 0;
202}
203
204static void multipath_status (struct seq_file *seq, mddev_t *mddev)
205{
206 multipath_conf_t *conf = mddev_to_conf(mddev);
207 int i;
208
209 seq_printf (seq, " [%d/%d] [", conf->raid_disks,
210 conf->working_disks);
211 for (i = 0; i < conf->raid_disks; i++)
212 seq_printf (seq, "%s",
213 conf->multipaths[i].rdev &&
214 conf->multipaths[i].rdev->in_sync ? "U" : "_");
215 seq_printf (seq, "]");
216}
217
218static int multipath_issue_flush(request_queue_t *q, struct gendisk *disk,
219 sector_t *error_sector)
220{
221 mddev_t *mddev = q->queuedata;
222 multipath_conf_t *conf = mddev_to_conf(mddev);
223 int i, ret = 0;
224
225 rcu_read_lock();
226 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800227 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 if (rdev && !rdev->faulty) {
229 struct block_device *bdev = rdev->bdev;
230 request_queue_t *r_queue = bdev_get_queue(bdev);
231
232 if (!r_queue->issue_flush_fn)
233 ret = -EOPNOTSUPP;
234 else {
235 atomic_inc(&rdev->nr_pending);
236 rcu_read_unlock();
237 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
238 error_sector);
239 rdev_dec_pending(rdev, mddev);
240 rcu_read_lock();
241 }
242 }
243 }
244 rcu_read_unlock();
245 return ret;
246}
247
248/*
249 * Careful, this can execute in IRQ contexts as well!
250 */
251static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
252{
253 multipath_conf_t *conf = mddev_to_conf(mddev);
254
255 if (conf->working_disks <= 1) {
256 /*
257 * Uh oh, we can do nothing if this is our last path, but
258 * first check if this is a queued request for a device
259 * which has just failed.
260 */
261 printk(KERN_ALERT
262 "multipath: only one IO path left and IO error.\n");
263 /* leave it active... it's all we have */
264 } else {
265 /*
266 * Mark disk as unusable
267 */
268 if (!rdev->faulty) {
269 char b[BDEVNAME_SIZE];
270 rdev->in_sync = 0;
271 rdev->faulty = 1;
272 mddev->sb_dirty = 1;
273 conf->working_disks--;
274 printk(KERN_ALERT "multipath: IO failure on %s,"
275 " disabling IO path. \n Operation continuing"
276 " on %d IO paths.\n",
277 bdevname (rdev->bdev,b),
278 conf->working_disks);
279 }
280 }
281}
282
283static void print_multipath_conf (multipath_conf_t *conf)
284{
285 int i;
286 struct multipath_info *tmp;
287
288 printk("MULTIPATH conf printout:\n");
289 if (!conf) {
290 printk("(conf==NULL)\n");
291 return;
292 }
293 printk(" --- wd:%d rd:%d\n", conf->working_disks,
294 conf->raid_disks);
295
296 for (i = 0; i < conf->raid_disks; i++) {
297 char b[BDEVNAME_SIZE];
298 tmp = conf->multipaths + i;
299 if (tmp->rdev)
300 printk(" disk%d, o:%d, dev:%s\n",
301 i,!tmp->rdev->faulty,
302 bdevname(tmp->rdev->bdev,b));
303 }
304}
305
306
307static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
308{
309 multipath_conf_t *conf = mddev->private;
310 int found = 0;
311 int path;
312 struct multipath_info *p;
313
314 print_multipath_conf(conf);
315
316 for (path=0; path<mddev->raid_disks; path++)
317 if ((p=conf->multipaths+path)->rdev == NULL) {
318 blk_queue_stack_limits(mddev->queue,
319 rdev->bdev->bd_disk->queue);
320
321 /* as we don't honour merge_bvec_fn, we must never risk
322 * violating it, so limit ->max_sector to one PAGE, as
323 * a one page request is never in violation.
324 * (Note: it is very unlikely that a device with
325 * merge_bvec_fn will be involved in multipath.)
326 */
327 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
328 mddev->queue->max_sectors > (PAGE_SIZE>>9))
329 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
330
331 conf->working_disks++;
332 rdev->raid_disk = path;
333 rdev->in_sync = 1;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800334 rcu_assign_pointer(p->rdev, rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 found = 1;
336 }
337
338 print_multipath_conf(conf);
339 return found;
340}
341
342static int multipath_remove_disk(mddev_t *mddev, int number)
343{
344 multipath_conf_t *conf = mddev->private;
345 int err = 0;
346 mdk_rdev_t *rdev;
347 struct multipath_info *p = conf->multipaths + number;
348
349 print_multipath_conf(conf);
350
351 rdev = p->rdev;
352 if (rdev) {
353 if (rdev->in_sync ||
354 atomic_read(&rdev->nr_pending)) {
355 printk(KERN_ERR "hot-remove-disk, slot %d is identified" " but is still operational!\n", number);
356 err = -EBUSY;
357 goto abort;
358 }
359 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -0700360 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 if (atomic_read(&rdev->nr_pending)) {
362 /* lost the race, try later */
363 err = -EBUSY;
364 p->rdev = rdev;
365 }
366 }
367abort:
368
369 print_multipath_conf(conf);
370 return err;
371}
372
373
374
375/*
376 * This is a kernel thread which:
377 *
378 * 1. Retries failed read operations on working multipaths.
379 * 2. Updates the raid superblock when problems encounter.
380 * 3. Performs writes following reads for array syncronising.
381 */
382
383static void multipathd (mddev_t *mddev)
384{
385 struct multipath_bh *mp_bh;
386 struct bio *bio;
387 unsigned long flags;
388 multipath_conf_t *conf = mddev_to_conf(mddev);
389 struct list_head *head = &conf->retry_list;
390
391 md_check_recovery(mddev);
392 for (;;) {
393 char b[BDEVNAME_SIZE];
394 spin_lock_irqsave(&conf->device_lock, flags);
395 if (list_empty(head))
396 break;
397 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
398 list_del(head->prev);
399 spin_unlock_irqrestore(&conf->device_lock, flags);
400
401 bio = &mp_bh->bio;
402 bio->bi_sector = mp_bh->master_bio->bi_sector;
403
404 if ((mp_bh->path = multipath_map (conf))<0) {
405 printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
406 " error for block %llu\n",
407 bdevname(bio->bi_bdev,b),
408 (unsigned long long)bio->bi_sector);
409 multipath_end_bh_io(mp_bh, -EIO);
410 } else {
411 printk(KERN_ERR "multipath: %s: redirecting sector %llu"
412 " to another IO path\n",
413 bdevname(bio->bi_bdev,b),
414 (unsigned long long)bio->bi_sector);
415 *bio = *(mp_bh->master_bio);
416 bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
417 bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
418 bio->bi_rw |= (1 << BIO_RW_FAILFAST);
419 bio->bi_end_io = multipath_end_request;
420 bio->bi_private = mp_bh;
421 generic_make_request(bio);
422 }
423 }
424 spin_unlock_irqrestore(&conf->device_lock, flags);
425}
426
427static int multipath_run (mddev_t *mddev)
428{
429 multipath_conf_t *conf;
430 int disk_idx;
431 struct multipath_info *disk;
432 mdk_rdev_t *rdev;
433 struct list_head *tmp;
434
435 if (mddev->level != LEVEL_MULTIPATH) {
436 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
437 mdname(mddev), mddev->level);
438 goto out;
439 }
440 /*
441 * copy the already verified devices into our private MULTIPATH
442 * bookkeeping area. [whatever we allocate in multipath_run(),
443 * should be freed in multipath_stop()]
444 */
445
446 conf = kmalloc(sizeof(multipath_conf_t), GFP_KERNEL);
447 mddev->private = conf;
448 if (!conf) {
449 printk(KERN_ERR
450 "multipath: couldn't allocate memory for %s\n",
451 mdname(mddev));
452 goto out;
453 }
454 memset(conf, 0, sizeof(*conf));
455
456 conf->multipaths = kmalloc(sizeof(struct multipath_info)*mddev->raid_disks,
457 GFP_KERNEL);
458 if (!conf->multipaths) {
459 printk(KERN_ERR
460 "multipath: couldn't allocate memory for %s\n",
461 mdname(mddev));
462 goto out_free_conf;
463 }
464 memset(conf->multipaths, 0, sizeof(struct multipath_info)*mddev->raid_disks);
465
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 conf->working_disks = 0;
467 ITERATE_RDEV(mddev,rdev,tmp) {
468 disk_idx = rdev->raid_disk;
469 if (disk_idx < 0 ||
470 disk_idx >= mddev->raid_disks)
471 continue;
472
473 disk = conf->multipaths + disk_idx;
474 disk->rdev = rdev;
475
476 blk_queue_stack_limits(mddev->queue,
477 rdev->bdev->bd_disk->queue);
478 /* as we don't honour merge_bvec_fn, we must never risk
479 * violating it, not that we ever expect a device with
480 * a merge_bvec_fn to be involved in multipath */
481 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
482 mddev->queue->max_sectors > (PAGE_SIZE>>9))
483 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
484
485 if (!rdev->faulty)
486 conf->working_disks++;
487 }
488
489 conf->raid_disks = mddev->raid_disks;
490 mddev->sb_dirty = 1;
491 conf->mddev = mddev;
492 spin_lock_init(&conf->device_lock);
493 INIT_LIST_HEAD(&conf->retry_list);
494
495 if (!conf->working_disks) {
496 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
497 mdname(mddev));
498 goto out_free_conf;
499 }
500 mddev->degraded = conf->raid_disks = conf->working_disks;
501
502 conf->pool = mempool_create(NR_RESERVED_BUFS,
503 mp_pool_alloc, mp_pool_free,
504 NULL);
505 if (conf->pool == NULL) {
506 printk(KERN_ERR
507 "multipath: couldn't allocate memory for %s\n",
508 mdname(mddev));
509 goto out_free_conf;
510 }
511
512 {
513 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
514 if (!mddev->thread) {
515 printk(KERN_ERR "multipath: couldn't allocate thread"
516 " for %s\n", mdname(mddev));
517 goto out_free_conf;
518 }
519 }
520
521 printk(KERN_INFO
522 "multipath: array %s active with %d out of %d IO paths\n",
523 mdname(mddev), conf->working_disks, mddev->raid_disks);
524 /*
525 * Ok, everything is just fine now
526 */
527 mddev->array_size = mddev->size;
NeilBrown7a5febe2005-05-16 21:53:16 -0700528
529 mddev->queue->unplug_fn = multipath_unplug;
530 mddev->queue->issue_flush_fn = multipath_issue_flush;
531
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 return 0;
533
534out_free_conf:
535 if (conf->pool)
536 mempool_destroy(conf->pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -0700537 kfree(conf->multipaths);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 kfree(conf);
539 mddev->private = NULL;
540out:
541 return -EIO;
542}
543
544
545static int multipath_stop (mddev_t *mddev)
546{
547 multipath_conf_t *conf = mddev_to_conf(mddev);
548
549 md_unregister_thread(mddev->thread);
550 mddev->thread = NULL;
551 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
552 mempool_destroy(conf->pool);
553 kfree(conf->multipaths);
554 kfree(conf);
555 mddev->private = NULL;
556 return 0;
557}
558
559static mdk_personality_t multipath_personality=
560{
561 .name = "multipath",
562 .owner = THIS_MODULE,
563 .make_request = multipath_make_request,
564 .run = multipath_run,
565 .stop = multipath_stop,
566 .status = multipath_status,
567 .error_handler = multipath_error,
568 .hot_add_disk = multipath_add_disk,
569 .hot_remove_disk= multipath_remove_disk,
570};
571
572static int __init multipath_init (void)
573{
574 return register_md_personality (MULTIPATH, &multipath_personality);
575}
576
577static void __exit multipath_exit (void)
578{
579 unregister_md_personality (MULTIPATH);
580}
581
582module_init(multipath_init);
583module_exit(multipath_exit);
584MODULE_LICENSE("GPL");
585MODULE_ALIAS("md-personality-7"); /* MULTIPATH */