xref: /openbmc/linux/drivers/md/md-multipath.c (revision 55eb9a6c)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * multipath.c : Multiple Devices driver for Linux
4  *
5  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6  *
7  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8  *
9  * MULTIPATH management functions.
10  *
11  * derived from raid1.c.
12  */
13 
14 #include <linux/blkdev.h>
15 #include <linux/module.h>
16 #include <linux/raid/md_u.h>
17 #include <linux/seq_file.h>
18 #include <linux/slab.h>
19 #include "md.h"
20 #include "md-multipath.h"
21 
22 #define MAX_WORK_PER_DISK 128
23 
24 #define	NR_RESERVED_BUFS	32
25 
26 static int multipath_map (struct mpconf *conf)
27 {
28 	int i, disks = conf->raid_disks;
29 
30 	/*
31 	 * Later we do read balancing on the read side
32 	 * now we use the first available disk.
33 	 */
34 
35 	rcu_read_lock();
36 	for (i = 0; i < disks; i++) {
37 		struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
38 		if (rdev && test_bit(In_sync, &rdev->flags) &&
39 		    !test_bit(Faulty, &rdev->flags)) {
40 			atomic_inc(&rdev->nr_pending);
41 			rcu_read_unlock();
42 			return i;
43 		}
44 	}
45 	rcu_read_unlock();
46 
47 	pr_crit_ratelimited("multipath_map(): no more operational IO paths?\n");
48 	return (-1);
49 }
50 
51 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
52 {
53 	unsigned long flags;
54 	struct mddev *mddev = mp_bh->mddev;
55 	struct mpconf *conf = mddev->private;
56 
57 	spin_lock_irqsave(&conf->device_lock, flags);
58 	list_add(&mp_bh->retry_list, &conf->retry_list);
59 	spin_unlock_irqrestore(&conf->device_lock, flags);
60 	md_wakeup_thread(mddev->thread);
61 }
62 
63 /*
64  * multipath_end_bh_io() is called when we have finished servicing a multipathed
65  * operation and are ready to return a success/failure code to the buffer
66  * cache layer.
67  */
68 static void multipath_end_bh_io(struct multipath_bh *mp_bh, blk_status_t status)
69 {
70 	struct bio *bio = mp_bh->master_bio;
71 	struct mpconf *conf = mp_bh->mddev->private;
72 
73 	bio->bi_status = status;
74 	bio_endio(bio);
75 	mempool_free(mp_bh, &conf->pool);
76 }
77 
78 static void multipath_end_request(struct bio *bio)
79 {
80 	struct multipath_bh *mp_bh = bio->bi_private;
81 	struct mpconf *conf = mp_bh->mddev->private;
82 	struct md_rdev *rdev = conf->multipaths[mp_bh->path].rdev;
83 
84 	if (!bio->bi_status)
85 		multipath_end_bh_io(mp_bh, 0);
86 	else if (!(bio->bi_opf & REQ_RAHEAD)) {
87 		/*
88 		 * oops, IO error:
89 		 */
90 		char b[BDEVNAME_SIZE];
91 		md_error (mp_bh->mddev, rdev);
92 		pr_info("multipath: %s: rescheduling sector %llu\n",
93 			bdevname(rdev->bdev,b),
94 			(unsigned long long)bio->bi_iter.bi_sector);
95 		multipath_reschedule_retry(mp_bh);
96 	} else
97 		multipath_end_bh_io(mp_bh, bio->bi_status);
98 	rdev_dec_pending(rdev, conf->mddev);
99 }
100 
101 static bool multipath_make_request(struct mddev *mddev, struct bio * bio)
102 {
103 	struct mpconf *conf = mddev->private;
104 	struct multipath_bh * mp_bh;
105 	struct multipath_info *multipath;
106 
107 	if (unlikely(bio->bi_opf & REQ_PREFLUSH)
108 	    && md_flush_request(mddev, bio))
109 		return true;
110 
111 	mp_bh = mempool_alloc(&conf->pool, GFP_NOIO);
112 
113 	mp_bh->master_bio = bio;
114 	mp_bh->mddev = mddev;
115 
116 	mp_bh->path = multipath_map(conf);
117 	if (mp_bh->path < 0) {
118 		bio_io_error(bio);
119 		mempool_free(mp_bh, &conf->pool);
120 		return true;
121 	}
122 	multipath = conf->multipaths + mp_bh->path;
123 
124 	bio_init_clone(multipath->rdev->bdev, &mp_bh->bio, bio, GFP_NOIO);
125 
126 	mp_bh->bio.bi_iter.bi_sector += multipath->rdev->data_offset;
127 	mp_bh->bio.bi_opf |= REQ_FAILFAST_TRANSPORT;
128 	mp_bh->bio.bi_end_io = multipath_end_request;
129 	mp_bh->bio.bi_private = mp_bh;
130 	mddev_check_write_zeroes(mddev, &mp_bh->bio);
131 	submit_bio_noacct(&mp_bh->bio);
132 	return true;
133 }
134 
135 static void multipath_status(struct seq_file *seq, struct mddev *mddev)
136 {
137 	struct mpconf *conf = mddev->private;
138 	int i;
139 
140 	seq_printf (seq, " [%d/%d] [", conf->raid_disks,
141 		    conf->raid_disks - mddev->degraded);
142 	rcu_read_lock();
143 	for (i = 0; i < conf->raid_disks; i++) {
144 		struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
145 		seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
146 	}
147 	rcu_read_unlock();
148 	seq_putc(seq, ']');
149 }
150 
151 /*
152  * Careful, this can execute in IRQ contexts as well!
153  */
154 static void multipath_error (struct mddev *mddev, struct md_rdev *rdev)
155 {
156 	struct mpconf *conf = mddev->private;
157 	char b[BDEVNAME_SIZE];
158 
159 	if (conf->raid_disks - mddev->degraded <= 1) {
160 		/*
161 		 * Uh oh, we can do nothing if this is our last path, but
162 		 * first check if this is a queued request for a device
163 		 * which has just failed.
164 		 */
165 		pr_warn("multipath: only one IO path left and IO error.\n");
166 		/* leave it active... it's all we have */
167 		return;
168 	}
169 	/*
170 	 * Mark disk as unusable
171 	 */
172 	if (test_and_clear_bit(In_sync, &rdev->flags)) {
173 		unsigned long flags;
174 		spin_lock_irqsave(&conf->device_lock, flags);
175 		mddev->degraded++;
176 		spin_unlock_irqrestore(&conf->device_lock, flags);
177 	}
178 	set_bit(Faulty, &rdev->flags);
179 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
180 	pr_err("multipath: IO failure on %s, disabling IO path.\n"
181 	       "multipath: Operation continuing on %d IO paths.\n",
182 	       bdevname(rdev->bdev, b),
183 	       conf->raid_disks - mddev->degraded);
184 }
185 
186 static void print_multipath_conf (struct mpconf *conf)
187 {
188 	int i;
189 	struct multipath_info *tmp;
190 
191 	pr_debug("MULTIPATH conf printout:\n");
192 	if (!conf) {
193 		pr_debug("(conf==NULL)\n");
194 		return;
195 	}
196 	pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
197 		 conf->raid_disks);
198 
199 	for (i = 0; i < conf->raid_disks; i++) {
200 		char b[BDEVNAME_SIZE];
201 		tmp = conf->multipaths + i;
202 		if (tmp->rdev)
203 			pr_debug(" disk%d, o:%d, dev:%s\n",
204 				 i,!test_bit(Faulty, &tmp->rdev->flags),
205 				 bdevname(tmp->rdev->bdev,b));
206 	}
207 }
208 
209 static int multipath_add_disk(struct mddev *mddev, struct md_rdev *rdev)
210 {
211 	struct mpconf *conf = mddev->private;
212 	int err = -EEXIST;
213 	int path;
214 	struct multipath_info *p;
215 	int first = 0;
216 	int last = mddev->raid_disks - 1;
217 
218 	if (rdev->raid_disk >= 0)
219 		first = last = rdev->raid_disk;
220 
221 	print_multipath_conf(conf);
222 
223 	for (path = first; path <= last; path++)
224 		if ((p=conf->multipaths+path)->rdev == NULL) {
225 			disk_stack_limits(mddev->gendisk, rdev->bdev,
226 					  rdev->data_offset << 9);
227 
228 			err = md_integrity_add_rdev(rdev, mddev);
229 			if (err)
230 				break;
231 			spin_lock_irq(&conf->device_lock);
232 			mddev->degraded--;
233 			rdev->raid_disk = path;
234 			set_bit(In_sync, &rdev->flags);
235 			spin_unlock_irq(&conf->device_lock);
236 			rcu_assign_pointer(p->rdev, rdev);
237 			err = 0;
238 			break;
239 		}
240 
241 	print_multipath_conf(conf);
242 
243 	return err;
244 }
245 
246 static int multipath_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
247 {
248 	struct mpconf *conf = mddev->private;
249 	int err = 0;
250 	int number = rdev->raid_disk;
251 	struct multipath_info *p = conf->multipaths + number;
252 
253 	print_multipath_conf(conf);
254 
255 	if (rdev == p->rdev) {
256 		if (test_bit(In_sync, &rdev->flags) ||
257 		    atomic_read(&rdev->nr_pending)) {
258 			pr_warn("hot-remove-disk, slot %d is identified but is still operational!\n", number);
259 			err = -EBUSY;
260 			goto abort;
261 		}
262 		p->rdev = NULL;
263 		if (!test_bit(RemoveSynchronized, &rdev->flags)) {
264 			synchronize_rcu();
265 			if (atomic_read(&rdev->nr_pending)) {
266 				/* lost the race, try later */
267 				err = -EBUSY;
268 				p->rdev = rdev;
269 				goto abort;
270 			}
271 		}
272 		err = md_integrity_register(mddev);
273 	}
274 abort:
275 
276 	print_multipath_conf(conf);
277 	return err;
278 }
279 
280 /*
281  * This is a kernel thread which:
282  *
283  *	1.	Retries failed read operations on working multipaths.
284  *	2.	Updates the raid superblock when problems encounter.
285  *	3.	Performs writes following reads for array syncronising.
286  */
287 
288 static void multipathd(struct md_thread *thread)
289 {
290 	struct mddev *mddev = thread->mddev;
291 	struct multipath_bh *mp_bh;
292 	struct bio *bio;
293 	unsigned long flags;
294 	struct mpconf *conf = mddev->private;
295 	struct list_head *head = &conf->retry_list;
296 
297 	md_check_recovery(mddev);
298 	for (;;) {
299 		spin_lock_irqsave(&conf->device_lock, flags);
300 		if (list_empty(head))
301 			break;
302 		mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
303 		list_del(head->prev);
304 		spin_unlock_irqrestore(&conf->device_lock, flags);
305 
306 		bio = &mp_bh->bio;
307 		bio->bi_iter.bi_sector = mp_bh->master_bio->bi_iter.bi_sector;
308 
309 		if ((mp_bh->path = multipath_map (conf))<0) {
310 			pr_err("multipath: %pg: unrecoverable IO read error for block %llu\n",
311 			       bio->bi_bdev,
312 			       (unsigned long long)bio->bi_iter.bi_sector);
313 			multipath_end_bh_io(mp_bh, BLK_STS_IOERR);
314 		} else {
315 			pr_err("multipath: %pg: redirecting sector %llu to another IO path\n",
316 			       bio->bi_bdev,
317 			       (unsigned long long)bio->bi_iter.bi_sector);
318 			*bio = *(mp_bh->master_bio);
319 			bio->bi_iter.bi_sector +=
320 				conf->multipaths[mp_bh->path].rdev->data_offset;
321 			bio_set_dev(bio, conf->multipaths[mp_bh->path].rdev->bdev);
322 			bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
323 			bio->bi_end_io = multipath_end_request;
324 			bio->bi_private = mp_bh;
325 			submit_bio_noacct(bio);
326 		}
327 	}
328 	spin_unlock_irqrestore(&conf->device_lock, flags);
329 }
330 
331 static sector_t multipath_size(struct mddev *mddev, sector_t sectors, int raid_disks)
332 {
333 	WARN_ONCE(sectors || raid_disks,
334 		  "%s does not support generic reshape\n", __func__);
335 
336 	return mddev->dev_sectors;
337 }
338 
339 static int multipath_run (struct mddev *mddev)
340 {
341 	struct mpconf *conf;
342 	int disk_idx;
343 	struct multipath_info *disk;
344 	struct md_rdev *rdev;
345 	int working_disks;
346 	int ret;
347 
348 	if (md_check_no_bitmap(mddev))
349 		return -EINVAL;
350 
351 	if (mddev->level != LEVEL_MULTIPATH) {
352 		pr_warn("multipath: %s: raid level not set to multipath IO (%d)\n",
353 			mdname(mddev), mddev->level);
354 		goto out;
355 	}
356 	/*
357 	 * copy the already verified devices into our private MULTIPATH
358 	 * bookkeeping area. [whatever we allocate in multipath_run(),
359 	 * should be freed in multipath_free()]
360 	 */
361 
362 	conf = kzalloc(sizeof(struct mpconf), GFP_KERNEL);
363 	mddev->private = conf;
364 	if (!conf)
365 		goto out;
366 
367 	conf->multipaths = kcalloc(mddev->raid_disks,
368 				   sizeof(struct multipath_info),
369 				   GFP_KERNEL);
370 	if (!conf->multipaths)
371 		goto out_free_conf;
372 
373 	working_disks = 0;
374 	rdev_for_each(rdev, mddev) {
375 		disk_idx = rdev->raid_disk;
376 		if (disk_idx < 0 ||
377 		    disk_idx >= mddev->raid_disks)
378 			continue;
379 
380 		disk = conf->multipaths + disk_idx;
381 		disk->rdev = rdev;
382 		disk_stack_limits(mddev->gendisk, rdev->bdev,
383 				  rdev->data_offset << 9);
384 
385 		if (!test_bit(Faulty, &rdev->flags))
386 			working_disks++;
387 	}
388 
389 	conf->raid_disks = mddev->raid_disks;
390 	conf->mddev = mddev;
391 	spin_lock_init(&conf->device_lock);
392 	INIT_LIST_HEAD(&conf->retry_list);
393 
394 	if (!working_disks) {
395 		pr_warn("multipath: no operational IO paths for %s\n",
396 			mdname(mddev));
397 		goto out_free_conf;
398 	}
399 	mddev->degraded = conf->raid_disks - working_disks;
400 
401 	ret = mempool_init_kmalloc_pool(&conf->pool, NR_RESERVED_BUFS,
402 					sizeof(struct multipath_bh));
403 	if (ret)
404 		goto out_free_conf;
405 
406 	mddev->thread = md_register_thread(multipathd, mddev,
407 					   "multipath");
408 	if (!mddev->thread)
409 		goto out_free_conf;
410 
411 	pr_info("multipath: array %s active with %d out of %d IO paths\n",
412 		mdname(mddev), conf->raid_disks - mddev->degraded,
413 		mddev->raid_disks);
414 	/*
415 	 * Ok, everything is just fine now
416 	 */
417 	md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
418 
419 	if (md_integrity_register(mddev))
420 		goto out_free_conf;
421 
422 	return 0;
423 
424 out_free_conf:
425 	mempool_exit(&conf->pool);
426 	kfree(conf->multipaths);
427 	kfree(conf);
428 	mddev->private = NULL;
429 out:
430 	return -EIO;
431 }
432 
433 static void multipath_free(struct mddev *mddev, void *priv)
434 {
435 	struct mpconf *conf = priv;
436 
437 	mempool_exit(&conf->pool);
438 	kfree(conf->multipaths);
439 	kfree(conf);
440 }
441 
442 static struct md_personality multipath_personality =
443 {
444 	.name		= "multipath",
445 	.level		= LEVEL_MULTIPATH,
446 	.owner		= THIS_MODULE,
447 	.make_request	= multipath_make_request,
448 	.run		= multipath_run,
449 	.free		= multipath_free,
450 	.status		= multipath_status,
451 	.error_handler	= multipath_error,
452 	.hot_add_disk	= multipath_add_disk,
453 	.hot_remove_disk= multipath_remove_disk,
454 	.size		= multipath_size,
455 };
456 
457 static int __init multipath_init (void)
458 {
459 	return register_md_personality (&multipath_personality);
460 }
461 
462 static void __exit multipath_exit (void)
463 {
464 	unregister_md_personality (&multipath_personality);
465 }
466 
467 module_init(multipath_init);
468 module_exit(multipath_exit);
469 MODULE_LICENSE("GPL");
470 MODULE_DESCRIPTION("simple multi-path personality for MD (deprecated)");
471 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
472 MODULE_ALIAS("md-multipath");
473 MODULE_ALIAS("md-level--4");
474