Loading...
1/*
2 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm-core.h"
9#include "dm-rq.h"
10#include "dm-uevent.h"
11
12#include <linux/init.h>
13#include <linux/module.h>
14#include <linux/mutex.h>
15#include <linux/blkpg.h>
16#include <linux/bio.h>
17#include <linux/mempool.h>
18#include <linux/slab.h>
19#include <linux/idr.h>
20#include <linux/hdreg.h>
21#include <linux/delay.h>
22#include <linux/wait.h>
23#include <linux/pr.h>
24
25#define DM_MSG_PREFIX "core"
26
27#ifdef CONFIG_PRINTK
28/*
29 * ratelimit state to be used in DMXXX_LIMIT().
30 */
31DEFINE_RATELIMIT_STATE(dm_ratelimit_state,
32 DEFAULT_RATELIMIT_INTERVAL,
33 DEFAULT_RATELIMIT_BURST);
34EXPORT_SYMBOL(dm_ratelimit_state);
35#endif
36
37/*
38 * Cookies are numeric values sent with CHANGE and REMOVE
39 * uevents while resuming, removing or renaming the device.
40 */
41#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
42#define DM_COOKIE_LENGTH 24
43
44static const char *_name = DM_NAME;
45
46static unsigned int major = 0;
47static unsigned int _major = 0;
48
49static DEFINE_IDR(_minor_idr);
50
51static DEFINE_SPINLOCK(_minor_lock);
52
53static void do_deferred_remove(struct work_struct *w);
54
55static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
56
57static struct workqueue_struct *deferred_remove_workqueue;
58
59/*
60 * One of these is allocated per bio.
61 */
62struct dm_io {
63 struct mapped_device *md;
64 int error;
65 atomic_t io_count;
66 struct bio *bio;
67 unsigned long start_time;
68 spinlock_t endio_lock;
69 struct dm_stats_aux stats_aux;
70};
71
72#define MINOR_ALLOCED ((void *)-1)
73
74/*
75 * Bits for the md->flags field.
76 */
77#define DMF_BLOCK_IO_FOR_SUSPEND 0
78#define DMF_SUSPENDED 1
79#define DMF_FROZEN 2
80#define DMF_FREEING 3
81#define DMF_DELETING 4
82#define DMF_NOFLUSH_SUSPENDING 5
83#define DMF_DEFERRED_REMOVE 6
84#define DMF_SUSPENDED_INTERNALLY 7
85
86#define DM_NUMA_NODE NUMA_NO_NODE
87static int dm_numa_node = DM_NUMA_NODE;
88
89/*
90 * For mempools pre-allocation at the table loading time.
91 */
92struct dm_md_mempools {
93 mempool_t *io_pool;
94 mempool_t *rq_pool;
95 struct bio_set *bs;
96};
97
98struct table_device {
99 struct list_head list;
100 atomic_t count;
101 struct dm_dev dm_dev;
102};
103
104static struct kmem_cache *_io_cache;
105static struct kmem_cache *_rq_tio_cache;
106static struct kmem_cache *_rq_cache;
107
108/*
109 * Bio-based DM's mempools' reserved IOs set by the user.
110 */
111#define RESERVED_BIO_BASED_IOS 16
112static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
113
114static int __dm_get_module_param_int(int *module_param, int min, int max)
115{
116 int param = ACCESS_ONCE(*module_param);
117 int modified_param = 0;
118 bool modified = true;
119
120 if (param < min)
121 modified_param = min;
122 else if (param > max)
123 modified_param = max;
124 else
125 modified = false;
126
127 if (modified) {
128 (void)cmpxchg(module_param, param, modified_param);
129 param = modified_param;
130 }
131
132 return param;
133}
134
135unsigned __dm_get_module_param(unsigned *module_param,
136 unsigned def, unsigned max)
137{
138 unsigned param = ACCESS_ONCE(*module_param);
139 unsigned modified_param = 0;
140
141 if (!param)
142 modified_param = def;
143 else if (param > max)
144 modified_param = max;
145
146 if (modified_param) {
147 (void)cmpxchg(module_param, param, modified_param);
148 param = modified_param;
149 }
150
151 return param;
152}
153
154unsigned dm_get_reserved_bio_based_ios(void)
155{
156 return __dm_get_module_param(&reserved_bio_based_ios,
157 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
158}
159EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
160
161static unsigned dm_get_numa_node(void)
162{
163 return __dm_get_module_param_int(&dm_numa_node,
164 DM_NUMA_NODE, num_online_nodes() - 1);
165}
166
167static int __init local_init(void)
168{
169 int r = -ENOMEM;
170
171 /* allocate a slab for the dm_ios */
172 _io_cache = KMEM_CACHE(dm_io, 0);
173 if (!_io_cache)
174 return r;
175
176 _rq_tio_cache = KMEM_CACHE(dm_rq_target_io, 0);
177 if (!_rq_tio_cache)
178 goto out_free_io_cache;
179
180 _rq_cache = kmem_cache_create("dm_old_clone_request", sizeof(struct request),
181 __alignof__(struct request), 0, NULL);
182 if (!_rq_cache)
183 goto out_free_rq_tio_cache;
184
185 r = dm_uevent_init();
186 if (r)
187 goto out_free_rq_cache;
188
189 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
190 if (!deferred_remove_workqueue) {
191 r = -ENOMEM;
192 goto out_uevent_exit;
193 }
194
195 _major = major;
196 r = register_blkdev(_major, _name);
197 if (r < 0)
198 goto out_free_workqueue;
199
200 if (!_major)
201 _major = r;
202
203 return 0;
204
205out_free_workqueue:
206 destroy_workqueue(deferred_remove_workqueue);
207out_uevent_exit:
208 dm_uevent_exit();
209out_free_rq_cache:
210 kmem_cache_destroy(_rq_cache);
211out_free_rq_tio_cache:
212 kmem_cache_destroy(_rq_tio_cache);
213out_free_io_cache:
214 kmem_cache_destroy(_io_cache);
215
216 return r;
217}
218
219static void local_exit(void)
220{
221 flush_scheduled_work();
222 destroy_workqueue(deferred_remove_workqueue);
223
224 kmem_cache_destroy(_rq_cache);
225 kmem_cache_destroy(_rq_tio_cache);
226 kmem_cache_destroy(_io_cache);
227 unregister_blkdev(_major, _name);
228 dm_uevent_exit();
229
230 _major = 0;
231
232 DMINFO("cleaned up");
233}
234
235static int (*_inits[])(void) __initdata = {
236 local_init,
237 dm_target_init,
238 dm_linear_init,
239 dm_stripe_init,
240 dm_io_init,
241 dm_kcopyd_init,
242 dm_interface_init,
243 dm_statistics_init,
244};
245
246static void (*_exits[])(void) = {
247 local_exit,
248 dm_target_exit,
249 dm_linear_exit,
250 dm_stripe_exit,
251 dm_io_exit,
252 dm_kcopyd_exit,
253 dm_interface_exit,
254 dm_statistics_exit,
255};
256
257static int __init dm_init(void)
258{
259 const int count = ARRAY_SIZE(_inits);
260
261 int r, i;
262
263 for (i = 0; i < count; i++) {
264 r = _inits[i]();
265 if (r)
266 goto bad;
267 }
268
269 return 0;
270
271 bad:
272 while (i--)
273 _exits[i]();
274
275 return r;
276}
277
278static void __exit dm_exit(void)
279{
280 int i = ARRAY_SIZE(_exits);
281
282 while (i--)
283 _exits[i]();
284
285 /*
286 * Should be empty by this point.
287 */
288 idr_destroy(&_minor_idr);
289}
290
291/*
292 * Block device functions
293 */
294int dm_deleting_md(struct mapped_device *md)
295{
296 return test_bit(DMF_DELETING, &md->flags);
297}
298
299static int dm_blk_open(struct block_device *bdev, fmode_t mode)
300{
301 struct mapped_device *md;
302
303 spin_lock(&_minor_lock);
304
305 md = bdev->bd_disk->private_data;
306 if (!md)
307 goto out;
308
309 if (test_bit(DMF_FREEING, &md->flags) ||
310 dm_deleting_md(md)) {
311 md = NULL;
312 goto out;
313 }
314
315 dm_get(md);
316 atomic_inc(&md->open_count);
317out:
318 spin_unlock(&_minor_lock);
319
320 return md ? 0 : -ENXIO;
321}
322
323static void dm_blk_close(struct gendisk *disk, fmode_t mode)
324{
325 struct mapped_device *md;
326
327 spin_lock(&_minor_lock);
328
329 md = disk->private_data;
330 if (WARN_ON(!md))
331 goto out;
332
333 if (atomic_dec_and_test(&md->open_count) &&
334 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
335 queue_work(deferred_remove_workqueue, &deferred_remove_work);
336
337 dm_put(md);
338out:
339 spin_unlock(&_minor_lock);
340}
341
342int dm_open_count(struct mapped_device *md)
343{
344 return atomic_read(&md->open_count);
345}
346
347/*
348 * Guarantees nothing is using the device before it's deleted.
349 */
350int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
351{
352 int r = 0;
353
354 spin_lock(&_minor_lock);
355
356 if (dm_open_count(md)) {
357 r = -EBUSY;
358 if (mark_deferred)
359 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
360 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
361 r = -EEXIST;
362 else
363 set_bit(DMF_DELETING, &md->flags);
364
365 spin_unlock(&_minor_lock);
366
367 return r;
368}
369
370int dm_cancel_deferred_remove(struct mapped_device *md)
371{
372 int r = 0;
373
374 spin_lock(&_minor_lock);
375
376 if (test_bit(DMF_DELETING, &md->flags))
377 r = -EBUSY;
378 else
379 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
380
381 spin_unlock(&_minor_lock);
382
383 return r;
384}
385
386static void do_deferred_remove(struct work_struct *w)
387{
388 dm_deferred_remove();
389}
390
391sector_t dm_get_size(struct mapped_device *md)
392{
393 return get_capacity(md->disk);
394}
395
396struct request_queue *dm_get_md_queue(struct mapped_device *md)
397{
398 return md->queue;
399}
400
401struct dm_stats *dm_get_stats(struct mapped_device *md)
402{
403 return &md->stats;
404}
405
406static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
407{
408 struct mapped_device *md = bdev->bd_disk->private_data;
409
410 return dm_get_geometry(md, geo);
411}
412
413static int dm_grab_bdev_for_ioctl(struct mapped_device *md,
414 struct block_device **bdev,
415 fmode_t *mode)
416{
417 struct dm_target *tgt;
418 struct dm_table *map;
419 int srcu_idx, r;
420
421retry:
422 r = -ENOTTY;
423 map = dm_get_live_table(md, &srcu_idx);
424 if (!map || !dm_table_get_size(map))
425 goto out;
426
427 /* We only support devices that have a single target */
428 if (dm_table_get_num_targets(map) != 1)
429 goto out;
430
431 tgt = dm_table_get_target(map, 0);
432 if (!tgt->type->prepare_ioctl)
433 goto out;
434
435 if (dm_suspended_md(md)) {
436 r = -EAGAIN;
437 goto out;
438 }
439
440 r = tgt->type->prepare_ioctl(tgt, bdev, mode);
441 if (r < 0)
442 goto out;
443
444 bdgrab(*bdev);
445 dm_put_live_table(md, srcu_idx);
446 return r;
447
448out:
449 dm_put_live_table(md, srcu_idx);
450 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
451 msleep(10);
452 goto retry;
453 }
454 return r;
455}
456
457static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
458 unsigned int cmd, unsigned long arg)
459{
460 struct mapped_device *md = bdev->bd_disk->private_data;
461 int r;
462
463 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
464 if (r < 0)
465 return r;
466
467 if (r > 0) {
468 /*
469 * Target determined this ioctl is being issued against
470 * a logical partition of the parent bdev; so extra
471 * validation is needed.
472 */
473 r = scsi_verify_blk_ioctl(NULL, cmd);
474 if (r)
475 goto out;
476 }
477
478 r = __blkdev_driver_ioctl(bdev, mode, cmd, arg);
479out:
480 bdput(bdev);
481 return r;
482}
483
484static struct dm_io *alloc_io(struct mapped_device *md)
485{
486 return mempool_alloc(md->io_pool, GFP_NOIO);
487}
488
489static void free_io(struct mapped_device *md, struct dm_io *io)
490{
491 mempool_free(io, md->io_pool);
492}
493
494static void free_tio(struct dm_target_io *tio)
495{
496 bio_put(&tio->clone);
497}
498
499int md_in_flight(struct mapped_device *md)
500{
501 return atomic_read(&md->pending[READ]) +
502 atomic_read(&md->pending[WRITE]);
503}
504
505static void start_io_acct(struct dm_io *io)
506{
507 struct mapped_device *md = io->md;
508 struct bio *bio = io->bio;
509 int cpu;
510 int rw = bio_data_dir(bio);
511
512 io->start_time = jiffies;
513
514 cpu = part_stat_lock();
515 part_round_stats(cpu, &dm_disk(md)->part0);
516 part_stat_unlock();
517 atomic_set(&dm_disk(md)->part0.in_flight[rw],
518 atomic_inc_return(&md->pending[rw]));
519
520 if (unlikely(dm_stats_used(&md->stats)))
521 dm_stats_account_io(&md->stats, bio_data_dir(bio),
522 bio->bi_iter.bi_sector, bio_sectors(bio),
523 false, 0, &io->stats_aux);
524}
525
526static void end_io_acct(struct dm_io *io)
527{
528 struct mapped_device *md = io->md;
529 struct bio *bio = io->bio;
530 unsigned long duration = jiffies - io->start_time;
531 int pending;
532 int rw = bio_data_dir(bio);
533
534 generic_end_io_acct(rw, &dm_disk(md)->part0, io->start_time);
535
536 if (unlikely(dm_stats_used(&md->stats)))
537 dm_stats_account_io(&md->stats, bio_data_dir(bio),
538 bio->bi_iter.bi_sector, bio_sectors(bio),
539 true, duration, &io->stats_aux);
540
541 /*
542 * After this is decremented the bio must not be touched if it is
543 * a flush.
544 */
545 pending = atomic_dec_return(&md->pending[rw]);
546 atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
547 pending += atomic_read(&md->pending[rw^0x1]);
548
549 /* nudge anyone waiting on suspend queue */
550 if (!pending)
551 wake_up(&md->wait);
552}
553
554/*
555 * Add the bio to the list of deferred io.
556 */
557static void queue_io(struct mapped_device *md, struct bio *bio)
558{
559 unsigned long flags;
560
561 spin_lock_irqsave(&md->deferred_lock, flags);
562 bio_list_add(&md->deferred, bio);
563 spin_unlock_irqrestore(&md->deferred_lock, flags);
564 queue_work(md->wq, &md->work);
565}
566
567/*
568 * Everyone (including functions in this file), should use this
569 * function to access the md->map field, and make sure they call
570 * dm_put_live_table() when finished.
571 */
572struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
573{
574 *srcu_idx = srcu_read_lock(&md->io_barrier);
575
576 return srcu_dereference(md->map, &md->io_barrier);
577}
578
579void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
580{
581 srcu_read_unlock(&md->io_barrier, srcu_idx);
582}
583
584void dm_sync_table(struct mapped_device *md)
585{
586 synchronize_srcu(&md->io_barrier);
587 synchronize_rcu_expedited();
588}
589
590/*
591 * A fast alternative to dm_get_live_table/dm_put_live_table.
592 * The caller must not block between these two functions.
593 */
594static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
595{
596 rcu_read_lock();
597 return rcu_dereference(md->map);
598}
599
600static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
601{
602 rcu_read_unlock();
603}
604
605/*
606 * Open a table device so we can use it as a map destination.
607 */
608static int open_table_device(struct table_device *td, dev_t dev,
609 struct mapped_device *md)
610{
611 static char *_claim_ptr = "I belong to device-mapper";
612 struct block_device *bdev;
613
614 int r;
615
616 BUG_ON(td->dm_dev.bdev);
617
618 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _claim_ptr);
619 if (IS_ERR(bdev))
620 return PTR_ERR(bdev);
621
622 r = bd_link_disk_holder(bdev, dm_disk(md));
623 if (r) {
624 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
625 return r;
626 }
627
628 td->dm_dev.bdev = bdev;
629 return 0;
630}
631
632/*
633 * Close a table device that we've been using.
634 */
635static void close_table_device(struct table_device *td, struct mapped_device *md)
636{
637 if (!td->dm_dev.bdev)
638 return;
639
640 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
641 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
642 td->dm_dev.bdev = NULL;
643}
644
645static struct table_device *find_table_device(struct list_head *l, dev_t dev,
646 fmode_t mode) {
647 struct table_device *td;
648
649 list_for_each_entry(td, l, list)
650 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
651 return td;
652
653 return NULL;
654}
655
656int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
657 struct dm_dev **result) {
658 int r;
659 struct table_device *td;
660
661 mutex_lock(&md->table_devices_lock);
662 td = find_table_device(&md->table_devices, dev, mode);
663 if (!td) {
664 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
665 if (!td) {
666 mutex_unlock(&md->table_devices_lock);
667 return -ENOMEM;
668 }
669
670 td->dm_dev.mode = mode;
671 td->dm_dev.bdev = NULL;
672
673 if ((r = open_table_device(td, dev, md))) {
674 mutex_unlock(&md->table_devices_lock);
675 kfree(td);
676 return r;
677 }
678
679 format_dev_t(td->dm_dev.name, dev);
680
681 atomic_set(&td->count, 0);
682 list_add(&td->list, &md->table_devices);
683 }
684 atomic_inc(&td->count);
685 mutex_unlock(&md->table_devices_lock);
686
687 *result = &td->dm_dev;
688 return 0;
689}
690EXPORT_SYMBOL_GPL(dm_get_table_device);
691
692void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
693{
694 struct table_device *td = container_of(d, struct table_device, dm_dev);
695
696 mutex_lock(&md->table_devices_lock);
697 if (atomic_dec_and_test(&td->count)) {
698 close_table_device(td, md);
699 list_del(&td->list);
700 kfree(td);
701 }
702 mutex_unlock(&md->table_devices_lock);
703}
704EXPORT_SYMBOL(dm_put_table_device);
705
706static void free_table_devices(struct list_head *devices)
707{
708 struct list_head *tmp, *next;
709
710 list_for_each_safe(tmp, next, devices) {
711 struct table_device *td = list_entry(tmp, struct table_device, list);
712
713 DMWARN("dm_destroy: %s still exists with %d references",
714 td->dm_dev.name, atomic_read(&td->count));
715 kfree(td);
716 }
717}
718
719/*
720 * Get the geometry associated with a dm device
721 */
722int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
723{
724 *geo = md->geometry;
725
726 return 0;
727}
728
729/*
730 * Set the geometry of a device.
731 */
732int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
733{
734 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
735
736 if (geo->start > sz) {
737 DMWARN("Start sector is beyond the geometry limits.");
738 return -EINVAL;
739 }
740
741 md->geometry = *geo;
742
743 return 0;
744}
745
746/*-----------------------------------------------------------------
747 * CRUD START:
748 * A more elegant soln is in the works that uses the queue
749 * merge fn, unfortunately there are a couple of changes to
750 * the block layer that I want to make for this. So in the
751 * interests of getting something for people to use I give
752 * you this clearly demarcated crap.
753 *---------------------------------------------------------------*/
754
755static int __noflush_suspending(struct mapped_device *md)
756{
757 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
758}
759
760/*
761 * Decrements the number of outstanding ios that a bio has been
762 * cloned into, completing the original io if necc.
763 */
764static void dec_pending(struct dm_io *io, int error)
765{
766 unsigned long flags;
767 int io_error;
768 struct bio *bio;
769 struct mapped_device *md = io->md;
770
771 /* Push-back supersedes any I/O errors */
772 if (unlikely(error)) {
773 spin_lock_irqsave(&io->endio_lock, flags);
774 if (!(io->error > 0 && __noflush_suspending(md)))
775 io->error = error;
776 spin_unlock_irqrestore(&io->endio_lock, flags);
777 }
778
779 if (atomic_dec_and_test(&io->io_count)) {
780 if (io->error == DM_ENDIO_REQUEUE) {
781 /*
782 * Target requested pushing back the I/O.
783 */
784 spin_lock_irqsave(&md->deferred_lock, flags);
785 if (__noflush_suspending(md))
786 bio_list_add_head(&md->deferred, io->bio);
787 else
788 /* noflush suspend was interrupted. */
789 io->error = -EIO;
790 spin_unlock_irqrestore(&md->deferred_lock, flags);
791 }
792
793 io_error = io->error;
794 bio = io->bio;
795 end_io_acct(io);
796 free_io(md, io);
797
798 if (io_error == DM_ENDIO_REQUEUE)
799 return;
800
801 if ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size) {
802 /*
803 * Preflush done for flush with data, reissue
804 * without REQ_PREFLUSH.
805 */
806 bio->bi_opf &= ~REQ_PREFLUSH;
807 queue_io(md, bio);
808 } else {
809 /* done with normal IO or empty flush */
810 trace_block_bio_complete(md->queue, bio, io_error);
811 bio->bi_error = io_error;
812 bio_endio(bio);
813 }
814 }
815}
816
817void disable_write_same(struct mapped_device *md)
818{
819 struct queue_limits *limits = dm_get_queue_limits(md);
820
821 /* device doesn't really support WRITE SAME, disable it */
822 limits->max_write_same_sectors = 0;
823}
824
825static void clone_endio(struct bio *bio)
826{
827 int error = bio->bi_error;
828 int r = error;
829 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
830 struct dm_io *io = tio->io;
831 struct mapped_device *md = tio->io->md;
832 dm_endio_fn endio = tio->ti->type->end_io;
833
834 if (endio) {
835 r = endio(tio->ti, bio, error);
836 if (r < 0 || r == DM_ENDIO_REQUEUE)
837 /*
838 * error and requeue request are handled
839 * in dec_pending().
840 */
841 error = r;
842 else if (r == DM_ENDIO_INCOMPLETE)
843 /* The target will handle the io */
844 return;
845 else if (r) {
846 DMWARN("unimplemented target endio return value: %d", r);
847 BUG();
848 }
849 }
850
851 if (unlikely(r == -EREMOTEIO && (bio_op(bio) == REQ_OP_WRITE_SAME) &&
852 !bdev_get_queue(bio->bi_bdev)->limits.max_write_same_sectors))
853 disable_write_same(md);
854
855 free_tio(tio);
856 dec_pending(io, error);
857}
858
859/*
860 * Return maximum size of I/O possible at the supplied sector up to the current
861 * target boundary.
862 */
863static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
864{
865 sector_t target_offset = dm_target_offset(ti, sector);
866
867 return ti->len - target_offset;
868}
869
870static sector_t max_io_len(sector_t sector, struct dm_target *ti)
871{
872 sector_t len = max_io_len_target_boundary(sector, ti);
873 sector_t offset, max_len;
874
875 /*
876 * Does the target need to split even further?
877 */
878 if (ti->max_io_len) {
879 offset = dm_target_offset(ti, sector);
880 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
881 max_len = sector_div(offset, ti->max_io_len);
882 else
883 max_len = offset & (ti->max_io_len - 1);
884 max_len = ti->max_io_len - max_len;
885
886 if (len > max_len)
887 len = max_len;
888 }
889
890 return len;
891}
892
893int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
894{
895 if (len > UINT_MAX) {
896 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
897 (unsigned long long)len, UINT_MAX);
898 ti->error = "Maximum size of target IO is too large";
899 return -EINVAL;
900 }
901
902 ti->max_io_len = (uint32_t) len;
903
904 return 0;
905}
906EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
907
908static long dm_blk_direct_access(struct block_device *bdev, sector_t sector,
909 void **kaddr, pfn_t *pfn, long size)
910{
911 struct mapped_device *md = bdev->bd_disk->private_data;
912 struct dm_table *map;
913 struct dm_target *ti;
914 int srcu_idx;
915 long len, ret = -EIO;
916
917 map = dm_get_live_table(md, &srcu_idx);
918 if (!map)
919 goto out;
920
921 ti = dm_table_find_target(map, sector);
922 if (!dm_target_is_valid(ti))
923 goto out;
924
925 len = max_io_len(sector, ti) << SECTOR_SHIFT;
926 size = min(len, size);
927
928 if (ti->type->direct_access)
929 ret = ti->type->direct_access(ti, sector, kaddr, pfn, size);
930out:
931 dm_put_live_table(md, srcu_idx);
932 return min(ret, size);
933}
934
935/*
936 * A target may call dm_accept_partial_bio only from the map routine. It is
937 * allowed for all bio types except REQ_PREFLUSH.
938 *
939 * dm_accept_partial_bio informs the dm that the target only wants to process
940 * additional n_sectors sectors of the bio and the rest of the data should be
941 * sent in a next bio.
942 *
943 * A diagram that explains the arithmetics:
944 * +--------------------+---------------+-------+
945 * | 1 | 2 | 3 |
946 * +--------------------+---------------+-------+
947 *
948 * <-------------- *tio->len_ptr --------------->
949 * <------- bi_size ------->
950 * <-- n_sectors -->
951 *
952 * Region 1 was already iterated over with bio_advance or similar function.
953 * (it may be empty if the target doesn't use bio_advance)
954 * Region 2 is the remaining bio size that the target wants to process.
955 * (it may be empty if region 1 is non-empty, although there is no reason
956 * to make it empty)
957 * The target requires that region 3 is to be sent in the next bio.
958 *
959 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
960 * the partially processed part (the sum of regions 1+2) must be the same for all
961 * copies of the bio.
962 */
963void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
964{
965 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
966 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
967 BUG_ON(bio->bi_opf & REQ_PREFLUSH);
968 BUG_ON(bi_size > *tio->len_ptr);
969 BUG_ON(n_sectors > bi_size);
970 *tio->len_ptr -= bi_size - n_sectors;
971 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
972}
973EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
974
975/*
976 * Flush current->bio_list when the target map method blocks.
977 * This fixes deadlocks in snapshot and possibly in other targets.
978 */
979struct dm_offload {
980 struct blk_plug plug;
981 struct blk_plug_cb cb;
982};
983
984static void flush_current_bio_list(struct blk_plug_cb *cb, bool from_schedule)
985{
986 struct dm_offload *o = container_of(cb, struct dm_offload, cb);
987 struct bio_list list;
988 struct bio *bio;
989 int i;
990
991 INIT_LIST_HEAD(&o->cb.list);
992
993 if (unlikely(!current->bio_list))
994 return;
995
996 for (i = 0; i < 2; i++) {
997 list = current->bio_list[i];
998 bio_list_init(¤t->bio_list[i]);
999
1000 while ((bio = bio_list_pop(&list))) {
1001 struct bio_set *bs = bio->bi_pool;
1002 if (unlikely(!bs) || bs == fs_bio_set) {
1003 bio_list_add(¤t->bio_list[i], bio);
1004 continue;
1005 }
1006
1007 spin_lock(&bs->rescue_lock);
1008 bio_list_add(&bs->rescue_list, bio);
1009 queue_work(bs->rescue_workqueue, &bs->rescue_work);
1010 spin_unlock(&bs->rescue_lock);
1011 }
1012 }
1013}
1014
1015static void dm_offload_start(struct dm_offload *o)
1016{
1017 blk_start_plug(&o->plug);
1018 o->cb.callback = flush_current_bio_list;
1019 list_add(&o->cb.list, ¤t->plug->cb_list);
1020}
1021
1022static void dm_offload_end(struct dm_offload *o)
1023{
1024 list_del(&o->cb.list);
1025 blk_finish_plug(&o->plug);
1026}
1027
1028static void __map_bio(struct dm_target_io *tio)
1029{
1030 int r;
1031 sector_t sector;
1032 struct dm_offload o;
1033 struct bio *clone = &tio->clone;
1034 struct dm_target *ti = tio->ti;
1035
1036 clone->bi_end_io = clone_endio;
1037
1038 /*
1039 * Map the clone. If r == 0 we don't need to do
1040 * anything, the target has assumed ownership of
1041 * this io.
1042 */
1043 atomic_inc(&tio->io->io_count);
1044 sector = clone->bi_iter.bi_sector;
1045
1046 dm_offload_start(&o);
1047 r = ti->type->map(ti, clone);
1048 dm_offload_end(&o);
1049
1050 if (r == DM_MAPIO_REMAPPED) {
1051 /* the bio has been remapped so dispatch it */
1052
1053 trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1054 tio->io->bio->bi_bdev->bd_dev, sector);
1055
1056 generic_make_request(clone);
1057 } else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1058 /* error the io and bail out, or requeue it if needed */
1059 dec_pending(tio->io, r);
1060 free_tio(tio);
1061 } else if (r != DM_MAPIO_SUBMITTED) {
1062 DMWARN("unimplemented target map return value: %d", r);
1063 BUG();
1064 }
1065}
1066
1067struct clone_info {
1068 struct mapped_device *md;
1069 struct dm_table *map;
1070 struct bio *bio;
1071 struct dm_io *io;
1072 sector_t sector;
1073 unsigned sector_count;
1074};
1075
1076static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1077{
1078 bio->bi_iter.bi_sector = sector;
1079 bio->bi_iter.bi_size = to_bytes(len);
1080}
1081
1082/*
1083 * Creates a bio that consists of range of complete bvecs.
1084 */
1085static int clone_bio(struct dm_target_io *tio, struct bio *bio,
1086 sector_t sector, unsigned len)
1087{
1088 struct bio *clone = &tio->clone;
1089
1090 __bio_clone_fast(clone, bio);
1091
1092 if (bio_integrity(bio)) {
1093 int r = bio_integrity_clone(clone, bio, GFP_NOIO);
1094 if (r < 0)
1095 return r;
1096 }
1097
1098 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1099 clone->bi_iter.bi_size = to_bytes(len);
1100
1101 if (bio_integrity(bio))
1102 bio_integrity_trim(clone, 0, len);
1103
1104 return 0;
1105}
1106
1107static struct dm_target_io *alloc_tio(struct clone_info *ci,
1108 struct dm_target *ti,
1109 unsigned target_bio_nr)
1110{
1111 struct dm_target_io *tio;
1112 struct bio *clone;
1113
1114 clone = bio_alloc_bioset(GFP_NOIO, 0, ci->md->bs);
1115 tio = container_of(clone, struct dm_target_io, clone);
1116
1117 tio->io = ci->io;
1118 tio->ti = ti;
1119 tio->target_bio_nr = target_bio_nr;
1120
1121 return tio;
1122}
1123
1124static void __clone_and_map_simple_bio(struct clone_info *ci,
1125 struct dm_target *ti,
1126 unsigned target_bio_nr, unsigned *len)
1127{
1128 struct dm_target_io *tio = alloc_tio(ci, ti, target_bio_nr);
1129 struct bio *clone = &tio->clone;
1130
1131 tio->len_ptr = len;
1132
1133 __bio_clone_fast(clone, ci->bio);
1134 if (len)
1135 bio_setup_sector(clone, ci->sector, *len);
1136
1137 __map_bio(tio);
1138}
1139
1140static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1141 unsigned num_bios, unsigned *len)
1142{
1143 unsigned target_bio_nr;
1144
1145 for (target_bio_nr = 0; target_bio_nr < num_bios; target_bio_nr++)
1146 __clone_and_map_simple_bio(ci, ti, target_bio_nr, len);
1147}
1148
1149static int __send_empty_flush(struct clone_info *ci)
1150{
1151 unsigned target_nr = 0;
1152 struct dm_target *ti;
1153
1154 BUG_ON(bio_has_data(ci->bio));
1155 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1156 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1157
1158 return 0;
1159}
1160
1161static int __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1162 sector_t sector, unsigned *len)
1163{
1164 struct bio *bio = ci->bio;
1165 struct dm_target_io *tio;
1166 unsigned target_bio_nr;
1167 unsigned num_target_bios = 1;
1168 int r = 0;
1169
1170 /*
1171 * Does the target want to receive duplicate copies of the bio?
1172 */
1173 if (bio_data_dir(bio) == WRITE && ti->num_write_bios)
1174 num_target_bios = ti->num_write_bios(ti, bio);
1175
1176 for (target_bio_nr = 0; target_bio_nr < num_target_bios; target_bio_nr++) {
1177 tio = alloc_tio(ci, ti, target_bio_nr);
1178 tio->len_ptr = len;
1179 r = clone_bio(tio, bio, sector, *len);
1180 if (r < 0) {
1181 free_tio(tio);
1182 break;
1183 }
1184 __map_bio(tio);
1185 }
1186
1187 return r;
1188}
1189
1190typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1191
1192static unsigned get_num_discard_bios(struct dm_target *ti)
1193{
1194 return ti->num_discard_bios;
1195}
1196
1197static unsigned get_num_write_same_bios(struct dm_target *ti)
1198{
1199 return ti->num_write_same_bios;
1200}
1201
1202typedef bool (*is_split_required_fn)(struct dm_target *ti);
1203
1204static bool is_split_required_for_discard(struct dm_target *ti)
1205{
1206 return ti->split_discard_bios;
1207}
1208
1209static int __send_changing_extent_only(struct clone_info *ci,
1210 get_num_bios_fn get_num_bios,
1211 is_split_required_fn is_split_required)
1212{
1213 struct dm_target *ti;
1214 unsigned len;
1215 unsigned num_bios;
1216
1217 do {
1218 ti = dm_table_find_target(ci->map, ci->sector);
1219 if (!dm_target_is_valid(ti))
1220 return -EIO;
1221
1222 /*
1223 * Even though the device advertised support for this type of
1224 * request, that does not mean every target supports it, and
1225 * reconfiguration might also have changed that since the
1226 * check was performed.
1227 */
1228 num_bios = get_num_bios ? get_num_bios(ti) : 0;
1229 if (!num_bios)
1230 return -EOPNOTSUPP;
1231
1232 if (is_split_required && !is_split_required(ti))
1233 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1234 else
1235 len = min((sector_t)ci->sector_count, max_io_len(ci->sector, ti));
1236
1237 __send_duplicate_bios(ci, ti, num_bios, &len);
1238
1239 ci->sector += len;
1240 } while (ci->sector_count -= len);
1241
1242 return 0;
1243}
1244
1245static int __send_discard(struct clone_info *ci)
1246{
1247 return __send_changing_extent_only(ci, get_num_discard_bios,
1248 is_split_required_for_discard);
1249}
1250
1251static int __send_write_same(struct clone_info *ci)
1252{
1253 return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
1254}
1255
1256/*
1257 * Select the correct strategy for processing a non-flush bio.
1258 */
1259static int __split_and_process_non_flush(struct clone_info *ci)
1260{
1261 struct bio *bio = ci->bio;
1262 struct dm_target *ti;
1263 unsigned len;
1264 int r;
1265
1266 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
1267 return __send_discard(ci);
1268 else if (unlikely(bio_op(bio) == REQ_OP_WRITE_SAME))
1269 return __send_write_same(ci);
1270
1271 ti = dm_table_find_target(ci->map, ci->sector);
1272 if (!dm_target_is_valid(ti))
1273 return -EIO;
1274
1275 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1276
1277 r = __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1278 if (r < 0)
1279 return r;
1280
1281 ci->sector += len;
1282 ci->sector_count -= len;
1283
1284 return 0;
1285}
1286
1287/*
1288 * Entry point to split a bio into clones and submit them to the targets.
1289 */
1290static void __split_and_process_bio(struct mapped_device *md,
1291 struct dm_table *map, struct bio *bio)
1292{
1293 struct clone_info ci;
1294 int error = 0;
1295
1296 if (unlikely(!map)) {
1297 bio_io_error(bio);
1298 return;
1299 }
1300
1301 ci.map = map;
1302 ci.md = md;
1303 ci.io = alloc_io(md);
1304 ci.io->error = 0;
1305 atomic_set(&ci.io->io_count, 1);
1306 ci.io->bio = bio;
1307 ci.io->md = md;
1308 spin_lock_init(&ci.io->endio_lock);
1309 ci.sector = bio->bi_iter.bi_sector;
1310
1311 start_io_acct(ci.io);
1312
1313 if (bio->bi_opf & REQ_PREFLUSH) {
1314 ci.bio = &ci.md->flush_bio;
1315 ci.sector_count = 0;
1316 error = __send_empty_flush(&ci);
1317 /* dec_pending submits any data associated with flush */
1318 } else {
1319 ci.bio = bio;
1320 ci.sector_count = bio_sectors(bio);
1321 while (ci.sector_count && !error)
1322 error = __split_and_process_non_flush(&ci);
1323 }
1324
1325 /* drop the extra reference count */
1326 dec_pending(ci.io, error);
1327}
1328/*-----------------------------------------------------------------
1329 * CRUD END
1330 *---------------------------------------------------------------*/
1331
1332/*
1333 * The request function that just remaps the bio built up by
1334 * dm_merge_bvec.
1335 */
1336static blk_qc_t dm_make_request(struct request_queue *q, struct bio *bio)
1337{
1338 int rw = bio_data_dir(bio);
1339 struct mapped_device *md = q->queuedata;
1340 int srcu_idx;
1341 struct dm_table *map;
1342
1343 map = dm_get_live_table(md, &srcu_idx);
1344
1345 generic_start_io_acct(rw, bio_sectors(bio), &dm_disk(md)->part0);
1346
1347 /* if we're suspended, we have to queue this io for later */
1348 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1349 dm_put_live_table(md, srcu_idx);
1350
1351 if (!(bio->bi_opf & REQ_RAHEAD))
1352 queue_io(md, bio);
1353 else
1354 bio_io_error(bio);
1355 return BLK_QC_T_NONE;
1356 }
1357
1358 __split_and_process_bio(md, map, bio);
1359 dm_put_live_table(md, srcu_idx);
1360 return BLK_QC_T_NONE;
1361}
1362
1363static int dm_any_congested(void *congested_data, int bdi_bits)
1364{
1365 int r = bdi_bits;
1366 struct mapped_device *md = congested_data;
1367 struct dm_table *map;
1368
1369 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
1370 if (dm_request_based(md)) {
1371 /*
1372 * With request-based DM we only need to check the
1373 * top-level queue for congestion.
1374 */
1375 r = md->queue->backing_dev_info.wb.state & bdi_bits;
1376 } else {
1377 map = dm_get_live_table_fast(md);
1378 if (map)
1379 r = dm_table_any_congested(map, bdi_bits);
1380 dm_put_live_table_fast(md);
1381 }
1382 }
1383
1384 return r;
1385}
1386
1387/*-----------------------------------------------------------------
1388 * An IDR is used to keep track of allocated minor numbers.
1389 *---------------------------------------------------------------*/
1390static void free_minor(int minor)
1391{
1392 spin_lock(&_minor_lock);
1393 idr_remove(&_minor_idr, minor);
1394 spin_unlock(&_minor_lock);
1395}
1396
1397/*
1398 * See if the device with a specific minor # is free.
1399 */
1400static int specific_minor(int minor)
1401{
1402 int r;
1403
1404 if (minor >= (1 << MINORBITS))
1405 return -EINVAL;
1406
1407 idr_preload(GFP_KERNEL);
1408 spin_lock(&_minor_lock);
1409
1410 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
1411
1412 spin_unlock(&_minor_lock);
1413 idr_preload_end();
1414 if (r < 0)
1415 return r == -ENOSPC ? -EBUSY : r;
1416 return 0;
1417}
1418
1419static int next_free_minor(int *minor)
1420{
1421 int r;
1422
1423 idr_preload(GFP_KERNEL);
1424 spin_lock(&_minor_lock);
1425
1426 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
1427
1428 spin_unlock(&_minor_lock);
1429 idr_preload_end();
1430 if (r < 0)
1431 return r;
1432 *minor = r;
1433 return 0;
1434}
1435
1436static const struct block_device_operations dm_blk_dops;
1437
1438static void dm_wq_work(struct work_struct *work);
1439
1440void dm_init_md_queue(struct mapped_device *md)
1441{
1442 /*
1443 * Request-based dm devices cannot be stacked on top of bio-based dm
1444 * devices. The type of this dm device may not have been decided yet.
1445 * The type is decided at the first table loading time.
1446 * To prevent problematic device stacking, clear the queue flag
1447 * for request stacking support until then.
1448 *
1449 * This queue is new, so no concurrency on the queue_flags.
1450 */
1451 queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
1452
1453 /*
1454 * Initialize data that will only be used by a non-blk-mq DM queue
1455 * - must do so here (in alloc_dev callchain) before queue is used
1456 */
1457 md->queue->queuedata = md;
1458 md->queue->backing_dev_info.congested_data = md;
1459}
1460
1461void dm_init_normal_md_queue(struct mapped_device *md)
1462{
1463 md->use_blk_mq = false;
1464 dm_init_md_queue(md);
1465
1466 /*
1467 * Initialize aspects of queue that aren't relevant for blk-mq
1468 */
1469 md->queue->backing_dev_info.congested_fn = dm_any_congested;
1470 blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
1471}
1472
1473static void cleanup_mapped_device(struct mapped_device *md)
1474{
1475 if (md->wq)
1476 destroy_workqueue(md->wq);
1477 if (md->kworker_task)
1478 kthread_stop(md->kworker_task);
1479 mempool_destroy(md->io_pool);
1480 mempool_destroy(md->rq_pool);
1481 if (md->bs)
1482 bioset_free(md->bs);
1483
1484 if (md->disk) {
1485 spin_lock(&_minor_lock);
1486 md->disk->private_data = NULL;
1487 spin_unlock(&_minor_lock);
1488 del_gendisk(md->disk);
1489 put_disk(md->disk);
1490 }
1491
1492 if (md->queue)
1493 blk_cleanup_queue(md->queue);
1494
1495 cleanup_srcu_struct(&md->io_barrier);
1496
1497 if (md->bdev) {
1498 bdput(md->bdev);
1499 md->bdev = NULL;
1500 }
1501
1502 dm_mq_cleanup_mapped_device(md);
1503}
1504
1505/*
1506 * Allocate and initialise a blank device with a given minor.
1507 */
1508static struct mapped_device *alloc_dev(int minor)
1509{
1510 int r, numa_node_id = dm_get_numa_node();
1511 struct mapped_device *md;
1512 void *old_md;
1513
1514 md = kzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
1515 if (!md) {
1516 DMWARN("unable to allocate device, out of memory.");
1517 return NULL;
1518 }
1519
1520 if (!try_module_get(THIS_MODULE))
1521 goto bad_module_get;
1522
1523 /* get a minor number for the dev */
1524 if (minor == DM_ANY_MINOR)
1525 r = next_free_minor(&minor);
1526 else
1527 r = specific_minor(minor);
1528 if (r < 0)
1529 goto bad_minor;
1530
1531 r = init_srcu_struct(&md->io_barrier);
1532 if (r < 0)
1533 goto bad_io_barrier;
1534
1535 md->numa_node_id = numa_node_id;
1536 md->use_blk_mq = dm_use_blk_mq_default();
1537 md->init_tio_pdu = false;
1538 md->type = DM_TYPE_NONE;
1539 mutex_init(&md->suspend_lock);
1540 mutex_init(&md->type_lock);
1541 mutex_init(&md->table_devices_lock);
1542 spin_lock_init(&md->deferred_lock);
1543 atomic_set(&md->holders, 1);
1544 atomic_set(&md->open_count, 0);
1545 atomic_set(&md->event_nr, 0);
1546 atomic_set(&md->uevent_seq, 0);
1547 INIT_LIST_HEAD(&md->uevent_list);
1548 INIT_LIST_HEAD(&md->table_devices);
1549 spin_lock_init(&md->uevent_lock);
1550
1551 md->queue = blk_alloc_queue_node(GFP_KERNEL, numa_node_id);
1552 if (!md->queue)
1553 goto bad;
1554
1555 dm_init_md_queue(md);
1556
1557 md->disk = alloc_disk_node(1, numa_node_id);
1558 if (!md->disk)
1559 goto bad;
1560
1561 atomic_set(&md->pending[0], 0);
1562 atomic_set(&md->pending[1], 0);
1563 init_waitqueue_head(&md->wait);
1564 INIT_WORK(&md->work, dm_wq_work);
1565 init_waitqueue_head(&md->eventq);
1566 init_completion(&md->kobj_holder.completion);
1567 md->kworker_task = NULL;
1568
1569 md->disk->major = _major;
1570 md->disk->first_minor = minor;
1571 md->disk->fops = &dm_blk_dops;
1572 md->disk->queue = md->queue;
1573 md->disk->private_data = md;
1574 sprintf(md->disk->disk_name, "dm-%d", minor);
1575 add_disk(md->disk);
1576 format_dev_t(md->name, MKDEV(_major, minor));
1577
1578 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
1579 if (!md->wq)
1580 goto bad;
1581
1582 md->bdev = bdget_disk(md->disk, 0);
1583 if (!md->bdev)
1584 goto bad;
1585
1586 bio_init(&md->flush_bio, NULL, 0);
1587 md->flush_bio.bi_bdev = md->bdev;
1588 md->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1589
1590 dm_stats_init(&md->stats);
1591
1592 /* Populate the mapping, nobody knows we exist yet */
1593 spin_lock(&_minor_lock);
1594 old_md = idr_replace(&_minor_idr, md, minor);
1595 spin_unlock(&_minor_lock);
1596
1597 BUG_ON(old_md != MINOR_ALLOCED);
1598
1599 return md;
1600
1601bad:
1602 cleanup_mapped_device(md);
1603bad_io_barrier:
1604 free_minor(minor);
1605bad_minor:
1606 module_put(THIS_MODULE);
1607bad_module_get:
1608 kfree(md);
1609 return NULL;
1610}
1611
1612static void unlock_fs(struct mapped_device *md);
1613
1614static void free_dev(struct mapped_device *md)
1615{
1616 int minor = MINOR(disk_devt(md->disk));
1617
1618 unlock_fs(md);
1619
1620 cleanup_mapped_device(md);
1621
1622 free_table_devices(&md->table_devices);
1623 dm_stats_cleanup(&md->stats);
1624 free_minor(minor);
1625
1626 module_put(THIS_MODULE);
1627 kfree(md);
1628}
1629
1630static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
1631{
1632 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
1633
1634 if (md->bs) {
1635 /* The md already has necessary mempools. */
1636 if (dm_table_bio_based(t)) {
1637 /*
1638 * Reload bioset because front_pad may have changed
1639 * because a different table was loaded.
1640 */
1641 bioset_free(md->bs);
1642 md->bs = p->bs;
1643 p->bs = NULL;
1644 }
1645 /*
1646 * There's no need to reload with request-based dm
1647 * because the size of front_pad doesn't change.
1648 * Note for future: If you are to reload bioset,
1649 * prep-ed requests in the queue may refer
1650 * to bio from the old bioset, so you must walk
1651 * through the queue to unprep.
1652 */
1653 goto out;
1654 }
1655
1656 BUG_ON(!p || md->io_pool || md->rq_pool || md->bs);
1657
1658 md->io_pool = p->io_pool;
1659 p->io_pool = NULL;
1660 md->rq_pool = p->rq_pool;
1661 p->rq_pool = NULL;
1662 md->bs = p->bs;
1663 p->bs = NULL;
1664
1665out:
1666 /* mempool bind completed, no longer need any mempools in the table */
1667 dm_table_free_md_mempools(t);
1668}
1669
1670/*
1671 * Bind a table to the device.
1672 */
1673static void event_callback(void *context)
1674{
1675 unsigned long flags;
1676 LIST_HEAD(uevents);
1677 struct mapped_device *md = (struct mapped_device *) context;
1678
1679 spin_lock_irqsave(&md->uevent_lock, flags);
1680 list_splice_init(&md->uevent_list, &uevents);
1681 spin_unlock_irqrestore(&md->uevent_lock, flags);
1682
1683 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
1684
1685 atomic_inc(&md->event_nr);
1686 wake_up(&md->eventq);
1687}
1688
1689/*
1690 * Protected by md->suspend_lock obtained by dm_swap_table().
1691 */
1692static void __set_size(struct mapped_device *md, sector_t size)
1693{
1694 set_capacity(md->disk, size);
1695
1696 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
1697}
1698
1699/*
1700 * Returns old map, which caller must destroy.
1701 */
1702static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
1703 struct queue_limits *limits)
1704{
1705 struct dm_table *old_map;
1706 struct request_queue *q = md->queue;
1707 sector_t size;
1708
1709 lockdep_assert_held(&md->suspend_lock);
1710
1711 size = dm_table_get_size(t);
1712
1713 /*
1714 * Wipe any geometry if the size of the table changed.
1715 */
1716 if (size != dm_get_size(md))
1717 memset(&md->geometry, 0, sizeof(md->geometry));
1718
1719 __set_size(md, size);
1720
1721 dm_table_event_callback(t, event_callback, md);
1722
1723 /*
1724 * The queue hasn't been stopped yet, if the old table type wasn't
1725 * for request-based during suspension. So stop it to prevent
1726 * I/O mapping before resume.
1727 * This must be done before setting the queue restrictions,
1728 * because request-based dm may be run just after the setting.
1729 */
1730 if (dm_table_request_based(t)) {
1731 dm_stop_queue(q);
1732 /*
1733 * Leverage the fact that request-based DM targets are
1734 * immutable singletons and establish md->immutable_target
1735 * - used to optimize both dm_request_fn and dm_mq_queue_rq
1736 */
1737 md->immutable_target = dm_table_get_immutable_target(t);
1738 }
1739
1740 __bind_mempools(md, t);
1741
1742 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
1743 rcu_assign_pointer(md->map, (void *)t);
1744 md->immutable_target_type = dm_table_get_immutable_target_type(t);
1745
1746 dm_table_set_restrictions(t, q, limits);
1747 if (old_map)
1748 dm_sync_table(md);
1749
1750 return old_map;
1751}
1752
1753/*
1754 * Returns unbound table for the caller to free.
1755 */
1756static struct dm_table *__unbind(struct mapped_device *md)
1757{
1758 struct dm_table *map = rcu_dereference_protected(md->map, 1);
1759
1760 if (!map)
1761 return NULL;
1762
1763 dm_table_event_callback(map, NULL, NULL);
1764 RCU_INIT_POINTER(md->map, NULL);
1765 dm_sync_table(md);
1766
1767 return map;
1768}
1769
1770/*
1771 * Constructor for a new device.
1772 */
1773int dm_create(int minor, struct mapped_device **result)
1774{
1775 struct mapped_device *md;
1776
1777 md = alloc_dev(minor);
1778 if (!md)
1779 return -ENXIO;
1780
1781 dm_sysfs_init(md);
1782
1783 *result = md;
1784 return 0;
1785}
1786
1787/*
1788 * Functions to manage md->type.
1789 * All are required to hold md->type_lock.
1790 */
1791void dm_lock_md_type(struct mapped_device *md)
1792{
1793 mutex_lock(&md->type_lock);
1794}
1795
1796void dm_unlock_md_type(struct mapped_device *md)
1797{
1798 mutex_unlock(&md->type_lock);
1799}
1800
1801void dm_set_md_type(struct mapped_device *md, unsigned type)
1802{
1803 BUG_ON(!mutex_is_locked(&md->type_lock));
1804 md->type = type;
1805}
1806
1807unsigned dm_get_md_type(struct mapped_device *md)
1808{
1809 return md->type;
1810}
1811
1812struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
1813{
1814 return md->immutable_target_type;
1815}
1816
1817/*
1818 * The queue_limits are only valid as long as you have a reference
1819 * count on 'md'.
1820 */
1821struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
1822{
1823 BUG_ON(!atomic_read(&md->holders));
1824 return &md->queue->limits;
1825}
1826EXPORT_SYMBOL_GPL(dm_get_queue_limits);
1827
1828/*
1829 * Setup the DM device's queue based on md's type
1830 */
1831int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
1832{
1833 int r;
1834 unsigned type = dm_get_md_type(md);
1835
1836 switch (type) {
1837 case DM_TYPE_REQUEST_BASED:
1838 r = dm_old_init_request_queue(md);
1839 if (r) {
1840 DMERR("Cannot initialize queue for request-based mapped device");
1841 return r;
1842 }
1843 break;
1844 case DM_TYPE_MQ_REQUEST_BASED:
1845 r = dm_mq_init_request_queue(md, t);
1846 if (r) {
1847 DMERR("Cannot initialize queue for request-based dm-mq mapped device");
1848 return r;
1849 }
1850 break;
1851 case DM_TYPE_BIO_BASED:
1852 case DM_TYPE_DAX_BIO_BASED:
1853 dm_init_normal_md_queue(md);
1854 blk_queue_make_request(md->queue, dm_make_request);
1855 /*
1856 * DM handles splitting bios as needed. Free the bio_split bioset
1857 * since it won't be used (saves 1 process per bio-based DM device).
1858 */
1859 bioset_free(md->queue->bio_split);
1860 md->queue->bio_split = NULL;
1861
1862 if (type == DM_TYPE_DAX_BIO_BASED)
1863 queue_flag_set_unlocked(QUEUE_FLAG_DAX, md->queue);
1864 break;
1865 }
1866
1867 return 0;
1868}
1869
1870struct mapped_device *dm_get_md(dev_t dev)
1871{
1872 struct mapped_device *md;
1873 unsigned minor = MINOR(dev);
1874
1875 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
1876 return NULL;
1877
1878 spin_lock(&_minor_lock);
1879
1880 md = idr_find(&_minor_idr, minor);
1881 if (md) {
1882 if ((md == MINOR_ALLOCED ||
1883 (MINOR(disk_devt(dm_disk(md))) != minor) ||
1884 dm_deleting_md(md) ||
1885 test_bit(DMF_FREEING, &md->flags))) {
1886 md = NULL;
1887 goto out;
1888 }
1889 dm_get(md);
1890 }
1891
1892out:
1893 spin_unlock(&_minor_lock);
1894
1895 return md;
1896}
1897EXPORT_SYMBOL_GPL(dm_get_md);
1898
1899void *dm_get_mdptr(struct mapped_device *md)
1900{
1901 return md->interface_ptr;
1902}
1903
1904void dm_set_mdptr(struct mapped_device *md, void *ptr)
1905{
1906 md->interface_ptr = ptr;
1907}
1908
1909void dm_get(struct mapped_device *md)
1910{
1911 atomic_inc(&md->holders);
1912 BUG_ON(test_bit(DMF_FREEING, &md->flags));
1913}
1914
1915int dm_hold(struct mapped_device *md)
1916{
1917 spin_lock(&_minor_lock);
1918 if (test_bit(DMF_FREEING, &md->flags)) {
1919 spin_unlock(&_minor_lock);
1920 return -EBUSY;
1921 }
1922 dm_get(md);
1923 spin_unlock(&_minor_lock);
1924 return 0;
1925}
1926EXPORT_SYMBOL_GPL(dm_hold);
1927
1928const char *dm_device_name(struct mapped_device *md)
1929{
1930 return md->name;
1931}
1932EXPORT_SYMBOL_GPL(dm_device_name);
1933
1934static void __dm_destroy(struct mapped_device *md, bool wait)
1935{
1936 struct request_queue *q = dm_get_md_queue(md);
1937 struct dm_table *map;
1938 int srcu_idx;
1939
1940 might_sleep();
1941
1942 spin_lock(&_minor_lock);
1943 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
1944 set_bit(DMF_FREEING, &md->flags);
1945 spin_unlock(&_minor_lock);
1946
1947 blk_set_queue_dying(q);
1948
1949 if (dm_request_based(md) && md->kworker_task)
1950 kthread_flush_worker(&md->kworker);
1951
1952 /*
1953 * Take suspend_lock so that presuspend and postsuspend methods
1954 * do not race with internal suspend.
1955 */
1956 mutex_lock(&md->suspend_lock);
1957 map = dm_get_live_table(md, &srcu_idx);
1958 if (!dm_suspended_md(md)) {
1959 dm_table_presuspend_targets(map);
1960 dm_table_postsuspend_targets(map);
1961 }
1962 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
1963 dm_put_live_table(md, srcu_idx);
1964 mutex_unlock(&md->suspend_lock);
1965
1966 /*
1967 * Rare, but there may be I/O requests still going to complete,
1968 * for example. Wait for all references to disappear.
1969 * No one should increment the reference count of the mapped_device,
1970 * after the mapped_device state becomes DMF_FREEING.
1971 */
1972 if (wait)
1973 while (atomic_read(&md->holders))
1974 msleep(1);
1975 else if (atomic_read(&md->holders))
1976 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
1977 dm_device_name(md), atomic_read(&md->holders));
1978
1979 dm_sysfs_exit(md);
1980 dm_table_destroy(__unbind(md));
1981 free_dev(md);
1982}
1983
1984void dm_destroy(struct mapped_device *md)
1985{
1986 __dm_destroy(md, true);
1987}
1988
1989void dm_destroy_immediate(struct mapped_device *md)
1990{
1991 __dm_destroy(md, false);
1992}
1993
1994void dm_put(struct mapped_device *md)
1995{
1996 atomic_dec(&md->holders);
1997}
1998EXPORT_SYMBOL_GPL(dm_put);
1999
2000static int dm_wait_for_completion(struct mapped_device *md, long task_state)
2001{
2002 int r = 0;
2003 DEFINE_WAIT(wait);
2004
2005 while (1) {
2006 prepare_to_wait(&md->wait, &wait, task_state);
2007
2008 if (!md_in_flight(md))
2009 break;
2010
2011 if (signal_pending_state(task_state, current)) {
2012 r = -EINTR;
2013 break;
2014 }
2015
2016 io_schedule();
2017 }
2018 finish_wait(&md->wait, &wait);
2019
2020 return r;
2021}
2022
2023/*
2024 * Process the deferred bios
2025 */
2026static void dm_wq_work(struct work_struct *work)
2027{
2028 struct mapped_device *md = container_of(work, struct mapped_device,
2029 work);
2030 struct bio *c;
2031 int srcu_idx;
2032 struct dm_table *map;
2033
2034 map = dm_get_live_table(md, &srcu_idx);
2035
2036 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2037 spin_lock_irq(&md->deferred_lock);
2038 c = bio_list_pop(&md->deferred);
2039 spin_unlock_irq(&md->deferred_lock);
2040
2041 if (!c)
2042 break;
2043
2044 if (dm_request_based(md))
2045 generic_make_request(c);
2046 else
2047 __split_and_process_bio(md, map, c);
2048 }
2049
2050 dm_put_live_table(md, srcu_idx);
2051}
2052
2053static void dm_queue_flush(struct mapped_device *md)
2054{
2055 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2056 smp_mb__after_atomic();
2057 queue_work(md->wq, &md->work);
2058}
2059
2060/*
2061 * Swap in a new table, returning the old one for the caller to destroy.
2062 */
2063struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2064{
2065 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2066 struct queue_limits limits;
2067 int r;
2068
2069 mutex_lock(&md->suspend_lock);
2070
2071 /* device must be suspended */
2072 if (!dm_suspended_md(md))
2073 goto out;
2074
2075 /*
2076 * If the new table has no data devices, retain the existing limits.
2077 * This helps multipath with queue_if_no_path if all paths disappear,
2078 * then new I/O is queued based on these limits, and then some paths
2079 * reappear.
2080 */
2081 if (dm_table_has_no_data_devices(table)) {
2082 live_map = dm_get_live_table_fast(md);
2083 if (live_map)
2084 limits = md->queue->limits;
2085 dm_put_live_table_fast(md);
2086 }
2087
2088 if (!live_map) {
2089 r = dm_calculate_queue_limits(table, &limits);
2090 if (r) {
2091 map = ERR_PTR(r);
2092 goto out;
2093 }
2094 }
2095
2096 map = __bind(md, table, &limits);
2097
2098out:
2099 mutex_unlock(&md->suspend_lock);
2100 return map;
2101}
2102
2103/*
2104 * Functions to lock and unlock any filesystem running on the
2105 * device.
2106 */
2107static int lock_fs(struct mapped_device *md)
2108{
2109 int r;
2110
2111 WARN_ON(md->frozen_sb);
2112
2113 md->frozen_sb = freeze_bdev(md->bdev);
2114 if (IS_ERR(md->frozen_sb)) {
2115 r = PTR_ERR(md->frozen_sb);
2116 md->frozen_sb = NULL;
2117 return r;
2118 }
2119
2120 set_bit(DMF_FROZEN, &md->flags);
2121
2122 return 0;
2123}
2124
2125static void unlock_fs(struct mapped_device *md)
2126{
2127 if (!test_bit(DMF_FROZEN, &md->flags))
2128 return;
2129
2130 thaw_bdev(md->bdev, md->frozen_sb);
2131 md->frozen_sb = NULL;
2132 clear_bit(DMF_FROZEN, &md->flags);
2133}
2134
2135/*
2136 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2137 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2138 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2139 *
2140 * If __dm_suspend returns 0, the device is completely quiescent
2141 * now. There is no request-processing activity. All new requests
2142 * are being added to md->deferred list.
2143 *
2144 * Caller must hold md->suspend_lock
2145 */
2146static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2147 unsigned suspend_flags, long task_state,
2148 int dmf_suspended_flag)
2149{
2150 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2151 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2152 int r;
2153
2154 lockdep_assert_held(&md->suspend_lock);
2155
2156 /*
2157 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2158 * This flag is cleared before dm_suspend returns.
2159 */
2160 if (noflush)
2161 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2162
2163 /*
2164 * This gets reverted if there's an error later and the targets
2165 * provide the .presuspend_undo hook.
2166 */
2167 dm_table_presuspend_targets(map);
2168
2169 /*
2170 * Flush I/O to the device.
2171 * Any I/O submitted after lock_fs() may not be flushed.
2172 * noflush takes precedence over do_lockfs.
2173 * (lock_fs() flushes I/Os and waits for them to complete.)
2174 */
2175 if (!noflush && do_lockfs) {
2176 r = lock_fs(md);
2177 if (r) {
2178 dm_table_presuspend_undo_targets(map);
2179 return r;
2180 }
2181 }
2182
2183 /*
2184 * Here we must make sure that no processes are submitting requests
2185 * to target drivers i.e. no one may be executing
2186 * __split_and_process_bio. This is called from dm_request and
2187 * dm_wq_work.
2188 *
2189 * To get all processes out of __split_and_process_bio in dm_request,
2190 * we take the write lock. To prevent any process from reentering
2191 * __split_and_process_bio from dm_request and quiesce the thread
2192 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
2193 * flush_workqueue(md->wq).
2194 */
2195 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2196 if (map)
2197 synchronize_srcu(&md->io_barrier);
2198
2199 /*
2200 * Stop md->queue before flushing md->wq in case request-based
2201 * dm defers requests to md->wq from md->queue.
2202 */
2203 if (dm_request_based(md)) {
2204 dm_stop_queue(md->queue);
2205 if (md->kworker_task)
2206 kthread_flush_worker(&md->kworker);
2207 }
2208
2209 flush_workqueue(md->wq);
2210
2211 /*
2212 * At this point no more requests are entering target request routines.
2213 * We call dm_wait_for_completion to wait for all existing requests
2214 * to finish.
2215 */
2216 r = dm_wait_for_completion(md, task_state);
2217 if (!r)
2218 set_bit(dmf_suspended_flag, &md->flags);
2219
2220 if (noflush)
2221 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2222 if (map)
2223 synchronize_srcu(&md->io_barrier);
2224
2225 /* were we interrupted ? */
2226 if (r < 0) {
2227 dm_queue_flush(md);
2228
2229 if (dm_request_based(md))
2230 dm_start_queue(md->queue);
2231
2232 unlock_fs(md);
2233 dm_table_presuspend_undo_targets(map);
2234 /* pushback list is already flushed, so skip flush */
2235 }
2236
2237 return r;
2238}
2239
2240/*
2241 * We need to be able to change a mapping table under a mounted
2242 * filesystem. For example we might want to move some data in
2243 * the background. Before the table can be swapped with
2244 * dm_bind_table, dm_suspend must be called to flush any in
2245 * flight bios and ensure that any further io gets deferred.
2246 */
2247/*
2248 * Suspend mechanism in request-based dm.
2249 *
2250 * 1. Flush all I/Os by lock_fs() if needed.
2251 * 2. Stop dispatching any I/O by stopping the request_queue.
2252 * 3. Wait for all in-flight I/Os to be completed or requeued.
2253 *
2254 * To abort suspend, start the request_queue.
2255 */
2256int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
2257{
2258 struct dm_table *map = NULL;
2259 int r = 0;
2260
2261retry:
2262 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2263
2264 if (dm_suspended_md(md)) {
2265 r = -EINVAL;
2266 goto out_unlock;
2267 }
2268
2269 if (dm_suspended_internally_md(md)) {
2270 /* already internally suspended, wait for internal resume */
2271 mutex_unlock(&md->suspend_lock);
2272 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2273 if (r)
2274 return r;
2275 goto retry;
2276 }
2277
2278 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2279
2280 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2281 if (r)
2282 goto out_unlock;
2283
2284 dm_table_postsuspend_targets(map);
2285
2286out_unlock:
2287 mutex_unlock(&md->suspend_lock);
2288 return r;
2289}
2290
2291static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2292{
2293 if (map) {
2294 int r = dm_table_resume_targets(map);
2295 if (r)
2296 return r;
2297 }
2298
2299 dm_queue_flush(md);
2300
2301 /*
2302 * Flushing deferred I/Os must be done after targets are resumed
2303 * so that mapping of targets can work correctly.
2304 * Request-based dm is queueing the deferred I/Os in its request_queue.
2305 */
2306 if (dm_request_based(md))
2307 dm_start_queue(md->queue);
2308
2309 unlock_fs(md);
2310
2311 return 0;
2312}
2313
2314int dm_resume(struct mapped_device *md)
2315{
2316 int r;
2317 struct dm_table *map = NULL;
2318
2319retry:
2320 r = -EINVAL;
2321 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2322
2323 if (!dm_suspended_md(md))
2324 goto out;
2325
2326 if (dm_suspended_internally_md(md)) {
2327 /* already internally suspended, wait for internal resume */
2328 mutex_unlock(&md->suspend_lock);
2329 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2330 if (r)
2331 return r;
2332 goto retry;
2333 }
2334
2335 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2336 if (!map || !dm_table_get_size(map))
2337 goto out;
2338
2339 r = __dm_resume(md, map);
2340 if (r)
2341 goto out;
2342
2343 clear_bit(DMF_SUSPENDED, &md->flags);
2344out:
2345 mutex_unlock(&md->suspend_lock);
2346
2347 return r;
2348}
2349
2350/*
2351 * Internal suspend/resume works like userspace-driven suspend. It waits
2352 * until all bios finish and prevents issuing new bios to the target drivers.
2353 * It may be used only from the kernel.
2354 */
2355
2356static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
2357{
2358 struct dm_table *map = NULL;
2359
2360 if (md->internal_suspend_count++)
2361 return; /* nested internal suspend */
2362
2363 if (dm_suspended_md(md)) {
2364 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2365 return; /* nest suspend */
2366 }
2367
2368 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2369
2370 /*
2371 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2372 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2373 * would require changing .presuspend to return an error -- avoid this
2374 * until there is a need for more elaborate variants of internal suspend.
2375 */
2376 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2377 DMF_SUSPENDED_INTERNALLY);
2378
2379 dm_table_postsuspend_targets(map);
2380}
2381
2382static void __dm_internal_resume(struct mapped_device *md)
2383{
2384 BUG_ON(!md->internal_suspend_count);
2385
2386 if (--md->internal_suspend_count)
2387 return; /* resume from nested internal suspend */
2388
2389 if (dm_suspended_md(md))
2390 goto done; /* resume from nested suspend */
2391
2392 /*
2393 * NOTE: existing callers don't need to call dm_table_resume_targets
2394 * (which may fail -- so best to avoid it for now by passing NULL map)
2395 */
2396 (void) __dm_resume(md, NULL);
2397
2398done:
2399 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2400 smp_mb__after_atomic();
2401 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
2402}
2403
2404void dm_internal_suspend_noflush(struct mapped_device *md)
2405{
2406 mutex_lock(&md->suspend_lock);
2407 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2408 mutex_unlock(&md->suspend_lock);
2409}
2410EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2411
2412void dm_internal_resume(struct mapped_device *md)
2413{
2414 mutex_lock(&md->suspend_lock);
2415 __dm_internal_resume(md);
2416 mutex_unlock(&md->suspend_lock);
2417}
2418EXPORT_SYMBOL_GPL(dm_internal_resume);
2419
2420/*
2421 * Fast variants of internal suspend/resume hold md->suspend_lock,
2422 * which prevents interaction with userspace-driven suspend.
2423 */
2424
2425void dm_internal_suspend_fast(struct mapped_device *md)
2426{
2427 mutex_lock(&md->suspend_lock);
2428 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2429 return;
2430
2431 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2432 synchronize_srcu(&md->io_barrier);
2433 flush_workqueue(md->wq);
2434 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
2435}
2436EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
2437
2438void dm_internal_resume_fast(struct mapped_device *md)
2439{
2440 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2441 goto done;
2442
2443 dm_queue_flush(md);
2444
2445done:
2446 mutex_unlock(&md->suspend_lock);
2447}
2448EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
2449
2450/*-----------------------------------------------------------------
2451 * Event notification.
2452 *---------------------------------------------------------------*/
2453int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
2454 unsigned cookie)
2455{
2456 char udev_cookie[DM_COOKIE_LENGTH];
2457 char *envp[] = { udev_cookie, NULL };
2458
2459 if (!cookie)
2460 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
2461 else {
2462 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
2463 DM_COOKIE_ENV_VAR_NAME, cookie);
2464 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
2465 action, envp);
2466 }
2467}
2468
2469uint32_t dm_next_uevent_seq(struct mapped_device *md)
2470{
2471 return atomic_add_return(1, &md->uevent_seq);
2472}
2473
2474uint32_t dm_get_event_nr(struct mapped_device *md)
2475{
2476 return atomic_read(&md->event_nr);
2477}
2478
2479int dm_wait_event(struct mapped_device *md, int event_nr)
2480{
2481 return wait_event_interruptible(md->eventq,
2482 (event_nr != atomic_read(&md->event_nr)));
2483}
2484
2485void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
2486{
2487 unsigned long flags;
2488
2489 spin_lock_irqsave(&md->uevent_lock, flags);
2490 list_add(elist, &md->uevent_list);
2491 spin_unlock_irqrestore(&md->uevent_lock, flags);
2492}
2493
2494/*
2495 * The gendisk is only valid as long as you have a reference
2496 * count on 'md'.
2497 */
2498struct gendisk *dm_disk(struct mapped_device *md)
2499{
2500 return md->disk;
2501}
2502EXPORT_SYMBOL_GPL(dm_disk);
2503
2504struct kobject *dm_kobject(struct mapped_device *md)
2505{
2506 return &md->kobj_holder.kobj;
2507}
2508
2509struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
2510{
2511 struct mapped_device *md;
2512
2513 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
2514
2515 if (test_bit(DMF_FREEING, &md->flags) ||
2516 dm_deleting_md(md))
2517 return NULL;
2518
2519 dm_get(md);
2520 return md;
2521}
2522
2523int dm_suspended_md(struct mapped_device *md)
2524{
2525 return test_bit(DMF_SUSPENDED, &md->flags);
2526}
2527
2528int dm_suspended_internally_md(struct mapped_device *md)
2529{
2530 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2531}
2532
2533int dm_test_deferred_remove_flag(struct mapped_device *md)
2534{
2535 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
2536}
2537
2538int dm_suspended(struct dm_target *ti)
2539{
2540 return dm_suspended_md(dm_table_get_md(ti->table));
2541}
2542EXPORT_SYMBOL_GPL(dm_suspended);
2543
2544int dm_noflush_suspending(struct dm_target *ti)
2545{
2546 return __noflush_suspending(dm_table_get_md(ti->table));
2547}
2548EXPORT_SYMBOL_GPL(dm_noflush_suspending);
2549
2550struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
2551 unsigned integrity, unsigned per_io_data_size)
2552{
2553 struct dm_md_mempools *pools = kzalloc_node(sizeof(*pools), GFP_KERNEL, md->numa_node_id);
2554 struct kmem_cache *cachep = NULL;
2555 unsigned int pool_size = 0;
2556 unsigned int front_pad;
2557
2558 if (!pools)
2559 return NULL;
2560
2561 switch (type) {
2562 case DM_TYPE_BIO_BASED:
2563 case DM_TYPE_DAX_BIO_BASED:
2564 cachep = _io_cache;
2565 pool_size = dm_get_reserved_bio_based_ios();
2566 front_pad = roundup(per_io_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
2567 break;
2568 case DM_TYPE_REQUEST_BASED:
2569 cachep = _rq_tio_cache;
2570 pool_size = dm_get_reserved_rq_based_ios();
2571 pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
2572 if (!pools->rq_pool)
2573 goto out;
2574 /* fall through to setup remaining rq-based pools */
2575 case DM_TYPE_MQ_REQUEST_BASED:
2576 if (!pool_size)
2577 pool_size = dm_get_reserved_rq_based_ios();
2578 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
2579 /* per_io_data_size is used for blk-mq pdu at queue allocation */
2580 break;
2581 default:
2582 BUG();
2583 }
2584
2585 if (cachep) {
2586 pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
2587 if (!pools->io_pool)
2588 goto out;
2589 }
2590
2591 pools->bs = bioset_create_nobvec(pool_size, front_pad);
2592 if (!pools->bs)
2593 goto out;
2594
2595 if (integrity && bioset_integrity_create(pools->bs, pool_size))
2596 goto out;
2597
2598 return pools;
2599
2600out:
2601 dm_free_md_mempools(pools);
2602
2603 return NULL;
2604}
2605
2606void dm_free_md_mempools(struct dm_md_mempools *pools)
2607{
2608 if (!pools)
2609 return;
2610
2611 mempool_destroy(pools->io_pool);
2612 mempool_destroy(pools->rq_pool);
2613
2614 if (pools->bs)
2615 bioset_free(pools->bs);
2616
2617 kfree(pools);
2618}
2619
2620struct dm_pr {
2621 u64 old_key;
2622 u64 new_key;
2623 u32 flags;
2624 bool fail_early;
2625};
2626
2627static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
2628 void *data)
2629{
2630 struct mapped_device *md = bdev->bd_disk->private_data;
2631 struct dm_table *table;
2632 struct dm_target *ti;
2633 int ret = -ENOTTY, srcu_idx;
2634
2635 table = dm_get_live_table(md, &srcu_idx);
2636 if (!table || !dm_table_get_size(table))
2637 goto out;
2638
2639 /* We only support devices that have a single target */
2640 if (dm_table_get_num_targets(table) != 1)
2641 goto out;
2642 ti = dm_table_get_target(table, 0);
2643
2644 ret = -EINVAL;
2645 if (!ti->type->iterate_devices)
2646 goto out;
2647
2648 ret = ti->type->iterate_devices(ti, fn, data);
2649out:
2650 dm_put_live_table(md, srcu_idx);
2651 return ret;
2652}
2653
2654/*
2655 * For register / unregister we need to manually call out to every path.
2656 */
2657static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
2658 sector_t start, sector_t len, void *data)
2659{
2660 struct dm_pr *pr = data;
2661 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
2662
2663 if (!ops || !ops->pr_register)
2664 return -EOPNOTSUPP;
2665 return ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
2666}
2667
2668static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
2669 u32 flags)
2670{
2671 struct dm_pr pr = {
2672 .old_key = old_key,
2673 .new_key = new_key,
2674 .flags = flags,
2675 .fail_early = true,
2676 };
2677 int ret;
2678
2679 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
2680 if (ret && new_key) {
2681 /* unregister all paths if we failed to register any path */
2682 pr.old_key = new_key;
2683 pr.new_key = 0;
2684 pr.flags = 0;
2685 pr.fail_early = false;
2686 dm_call_pr(bdev, __dm_pr_register, &pr);
2687 }
2688
2689 return ret;
2690}
2691
2692static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
2693 u32 flags)
2694{
2695 struct mapped_device *md = bdev->bd_disk->private_data;
2696 const struct pr_ops *ops;
2697 fmode_t mode;
2698 int r;
2699
2700 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
2701 if (r < 0)
2702 return r;
2703
2704 ops = bdev->bd_disk->fops->pr_ops;
2705 if (ops && ops->pr_reserve)
2706 r = ops->pr_reserve(bdev, key, type, flags);
2707 else
2708 r = -EOPNOTSUPP;
2709
2710 bdput(bdev);
2711 return r;
2712}
2713
2714static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
2715{
2716 struct mapped_device *md = bdev->bd_disk->private_data;
2717 const struct pr_ops *ops;
2718 fmode_t mode;
2719 int r;
2720
2721 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
2722 if (r < 0)
2723 return r;
2724
2725 ops = bdev->bd_disk->fops->pr_ops;
2726 if (ops && ops->pr_release)
2727 r = ops->pr_release(bdev, key, type);
2728 else
2729 r = -EOPNOTSUPP;
2730
2731 bdput(bdev);
2732 return r;
2733}
2734
2735static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
2736 enum pr_type type, bool abort)
2737{
2738 struct mapped_device *md = bdev->bd_disk->private_data;
2739 const struct pr_ops *ops;
2740 fmode_t mode;
2741 int r;
2742
2743 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
2744 if (r < 0)
2745 return r;
2746
2747 ops = bdev->bd_disk->fops->pr_ops;
2748 if (ops && ops->pr_preempt)
2749 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
2750 else
2751 r = -EOPNOTSUPP;
2752
2753 bdput(bdev);
2754 return r;
2755}
2756
2757static int dm_pr_clear(struct block_device *bdev, u64 key)
2758{
2759 struct mapped_device *md = bdev->bd_disk->private_data;
2760 const struct pr_ops *ops;
2761 fmode_t mode;
2762 int r;
2763
2764 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
2765 if (r < 0)
2766 return r;
2767
2768 ops = bdev->bd_disk->fops->pr_ops;
2769 if (ops && ops->pr_clear)
2770 r = ops->pr_clear(bdev, key);
2771 else
2772 r = -EOPNOTSUPP;
2773
2774 bdput(bdev);
2775 return r;
2776}
2777
2778static const struct pr_ops dm_pr_ops = {
2779 .pr_register = dm_pr_register,
2780 .pr_reserve = dm_pr_reserve,
2781 .pr_release = dm_pr_release,
2782 .pr_preempt = dm_pr_preempt,
2783 .pr_clear = dm_pr_clear,
2784};
2785
2786static const struct block_device_operations dm_blk_dops = {
2787 .open = dm_blk_open,
2788 .release = dm_blk_close,
2789 .ioctl = dm_blk_ioctl,
2790 .direct_access = dm_blk_direct_access,
2791 .getgeo = dm_blk_getgeo,
2792 .pr_ops = &dm_pr_ops,
2793 .owner = THIS_MODULE
2794};
2795
2796/*
2797 * module hooks
2798 */
2799module_init(dm_init);
2800module_exit(dm_exit);
2801
2802module_param(major, uint, 0);
2803MODULE_PARM_DESC(major, "The major number of the device mapper");
2804
2805module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
2806MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
2807
2808module_param(dm_numa_node, int, S_IRUGO | S_IWUSR);
2809MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
2810
2811MODULE_DESCRIPTION(DM_NAME " driver");
2812MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2813MODULE_LICENSE("GPL");
1/*
2 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm-core.h"
9#include "dm-rq.h"
10#include "dm-uevent.h"
11
12#include <linux/init.h>
13#include <linux/module.h>
14#include <linux/mutex.h>
15#include <linux/sched/signal.h>
16#include <linux/blkpg.h>
17#include <linux/bio.h>
18#include <linux/mempool.h>
19#include <linux/dax.h>
20#include <linux/slab.h>
21#include <linux/idr.h>
22#include <linux/uio.h>
23#include <linux/hdreg.h>
24#include <linux/delay.h>
25#include <linux/wait.h>
26#include <linux/pr.h>
27#include <linux/refcount.h>
28
29#define DM_MSG_PREFIX "core"
30
31/*
32 * Cookies are numeric values sent with CHANGE and REMOVE
33 * uevents while resuming, removing or renaming the device.
34 */
35#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
36#define DM_COOKIE_LENGTH 24
37
38static const char *_name = DM_NAME;
39
40static unsigned int major = 0;
41static unsigned int _major = 0;
42
43static DEFINE_IDR(_minor_idr);
44
45static DEFINE_SPINLOCK(_minor_lock);
46
47static void do_deferred_remove(struct work_struct *w);
48
49static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
50
51static struct workqueue_struct *deferred_remove_workqueue;
52
53atomic_t dm_global_event_nr = ATOMIC_INIT(0);
54DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
55
56void dm_issue_global_event(void)
57{
58 atomic_inc(&dm_global_event_nr);
59 wake_up(&dm_global_eventq);
60}
61
62/*
63 * One of these is allocated (on-stack) per original bio.
64 */
65struct clone_info {
66 struct dm_table *map;
67 struct bio *bio;
68 struct dm_io *io;
69 sector_t sector;
70 unsigned sector_count;
71};
72
73/*
74 * One of these is allocated per clone bio.
75 */
76#define DM_TIO_MAGIC 7282014
77struct dm_target_io {
78 unsigned magic;
79 struct dm_io *io;
80 struct dm_target *ti;
81 unsigned target_bio_nr;
82 unsigned *len_ptr;
83 bool inside_dm_io;
84 struct bio clone;
85};
86
87/*
88 * One of these is allocated per original bio.
89 * It contains the first clone used for that original.
90 */
91#define DM_IO_MAGIC 5191977
92struct dm_io {
93 unsigned magic;
94 struct mapped_device *md;
95 blk_status_t status;
96 atomic_t io_count;
97 struct bio *orig_bio;
98 unsigned long start_time;
99 spinlock_t endio_lock;
100 struct dm_stats_aux stats_aux;
101 /* last member of dm_target_io is 'struct bio' */
102 struct dm_target_io tio;
103};
104
105void *dm_per_bio_data(struct bio *bio, size_t data_size)
106{
107 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
108 if (!tio->inside_dm_io)
109 return (char *)bio - offsetof(struct dm_target_io, clone) - data_size;
110 return (char *)bio - offsetof(struct dm_target_io, clone) - offsetof(struct dm_io, tio) - data_size;
111}
112EXPORT_SYMBOL_GPL(dm_per_bio_data);
113
114struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
115{
116 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
117 if (io->magic == DM_IO_MAGIC)
118 return (struct bio *)((char *)io + offsetof(struct dm_io, tio) + offsetof(struct dm_target_io, clone));
119 BUG_ON(io->magic != DM_TIO_MAGIC);
120 return (struct bio *)((char *)io + offsetof(struct dm_target_io, clone));
121}
122EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
123
124unsigned dm_bio_get_target_bio_nr(const struct bio *bio)
125{
126 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
127}
128EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
129
130#define MINOR_ALLOCED ((void *)-1)
131
132/*
133 * Bits for the md->flags field.
134 */
135#define DMF_BLOCK_IO_FOR_SUSPEND 0
136#define DMF_SUSPENDED 1
137#define DMF_FROZEN 2
138#define DMF_FREEING 3
139#define DMF_DELETING 4
140#define DMF_NOFLUSH_SUSPENDING 5
141#define DMF_DEFERRED_REMOVE 6
142#define DMF_SUSPENDED_INTERNALLY 7
143
144#define DM_NUMA_NODE NUMA_NO_NODE
145static int dm_numa_node = DM_NUMA_NODE;
146
147/*
148 * For mempools pre-allocation at the table loading time.
149 */
150struct dm_md_mempools {
151 struct bio_set *bs;
152 struct bio_set *io_bs;
153};
154
155struct table_device {
156 struct list_head list;
157 refcount_t count;
158 struct dm_dev dm_dev;
159};
160
161static struct kmem_cache *_rq_tio_cache;
162static struct kmem_cache *_rq_cache;
163
164/*
165 * Bio-based DM's mempools' reserved IOs set by the user.
166 */
167#define RESERVED_BIO_BASED_IOS 16
168static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
169
170static int __dm_get_module_param_int(int *module_param, int min, int max)
171{
172 int param = READ_ONCE(*module_param);
173 int modified_param = 0;
174 bool modified = true;
175
176 if (param < min)
177 modified_param = min;
178 else if (param > max)
179 modified_param = max;
180 else
181 modified = false;
182
183 if (modified) {
184 (void)cmpxchg(module_param, param, modified_param);
185 param = modified_param;
186 }
187
188 return param;
189}
190
191unsigned __dm_get_module_param(unsigned *module_param,
192 unsigned def, unsigned max)
193{
194 unsigned param = READ_ONCE(*module_param);
195 unsigned modified_param = 0;
196
197 if (!param)
198 modified_param = def;
199 else if (param > max)
200 modified_param = max;
201
202 if (modified_param) {
203 (void)cmpxchg(module_param, param, modified_param);
204 param = modified_param;
205 }
206
207 return param;
208}
209
210unsigned dm_get_reserved_bio_based_ios(void)
211{
212 return __dm_get_module_param(&reserved_bio_based_ios,
213 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
214}
215EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
216
217static unsigned dm_get_numa_node(void)
218{
219 return __dm_get_module_param_int(&dm_numa_node,
220 DM_NUMA_NODE, num_online_nodes() - 1);
221}
222
223static int __init local_init(void)
224{
225 int r = -ENOMEM;
226
227 _rq_tio_cache = KMEM_CACHE(dm_rq_target_io, 0);
228 if (!_rq_tio_cache)
229 return r;
230
231 _rq_cache = kmem_cache_create("dm_old_clone_request", sizeof(struct request),
232 __alignof__(struct request), 0, NULL);
233 if (!_rq_cache)
234 goto out_free_rq_tio_cache;
235
236 r = dm_uevent_init();
237 if (r)
238 goto out_free_rq_cache;
239
240 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
241 if (!deferred_remove_workqueue) {
242 r = -ENOMEM;
243 goto out_uevent_exit;
244 }
245
246 _major = major;
247 r = register_blkdev(_major, _name);
248 if (r < 0)
249 goto out_free_workqueue;
250
251 if (!_major)
252 _major = r;
253
254 return 0;
255
256out_free_workqueue:
257 destroy_workqueue(deferred_remove_workqueue);
258out_uevent_exit:
259 dm_uevent_exit();
260out_free_rq_cache:
261 kmem_cache_destroy(_rq_cache);
262out_free_rq_tio_cache:
263 kmem_cache_destroy(_rq_tio_cache);
264
265 return r;
266}
267
268static void local_exit(void)
269{
270 flush_scheduled_work();
271 destroy_workqueue(deferred_remove_workqueue);
272
273 kmem_cache_destroy(_rq_cache);
274 kmem_cache_destroy(_rq_tio_cache);
275 unregister_blkdev(_major, _name);
276 dm_uevent_exit();
277
278 _major = 0;
279
280 DMINFO("cleaned up");
281}
282
283static int (*_inits[])(void) __initdata = {
284 local_init,
285 dm_target_init,
286 dm_linear_init,
287 dm_stripe_init,
288 dm_io_init,
289 dm_kcopyd_init,
290 dm_interface_init,
291 dm_statistics_init,
292};
293
294static void (*_exits[])(void) = {
295 local_exit,
296 dm_target_exit,
297 dm_linear_exit,
298 dm_stripe_exit,
299 dm_io_exit,
300 dm_kcopyd_exit,
301 dm_interface_exit,
302 dm_statistics_exit,
303};
304
305static int __init dm_init(void)
306{
307 const int count = ARRAY_SIZE(_inits);
308
309 int r, i;
310
311 for (i = 0; i < count; i++) {
312 r = _inits[i]();
313 if (r)
314 goto bad;
315 }
316
317 return 0;
318
319 bad:
320 while (i--)
321 _exits[i]();
322
323 return r;
324}
325
326static void __exit dm_exit(void)
327{
328 int i = ARRAY_SIZE(_exits);
329
330 while (i--)
331 _exits[i]();
332
333 /*
334 * Should be empty by this point.
335 */
336 idr_destroy(&_minor_idr);
337}
338
339/*
340 * Block device functions
341 */
342int dm_deleting_md(struct mapped_device *md)
343{
344 return test_bit(DMF_DELETING, &md->flags);
345}
346
347static int dm_blk_open(struct block_device *bdev, fmode_t mode)
348{
349 struct mapped_device *md;
350
351 spin_lock(&_minor_lock);
352
353 md = bdev->bd_disk->private_data;
354 if (!md)
355 goto out;
356
357 if (test_bit(DMF_FREEING, &md->flags) ||
358 dm_deleting_md(md)) {
359 md = NULL;
360 goto out;
361 }
362
363 dm_get(md);
364 atomic_inc(&md->open_count);
365out:
366 spin_unlock(&_minor_lock);
367
368 return md ? 0 : -ENXIO;
369}
370
371static void dm_blk_close(struct gendisk *disk, fmode_t mode)
372{
373 struct mapped_device *md;
374
375 spin_lock(&_minor_lock);
376
377 md = disk->private_data;
378 if (WARN_ON(!md))
379 goto out;
380
381 if (atomic_dec_and_test(&md->open_count) &&
382 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
383 queue_work(deferred_remove_workqueue, &deferred_remove_work);
384
385 dm_put(md);
386out:
387 spin_unlock(&_minor_lock);
388}
389
390int dm_open_count(struct mapped_device *md)
391{
392 return atomic_read(&md->open_count);
393}
394
395/*
396 * Guarantees nothing is using the device before it's deleted.
397 */
398int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
399{
400 int r = 0;
401
402 spin_lock(&_minor_lock);
403
404 if (dm_open_count(md)) {
405 r = -EBUSY;
406 if (mark_deferred)
407 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
408 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
409 r = -EEXIST;
410 else
411 set_bit(DMF_DELETING, &md->flags);
412
413 spin_unlock(&_minor_lock);
414
415 return r;
416}
417
418int dm_cancel_deferred_remove(struct mapped_device *md)
419{
420 int r = 0;
421
422 spin_lock(&_minor_lock);
423
424 if (test_bit(DMF_DELETING, &md->flags))
425 r = -EBUSY;
426 else
427 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
428
429 spin_unlock(&_minor_lock);
430
431 return r;
432}
433
434static void do_deferred_remove(struct work_struct *w)
435{
436 dm_deferred_remove();
437}
438
439sector_t dm_get_size(struct mapped_device *md)
440{
441 return get_capacity(md->disk);
442}
443
444struct request_queue *dm_get_md_queue(struct mapped_device *md)
445{
446 return md->queue;
447}
448
449struct dm_stats *dm_get_stats(struct mapped_device *md)
450{
451 return &md->stats;
452}
453
454static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
455{
456 struct mapped_device *md = bdev->bd_disk->private_data;
457
458 return dm_get_geometry(md, geo);
459}
460
461static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
462 struct block_device **bdev)
463 __acquires(md->io_barrier)
464{
465 struct dm_target *tgt;
466 struct dm_table *map;
467 int r;
468
469retry:
470 r = -ENOTTY;
471 map = dm_get_live_table(md, srcu_idx);
472 if (!map || !dm_table_get_size(map))
473 return r;
474
475 /* We only support devices that have a single target */
476 if (dm_table_get_num_targets(map) != 1)
477 return r;
478
479 tgt = dm_table_get_target(map, 0);
480 if (!tgt->type->prepare_ioctl)
481 return r;
482
483 if (dm_suspended_md(md))
484 return -EAGAIN;
485
486 r = tgt->type->prepare_ioctl(tgt, bdev);
487 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
488 dm_put_live_table(md, *srcu_idx);
489 msleep(10);
490 goto retry;
491 }
492
493 return r;
494}
495
496static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
497 __releases(md->io_barrier)
498{
499 dm_put_live_table(md, srcu_idx);
500}
501
502static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
503 unsigned int cmd, unsigned long arg)
504{
505 struct mapped_device *md = bdev->bd_disk->private_data;
506 int r, srcu_idx;
507
508 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
509 if (r < 0)
510 goto out;
511
512 if (r > 0) {
513 /*
514 * Target determined this ioctl is being issued against a
515 * subset of the parent bdev; require extra privileges.
516 */
517 if (!capable(CAP_SYS_RAWIO)) {
518 DMWARN_LIMIT(
519 "%s: sending ioctl %x to DM device without required privilege.",
520 current->comm, cmd);
521 r = -ENOIOCTLCMD;
522 goto out;
523 }
524 }
525
526 r = __blkdev_driver_ioctl(bdev, mode, cmd, arg);
527out:
528 dm_unprepare_ioctl(md, srcu_idx);
529 return r;
530}
531
532static void start_io_acct(struct dm_io *io);
533
534static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio)
535{
536 struct dm_io *io;
537 struct dm_target_io *tio;
538 struct bio *clone;
539
540 clone = bio_alloc_bioset(GFP_NOIO, 0, md->io_bs);
541 if (!clone)
542 return NULL;
543
544 tio = container_of(clone, struct dm_target_io, clone);
545 tio->inside_dm_io = true;
546 tio->io = NULL;
547
548 io = container_of(tio, struct dm_io, tio);
549 io->magic = DM_IO_MAGIC;
550 io->status = 0;
551 atomic_set(&io->io_count, 1);
552 io->orig_bio = bio;
553 io->md = md;
554 spin_lock_init(&io->endio_lock);
555
556 start_io_acct(io);
557
558 return io;
559}
560
561static void free_io(struct mapped_device *md, struct dm_io *io)
562{
563 bio_put(&io->tio.clone);
564}
565
566static struct dm_target_io *alloc_tio(struct clone_info *ci, struct dm_target *ti,
567 unsigned target_bio_nr, gfp_t gfp_mask)
568{
569 struct dm_target_io *tio;
570
571 if (!ci->io->tio.io) {
572 /* the dm_target_io embedded in ci->io is available */
573 tio = &ci->io->tio;
574 } else {
575 struct bio *clone = bio_alloc_bioset(gfp_mask, 0, ci->io->md->bs);
576 if (!clone)
577 return NULL;
578
579 tio = container_of(clone, struct dm_target_io, clone);
580 tio->inside_dm_io = false;
581 }
582
583 tio->magic = DM_TIO_MAGIC;
584 tio->io = ci->io;
585 tio->ti = ti;
586 tio->target_bio_nr = target_bio_nr;
587
588 return tio;
589}
590
591static void free_tio(struct dm_target_io *tio)
592{
593 if (tio->inside_dm_io)
594 return;
595 bio_put(&tio->clone);
596}
597
598int md_in_flight(struct mapped_device *md)
599{
600 return atomic_read(&md->pending[READ]) +
601 atomic_read(&md->pending[WRITE]);
602}
603
604static void start_io_acct(struct dm_io *io)
605{
606 struct mapped_device *md = io->md;
607 struct bio *bio = io->orig_bio;
608 int rw = bio_data_dir(bio);
609
610 io->start_time = jiffies;
611
612 generic_start_io_acct(md->queue, rw, bio_sectors(bio), &dm_disk(md)->part0);
613
614 atomic_set(&dm_disk(md)->part0.in_flight[rw],
615 atomic_inc_return(&md->pending[rw]));
616
617 if (unlikely(dm_stats_used(&md->stats)))
618 dm_stats_account_io(&md->stats, bio_data_dir(bio),
619 bio->bi_iter.bi_sector, bio_sectors(bio),
620 false, 0, &io->stats_aux);
621}
622
623static void end_io_acct(struct dm_io *io)
624{
625 struct mapped_device *md = io->md;
626 struct bio *bio = io->orig_bio;
627 unsigned long duration = jiffies - io->start_time;
628 int pending;
629 int rw = bio_data_dir(bio);
630
631 generic_end_io_acct(md->queue, rw, &dm_disk(md)->part0, io->start_time);
632
633 if (unlikely(dm_stats_used(&md->stats)))
634 dm_stats_account_io(&md->stats, bio_data_dir(bio),
635 bio->bi_iter.bi_sector, bio_sectors(bio),
636 true, duration, &io->stats_aux);
637
638 /*
639 * After this is decremented the bio must not be touched if it is
640 * a flush.
641 */
642 pending = atomic_dec_return(&md->pending[rw]);
643 atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
644 pending += atomic_read(&md->pending[rw^0x1]);
645
646 /* nudge anyone waiting on suspend queue */
647 if (!pending)
648 wake_up(&md->wait);
649}
650
651/*
652 * Add the bio to the list of deferred io.
653 */
654static void queue_io(struct mapped_device *md, struct bio *bio)
655{
656 unsigned long flags;
657
658 spin_lock_irqsave(&md->deferred_lock, flags);
659 bio_list_add(&md->deferred, bio);
660 spin_unlock_irqrestore(&md->deferred_lock, flags);
661 queue_work(md->wq, &md->work);
662}
663
664/*
665 * Everyone (including functions in this file), should use this
666 * function to access the md->map field, and make sure they call
667 * dm_put_live_table() when finished.
668 */
669struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
670{
671 *srcu_idx = srcu_read_lock(&md->io_barrier);
672
673 return srcu_dereference(md->map, &md->io_barrier);
674}
675
676void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
677{
678 srcu_read_unlock(&md->io_barrier, srcu_idx);
679}
680
681void dm_sync_table(struct mapped_device *md)
682{
683 synchronize_srcu(&md->io_barrier);
684 synchronize_rcu_expedited();
685}
686
687/*
688 * A fast alternative to dm_get_live_table/dm_put_live_table.
689 * The caller must not block between these two functions.
690 */
691static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
692{
693 rcu_read_lock();
694 return rcu_dereference(md->map);
695}
696
697static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
698{
699 rcu_read_unlock();
700}
701
702static char *_dm_claim_ptr = "I belong to device-mapper";
703
704/*
705 * Open a table device so we can use it as a map destination.
706 */
707static int open_table_device(struct table_device *td, dev_t dev,
708 struct mapped_device *md)
709{
710 struct block_device *bdev;
711
712 int r;
713
714 BUG_ON(td->dm_dev.bdev);
715
716 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _dm_claim_ptr);
717 if (IS_ERR(bdev))
718 return PTR_ERR(bdev);
719
720 r = bd_link_disk_holder(bdev, dm_disk(md));
721 if (r) {
722 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
723 return r;
724 }
725
726 td->dm_dev.bdev = bdev;
727 td->dm_dev.dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
728 return 0;
729}
730
731/*
732 * Close a table device that we've been using.
733 */
734static void close_table_device(struct table_device *td, struct mapped_device *md)
735{
736 if (!td->dm_dev.bdev)
737 return;
738
739 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
740 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
741 put_dax(td->dm_dev.dax_dev);
742 td->dm_dev.bdev = NULL;
743 td->dm_dev.dax_dev = NULL;
744}
745
746static struct table_device *find_table_device(struct list_head *l, dev_t dev,
747 fmode_t mode) {
748 struct table_device *td;
749
750 list_for_each_entry(td, l, list)
751 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
752 return td;
753
754 return NULL;
755}
756
757int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
758 struct dm_dev **result) {
759 int r;
760 struct table_device *td;
761
762 mutex_lock(&md->table_devices_lock);
763 td = find_table_device(&md->table_devices, dev, mode);
764 if (!td) {
765 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
766 if (!td) {
767 mutex_unlock(&md->table_devices_lock);
768 return -ENOMEM;
769 }
770
771 td->dm_dev.mode = mode;
772 td->dm_dev.bdev = NULL;
773
774 if ((r = open_table_device(td, dev, md))) {
775 mutex_unlock(&md->table_devices_lock);
776 kfree(td);
777 return r;
778 }
779
780 format_dev_t(td->dm_dev.name, dev);
781
782 refcount_set(&td->count, 1);
783 list_add(&td->list, &md->table_devices);
784 } else {
785 refcount_inc(&td->count);
786 }
787 mutex_unlock(&md->table_devices_lock);
788
789 *result = &td->dm_dev;
790 return 0;
791}
792EXPORT_SYMBOL_GPL(dm_get_table_device);
793
794void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
795{
796 struct table_device *td = container_of(d, struct table_device, dm_dev);
797
798 mutex_lock(&md->table_devices_lock);
799 if (refcount_dec_and_test(&td->count)) {
800 close_table_device(td, md);
801 list_del(&td->list);
802 kfree(td);
803 }
804 mutex_unlock(&md->table_devices_lock);
805}
806EXPORT_SYMBOL(dm_put_table_device);
807
808static void free_table_devices(struct list_head *devices)
809{
810 struct list_head *tmp, *next;
811
812 list_for_each_safe(tmp, next, devices) {
813 struct table_device *td = list_entry(tmp, struct table_device, list);
814
815 DMWARN("dm_destroy: %s still exists with %d references",
816 td->dm_dev.name, refcount_read(&td->count));
817 kfree(td);
818 }
819}
820
821/*
822 * Get the geometry associated with a dm device
823 */
824int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
825{
826 *geo = md->geometry;
827
828 return 0;
829}
830
831/*
832 * Set the geometry of a device.
833 */
834int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
835{
836 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
837
838 if (geo->start > sz) {
839 DMWARN("Start sector is beyond the geometry limits.");
840 return -EINVAL;
841 }
842
843 md->geometry = *geo;
844
845 return 0;
846}
847
848static int __noflush_suspending(struct mapped_device *md)
849{
850 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
851}
852
853/*
854 * Decrements the number of outstanding ios that a bio has been
855 * cloned into, completing the original io if necc.
856 */
857static void dec_pending(struct dm_io *io, blk_status_t error)
858{
859 unsigned long flags;
860 blk_status_t io_error;
861 struct bio *bio;
862 struct mapped_device *md = io->md;
863
864 /* Push-back supersedes any I/O errors */
865 if (unlikely(error)) {
866 spin_lock_irqsave(&io->endio_lock, flags);
867 if (!(io->status == BLK_STS_DM_REQUEUE && __noflush_suspending(md)))
868 io->status = error;
869 spin_unlock_irqrestore(&io->endio_lock, flags);
870 }
871
872 if (atomic_dec_and_test(&io->io_count)) {
873 if (io->status == BLK_STS_DM_REQUEUE) {
874 /*
875 * Target requested pushing back the I/O.
876 */
877 spin_lock_irqsave(&md->deferred_lock, flags);
878 if (__noflush_suspending(md))
879 /* NOTE early return due to BLK_STS_DM_REQUEUE below */
880 bio_list_add_head(&md->deferred, io->orig_bio);
881 else
882 /* noflush suspend was interrupted. */
883 io->status = BLK_STS_IOERR;
884 spin_unlock_irqrestore(&md->deferred_lock, flags);
885 }
886
887 io_error = io->status;
888 bio = io->orig_bio;
889 end_io_acct(io);
890 free_io(md, io);
891
892 if (io_error == BLK_STS_DM_REQUEUE)
893 return;
894
895 if ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size) {
896 /*
897 * Preflush done for flush with data, reissue
898 * without REQ_PREFLUSH.
899 */
900 bio->bi_opf &= ~REQ_PREFLUSH;
901 queue_io(md, bio);
902 } else {
903 /* done with normal IO or empty flush */
904 if (io_error)
905 bio->bi_status = io_error;
906 bio_endio(bio);
907 }
908 }
909}
910
911void disable_write_same(struct mapped_device *md)
912{
913 struct queue_limits *limits = dm_get_queue_limits(md);
914
915 /* device doesn't really support WRITE SAME, disable it */
916 limits->max_write_same_sectors = 0;
917}
918
919void disable_write_zeroes(struct mapped_device *md)
920{
921 struct queue_limits *limits = dm_get_queue_limits(md);
922
923 /* device doesn't really support WRITE ZEROES, disable it */
924 limits->max_write_zeroes_sectors = 0;
925}
926
927static void clone_endio(struct bio *bio)
928{
929 blk_status_t error = bio->bi_status;
930 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
931 struct dm_io *io = tio->io;
932 struct mapped_device *md = tio->io->md;
933 dm_endio_fn endio = tio->ti->type->end_io;
934
935 if (unlikely(error == BLK_STS_TARGET) && md->type != DM_TYPE_NVME_BIO_BASED) {
936 if (bio_op(bio) == REQ_OP_WRITE_SAME &&
937 !bio->bi_disk->queue->limits.max_write_same_sectors)
938 disable_write_same(md);
939 if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
940 !bio->bi_disk->queue->limits.max_write_zeroes_sectors)
941 disable_write_zeroes(md);
942 }
943
944 if (endio) {
945 int r = endio(tio->ti, bio, &error);
946 switch (r) {
947 case DM_ENDIO_REQUEUE:
948 error = BLK_STS_DM_REQUEUE;
949 /*FALLTHRU*/
950 case DM_ENDIO_DONE:
951 break;
952 case DM_ENDIO_INCOMPLETE:
953 /* The target will handle the io */
954 return;
955 default:
956 DMWARN("unimplemented target endio return value: %d", r);
957 BUG();
958 }
959 }
960
961 free_tio(tio);
962 dec_pending(io, error);
963}
964
965/*
966 * Return maximum size of I/O possible at the supplied sector up to the current
967 * target boundary.
968 */
969static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
970{
971 sector_t target_offset = dm_target_offset(ti, sector);
972
973 return ti->len - target_offset;
974}
975
976static sector_t max_io_len(sector_t sector, struct dm_target *ti)
977{
978 sector_t len = max_io_len_target_boundary(sector, ti);
979 sector_t offset, max_len;
980
981 /*
982 * Does the target need to split even further?
983 */
984 if (ti->max_io_len) {
985 offset = dm_target_offset(ti, sector);
986 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
987 max_len = sector_div(offset, ti->max_io_len);
988 else
989 max_len = offset & (ti->max_io_len - 1);
990 max_len = ti->max_io_len - max_len;
991
992 if (len > max_len)
993 len = max_len;
994 }
995
996 return len;
997}
998
999int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1000{
1001 if (len > UINT_MAX) {
1002 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1003 (unsigned long long)len, UINT_MAX);
1004 ti->error = "Maximum size of target IO is too large";
1005 return -EINVAL;
1006 }
1007
1008 /*
1009 * BIO based queue uses its own splitting. When multipage bvecs
1010 * is switched on, size of the incoming bio may be too big to
1011 * be handled in some targets, such as crypt.
1012 *
1013 * When these targets are ready for the big bio, we can remove
1014 * the limit.
1015 */
1016 ti->max_io_len = min_t(uint32_t, len, BIO_MAX_PAGES * PAGE_SIZE);
1017
1018 return 0;
1019}
1020EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1021
1022static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1023 sector_t sector, int *srcu_idx)
1024 __acquires(md->io_barrier)
1025{
1026 struct dm_table *map;
1027 struct dm_target *ti;
1028
1029 map = dm_get_live_table(md, srcu_idx);
1030 if (!map)
1031 return NULL;
1032
1033 ti = dm_table_find_target(map, sector);
1034 if (!dm_target_is_valid(ti))
1035 return NULL;
1036
1037 return ti;
1038}
1039
1040static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1041 long nr_pages, void **kaddr, pfn_t *pfn)
1042{
1043 struct mapped_device *md = dax_get_private(dax_dev);
1044 sector_t sector = pgoff * PAGE_SECTORS;
1045 struct dm_target *ti;
1046 long len, ret = -EIO;
1047 int srcu_idx;
1048
1049 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1050
1051 if (!ti)
1052 goto out;
1053 if (!ti->type->direct_access)
1054 goto out;
1055 len = max_io_len(sector, ti) / PAGE_SECTORS;
1056 if (len < 1)
1057 goto out;
1058 nr_pages = min(len, nr_pages);
1059 if (ti->type->direct_access)
1060 ret = ti->type->direct_access(ti, pgoff, nr_pages, kaddr, pfn);
1061
1062 out:
1063 dm_put_live_table(md, srcu_idx);
1064
1065 return ret;
1066}
1067
1068static size_t dm_dax_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
1069 void *addr, size_t bytes, struct iov_iter *i)
1070{
1071 struct mapped_device *md = dax_get_private(dax_dev);
1072 sector_t sector = pgoff * PAGE_SECTORS;
1073 struct dm_target *ti;
1074 long ret = 0;
1075 int srcu_idx;
1076
1077 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1078
1079 if (!ti)
1080 goto out;
1081 if (!ti->type->dax_copy_from_iter) {
1082 ret = copy_from_iter(addr, bytes, i);
1083 goto out;
1084 }
1085 ret = ti->type->dax_copy_from_iter(ti, pgoff, addr, bytes, i);
1086 out:
1087 dm_put_live_table(md, srcu_idx);
1088
1089 return ret;
1090}
1091
1092/*
1093 * A target may call dm_accept_partial_bio only from the map routine. It is
1094 * allowed for all bio types except REQ_PREFLUSH and REQ_OP_ZONE_RESET.
1095 *
1096 * dm_accept_partial_bio informs the dm that the target only wants to process
1097 * additional n_sectors sectors of the bio and the rest of the data should be
1098 * sent in a next bio.
1099 *
1100 * A diagram that explains the arithmetics:
1101 * +--------------------+---------------+-------+
1102 * | 1 | 2 | 3 |
1103 * +--------------------+---------------+-------+
1104 *
1105 * <-------------- *tio->len_ptr --------------->
1106 * <------- bi_size ------->
1107 * <-- n_sectors -->
1108 *
1109 * Region 1 was already iterated over with bio_advance or similar function.
1110 * (it may be empty if the target doesn't use bio_advance)
1111 * Region 2 is the remaining bio size that the target wants to process.
1112 * (it may be empty if region 1 is non-empty, although there is no reason
1113 * to make it empty)
1114 * The target requires that region 3 is to be sent in the next bio.
1115 *
1116 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1117 * the partially processed part (the sum of regions 1+2) must be the same for all
1118 * copies of the bio.
1119 */
1120void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1121{
1122 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1123 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1124 BUG_ON(bio->bi_opf & REQ_PREFLUSH);
1125 BUG_ON(bi_size > *tio->len_ptr);
1126 BUG_ON(n_sectors > bi_size);
1127 *tio->len_ptr -= bi_size - n_sectors;
1128 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1129}
1130EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1131
1132/*
1133 * The zone descriptors obtained with a zone report indicate
1134 * zone positions within the target device. The zone descriptors
1135 * must be remapped to match their position within the dm device.
1136 * A target may call dm_remap_zone_report after completion of a
1137 * REQ_OP_ZONE_REPORT bio to remap the zone descriptors obtained
1138 * from the target device mapping to the dm device.
1139 */
1140void dm_remap_zone_report(struct dm_target *ti, struct bio *bio, sector_t start)
1141{
1142#ifdef CONFIG_BLK_DEV_ZONED
1143 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1144 struct bio *report_bio = tio->io->orig_bio;
1145 struct blk_zone_report_hdr *hdr = NULL;
1146 struct blk_zone *zone;
1147 unsigned int nr_rep = 0;
1148 unsigned int ofst;
1149 struct bio_vec bvec;
1150 struct bvec_iter iter;
1151 void *addr;
1152
1153 if (bio->bi_status)
1154 return;
1155
1156 /*
1157 * Remap the start sector of the reported zones. For sequential zones,
1158 * also remap the write pointer position.
1159 */
1160 bio_for_each_segment(bvec, report_bio, iter) {
1161 addr = kmap_atomic(bvec.bv_page);
1162
1163 /* Remember the report header in the first page */
1164 if (!hdr) {
1165 hdr = addr;
1166 ofst = sizeof(struct blk_zone_report_hdr);
1167 } else
1168 ofst = 0;
1169
1170 /* Set zones start sector */
1171 while (hdr->nr_zones && ofst < bvec.bv_len) {
1172 zone = addr + ofst;
1173 if (zone->start >= start + ti->len) {
1174 hdr->nr_zones = 0;
1175 break;
1176 }
1177 zone->start = zone->start + ti->begin - start;
1178 if (zone->type != BLK_ZONE_TYPE_CONVENTIONAL) {
1179 if (zone->cond == BLK_ZONE_COND_FULL)
1180 zone->wp = zone->start + zone->len;
1181 else if (zone->cond == BLK_ZONE_COND_EMPTY)
1182 zone->wp = zone->start;
1183 else
1184 zone->wp = zone->wp + ti->begin - start;
1185 }
1186 ofst += sizeof(struct blk_zone);
1187 hdr->nr_zones--;
1188 nr_rep++;
1189 }
1190
1191 if (addr != hdr)
1192 kunmap_atomic(addr);
1193
1194 if (!hdr->nr_zones)
1195 break;
1196 }
1197
1198 if (hdr) {
1199 hdr->nr_zones = nr_rep;
1200 kunmap_atomic(hdr);
1201 }
1202
1203 bio_advance(report_bio, report_bio->bi_iter.bi_size);
1204
1205#else /* !CONFIG_BLK_DEV_ZONED */
1206 bio->bi_status = BLK_STS_NOTSUPP;
1207#endif
1208}
1209EXPORT_SYMBOL_GPL(dm_remap_zone_report);
1210
1211static blk_qc_t __map_bio(struct dm_target_io *tio)
1212{
1213 int r;
1214 sector_t sector;
1215 struct bio *clone = &tio->clone;
1216 struct dm_io *io = tio->io;
1217 struct mapped_device *md = io->md;
1218 struct dm_target *ti = tio->ti;
1219 blk_qc_t ret = BLK_QC_T_NONE;
1220
1221 clone->bi_end_io = clone_endio;
1222
1223 /*
1224 * Map the clone. If r == 0 we don't need to do
1225 * anything, the target has assumed ownership of
1226 * this io.
1227 */
1228 atomic_inc(&io->io_count);
1229 sector = clone->bi_iter.bi_sector;
1230
1231 r = ti->type->map(ti, clone);
1232 switch (r) {
1233 case DM_MAPIO_SUBMITTED:
1234 break;
1235 case DM_MAPIO_REMAPPED:
1236 /* the bio has been remapped so dispatch it */
1237 trace_block_bio_remap(clone->bi_disk->queue, clone,
1238 bio_dev(io->orig_bio), sector);
1239 if (md->type == DM_TYPE_NVME_BIO_BASED)
1240 ret = direct_make_request(clone);
1241 else
1242 ret = generic_make_request(clone);
1243 break;
1244 case DM_MAPIO_KILL:
1245 free_tio(tio);
1246 dec_pending(io, BLK_STS_IOERR);
1247 break;
1248 case DM_MAPIO_REQUEUE:
1249 free_tio(tio);
1250 dec_pending(io, BLK_STS_DM_REQUEUE);
1251 break;
1252 default:
1253 DMWARN("unimplemented target map return value: %d", r);
1254 BUG();
1255 }
1256
1257 return ret;
1258}
1259
1260static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1261{
1262 bio->bi_iter.bi_sector = sector;
1263 bio->bi_iter.bi_size = to_bytes(len);
1264}
1265
1266/*
1267 * Creates a bio that consists of range of complete bvecs.
1268 */
1269static int clone_bio(struct dm_target_io *tio, struct bio *bio,
1270 sector_t sector, unsigned len)
1271{
1272 struct bio *clone = &tio->clone;
1273
1274 __bio_clone_fast(clone, bio);
1275
1276 if (unlikely(bio_integrity(bio) != NULL)) {
1277 int r;
1278
1279 if (unlikely(!dm_target_has_integrity(tio->ti->type) &&
1280 !dm_target_passes_integrity(tio->ti->type))) {
1281 DMWARN("%s: the target %s doesn't support integrity data.",
1282 dm_device_name(tio->io->md),
1283 tio->ti->type->name);
1284 return -EIO;
1285 }
1286
1287 r = bio_integrity_clone(clone, bio, GFP_NOIO);
1288 if (r < 0)
1289 return r;
1290 }
1291
1292 if (bio_op(bio) != REQ_OP_ZONE_REPORT)
1293 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1294 clone->bi_iter.bi_size = to_bytes(len);
1295
1296 if (unlikely(bio_integrity(bio) != NULL))
1297 bio_integrity_trim(clone);
1298
1299 return 0;
1300}
1301
1302static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1303 struct dm_target *ti, unsigned num_bios)
1304{
1305 struct dm_target_io *tio;
1306 int try;
1307
1308 if (!num_bios)
1309 return;
1310
1311 if (num_bios == 1) {
1312 tio = alloc_tio(ci, ti, 0, GFP_NOIO);
1313 bio_list_add(blist, &tio->clone);
1314 return;
1315 }
1316
1317 for (try = 0; try < 2; try++) {
1318 int bio_nr;
1319 struct bio *bio;
1320
1321 if (try)
1322 mutex_lock(&ci->io->md->table_devices_lock);
1323 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1324 tio = alloc_tio(ci, ti, bio_nr, try ? GFP_NOIO : GFP_NOWAIT);
1325 if (!tio)
1326 break;
1327
1328 bio_list_add(blist, &tio->clone);
1329 }
1330 if (try)
1331 mutex_unlock(&ci->io->md->table_devices_lock);
1332 if (bio_nr == num_bios)
1333 return;
1334
1335 while ((bio = bio_list_pop(blist))) {
1336 tio = container_of(bio, struct dm_target_io, clone);
1337 free_tio(tio);
1338 }
1339 }
1340}
1341
1342static blk_qc_t __clone_and_map_simple_bio(struct clone_info *ci,
1343 struct dm_target_io *tio, unsigned *len)
1344{
1345 struct bio *clone = &tio->clone;
1346
1347 tio->len_ptr = len;
1348
1349 __bio_clone_fast(clone, ci->bio);
1350 if (len)
1351 bio_setup_sector(clone, ci->sector, *len);
1352
1353 return __map_bio(tio);
1354}
1355
1356static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1357 unsigned num_bios, unsigned *len)
1358{
1359 struct bio_list blist = BIO_EMPTY_LIST;
1360 struct bio *bio;
1361 struct dm_target_io *tio;
1362
1363 alloc_multiple_bios(&blist, ci, ti, num_bios);
1364
1365 while ((bio = bio_list_pop(&blist))) {
1366 tio = container_of(bio, struct dm_target_io, clone);
1367 (void) __clone_and_map_simple_bio(ci, tio, len);
1368 }
1369}
1370
1371static int __send_empty_flush(struct clone_info *ci)
1372{
1373 unsigned target_nr = 0;
1374 struct dm_target *ti;
1375
1376 BUG_ON(bio_has_data(ci->bio));
1377 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1378 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1379
1380 return 0;
1381}
1382
1383static int __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1384 sector_t sector, unsigned *len)
1385{
1386 struct bio *bio = ci->bio;
1387 struct dm_target_io *tio;
1388 int r;
1389
1390 tio = alloc_tio(ci, ti, 0, GFP_NOIO);
1391 tio->len_ptr = len;
1392 r = clone_bio(tio, bio, sector, *len);
1393 if (r < 0) {
1394 free_tio(tio);
1395 return r;
1396 }
1397 (void) __map_bio(tio);
1398
1399 return 0;
1400}
1401
1402typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1403
1404static unsigned get_num_discard_bios(struct dm_target *ti)
1405{
1406 return ti->num_discard_bios;
1407}
1408
1409static unsigned get_num_secure_erase_bios(struct dm_target *ti)
1410{
1411 return ti->num_secure_erase_bios;
1412}
1413
1414static unsigned get_num_write_same_bios(struct dm_target *ti)
1415{
1416 return ti->num_write_same_bios;
1417}
1418
1419static unsigned get_num_write_zeroes_bios(struct dm_target *ti)
1420{
1421 return ti->num_write_zeroes_bios;
1422}
1423
1424typedef bool (*is_split_required_fn)(struct dm_target *ti);
1425
1426static bool is_split_required_for_discard(struct dm_target *ti)
1427{
1428 return ti->split_discard_bios;
1429}
1430
1431static int __send_changing_extent_only(struct clone_info *ci, struct dm_target *ti,
1432 get_num_bios_fn get_num_bios,
1433 is_split_required_fn is_split_required)
1434{
1435 unsigned len;
1436 unsigned num_bios;
1437
1438 /*
1439 * Even though the device advertised support for this type of
1440 * request, that does not mean every target supports it, and
1441 * reconfiguration might also have changed that since the
1442 * check was performed.
1443 */
1444 num_bios = get_num_bios ? get_num_bios(ti) : 0;
1445 if (!num_bios)
1446 return -EOPNOTSUPP;
1447
1448 if (is_split_required && !is_split_required(ti))
1449 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1450 else
1451 len = min((sector_t)ci->sector_count, max_io_len(ci->sector, ti));
1452
1453 __send_duplicate_bios(ci, ti, num_bios, &len);
1454
1455 ci->sector += len;
1456 ci->sector_count -= len;
1457
1458 return 0;
1459}
1460
1461static int __send_discard(struct clone_info *ci, struct dm_target *ti)
1462{
1463 return __send_changing_extent_only(ci, ti, get_num_discard_bios,
1464 is_split_required_for_discard);
1465}
1466
1467static int __send_secure_erase(struct clone_info *ci, struct dm_target *ti)
1468{
1469 return __send_changing_extent_only(ci, ti, get_num_secure_erase_bios, NULL);
1470}
1471
1472static int __send_write_same(struct clone_info *ci, struct dm_target *ti)
1473{
1474 return __send_changing_extent_only(ci, ti, get_num_write_same_bios, NULL);
1475}
1476
1477static int __send_write_zeroes(struct clone_info *ci, struct dm_target *ti)
1478{
1479 return __send_changing_extent_only(ci, ti, get_num_write_zeroes_bios, NULL);
1480}
1481
1482static bool __process_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1483 int *result)
1484{
1485 struct bio *bio = ci->bio;
1486
1487 if (bio_op(bio) == REQ_OP_DISCARD)
1488 *result = __send_discard(ci, ti);
1489 else if (bio_op(bio) == REQ_OP_SECURE_ERASE)
1490 *result = __send_secure_erase(ci, ti);
1491 else if (bio_op(bio) == REQ_OP_WRITE_SAME)
1492 *result = __send_write_same(ci, ti);
1493 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
1494 *result = __send_write_zeroes(ci, ti);
1495 else
1496 return false;
1497
1498 return true;
1499}
1500
1501/*
1502 * Select the correct strategy for processing a non-flush bio.
1503 */
1504static int __split_and_process_non_flush(struct clone_info *ci)
1505{
1506 struct bio *bio = ci->bio;
1507 struct dm_target *ti;
1508 unsigned len;
1509 int r;
1510
1511 ti = dm_table_find_target(ci->map, ci->sector);
1512 if (!dm_target_is_valid(ti))
1513 return -EIO;
1514
1515 if (unlikely(__process_abnormal_io(ci, ti, &r)))
1516 return r;
1517
1518 if (bio_op(bio) == REQ_OP_ZONE_REPORT)
1519 len = ci->sector_count;
1520 else
1521 len = min_t(sector_t, max_io_len(ci->sector, ti),
1522 ci->sector_count);
1523
1524 r = __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1525 if (r < 0)
1526 return r;
1527
1528 ci->sector += len;
1529 ci->sector_count -= len;
1530
1531 return 0;
1532}
1533
1534static void init_clone_info(struct clone_info *ci, struct mapped_device *md,
1535 struct dm_table *map, struct bio *bio)
1536{
1537 ci->map = map;
1538 ci->io = alloc_io(md, bio);
1539 ci->sector = bio->bi_iter.bi_sector;
1540}
1541
1542/*
1543 * Entry point to split a bio into clones and submit them to the targets.
1544 */
1545static blk_qc_t __split_and_process_bio(struct mapped_device *md,
1546 struct dm_table *map, struct bio *bio)
1547{
1548 struct clone_info ci;
1549 blk_qc_t ret = BLK_QC_T_NONE;
1550 int error = 0;
1551
1552 if (unlikely(!map)) {
1553 bio_io_error(bio);
1554 return ret;
1555 }
1556
1557 init_clone_info(&ci, md, map, bio);
1558
1559 if (bio->bi_opf & REQ_PREFLUSH) {
1560 ci.bio = &ci.io->md->flush_bio;
1561 ci.sector_count = 0;
1562 error = __send_empty_flush(&ci);
1563 /* dec_pending submits any data associated with flush */
1564 } else if (bio_op(bio) == REQ_OP_ZONE_RESET) {
1565 ci.bio = bio;
1566 ci.sector_count = 0;
1567 error = __split_and_process_non_flush(&ci);
1568 } else {
1569 ci.bio = bio;
1570 ci.sector_count = bio_sectors(bio);
1571 while (ci.sector_count && !error) {
1572 error = __split_and_process_non_flush(&ci);
1573 if (current->bio_list && ci.sector_count && !error) {
1574 /*
1575 * Remainder must be passed to generic_make_request()
1576 * so that it gets handled *after* bios already submitted
1577 * have been completely processed.
1578 * We take a clone of the original to store in
1579 * ci.io->orig_bio to be used by end_io_acct() and
1580 * for dec_pending to use for completion handling.
1581 * As this path is not used for REQ_OP_ZONE_REPORT,
1582 * the usage of io->orig_bio in dm_remap_zone_report()
1583 * won't be affected by this reassignment.
1584 */
1585 struct bio *b = bio_clone_bioset(bio, GFP_NOIO,
1586 md->queue->bio_split);
1587 ci.io->orig_bio = b;
1588 bio_advance(bio, (bio_sectors(bio) - ci.sector_count) << 9);
1589 bio_chain(b, bio);
1590 ret = generic_make_request(bio);
1591 break;
1592 }
1593 }
1594 }
1595
1596 /* drop the extra reference count */
1597 dec_pending(ci.io, errno_to_blk_status(error));
1598 return ret;
1599}
1600
1601/*
1602 * Optimized variant of __split_and_process_bio that leverages the
1603 * fact that targets that use it do _not_ have a need to split bios.
1604 */
1605static blk_qc_t __process_bio(struct mapped_device *md,
1606 struct dm_table *map, struct bio *bio)
1607{
1608 struct clone_info ci;
1609 blk_qc_t ret = BLK_QC_T_NONE;
1610 int error = 0;
1611
1612 if (unlikely(!map)) {
1613 bio_io_error(bio);
1614 return ret;
1615 }
1616
1617 init_clone_info(&ci, md, map, bio);
1618
1619 if (bio->bi_opf & REQ_PREFLUSH) {
1620 ci.bio = &ci.io->md->flush_bio;
1621 ci.sector_count = 0;
1622 error = __send_empty_flush(&ci);
1623 /* dec_pending submits any data associated with flush */
1624 } else {
1625 struct dm_target *ti = md->immutable_target;
1626 struct dm_target_io *tio;
1627
1628 /*
1629 * Defend against IO still getting in during teardown
1630 * - as was seen for a time with nvme-fcloop
1631 */
1632 if (unlikely(WARN_ON_ONCE(!ti || !dm_target_is_valid(ti)))) {
1633 error = -EIO;
1634 goto out;
1635 }
1636
1637 ci.bio = bio;
1638 ci.sector_count = bio_sectors(bio);
1639 if (unlikely(__process_abnormal_io(&ci, ti, &error)))
1640 goto out;
1641
1642 tio = alloc_tio(&ci, ti, 0, GFP_NOIO);
1643 ret = __clone_and_map_simple_bio(&ci, tio, NULL);
1644 }
1645out:
1646 /* drop the extra reference count */
1647 dec_pending(ci.io, errno_to_blk_status(error));
1648 return ret;
1649}
1650
1651typedef blk_qc_t (process_bio_fn)(struct mapped_device *, struct dm_table *, struct bio *);
1652
1653static blk_qc_t __dm_make_request(struct request_queue *q, struct bio *bio,
1654 process_bio_fn process_bio)
1655{
1656 struct mapped_device *md = q->queuedata;
1657 blk_qc_t ret = BLK_QC_T_NONE;
1658 int srcu_idx;
1659 struct dm_table *map;
1660
1661 map = dm_get_live_table(md, &srcu_idx);
1662
1663 /* if we're suspended, we have to queue this io for later */
1664 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1665 dm_put_live_table(md, srcu_idx);
1666
1667 if (!(bio->bi_opf & REQ_RAHEAD))
1668 queue_io(md, bio);
1669 else
1670 bio_io_error(bio);
1671 return ret;
1672 }
1673
1674 ret = process_bio(md, map, bio);
1675
1676 dm_put_live_table(md, srcu_idx);
1677 return ret;
1678}
1679
1680/*
1681 * The request function that remaps the bio to one target and
1682 * splits off any remainder.
1683 */
1684static blk_qc_t dm_make_request(struct request_queue *q, struct bio *bio)
1685{
1686 return __dm_make_request(q, bio, __split_and_process_bio);
1687}
1688
1689static blk_qc_t dm_make_request_nvme(struct request_queue *q, struct bio *bio)
1690{
1691 return __dm_make_request(q, bio, __process_bio);
1692}
1693
1694static int dm_any_congested(void *congested_data, int bdi_bits)
1695{
1696 int r = bdi_bits;
1697 struct mapped_device *md = congested_data;
1698 struct dm_table *map;
1699
1700 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
1701 if (dm_request_based(md)) {
1702 /*
1703 * With request-based DM we only need to check the
1704 * top-level queue for congestion.
1705 */
1706 r = md->queue->backing_dev_info->wb.state & bdi_bits;
1707 } else {
1708 map = dm_get_live_table_fast(md);
1709 if (map)
1710 r = dm_table_any_congested(map, bdi_bits);
1711 dm_put_live_table_fast(md);
1712 }
1713 }
1714
1715 return r;
1716}
1717
1718/*-----------------------------------------------------------------
1719 * An IDR is used to keep track of allocated minor numbers.
1720 *---------------------------------------------------------------*/
1721static void free_minor(int minor)
1722{
1723 spin_lock(&_minor_lock);
1724 idr_remove(&_minor_idr, minor);
1725 spin_unlock(&_minor_lock);
1726}
1727
1728/*
1729 * See if the device with a specific minor # is free.
1730 */
1731static int specific_minor(int minor)
1732{
1733 int r;
1734
1735 if (minor >= (1 << MINORBITS))
1736 return -EINVAL;
1737
1738 idr_preload(GFP_KERNEL);
1739 spin_lock(&_minor_lock);
1740
1741 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
1742
1743 spin_unlock(&_minor_lock);
1744 idr_preload_end();
1745 if (r < 0)
1746 return r == -ENOSPC ? -EBUSY : r;
1747 return 0;
1748}
1749
1750static int next_free_minor(int *minor)
1751{
1752 int r;
1753
1754 idr_preload(GFP_KERNEL);
1755 spin_lock(&_minor_lock);
1756
1757 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
1758
1759 spin_unlock(&_minor_lock);
1760 idr_preload_end();
1761 if (r < 0)
1762 return r;
1763 *minor = r;
1764 return 0;
1765}
1766
1767static const struct block_device_operations dm_blk_dops;
1768static const struct dax_operations dm_dax_ops;
1769
1770static void dm_wq_work(struct work_struct *work);
1771
1772static void dm_init_normal_md_queue(struct mapped_device *md)
1773{
1774 md->use_blk_mq = false;
1775
1776 /*
1777 * Initialize aspects of queue that aren't relevant for blk-mq
1778 */
1779 md->queue->backing_dev_info->congested_fn = dm_any_congested;
1780}
1781
1782static void cleanup_mapped_device(struct mapped_device *md)
1783{
1784 if (md->wq)
1785 destroy_workqueue(md->wq);
1786 if (md->kworker_task)
1787 kthread_stop(md->kworker_task);
1788 if (md->bs)
1789 bioset_free(md->bs);
1790 if (md->io_bs)
1791 bioset_free(md->io_bs);
1792
1793 if (md->dax_dev) {
1794 kill_dax(md->dax_dev);
1795 put_dax(md->dax_dev);
1796 md->dax_dev = NULL;
1797 }
1798
1799 if (md->disk) {
1800 spin_lock(&_minor_lock);
1801 md->disk->private_data = NULL;
1802 spin_unlock(&_minor_lock);
1803 del_gendisk(md->disk);
1804 put_disk(md->disk);
1805 }
1806
1807 if (md->queue)
1808 blk_cleanup_queue(md->queue);
1809
1810 cleanup_srcu_struct(&md->io_barrier);
1811
1812 if (md->bdev) {
1813 bdput(md->bdev);
1814 md->bdev = NULL;
1815 }
1816
1817 mutex_destroy(&md->suspend_lock);
1818 mutex_destroy(&md->type_lock);
1819 mutex_destroy(&md->table_devices_lock);
1820
1821 dm_mq_cleanup_mapped_device(md);
1822}
1823
1824/*
1825 * Allocate and initialise a blank device with a given minor.
1826 */
1827static struct mapped_device *alloc_dev(int minor)
1828{
1829 int r, numa_node_id = dm_get_numa_node();
1830 struct dax_device *dax_dev = NULL;
1831 struct mapped_device *md;
1832 void *old_md;
1833
1834 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
1835 if (!md) {
1836 DMWARN("unable to allocate device, out of memory.");
1837 return NULL;
1838 }
1839
1840 if (!try_module_get(THIS_MODULE))
1841 goto bad_module_get;
1842
1843 /* get a minor number for the dev */
1844 if (minor == DM_ANY_MINOR)
1845 r = next_free_minor(&minor);
1846 else
1847 r = specific_minor(minor);
1848 if (r < 0)
1849 goto bad_minor;
1850
1851 r = init_srcu_struct(&md->io_barrier);
1852 if (r < 0)
1853 goto bad_io_barrier;
1854
1855 md->numa_node_id = numa_node_id;
1856 md->use_blk_mq = dm_use_blk_mq_default();
1857 md->init_tio_pdu = false;
1858 md->type = DM_TYPE_NONE;
1859 mutex_init(&md->suspend_lock);
1860 mutex_init(&md->type_lock);
1861 mutex_init(&md->table_devices_lock);
1862 spin_lock_init(&md->deferred_lock);
1863 atomic_set(&md->holders, 1);
1864 atomic_set(&md->open_count, 0);
1865 atomic_set(&md->event_nr, 0);
1866 atomic_set(&md->uevent_seq, 0);
1867 INIT_LIST_HEAD(&md->uevent_list);
1868 INIT_LIST_HEAD(&md->table_devices);
1869 spin_lock_init(&md->uevent_lock);
1870
1871 md->queue = blk_alloc_queue_node(GFP_KERNEL, numa_node_id, NULL);
1872 if (!md->queue)
1873 goto bad;
1874 md->queue->queuedata = md;
1875 md->queue->backing_dev_info->congested_data = md;
1876
1877 md->disk = alloc_disk_node(1, md->numa_node_id);
1878 if (!md->disk)
1879 goto bad;
1880
1881 atomic_set(&md->pending[0], 0);
1882 atomic_set(&md->pending[1], 0);
1883 init_waitqueue_head(&md->wait);
1884 INIT_WORK(&md->work, dm_wq_work);
1885 init_waitqueue_head(&md->eventq);
1886 init_completion(&md->kobj_holder.completion);
1887 md->kworker_task = NULL;
1888
1889 md->disk->major = _major;
1890 md->disk->first_minor = minor;
1891 md->disk->fops = &dm_blk_dops;
1892 md->disk->queue = md->queue;
1893 md->disk->private_data = md;
1894 sprintf(md->disk->disk_name, "dm-%d", minor);
1895
1896 if (IS_ENABLED(CONFIG_DAX_DRIVER)) {
1897 dax_dev = alloc_dax(md, md->disk->disk_name, &dm_dax_ops);
1898 if (!dax_dev)
1899 goto bad;
1900 }
1901 md->dax_dev = dax_dev;
1902
1903 add_disk_no_queue_reg(md->disk);
1904 format_dev_t(md->name, MKDEV(_major, minor));
1905
1906 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
1907 if (!md->wq)
1908 goto bad;
1909
1910 md->bdev = bdget_disk(md->disk, 0);
1911 if (!md->bdev)
1912 goto bad;
1913
1914 bio_init(&md->flush_bio, NULL, 0);
1915 bio_set_dev(&md->flush_bio, md->bdev);
1916 md->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC;
1917
1918 dm_stats_init(&md->stats);
1919
1920 /* Populate the mapping, nobody knows we exist yet */
1921 spin_lock(&_minor_lock);
1922 old_md = idr_replace(&_minor_idr, md, minor);
1923 spin_unlock(&_minor_lock);
1924
1925 BUG_ON(old_md != MINOR_ALLOCED);
1926
1927 return md;
1928
1929bad:
1930 cleanup_mapped_device(md);
1931bad_io_barrier:
1932 free_minor(minor);
1933bad_minor:
1934 module_put(THIS_MODULE);
1935bad_module_get:
1936 kvfree(md);
1937 return NULL;
1938}
1939
1940static void unlock_fs(struct mapped_device *md);
1941
1942static void free_dev(struct mapped_device *md)
1943{
1944 int minor = MINOR(disk_devt(md->disk));
1945
1946 unlock_fs(md);
1947
1948 cleanup_mapped_device(md);
1949
1950 free_table_devices(&md->table_devices);
1951 dm_stats_cleanup(&md->stats);
1952 free_minor(minor);
1953
1954 module_put(THIS_MODULE);
1955 kvfree(md);
1956}
1957
1958static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
1959{
1960 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
1961
1962 if (dm_table_bio_based(t)) {
1963 /*
1964 * The md may already have mempools that need changing.
1965 * If so, reload bioset because front_pad may have changed
1966 * because a different table was loaded.
1967 */
1968 if (md->bs) {
1969 bioset_free(md->bs);
1970 md->bs = NULL;
1971 }
1972 if (md->io_bs) {
1973 bioset_free(md->io_bs);
1974 md->io_bs = NULL;
1975 }
1976
1977 } else if (md->bs) {
1978 /*
1979 * There's no need to reload with request-based dm
1980 * because the size of front_pad doesn't change.
1981 * Note for future: If you are to reload bioset,
1982 * prep-ed requests in the queue may refer
1983 * to bio from the old bioset, so you must walk
1984 * through the queue to unprep.
1985 */
1986 goto out;
1987 }
1988
1989 BUG_ON(!p || md->bs || md->io_bs);
1990
1991 md->bs = p->bs;
1992 p->bs = NULL;
1993 md->io_bs = p->io_bs;
1994 p->io_bs = NULL;
1995out:
1996 /* mempool bind completed, no longer need any mempools in the table */
1997 dm_table_free_md_mempools(t);
1998}
1999
2000/*
2001 * Bind a table to the device.
2002 */
2003static void event_callback(void *context)
2004{
2005 unsigned long flags;
2006 LIST_HEAD(uevents);
2007 struct mapped_device *md = (struct mapped_device *) context;
2008
2009 spin_lock_irqsave(&md->uevent_lock, flags);
2010 list_splice_init(&md->uevent_list, &uevents);
2011 spin_unlock_irqrestore(&md->uevent_lock, flags);
2012
2013 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2014
2015 atomic_inc(&md->event_nr);
2016 wake_up(&md->eventq);
2017 dm_issue_global_event();
2018}
2019
2020/*
2021 * Protected by md->suspend_lock obtained by dm_swap_table().
2022 */
2023static void __set_size(struct mapped_device *md, sector_t size)
2024{
2025 lockdep_assert_held(&md->suspend_lock);
2026
2027 set_capacity(md->disk, size);
2028
2029 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2030}
2031
2032/*
2033 * Returns old map, which caller must destroy.
2034 */
2035static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2036 struct queue_limits *limits)
2037{
2038 struct dm_table *old_map;
2039 struct request_queue *q = md->queue;
2040 bool request_based = dm_table_request_based(t);
2041 sector_t size;
2042
2043 lockdep_assert_held(&md->suspend_lock);
2044
2045 size = dm_table_get_size(t);
2046
2047 /*
2048 * Wipe any geometry if the size of the table changed.
2049 */
2050 if (size != dm_get_size(md))
2051 memset(&md->geometry, 0, sizeof(md->geometry));
2052
2053 __set_size(md, size);
2054
2055 dm_table_event_callback(t, event_callback, md);
2056
2057 /*
2058 * The queue hasn't been stopped yet, if the old table type wasn't
2059 * for request-based during suspension. So stop it to prevent
2060 * I/O mapping before resume.
2061 * This must be done before setting the queue restrictions,
2062 * because request-based dm may be run just after the setting.
2063 */
2064 if (request_based)
2065 dm_stop_queue(q);
2066
2067 if (request_based || md->type == DM_TYPE_NVME_BIO_BASED) {
2068 /*
2069 * Leverage the fact that request-based DM targets and
2070 * NVMe bio based targets are immutable singletons
2071 * - used to optimize both dm_request_fn and dm_mq_queue_rq;
2072 * and __process_bio.
2073 */
2074 md->immutable_target = dm_table_get_immutable_target(t);
2075 }
2076
2077 __bind_mempools(md, t);
2078
2079 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2080 rcu_assign_pointer(md->map, (void *)t);
2081 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2082
2083 dm_table_set_restrictions(t, q, limits);
2084 if (old_map)
2085 dm_sync_table(md);
2086
2087 return old_map;
2088}
2089
2090/*
2091 * Returns unbound table for the caller to free.
2092 */
2093static struct dm_table *__unbind(struct mapped_device *md)
2094{
2095 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2096
2097 if (!map)
2098 return NULL;
2099
2100 dm_table_event_callback(map, NULL, NULL);
2101 RCU_INIT_POINTER(md->map, NULL);
2102 dm_sync_table(md);
2103
2104 return map;
2105}
2106
2107/*
2108 * Constructor for a new device.
2109 */
2110int dm_create(int minor, struct mapped_device **result)
2111{
2112 int r;
2113 struct mapped_device *md;
2114
2115 md = alloc_dev(minor);
2116 if (!md)
2117 return -ENXIO;
2118
2119 r = dm_sysfs_init(md);
2120 if (r) {
2121 free_dev(md);
2122 return r;
2123 }
2124
2125 *result = md;
2126 return 0;
2127}
2128
2129/*
2130 * Functions to manage md->type.
2131 * All are required to hold md->type_lock.
2132 */
2133void dm_lock_md_type(struct mapped_device *md)
2134{
2135 mutex_lock(&md->type_lock);
2136}
2137
2138void dm_unlock_md_type(struct mapped_device *md)
2139{
2140 mutex_unlock(&md->type_lock);
2141}
2142
2143void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2144{
2145 BUG_ON(!mutex_is_locked(&md->type_lock));
2146 md->type = type;
2147}
2148
2149enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2150{
2151 return md->type;
2152}
2153
2154struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2155{
2156 return md->immutable_target_type;
2157}
2158
2159/*
2160 * The queue_limits are only valid as long as you have a reference
2161 * count on 'md'.
2162 */
2163struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2164{
2165 BUG_ON(!atomic_read(&md->holders));
2166 return &md->queue->limits;
2167}
2168EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2169
2170/*
2171 * Setup the DM device's queue based on md's type
2172 */
2173int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2174{
2175 int r;
2176 struct queue_limits limits;
2177 enum dm_queue_mode type = dm_get_md_type(md);
2178
2179 switch (type) {
2180 case DM_TYPE_REQUEST_BASED:
2181 dm_init_normal_md_queue(md);
2182 r = dm_old_init_request_queue(md, t);
2183 if (r) {
2184 DMERR("Cannot initialize queue for request-based mapped device");
2185 return r;
2186 }
2187 break;
2188 case DM_TYPE_MQ_REQUEST_BASED:
2189 r = dm_mq_init_request_queue(md, t);
2190 if (r) {
2191 DMERR("Cannot initialize queue for request-based dm-mq mapped device");
2192 return r;
2193 }
2194 break;
2195 case DM_TYPE_BIO_BASED:
2196 case DM_TYPE_DAX_BIO_BASED:
2197 dm_init_normal_md_queue(md);
2198 blk_queue_make_request(md->queue, dm_make_request);
2199 break;
2200 case DM_TYPE_NVME_BIO_BASED:
2201 dm_init_normal_md_queue(md);
2202 blk_queue_make_request(md->queue, dm_make_request_nvme);
2203 break;
2204 case DM_TYPE_NONE:
2205 WARN_ON_ONCE(true);
2206 break;
2207 }
2208
2209 r = dm_calculate_queue_limits(t, &limits);
2210 if (r) {
2211 DMERR("Cannot calculate initial queue limits");
2212 return r;
2213 }
2214 dm_table_set_restrictions(t, md->queue, &limits);
2215 blk_register_queue(md->disk);
2216
2217 return 0;
2218}
2219
2220struct mapped_device *dm_get_md(dev_t dev)
2221{
2222 struct mapped_device *md;
2223 unsigned minor = MINOR(dev);
2224
2225 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2226 return NULL;
2227
2228 spin_lock(&_minor_lock);
2229
2230 md = idr_find(&_minor_idr, minor);
2231 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2232 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2233 md = NULL;
2234 goto out;
2235 }
2236 dm_get(md);
2237out:
2238 spin_unlock(&_minor_lock);
2239
2240 return md;
2241}
2242EXPORT_SYMBOL_GPL(dm_get_md);
2243
2244void *dm_get_mdptr(struct mapped_device *md)
2245{
2246 return md->interface_ptr;
2247}
2248
2249void dm_set_mdptr(struct mapped_device *md, void *ptr)
2250{
2251 md->interface_ptr = ptr;
2252}
2253
2254void dm_get(struct mapped_device *md)
2255{
2256 atomic_inc(&md->holders);
2257 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2258}
2259
2260int dm_hold(struct mapped_device *md)
2261{
2262 spin_lock(&_minor_lock);
2263 if (test_bit(DMF_FREEING, &md->flags)) {
2264 spin_unlock(&_minor_lock);
2265 return -EBUSY;
2266 }
2267 dm_get(md);
2268 spin_unlock(&_minor_lock);
2269 return 0;
2270}
2271EXPORT_SYMBOL_GPL(dm_hold);
2272
2273const char *dm_device_name(struct mapped_device *md)
2274{
2275 return md->name;
2276}
2277EXPORT_SYMBOL_GPL(dm_device_name);
2278
2279static void __dm_destroy(struct mapped_device *md, bool wait)
2280{
2281 struct dm_table *map;
2282 int srcu_idx;
2283
2284 might_sleep();
2285
2286 spin_lock(&_minor_lock);
2287 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2288 set_bit(DMF_FREEING, &md->flags);
2289 spin_unlock(&_minor_lock);
2290
2291 blk_set_queue_dying(md->queue);
2292
2293 if (dm_request_based(md) && md->kworker_task)
2294 kthread_flush_worker(&md->kworker);
2295
2296 /*
2297 * Take suspend_lock so that presuspend and postsuspend methods
2298 * do not race with internal suspend.
2299 */
2300 mutex_lock(&md->suspend_lock);
2301 map = dm_get_live_table(md, &srcu_idx);
2302 if (!dm_suspended_md(md)) {
2303 dm_table_presuspend_targets(map);
2304 dm_table_postsuspend_targets(map);
2305 }
2306 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2307 dm_put_live_table(md, srcu_idx);
2308 mutex_unlock(&md->suspend_lock);
2309
2310 /*
2311 * Rare, but there may be I/O requests still going to complete,
2312 * for example. Wait for all references to disappear.
2313 * No one should increment the reference count of the mapped_device,
2314 * after the mapped_device state becomes DMF_FREEING.
2315 */
2316 if (wait)
2317 while (atomic_read(&md->holders))
2318 msleep(1);
2319 else if (atomic_read(&md->holders))
2320 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2321 dm_device_name(md), atomic_read(&md->holders));
2322
2323 dm_sysfs_exit(md);
2324 dm_table_destroy(__unbind(md));
2325 free_dev(md);
2326}
2327
2328void dm_destroy(struct mapped_device *md)
2329{
2330 __dm_destroy(md, true);
2331}
2332
2333void dm_destroy_immediate(struct mapped_device *md)
2334{
2335 __dm_destroy(md, false);
2336}
2337
2338void dm_put(struct mapped_device *md)
2339{
2340 atomic_dec(&md->holders);
2341}
2342EXPORT_SYMBOL_GPL(dm_put);
2343
2344static int dm_wait_for_completion(struct mapped_device *md, long task_state)
2345{
2346 int r = 0;
2347 DEFINE_WAIT(wait);
2348
2349 while (1) {
2350 prepare_to_wait(&md->wait, &wait, task_state);
2351
2352 if (!md_in_flight(md))
2353 break;
2354
2355 if (signal_pending_state(task_state, current)) {
2356 r = -EINTR;
2357 break;
2358 }
2359
2360 io_schedule();
2361 }
2362 finish_wait(&md->wait, &wait);
2363
2364 return r;
2365}
2366
2367/*
2368 * Process the deferred bios
2369 */
2370static void dm_wq_work(struct work_struct *work)
2371{
2372 struct mapped_device *md = container_of(work, struct mapped_device,
2373 work);
2374 struct bio *c;
2375 int srcu_idx;
2376 struct dm_table *map;
2377
2378 map = dm_get_live_table(md, &srcu_idx);
2379
2380 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2381 spin_lock_irq(&md->deferred_lock);
2382 c = bio_list_pop(&md->deferred);
2383 spin_unlock_irq(&md->deferred_lock);
2384
2385 if (!c)
2386 break;
2387
2388 if (dm_request_based(md))
2389 generic_make_request(c);
2390 else
2391 __split_and_process_bio(md, map, c);
2392 }
2393
2394 dm_put_live_table(md, srcu_idx);
2395}
2396
2397static void dm_queue_flush(struct mapped_device *md)
2398{
2399 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2400 smp_mb__after_atomic();
2401 queue_work(md->wq, &md->work);
2402}
2403
2404/*
2405 * Swap in a new table, returning the old one for the caller to destroy.
2406 */
2407struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2408{
2409 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2410 struct queue_limits limits;
2411 int r;
2412
2413 mutex_lock(&md->suspend_lock);
2414
2415 /* device must be suspended */
2416 if (!dm_suspended_md(md))
2417 goto out;
2418
2419 /*
2420 * If the new table has no data devices, retain the existing limits.
2421 * This helps multipath with queue_if_no_path if all paths disappear,
2422 * then new I/O is queued based on these limits, and then some paths
2423 * reappear.
2424 */
2425 if (dm_table_has_no_data_devices(table)) {
2426 live_map = dm_get_live_table_fast(md);
2427 if (live_map)
2428 limits = md->queue->limits;
2429 dm_put_live_table_fast(md);
2430 }
2431
2432 if (!live_map) {
2433 r = dm_calculate_queue_limits(table, &limits);
2434 if (r) {
2435 map = ERR_PTR(r);
2436 goto out;
2437 }
2438 }
2439
2440 map = __bind(md, table, &limits);
2441 dm_issue_global_event();
2442
2443out:
2444 mutex_unlock(&md->suspend_lock);
2445 return map;
2446}
2447
2448/*
2449 * Functions to lock and unlock any filesystem running on the
2450 * device.
2451 */
2452static int lock_fs(struct mapped_device *md)
2453{
2454 int r;
2455
2456 WARN_ON(md->frozen_sb);
2457
2458 md->frozen_sb = freeze_bdev(md->bdev);
2459 if (IS_ERR(md->frozen_sb)) {
2460 r = PTR_ERR(md->frozen_sb);
2461 md->frozen_sb = NULL;
2462 return r;
2463 }
2464
2465 set_bit(DMF_FROZEN, &md->flags);
2466
2467 return 0;
2468}
2469
2470static void unlock_fs(struct mapped_device *md)
2471{
2472 if (!test_bit(DMF_FROZEN, &md->flags))
2473 return;
2474
2475 thaw_bdev(md->bdev, md->frozen_sb);
2476 md->frozen_sb = NULL;
2477 clear_bit(DMF_FROZEN, &md->flags);
2478}
2479
2480/*
2481 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2482 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2483 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2484 *
2485 * If __dm_suspend returns 0, the device is completely quiescent
2486 * now. There is no request-processing activity. All new requests
2487 * are being added to md->deferred list.
2488 */
2489static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2490 unsigned suspend_flags, long task_state,
2491 int dmf_suspended_flag)
2492{
2493 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2494 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2495 int r;
2496
2497 lockdep_assert_held(&md->suspend_lock);
2498
2499 /*
2500 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2501 * This flag is cleared before dm_suspend returns.
2502 */
2503 if (noflush)
2504 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2505 else
2506 pr_debug("%s: suspending with flush\n", dm_device_name(md));
2507
2508 /*
2509 * This gets reverted if there's an error later and the targets
2510 * provide the .presuspend_undo hook.
2511 */
2512 dm_table_presuspend_targets(map);
2513
2514 /*
2515 * Flush I/O to the device.
2516 * Any I/O submitted after lock_fs() may not be flushed.
2517 * noflush takes precedence over do_lockfs.
2518 * (lock_fs() flushes I/Os and waits for them to complete.)
2519 */
2520 if (!noflush && do_lockfs) {
2521 r = lock_fs(md);
2522 if (r) {
2523 dm_table_presuspend_undo_targets(map);
2524 return r;
2525 }
2526 }
2527
2528 /*
2529 * Here we must make sure that no processes are submitting requests
2530 * to target drivers i.e. no one may be executing
2531 * __split_and_process_bio. This is called from dm_request and
2532 * dm_wq_work.
2533 *
2534 * To get all processes out of __split_and_process_bio in dm_request,
2535 * we take the write lock. To prevent any process from reentering
2536 * __split_and_process_bio from dm_request and quiesce the thread
2537 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
2538 * flush_workqueue(md->wq).
2539 */
2540 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2541 if (map)
2542 synchronize_srcu(&md->io_barrier);
2543
2544 /*
2545 * Stop md->queue before flushing md->wq in case request-based
2546 * dm defers requests to md->wq from md->queue.
2547 */
2548 if (dm_request_based(md)) {
2549 dm_stop_queue(md->queue);
2550 if (md->kworker_task)
2551 kthread_flush_worker(&md->kworker);
2552 }
2553
2554 flush_workqueue(md->wq);
2555
2556 /*
2557 * At this point no more requests are entering target request routines.
2558 * We call dm_wait_for_completion to wait for all existing requests
2559 * to finish.
2560 */
2561 r = dm_wait_for_completion(md, task_state);
2562 if (!r)
2563 set_bit(dmf_suspended_flag, &md->flags);
2564
2565 if (noflush)
2566 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2567 if (map)
2568 synchronize_srcu(&md->io_barrier);
2569
2570 /* were we interrupted ? */
2571 if (r < 0) {
2572 dm_queue_flush(md);
2573
2574 if (dm_request_based(md))
2575 dm_start_queue(md->queue);
2576
2577 unlock_fs(md);
2578 dm_table_presuspend_undo_targets(map);
2579 /* pushback list is already flushed, so skip flush */
2580 }
2581
2582 return r;
2583}
2584
2585/*
2586 * We need to be able to change a mapping table under a mounted
2587 * filesystem. For example we might want to move some data in
2588 * the background. Before the table can be swapped with
2589 * dm_bind_table, dm_suspend must be called to flush any in
2590 * flight bios and ensure that any further io gets deferred.
2591 */
2592/*
2593 * Suspend mechanism in request-based dm.
2594 *
2595 * 1. Flush all I/Os by lock_fs() if needed.
2596 * 2. Stop dispatching any I/O by stopping the request_queue.
2597 * 3. Wait for all in-flight I/Os to be completed or requeued.
2598 *
2599 * To abort suspend, start the request_queue.
2600 */
2601int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
2602{
2603 struct dm_table *map = NULL;
2604 int r = 0;
2605
2606retry:
2607 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2608
2609 if (dm_suspended_md(md)) {
2610 r = -EINVAL;
2611 goto out_unlock;
2612 }
2613
2614 if (dm_suspended_internally_md(md)) {
2615 /* already internally suspended, wait for internal resume */
2616 mutex_unlock(&md->suspend_lock);
2617 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2618 if (r)
2619 return r;
2620 goto retry;
2621 }
2622
2623 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2624
2625 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2626 if (r)
2627 goto out_unlock;
2628
2629 dm_table_postsuspend_targets(map);
2630
2631out_unlock:
2632 mutex_unlock(&md->suspend_lock);
2633 return r;
2634}
2635
2636static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2637{
2638 if (map) {
2639 int r = dm_table_resume_targets(map);
2640 if (r)
2641 return r;
2642 }
2643
2644 dm_queue_flush(md);
2645
2646 /*
2647 * Flushing deferred I/Os must be done after targets are resumed
2648 * so that mapping of targets can work correctly.
2649 * Request-based dm is queueing the deferred I/Os in its request_queue.
2650 */
2651 if (dm_request_based(md))
2652 dm_start_queue(md->queue);
2653
2654 unlock_fs(md);
2655
2656 return 0;
2657}
2658
2659int dm_resume(struct mapped_device *md)
2660{
2661 int r;
2662 struct dm_table *map = NULL;
2663
2664retry:
2665 r = -EINVAL;
2666 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2667
2668 if (!dm_suspended_md(md))
2669 goto out;
2670
2671 if (dm_suspended_internally_md(md)) {
2672 /* already internally suspended, wait for internal resume */
2673 mutex_unlock(&md->suspend_lock);
2674 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2675 if (r)
2676 return r;
2677 goto retry;
2678 }
2679
2680 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2681 if (!map || !dm_table_get_size(map))
2682 goto out;
2683
2684 r = __dm_resume(md, map);
2685 if (r)
2686 goto out;
2687
2688 clear_bit(DMF_SUSPENDED, &md->flags);
2689out:
2690 mutex_unlock(&md->suspend_lock);
2691
2692 return r;
2693}
2694
2695/*
2696 * Internal suspend/resume works like userspace-driven suspend. It waits
2697 * until all bios finish and prevents issuing new bios to the target drivers.
2698 * It may be used only from the kernel.
2699 */
2700
2701static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
2702{
2703 struct dm_table *map = NULL;
2704
2705 lockdep_assert_held(&md->suspend_lock);
2706
2707 if (md->internal_suspend_count++)
2708 return; /* nested internal suspend */
2709
2710 if (dm_suspended_md(md)) {
2711 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2712 return; /* nest suspend */
2713 }
2714
2715 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2716
2717 /*
2718 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2719 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2720 * would require changing .presuspend to return an error -- avoid this
2721 * until there is a need for more elaborate variants of internal suspend.
2722 */
2723 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2724 DMF_SUSPENDED_INTERNALLY);
2725
2726 dm_table_postsuspend_targets(map);
2727}
2728
2729static void __dm_internal_resume(struct mapped_device *md)
2730{
2731 BUG_ON(!md->internal_suspend_count);
2732
2733 if (--md->internal_suspend_count)
2734 return; /* resume from nested internal suspend */
2735
2736 if (dm_suspended_md(md))
2737 goto done; /* resume from nested suspend */
2738
2739 /*
2740 * NOTE: existing callers don't need to call dm_table_resume_targets
2741 * (which may fail -- so best to avoid it for now by passing NULL map)
2742 */
2743 (void) __dm_resume(md, NULL);
2744
2745done:
2746 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2747 smp_mb__after_atomic();
2748 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
2749}
2750
2751void dm_internal_suspend_noflush(struct mapped_device *md)
2752{
2753 mutex_lock(&md->suspend_lock);
2754 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2755 mutex_unlock(&md->suspend_lock);
2756}
2757EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2758
2759void dm_internal_resume(struct mapped_device *md)
2760{
2761 mutex_lock(&md->suspend_lock);
2762 __dm_internal_resume(md);
2763 mutex_unlock(&md->suspend_lock);
2764}
2765EXPORT_SYMBOL_GPL(dm_internal_resume);
2766
2767/*
2768 * Fast variants of internal suspend/resume hold md->suspend_lock,
2769 * which prevents interaction with userspace-driven suspend.
2770 */
2771
2772void dm_internal_suspend_fast(struct mapped_device *md)
2773{
2774 mutex_lock(&md->suspend_lock);
2775 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2776 return;
2777
2778 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2779 synchronize_srcu(&md->io_barrier);
2780 flush_workqueue(md->wq);
2781 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
2782}
2783EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
2784
2785void dm_internal_resume_fast(struct mapped_device *md)
2786{
2787 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2788 goto done;
2789
2790 dm_queue_flush(md);
2791
2792done:
2793 mutex_unlock(&md->suspend_lock);
2794}
2795EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
2796
2797/*-----------------------------------------------------------------
2798 * Event notification.
2799 *---------------------------------------------------------------*/
2800int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
2801 unsigned cookie)
2802{
2803 char udev_cookie[DM_COOKIE_LENGTH];
2804 char *envp[] = { udev_cookie, NULL };
2805
2806 if (!cookie)
2807 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
2808 else {
2809 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
2810 DM_COOKIE_ENV_VAR_NAME, cookie);
2811 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
2812 action, envp);
2813 }
2814}
2815
2816uint32_t dm_next_uevent_seq(struct mapped_device *md)
2817{
2818 return atomic_add_return(1, &md->uevent_seq);
2819}
2820
2821uint32_t dm_get_event_nr(struct mapped_device *md)
2822{
2823 return atomic_read(&md->event_nr);
2824}
2825
2826int dm_wait_event(struct mapped_device *md, int event_nr)
2827{
2828 return wait_event_interruptible(md->eventq,
2829 (event_nr != atomic_read(&md->event_nr)));
2830}
2831
2832void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
2833{
2834 unsigned long flags;
2835
2836 spin_lock_irqsave(&md->uevent_lock, flags);
2837 list_add(elist, &md->uevent_list);
2838 spin_unlock_irqrestore(&md->uevent_lock, flags);
2839}
2840
2841/*
2842 * The gendisk is only valid as long as you have a reference
2843 * count on 'md'.
2844 */
2845struct gendisk *dm_disk(struct mapped_device *md)
2846{
2847 return md->disk;
2848}
2849EXPORT_SYMBOL_GPL(dm_disk);
2850
2851struct kobject *dm_kobject(struct mapped_device *md)
2852{
2853 return &md->kobj_holder.kobj;
2854}
2855
2856struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
2857{
2858 struct mapped_device *md;
2859
2860 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
2861
2862 spin_lock(&_minor_lock);
2863 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2864 md = NULL;
2865 goto out;
2866 }
2867 dm_get(md);
2868out:
2869 spin_unlock(&_minor_lock);
2870
2871 return md;
2872}
2873
2874int dm_suspended_md(struct mapped_device *md)
2875{
2876 return test_bit(DMF_SUSPENDED, &md->flags);
2877}
2878
2879int dm_suspended_internally_md(struct mapped_device *md)
2880{
2881 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2882}
2883
2884int dm_test_deferred_remove_flag(struct mapped_device *md)
2885{
2886 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
2887}
2888
2889int dm_suspended(struct dm_target *ti)
2890{
2891 return dm_suspended_md(dm_table_get_md(ti->table));
2892}
2893EXPORT_SYMBOL_GPL(dm_suspended);
2894
2895int dm_noflush_suspending(struct dm_target *ti)
2896{
2897 return __noflush_suspending(dm_table_get_md(ti->table));
2898}
2899EXPORT_SYMBOL_GPL(dm_noflush_suspending);
2900
2901struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, enum dm_queue_mode type,
2902 unsigned integrity, unsigned per_io_data_size,
2903 unsigned min_pool_size)
2904{
2905 struct dm_md_mempools *pools = kzalloc_node(sizeof(*pools), GFP_KERNEL, md->numa_node_id);
2906 unsigned int pool_size = 0;
2907 unsigned int front_pad, io_front_pad;
2908
2909 if (!pools)
2910 return NULL;
2911
2912 switch (type) {
2913 case DM_TYPE_BIO_BASED:
2914 case DM_TYPE_DAX_BIO_BASED:
2915 case DM_TYPE_NVME_BIO_BASED:
2916 pool_size = max(dm_get_reserved_bio_based_ios(), min_pool_size);
2917 front_pad = roundup(per_io_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
2918 io_front_pad = roundup(front_pad, __alignof__(struct dm_io)) + offsetof(struct dm_io, tio);
2919 pools->io_bs = bioset_create(pool_size, io_front_pad, 0);
2920 if (!pools->io_bs)
2921 goto out;
2922 if (integrity && bioset_integrity_create(pools->io_bs, pool_size))
2923 goto out;
2924 break;
2925 case DM_TYPE_REQUEST_BASED:
2926 case DM_TYPE_MQ_REQUEST_BASED:
2927 pool_size = max(dm_get_reserved_rq_based_ios(), min_pool_size);
2928 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
2929 /* per_io_data_size is used for blk-mq pdu at queue allocation */
2930 break;
2931 default:
2932 BUG();
2933 }
2934
2935 pools->bs = bioset_create(pool_size, front_pad, 0);
2936 if (!pools->bs)
2937 goto out;
2938
2939 if (integrity && bioset_integrity_create(pools->bs, pool_size))
2940 goto out;
2941
2942 return pools;
2943
2944out:
2945 dm_free_md_mempools(pools);
2946
2947 return NULL;
2948}
2949
2950void dm_free_md_mempools(struct dm_md_mempools *pools)
2951{
2952 if (!pools)
2953 return;
2954
2955 if (pools->bs)
2956 bioset_free(pools->bs);
2957 if (pools->io_bs)
2958 bioset_free(pools->io_bs);
2959
2960 kfree(pools);
2961}
2962
2963struct dm_pr {
2964 u64 old_key;
2965 u64 new_key;
2966 u32 flags;
2967 bool fail_early;
2968};
2969
2970static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
2971 void *data)
2972{
2973 struct mapped_device *md = bdev->bd_disk->private_data;
2974 struct dm_table *table;
2975 struct dm_target *ti;
2976 int ret = -ENOTTY, srcu_idx;
2977
2978 table = dm_get_live_table(md, &srcu_idx);
2979 if (!table || !dm_table_get_size(table))
2980 goto out;
2981
2982 /* We only support devices that have a single target */
2983 if (dm_table_get_num_targets(table) != 1)
2984 goto out;
2985 ti = dm_table_get_target(table, 0);
2986
2987 ret = -EINVAL;
2988 if (!ti->type->iterate_devices)
2989 goto out;
2990
2991 ret = ti->type->iterate_devices(ti, fn, data);
2992out:
2993 dm_put_live_table(md, srcu_idx);
2994 return ret;
2995}
2996
2997/*
2998 * For register / unregister we need to manually call out to every path.
2999 */
3000static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3001 sector_t start, sector_t len, void *data)
3002{
3003 struct dm_pr *pr = data;
3004 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3005
3006 if (!ops || !ops->pr_register)
3007 return -EOPNOTSUPP;
3008 return ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3009}
3010
3011static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3012 u32 flags)
3013{
3014 struct dm_pr pr = {
3015 .old_key = old_key,
3016 .new_key = new_key,
3017 .flags = flags,
3018 .fail_early = true,
3019 };
3020 int ret;
3021
3022 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3023 if (ret && new_key) {
3024 /* unregister all paths if we failed to register any path */
3025 pr.old_key = new_key;
3026 pr.new_key = 0;
3027 pr.flags = 0;
3028 pr.fail_early = false;
3029 dm_call_pr(bdev, __dm_pr_register, &pr);
3030 }
3031
3032 return ret;
3033}
3034
3035static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3036 u32 flags)
3037{
3038 struct mapped_device *md = bdev->bd_disk->private_data;
3039 const struct pr_ops *ops;
3040 int r, srcu_idx;
3041
3042 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3043 if (r < 0)
3044 goto out;
3045
3046 ops = bdev->bd_disk->fops->pr_ops;
3047 if (ops && ops->pr_reserve)
3048 r = ops->pr_reserve(bdev, key, type, flags);
3049 else
3050 r = -EOPNOTSUPP;
3051out:
3052 dm_unprepare_ioctl(md, srcu_idx);
3053 return r;
3054}
3055
3056static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3057{
3058 struct mapped_device *md = bdev->bd_disk->private_data;
3059 const struct pr_ops *ops;
3060 int r, srcu_idx;
3061
3062 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3063 if (r < 0)
3064 goto out;
3065
3066 ops = bdev->bd_disk->fops->pr_ops;
3067 if (ops && ops->pr_release)
3068 r = ops->pr_release(bdev, key, type);
3069 else
3070 r = -EOPNOTSUPP;
3071out:
3072 dm_unprepare_ioctl(md, srcu_idx);
3073 return r;
3074}
3075
3076static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3077 enum pr_type type, bool abort)
3078{
3079 struct mapped_device *md = bdev->bd_disk->private_data;
3080 const struct pr_ops *ops;
3081 int r, srcu_idx;
3082
3083 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3084 if (r < 0)
3085 goto out;
3086
3087 ops = bdev->bd_disk->fops->pr_ops;
3088 if (ops && ops->pr_preempt)
3089 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
3090 else
3091 r = -EOPNOTSUPP;
3092out:
3093 dm_unprepare_ioctl(md, srcu_idx);
3094 return r;
3095}
3096
3097static int dm_pr_clear(struct block_device *bdev, u64 key)
3098{
3099 struct mapped_device *md = bdev->bd_disk->private_data;
3100 const struct pr_ops *ops;
3101 int r, srcu_idx;
3102
3103 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3104 if (r < 0)
3105 goto out;
3106
3107 ops = bdev->bd_disk->fops->pr_ops;
3108 if (ops && ops->pr_clear)
3109 r = ops->pr_clear(bdev, key);
3110 else
3111 r = -EOPNOTSUPP;
3112out:
3113 dm_unprepare_ioctl(md, srcu_idx);
3114 return r;
3115}
3116
3117static const struct pr_ops dm_pr_ops = {
3118 .pr_register = dm_pr_register,
3119 .pr_reserve = dm_pr_reserve,
3120 .pr_release = dm_pr_release,
3121 .pr_preempt = dm_pr_preempt,
3122 .pr_clear = dm_pr_clear,
3123};
3124
3125static const struct block_device_operations dm_blk_dops = {
3126 .open = dm_blk_open,
3127 .release = dm_blk_close,
3128 .ioctl = dm_blk_ioctl,
3129 .getgeo = dm_blk_getgeo,
3130 .pr_ops = &dm_pr_ops,
3131 .owner = THIS_MODULE
3132};
3133
3134static const struct dax_operations dm_dax_ops = {
3135 .direct_access = dm_dax_direct_access,
3136 .copy_from_iter = dm_dax_copy_from_iter,
3137};
3138
3139/*
3140 * module hooks
3141 */
3142module_init(dm_init);
3143module_exit(dm_exit);
3144
3145module_param(major, uint, 0);
3146MODULE_PARM_DESC(major, "The major number of the device mapper");
3147
3148module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3149MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3150
3151module_param(dm_numa_node, int, S_IRUGO | S_IWUSR);
3152MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3153
3154MODULE_DESCRIPTION(DM_NAME " driver");
3155MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3156MODULE_LICENSE("GPL");