xref: /openbmc/linux/drivers/md/md-multipath.c (revision ac3cb6de)
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 		md_error (mp_bh->mddev, rdev);
91 		pr_info("multipath: %pg: rescheduling sector %llu\n",
92 			rdev->bdev,
93 			(unsigned long long)bio->bi_iter.bi_sector);
94 		multipath_reschedule_retry(mp_bh);
95 	} else
96 		multipath_end_bh_io(mp_bh, bio->bi_status);
97 	rdev_dec_pending(rdev, conf->mddev);
98 }
99 
100 static bool multipath_make_request(struct mddev *mddev, struct bio * bio)
101 {
102 	struct mpconf *conf = mddev->private;
103 	struct multipath_bh * mp_bh;
104 	struct multipath_info *multipath;
105 
106 	if (unlikely(bio->bi_opf & REQ_PREFLUSH)
107 	    && md_flush_request(mddev, bio))
108 		return true;
109 
110 	mp_bh = mempool_alloc(&conf->pool, GFP_NOIO);
111 
112 	mp_bh->master_bio = bio;
113 	mp_bh->mddev = mddev;
114 
115 	mp_bh->path = multipath_map(conf);
116 	if (mp_bh->path < 0) {
117 		bio_io_error(bio);
118 		mempool_free(mp_bh, &conf->pool);
119 		return true;
120 	}
121 	multipath = conf->multipaths + mp_bh->path;
122 
123 	bio_init_clone(multipath->rdev->bdev, &mp_bh->bio, bio, GFP_NOIO);
124 
125 	mp_bh->bio.bi_iter.bi_sector += multipath->rdev->data_offset;
126 	mp_bh->bio.bi_opf |= REQ_FAILFAST_TRANSPORT;
127 	mp_bh->bio.bi_end_io = multipath_end_request;
128 	mp_bh->bio.bi_private = mp_bh;
129 	mddev_check_write_zeroes(mddev, &mp_bh->bio);
130 	submit_bio_noacct(&mp_bh->bio);
131 	return true;
132 }
133 
134 static void multipath_status(struct seq_file *seq, struct mddev *mddev)
135 {
136 	struct mpconf *conf = mddev->private;
137 	int i;
138 
139 	seq_printf (seq, " [%d/%d] [", conf->raid_disks,
140 		    conf->raid_disks - mddev->degraded);
141 	rcu_read_lock();
142 	for (i = 0; i < conf->raid_disks; i++) {
143 		struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
144 		seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
145 	}
146 	rcu_read_unlock();
147 	seq_putc(seq, ']');
148 }
149 
150 /*
151  * Careful, this can execute in IRQ contexts as well!
152  */
153 static void multipath_error (struct mddev *mddev, struct md_rdev *rdev)
154 {
155 	struct mpconf *conf = mddev->private;
156 
157 	if (conf->raid_disks - mddev->degraded <= 1) {
158 		/*
159 		 * Uh oh, we can do nothing if this is our last path, but
160 		 * first check if this is a queued request for a device
161 		 * which has just failed.
162 		 */
163 		pr_warn("multipath: only one IO path left and IO error.\n");
164 		/* leave it active... it's all we have */
165 		return;
166 	}
167 	/*
168 	 * Mark disk as unusable
169 	 */
170 	if (test_and_clear_bit(In_sync, &rdev->flags)) {
171 		unsigned long flags;
172 		spin_lock_irqsave(&conf->device_lock, flags);
173 		mddev->degraded++;
174 		spin_unlock_irqrestore(&conf->device_lock, flags);
175 	}
176 	set_bit(Faulty, &rdev->flags);
177 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
178 	pr_err("multipath: IO failure on %pg, disabling IO path.\n"
179 	       "multipath: Operation continuing on %d IO paths.\n",
180 	       rdev->bdev,
181 	       conf->raid_disks - mddev->degraded);
182 }
183 
184 static void print_multipath_conf (struct mpconf *conf)
185 {
186 	int i;
187 	struct multipath_info *tmp;
188 
189 	pr_debug("MULTIPATH conf printout:\n");
190 	if (!conf) {
191 		pr_debug("(conf==NULL)\n");
192 		return;
193 	}
194 	pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
195 		 conf->raid_disks);
196 
197 	for (i = 0; i < conf->raid_disks; i++) {
198 		tmp = conf->multipaths + i;
199 		if (tmp->rdev)
200 			pr_debug(" disk%d, o:%d, dev:%pg\n",
201 				 i,!test_bit(Faulty, &tmp->rdev->flags),
202 				 tmp->rdev->bdev);
203 	}
204 }
205 
206 static int multipath_add_disk(struct mddev *mddev, struct md_rdev *rdev)
207 {
208 	struct mpconf *conf = mddev->private;
209 	int err = -EEXIST;
210 	int path;
211 	struct multipath_info *p;
212 	int first = 0;
213 	int last = mddev->raid_disks - 1;
214 
215 	if (rdev->raid_disk >= 0)
216 		first = last = rdev->raid_disk;
217 
218 	print_multipath_conf(conf);
219 
220 	for (path = first; path <= last; path++)
221 		if ((p=conf->multipaths+path)->rdev == NULL) {
222 			disk_stack_limits(mddev->gendisk, rdev->bdev,
223 					  rdev->data_offset << 9);
224 
225 			err = md_integrity_add_rdev(rdev, mddev);
226 			if (err)
227 				break;
228 			spin_lock_irq(&conf->device_lock);
229 			mddev->degraded--;
230 			rdev->raid_disk = path;
231 			set_bit(In_sync, &rdev->flags);
232 			spin_unlock_irq(&conf->device_lock);
233 			rcu_assign_pointer(p->rdev, rdev);
234 			err = 0;
235 			break;
236 		}
237 
238 	print_multipath_conf(conf);
239 
240 	return err;
241 }
242 
243 static int multipath_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
244 {
245 	struct mpconf *conf = mddev->private;
246 	int err = 0;
247 	int number = rdev->raid_disk;
248 	struct multipath_info *p = conf->multipaths + number;
249 
250 	print_multipath_conf(conf);
251 
252 	if (rdev == p->rdev) {
253 		if (test_bit(In_sync, &rdev->flags) ||
254 		    atomic_read(&rdev->nr_pending)) {
255 			pr_warn("hot-remove-disk, slot %d is identified but is still operational!\n", number);
256 			err = -EBUSY;
257 			goto abort;
258 		}
259 		p->rdev = NULL;
260 		if (!test_bit(RemoveSynchronized, &rdev->flags)) {
261 			synchronize_rcu();
262 			if (atomic_read(&rdev->nr_pending)) {
263 				/* lost the race, try later */
264 				err = -EBUSY;
265 				p->rdev = rdev;
266 				goto abort;
267 			}
268 		}
269 		err = md_integrity_register(mddev);
270 	}
271 abort:
272 
273 	print_multipath_conf(conf);
274 	return err;
275 }
276 
277 /*
278  * This is a kernel thread which:
279  *
280  *	1.	Retries failed read operations on working multipaths.
281  *	2.	Updates the raid superblock when problems encounter.
282  *	3.	Performs writes following reads for array syncronising.
283  */
284 
285 static void multipathd(struct md_thread *thread)
286 {
287 	struct mddev *mddev = thread->mddev;
288 	struct multipath_bh *mp_bh;
289 	struct bio *bio;
290 	unsigned long flags;
291 	struct mpconf *conf = mddev->private;
292 	struct list_head *head = &conf->retry_list;
293 
294 	md_check_recovery(mddev);
295 	for (;;) {
296 		spin_lock_irqsave(&conf->device_lock, flags);
297 		if (list_empty(head))
298 			break;
299 		mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
300 		list_del(head->prev);
301 		spin_unlock_irqrestore(&conf->device_lock, flags);
302 
303 		bio = &mp_bh->bio;
304 		bio->bi_iter.bi_sector = mp_bh->master_bio->bi_iter.bi_sector;
305 
306 		if ((mp_bh->path = multipath_map (conf))<0) {
307 			pr_err("multipath: %pg: unrecoverable IO read error for block %llu\n",
308 			       bio->bi_bdev,
309 			       (unsigned long long)bio->bi_iter.bi_sector);
310 			multipath_end_bh_io(mp_bh, BLK_STS_IOERR);
311 		} else {
312 			pr_err("multipath: %pg: redirecting sector %llu to another IO path\n",
313 			       bio->bi_bdev,
314 			       (unsigned long long)bio->bi_iter.bi_sector);
315 			*bio = *(mp_bh->master_bio);
316 			bio->bi_iter.bi_sector +=
317 				conf->multipaths[mp_bh->path].rdev->data_offset;
318 			bio_set_dev(bio, conf->multipaths[mp_bh->path].rdev->bdev);
319 			bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
320 			bio->bi_end_io = multipath_end_request;
321 			bio->bi_private = mp_bh;
322 			submit_bio_noacct(bio);
323 		}
324 	}
325 	spin_unlock_irqrestore(&conf->device_lock, flags);
326 }
327 
328 static sector_t multipath_size(struct mddev *mddev, sector_t sectors, int raid_disks)
329 {
330 	WARN_ONCE(sectors || raid_disks,
331 		  "%s does not support generic reshape\n", __func__);
332 
333 	return mddev->dev_sectors;
334 }
335 
336 static int multipath_run (struct mddev *mddev)
337 {
338 	struct mpconf *conf;
339 	int disk_idx;
340 	struct multipath_info *disk;
341 	struct md_rdev *rdev;
342 	int working_disks;
343 	int ret;
344 
345 	if (md_check_no_bitmap(mddev))
346 		return -EINVAL;
347 
348 	if (mddev->level != LEVEL_MULTIPATH) {
349 		pr_warn("multipath: %s: raid level not set to multipath IO (%d)\n",
350 			mdname(mddev), mddev->level);
351 		goto out;
352 	}
353 	/*
354 	 * copy the already verified devices into our private MULTIPATH
355 	 * bookkeeping area. [whatever we allocate in multipath_run(),
356 	 * should be freed in multipath_free()]
357 	 */
358 
359 	conf = kzalloc(sizeof(struct mpconf), GFP_KERNEL);
360 	mddev->private = conf;
361 	if (!conf)
362 		goto out;
363 
364 	conf->multipaths = kcalloc(mddev->raid_disks,
365 				   sizeof(struct multipath_info),
366 				   GFP_KERNEL);
367 	if (!conf->multipaths)
368 		goto out_free_conf;
369 
370 	working_disks = 0;
371 	rdev_for_each(rdev, mddev) {
372 		disk_idx = rdev->raid_disk;
373 		if (disk_idx < 0 ||
374 		    disk_idx >= mddev->raid_disks)
375 			continue;
376 
377 		disk = conf->multipaths + disk_idx;
378 		disk->rdev = rdev;
379 		disk_stack_limits(mddev->gendisk, rdev->bdev,
380 				  rdev->data_offset << 9);
381 
382 		if (!test_bit(Faulty, &rdev->flags))
383 			working_disks++;
384 	}
385 
386 	conf->raid_disks = mddev->raid_disks;
387 	conf->mddev = mddev;
388 	spin_lock_init(&conf->device_lock);
389 	INIT_LIST_HEAD(&conf->retry_list);
390 
391 	if (!working_disks) {
392 		pr_warn("multipath: no operational IO paths for %s\n",
393 			mdname(mddev));
394 		goto out_free_conf;
395 	}
396 	mddev->degraded = conf->raid_disks - working_disks;
397 
398 	ret = mempool_init_kmalloc_pool(&conf->pool, NR_RESERVED_BUFS,
399 					sizeof(struct multipath_bh));
400 	if (ret)
401 		goto out_free_conf;
402 
403 	rcu_assign_pointer(mddev->thread,
404 			   md_register_thread(multipathd, mddev, "multipath"));
405 	if (!mddev->thread)
406 		goto out_free_conf;
407 
408 	pr_info("multipath: array %s active with %d out of %d IO paths\n",
409 		mdname(mddev), conf->raid_disks - mddev->degraded,
410 		mddev->raid_disks);
411 	/*
412 	 * Ok, everything is just fine now
413 	 */
414 	md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
415 
416 	if (md_integrity_register(mddev))
417 		goto out_free_conf;
418 
419 	return 0;
420 
421 out_free_conf:
422 	mempool_exit(&conf->pool);
423 	kfree(conf->multipaths);
424 	kfree(conf);
425 	mddev->private = NULL;
426 out:
427 	return -EIO;
428 }
429 
430 static void multipath_free(struct mddev *mddev, void *priv)
431 {
432 	struct mpconf *conf = priv;
433 
434 	mempool_exit(&conf->pool);
435 	kfree(conf->multipaths);
436 	kfree(conf);
437 }
438 
439 static struct md_personality multipath_personality =
440 {
441 	.name		= "multipath",
442 	.level		= LEVEL_MULTIPATH,
443 	.owner		= THIS_MODULE,
444 	.make_request	= multipath_make_request,
445 	.run		= multipath_run,
446 	.free		= multipath_free,
447 	.status		= multipath_status,
448 	.error_handler	= multipath_error,
449 	.hot_add_disk	= multipath_add_disk,
450 	.hot_remove_disk= multipath_remove_disk,
451 	.size		= multipath_size,
452 };
453 
454 static int __init multipath_init (void)
455 {
456 	return register_md_personality (&multipath_personality);
457 }
458 
459 static void __exit multipath_exit (void)
460 {
461 	unregister_md_personality (&multipath_personality);
462 }
463 
464 module_init(multipath_init);
465 module_exit(multipath_exit);
466 MODULE_LICENSE("GPL");
467 MODULE_DESCRIPTION("simple multi-path personality for MD (deprecated)");
468 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
469 MODULE_ALIAS("md-multipath");
470 MODULE_ALIAS("md-level--4");
471