6d7ddf32ef2ec932040d0611cbf109b82a7a656c
[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/blkdev.h>
23 #include <linux/raid/md_u.h>
24 #include <linux/seq_file.h>
25 #include <linux/slab.h>
26 #include "md.h"
27 #include "multipath.h"
28
29 #define MAX_WORK_PER_DISK 128
30
31 #define NR_RESERVED_BUFS        32
32
33
34 static int multipath_map (multipath_conf_t *conf)
35 {
36         int i, disks = conf->raid_disks;
37
38         /*
39          * Later we do read balancing on the read side 
40          * now we use the first available disk.
41          */
42
43         rcu_read_lock();
44         for (i = 0; i < disks; i++) {
45                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
46                 if (rdev && test_bit(In_sync, &rdev->flags)) {
47                         atomic_inc(&rdev->nr_pending);
48                         rcu_read_unlock();
49                         return i;
50                 }
51         }
52         rcu_read_unlock();
53
54         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
55         return (-1);
56 }
57
58 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
59 {
60         unsigned long flags;
61         mddev_t *mddev = mp_bh->mddev;
62         multipath_conf_t *conf = mddev->private;
63
64         spin_lock_irqsave(&conf->device_lock, flags);
65         list_add(&mp_bh->retry_list, &conf->retry_list);
66         spin_unlock_irqrestore(&conf->device_lock, flags);
67         md_wakeup_thread(mddev->thread);
68 }
69
70
71 /*
72  * multipath_end_bh_io() is called when we have finished servicing a multipathed
73  * operation and are ready to return a success/failure code to the buffer
74  * cache layer.
75  */
76 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
77 {
78         struct bio *bio = mp_bh->master_bio;
79         multipath_conf_t *conf = mp_bh->mddev->private;
80
81         bio_endio(bio, err);
82         mempool_free(mp_bh, conf->pool);
83 }
84
85 static void multipath_end_request(struct bio *bio, int error)
86 {
87         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
88         struct multipath_bh *mp_bh = bio->bi_private;
89         multipath_conf_t *conf = mp_bh->mddev->private;
90         mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
91
92         if (uptodate)
93                 multipath_end_bh_io(mp_bh, 0);
94         else if (!(bio->bi_rw & REQ_RAHEAD)) {
95                 /*
96                  * oops, IO error:
97                  */
98                 char b[BDEVNAME_SIZE];
99                 md_error (mp_bh->mddev, rdev);
100                 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n", 
101                        bdevname(rdev->bdev,b), 
102                        (unsigned long long)bio->bi_sector);
103                 multipath_reschedule_retry(mp_bh);
104         } else
105                 multipath_end_bh_io(mp_bh, error);
106         rdev_dec_pending(rdev, conf->mddev);
107 }
108
109 static void unplug_slaves(mddev_t *mddev)
110 {
111         multipath_conf_t *conf = mddev->private;
112         int i;
113
114         rcu_read_lock();
115         for (i=0; i<mddev->raid_disks; i++) {
116                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
117                 if (rdev && !test_bit(Faulty, &rdev->flags)
118                     && atomic_read(&rdev->nr_pending)) {
119                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
120
121                         atomic_inc(&rdev->nr_pending);
122                         rcu_read_unlock();
123
124                         blk_unplug(r_queue);
125
126                         rdev_dec_pending(rdev, mddev);
127                         rcu_read_lock();
128                 }
129         }
130         rcu_read_unlock();
131 }
132
133 static void multipath_unplug(struct request_queue *q)
134 {
135         unplug_slaves(q->queuedata);
136 }
137
138
139 static int multipath_make_request(mddev_t *mddev, struct bio * bio)
140 {
141         multipath_conf_t *conf = mddev->private;
142         struct multipath_bh * mp_bh;
143         struct multipath_info *multipath;
144
145         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
146                 md_flush_request(mddev, bio);
147                 return 0;
148         }
149
150         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
151
152         mp_bh->master_bio = bio;
153         mp_bh->mddev = mddev;
154
155         mp_bh->path = multipath_map(conf);
156         if (mp_bh->path < 0) {
157                 bio_endio(bio, -EIO);
158                 mempool_free(mp_bh, conf->pool);
159                 return 0;
160         }
161         multipath = conf->multipaths + mp_bh->path;
162
163         mp_bh->bio = *bio;
164         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
165         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
166         mp_bh->bio.bi_rw |= REQ_FAILFAST_TRANSPORT;
167         mp_bh->bio.bi_end_io = multipath_end_request;
168         mp_bh->bio.bi_private = mp_bh;
169         generic_make_request(&mp_bh->bio);
170         return 0;
171 }
172
173 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
174 {
175         multipath_conf_t *conf = mddev->private;
176         int i;
177         
178         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
179                                                  conf->working_disks);
180         for (i = 0; i < conf->raid_disks; i++)
181                 seq_printf (seq, "%s",
182                                conf->multipaths[i].rdev && 
183                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
184         seq_printf (seq, "]");
185 }
186
187 static int multipath_congested(void *data, int bits)
188 {
189         mddev_t *mddev = data;
190         multipath_conf_t *conf = mddev->private;
191         int i, ret = 0;
192
193         if (mddev_congested(mddev, bits))
194                 return 1;
195
196         rcu_read_lock();
197         for (i = 0; i < mddev->raid_disks ; i++) {
198                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
199                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
200                         struct request_queue *q = bdev_get_queue(rdev->bdev);
201
202                         ret |= bdi_congested(&q->backing_dev_info, bits);
203                         /* Just like multipath_map, we just check the
204                          * first available device
205                          */
206                         break;
207                 }
208         }
209         rcu_read_unlock();
210         return ret;
211 }
212
213 /*
214  * Careful, this can execute in IRQ contexts as well!
215  */
216 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
217 {
218         multipath_conf_t *conf = mddev->private;
219
220         if (conf->working_disks <= 1) {
221                 /*
222                  * Uh oh, we can do nothing if this is our last path, but
223                  * first check if this is a queued request for a device
224                  * which has just failed.
225                  */
226                 printk(KERN_ALERT 
227                         "multipath: only one IO path left and IO error.\n");
228                 /* leave it active... it's all we have */
229         } else {
230                 /*
231                  * Mark disk as unusable
232                  */
233                 if (!test_bit(Faulty, &rdev->flags)) {
234                         char b[BDEVNAME_SIZE];
235                         clear_bit(In_sync, &rdev->flags);
236                         set_bit(Faulty, &rdev->flags);
237                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
238                         conf->working_disks--;
239                         mddev->degraded++;
240                         printk(KERN_ALERT "multipath: IO failure on %s,"
241                                 " disabling IO path.\n"
242                                 "multipath: Operation continuing"
243                                 " on %d IO paths.\n",
244                                 bdevname (rdev->bdev,b),
245                                 conf->working_disks);
246                 }
247         }
248 }
249
250 static void print_multipath_conf (multipath_conf_t *conf)
251 {
252         int i;
253         struct multipath_info *tmp;
254
255         printk("MULTIPATH conf printout:\n");
256         if (!conf) {
257                 printk("(conf==NULL)\n");
258                 return;
259         }
260         printk(" --- wd:%d rd:%d\n", conf->working_disks,
261                          conf->raid_disks);
262
263         for (i = 0; i < conf->raid_disks; i++) {
264                 char b[BDEVNAME_SIZE];
265                 tmp = conf->multipaths + i;
266                 if (tmp->rdev)
267                         printk(" disk%d, o:%d, dev:%s\n",
268                                 i,!test_bit(Faulty, &tmp->rdev->flags),
269                                bdevname(tmp->rdev->bdev,b));
270         }
271 }
272
273
274 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
275 {
276         multipath_conf_t *conf = mddev->private;
277         struct request_queue *q;
278         int err = -EEXIST;
279         int path;
280         struct multipath_info *p;
281         int first = 0;
282         int last = mddev->raid_disks - 1;
283
284         if (rdev->raid_disk >= 0)
285                 first = last = rdev->raid_disk;
286
287         print_multipath_conf(conf);
288
289         for (path = first; path <= last; path++)
290                 if ((p=conf->multipaths+path)->rdev == NULL) {
291                         q = rdev->bdev->bd_disk->queue;
292                         disk_stack_limits(mddev->gendisk, rdev->bdev,
293                                           rdev->data_offset << 9);
294
295                 /* as we don't honour merge_bvec_fn, we must never risk
296                  * violating it, so limit ->max_segments to one, lying
297                  * within a single page.
298                  * (Note: it is very unlikely that a device with
299                  * merge_bvec_fn will be involved in multipath.)
300                  */
301                         if (q->merge_bvec_fn) {
302                                 blk_queue_max_segments(mddev->queue, 1);
303                                 blk_queue_segment_boundary(mddev->queue,
304                                                            PAGE_CACHE_SIZE - 1);
305                         }
306
307                         conf->working_disks++;
308                         mddev->degraded--;
309                         rdev->raid_disk = path;
310                         set_bit(In_sync, &rdev->flags);
311                         rcu_assign_pointer(p->rdev, rdev);
312                         err = 0;
313                         md_integrity_add_rdev(rdev, mddev);
314                         break;
315                 }
316
317         print_multipath_conf(conf);
318
319         return err;
320 }
321
322 static int multipath_remove_disk(mddev_t *mddev, int number)
323 {
324         multipath_conf_t *conf = mddev->private;
325         int err = 0;
326         mdk_rdev_t *rdev;
327         struct multipath_info *p = conf->multipaths + number;
328
329         print_multipath_conf(conf);
330
331         rdev = p->rdev;
332         if (rdev) {
333                 if (test_bit(In_sync, &rdev->flags) ||
334                     atomic_read(&rdev->nr_pending)) {
335                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"
336                                " but is still operational!\n", number);
337                         err = -EBUSY;
338                         goto abort;
339                 }
340                 p->rdev = NULL;
341                 synchronize_rcu();
342                 if (atomic_read(&rdev->nr_pending)) {
343                         /* lost the race, try later */
344                         err = -EBUSY;
345                         p->rdev = rdev;
346                         goto abort;
347                 }
348                 md_integrity_register(mddev);
349         }
350 abort:
351
352         print_multipath_conf(conf);
353         return err;
354 }
355
356
357
358 /*
359  * This is a kernel thread which:
360  *
361  *      1.      Retries failed read operations on working multipaths.
362  *      2.      Updates the raid superblock when problems encounter.
363  *      3.      Performs writes following reads for array syncronising.
364  */
365
366 static void multipathd (mddev_t *mddev)
367 {
368         struct multipath_bh *mp_bh;
369         struct bio *bio;
370         unsigned long flags;
371         multipath_conf_t *conf = mddev->private;
372         struct list_head *head = &conf->retry_list;
373
374         md_check_recovery(mddev);
375         for (;;) {
376                 char b[BDEVNAME_SIZE];
377                 spin_lock_irqsave(&conf->device_lock, flags);
378                 if (list_empty(head))
379                         break;
380                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
381                 list_del(head->prev);
382                 spin_unlock_irqrestore(&conf->device_lock, flags);
383
384                 bio = &mp_bh->bio;
385                 bio->bi_sector = mp_bh->master_bio->bi_sector;
386                 
387                 if ((mp_bh->path = multipath_map (conf))<0) {
388                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
389                                 " error for block %llu\n",
390                                 bdevname(bio->bi_bdev,b),
391                                 (unsigned long long)bio->bi_sector);
392                         multipath_end_bh_io(mp_bh, -EIO);
393                 } else {
394                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
395                                 " to another IO path\n",
396                                 bdevname(bio->bi_bdev,b),
397                                 (unsigned long long)bio->bi_sector);
398                         *bio = *(mp_bh->master_bio);
399                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
400                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
401                         bio->bi_rw |= REQ_FAILFAST_TRANSPORT;
402                         bio->bi_end_io = multipath_end_request;
403                         bio->bi_private = mp_bh;
404                         generic_make_request(bio);
405                 }
406         }
407         spin_unlock_irqrestore(&conf->device_lock, flags);
408 }
409
410 static sector_t multipath_size(mddev_t *mddev, sector_t sectors, int raid_disks)
411 {
412         WARN_ONCE(sectors || raid_disks,
413                   "%s does not support generic reshape\n", __func__);
414
415         return mddev->dev_sectors;
416 }
417
418 static int multipath_run (mddev_t *mddev)
419 {
420         multipath_conf_t *conf;
421         int disk_idx;
422         struct multipath_info *disk;
423         mdk_rdev_t *rdev;
424
425         if (md_check_no_bitmap(mddev))
426                 return -EINVAL;
427
428         if (mddev->level != LEVEL_MULTIPATH) {
429                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
430                        mdname(mddev), mddev->level);
431                 goto out;
432         }
433         /*
434          * copy the already verified devices into our private MULTIPATH
435          * bookkeeping area. [whatever we allocate in multipath_run(),
436          * should be freed in multipath_stop()]
437          */
438         mddev->queue->queue_lock = &mddev->queue->__queue_lock;
439
440         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
441         mddev->private = conf;
442         if (!conf) {
443                 printk(KERN_ERR 
444                         "multipath: couldn't allocate memory for %s\n",
445                         mdname(mddev));
446                 goto out;
447         }
448
449         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
450                                    GFP_KERNEL);
451         if (!conf->multipaths) {
452                 printk(KERN_ERR 
453                         "multipath: couldn't allocate memory for %s\n",
454                         mdname(mddev));
455                 goto out_free_conf;
456         }
457
458         conf->working_disks = 0;
459         list_for_each_entry(rdev, &mddev->disks, same_set) {
460                 disk_idx = rdev->raid_disk;
461                 if (disk_idx < 0 ||
462                     disk_idx >= mddev->raid_disks)
463                         continue;
464
465                 disk = conf->multipaths + disk_idx;
466                 disk->rdev = rdev;
467                 disk_stack_limits(mddev->gendisk, rdev->bdev,
468                                   rdev->data_offset << 9);
469
470                 /* as we don't honour merge_bvec_fn, we must never risk
471                  * violating it, not that we ever expect a device with
472                  * a merge_bvec_fn to be involved in multipath */
473                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
474                         blk_queue_max_segments(mddev->queue, 1);
475                         blk_queue_segment_boundary(mddev->queue,
476                                                    PAGE_CACHE_SIZE - 1);
477                 }
478
479                 if (!test_bit(Faulty, &rdev->flags))
480                         conf->working_disks++;
481         }
482
483         conf->raid_disks = mddev->raid_disks;
484         conf->mddev = mddev;
485         spin_lock_init(&conf->device_lock);
486         INIT_LIST_HEAD(&conf->retry_list);
487
488         if (!conf->working_disks) {
489                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
490                         mdname(mddev));
491                 goto out_free_conf;
492         }
493         mddev->degraded = conf->raid_disks - conf->working_disks;
494
495         conf->pool = mempool_create_kmalloc_pool(NR_RESERVED_BUFS,
496                                                  sizeof(struct multipath_bh));
497         if (conf->pool == NULL) {
498                 printk(KERN_ERR 
499                         "multipath: couldn't allocate memory for %s\n",
500                         mdname(mddev));
501                 goto out_free_conf;
502         }
503
504         {
505                 mddev->thread = md_register_thread(multipathd, mddev, NULL);
506                 if (!mddev->thread) {
507                         printk(KERN_ERR "multipath: couldn't allocate thread"
508                                 " for %s\n", mdname(mddev));
509                         goto out_free_conf;
510                 }
511         }
512
513         printk(KERN_INFO 
514                 "multipath: array %s active with %d out of %d IO paths\n",
515                 mdname(mddev), conf->working_disks, mddev->raid_disks);
516         /*
517          * Ok, everything is just fine now
518          */
519         md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
520
521         mddev->queue->unplug_fn = multipath_unplug;
522         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
523         mddev->queue->backing_dev_info.congested_data = mddev;
524         md_integrity_register(mddev);
525         return 0;
526
527 out_free_conf:
528         if (conf->pool)
529                 mempool_destroy(conf->pool);
530         kfree(conf->multipaths);
531         kfree(conf);
532         mddev->private = NULL;
533 out:
534         return -EIO;
535 }
536
537
538 static int multipath_stop (mddev_t *mddev)
539 {
540         multipath_conf_t *conf = mddev->private;
541
542         md_unregister_thread(mddev->thread);
543         mddev->thread = NULL;
544         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
545         mempool_destroy(conf->pool);
546         kfree(conf->multipaths);
547         kfree(conf);
548         mddev->private = NULL;
549         return 0;
550 }
551
552 static struct mdk_personality multipath_personality =
553 {
554         .name           = "multipath",
555         .level          = LEVEL_MULTIPATH,
556         .owner          = THIS_MODULE,
557         .make_request   = multipath_make_request,
558         .run            = multipath_run,
559         .stop           = multipath_stop,
560         .status         = multipath_status,
561         .error_handler  = multipath_error,
562         .hot_add_disk   = multipath_add_disk,
563         .hot_remove_disk= multipath_remove_disk,
564         .size           = multipath_size,
565 };
566
567 static int __init multipath_init (void)
568 {
569         return register_md_personality (&multipath_personality);
570 }
571
572 static void __exit multipath_exit (void)
573 {
574         unregister_md_personality (&multipath_personality);
575 }
576
577 module_init(multipath_init);
578 module_exit(multipath_exit);
579 MODULE_LICENSE("GPL");
580 MODULE_DESCRIPTION("simple multi-path personality for MD");
581 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
582 MODULE_ALIAS("md-multipath");
583 MODULE_ALIAS("md-level--4");