Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
4 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
5 *
6 * This file is released under the GPL.
7 */
8
9#include "dm-core.h"
10#include "dm-rq.h"
11#include "dm-uevent.h"
12#include "dm-ima.h"
13
14#include <linux/init.h>
15#include <linux/module.h>
16#include <linux/mutex.h>
17#include <linux/sched/mm.h>
18#include <linux/sched/signal.h>
19#include <linux/blkpg.h>
20#include <linux/bio.h>
21#include <linux/mempool.h>
22#include <linux/dax.h>
23#include <linux/slab.h>
24#include <linux/idr.h>
25#include <linux/uio.h>
26#include <linux/hdreg.h>
27#include <linux/delay.h>
28#include <linux/wait.h>
29#include <linux/pr.h>
30#include <linux/refcount.h>
31#include <linux/part_stat.h>
32#include <linux/blk-crypto.h>
33#include <linux/blk-crypto-profile.h>
34
35#define DM_MSG_PREFIX "core"
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
44/*
45 * For REQ_POLLED fs bio, this flag is set if we link mapped underlying
46 * dm_io into one list, and reuse bio->bi_private as the list head. Before
47 * ending this fs bio, we will recover its ->bi_private.
48 */
49#define REQ_DM_POLL_LIST REQ_DRV
50
51static const char *_name = DM_NAME;
52
53static unsigned int major;
54static unsigned int _major;
55
56static DEFINE_IDR(_minor_idr);
57
58static DEFINE_SPINLOCK(_minor_lock);
59
60static void do_deferred_remove(struct work_struct *w);
61
62static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
63
64static struct workqueue_struct *deferred_remove_workqueue;
65
66atomic_t dm_global_event_nr = ATOMIC_INIT(0);
67DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
68
69void dm_issue_global_event(void)
70{
71 atomic_inc(&dm_global_event_nr);
72 wake_up(&dm_global_eventq);
73}
74
75DEFINE_STATIC_KEY_FALSE(stats_enabled);
76DEFINE_STATIC_KEY_FALSE(swap_bios_enabled);
77DEFINE_STATIC_KEY_FALSE(zoned_enabled);
78
79/*
80 * One of these is allocated (on-stack) per original bio.
81 */
82struct clone_info {
83 struct dm_table *map;
84 struct bio *bio;
85 struct dm_io *io;
86 sector_t sector;
87 unsigned int sector_count;
88 bool is_abnormal_io:1;
89 bool submit_as_polled:1;
90};
91
92static inline struct dm_target_io *clone_to_tio(struct bio *clone)
93{
94 return container_of(clone, struct dm_target_io, clone);
95}
96
97void *dm_per_bio_data(struct bio *bio, size_t data_size)
98{
99 if (!dm_tio_flagged(clone_to_tio(bio), DM_TIO_INSIDE_DM_IO))
100 return (char *)bio - DM_TARGET_IO_BIO_OFFSET - data_size;
101 return (char *)bio - DM_IO_BIO_OFFSET - data_size;
102}
103EXPORT_SYMBOL_GPL(dm_per_bio_data);
104
105struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
106{
107 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
108
109 if (io->magic == DM_IO_MAGIC)
110 return (struct bio *)((char *)io + DM_IO_BIO_OFFSET);
111 BUG_ON(io->magic != DM_TIO_MAGIC);
112 return (struct bio *)((char *)io + DM_TARGET_IO_BIO_OFFSET);
113}
114EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
115
116unsigned int dm_bio_get_target_bio_nr(const struct bio *bio)
117{
118 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
119}
120EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
121
122#define MINOR_ALLOCED ((void *)-1)
123
124#define DM_NUMA_NODE NUMA_NO_NODE
125static int dm_numa_node = DM_NUMA_NODE;
126
127#define DEFAULT_SWAP_BIOS (8 * 1048576 / PAGE_SIZE)
128static int swap_bios = DEFAULT_SWAP_BIOS;
129static int get_swap_bios(void)
130{
131 int latch = READ_ONCE(swap_bios);
132
133 if (unlikely(latch <= 0))
134 latch = DEFAULT_SWAP_BIOS;
135 return latch;
136}
137
138struct table_device {
139 struct list_head list;
140 refcount_t count;
141 struct dm_dev dm_dev;
142};
143
144/*
145 * Bio-based DM's mempools' reserved IOs set by the user.
146 */
147#define RESERVED_BIO_BASED_IOS 16
148static unsigned int reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
149
150static int __dm_get_module_param_int(int *module_param, int min, int max)
151{
152 int param = READ_ONCE(*module_param);
153 int modified_param = 0;
154 bool modified = true;
155
156 if (param < min)
157 modified_param = min;
158 else if (param > max)
159 modified_param = max;
160 else
161 modified = false;
162
163 if (modified) {
164 (void)cmpxchg(module_param, param, modified_param);
165 param = modified_param;
166 }
167
168 return param;
169}
170
171unsigned int __dm_get_module_param(unsigned int *module_param, unsigned int def, unsigned int max)
172{
173 unsigned int param = READ_ONCE(*module_param);
174 unsigned int modified_param = 0;
175
176 if (!param)
177 modified_param = def;
178 else if (param > max)
179 modified_param = max;
180
181 if (modified_param) {
182 (void)cmpxchg(module_param, param, modified_param);
183 param = modified_param;
184 }
185
186 return param;
187}
188
189unsigned int dm_get_reserved_bio_based_ios(void)
190{
191 return __dm_get_module_param(&reserved_bio_based_ios,
192 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
193}
194EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
195
196static unsigned int dm_get_numa_node(void)
197{
198 return __dm_get_module_param_int(&dm_numa_node,
199 DM_NUMA_NODE, num_online_nodes() - 1);
200}
201
202static int __init local_init(void)
203{
204 int r;
205
206 r = dm_uevent_init();
207 if (r)
208 return r;
209
210 deferred_remove_workqueue = alloc_ordered_workqueue("kdmremove", 0);
211 if (!deferred_remove_workqueue) {
212 r = -ENOMEM;
213 goto out_uevent_exit;
214 }
215
216 _major = major;
217 r = register_blkdev(_major, _name);
218 if (r < 0)
219 goto out_free_workqueue;
220
221 if (!_major)
222 _major = r;
223
224 return 0;
225
226out_free_workqueue:
227 destroy_workqueue(deferred_remove_workqueue);
228out_uevent_exit:
229 dm_uevent_exit();
230
231 return r;
232}
233
234static void local_exit(void)
235{
236 destroy_workqueue(deferred_remove_workqueue);
237
238 unregister_blkdev(_major, _name);
239 dm_uevent_exit();
240
241 _major = 0;
242
243 DMINFO("cleaned up");
244}
245
246static int (*_inits[])(void) __initdata = {
247 local_init,
248 dm_target_init,
249 dm_linear_init,
250 dm_stripe_init,
251 dm_io_init,
252 dm_kcopyd_init,
253 dm_interface_init,
254 dm_statistics_init,
255};
256
257static void (*_exits[])(void) = {
258 local_exit,
259 dm_target_exit,
260 dm_linear_exit,
261 dm_stripe_exit,
262 dm_io_exit,
263 dm_kcopyd_exit,
264 dm_interface_exit,
265 dm_statistics_exit,
266};
267
268static int __init dm_init(void)
269{
270 const int count = ARRAY_SIZE(_inits);
271 int r, i;
272
273#if (IS_ENABLED(CONFIG_IMA) && !IS_ENABLED(CONFIG_IMA_DISABLE_HTABLE))
274 DMWARN("CONFIG_IMA_DISABLE_HTABLE is disabled."
275 " Duplicate IMA measurements will not be recorded in the IMA log.");
276#endif
277
278 for (i = 0; i < count; i++) {
279 r = _inits[i]();
280 if (r)
281 goto bad;
282 }
283
284 return 0;
285bad:
286 while (i--)
287 _exits[i]();
288
289 return r;
290}
291
292static void __exit dm_exit(void)
293{
294 int i = ARRAY_SIZE(_exits);
295
296 while (i--)
297 _exits[i]();
298
299 /*
300 * Should be empty by this point.
301 */
302 idr_destroy(&_minor_idr);
303}
304
305/*
306 * Block device functions
307 */
308int dm_deleting_md(struct mapped_device *md)
309{
310 return test_bit(DMF_DELETING, &md->flags);
311}
312
313static int dm_blk_open(struct gendisk *disk, blk_mode_t mode)
314{
315 struct mapped_device *md;
316
317 spin_lock(&_minor_lock);
318
319 md = disk->private_data;
320 if (!md)
321 goto out;
322
323 if (test_bit(DMF_FREEING, &md->flags) ||
324 dm_deleting_md(md)) {
325 md = NULL;
326 goto out;
327 }
328
329 dm_get(md);
330 atomic_inc(&md->open_count);
331out:
332 spin_unlock(&_minor_lock);
333
334 return md ? 0 : -ENXIO;
335}
336
337static void dm_blk_close(struct gendisk *disk)
338{
339 struct mapped_device *md;
340
341 spin_lock(&_minor_lock);
342
343 md = disk->private_data;
344 if (WARN_ON(!md))
345 goto out;
346
347 if (atomic_dec_and_test(&md->open_count) &&
348 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
349 queue_work(deferred_remove_workqueue, &deferred_remove_work);
350
351 dm_put(md);
352out:
353 spin_unlock(&_minor_lock);
354}
355
356int dm_open_count(struct mapped_device *md)
357{
358 return atomic_read(&md->open_count);
359}
360
361/*
362 * Guarantees nothing is using the device before it's deleted.
363 */
364int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
365{
366 int r = 0;
367
368 spin_lock(&_minor_lock);
369
370 if (dm_open_count(md)) {
371 r = -EBUSY;
372 if (mark_deferred)
373 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
374 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
375 r = -EEXIST;
376 else
377 set_bit(DMF_DELETING, &md->flags);
378
379 spin_unlock(&_minor_lock);
380
381 return r;
382}
383
384int dm_cancel_deferred_remove(struct mapped_device *md)
385{
386 int r = 0;
387
388 spin_lock(&_minor_lock);
389
390 if (test_bit(DMF_DELETING, &md->flags))
391 r = -EBUSY;
392 else
393 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
394
395 spin_unlock(&_minor_lock);
396
397 return r;
398}
399
400static void do_deferred_remove(struct work_struct *w)
401{
402 dm_deferred_remove();
403}
404
405static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
406{
407 struct mapped_device *md = bdev->bd_disk->private_data;
408
409 return dm_get_geometry(md, geo);
410}
411
412static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
413 struct block_device **bdev)
414{
415 struct dm_target *ti;
416 struct dm_table *map;
417 int r;
418
419retry:
420 r = -ENOTTY;
421 map = dm_get_live_table(md, srcu_idx);
422 if (!map || !dm_table_get_size(map))
423 return r;
424
425 /* We only support devices that have a single target */
426 if (map->num_targets != 1)
427 return r;
428
429 ti = dm_table_get_target(map, 0);
430 if (!ti->type->prepare_ioctl)
431 return r;
432
433 if (dm_suspended_md(md))
434 return -EAGAIN;
435
436 r = ti->type->prepare_ioctl(ti, bdev);
437 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
438 dm_put_live_table(md, *srcu_idx);
439 fsleep(10000);
440 goto retry;
441 }
442
443 return r;
444}
445
446static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
447{
448 dm_put_live_table(md, srcu_idx);
449}
450
451static int dm_blk_ioctl(struct block_device *bdev, blk_mode_t mode,
452 unsigned int cmd, unsigned long arg)
453{
454 struct mapped_device *md = bdev->bd_disk->private_data;
455 int r, srcu_idx;
456
457 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
458 if (r < 0)
459 goto out;
460
461 if (r > 0) {
462 /*
463 * Target determined this ioctl is being issued against a
464 * subset of the parent bdev; require extra privileges.
465 */
466 if (!capable(CAP_SYS_RAWIO)) {
467 DMDEBUG_LIMIT(
468 "%s: sending ioctl %x to DM device without required privilege.",
469 current->comm, cmd);
470 r = -ENOIOCTLCMD;
471 goto out;
472 }
473 }
474
475 if (!bdev->bd_disk->fops->ioctl)
476 r = -ENOTTY;
477 else
478 r = bdev->bd_disk->fops->ioctl(bdev, mode, cmd, arg);
479out:
480 dm_unprepare_ioctl(md, srcu_idx);
481 return r;
482}
483
484u64 dm_start_time_ns_from_clone(struct bio *bio)
485{
486 return jiffies_to_nsecs(clone_to_tio(bio)->io->start_time);
487}
488EXPORT_SYMBOL_GPL(dm_start_time_ns_from_clone);
489
490static inline bool bio_is_flush_with_data(struct bio *bio)
491{
492 return ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size);
493}
494
495static inline unsigned int dm_io_sectors(struct dm_io *io, struct bio *bio)
496{
497 /*
498 * If REQ_PREFLUSH set, don't account payload, it will be
499 * submitted (and accounted) after this flush completes.
500 */
501 if (bio_is_flush_with_data(bio))
502 return 0;
503 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
504 return io->sectors;
505 return bio_sectors(bio);
506}
507
508static void dm_io_acct(struct dm_io *io, bool end)
509{
510 struct bio *bio = io->orig_bio;
511
512 if (dm_io_flagged(io, DM_IO_BLK_STAT)) {
513 if (!end)
514 bdev_start_io_acct(bio->bi_bdev, bio_op(bio),
515 io->start_time);
516 else
517 bdev_end_io_acct(bio->bi_bdev, bio_op(bio),
518 dm_io_sectors(io, bio),
519 io->start_time);
520 }
521
522 if (static_branch_unlikely(&stats_enabled) &&
523 unlikely(dm_stats_used(&io->md->stats))) {
524 sector_t sector;
525
526 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
527 sector = bio_end_sector(bio) - io->sector_offset;
528 else
529 sector = bio->bi_iter.bi_sector;
530
531 dm_stats_account_io(&io->md->stats, bio_data_dir(bio),
532 sector, dm_io_sectors(io, bio),
533 end, io->start_time, &io->stats_aux);
534 }
535}
536
537static void __dm_start_io_acct(struct dm_io *io)
538{
539 dm_io_acct(io, false);
540}
541
542static void dm_start_io_acct(struct dm_io *io, struct bio *clone)
543{
544 /*
545 * Ensure IO accounting is only ever started once.
546 */
547 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
548 return;
549
550 /* Expect no possibility for race unless DM_TIO_IS_DUPLICATE_BIO. */
551 if (!clone || likely(dm_tio_is_normal(clone_to_tio(clone)))) {
552 dm_io_set_flag(io, DM_IO_ACCOUNTED);
553 } else {
554 unsigned long flags;
555 /* Can afford locking given DM_TIO_IS_DUPLICATE_BIO */
556 spin_lock_irqsave(&io->lock, flags);
557 if (dm_io_flagged(io, DM_IO_ACCOUNTED)) {
558 spin_unlock_irqrestore(&io->lock, flags);
559 return;
560 }
561 dm_io_set_flag(io, DM_IO_ACCOUNTED);
562 spin_unlock_irqrestore(&io->lock, flags);
563 }
564
565 __dm_start_io_acct(io);
566}
567
568static void dm_end_io_acct(struct dm_io *io)
569{
570 dm_io_acct(io, true);
571}
572
573static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio, gfp_t gfp_mask)
574{
575 struct dm_io *io;
576 struct dm_target_io *tio;
577 struct bio *clone;
578
579 clone = bio_alloc_clone(NULL, bio, gfp_mask, &md->mempools->io_bs);
580 if (unlikely(!clone))
581 return NULL;
582 tio = clone_to_tio(clone);
583 tio->flags = 0;
584 dm_tio_set_flag(tio, DM_TIO_INSIDE_DM_IO);
585 tio->io = NULL;
586
587 io = container_of(tio, struct dm_io, tio);
588 io->magic = DM_IO_MAGIC;
589 io->status = BLK_STS_OK;
590
591 /* one ref is for submission, the other is for completion */
592 atomic_set(&io->io_count, 2);
593 this_cpu_inc(*md->pending_io);
594 io->orig_bio = bio;
595 io->md = md;
596 spin_lock_init(&io->lock);
597 io->start_time = jiffies;
598 io->flags = 0;
599 if (blk_queue_io_stat(md->queue))
600 dm_io_set_flag(io, DM_IO_BLK_STAT);
601
602 if (static_branch_unlikely(&stats_enabled) &&
603 unlikely(dm_stats_used(&md->stats)))
604 dm_stats_record_start(&md->stats, &io->stats_aux);
605
606 return io;
607}
608
609static void free_io(struct dm_io *io)
610{
611 bio_put(&io->tio.clone);
612}
613
614static struct bio *alloc_tio(struct clone_info *ci, struct dm_target *ti,
615 unsigned int target_bio_nr, unsigned int *len, gfp_t gfp_mask)
616{
617 struct mapped_device *md = ci->io->md;
618 struct dm_target_io *tio;
619 struct bio *clone;
620
621 if (!ci->io->tio.io) {
622 /* the dm_target_io embedded in ci->io is available */
623 tio = &ci->io->tio;
624 /* alloc_io() already initialized embedded clone */
625 clone = &tio->clone;
626 } else {
627 clone = bio_alloc_clone(NULL, ci->bio, gfp_mask,
628 &md->mempools->bs);
629 if (!clone)
630 return NULL;
631
632 /* REQ_DM_POLL_LIST shouldn't be inherited */
633 clone->bi_opf &= ~REQ_DM_POLL_LIST;
634
635 tio = clone_to_tio(clone);
636 tio->flags = 0; /* also clears DM_TIO_INSIDE_DM_IO */
637 }
638
639 tio->magic = DM_TIO_MAGIC;
640 tio->io = ci->io;
641 tio->ti = ti;
642 tio->target_bio_nr = target_bio_nr;
643 tio->len_ptr = len;
644 tio->old_sector = 0;
645
646 /* Set default bdev, but target must bio_set_dev() before issuing IO */
647 clone->bi_bdev = md->disk->part0;
648 if (unlikely(ti->needs_bio_set_dev))
649 bio_set_dev(clone, md->disk->part0);
650
651 if (len) {
652 clone->bi_iter.bi_size = to_bytes(*len);
653 if (bio_integrity(clone))
654 bio_integrity_trim(clone);
655 }
656
657 return clone;
658}
659
660static void free_tio(struct bio *clone)
661{
662 if (dm_tio_flagged(clone_to_tio(clone), DM_TIO_INSIDE_DM_IO))
663 return;
664 bio_put(clone);
665}
666
667/*
668 * Add the bio to the list of deferred io.
669 */
670static void queue_io(struct mapped_device *md, struct bio *bio)
671{
672 unsigned long flags;
673
674 spin_lock_irqsave(&md->deferred_lock, flags);
675 bio_list_add(&md->deferred, bio);
676 spin_unlock_irqrestore(&md->deferred_lock, flags);
677 queue_work(md->wq, &md->work);
678}
679
680/*
681 * Everyone (including functions in this file), should use this
682 * function to access the md->map field, and make sure they call
683 * dm_put_live_table() when finished.
684 */
685struct dm_table *dm_get_live_table(struct mapped_device *md,
686 int *srcu_idx) __acquires(md->io_barrier)
687{
688 *srcu_idx = srcu_read_lock(&md->io_barrier);
689
690 return srcu_dereference(md->map, &md->io_barrier);
691}
692
693void dm_put_live_table(struct mapped_device *md,
694 int srcu_idx) __releases(md->io_barrier)
695{
696 srcu_read_unlock(&md->io_barrier, srcu_idx);
697}
698
699void dm_sync_table(struct mapped_device *md)
700{
701 synchronize_srcu(&md->io_barrier);
702 synchronize_rcu_expedited();
703}
704
705/*
706 * A fast alternative to dm_get_live_table/dm_put_live_table.
707 * The caller must not block between these two functions.
708 */
709static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
710{
711 rcu_read_lock();
712 return rcu_dereference(md->map);
713}
714
715static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
716{
717 rcu_read_unlock();
718}
719
720static char *_dm_claim_ptr = "I belong to device-mapper";
721
722/*
723 * Open a table device so we can use it as a map destination.
724 */
725static struct table_device *open_table_device(struct mapped_device *md,
726 dev_t dev, blk_mode_t mode)
727{
728 struct table_device *td;
729 struct bdev_handle *bdev_handle;
730 u64 part_off;
731 int r;
732
733 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
734 if (!td)
735 return ERR_PTR(-ENOMEM);
736 refcount_set(&td->count, 1);
737
738 bdev_handle = bdev_open_by_dev(dev, mode, _dm_claim_ptr, NULL);
739 if (IS_ERR(bdev_handle)) {
740 r = PTR_ERR(bdev_handle);
741 goto out_free_td;
742 }
743
744 /*
745 * We can be called before the dm disk is added. In that case we can't
746 * register the holder relation here. It will be done once add_disk was
747 * called.
748 */
749 if (md->disk->slave_dir) {
750 r = bd_link_disk_holder(bdev_handle->bdev, md->disk);
751 if (r)
752 goto out_blkdev_put;
753 }
754
755 td->dm_dev.mode = mode;
756 td->dm_dev.bdev = bdev_handle->bdev;
757 td->dm_dev.bdev_handle = bdev_handle;
758 td->dm_dev.dax_dev = fs_dax_get_by_bdev(bdev_handle->bdev, &part_off,
759 NULL, NULL);
760 format_dev_t(td->dm_dev.name, dev);
761 list_add(&td->list, &md->table_devices);
762 return td;
763
764out_blkdev_put:
765 bdev_release(bdev_handle);
766out_free_td:
767 kfree(td);
768 return ERR_PTR(r);
769}
770
771/*
772 * Close a table device that we've been using.
773 */
774static void close_table_device(struct table_device *td, struct mapped_device *md)
775{
776 if (md->disk->slave_dir)
777 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
778 bdev_release(td->dm_dev.bdev_handle);
779 put_dax(td->dm_dev.dax_dev);
780 list_del(&td->list);
781 kfree(td);
782}
783
784static struct table_device *find_table_device(struct list_head *l, dev_t dev,
785 blk_mode_t mode)
786{
787 struct table_device *td;
788
789 list_for_each_entry(td, l, list)
790 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
791 return td;
792
793 return NULL;
794}
795
796int dm_get_table_device(struct mapped_device *md, dev_t dev, blk_mode_t mode,
797 struct dm_dev **result)
798{
799 struct table_device *td;
800
801 mutex_lock(&md->table_devices_lock);
802 td = find_table_device(&md->table_devices, dev, mode);
803 if (!td) {
804 td = open_table_device(md, dev, mode);
805 if (IS_ERR(td)) {
806 mutex_unlock(&md->table_devices_lock);
807 return PTR_ERR(td);
808 }
809 } else {
810 refcount_inc(&td->count);
811 }
812 mutex_unlock(&md->table_devices_lock);
813
814 *result = &td->dm_dev;
815 return 0;
816}
817
818void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
819{
820 struct table_device *td = container_of(d, struct table_device, dm_dev);
821
822 mutex_lock(&md->table_devices_lock);
823 if (refcount_dec_and_test(&td->count))
824 close_table_device(td, md);
825 mutex_unlock(&md->table_devices_lock);
826}
827
828/*
829 * Get the geometry associated with a dm device
830 */
831int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
832{
833 *geo = md->geometry;
834
835 return 0;
836}
837
838/*
839 * Set the geometry of a device.
840 */
841int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
842{
843 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
844
845 if (geo->start > sz) {
846 DMERR("Start sector is beyond the geometry limits.");
847 return -EINVAL;
848 }
849
850 md->geometry = *geo;
851
852 return 0;
853}
854
855static int __noflush_suspending(struct mapped_device *md)
856{
857 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
858}
859
860static void dm_requeue_add_io(struct dm_io *io, bool first_stage)
861{
862 struct mapped_device *md = io->md;
863
864 if (first_stage) {
865 struct dm_io *next = md->requeue_list;
866
867 md->requeue_list = io;
868 io->next = next;
869 } else {
870 bio_list_add_head(&md->deferred, io->orig_bio);
871 }
872}
873
874static void dm_kick_requeue(struct mapped_device *md, bool first_stage)
875{
876 if (first_stage)
877 queue_work(md->wq, &md->requeue_work);
878 else
879 queue_work(md->wq, &md->work);
880}
881
882/*
883 * Return true if the dm_io's original bio is requeued.
884 * io->status is updated with error if requeue disallowed.
885 */
886static bool dm_handle_requeue(struct dm_io *io, bool first_stage)
887{
888 struct bio *bio = io->orig_bio;
889 bool handle_requeue = (io->status == BLK_STS_DM_REQUEUE);
890 bool handle_polled_eagain = ((io->status == BLK_STS_AGAIN) &&
891 (bio->bi_opf & REQ_POLLED));
892 struct mapped_device *md = io->md;
893 bool requeued = false;
894
895 if (handle_requeue || handle_polled_eagain) {
896 unsigned long flags;
897
898 if (bio->bi_opf & REQ_POLLED) {
899 /*
900 * Upper layer won't help us poll split bio
901 * (io->orig_bio may only reflect a subset of the
902 * pre-split original) so clear REQ_POLLED.
903 */
904 bio_clear_polled(bio);
905 }
906
907 /*
908 * Target requested pushing back the I/O or
909 * polled IO hit BLK_STS_AGAIN.
910 */
911 spin_lock_irqsave(&md->deferred_lock, flags);
912 if ((__noflush_suspending(md) &&
913 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) ||
914 handle_polled_eagain || first_stage) {
915 dm_requeue_add_io(io, first_stage);
916 requeued = true;
917 } else {
918 /*
919 * noflush suspend was interrupted or this is
920 * a write to a zoned target.
921 */
922 io->status = BLK_STS_IOERR;
923 }
924 spin_unlock_irqrestore(&md->deferred_lock, flags);
925 }
926
927 if (requeued)
928 dm_kick_requeue(md, first_stage);
929
930 return requeued;
931}
932
933static void __dm_io_complete(struct dm_io *io, bool first_stage)
934{
935 struct bio *bio = io->orig_bio;
936 struct mapped_device *md = io->md;
937 blk_status_t io_error;
938 bool requeued;
939
940 requeued = dm_handle_requeue(io, first_stage);
941 if (requeued && first_stage)
942 return;
943
944 io_error = io->status;
945 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
946 dm_end_io_acct(io);
947 else if (!io_error) {
948 /*
949 * Must handle target that DM_MAPIO_SUBMITTED only to
950 * then bio_endio() rather than dm_submit_bio_remap()
951 */
952 __dm_start_io_acct(io);
953 dm_end_io_acct(io);
954 }
955 free_io(io);
956 smp_wmb();
957 this_cpu_dec(*md->pending_io);
958
959 /* nudge anyone waiting on suspend queue */
960 if (unlikely(wq_has_sleeper(&md->wait)))
961 wake_up(&md->wait);
962
963 /* Return early if the original bio was requeued */
964 if (requeued)
965 return;
966
967 if (bio_is_flush_with_data(bio)) {
968 /*
969 * Preflush done for flush with data, reissue
970 * without REQ_PREFLUSH.
971 */
972 bio->bi_opf &= ~REQ_PREFLUSH;
973 queue_io(md, bio);
974 } else {
975 /* done with normal IO or empty flush */
976 if (io_error)
977 bio->bi_status = io_error;
978 bio_endio(bio);
979 }
980}
981
982static void dm_wq_requeue_work(struct work_struct *work)
983{
984 struct mapped_device *md = container_of(work, struct mapped_device,
985 requeue_work);
986 unsigned long flags;
987 struct dm_io *io;
988
989 /* reuse deferred lock to simplify dm_handle_requeue */
990 spin_lock_irqsave(&md->deferred_lock, flags);
991 io = md->requeue_list;
992 md->requeue_list = NULL;
993 spin_unlock_irqrestore(&md->deferred_lock, flags);
994
995 while (io) {
996 struct dm_io *next = io->next;
997
998 dm_io_rewind(io, &md->disk->bio_split);
999
1000 io->next = NULL;
1001 __dm_io_complete(io, false);
1002 io = next;
1003 cond_resched();
1004 }
1005}
1006
1007/*
1008 * Two staged requeue:
1009 *
1010 * 1) io->orig_bio points to the real original bio, and the part mapped to
1011 * this io must be requeued, instead of other parts of the original bio.
1012 *
1013 * 2) io->orig_bio points to new cloned bio which matches the requeued dm_io.
1014 */
1015static void dm_io_complete(struct dm_io *io)
1016{
1017 bool first_requeue;
1018
1019 /*
1020 * Only dm_io that has been split needs two stage requeue, otherwise
1021 * we may run into long bio clone chain during suspend and OOM could
1022 * be triggered.
1023 *
1024 * Also flush data dm_io won't be marked as DM_IO_WAS_SPLIT, so they
1025 * also aren't handled via the first stage requeue.
1026 */
1027 if (dm_io_flagged(io, DM_IO_WAS_SPLIT))
1028 first_requeue = true;
1029 else
1030 first_requeue = false;
1031
1032 __dm_io_complete(io, first_requeue);
1033}
1034
1035/*
1036 * Decrements the number of outstanding ios that a bio has been
1037 * cloned into, completing the original io if necc.
1038 */
1039static inline void __dm_io_dec_pending(struct dm_io *io)
1040{
1041 if (atomic_dec_and_test(&io->io_count))
1042 dm_io_complete(io);
1043}
1044
1045static void dm_io_set_error(struct dm_io *io, blk_status_t error)
1046{
1047 unsigned long flags;
1048
1049 /* Push-back supersedes any I/O errors */
1050 spin_lock_irqsave(&io->lock, flags);
1051 if (!(io->status == BLK_STS_DM_REQUEUE &&
1052 __noflush_suspending(io->md))) {
1053 io->status = error;
1054 }
1055 spin_unlock_irqrestore(&io->lock, flags);
1056}
1057
1058static void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
1059{
1060 if (unlikely(error))
1061 dm_io_set_error(io, error);
1062
1063 __dm_io_dec_pending(io);
1064}
1065
1066/*
1067 * The queue_limits are only valid as long as you have a reference
1068 * count on 'md'. But _not_ imposing verification to avoid atomic_read(),
1069 */
1070static inline struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
1071{
1072 return &md->queue->limits;
1073}
1074
1075void disable_discard(struct mapped_device *md)
1076{
1077 struct queue_limits *limits = dm_get_queue_limits(md);
1078
1079 /* device doesn't really support DISCARD, disable it */
1080 limits->max_discard_sectors = 0;
1081}
1082
1083void disable_write_zeroes(struct mapped_device *md)
1084{
1085 struct queue_limits *limits = dm_get_queue_limits(md);
1086
1087 /* device doesn't really support WRITE ZEROES, disable it */
1088 limits->max_write_zeroes_sectors = 0;
1089}
1090
1091static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
1092{
1093 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
1094}
1095
1096static void clone_endio(struct bio *bio)
1097{
1098 blk_status_t error = bio->bi_status;
1099 struct dm_target_io *tio = clone_to_tio(bio);
1100 struct dm_target *ti = tio->ti;
1101 dm_endio_fn endio = ti->type->end_io;
1102 struct dm_io *io = tio->io;
1103 struct mapped_device *md = io->md;
1104
1105 if (unlikely(error == BLK_STS_TARGET)) {
1106 if (bio_op(bio) == REQ_OP_DISCARD &&
1107 !bdev_max_discard_sectors(bio->bi_bdev))
1108 disable_discard(md);
1109 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
1110 !bdev_write_zeroes_sectors(bio->bi_bdev))
1111 disable_write_zeroes(md);
1112 }
1113
1114 if (static_branch_unlikely(&zoned_enabled) &&
1115 unlikely(bdev_is_zoned(bio->bi_bdev)))
1116 dm_zone_endio(io, bio);
1117
1118 if (endio) {
1119 int r = endio(ti, bio, &error);
1120
1121 switch (r) {
1122 case DM_ENDIO_REQUEUE:
1123 if (static_branch_unlikely(&zoned_enabled)) {
1124 /*
1125 * Requeuing writes to a sequential zone of a zoned
1126 * target will break the sequential write pattern:
1127 * fail such IO.
1128 */
1129 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
1130 error = BLK_STS_IOERR;
1131 else
1132 error = BLK_STS_DM_REQUEUE;
1133 } else
1134 error = BLK_STS_DM_REQUEUE;
1135 fallthrough;
1136 case DM_ENDIO_DONE:
1137 break;
1138 case DM_ENDIO_INCOMPLETE:
1139 /* The target will handle the io */
1140 return;
1141 default:
1142 DMCRIT("unimplemented target endio return value: %d", r);
1143 BUG();
1144 }
1145 }
1146
1147 if (static_branch_unlikely(&swap_bios_enabled) &&
1148 unlikely(swap_bios_limit(ti, bio)))
1149 up(&md->swap_bios_semaphore);
1150
1151 free_tio(bio);
1152 dm_io_dec_pending(io, error);
1153}
1154
1155/*
1156 * Return maximum size of I/O possible at the supplied sector up to the current
1157 * target boundary.
1158 */
1159static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
1160 sector_t target_offset)
1161{
1162 return ti->len - target_offset;
1163}
1164
1165static sector_t __max_io_len(struct dm_target *ti, sector_t sector,
1166 unsigned int max_granularity,
1167 unsigned int max_sectors)
1168{
1169 sector_t target_offset = dm_target_offset(ti, sector);
1170 sector_t len = max_io_len_target_boundary(ti, target_offset);
1171
1172 /*
1173 * Does the target need to split IO even further?
1174 * - varied (per target) IO splitting is a tenet of DM; this
1175 * explains why stacked chunk_sectors based splitting via
1176 * bio_split_to_limits() isn't possible here.
1177 */
1178 if (!max_granularity)
1179 return len;
1180 return min_t(sector_t, len,
1181 min(max_sectors ? : queue_max_sectors(ti->table->md->queue),
1182 blk_chunk_sectors_left(target_offset, max_granularity)));
1183}
1184
1185static inline sector_t max_io_len(struct dm_target *ti, sector_t sector)
1186{
1187 return __max_io_len(ti, sector, ti->max_io_len, 0);
1188}
1189
1190int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1191{
1192 if (len > UINT_MAX) {
1193 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1194 (unsigned long long)len, UINT_MAX);
1195 ti->error = "Maximum size of target IO is too large";
1196 return -EINVAL;
1197 }
1198
1199 ti->max_io_len = (uint32_t) len;
1200
1201 return 0;
1202}
1203EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1204
1205static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1206 sector_t sector, int *srcu_idx)
1207 __acquires(md->io_barrier)
1208{
1209 struct dm_table *map;
1210 struct dm_target *ti;
1211
1212 map = dm_get_live_table(md, srcu_idx);
1213 if (!map)
1214 return NULL;
1215
1216 ti = dm_table_find_target(map, sector);
1217 if (!ti)
1218 return NULL;
1219
1220 return ti;
1221}
1222
1223static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1224 long nr_pages, enum dax_access_mode mode, void **kaddr,
1225 pfn_t *pfn)
1226{
1227 struct mapped_device *md = dax_get_private(dax_dev);
1228 sector_t sector = pgoff * PAGE_SECTORS;
1229 struct dm_target *ti;
1230 long len, ret = -EIO;
1231 int srcu_idx;
1232
1233 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1234
1235 if (!ti)
1236 goto out;
1237 if (!ti->type->direct_access)
1238 goto out;
1239 len = max_io_len(ti, sector) / PAGE_SECTORS;
1240 if (len < 1)
1241 goto out;
1242 nr_pages = min(len, nr_pages);
1243 ret = ti->type->direct_access(ti, pgoff, nr_pages, mode, kaddr, pfn);
1244
1245 out:
1246 dm_put_live_table(md, srcu_idx);
1247
1248 return ret;
1249}
1250
1251static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1252 size_t nr_pages)
1253{
1254 struct mapped_device *md = dax_get_private(dax_dev);
1255 sector_t sector = pgoff * PAGE_SECTORS;
1256 struct dm_target *ti;
1257 int ret = -EIO;
1258 int srcu_idx;
1259
1260 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1261
1262 if (!ti)
1263 goto out;
1264 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1265 /*
1266 * ->zero_page_range() is mandatory dax operation. If we are
1267 * here, something is wrong.
1268 */
1269 goto out;
1270 }
1271 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1272 out:
1273 dm_put_live_table(md, srcu_idx);
1274
1275 return ret;
1276}
1277
1278static size_t dm_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
1279 void *addr, size_t bytes, struct iov_iter *i)
1280{
1281 struct mapped_device *md = dax_get_private(dax_dev);
1282 sector_t sector = pgoff * PAGE_SECTORS;
1283 struct dm_target *ti;
1284 int srcu_idx;
1285 long ret = 0;
1286
1287 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1288 if (!ti || !ti->type->dax_recovery_write)
1289 goto out;
1290
1291 ret = ti->type->dax_recovery_write(ti, pgoff, addr, bytes, i);
1292out:
1293 dm_put_live_table(md, srcu_idx);
1294 return ret;
1295}
1296
1297/*
1298 * A target may call dm_accept_partial_bio only from the map routine. It is
1299 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1300 * operations, REQ_OP_ZONE_APPEND (zone append writes) and any bio serviced by
1301 * __send_duplicate_bios().
1302 *
1303 * dm_accept_partial_bio informs the dm that the target only wants to process
1304 * additional n_sectors sectors of the bio and the rest of the data should be
1305 * sent in a next bio.
1306 *
1307 * A diagram that explains the arithmetics:
1308 * +--------------------+---------------+-------+
1309 * | 1 | 2 | 3 |
1310 * +--------------------+---------------+-------+
1311 *
1312 * <-------------- *tio->len_ptr --------------->
1313 * <----- bio_sectors ----->
1314 * <-- n_sectors -->
1315 *
1316 * Region 1 was already iterated over with bio_advance or similar function.
1317 * (it may be empty if the target doesn't use bio_advance)
1318 * Region 2 is the remaining bio size that the target wants to process.
1319 * (it may be empty if region 1 is non-empty, although there is no reason
1320 * to make it empty)
1321 * The target requires that region 3 is to be sent in the next bio.
1322 *
1323 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1324 * the partially processed part (the sum of regions 1+2) must be the same for all
1325 * copies of the bio.
1326 */
1327void dm_accept_partial_bio(struct bio *bio, unsigned int n_sectors)
1328{
1329 struct dm_target_io *tio = clone_to_tio(bio);
1330 struct dm_io *io = tio->io;
1331 unsigned int bio_sectors = bio_sectors(bio);
1332
1333 BUG_ON(dm_tio_flagged(tio, DM_TIO_IS_DUPLICATE_BIO));
1334 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1335 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1336 BUG_ON(bio_sectors > *tio->len_ptr);
1337 BUG_ON(n_sectors > bio_sectors);
1338
1339 *tio->len_ptr -= bio_sectors - n_sectors;
1340 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1341
1342 /*
1343 * __split_and_process_bio() may have already saved mapped part
1344 * for accounting but it is being reduced so update accordingly.
1345 */
1346 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1347 io->sectors = n_sectors;
1348 io->sector_offset = bio_sectors(io->orig_bio);
1349}
1350EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1351
1352/*
1353 * @clone: clone bio that DM core passed to target's .map function
1354 * @tgt_clone: clone of @clone bio that target needs submitted
1355 *
1356 * Targets should use this interface to submit bios they take
1357 * ownership of when returning DM_MAPIO_SUBMITTED.
1358 *
1359 * Target should also enable ti->accounts_remapped_io
1360 */
1361void dm_submit_bio_remap(struct bio *clone, struct bio *tgt_clone)
1362{
1363 struct dm_target_io *tio = clone_to_tio(clone);
1364 struct dm_io *io = tio->io;
1365
1366 /* establish bio that will get submitted */
1367 if (!tgt_clone)
1368 tgt_clone = clone;
1369
1370 /*
1371 * Account io->origin_bio to DM dev on behalf of target
1372 * that took ownership of IO with DM_MAPIO_SUBMITTED.
1373 */
1374 dm_start_io_acct(io, clone);
1375
1376 trace_block_bio_remap(tgt_clone, disk_devt(io->md->disk),
1377 tio->old_sector);
1378 submit_bio_noacct(tgt_clone);
1379}
1380EXPORT_SYMBOL_GPL(dm_submit_bio_remap);
1381
1382static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1383{
1384 mutex_lock(&md->swap_bios_lock);
1385 while (latch < md->swap_bios) {
1386 cond_resched();
1387 down(&md->swap_bios_semaphore);
1388 md->swap_bios--;
1389 }
1390 while (latch > md->swap_bios) {
1391 cond_resched();
1392 up(&md->swap_bios_semaphore);
1393 md->swap_bios++;
1394 }
1395 mutex_unlock(&md->swap_bios_lock);
1396}
1397
1398static void __map_bio(struct bio *clone)
1399{
1400 struct dm_target_io *tio = clone_to_tio(clone);
1401 struct dm_target *ti = tio->ti;
1402 struct dm_io *io = tio->io;
1403 struct mapped_device *md = io->md;
1404 int r;
1405
1406 clone->bi_end_io = clone_endio;
1407
1408 /*
1409 * Map the clone.
1410 */
1411 tio->old_sector = clone->bi_iter.bi_sector;
1412
1413 if (static_branch_unlikely(&swap_bios_enabled) &&
1414 unlikely(swap_bios_limit(ti, clone))) {
1415 int latch = get_swap_bios();
1416
1417 if (unlikely(latch != md->swap_bios))
1418 __set_swap_bios_limit(md, latch);
1419 down(&md->swap_bios_semaphore);
1420 }
1421
1422 if (static_branch_unlikely(&zoned_enabled)) {
1423 /*
1424 * Check if the IO needs a special mapping due to zone append
1425 * emulation on zoned target. In this case, dm_zone_map_bio()
1426 * calls the target map operation.
1427 */
1428 if (unlikely(dm_emulate_zone_append(md)))
1429 r = dm_zone_map_bio(tio);
1430 else
1431 goto do_map;
1432 } else {
1433do_map:
1434 if (likely(ti->type->map == linear_map))
1435 r = linear_map(ti, clone);
1436 else if (ti->type->map == stripe_map)
1437 r = stripe_map(ti, clone);
1438 else
1439 r = ti->type->map(ti, clone);
1440 }
1441
1442 switch (r) {
1443 case DM_MAPIO_SUBMITTED:
1444 /* target has assumed ownership of this io */
1445 if (!ti->accounts_remapped_io)
1446 dm_start_io_acct(io, clone);
1447 break;
1448 case DM_MAPIO_REMAPPED:
1449 dm_submit_bio_remap(clone, NULL);
1450 break;
1451 case DM_MAPIO_KILL:
1452 case DM_MAPIO_REQUEUE:
1453 if (static_branch_unlikely(&swap_bios_enabled) &&
1454 unlikely(swap_bios_limit(ti, clone)))
1455 up(&md->swap_bios_semaphore);
1456 free_tio(clone);
1457 if (r == DM_MAPIO_KILL)
1458 dm_io_dec_pending(io, BLK_STS_IOERR);
1459 else
1460 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1461 break;
1462 default:
1463 DMCRIT("unimplemented target map return value: %d", r);
1464 BUG();
1465 }
1466}
1467
1468static void setup_split_accounting(struct clone_info *ci, unsigned int len)
1469{
1470 struct dm_io *io = ci->io;
1471
1472 if (ci->sector_count > len) {
1473 /*
1474 * Split needed, save the mapped part for accounting.
1475 * NOTE: dm_accept_partial_bio() will update accordingly.
1476 */
1477 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1478 io->sectors = len;
1479 io->sector_offset = bio_sectors(ci->bio);
1480 }
1481}
1482
1483static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1484 struct dm_target *ti, unsigned int num_bios,
1485 unsigned *len, gfp_t gfp_flag)
1486{
1487 struct bio *bio;
1488 int try = (gfp_flag & GFP_NOWAIT) ? 0 : 1;
1489
1490 for (; try < 2; try++) {
1491 int bio_nr;
1492
1493 if (try && num_bios > 1)
1494 mutex_lock(&ci->io->md->table_devices_lock);
1495 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1496 bio = alloc_tio(ci, ti, bio_nr, len,
1497 try ? GFP_NOIO : GFP_NOWAIT);
1498 if (!bio)
1499 break;
1500
1501 bio_list_add(blist, bio);
1502 }
1503 if (try && num_bios > 1)
1504 mutex_unlock(&ci->io->md->table_devices_lock);
1505 if (bio_nr == num_bios)
1506 return;
1507
1508 while ((bio = bio_list_pop(blist)))
1509 free_tio(bio);
1510 }
1511}
1512
1513static unsigned int __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1514 unsigned int num_bios, unsigned int *len,
1515 gfp_t gfp_flag)
1516{
1517 struct bio_list blist = BIO_EMPTY_LIST;
1518 struct bio *clone;
1519 unsigned int ret = 0;
1520
1521 if (WARN_ON_ONCE(num_bios == 0)) /* num_bios = 0 is a bug in caller */
1522 return 0;
1523
1524 /* dm_accept_partial_bio() is not supported with shared tio->len_ptr */
1525 if (len)
1526 setup_split_accounting(ci, *len);
1527
1528 /*
1529 * Using alloc_multiple_bios(), even if num_bios is 1, to consistently
1530 * support allocating using GFP_NOWAIT with GFP_NOIO fallback.
1531 */
1532 alloc_multiple_bios(&blist, ci, ti, num_bios, len, gfp_flag);
1533 while ((clone = bio_list_pop(&blist))) {
1534 if (num_bios > 1)
1535 dm_tio_set_flag(clone_to_tio(clone), DM_TIO_IS_DUPLICATE_BIO);
1536 __map_bio(clone);
1537 ret += 1;
1538 }
1539
1540 return ret;
1541}
1542
1543static void __send_empty_flush(struct clone_info *ci)
1544{
1545 struct dm_table *t = ci->map;
1546 struct bio flush_bio;
1547
1548 /*
1549 * Use an on-stack bio for this, it's safe since we don't
1550 * need to reference it after submit. It's just used as
1551 * the basis for the clone(s).
1552 */
1553 bio_init(&flush_bio, ci->io->md->disk->part0, NULL, 0,
1554 REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
1555
1556 ci->bio = &flush_bio;
1557 ci->sector_count = 0;
1558 ci->io->tio.clone.bi_iter.bi_size = 0;
1559
1560 for (unsigned int i = 0; i < t->num_targets; i++) {
1561 unsigned int bios;
1562 struct dm_target *ti = dm_table_get_target(t, i);
1563
1564 if (unlikely(ti->num_flush_bios == 0))
1565 continue;
1566
1567 atomic_add(ti->num_flush_bios, &ci->io->io_count);
1568 bios = __send_duplicate_bios(ci, ti, ti->num_flush_bios,
1569 NULL, GFP_NOWAIT);
1570 atomic_sub(ti->num_flush_bios - bios, &ci->io->io_count);
1571 }
1572
1573 /*
1574 * alloc_io() takes one extra reference for submission, so the
1575 * reference won't reach 0 without the following subtraction
1576 */
1577 atomic_sub(1, &ci->io->io_count);
1578
1579 bio_uninit(ci->bio);
1580}
1581
1582static void __send_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1583 unsigned int num_bios, unsigned int max_granularity,
1584 unsigned int max_sectors)
1585{
1586 unsigned int len, bios;
1587
1588 len = min_t(sector_t, ci->sector_count,
1589 __max_io_len(ti, ci->sector, max_granularity, max_sectors));
1590
1591 atomic_add(num_bios, &ci->io->io_count);
1592 bios = __send_duplicate_bios(ci, ti, num_bios, &len, GFP_NOIO);
1593 /*
1594 * alloc_io() takes one extra reference for submission, so the
1595 * reference won't reach 0 without the following (+1) subtraction
1596 */
1597 atomic_sub(num_bios - bios + 1, &ci->io->io_count);
1598
1599 ci->sector += len;
1600 ci->sector_count -= len;
1601}
1602
1603static bool is_abnormal_io(struct bio *bio)
1604{
1605 enum req_op op = bio_op(bio);
1606
1607 if (op != REQ_OP_READ && op != REQ_OP_WRITE && op != REQ_OP_FLUSH) {
1608 switch (op) {
1609 case REQ_OP_DISCARD:
1610 case REQ_OP_SECURE_ERASE:
1611 case REQ_OP_WRITE_ZEROES:
1612 return true;
1613 default:
1614 break;
1615 }
1616 }
1617
1618 return false;
1619}
1620
1621static blk_status_t __process_abnormal_io(struct clone_info *ci,
1622 struct dm_target *ti)
1623{
1624 unsigned int num_bios = 0;
1625 unsigned int max_granularity = 0;
1626 unsigned int max_sectors = 0;
1627 struct queue_limits *limits = dm_get_queue_limits(ti->table->md);
1628
1629 switch (bio_op(ci->bio)) {
1630 case REQ_OP_DISCARD:
1631 num_bios = ti->num_discard_bios;
1632 max_sectors = limits->max_discard_sectors;
1633 if (ti->max_discard_granularity)
1634 max_granularity = max_sectors;
1635 break;
1636 case REQ_OP_SECURE_ERASE:
1637 num_bios = ti->num_secure_erase_bios;
1638 max_sectors = limits->max_secure_erase_sectors;
1639 if (ti->max_secure_erase_granularity)
1640 max_granularity = max_sectors;
1641 break;
1642 case REQ_OP_WRITE_ZEROES:
1643 num_bios = ti->num_write_zeroes_bios;
1644 max_sectors = limits->max_write_zeroes_sectors;
1645 if (ti->max_write_zeroes_granularity)
1646 max_granularity = max_sectors;
1647 break;
1648 default:
1649 break;
1650 }
1651
1652 /*
1653 * Even though the device advertised support for this type of
1654 * request, that does not mean every target supports it, and
1655 * reconfiguration might also have changed that since the
1656 * check was performed.
1657 */
1658 if (unlikely(!num_bios))
1659 return BLK_STS_NOTSUPP;
1660
1661 __send_abnormal_io(ci, ti, num_bios, max_granularity, max_sectors);
1662
1663 return BLK_STS_OK;
1664}
1665
1666/*
1667 * Reuse ->bi_private as dm_io list head for storing all dm_io instances
1668 * associated with this bio, and this bio's bi_private needs to be
1669 * stored in dm_io->data before the reuse.
1670 *
1671 * bio->bi_private is owned by fs or upper layer, so block layer won't
1672 * touch it after splitting. Meantime it won't be changed by anyone after
1673 * bio is submitted. So this reuse is safe.
1674 */
1675static inline struct dm_io **dm_poll_list_head(struct bio *bio)
1676{
1677 return (struct dm_io **)&bio->bi_private;
1678}
1679
1680static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
1681{
1682 struct dm_io **head = dm_poll_list_head(bio);
1683
1684 if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
1685 bio->bi_opf |= REQ_DM_POLL_LIST;
1686 /*
1687 * Save .bi_private into dm_io, so that we can reuse
1688 * .bi_private as dm_io list head for storing dm_io list
1689 */
1690 io->data = bio->bi_private;
1691
1692 /* tell block layer to poll for completion */
1693 bio->bi_cookie = ~BLK_QC_T_NONE;
1694
1695 io->next = NULL;
1696 } else {
1697 /*
1698 * bio recursed due to split, reuse original poll list,
1699 * and save bio->bi_private too.
1700 */
1701 io->data = (*head)->data;
1702 io->next = *head;
1703 }
1704
1705 *head = io;
1706}
1707
1708/*
1709 * Select the correct strategy for processing a non-flush bio.
1710 */
1711static blk_status_t __split_and_process_bio(struct clone_info *ci)
1712{
1713 struct bio *clone;
1714 struct dm_target *ti;
1715 unsigned int len;
1716
1717 ti = dm_table_find_target(ci->map, ci->sector);
1718 if (unlikely(!ti))
1719 return BLK_STS_IOERR;
1720
1721 if (unlikely(ci->is_abnormal_io))
1722 return __process_abnormal_io(ci, ti);
1723
1724 /*
1725 * Only support bio polling for normal IO, and the target io is
1726 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
1727 */
1728 ci->submit_as_polled = !!(ci->bio->bi_opf & REQ_POLLED);
1729
1730 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1731 setup_split_accounting(ci, len);
1732
1733 if (unlikely(ci->bio->bi_opf & REQ_NOWAIT)) {
1734 if (unlikely(!dm_target_supports_nowait(ti->type)))
1735 return BLK_STS_NOTSUPP;
1736
1737 clone = alloc_tio(ci, ti, 0, &len, GFP_NOWAIT);
1738 if (unlikely(!clone))
1739 return BLK_STS_AGAIN;
1740 } else {
1741 clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
1742 }
1743 __map_bio(clone);
1744
1745 ci->sector += len;
1746 ci->sector_count -= len;
1747
1748 return BLK_STS_OK;
1749}
1750
1751static void init_clone_info(struct clone_info *ci, struct dm_io *io,
1752 struct dm_table *map, struct bio *bio, bool is_abnormal)
1753{
1754 ci->map = map;
1755 ci->io = io;
1756 ci->bio = bio;
1757 ci->is_abnormal_io = is_abnormal;
1758 ci->submit_as_polled = false;
1759 ci->sector = bio->bi_iter.bi_sector;
1760 ci->sector_count = bio_sectors(bio);
1761
1762 /* Shouldn't happen but sector_count was being set to 0 so... */
1763 if (static_branch_unlikely(&zoned_enabled) &&
1764 WARN_ON_ONCE(op_is_zone_mgmt(bio_op(bio)) && ci->sector_count))
1765 ci->sector_count = 0;
1766}
1767
1768/*
1769 * Entry point to split a bio into clones and submit them to the targets.
1770 */
1771static void dm_split_and_process_bio(struct mapped_device *md,
1772 struct dm_table *map, struct bio *bio)
1773{
1774 struct clone_info ci;
1775 struct dm_io *io;
1776 blk_status_t error = BLK_STS_OK;
1777 bool is_abnormal;
1778
1779 is_abnormal = is_abnormal_io(bio);
1780 if (unlikely(is_abnormal)) {
1781 /*
1782 * Use bio_split_to_limits() for abnormal IO (e.g. discard, etc)
1783 * otherwise associated queue_limits won't be imposed.
1784 */
1785 bio = bio_split_to_limits(bio);
1786 if (!bio)
1787 return;
1788 }
1789
1790 /* Only support nowait for normal IO */
1791 if (unlikely(bio->bi_opf & REQ_NOWAIT) && !is_abnormal) {
1792 io = alloc_io(md, bio, GFP_NOWAIT);
1793 if (unlikely(!io)) {
1794 /* Unable to do anything without dm_io. */
1795 bio_wouldblock_error(bio);
1796 return;
1797 }
1798 } else {
1799 io = alloc_io(md, bio, GFP_NOIO);
1800 }
1801 init_clone_info(&ci, io, map, bio, is_abnormal);
1802
1803 if (bio->bi_opf & REQ_PREFLUSH) {
1804 __send_empty_flush(&ci);
1805 /* dm_io_complete submits any data associated with flush */
1806 goto out;
1807 }
1808
1809 error = __split_and_process_bio(&ci);
1810 if (error || !ci.sector_count)
1811 goto out;
1812 /*
1813 * Remainder must be passed to submit_bio_noacct() so it gets handled
1814 * *after* bios already submitted have been completely processed.
1815 */
1816 bio_trim(bio, io->sectors, ci.sector_count);
1817 trace_block_split(bio, bio->bi_iter.bi_sector);
1818 bio_inc_remaining(bio);
1819 submit_bio_noacct(bio);
1820out:
1821 /*
1822 * Drop the extra reference count for non-POLLED bio, and hold one
1823 * reference for POLLED bio, which will be released in dm_poll_bio
1824 *
1825 * Add every dm_io instance into the dm_io list head which is stored
1826 * in bio->bi_private, so that dm_poll_bio can poll them all.
1827 */
1828 if (error || !ci.submit_as_polled) {
1829 /*
1830 * In case of submission failure, the extra reference for
1831 * submitting io isn't consumed yet
1832 */
1833 if (error)
1834 atomic_dec(&io->io_count);
1835 dm_io_dec_pending(io, error);
1836 } else
1837 dm_queue_poll_io(bio, io);
1838}
1839
1840static void dm_submit_bio(struct bio *bio)
1841{
1842 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
1843 int srcu_idx;
1844 struct dm_table *map;
1845
1846 map = dm_get_live_table(md, &srcu_idx);
1847
1848 /* If suspended, or map not yet available, queue this IO for later */
1849 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) ||
1850 unlikely(!map)) {
1851 if (bio->bi_opf & REQ_NOWAIT)
1852 bio_wouldblock_error(bio);
1853 else if (bio->bi_opf & REQ_RAHEAD)
1854 bio_io_error(bio);
1855 else
1856 queue_io(md, bio);
1857 goto out;
1858 }
1859
1860 dm_split_and_process_bio(md, map, bio);
1861out:
1862 dm_put_live_table(md, srcu_idx);
1863}
1864
1865static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
1866 unsigned int flags)
1867{
1868 WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
1869
1870 /* don't poll if the mapped io is done */
1871 if (atomic_read(&io->io_count) > 1)
1872 bio_poll(&io->tio.clone, iob, flags);
1873
1874 /* bio_poll holds the last reference */
1875 return atomic_read(&io->io_count) == 1;
1876}
1877
1878static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
1879 unsigned int flags)
1880{
1881 struct dm_io **head = dm_poll_list_head(bio);
1882 struct dm_io *list = *head;
1883 struct dm_io *tmp = NULL;
1884 struct dm_io *curr, *next;
1885
1886 /* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
1887 if (!(bio->bi_opf & REQ_DM_POLL_LIST))
1888 return 0;
1889
1890 WARN_ON_ONCE(!list);
1891
1892 /*
1893 * Restore .bi_private before possibly completing dm_io.
1894 *
1895 * bio_poll() is only possible once @bio has been completely
1896 * submitted via submit_bio_noacct()'s depth-first submission.
1897 * So there is no dm_queue_poll_io() race associated with
1898 * clearing REQ_DM_POLL_LIST here.
1899 */
1900 bio->bi_opf &= ~REQ_DM_POLL_LIST;
1901 bio->bi_private = list->data;
1902
1903 for (curr = list, next = curr->next; curr; curr = next, next =
1904 curr ? curr->next : NULL) {
1905 if (dm_poll_dm_io(curr, iob, flags)) {
1906 /*
1907 * clone_endio() has already occurred, so no
1908 * error handling is needed here.
1909 */
1910 __dm_io_dec_pending(curr);
1911 } else {
1912 curr->next = tmp;
1913 tmp = curr;
1914 }
1915 }
1916
1917 /* Not done? */
1918 if (tmp) {
1919 bio->bi_opf |= REQ_DM_POLL_LIST;
1920 /* Reset bio->bi_private to dm_io list head */
1921 *head = tmp;
1922 return 0;
1923 }
1924 return 1;
1925}
1926
1927/*
1928 *---------------------------------------------------------------
1929 * An IDR is used to keep track of allocated minor numbers.
1930 *---------------------------------------------------------------
1931 */
1932static void free_minor(int minor)
1933{
1934 spin_lock(&_minor_lock);
1935 idr_remove(&_minor_idr, minor);
1936 spin_unlock(&_minor_lock);
1937}
1938
1939/*
1940 * See if the device with a specific minor # is free.
1941 */
1942static int specific_minor(int minor)
1943{
1944 int r;
1945
1946 if (minor >= (1 << MINORBITS))
1947 return -EINVAL;
1948
1949 idr_preload(GFP_KERNEL);
1950 spin_lock(&_minor_lock);
1951
1952 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
1953
1954 spin_unlock(&_minor_lock);
1955 idr_preload_end();
1956 if (r < 0)
1957 return r == -ENOSPC ? -EBUSY : r;
1958 return 0;
1959}
1960
1961static int next_free_minor(int *minor)
1962{
1963 int r;
1964
1965 idr_preload(GFP_KERNEL);
1966 spin_lock(&_minor_lock);
1967
1968 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
1969
1970 spin_unlock(&_minor_lock);
1971 idr_preload_end();
1972 if (r < 0)
1973 return r;
1974 *minor = r;
1975 return 0;
1976}
1977
1978static const struct block_device_operations dm_blk_dops;
1979static const struct block_device_operations dm_rq_blk_dops;
1980static const struct dax_operations dm_dax_ops;
1981
1982static void dm_wq_work(struct work_struct *work);
1983
1984#ifdef CONFIG_BLK_INLINE_ENCRYPTION
1985static void dm_queue_destroy_crypto_profile(struct request_queue *q)
1986{
1987 dm_destroy_crypto_profile(q->crypto_profile);
1988}
1989
1990#else /* CONFIG_BLK_INLINE_ENCRYPTION */
1991
1992static inline void dm_queue_destroy_crypto_profile(struct request_queue *q)
1993{
1994}
1995#endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
1996
1997static void cleanup_mapped_device(struct mapped_device *md)
1998{
1999 if (md->wq)
2000 destroy_workqueue(md->wq);
2001 dm_free_md_mempools(md->mempools);
2002
2003 if (md->dax_dev) {
2004 dax_remove_host(md->disk);
2005 kill_dax(md->dax_dev);
2006 put_dax(md->dax_dev);
2007 md->dax_dev = NULL;
2008 }
2009
2010 dm_cleanup_zoned_dev(md);
2011 if (md->disk) {
2012 spin_lock(&_minor_lock);
2013 md->disk->private_data = NULL;
2014 spin_unlock(&_minor_lock);
2015 if (dm_get_md_type(md) != DM_TYPE_NONE) {
2016 struct table_device *td;
2017
2018 dm_sysfs_exit(md);
2019 list_for_each_entry(td, &md->table_devices, list) {
2020 bd_unlink_disk_holder(td->dm_dev.bdev,
2021 md->disk);
2022 }
2023
2024 /*
2025 * Hold lock to make sure del_gendisk() won't concurrent
2026 * with open/close_table_device().
2027 */
2028 mutex_lock(&md->table_devices_lock);
2029 del_gendisk(md->disk);
2030 mutex_unlock(&md->table_devices_lock);
2031 }
2032 dm_queue_destroy_crypto_profile(md->queue);
2033 put_disk(md->disk);
2034 }
2035
2036 if (md->pending_io) {
2037 free_percpu(md->pending_io);
2038 md->pending_io = NULL;
2039 }
2040
2041 cleanup_srcu_struct(&md->io_barrier);
2042
2043 mutex_destroy(&md->suspend_lock);
2044 mutex_destroy(&md->type_lock);
2045 mutex_destroy(&md->table_devices_lock);
2046 mutex_destroy(&md->swap_bios_lock);
2047
2048 dm_mq_cleanup_mapped_device(md);
2049}
2050
2051/*
2052 * Allocate and initialise a blank device with a given minor.
2053 */
2054static struct mapped_device *alloc_dev(int minor)
2055{
2056 int r, numa_node_id = dm_get_numa_node();
2057 struct mapped_device *md;
2058 void *old_md;
2059
2060 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
2061 if (!md) {
2062 DMERR("unable to allocate device, out of memory.");
2063 return NULL;
2064 }
2065
2066 if (!try_module_get(THIS_MODULE))
2067 goto bad_module_get;
2068
2069 /* get a minor number for the dev */
2070 if (minor == DM_ANY_MINOR)
2071 r = next_free_minor(&minor);
2072 else
2073 r = specific_minor(minor);
2074 if (r < 0)
2075 goto bad_minor;
2076
2077 r = init_srcu_struct(&md->io_barrier);
2078 if (r < 0)
2079 goto bad_io_barrier;
2080
2081 md->numa_node_id = numa_node_id;
2082 md->init_tio_pdu = false;
2083 md->type = DM_TYPE_NONE;
2084 mutex_init(&md->suspend_lock);
2085 mutex_init(&md->type_lock);
2086 mutex_init(&md->table_devices_lock);
2087 spin_lock_init(&md->deferred_lock);
2088 atomic_set(&md->holders, 1);
2089 atomic_set(&md->open_count, 0);
2090 atomic_set(&md->event_nr, 0);
2091 atomic_set(&md->uevent_seq, 0);
2092 INIT_LIST_HEAD(&md->uevent_list);
2093 INIT_LIST_HEAD(&md->table_devices);
2094 spin_lock_init(&md->uevent_lock);
2095
2096 /*
2097 * default to bio-based until DM table is loaded and md->type
2098 * established. If request-based table is loaded: blk-mq will
2099 * override accordingly.
2100 */
2101 md->disk = blk_alloc_disk(md->numa_node_id);
2102 if (!md->disk)
2103 goto bad;
2104 md->queue = md->disk->queue;
2105
2106 init_waitqueue_head(&md->wait);
2107 INIT_WORK(&md->work, dm_wq_work);
2108 INIT_WORK(&md->requeue_work, dm_wq_requeue_work);
2109 init_waitqueue_head(&md->eventq);
2110 init_completion(&md->kobj_holder.completion);
2111
2112 md->requeue_list = NULL;
2113 md->swap_bios = get_swap_bios();
2114 sema_init(&md->swap_bios_semaphore, md->swap_bios);
2115 mutex_init(&md->swap_bios_lock);
2116
2117 md->disk->major = _major;
2118 md->disk->first_minor = minor;
2119 md->disk->minors = 1;
2120 md->disk->flags |= GENHD_FL_NO_PART;
2121 md->disk->fops = &dm_blk_dops;
2122 md->disk->private_data = md;
2123 sprintf(md->disk->disk_name, "dm-%d", minor);
2124
2125 if (IS_ENABLED(CONFIG_FS_DAX)) {
2126 md->dax_dev = alloc_dax(md, &dm_dax_ops);
2127 if (IS_ERR(md->dax_dev)) {
2128 md->dax_dev = NULL;
2129 goto bad;
2130 }
2131 set_dax_nocache(md->dax_dev);
2132 set_dax_nomc(md->dax_dev);
2133 if (dax_add_host(md->dax_dev, md->disk))
2134 goto bad;
2135 }
2136
2137 format_dev_t(md->name, MKDEV(_major, minor));
2138
2139 md->wq = alloc_workqueue("kdmflush/%s", WQ_MEM_RECLAIM, 0, md->name);
2140 if (!md->wq)
2141 goto bad;
2142
2143 md->pending_io = alloc_percpu(unsigned long);
2144 if (!md->pending_io)
2145 goto bad;
2146
2147 r = dm_stats_init(&md->stats);
2148 if (r < 0)
2149 goto bad;
2150
2151 /* Populate the mapping, nobody knows we exist yet */
2152 spin_lock(&_minor_lock);
2153 old_md = idr_replace(&_minor_idr, md, minor);
2154 spin_unlock(&_minor_lock);
2155
2156 BUG_ON(old_md != MINOR_ALLOCED);
2157
2158 return md;
2159
2160bad:
2161 cleanup_mapped_device(md);
2162bad_io_barrier:
2163 free_minor(minor);
2164bad_minor:
2165 module_put(THIS_MODULE);
2166bad_module_get:
2167 kvfree(md);
2168 return NULL;
2169}
2170
2171static void unlock_fs(struct mapped_device *md);
2172
2173static void free_dev(struct mapped_device *md)
2174{
2175 int minor = MINOR(disk_devt(md->disk));
2176
2177 unlock_fs(md);
2178
2179 cleanup_mapped_device(md);
2180
2181 WARN_ON_ONCE(!list_empty(&md->table_devices));
2182 dm_stats_cleanup(&md->stats);
2183 free_minor(minor);
2184
2185 module_put(THIS_MODULE);
2186 kvfree(md);
2187}
2188
2189/*
2190 * Bind a table to the device.
2191 */
2192static void event_callback(void *context)
2193{
2194 unsigned long flags;
2195 LIST_HEAD(uevents);
2196 struct mapped_device *md = context;
2197
2198 spin_lock_irqsave(&md->uevent_lock, flags);
2199 list_splice_init(&md->uevent_list, &uevents);
2200 spin_unlock_irqrestore(&md->uevent_lock, flags);
2201
2202 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2203
2204 atomic_inc(&md->event_nr);
2205 wake_up(&md->eventq);
2206 dm_issue_global_event();
2207}
2208
2209/*
2210 * Returns old map, which caller must destroy.
2211 */
2212static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2213 struct queue_limits *limits)
2214{
2215 struct dm_table *old_map;
2216 sector_t size;
2217 int ret;
2218
2219 lockdep_assert_held(&md->suspend_lock);
2220
2221 size = dm_table_get_size(t);
2222
2223 /*
2224 * Wipe any geometry if the size of the table changed.
2225 */
2226 if (size != dm_get_size(md))
2227 memset(&md->geometry, 0, sizeof(md->geometry));
2228
2229 set_capacity(md->disk, size);
2230
2231 dm_table_event_callback(t, event_callback, md);
2232
2233 if (dm_table_request_based(t)) {
2234 /*
2235 * Leverage the fact that request-based DM targets are
2236 * immutable singletons - used to optimize dm_mq_queue_rq.
2237 */
2238 md->immutable_target = dm_table_get_immutable_target(t);
2239
2240 /*
2241 * There is no need to reload with request-based dm because the
2242 * size of front_pad doesn't change.
2243 *
2244 * Note for future: If you are to reload bioset, prep-ed
2245 * requests in the queue may refer to bio from the old bioset,
2246 * so you must walk through the queue to unprep.
2247 */
2248 if (!md->mempools) {
2249 md->mempools = t->mempools;
2250 t->mempools = NULL;
2251 }
2252 } else {
2253 /*
2254 * The md may already have mempools that need changing.
2255 * If so, reload bioset because front_pad may have changed
2256 * because a different table was loaded.
2257 */
2258 dm_free_md_mempools(md->mempools);
2259 md->mempools = t->mempools;
2260 t->mempools = NULL;
2261 }
2262
2263 ret = dm_table_set_restrictions(t, md->queue, limits);
2264 if (ret) {
2265 old_map = ERR_PTR(ret);
2266 goto out;
2267 }
2268
2269 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2270 rcu_assign_pointer(md->map, (void *)t);
2271 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2272
2273 if (old_map)
2274 dm_sync_table(md);
2275out:
2276 return old_map;
2277}
2278
2279/*
2280 * Returns unbound table for the caller to free.
2281 */
2282static struct dm_table *__unbind(struct mapped_device *md)
2283{
2284 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2285
2286 if (!map)
2287 return NULL;
2288
2289 dm_table_event_callback(map, NULL, NULL);
2290 RCU_INIT_POINTER(md->map, NULL);
2291 dm_sync_table(md);
2292
2293 return map;
2294}
2295
2296/*
2297 * Constructor for a new device.
2298 */
2299int dm_create(int minor, struct mapped_device **result)
2300{
2301 struct mapped_device *md;
2302
2303 md = alloc_dev(minor);
2304 if (!md)
2305 return -ENXIO;
2306
2307 dm_ima_reset_data(md);
2308
2309 *result = md;
2310 return 0;
2311}
2312
2313/*
2314 * Functions to manage md->type.
2315 * All are required to hold md->type_lock.
2316 */
2317void dm_lock_md_type(struct mapped_device *md)
2318{
2319 mutex_lock(&md->type_lock);
2320}
2321
2322void dm_unlock_md_type(struct mapped_device *md)
2323{
2324 mutex_unlock(&md->type_lock);
2325}
2326
2327void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2328{
2329 BUG_ON(!mutex_is_locked(&md->type_lock));
2330 md->type = type;
2331}
2332
2333enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2334{
2335 return md->type;
2336}
2337
2338struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2339{
2340 return md->immutable_target_type;
2341}
2342
2343/*
2344 * Setup the DM device's queue based on md's type
2345 */
2346int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2347{
2348 enum dm_queue_mode type = dm_table_get_type(t);
2349 struct queue_limits limits;
2350 struct table_device *td;
2351 int r;
2352
2353 switch (type) {
2354 case DM_TYPE_REQUEST_BASED:
2355 md->disk->fops = &dm_rq_blk_dops;
2356 r = dm_mq_init_request_queue(md, t);
2357 if (r) {
2358 DMERR("Cannot initialize queue for request-based dm mapped device");
2359 return r;
2360 }
2361 break;
2362 case DM_TYPE_BIO_BASED:
2363 case DM_TYPE_DAX_BIO_BASED:
2364 blk_queue_flag_set(QUEUE_FLAG_IO_STAT, md->queue);
2365 break;
2366 case DM_TYPE_NONE:
2367 WARN_ON_ONCE(true);
2368 break;
2369 }
2370
2371 r = dm_calculate_queue_limits(t, &limits);
2372 if (r) {
2373 DMERR("Cannot calculate initial queue limits");
2374 return r;
2375 }
2376 r = dm_table_set_restrictions(t, md->queue, &limits);
2377 if (r)
2378 return r;
2379
2380 /*
2381 * Hold lock to make sure add_disk() and del_gendisk() won't concurrent
2382 * with open_table_device() and close_table_device().
2383 */
2384 mutex_lock(&md->table_devices_lock);
2385 r = add_disk(md->disk);
2386 mutex_unlock(&md->table_devices_lock);
2387 if (r)
2388 return r;
2389
2390 /*
2391 * Register the holder relationship for devices added before the disk
2392 * was live.
2393 */
2394 list_for_each_entry(td, &md->table_devices, list) {
2395 r = bd_link_disk_holder(td->dm_dev.bdev, md->disk);
2396 if (r)
2397 goto out_undo_holders;
2398 }
2399
2400 r = dm_sysfs_init(md);
2401 if (r)
2402 goto out_undo_holders;
2403
2404 md->type = type;
2405 return 0;
2406
2407out_undo_holders:
2408 list_for_each_entry_continue_reverse(td, &md->table_devices, list)
2409 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
2410 mutex_lock(&md->table_devices_lock);
2411 del_gendisk(md->disk);
2412 mutex_unlock(&md->table_devices_lock);
2413 return r;
2414}
2415
2416struct mapped_device *dm_get_md(dev_t dev)
2417{
2418 struct mapped_device *md;
2419 unsigned int minor = MINOR(dev);
2420
2421 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2422 return NULL;
2423
2424 spin_lock(&_minor_lock);
2425
2426 md = idr_find(&_minor_idr, minor);
2427 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2428 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2429 md = NULL;
2430 goto out;
2431 }
2432 dm_get(md);
2433out:
2434 spin_unlock(&_minor_lock);
2435
2436 return md;
2437}
2438EXPORT_SYMBOL_GPL(dm_get_md);
2439
2440void *dm_get_mdptr(struct mapped_device *md)
2441{
2442 return md->interface_ptr;
2443}
2444
2445void dm_set_mdptr(struct mapped_device *md, void *ptr)
2446{
2447 md->interface_ptr = ptr;
2448}
2449
2450void dm_get(struct mapped_device *md)
2451{
2452 atomic_inc(&md->holders);
2453 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2454}
2455
2456int dm_hold(struct mapped_device *md)
2457{
2458 spin_lock(&_minor_lock);
2459 if (test_bit(DMF_FREEING, &md->flags)) {
2460 spin_unlock(&_minor_lock);
2461 return -EBUSY;
2462 }
2463 dm_get(md);
2464 spin_unlock(&_minor_lock);
2465 return 0;
2466}
2467EXPORT_SYMBOL_GPL(dm_hold);
2468
2469const char *dm_device_name(struct mapped_device *md)
2470{
2471 return md->name;
2472}
2473EXPORT_SYMBOL_GPL(dm_device_name);
2474
2475static void __dm_destroy(struct mapped_device *md, bool wait)
2476{
2477 struct dm_table *map;
2478 int srcu_idx;
2479
2480 might_sleep();
2481
2482 spin_lock(&_minor_lock);
2483 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2484 set_bit(DMF_FREEING, &md->flags);
2485 spin_unlock(&_minor_lock);
2486
2487 blk_mark_disk_dead(md->disk);
2488
2489 /*
2490 * Take suspend_lock so that presuspend and postsuspend methods
2491 * do not race with internal suspend.
2492 */
2493 mutex_lock(&md->suspend_lock);
2494 map = dm_get_live_table(md, &srcu_idx);
2495 if (!dm_suspended_md(md)) {
2496 dm_table_presuspend_targets(map);
2497 set_bit(DMF_SUSPENDED, &md->flags);
2498 set_bit(DMF_POST_SUSPENDING, &md->flags);
2499 dm_table_postsuspend_targets(map);
2500 }
2501 /* dm_put_live_table must be before fsleep, otherwise deadlock is possible */
2502 dm_put_live_table(md, srcu_idx);
2503 mutex_unlock(&md->suspend_lock);
2504
2505 /*
2506 * Rare, but there may be I/O requests still going to complete,
2507 * for example. Wait for all references to disappear.
2508 * No one should increment the reference count of the mapped_device,
2509 * after the mapped_device state becomes DMF_FREEING.
2510 */
2511 if (wait)
2512 while (atomic_read(&md->holders))
2513 fsleep(1000);
2514 else if (atomic_read(&md->holders))
2515 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2516 dm_device_name(md), atomic_read(&md->holders));
2517
2518 dm_table_destroy(__unbind(md));
2519 free_dev(md);
2520}
2521
2522void dm_destroy(struct mapped_device *md)
2523{
2524 __dm_destroy(md, true);
2525}
2526
2527void dm_destroy_immediate(struct mapped_device *md)
2528{
2529 __dm_destroy(md, false);
2530}
2531
2532void dm_put(struct mapped_device *md)
2533{
2534 atomic_dec(&md->holders);
2535}
2536EXPORT_SYMBOL_GPL(dm_put);
2537
2538static bool dm_in_flight_bios(struct mapped_device *md)
2539{
2540 int cpu;
2541 unsigned long sum = 0;
2542
2543 for_each_possible_cpu(cpu)
2544 sum += *per_cpu_ptr(md->pending_io, cpu);
2545
2546 return sum != 0;
2547}
2548
2549static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2550{
2551 int r = 0;
2552 DEFINE_WAIT(wait);
2553
2554 while (true) {
2555 prepare_to_wait(&md->wait, &wait, task_state);
2556
2557 if (!dm_in_flight_bios(md))
2558 break;
2559
2560 if (signal_pending_state(task_state, current)) {
2561 r = -EINTR;
2562 break;
2563 }
2564
2565 io_schedule();
2566 }
2567 finish_wait(&md->wait, &wait);
2568
2569 smp_rmb();
2570
2571 return r;
2572}
2573
2574static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2575{
2576 int r = 0;
2577
2578 if (!queue_is_mq(md->queue))
2579 return dm_wait_for_bios_completion(md, task_state);
2580
2581 while (true) {
2582 if (!blk_mq_queue_inflight(md->queue))
2583 break;
2584
2585 if (signal_pending_state(task_state, current)) {
2586 r = -EINTR;
2587 break;
2588 }
2589
2590 fsleep(5000);
2591 }
2592
2593 return r;
2594}
2595
2596/*
2597 * Process the deferred bios
2598 */
2599static void dm_wq_work(struct work_struct *work)
2600{
2601 struct mapped_device *md = container_of(work, struct mapped_device, work);
2602 struct bio *bio;
2603
2604 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2605 spin_lock_irq(&md->deferred_lock);
2606 bio = bio_list_pop(&md->deferred);
2607 spin_unlock_irq(&md->deferred_lock);
2608
2609 if (!bio)
2610 break;
2611
2612 submit_bio_noacct(bio);
2613 cond_resched();
2614 }
2615}
2616
2617static void dm_queue_flush(struct mapped_device *md)
2618{
2619 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2620 smp_mb__after_atomic();
2621 queue_work(md->wq, &md->work);
2622}
2623
2624/*
2625 * Swap in a new table, returning the old one for the caller to destroy.
2626 */
2627struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2628{
2629 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2630 struct queue_limits limits;
2631 int r;
2632
2633 mutex_lock(&md->suspend_lock);
2634
2635 /* device must be suspended */
2636 if (!dm_suspended_md(md))
2637 goto out;
2638
2639 /*
2640 * If the new table has no data devices, retain the existing limits.
2641 * This helps multipath with queue_if_no_path if all paths disappear,
2642 * then new I/O is queued based on these limits, and then some paths
2643 * reappear.
2644 */
2645 if (dm_table_has_no_data_devices(table)) {
2646 live_map = dm_get_live_table_fast(md);
2647 if (live_map)
2648 limits = md->queue->limits;
2649 dm_put_live_table_fast(md);
2650 }
2651
2652 if (!live_map) {
2653 r = dm_calculate_queue_limits(table, &limits);
2654 if (r) {
2655 map = ERR_PTR(r);
2656 goto out;
2657 }
2658 }
2659
2660 map = __bind(md, table, &limits);
2661 dm_issue_global_event();
2662
2663out:
2664 mutex_unlock(&md->suspend_lock);
2665 return map;
2666}
2667
2668/*
2669 * Functions to lock and unlock any filesystem running on the
2670 * device.
2671 */
2672static int lock_fs(struct mapped_device *md)
2673{
2674 int r;
2675
2676 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2677
2678 r = bdev_freeze(md->disk->part0);
2679 if (!r)
2680 set_bit(DMF_FROZEN, &md->flags);
2681 return r;
2682}
2683
2684static void unlock_fs(struct mapped_device *md)
2685{
2686 if (!test_bit(DMF_FROZEN, &md->flags))
2687 return;
2688 bdev_thaw(md->disk->part0);
2689 clear_bit(DMF_FROZEN, &md->flags);
2690}
2691
2692/*
2693 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2694 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2695 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2696 *
2697 * If __dm_suspend returns 0, the device is completely quiescent
2698 * now. There is no request-processing activity. All new requests
2699 * are being added to md->deferred list.
2700 */
2701static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2702 unsigned int suspend_flags, unsigned int task_state,
2703 int dmf_suspended_flag)
2704{
2705 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2706 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2707 int r;
2708
2709 lockdep_assert_held(&md->suspend_lock);
2710
2711 /*
2712 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2713 * This flag is cleared before dm_suspend returns.
2714 */
2715 if (noflush)
2716 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2717 else
2718 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2719
2720 /*
2721 * This gets reverted if there's an error later and the targets
2722 * provide the .presuspend_undo hook.
2723 */
2724 dm_table_presuspend_targets(map);
2725
2726 /*
2727 * Flush I/O to the device.
2728 * Any I/O submitted after lock_fs() may not be flushed.
2729 * noflush takes precedence over do_lockfs.
2730 * (lock_fs() flushes I/Os and waits for them to complete.)
2731 */
2732 if (!noflush && do_lockfs) {
2733 r = lock_fs(md);
2734 if (r) {
2735 dm_table_presuspend_undo_targets(map);
2736 return r;
2737 }
2738 }
2739
2740 /*
2741 * Here we must make sure that no processes are submitting requests
2742 * to target drivers i.e. no one may be executing
2743 * dm_split_and_process_bio from dm_submit_bio.
2744 *
2745 * To get all processes out of dm_split_and_process_bio in dm_submit_bio,
2746 * we take the write lock. To prevent any process from reentering
2747 * dm_split_and_process_bio from dm_submit_bio and quiesce the thread
2748 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2749 * flush_workqueue(md->wq).
2750 */
2751 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2752 if (map)
2753 synchronize_srcu(&md->io_barrier);
2754
2755 /*
2756 * Stop md->queue before flushing md->wq in case request-based
2757 * dm defers requests to md->wq from md->queue.
2758 */
2759 if (dm_request_based(md))
2760 dm_stop_queue(md->queue);
2761
2762 flush_workqueue(md->wq);
2763
2764 /*
2765 * At this point no more requests are entering target request routines.
2766 * We call dm_wait_for_completion to wait for all existing requests
2767 * to finish.
2768 */
2769 r = dm_wait_for_completion(md, task_state);
2770 if (!r)
2771 set_bit(dmf_suspended_flag, &md->flags);
2772
2773 if (noflush)
2774 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2775 if (map)
2776 synchronize_srcu(&md->io_barrier);
2777
2778 /* were we interrupted ? */
2779 if (r < 0) {
2780 dm_queue_flush(md);
2781
2782 if (dm_request_based(md))
2783 dm_start_queue(md->queue);
2784
2785 unlock_fs(md);
2786 dm_table_presuspend_undo_targets(map);
2787 /* pushback list is already flushed, so skip flush */
2788 }
2789
2790 return r;
2791}
2792
2793/*
2794 * We need to be able to change a mapping table under a mounted
2795 * filesystem. For example we might want to move some data in
2796 * the background. Before the table can be swapped with
2797 * dm_bind_table, dm_suspend must be called to flush any in
2798 * flight bios and ensure that any further io gets deferred.
2799 */
2800/*
2801 * Suspend mechanism in request-based dm.
2802 *
2803 * 1. Flush all I/Os by lock_fs() if needed.
2804 * 2. Stop dispatching any I/O by stopping the request_queue.
2805 * 3. Wait for all in-flight I/Os to be completed or requeued.
2806 *
2807 * To abort suspend, start the request_queue.
2808 */
2809int dm_suspend(struct mapped_device *md, unsigned int suspend_flags)
2810{
2811 struct dm_table *map = NULL;
2812 int r = 0;
2813
2814retry:
2815 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2816
2817 if (dm_suspended_md(md)) {
2818 r = -EINVAL;
2819 goto out_unlock;
2820 }
2821
2822 if (dm_suspended_internally_md(md)) {
2823 /* already internally suspended, wait for internal resume */
2824 mutex_unlock(&md->suspend_lock);
2825 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2826 if (r)
2827 return r;
2828 goto retry;
2829 }
2830
2831 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2832 if (!map) {
2833 /* avoid deadlock with fs/namespace.c:do_mount() */
2834 suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
2835 }
2836
2837 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2838 if (r)
2839 goto out_unlock;
2840
2841 set_bit(DMF_POST_SUSPENDING, &md->flags);
2842 dm_table_postsuspend_targets(map);
2843 clear_bit(DMF_POST_SUSPENDING, &md->flags);
2844
2845out_unlock:
2846 mutex_unlock(&md->suspend_lock);
2847 return r;
2848}
2849
2850static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2851{
2852 if (map) {
2853 int r = dm_table_resume_targets(map);
2854
2855 if (r)
2856 return r;
2857 }
2858
2859 dm_queue_flush(md);
2860
2861 /*
2862 * Flushing deferred I/Os must be done after targets are resumed
2863 * so that mapping of targets can work correctly.
2864 * Request-based dm is queueing the deferred I/Os in its request_queue.
2865 */
2866 if (dm_request_based(md))
2867 dm_start_queue(md->queue);
2868
2869 unlock_fs(md);
2870
2871 return 0;
2872}
2873
2874int dm_resume(struct mapped_device *md)
2875{
2876 int r;
2877 struct dm_table *map = NULL;
2878
2879retry:
2880 r = -EINVAL;
2881 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2882
2883 if (!dm_suspended_md(md))
2884 goto out;
2885
2886 if (dm_suspended_internally_md(md)) {
2887 /* already internally suspended, wait for internal resume */
2888 mutex_unlock(&md->suspend_lock);
2889 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2890 if (r)
2891 return r;
2892 goto retry;
2893 }
2894
2895 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2896 if (!map || !dm_table_get_size(map))
2897 goto out;
2898
2899 r = __dm_resume(md, map);
2900 if (r)
2901 goto out;
2902
2903 clear_bit(DMF_SUSPENDED, &md->flags);
2904out:
2905 mutex_unlock(&md->suspend_lock);
2906
2907 return r;
2908}
2909
2910/*
2911 * Internal suspend/resume works like userspace-driven suspend. It waits
2912 * until all bios finish and prevents issuing new bios to the target drivers.
2913 * It may be used only from the kernel.
2914 */
2915
2916static void __dm_internal_suspend(struct mapped_device *md, unsigned int suspend_flags)
2917{
2918 struct dm_table *map = NULL;
2919
2920 lockdep_assert_held(&md->suspend_lock);
2921
2922 if (md->internal_suspend_count++)
2923 return; /* nested internal suspend */
2924
2925 if (dm_suspended_md(md)) {
2926 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2927 return; /* nest suspend */
2928 }
2929
2930 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2931
2932 /*
2933 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2934 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2935 * would require changing .presuspend to return an error -- avoid this
2936 * until there is a need for more elaborate variants of internal suspend.
2937 */
2938 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2939 DMF_SUSPENDED_INTERNALLY);
2940
2941 set_bit(DMF_POST_SUSPENDING, &md->flags);
2942 dm_table_postsuspend_targets(map);
2943 clear_bit(DMF_POST_SUSPENDING, &md->flags);
2944}
2945
2946static void __dm_internal_resume(struct mapped_device *md)
2947{
2948 BUG_ON(!md->internal_suspend_count);
2949
2950 if (--md->internal_suspend_count)
2951 return; /* resume from nested internal suspend */
2952
2953 if (dm_suspended_md(md))
2954 goto done; /* resume from nested suspend */
2955
2956 /*
2957 * NOTE: existing callers don't need to call dm_table_resume_targets
2958 * (which may fail -- so best to avoid it for now by passing NULL map)
2959 */
2960 (void) __dm_resume(md, NULL);
2961
2962done:
2963 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2964 smp_mb__after_atomic();
2965 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
2966}
2967
2968void dm_internal_suspend_noflush(struct mapped_device *md)
2969{
2970 mutex_lock(&md->suspend_lock);
2971 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2972 mutex_unlock(&md->suspend_lock);
2973}
2974EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2975
2976void dm_internal_resume(struct mapped_device *md)
2977{
2978 mutex_lock(&md->suspend_lock);
2979 __dm_internal_resume(md);
2980 mutex_unlock(&md->suspend_lock);
2981}
2982EXPORT_SYMBOL_GPL(dm_internal_resume);
2983
2984/*
2985 * Fast variants of internal suspend/resume hold md->suspend_lock,
2986 * which prevents interaction with userspace-driven suspend.
2987 */
2988
2989void dm_internal_suspend_fast(struct mapped_device *md)
2990{
2991 mutex_lock(&md->suspend_lock);
2992 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2993 return;
2994
2995 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2996 synchronize_srcu(&md->io_barrier);
2997 flush_workqueue(md->wq);
2998 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
2999}
3000EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3001
3002void dm_internal_resume_fast(struct mapped_device *md)
3003{
3004 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3005 goto done;
3006
3007 dm_queue_flush(md);
3008
3009done:
3010 mutex_unlock(&md->suspend_lock);
3011}
3012EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3013
3014/*
3015 *---------------------------------------------------------------
3016 * Event notification.
3017 *---------------------------------------------------------------
3018 */
3019int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3020 unsigned int cookie, bool need_resize_uevent)
3021{
3022 int r;
3023 unsigned int noio_flag;
3024 char udev_cookie[DM_COOKIE_LENGTH];
3025 char *envp[3] = { NULL, NULL, NULL };
3026 char **envpp = envp;
3027 if (cookie) {
3028 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3029 DM_COOKIE_ENV_VAR_NAME, cookie);
3030 *envpp++ = udev_cookie;
3031 }
3032 if (need_resize_uevent) {
3033 *envpp++ = "RESIZE=1";
3034 }
3035
3036 noio_flag = memalloc_noio_save();
3037
3038 r = kobject_uevent_env(&disk_to_dev(md->disk)->kobj, action, envp);
3039
3040 memalloc_noio_restore(noio_flag);
3041
3042 return r;
3043}
3044
3045uint32_t dm_next_uevent_seq(struct mapped_device *md)
3046{
3047 return atomic_add_return(1, &md->uevent_seq);
3048}
3049
3050uint32_t dm_get_event_nr(struct mapped_device *md)
3051{
3052 return atomic_read(&md->event_nr);
3053}
3054
3055int dm_wait_event(struct mapped_device *md, int event_nr)
3056{
3057 return wait_event_interruptible(md->eventq,
3058 (event_nr != atomic_read(&md->event_nr)));
3059}
3060
3061void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3062{
3063 unsigned long flags;
3064
3065 spin_lock_irqsave(&md->uevent_lock, flags);
3066 list_add(elist, &md->uevent_list);
3067 spin_unlock_irqrestore(&md->uevent_lock, flags);
3068}
3069
3070/*
3071 * The gendisk is only valid as long as you have a reference
3072 * count on 'md'.
3073 */
3074struct gendisk *dm_disk(struct mapped_device *md)
3075{
3076 return md->disk;
3077}
3078EXPORT_SYMBOL_GPL(dm_disk);
3079
3080struct kobject *dm_kobject(struct mapped_device *md)
3081{
3082 return &md->kobj_holder.kobj;
3083}
3084
3085struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3086{
3087 struct mapped_device *md;
3088
3089 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3090
3091 spin_lock(&_minor_lock);
3092 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
3093 md = NULL;
3094 goto out;
3095 }
3096 dm_get(md);
3097out:
3098 spin_unlock(&_minor_lock);
3099
3100 return md;
3101}
3102
3103int dm_suspended_md(struct mapped_device *md)
3104{
3105 return test_bit(DMF_SUSPENDED, &md->flags);
3106}
3107
3108static int dm_post_suspending_md(struct mapped_device *md)
3109{
3110 return test_bit(DMF_POST_SUSPENDING, &md->flags);
3111}
3112
3113int dm_suspended_internally_md(struct mapped_device *md)
3114{
3115 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3116}
3117
3118int dm_test_deferred_remove_flag(struct mapped_device *md)
3119{
3120 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3121}
3122
3123int dm_suspended(struct dm_target *ti)
3124{
3125 return dm_suspended_md(ti->table->md);
3126}
3127EXPORT_SYMBOL_GPL(dm_suspended);
3128
3129int dm_post_suspending(struct dm_target *ti)
3130{
3131 return dm_post_suspending_md(ti->table->md);
3132}
3133EXPORT_SYMBOL_GPL(dm_post_suspending);
3134
3135int dm_noflush_suspending(struct dm_target *ti)
3136{
3137 return __noflush_suspending(ti->table->md);
3138}
3139EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3140
3141void dm_free_md_mempools(struct dm_md_mempools *pools)
3142{
3143 if (!pools)
3144 return;
3145
3146 bioset_exit(&pools->bs);
3147 bioset_exit(&pools->io_bs);
3148
3149 kfree(pools);
3150}
3151
3152struct dm_pr {
3153 u64 old_key;
3154 u64 new_key;
3155 u32 flags;
3156 bool abort;
3157 bool fail_early;
3158 int ret;
3159 enum pr_type type;
3160 struct pr_keys *read_keys;
3161 struct pr_held_reservation *rsv;
3162};
3163
3164static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3165 struct dm_pr *pr)
3166{
3167 struct mapped_device *md = bdev->bd_disk->private_data;
3168 struct dm_table *table;
3169 struct dm_target *ti;
3170 int ret = -ENOTTY, srcu_idx;
3171
3172 table = dm_get_live_table(md, &srcu_idx);
3173 if (!table || !dm_table_get_size(table))
3174 goto out;
3175
3176 /* We only support devices that have a single target */
3177 if (table->num_targets != 1)
3178 goto out;
3179 ti = dm_table_get_target(table, 0);
3180
3181 if (dm_suspended_md(md)) {
3182 ret = -EAGAIN;
3183 goto out;
3184 }
3185
3186 ret = -EINVAL;
3187 if (!ti->type->iterate_devices)
3188 goto out;
3189
3190 ti->type->iterate_devices(ti, fn, pr);
3191 ret = 0;
3192out:
3193 dm_put_live_table(md, srcu_idx);
3194 return ret;
3195}
3196
3197/*
3198 * For register / unregister we need to manually call out to every path.
3199 */
3200static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3201 sector_t start, sector_t len, void *data)
3202{
3203 struct dm_pr *pr = data;
3204 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3205 int ret;
3206
3207 if (!ops || !ops->pr_register) {
3208 pr->ret = -EOPNOTSUPP;
3209 return -1;
3210 }
3211
3212 ret = ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3213 if (!ret)
3214 return 0;
3215
3216 if (!pr->ret)
3217 pr->ret = ret;
3218
3219 if (pr->fail_early)
3220 return -1;
3221
3222 return 0;
3223}
3224
3225static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3226 u32 flags)
3227{
3228 struct dm_pr pr = {
3229 .old_key = old_key,
3230 .new_key = new_key,
3231 .flags = flags,
3232 .fail_early = true,
3233 .ret = 0,
3234 };
3235 int ret;
3236
3237 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3238 if (ret) {
3239 /* Didn't even get to register a path */
3240 return ret;
3241 }
3242
3243 if (!pr.ret)
3244 return 0;
3245 ret = pr.ret;
3246
3247 if (!new_key)
3248 return ret;
3249
3250 /* unregister all paths if we failed to register any path */
3251 pr.old_key = new_key;
3252 pr.new_key = 0;
3253 pr.flags = 0;
3254 pr.fail_early = false;
3255 (void) dm_call_pr(bdev, __dm_pr_register, &pr);
3256 return ret;
3257}
3258
3259
3260static int __dm_pr_reserve(struct dm_target *ti, struct dm_dev *dev,
3261 sector_t start, sector_t len, void *data)
3262{
3263 struct dm_pr *pr = data;
3264 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3265
3266 if (!ops || !ops->pr_reserve) {
3267 pr->ret = -EOPNOTSUPP;
3268 return -1;
3269 }
3270
3271 pr->ret = ops->pr_reserve(dev->bdev, pr->old_key, pr->type, pr->flags);
3272 if (!pr->ret)
3273 return -1;
3274
3275 return 0;
3276}
3277
3278static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3279 u32 flags)
3280{
3281 struct dm_pr pr = {
3282 .old_key = key,
3283 .flags = flags,
3284 .type = type,
3285 .fail_early = false,
3286 .ret = 0,
3287 };
3288 int ret;
3289
3290 ret = dm_call_pr(bdev, __dm_pr_reserve, &pr);
3291 if (ret)
3292 return ret;
3293
3294 return pr.ret;
3295}
3296
3297/*
3298 * If there is a non-All Registrants type of reservation, the release must be
3299 * sent down the holding path. For the cases where there is no reservation or
3300 * the path is not the holder the device will also return success, so we must
3301 * try each path to make sure we got the correct path.
3302 */
3303static int __dm_pr_release(struct dm_target *ti, struct dm_dev *dev,
3304 sector_t start, sector_t len, void *data)
3305{
3306 struct dm_pr *pr = data;
3307 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3308
3309 if (!ops || !ops->pr_release) {
3310 pr->ret = -EOPNOTSUPP;
3311 return -1;
3312 }
3313
3314 pr->ret = ops->pr_release(dev->bdev, pr->old_key, pr->type);
3315 if (pr->ret)
3316 return -1;
3317
3318 return 0;
3319}
3320
3321static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3322{
3323 struct dm_pr pr = {
3324 .old_key = key,
3325 .type = type,
3326 .fail_early = false,
3327 };
3328 int ret;
3329
3330 ret = dm_call_pr(bdev, __dm_pr_release, &pr);
3331 if (ret)
3332 return ret;
3333
3334 return pr.ret;
3335}
3336
3337static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
3338 sector_t start, sector_t len, void *data)
3339{
3340 struct dm_pr *pr = data;
3341 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3342
3343 if (!ops || !ops->pr_preempt) {
3344 pr->ret = -EOPNOTSUPP;
3345 return -1;
3346 }
3347
3348 pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
3349 pr->abort);
3350 if (!pr->ret)
3351 return -1;
3352
3353 return 0;
3354}
3355
3356static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3357 enum pr_type type, bool abort)
3358{
3359 struct dm_pr pr = {
3360 .new_key = new_key,
3361 .old_key = old_key,
3362 .type = type,
3363 .fail_early = false,
3364 };
3365 int ret;
3366
3367 ret = dm_call_pr(bdev, __dm_pr_preempt, &pr);
3368 if (ret)
3369 return ret;
3370
3371 return pr.ret;
3372}
3373
3374static int dm_pr_clear(struct block_device *bdev, u64 key)
3375{
3376 struct mapped_device *md = bdev->bd_disk->private_data;
3377 const struct pr_ops *ops;
3378 int r, srcu_idx;
3379
3380 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3381 if (r < 0)
3382 goto out;
3383
3384 ops = bdev->bd_disk->fops->pr_ops;
3385 if (ops && ops->pr_clear)
3386 r = ops->pr_clear(bdev, key);
3387 else
3388 r = -EOPNOTSUPP;
3389out:
3390 dm_unprepare_ioctl(md, srcu_idx);
3391 return r;
3392}
3393
3394static int __dm_pr_read_keys(struct dm_target *ti, struct dm_dev *dev,
3395 sector_t start, sector_t len, void *data)
3396{
3397 struct dm_pr *pr = data;
3398 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3399
3400 if (!ops || !ops->pr_read_keys) {
3401 pr->ret = -EOPNOTSUPP;
3402 return -1;
3403 }
3404
3405 pr->ret = ops->pr_read_keys(dev->bdev, pr->read_keys);
3406 if (!pr->ret)
3407 return -1;
3408
3409 return 0;
3410}
3411
3412static int dm_pr_read_keys(struct block_device *bdev, struct pr_keys *keys)
3413{
3414 struct dm_pr pr = {
3415 .read_keys = keys,
3416 };
3417 int ret;
3418
3419 ret = dm_call_pr(bdev, __dm_pr_read_keys, &pr);
3420 if (ret)
3421 return ret;
3422
3423 return pr.ret;
3424}
3425
3426static int __dm_pr_read_reservation(struct dm_target *ti, struct dm_dev *dev,
3427 sector_t start, sector_t len, void *data)
3428{
3429 struct dm_pr *pr = data;
3430 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3431
3432 if (!ops || !ops->pr_read_reservation) {
3433 pr->ret = -EOPNOTSUPP;
3434 return -1;
3435 }
3436
3437 pr->ret = ops->pr_read_reservation(dev->bdev, pr->rsv);
3438 if (!pr->ret)
3439 return -1;
3440
3441 return 0;
3442}
3443
3444static int dm_pr_read_reservation(struct block_device *bdev,
3445 struct pr_held_reservation *rsv)
3446{
3447 struct dm_pr pr = {
3448 .rsv = rsv,
3449 };
3450 int ret;
3451
3452 ret = dm_call_pr(bdev, __dm_pr_read_reservation, &pr);
3453 if (ret)
3454 return ret;
3455
3456 return pr.ret;
3457}
3458
3459static const struct pr_ops dm_pr_ops = {
3460 .pr_register = dm_pr_register,
3461 .pr_reserve = dm_pr_reserve,
3462 .pr_release = dm_pr_release,
3463 .pr_preempt = dm_pr_preempt,
3464 .pr_clear = dm_pr_clear,
3465 .pr_read_keys = dm_pr_read_keys,
3466 .pr_read_reservation = dm_pr_read_reservation,
3467};
3468
3469static const struct block_device_operations dm_blk_dops = {
3470 .submit_bio = dm_submit_bio,
3471 .poll_bio = dm_poll_bio,
3472 .open = dm_blk_open,
3473 .release = dm_blk_close,
3474 .ioctl = dm_blk_ioctl,
3475 .getgeo = dm_blk_getgeo,
3476 .report_zones = dm_blk_report_zones,
3477 .pr_ops = &dm_pr_ops,
3478 .owner = THIS_MODULE
3479};
3480
3481static const struct block_device_operations dm_rq_blk_dops = {
3482 .open = dm_blk_open,
3483 .release = dm_blk_close,
3484 .ioctl = dm_blk_ioctl,
3485 .getgeo = dm_blk_getgeo,
3486 .pr_ops = &dm_pr_ops,
3487 .owner = THIS_MODULE
3488};
3489
3490static const struct dax_operations dm_dax_ops = {
3491 .direct_access = dm_dax_direct_access,
3492 .zero_page_range = dm_dax_zero_page_range,
3493 .recovery_write = dm_dax_recovery_write,
3494};
3495
3496/*
3497 * module hooks
3498 */
3499module_init(dm_init);
3500module_exit(dm_exit);
3501
3502module_param(major, uint, 0);
3503MODULE_PARM_DESC(major, "The major number of the device mapper");
3504
3505module_param(reserved_bio_based_ios, uint, 0644);
3506MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3507
3508module_param(dm_numa_node, int, 0644);
3509MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3510
3511module_param(swap_bios, int, 0644);
3512MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3513
3514MODULE_DESCRIPTION(DM_NAME " driver");
3515MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3516MODULE_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/mm.h>
16#include <linux/sched/signal.h>
17#include <linux/blkpg.h>
18#include <linux/bio.h>
19#include <linux/mempool.h>
20#include <linux/dax.h>
21#include <linux/slab.h>
22#include <linux/idr.h>
23#include <linux/uio.h>
24#include <linux/hdreg.h>
25#include <linux/delay.h>
26#include <linux/wait.h>
27#include <linux/pr.h>
28#include <linux/refcount.h>
29#include <linux/part_stat.h>
30#include <linux/blk-crypto.h>
31#include <linux/keyslot-manager.h>
32
33#define DM_MSG_PREFIX "core"
34
35/*
36 * Cookies are numeric values sent with CHANGE and REMOVE
37 * uevents while resuming, removing or renaming the device.
38 */
39#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
40#define DM_COOKIE_LENGTH 24
41
42static const char *_name = DM_NAME;
43
44static unsigned int major = 0;
45static unsigned int _major = 0;
46
47static DEFINE_IDR(_minor_idr);
48
49static DEFINE_SPINLOCK(_minor_lock);
50
51static void do_deferred_remove(struct work_struct *w);
52
53static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
54
55static struct workqueue_struct *deferred_remove_workqueue;
56
57atomic_t dm_global_event_nr = ATOMIC_INIT(0);
58DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
59
60void dm_issue_global_event(void)
61{
62 atomic_inc(&dm_global_event_nr);
63 wake_up(&dm_global_eventq);
64}
65
66/*
67 * One of these is allocated (on-stack) per original bio.
68 */
69struct clone_info {
70 struct dm_table *map;
71 struct bio *bio;
72 struct dm_io *io;
73 sector_t sector;
74 unsigned sector_count;
75};
76
77#define DM_TARGET_IO_BIO_OFFSET (offsetof(struct dm_target_io, clone))
78#define DM_IO_BIO_OFFSET \
79 (offsetof(struct dm_target_io, clone) + offsetof(struct dm_io, tio))
80
81void *dm_per_bio_data(struct bio *bio, size_t data_size)
82{
83 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
84 if (!tio->inside_dm_io)
85 return (char *)bio - DM_TARGET_IO_BIO_OFFSET - data_size;
86 return (char *)bio - DM_IO_BIO_OFFSET - data_size;
87}
88EXPORT_SYMBOL_GPL(dm_per_bio_data);
89
90struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
91{
92 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
93 if (io->magic == DM_IO_MAGIC)
94 return (struct bio *)((char *)io + DM_IO_BIO_OFFSET);
95 BUG_ON(io->magic != DM_TIO_MAGIC);
96 return (struct bio *)((char *)io + DM_TARGET_IO_BIO_OFFSET);
97}
98EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
99
100unsigned dm_bio_get_target_bio_nr(const struct bio *bio)
101{
102 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
103}
104EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
105
106#define MINOR_ALLOCED ((void *)-1)
107
108#define DM_NUMA_NODE NUMA_NO_NODE
109static int dm_numa_node = DM_NUMA_NODE;
110
111#define DEFAULT_SWAP_BIOS (8 * 1048576 / PAGE_SIZE)
112static int swap_bios = DEFAULT_SWAP_BIOS;
113static int get_swap_bios(void)
114{
115 int latch = READ_ONCE(swap_bios);
116 if (unlikely(latch <= 0))
117 latch = DEFAULT_SWAP_BIOS;
118 return latch;
119}
120
121/*
122 * For mempools pre-allocation at the table loading time.
123 */
124struct dm_md_mempools {
125 struct bio_set bs;
126 struct bio_set io_bs;
127};
128
129struct table_device {
130 struct list_head list;
131 refcount_t count;
132 struct dm_dev dm_dev;
133};
134
135/*
136 * Bio-based DM's mempools' reserved IOs set by the user.
137 */
138#define RESERVED_BIO_BASED_IOS 16
139static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
140
141static int __dm_get_module_param_int(int *module_param, int min, int max)
142{
143 int param = READ_ONCE(*module_param);
144 int modified_param = 0;
145 bool modified = true;
146
147 if (param < min)
148 modified_param = min;
149 else if (param > max)
150 modified_param = max;
151 else
152 modified = false;
153
154 if (modified) {
155 (void)cmpxchg(module_param, param, modified_param);
156 param = modified_param;
157 }
158
159 return param;
160}
161
162unsigned __dm_get_module_param(unsigned *module_param,
163 unsigned def, unsigned max)
164{
165 unsigned param = READ_ONCE(*module_param);
166 unsigned modified_param = 0;
167
168 if (!param)
169 modified_param = def;
170 else if (param > max)
171 modified_param = max;
172
173 if (modified_param) {
174 (void)cmpxchg(module_param, param, modified_param);
175 param = modified_param;
176 }
177
178 return param;
179}
180
181unsigned dm_get_reserved_bio_based_ios(void)
182{
183 return __dm_get_module_param(&reserved_bio_based_ios,
184 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
185}
186EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
187
188static unsigned dm_get_numa_node(void)
189{
190 return __dm_get_module_param_int(&dm_numa_node,
191 DM_NUMA_NODE, num_online_nodes() - 1);
192}
193
194static int __init local_init(void)
195{
196 int r;
197
198 r = dm_uevent_init();
199 if (r)
200 return r;
201
202 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
203 if (!deferred_remove_workqueue) {
204 r = -ENOMEM;
205 goto out_uevent_exit;
206 }
207
208 _major = major;
209 r = register_blkdev(_major, _name);
210 if (r < 0)
211 goto out_free_workqueue;
212
213 if (!_major)
214 _major = r;
215
216 return 0;
217
218out_free_workqueue:
219 destroy_workqueue(deferred_remove_workqueue);
220out_uevent_exit:
221 dm_uevent_exit();
222
223 return r;
224}
225
226static void local_exit(void)
227{
228 flush_scheduled_work();
229 destroy_workqueue(deferred_remove_workqueue);
230
231 unregister_blkdev(_major, _name);
232 dm_uevent_exit();
233
234 _major = 0;
235
236 DMINFO("cleaned up");
237}
238
239static int (*_inits[])(void) __initdata = {
240 local_init,
241 dm_target_init,
242 dm_linear_init,
243 dm_stripe_init,
244 dm_io_init,
245 dm_kcopyd_init,
246 dm_interface_init,
247 dm_statistics_init,
248};
249
250static void (*_exits[])(void) = {
251 local_exit,
252 dm_target_exit,
253 dm_linear_exit,
254 dm_stripe_exit,
255 dm_io_exit,
256 dm_kcopyd_exit,
257 dm_interface_exit,
258 dm_statistics_exit,
259};
260
261static int __init dm_init(void)
262{
263 const int count = ARRAY_SIZE(_inits);
264
265 int r, i;
266
267 for (i = 0; i < count; i++) {
268 r = _inits[i]();
269 if (r)
270 goto bad;
271 }
272
273 return 0;
274
275 bad:
276 while (i--)
277 _exits[i]();
278
279 return r;
280}
281
282static void __exit dm_exit(void)
283{
284 int i = ARRAY_SIZE(_exits);
285
286 while (i--)
287 _exits[i]();
288
289 /*
290 * Should be empty by this point.
291 */
292 idr_destroy(&_minor_idr);
293}
294
295/*
296 * Block device functions
297 */
298int dm_deleting_md(struct mapped_device *md)
299{
300 return test_bit(DMF_DELETING, &md->flags);
301}
302
303static int dm_blk_open(struct block_device *bdev, fmode_t mode)
304{
305 struct mapped_device *md;
306
307 spin_lock(&_minor_lock);
308
309 md = bdev->bd_disk->private_data;
310 if (!md)
311 goto out;
312
313 if (test_bit(DMF_FREEING, &md->flags) ||
314 dm_deleting_md(md)) {
315 md = NULL;
316 goto out;
317 }
318
319 dm_get(md);
320 atomic_inc(&md->open_count);
321out:
322 spin_unlock(&_minor_lock);
323
324 return md ? 0 : -ENXIO;
325}
326
327static void dm_blk_close(struct gendisk *disk, fmode_t mode)
328{
329 struct mapped_device *md;
330
331 spin_lock(&_minor_lock);
332
333 md = disk->private_data;
334 if (WARN_ON(!md))
335 goto out;
336
337 if (atomic_dec_and_test(&md->open_count) &&
338 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
339 queue_work(deferred_remove_workqueue, &deferred_remove_work);
340
341 dm_put(md);
342out:
343 spin_unlock(&_minor_lock);
344}
345
346int dm_open_count(struct mapped_device *md)
347{
348 return atomic_read(&md->open_count);
349}
350
351/*
352 * Guarantees nothing is using the device before it's deleted.
353 */
354int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
355{
356 int r = 0;
357
358 spin_lock(&_minor_lock);
359
360 if (dm_open_count(md)) {
361 r = -EBUSY;
362 if (mark_deferred)
363 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
364 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
365 r = -EEXIST;
366 else
367 set_bit(DMF_DELETING, &md->flags);
368
369 spin_unlock(&_minor_lock);
370
371 return r;
372}
373
374int dm_cancel_deferred_remove(struct mapped_device *md)
375{
376 int r = 0;
377
378 spin_lock(&_minor_lock);
379
380 if (test_bit(DMF_DELETING, &md->flags))
381 r = -EBUSY;
382 else
383 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
384
385 spin_unlock(&_minor_lock);
386
387 return r;
388}
389
390static void do_deferred_remove(struct work_struct *w)
391{
392 dm_deferred_remove();
393}
394
395static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
396{
397 struct mapped_device *md = bdev->bd_disk->private_data;
398
399 return dm_get_geometry(md, geo);
400}
401
402static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
403 struct block_device **bdev)
404{
405 struct dm_target *tgt;
406 struct dm_table *map;
407 int r;
408
409retry:
410 r = -ENOTTY;
411 map = dm_get_live_table(md, srcu_idx);
412 if (!map || !dm_table_get_size(map))
413 return r;
414
415 /* We only support devices that have a single target */
416 if (dm_table_get_num_targets(map) != 1)
417 return r;
418
419 tgt = dm_table_get_target(map, 0);
420 if (!tgt->type->prepare_ioctl)
421 return r;
422
423 if (dm_suspended_md(md))
424 return -EAGAIN;
425
426 r = tgt->type->prepare_ioctl(tgt, bdev);
427 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
428 dm_put_live_table(md, *srcu_idx);
429 msleep(10);
430 goto retry;
431 }
432
433 return r;
434}
435
436static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
437{
438 dm_put_live_table(md, srcu_idx);
439}
440
441static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
442 unsigned int cmd, unsigned long arg)
443{
444 struct mapped_device *md = bdev->bd_disk->private_data;
445 int r, srcu_idx;
446
447 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
448 if (r < 0)
449 goto out;
450
451 if (r > 0) {
452 /*
453 * Target determined this ioctl is being issued against a
454 * subset of the parent bdev; require extra privileges.
455 */
456 if (!capable(CAP_SYS_RAWIO)) {
457 DMDEBUG_LIMIT(
458 "%s: sending ioctl %x to DM device without required privilege.",
459 current->comm, cmd);
460 r = -ENOIOCTLCMD;
461 goto out;
462 }
463 }
464
465 if (!bdev->bd_disk->fops->ioctl)
466 r = -ENOTTY;
467 else
468 r = bdev->bd_disk->fops->ioctl(bdev, mode, cmd, arg);
469out:
470 dm_unprepare_ioctl(md, srcu_idx);
471 return r;
472}
473
474u64 dm_start_time_ns_from_clone(struct bio *bio)
475{
476 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
477 struct dm_io *io = tio->io;
478
479 return jiffies_to_nsecs(io->start_time);
480}
481EXPORT_SYMBOL_GPL(dm_start_time_ns_from_clone);
482
483static void start_io_acct(struct dm_io *io)
484{
485 struct mapped_device *md = io->md;
486 struct bio *bio = io->orig_bio;
487
488 io->start_time = bio_start_io_acct(bio);
489 if (unlikely(dm_stats_used(&md->stats)))
490 dm_stats_account_io(&md->stats, bio_data_dir(bio),
491 bio->bi_iter.bi_sector, bio_sectors(bio),
492 false, 0, &io->stats_aux);
493}
494
495static void end_io_acct(struct mapped_device *md, struct bio *bio,
496 unsigned long start_time, struct dm_stats_aux *stats_aux)
497{
498 unsigned long duration = jiffies - start_time;
499
500 bio_end_io_acct(bio, start_time);
501
502 if (unlikely(dm_stats_used(&md->stats)))
503 dm_stats_account_io(&md->stats, bio_data_dir(bio),
504 bio->bi_iter.bi_sector, bio_sectors(bio),
505 true, duration, stats_aux);
506
507 /* nudge anyone waiting on suspend queue */
508 if (unlikely(wq_has_sleeper(&md->wait)))
509 wake_up(&md->wait);
510}
511
512static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio)
513{
514 struct dm_io *io;
515 struct dm_target_io *tio;
516 struct bio *clone;
517
518 clone = bio_alloc_bioset(GFP_NOIO, 0, &md->io_bs);
519 if (!clone)
520 return NULL;
521
522 tio = container_of(clone, struct dm_target_io, clone);
523 tio->inside_dm_io = true;
524 tio->io = NULL;
525
526 io = container_of(tio, struct dm_io, tio);
527 io->magic = DM_IO_MAGIC;
528 io->status = 0;
529 atomic_set(&io->io_count, 1);
530 io->orig_bio = bio;
531 io->md = md;
532 spin_lock_init(&io->endio_lock);
533
534 start_io_acct(io);
535
536 return io;
537}
538
539static void free_io(struct mapped_device *md, struct dm_io *io)
540{
541 bio_put(&io->tio.clone);
542}
543
544static struct dm_target_io *alloc_tio(struct clone_info *ci, struct dm_target *ti,
545 unsigned target_bio_nr, gfp_t gfp_mask)
546{
547 struct dm_target_io *tio;
548
549 if (!ci->io->tio.io) {
550 /* the dm_target_io embedded in ci->io is available */
551 tio = &ci->io->tio;
552 } else {
553 struct bio *clone = bio_alloc_bioset(gfp_mask, 0, &ci->io->md->bs);
554 if (!clone)
555 return NULL;
556
557 tio = container_of(clone, struct dm_target_io, clone);
558 tio->inside_dm_io = false;
559 }
560
561 tio->magic = DM_TIO_MAGIC;
562 tio->io = ci->io;
563 tio->ti = ti;
564 tio->target_bio_nr = target_bio_nr;
565
566 return tio;
567}
568
569static void free_tio(struct dm_target_io *tio)
570{
571 if (tio->inside_dm_io)
572 return;
573 bio_put(&tio->clone);
574}
575
576/*
577 * Add the bio to the list of deferred io.
578 */
579static void queue_io(struct mapped_device *md, struct bio *bio)
580{
581 unsigned long flags;
582
583 spin_lock_irqsave(&md->deferred_lock, flags);
584 bio_list_add(&md->deferred, bio);
585 spin_unlock_irqrestore(&md->deferred_lock, flags);
586 queue_work(md->wq, &md->work);
587}
588
589/*
590 * Everyone (including functions in this file), should use this
591 * function to access the md->map field, and make sure they call
592 * dm_put_live_table() when finished.
593 */
594struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
595{
596 *srcu_idx = srcu_read_lock(&md->io_barrier);
597
598 return srcu_dereference(md->map, &md->io_barrier);
599}
600
601void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
602{
603 srcu_read_unlock(&md->io_barrier, srcu_idx);
604}
605
606void dm_sync_table(struct mapped_device *md)
607{
608 synchronize_srcu(&md->io_barrier);
609 synchronize_rcu_expedited();
610}
611
612/*
613 * A fast alternative to dm_get_live_table/dm_put_live_table.
614 * The caller must not block between these two functions.
615 */
616static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
617{
618 rcu_read_lock();
619 return rcu_dereference(md->map);
620}
621
622static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
623{
624 rcu_read_unlock();
625}
626
627static char *_dm_claim_ptr = "I belong to device-mapper";
628
629/*
630 * Open a table device so we can use it as a map destination.
631 */
632static int open_table_device(struct table_device *td, dev_t dev,
633 struct mapped_device *md)
634{
635 struct block_device *bdev;
636
637 int r;
638
639 BUG_ON(td->dm_dev.bdev);
640
641 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _dm_claim_ptr);
642 if (IS_ERR(bdev))
643 return PTR_ERR(bdev);
644
645 r = bd_link_disk_holder(bdev, dm_disk(md));
646 if (r) {
647 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
648 return r;
649 }
650
651 td->dm_dev.bdev = bdev;
652 td->dm_dev.dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
653 return 0;
654}
655
656/*
657 * Close a table device that we've been using.
658 */
659static void close_table_device(struct table_device *td, struct mapped_device *md)
660{
661 if (!td->dm_dev.bdev)
662 return;
663
664 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
665 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
666 put_dax(td->dm_dev.dax_dev);
667 td->dm_dev.bdev = NULL;
668 td->dm_dev.dax_dev = NULL;
669}
670
671static struct table_device *find_table_device(struct list_head *l, dev_t dev,
672 fmode_t mode)
673{
674 struct table_device *td;
675
676 list_for_each_entry(td, l, list)
677 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
678 return td;
679
680 return NULL;
681}
682
683int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
684 struct dm_dev **result)
685{
686 int r;
687 struct table_device *td;
688
689 mutex_lock(&md->table_devices_lock);
690 td = find_table_device(&md->table_devices, dev, mode);
691 if (!td) {
692 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
693 if (!td) {
694 mutex_unlock(&md->table_devices_lock);
695 return -ENOMEM;
696 }
697
698 td->dm_dev.mode = mode;
699 td->dm_dev.bdev = NULL;
700
701 if ((r = open_table_device(td, dev, md))) {
702 mutex_unlock(&md->table_devices_lock);
703 kfree(td);
704 return r;
705 }
706
707 format_dev_t(td->dm_dev.name, dev);
708
709 refcount_set(&td->count, 1);
710 list_add(&td->list, &md->table_devices);
711 } else {
712 refcount_inc(&td->count);
713 }
714 mutex_unlock(&md->table_devices_lock);
715
716 *result = &td->dm_dev;
717 return 0;
718}
719
720void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
721{
722 struct table_device *td = container_of(d, struct table_device, dm_dev);
723
724 mutex_lock(&md->table_devices_lock);
725 if (refcount_dec_and_test(&td->count)) {
726 close_table_device(td, md);
727 list_del(&td->list);
728 kfree(td);
729 }
730 mutex_unlock(&md->table_devices_lock);
731}
732
733static void free_table_devices(struct list_head *devices)
734{
735 struct list_head *tmp, *next;
736
737 list_for_each_safe(tmp, next, devices) {
738 struct table_device *td = list_entry(tmp, struct table_device, list);
739
740 DMWARN("dm_destroy: %s still exists with %d references",
741 td->dm_dev.name, refcount_read(&td->count));
742 kfree(td);
743 }
744}
745
746/*
747 * Get the geometry associated with a dm device
748 */
749int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
750{
751 *geo = md->geometry;
752
753 return 0;
754}
755
756/*
757 * Set the geometry of a device.
758 */
759int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
760{
761 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
762
763 if (geo->start > sz) {
764 DMWARN("Start sector is beyond the geometry limits.");
765 return -EINVAL;
766 }
767
768 md->geometry = *geo;
769
770 return 0;
771}
772
773static int __noflush_suspending(struct mapped_device *md)
774{
775 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
776}
777
778/*
779 * Decrements the number of outstanding ios that a bio has been
780 * cloned into, completing the original io if necc.
781 */
782void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
783{
784 unsigned long flags;
785 blk_status_t io_error;
786 struct bio *bio;
787 struct mapped_device *md = io->md;
788 unsigned long start_time = 0;
789 struct dm_stats_aux stats_aux;
790
791 /* Push-back supersedes any I/O errors */
792 if (unlikely(error)) {
793 spin_lock_irqsave(&io->endio_lock, flags);
794 if (!(io->status == BLK_STS_DM_REQUEUE && __noflush_suspending(md)))
795 io->status = error;
796 spin_unlock_irqrestore(&io->endio_lock, flags);
797 }
798
799 if (atomic_dec_and_test(&io->io_count)) {
800 bio = io->orig_bio;
801 if (io->status == BLK_STS_DM_REQUEUE) {
802 /*
803 * Target requested pushing back the I/O.
804 */
805 spin_lock_irqsave(&md->deferred_lock, flags);
806 if (__noflush_suspending(md) &&
807 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) {
808 /* NOTE early return due to BLK_STS_DM_REQUEUE below */
809 bio_list_add_head(&md->deferred, bio);
810 } else {
811 /*
812 * noflush suspend was interrupted or this is
813 * a write to a zoned target.
814 */
815 io->status = BLK_STS_IOERR;
816 }
817 spin_unlock_irqrestore(&md->deferred_lock, flags);
818 }
819
820 io_error = io->status;
821 start_time = io->start_time;
822 stats_aux = io->stats_aux;
823 free_io(md, io);
824 end_io_acct(md, bio, start_time, &stats_aux);
825
826 if (io_error == BLK_STS_DM_REQUEUE)
827 return;
828
829 if ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size) {
830 /*
831 * Preflush done for flush with data, reissue
832 * without REQ_PREFLUSH.
833 */
834 bio->bi_opf &= ~REQ_PREFLUSH;
835 queue_io(md, bio);
836 } else {
837 /* done with normal IO or empty flush */
838 if (io_error)
839 bio->bi_status = io_error;
840 bio_endio(bio);
841 }
842 }
843}
844
845void disable_discard(struct mapped_device *md)
846{
847 struct queue_limits *limits = dm_get_queue_limits(md);
848
849 /* device doesn't really support DISCARD, disable it */
850 limits->max_discard_sectors = 0;
851 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, md->queue);
852}
853
854void disable_write_same(struct mapped_device *md)
855{
856 struct queue_limits *limits = dm_get_queue_limits(md);
857
858 /* device doesn't really support WRITE SAME, disable it */
859 limits->max_write_same_sectors = 0;
860}
861
862void disable_write_zeroes(struct mapped_device *md)
863{
864 struct queue_limits *limits = dm_get_queue_limits(md);
865
866 /* device doesn't really support WRITE ZEROES, disable it */
867 limits->max_write_zeroes_sectors = 0;
868}
869
870static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
871{
872 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
873}
874
875static void clone_endio(struct bio *bio)
876{
877 blk_status_t error = bio->bi_status;
878 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
879 struct dm_io *io = tio->io;
880 struct mapped_device *md = tio->io->md;
881 dm_endio_fn endio = tio->ti->type->end_io;
882 struct request_queue *q = bio->bi_bdev->bd_disk->queue;
883
884 if (unlikely(error == BLK_STS_TARGET)) {
885 if (bio_op(bio) == REQ_OP_DISCARD &&
886 !q->limits.max_discard_sectors)
887 disable_discard(md);
888 else if (bio_op(bio) == REQ_OP_WRITE_SAME &&
889 !q->limits.max_write_same_sectors)
890 disable_write_same(md);
891 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
892 !q->limits.max_write_zeroes_sectors)
893 disable_write_zeroes(md);
894 }
895
896 if (blk_queue_is_zoned(q))
897 dm_zone_endio(io, bio);
898
899 if (endio) {
900 int r = endio(tio->ti, bio, &error);
901 switch (r) {
902 case DM_ENDIO_REQUEUE:
903 /*
904 * Requeuing writes to a sequential zone of a zoned
905 * target will break the sequential write pattern:
906 * fail such IO.
907 */
908 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
909 error = BLK_STS_IOERR;
910 else
911 error = BLK_STS_DM_REQUEUE;
912 fallthrough;
913 case DM_ENDIO_DONE:
914 break;
915 case DM_ENDIO_INCOMPLETE:
916 /* The target will handle the io */
917 return;
918 default:
919 DMWARN("unimplemented target endio return value: %d", r);
920 BUG();
921 }
922 }
923
924 if (unlikely(swap_bios_limit(tio->ti, bio))) {
925 struct mapped_device *md = io->md;
926 up(&md->swap_bios_semaphore);
927 }
928
929 free_tio(tio);
930 dm_io_dec_pending(io, error);
931}
932
933/*
934 * Return maximum size of I/O possible at the supplied sector up to the current
935 * target boundary.
936 */
937static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
938 sector_t target_offset)
939{
940 return ti->len - target_offset;
941}
942
943static sector_t max_io_len(struct dm_target *ti, sector_t sector)
944{
945 sector_t target_offset = dm_target_offset(ti, sector);
946 sector_t len = max_io_len_target_boundary(ti, target_offset);
947 sector_t max_len;
948
949 /*
950 * Does the target need to split IO even further?
951 * - varied (per target) IO splitting is a tenet of DM; this
952 * explains why stacked chunk_sectors based splitting via
953 * blk_max_size_offset() isn't possible here. So pass in
954 * ti->max_io_len to override stacked chunk_sectors.
955 */
956 if (ti->max_io_len) {
957 max_len = blk_max_size_offset(ti->table->md->queue,
958 target_offset, ti->max_io_len);
959 if (len > max_len)
960 len = max_len;
961 }
962
963 return len;
964}
965
966int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
967{
968 if (len > UINT_MAX) {
969 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
970 (unsigned long long)len, UINT_MAX);
971 ti->error = "Maximum size of target IO is too large";
972 return -EINVAL;
973 }
974
975 ti->max_io_len = (uint32_t) len;
976
977 return 0;
978}
979EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
980
981static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
982 sector_t sector, int *srcu_idx)
983 __acquires(md->io_barrier)
984{
985 struct dm_table *map;
986 struct dm_target *ti;
987
988 map = dm_get_live_table(md, srcu_idx);
989 if (!map)
990 return NULL;
991
992 ti = dm_table_find_target(map, sector);
993 if (!ti)
994 return NULL;
995
996 return ti;
997}
998
999static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1000 long nr_pages, void **kaddr, pfn_t *pfn)
1001{
1002 struct mapped_device *md = dax_get_private(dax_dev);
1003 sector_t sector = pgoff * PAGE_SECTORS;
1004 struct dm_target *ti;
1005 long len, ret = -EIO;
1006 int srcu_idx;
1007
1008 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1009
1010 if (!ti)
1011 goto out;
1012 if (!ti->type->direct_access)
1013 goto out;
1014 len = max_io_len(ti, sector) / PAGE_SECTORS;
1015 if (len < 1)
1016 goto out;
1017 nr_pages = min(len, nr_pages);
1018 ret = ti->type->direct_access(ti, pgoff, nr_pages, kaddr, pfn);
1019
1020 out:
1021 dm_put_live_table(md, srcu_idx);
1022
1023 return ret;
1024}
1025
1026static bool dm_dax_supported(struct dax_device *dax_dev, struct block_device *bdev,
1027 int blocksize, sector_t start, sector_t len)
1028{
1029 struct mapped_device *md = dax_get_private(dax_dev);
1030 struct dm_table *map;
1031 bool ret = false;
1032 int srcu_idx;
1033
1034 map = dm_get_live_table(md, &srcu_idx);
1035 if (!map)
1036 goto out;
1037
1038 ret = dm_table_supports_dax(map, device_not_dax_capable, &blocksize);
1039
1040out:
1041 dm_put_live_table(md, srcu_idx);
1042
1043 return ret;
1044}
1045
1046static size_t dm_dax_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
1047 void *addr, size_t bytes, struct iov_iter *i)
1048{
1049 struct mapped_device *md = dax_get_private(dax_dev);
1050 sector_t sector = pgoff * PAGE_SECTORS;
1051 struct dm_target *ti;
1052 long ret = 0;
1053 int srcu_idx;
1054
1055 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1056
1057 if (!ti)
1058 goto out;
1059 if (!ti->type->dax_copy_from_iter) {
1060 ret = copy_from_iter(addr, bytes, i);
1061 goto out;
1062 }
1063 ret = ti->type->dax_copy_from_iter(ti, pgoff, addr, bytes, i);
1064 out:
1065 dm_put_live_table(md, srcu_idx);
1066
1067 return ret;
1068}
1069
1070static size_t dm_dax_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
1071 void *addr, size_t bytes, struct iov_iter *i)
1072{
1073 struct mapped_device *md = dax_get_private(dax_dev);
1074 sector_t sector = pgoff * PAGE_SECTORS;
1075 struct dm_target *ti;
1076 long ret = 0;
1077 int srcu_idx;
1078
1079 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1080
1081 if (!ti)
1082 goto out;
1083 if (!ti->type->dax_copy_to_iter) {
1084 ret = copy_to_iter(addr, bytes, i);
1085 goto out;
1086 }
1087 ret = ti->type->dax_copy_to_iter(ti, pgoff, addr, bytes, i);
1088 out:
1089 dm_put_live_table(md, srcu_idx);
1090
1091 return ret;
1092}
1093
1094static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1095 size_t nr_pages)
1096{
1097 struct mapped_device *md = dax_get_private(dax_dev);
1098 sector_t sector = pgoff * PAGE_SECTORS;
1099 struct dm_target *ti;
1100 int ret = -EIO;
1101 int srcu_idx;
1102
1103 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1104
1105 if (!ti)
1106 goto out;
1107 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1108 /*
1109 * ->zero_page_range() is mandatory dax operation. If we are
1110 * here, something is wrong.
1111 */
1112 goto out;
1113 }
1114 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1115 out:
1116 dm_put_live_table(md, srcu_idx);
1117
1118 return ret;
1119}
1120
1121/*
1122 * A target may call dm_accept_partial_bio only from the map routine. It is
1123 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1124 * operations and REQ_OP_ZONE_APPEND (zone append writes).
1125 *
1126 * dm_accept_partial_bio informs the dm that the target only wants to process
1127 * additional n_sectors sectors of the bio and the rest of the data should be
1128 * sent in a next bio.
1129 *
1130 * A diagram that explains the arithmetics:
1131 * +--------------------+---------------+-------+
1132 * | 1 | 2 | 3 |
1133 * +--------------------+---------------+-------+
1134 *
1135 * <-------------- *tio->len_ptr --------------->
1136 * <------- bi_size ------->
1137 * <-- n_sectors -->
1138 *
1139 * Region 1 was already iterated over with bio_advance or similar function.
1140 * (it may be empty if the target doesn't use bio_advance)
1141 * Region 2 is the remaining bio size that the target wants to process.
1142 * (it may be empty if region 1 is non-empty, although there is no reason
1143 * to make it empty)
1144 * The target requires that region 3 is to be sent in the next bio.
1145 *
1146 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1147 * the partially processed part (the sum of regions 1+2) must be the same for all
1148 * copies of the bio.
1149 */
1150void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1151{
1152 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1153 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1154
1155 BUG_ON(bio->bi_opf & REQ_PREFLUSH);
1156 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1157 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1158 BUG_ON(bi_size > *tio->len_ptr);
1159 BUG_ON(n_sectors > bi_size);
1160
1161 *tio->len_ptr -= bi_size - n_sectors;
1162 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1163}
1164EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1165
1166static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1167{
1168 mutex_lock(&md->swap_bios_lock);
1169 while (latch < md->swap_bios) {
1170 cond_resched();
1171 down(&md->swap_bios_semaphore);
1172 md->swap_bios--;
1173 }
1174 while (latch > md->swap_bios) {
1175 cond_resched();
1176 up(&md->swap_bios_semaphore);
1177 md->swap_bios++;
1178 }
1179 mutex_unlock(&md->swap_bios_lock);
1180}
1181
1182static blk_qc_t __map_bio(struct dm_target_io *tio)
1183{
1184 int r;
1185 sector_t sector;
1186 struct bio *clone = &tio->clone;
1187 struct dm_io *io = tio->io;
1188 struct dm_target *ti = tio->ti;
1189 blk_qc_t ret = BLK_QC_T_NONE;
1190
1191 clone->bi_end_io = clone_endio;
1192
1193 /*
1194 * Map the clone. If r == 0 we don't need to do
1195 * anything, the target has assumed ownership of
1196 * this io.
1197 */
1198 dm_io_inc_pending(io);
1199 sector = clone->bi_iter.bi_sector;
1200
1201 if (unlikely(swap_bios_limit(ti, clone))) {
1202 struct mapped_device *md = io->md;
1203 int latch = get_swap_bios();
1204 if (unlikely(latch != md->swap_bios))
1205 __set_swap_bios_limit(md, latch);
1206 down(&md->swap_bios_semaphore);
1207 }
1208
1209 /*
1210 * Check if the IO needs a special mapping due to zone append emulation
1211 * on zoned target. In this case, dm_zone_map_bio() calls the target
1212 * map operation.
1213 */
1214 if (dm_emulate_zone_append(io->md))
1215 r = dm_zone_map_bio(tio);
1216 else
1217 r = ti->type->map(ti, clone);
1218
1219 switch (r) {
1220 case DM_MAPIO_SUBMITTED:
1221 break;
1222 case DM_MAPIO_REMAPPED:
1223 /* the bio has been remapped so dispatch it */
1224 trace_block_bio_remap(clone, bio_dev(io->orig_bio), sector);
1225 ret = submit_bio_noacct(clone);
1226 break;
1227 case DM_MAPIO_KILL:
1228 if (unlikely(swap_bios_limit(ti, clone))) {
1229 struct mapped_device *md = io->md;
1230 up(&md->swap_bios_semaphore);
1231 }
1232 free_tio(tio);
1233 dm_io_dec_pending(io, BLK_STS_IOERR);
1234 break;
1235 case DM_MAPIO_REQUEUE:
1236 if (unlikely(swap_bios_limit(ti, clone))) {
1237 struct mapped_device *md = io->md;
1238 up(&md->swap_bios_semaphore);
1239 }
1240 free_tio(tio);
1241 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1242 break;
1243 default:
1244 DMWARN("unimplemented target map return value: %d", r);
1245 BUG();
1246 }
1247
1248 return ret;
1249}
1250
1251static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1252{
1253 bio->bi_iter.bi_sector = sector;
1254 bio->bi_iter.bi_size = to_bytes(len);
1255}
1256
1257/*
1258 * Creates a bio that consists of range of complete bvecs.
1259 */
1260static int clone_bio(struct dm_target_io *tio, struct bio *bio,
1261 sector_t sector, unsigned len)
1262{
1263 struct bio *clone = &tio->clone;
1264 int r;
1265
1266 __bio_clone_fast(clone, bio);
1267
1268 r = bio_crypt_clone(clone, bio, GFP_NOIO);
1269 if (r < 0)
1270 return r;
1271
1272 if (bio_integrity(bio)) {
1273 if (unlikely(!dm_target_has_integrity(tio->ti->type) &&
1274 !dm_target_passes_integrity(tio->ti->type))) {
1275 DMWARN("%s: the target %s doesn't support integrity data.",
1276 dm_device_name(tio->io->md),
1277 tio->ti->type->name);
1278 return -EIO;
1279 }
1280
1281 r = bio_integrity_clone(clone, bio, GFP_NOIO);
1282 if (r < 0)
1283 return r;
1284 }
1285
1286 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1287 clone->bi_iter.bi_size = to_bytes(len);
1288
1289 if (bio_integrity(bio))
1290 bio_integrity_trim(clone);
1291
1292 return 0;
1293}
1294
1295static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1296 struct dm_target *ti, unsigned num_bios)
1297{
1298 struct dm_target_io *tio;
1299 int try;
1300
1301 if (!num_bios)
1302 return;
1303
1304 if (num_bios == 1) {
1305 tio = alloc_tio(ci, ti, 0, GFP_NOIO);
1306 bio_list_add(blist, &tio->clone);
1307 return;
1308 }
1309
1310 for (try = 0; try < 2; try++) {
1311 int bio_nr;
1312 struct bio *bio;
1313
1314 if (try)
1315 mutex_lock(&ci->io->md->table_devices_lock);
1316 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1317 tio = alloc_tio(ci, ti, bio_nr, try ? GFP_NOIO : GFP_NOWAIT);
1318 if (!tio)
1319 break;
1320
1321 bio_list_add(blist, &tio->clone);
1322 }
1323 if (try)
1324 mutex_unlock(&ci->io->md->table_devices_lock);
1325 if (bio_nr == num_bios)
1326 return;
1327
1328 while ((bio = bio_list_pop(blist))) {
1329 tio = container_of(bio, struct dm_target_io, clone);
1330 free_tio(tio);
1331 }
1332 }
1333}
1334
1335static blk_qc_t __clone_and_map_simple_bio(struct clone_info *ci,
1336 struct dm_target_io *tio, unsigned *len)
1337{
1338 struct bio *clone = &tio->clone;
1339
1340 tio->len_ptr = len;
1341
1342 __bio_clone_fast(clone, ci->bio);
1343 if (len)
1344 bio_setup_sector(clone, ci->sector, *len);
1345
1346 return __map_bio(tio);
1347}
1348
1349static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1350 unsigned num_bios, unsigned *len)
1351{
1352 struct bio_list blist = BIO_EMPTY_LIST;
1353 struct bio *bio;
1354 struct dm_target_io *tio;
1355
1356 alloc_multiple_bios(&blist, ci, ti, num_bios);
1357
1358 while ((bio = bio_list_pop(&blist))) {
1359 tio = container_of(bio, struct dm_target_io, clone);
1360 (void) __clone_and_map_simple_bio(ci, tio, len);
1361 }
1362}
1363
1364static int __send_empty_flush(struct clone_info *ci)
1365{
1366 unsigned target_nr = 0;
1367 struct dm_target *ti;
1368 struct bio flush_bio;
1369
1370 /*
1371 * Use an on-stack bio for this, it's safe since we don't
1372 * need to reference it after submit. It's just used as
1373 * the basis for the clone(s).
1374 */
1375 bio_init(&flush_bio, NULL, 0);
1376 flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC;
1377 bio_set_dev(&flush_bio, ci->io->md->disk->part0);
1378
1379 ci->bio = &flush_bio;
1380 ci->sector_count = 0;
1381
1382 BUG_ON(bio_has_data(ci->bio));
1383 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1384 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1385
1386 bio_uninit(ci->bio);
1387 return 0;
1388}
1389
1390static int __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1391 sector_t sector, unsigned *len)
1392{
1393 struct bio *bio = ci->bio;
1394 struct dm_target_io *tio;
1395 int r;
1396
1397 tio = alloc_tio(ci, ti, 0, GFP_NOIO);
1398 tio->len_ptr = len;
1399 r = clone_bio(tio, bio, sector, *len);
1400 if (r < 0) {
1401 free_tio(tio);
1402 return r;
1403 }
1404 (void) __map_bio(tio);
1405
1406 return 0;
1407}
1408
1409static int __send_changing_extent_only(struct clone_info *ci, struct dm_target *ti,
1410 unsigned num_bios)
1411{
1412 unsigned len;
1413
1414 /*
1415 * Even though the device advertised support for this type of
1416 * request, that does not mean every target supports it, and
1417 * reconfiguration might also have changed that since the
1418 * check was performed.
1419 */
1420 if (!num_bios)
1421 return -EOPNOTSUPP;
1422
1423 len = min_t(sector_t, ci->sector_count,
1424 max_io_len_target_boundary(ti, dm_target_offset(ti, ci->sector)));
1425
1426 __send_duplicate_bios(ci, ti, num_bios, &len);
1427
1428 ci->sector += len;
1429 ci->sector_count -= len;
1430
1431 return 0;
1432}
1433
1434static bool is_abnormal_io(struct bio *bio)
1435{
1436 bool r = false;
1437
1438 switch (bio_op(bio)) {
1439 case REQ_OP_DISCARD:
1440 case REQ_OP_SECURE_ERASE:
1441 case REQ_OP_WRITE_SAME:
1442 case REQ_OP_WRITE_ZEROES:
1443 r = true;
1444 break;
1445 }
1446
1447 return r;
1448}
1449
1450static bool __process_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1451 int *result)
1452{
1453 struct bio *bio = ci->bio;
1454 unsigned num_bios = 0;
1455
1456 switch (bio_op(bio)) {
1457 case REQ_OP_DISCARD:
1458 num_bios = ti->num_discard_bios;
1459 break;
1460 case REQ_OP_SECURE_ERASE:
1461 num_bios = ti->num_secure_erase_bios;
1462 break;
1463 case REQ_OP_WRITE_SAME:
1464 num_bios = ti->num_write_same_bios;
1465 break;
1466 case REQ_OP_WRITE_ZEROES:
1467 num_bios = ti->num_write_zeroes_bios;
1468 break;
1469 default:
1470 return false;
1471 }
1472
1473 *result = __send_changing_extent_only(ci, ti, num_bios);
1474 return true;
1475}
1476
1477/*
1478 * Select the correct strategy for processing a non-flush bio.
1479 */
1480static int __split_and_process_non_flush(struct clone_info *ci)
1481{
1482 struct dm_target *ti;
1483 unsigned len;
1484 int r;
1485
1486 ti = dm_table_find_target(ci->map, ci->sector);
1487 if (!ti)
1488 return -EIO;
1489
1490 if (__process_abnormal_io(ci, ti, &r))
1491 return r;
1492
1493 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1494
1495 r = __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1496 if (r < 0)
1497 return r;
1498
1499 ci->sector += len;
1500 ci->sector_count -= len;
1501
1502 return 0;
1503}
1504
1505static void init_clone_info(struct clone_info *ci, struct mapped_device *md,
1506 struct dm_table *map, struct bio *bio)
1507{
1508 ci->map = map;
1509 ci->io = alloc_io(md, bio);
1510 ci->sector = bio->bi_iter.bi_sector;
1511}
1512
1513#define __dm_part_stat_sub(part, field, subnd) \
1514 (part_stat_get(part, field) -= (subnd))
1515
1516/*
1517 * Entry point to split a bio into clones and submit them to the targets.
1518 */
1519static blk_qc_t __split_and_process_bio(struct mapped_device *md,
1520 struct dm_table *map, struct bio *bio)
1521{
1522 struct clone_info ci;
1523 blk_qc_t ret = BLK_QC_T_NONE;
1524 int error = 0;
1525
1526 init_clone_info(&ci, md, map, bio);
1527
1528 if (bio->bi_opf & REQ_PREFLUSH) {
1529 error = __send_empty_flush(&ci);
1530 /* dm_io_dec_pending submits any data associated with flush */
1531 } else if (op_is_zone_mgmt(bio_op(bio))) {
1532 ci.bio = bio;
1533 ci.sector_count = 0;
1534 error = __split_and_process_non_flush(&ci);
1535 } else {
1536 ci.bio = bio;
1537 ci.sector_count = bio_sectors(bio);
1538 error = __split_and_process_non_flush(&ci);
1539 if (ci.sector_count && !error) {
1540 /*
1541 * Remainder must be passed to submit_bio_noacct()
1542 * so that it gets handled *after* bios already submitted
1543 * have been completely processed.
1544 * We take a clone of the original to store in
1545 * ci.io->orig_bio to be used by end_io_acct() and
1546 * for dec_pending to use for completion handling.
1547 */
1548 struct bio *b = bio_split(bio, bio_sectors(bio) - ci.sector_count,
1549 GFP_NOIO, &md->queue->bio_split);
1550 ci.io->orig_bio = b;
1551
1552 /*
1553 * Adjust IO stats for each split, otherwise upon queue
1554 * reentry there will be redundant IO accounting.
1555 * NOTE: this is a stop-gap fix, a proper fix involves
1556 * significant refactoring of DM core's bio splitting
1557 * (by eliminating DM's splitting and just using bio_split)
1558 */
1559 part_stat_lock();
1560 __dm_part_stat_sub(dm_disk(md)->part0,
1561 sectors[op_stat_group(bio_op(bio))], ci.sector_count);
1562 part_stat_unlock();
1563
1564 bio_chain(b, bio);
1565 trace_block_split(b, bio->bi_iter.bi_sector);
1566 ret = submit_bio_noacct(bio);
1567 }
1568 }
1569
1570 /* drop the extra reference count */
1571 dm_io_dec_pending(ci.io, errno_to_blk_status(error));
1572 return ret;
1573}
1574
1575static blk_qc_t dm_submit_bio(struct bio *bio)
1576{
1577 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
1578 blk_qc_t ret = BLK_QC_T_NONE;
1579 int srcu_idx;
1580 struct dm_table *map;
1581
1582 map = dm_get_live_table(md, &srcu_idx);
1583 if (unlikely(!map)) {
1584 DMERR_LIMIT("%s: mapping table unavailable, erroring io",
1585 dm_device_name(md));
1586 bio_io_error(bio);
1587 goto out;
1588 }
1589
1590 /* If suspended, queue this IO for later */
1591 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1592 if (bio->bi_opf & REQ_NOWAIT)
1593 bio_wouldblock_error(bio);
1594 else if (bio->bi_opf & REQ_RAHEAD)
1595 bio_io_error(bio);
1596 else
1597 queue_io(md, bio);
1598 goto out;
1599 }
1600
1601 /*
1602 * Use blk_queue_split() for abnormal IO (e.g. discard, writesame, etc)
1603 * otherwise associated queue_limits won't be imposed.
1604 */
1605 if (is_abnormal_io(bio))
1606 blk_queue_split(&bio);
1607
1608 ret = __split_and_process_bio(md, map, bio);
1609out:
1610 dm_put_live_table(md, srcu_idx);
1611 return ret;
1612}
1613
1614/*-----------------------------------------------------------------
1615 * An IDR is used to keep track of allocated minor numbers.
1616 *---------------------------------------------------------------*/
1617static void free_minor(int minor)
1618{
1619 spin_lock(&_minor_lock);
1620 idr_remove(&_minor_idr, minor);
1621 spin_unlock(&_minor_lock);
1622}
1623
1624/*
1625 * See if the device with a specific minor # is free.
1626 */
1627static int specific_minor(int minor)
1628{
1629 int r;
1630
1631 if (minor >= (1 << MINORBITS))
1632 return -EINVAL;
1633
1634 idr_preload(GFP_KERNEL);
1635 spin_lock(&_minor_lock);
1636
1637 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
1638
1639 spin_unlock(&_minor_lock);
1640 idr_preload_end();
1641 if (r < 0)
1642 return r == -ENOSPC ? -EBUSY : r;
1643 return 0;
1644}
1645
1646static int next_free_minor(int *minor)
1647{
1648 int r;
1649
1650 idr_preload(GFP_KERNEL);
1651 spin_lock(&_minor_lock);
1652
1653 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
1654
1655 spin_unlock(&_minor_lock);
1656 idr_preload_end();
1657 if (r < 0)
1658 return r;
1659 *minor = r;
1660 return 0;
1661}
1662
1663static const struct block_device_operations dm_blk_dops;
1664static const struct block_device_operations dm_rq_blk_dops;
1665static const struct dax_operations dm_dax_ops;
1666
1667static void dm_wq_work(struct work_struct *work);
1668
1669#ifdef CONFIG_BLK_INLINE_ENCRYPTION
1670static void dm_queue_destroy_keyslot_manager(struct request_queue *q)
1671{
1672 dm_destroy_keyslot_manager(q->ksm);
1673}
1674
1675#else /* CONFIG_BLK_INLINE_ENCRYPTION */
1676
1677static inline void dm_queue_destroy_keyslot_manager(struct request_queue *q)
1678{
1679}
1680#endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
1681
1682static void cleanup_mapped_device(struct mapped_device *md)
1683{
1684 if (md->wq)
1685 destroy_workqueue(md->wq);
1686 bioset_exit(&md->bs);
1687 bioset_exit(&md->io_bs);
1688
1689 if (md->dax_dev) {
1690 kill_dax(md->dax_dev);
1691 put_dax(md->dax_dev);
1692 md->dax_dev = NULL;
1693 }
1694
1695 if (md->disk) {
1696 spin_lock(&_minor_lock);
1697 md->disk->private_data = NULL;
1698 spin_unlock(&_minor_lock);
1699 del_gendisk(md->disk);
1700 }
1701
1702 if (md->queue)
1703 dm_queue_destroy_keyslot_manager(md->queue);
1704
1705 if (md->disk)
1706 blk_cleanup_disk(md->disk);
1707
1708 cleanup_srcu_struct(&md->io_barrier);
1709
1710 mutex_destroy(&md->suspend_lock);
1711 mutex_destroy(&md->type_lock);
1712 mutex_destroy(&md->table_devices_lock);
1713 mutex_destroy(&md->swap_bios_lock);
1714
1715 dm_mq_cleanup_mapped_device(md);
1716 dm_cleanup_zoned_dev(md);
1717}
1718
1719/*
1720 * Allocate and initialise a blank device with a given minor.
1721 */
1722static struct mapped_device *alloc_dev(int minor)
1723{
1724 int r, numa_node_id = dm_get_numa_node();
1725 struct mapped_device *md;
1726 void *old_md;
1727
1728 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
1729 if (!md) {
1730 DMWARN("unable to allocate device, out of memory.");
1731 return NULL;
1732 }
1733
1734 if (!try_module_get(THIS_MODULE))
1735 goto bad_module_get;
1736
1737 /* get a minor number for the dev */
1738 if (minor == DM_ANY_MINOR)
1739 r = next_free_minor(&minor);
1740 else
1741 r = specific_minor(minor);
1742 if (r < 0)
1743 goto bad_minor;
1744
1745 r = init_srcu_struct(&md->io_barrier);
1746 if (r < 0)
1747 goto bad_io_barrier;
1748
1749 md->numa_node_id = numa_node_id;
1750 md->init_tio_pdu = false;
1751 md->type = DM_TYPE_NONE;
1752 mutex_init(&md->suspend_lock);
1753 mutex_init(&md->type_lock);
1754 mutex_init(&md->table_devices_lock);
1755 spin_lock_init(&md->deferred_lock);
1756 atomic_set(&md->holders, 1);
1757 atomic_set(&md->open_count, 0);
1758 atomic_set(&md->event_nr, 0);
1759 atomic_set(&md->uevent_seq, 0);
1760 INIT_LIST_HEAD(&md->uevent_list);
1761 INIT_LIST_HEAD(&md->table_devices);
1762 spin_lock_init(&md->uevent_lock);
1763
1764 /*
1765 * default to bio-based until DM table is loaded and md->type
1766 * established. If request-based table is loaded: blk-mq will
1767 * override accordingly.
1768 */
1769 md->disk = blk_alloc_disk(md->numa_node_id);
1770 if (!md->disk)
1771 goto bad;
1772 md->queue = md->disk->queue;
1773
1774 init_waitqueue_head(&md->wait);
1775 INIT_WORK(&md->work, dm_wq_work);
1776 init_waitqueue_head(&md->eventq);
1777 init_completion(&md->kobj_holder.completion);
1778
1779 md->swap_bios = get_swap_bios();
1780 sema_init(&md->swap_bios_semaphore, md->swap_bios);
1781 mutex_init(&md->swap_bios_lock);
1782
1783 md->disk->major = _major;
1784 md->disk->first_minor = minor;
1785 md->disk->minors = 1;
1786 md->disk->fops = &dm_blk_dops;
1787 md->disk->queue = md->queue;
1788 md->disk->private_data = md;
1789 sprintf(md->disk->disk_name, "dm-%d", minor);
1790
1791 if (IS_ENABLED(CONFIG_DAX_DRIVER)) {
1792 md->dax_dev = alloc_dax(md, md->disk->disk_name,
1793 &dm_dax_ops, 0);
1794 if (IS_ERR(md->dax_dev))
1795 goto bad;
1796 }
1797
1798 add_disk_no_queue_reg(md->disk);
1799 format_dev_t(md->name, MKDEV(_major, minor));
1800
1801 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
1802 if (!md->wq)
1803 goto bad;
1804
1805 dm_stats_init(&md->stats);
1806
1807 /* Populate the mapping, nobody knows we exist yet */
1808 spin_lock(&_minor_lock);
1809 old_md = idr_replace(&_minor_idr, md, minor);
1810 spin_unlock(&_minor_lock);
1811
1812 BUG_ON(old_md != MINOR_ALLOCED);
1813
1814 return md;
1815
1816bad:
1817 cleanup_mapped_device(md);
1818bad_io_barrier:
1819 free_minor(minor);
1820bad_minor:
1821 module_put(THIS_MODULE);
1822bad_module_get:
1823 kvfree(md);
1824 return NULL;
1825}
1826
1827static void unlock_fs(struct mapped_device *md);
1828
1829static void free_dev(struct mapped_device *md)
1830{
1831 int minor = MINOR(disk_devt(md->disk));
1832
1833 unlock_fs(md);
1834
1835 cleanup_mapped_device(md);
1836
1837 free_table_devices(&md->table_devices);
1838 dm_stats_cleanup(&md->stats);
1839 free_minor(minor);
1840
1841 module_put(THIS_MODULE);
1842 kvfree(md);
1843}
1844
1845static int __bind_mempools(struct mapped_device *md, struct dm_table *t)
1846{
1847 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
1848 int ret = 0;
1849
1850 if (dm_table_bio_based(t)) {
1851 /*
1852 * The md may already have mempools that need changing.
1853 * If so, reload bioset because front_pad may have changed
1854 * because a different table was loaded.
1855 */
1856 bioset_exit(&md->bs);
1857 bioset_exit(&md->io_bs);
1858
1859 } else if (bioset_initialized(&md->bs)) {
1860 /*
1861 * There's no need to reload with request-based dm
1862 * because the size of front_pad doesn't change.
1863 * Note for future: If you are to reload bioset,
1864 * prep-ed requests in the queue may refer
1865 * to bio from the old bioset, so you must walk
1866 * through the queue to unprep.
1867 */
1868 goto out;
1869 }
1870
1871 BUG_ON(!p ||
1872 bioset_initialized(&md->bs) ||
1873 bioset_initialized(&md->io_bs));
1874
1875 ret = bioset_init_from_src(&md->bs, &p->bs);
1876 if (ret)
1877 goto out;
1878 ret = bioset_init_from_src(&md->io_bs, &p->io_bs);
1879 if (ret)
1880 bioset_exit(&md->bs);
1881out:
1882 /* mempool bind completed, no longer need any mempools in the table */
1883 dm_table_free_md_mempools(t);
1884 return ret;
1885}
1886
1887/*
1888 * Bind a table to the device.
1889 */
1890static void event_callback(void *context)
1891{
1892 unsigned long flags;
1893 LIST_HEAD(uevents);
1894 struct mapped_device *md = (struct mapped_device *) context;
1895
1896 spin_lock_irqsave(&md->uevent_lock, flags);
1897 list_splice_init(&md->uevent_list, &uevents);
1898 spin_unlock_irqrestore(&md->uevent_lock, flags);
1899
1900 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
1901
1902 atomic_inc(&md->event_nr);
1903 wake_up(&md->eventq);
1904 dm_issue_global_event();
1905}
1906
1907/*
1908 * Returns old map, which caller must destroy.
1909 */
1910static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
1911 struct queue_limits *limits)
1912{
1913 struct dm_table *old_map;
1914 struct request_queue *q = md->queue;
1915 bool request_based = dm_table_request_based(t);
1916 sector_t size;
1917 int ret;
1918
1919 lockdep_assert_held(&md->suspend_lock);
1920
1921 size = dm_table_get_size(t);
1922
1923 /*
1924 * Wipe any geometry if the size of the table changed.
1925 */
1926 if (size != dm_get_size(md))
1927 memset(&md->geometry, 0, sizeof(md->geometry));
1928
1929 if (!get_capacity(md->disk))
1930 set_capacity(md->disk, size);
1931 else
1932 set_capacity_and_notify(md->disk, size);
1933
1934 dm_table_event_callback(t, event_callback, md);
1935
1936 /*
1937 * The queue hasn't been stopped yet, if the old table type wasn't
1938 * for request-based during suspension. So stop it to prevent
1939 * I/O mapping before resume.
1940 * This must be done before setting the queue restrictions,
1941 * because request-based dm may be run just after the setting.
1942 */
1943 if (request_based)
1944 dm_stop_queue(q);
1945
1946 if (request_based) {
1947 /*
1948 * Leverage the fact that request-based DM targets are
1949 * immutable singletons - used to optimize dm_mq_queue_rq.
1950 */
1951 md->immutable_target = dm_table_get_immutable_target(t);
1952 }
1953
1954 ret = __bind_mempools(md, t);
1955 if (ret) {
1956 old_map = ERR_PTR(ret);
1957 goto out;
1958 }
1959
1960 ret = dm_table_set_restrictions(t, q, limits);
1961 if (ret) {
1962 old_map = ERR_PTR(ret);
1963 goto out;
1964 }
1965
1966 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
1967 rcu_assign_pointer(md->map, (void *)t);
1968 md->immutable_target_type = dm_table_get_immutable_target_type(t);
1969
1970 if (old_map)
1971 dm_sync_table(md);
1972
1973out:
1974 return old_map;
1975}
1976
1977/*
1978 * Returns unbound table for the caller to free.
1979 */
1980static struct dm_table *__unbind(struct mapped_device *md)
1981{
1982 struct dm_table *map = rcu_dereference_protected(md->map, 1);
1983
1984 if (!map)
1985 return NULL;
1986
1987 dm_table_event_callback(map, NULL, NULL);
1988 RCU_INIT_POINTER(md->map, NULL);
1989 dm_sync_table(md);
1990
1991 return map;
1992}
1993
1994/*
1995 * Constructor for a new device.
1996 */
1997int dm_create(int minor, struct mapped_device **result)
1998{
1999 int r;
2000 struct mapped_device *md;
2001
2002 md = alloc_dev(minor);
2003 if (!md)
2004 return -ENXIO;
2005
2006 r = dm_sysfs_init(md);
2007 if (r) {
2008 free_dev(md);
2009 return r;
2010 }
2011
2012 *result = md;
2013 return 0;
2014}
2015
2016/*
2017 * Functions to manage md->type.
2018 * All are required to hold md->type_lock.
2019 */
2020void dm_lock_md_type(struct mapped_device *md)
2021{
2022 mutex_lock(&md->type_lock);
2023}
2024
2025void dm_unlock_md_type(struct mapped_device *md)
2026{
2027 mutex_unlock(&md->type_lock);
2028}
2029
2030void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2031{
2032 BUG_ON(!mutex_is_locked(&md->type_lock));
2033 md->type = type;
2034}
2035
2036enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2037{
2038 return md->type;
2039}
2040
2041struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2042{
2043 return md->immutable_target_type;
2044}
2045
2046/*
2047 * The queue_limits are only valid as long as you have a reference
2048 * count on 'md'.
2049 */
2050struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2051{
2052 BUG_ON(!atomic_read(&md->holders));
2053 return &md->queue->limits;
2054}
2055EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2056
2057/*
2058 * Setup the DM device's queue based on md's type
2059 */
2060int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2061{
2062 int r;
2063 struct queue_limits limits;
2064 enum dm_queue_mode type = dm_get_md_type(md);
2065
2066 switch (type) {
2067 case DM_TYPE_REQUEST_BASED:
2068 md->disk->fops = &dm_rq_blk_dops;
2069 r = dm_mq_init_request_queue(md, t);
2070 if (r) {
2071 DMERR("Cannot initialize queue for request-based dm mapped device");
2072 return r;
2073 }
2074 break;
2075 case DM_TYPE_BIO_BASED:
2076 case DM_TYPE_DAX_BIO_BASED:
2077 break;
2078 case DM_TYPE_NONE:
2079 WARN_ON_ONCE(true);
2080 break;
2081 }
2082
2083 r = dm_calculate_queue_limits(t, &limits);
2084 if (r) {
2085 DMERR("Cannot calculate initial queue limits");
2086 return r;
2087 }
2088 r = dm_table_set_restrictions(t, md->queue, &limits);
2089 if (r)
2090 return r;
2091
2092 blk_register_queue(md->disk);
2093
2094 return 0;
2095}
2096
2097struct mapped_device *dm_get_md(dev_t dev)
2098{
2099 struct mapped_device *md;
2100 unsigned minor = MINOR(dev);
2101
2102 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2103 return NULL;
2104
2105 spin_lock(&_minor_lock);
2106
2107 md = idr_find(&_minor_idr, minor);
2108 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2109 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2110 md = NULL;
2111 goto out;
2112 }
2113 dm_get(md);
2114out:
2115 spin_unlock(&_minor_lock);
2116
2117 return md;
2118}
2119EXPORT_SYMBOL_GPL(dm_get_md);
2120
2121void *dm_get_mdptr(struct mapped_device *md)
2122{
2123 return md->interface_ptr;
2124}
2125
2126void dm_set_mdptr(struct mapped_device *md, void *ptr)
2127{
2128 md->interface_ptr = ptr;
2129}
2130
2131void dm_get(struct mapped_device *md)
2132{
2133 atomic_inc(&md->holders);
2134 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2135}
2136
2137int dm_hold(struct mapped_device *md)
2138{
2139 spin_lock(&_minor_lock);
2140 if (test_bit(DMF_FREEING, &md->flags)) {
2141 spin_unlock(&_minor_lock);
2142 return -EBUSY;
2143 }
2144 dm_get(md);
2145 spin_unlock(&_minor_lock);
2146 return 0;
2147}
2148EXPORT_SYMBOL_GPL(dm_hold);
2149
2150const char *dm_device_name(struct mapped_device *md)
2151{
2152 return md->name;
2153}
2154EXPORT_SYMBOL_GPL(dm_device_name);
2155
2156static void __dm_destroy(struct mapped_device *md, bool wait)
2157{
2158 struct dm_table *map;
2159 int srcu_idx;
2160
2161 might_sleep();
2162
2163 spin_lock(&_minor_lock);
2164 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2165 set_bit(DMF_FREEING, &md->flags);
2166 spin_unlock(&_minor_lock);
2167
2168 blk_set_queue_dying(md->queue);
2169
2170 /*
2171 * Take suspend_lock so that presuspend and postsuspend methods
2172 * do not race with internal suspend.
2173 */
2174 mutex_lock(&md->suspend_lock);
2175 map = dm_get_live_table(md, &srcu_idx);
2176 if (!dm_suspended_md(md)) {
2177 dm_table_presuspend_targets(map);
2178 set_bit(DMF_SUSPENDED, &md->flags);
2179 set_bit(DMF_POST_SUSPENDING, &md->flags);
2180 dm_table_postsuspend_targets(map);
2181 }
2182 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2183 dm_put_live_table(md, srcu_idx);
2184 mutex_unlock(&md->suspend_lock);
2185
2186 /*
2187 * Rare, but there may be I/O requests still going to complete,
2188 * for example. Wait for all references to disappear.
2189 * No one should increment the reference count of the mapped_device,
2190 * after the mapped_device state becomes DMF_FREEING.
2191 */
2192 if (wait)
2193 while (atomic_read(&md->holders))
2194 msleep(1);
2195 else if (atomic_read(&md->holders))
2196 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2197 dm_device_name(md), atomic_read(&md->holders));
2198
2199 dm_sysfs_exit(md);
2200 dm_table_destroy(__unbind(md));
2201 free_dev(md);
2202}
2203
2204void dm_destroy(struct mapped_device *md)
2205{
2206 __dm_destroy(md, true);
2207}
2208
2209void dm_destroy_immediate(struct mapped_device *md)
2210{
2211 __dm_destroy(md, false);
2212}
2213
2214void dm_put(struct mapped_device *md)
2215{
2216 atomic_dec(&md->holders);
2217}
2218EXPORT_SYMBOL_GPL(dm_put);
2219
2220static bool md_in_flight_bios(struct mapped_device *md)
2221{
2222 int cpu;
2223 struct block_device *part = dm_disk(md)->part0;
2224 long sum = 0;
2225
2226 for_each_possible_cpu(cpu) {
2227 sum += part_stat_local_read_cpu(part, in_flight[0], cpu);
2228 sum += part_stat_local_read_cpu(part, in_flight[1], cpu);
2229 }
2230
2231 return sum != 0;
2232}
2233
2234static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2235{
2236 int r = 0;
2237 DEFINE_WAIT(wait);
2238
2239 while (true) {
2240 prepare_to_wait(&md->wait, &wait, task_state);
2241
2242 if (!md_in_flight_bios(md))
2243 break;
2244
2245 if (signal_pending_state(task_state, current)) {
2246 r = -EINTR;
2247 break;
2248 }
2249
2250 io_schedule();
2251 }
2252 finish_wait(&md->wait, &wait);
2253
2254 return r;
2255}
2256
2257static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2258{
2259 int r = 0;
2260
2261 if (!queue_is_mq(md->queue))
2262 return dm_wait_for_bios_completion(md, task_state);
2263
2264 while (true) {
2265 if (!blk_mq_queue_inflight(md->queue))
2266 break;
2267
2268 if (signal_pending_state(task_state, current)) {
2269 r = -EINTR;
2270 break;
2271 }
2272
2273 msleep(5);
2274 }
2275
2276 return r;
2277}
2278
2279/*
2280 * Process the deferred bios
2281 */
2282static void dm_wq_work(struct work_struct *work)
2283{
2284 struct mapped_device *md = container_of(work, struct mapped_device, work);
2285 struct bio *bio;
2286
2287 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2288 spin_lock_irq(&md->deferred_lock);
2289 bio = bio_list_pop(&md->deferred);
2290 spin_unlock_irq(&md->deferred_lock);
2291
2292 if (!bio)
2293 break;
2294
2295 submit_bio_noacct(bio);
2296 }
2297}
2298
2299static void dm_queue_flush(struct mapped_device *md)
2300{
2301 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2302 smp_mb__after_atomic();
2303 queue_work(md->wq, &md->work);
2304}
2305
2306/*
2307 * Swap in a new table, returning the old one for the caller to destroy.
2308 */
2309struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2310{
2311 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2312 struct queue_limits limits;
2313 int r;
2314
2315 mutex_lock(&md->suspend_lock);
2316
2317 /* device must be suspended */
2318 if (!dm_suspended_md(md))
2319 goto out;
2320
2321 /*
2322 * If the new table has no data devices, retain the existing limits.
2323 * This helps multipath with queue_if_no_path if all paths disappear,
2324 * then new I/O is queued based on these limits, and then some paths
2325 * reappear.
2326 */
2327 if (dm_table_has_no_data_devices(table)) {
2328 live_map = dm_get_live_table_fast(md);
2329 if (live_map)
2330 limits = md->queue->limits;
2331 dm_put_live_table_fast(md);
2332 }
2333
2334 if (!live_map) {
2335 r = dm_calculate_queue_limits(table, &limits);
2336 if (r) {
2337 map = ERR_PTR(r);
2338 goto out;
2339 }
2340 }
2341
2342 map = __bind(md, table, &limits);
2343 dm_issue_global_event();
2344
2345out:
2346 mutex_unlock(&md->suspend_lock);
2347 return map;
2348}
2349
2350/*
2351 * Functions to lock and unlock any filesystem running on the
2352 * device.
2353 */
2354static int lock_fs(struct mapped_device *md)
2355{
2356 int r;
2357
2358 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2359
2360 r = freeze_bdev(md->disk->part0);
2361 if (!r)
2362 set_bit(DMF_FROZEN, &md->flags);
2363 return r;
2364}
2365
2366static void unlock_fs(struct mapped_device *md)
2367{
2368 if (!test_bit(DMF_FROZEN, &md->flags))
2369 return;
2370 thaw_bdev(md->disk->part0);
2371 clear_bit(DMF_FROZEN, &md->flags);
2372}
2373
2374/*
2375 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2376 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2377 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2378 *
2379 * If __dm_suspend returns 0, the device is completely quiescent
2380 * now. There is no request-processing activity. All new requests
2381 * are being added to md->deferred list.
2382 */
2383static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2384 unsigned suspend_flags, unsigned int task_state,
2385 int dmf_suspended_flag)
2386{
2387 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2388 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2389 int r;
2390
2391 lockdep_assert_held(&md->suspend_lock);
2392
2393 /*
2394 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2395 * This flag is cleared before dm_suspend returns.
2396 */
2397 if (noflush)
2398 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2399 else
2400 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2401
2402 /*
2403 * This gets reverted if there's an error later and the targets
2404 * provide the .presuspend_undo hook.
2405 */
2406 dm_table_presuspend_targets(map);
2407
2408 /*
2409 * Flush I/O to the device.
2410 * Any I/O submitted after lock_fs() may not be flushed.
2411 * noflush takes precedence over do_lockfs.
2412 * (lock_fs() flushes I/Os and waits for them to complete.)
2413 */
2414 if (!noflush && do_lockfs) {
2415 r = lock_fs(md);
2416 if (r) {
2417 dm_table_presuspend_undo_targets(map);
2418 return r;
2419 }
2420 }
2421
2422 /*
2423 * Here we must make sure that no processes are submitting requests
2424 * to target drivers i.e. no one may be executing
2425 * __split_and_process_bio from dm_submit_bio.
2426 *
2427 * To get all processes out of __split_and_process_bio in dm_submit_bio,
2428 * we take the write lock. To prevent any process from reentering
2429 * __split_and_process_bio from dm_submit_bio and quiesce the thread
2430 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2431 * flush_workqueue(md->wq).
2432 */
2433 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2434 if (map)
2435 synchronize_srcu(&md->io_barrier);
2436
2437 /*
2438 * Stop md->queue before flushing md->wq in case request-based
2439 * dm defers requests to md->wq from md->queue.
2440 */
2441 if (dm_request_based(md))
2442 dm_stop_queue(md->queue);
2443
2444 flush_workqueue(md->wq);
2445
2446 /*
2447 * At this point no more requests are entering target request routines.
2448 * We call dm_wait_for_completion to wait for all existing requests
2449 * to finish.
2450 */
2451 r = dm_wait_for_completion(md, task_state);
2452 if (!r)
2453 set_bit(dmf_suspended_flag, &md->flags);
2454
2455 if (noflush)
2456 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2457 if (map)
2458 synchronize_srcu(&md->io_barrier);
2459
2460 /* were we interrupted ? */
2461 if (r < 0) {
2462 dm_queue_flush(md);
2463
2464 if (dm_request_based(md))
2465 dm_start_queue(md->queue);
2466
2467 unlock_fs(md);
2468 dm_table_presuspend_undo_targets(map);
2469 /* pushback list is already flushed, so skip flush */
2470 }
2471
2472 return r;
2473}
2474
2475/*
2476 * We need to be able to change a mapping table under a mounted
2477 * filesystem. For example we might want to move some data in
2478 * the background. Before the table can be swapped with
2479 * dm_bind_table, dm_suspend must be called to flush any in
2480 * flight bios and ensure that any further io gets deferred.
2481 */
2482/*
2483 * Suspend mechanism in request-based dm.
2484 *
2485 * 1. Flush all I/Os by lock_fs() if needed.
2486 * 2. Stop dispatching any I/O by stopping the request_queue.
2487 * 3. Wait for all in-flight I/Os to be completed or requeued.
2488 *
2489 * To abort suspend, start the request_queue.
2490 */
2491int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
2492{
2493 struct dm_table *map = NULL;
2494 int r = 0;
2495
2496retry:
2497 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2498
2499 if (dm_suspended_md(md)) {
2500 r = -EINVAL;
2501 goto out_unlock;
2502 }
2503
2504 if (dm_suspended_internally_md(md)) {
2505 /* already internally suspended, wait for internal resume */
2506 mutex_unlock(&md->suspend_lock);
2507 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2508 if (r)
2509 return r;
2510 goto retry;
2511 }
2512
2513 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2514
2515 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2516 if (r)
2517 goto out_unlock;
2518
2519 set_bit(DMF_POST_SUSPENDING, &md->flags);
2520 dm_table_postsuspend_targets(map);
2521 clear_bit(DMF_POST_SUSPENDING, &md->flags);
2522
2523out_unlock:
2524 mutex_unlock(&md->suspend_lock);
2525 return r;
2526}
2527
2528static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2529{
2530 if (map) {
2531 int r = dm_table_resume_targets(map);
2532 if (r)
2533 return r;
2534 }
2535
2536 dm_queue_flush(md);
2537
2538 /*
2539 * Flushing deferred I/Os must be done after targets are resumed
2540 * so that mapping of targets can work correctly.
2541 * Request-based dm is queueing the deferred I/Os in its request_queue.
2542 */
2543 if (dm_request_based(md))
2544 dm_start_queue(md->queue);
2545
2546 unlock_fs(md);
2547
2548 return 0;
2549}
2550
2551int dm_resume(struct mapped_device *md)
2552{
2553 int r;
2554 struct dm_table *map = NULL;
2555
2556retry:
2557 r = -EINVAL;
2558 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2559
2560 if (!dm_suspended_md(md))
2561 goto out;
2562
2563 if (dm_suspended_internally_md(md)) {
2564 /* already internally suspended, wait for internal resume */
2565 mutex_unlock(&md->suspend_lock);
2566 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2567 if (r)
2568 return r;
2569 goto retry;
2570 }
2571
2572 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2573 if (!map || !dm_table_get_size(map))
2574 goto out;
2575
2576 r = __dm_resume(md, map);
2577 if (r)
2578 goto out;
2579
2580 clear_bit(DMF_SUSPENDED, &md->flags);
2581out:
2582 mutex_unlock(&md->suspend_lock);
2583
2584 return r;
2585}
2586
2587/*
2588 * Internal suspend/resume works like userspace-driven suspend. It waits
2589 * until all bios finish and prevents issuing new bios to the target drivers.
2590 * It may be used only from the kernel.
2591 */
2592
2593static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
2594{
2595 struct dm_table *map = NULL;
2596
2597 lockdep_assert_held(&md->suspend_lock);
2598
2599 if (md->internal_suspend_count++)
2600 return; /* nested internal suspend */
2601
2602 if (dm_suspended_md(md)) {
2603 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2604 return; /* nest suspend */
2605 }
2606
2607 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2608
2609 /*
2610 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2611 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2612 * would require changing .presuspend to return an error -- avoid this
2613 * until there is a need for more elaborate variants of internal suspend.
2614 */
2615 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2616 DMF_SUSPENDED_INTERNALLY);
2617
2618 set_bit(DMF_POST_SUSPENDING, &md->flags);
2619 dm_table_postsuspend_targets(map);
2620 clear_bit(DMF_POST_SUSPENDING, &md->flags);
2621}
2622
2623static void __dm_internal_resume(struct mapped_device *md)
2624{
2625 BUG_ON(!md->internal_suspend_count);
2626
2627 if (--md->internal_suspend_count)
2628 return; /* resume from nested internal suspend */
2629
2630 if (dm_suspended_md(md))
2631 goto done; /* resume from nested suspend */
2632
2633 /*
2634 * NOTE: existing callers don't need to call dm_table_resume_targets
2635 * (which may fail -- so best to avoid it for now by passing NULL map)
2636 */
2637 (void) __dm_resume(md, NULL);
2638
2639done:
2640 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2641 smp_mb__after_atomic();
2642 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
2643}
2644
2645void dm_internal_suspend_noflush(struct mapped_device *md)
2646{
2647 mutex_lock(&md->suspend_lock);
2648 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2649 mutex_unlock(&md->suspend_lock);
2650}
2651EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2652
2653void dm_internal_resume(struct mapped_device *md)
2654{
2655 mutex_lock(&md->suspend_lock);
2656 __dm_internal_resume(md);
2657 mutex_unlock(&md->suspend_lock);
2658}
2659EXPORT_SYMBOL_GPL(dm_internal_resume);
2660
2661/*
2662 * Fast variants of internal suspend/resume hold md->suspend_lock,
2663 * which prevents interaction with userspace-driven suspend.
2664 */
2665
2666void dm_internal_suspend_fast(struct mapped_device *md)
2667{
2668 mutex_lock(&md->suspend_lock);
2669 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2670 return;
2671
2672 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2673 synchronize_srcu(&md->io_barrier);
2674 flush_workqueue(md->wq);
2675 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
2676}
2677EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
2678
2679void dm_internal_resume_fast(struct mapped_device *md)
2680{
2681 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2682 goto done;
2683
2684 dm_queue_flush(md);
2685
2686done:
2687 mutex_unlock(&md->suspend_lock);
2688}
2689EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
2690
2691/*-----------------------------------------------------------------
2692 * Event notification.
2693 *---------------------------------------------------------------*/
2694int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
2695 unsigned cookie)
2696{
2697 int r;
2698 unsigned noio_flag;
2699 char udev_cookie[DM_COOKIE_LENGTH];
2700 char *envp[] = { udev_cookie, NULL };
2701
2702 noio_flag = memalloc_noio_save();
2703
2704 if (!cookie)
2705 r = kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
2706 else {
2707 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
2708 DM_COOKIE_ENV_VAR_NAME, cookie);
2709 r = kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
2710 action, envp);
2711 }
2712
2713 memalloc_noio_restore(noio_flag);
2714
2715 return r;
2716}
2717
2718uint32_t dm_next_uevent_seq(struct mapped_device *md)
2719{
2720 return atomic_add_return(1, &md->uevent_seq);
2721}
2722
2723uint32_t dm_get_event_nr(struct mapped_device *md)
2724{
2725 return atomic_read(&md->event_nr);
2726}
2727
2728int dm_wait_event(struct mapped_device *md, int event_nr)
2729{
2730 return wait_event_interruptible(md->eventq,
2731 (event_nr != atomic_read(&md->event_nr)));
2732}
2733
2734void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
2735{
2736 unsigned long flags;
2737
2738 spin_lock_irqsave(&md->uevent_lock, flags);
2739 list_add(elist, &md->uevent_list);
2740 spin_unlock_irqrestore(&md->uevent_lock, flags);
2741}
2742
2743/*
2744 * The gendisk is only valid as long as you have a reference
2745 * count on 'md'.
2746 */
2747struct gendisk *dm_disk(struct mapped_device *md)
2748{
2749 return md->disk;
2750}
2751EXPORT_SYMBOL_GPL(dm_disk);
2752
2753struct kobject *dm_kobject(struct mapped_device *md)
2754{
2755 return &md->kobj_holder.kobj;
2756}
2757
2758struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
2759{
2760 struct mapped_device *md;
2761
2762 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
2763
2764 spin_lock(&_minor_lock);
2765 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2766 md = NULL;
2767 goto out;
2768 }
2769 dm_get(md);
2770out:
2771 spin_unlock(&_minor_lock);
2772
2773 return md;
2774}
2775
2776int dm_suspended_md(struct mapped_device *md)
2777{
2778 return test_bit(DMF_SUSPENDED, &md->flags);
2779}
2780
2781static int dm_post_suspending_md(struct mapped_device *md)
2782{
2783 return test_bit(DMF_POST_SUSPENDING, &md->flags);
2784}
2785
2786int dm_suspended_internally_md(struct mapped_device *md)
2787{
2788 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2789}
2790
2791int dm_test_deferred_remove_flag(struct mapped_device *md)
2792{
2793 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
2794}
2795
2796int dm_suspended(struct dm_target *ti)
2797{
2798 return dm_suspended_md(ti->table->md);
2799}
2800EXPORT_SYMBOL_GPL(dm_suspended);
2801
2802int dm_post_suspending(struct dm_target *ti)
2803{
2804 return dm_post_suspending_md(ti->table->md);
2805}
2806EXPORT_SYMBOL_GPL(dm_post_suspending);
2807
2808int dm_noflush_suspending(struct dm_target *ti)
2809{
2810 return __noflush_suspending(ti->table->md);
2811}
2812EXPORT_SYMBOL_GPL(dm_noflush_suspending);
2813
2814struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, enum dm_queue_mode type,
2815 unsigned integrity, unsigned per_io_data_size,
2816 unsigned min_pool_size)
2817{
2818 struct dm_md_mempools *pools = kzalloc_node(sizeof(*pools), GFP_KERNEL, md->numa_node_id);
2819 unsigned int pool_size = 0;
2820 unsigned int front_pad, io_front_pad;
2821 int ret;
2822
2823 if (!pools)
2824 return NULL;
2825
2826 switch (type) {
2827 case DM_TYPE_BIO_BASED:
2828 case DM_TYPE_DAX_BIO_BASED:
2829 pool_size = max(dm_get_reserved_bio_based_ios(), min_pool_size);
2830 front_pad = roundup(per_io_data_size, __alignof__(struct dm_target_io)) + DM_TARGET_IO_BIO_OFFSET;
2831 io_front_pad = roundup(per_io_data_size, __alignof__(struct dm_io)) + DM_IO_BIO_OFFSET;
2832 ret = bioset_init(&pools->io_bs, pool_size, io_front_pad, 0);
2833 if (ret)
2834 goto out;
2835 if (integrity && bioset_integrity_create(&pools->io_bs, pool_size))
2836 goto out;
2837 break;
2838 case DM_TYPE_REQUEST_BASED:
2839 pool_size = max(dm_get_reserved_rq_based_ios(), min_pool_size);
2840 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
2841 /* per_io_data_size is used for blk-mq pdu at queue allocation */
2842 break;
2843 default:
2844 BUG();
2845 }
2846
2847 ret = bioset_init(&pools->bs, pool_size, front_pad, 0);
2848 if (ret)
2849 goto out;
2850
2851 if (integrity && bioset_integrity_create(&pools->bs, pool_size))
2852 goto out;
2853
2854 return pools;
2855
2856out:
2857 dm_free_md_mempools(pools);
2858
2859 return NULL;
2860}
2861
2862void dm_free_md_mempools(struct dm_md_mempools *pools)
2863{
2864 if (!pools)
2865 return;
2866
2867 bioset_exit(&pools->bs);
2868 bioset_exit(&pools->io_bs);
2869
2870 kfree(pools);
2871}
2872
2873struct dm_pr {
2874 u64 old_key;
2875 u64 new_key;
2876 u32 flags;
2877 bool fail_early;
2878};
2879
2880static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
2881 void *data)
2882{
2883 struct mapped_device *md = bdev->bd_disk->private_data;
2884 struct dm_table *table;
2885 struct dm_target *ti;
2886 int ret = -ENOTTY, srcu_idx;
2887
2888 table = dm_get_live_table(md, &srcu_idx);
2889 if (!table || !dm_table_get_size(table))
2890 goto out;
2891
2892 /* We only support devices that have a single target */
2893 if (dm_table_get_num_targets(table) != 1)
2894 goto out;
2895 ti = dm_table_get_target(table, 0);
2896
2897 ret = -EINVAL;
2898 if (!ti->type->iterate_devices)
2899 goto out;
2900
2901 ret = ti->type->iterate_devices(ti, fn, data);
2902out:
2903 dm_put_live_table(md, srcu_idx);
2904 return ret;
2905}
2906
2907/*
2908 * For register / unregister we need to manually call out to every path.
2909 */
2910static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
2911 sector_t start, sector_t len, void *data)
2912{
2913 struct dm_pr *pr = data;
2914 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
2915
2916 if (!ops || !ops->pr_register)
2917 return -EOPNOTSUPP;
2918 return ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
2919}
2920
2921static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
2922 u32 flags)
2923{
2924 struct dm_pr pr = {
2925 .old_key = old_key,
2926 .new_key = new_key,
2927 .flags = flags,
2928 .fail_early = true,
2929 };
2930 int ret;
2931
2932 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
2933 if (ret && new_key) {
2934 /* unregister all paths if we failed to register any path */
2935 pr.old_key = new_key;
2936 pr.new_key = 0;
2937 pr.flags = 0;
2938 pr.fail_early = false;
2939 dm_call_pr(bdev, __dm_pr_register, &pr);
2940 }
2941
2942 return ret;
2943}
2944
2945static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
2946 u32 flags)
2947{
2948 struct mapped_device *md = bdev->bd_disk->private_data;
2949 const struct pr_ops *ops;
2950 int r, srcu_idx;
2951
2952 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
2953 if (r < 0)
2954 goto out;
2955
2956 ops = bdev->bd_disk->fops->pr_ops;
2957 if (ops && ops->pr_reserve)
2958 r = ops->pr_reserve(bdev, key, type, flags);
2959 else
2960 r = -EOPNOTSUPP;
2961out:
2962 dm_unprepare_ioctl(md, srcu_idx);
2963 return r;
2964}
2965
2966static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
2967{
2968 struct mapped_device *md = bdev->bd_disk->private_data;
2969 const struct pr_ops *ops;
2970 int r, srcu_idx;
2971
2972 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
2973 if (r < 0)
2974 goto out;
2975
2976 ops = bdev->bd_disk->fops->pr_ops;
2977 if (ops && ops->pr_release)
2978 r = ops->pr_release(bdev, key, type);
2979 else
2980 r = -EOPNOTSUPP;
2981out:
2982 dm_unprepare_ioctl(md, srcu_idx);
2983 return r;
2984}
2985
2986static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
2987 enum pr_type type, bool abort)
2988{
2989 struct mapped_device *md = bdev->bd_disk->private_data;
2990 const struct pr_ops *ops;
2991 int r, srcu_idx;
2992
2993 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
2994 if (r < 0)
2995 goto out;
2996
2997 ops = bdev->bd_disk->fops->pr_ops;
2998 if (ops && ops->pr_preempt)
2999 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
3000 else
3001 r = -EOPNOTSUPP;
3002out:
3003 dm_unprepare_ioctl(md, srcu_idx);
3004 return r;
3005}
3006
3007static int dm_pr_clear(struct block_device *bdev, u64 key)
3008{
3009 struct mapped_device *md = bdev->bd_disk->private_data;
3010 const struct pr_ops *ops;
3011 int r, srcu_idx;
3012
3013 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3014 if (r < 0)
3015 goto out;
3016
3017 ops = bdev->bd_disk->fops->pr_ops;
3018 if (ops && ops->pr_clear)
3019 r = ops->pr_clear(bdev, key);
3020 else
3021 r = -EOPNOTSUPP;
3022out:
3023 dm_unprepare_ioctl(md, srcu_idx);
3024 return r;
3025}
3026
3027static const struct pr_ops dm_pr_ops = {
3028 .pr_register = dm_pr_register,
3029 .pr_reserve = dm_pr_reserve,
3030 .pr_release = dm_pr_release,
3031 .pr_preempt = dm_pr_preempt,
3032 .pr_clear = dm_pr_clear,
3033};
3034
3035static const struct block_device_operations dm_blk_dops = {
3036 .submit_bio = dm_submit_bio,
3037 .open = dm_blk_open,
3038 .release = dm_blk_close,
3039 .ioctl = dm_blk_ioctl,
3040 .getgeo = dm_blk_getgeo,
3041 .report_zones = dm_blk_report_zones,
3042 .pr_ops = &dm_pr_ops,
3043 .owner = THIS_MODULE
3044};
3045
3046static const struct block_device_operations dm_rq_blk_dops = {
3047 .open = dm_blk_open,
3048 .release = dm_blk_close,
3049 .ioctl = dm_blk_ioctl,
3050 .getgeo = dm_blk_getgeo,
3051 .pr_ops = &dm_pr_ops,
3052 .owner = THIS_MODULE
3053};
3054
3055static const struct dax_operations dm_dax_ops = {
3056 .direct_access = dm_dax_direct_access,
3057 .dax_supported = dm_dax_supported,
3058 .copy_from_iter = dm_dax_copy_from_iter,
3059 .copy_to_iter = dm_dax_copy_to_iter,
3060 .zero_page_range = dm_dax_zero_page_range,
3061};
3062
3063/*
3064 * module hooks
3065 */
3066module_init(dm_init);
3067module_exit(dm_exit);
3068
3069module_param(major, uint, 0);
3070MODULE_PARM_DESC(major, "The major number of the device mapper");
3071
3072module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3073MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3074
3075module_param(dm_numa_node, int, S_IRUGO | S_IWUSR);
3076MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3077
3078module_param(swap_bios, int, S_IRUGO | S_IWUSR);
3079MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3080
3081MODULE_DESCRIPTION(DM_NAME " driver");
3082MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3083MODULE_LICENSE("GPL");