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
26static 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(v: &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
51static 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(new: &mp_bh->retry_list, head: &conf->retry_list);
59 spin_unlock_irqrestore(lock: &conf->device_lock, flags);
60 md_wakeup_thread(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 */
68static 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(element: mp_bh, pool: &conf->pool);
76}
77
78static 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, status: 0);
86 else if (!(bio->bi_opf & REQ_RAHEAD)) {
87 /*
88 * oops, IO error:
89 */
90 md_error (mddev: 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, status: bio->bi_status);
97 rdev_dec_pending(rdev, mddev: conf->mddev);
98}
99
100static 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 md_account_bio(mddev, bio: &bio);
111 mp_bh = mempool_alloc(pool: &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(element: mp_bh, pool: &conf->pool);
120 return true;
121 }
122 multipath = conf->multipaths + mp_bh->path;
123
124 bio_init_clone(bdev: multipath->rdev->bdev, bio: &mp_bh->bio, bio_src: 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, bio: &mp_bh->bio);
131 submit_bio_noacct(bio: &mp_bh->bio);
132 return true;
133}
134
135static void multipath_status(struct seq_file *seq, struct mddev *mddev)
136{
137 struct mpconf *conf = mddev->private;
138 int i;
139
140 seq_printf (m: seq, fmt: " [%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 (m: seq, fmt: "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
146 }
147 rcu_read_unlock();
148 seq_putc(m: seq, c: ']');
149}
150
151/*
152 * Careful, this can execute in IRQ contexts as well!
153 */
154static void multipath_error (struct mddev *mddev, struct md_rdev *rdev)
155{
156 struct mpconf *conf = mddev->private;
157
158 if (conf->raid_disks - mddev->degraded <= 1) {
159 /*
160 * Uh oh, we can do nothing if this is our last path, but
161 * first check if this is a queued request for a device
162 * which has just failed.
163 */
164 pr_warn("multipath: only one IO path left and IO error.\n");
165 /* leave it active... it's all we have */
166 return;
167 }
168 /*
169 * Mark disk as unusable
170 */
171 if (test_and_clear_bit(nr: In_sync, addr: &rdev->flags)) {
172 unsigned long flags;
173 spin_lock_irqsave(&conf->device_lock, flags);
174 mddev->degraded++;
175 spin_unlock_irqrestore(lock: &conf->device_lock, flags);
176 }
177 set_bit(nr: Faulty, addr: &rdev->flags);
178 set_bit(nr: MD_SB_CHANGE_DEVS, addr: &mddev->sb_flags);
179 pr_err("multipath: IO failure on %pg, disabling IO path.\n"
180 "multipath: Operation continuing on %d IO paths.\n",
181 rdev->bdev,
182 conf->raid_disks - mddev->degraded);
183}
184
185static void print_multipath_conf (struct mpconf *conf)
186{
187 int i;
188 struct multipath_info *tmp;
189
190 pr_debug("MULTIPATH conf printout:\n");
191 if (!conf) {
192 pr_debug("(conf==NULL)\n");
193 return;
194 }
195 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
196 conf->raid_disks);
197
198 for (i = 0; i < conf->raid_disks; i++) {
199 tmp = conf->multipaths + i;
200 if (tmp->rdev)
201 pr_debug(" disk%d, o:%d, dev:%pg\n",
202 i,!test_bit(Faulty, &tmp->rdev->flags),
203 tmp->rdev->bdev);
204 }
205}
206
207static int multipath_add_disk(struct mddev *mddev, struct md_rdev *rdev)
208{
209 struct mpconf *conf = mddev->private;
210 int err = -EEXIST;
211 int path;
212 struct multipath_info *p;
213 int first = 0;
214 int last = mddev->raid_disks - 1;
215
216 if (rdev->raid_disk >= 0)
217 first = last = rdev->raid_disk;
218
219 print_multipath_conf(conf);
220
221 for (path = first; path <= last; path++)
222 if ((p=conf->multipaths+path)->rdev == NULL) {
223 disk_stack_limits(disk: mddev->gendisk, bdev: rdev->bdev,
224 offset: rdev->data_offset << 9);
225
226 err = md_integrity_add_rdev(rdev, mddev);
227 if (err)
228 break;
229 spin_lock_irq(lock: &conf->device_lock);
230 mddev->degraded--;
231 rdev->raid_disk = path;
232 set_bit(nr: In_sync, addr: &rdev->flags);
233 spin_unlock_irq(lock: &conf->device_lock);
234 rcu_assign_pointer(p->rdev, rdev);
235 err = 0;
236 break;
237 }
238
239 print_multipath_conf(conf);
240
241 return err;
242}
243
244static int multipath_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
245{
246 struct mpconf *conf = mddev->private;
247 int err = 0;
248 int number = rdev->raid_disk;
249 struct multipath_info *p = conf->multipaths + number;
250
251 print_multipath_conf(conf);
252
253 if (rdev == p->rdev) {
254 if (test_bit(In_sync, &rdev->flags) ||
255 atomic_read(v: &rdev->nr_pending)) {
256 pr_warn("hot-remove-disk, slot %d is identified but is still operational!\n", number);
257 err = -EBUSY;
258 goto abort;
259 }
260 p->rdev = NULL;
261 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
262 synchronize_rcu();
263 if (atomic_read(v: &rdev->nr_pending)) {
264 /* lost the race, try later */
265 err = -EBUSY;
266 p->rdev = rdev;
267 goto abort;
268 }
269 }
270 err = md_integrity_register(mddev);
271 }
272abort:
273
274 print_multipath_conf(conf);
275 return err;
276}
277
278/*
279 * This is a kernel thread which:
280 *
281 * 1. Retries failed read operations on working multipaths.
282 * 2. Updates the raid superblock when problems encounter.
283 * 3. Performs writes following reads for array syncronising.
284 */
285
286static void multipathd(struct md_thread *thread)
287{
288 struct mddev *mddev = thread->mddev;
289 struct multipath_bh *mp_bh;
290 struct bio *bio;
291 unsigned long flags;
292 struct mpconf *conf = mddev->private;
293 struct list_head *head = &conf->retry_list;
294
295 md_check_recovery(mddev);
296 for (;;) {
297 spin_lock_irqsave(&conf->device_lock, flags);
298 if (list_empty(head))
299 break;
300 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
301 list_del(entry: head->prev);
302 spin_unlock_irqrestore(lock: &conf->device_lock, flags);
303
304 bio = &mp_bh->bio;
305 bio->bi_iter.bi_sector = mp_bh->master_bio->bi_iter.bi_sector;
306
307 if ((mp_bh->path = multipath_map (conf))<0) {
308 pr_err("multipath: %pg: unrecoverable IO read error for block %llu\n",
309 bio->bi_bdev,
310 (unsigned long long)bio->bi_iter.bi_sector);
311 multipath_end_bh_io(mp_bh, BLK_STS_IOERR);
312 } else {
313 pr_err("multipath: %pg: redirecting sector %llu to another IO path\n",
314 bio->bi_bdev,
315 (unsigned long long)bio->bi_iter.bi_sector);
316 *bio = *(mp_bh->master_bio);
317 bio->bi_iter.bi_sector +=
318 conf->multipaths[mp_bh->path].rdev->data_offset;
319 bio_set_dev(bio, bdev: conf->multipaths[mp_bh->path].rdev->bdev);
320 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
321 bio->bi_end_io = multipath_end_request;
322 bio->bi_private = mp_bh;
323 submit_bio_noacct(bio);
324 }
325 }
326 spin_unlock_irqrestore(lock: &conf->device_lock, flags);
327}
328
329static sector_t multipath_size(struct mddev *mddev, sector_t sectors, int raid_disks)
330{
331 WARN_ONCE(sectors || raid_disks,
332 "%s does not support generic reshape\n", __func__);
333
334 return mddev->dev_sectors;
335}
336
337static int multipath_run (struct mddev *mddev)
338{
339 struct mpconf *conf;
340 int disk_idx;
341 struct multipath_info *disk;
342 struct md_rdev *rdev;
343 int working_disks;
344 int ret;
345
346 if (md_check_no_bitmap(mddev))
347 return -EINVAL;
348
349 if (mddev->level != LEVEL_MULTIPATH) {
350 pr_warn("multipath: %s: raid level not set to multipath IO (%d)\n",
351 mdname(mddev), mddev->level);
352 goto out;
353 }
354 /*
355 * copy the already verified devices into our private MULTIPATH
356 * bookkeeping area. [whatever we allocate in multipath_run(),
357 * should be freed in multipath_free()]
358 */
359
360 conf = kzalloc(size: sizeof(struct mpconf), GFP_KERNEL);
361 mddev->private = conf;
362 if (!conf)
363 goto out;
364
365 conf->multipaths = kcalloc(n: mddev->raid_disks,
366 size: sizeof(struct multipath_info),
367 GFP_KERNEL);
368 if (!conf->multipaths)
369 goto out_free_conf;
370
371 working_disks = 0;
372 rdev_for_each(rdev, mddev) {
373 disk_idx = rdev->raid_disk;
374 if (disk_idx < 0 ||
375 disk_idx >= mddev->raid_disks)
376 continue;
377
378 disk = conf->multipaths + disk_idx;
379 disk->rdev = rdev;
380 disk_stack_limits(disk: mddev->gendisk, bdev: rdev->bdev,
381 offset: rdev->data_offset << 9);
382
383 if (!test_bit(Faulty, &rdev->flags))
384 working_disks++;
385 }
386
387 conf->raid_disks = mddev->raid_disks;
388 conf->mddev = mddev;
389 spin_lock_init(&conf->device_lock);
390 INIT_LIST_HEAD(list: &conf->retry_list);
391
392 if (!working_disks) {
393 pr_warn("multipath: no operational IO paths for %s\n",
394 mdname(mddev));
395 goto out_free_conf;
396 }
397 mddev->degraded = conf->raid_disks - working_disks;
398
399 ret = mempool_init_kmalloc_pool(pool: &conf->pool, NR_RESERVED_BUFS,
400 size: sizeof(struct multipath_bh));
401 if (ret)
402 goto out_free_conf;
403
404 rcu_assign_pointer(mddev->thread,
405 md_register_thread(multipathd, mddev, "multipath"));
406 if (!mddev->thread)
407 goto out_free_conf;
408
409 pr_info("multipath: array %s active with %d out of %d IO paths\n",
410 mdname(mddev), conf->raid_disks - mddev->degraded,
411 mddev->raid_disks);
412 /*
413 * Ok, everything is just fine now
414 */
415 md_set_array_sectors(mddev, array_sectors: multipath_size(mddev, sectors: 0, raid_disks: 0));
416
417 if (md_integrity_register(mddev))
418 goto out_free_conf;
419
420 return 0;
421
422out_free_conf:
423 mempool_exit(pool: &conf->pool);
424 kfree(objp: conf->multipaths);
425 kfree(objp: conf);
426 mddev->private = NULL;
427out:
428 return -EIO;
429}
430
431static void multipath_free(struct mddev *mddev, void *priv)
432{
433 struct mpconf *conf = priv;
434
435 mempool_exit(pool: &conf->pool);
436 kfree(objp: conf->multipaths);
437 kfree(objp: conf);
438}
439
440static struct md_personality multipath_personality =
441{
442 .name = "multipath",
443 .level = LEVEL_MULTIPATH,
444 .owner = THIS_MODULE,
445 .make_request = multipath_make_request,
446 .run = multipath_run,
447 .free = multipath_free,
448 .status = multipath_status,
449 .error_handler = multipath_error,
450 .hot_add_disk = multipath_add_disk,
451 .hot_remove_disk= multipath_remove_disk,
452 .size = multipath_size,
453};
454
455static int __init multipath_init (void)
456{
457 return register_md_personality (p: &multipath_personality);
458}
459
460static void __exit multipath_exit (void)
461{
462 unregister_md_personality (p: &multipath_personality);
463}
464
465module_init(multipath_init);
466module_exit(multipath_exit);
467MODULE_LICENSE("GPL");
468MODULE_DESCRIPTION("simple multi-path personality for MD (deprecated)");
469MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
470MODULE_ALIAS("md-multipath");
471MODULE_ALIAS("md-level--4");
472

source code of linux/drivers/md/md-multipath.c