f2a63f394ad980187ba9fc319ce71b57d27f349a
[linux-2.6.git] / 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
33 #define MAX_WORK_PER_DISK 128
34
35 #define NR_RESERVED_BUFS        32
36
37
38 static int multipath_map (multipath_conf_t *conf)
39 {
40         int i, disks = conf->raid_disks;
41
42         /*
43          * Later we do read balancing on the read side 
44          * now we use the first available disk.
45          */
46
47         rcu_read_lock();
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);
52                         rcu_read_unlock();
53                         return i;
54                 }
55         }
56         rcu_read_unlock();
57
58         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
59         return (-1);
60 }
61
62 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
63 {
64         unsigned long flags;
65         mddev_t *mddev = mp_bh->mddev;
66         multipath_conf_t *conf = mddev_to_conf(mddev);
67
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);
72 }
73
74
75 /*
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
78  * cache layer.
79  */
80 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
81 {
82         struct bio *bio = mp_bh->master_bio;
83         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
84
85         bio_endio(bio, err);
86         mempool_free(mp_bh, conf->pool);
87 }
88
89 static void multipath_end_request(struct bio *bio, int error)
90 {
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;
95
96         if (uptodate)
97                 multipath_end_bh_io(mp_bh, 0);
98         else if (!bio_rw_ahead(bio)) {
99                 /*
100                  * oops, IO error:
101                  */
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);
108         } else
109                 multipath_end_bh_io(mp_bh, error);
110         rdev_dec_pending(rdev, conf->mddev);
111 }
112
113 static void unplug_slaves(mddev_t *mddev)
114 {
115         multipath_conf_t *conf = mddev_to_conf(mddev);
116         int i;
117
118         rcu_read_lock();
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);
124
125                         atomic_inc(&rdev->nr_pending);
126                         rcu_read_unlock();
127
128                         if (r_queue->unplug_fn)
129                                 r_queue->unplug_fn(r_queue);
130
131                         rdev_dec_pending(rdev, mddev);
132                         rcu_read_lock();
133                 }
134         }
135         rcu_read_unlock();
136 }
137
138 static void multipath_unplug(struct request_queue *q)
139 {
140         unplug_slaves(q->queuedata);
141 }
142
143
144 static int multipath_make_request (struct request_queue *q, struct bio * bio)
145 {
146         mddev_t *mddev = q->queuedata;
147         multipath_conf_t *conf = mddev_to_conf(mddev);
148         struct multipath_bh * mp_bh;
149         struct multipath_info *multipath;
150         const int rw = bio_data_dir(bio);
151
152         if (unlikely(bio_barrier(bio))) {
153                 bio_endio(bio, -EOPNOTSUPP);
154                 return 0;
155         }
156
157         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
158
159         mp_bh->master_bio = bio;
160         mp_bh->mddev = mddev;
161
162         disk_stat_inc(mddev->gendisk, ios[rw]);
163         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
164
165         mp_bh->path = multipath_map(conf);
166         if (mp_bh->path < 0) {
167                 bio_endio(bio, -EIO);
168                 mempool_free(mp_bh, conf->pool);
169                 return 0;
170         }
171         multipath = conf->multipaths + mp_bh->path;
172
173         mp_bh->bio = *bio;
174         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
175         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
176         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
177         mp_bh->bio.bi_end_io = multipath_end_request;
178         mp_bh->bio.bi_private = mp_bh;
179         generic_make_request(&mp_bh->bio);
180         return 0;
181 }
182
183 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
184 {
185         multipath_conf_t *conf = mddev_to_conf(mddev);
186         int i;
187         
188         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
189                                                  conf->working_disks);
190         for (i = 0; i < conf->raid_disks; i++)
191                 seq_printf (seq, "%s",
192                                conf->multipaths[i].rdev && 
193                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
194         seq_printf (seq, "]");
195 }
196
197 static int multipath_issue_flush(struct request_queue *q, struct gendisk *disk,
198                                  sector_t *error_sector)
199 {
200         mddev_t *mddev = q->queuedata;
201         multipath_conf_t *conf = mddev_to_conf(mddev);
202         int i, ret = 0;
203
204         rcu_read_lock();
205         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
206                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
207                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
208                         struct block_device *bdev = rdev->bdev;
209                         struct request_queue *r_queue = bdev_get_queue(bdev);
210
211                         if (!r_queue->issue_flush_fn)
212                                 ret = -EOPNOTSUPP;
213                         else {
214                                 atomic_inc(&rdev->nr_pending);
215                                 rcu_read_unlock();
216                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
217                                                               error_sector);
218                                 rdev_dec_pending(rdev, mddev);
219                                 rcu_read_lock();
220                         }
221                 }
222         }
223         rcu_read_unlock();
224         return ret;
225 }
226 static int multipath_congested(void *data, int bits)
227 {
228         mddev_t *mddev = data;
229         multipath_conf_t *conf = mddev_to_conf(mddev);
230         int i, ret = 0;
231
232         rcu_read_lock();
233         for (i = 0; i < mddev->raid_disks ; i++) {
234                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
235                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
236                         struct request_queue *q = bdev_get_queue(rdev->bdev);
237
238                         ret |= bdi_congested(&q->backing_dev_info, bits);
239                         /* Just like multipath_map, we just check the
240                          * first available device
241                          */
242                         break;
243                 }
244         }
245         rcu_read_unlock();
246         return ret;
247 }
248
249 /*
250  * Careful, this can execute in IRQ contexts as well!
251  */
252 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
253 {
254         multipath_conf_t *conf = mddev_to_conf(mddev);
255
256         if (conf->working_disks <= 1) {
257                 /*
258                  * Uh oh, we can do nothing if this is our last path, but
259                  * first check if this is a queued request for a device
260                  * which has just failed.
261                  */
262                 printk(KERN_ALERT 
263                         "multipath: only one IO path left and IO error.\n");
264                 /* leave it active... it's all we have */
265         } else {
266                 /*
267                  * Mark disk as unusable
268                  */
269                 if (!test_bit(Faulty, &rdev->flags)) {
270                         char b[BDEVNAME_SIZE];
271                         clear_bit(In_sync, &rdev->flags);
272                         set_bit(Faulty, &rdev->flags);
273                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
274                         conf->working_disks--;
275                         mddev->degraded++;
276                         printk(KERN_ALERT "multipath: IO failure on %s,"
277                                 " disabling IO path. \n Operation continuing"
278                                 " on %d IO paths.\n",
279                                 bdevname (rdev->bdev,b),
280                                 conf->working_disks);
281                 }
282         }
283 }
284
285 static void print_multipath_conf (multipath_conf_t *conf)
286 {
287         int i;
288         struct multipath_info *tmp;
289
290         printk("MULTIPATH conf printout:\n");
291         if (!conf) {
292                 printk("(conf==NULL)\n");
293                 return;
294         }
295         printk(" --- wd:%d rd:%d\n", conf->working_disks,
296                          conf->raid_disks);
297
298         for (i = 0; i < conf->raid_disks; i++) {
299                 char b[BDEVNAME_SIZE];
300                 tmp = conf->multipaths + i;
301                 if (tmp->rdev)
302                         printk(" disk%d, o:%d, dev:%s\n",
303                                 i,!test_bit(Faulty, &tmp->rdev->flags),
304                                bdevname(tmp->rdev->bdev,b));
305         }
306 }
307
308
309 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
310 {
311         multipath_conf_t *conf = mddev->private;
312         struct request_queue *q;
313         int found = 0;
314         int path;
315         struct multipath_info *p;
316
317         print_multipath_conf(conf);
318
319         for (path=0; path<mddev->raid_disks; path++) 
320                 if ((p=conf->multipaths+path)->rdev == NULL) {
321                         q = rdev->bdev->bd_disk->queue;
322                         blk_queue_stack_limits(mddev->queue, q);
323
324                 /* as we don't honour merge_bvec_fn, we must never risk
325                  * violating it, so limit ->max_sector to one PAGE, as
326                  * a one page request is never in violation.
327                  * (Note: it is very unlikely that a device with
328                  * merge_bvec_fn will be involved in multipath.)
329                  */
330                         if (q->merge_bvec_fn &&
331                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
332                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
333
334                         conf->working_disks++;
335                         mddev->degraded--;
336                         rdev->raid_disk = path;
337                         set_bit(In_sync, &rdev->flags);
338                         rcu_assign_pointer(p->rdev, rdev);
339                         found = 1;
340                 }
341
342         print_multipath_conf(conf);
343         return found;
344 }
345
346 static int multipath_remove_disk(mddev_t *mddev, int number)
347 {
348         multipath_conf_t *conf = mddev->private;
349         int err = 0;
350         mdk_rdev_t *rdev;
351         struct multipath_info *p = conf->multipaths + number;
352
353         print_multipath_conf(conf);
354
355         rdev = p->rdev;
356         if (rdev) {
357                 if (test_bit(In_sync, &rdev->flags) ||
358                     atomic_read(&rdev->nr_pending)) {
359                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"                                " but is still operational!\n", number);
360                         err = -EBUSY;
361                         goto abort;
362                 }
363                 p->rdev = NULL;
364                 synchronize_rcu();
365                 if (atomic_read(&rdev->nr_pending)) {
366                         /* lost the race, try later */
367                         err = -EBUSY;
368                         p->rdev = rdev;
369                 }
370         }
371 abort:
372
373         print_multipath_conf(conf);
374         return err;
375 }
376
377
378
379 /*
380  * This is a kernel thread which:
381  *
382  *      1.      Retries failed read operations on working multipaths.
383  *      2.      Updates the raid superblock when problems encounter.
384  *      3.      Performs writes following reads for array syncronising.
385  */
386
387 static void multipathd (mddev_t *mddev)
388 {
389         struct multipath_bh *mp_bh;
390         struct bio *bio;
391         unsigned long flags;
392         multipath_conf_t *conf = mddev_to_conf(mddev);
393         struct list_head *head = &conf->retry_list;
394
395         md_check_recovery(mddev);
396         for (;;) {
397                 char b[BDEVNAME_SIZE];
398                 spin_lock_irqsave(&conf->device_lock, flags);
399                 if (list_empty(head))
400                         break;
401                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
402                 list_del(head->prev);
403                 spin_unlock_irqrestore(&conf->device_lock, flags);
404
405                 bio = &mp_bh->bio;
406                 bio->bi_sector = mp_bh->master_bio->bi_sector;
407                 
408                 if ((mp_bh->path = multipath_map (conf))<0) {
409                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
410                                 " error for block %llu\n",
411                                 bdevname(bio->bi_bdev,b),
412                                 (unsigned long long)bio->bi_sector);
413                         multipath_end_bh_io(mp_bh, -EIO);
414                 } else {
415                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
416                                 " to another IO path\n",
417                                 bdevname(bio->bi_bdev,b),
418                                 (unsigned long long)bio->bi_sector);
419                         *bio = *(mp_bh->master_bio);
420                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
421                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
422                         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
423                         bio->bi_end_io = multipath_end_request;
424                         bio->bi_private = mp_bh;
425                         generic_make_request(bio);
426                 }
427         }
428         spin_unlock_irqrestore(&conf->device_lock, flags);
429 }
430
431 static int multipath_run (mddev_t *mddev)
432 {
433         multipath_conf_t *conf;
434         int disk_idx;
435         struct multipath_info *disk;
436         mdk_rdev_t *rdev;
437         struct list_head *tmp;
438
439         if (mddev->level != LEVEL_MULTIPATH) {
440                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
441                        mdname(mddev), mddev->level);
442                 goto out;
443         }
444         /*
445          * copy the already verified devices into our private MULTIPATH
446          * bookkeeping area. [whatever we allocate in multipath_run(),
447          * should be freed in multipath_stop()]
448          */
449
450         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
451         mddev->private = conf;
452         if (!conf) {
453                 printk(KERN_ERR 
454                         "multipath: couldn't allocate memory for %s\n",
455                         mdname(mddev));
456                 goto out;
457         }
458
459         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
460                                    GFP_KERNEL);
461         if (!conf->multipaths) {
462                 printk(KERN_ERR 
463                         "multipath: couldn't allocate memory for %s\n",
464                         mdname(mddev));
465                 goto out_free_conf;
466         }
467
468         conf->working_disks = 0;
469         ITERATE_RDEV(mddev,rdev,tmp) {
470                 disk_idx = rdev->raid_disk;
471                 if (disk_idx < 0 ||
472                     disk_idx >= mddev->raid_disks)
473                         continue;
474
475                 disk = conf->multipaths + disk_idx;
476                 disk->rdev = rdev;
477
478                 blk_queue_stack_limits(mddev->queue,
479                                        rdev->bdev->bd_disk->queue);
480                 /* as we don't honour merge_bvec_fn, we must never risk
481                  * violating it, not that we ever expect a device with
482                  * a merge_bvec_fn to be involved in multipath */
483                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
484                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
485                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
486
487                 if (!test_bit(Faulty, &rdev->flags))
488                         conf->working_disks++;
489         }
490
491         conf->raid_disks = mddev->raid_disks;
492         conf->mddev = mddev;
493         spin_lock_init(&conf->device_lock);
494         INIT_LIST_HEAD(&conf->retry_list);
495
496         if (!conf->working_disks) {
497                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
498                         mdname(mddev));
499                 goto out_free_conf;
500         }
501         mddev->degraded = conf->raid_disks - conf->working_disks;
502
503         conf->pool = mempool_create_kzalloc_pool(NR_RESERVED_BUFS,
504                                                  sizeof(struct multipath_bh));
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;
528
529         mddev->queue->unplug_fn = multipath_unplug;
530         mddev->queue->issue_flush_fn = multipath_issue_flush;
531         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
532         mddev->queue->backing_dev_info.congested_data = mddev;
533
534         return 0;
535
536 out_free_conf:
537         if (conf->pool)
538                 mempool_destroy(conf->pool);
539         kfree(conf->multipaths);
540         kfree(conf);
541         mddev->private = NULL;
542 out:
543         return -EIO;
544 }
545
546
547 static int multipath_stop (mddev_t *mddev)
548 {
549         multipath_conf_t *conf = mddev_to_conf(mddev);
550
551         md_unregister_thread(mddev->thread);
552         mddev->thread = NULL;
553         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
554         mempool_destroy(conf->pool);
555         kfree(conf->multipaths);
556         kfree(conf);
557         mddev->private = NULL;
558         return 0;
559 }
560
561 static struct mdk_personality multipath_personality =
562 {
563         .name           = "multipath",
564         .level          = LEVEL_MULTIPATH,
565         .owner          = THIS_MODULE,
566         .make_request   = multipath_make_request,
567         .run            = multipath_run,
568         .stop           = multipath_stop,
569         .status         = multipath_status,
570         .error_handler  = multipath_error,
571         .hot_add_disk   = multipath_add_disk,
572         .hot_remove_disk= multipath_remove_disk,
573 };
574
575 static int __init multipath_init (void)
576 {
577         return register_md_personality (&multipath_personality);
578 }
579
580 static void __exit multipath_exit (void)
581 {
582         unregister_md_personality (&multipath_personality);
583 }
584
585 module_init(multipath_init);
586 module_exit(multipath_exit);
587 MODULE_LICENSE("GPL");
588 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
589 MODULE_ALIAS("md-multipath");
590 MODULE_ALIAS("md-level--4");