Merge master.kernel.org:/pub/scm/linux/kernel/git/roland/infiniband
[linux-2.6] / 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 mdk_personality_t multipath_personality;
39
40
41 static void *mp_pool_alloc(unsigned int __nocast gfp_flags, void *data)
42 {
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
50 static void mp_pool_free(void *mpb, void *data)
51 {
52         kfree(mpb);
53 }
54
55 static 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++) {
66                 mdk_rdev_t *rdev = conf->multipaths[i].rdev;
67                 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
79 static 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  */
97 static 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
106 static int multipath_end_request(struct bio *bio, unsigned int bytes_done,
107                                  int error)
108 {
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
135 static 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++) {
142                 mdk_rdev_t *rdev = conf->multipaths[i].rdev;
143                 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
159 static void multipath_unplug(request_queue_t *q)
160 {
161         unplug_slaves(q->queuedata);
162 }
163
164
165 static 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;
171
172         if (unlikely(bio_barrier(bio))) {
173                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
174                 return 0;
175         }
176
177         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
178
179         mp_bh->master_bio = bio;
180         mp_bh->mddev = mddev;
181
182         if (bio_data_dir(bio)==WRITE) {
183                 disk_stat_inc(mddev->gendisk, writes);
184                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
185         } else {
186                 disk_stat_inc(mddev->gendisk, reads);
187                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
188         }
189
190         mp_bh->path = multipath_map(conf);
191         if (mp_bh->path < 0) {
192                 bio_endio(bio, bio->bi_size, -EIO);
193                 mempool_free(mp_bh, conf->pool);
194                 return 0;
195         }
196         multipath = conf->multipaths + mp_bh->path;
197
198         mp_bh->bio = *bio;
199         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
200         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
201         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
202         mp_bh->bio.bi_end_io = multipath_end_request;
203         mp_bh->bio.bi_private = mp_bh;
204         generic_make_request(&mp_bh->bio);
205         return 0;
206 }
207
208 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
209 {
210         multipath_conf_t *conf = mddev_to_conf(mddev);
211         int i;
212         
213         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
214                                                  conf->working_disks);
215         for (i = 0; i < conf->raid_disks; i++)
216                 seq_printf (seq, "%s",
217                                conf->multipaths[i].rdev && 
218                                conf->multipaths[i].rdev->in_sync ? "U" : "_");
219         seq_printf (seq, "]");
220 }
221
222 static int multipath_issue_flush(request_queue_t *q, struct gendisk *disk,
223                                  sector_t *error_sector)
224 {
225         mddev_t *mddev = q->queuedata;
226         multipath_conf_t *conf = mddev_to_conf(mddev);
227         int i, ret = 0;
228
229         rcu_read_lock();
230         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
231                 mdk_rdev_t *rdev = conf->multipaths[i].rdev;
232                 if (rdev && !rdev->faulty) {
233                         struct block_device *bdev = rdev->bdev;
234                         request_queue_t *r_queue = bdev_get_queue(bdev);
235
236                         if (!r_queue->issue_flush_fn)
237                                 ret = -EOPNOTSUPP;
238                         else {
239                                 atomic_inc(&rdev->nr_pending);
240                                 rcu_read_unlock();
241                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
242                                                               error_sector);
243                                 rdev_dec_pending(rdev, mddev);
244                                 rcu_read_lock();
245                         }
246                 }
247         }
248         rcu_read_unlock();
249         return ret;
250 }
251
252 /*
253  * Careful, this can execute in IRQ contexts as well!
254  */
255 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
256 {
257         multipath_conf_t *conf = mddev_to_conf(mddev);
258
259         if (conf->working_disks <= 1) {
260                 /*
261                  * Uh oh, we can do nothing if this is our last path, but
262                  * first check if this is a queued request for a device
263                  * which has just failed.
264                  */
265                 printk(KERN_ALERT 
266                         "multipath: only one IO path left and IO error.\n");
267                 /* leave it active... it's all we have */
268         } else {
269                 /*
270                  * Mark disk as unusable
271                  */
272                 if (!rdev->faulty) {
273                         char b[BDEVNAME_SIZE];
274                         rdev->in_sync = 0;
275                         rdev->faulty = 1;
276                         mddev->sb_dirty = 1;
277                         conf->working_disks--;
278                         printk(KERN_ALERT "multipath: IO failure on %s,"
279                                 " disabling IO path. \n Operation continuing"
280                                 " on %d IO paths.\n",
281                                 bdevname (rdev->bdev,b),
282                                 conf->working_disks);
283                 }
284         }
285 }
286
287 static void print_multipath_conf (multipath_conf_t *conf)
288 {
289         int i;
290         struct multipath_info *tmp;
291
292         printk("MULTIPATH conf printout:\n");
293         if (!conf) {
294                 printk("(conf==NULL)\n");
295                 return;
296         }
297         printk(" --- wd:%d rd:%d\n", conf->working_disks,
298                          conf->raid_disks);
299
300         for (i = 0; i < conf->raid_disks; i++) {
301                 char b[BDEVNAME_SIZE];
302                 tmp = conf->multipaths + i;
303                 if (tmp->rdev)
304                         printk(" disk%d, o:%d, dev:%s\n",
305                                 i,!tmp->rdev->faulty,
306                                bdevname(tmp->rdev->bdev,b));
307         }
308 }
309
310
311 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
312 {
313         multipath_conf_t *conf = mddev->private;
314         int found = 0;
315         int path;
316         struct multipath_info *p;
317
318         print_multipath_conf(conf);
319
320         for (path=0; path<mddev->raid_disks; path++) 
321                 if ((p=conf->multipaths+path)->rdev == NULL) {
322                         blk_queue_stack_limits(mddev->queue,
323                                                rdev->bdev->bd_disk->queue);
324
325                 /* as we don't honour merge_bvec_fn, we must never risk
326                  * violating it, so limit ->max_sector to one PAGE, as
327                  * a one page request is never in violation.
328                  * (Note: it is very unlikely that a device with
329                  * merge_bvec_fn will be involved in multipath.)
330                  */
331                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
332                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
333                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
334
335                         conf->working_disks++;
336                         rdev->raid_disk = path;
337                         rdev->in_sync = 1;
338                         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 (rdev->in_sync ||
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 = kmalloc(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         memset(conf, 0, sizeof(*conf));
459
460         conf->multipaths = kmalloc(sizeof(struct multipath_info)*mddev->raid_disks,
461                                    GFP_KERNEL);
462         if (!conf->multipaths) {
463                 printk(KERN_ERR 
464                         "multipath: couldn't allocate memory for %s\n",
465                         mdname(mddev));
466                 goto out_free_conf;
467         }
468         memset(conf->multipaths, 0, sizeof(struct multipath_info)*mddev->raid_disks);
469
470         conf->working_disks = 0;
471         ITERATE_RDEV(mddev,rdev,tmp) {
472                 disk_idx = rdev->raid_disk;
473                 if (disk_idx < 0 ||
474                     disk_idx >= mddev->raid_disks)
475                         continue;
476
477                 disk = conf->multipaths + disk_idx;
478                 disk->rdev = rdev;
479
480                 blk_queue_stack_limits(mddev->queue,
481                                        rdev->bdev->bd_disk->queue);
482                 /* as we don't honour merge_bvec_fn, we must never risk
483                  * violating it, not that we ever expect a device with
484                  * a merge_bvec_fn to be involved in multipath */
485                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
486                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
487                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
488
489                 if (!rdev->faulty) 
490                         conf->working_disks++;
491         }
492
493         conf->raid_disks = mddev->raid_disks;
494         mddev->sb_dirty = 1;
495         conf->mddev = mddev;
496         spin_lock_init(&conf->device_lock);
497         INIT_LIST_HEAD(&conf->retry_list);
498
499         if (!conf->working_disks) {
500                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
501                         mdname(mddev));
502                 goto out_free_conf;
503         }
504         mddev->degraded = conf->raid_disks = conf->working_disks;
505
506         conf->pool = mempool_create(NR_RESERVED_BUFS,
507                                     mp_pool_alloc, mp_pool_free,
508                                     NULL);
509         if (conf->pool == NULL) {
510                 printk(KERN_ERR 
511                         "multipath: couldn't allocate memory for %s\n",
512                         mdname(mddev));
513                 goto out_free_conf;
514         }
515
516         {
517                 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
518                 if (!mddev->thread) {
519                         printk(KERN_ERR "multipath: couldn't allocate thread"
520                                 " for %s\n", mdname(mddev));
521                         goto out_free_conf;
522                 }
523         }
524
525         printk(KERN_INFO 
526                 "multipath: array %s active with %d out of %d IO paths\n",
527                 mdname(mddev), conf->working_disks, mddev->raid_disks);
528         /*
529          * Ok, everything is just fine now
530          */
531         mddev->array_size = mddev->size;
532
533         mddev->queue->unplug_fn = multipath_unplug;
534         mddev->queue->issue_flush_fn = multipath_issue_flush;
535
536         return 0;
537
538 out_free_conf:
539         if (conf->pool)
540                 mempool_destroy(conf->pool);
541         kfree(conf->multipaths);
542         kfree(conf);
543         mddev->private = NULL;
544 out:
545         return -EIO;
546 }
547
548
549 static int multipath_stop (mddev_t *mddev)
550 {
551         multipath_conf_t *conf = mddev_to_conf(mddev);
552
553         md_unregister_thread(mddev->thread);
554         mddev->thread = NULL;
555         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
556         mempool_destroy(conf->pool);
557         kfree(conf->multipaths);
558         kfree(conf);
559         mddev->private = NULL;
560         return 0;
561 }
562
563 static mdk_personality_t multipath_personality=
564 {
565         .name           = "multipath",
566         .owner          = THIS_MODULE,
567         .make_request   = multipath_make_request,
568         .run            = multipath_run,
569         .stop           = multipath_stop,
570         .status         = multipath_status,
571         .error_handler  = multipath_error,
572         .hot_add_disk   = multipath_add_disk,
573         .hot_remove_disk= multipath_remove_disk,
574 };
575
576 static int __init multipath_init (void)
577 {
578         return register_md_personality (MULTIPATH, &multipath_personality);
579 }
580
581 static void __exit multipath_exit (void)
582 {
583         unregister_md_personality (MULTIPATH);
584 }
585
586 module_init(multipath_init);
587 module_exit(multipath_exit);
588 MODULE_LICENSE("GPL");
589 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */