182f5a94cdc54cf1886ea7ae2347e3020a66541e
[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                         blk_unplug(r_queue);
129
130                         rdev_dec_pending(rdev, mddev);
131                         rcu_read_lock();
132                 }
133         }
134         rcu_read_unlock();
135 }
136
137 static void multipath_unplug(struct request_queue *q)
138 {
139         unplug_slaves(q->queuedata);
140 }
141
142
143 static int multipath_make_request (struct request_queue *q, struct bio * bio)
144 {
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);
150         int cpu;
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         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));
165         disk_stat_unlock();
166
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);
171                 return 0;
172         }
173         multipath = conf->multipaths + mp_bh->path;
174
175         mp_bh->bio = *bio;
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);
182         return 0;
183 }
184
185 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
186 {
187         multipath_conf_t *conf = mddev_to_conf(mddev);
188         int i;
189         
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, "]");
197 }
198
199 static int multipath_congested(void *data, int bits)
200 {
201         mddev_t *mddev = data;
202         multipath_conf_t *conf = mddev_to_conf(mddev);
203         int i, ret = 0;
204
205         rcu_read_lock();
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);
210
211                         ret |= bdi_congested(&q->backing_dev_info, bits);
212                         /* Just like multipath_map, we just check the
213                          * first available device
214                          */
215                         break;
216                 }
217         }
218         rcu_read_unlock();
219         return ret;
220 }
221
222 /*
223  * Careful, this can execute in IRQ contexts as well!
224  */
225 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
226 {
227         multipath_conf_t *conf = mddev_to_conf(mddev);
228
229         if (conf->working_disks <= 1) {
230                 /*
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.
234                  */
235                 printk(KERN_ALERT 
236                         "multipath: only one IO path left and IO error.\n");
237                 /* leave it active... it's all we have */
238         } else {
239                 /*
240                  * Mark disk as unusable
241                  */
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--;
248                         mddev->degraded++;
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);
255                 }
256         }
257 }
258
259 static void print_multipath_conf (multipath_conf_t *conf)
260 {
261         int i;
262         struct multipath_info *tmp;
263
264         printk("MULTIPATH conf printout:\n");
265         if (!conf) {
266                 printk("(conf==NULL)\n");
267                 return;
268         }
269         printk(" --- wd:%d rd:%d\n", conf->working_disks,
270                          conf->raid_disks);
271
272         for (i = 0; i < conf->raid_disks; i++) {
273                 char b[BDEVNAME_SIZE];
274                 tmp = conf->multipaths + i;
275                 if (tmp->rdev)
276                         printk(" disk%d, o:%d, dev:%s\n",
277                                 i,!test_bit(Faulty, &tmp->rdev->flags),
278                                bdevname(tmp->rdev->bdev,b));
279         }
280 }
281
282
283 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
284 {
285         multipath_conf_t *conf = mddev->private;
286         struct request_queue *q;
287         int err = -EEXIST;
288         int path;
289         struct multipath_info *p;
290         int first = 0;
291         int last = mddev->raid_disks - 1;
292
293         if (rdev->raid_disk >= 0)
294                 first = last = rdev->raid_disk;
295
296         print_multipath_conf(conf);
297
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);
302
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.)
308                  */
309                         if (q->merge_bvec_fn &&
310                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
311                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
312
313                         conf->working_disks++;
314                         mddev->degraded--;
315                         rdev->raid_disk = path;
316                         set_bit(In_sync, &rdev->flags);
317                         rcu_assign_pointer(p->rdev, rdev);
318                         err = 0;
319                         break;
320                 }
321
322         print_multipath_conf(conf);
323
324         return err;
325 }
326
327 static int multipath_remove_disk(mddev_t *mddev, int number)
328 {
329         multipath_conf_t *conf = mddev->private;
330         int err = 0;
331         mdk_rdev_t *rdev;
332         struct multipath_info *p = conf->multipaths + number;
333
334         print_multipath_conf(conf);
335
336         rdev = p->rdev;
337         if (rdev) {
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);
342                         err = -EBUSY;
343                         goto abort;
344                 }
345                 p->rdev = NULL;
346                 synchronize_rcu();
347                 if (atomic_read(&rdev->nr_pending)) {
348                         /* lost the race, try later */
349                         err = -EBUSY;
350                         p->rdev = rdev;
351                 }
352         }
353 abort:
354
355         print_multipath_conf(conf);
356         return err;
357 }
358
359
360
361 /*
362  * This is a kernel thread which:
363  *
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.
367  */
368
369 static void multipathd (mddev_t *mddev)
370 {
371         struct multipath_bh *mp_bh;
372         struct bio *bio;
373         unsigned long flags;
374         multipath_conf_t *conf = mddev_to_conf(mddev);
375         struct list_head *head = &conf->retry_list;
376
377         md_check_recovery(mddev);
378         for (;;) {
379                 char b[BDEVNAME_SIZE];
380                 spin_lock_irqsave(&conf->device_lock, flags);
381                 if (list_empty(head))
382                         break;
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);
386
387                 bio = &mp_bh->bio;
388                 bio->bi_sector = mp_bh->master_bio->bi_sector;
389                 
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);
396                 } else {
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);
408                 }
409         }
410         spin_unlock_irqrestore(&conf->device_lock, flags);
411 }
412
413 static int multipath_run (mddev_t *mddev)
414 {
415         multipath_conf_t *conf;
416         int disk_idx;
417         struct multipath_info *disk;
418         mdk_rdev_t *rdev;
419         struct list_head *tmp;
420
421         if (mddev->level != LEVEL_MULTIPATH) {
422                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
423                        mdname(mddev), mddev->level);
424                 goto out;
425         }
426         /*
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()]
430          */
431         mddev->queue->queue_lock = &mddev->queue->__queue_lock;
432
433         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
434         mddev->private = conf;
435         if (!conf) {
436                 printk(KERN_ERR 
437                         "multipath: couldn't allocate memory for %s\n",
438                         mdname(mddev));
439                 goto out;
440         }
441
442         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
443                                    GFP_KERNEL);
444         if (!conf->multipaths) {
445                 printk(KERN_ERR 
446                         "multipath: couldn't allocate memory for %s\n",
447                         mdname(mddev));
448                 goto out_free_conf;
449         }
450
451         conf->working_disks = 0;
452         rdev_for_each(rdev, tmp, mddev) {
453                 disk_idx = rdev->raid_disk;
454                 if (disk_idx < 0 ||
455                     disk_idx >= mddev->raid_disks)
456                         continue;
457
458                 disk = conf->multipaths + disk_idx;
459                 disk->rdev = rdev;
460
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);
469
470                 if (!test_bit(Faulty, &rdev->flags))
471                         conf->working_disks++;
472         }
473
474         conf->raid_disks = mddev->raid_disks;
475         conf->mddev = mddev;
476         spin_lock_init(&conf->device_lock);
477         INIT_LIST_HEAD(&conf->retry_list);
478
479         if (!conf->working_disks) {
480                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
481                         mdname(mddev));
482                 goto out_free_conf;
483         }
484         mddev->degraded = conf->raid_disks - conf->working_disks;
485
486         conf->pool = mempool_create_kzalloc_pool(NR_RESERVED_BUFS,
487                                                  sizeof(struct multipath_bh));
488         if (conf->pool == NULL) {
489                 printk(KERN_ERR 
490                         "multipath: couldn't allocate memory for %s\n",
491                         mdname(mddev));
492                 goto out_free_conf;
493         }
494
495         {
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));
500                         goto out_free_conf;
501                 }
502         }
503
504         printk(KERN_INFO 
505                 "multipath: array %s active with %d out of %d IO paths\n",
506                 mdname(mddev), conf->working_disks, mddev->raid_disks);
507         /*
508          * Ok, everything is just fine now
509          */
510         mddev->array_sectors = mddev->size * 2;
511
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;
515
516         return 0;
517
518 out_free_conf:
519         if (conf->pool)
520                 mempool_destroy(conf->pool);
521         kfree(conf->multipaths);
522         kfree(conf);
523         mddev->private = NULL;
524 out:
525         return -EIO;
526 }
527
528
529 static int multipath_stop (mddev_t *mddev)
530 {
531         multipath_conf_t *conf = mddev_to_conf(mddev);
532
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);
538         kfree(conf);
539         mddev->private = NULL;
540         return 0;
541 }
542
543 static struct mdk_personality multipath_personality =
544 {
545         .name           = "multipath",
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,
555 };
556
557 static int __init multipath_init (void)
558 {
559         return register_md_personality (&multipath_personality);
560 }
561
562 static void __exit multipath_exit (void)
563 {
564         unregister_md_personality (&multipath_personality);
565 }
566
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");