xref: /openbmc/linux/drivers/md/md-multipath.c (revision 4a3fad70)
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 "md-multipath.h"
29 
30 #define MAX_WORK_PER_DISK 128
31 
32 #define	NR_RESERVED_BUFS	32
33 
34 static 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 		    !test_bit(Faulty, &rdev->flags)) {
48 			atomic_inc(&rdev->nr_pending);
49 			rcu_read_unlock();
50 			return i;
51 		}
52 	}
53 	rcu_read_unlock();
54 
55 	pr_crit_ratelimited("multipath_map(): no more operational IO paths?\n");
56 	return (-1);
57 }
58 
59 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
60 {
61 	unsigned long flags;
62 	struct mddev *mddev = mp_bh->mddev;
63 	struct mpconf *conf = mddev->private;
64 
65 	spin_lock_irqsave(&conf->device_lock, flags);
66 	list_add(&mp_bh->retry_list, &conf->retry_list);
67 	spin_unlock_irqrestore(&conf->device_lock, flags);
68 	md_wakeup_thread(mddev->thread);
69 }
70 
71 /*
72  * multipath_end_bh_io() is called when we have finished servicing a multipathed
73  * operation and are ready to return a success/failure code to the buffer
74  * cache layer.
75  */
76 static void multipath_end_bh_io(struct multipath_bh *mp_bh, blk_status_t status)
77 {
78 	struct bio *bio = mp_bh->master_bio;
79 	struct mpconf *conf = mp_bh->mddev->private;
80 
81 	bio->bi_status = status;
82 	bio_endio(bio);
83 	mempool_free(mp_bh, conf->pool);
84 }
85 
86 static void multipath_end_request(struct bio *bio)
87 {
88 	struct multipath_bh *mp_bh = bio->bi_private;
89 	struct mpconf *conf = mp_bh->mddev->private;
90 	struct md_rdev *rdev = conf->multipaths[mp_bh->path].rdev;
91 
92 	if (!bio->bi_status)
93 		multipath_end_bh_io(mp_bh, 0);
94 	else if (!(bio->bi_opf & REQ_RAHEAD)) {
95 		/*
96 		 * oops, IO error:
97 		 */
98 		char b[BDEVNAME_SIZE];
99 		md_error (mp_bh->mddev, rdev);
100 		pr_info("multipath: %s: rescheduling sector %llu\n",
101 			bdevname(rdev->bdev,b),
102 			(unsigned long long)bio->bi_iter.bi_sector);
103 		multipath_reschedule_retry(mp_bh);
104 	} else
105 		multipath_end_bh_io(mp_bh, bio->bi_status);
106 	rdev_dec_pending(rdev, conf->mddev);
107 }
108 
109 static bool multipath_make_request(struct mddev *mddev, struct bio * bio)
110 {
111 	struct mpconf *conf = mddev->private;
112 	struct multipath_bh * mp_bh;
113 	struct multipath_info *multipath;
114 
115 	if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
116 		md_flush_request(mddev, bio);
117 		return true;
118 	}
119 
120 	mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
121 
122 	mp_bh->master_bio = bio;
123 	mp_bh->mddev = mddev;
124 
125 	mp_bh->path = multipath_map(conf);
126 	if (mp_bh->path < 0) {
127 		bio_io_error(bio);
128 		mempool_free(mp_bh, conf->pool);
129 		return true;
130 	}
131 	multipath = conf->multipaths + mp_bh->path;
132 
133 	bio_init(&mp_bh->bio, NULL, 0);
134 	__bio_clone_fast(&mp_bh->bio, bio);
135 
136 	mp_bh->bio.bi_iter.bi_sector += multipath->rdev->data_offset;
137 	bio_set_dev(&mp_bh->bio, multipath->rdev->bdev);
138 	mp_bh->bio.bi_opf |= REQ_FAILFAST_TRANSPORT;
139 	mp_bh->bio.bi_end_io = multipath_end_request;
140 	mp_bh->bio.bi_private = mp_bh;
141 	mddev_check_writesame(mddev, &mp_bh->bio);
142 	mddev_check_write_zeroes(mddev, &mp_bh->bio);
143 	generic_make_request(&mp_bh->bio);
144 	return true;
145 }
146 
147 static void multipath_status(struct seq_file *seq, struct mddev *mddev)
148 {
149 	struct mpconf *conf = mddev->private;
150 	int i;
151 
152 	seq_printf (seq, " [%d/%d] [", conf->raid_disks,
153 		    conf->raid_disks - mddev->degraded);
154 	rcu_read_lock();
155 	for (i = 0; i < conf->raid_disks; i++) {
156 		struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
157 		seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
158 	}
159 	rcu_read_unlock();
160 	seq_printf (seq, "]");
161 }
162 
163 static int multipath_congested(struct mddev *mddev, int bits)
164 {
165 	struct mpconf *conf = mddev->private;
166 	int i, ret = 0;
167 
168 	rcu_read_lock();
169 	for (i = 0; i < mddev->raid_disks ; i++) {
170 		struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
171 		if (rdev && !test_bit(Faulty, &rdev->flags)) {
172 			struct request_queue *q = bdev_get_queue(rdev->bdev);
173 
174 			ret |= bdi_congested(q->backing_dev_info, bits);
175 			/* Just like multipath_map, we just check the
176 			 * first available device
177 			 */
178 			break;
179 		}
180 	}
181 	rcu_read_unlock();
182 	return ret;
183 }
184 
185 /*
186  * Careful, this can execute in IRQ contexts as well!
187  */
188 static void multipath_error (struct mddev *mddev, struct md_rdev *rdev)
189 {
190 	struct mpconf *conf = mddev->private;
191 	char b[BDEVNAME_SIZE];
192 
193 	if (conf->raid_disks - mddev->degraded <= 1) {
194 		/*
195 		 * Uh oh, we can do nothing if this is our last path, but
196 		 * first check if this is a queued request for a device
197 		 * which has just failed.
198 		 */
199 		pr_warn("multipath: only one IO path left and IO error.\n");
200 		/* leave it active... it's all we have */
201 		return;
202 	}
203 	/*
204 	 * Mark disk as unusable
205 	 */
206 	if (test_and_clear_bit(In_sync, &rdev->flags)) {
207 		unsigned long flags;
208 		spin_lock_irqsave(&conf->device_lock, flags);
209 		mddev->degraded++;
210 		spin_unlock_irqrestore(&conf->device_lock, flags);
211 	}
212 	set_bit(Faulty, &rdev->flags);
213 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
214 	pr_err("multipath: IO failure on %s, disabling IO path.\n"
215 	       "multipath: Operation continuing on %d IO paths.\n",
216 	       bdevname(rdev->bdev, b),
217 	       conf->raid_disks - mddev->degraded);
218 }
219 
220 static void print_multipath_conf (struct mpconf *conf)
221 {
222 	int i;
223 	struct multipath_info *tmp;
224 
225 	pr_debug("MULTIPATH conf printout:\n");
226 	if (!conf) {
227 		pr_debug("(conf==NULL)\n");
228 		return;
229 	}
230 	pr_debug(" --- 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 			pr_debug(" disk%d, o:%d, dev:%s\n",
238 				 i,!test_bit(Faulty, &tmp->rdev->flags),
239 				 bdevname(tmp->rdev->bdev,b));
240 	}
241 }
242 
243 static int multipath_add_disk(struct mddev *mddev, struct md_rdev *rdev)
244 {
245 	struct mpconf *conf = mddev->private;
246 	int err = -EEXIST;
247 	int path;
248 	struct multipath_info *p;
249 	int first = 0;
250 	int last = mddev->raid_disks - 1;
251 
252 	if (rdev->raid_disk >= 0)
253 		first = last = rdev->raid_disk;
254 
255 	print_multipath_conf(conf);
256 
257 	for (path = first; path <= last; path++)
258 		if ((p=conf->multipaths+path)->rdev == NULL) {
259 			disk_stack_limits(mddev->gendisk, rdev->bdev,
260 					  rdev->data_offset << 9);
261 
262 			err = md_integrity_add_rdev(rdev, mddev);
263 			if (err)
264 				break;
265 			spin_lock_irq(&conf->device_lock);
266 			mddev->degraded--;
267 			rdev->raid_disk = path;
268 			set_bit(In_sync, &rdev->flags);
269 			spin_unlock_irq(&conf->device_lock);
270 			rcu_assign_pointer(p->rdev, rdev);
271 			err = 0;
272 			break;
273 		}
274 
275 	print_multipath_conf(conf);
276 
277 	return err;
278 }
279 
280 static int multipath_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
281 {
282 	struct mpconf *conf = mddev->private;
283 	int err = 0;
284 	int number = rdev->raid_disk;
285 	struct multipath_info *p = conf->multipaths + number;
286 
287 	print_multipath_conf(conf);
288 
289 	if (rdev == p->rdev) {
290 		if (test_bit(In_sync, &rdev->flags) ||
291 		    atomic_read(&rdev->nr_pending)) {
292 			pr_warn("hot-remove-disk, slot %d is identified but is still operational!\n", number);
293 			err = -EBUSY;
294 			goto abort;
295 		}
296 		p->rdev = NULL;
297 		if (!test_bit(RemoveSynchronized, &rdev->flags)) {
298 			synchronize_rcu();
299 			if (atomic_read(&rdev->nr_pending)) {
300 				/* lost the race, try later */
301 				err = -EBUSY;
302 				p->rdev = rdev;
303 				goto abort;
304 			}
305 		}
306 		err = md_integrity_register(mddev);
307 	}
308 abort:
309 
310 	print_multipath_conf(conf);
311 	return err;
312 }
313 
314 /*
315  * This is a kernel thread which:
316  *
317  *	1.	Retries failed read operations on working multipaths.
318  *	2.	Updates the raid superblock when problems encounter.
319  *	3.	Performs writes following reads for array syncronising.
320  */
321 
322 static void multipathd(struct md_thread *thread)
323 {
324 	struct mddev *mddev = thread->mddev;
325 	struct multipath_bh *mp_bh;
326 	struct bio *bio;
327 	unsigned long flags;
328 	struct mpconf *conf = mddev->private;
329 	struct list_head *head = &conf->retry_list;
330 
331 	md_check_recovery(mddev);
332 	for (;;) {
333 		char b[BDEVNAME_SIZE];
334 		spin_lock_irqsave(&conf->device_lock, flags);
335 		if (list_empty(head))
336 			break;
337 		mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
338 		list_del(head->prev);
339 		spin_unlock_irqrestore(&conf->device_lock, flags);
340 
341 		bio = &mp_bh->bio;
342 		bio->bi_iter.bi_sector = mp_bh->master_bio->bi_iter.bi_sector;
343 
344 		if ((mp_bh->path = multipath_map (conf))<0) {
345 			pr_err("multipath: %s: unrecoverable IO read error for block %llu\n",
346 			       bio_devname(bio, b),
347 			       (unsigned long long)bio->bi_iter.bi_sector);
348 			multipath_end_bh_io(mp_bh, BLK_STS_IOERR);
349 		} else {
350 			pr_err("multipath: %s: redirecting sector %llu to another IO path\n",
351 			       bio_devname(bio, b),
352 			       (unsigned long long)bio->bi_iter.bi_sector);
353 			*bio = *(mp_bh->master_bio);
354 			bio->bi_iter.bi_sector +=
355 				conf->multipaths[mp_bh->path].rdev->data_offset;
356 			bio_set_dev(bio, conf->multipaths[mp_bh->path].rdev->bdev);
357 			bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
358 			bio->bi_end_io = multipath_end_request;
359 			bio->bi_private = mp_bh;
360 			generic_make_request(bio);
361 		}
362 	}
363 	spin_unlock_irqrestore(&conf->device_lock, flags);
364 }
365 
366 static sector_t multipath_size(struct mddev *mddev, sector_t sectors, int raid_disks)
367 {
368 	WARN_ONCE(sectors || raid_disks,
369 		  "%s does not support generic reshape\n", __func__);
370 
371 	return mddev->dev_sectors;
372 }
373 
374 static int multipath_run (struct mddev *mddev)
375 {
376 	struct mpconf *conf;
377 	int disk_idx;
378 	struct multipath_info *disk;
379 	struct md_rdev *rdev;
380 	int working_disks;
381 
382 	if (md_check_no_bitmap(mddev))
383 		return -EINVAL;
384 
385 	if (mddev->level != LEVEL_MULTIPATH) {
386 		pr_warn("multipath: %s: raid level not set to multipath IO (%d)\n",
387 			mdname(mddev), mddev->level);
388 		goto out;
389 	}
390 	/*
391 	 * copy the already verified devices into our private MULTIPATH
392 	 * bookkeeping area. [whatever we allocate in multipath_run(),
393 	 * should be freed in multipath_free()]
394 	 */
395 
396 	conf = kzalloc(sizeof(struct mpconf), GFP_KERNEL);
397 	mddev->private = conf;
398 	if (!conf)
399 		goto out;
400 
401 	conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
402 				   GFP_KERNEL);
403 	if (!conf->multipaths)
404 		goto out_free_conf;
405 
406 	working_disks = 0;
407 	rdev_for_each(rdev, mddev) {
408 		disk_idx = rdev->raid_disk;
409 		if (disk_idx < 0 ||
410 		    disk_idx >= mddev->raid_disks)
411 			continue;
412 
413 		disk = conf->multipaths + disk_idx;
414 		disk->rdev = rdev;
415 		disk_stack_limits(mddev->gendisk, rdev->bdev,
416 				  rdev->data_offset << 9);
417 
418 		if (!test_bit(Faulty, &rdev->flags))
419 			working_disks++;
420 	}
421 
422 	conf->raid_disks = mddev->raid_disks;
423 	conf->mddev = mddev;
424 	spin_lock_init(&conf->device_lock);
425 	INIT_LIST_HEAD(&conf->retry_list);
426 
427 	if (!working_disks) {
428 		pr_warn("multipath: no operational IO paths for %s\n",
429 			mdname(mddev));
430 		goto out_free_conf;
431 	}
432 	mddev->degraded = conf->raid_disks - working_disks;
433 
434 	conf->pool = mempool_create_kmalloc_pool(NR_RESERVED_BUFS,
435 						 sizeof(struct multipath_bh));
436 	if (conf->pool == NULL)
437 		goto out_free_conf;
438 
439 	mddev->thread = md_register_thread(multipathd, mddev,
440 					   "multipath");
441 	if (!mddev->thread)
442 		goto out_free_conf;
443 
444 	pr_info("multipath: array %s active with %d out of %d IO paths\n",
445 		mdname(mddev), conf->raid_disks - mddev->degraded,
446 		mddev->raid_disks);
447 	/*
448 	 * Ok, everything is just fine now
449 	 */
450 	md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
451 
452 	if (md_integrity_register(mddev))
453 		goto out_free_conf;
454 
455 	return 0;
456 
457 out_free_conf:
458 	mempool_destroy(conf->pool);
459 	kfree(conf->multipaths);
460 	kfree(conf);
461 	mddev->private = NULL;
462 out:
463 	return -EIO;
464 }
465 
466 static void multipath_free(struct mddev *mddev, void *priv)
467 {
468 	struct mpconf *conf = priv;
469 
470 	mempool_destroy(conf->pool);
471 	kfree(conf->multipaths);
472 	kfree(conf);
473 }
474 
475 static struct md_personality multipath_personality =
476 {
477 	.name		= "multipath",
478 	.level		= LEVEL_MULTIPATH,
479 	.owner		= THIS_MODULE,
480 	.make_request	= multipath_make_request,
481 	.run		= multipath_run,
482 	.free		= multipath_free,
483 	.status		= multipath_status,
484 	.error_handler	= multipath_error,
485 	.hot_add_disk	= multipath_add_disk,
486 	.hot_remove_disk= multipath_remove_disk,
487 	.size		= multipath_size,
488 	.congested	= multipath_congested,
489 };
490 
491 static int __init multipath_init (void)
492 {
493 	return register_md_personality (&multipath_personality);
494 }
495 
496 static void __exit multipath_exit (void)
497 {
498 	unregister_md_personality (&multipath_personality);
499 }
500 
501 module_init(multipath_init);
502 module_exit(multipath_exit);
503 MODULE_LICENSE("GPL");
504 MODULE_DESCRIPTION("simple multi-path personality for MD");
505 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
506 MODULE_ALIAS("md-multipath");
507 MODULE_ALIAS("md-level--4");
508