Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[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/raid/multipath.h>
23
24 #define MAX_WORK_PER_DISK 128
25
26 #define NR_RESERVED_BUFS        32
27
28
29 static int multipath_map (multipath_conf_t *conf)
30 {
31         int i, disks = conf->raid_disks;
32
33         /*
34          * Later we do read balancing on the read side 
35          * now we use the first available disk.
36          */
37
38         rcu_read_lock();
39         for (i = 0; i < disks; i++) {
40                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
41                 if (rdev && test_bit(In_sync, &rdev->flags)) {
42                         atomic_inc(&rdev->nr_pending);
43                         rcu_read_unlock();
44                         return i;
45                 }
46         }
47         rcu_read_unlock();
48
49         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
50         return (-1);
51 }
52
53 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
54 {
55         unsigned long flags;
56         mddev_t *mddev = mp_bh->mddev;
57         multipath_conf_t *conf = mddev_to_conf(mddev);
58
59         spin_lock_irqsave(&conf->device_lock, flags);
60         list_add(&mp_bh->retry_list, &conf->retry_list);
61         spin_unlock_irqrestore(&conf->device_lock, flags);
62         md_wakeup_thread(mddev->thread);
63 }
64
65
66 /*
67  * multipath_end_bh_io() is called when we have finished servicing a multipathed
68  * operation and are ready to return a success/failure code to the buffer
69  * cache layer.
70  */
71 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
72 {
73         struct bio *bio = mp_bh->master_bio;
74         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
75
76         bio_endio(bio, err);
77         mempool_free(mp_bh, conf->pool);
78 }
79
80 static void multipath_end_request(struct bio *bio, int error)
81 {
82         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
83         struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
84         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
85         mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
86
87         if (uptodate)
88                 multipath_end_bh_io(mp_bh, 0);
89         else if (!bio_rw_ahead(bio)) {
90                 /*
91                  * oops, IO error:
92                  */
93                 char b[BDEVNAME_SIZE];
94                 md_error (mp_bh->mddev, rdev);
95                 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n", 
96                        bdevname(rdev->bdev,b), 
97                        (unsigned long long)bio->bi_sector);
98                 multipath_reschedule_retry(mp_bh);
99         } else
100                 multipath_end_bh_io(mp_bh, error);
101         rdev_dec_pending(rdev, conf->mddev);
102 }
103
104 static void unplug_slaves(mddev_t *mddev)
105 {
106         multipath_conf_t *conf = mddev_to_conf(mddev);
107         int i;
108
109         rcu_read_lock();
110         for (i=0; i<mddev->raid_disks; i++) {
111                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
112                 if (rdev && !test_bit(Faulty, &rdev->flags)
113                     && atomic_read(&rdev->nr_pending)) {
114                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
115
116                         atomic_inc(&rdev->nr_pending);
117                         rcu_read_unlock();
118
119                         blk_unplug(r_queue);
120
121                         rdev_dec_pending(rdev, mddev);
122                         rcu_read_lock();
123                 }
124         }
125         rcu_read_unlock();
126 }
127
128 static void multipath_unplug(struct request_queue *q)
129 {
130         unplug_slaves(q->queuedata);
131 }
132
133
134 static int multipath_make_request (struct request_queue *q, struct bio * bio)
135 {
136         mddev_t *mddev = q->queuedata;
137         multipath_conf_t *conf = mddev_to_conf(mddev);
138         struct multipath_bh * mp_bh;
139         struct multipath_info *multipath;
140         const int rw = bio_data_dir(bio);
141         int cpu;
142
143         if (unlikely(bio_barrier(bio))) {
144                 bio_endio(bio, -EOPNOTSUPP);
145                 return 0;
146         }
147
148         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
149
150         mp_bh->master_bio = bio;
151         mp_bh->mddev = mddev;
152
153         cpu = part_stat_lock();
154         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
155         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
156                       bio_sectors(bio));
157         part_stat_unlock();
158
159         mp_bh->path = multipath_map(conf);
160         if (mp_bh->path < 0) {
161                 bio_endio(bio, -EIO);
162                 mempool_free(mp_bh, conf->pool);
163                 return 0;
164         }
165         multipath = conf->multipaths + mp_bh->path;
166
167         mp_bh->bio = *bio;
168         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
169         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
170         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
171         mp_bh->bio.bi_end_io = multipath_end_request;
172         mp_bh->bio.bi_private = mp_bh;
173         generic_make_request(&mp_bh->bio);
174         return 0;
175 }
176
177 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
178 {
179         multipath_conf_t *conf = mddev_to_conf(mddev);
180         int i;
181         
182         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
183                                                  conf->working_disks);
184         for (i = 0; i < conf->raid_disks; i++)
185                 seq_printf (seq, "%s",
186                                conf->multipaths[i].rdev && 
187                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
188         seq_printf (seq, "]");
189 }
190
191 static int multipath_congested(void *data, int bits)
192 {
193         mddev_t *mddev = data;
194         multipath_conf_t *conf = mddev_to_conf(mddev);
195         int i, ret = 0;
196
197         rcu_read_lock();
198         for (i = 0; i < mddev->raid_disks ; i++) {
199                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
200                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
201                         struct request_queue *q = bdev_get_queue(rdev->bdev);
202
203                         ret |= bdi_congested(&q->backing_dev_info, bits);
204                         /* Just like multipath_map, we just check the
205                          * first available device
206                          */
207                         break;
208                 }
209         }
210         rcu_read_unlock();
211         return ret;
212 }
213
214 /*
215  * Careful, this can execute in IRQ contexts as well!
216  */
217 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
218 {
219         multipath_conf_t *conf = mddev_to_conf(mddev);
220
221         if (conf->working_disks <= 1) {
222                 /*
223                  * Uh oh, we can do nothing if this is our last path, but
224                  * first check if this is a queued request for a device
225                  * which has just failed.
226                  */
227                 printk(KERN_ALERT 
228                         "multipath: only one IO path left and IO error.\n");
229                 /* leave it active... it's all we have */
230         } else {
231                 /*
232                  * Mark disk as unusable
233                  */
234                 if (!test_bit(Faulty, &rdev->flags)) {
235                         char b[BDEVNAME_SIZE];
236                         clear_bit(In_sync, &rdev->flags);
237                         set_bit(Faulty, &rdev->flags);
238                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
239                         conf->working_disks--;
240                         mddev->degraded++;
241                         printk(KERN_ALERT "multipath: IO failure on %s,"
242                                 " disabling IO path.\n"
243                                 "multipath: Operation continuing"
244                                 " on %d IO paths.\n",
245                                 bdevname (rdev->bdev,b),
246                                 conf->working_disks);
247                 }
248         }
249 }
250
251 static void print_multipath_conf (multipath_conf_t *conf)
252 {
253         int i;
254         struct multipath_info *tmp;
255
256         printk("MULTIPATH conf printout:\n");
257         if (!conf) {
258                 printk("(conf==NULL)\n");
259                 return;
260         }
261         printk(" --- wd:%d rd:%d\n", conf->working_disks,
262                          conf->raid_disks);
263
264         for (i = 0; i < conf->raid_disks; i++) {
265                 char b[BDEVNAME_SIZE];
266                 tmp = conf->multipaths + i;
267                 if (tmp->rdev)
268                         printk(" disk%d, o:%d, dev:%s\n",
269                                 i,!test_bit(Faulty, &tmp->rdev->flags),
270                                bdevname(tmp->rdev->bdev,b));
271         }
272 }
273
274
275 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
276 {
277         multipath_conf_t *conf = mddev->private;
278         struct request_queue *q;
279         int err = -EEXIST;
280         int path;
281         struct multipath_info *p;
282         int first = 0;
283         int last = mddev->raid_disks - 1;
284
285         if (rdev->raid_disk >= 0)
286                 first = last = rdev->raid_disk;
287
288         print_multipath_conf(conf);
289
290         for (path = first; path <= last; path++)
291                 if ((p=conf->multipaths+path)->rdev == NULL) {
292                         q = rdev->bdev->bd_disk->queue;
293                         blk_queue_stack_limits(mddev->queue, q);
294
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.)
300                  */
301                         if (q->merge_bvec_fn &&
302                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
303                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
304
305                         conf->working_disks++;
306                         mddev->degraded--;
307                         rdev->raid_disk = path;
308                         set_bit(In_sync, &rdev->flags);
309                         rcu_assign_pointer(p->rdev, rdev);
310                         err = 0;
311                         break;
312                 }
313
314         print_multipath_conf(conf);
315
316         return err;
317 }
318
319 static int multipath_remove_disk(mddev_t *mddev, int number)
320 {
321         multipath_conf_t *conf = mddev->private;
322         int err = 0;
323         mdk_rdev_t *rdev;
324         struct multipath_info *p = conf->multipaths + number;
325
326         print_multipath_conf(conf);
327
328         rdev = p->rdev;
329         if (rdev) {
330                 if (test_bit(In_sync, &rdev->flags) ||
331                     atomic_read(&rdev->nr_pending)) {
332                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"
333                                " but is still operational!\n", number);
334                         err = -EBUSY;
335                         goto abort;
336                 }
337                 p->rdev = NULL;
338                 synchronize_rcu();
339                 if (atomic_read(&rdev->nr_pending)) {
340                         /* lost the race, try later */
341                         err = -EBUSY;
342                         p->rdev = rdev;
343                 }
344         }
345 abort:
346
347         print_multipath_conf(conf);
348         return err;
349 }
350
351
352
353 /*
354  * This is a kernel thread which:
355  *
356  *      1.      Retries failed read operations on working multipaths.
357  *      2.      Updates the raid superblock when problems encounter.
358  *      3.      Performs writes following reads for array syncronising.
359  */
360
361 static void multipathd (mddev_t *mddev)
362 {
363         struct multipath_bh *mp_bh;
364         struct bio *bio;
365         unsigned long flags;
366         multipath_conf_t *conf = mddev_to_conf(mddev);
367         struct list_head *head = &conf->retry_list;
368
369         md_check_recovery(mddev);
370         for (;;) {
371                 char b[BDEVNAME_SIZE];
372                 spin_lock_irqsave(&conf->device_lock, flags);
373                 if (list_empty(head))
374                         break;
375                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
376                 list_del(head->prev);
377                 spin_unlock_irqrestore(&conf->device_lock, flags);
378
379                 bio = &mp_bh->bio;
380                 bio->bi_sector = mp_bh->master_bio->bi_sector;
381                 
382                 if ((mp_bh->path = multipath_map (conf))<0) {
383                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
384                                 " error for block %llu\n",
385                                 bdevname(bio->bi_bdev,b),
386                                 (unsigned long long)bio->bi_sector);
387                         multipath_end_bh_io(mp_bh, -EIO);
388                 } else {
389                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
390                                 " to another IO path\n",
391                                 bdevname(bio->bi_bdev,b),
392                                 (unsigned long long)bio->bi_sector);
393                         *bio = *(mp_bh->master_bio);
394                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
395                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
396                         bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
397                         bio->bi_end_io = multipath_end_request;
398                         bio->bi_private = mp_bh;
399                         generic_make_request(bio);
400                 }
401         }
402         spin_unlock_irqrestore(&conf->device_lock, flags);
403 }
404
405 static int multipath_run (mddev_t *mddev)
406 {
407         multipath_conf_t *conf;
408         int disk_idx;
409         struct multipath_info *disk;
410         mdk_rdev_t *rdev;
411
412         if (mddev->level != LEVEL_MULTIPATH) {
413                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
414                        mdname(mddev), mddev->level);
415                 goto out;
416         }
417         /*
418          * copy the already verified devices into our private MULTIPATH
419          * bookkeeping area. [whatever we allocate in multipath_run(),
420          * should be freed in multipath_stop()]
421          */
422         mddev->queue->queue_lock = &mddev->queue->__queue_lock;
423
424         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
425         mddev->private = conf;
426         if (!conf) {
427                 printk(KERN_ERR 
428                         "multipath: couldn't allocate memory for %s\n",
429                         mdname(mddev));
430                 goto out;
431         }
432
433         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
434                                    GFP_KERNEL);
435         if (!conf->multipaths) {
436                 printk(KERN_ERR 
437                         "multipath: couldn't allocate memory for %s\n",
438                         mdname(mddev));
439                 goto out_free_conf;
440         }
441
442         conf->working_disks = 0;
443         list_for_each_entry(rdev, &mddev->disks, same_set) {
444                 disk_idx = rdev->raid_disk;
445                 if (disk_idx < 0 ||
446                     disk_idx >= mddev->raid_disks)
447                         continue;
448
449                 disk = conf->multipaths + disk_idx;
450                 disk->rdev = rdev;
451
452                 blk_queue_stack_limits(mddev->queue,
453                                        rdev->bdev->bd_disk->queue);
454                 /* as we don't honour merge_bvec_fn, we must never risk
455                  * violating it, not that we ever expect a device with
456                  * a merge_bvec_fn to be involved in multipath */
457                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
458                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
459                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
460
461                 if (!test_bit(Faulty, &rdev->flags))
462                         conf->working_disks++;
463         }
464
465         conf->raid_disks = mddev->raid_disks;
466         conf->mddev = mddev;
467         spin_lock_init(&conf->device_lock);
468         INIT_LIST_HEAD(&conf->retry_list);
469
470         if (!conf->working_disks) {
471                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
472                         mdname(mddev));
473                 goto out_free_conf;
474         }
475         mddev->degraded = conf->raid_disks - conf->working_disks;
476
477         conf->pool = mempool_create_kzalloc_pool(NR_RESERVED_BUFS,
478                                                  sizeof(struct multipath_bh));
479         if (conf->pool == NULL) {
480                 printk(KERN_ERR 
481                         "multipath: couldn't allocate memory for %s\n",
482                         mdname(mddev));
483                 goto out_free_conf;
484         }
485
486         {
487                 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
488                 if (!mddev->thread) {
489                         printk(KERN_ERR "multipath: couldn't allocate thread"
490                                 " for %s\n", mdname(mddev));
491                         goto out_free_conf;
492                 }
493         }
494
495         printk(KERN_INFO 
496                 "multipath: array %s active with %d out of %d IO paths\n",
497                 mdname(mddev), conf->working_disks, mddev->raid_disks);
498         /*
499          * Ok, everything is just fine now
500          */
501         mddev->array_sectors = mddev->size * 2;
502
503         mddev->queue->unplug_fn = multipath_unplug;
504         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
505         mddev->queue->backing_dev_info.congested_data = mddev;
506
507         return 0;
508
509 out_free_conf:
510         if (conf->pool)
511                 mempool_destroy(conf->pool);
512         kfree(conf->multipaths);
513         kfree(conf);
514         mddev->private = NULL;
515 out:
516         return -EIO;
517 }
518
519
520 static int multipath_stop (mddev_t *mddev)
521 {
522         multipath_conf_t *conf = mddev_to_conf(mddev);
523
524         md_unregister_thread(mddev->thread);
525         mddev->thread = NULL;
526         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
527         mempool_destroy(conf->pool);
528         kfree(conf->multipaths);
529         kfree(conf);
530         mddev->private = NULL;
531         return 0;
532 }
533
534 static struct mdk_personality multipath_personality =
535 {
536         .name           = "multipath",
537         .level          = LEVEL_MULTIPATH,
538         .owner          = THIS_MODULE,
539         .make_request   = multipath_make_request,
540         .run            = multipath_run,
541         .stop           = multipath_stop,
542         .status         = multipath_status,
543         .error_handler  = multipath_error,
544         .hot_add_disk   = multipath_add_disk,
545         .hot_remove_disk= multipath_remove_disk,
546 };
547
548 static int __init multipath_init (void)
549 {
550         return register_md_personality (&multipath_personality);
551 }
552
553 static void __exit multipath_exit (void)
554 {
555         unregister_md_personality (&multipath_personality);
556 }
557
558 module_init(multipath_init);
559 module_exit(multipath_exit);
560 MODULE_LICENSE("GPL");
561 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
562 MODULE_ALIAS("md-multipath");
563 MODULE_ALIAS("md-level--4");