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