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