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
);
151 if (unlikely(bio_barrier(bio
))) {
152 bio_endio(bio
, -EOPNOTSUPP
);
156 mp_bh
= mempool_alloc(conf
->pool
, GFP_NOIO
);
158 mp_bh
->master_bio
= bio
;
159 mp_bh
->mddev
= mddev
;
161 disk_stat_inc(mddev
->gendisk
, ios
[rw
]);
162 disk_stat_add(mddev
->gendisk
, sectors
[rw
], bio_sectors(bio
));
164 mp_bh
->path
= multipath_map(conf
);
165 if (mp_bh
->path
< 0) {
166 bio_endio(bio
, -EIO
);
167 mempool_free(mp_bh
, conf
->pool
);
170 multipath
= conf
->multipaths
+ mp_bh
->path
;
173 mp_bh
->bio
.bi_sector
+= multipath
->rdev
->data_offset
;
174 mp_bh
->bio
.bi_bdev
= multipath
->rdev
->bdev
;
175 mp_bh
->bio
.bi_rw
|= (1 << BIO_RW_FAILFAST
);
176 mp_bh
->bio
.bi_end_io
= multipath_end_request
;
177 mp_bh
->bio
.bi_private
= mp_bh
;
178 generic_make_request(&mp_bh
->bio
);
182 static void multipath_status (struct seq_file
*seq
, mddev_t
*mddev
)
184 multipath_conf_t
*conf
= mddev_to_conf(mddev
);
187 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
,
188 conf
->working_disks
);
189 for (i
= 0; i
< conf
->raid_disks
; i
++)
190 seq_printf (seq
, "%s",
191 conf
->multipaths
[i
].rdev
&&
192 test_bit(In_sync
, &conf
->multipaths
[i
].rdev
->flags
) ? "U" : "_");
193 seq_printf (seq
, "]");
196 static int multipath_congested(void *data
, int bits
)
198 mddev_t
*mddev
= data
;
199 multipath_conf_t
*conf
= mddev_to_conf(mddev
);
203 for (i
= 0; i
< mddev
->raid_disks
; i
++) {
204 mdk_rdev_t
*rdev
= rcu_dereference(conf
->multipaths
[i
].rdev
);
205 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
206 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
208 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
209 /* Just like multipath_map, we just check the
210 * first available device
220 * Careful, this can execute in IRQ contexts as well!
222 static void multipath_error (mddev_t
*mddev
, mdk_rdev_t
*rdev
)
224 multipath_conf_t
*conf
= mddev_to_conf(mddev
);
226 if (conf
->working_disks
<= 1) {
228 * Uh oh, we can do nothing if this is our last path, but
229 * first check if this is a queued request for a device
230 * which has just failed.
233 "multipath: only one IO path left and IO error.\n");
234 /* leave it active... it's all we have */
237 * Mark disk as unusable
239 if (!test_bit(Faulty
, &rdev
->flags
)) {
240 char b
[BDEVNAME_SIZE
];
241 clear_bit(In_sync
, &rdev
->flags
);
242 set_bit(Faulty
, &rdev
->flags
);
243 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
244 conf
->working_disks
--;
246 printk(KERN_ALERT
"multipath: IO failure on %s,"
247 " disabling IO path.\n"
248 "multipath: Operation continuing"
249 " on %d IO paths.\n",
250 bdevname (rdev
->bdev
,b
),
251 conf
->working_disks
);
256 static void print_multipath_conf (multipath_conf_t
*conf
)
259 struct multipath_info
*tmp
;
261 printk("MULTIPATH conf printout:\n");
263 printk("(conf==NULL)\n");
266 printk(" --- wd:%d rd:%d\n", conf
->working_disks
,
269 for (i
= 0; i
< conf
->raid_disks
; i
++) {
270 char b
[BDEVNAME_SIZE
];
271 tmp
= conf
->multipaths
+ i
;
273 printk(" disk%d, o:%d, dev:%s\n",
274 i
,!test_bit(Faulty
, &tmp
->rdev
->flags
),
275 bdevname(tmp
->rdev
->bdev
,b
));
280 static int multipath_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
282 multipath_conf_t
*conf
= mddev
->private;
283 struct request_queue
*q
;
286 struct multipath_info
*p
;
288 print_multipath_conf(conf
);
290 for (path
=0; path
<mddev
->raid_disks
; path
++)
291 if ((p
=conf
->multipaths
+path
)->rdev
== NULL
) {
292 q
= rdev
->bdev
->bd_disk
->queue
;
293 blk_queue_stack_limits(mddev
->queue
, q
);
295 /* as we don't honour merge_bvec_fn, we must never risk
296 * violating it, so limit ->max_sector to one PAGE, as
297 * a one page request is never in violation.
298 * (Note: it is very unlikely that a device with
299 * merge_bvec_fn will be involved in multipath.)
301 if (q
->merge_bvec_fn
&&
302 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
303 blk_queue_max_sectors(mddev
->queue
, PAGE_SIZE
>>9);
305 conf
->working_disks
++;
307 rdev
->raid_disk
= path
;
308 set_bit(In_sync
, &rdev
->flags
);
309 rcu_assign_pointer(p
->rdev
, rdev
);
313 print_multipath_conf(conf
);
317 static int multipath_remove_disk(mddev_t
*mddev
, int number
)
319 multipath_conf_t
*conf
= mddev
->private;
322 struct multipath_info
*p
= conf
->multipaths
+ number
;
324 print_multipath_conf(conf
);
328 if (test_bit(In_sync
, &rdev
->flags
) ||
329 atomic_read(&rdev
->nr_pending
)) {
330 printk(KERN_ERR
"hot-remove-disk, slot %d is identified" " but is still operational!\n", number
);
336 if (atomic_read(&rdev
->nr_pending
)) {
337 /* lost the race, try later */
344 print_multipath_conf(conf
);
351 * This is a kernel thread which:
353 * 1. Retries failed read operations on working multipaths.
354 * 2. Updates the raid superblock when problems encounter.
355 * 3. Performs writes following reads for array syncronising.
358 static void multipathd (mddev_t
*mddev
)
360 struct multipath_bh
*mp_bh
;
363 multipath_conf_t
*conf
= mddev_to_conf(mddev
);
364 struct list_head
*head
= &conf
->retry_list
;
366 md_check_recovery(mddev
);
368 char b
[BDEVNAME_SIZE
];
369 spin_lock_irqsave(&conf
->device_lock
, flags
);
370 if (list_empty(head
))
372 mp_bh
= list_entry(head
->prev
, struct multipath_bh
, retry_list
);
373 list_del(head
->prev
);
374 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
377 bio
->bi_sector
= mp_bh
->master_bio
->bi_sector
;
379 if ((mp_bh
->path
= multipath_map (conf
))<0) {
380 printk(KERN_ALERT
"multipath: %s: unrecoverable IO read"
381 " error for block %llu\n",
382 bdevname(bio
->bi_bdev
,b
),
383 (unsigned long long)bio
->bi_sector
);
384 multipath_end_bh_io(mp_bh
, -EIO
);
386 printk(KERN_ERR
"multipath: %s: redirecting sector %llu"
387 " to another IO path\n",
388 bdevname(bio
->bi_bdev
,b
),
389 (unsigned long long)bio
->bi_sector
);
390 *bio
= *(mp_bh
->master_bio
);
391 bio
->bi_sector
+= conf
->multipaths
[mp_bh
->path
].rdev
->data_offset
;
392 bio
->bi_bdev
= conf
->multipaths
[mp_bh
->path
].rdev
->bdev
;
393 bio
->bi_rw
|= (1 << BIO_RW_FAILFAST
);
394 bio
->bi_end_io
= multipath_end_request
;
395 bio
->bi_private
= mp_bh
;
396 generic_make_request(bio
);
399 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
402 static int multipath_run (mddev_t
*mddev
)
404 multipath_conf_t
*conf
;
406 struct multipath_info
*disk
;
408 struct list_head
*tmp
;
410 if (mddev
->level
!= LEVEL_MULTIPATH
) {
411 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
412 mdname(mddev
), mddev
->level
);
416 * copy the already verified devices into our private MULTIPATH
417 * bookkeeping area. [whatever we allocate in multipath_run(),
418 * should be freed in multipath_stop()]
420 mddev
->queue
->queue_lock
= &mddev
->queue
->__queue_lock
;
422 conf
= kzalloc(sizeof(multipath_conf_t
), GFP_KERNEL
);
423 mddev
->private = conf
;
426 "multipath: couldn't allocate memory for %s\n",
431 conf
->multipaths
= kzalloc(sizeof(struct multipath_info
)*mddev
->raid_disks
,
433 if (!conf
->multipaths
) {
435 "multipath: couldn't allocate memory for %s\n",
440 conf
->working_disks
= 0;
441 rdev_for_each(rdev
, tmp
, mddev
) {
442 disk_idx
= rdev
->raid_disk
;
444 disk_idx
>= mddev
->raid_disks
)
447 disk
= conf
->multipaths
+ disk_idx
;
450 blk_queue_stack_limits(mddev
->queue
,
451 rdev
->bdev
->bd_disk
->queue
);
452 /* as we don't honour merge_bvec_fn, we must never risk
453 * violating it, not that we ever expect a device with
454 * a merge_bvec_fn to be involved in multipath */
455 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
456 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
457 blk_queue_max_sectors(mddev
->queue
, PAGE_SIZE
>>9);
459 if (!test_bit(Faulty
, &rdev
->flags
))
460 conf
->working_disks
++;
463 conf
->raid_disks
= mddev
->raid_disks
;
465 spin_lock_init(&conf
->device_lock
);
466 INIT_LIST_HEAD(&conf
->retry_list
);
468 if (!conf
->working_disks
) {
469 printk(KERN_ERR
"multipath: no operational IO paths for %s\n",
473 mddev
->degraded
= conf
->raid_disks
- conf
->working_disks
;
475 conf
->pool
= mempool_create_kzalloc_pool(NR_RESERVED_BUFS
,
476 sizeof(struct multipath_bh
));
477 if (conf
->pool
== NULL
) {
479 "multipath: couldn't allocate memory for %s\n",
485 mddev
->thread
= md_register_thread(multipathd
, mddev
, "%s_multipath");
486 if (!mddev
->thread
) {
487 printk(KERN_ERR
"multipath: couldn't allocate thread"
488 " for %s\n", mdname(mddev
));
494 "multipath: array %s active with %d out of %d IO paths\n",
495 mdname(mddev
), conf
->working_disks
, mddev
->raid_disks
);
497 * Ok, everything is just fine now
499 mddev
->array_size
= mddev
->size
;
501 mddev
->queue
->unplug_fn
= multipath_unplug
;
502 mddev
->queue
->backing_dev_info
.congested_fn
= multipath_congested
;
503 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
509 mempool_destroy(conf
->pool
);
510 kfree(conf
->multipaths
);
512 mddev
->private = NULL
;
518 static int multipath_stop (mddev_t
*mddev
)
520 multipath_conf_t
*conf
= mddev_to_conf(mddev
);
522 md_unregister_thread(mddev
->thread
);
523 mddev
->thread
= NULL
;
524 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
525 mempool_destroy(conf
->pool
);
526 kfree(conf
->multipaths
);
528 mddev
->private = NULL
;
532 static struct mdk_personality multipath_personality
=
535 .level
= LEVEL_MULTIPATH
,
536 .owner
= THIS_MODULE
,
537 .make_request
= multipath_make_request
,
538 .run
= multipath_run
,
539 .stop
= multipath_stop
,
540 .status
= multipath_status
,
541 .error_handler
= multipath_error
,
542 .hot_add_disk
= multipath_add_disk
,
543 .hot_remove_disk
= multipath_remove_disk
,
546 static int __init
multipath_init (void)
548 return register_md_personality (&multipath_personality
);
551 static void __exit
multipath_exit (void)
553 unregister_md_personality (&multipath_personality
);
556 module_init(multipath_init
);
557 module_exit(multipath_exit
);
558 MODULE_LICENSE("GPL");
559 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
560 MODULE_ALIAS("md-multipath");
561 MODULE_ALIAS("md-level--4");