xref: /openbmc/linux/drivers/md/md-multipath.c (revision 4eb5928d)
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(&mp_bh->bio, NULL, 0);
125 	__bio_clone_fast(&mp_bh->bio, bio);
126 
127 	mp_bh->bio.bi_iter.bi_sector += multipath->rdev->data_offset;
128 	bio_set_dev(&mp_bh->bio, multipath->rdev->bdev);
129 	mp_bh->bio.bi_opf |= REQ_FAILFAST_TRANSPORT;
130 	mp_bh->bio.bi_end_io = multipath_end_request;
131 	mp_bh->bio.bi_private = mp_bh;
132 	mddev_check_writesame(mddev, &mp_bh->bio);
133 	mddev_check_write_zeroes(mddev, &mp_bh->bio);
134 	submit_bio_noacct(&mp_bh->bio);
135 	return true;
136 }
137 
138 static void multipath_status(struct seq_file *seq, struct mddev *mddev)
139 {
140 	struct mpconf *conf = mddev->private;
141 	int i;
142 
143 	seq_printf (seq, " [%d/%d] [", conf->raid_disks,
144 		    conf->raid_disks - mddev->degraded);
145 	rcu_read_lock();
146 	for (i = 0; i < conf->raid_disks; i++) {
147 		struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
148 		seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
149 	}
150 	rcu_read_unlock();
151 	seq_putc(seq, ']');
152 }
153 
154 /*
155  * Careful, this can execute in IRQ contexts as well!
156  */
157 static void multipath_error (struct mddev *mddev, struct md_rdev *rdev)
158 {
159 	struct mpconf *conf = mddev->private;
160 	char b[BDEVNAME_SIZE];
161 
162 	if (conf->raid_disks - mddev->degraded <= 1) {
163 		/*
164 		 * Uh oh, we can do nothing if this is our last path, but
165 		 * first check if this is a queued request for a device
166 		 * which has just failed.
167 		 */
168 		pr_warn("multipath: only one IO path left and IO error.\n");
169 		/* leave it active... it's all we have */
170 		return;
171 	}
172 	/*
173 	 * Mark disk as unusable
174 	 */
175 	if (test_and_clear_bit(In_sync, &rdev->flags)) {
176 		unsigned long flags;
177 		spin_lock_irqsave(&conf->device_lock, flags);
178 		mddev->degraded++;
179 		spin_unlock_irqrestore(&conf->device_lock, flags);
180 	}
181 	set_bit(Faulty, &rdev->flags);
182 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
183 	pr_err("multipath: IO failure on %s, disabling IO path.\n"
184 	       "multipath: Operation continuing on %d IO paths.\n",
185 	       bdevname(rdev->bdev, b),
186 	       conf->raid_disks - mddev->degraded);
187 }
188 
189 static void print_multipath_conf (struct mpconf *conf)
190 {
191 	int i;
192 	struct multipath_info *tmp;
193 
194 	pr_debug("MULTIPATH conf printout:\n");
195 	if (!conf) {
196 		pr_debug("(conf==NULL)\n");
197 		return;
198 	}
199 	pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
200 		 conf->raid_disks);
201 
202 	for (i = 0; i < conf->raid_disks; i++) {
203 		char b[BDEVNAME_SIZE];
204 		tmp = conf->multipaths + i;
205 		if (tmp->rdev)
206 			pr_debug(" disk%d, o:%d, dev:%s\n",
207 				 i,!test_bit(Faulty, &tmp->rdev->flags),
208 				 bdevname(tmp->rdev->bdev,b));
209 	}
210 }
211 
212 static int multipath_add_disk(struct mddev *mddev, struct md_rdev *rdev)
213 {
214 	struct mpconf *conf = mddev->private;
215 	int err = -EEXIST;
216 	int path;
217 	struct multipath_info *p;
218 	int first = 0;
219 	int last = mddev->raid_disks - 1;
220 
221 	if (rdev->raid_disk >= 0)
222 		first = last = rdev->raid_disk;
223 
224 	print_multipath_conf(conf);
225 
226 	for (path = first; path <= last; path++)
227 		if ((p=conf->multipaths+path)->rdev == NULL) {
228 			disk_stack_limits(mddev->gendisk, rdev->bdev,
229 					  rdev->data_offset << 9);
230 
231 			err = md_integrity_add_rdev(rdev, mddev);
232 			if (err)
233 				break;
234 			spin_lock_irq(&conf->device_lock);
235 			mddev->degraded--;
236 			rdev->raid_disk = path;
237 			set_bit(In_sync, &rdev->flags);
238 			spin_unlock_irq(&conf->device_lock);
239 			rcu_assign_pointer(p->rdev, rdev);
240 			err = 0;
241 			break;
242 		}
243 
244 	print_multipath_conf(conf);
245 
246 	return err;
247 }
248 
249 static int multipath_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
250 {
251 	struct mpconf *conf = mddev->private;
252 	int err = 0;
253 	int number = rdev->raid_disk;
254 	struct multipath_info *p = conf->multipaths + number;
255 
256 	print_multipath_conf(conf);
257 
258 	if (rdev == p->rdev) {
259 		if (test_bit(In_sync, &rdev->flags) ||
260 		    atomic_read(&rdev->nr_pending)) {
261 			pr_warn("hot-remove-disk, slot %d is identified but is still operational!\n", number);
262 			err = -EBUSY;
263 			goto abort;
264 		}
265 		p->rdev = NULL;
266 		if (!test_bit(RemoveSynchronized, &rdev->flags)) {
267 			synchronize_rcu();
268 			if (atomic_read(&rdev->nr_pending)) {
269 				/* lost the race, try later */
270 				err = -EBUSY;
271 				p->rdev = rdev;
272 				goto abort;
273 			}
274 		}
275 		err = md_integrity_register(mddev);
276 	}
277 abort:
278 
279 	print_multipath_conf(conf);
280 	return err;
281 }
282 
283 /*
284  * This is a kernel thread which:
285  *
286  *	1.	Retries failed read operations on working multipaths.
287  *	2.	Updates the raid superblock when problems encounter.
288  *	3.	Performs writes following reads for array syncronising.
289  */
290 
291 static void multipathd(struct md_thread *thread)
292 {
293 	struct mddev *mddev = thread->mddev;
294 	struct multipath_bh *mp_bh;
295 	struct bio *bio;
296 	unsigned long flags;
297 	struct mpconf *conf = mddev->private;
298 	struct list_head *head = &conf->retry_list;
299 
300 	md_check_recovery(mddev);
301 	for (;;) {
302 		char b[BDEVNAME_SIZE];
303 		spin_lock_irqsave(&conf->device_lock, flags);
304 		if (list_empty(head))
305 			break;
306 		mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
307 		list_del(head->prev);
308 		spin_unlock_irqrestore(&conf->device_lock, flags);
309 
310 		bio = &mp_bh->bio;
311 		bio->bi_iter.bi_sector = mp_bh->master_bio->bi_iter.bi_sector;
312 
313 		if ((mp_bh->path = multipath_map (conf))<0) {
314 			pr_err("multipath: %s: unrecoverable IO read error for block %llu\n",
315 			       bio_devname(bio, b),
316 			       (unsigned long long)bio->bi_iter.bi_sector);
317 			multipath_end_bh_io(mp_bh, BLK_STS_IOERR);
318 		} else {
319 			pr_err("multipath: %s: redirecting sector %llu to another IO path\n",
320 			       bio_devname(bio, b),
321 			       (unsigned long long)bio->bi_iter.bi_sector);
322 			*bio = *(mp_bh->master_bio);
323 			bio->bi_iter.bi_sector +=
324 				conf->multipaths[mp_bh->path].rdev->data_offset;
325 			bio_set_dev(bio, conf->multipaths[mp_bh->path].rdev->bdev);
326 			bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
327 			bio->bi_end_io = multipath_end_request;
328 			bio->bi_private = mp_bh;
329 			submit_bio_noacct(bio);
330 		}
331 	}
332 	spin_unlock_irqrestore(&conf->device_lock, flags);
333 }
334 
335 static sector_t multipath_size(struct mddev *mddev, sector_t sectors, int raid_disks)
336 {
337 	WARN_ONCE(sectors || raid_disks,
338 		  "%s does not support generic reshape\n", __func__);
339 
340 	return mddev->dev_sectors;
341 }
342 
343 static int multipath_run (struct mddev *mddev)
344 {
345 	struct mpconf *conf;
346 	int disk_idx;
347 	struct multipath_info *disk;
348 	struct md_rdev *rdev;
349 	int working_disks;
350 	int ret;
351 
352 	if (md_check_no_bitmap(mddev))
353 		return -EINVAL;
354 
355 	if (mddev->level != LEVEL_MULTIPATH) {
356 		pr_warn("multipath: %s: raid level not set to multipath IO (%d)\n",
357 			mdname(mddev), mddev->level);
358 		goto out;
359 	}
360 	/*
361 	 * copy the already verified devices into our private MULTIPATH
362 	 * bookkeeping area. [whatever we allocate in multipath_run(),
363 	 * should be freed in multipath_free()]
364 	 */
365 
366 	conf = kzalloc(sizeof(struct mpconf), GFP_KERNEL);
367 	mddev->private = conf;
368 	if (!conf)
369 		goto out;
370 
371 	conf->multipaths = kcalloc(mddev->raid_disks,
372 				   sizeof(struct multipath_info),
373 				   GFP_KERNEL);
374 	if (!conf->multipaths)
375 		goto out_free_conf;
376 
377 	working_disks = 0;
378 	rdev_for_each(rdev, mddev) {
379 		disk_idx = rdev->raid_disk;
380 		if (disk_idx < 0 ||
381 		    disk_idx >= mddev->raid_disks)
382 			continue;
383 
384 		disk = conf->multipaths + disk_idx;
385 		disk->rdev = rdev;
386 		disk_stack_limits(mddev->gendisk, rdev->bdev,
387 				  rdev->data_offset << 9);
388 
389 		if (!test_bit(Faulty, &rdev->flags))
390 			working_disks++;
391 	}
392 
393 	conf->raid_disks = mddev->raid_disks;
394 	conf->mddev = mddev;
395 	spin_lock_init(&conf->device_lock);
396 	INIT_LIST_HEAD(&conf->retry_list);
397 
398 	if (!working_disks) {
399 		pr_warn("multipath: no operational IO paths for %s\n",
400 			mdname(mddev));
401 		goto out_free_conf;
402 	}
403 	mddev->degraded = conf->raid_disks - working_disks;
404 
405 	ret = mempool_init_kmalloc_pool(&conf->pool, NR_RESERVED_BUFS,
406 					sizeof(struct multipath_bh));
407 	if (ret)
408 		goto out_free_conf;
409 
410 	mddev->thread = md_register_thread(multipathd, mddev,
411 					   "multipath");
412 	if (!mddev->thread)
413 		goto out_free_conf;
414 
415 	pr_info("multipath: array %s active with %d out of %d IO paths\n",
416 		mdname(mddev), conf->raid_disks - mddev->degraded,
417 		mddev->raid_disks);
418 	/*
419 	 * Ok, everything is just fine now
420 	 */
421 	md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
422 
423 	if (md_integrity_register(mddev))
424 		goto out_free_conf;
425 
426 	return 0;
427 
428 out_free_conf:
429 	mempool_exit(&conf->pool);
430 	kfree(conf->multipaths);
431 	kfree(conf);
432 	mddev->private = NULL;
433 out:
434 	return -EIO;
435 }
436 
437 static void multipath_free(struct mddev *mddev, void *priv)
438 {
439 	struct mpconf *conf = priv;
440 
441 	mempool_exit(&conf->pool);
442 	kfree(conf->multipaths);
443 	kfree(conf);
444 }
445 
446 static struct md_personality multipath_personality =
447 {
448 	.name		= "multipath",
449 	.level		= LEVEL_MULTIPATH,
450 	.owner		= THIS_MODULE,
451 	.make_request	= multipath_make_request,
452 	.run		= multipath_run,
453 	.free		= multipath_free,
454 	.status		= multipath_status,
455 	.error_handler	= multipath_error,
456 	.hot_add_disk	= multipath_add_disk,
457 	.hot_remove_disk= multipath_remove_disk,
458 	.size		= multipath_size,
459 };
460 
461 static int __init multipath_init (void)
462 {
463 	return register_md_personality (&multipath_personality);
464 }
465 
466 static void __exit multipath_exit (void)
467 {
468 	unregister_md_personality (&multipath_personality);
469 }
470 
471 module_init(multipath_init);
472 module_exit(multipath_exit);
473 MODULE_LICENSE("GPL");
474 MODULE_DESCRIPTION("simple multi-path personality for MD");
475 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
476 MODULE_ALIAS("md-multipath");
477 MODULE_ALIAS("md-level--4");
478