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/bio-integrity.h>
15#include <linux/init.h>
16#include <linux/module.h>
17#include <linux/mutex.h>
18#include <linux/sched/mm.h>
19#include <linux/sched/signal.h>
20#include <linux/blkpg.h>
21#include <linux/bio.h>
22#include <linux/mempool.h>
23#include <linux/dax.h>
24#include <linux/slab.h>
25#include <linux/idr.h>
26#include <linux/uio.h>
27#include <linux/hdreg.h>
28#include <linux/delay.h>
29#include <linux/wait.h>
30#include <linux/pr.h>
31#include <linux/refcount.h>
32#include <linux/part_stat.h>
33#include <linux/blk-crypto.h>
34#include <linux/blk-crypto-profile.h>
35
36#define DM_MSG_PREFIX "core"
37
38/*
39 * Cookies are numeric values sent with CHANGE and REMOVE
40 * uevents while resuming, removing or renaming the device.
41 */
42#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
43#define DM_COOKIE_LENGTH 24
44
45/*
46 * For REQ_POLLED fs bio, this flag is set if we link mapped underlying
47 * dm_io into one list, and reuse bio->bi_private as the list head. Before
48 * ending this fs bio, we will recover its ->bi_private.
49 */
50#define REQ_DM_POLL_LIST REQ_DRV
51
52static const char *_name = DM_NAME;
53
54static unsigned int major;
55static unsigned int _major;
56
57static DEFINE_IDR(_minor_idr);
58
59static DEFINE_SPINLOCK(_minor_lock);
60
61static void do_deferred_remove(struct work_struct *w);
62
63static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
64
65static struct workqueue_struct *deferred_remove_workqueue;
66
67atomic_t dm_global_event_nr = ATOMIC_INIT(0);
68DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
69
70void dm_issue_global_event(void)
71{
72 atomic_inc(&dm_global_event_nr);
73 wake_up(&dm_global_eventq);
74}
75
76DEFINE_STATIC_KEY_FALSE(stats_enabled);
77DEFINE_STATIC_KEY_FALSE(swap_bios_enabled);
78DEFINE_STATIC_KEY_FALSE(zoned_enabled);
79
80/*
81 * One of these is allocated (on-stack) per original bio.
82 */
83struct clone_info {
84 struct dm_table *map;
85 struct bio *bio;
86 struct dm_io *io;
87 sector_t sector;
88 unsigned int sector_count;
89 bool is_abnormal_io:1;
90 bool submit_as_polled:1;
91};
92
93static inline struct dm_target_io *clone_to_tio(struct bio *clone)
94{
95 return container_of(clone, struct dm_target_io, clone);
96}
97
98void *dm_per_bio_data(struct bio *bio, size_t data_size)
99{
100 if (!dm_tio_flagged(clone_to_tio(bio), DM_TIO_INSIDE_DM_IO))
101 return (char *)bio - DM_TARGET_IO_BIO_OFFSET - data_size;
102 return (char *)bio - DM_IO_BIO_OFFSET - data_size;
103}
104EXPORT_SYMBOL_GPL(dm_per_bio_data);
105
106struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
107{
108 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
109
110 if (io->magic == DM_IO_MAGIC)
111 return (struct bio *)((char *)io + DM_IO_BIO_OFFSET);
112 BUG_ON(io->magic != DM_TIO_MAGIC);
113 return (struct bio *)((char *)io + DM_TARGET_IO_BIO_OFFSET);
114}
115EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
116
117unsigned int dm_bio_get_target_bio_nr(const struct bio *bio)
118{
119 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
120}
121EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
122
123#define MINOR_ALLOCED ((void *)-1)
124
125#define DM_NUMA_NODE NUMA_NO_NODE
126static int dm_numa_node = DM_NUMA_NODE;
127
128#define DEFAULT_SWAP_BIOS (8 * 1048576 / PAGE_SIZE)
129static int swap_bios = DEFAULT_SWAP_BIOS;
130static int get_swap_bios(void)
131{
132 int latch = READ_ONCE(swap_bios);
133
134 if (unlikely(latch <= 0))
135 latch = DEFAULT_SWAP_BIOS;
136 return latch;
137}
138
139struct table_device {
140 struct list_head list;
141 refcount_t count;
142 struct dm_dev dm_dev;
143};
144
145/*
146 * Bio-based DM's mempools' reserved IOs set by the user.
147 */
148#define RESERVED_BIO_BASED_IOS 16
149static unsigned int reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
150
151static int __dm_get_module_param_int(int *module_param, int min, int max)
152{
153 int param = READ_ONCE(*module_param);
154 int modified_param = 0;
155 bool modified = true;
156
157 if (param < min)
158 modified_param = min;
159 else if (param > max)
160 modified_param = max;
161 else
162 modified = false;
163
164 if (modified) {
165 (void)cmpxchg(module_param, param, modified_param);
166 param = modified_param;
167 }
168
169 return param;
170}
171
172unsigned int __dm_get_module_param(unsigned int *module_param, unsigned int def, unsigned int max)
173{
174 unsigned int param = READ_ONCE(*module_param);
175 unsigned int modified_param = 0;
176
177 if (!param)
178 modified_param = def;
179 else if (param > max)
180 modified_param = max;
181
182 if (modified_param) {
183 (void)cmpxchg(module_param, param, modified_param);
184 param = modified_param;
185 }
186
187 return param;
188}
189
190unsigned int dm_get_reserved_bio_based_ios(void)
191{
192 return __dm_get_module_param(&reserved_bio_based_ios,
193 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
194}
195EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
196
197static unsigned int dm_get_numa_node(void)
198{
199 return __dm_get_module_param_int(&dm_numa_node,
200 DM_NUMA_NODE, num_online_nodes() - 1);
201}
202
203static int __init local_init(void)
204{
205 int r;
206
207 r = dm_uevent_init();
208 if (r)
209 return r;
210
211 deferred_remove_workqueue = alloc_ordered_workqueue("kdmremove", 0);
212 if (!deferred_remove_workqueue) {
213 r = -ENOMEM;
214 goto out_uevent_exit;
215 }
216
217 _major = major;
218 r = register_blkdev(_major, _name);
219 if (r < 0)
220 goto out_free_workqueue;
221
222 if (!_major)
223 _major = r;
224
225 return 0;
226
227out_free_workqueue:
228 destroy_workqueue(deferred_remove_workqueue);
229out_uevent_exit:
230 dm_uevent_exit();
231
232 return r;
233}
234
235static void local_exit(void)
236{
237 destroy_workqueue(deferred_remove_workqueue);
238
239 unregister_blkdev(_major, _name);
240 dm_uevent_exit();
241
242 _major = 0;
243
244 DMINFO("cleaned up");
245}
246
247static int (*_inits[])(void) __initdata = {
248 local_init,
249 dm_target_init,
250 dm_linear_init,
251 dm_stripe_init,
252 dm_io_init,
253 dm_kcopyd_init,
254 dm_interface_init,
255 dm_statistics_init,
256};
257
258static void (*_exits[])(void) = {
259 local_exit,
260 dm_target_exit,
261 dm_linear_exit,
262 dm_stripe_exit,
263 dm_io_exit,
264 dm_kcopyd_exit,
265 dm_interface_exit,
266 dm_statistics_exit,
267};
268
269static int __init dm_init(void)
270{
271 const int count = ARRAY_SIZE(_inits);
272 int r, i;
273
274#if (IS_ENABLED(CONFIG_IMA) && !IS_ENABLED(CONFIG_IMA_DISABLE_HTABLE))
275 DMWARN("CONFIG_IMA_DISABLE_HTABLE is disabled."
276 " Duplicate IMA measurements will not be recorded in the IMA log.");
277#endif
278
279 for (i = 0; i < count; i++) {
280 r = _inits[i]();
281 if (r)
282 goto bad;
283 }
284
285 return 0;
286bad:
287 while (i--)
288 _exits[i]();
289
290 return r;
291}
292
293static void __exit dm_exit(void)
294{
295 int i = ARRAY_SIZE(_exits);
296
297 while (i--)
298 _exits[i]();
299
300 /*
301 * Should be empty by this point.
302 */
303 idr_destroy(&_minor_idr);
304}
305
306/*
307 * Block device functions
308 */
309int dm_deleting_md(struct mapped_device *md)
310{
311 return test_bit(DMF_DELETING, &md->flags);
312}
313
314static int dm_blk_open(struct gendisk *disk, blk_mode_t mode)
315{
316 struct mapped_device *md;
317
318 spin_lock(&_minor_lock);
319
320 md = disk->private_data;
321 if (!md)
322 goto out;
323
324 if (test_bit(DMF_FREEING, &md->flags) ||
325 dm_deleting_md(md)) {
326 md = NULL;
327 goto out;
328 }
329
330 dm_get(md);
331 atomic_inc(&md->open_count);
332out:
333 spin_unlock(&_minor_lock);
334
335 return md ? 0 : -ENXIO;
336}
337
338static void dm_blk_close(struct gendisk *disk)
339{
340 struct mapped_device *md;
341
342 spin_lock(&_minor_lock);
343
344 md = disk->private_data;
345 if (WARN_ON(!md))
346 goto out;
347
348 if (atomic_dec_and_test(&md->open_count) &&
349 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
350 queue_work(deferred_remove_workqueue, &deferred_remove_work);
351
352 dm_put(md);
353out:
354 spin_unlock(&_minor_lock);
355}
356
357int dm_open_count(struct mapped_device *md)
358{
359 return atomic_read(&md->open_count);
360}
361
362/*
363 * Guarantees nothing is using the device before it's deleted.
364 */
365int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
366{
367 int r = 0;
368
369 spin_lock(&_minor_lock);
370
371 if (dm_open_count(md)) {
372 r = -EBUSY;
373 if (mark_deferred)
374 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
375 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
376 r = -EEXIST;
377 else
378 set_bit(DMF_DELETING, &md->flags);
379
380 spin_unlock(&_minor_lock);
381
382 return r;
383}
384
385int dm_cancel_deferred_remove(struct mapped_device *md)
386{
387 int r = 0;
388
389 spin_lock(&_minor_lock);
390
391 if (test_bit(DMF_DELETING, &md->flags))
392 r = -EBUSY;
393 else
394 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
395
396 spin_unlock(&_minor_lock);
397
398 return r;
399}
400
401static void do_deferred_remove(struct work_struct *w)
402{
403 dm_deferred_remove();
404}
405
406static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
407{
408 struct mapped_device *md = bdev->bd_disk->private_data;
409
410 return dm_get_geometry(md, geo);
411}
412
413static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
414 struct block_device **bdev)
415{
416 struct dm_target *ti;
417 struct dm_table *map;
418 int r;
419
420retry:
421 r = -ENOTTY;
422 map = dm_get_live_table(md, srcu_idx);
423 if (!map || !dm_table_get_size(map))
424 return r;
425
426 /* We only support devices that have a single target */
427 if (map->num_targets != 1)
428 return r;
429
430 ti = dm_table_get_target(map, 0);
431 if (!ti->type->prepare_ioctl)
432 return r;
433
434 if (dm_suspended_md(md))
435 return -EAGAIN;
436
437 r = ti->type->prepare_ioctl(ti, bdev);
438 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
439 dm_put_live_table(md, *srcu_idx);
440 fsleep(10000);
441 goto retry;
442 }
443
444 return r;
445}
446
447static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
448{
449 dm_put_live_table(md, srcu_idx);
450}
451
452static int dm_blk_ioctl(struct block_device *bdev, blk_mode_t mode,
453 unsigned int cmd, unsigned long arg)
454{
455 struct mapped_device *md = bdev->bd_disk->private_data;
456 int r, srcu_idx;
457
458 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
459 if (r < 0)
460 goto out;
461
462 if (r > 0) {
463 /*
464 * Target determined this ioctl is being issued against a
465 * subset of the parent bdev; require extra privileges.
466 */
467 if (!capable(CAP_SYS_RAWIO)) {
468 DMDEBUG_LIMIT(
469 "%s: sending ioctl %x to DM device without required privilege.",
470 current->comm, cmd);
471 r = -ENOIOCTLCMD;
472 goto out;
473 }
474 }
475
476 if (!bdev->bd_disk->fops->ioctl)
477 r = -ENOTTY;
478 else
479 r = bdev->bd_disk->fops->ioctl(bdev, mode, cmd, arg);
480out:
481 dm_unprepare_ioctl(md, srcu_idx);
482 return r;
483}
484
485u64 dm_start_time_ns_from_clone(struct bio *bio)
486{
487 return jiffies_to_nsecs(clone_to_tio(bio)->io->start_time);
488}
489EXPORT_SYMBOL_GPL(dm_start_time_ns_from_clone);
490
491static inline bool bio_is_flush_with_data(struct bio *bio)
492{
493 return ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size);
494}
495
496static inline unsigned int dm_io_sectors(struct dm_io *io, struct bio *bio)
497{
498 /*
499 * If REQ_PREFLUSH set, don't account payload, it will be
500 * submitted (and accounted) after this flush completes.
501 */
502 if (bio_is_flush_with_data(bio))
503 return 0;
504 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
505 return io->sectors;
506 return bio_sectors(bio);
507}
508
509static void dm_io_acct(struct dm_io *io, bool end)
510{
511 struct bio *bio = io->orig_bio;
512
513 if (dm_io_flagged(io, DM_IO_BLK_STAT)) {
514 if (!end)
515 bdev_start_io_acct(bio->bi_bdev, bio_op(bio),
516 io->start_time);
517 else
518 bdev_end_io_acct(bio->bi_bdev, bio_op(bio),
519 dm_io_sectors(io, bio),
520 io->start_time);
521 }
522
523 if (static_branch_unlikely(&stats_enabled) &&
524 unlikely(dm_stats_used(&io->md->stats))) {
525 sector_t sector;
526
527 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
528 sector = bio_end_sector(bio) - io->sector_offset;
529 else
530 sector = bio->bi_iter.bi_sector;
531
532 dm_stats_account_io(&io->md->stats, bio_data_dir(bio),
533 sector, dm_io_sectors(io, bio),
534 end, io->start_time, &io->stats_aux);
535 }
536}
537
538static void __dm_start_io_acct(struct dm_io *io)
539{
540 dm_io_acct(io, false);
541}
542
543static void dm_start_io_acct(struct dm_io *io, struct bio *clone)
544{
545 /*
546 * Ensure IO accounting is only ever started once.
547 */
548 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
549 return;
550
551 /* Expect no possibility for race unless DM_TIO_IS_DUPLICATE_BIO. */
552 if (!clone || likely(dm_tio_is_normal(clone_to_tio(clone)))) {
553 dm_io_set_flag(io, DM_IO_ACCOUNTED);
554 } else {
555 unsigned long flags;
556 /* Can afford locking given DM_TIO_IS_DUPLICATE_BIO */
557 spin_lock_irqsave(&io->lock, flags);
558 if (dm_io_flagged(io, DM_IO_ACCOUNTED)) {
559 spin_unlock_irqrestore(&io->lock, flags);
560 return;
561 }
562 dm_io_set_flag(io, DM_IO_ACCOUNTED);
563 spin_unlock_irqrestore(&io->lock, flags);
564 }
565
566 __dm_start_io_acct(io);
567}
568
569static void dm_end_io_acct(struct dm_io *io)
570{
571 dm_io_acct(io, true);
572}
573
574static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio, gfp_t gfp_mask)
575{
576 struct dm_io *io;
577 struct dm_target_io *tio;
578 struct bio *clone;
579
580 clone = bio_alloc_clone(NULL, bio, gfp_mask, &md->mempools->io_bs);
581 if (unlikely(!clone))
582 return NULL;
583 tio = clone_to_tio(clone);
584 tio->flags = 0;
585 dm_tio_set_flag(tio, DM_TIO_INSIDE_DM_IO);
586 tio->io = NULL;
587
588 io = container_of(tio, struct dm_io, tio);
589 io->magic = DM_IO_MAGIC;
590 io->status = BLK_STS_OK;
591
592 /* one ref is for submission, the other is for completion */
593 atomic_set(&io->io_count, 2);
594 this_cpu_inc(*md->pending_io);
595 io->orig_bio = bio;
596 io->md = md;
597 spin_lock_init(&io->lock);
598 io->start_time = jiffies;
599 io->flags = 0;
600 if (blk_queue_io_stat(md->queue))
601 dm_io_set_flag(io, DM_IO_BLK_STAT);
602
603 if (static_branch_unlikely(&stats_enabled) &&
604 unlikely(dm_stats_used(&md->stats)))
605 dm_stats_record_start(&md->stats, &io->stats_aux);
606
607 return io;
608}
609
610static void free_io(struct dm_io *io)
611{
612 bio_put(&io->tio.clone);
613}
614
615static struct bio *alloc_tio(struct clone_info *ci, struct dm_target *ti,
616 unsigned int target_bio_nr, unsigned int *len, gfp_t gfp_mask)
617{
618 struct mapped_device *md = ci->io->md;
619 struct dm_target_io *tio;
620 struct bio *clone;
621
622 if (!ci->io->tio.io) {
623 /* the dm_target_io embedded in ci->io is available */
624 tio = &ci->io->tio;
625 /* alloc_io() already initialized embedded clone */
626 clone = &tio->clone;
627 } else {
628 clone = bio_alloc_clone(NULL, ci->bio, gfp_mask,
629 &md->mempools->bs);
630 if (!clone)
631 return NULL;
632
633 /* REQ_DM_POLL_LIST shouldn't be inherited */
634 clone->bi_opf &= ~REQ_DM_POLL_LIST;
635
636 tio = clone_to_tio(clone);
637 tio->flags = 0; /* also clears DM_TIO_INSIDE_DM_IO */
638 }
639
640 tio->magic = DM_TIO_MAGIC;
641 tio->io = ci->io;
642 tio->ti = ti;
643 tio->target_bio_nr = target_bio_nr;
644 tio->len_ptr = len;
645 tio->old_sector = 0;
646
647 /* Set default bdev, but target must bio_set_dev() before issuing IO */
648 clone->bi_bdev = md->disk->part0;
649 if (likely(ti != NULL) && unlikely(ti->needs_bio_set_dev))
650 bio_set_dev(clone, md->disk->part0);
651
652 if (len) {
653 clone->bi_iter.bi_size = to_bytes(*len);
654 if (bio_integrity(clone))
655 bio_integrity_trim(clone);
656 }
657
658 return clone;
659}
660
661static void free_tio(struct bio *clone)
662{
663 if (dm_tio_flagged(clone_to_tio(clone), DM_TIO_INSIDE_DM_IO))
664 return;
665 bio_put(clone);
666}
667
668/*
669 * Add the bio to the list of deferred io.
670 */
671static void queue_io(struct mapped_device *md, struct bio *bio)
672{
673 unsigned long flags;
674
675 spin_lock_irqsave(&md->deferred_lock, flags);
676 bio_list_add(&md->deferred, bio);
677 spin_unlock_irqrestore(&md->deferred_lock, flags);
678 queue_work(md->wq, &md->work);
679}
680
681/*
682 * Everyone (including functions in this file), should use this
683 * function to access the md->map field, and make sure they call
684 * dm_put_live_table() when finished.
685 */
686struct dm_table *dm_get_live_table(struct mapped_device *md,
687 int *srcu_idx) __acquires(md->io_barrier)
688{
689 *srcu_idx = srcu_read_lock(&md->io_barrier);
690
691 return srcu_dereference(md->map, &md->io_barrier);
692}
693
694void dm_put_live_table(struct mapped_device *md,
695 int srcu_idx) __releases(md->io_barrier)
696{
697 srcu_read_unlock(&md->io_barrier, srcu_idx);
698}
699
700void dm_sync_table(struct mapped_device *md)
701{
702 synchronize_srcu(&md->io_barrier);
703 synchronize_rcu_expedited();
704}
705
706/*
707 * A fast alternative to dm_get_live_table/dm_put_live_table.
708 * The caller must not block between these two functions.
709 */
710static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
711{
712 rcu_read_lock();
713 return rcu_dereference(md->map);
714}
715
716static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
717{
718 rcu_read_unlock();
719}
720
721static char *_dm_claim_ptr = "I belong to device-mapper";
722
723/*
724 * Open a table device so we can use it as a map destination.
725 */
726static struct table_device *open_table_device(struct mapped_device *md,
727 dev_t dev, blk_mode_t mode)
728{
729 struct table_device *td;
730 struct file *bdev_file;
731 struct block_device *bdev;
732 u64 part_off;
733 int r;
734
735 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
736 if (!td)
737 return ERR_PTR(-ENOMEM);
738 refcount_set(&td->count, 1);
739
740 bdev_file = bdev_file_open_by_dev(dev, mode, _dm_claim_ptr, NULL);
741 if (IS_ERR(bdev_file)) {
742 r = PTR_ERR(bdev_file);
743 goto out_free_td;
744 }
745
746 bdev = file_bdev(bdev_file);
747
748 /*
749 * We can be called before the dm disk is added. In that case we can't
750 * register the holder relation here. It will be done once add_disk was
751 * called.
752 */
753 if (md->disk->slave_dir) {
754 r = bd_link_disk_holder(bdev, md->disk);
755 if (r)
756 goto out_blkdev_put;
757 }
758
759 td->dm_dev.mode = mode;
760 td->dm_dev.bdev = bdev;
761 td->dm_dev.bdev_file = bdev_file;
762 td->dm_dev.dax_dev = fs_dax_get_by_bdev(bdev, &part_off,
763 NULL, NULL);
764 format_dev_t(td->dm_dev.name, dev);
765 list_add(&td->list, &md->table_devices);
766 return td;
767
768out_blkdev_put:
769 __fput_sync(bdev_file);
770out_free_td:
771 kfree(td);
772 return ERR_PTR(r);
773}
774
775/*
776 * Close a table device that we've been using.
777 */
778static void close_table_device(struct table_device *td, struct mapped_device *md)
779{
780 if (md->disk->slave_dir)
781 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
782
783 /* Leverage async fput() if DMF_DEFERRED_REMOVE set */
784 if (unlikely(test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
785 fput(td->dm_dev.bdev_file);
786 else
787 __fput_sync(td->dm_dev.bdev_file);
788
789 put_dax(td->dm_dev.dax_dev);
790 list_del(&td->list);
791 kfree(td);
792}
793
794static struct table_device *find_table_device(struct list_head *l, dev_t dev,
795 blk_mode_t mode)
796{
797 struct table_device *td;
798
799 list_for_each_entry(td, l, list)
800 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
801 return td;
802
803 return NULL;
804}
805
806int dm_get_table_device(struct mapped_device *md, dev_t dev, blk_mode_t mode,
807 struct dm_dev **result)
808{
809 struct table_device *td;
810
811 mutex_lock(&md->table_devices_lock);
812 td = find_table_device(&md->table_devices, dev, mode);
813 if (!td) {
814 td = open_table_device(md, dev, mode);
815 if (IS_ERR(td)) {
816 mutex_unlock(&md->table_devices_lock);
817 return PTR_ERR(td);
818 }
819 } else {
820 refcount_inc(&td->count);
821 }
822 mutex_unlock(&md->table_devices_lock);
823
824 *result = &td->dm_dev;
825 return 0;
826}
827
828void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
829{
830 struct table_device *td = container_of(d, struct table_device, dm_dev);
831
832 mutex_lock(&md->table_devices_lock);
833 if (refcount_dec_and_test(&td->count))
834 close_table_device(td, md);
835 mutex_unlock(&md->table_devices_lock);
836}
837
838/*
839 * Get the geometry associated with a dm device
840 */
841int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
842{
843 *geo = md->geometry;
844
845 return 0;
846}
847
848/*
849 * Set the geometry of a device.
850 */
851int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
852{
853 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
854
855 if (geo->start > sz) {
856 DMERR("Start sector is beyond the geometry limits.");
857 return -EINVAL;
858 }
859
860 md->geometry = *geo;
861
862 return 0;
863}
864
865static int __noflush_suspending(struct mapped_device *md)
866{
867 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
868}
869
870static void dm_requeue_add_io(struct dm_io *io, bool first_stage)
871{
872 struct mapped_device *md = io->md;
873
874 if (first_stage) {
875 struct dm_io *next = md->requeue_list;
876
877 md->requeue_list = io;
878 io->next = next;
879 } else {
880 bio_list_add_head(&md->deferred, io->orig_bio);
881 }
882}
883
884static void dm_kick_requeue(struct mapped_device *md, bool first_stage)
885{
886 if (first_stage)
887 queue_work(md->wq, &md->requeue_work);
888 else
889 queue_work(md->wq, &md->work);
890}
891
892/*
893 * Return true if the dm_io's original bio is requeued.
894 * io->status is updated with error if requeue disallowed.
895 */
896static bool dm_handle_requeue(struct dm_io *io, bool first_stage)
897{
898 struct bio *bio = io->orig_bio;
899 bool handle_requeue = (io->status == BLK_STS_DM_REQUEUE);
900 bool handle_polled_eagain = ((io->status == BLK_STS_AGAIN) &&
901 (bio->bi_opf & REQ_POLLED));
902 struct mapped_device *md = io->md;
903 bool requeued = false;
904
905 if (handle_requeue || handle_polled_eagain) {
906 unsigned long flags;
907
908 if (bio->bi_opf & REQ_POLLED) {
909 /*
910 * Upper layer won't help us poll split bio
911 * (io->orig_bio may only reflect a subset of the
912 * pre-split original) so clear REQ_POLLED.
913 */
914 bio_clear_polled(bio);
915 }
916
917 /*
918 * Target requested pushing back the I/O or
919 * polled IO hit BLK_STS_AGAIN.
920 */
921 spin_lock_irqsave(&md->deferred_lock, flags);
922 if ((__noflush_suspending(md) &&
923 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) ||
924 handle_polled_eagain || first_stage) {
925 dm_requeue_add_io(io, first_stage);
926 requeued = true;
927 } else {
928 /*
929 * noflush suspend was interrupted or this is
930 * a write to a zoned target.
931 */
932 io->status = BLK_STS_IOERR;
933 }
934 spin_unlock_irqrestore(&md->deferred_lock, flags);
935 }
936
937 if (requeued)
938 dm_kick_requeue(md, first_stage);
939
940 return requeued;
941}
942
943static void __dm_io_complete(struct dm_io *io, bool first_stage)
944{
945 struct bio *bio = io->orig_bio;
946 struct mapped_device *md = io->md;
947 blk_status_t io_error;
948 bool requeued;
949
950 requeued = dm_handle_requeue(io, first_stage);
951 if (requeued && first_stage)
952 return;
953
954 io_error = io->status;
955 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
956 dm_end_io_acct(io);
957 else if (!io_error) {
958 /*
959 * Must handle target that DM_MAPIO_SUBMITTED only to
960 * then bio_endio() rather than dm_submit_bio_remap()
961 */
962 __dm_start_io_acct(io);
963 dm_end_io_acct(io);
964 }
965 free_io(io);
966 smp_wmb();
967 this_cpu_dec(*md->pending_io);
968
969 /* nudge anyone waiting on suspend queue */
970 if (unlikely(wq_has_sleeper(&md->wait)))
971 wake_up(&md->wait);
972
973 /* Return early if the original bio was requeued */
974 if (requeued)
975 return;
976
977 if (bio_is_flush_with_data(bio)) {
978 /*
979 * Preflush done for flush with data, reissue
980 * without REQ_PREFLUSH.
981 */
982 bio->bi_opf &= ~REQ_PREFLUSH;
983 queue_io(md, bio);
984 } else {
985 /* done with normal IO or empty flush */
986 if (io_error)
987 bio->bi_status = io_error;
988 bio_endio(bio);
989 }
990}
991
992static void dm_wq_requeue_work(struct work_struct *work)
993{
994 struct mapped_device *md = container_of(work, struct mapped_device,
995 requeue_work);
996 unsigned long flags;
997 struct dm_io *io;
998
999 /* reuse deferred lock to simplify dm_handle_requeue */
1000 spin_lock_irqsave(&md->deferred_lock, flags);
1001 io = md->requeue_list;
1002 md->requeue_list = NULL;
1003 spin_unlock_irqrestore(&md->deferred_lock, flags);
1004
1005 while (io) {
1006 struct dm_io *next = io->next;
1007
1008 dm_io_rewind(io, &md->disk->bio_split);
1009
1010 io->next = NULL;
1011 __dm_io_complete(io, false);
1012 io = next;
1013 cond_resched();
1014 }
1015}
1016
1017/*
1018 * Two staged requeue:
1019 *
1020 * 1) io->orig_bio points to the real original bio, and the part mapped to
1021 * this io must be requeued, instead of other parts of the original bio.
1022 *
1023 * 2) io->orig_bio points to new cloned bio which matches the requeued dm_io.
1024 */
1025static void dm_io_complete(struct dm_io *io)
1026{
1027 bool first_requeue;
1028
1029 /*
1030 * Only dm_io that has been split needs two stage requeue, otherwise
1031 * we may run into long bio clone chain during suspend and OOM could
1032 * be triggered.
1033 *
1034 * Also flush data dm_io won't be marked as DM_IO_WAS_SPLIT, so they
1035 * also aren't handled via the first stage requeue.
1036 */
1037 if (dm_io_flagged(io, DM_IO_WAS_SPLIT))
1038 first_requeue = true;
1039 else
1040 first_requeue = false;
1041
1042 __dm_io_complete(io, first_requeue);
1043}
1044
1045/*
1046 * Decrements the number of outstanding ios that a bio has been
1047 * cloned into, completing the original io if necc.
1048 */
1049static inline void __dm_io_dec_pending(struct dm_io *io)
1050{
1051 if (atomic_dec_and_test(&io->io_count))
1052 dm_io_complete(io);
1053}
1054
1055static void dm_io_set_error(struct dm_io *io, blk_status_t error)
1056{
1057 unsigned long flags;
1058
1059 /* Push-back supersedes any I/O errors */
1060 spin_lock_irqsave(&io->lock, flags);
1061 if (!(io->status == BLK_STS_DM_REQUEUE &&
1062 __noflush_suspending(io->md))) {
1063 io->status = error;
1064 }
1065 spin_unlock_irqrestore(&io->lock, flags);
1066}
1067
1068static void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
1069{
1070 if (unlikely(error))
1071 dm_io_set_error(io, error);
1072
1073 __dm_io_dec_pending(io);
1074}
1075
1076/*
1077 * The queue_limits are only valid as long as you have a reference
1078 * count on 'md'. But _not_ imposing verification to avoid atomic_read(),
1079 */
1080static inline struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
1081{
1082 return &md->queue->limits;
1083}
1084
1085void disable_discard(struct mapped_device *md)
1086{
1087 struct queue_limits *limits = dm_get_queue_limits(md);
1088
1089 /* device doesn't really support DISCARD, disable it */
1090 limits->max_hw_discard_sectors = 0;
1091}
1092
1093void disable_write_zeroes(struct mapped_device *md)
1094{
1095 struct queue_limits *limits = dm_get_queue_limits(md);
1096
1097 /* device doesn't really support WRITE ZEROES, disable it */
1098 limits->max_write_zeroes_sectors = 0;
1099}
1100
1101static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
1102{
1103 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
1104}
1105
1106static void clone_endio(struct bio *bio)
1107{
1108 blk_status_t error = bio->bi_status;
1109 struct dm_target_io *tio = clone_to_tio(bio);
1110 struct dm_target *ti = tio->ti;
1111 dm_endio_fn endio = likely(ti != NULL) ? ti->type->end_io : NULL;
1112 struct dm_io *io = tio->io;
1113 struct mapped_device *md = io->md;
1114
1115 if (unlikely(error == BLK_STS_TARGET)) {
1116 if (bio_op(bio) == REQ_OP_DISCARD &&
1117 !bdev_max_discard_sectors(bio->bi_bdev))
1118 disable_discard(md);
1119 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
1120 !bdev_write_zeroes_sectors(bio->bi_bdev))
1121 disable_write_zeroes(md);
1122 }
1123
1124 if (static_branch_unlikely(&zoned_enabled) &&
1125 unlikely(bdev_is_zoned(bio->bi_bdev)))
1126 dm_zone_endio(io, bio);
1127
1128 if (endio) {
1129 int r = endio(ti, bio, &error);
1130
1131 switch (r) {
1132 case DM_ENDIO_REQUEUE:
1133 if (static_branch_unlikely(&zoned_enabled)) {
1134 /*
1135 * Requeuing writes to a sequential zone of a zoned
1136 * target will break the sequential write pattern:
1137 * fail such IO.
1138 */
1139 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
1140 error = BLK_STS_IOERR;
1141 else
1142 error = BLK_STS_DM_REQUEUE;
1143 } else
1144 error = BLK_STS_DM_REQUEUE;
1145 fallthrough;
1146 case DM_ENDIO_DONE:
1147 break;
1148 case DM_ENDIO_INCOMPLETE:
1149 /* The target will handle the io */
1150 return;
1151 default:
1152 DMCRIT("unimplemented target endio return value: %d", r);
1153 BUG();
1154 }
1155 }
1156
1157 if (static_branch_unlikely(&swap_bios_enabled) &&
1158 likely(ti != NULL) && unlikely(swap_bios_limit(ti, bio)))
1159 up(&md->swap_bios_semaphore);
1160
1161 free_tio(bio);
1162 dm_io_dec_pending(io, error);
1163}
1164
1165/*
1166 * Return maximum size of I/O possible at the supplied sector up to the current
1167 * target boundary.
1168 */
1169static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
1170 sector_t target_offset)
1171{
1172 return ti->len - target_offset;
1173}
1174
1175static sector_t __max_io_len(struct dm_target *ti, sector_t sector,
1176 unsigned int max_granularity,
1177 unsigned int max_sectors)
1178{
1179 sector_t target_offset = dm_target_offset(ti, sector);
1180 sector_t len = max_io_len_target_boundary(ti, target_offset);
1181
1182 /*
1183 * Does the target need to split IO even further?
1184 * - varied (per target) IO splitting is a tenet of DM; this
1185 * explains why stacked chunk_sectors based splitting via
1186 * bio_split_to_limits() isn't possible here.
1187 */
1188 if (!max_granularity)
1189 return len;
1190 return min_t(sector_t, len,
1191 min(max_sectors ? : queue_max_sectors(ti->table->md->queue),
1192 blk_boundary_sectors_left(target_offset, max_granularity)));
1193}
1194
1195static inline sector_t max_io_len(struct dm_target *ti, sector_t sector)
1196{
1197 return __max_io_len(ti, sector, ti->max_io_len, 0);
1198}
1199
1200int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1201{
1202 if (len > UINT_MAX) {
1203 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1204 (unsigned long long)len, UINT_MAX);
1205 ti->error = "Maximum size of target IO is too large";
1206 return -EINVAL;
1207 }
1208
1209 ti->max_io_len = (uint32_t) len;
1210
1211 return 0;
1212}
1213EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1214
1215static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1216 sector_t sector, int *srcu_idx)
1217 __acquires(md->io_barrier)
1218{
1219 struct dm_table *map;
1220 struct dm_target *ti;
1221
1222 map = dm_get_live_table(md, srcu_idx);
1223 if (!map)
1224 return NULL;
1225
1226 ti = dm_table_find_target(map, sector);
1227 if (!ti)
1228 return NULL;
1229
1230 return ti;
1231}
1232
1233static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1234 long nr_pages, enum dax_access_mode mode, void **kaddr,
1235 pfn_t *pfn)
1236{
1237 struct mapped_device *md = dax_get_private(dax_dev);
1238 sector_t sector = pgoff * PAGE_SECTORS;
1239 struct dm_target *ti;
1240 long len, ret = -EIO;
1241 int srcu_idx;
1242
1243 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1244
1245 if (!ti)
1246 goto out;
1247 if (!ti->type->direct_access)
1248 goto out;
1249 len = max_io_len(ti, sector) / PAGE_SECTORS;
1250 if (len < 1)
1251 goto out;
1252 nr_pages = min(len, nr_pages);
1253 ret = ti->type->direct_access(ti, pgoff, nr_pages, mode, kaddr, pfn);
1254
1255 out:
1256 dm_put_live_table(md, srcu_idx);
1257
1258 return ret;
1259}
1260
1261static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1262 size_t nr_pages)
1263{
1264 struct mapped_device *md = dax_get_private(dax_dev);
1265 sector_t sector = pgoff * PAGE_SECTORS;
1266 struct dm_target *ti;
1267 int ret = -EIO;
1268 int srcu_idx;
1269
1270 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1271
1272 if (!ti)
1273 goto out;
1274 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1275 /*
1276 * ->zero_page_range() is mandatory dax operation. If we are
1277 * here, something is wrong.
1278 */
1279 goto out;
1280 }
1281 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1282 out:
1283 dm_put_live_table(md, srcu_idx);
1284
1285 return ret;
1286}
1287
1288static size_t dm_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
1289 void *addr, size_t bytes, struct iov_iter *i)
1290{
1291 struct mapped_device *md = dax_get_private(dax_dev);
1292 sector_t sector = pgoff * PAGE_SECTORS;
1293 struct dm_target *ti;
1294 int srcu_idx;
1295 long ret = 0;
1296
1297 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1298 if (!ti || !ti->type->dax_recovery_write)
1299 goto out;
1300
1301 ret = ti->type->dax_recovery_write(ti, pgoff, addr, bytes, i);
1302out:
1303 dm_put_live_table(md, srcu_idx);
1304 return ret;
1305}
1306
1307/*
1308 * A target may call dm_accept_partial_bio only from the map routine. It is
1309 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1310 * operations, REQ_OP_ZONE_APPEND (zone append writes) and any bio serviced by
1311 * __send_duplicate_bios().
1312 *
1313 * dm_accept_partial_bio informs the dm that the target only wants to process
1314 * additional n_sectors sectors of the bio and the rest of the data should be
1315 * sent in a next bio.
1316 *
1317 * A diagram that explains the arithmetics:
1318 * +--------------------+---------------+-------+
1319 * | 1 | 2 | 3 |
1320 * +--------------------+---------------+-------+
1321 *
1322 * <-------------- *tio->len_ptr --------------->
1323 * <----- bio_sectors ----->
1324 * <-- n_sectors -->
1325 *
1326 * Region 1 was already iterated over with bio_advance or similar function.
1327 * (it may be empty if the target doesn't use bio_advance)
1328 * Region 2 is the remaining bio size that the target wants to process.
1329 * (it may be empty if region 1 is non-empty, although there is no reason
1330 * to make it empty)
1331 * The target requires that region 3 is to be sent in the next bio.
1332 *
1333 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1334 * the partially processed part (the sum of regions 1+2) must be the same for all
1335 * copies of the bio.
1336 */
1337void dm_accept_partial_bio(struct bio *bio, unsigned int n_sectors)
1338{
1339 struct dm_target_io *tio = clone_to_tio(bio);
1340 struct dm_io *io = tio->io;
1341 unsigned int bio_sectors = bio_sectors(bio);
1342
1343 BUG_ON(dm_tio_flagged(tio, DM_TIO_IS_DUPLICATE_BIO));
1344 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1345 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1346 BUG_ON(bio_sectors > *tio->len_ptr);
1347 BUG_ON(n_sectors > bio_sectors);
1348
1349 *tio->len_ptr -= bio_sectors - n_sectors;
1350 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1351
1352 /*
1353 * __split_and_process_bio() may have already saved mapped part
1354 * for accounting but it is being reduced so update accordingly.
1355 */
1356 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1357 io->sectors = n_sectors;
1358 io->sector_offset = bio_sectors(io->orig_bio);
1359}
1360EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1361
1362/*
1363 * @clone: clone bio that DM core passed to target's .map function
1364 * @tgt_clone: clone of @clone bio that target needs submitted
1365 *
1366 * Targets should use this interface to submit bios they take
1367 * ownership of when returning DM_MAPIO_SUBMITTED.
1368 *
1369 * Target should also enable ti->accounts_remapped_io
1370 */
1371void dm_submit_bio_remap(struct bio *clone, struct bio *tgt_clone)
1372{
1373 struct dm_target_io *tio = clone_to_tio(clone);
1374 struct dm_io *io = tio->io;
1375
1376 /* establish bio that will get submitted */
1377 if (!tgt_clone)
1378 tgt_clone = clone;
1379
1380 /*
1381 * Account io->origin_bio to DM dev on behalf of target
1382 * that took ownership of IO with DM_MAPIO_SUBMITTED.
1383 */
1384 dm_start_io_acct(io, clone);
1385
1386 trace_block_bio_remap(tgt_clone, disk_devt(io->md->disk),
1387 tio->old_sector);
1388 submit_bio_noacct(tgt_clone);
1389}
1390EXPORT_SYMBOL_GPL(dm_submit_bio_remap);
1391
1392static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1393{
1394 mutex_lock(&md->swap_bios_lock);
1395 while (latch < md->swap_bios) {
1396 cond_resched();
1397 down(&md->swap_bios_semaphore);
1398 md->swap_bios--;
1399 }
1400 while (latch > md->swap_bios) {
1401 cond_resched();
1402 up(&md->swap_bios_semaphore);
1403 md->swap_bios++;
1404 }
1405 mutex_unlock(&md->swap_bios_lock);
1406}
1407
1408static void __map_bio(struct bio *clone)
1409{
1410 struct dm_target_io *tio = clone_to_tio(clone);
1411 struct dm_target *ti = tio->ti;
1412 struct dm_io *io = tio->io;
1413 struct mapped_device *md = io->md;
1414 int r;
1415
1416 clone->bi_end_io = clone_endio;
1417
1418 /*
1419 * Map the clone.
1420 */
1421 tio->old_sector = clone->bi_iter.bi_sector;
1422
1423 if (static_branch_unlikely(&swap_bios_enabled) &&
1424 unlikely(swap_bios_limit(ti, clone))) {
1425 int latch = get_swap_bios();
1426
1427 if (unlikely(latch != md->swap_bios))
1428 __set_swap_bios_limit(md, latch);
1429 down(&md->swap_bios_semaphore);
1430 }
1431
1432 if (likely(ti->type->map == linear_map))
1433 r = linear_map(ti, clone);
1434 else if (ti->type->map == stripe_map)
1435 r = stripe_map(ti, clone);
1436 else
1437 r = ti->type->map(ti, clone);
1438
1439 switch (r) {
1440 case DM_MAPIO_SUBMITTED:
1441 /* target has assumed ownership of this io */
1442 if (!ti->accounts_remapped_io)
1443 dm_start_io_acct(io, clone);
1444 break;
1445 case DM_MAPIO_REMAPPED:
1446 dm_submit_bio_remap(clone, NULL);
1447 break;
1448 case DM_MAPIO_KILL:
1449 case DM_MAPIO_REQUEUE:
1450 if (static_branch_unlikely(&swap_bios_enabled) &&
1451 unlikely(swap_bios_limit(ti, clone)))
1452 up(&md->swap_bios_semaphore);
1453 free_tio(clone);
1454 if (r == DM_MAPIO_KILL)
1455 dm_io_dec_pending(io, BLK_STS_IOERR);
1456 else
1457 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1458 break;
1459 default:
1460 DMCRIT("unimplemented target map return value: %d", r);
1461 BUG();
1462 }
1463}
1464
1465static void setup_split_accounting(struct clone_info *ci, unsigned int len)
1466{
1467 struct dm_io *io = ci->io;
1468
1469 if (ci->sector_count > len) {
1470 /*
1471 * Split needed, save the mapped part for accounting.
1472 * NOTE: dm_accept_partial_bio() will update accordingly.
1473 */
1474 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1475 io->sectors = len;
1476 io->sector_offset = bio_sectors(ci->bio);
1477 }
1478}
1479
1480static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1481 struct dm_target *ti, unsigned int num_bios,
1482 unsigned *len, gfp_t gfp_flag)
1483{
1484 struct bio *bio;
1485 int try = (gfp_flag & GFP_NOWAIT) ? 0 : 1;
1486
1487 for (; try < 2; try++) {
1488 int bio_nr;
1489
1490 if (try && num_bios > 1)
1491 mutex_lock(&ci->io->md->table_devices_lock);
1492 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1493 bio = alloc_tio(ci, ti, bio_nr, len,
1494 try ? GFP_NOIO : GFP_NOWAIT);
1495 if (!bio)
1496 break;
1497
1498 bio_list_add(blist, bio);
1499 }
1500 if (try && num_bios > 1)
1501 mutex_unlock(&ci->io->md->table_devices_lock);
1502 if (bio_nr == num_bios)
1503 return;
1504
1505 while ((bio = bio_list_pop(blist)))
1506 free_tio(bio);
1507 }
1508}
1509
1510static unsigned int __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1511 unsigned int num_bios, unsigned int *len,
1512 gfp_t gfp_flag)
1513{
1514 struct bio_list blist = BIO_EMPTY_LIST;
1515 struct bio *clone;
1516 unsigned int ret = 0;
1517
1518 if (WARN_ON_ONCE(num_bios == 0)) /* num_bios = 0 is a bug in caller */
1519 return 0;
1520
1521 /* dm_accept_partial_bio() is not supported with shared tio->len_ptr */
1522 if (len)
1523 setup_split_accounting(ci, *len);
1524
1525 /*
1526 * Using alloc_multiple_bios(), even if num_bios is 1, to consistently
1527 * support allocating using GFP_NOWAIT with GFP_NOIO fallback.
1528 */
1529 alloc_multiple_bios(&blist, ci, ti, num_bios, len, gfp_flag);
1530 while ((clone = bio_list_pop(&blist))) {
1531 if (num_bios > 1)
1532 dm_tio_set_flag(clone_to_tio(clone), DM_TIO_IS_DUPLICATE_BIO);
1533 __map_bio(clone);
1534 ret += 1;
1535 }
1536
1537 return ret;
1538}
1539
1540static void __send_empty_flush(struct clone_info *ci)
1541{
1542 struct dm_table *t = ci->map;
1543 struct bio flush_bio;
1544
1545 /*
1546 * Use an on-stack bio for this, it's safe since we don't
1547 * need to reference it after submit. It's just used as
1548 * the basis for the clone(s).
1549 */
1550 bio_init(&flush_bio, ci->io->md->disk->part0, NULL, 0,
1551 REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
1552
1553 ci->bio = &flush_bio;
1554 ci->sector_count = 0;
1555 ci->io->tio.clone.bi_iter.bi_size = 0;
1556
1557 if (!t->flush_bypasses_map) {
1558 for (unsigned int i = 0; i < t->num_targets; i++) {
1559 unsigned int bios;
1560 struct dm_target *ti = dm_table_get_target(t, i);
1561
1562 if (unlikely(ti->num_flush_bios == 0))
1563 continue;
1564
1565 atomic_add(ti->num_flush_bios, &ci->io->io_count);
1566 bios = __send_duplicate_bios(ci, ti, ti->num_flush_bios,
1567 NULL, GFP_NOWAIT);
1568 atomic_sub(ti->num_flush_bios - bios, &ci->io->io_count);
1569 }
1570 } else {
1571 /*
1572 * Note that there's no need to grab t->devices_lock here
1573 * because the targets that support flush optimization don't
1574 * modify the list of devices.
1575 */
1576 struct list_head *devices = dm_table_get_devices(t);
1577 unsigned int len = 0;
1578 struct dm_dev_internal *dd;
1579 list_for_each_entry(dd, devices, list) {
1580 struct bio *clone;
1581 /*
1582 * Note that the structure dm_target_io is not
1583 * associated with any target (because the device may be
1584 * used by multiple targets), so we set tio->ti = NULL.
1585 * We must check for NULL in the I/O processing path, to
1586 * avoid NULL pointer dereference.
1587 */
1588 clone = alloc_tio(ci, NULL, 0, &len, GFP_NOIO);
1589 atomic_add(1, &ci->io->io_count);
1590 bio_set_dev(clone, dd->dm_dev->bdev);
1591 clone->bi_end_io = clone_endio;
1592 dm_submit_bio_remap(clone, NULL);
1593 }
1594 }
1595
1596 /*
1597 * alloc_io() takes one extra reference for submission, so the
1598 * reference won't reach 0 without the following subtraction
1599 */
1600 atomic_sub(1, &ci->io->io_count);
1601
1602 bio_uninit(ci->bio);
1603}
1604
1605static void __send_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1606 unsigned int num_bios, unsigned int max_granularity,
1607 unsigned int max_sectors)
1608{
1609 unsigned int len, bios;
1610
1611 len = min_t(sector_t, ci->sector_count,
1612 __max_io_len(ti, ci->sector, max_granularity, max_sectors));
1613
1614 atomic_add(num_bios, &ci->io->io_count);
1615 bios = __send_duplicate_bios(ci, ti, num_bios, &len, GFP_NOIO);
1616 /*
1617 * alloc_io() takes one extra reference for submission, so the
1618 * reference won't reach 0 without the following (+1) subtraction
1619 */
1620 atomic_sub(num_bios - bios + 1, &ci->io->io_count);
1621
1622 ci->sector += len;
1623 ci->sector_count -= len;
1624}
1625
1626static bool is_abnormal_io(struct bio *bio)
1627{
1628 switch (bio_op(bio)) {
1629 case REQ_OP_READ:
1630 case REQ_OP_WRITE:
1631 case REQ_OP_FLUSH:
1632 return false;
1633 case REQ_OP_DISCARD:
1634 case REQ_OP_SECURE_ERASE:
1635 case REQ_OP_WRITE_ZEROES:
1636 case REQ_OP_ZONE_RESET_ALL:
1637 return true;
1638 default:
1639 return false;
1640 }
1641}
1642
1643static blk_status_t __process_abnormal_io(struct clone_info *ci,
1644 struct dm_target *ti)
1645{
1646 unsigned int num_bios = 0;
1647 unsigned int max_granularity = 0;
1648 unsigned int max_sectors = 0;
1649 struct queue_limits *limits = dm_get_queue_limits(ti->table->md);
1650
1651 switch (bio_op(ci->bio)) {
1652 case REQ_OP_DISCARD:
1653 num_bios = ti->num_discard_bios;
1654 max_sectors = limits->max_discard_sectors;
1655 if (ti->max_discard_granularity)
1656 max_granularity = max_sectors;
1657 break;
1658 case REQ_OP_SECURE_ERASE:
1659 num_bios = ti->num_secure_erase_bios;
1660 max_sectors = limits->max_secure_erase_sectors;
1661 break;
1662 case REQ_OP_WRITE_ZEROES:
1663 num_bios = ti->num_write_zeroes_bios;
1664 max_sectors = limits->max_write_zeroes_sectors;
1665 break;
1666 default:
1667 break;
1668 }
1669
1670 /*
1671 * Even though the device advertised support for this type of
1672 * request, that does not mean every target supports it, and
1673 * reconfiguration might also have changed that since the
1674 * check was performed.
1675 */
1676 if (unlikely(!num_bios))
1677 return BLK_STS_NOTSUPP;
1678
1679 __send_abnormal_io(ci, ti, num_bios, max_granularity, max_sectors);
1680
1681 return BLK_STS_OK;
1682}
1683
1684/*
1685 * Reuse ->bi_private as dm_io list head for storing all dm_io instances
1686 * associated with this bio, and this bio's bi_private needs to be
1687 * stored in dm_io->data before the reuse.
1688 *
1689 * bio->bi_private is owned by fs or upper layer, so block layer won't
1690 * touch it after splitting. Meantime it won't be changed by anyone after
1691 * bio is submitted. So this reuse is safe.
1692 */
1693static inline struct dm_io **dm_poll_list_head(struct bio *bio)
1694{
1695 return (struct dm_io **)&bio->bi_private;
1696}
1697
1698static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
1699{
1700 struct dm_io **head = dm_poll_list_head(bio);
1701
1702 if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
1703 bio->bi_opf |= REQ_DM_POLL_LIST;
1704 /*
1705 * Save .bi_private into dm_io, so that we can reuse
1706 * .bi_private as dm_io list head for storing dm_io list
1707 */
1708 io->data = bio->bi_private;
1709
1710 /* tell block layer to poll for completion */
1711 bio->bi_cookie = ~BLK_QC_T_NONE;
1712
1713 io->next = NULL;
1714 } else {
1715 /*
1716 * bio recursed due to split, reuse original poll list,
1717 * and save bio->bi_private too.
1718 */
1719 io->data = (*head)->data;
1720 io->next = *head;
1721 }
1722
1723 *head = io;
1724}
1725
1726/*
1727 * Select the correct strategy for processing a non-flush bio.
1728 */
1729static blk_status_t __split_and_process_bio(struct clone_info *ci)
1730{
1731 struct bio *clone;
1732 struct dm_target *ti;
1733 unsigned int len;
1734
1735 ti = dm_table_find_target(ci->map, ci->sector);
1736 if (unlikely(!ti))
1737 return BLK_STS_IOERR;
1738
1739 if (unlikely(ci->is_abnormal_io))
1740 return __process_abnormal_io(ci, ti);
1741
1742 /*
1743 * Only support bio polling for normal IO, and the target io is
1744 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
1745 */
1746 ci->submit_as_polled = !!(ci->bio->bi_opf & REQ_POLLED);
1747
1748 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1749 setup_split_accounting(ci, len);
1750
1751 if (unlikely(ci->bio->bi_opf & REQ_NOWAIT)) {
1752 if (unlikely(!dm_target_supports_nowait(ti->type)))
1753 return BLK_STS_NOTSUPP;
1754
1755 clone = alloc_tio(ci, ti, 0, &len, GFP_NOWAIT);
1756 if (unlikely(!clone))
1757 return BLK_STS_AGAIN;
1758 } else {
1759 clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
1760 }
1761 __map_bio(clone);
1762
1763 ci->sector += len;
1764 ci->sector_count -= len;
1765
1766 return BLK_STS_OK;
1767}
1768
1769static void init_clone_info(struct clone_info *ci, struct dm_io *io,
1770 struct dm_table *map, struct bio *bio, bool is_abnormal)
1771{
1772 ci->map = map;
1773 ci->io = io;
1774 ci->bio = bio;
1775 ci->is_abnormal_io = is_abnormal;
1776 ci->submit_as_polled = false;
1777 ci->sector = bio->bi_iter.bi_sector;
1778 ci->sector_count = bio_sectors(bio);
1779
1780 /* Shouldn't happen but sector_count was being set to 0 so... */
1781 if (static_branch_unlikely(&zoned_enabled) &&
1782 WARN_ON_ONCE(op_is_zone_mgmt(bio_op(bio)) && ci->sector_count))
1783 ci->sector_count = 0;
1784}
1785
1786#ifdef CONFIG_BLK_DEV_ZONED
1787static inline bool dm_zone_bio_needs_split(struct mapped_device *md,
1788 struct bio *bio)
1789{
1790 /*
1791 * For mapped device that need zone append emulation, we must
1792 * split any large BIO that straddles zone boundaries.
1793 */
1794 return dm_emulate_zone_append(md) && bio_straddles_zones(bio) &&
1795 !bio_flagged(bio, BIO_ZONE_WRITE_PLUGGING);
1796}
1797static inline bool dm_zone_plug_bio(struct mapped_device *md, struct bio *bio)
1798{
1799 return dm_emulate_zone_append(md) && blk_zone_plug_bio(bio, 0);
1800}
1801
1802static blk_status_t __send_zone_reset_all_emulated(struct clone_info *ci,
1803 struct dm_target *ti)
1804{
1805 struct bio_list blist = BIO_EMPTY_LIST;
1806 struct mapped_device *md = ci->io->md;
1807 unsigned int zone_sectors = md->disk->queue->limits.chunk_sectors;
1808 unsigned long *need_reset;
1809 unsigned int i, nr_zones, nr_reset;
1810 unsigned int num_bios = 0;
1811 blk_status_t sts = BLK_STS_OK;
1812 sector_t sector = ti->begin;
1813 struct bio *clone;
1814 int ret;
1815
1816 nr_zones = ti->len >> ilog2(zone_sectors);
1817 need_reset = bitmap_zalloc(nr_zones, GFP_NOIO);
1818 if (!need_reset)
1819 return BLK_STS_RESOURCE;
1820
1821 ret = dm_zone_get_reset_bitmap(md, ci->map, ti->begin,
1822 nr_zones, need_reset);
1823 if (ret) {
1824 sts = BLK_STS_IOERR;
1825 goto free_bitmap;
1826 }
1827
1828 /* If we have no zone to reset, we are done. */
1829 nr_reset = bitmap_weight(need_reset, nr_zones);
1830 if (!nr_reset)
1831 goto free_bitmap;
1832
1833 atomic_add(nr_zones, &ci->io->io_count);
1834
1835 for (i = 0; i < nr_zones; i++) {
1836
1837 if (!test_bit(i, need_reset)) {
1838 sector += zone_sectors;
1839 continue;
1840 }
1841
1842 if (bio_list_empty(&blist)) {
1843 /* This may take a while, so be nice to others */
1844 if (num_bios)
1845 cond_resched();
1846
1847 /*
1848 * We may need to reset thousands of zones, so let's
1849 * not go crazy with the clone allocation.
1850 */
1851 alloc_multiple_bios(&blist, ci, ti, min(nr_reset, 32),
1852 NULL, GFP_NOIO);
1853 }
1854
1855 /* Get a clone and change it to a regular reset operation. */
1856 clone = bio_list_pop(&blist);
1857 clone->bi_opf &= ~REQ_OP_MASK;
1858 clone->bi_opf |= REQ_OP_ZONE_RESET | REQ_SYNC;
1859 clone->bi_iter.bi_sector = sector;
1860 clone->bi_iter.bi_size = 0;
1861 __map_bio(clone);
1862
1863 sector += zone_sectors;
1864 num_bios++;
1865 nr_reset--;
1866 }
1867
1868 WARN_ON_ONCE(!bio_list_empty(&blist));
1869 atomic_sub(nr_zones - num_bios, &ci->io->io_count);
1870 ci->sector_count = 0;
1871
1872free_bitmap:
1873 bitmap_free(need_reset);
1874
1875 return sts;
1876}
1877
1878static void __send_zone_reset_all_native(struct clone_info *ci,
1879 struct dm_target *ti)
1880{
1881 unsigned int bios;
1882
1883 atomic_add(1, &ci->io->io_count);
1884 bios = __send_duplicate_bios(ci, ti, 1, NULL, GFP_NOIO);
1885 atomic_sub(1 - bios, &ci->io->io_count);
1886
1887 ci->sector_count = 0;
1888}
1889
1890static blk_status_t __send_zone_reset_all(struct clone_info *ci)
1891{
1892 struct dm_table *t = ci->map;
1893 blk_status_t sts = BLK_STS_OK;
1894
1895 for (unsigned int i = 0; i < t->num_targets; i++) {
1896 struct dm_target *ti = dm_table_get_target(t, i);
1897
1898 if (ti->zone_reset_all_supported) {
1899 __send_zone_reset_all_native(ci, ti);
1900 continue;
1901 }
1902
1903 sts = __send_zone_reset_all_emulated(ci, ti);
1904 if (sts != BLK_STS_OK)
1905 break;
1906 }
1907
1908 /* Release the reference that alloc_io() took for submission. */
1909 atomic_sub(1, &ci->io->io_count);
1910
1911 return sts;
1912}
1913
1914#else
1915static inline bool dm_zone_bio_needs_split(struct mapped_device *md,
1916 struct bio *bio)
1917{
1918 return false;
1919}
1920static inline bool dm_zone_plug_bio(struct mapped_device *md, struct bio *bio)
1921{
1922 return false;
1923}
1924static blk_status_t __send_zone_reset_all(struct clone_info *ci)
1925{
1926 return BLK_STS_NOTSUPP;
1927}
1928#endif
1929
1930/*
1931 * Entry point to split a bio into clones and submit them to the targets.
1932 */
1933static void dm_split_and_process_bio(struct mapped_device *md,
1934 struct dm_table *map, struct bio *bio)
1935{
1936 struct clone_info ci;
1937 struct dm_io *io;
1938 blk_status_t error = BLK_STS_OK;
1939 bool is_abnormal, need_split;
1940
1941 is_abnormal = is_abnormal_io(bio);
1942 if (static_branch_unlikely(&zoned_enabled)) {
1943 /* Special case REQ_OP_ZONE_RESET_ALL as it cannot be split. */
1944 need_split = (bio_op(bio) != REQ_OP_ZONE_RESET_ALL) &&
1945 (is_abnormal || dm_zone_bio_needs_split(md, bio));
1946 } else {
1947 need_split = is_abnormal;
1948 }
1949
1950 if (unlikely(need_split)) {
1951 /*
1952 * Use bio_split_to_limits() for abnormal IO (e.g. discard, etc)
1953 * otherwise associated queue_limits won't be imposed.
1954 * Also split the BIO for mapped devices needing zone append
1955 * emulation to ensure that the BIO does not cross zone
1956 * boundaries.
1957 */
1958 bio = bio_split_to_limits(bio);
1959 if (!bio)
1960 return;
1961 }
1962
1963 /*
1964 * Use the block layer zone write plugging for mapped devices that
1965 * need zone append emulation (e.g. dm-crypt).
1966 */
1967 if (static_branch_unlikely(&zoned_enabled) && dm_zone_plug_bio(md, bio))
1968 return;
1969
1970 /* Only support nowait for normal IO */
1971 if (unlikely(bio->bi_opf & REQ_NOWAIT) && !is_abnormal) {
1972 io = alloc_io(md, bio, GFP_NOWAIT);
1973 if (unlikely(!io)) {
1974 /* Unable to do anything without dm_io. */
1975 bio_wouldblock_error(bio);
1976 return;
1977 }
1978 } else {
1979 io = alloc_io(md, bio, GFP_NOIO);
1980 }
1981 init_clone_info(&ci, io, map, bio, is_abnormal);
1982
1983 if (bio->bi_opf & REQ_PREFLUSH) {
1984 __send_empty_flush(&ci);
1985 /* dm_io_complete submits any data associated with flush */
1986 goto out;
1987 }
1988
1989 if (static_branch_unlikely(&zoned_enabled) &&
1990 (bio_op(bio) == REQ_OP_ZONE_RESET_ALL)) {
1991 error = __send_zone_reset_all(&ci);
1992 goto out;
1993 }
1994
1995 error = __split_and_process_bio(&ci);
1996 if (error || !ci.sector_count)
1997 goto out;
1998 /*
1999 * Remainder must be passed to submit_bio_noacct() so it gets handled
2000 * *after* bios already submitted have been completely processed.
2001 */
2002 bio_trim(bio, io->sectors, ci.sector_count);
2003 trace_block_split(bio, bio->bi_iter.bi_sector);
2004 bio_inc_remaining(bio);
2005 submit_bio_noacct(bio);
2006out:
2007 /*
2008 * Drop the extra reference count for non-POLLED bio, and hold one
2009 * reference for POLLED bio, which will be released in dm_poll_bio
2010 *
2011 * Add every dm_io instance into the dm_io list head which is stored
2012 * in bio->bi_private, so that dm_poll_bio can poll them all.
2013 */
2014 if (error || !ci.submit_as_polled) {
2015 /*
2016 * In case of submission failure, the extra reference for
2017 * submitting io isn't consumed yet
2018 */
2019 if (error)
2020 atomic_dec(&io->io_count);
2021 dm_io_dec_pending(io, error);
2022 } else
2023 dm_queue_poll_io(bio, io);
2024}
2025
2026static void dm_submit_bio(struct bio *bio)
2027{
2028 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
2029 int srcu_idx;
2030 struct dm_table *map;
2031
2032 map = dm_get_live_table(md, &srcu_idx);
2033 if (unlikely(!map)) {
2034 DMERR_LIMIT("%s: mapping table unavailable, erroring io",
2035 dm_device_name(md));
2036 bio_io_error(bio);
2037 goto out;
2038 }
2039
2040 /* If suspended, queue this IO for later */
2041 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
2042 if (bio->bi_opf & REQ_NOWAIT)
2043 bio_wouldblock_error(bio);
2044 else if (bio->bi_opf & REQ_RAHEAD)
2045 bio_io_error(bio);
2046 else
2047 queue_io(md, bio);
2048 goto out;
2049 }
2050
2051 dm_split_and_process_bio(md, map, bio);
2052out:
2053 dm_put_live_table(md, srcu_idx);
2054}
2055
2056static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
2057 unsigned int flags)
2058{
2059 WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
2060
2061 /* don't poll if the mapped io is done */
2062 if (atomic_read(&io->io_count) > 1)
2063 bio_poll(&io->tio.clone, iob, flags);
2064
2065 /* bio_poll holds the last reference */
2066 return atomic_read(&io->io_count) == 1;
2067}
2068
2069static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
2070 unsigned int flags)
2071{
2072 struct dm_io **head = dm_poll_list_head(bio);
2073 struct dm_io *list = *head;
2074 struct dm_io *tmp = NULL;
2075 struct dm_io *curr, *next;
2076
2077 /* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
2078 if (!(bio->bi_opf & REQ_DM_POLL_LIST))
2079 return 0;
2080
2081 WARN_ON_ONCE(!list);
2082
2083 /*
2084 * Restore .bi_private before possibly completing dm_io.
2085 *
2086 * bio_poll() is only possible once @bio has been completely
2087 * submitted via submit_bio_noacct()'s depth-first submission.
2088 * So there is no dm_queue_poll_io() race associated with
2089 * clearing REQ_DM_POLL_LIST here.
2090 */
2091 bio->bi_opf &= ~REQ_DM_POLL_LIST;
2092 bio->bi_private = list->data;
2093
2094 for (curr = list, next = curr->next; curr; curr = next, next =
2095 curr ? curr->next : NULL) {
2096 if (dm_poll_dm_io(curr, iob, flags)) {
2097 /*
2098 * clone_endio() has already occurred, so no
2099 * error handling is needed here.
2100 */
2101 __dm_io_dec_pending(curr);
2102 } else {
2103 curr->next = tmp;
2104 tmp = curr;
2105 }
2106 }
2107
2108 /* Not done? */
2109 if (tmp) {
2110 bio->bi_opf |= REQ_DM_POLL_LIST;
2111 /* Reset bio->bi_private to dm_io list head */
2112 *head = tmp;
2113 return 0;
2114 }
2115 return 1;
2116}
2117
2118/*
2119 *---------------------------------------------------------------
2120 * An IDR is used to keep track of allocated minor numbers.
2121 *---------------------------------------------------------------
2122 */
2123static void free_minor(int minor)
2124{
2125 spin_lock(&_minor_lock);
2126 idr_remove(&_minor_idr, minor);
2127 spin_unlock(&_minor_lock);
2128}
2129
2130/*
2131 * See if the device with a specific minor # is free.
2132 */
2133static int specific_minor(int minor)
2134{
2135 int r;
2136
2137 if (minor >= (1 << MINORBITS))
2138 return -EINVAL;
2139
2140 idr_preload(GFP_KERNEL);
2141 spin_lock(&_minor_lock);
2142
2143 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2144
2145 spin_unlock(&_minor_lock);
2146 idr_preload_end();
2147 if (r < 0)
2148 return r == -ENOSPC ? -EBUSY : r;
2149 return 0;
2150}
2151
2152static int next_free_minor(int *minor)
2153{
2154 int r;
2155
2156 idr_preload(GFP_KERNEL);
2157 spin_lock(&_minor_lock);
2158
2159 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2160
2161 spin_unlock(&_minor_lock);
2162 idr_preload_end();
2163 if (r < 0)
2164 return r;
2165 *minor = r;
2166 return 0;
2167}
2168
2169static const struct block_device_operations dm_blk_dops;
2170static const struct block_device_operations dm_rq_blk_dops;
2171static const struct dax_operations dm_dax_ops;
2172
2173static void dm_wq_work(struct work_struct *work);
2174
2175#ifdef CONFIG_BLK_INLINE_ENCRYPTION
2176static void dm_queue_destroy_crypto_profile(struct request_queue *q)
2177{
2178 dm_destroy_crypto_profile(q->crypto_profile);
2179}
2180
2181#else /* CONFIG_BLK_INLINE_ENCRYPTION */
2182
2183static inline void dm_queue_destroy_crypto_profile(struct request_queue *q)
2184{
2185}
2186#endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
2187
2188static void cleanup_mapped_device(struct mapped_device *md)
2189{
2190 if (md->wq)
2191 destroy_workqueue(md->wq);
2192 dm_free_md_mempools(md->mempools);
2193
2194 if (md->dax_dev) {
2195 dax_remove_host(md->disk);
2196 kill_dax(md->dax_dev);
2197 put_dax(md->dax_dev);
2198 md->dax_dev = NULL;
2199 }
2200
2201 if (md->disk) {
2202 spin_lock(&_minor_lock);
2203 md->disk->private_data = NULL;
2204 spin_unlock(&_minor_lock);
2205 if (dm_get_md_type(md) != DM_TYPE_NONE) {
2206 struct table_device *td;
2207
2208 dm_sysfs_exit(md);
2209 list_for_each_entry(td, &md->table_devices, list) {
2210 bd_unlink_disk_holder(td->dm_dev.bdev,
2211 md->disk);
2212 }
2213
2214 /*
2215 * Hold lock to make sure del_gendisk() won't concurrent
2216 * with open/close_table_device().
2217 */
2218 mutex_lock(&md->table_devices_lock);
2219 del_gendisk(md->disk);
2220 mutex_unlock(&md->table_devices_lock);
2221 }
2222 dm_queue_destroy_crypto_profile(md->queue);
2223 put_disk(md->disk);
2224 }
2225
2226 if (md->pending_io) {
2227 free_percpu(md->pending_io);
2228 md->pending_io = NULL;
2229 }
2230
2231 cleanup_srcu_struct(&md->io_barrier);
2232
2233 mutex_destroy(&md->suspend_lock);
2234 mutex_destroy(&md->type_lock);
2235 mutex_destroy(&md->table_devices_lock);
2236 mutex_destroy(&md->swap_bios_lock);
2237
2238 dm_mq_cleanup_mapped_device(md);
2239}
2240
2241/*
2242 * Allocate and initialise a blank device with a given minor.
2243 */
2244static struct mapped_device *alloc_dev(int minor)
2245{
2246 int r, numa_node_id = dm_get_numa_node();
2247 struct dax_device *dax_dev;
2248 struct mapped_device *md;
2249 void *old_md;
2250
2251 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
2252 if (!md) {
2253 DMERR("unable to allocate device, out of memory.");
2254 return NULL;
2255 }
2256
2257 if (!try_module_get(THIS_MODULE))
2258 goto bad_module_get;
2259
2260 /* get a minor number for the dev */
2261 if (minor == DM_ANY_MINOR)
2262 r = next_free_minor(&minor);
2263 else
2264 r = specific_minor(minor);
2265 if (r < 0)
2266 goto bad_minor;
2267
2268 r = init_srcu_struct(&md->io_barrier);
2269 if (r < 0)
2270 goto bad_io_barrier;
2271
2272 md->numa_node_id = numa_node_id;
2273 md->init_tio_pdu = false;
2274 md->type = DM_TYPE_NONE;
2275 mutex_init(&md->suspend_lock);
2276 mutex_init(&md->type_lock);
2277 mutex_init(&md->table_devices_lock);
2278 spin_lock_init(&md->deferred_lock);
2279 atomic_set(&md->holders, 1);
2280 atomic_set(&md->open_count, 0);
2281 atomic_set(&md->event_nr, 0);
2282 atomic_set(&md->uevent_seq, 0);
2283 INIT_LIST_HEAD(&md->uevent_list);
2284 INIT_LIST_HEAD(&md->table_devices);
2285 spin_lock_init(&md->uevent_lock);
2286
2287 /*
2288 * default to bio-based until DM table is loaded and md->type
2289 * established. If request-based table is loaded: blk-mq will
2290 * override accordingly.
2291 */
2292 md->disk = blk_alloc_disk(NULL, md->numa_node_id);
2293 if (IS_ERR(md->disk)) {
2294 md->disk = NULL;
2295 goto bad;
2296 }
2297 md->queue = md->disk->queue;
2298
2299 init_waitqueue_head(&md->wait);
2300 INIT_WORK(&md->work, dm_wq_work);
2301 INIT_WORK(&md->requeue_work, dm_wq_requeue_work);
2302 init_waitqueue_head(&md->eventq);
2303 init_completion(&md->kobj_holder.completion);
2304
2305 md->requeue_list = NULL;
2306 md->swap_bios = get_swap_bios();
2307 sema_init(&md->swap_bios_semaphore, md->swap_bios);
2308 mutex_init(&md->swap_bios_lock);
2309
2310 md->disk->major = _major;
2311 md->disk->first_minor = minor;
2312 md->disk->minors = 1;
2313 md->disk->flags |= GENHD_FL_NO_PART;
2314 md->disk->fops = &dm_blk_dops;
2315 md->disk->private_data = md;
2316 sprintf(md->disk->disk_name, "dm-%d", minor);
2317
2318 dax_dev = alloc_dax(md, &dm_dax_ops);
2319 if (IS_ERR(dax_dev)) {
2320 if (PTR_ERR(dax_dev) != -EOPNOTSUPP)
2321 goto bad;
2322 } else {
2323 set_dax_nocache(dax_dev);
2324 set_dax_nomc(dax_dev);
2325 md->dax_dev = dax_dev;
2326 if (dax_add_host(dax_dev, md->disk))
2327 goto bad;
2328 }
2329
2330 format_dev_t(md->name, MKDEV(_major, minor));
2331
2332 md->wq = alloc_workqueue("kdmflush/%s", WQ_MEM_RECLAIM, 0, md->name);
2333 if (!md->wq)
2334 goto bad;
2335
2336 md->pending_io = alloc_percpu(unsigned long);
2337 if (!md->pending_io)
2338 goto bad;
2339
2340 r = dm_stats_init(&md->stats);
2341 if (r < 0)
2342 goto bad;
2343
2344 /* Populate the mapping, nobody knows we exist yet */
2345 spin_lock(&_minor_lock);
2346 old_md = idr_replace(&_minor_idr, md, minor);
2347 spin_unlock(&_minor_lock);
2348
2349 BUG_ON(old_md != MINOR_ALLOCED);
2350
2351 return md;
2352
2353bad:
2354 cleanup_mapped_device(md);
2355bad_io_barrier:
2356 free_minor(minor);
2357bad_minor:
2358 module_put(THIS_MODULE);
2359bad_module_get:
2360 kvfree(md);
2361 return NULL;
2362}
2363
2364static void unlock_fs(struct mapped_device *md);
2365
2366static void free_dev(struct mapped_device *md)
2367{
2368 int minor = MINOR(disk_devt(md->disk));
2369
2370 unlock_fs(md);
2371
2372 cleanup_mapped_device(md);
2373
2374 WARN_ON_ONCE(!list_empty(&md->table_devices));
2375 dm_stats_cleanup(&md->stats);
2376 free_minor(minor);
2377
2378 module_put(THIS_MODULE);
2379 kvfree(md);
2380}
2381
2382/*
2383 * Bind a table to the device.
2384 */
2385static void event_callback(void *context)
2386{
2387 unsigned long flags;
2388 LIST_HEAD(uevents);
2389 struct mapped_device *md = context;
2390
2391 spin_lock_irqsave(&md->uevent_lock, flags);
2392 list_splice_init(&md->uevent_list, &uevents);
2393 spin_unlock_irqrestore(&md->uevent_lock, flags);
2394
2395 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2396
2397 atomic_inc(&md->event_nr);
2398 wake_up(&md->eventq);
2399 dm_issue_global_event();
2400}
2401
2402/*
2403 * Returns old map, which caller must destroy.
2404 */
2405static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2406 struct queue_limits *limits)
2407{
2408 struct dm_table *old_map;
2409 sector_t size;
2410 int ret;
2411
2412 lockdep_assert_held(&md->suspend_lock);
2413
2414 size = dm_table_get_size(t);
2415
2416 /*
2417 * Wipe any geometry if the size of the table changed.
2418 */
2419 if (size != dm_get_size(md))
2420 memset(&md->geometry, 0, sizeof(md->geometry));
2421
2422 set_capacity(md->disk, size);
2423
2424 dm_table_event_callback(t, event_callback, md);
2425
2426 if (dm_table_request_based(t)) {
2427 /*
2428 * Leverage the fact that request-based DM targets are
2429 * immutable singletons - used to optimize dm_mq_queue_rq.
2430 */
2431 md->immutable_target = dm_table_get_immutable_target(t);
2432
2433 /*
2434 * There is no need to reload with request-based dm because the
2435 * size of front_pad doesn't change.
2436 *
2437 * Note for future: If you are to reload bioset, prep-ed
2438 * requests in the queue may refer to bio from the old bioset,
2439 * so you must walk through the queue to unprep.
2440 */
2441 if (!md->mempools) {
2442 md->mempools = t->mempools;
2443 t->mempools = NULL;
2444 }
2445 } else {
2446 /*
2447 * The md may already have mempools that need changing.
2448 * If so, reload bioset because front_pad may have changed
2449 * because a different table was loaded.
2450 */
2451 dm_free_md_mempools(md->mempools);
2452 md->mempools = t->mempools;
2453 t->mempools = NULL;
2454 }
2455
2456 ret = dm_table_set_restrictions(t, md->queue, limits);
2457 if (ret) {
2458 old_map = ERR_PTR(ret);
2459 goto out;
2460 }
2461
2462 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2463 rcu_assign_pointer(md->map, (void *)t);
2464 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2465
2466 if (old_map)
2467 dm_sync_table(md);
2468out:
2469 return old_map;
2470}
2471
2472/*
2473 * Returns unbound table for the caller to free.
2474 */
2475static struct dm_table *__unbind(struct mapped_device *md)
2476{
2477 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2478
2479 if (!map)
2480 return NULL;
2481
2482 dm_table_event_callback(map, NULL, NULL);
2483 RCU_INIT_POINTER(md->map, NULL);
2484 dm_sync_table(md);
2485
2486 return map;
2487}
2488
2489/*
2490 * Constructor for a new device.
2491 */
2492int dm_create(int minor, struct mapped_device **result)
2493{
2494 struct mapped_device *md;
2495
2496 md = alloc_dev(minor);
2497 if (!md)
2498 return -ENXIO;
2499
2500 dm_ima_reset_data(md);
2501
2502 *result = md;
2503 return 0;
2504}
2505
2506/*
2507 * Functions to manage md->type.
2508 * All are required to hold md->type_lock.
2509 */
2510void dm_lock_md_type(struct mapped_device *md)
2511{
2512 mutex_lock(&md->type_lock);
2513}
2514
2515void dm_unlock_md_type(struct mapped_device *md)
2516{
2517 mutex_unlock(&md->type_lock);
2518}
2519
2520enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2521{
2522 return md->type;
2523}
2524
2525struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2526{
2527 return md->immutable_target_type;
2528}
2529
2530/*
2531 * Setup the DM device's queue based on md's type
2532 */
2533int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2534{
2535 enum dm_queue_mode type = dm_table_get_type(t);
2536 struct queue_limits limits;
2537 struct table_device *td;
2538 int r;
2539
2540 WARN_ON_ONCE(type == DM_TYPE_NONE);
2541
2542 if (type == DM_TYPE_REQUEST_BASED) {
2543 md->disk->fops = &dm_rq_blk_dops;
2544 r = dm_mq_init_request_queue(md, t);
2545 if (r) {
2546 DMERR("Cannot initialize queue for request-based dm mapped device");
2547 return r;
2548 }
2549 }
2550
2551 r = dm_calculate_queue_limits(t, &limits);
2552 if (r) {
2553 DMERR("Cannot calculate initial queue limits");
2554 return r;
2555 }
2556 r = dm_table_set_restrictions(t, md->queue, &limits);
2557 if (r)
2558 return r;
2559
2560 /*
2561 * Hold lock to make sure add_disk() and del_gendisk() won't concurrent
2562 * with open_table_device() and close_table_device().
2563 */
2564 mutex_lock(&md->table_devices_lock);
2565 r = add_disk(md->disk);
2566 mutex_unlock(&md->table_devices_lock);
2567 if (r)
2568 return r;
2569
2570 /*
2571 * Register the holder relationship for devices added before the disk
2572 * was live.
2573 */
2574 list_for_each_entry(td, &md->table_devices, list) {
2575 r = bd_link_disk_holder(td->dm_dev.bdev, md->disk);
2576 if (r)
2577 goto out_undo_holders;
2578 }
2579
2580 r = dm_sysfs_init(md);
2581 if (r)
2582 goto out_undo_holders;
2583
2584 md->type = type;
2585 return 0;
2586
2587out_undo_holders:
2588 list_for_each_entry_continue_reverse(td, &md->table_devices, list)
2589 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
2590 mutex_lock(&md->table_devices_lock);
2591 del_gendisk(md->disk);
2592 mutex_unlock(&md->table_devices_lock);
2593 return r;
2594}
2595
2596struct mapped_device *dm_get_md(dev_t dev)
2597{
2598 struct mapped_device *md;
2599 unsigned int minor = MINOR(dev);
2600
2601 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2602 return NULL;
2603
2604 spin_lock(&_minor_lock);
2605
2606 md = idr_find(&_minor_idr, minor);
2607 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2608 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2609 md = NULL;
2610 goto out;
2611 }
2612 dm_get(md);
2613out:
2614 spin_unlock(&_minor_lock);
2615
2616 return md;
2617}
2618EXPORT_SYMBOL_GPL(dm_get_md);
2619
2620void *dm_get_mdptr(struct mapped_device *md)
2621{
2622 return md->interface_ptr;
2623}
2624
2625void dm_set_mdptr(struct mapped_device *md, void *ptr)
2626{
2627 md->interface_ptr = ptr;
2628}
2629
2630void dm_get(struct mapped_device *md)
2631{
2632 atomic_inc(&md->holders);
2633 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2634}
2635
2636int dm_hold(struct mapped_device *md)
2637{
2638 spin_lock(&_minor_lock);
2639 if (test_bit(DMF_FREEING, &md->flags)) {
2640 spin_unlock(&_minor_lock);
2641 return -EBUSY;
2642 }
2643 dm_get(md);
2644 spin_unlock(&_minor_lock);
2645 return 0;
2646}
2647EXPORT_SYMBOL_GPL(dm_hold);
2648
2649const char *dm_device_name(struct mapped_device *md)
2650{
2651 return md->name;
2652}
2653EXPORT_SYMBOL_GPL(dm_device_name);
2654
2655static void __dm_destroy(struct mapped_device *md, bool wait)
2656{
2657 struct dm_table *map;
2658 int srcu_idx;
2659
2660 might_sleep();
2661
2662 spin_lock(&_minor_lock);
2663 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2664 set_bit(DMF_FREEING, &md->flags);
2665 spin_unlock(&_minor_lock);
2666
2667 blk_mark_disk_dead(md->disk);
2668
2669 /*
2670 * Take suspend_lock so that presuspend and postsuspend methods
2671 * do not race with internal suspend.
2672 */
2673 mutex_lock(&md->suspend_lock);
2674 map = dm_get_live_table(md, &srcu_idx);
2675 if (!dm_suspended_md(md)) {
2676 dm_table_presuspend_targets(map);
2677 set_bit(DMF_SUSPENDED, &md->flags);
2678 set_bit(DMF_POST_SUSPENDING, &md->flags);
2679 dm_table_postsuspend_targets(map);
2680 }
2681 /* dm_put_live_table must be before fsleep, otherwise deadlock is possible */
2682 dm_put_live_table(md, srcu_idx);
2683 mutex_unlock(&md->suspend_lock);
2684
2685 /*
2686 * Rare, but there may be I/O requests still going to complete,
2687 * for example. Wait for all references to disappear.
2688 * No one should increment the reference count of the mapped_device,
2689 * after the mapped_device state becomes DMF_FREEING.
2690 */
2691 if (wait)
2692 while (atomic_read(&md->holders))
2693 fsleep(1000);
2694 else if (atomic_read(&md->holders))
2695 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2696 dm_device_name(md), atomic_read(&md->holders));
2697
2698 dm_table_destroy(__unbind(md));
2699 free_dev(md);
2700}
2701
2702void dm_destroy(struct mapped_device *md)
2703{
2704 __dm_destroy(md, true);
2705}
2706
2707void dm_destroy_immediate(struct mapped_device *md)
2708{
2709 __dm_destroy(md, false);
2710}
2711
2712void dm_put(struct mapped_device *md)
2713{
2714 atomic_dec(&md->holders);
2715}
2716EXPORT_SYMBOL_GPL(dm_put);
2717
2718static bool dm_in_flight_bios(struct mapped_device *md)
2719{
2720 int cpu;
2721 unsigned long sum = 0;
2722
2723 for_each_possible_cpu(cpu)
2724 sum += *per_cpu_ptr(md->pending_io, cpu);
2725
2726 return sum != 0;
2727}
2728
2729static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2730{
2731 int r = 0;
2732 DEFINE_WAIT(wait);
2733
2734 while (true) {
2735 prepare_to_wait(&md->wait, &wait, task_state);
2736
2737 if (!dm_in_flight_bios(md))
2738 break;
2739
2740 if (signal_pending_state(task_state, current)) {
2741 r = -ERESTARTSYS;
2742 break;
2743 }
2744
2745 io_schedule();
2746 }
2747 finish_wait(&md->wait, &wait);
2748
2749 smp_rmb();
2750
2751 return r;
2752}
2753
2754static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2755{
2756 int r = 0;
2757
2758 if (!queue_is_mq(md->queue))
2759 return dm_wait_for_bios_completion(md, task_state);
2760
2761 while (true) {
2762 if (!blk_mq_queue_inflight(md->queue))
2763 break;
2764
2765 if (signal_pending_state(task_state, current)) {
2766 r = -ERESTARTSYS;
2767 break;
2768 }
2769
2770 fsleep(5000);
2771 }
2772
2773 return r;
2774}
2775
2776/*
2777 * Process the deferred bios
2778 */
2779static void dm_wq_work(struct work_struct *work)
2780{
2781 struct mapped_device *md = container_of(work, struct mapped_device, work);
2782 struct bio *bio;
2783
2784 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2785 spin_lock_irq(&md->deferred_lock);
2786 bio = bio_list_pop(&md->deferred);
2787 spin_unlock_irq(&md->deferred_lock);
2788
2789 if (!bio)
2790 break;
2791
2792 submit_bio_noacct(bio);
2793 cond_resched();
2794 }
2795}
2796
2797static void dm_queue_flush(struct mapped_device *md)
2798{
2799 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2800 smp_mb__after_atomic();
2801 queue_work(md->wq, &md->work);
2802}
2803
2804/*
2805 * Swap in a new table, returning the old one for the caller to destroy.
2806 */
2807struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2808{
2809 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2810 struct queue_limits limits;
2811 int r;
2812
2813 mutex_lock(&md->suspend_lock);
2814
2815 /* device must be suspended */
2816 if (!dm_suspended_md(md))
2817 goto out;
2818
2819 /*
2820 * If the new table has no data devices, retain the existing limits.
2821 * This helps multipath with queue_if_no_path if all paths disappear,
2822 * then new I/O is queued based on these limits, and then some paths
2823 * reappear.
2824 */
2825 if (dm_table_has_no_data_devices(table)) {
2826 live_map = dm_get_live_table_fast(md);
2827 if (live_map)
2828 limits = md->queue->limits;
2829 dm_put_live_table_fast(md);
2830 }
2831
2832 if (!live_map) {
2833 r = dm_calculate_queue_limits(table, &limits);
2834 if (r) {
2835 map = ERR_PTR(r);
2836 goto out;
2837 }
2838 }
2839
2840 map = __bind(md, table, &limits);
2841 dm_issue_global_event();
2842
2843out:
2844 mutex_unlock(&md->suspend_lock);
2845 return map;
2846}
2847
2848/*
2849 * Functions to lock and unlock any filesystem running on the
2850 * device.
2851 */
2852static int lock_fs(struct mapped_device *md)
2853{
2854 int r;
2855
2856 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2857
2858 r = bdev_freeze(md->disk->part0);
2859 if (!r)
2860 set_bit(DMF_FROZEN, &md->flags);
2861 return r;
2862}
2863
2864static void unlock_fs(struct mapped_device *md)
2865{
2866 if (!test_bit(DMF_FROZEN, &md->flags))
2867 return;
2868 bdev_thaw(md->disk->part0);
2869 clear_bit(DMF_FROZEN, &md->flags);
2870}
2871
2872/*
2873 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2874 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2875 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2876 *
2877 * If __dm_suspend returns 0, the device is completely quiescent
2878 * now. There is no request-processing activity. All new requests
2879 * are being added to md->deferred list.
2880 */
2881static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2882 unsigned int suspend_flags, unsigned int task_state,
2883 int dmf_suspended_flag)
2884{
2885 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2886 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2887 int r;
2888
2889 lockdep_assert_held(&md->suspend_lock);
2890
2891 /*
2892 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2893 * This flag is cleared before dm_suspend returns.
2894 */
2895 if (noflush)
2896 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2897 else
2898 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2899
2900 /*
2901 * This gets reverted if there's an error later and the targets
2902 * provide the .presuspend_undo hook.
2903 */
2904 dm_table_presuspend_targets(map);
2905
2906 /*
2907 * Flush I/O to the device.
2908 * Any I/O submitted after lock_fs() may not be flushed.
2909 * noflush takes precedence over do_lockfs.
2910 * (lock_fs() flushes I/Os and waits for them to complete.)
2911 */
2912 if (!noflush && do_lockfs) {
2913 r = lock_fs(md);
2914 if (r) {
2915 dm_table_presuspend_undo_targets(map);
2916 return r;
2917 }
2918 }
2919
2920 /*
2921 * Here we must make sure that no processes are submitting requests
2922 * to target drivers i.e. no one may be executing
2923 * dm_split_and_process_bio from dm_submit_bio.
2924 *
2925 * To get all processes out of dm_split_and_process_bio in dm_submit_bio,
2926 * we take the write lock. To prevent any process from reentering
2927 * dm_split_and_process_bio from dm_submit_bio and quiesce the thread
2928 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2929 * flush_workqueue(md->wq).
2930 */
2931 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2932 if (map)
2933 synchronize_srcu(&md->io_barrier);
2934
2935 /*
2936 * Stop md->queue before flushing md->wq in case request-based
2937 * dm defers requests to md->wq from md->queue.
2938 */
2939 if (dm_request_based(md))
2940 dm_stop_queue(md->queue);
2941
2942 flush_workqueue(md->wq);
2943
2944 /*
2945 * At this point no more requests are entering target request routines.
2946 * We call dm_wait_for_completion to wait for all existing requests
2947 * to finish.
2948 */
2949 r = dm_wait_for_completion(md, task_state);
2950 if (!r)
2951 set_bit(dmf_suspended_flag, &md->flags);
2952
2953 if (noflush)
2954 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2955 if (map)
2956 synchronize_srcu(&md->io_barrier);
2957
2958 /* were we interrupted ? */
2959 if (r < 0) {
2960 dm_queue_flush(md);
2961
2962 if (dm_request_based(md))
2963 dm_start_queue(md->queue);
2964
2965 unlock_fs(md);
2966 dm_table_presuspend_undo_targets(map);
2967 /* pushback list is already flushed, so skip flush */
2968 }
2969
2970 return r;
2971}
2972
2973/*
2974 * We need to be able to change a mapping table under a mounted
2975 * filesystem. For example we might want to move some data in
2976 * the background. Before the table can be swapped with
2977 * dm_bind_table, dm_suspend must be called to flush any in
2978 * flight bios and ensure that any further io gets deferred.
2979 */
2980/*
2981 * Suspend mechanism in request-based dm.
2982 *
2983 * 1. Flush all I/Os by lock_fs() if needed.
2984 * 2. Stop dispatching any I/O by stopping the request_queue.
2985 * 3. Wait for all in-flight I/Os to be completed or requeued.
2986 *
2987 * To abort suspend, start the request_queue.
2988 */
2989int dm_suspend(struct mapped_device *md, unsigned int suspend_flags)
2990{
2991 struct dm_table *map = NULL;
2992 int r = 0;
2993
2994retry:
2995 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2996
2997 if (dm_suspended_md(md)) {
2998 r = -EINVAL;
2999 goto out_unlock;
3000 }
3001
3002 if (dm_suspended_internally_md(md)) {
3003 /* already internally suspended, wait for internal resume */
3004 mutex_unlock(&md->suspend_lock);
3005 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3006 if (r)
3007 return r;
3008 goto retry;
3009 }
3010
3011 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3012 if (!map) {
3013 /* avoid deadlock with fs/namespace.c:do_mount() */
3014 suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
3015 }
3016
3017 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
3018 if (r)
3019 goto out_unlock;
3020
3021 set_bit(DMF_POST_SUSPENDING, &md->flags);
3022 dm_table_postsuspend_targets(map);
3023 clear_bit(DMF_POST_SUSPENDING, &md->flags);
3024
3025out_unlock:
3026 mutex_unlock(&md->suspend_lock);
3027 return r;
3028}
3029
3030static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3031{
3032 if (map) {
3033 int r = dm_table_resume_targets(map);
3034
3035 if (r)
3036 return r;
3037 }
3038
3039 dm_queue_flush(md);
3040
3041 /*
3042 * Flushing deferred I/Os must be done after targets are resumed
3043 * so that mapping of targets can work correctly.
3044 * Request-based dm is queueing the deferred I/Os in its request_queue.
3045 */
3046 if (dm_request_based(md))
3047 dm_start_queue(md->queue);
3048
3049 unlock_fs(md);
3050
3051 return 0;
3052}
3053
3054int dm_resume(struct mapped_device *md)
3055{
3056 int r;
3057 struct dm_table *map = NULL;
3058
3059retry:
3060 r = -EINVAL;
3061 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3062
3063 if (!dm_suspended_md(md))
3064 goto out;
3065
3066 if (dm_suspended_internally_md(md)) {
3067 /* already internally suspended, wait for internal resume */
3068 mutex_unlock(&md->suspend_lock);
3069 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3070 if (r)
3071 return r;
3072 goto retry;
3073 }
3074
3075 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3076 if (!map || !dm_table_get_size(map))
3077 goto out;
3078
3079 r = __dm_resume(md, map);
3080 if (r)
3081 goto out;
3082
3083 clear_bit(DMF_SUSPENDED, &md->flags);
3084out:
3085 mutex_unlock(&md->suspend_lock);
3086
3087 return r;
3088}
3089
3090/*
3091 * Internal suspend/resume works like userspace-driven suspend. It waits
3092 * until all bios finish and prevents issuing new bios to the target drivers.
3093 * It may be used only from the kernel.
3094 */
3095
3096static void __dm_internal_suspend(struct mapped_device *md, unsigned int suspend_flags)
3097{
3098 struct dm_table *map = NULL;
3099
3100 lockdep_assert_held(&md->suspend_lock);
3101
3102 if (md->internal_suspend_count++)
3103 return; /* nested internal suspend */
3104
3105 if (dm_suspended_md(md)) {
3106 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3107 return; /* nest suspend */
3108 }
3109
3110 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3111
3112 /*
3113 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3114 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3115 * would require changing .presuspend to return an error -- avoid this
3116 * until there is a need for more elaborate variants of internal suspend.
3117 */
3118 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
3119 DMF_SUSPENDED_INTERNALLY);
3120
3121 set_bit(DMF_POST_SUSPENDING, &md->flags);
3122 dm_table_postsuspend_targets(map);
3123 clear_bit(DMF_POST_SUSPENDING, &md->flags);
3124}
3125
3126static void __dm_internal_resume(struct mapped_device *md)
3127{
3128 int r;
3129 struct dm_table *map;
3130
3131 BUG_ON(!md->internal_suspend_count);
3132
3133 if (--md->internal_suspend_count)
3134 return; /* resume from nested internal suspend */
3135
3136 if (dm_suspended_md(md))
3137 goto done; /* resume from nested suspend */
3138
3139 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3140 r = __dm_resume(md, map);
3141 if (r) {
3142 /*
3143 * If a preresume method of some target failed, we are in a
3144 * tricky situation. We can't return an error to the caller. We
3145 * can't fake success because then the "resume" and
3146 * "postsuspend" methods would not be paired correctly, and it
3147 * would break various targets, for example it would cause list
3148 * corruption in the "origin" target.
3149 *
3150 * So, we fake normal suspend here, to make sure that the
3151 * "resume" and "postsuspend" methods will be paired correctly.
3152 */
3153 DMERR("Preresume method failed: %d", r);
3154 set_bit(DMF_SUSPENDED, &md->flags);
3155 }
3156done:
3157 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3158 smp_mb__after_atomic();
3159 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3160}
3161
3162void dm_internal_suspend_noflush(struct mapped_device *md)
3163{
3164 mutex_lock(&md->suspend_lock);
3165 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3166 mutex_unlock(&md->suspend_lock);
3167}
3168EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3169
3170void dm_internal_resume(struct mapped_device *md)
3171{
3172 mutex_lock(&md->suspend_lock);
3173 __dm_internal_resume(md);
3174 mutex_unlock(&md->suspend_lock);
3175}
3176EXPORT_SYMBOL_GPL(dm_internal_resume);
3177
3178/*
3179 * Fast variants of internal suspend/resume hold md->suspend_lock,
3180 * which prevents interaction with userspace-driven suspend.
3181 */
3182
3183void dm_internal_suspend_fast(struct mapped_device *md)
3184{
3185 mutex_lock(&md->suspend_lock);
3186 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3187 return;
3188
3189 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3190 synchronize_srcu(&md->io_barrier);
3191 flush_workqueue(md->wq);
3192 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3193}
3194EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3195
3196void dm_internal_resume_fast(struct mapped_device *md)
3197{
3198 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3199 goto done;
3200
3201 dm_queue_flush(md);
3202
3203done:
3204 mutex_unlock(&md->suspend_lock);
3205}
3206EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3207
3208/*
3209 *---------------------------------------------------------------
3210 * Event notification.
3211 *---------------------------------------------------------------
3212 */
3213int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3214 unsigned int cookie, bool need_resize_uevent)
3215{
3216 int r;
3217 unsigned int noio_flag;
3218 char udev_cookie[DM_COOKIE_LENGTH];
3219 char *envp[3] = { NULL, NULL, NULL };
3220 char **envpp = envp;
3221 if (cookie) {
3222 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3223 DM_COOKIE_ENV_VAR_NAME, cookie);
3224 *envpp++ = udev_cookie;
3225 }
3226 if (need_resize_uevent) {
3227 *envpp++ = "RESIZE=1";
3228 }
3229
3230 noio_flag = memalloc_noio_save();
3231
3232 r = kobject_uevent_env(&disk_to_dev(md->disk)->kobj, action, envp);
3233
3234 memalloc_noio_restore(noio_flag);
3235
3236 return r;
3237}
3238
3239uint32_t dm_next_uevent_seq(struct mapped_device *md)
3240{
3241 return atomic_add_return(1, &md->uevent_seq);
3242}
3243
3244uint32_t dm_get_event_nr(struct mapped_device *md)
3245{
3246 return atomic_read(&md->event_nr);
3247}
3248
3249int dm_wait_event(struct mapped_device *md, int event_nr)
3250{
3251 return wait_event_interruptible(md->eventq,
3252 (event_nr != atomic_read(&md->event_nr)));
3253}
3254
3255void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3256{
3257 unsigned long flags;
3258
3259 spin_lock_irqsave(&md->uevent_lock, flags);
3260 list_add(elist, &md->uevent_list);
3261 spin_unlock_irqrestore(&md->uevent_lock, flags);
3262}
3263
3264/*
3265 * The gendisk is only valid as long as you have a reference
3266 * count on 'md'.
3267 */
3268struct gendisk *dm_disk(struct mapped_device *md)
3269{
3270 return md->disk;
3271}
3272EXPORT_SYMBOL_GPL(dm_disk);
3273
3274struct kobject *dm_kobject(struct mapped_device *md)
3275{
3276 return &md->kobj_holder.kobj;
3277}
3278
3279struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3280{
3281 struct mapped_device *md;
3282
3283 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3284
3285 spin_lock(&_minor_lock);
3286 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
3287 md = NULL;
3288 goto out;
3289 }
3290 dm_get(md);
3291out:
3292 spin_unlock(&_minor_lock);
3293
3294 return md;
3295}
3296
3297int dm_suspended_md(struct mapped_device *md)
3298{
3299 return test_bit(DMF_SUSPENDED, &md->flags);
3300}
3301
3302static int dm_post_suspending_md(struct mapped_device *md)
3303{
3304 return test_bit(DMF_POST_SUSPENDING, &md->flags);
3305}
3306
3307int dm_suspended_internally_md(struct mapped_device *md)
3308{
3309 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3310}
3311
3312int dm_test_deferred_remove_flag(struct mapped_device *md)
3313{
3314 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3315}
3316
3317int dm_suspended(struct dm_target *ti)
3318{
3319 return dm_suspended_md(ti->table->md);
3320}
3321EXPORT_SYMBOL_GPL(dm_suspended);
3322
3323int dm_post_suspending(struct dm_target *ti)
3324{
3325 return dm_post_suspending_md(ti->table->md);
3326}
3327EXPORT_SYMBOL_GPL(dm_post_suspending);
3328
3329int dm_noflush_suspending(struct dm_target *ti)
3330{
3331 return __noflush_suspending(ti->table->md);
3332}
3333EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3334
3335void dm_free_md_mempools(struct dm_md_mempools *pools)
3336{
3337 if (!pools)
3338 return;
3339
3340 bioset_exit(&pools->bs);
3341 bioset_exit(&pools->io_bs);
3342
3343 kfree(pools);
3344}
3345
3346struct dm_blkdev_id {
3347 u8 *id;
3348 enum blk_unique_id type;
3349};
3350
3351static int __dm_get_unique_id(struct dm_target *ti, struct dm_dev *dev,
3352 sector_t start, sector_t len, void *data)
3353{
3354 struct dm_blkdev_id *dm_id = data;
3355 const struct block_device_operations *fops = dev->bdev->bd_disk->fops;
3356
3357 if (!fops->get_unique_id)
3358 return 0;
3359
3360 return fops->get_unique_id(dev->bdev->bd_disk, dm_id->id, dm_id->type);
3361}
3362
3363/*
3364 * Allow access to get_unique_id() for the first device returning a
3365 * non-zero result. Reasonable use expects all devices to have the
3366 * same unique id.
3367 */
3368static int dm_blk_get_unique_id(struct gendisk *disk, u8 *id,
3369 enum blk_unique_id type)
3370{
3371 struct mapped_device *md = disk->private_data;
3372 struct dm_table *table;
3373 struct dm_target *ti;
3374 int ret = 0, srcu_idx;
3375
3376 struct dm_blkdev_id dm_id = {
3377 .id = id,
3378 .type = type,
3379 };
3380
3381 table = dm_get_live_table(md, &srcu_idx);
3382 if (!table || !dm_table_get_size(table))
3383 goto out;
3384
3385 /* We only support devices that have a single target */
3386 if (table->num_targets != 1)
3387 goto out;
3388 ti = dm_table_get_target(table, 0);
3389
3390 if (!ti->type->iterate_devices)
3391 goto out;
3392
3393 ret = ti->type->iterate_devices(ti, __dm_get_unique_id, &dm_id);
3394out:
3395 dm_put_live_table(md, srcu_idx);
3396 return ret;
3397}
3398
3399struct dm_pr {
3400 u64 old_key;
3401 u64 new_key;
3402 u32 flags;
3403 bool abort;
3404 bool fail_early;
3405 int ret;
3406 enum pr_type type;
3407 struct pr_keys *read_keys;
3408 struct pr_held_reservation *rsv;
3409};
3410
3411static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3412 struct dm_pr *pr)
3413{
3414 struct mapped_device *md = bdev->bd_disk->private_data;
3415 struct dm_table *table;
3416 struct dm_target *ti;
3417 int ret = -ENOTTY, srcu_idx;
3418
3419 table = dm_get_live_table(md, &srcu_idx);
3420 if (!table || !dm_table_get_size(table))
3421 goto out;
3422
3423 /* We only support devices that have a single target */
3424 if (table->num_targets != 1)
3425 goto out;
3426 ti = dm_table_get_target(table, 0);
3427
3428 if (dm_suspended_md(md)) {
3429 ret = -EAGAIN;
3430 goto out;
3431 }
3432
3433 ret = -EINVAL;
3434 if (!ti->type->iterate_devices)
3435 goto out;
3436
3437 ti->type->iterate_devices(ti, fn, pr);
3438 ret = 0;
3439out:
3440 dm_put_live_table(md, srcu_idx);
3441 return ret;
3442}
3443
3444/*
3445 * For register / unregister we need to manually call out to every path.
3446 */
3447static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3448 sector_t start, sector_t len, void *data)
3449{
3450 struct dm_pr *pr = data;
3451 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3452 int ret;
3453
3454 if (!ops || !ops->pr_register) {
3455 pr->ret = -EOPNOTSUPP;
3456 return -1;
3457 }
3458
3459 ret = ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3460 if (!ret)
3461 return 0;
3462
3463 if (!pr->ret)
3464 pr->ret = ret;
3465
3466 if (pr->fail_early)
3467 return -1;
3468
3469 return 0;
3470}
3471
3472static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3473 u32 flags)
3474{
3475 struct dm_pr pr = {
3476 .old_key = old_key,
3477 .new_key = new_key,
3478 .flags = flags,
3479 .fail_early = true,
3480 .ret = 0,
3481 };
3482 int ret;
3483
3484 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3485 if (ret) {
3486 /* Didn't even get to register a path */
3487 return ret;
3488 }
3489
3490 if (!pr.ret)
3491 return 0;
3492 ret = pr.ret;
3493
3494 if (!new_key)
3495 return ret;
3496
3497 /* unregister all paths if we failed to register any path */
3498 pr.old_key = new_key;
3499 pr.new_key = 0;
3500 pr.flags = 0;
3501 pr.fail_early = false;
3502 (void) dm_call_pr(bdev, __dm_pr_register, &pr);
3503 return ret;
3504}
3505
3506
3507static int __dm_pr_reserve(struct dm_target *ti, struct dm_dev *dev,
3508 sector_t start, sector_t len, void *data)
3509{
3510 struct dm_pr *pr = data;
3511 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3512
3513 if (!ops || !ops->pr_reserve) {
3514 pr->ret = -EOPNOTSUPP;
3515 return -1;
3516 }
3517
3518 pr->ret = ops->pr_reserve(dev->bdev, pr->old_key, pr->type, pr->flags);
3519 if (!pr->ret)
3520 return -1;
3521
3522 return 0;
3523}
3524
3525static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3526 u32 flags)
3527{
3528 struct dm_pr pr = {
3529 .old_key = key,
3530 .flags = flags,
3531 .type = type,
3532 .fail_early = false,
3533 .ret = 0,
3534 };
3535 int ret;
3536
3537 ret = dm_call_pr(bdev, __dm_pr_reserve, &pr);
3538 if (ret)
3539 return ret;
3540
3541 return pr.ret;
3542}
3543
3544/*
3545 * If there is a non-All Registrants type of reservation, the release must be
3546 * sent down the holding path. For the cases where there is no reservation or
3547 * the path is not the holder the device will also return success, so we must
3548 * try each path to make sure we got the correct path.
3549 */
3550static int __dm_pr_release(struct dm_target *ti, struct dm_dev *dev,
3551 sector_t start, sector_t len, void *data)
3552{
3553 struct dm_pr *pr = data;
3554 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3555
3556 if (!ops || !ops->pr_release) {
3557 pr->ret = -EOPNOTSUPP;
3558 return -1;
3559 }
3560
3561 pr->ret = ops->pr_release(dev->bdev, pr->old_key, pr->type);
3562 if (pr->ret)
3563 return -1;
3564
3565 return 0;
3566}
3567
3568static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3569{
3570 struct dm_pr pr = {
3571 .old_key = key,
3572 .type = type,
3573 .fail_early = false,
3574 };
3575 int ret;
3576
3577 ret = dm_call_pr(bdev, __dm_pr_release, &pr);
3578 if (ret)
3579 return ret;
3580
3581 return pr.ret;
3582}
3583
3584static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
3585 sector_t start, sector_t len, void *data)
3586{
3587 struct dm_pr *pr = data;
3588 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3589
3590 if (!ops || !ops->pr_preempt) {
3591 pr->ret = -EOPNOTSUPP;
3592 return -1;
3593 }
3594
3595 pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
3596 pr->abort);
3597 if (!pr->ret)
3598 return -1;
3599
3600 return 0;
3601}
3602
3603static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3604 enum pr_type type, bool abort)
3605{
3606 struct dm_pr pr = {
3607 .new_key = new_key,
3608 .old_key = old_key,
3609 .type = type,
3610 .fail_early = false,
3611 };
3612 int ret;
3613
3614 ret = dm_call_pr(bdev, __dm_pr_preempt, &pr);
3615 if (ret)
3616 return ret;
3617
3618 return pr.ret;
3619}
3620
3621static int dm_pr_clear(struct block_device *bdev, u64 key)
3622{
3623 struct mapped_device *md = bdev->bd_disk->private_data;
3624 const struct pr_ops *ops;
3625 int r, srcu_idx;
3626
3627 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3628 if (r < 0)
3629 goto out;
3630
3631 ops = bdev->bd_disk->fops->pr_ops;
3632 if (ops && ops->pr_clear)
3633 r = ops->pr_clear(bdev, key);
3634 else
3635 r = -EOPNOTSUPP;
3636out:
3637 dm_unprepare_ioctl(md, srcu_idx);
3638 return r;
3639}
3640
3641static int __dm_pr_read_keys(struct dm_target *ti, struct dm_dev *dev,
3642 sector_t start, sector_t len, void *data)
3643{
3644 struct dm_pr *pr = data;
3645 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3646
3647 if (!ops || !ops->pr_read_keys) {
3648 pr->ret = -EOPNOTSUPP;
3649 return -1;
3650 }
3651
3652 pr->ret = ops->pr_read_keys(dev->bdev, pr->read_keys);
3653 if (!pr->ret)
3654 return -1;
3655
3656 return 0;
3657}
3658
3659static int dm_pr_read_keys(struct block_device *bdev, struct pr_keys *keys)
3660{
3661 struct dm_pr pr = {
3662 .read_keys = keys,
3663 };
3664 int ret;
3665
3666 ret = dm_call_pr(bdev, __dm_pr_read_keys, &pr);
3667 if (ret)
3668 return ret;
3669
3670 return pr.ret;
3671}
3672
3673static int __dm_pr_read_reservation(struct dm_target *ti, struct dm_dev *dev,
3674 sector_t start, sector_t len, void *data)
3675{
3676 struct dm_pr *pr = data;
3677 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3678
3679 if (!ops || !ops->pr_read_reservation) {
3680 pr->ret = -EOPNOTSUPP;
3681 return -1;
3682 }
3683
3684 pr->ret = ops->pr_read_reservation(dev->bdev, pr->rsv);
3685 if (!pr->ret)
3686 return -1;
3687
3688 return 0;
3689}
3690
3691static int dm_pr_read_reservation(struct block_device *bdev,
3692 struct pr_held_reservation *rsv)
3693{
3694 struct dm_pr pr = {
3695 .rsv = rsv,
3696 };
3697 int ret;
3698
3699 ret = dm_call_pr(bdev, __dm_pr_read_reservation, &pr);
3700 if (ret)
3701 return ret;
3702
3703 return pr.ret;
3704}
3705
3706static const struct pr_ops dm_pr_ops = {
3707 .pr_register = dm_pr_register,
3708 .pr_reserve = dm_pr_reserve,
3709 .pr_release = dm_pr_release,
3710 .pr_preempt = dm_pr_preempt,
3711 .pr_clear = dm_pr_clear,
3712 .pr_read_keys = dm_pr_read_keys,
3713 .pr_read_reservation = dm_pr_read_reservation,
3714};
3715
3716static const struct block_device_operations dm_blk_dops = {
3717 .submit_bio = dm_submit_bio,
3718 .poll_bio = dm_poll_bio,
3719 .open = dm_blk_open,
3720 .release = dm_blk_close,
3721 .ioctl = dm_blk_ioctl,
3722 .getgeo = dm_blk_getgeo,
3723 .report_zones = dm_blk_report_zones,
3724 .get_unique_id = dm_blk_get_unique_id,
3725 .pr_ops = &dm_pr_ops,
3726 .owner = THIS_MODULE
3727};
3728
3729static const struct block_device_operations dm_rq_blk_dops = {
3730 .open = dm_blk_open,
3731 .release = dm_blk_close,
3732 .ioctl = dm_blk_ioctl,
3733 .getgeo = dm_blk_getgeo,
3734 .get_unique_id = dm_blk_get_unique_id,
3735 .pr_ops = &dm_pr_ops,
3736 .owner = THIS_MODULE
3737};
3738
3739static const struct dax_operations dm_dax_ops = {
3740 .direct_access = dm_dax_direct_access,
3741 .zero_page_range = dm_dax_zero_page_range,
3742 .recovery_write = dm_dax_recovery_write,
3743};
3744
3745/*
3746 * module hooks
3747 */
3748module_init(dm_init);
3749module_exit(dm_exit);
3750
3751module_param(major, uint, 0);
3752MODULE_PARM_DESC(major, "The major number of the device mapper");
3753
3754module_param(reserved_bio_based_ios, uint, 0644);
3755MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3756
3757module_param(dm_numa_node, int, 0644);
3758MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3759
3760module_param(swap_bios, int, 0644);
3761MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3762
3763MODULE_DESCRIPTION(DM_NAME " driver");
3764MODULE_AUTHOR("Joe Thornber <dm-devel@lists.linux.dev>");
3765MODULE_LICENSE("GPL");
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 file *bdev_file;
730 struct block_device *bdev;
731 u64 part_off;
732 int r;
733
734 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
735 if (!td)
736 return ERR_PTR(-ENOMEM);
737 refcount_set(&td->count, 1);
738
739 bdev_file = bdev_file_open_by_dev(dev, mode, _dm_claim_ptr, NULL);
740 if (IS_ERR(bdev_file)) {
741 r = PTR_ERR(bdev_file);
742 goto out_free_td;
743 }
744
745 bdev = file_bdev(bdev_file);
746
747 /*
748 * We can be called before the dm disk is added. In that case we can't
749 * register the holder relation here. It will be done once add_disk was
750 * called.
751 */
752 if (md->disk->slave_dir) {
753 r = bd_link_disk_holder(bdev, md->disk);
754 if (r)
755 goto out_blkdev_put;
756 }
757
758 td->dm_dev.mode = mode;
759 td->dm_dev.bdev = bdev;
760 td->dm_dev.bdev_file = bdev_file;
761 td->dm_dev.dax_dev = fs_dax_get_by_bdev(bdev, &part_off,
762 NULL, NULL);
763 format_dev_t(td->dm_dev.name, dev);
764 list_add(&td->list, &md->table_devices);
765 return td;
766
767out_blkdev_put:
768 __fput_sync(bdev_file);
769out_free_td:
770 kfree(td);
771 return ERR_PTR(r);
772}
773
774/*
775 * Close a table device that we've been using.
776 */
777static void close_table_device(struct table_device *td, struct mapped_device *md)
778{
779 if (md->disk->slave_dir)
780 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
781
782 /* Leverage async fput() if DMF_DEFERRED_REMOVE set */
783 if (unlikely(test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
784 fput(td->dm_dev.bdev_file);
785 else
786 __fput_sync(td->dm_dev.bdev_file);
787
788 put_dax(td->dm_dev.dax_dev);
789 list_del(&td->list);
790 kfree(td);
791}
792
793static struct table_device *find_table_device(struct list_head *l, dev_t dev,
794 blk_mode_t mode)
795{
796 struct table_device *td;
797
798 list_for_each_entry(td, l, list)
799 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
800 return td;
801
802 return NULL;
803}
804
805int dm_get_table_device(struct mapped_device *md, dev_t dev, blk_mode_t mode,
806 struct dm_dev **result)
807{
808 struct table_device *td;
809
810 mutex_lock(&md->table_devices_lock);
811 td = find_table_device(&md->table_devices, dev, mode);
812 if (!td) {
813 td = open_table_device(md, dev, mode);
814 if (IS_ERR(td)) {
815 mutex_unlock(&md->table_devices_lock);
816 return PTR_ERR(td);
817 }
818 } else {
819 refcount_inc(&td->count);
820 }
821 mutex_unlock(&md->table_devices_lock);
822
823 *result = &td->dm_dev;
824 return 0;
825}
826
827void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
828{
829 struct table_device *td = container_of(d, struct table_device, dm_dev);
830
831 mutex_lock(&md->table_devices_lock);
832 if (refcount_dec_and_test(&td->count))
833 close_table_device(td, md);
834 mutex_unlock(&md->table_devices_lock);
835}
836
837/*
838 * Get the geometry associated with a dm device
839 */
840int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
841{
842 *geo = md->geometry;
843
844 return 0;
845}
846
847/*
848 * Set the geometry of a device.
849 */
850int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
851{
852 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
853
854 if (geo->start > sz) {
855 DMERR("Start sector is beyond the geometry limits.");
856 return -EINVAL;
857 }
858
859 md->geometry = *geo;
860
861 return 0;
862}
863
864static int __noflush_suspending(struct mapped_device *md)
865{
866 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
867}
868
869static void dm_requeue_add_io(struct dm_io *io, bool first_stage)
870{
871 struct mapped_device *md = io->md;
872
873 if (first_stage) {
874 struct dm_io *next = md->requeue_list;
875
876 md->requeue_list = io;
877 io->next = next;
878 } else {
879 bio_list_add_head(&md->deferred, io->orig_bio);
880 }
881}
882
883static void dm_kick_requeue(struct mapped_device *md, bool first_stage)
884{
885 if (first_stage)
886 queue_work(md->wq, &md->requeue_work);
887 else
888 queue_work(md->wq, &md->work);
889}
890
891/*
892 * Return true if the dm_io's original bio is requeued.
893 * io->status is updated with error if requeue disallowed.
894 */
895static bool dm_handle_requeue(struct dm_io *io, bool first_stage)
896{
897 struct bio *bio = io->orig_bio;
898 bool handle_requeue = (io->status == BLK_STS_DM_REQUEUE);
899 bool handle_polled_eagain = ((io->status == BLK_STS_AGAIN) &&
900 (bio->bi_opf & REQ_POLLED));
901 struct mapped_device *md = io->md;
902 bool requeued = false;
903
904 if (handle_requeue || handle_polled_eagain) {
905 unsigned long flags;
906
907 if (bio->bi_opf & REQ_POLLED) {
908 /*
909 * Upper layer won't help us poll split bio
910 * (io->orig_bio may only reflect a subset of the
911 * pre-split original) so clear REQ_POLLED.
912 */
913 bio_clear_polled(bio);
914 }
915
916 /*
917 * Target requested pushing back the I/O or
918 * polled IO hit BLK_STS_AGAIN.
919 */
920 spin_lock_irqsave(&md->deferred_lock, flags);
921 if ((__noflush_suspending(md) &&
922 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) ||
923 handle_polled_eagain || first_stage) {
924 dm_requeue_add_io(io, first_stage);
925 requeued = true;
926 } else {
927 /*
928 * noflush suspend was interrupted or this is
929 * a write to a zoned target.
930 */
931 io->status = BLK_STS_IOERR;
932 }
933 spin_unlock_irqrestore(&md->deferred_lock, flags);
934 }
935
936 if (requeued)
937 dm_kick_requeue(md, first_stage);
938
939 return requeued;
940}
941
942static void __dm_io_complete(struct dm_io *io, bool first_stage)
943{
944 struct bio *bio = io->orig_bio;
945 struct mapped_device *md = io->md;
946 blk_status_t io_error;
947 bool requeued;
948
949 requeued = dm_handle_requeue(io, first_stage);
950 if (requeued && first_stage)
951 return;
952
953 io_error = io->status;
954 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
955 dm_end_io_acct(io);
956 else if (!io_error) {
957 /*
958 * Must handle target that DM_MAPIO_SUBMITTED only to
959 * then bio_endio() rather than dm_submit_bio_remap()
960 */
961 __dm_start_io_acct(io);
962 dm_end_io_acct(io);
963 }
964 free_io(io);
965 smp_wmb();
966 this_cpu_dec(*md->pending_io);
967
968 /* nudge anyone waiting on suspend queue */
969 if (unlikely(wq_has_sleeper(&md->wait)))
970 wake_up(&md->wait);
971
972 /* Return early if the original bio was requeued */
973 if (requeued)
974 return;
975
976 if (bio_is_flush_with_data(bio)) {
977 /*
978 * Preflush done for flush with data, reissue
979 * without REQ_PREFLUSH.
980 */
981 bio->bi_opf &= ~REQ_PREFLUSH;
982 queue_io(md, bio);
983 } else {
984 /* done with normal IO or empty flush */
985 if (io_error)
986 bio->bi_status = io_error;
987 bio_endio(bio);
988 }
989}
990
991static void dm_wq_requeue_work(struct work_struct *work)
992{
993 struct mapped_device *md = container_of(work, struct mapped_device,
994 requeue_work);
995 unsigned long flags;
996 struct dm_io *io;
997
998 /* reuse deferred lock to simplify dm_handle_requeue */
999 spin_lock_irqsave(&md->deferred_lock, flags);
1000 io = md->requeue_list;
1001 md->requeue_list = NULL;
1002 spin_unlock_irqrestore(&md->deferred_lock, flags);
1003
1004 while (io) {
1005 struct dm_io *next = io->next;
1006
1007 dm_io_rewind(io, &md->disk->bio_split);
1008
1009 io->next = NULL;
1010 __dm_io_complete(io, false);
1011 io = next;
1012 cond_resched();
1013 }
1014}
1015
1016/*
1017 * Two staged requeue:
1018 *
1019 * 1) io->orig_bio points to the real original bio, and the part mapped to
1020 * this io must be requeued, instead of other parts of the original bio.
1021 *
1022 * 2) io->orig_bio points to new cloned bio which matches the requeued dm_io.
1023 */
1024static void dm_io_complete(struct dm_io *io)
1025{
1026 bool first_requeue;
1027
1028 /*
1029 * Only dm_io that has been split needs two stage requeue, otherwise
1030 * we may run into long bio clone chain during suspend and OOM could
1031 * be triggered.
1032 *
1033 * Also flush data dm_io won't be marked as DM_IO_WAS_SPLIT, so they
1034 * also aren't handled via the first stage requeue.
1035 */
1036 if (dm_io_flagged(io, DM_IO_WAS_SPLIT))
1037 first_requeue = true;
1038 else
1039 first_requeue = false;
1040
1041 __dm_io_complete(io, first_requeue);
1042}
1043
1044/*
1045 * Decrements the number of outstanding ios that a bio has been
1046 * cloned into, completing the original io if necc.
1047 */
1048static inline void __dm_io_dec_pending(struct dm_io *io)
1049{
1050 if (atomic_dec_and_test(&io->io_count))
1051 dm_io_complete(io);
1052}
1053
1054static void dm_io_set_error(struct dm_io *io, blk_status_t error)
1055{
1056 unsigned long flags;
1057
1058 /* Push-back supersedes any I/O errors */
1059 spin_lock_irqsave(&io->lock, flags);
1060 if (!(io->status == BLK_STS_DM_REQUEUE &&
1061 __noflush_suspending(io->md))) {
1062 io->status = error;
1063 }
1064 spin_unlock_irqrestore(&io->lock, flags);
1065}
1066
1067static void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
1068{
1069 if (unlikely(error))
1070 dm_io_set_error(io, error);
1071
1072 __dm_io_dec_pending(io);
1073}
1074
1075/*
1076 * The queue_limits are only valid as long as you have a reference
1077 * count on 'md'. But _not_ imposing verification to avoid atomic_read(),
1078 */
1079static inline struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
1080{
1081 return &md->queue->limits;
1082}
1083
1084void disable_discard(struct mapped_device *md)
1085{
1086 struct queue_limits *limits = dm_get_queue_limits(md);
1087
1088 /* device doesn't really support DISCARD, disable it */
1089 limits->max_discard_sectors = 0;
1090}
1091
1092void disable_write_zeroes(struct mapped_device *md)
1093{
1094 struct queue_limits *limits = dm_get_queue_limits(md);
1095
1096 /* device doesn't really support WRITE ZEROES, disable it */
1097 limits->max_write_zeroes_sectors = 0;
1098}
1099
1100static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
1101{
1102 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
1103}
1104
1105static void clone_endio(struct bio *bio)
1106{
1107 blk_status_t error = bio->bi_status;
1108 struct dm_target_io *tio = clone_to_tio(bio);
1109 struct dm_target *ti = tio->ti;
1110 dm_endio_fn endio = ti->type->end_io;
1111 struct dm_io *io = tio->io;
1112 struct mapped_device *md = io->md;
1113
1114 if (unlikely(error == BLK_STS_TARGET)) {
1115 if (bio_op(bio) == REQ_OP_DISCARD &&
1116 !bdev_max_discard_sectors(bio->bi_bdev))
1117 disable_discard(md);
1118 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
1119 !bdev_write_zeroes_sectors(bio->bi_bdev))
1120 disable_write_zeroes(md);
1121 }
1122
1123 if (static_branch_unlikely(&zoned_enabled) &&
1124 unlikely(bdev_is_zoned(bio->bi_bdev)))
1125 dm_zone_endio(io, bio);
1126
1127 if (endio) {
1128 int r = endio(ti, bio, &error);
1129
1130 switch (r) {
1131 case DM_ENDIO_REQUEUE:
1132 if (static_branch_unlikely(&zoned_enabled)) {
1133 /*
1134 * Requeuing writes to a sequential zone of a zoned
1135 * target will break the sequential write pattern:
1136 * fail such IO.
1137 */
1138 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
1139 error = BLK_STS_IOERR;
1140 else
1141 error = BLK_STS_DM_REQUEUE;
1142 } else
1143 error = BLK_STS_DM_REQUEUE;
1144 fallthrough;
1145 case DM_ENDIO_DONE:
1146 break;
1147 case DM_ENDIO_INCOMPLETE:
1148 /* The target will handle the io */
1149 return;
1150 default:
1151 DMCRIT("unimplemented target endio return value: %d", r);
1152 BUG();
1153 }
1154 }
1155
1156 if (static_branch_unlikely(&swap_bios_enabled) &&
1157 unlikely(swap_bios_limit(ti, bio)))
1158 up(&md->swap_bios_semaphore);
1159
1160 free_tio(bio);
1161 dm_io_dec_pending(io, error);
1162}
1163
1164/*
1165 * Return maximum size of I/O possible at the supplied sector up to the current
1166 * target boundary.
1167 */
1168static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
1169 sector_t target_offset)
1170{
1171 return ti->len - target_offset;
1172}
1173
1174static sector_t __max_io_len(struct dm_target *ti, sector_t sector,
1175 unsigned int max_granularity,
1176 unsigned int max_sectors)
1177{
1178 sector_t target_offset = dm_target_offset(ti, sector);
1179 sector_t len = max_io_len_target_boundary(ti, target_offset);
1180
1181 /*
1182 * Does the target need to split IO even further?
1183 * - varied (per target) IO splitting is a tenet of DM; this
1184 * explains why stacked chunk_sectors based splitting via
1185 * bio_split_to_limits() isn't possible here.
1186 */
1187 if (!max_granularity)
1188 return len;
1189 return min_t(sector_t, len,
1190 min(max_sectors ? : queue_max_sectors(ti->table->md->queue),
1191 blk_chunk_sectors_left(target_offset, max_granularity)));
1192}
1193
1194static inline sector_t max_io_len(struct dm_target *ti, sector_t sector)
1195{
1196 return __max_io_len(ti, sector, ti->max_io_len, 0);
1197}
1198
1199int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1200{
1201 if (len > UINT_MAX) {
1202 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1203 (unsigned long long)len, UINT_MAX);
1204 ti->error = "Maximum size of target IO is too large";
1205 return -EINVAL;
1206 }
1207
1208 ti->max_io_len = (uint32_t) len;
1209
1210 return 0;
1211}
1212EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1213
1214static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1215 sector_t sector, int *srcu_idx)
1216 __acquires(md->io_barrier)
1217{
1218 struct dm_table *map;
1219 struct dm_target *ti;
1220
1221 map = dm_get_live_table(md, srcu_idx);
1222 if (!map)
1223 return NULL;
1224
1225 ti = dm_table_find_target(map, sector);
1226 if (!ti)
1227 return NULL;
1228
1229 return ti;
1230}
1231
1232static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1233 long nr_pages, enum dax_access_mode mode, void **kaddr,
1234 pfn_t *pfn)
1235{
1236 struct mapped_device *md = dax_get_private(dax_dev);
1237 sector_t sector = pgoff * PAGE_SECTORS;
1238 struct dm_target *ti;
1239 long len, ret = -EIO;
1240 int srcu_idx;
1241
1242 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1243
1244 if (!ti)
1245 goto out;
1246 if (!ti->type->direct_access)
1247 goto out;
1248 len = max_io_len(ti, sector) / PAGE_SECTORS;
1249 if (len < 1)
1250 goto out;
1251 nr_pages = min(len, nr_pages);
1252 ret = ti->type->direct_access(ti, pgoff, nr_pages, mode, kaddr, pfn);
1253
1254 out:
1255 dm_put_live_table(md, srcu_idx);
1256
1257 return ret;
1258}
1259
1260static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1261 size_t nr_pages)
1262{
1263 struct mapped_device *md = dax_get_private(dax_dev);
1264 sector_t sector = pgoff * PAGE_SECTORS;
1265 struct dm_target *ti;
1266 int ret = -EIO;
1267 int srcu_idx;
1268
1269 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1270
1271 if (!ti)
1272 goto out;
1273 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1274 /*
1275 * ->zero_page_range() is mandatory dax operation. If we are
1276 * here, something is wrong.
1277 */
1278 goto out;
1279 }
1280 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1281 out:
1282 dm_put_live_table(md, srcu_idx);
1283
1284 return ret;
1285}
1286
1287static size_t dm_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
1288 void *addr, size_t bytes, struct iov_iter *i)
1289{
1290 struct mapped_device *md = dax_get_private(dax_dev);
1291 sector_t sector = pgoff * PAGE_SECTORS;
1292 struct dm_target *ti;
1293 int srcu_idx;
1294 long ret = 0;
1295
1296 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1297 if (!ti || !ti->type->dax_recovery_write)
1298 goto out;
1299
1300 ret = ti->type->dax_recovery_write(ti, pgoff, addr, bytes, i);
1301out:
1302 dm_put_live_table(md, srcu_idx);
1303 return ret;
1304}
1305
1306/*
1307 * A target may call dm_accept_partial_bio only from the map routine. It is
1308 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1309 * operations, REQ_OP_ZONE_APPEND (zone append writes) and any bio serviced by
1310 * __send_duplicate_bios().
1311 *
1312 * dm_accept_partial_bio informs the dm that the target only wants to process
1313 * additional n_sectors sectors of the bio and the rest of the data should be
1314 * sent in a next bio.
1315 *
1316 * A diagram that explains the arithmetics:
1317 * +--------------------+---------------+-------+
1318 * | 1 | 2 | 3 |
1319 * +--------------------+---------------+-------+
1320 *
1321 * <-------------- *tio->len_ptr --------------->
1322 * <----- bio_sectors ----->
1323 * <-- n_sectors -->
1324 *
1325 * Region 1 was already iterated over with bio_advance or similar function.
1326 * (it may be empty if the target doesn't use bio_advance)
1327 * Region 2 is the remaining bio size that the target wants to process.
1328 * (it may be empty if region 1 is non-empty, although there is no reason
1329 * to make it empty)
1330 * The target requires that region 3 is to be sent in the next bio.
1331 *
1332 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1333 * the partially processed part (the sum of regions 1+2) must be the same for all
1334 * copies of the bio.
1335 */
1336void dm_accept_partial_bio(struct bio *bio, unsigned int n_sectors)
1337{
1338 struct dm_target_io *tio = clone_to_tio(bio);
1339 struct dm_io *io = tio->io;
1340 unsigned int bio_sectors = bio_sectors(bio);
1341
1342 BUG_ON(dm_tio_flagged(tio, DM_TIO_IS_DUPLICATE_BIO));
1343 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1344 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1345 BUG_ON(bio_sectors > *tio->len_ptr);
1346 BUG_ON(n_sectors > bio_sectors);
1347
1348 *tio->len_ptr -= bio_sectors - n_sectors;
1349 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1350
1351 /*
1352 * __split_and_process_bio() may have already saved mapped part
1353 * for accounting but it is being reduced so update accordingly.
1354 */
1355 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1356 io->sectors = n_sectors;
1357 io->sector_offset = bio_sectors(io->orig_bio);
1358}
1359EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1360
1361/*
1362 * @clone: clone bio that DM core passed to target's .map function
1363 * @tgt_clone: clone of @clone bio that target needs submitted
1364 *
1365 * Targets should use this interface to submit bios they take
1366 * ownership of when returning DM_MAPIO_SUBMITTED.
1367 *
1368 * Target should also enable ti->accounts_remapped_io
1369 */
1370void dm_submit_bio_remap(struct bio *clone, struct bio *tgt_clone)
1371{
1372 struct dm_target_io *tio = clone_to_tio(clone);
1373 struct dm_io *io = tio->io;
1374
1375 /* establish bio that will get submitted */
1376 if (!tgt_clone)
1377 tgt_clone = clone;
1378
1379 /*
1380 * Account io->origin_bio to DM dev on behalf of target
1381 * that took ownership of IO with DM_MAPIO_SUBMITTED.
1382 */
1383 dm_start_io_acct(io, clone);
1384
1385 trace_block_bio_remap(tgt_clone, disk_devt(io->md->disk),
1386 tio->old_sector);
1387 submit_bio_noacct(tgt_clone);
1388}
1389EXPORT_SYMBOL_GPL(dm_submit_bio_remap);
1390
1391static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1392{
1393 mutex_lock(&md->swap_bios_lock);
1394 while (latch < md->swap_bios) {
1395 cond_resched();
1396 down(&md->swap_bios_semaphore);
1397 md->swap_bios--;
1398 }
1399 while (latch > md->swap_bios) {
1400 cond_resched();
1401 up(&md->swap_bios_semaphore);
1402 md->swap_bios++;
1403 }
1404 mutex_unlock(&md->swap_bios_lock);
1405}
1406
1407static void __map_bio(struct bio *clone)
1408{
1409 struct dm_target_io *tio = clone_to_tio(clone);
1410 struct dm_target *ti = tio->ti;
1411 struct dm_io *io = tio->io;
1412 struct mapped_device *md = io->md;
1413 int r;
1414
1415 clone->bi_end_io = clone_endio;
1416
1417 /*
1418 * Map the clone.
1419 */
1420 tio->old_sector = clone->bi_iter.bi_sector;
1421
1422 if (static_branch_unlikely(&swap_bios_enabled) &&
1423 unlikely(swap_bios_limit(ti, clone))) {
1424 int latch = get_swap_bios();
1425
1426 if (unlikely(latch != md->swap_bios))
1427 __set_swap_bios_limit(md, latch);
1428 down(&md->swap_bios_semaphore);
1429 }
1430
1431 if (static_branch_unlikely(&zoned_enabled)) {
1432 /*
1433 * Check if the IO needs a special mapping due to zone append
1434 * emulation on zoned target. In this case, dm_zone_map_bio()
1435 * calls the target map operation.
1436 */
1437 if (unlikely(dm_emulate_zone_append(md)))
1438 r = dm_zone_map_bio(tio);
1439 else
1440 goto do_map;
1441 } else {
1442do_map:
1443 if (likely(ti->type->map == linear_map))
1444 r = linear_map(ti, clone);
1445 else if (ti->type->map == stripe_map)
1446 r = stripe_map(ti, clone);
1447 else
1448 r = ti->type->map(ti, clone);
1449 }
1450
1451 switch (r) {
1452 case DM_MAPIO_SUBMITTED:
1453 /* target has assumed ownership of this io */
1454 if (!ti->accounts_remapped_io)
1455 dm_start_io_acct(io, clone);
1456 break;
1457 case DM_MAPIO_REMAPPED:
1458 dm_submit_bio_remap(clone, NULL);
1459 break;
1460 case DM_MAPIO_KILL:
1461 case DM_MAPIO_REQUEUE:
1462 if (static_branch_unlikely(&swap_bios_enabled) &&
1463 unlikely(swap_bios_limit(ti, clone)))
1464 up(&md->swap_bios_semaphore);
1465 free_tio(clone);
1466 if (r == DM_MAPIO_KILL)
1467 dm_io_dec_pending(io, BLK_STS_IOERR);
1468 else
1469 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1470 break;
1471 default:
1472 DMCRIT("unimplemented target map return value: %d", r);
1473 BUG();
1474 }
1475}
1476
1477static void setup_split_accounting(struct clone_info *ci, unsigned int len)
1478{
1479 struct dm_io *io = ci->io;
1480
1481 if (ci->sector_count > len) {
1482 /*
1483 * Split needed, save the mapped part for accounting.
1484 * NOTE: dm_accept_partial_bio() will update accordingly.
1485 */
1486 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1487 io->sectors = len;
1488 io->sector_offset = bio_sectors(ci->bio);
1489 }
1490}
1491
1492static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1493 struct dm_target *ti, unsigned int num_bios,
1494 unsigned *len, gfp_t gfp_flag)
1495{
1496 struct bio *bio;
1497 int try = (gfp_flag & GFP_NOWAIT) ? 0 : 1;
1498
1499 for (; try < 2; try++) {
1500 int bio_nr;
1501
1502 if (try && num_bios > 1)
1503 mutex_lock(&ci->io->md->table_devices_lock);
1504 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1505 bio = alloc_tio(ci, ti, bio_nr, len,
1506 try ? GFP_NOIO : GFP_NOWAIT);
1507 if (!bio)
1508 break;
1509
1510 bio_list_add(blist, bio);
1511 }
1512 if (try && num_bios > 1)
1513 mutex_unlock(&ci->io->md->table_devices_lock);
1514 if (bio_nr == num_bios)
1515 return;
1516
1517 while ((bio = bio_list_pop(blist)))
1518 free_tio(bio);
1519 }
1520}
1521
1522static unsigned int __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1523 unsigned int num_bios, unsigned int *len,
1524 gfp_t gfp_flag)
1525{
1526 struct bio_list blist = BIO_EMPTY_LIST;
1527 struct bio *clone;
1528 unsigned int ret = 0;
1529
1530 if (WARN_ON_ONCE(num_bios == 0)) /* num_bios = 0 is a bug in caller */
1531 return 0;
1532
1533 /* dm_accept_partial_bio() is not supported with shared tio->len_ptr */
1534 if (len)
1535 setup_split_accounting(ci, *len);
1536
1537 /*
1538 * Using alloc_multiple_bios(), even if num_bios is 1, to consistently
1539 * support allocating using GFP_NOWAIT with GFP_NOIO fallback.
1540 */
1541 alloc_multiple_bios(&blist, ci, ti, num_bios, len, gfp_flag);
1542 while ((clone = bio_list_pop(&blist))) {
1543 if (num_bios > 1)
1544 dm_tio_set_flag(clone_to_tio(clone), DM_TIO_IS_DUPLICATE_BIO);
1545 __map_bio(clone);
1546 ret += 1;
1547 }
1548
1549 return ret;
1550}
1551
1552static void __send_empty_flush(struct clone_info *ci)
1553{
1554 struct dm_table *t = ci->map;
1555 struct bio flush_bio;
1556
1557 /*
1558 * Use an on-stack bio for this, it's safe since we don't
1559 * need to reference it after submit. It's just used as
1560 * the basis for the clone(s).
1561 */
1562 bio_init(&flush_bio, ci->io->md->disk->part0, NULL, 0,
1563 REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
1564
1565 ci->bio = &flush_bio;
1566 ci->sector_count = 0;
1567 ci->io->tio.clone.bi_iter.bi_size = 0;
1568
1569 for (unsigned int i = 0; i < t->num_targets; i++) {
1570 unsigned int bios;
1571 struct dm_target *ti = dm_table_get_target(t, i);
1572
1573 if (unlikely(ti->num_flush_bios == 0))
1574 continue;
1575
1576 atomic_add(ti->num_flush_bios, &ci->io->io_count);
1577 bios = __send_duplicate_bios(ci, ti, ti->num_flush_bios,
1578 NULL, GFP_NOWAIT);
1579 atomic_sub(ti->num_flush_bios - bios, &ci->io->io_count);
1580 }
1581
1582 /*
1583 * alloc_io() takes one extra reference for submission, so the
1584 * reference won't reach 0 without the following subtraction
1585 */
1586 atomic_sub(1, &ci->io->io_count);
1587
1588 bio_uninit(ci->bio);
1589}
1590
1591static void __send_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1592 unsigned int num_bios, unsigned int max_granularity,
1593 unsigned int max_sectors)
1594{
1595 unsigned int len, bios;
1596
1597 len = min_t(sector_t, ci->sector_count,
1598 __max_io_len(ti, ci->sector, max_granularity, max_sectors));
1599
1600 atomic_add(num_bios, &ci->io->io_count);
1601 bios = __send_duplicate_bios(ci, ti, num_bios, &len, GFP_NOIO);
1602 /*
1603 * alloc_io() takes one extra reference for submission, so the
1604 * reference won't reach 0 without the following (+1) subtraction
1605 */
1606 atomic_sub(num_bios - bios + 1, &ci->io->io_count);
1607
1608 ci->sector += len;
1609 ci->sector_count -= len;
1610}
1611
1612static bool is_abnormal_io(struct bio *bio)
1613{
1614 enum req_op op = bio_op(bio);
1615
1616 if (op != REQ_OP_READ && op != REQ_OP_WRITE && op != REQ_OP_FLUSH) {
1617 switch (op) {
1618 case REQ_OP_DISCARD:
1619 case REQ_OP_SECURE_ERASE:
1620 case REQ_OP_WRITE_ZEROES:
1621 return true;
1622 default:
1623 break;
1624 }
1625 }
1626
1627 return false;
1628}
1629
1630static blk_status_t __process_abnormal_io(struct clone_info *ci,
1631 struct dm_target *ti)
1632{
1633 unsigned int num_bios = 0;
1634 unsigned int max_granularity = 0;
1635 unsigned int max_sectors = 0;
1636 struct queue_limits *limits = dm_get_queue_limits(ti->table->md);
1637
1638 switch (bio_op(ci->bio)) {
1639 case REQ_OP_DISCARD:
1640 num_bios = ti->num_discard_bios;
1641 max_sectors = limits->max_discard_sectors;
1642 if (ti->max_discard_granularity)
1643 max_granularity = max_sectors;
1644 break;
1645 case REQ_OP_SECURE_ERASE:
1646 num_bios = ti->num_secure_erase_bios;
1647 max_sectors = limits->max_secure_erase_sectors;
1648 if (ti->max_secure_erase_granularity)
1649 max_granularity = max_sectors;
1650 break;
1651 case REQ_OP_WRITE_ZEROES:
1652 num_bios = ti->num_write_zeroes_bios;
1653 max_sectors = limits->max_write_zeroes_sectors;
1654 if (ti->max_write_zeroes_granularity)
1655 max_granularity = max_sectors;
1656 break;
1657 default:
1658 break;
1659 }
1660
1661 /*
1662 * Even though the device advertised support for this type of
1663 * request, that does not mean every target supports it, and
1664 * reconfiguration might also have changed that since the
1665 * check was performed.
1666 */
1667 if (unlikely(!num_bios))
1668 return BLK_STS_NOTSUPP;
1669
1670 __send_abnormal_io(ci, ti, num_bios, max_granularity, max_sectors);
1671
1672 return BLK_STS_OK;
1673}
1674
1675/*
1676 * Reuse ->bi_private as dm_io list head for storing all dm_io instances
1677 * associated with this bio, and this bio's bi_private needs to be
1678 * stored in dm_io->data before the reuse.
1679 *
1680 * bio->bi_private is owned by fs or upper layer, so block layer won't
1681 * touch it after splitting. Meantime it won't be changed by anyone after
1682 * bio is submitted. So this reuse is safe.
1683 */
1684static inline struct dm_io **dm_poll_list_head(struct bio *bio)
1685{
1686 return (struct dm_io **)&bio->bi_private;
1687}
1688
1689static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
1690{
1691 struct dm_io **head = dm_poll_list_head(bio);
1692
1693 if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
1694 bio->bi_opf |= REQ_DM_POLL_LIST;
1695 /*
1696 * Save .bi_private into dm_io, so that we can reuse
1697 * .bi_private as dm_io list head for storing dm_io list
1698 */
1699 io->data = bio->bi_private;
1700
1701 /* tell block layer to poll for completion */
1702 bio->bi_cookie = ~BLK_QC_T_NONE;
1703
1704 io->next = NULL;
1705 } else {
1706 /*
1707 * bio recursed due to split, reuse original poll list,
1708 * and save bio->bi_private too.
1709 */
1710 io->data = (*head)->data;
1711 io->next = *head;
1712 }
1713
1714 *head = io;
1715}
1716
1717/*
1718 * Select the correct strategy for processing a non-flush bio.
1719 */
1720static blk_status_t __split_and_process_bio(struct clone_info *ci)
1721{
1722 struct bio *clone;
1723 struct dm_target *ti;
1724 unsigned int len;
1725
1726 ti = dm_table_find_target(ci->map, ci->sector);
1727 if (unlikely(!ti))
1728 return BLK_STS_IOERR;
1729
1730 if (unlikely(ci->is_abnormal_io))
1731 return __process_abnormal_io(ci, ti);
1732
1733 /*
1734 * Only support bio polling for normal IO, and the target io is
1735 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
1736 */
1737 ci->submit_as_polled = !!(ci->bio->bi_opf & REQ_POLLED);
1738
1739 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1740 setup_split_accounting(ci, len);
1741
1742 if (unlikely(ci->bio->bi_opf & REQ_NOWAIT)) {
1743 if (unlikely(!dm_target_supports_nowait(ti->type)))
1744 return BLK_STS_NOTSUPP;
1745
1746 clone = alloc_tio(ci, ti, 0, &len, GFP_NOWAIT);
1747 if (unlikely(!clone))
1748 return BLK_STS_AGAIN;
1749 } else {
1750 clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
1751 }
1752 __map_bio(clone);
1753
1754 ci->sector += len;
1755 ci->sector_count -= len;
1756
1757 return BLK_STS_OK;
1758}
1759
1760static void init_clone_info(struct clone_info *ci, struct dm_io *io,
1761 struct dm_table *map, struct bio *bio, bool is_abnormal)
1762{
1763 ci->map = map;
1764 ci->io = io;
1765 ci->bio = bio;
1766 ci->is_abnormal_io = is_abnormal;
1767 ci->submit_as_polled = false;
1768 ci->sector = bio->bi_iter.bi_sector;
1769 ci->sector_count = bio_sectors(bio);
1770
1771 /* Shouldn't happen but sector_count was being set to 0 so... */
1772 if (static_branch_unlikely(&zoned_enabled) &&
1773 WARN_ON_ONCE(op_is_zone_mgmt(bio_op(bio)) && ci->sector_count))
1774 ci->sector_count = 0;
1775}
1776
1777/*
1778 * Entry point to split a bio into clones and submit them to the targets.
1779 */
1780static void dm_split_and_process_bio(struct mapped_device *md,
1781 struct dm_table *map, struct bio *bio)
1782{
1783 struct clone_info ci;
1784 struct dm_io *io;
1785 blk_status_t error = BLK_STS_OK;
1786 bool is_abnormal;
1787
1788 is_abnormal = is_abnormal_io(bio);
1789 if (unlikely(is_abnormal)) {
1790 /*
1791 * Use bio_split_to_limits() for abnormal IO (e.g. discard, etc)
1792 * otherwise associated queue_limits won't be imposed.
1793 */
1794 bio = bio_split_to_limits(bio);
1795 if (!bio)
1796 return;
1797 }
1798
1799 /* Only support nowait for normal IO */
1800 if (unlikely(bio->bi_opf & REQ_NOWAIT) && !is_abnormal) {
1801 io = alloc_io(md, bio, GFP_NOWAIT);
1802 if (unlikely(!io)) {
1803 /* Unable to do anything without dm_io. */
1804 bio_wouldblock_error(bio);
1805 return;
1806 }
1807 } else {
1808 io = alloc_io(md, bio, GFP_NOIO);
1809 }
1810 init_clone_info(&ci, io, map, bio, is_abnormal);
1811
1812 if (bio->bi_opf & REQ_PREFLUSH) {
1813 __send_empty_flush(&ci);
1814 /* dm_io_complete submits any data associated with flush */
1815 goto out;
1816 }
1817
1818 error = __split_and_process_bio(&ci);
1819 if (error || !ci.sector_count)
1820 goto out;
1821 /*
1822 * Remainder must be passed to submit_bio_noacct() so it gets handled
1823 * *after* bios already submitted have been completely processed.
1824 */
1825 bio_trim(bio, io->sectors, ci.sector_count);
1826 trace_block_split(bio, bio->bi_iter.bi_sector);
1827 bio_inc_remaining(bio);
1828 submit_bio_noacct(bio);
1829out:
1830 /*
1831 * Drop the extra reference count for non-POLLED bio, and hold one
1832 * reference for POLLED bio, which will be released in dm_poll_bio
1833 *
1834 * Add every dm_io instance into the dm_io list head which is stored
1835 * in bio->bi_private, so that dm_poll_bio can poll them all.
1836 */
1837 if (error || !ci.submit_as_polled) {
1838 /*
1839 * In case of submission failure, the extra reference for
1840 * submitting io isn't consumed yet
1841 */
1842 if (error)
1843 atomic_dec(&io->io_count);
1844 dm_io_dec_pending(io, error);
1845 } else
1846 dm_queue_poll_io(bio, io);
1847}
1848
1849static void dm_submit_bio(struct bio *bio)
1850{
1851 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
1852 int srcu_idx;
1853 struct dm_table *map;
1854
1855 map = dm_get_live_table(md, &srcu_idx);
1856
1857 /* If suspended, or map not yet available, queue this IO for later */
1858 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) ||
1859 unlikely(!map)) {
1860 if (bio->bi_opf & REQ_NOWAIT)
1861 bio_wouldblock_error(bio);
1862 else if (bio->bi_opf & REQ_RAHEAD)
1863 bio_io_error(bio);
1864 else
1865 queue_io(md, bio);
1866 goto out;
1867 }
1868
1869 dm_split_and_process_bio(md, map, bio);
1870out:
1871 dm_put_live_table(md, srcu_idx);
1872}
1873
1874static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
1875 unsigned int flags)
1876{
1877 WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
1878
1879 /* don't poll if the mapped io is done */
1880 if (atomic_read(&io->io_count) > 1)
1881 bio_poll(&io->tio.clone, iob, flags);
1882
1883 /* bio_poll holds the last reference */
1884 return atomic_read(&io->io_count) == 1;
1885}
1886
1887static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
1888 unsigned int flags)
1889{
1890 struct dm_io **head = dm_poll_list_head(bio);
1891 struct dm_io *list = *head;
1892 struct dm_io *tmp = NULL;
1893 struct dm_io *curr, *next;
1894
1895 /* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
1896 if (!(bio->bi_opf & REQ_DM_POLL_LIST))
1897 return 0;
1898
1899 WARN_ON_ONCE(!list);
1900
1901 /*
1902 * Restore .bi_private before possibly completing dm_io.
1903 *
1904 * bio_poll() is only possible once @bio has been completely
1905 * submitted via submit_bio_noacct()'s depth-first submission.
1906 * So there is no dm_queue_poll_io() race associated with
1907 * clearing REQ_DM_POLL_LIST here.
1908 */
1909 bio->bi_opf &= ~REQ_DM_POLL_LIST;
1910 bio->bi_private = list->data;
1911
1912 for (curr = list, next = curr->next; curr; curr = next, next =
1913 curr ? curr->next : NULL) {
1914 if (dm_poll_dm_io(curr, iob, flags)) {
1915 /*
1916 * clone_endio() has already occurred, so no
1917 * error handling is needed here.
1918 */
1919 __dm_io_dec_pending(curr);
1920 } else {
1921 curr->next = tmp;
1922 tmp = curr;
1923 }
1924 }
1925
1926 /* Not done? */
1927 if (tmp) {
1928 bio->bi_opf |= REQ_DM_POLL_LIST;
1929 /* Reset bio->bi_private to dm_io list head */
1930 *head = tmp;
1931 return 0;
1932 }
1933 return 1;
1934}
1935
1936/*
1937 *---------------------------------------------------------------
1938 * An IDR is used to keep track of allocated minor numbers.
1939 *---------------------------------------------------------------
1940 */
1941static void free_minor(int minor)
1942{
1943 spin_lock(&_minor_lock);
1944 idr_remove(&_minor_idr, minor);
1945 spin_unlock(&_minor_lock);
1946}
1947
1948/*
1949 * See if the device with a specific minor # is free.
1950 */
1951static int specific_minor(int minor)
1952{
1953 int r;
1954
1955 if (minor >= (1 << MINORBITS))
1956 return -EINVAL;
1957
1958 idr_preload(GFP_KERNEL);
1959 spin_lock(&_minor_lock);
1960
1961 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
1962
1963 spin_unlock(&_minor_lock);
1964 idr_preload_end();
1965 if (r < 0)
1966 return r == -ENOSPC ? -EBUSY : r;
1967 return 0;
1968}
1969
1970static int next_free_minor(int *minor)
1971{
1972 int r;
1973
1974 idr_preload(GFP_KERNEL);
1975 spin_lock(&_minor_lock);
1976
1977 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
1978
1979 spin_unlock(&_minor_lock);
1980 idr_preload_end();
1981 if (r < 0)
1982 return r;
1983 *minor = r;
1984 return 0;
1985}
1986
1987static const struct block_device_operations dm_blk_dops;
1988static const struct block_device_operations dm_rq_blk_dops;
1989static const struct dax_operations dm_dax_ops;
1990
1991static void dm_wq_work(struct work_struct *work);
1992
1993#ifdef CONFIG_BLK_INLINE_ENCRYPTION
1994static void dm_queue_destroy_crypto_profile(struct request_queue *q)
1995{
1996 dm_destroy_crypto_profile(q->crypto_profile);
1997}
1998
1999#else /* CONFIG_BLK_INLINE_ENCRYPTION */
2000
2001static inline void dm_queue_destroy_crypto_profile(struct request_queue *q)
2002{
2003}
2004#endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
2005
2006static void cleanup_mapped_device(struct mapped_device *md)
2007{
2008 if (md->wq)
2009 destroy_workqueue(md->wq);
2010 dm_free_md_mempools(md->mempools);
2011
2012 if (md->dax_dev) {
2013 dax_remove_host(md->disk);
2014 kill_dax(md->dax_dev);
2015 put_dax(md->dax_dev);
2016 md->dax_dev = NULL;
2017 }
2018
2019 dm_cleanup_zoned_dev(md);
2020 if (md->disk) {
2021 spin_lock(&_minor_lock);
2022 md->disk->private_data = NULL;
2023 spin_unlock(&_minor_lock);
2024 if (dm_get_md_type(md) != DM_TYPE_NONE) {
2025 struct table_device *td;
2026
2027 dm_sysfs_exit(md);
2028 list_for_each_entry(td, &md->table_devices, list) {
2029 bd_unlink_disk_holder(td->dm_dev.bdev,
2030 md->disk);
2031 }
2032
2033 /*
2034 * Hold lock to make sure del_gendisk() won't concurrent
2035 * with open/close_table_device().
2036 */
2037 mutex_lock(&md->table_devices_lock);
2038 del_gendisk(md->disk);
2039 mutex_unlock(&md->table_devices_lock);
2040 }
2041 dm_queue_destroy_crypto_profile(md->queue);
2042 put_disk(md->disk);
2043 }
2044
2045 if (md->pending_io) {
2046 free_percpu(md->pending_io);
2047 md->pending_io = NULL;
2048 }
2049
2050 cleanup_srcu_struct(&md->io_barrier);
2051
2052 mutex_destroy(&md->suspend_lock);
2053 mutex_destroy(&md->type_lock);
2054 mutex_destroy(&md->table_devices_lock);
2055 mutex_destroy(&md->swap_bios_lock);
2056
2057 dm_mq_cleanup_mapped_device(md);
2058}
2059
2060/*
2061 * Allocate and initialise a blank device with a given minor.
2062 */
2063static struct mapped_device *alloc_dev(int minor)
2064{
2065 int r, numa_node_id = dm_get_numa_node();
2066 struct dax_device *dax_dev;
2067 struct mapped_device *md;
2068 void *old_md;
2069
2070 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
2071 if (!md) {
2072 DMERR("unable to allocate device, out of memory.");
2073 return NULL;
2074 }
2075
2076 if (!try_module_get(THIS_MODULE))
2077 goto bad_module_get;
2078
2079 /* get a minor number for the dev */
2080 if (minor == DM_ANY_MINOR)
2081 r = next_free_minor(&minor);
2082 else
2083 r = specific_minor(minor);
2084 if (r < 0)
2085 goto bad_minor;
2086
2087 r = init_srcu_struct(&md->io_barrier);
2088 if (r < 0)
2089 goto bad_io_barrier;
2090
2091 md->numa_node_id = numa_node_id;
2092 md->init_tio_pdu = false;
2093 md->type = DM_TYPE_NONE;
2094 mutex_init(&md->suspend_lock);
2095 mutex_init(&md->type_lock);
2096 mutex_init(&md->table_devices_lock);
2097 spin_lock_init(&md->deferred_lock);
2098 atomic_set(&md->holders, 1);
2099 atomic_set(&md->open_count, 0);
2100 atomic_set(&md->event_nr, 0);
2101 atomic_set(&md->uevent_seq, 0);
2102 INIT_LIST_HEAD(&md->uevent_list);
2103 INIT_LIST_HEAD(&md->table_devices);
2104 spin_lock_init(&md->uevent_lock);
2105
2106 /*
2107 * default to bio-based until DM table is loaded and md->type
2108 * established. If request-based table is loaded: blk-mq will
2109 * override accordingly.
2110 */
2111 md->disk = blk_alloc_disk(NULL, md->numa_node_id);
2112 if (IS_ERR(md->disk))
2113 goto bad;
2114 md->queue = md->disk->queue;
2115
2116 init_waitqueue_head(&md->wait);
2117 INIT_WORK(&md->work, dm_wq_work);
2118 INIT_WORK(&md->requeue_work, dm_wq_requeue_work);
2119 init_waitqueue_head(&md->eventq);
2120 init_completion(&md->kobj_holder.completion);
2121
2122 md->requeue_list = NULL;
2123 md->swap_bios = get_swap_bios();
2124 sema_init(&md->swap_bios_semaphore, md->swap_bios);
2125 mutex_init(&md->swap_bios_lock);
2126
2127 md->disk->major = _major;
2128 md->disk->first_minor = minor;
2129 md->disk->minors = 1;
2130 md->disk->flags |= GENHD_FL_NO_PART;
2131 md->disk->fops = &dm_blk_dops;
2132 md->disk->private_data = md;
2133 sprintf(md->disk->disk_name, "dm-%d", minor);
2134
2135 dax_dev = alloc_dax(md, &dm_dax_ops);
2136 if (IS_ERR(dax_dev)) {
2137 if (PTR_ERR(dax_dev) != -EOPNOTSUPP)
2138 goto bad;
2139 } else {
2140 set_dax_nocache(dax_dev);
2141 set_dax_nomc(dax_dev);
2142 md->dax_dev = dax_dev;
2143 if (dax_add_host(dax_dev, md->disk))
2144 goto bad;
2145 }
2146
2147 format_dev_t(md->name, MKDEV(_major, minor));
2148
2149 md->wq = alloc_workqueue("kdmflush/%s", WQ_MEM_RECLAIM, 0, md->name);
2150 if (!md->wq)
2151 goto bad;
2152
2153 md->pending_io = alloc_percpu(unsigned long);
2154 if (!md->pending_io)
2155 goto bad;
2156
2157 r = dm_stats_init(&md->stats);
2158 if (r < 0)
2159 goto bad;
2160
2161 /* Populate the mapping, nobody knows we exist yet */
2162 spin_lock(&_minor_lock);
2163 old_md = idr_replace(&_minor_idr, md, minor);
2164 spin_unlock(&_minor_lock);
2165
2166 BUG_ON(old_md != MINOR_ALLOCED);
2167
2168 return md;
2169
2170bad:
2171 cleanup_mapped_device(md);
2172bad_io_barrier:
2173 free_minor(minor);
2174bad_minor:
2175 module_put(THIS_MODULE);
2176bad_module_get:
2177 kvfree(md);
2178 return NULL;
2179}
2180
2181static void unlock_fs(struct mapped_device *md);
2182
2183static void free_dev(struct mapped_device *md)
2184{
2185 int minor = MINOR(disk_devt(md->disk));
2186
2187 unlock_fs(md);
2188
2189 cleanup_mapped_device(md);
2190
2191 WARN_ON_ONCE(!list_empty(&md->table_devices));
2192 dm_stats_cleanup(&md->stats);
2193 free_minor(minor);
2194
2195 module_put(THIS_MODULE);
2196 kvfree(md);
2197}
2198
2199/*
2200 * Bind a table to the device.
2201 */
2202static void event_callback(void *context)
2203{
2204 unsigned long flags;
2205 LIST_HEAD(uevents);
2206 struct mapped_device *md = context;
2207
2208 spin_lock_irqsave(&md->uevent_lock, flags);
2209 list_splice_init(&md->uevent_list, &uevents);
2210 spin_unlock_irqrestore(&md->uevent_lock, flags);
2211
2212 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2213
2214 atomic_inc(&md->event_nr);
2215 wake_up(&md->eventq);
2216 dm_issue_global_event();
2217}
2218
2219/*
2220 * Returns old map, which caller must destroy.
2221 */
2222static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2223 struct queue_limits *limits)
2224{
2225 struct dm_table *old_map;
2226 sector_t size;
2227 int ret;
2228
2229 lockdep_assert_held(&md->suspend_lock);
2230
2231 size = dm_table_get_size(t);
2232
2233 /*
2234 * Wipe any geometry if the size of the table changed.
2235 */
2236 if (size != dm_get_size(md))
2237 memset(&md->geometry, 0, sizeof(md->geometry));
2238
2239 set_capacity(md->disk, size);
2240
2241 dm_table_event_callback(t, event_callback, md);
2242
2243 if (dm_table_request_based(t)) {
2244 /*
2245 * Leverage the fact that request-based DM targets are
2246 * immutable singletons - used to optimize dm_mq_queue_rq.
2247 */
2248 md->immutable_target = dm_table_get_immutable_target(t);
2249
2250 /*
2251 * There is no need to reload with request-based dm because the
2252 * size of front_pad doesn't change.
2253 *
2254 * Note for future: If you are to reload bioset, prep-ed
2255 * requests in the queue may refer to bio from the old bioset,
2256 * so you must walk through the queue to unprep.
2257 */
2258 if (!md->mempools) {
2259 md->mempools = t->mempools;
2260 t->mempools = NULL;
2261 }
2262 } else {
2263 /*
2264 * The md may already have mempools that need changing.
2265 * If so, reload bioset because front_pad may have changed
2266 * because a different table was loaded.
2267 */
2268 dm_free_md_mempools(md->mempools);
2269 md->mempools = t->mempools;
2270 t->mempools = NULL;
2271 }
2272
2273 ret = dm_table_set_restrictions(t, md->queue, limits);
2274 if (ret) {
2275 old_map = ERR_PTR(ret);
2276 goto out;
2277 }
2278
2279 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2280 rcu_assign_pointer(md->map, (void *)t);
2281 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2282
2283 if (old_map)
2284 dm_sync_table(md);
2285out:
2286 return old_map;
2287}
2288
2289/*
2290 * Returns unbound table for the caller to free.
2291 */
2292static struct dm_table *__unbind(struct mapped_device *md)
2293{
2294 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2295
2296 if (!map)
2297 return NULL;
2298
2299 dm_table_event_callback(map, NULL, NULL);
2300 RCU_INIT_POINTER(md->map, NULL);
2301 dm_sync_table(md);
2302
2303 return map;
2304}
2305
2306/*
2307 * Constructor for a new device.
2308 */
2309int dm_create(int minor, struct mapped_device **result)
2310{
2311 struct mapped_device *md;
2312
2313 md = alloc_dev(minor);
2314 if (!md)
2315 return -ENXIO;
2316
2317 dm_ima_reset_data(md);
2318
2319 *result = md;
2320 return 0;
2321}
2322
2323/*
2324 * Functions to manage md->type.
2325 * All are required to hold md->type_lock.
2326 */
2327void dm_lock_md_type(struct mapped_device *md)
2328{
2329 mutex_lock(&md->type_lock);
2330}
2331
2332void dm_unlock_md_type(struct mapped_device *md)
2333{
2334 mutex_unlock(&md->type_lock);
2335}
2336
2337void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2338{
2339 BUG_ON(!mutex_is_locked(&md->type_lock));
2340 md->type = type;
2341}
2342
2343enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2344{
2345 return md->type;
2346}
2347
2348struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2349{
2350 return md->immutable_target_type;
2351}
2352
2353/*
2354 * Setup the DM device's queue based on md's type
2355 */
2356int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2357{
2358 enum dm_queue_mode type = dm_table_get_type(t);
2359 struct queue_limits limits;
2360 struct table_device *td;
2361 int r;
2362
2363 switch (type) {
2364 case DM_TYPE_REQUEST_BASED:
2365 md->disk->fops = &dm_rq_blk_dops;
2366 r = dm_mq_init_request_queue(md, t);
2367 if (r) {
2368 DMERR("Cannot initialize queue for request-based dm mapped device");
2369 return r;
2370 }
2371 break;
2372 case DM_TYPE_BIO_BASED:
2373 case DM_TYPE_DAX_BIO_BASED:
2374 blk_queue_flag_set(QUEUE_FLAG_IO_STAT, md->queue);
2375 break;
2376 case DM_TYPE_NONE:
2377 WARN_ON_ONCE(true);
2378 break;
2379 }
2380
2381 r = dm_calculate_queue_limits(t, &limits);
2382 if (r) {
2383 DMERR("Cannot calculate initial queue limits");
2384 return r;
2385 }
2386 r = dm_table_set_restrictions(t, md->queue, &limits);
2387 if (r)
2388 return r;
2389
2390 /*
2391 * Hold lock to make sure add_disk() and del_gendisk() won't concurrent
2392 * with open_table_device() and close_table_device().
2393 */
2394 mutex_lock(&md->table_devices_lock);
2395 r = add_disk(md->disk);
2396 mutex_unlock(&md->table_devices_lock);
2397 if (r)
2398 return r;
2399
2400 /*
2401 * Register the holder relationship for devices added before the disk
2402 * was live.
2403 */
2404 list_for_each_entry(td, &md->table_devices, list) {
2405 r = bd_link_disk_holder(td->dm_dev.bdev, md->disk);
2406 if (r)
2407 goto out_undo_holders;
2408 }
2409
2410 r = dm_sysfs_init(md);
2411 if (r)
2412 goto out_undo_holders;
2413
2414 md->type = type;
2415 return 0;
2416
2417out_undo_holders:
2418 list_for_each_entry_continue_reverse(td, &md->table_devices, list)
2419 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
2420 mutex_lock(&md->table_devices_lock);
2421 del_gendisk(md->disk);
2422 mutex_unlock(&md->table_devices_lock);
2423 return r;
2424}
2425
2426struct mapped_device *dm_get_md(dev_t dev)
2427{
2428 struct mapped_device *md;
2429 unsigned int minor = MINOR(dev);
2430
2431 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2432 return NULL;
2433
2434 spin_lock(&_minor_lock);
2435
2436 md = idr_find(&_minor_idr, minor);
2437 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2438 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2439 md = NULL;
2440 goto out;
2441 }
2442 dm_get(md);
2443out:
2444 spin_unlock(&_minor_lock);
2445
2446 return md;
2447}
2448EXPORT_SYMBOL_GPL(dm_get_md);
2449
2450void *dm_get_mdptr(struct mapped_device *md)
2451{
2452 return md->interface_ptr;
2453}
2454
2455void dm_set_mdptr(struct mapped_device *md, void *ptr)
2456{
2457 md->interface_ptr = ptr;
2458}
2459
2460void dm_get(struct mapped_device *md)
2461{
2462 atomic_inc(&md->holders);
2463 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2464}
2465
2466int dm_hold(struct mapped_device *md)
2467{
2468 spin_lock(&_minor_lock);
2469 if (test_bit(DMF_FREEING, &md->flags)) {
2470 spin_unlock(&_minor_lock);
2471 return -EBUSY;
2472 }
2473 dm_get(md);
2474 spin_unlock(&_minor_lock);
2475 return 0;
2476}
2477EXPORT_SYMBOL_GPL(dm_hold);
2478
2479const char *dm_device_name(struct mapped_device *md)
2480{
2481 return md->name;
2482}
2483EXPORT_SYMBOL_GPL(dm_device_name);
2484
2485static void __dm_destroy(struct mapped_device *md, bool wait)
2486{
2487 struct dm_table *map;
2488 int srcu_idx;
2489
2490 might_sleep();
2491
2492 spin_lock(&_minor_lock);
2493 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2494 set_bit(DMF_FREEING, &md->flags);
2495 spin_unlock(&_minor_lock);
2496
2497 blk_mark_disk_dead(md->disk);
2498
2499 /*
2500 * Take suspend_lock so that presuspend and postsuspend methods
2501 * do not race with internal suspend.
2502 */
2503 mutex_lock(&md->suspend_lock);
2504 map = dm_get_live_table(md, &srcu_idx);
2505 if (!dm_suspended_md(md)) {
2506 dm_table_presuspend_targets(map);
2507 set_bit(DMF_SUSPENDED, &md->flags);
2508 set_bit(DMF_POST_SUSPENDING, &md->flags);
2509 dm_table_postsuspend_targets(map);
2510 }
2511 /* dm_put_live_table must be before fsleep, otherwise deadlock is possible */
2512 dm_put_live_table(md, srcu_idx);
2513 mutex_unlock(&md->suspend_lock);
2514
2515 /*
2516 * Rare, but there may be I/O requests still going to complete,
2517 * for example. Wait for all references to disappear.
2518 * No one should increment the reference count of the mapped_device,
2519 * after the mapped_device state becomes DMF_FREEING.
2520 */
2521 if (wait)
2522 while (atomic_read(&md->holders))
2523 fsleep(1000);
2524 else if (atomic_read(&md->holders))
2525 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2526 dm_device_name(md), atomic_read(&md->holders));
2527
2528 dm_table_destroy(__unbind(md));
2529 free_dev(md);
2530}
2531
2532void dm_destroy(struct mapped_device *md)
2533{
2534 __dm_destroy(md, true);
2535}
2536
2537void dm_destroy_immediate(struct mapped_device *md)
2538{
2539 __dm_destroy(md, false);
2540}
2541
2542void dm_put(struct mapped_device *md)
2543{
2544 atomic_dec(&md->holders);
2545}
2546EXPORT_SYMBOL_GPL(dm_put);
2547
2548static bool dm_in_flight_bios(struct mapped_device *md)
2549{
2550 int cpu;
2551 unsigned long sum = 0;
2552
2553 for_each_possible_cpu(cpu)
2554 sum += *per_cpu_ptr(md->pending_io, cpu);
2555
2556 return sum != 0;
2557}
2558
2559static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2560{
2561 int r = 0;
2562 DEFINE_WAIT(wait);
2563
2564 while (true) {
2565 prepare_to_wait(&md->wait, &wait, task_state);
2566
2567 if (!dm_in_flight_bios(md))
2568 break;
2569
2570 if (signal_pending_state(task_state, current)) {
2571 r = -EINTR;
2572 break;
2573 }
2574
2575 io_schedule();
2576 }
2577 finish_wait(&md->wait, &wait);
2578
2579 smp_rmb();
2580
2581 return r;
2582}
2583
2584static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2585{
2586 int r = 0;
2587
2588 if (!queue_is_mq(md->queue))
2589 return dm_wait_for_bios_completion(md, task_state);
2590
2591 while (true) {
2592 if (!blk_mq_queue_inflight(md->queue))
2593 break;
2594
2595 if (signal_pending_state(task_state, current)) {
2596 r = -EINTR;
2597 break;
2598 }
2599
2600 fsleep(5000);
2601 }
2602
2603 return r;
2604}
2605
2606/*
2607 * Process the deferred bios
2608 */
2609static void dm_wq_work(struct work_struct *work)
2610{
2611 struct mapped_device *md = container_of(work, struct mapped_device, work);
2612 struct bio *bio;
2613
2614 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2615 spin_lock_irq(&md->deferred_lock);
2616 bio = bio_list_pop(&md->deferred);
2617 spin_unlock_irq(&md->deferred_lock);
2618
2619 if (!bio)
2620 break;
2621
2622 submit_bio_noacct(bio);
2623 cond_resched();
2624 }
2625}
2626
2627static void dm_queue_flush(struct mapped_device *md)
2628{
2629 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2630 smp_mb__after_atomic();
2631 queue_work(md->wq, &md->work);
2632}
2633
2634/*
2635 * Swap in a new table, returning the old one for the caller to destroy.
2636 */
2637struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2638{
2639 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2640 struct queue_limits limits;
2641 int r;
2642
2643 mutex_lock(&md->suspend_lock);
2644
2645 /* device must be suspended */
2646 if (!dm_suspended_md(md))
2647 goto out;
2648
2649 /*
2650 * If the new table has no data devices, retain the existing limits.
2651 * This helps multipath with queue_if_no_path if all paths disappear,
2652 * then new I/O is queued based on these limits, and then some paths
2653 * reappear.
2654 */
2655 if (dm_table_has_no_data_devices(table)) {
2656 live_map = dm_get_live_table_fast(md);
2657 if (live_map)
2658 limits = md->queue->limits;
2659 dm_put_live_table_fast(md);
2660 }
2661
2662 if (!live_map) {
2663 r = dm_calculate_queue_limits(table, &limits);
2664 if (r) {
2665 map = ERR_PTR(r);
2666 goto out;
2667 }
2668 }
2669
2670 map = __bind(md, table, &limits);
2671 dm_issue_global_event();
2672
2673out:
2674 mutex_unlock(&md->suspend_lock);
2675 return map;
2676}
2677
2678/*
2679 * Functions to lock and unlock any filesystem running on the
2680 * device.
2681 */
2682static int lock_fs(struct mapped_device *md)
2683{
2684 int r;
2685
2686 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2687
2688 r = bdev_freeze(md->disk->part0);
2689 if (!r)
2690 set_bit(DMF_FROZEN, &md->flags);
2691 return r;
2692}
2693
2694static void unlock_fs(struct mapped_device *md)
2695{
2696 if (!test_bit(DMF_FROZEN, &md->flags))
2697 return;
2698 bdev_thaw(md->disk->part0);
2699 clear_bit(DMF_FROZEN, &md->flags);
2700}
2701
2702/*
2703 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2704 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2705 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2706 *
2707 * If __dm_suspend returns 0, the device is completely quiescent
2708 * now. There is no request-processing activity. All new requests
2709 * are being added to md->deferred list.
2710 */
2711static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2712 unsigned int suspend_flags, unsigned int task_state,
2713 int dmf_suspended_flag)
2714{
2715 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2716 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2717 int r;
2718
2719 lockdep_assert_held(&md->suspend_lock);
2720
2721 /*
2722 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2723 * This flag is cleared before dm_suspend returns.
2724 */
2725 if (noflush)
2726 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2727 else
2728 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2729
2730 /*
2731 * This gets reverted if there's an error later and the targets
2732 * provide the .presuspend_undo hook.
2733 */
2734 dm_table_presuspend_targets(map);
2735
2736 /*
2737 * Flush I/O to the device.
2738 * Any I/O submitted after lock_fs() may not be flushed.
2739 * noflush takes precedence over do_lockfs.
2740 * (lock_fs() flushes I/Os and waits for them to complete.)
2741 */
2742 if (!noflush && do_lockfs) {
2743 r = lock_fs(md);
2744 if (r) {
2745 dm_table_presuspend_undo_targets(map);
2746 return r;
2747 }
2748 }
2749
2750 /*
2751 * Here we must make sure that no processes are submitting requests
2752 * to target drivers i.e. no one may be executing
2753 * dm_split_and_process_bio from dm_submit_bio.
2754 *
2755 * To get all processes out of dm_split_and_process_bio in dm_submit_bio,
2756 * we take the write lock. To prevent any process from reentering
2757 * dm_split_and_process_bio from dm_submit_bio and quiesce the thread
2758 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2759 * flush_workqueue(md->wq).
2760 */
2761 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2762 if (map)
2763 synchronize_srcu(&md->io_barrier);
2764
2765 /*
2766 * Stop md->queue before flushing md->wq in case request-based
2767 * dm defers requests to md->wq from md->queue.
2768 */
2769 if (dm_request_based(md))
2770 dm_stop_queue(md->queue);
2771
2772 flush_workqueue(md->wq);
2773
2774 /*
2775 * At this point no more requests are entering target request routines.
2776 * We call dm_wait_for_completion to wait for all existing requests
2777 * to finish.
2778 */
2779 r = dm_wait_for_completion(md, task_state);
2780 if (!r)
2781 set_bit(dmf_suspended_flag, &md->flags);
2782
2783 if (noflush)
2784 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2785 if (map)
2786 synchronize_srcu(&md->io_barrier);
2787
2788 /* were we interrupted ? */
2789 if (r < 0) {
2790 dm_queue_flush(md);
2791
2792 if (dm_request_based(md))
2793 dm_start_queue(md->queue);
2794
2795 unlock_fs(md);
2796 dm_table_presuspend_undo_targets(map);
2797 /* pushback list is already flushed, so skip flush */
2798 }
2799
2800 return r;
2801}
2802
2803/*
2804 * We need to be able to change a mapping table under a mounted
2805 * filesystem. For example we might want to move some data in
2806 * the background. Before the table can be swapped with
2807 * dm_bind_table, dm_suspend must be called to flush any in
2808 * flight bios and ensure that any further io gets deferred.
2809 */
2810/*
2811 * Suspend mechanism in request-based dm.
2812 *
2813 * 1. Flush all I/Os by lock_fs() if needed.
2814 * 2. Stop dispatching any I/O by stopping the request_queue.
2815 * 3. Wait for all in-flight I/Os to be completed or requeued.
2816 *
2817 * To abort suspend, start the request_queue.
2818 */
2819int dm_suspend(struct mapped_device *md, unsigned int suspend_flags)
2820{
2821 struct dm_table *map = NULL;
2822 int r = 0;
2823
2824retry:
2825 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2826
2827 if (dm_suspended_md(md)) {
2828 r = -EINVAL;
2829 goto out_unlock;
2830 }
2831
2832 if (dm_suspended_internally_md(md)) {
2833 /* already internally suspended, wait for internal resume */
2834 mutex_unlock(&md->suspend_lock);
2835 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2836 if (r)
2837 return r;
2838 goto retry;
2839 }
2840
2841 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2842 if (!map) {
2843 /* avoid deadlock with fs/namespace.c:do_mount() */
2844 suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
2845 }
2846
2847 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2848 if (r)
2849 goto out_unlock;
2850
2851 set_bit(DMF_POST_SUSPENDING, &md->flags);
2852 dm_table_postsuspend_targets(map);
2853 clear_bit(DMF_POST_SUSPENDING, &md->flags);
2854
2855out_unlock:
2856 mutex_unlock(&md->suspend_lock);
2857 return r;
2858}
2859
2860static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2861{
2862 if (map) {
2863 int r = dm_table_resume_targets(map);
2864
2865 if (r)
2866 return r;
2867 }
2868
2869 dm_queue_flush(md);
2870
2871 /*
2872 * Flushing deferred I/Os must be done after targets are resumed
2873 * so that mapping of targets can work correctly.
2874 * Request-based dm is queueing the deferred I/Os in its request_queue.
2875 */
2876 if (dm_request_based(md))
2877 dm_start_queue(md->queue);
2878
2879 unlock_fs(md);
2880
2881 return 0;
2882}
2883
2884int dm_resume(struct mapped_device *md)
2885{
2886 int r;
2887 struct dm_table *map = NULL;
2888
2889retry:
2890 r = -EINVAL;
2891 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2892
2893 if (!dm_suspended_md(md))
2894 goto out;
2895
2896 if (dm_suspended_internally_md(md)) {
2897 /* already internally suspended, wait for internal resume */
2898 mutex_unlock(&md->suspend_lock);
2899 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2900 if (r)
2901 return r;
2902 goto retry;
2903 }
2904
2905 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2906 if (!map || !dm_table_get_size(map))
2907 goto out;
2908
2909 r = __dm_resume(md, map);
2910 if (r)
2911 goto out;
2912
2913 clear_bit(DMF_SUSPENDED, &md->flags);
2914out:
2915 mutex_unlock(&md->suspend_lock);
2916
2917 return r;
2918}
2919
2920/*
2921 * Internal suspend/resume works like userspace-driven suspend. It waits
2922 * until all bios finish and prevents issuing new bios to the target drivers.
2923 * It may be used only from the kernel.
2924 */
2925
2926static void __dm_internal_suspend(struct mapped_device *md, unsigned int suspend_flags)
2927{
2928 struct dm_table *map = NULL;
2929
2930 lockdep_assert_held(&md->suspend_lock);
2931
2932 if (md->internal_suspend_count++)
2933 return; /* nested internal suspend */
2934
2935 if (dm_suspended_md(md)) {
2936 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2937 return; /* nest suspend */
2938 }
2939
2940 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2941
2942 /*
2943 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2944 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2945 * would require changing .presuspend to return an error -- avoid this
2946 * until there is a need for more elaborate variants of internal suspend.
2947 */
2948 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2949 DMF_SUSPENDED_INTERNALLY);
2950
2951 set_bit(DMF_POST_SUSPENDING, &md->flags);
2952 dm_table_postsuspend_targets(map);
2953 clear_bit(DMF_POST_SUSPENDING, &md->flags);
2954}
2955
2956static void __dm_internal_resume(struct mapped_device *md)
2957{
2958 int r;
2959 struct dm_table *map;
2960
2961 BUG_ON(!md->internal_suspend_count);
2962
2963 if (--md->internal_suspend_count)
2964 return; /* resume from nested internal suspend */
2965
2966 if (dm_suspended_md(md))
2967 goto done; /* resume from nested suspend */
2968
2969 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2970 r = __dm_resume(md, map);
2971 if (r) {
2972 /*
2973 * If a preresume method of some target failed, we are in a
2974 * tricky situation. We can't return an error to the caller. We
2975 * can't fake success because then the "resume" and
2976 * "postsuspend" methods would not be paired correctly, and it
2977 * would break various targets, for example it would cause list
2978 * corruption in the "origin" target.
2979 *
2980 * So, we fake normal suspend here, to make sure that the
2981 * "resume" and "postsuspend" methods will be paired correctly.
2982 */
2983 DMERR("Preresume method failed: %d", r);
2984 set_bit(DMF_SUSPENDED, &md->flags);
2985 }
2986done:
2987 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2988 smp_mb__after_atomic();
2989 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
2990}
2991
2992void dm_internal_suspend_noflush(struct mapped_device *md)
2993{
2994 mutex_lock(&md->suspend_lock);
2995 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2996 mutex_unlock(&md->suspend_lock);
2997}
2998EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2999
3000void dm_internal_resume(struct mapped_device *md)
3001{
3002 mutex_lock(&md->suspend_lock);
3003 __dm_internal_resume(md);
3004 mutex_unlock(&md->suspend_lock);
3005}
3006EXPORT_SYMBOL_GPL(dm_internal_resume);
3007
3008/*
3009 * Fast variants of internal suspend/resume hold md->suspend_lock,
3010 * which prevents interaction with userspace-driven suspend.
3011 */
3012
3013void dm_internal_suspend_fast(struct mapped_device *md)
3014{
3015 mutex_lock(&md->suspend_lock);
3016 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3017 return;
3018
3019 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3020 synchronize_srcu(&md->io_barrier);
3021 flush_workqueue(md->wq);
3022 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3023}
3024EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3025
3026void dm_internal_resume_fast(struct mapped_device *md)
3027{
3028 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3029 goto done;
3030
3031 dm_queue_flush(md);
3032
3033done:
3034 mutex_unlock(&md->suspend_lock);
3035}
3036EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3037
3038/*
3039 *---------------------------------------------------------------
3040 * Event notification.
3041 *---------------------------------------------------------------
3042 */
3043int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3044 unsigned int cookie, bool need_resize_uevent)
3045{
3046 int r;
3047 unsigned int noio_flag;
3048 char udev_cookie[DM_COOKIE_LENGTH];
3049 char *envp[3] = { NULL, NULL, NULL };
3050 char **envpp = envp;
3051 if (cookie) {
3052 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3053 DM_COOKIE_ENV_VAR_NAME, cookie);
3054 *envpp++ = udev_cookie;
3055 }
3056 if (need_resize_uevent) {
3057 *envpp++ = "RESIZE=1";
3058 }
3059
3060 noio_flag = memalloc_noio_save();
3061
3062 r = kobject_uevent_env(&disk_to_dev(md->disk)->kobj, action, envp);
3063
3064 memalloc_noio_restore(noio_flag);
3065
3066 return r;
3067}
3068
3069uint32_t dm_next_uevent_seq(struct mapped_device *md)
3070{
3071 return atomic_add_return(1, &md->uevent_seq);
3072}
3073
3074uint32_t dm_get_event_nr(struct mapped_device *md)
3075{
3076 return atomic_read(&md->event_nr);
3077}
3078
3079int dm_wait_event(struct mapped_device *md, int event_nr)
3080{
3081 return wait_event_interruptible(md->eventq,
3082 (event_nr != atomic_read(&md->event_nr)));
3083}
3084
3085void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3086{
3087 unsigned long flags;
3088
3089 spin_lock_irqsave(&md->uevent_lock, flags);
3090 list_add(elist, &md->uevent_list);
3091 spin_unlock_irqrestore(&md->uevent_lock, flags);
3092}
3093
3094/*
3095 * The gendisk is only valid as long as you have a reference
3096 * count on 'md'.
3097 */
3098struct gendisk *dm_disk(struct mapped_device *md)
3099{
3100 return md->disk;
3101}
3102EXPORT_SYMBOL_GPL(dm_disk);
3103
3104struct kobject *dm_kobject(struct mapped_device *md)
3105{
3106 return &md->kobj_holder.kobj;
3107}
3108
3109struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3110{
3111 struct mapped_device *md;
3112
3113 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3114
3115 spin_lock(&_minor_lock);
3116 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
3117 md = NULL;
3118 goto out;
3119 }
3120 dm_get(md);
3121out:
3122 spin_unlock(&_minor_lock);
3123
3124 return md;
3125}
3126
3127int dm_suspended_md(struct mapped_device *md)
3128{
3129 return test_bit(DMF_SUSPENDED, &md->flags);
3130}
3131
3132static int dm_post_suspending_md(struct mapped_device *md)
3133{
3134 return test_bit(DMF_POST_SUSPENDING, &md->flags);
3135}
3136
3137int dm_suspended_internally_md(struct mapped_device *md)
3138{
3139 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3140}
3141
3142int dm_test_deferred_remove_flag(struct mapped_device *md)
3143{
3144 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3145}
3146
3147int dm_suspended(struct dm_target *ti)
3148{
3149 return dm_suspended_md(ti->table->md);
3150}
3151EXPORT_SYMBOL_GPL(dm_suspended);
3152
3153int dm_post_suspending(struct dm_target *ti)
3154{
3155 return dm_post_suspending_md(ti->table->md);
3156}
3157EXPORT_SYMBOL_GPL(dm_post_suspending);
3158
3159int dm_noflush_suspending(struct dm_target *ti)
3160{
3161 return __noflush_suspending(ti->table->md);
3162}
3163EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3164
3165void dm_free_md_mempools(struct dm_md_mempools *pools)
3166{
3167 if (!pools)
3168 return;
3169
3170 bioset_exit(&pools->bs);
3171 bioset_exit(&pools->io_bs);
3172
3173 kfree(pools);
3174}
3175
3176struct dm_pr {
3177 u64 old_key;
3178 u64 new_key;
3179 u32 flags;
3180 bool abort;
3181 bool fail_early;
3182 int ret;
3183 enum pr_type type;
3184 struct pr_keys *read_keys;
3185 struct pr_held_reservation *rsv;
3186};
3187
3188static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3189 struct dm_pr *pr)
3190{
3191 struct mapped_device *md = bdev->bd_disk->private_data;
3192 struct dm_table *table;
3193 struct dm_target *ti;
3194 int ret = -ENOTTY, srcu_idx;
3195
3196 table = dm_get_live_table(md, &srcu_idx);
3197 if (!table || !dm_table_get_size(table))
3198 goto out;
3199
3200 /* We only support devices that have a single target */
3201 if (table->num_targets != 1)
3202 goto out;
3203 ti = dm_table_get_target(table, 0);
3204
3205 if (dm_suspended_md(md)) {
3206 ret = -EAGAIN;
3207 goto out;
3208 }
3209
3210 ret = -EINVAL;
3211 if (!ti->type->iterate_devices)
3212 goto out;
3213
3214 ti->type->iterate_devices(ti, fn, pr);
3215 ret = 0;
3216out:
3217 dm_put_live_table(md, srcu_idx);
3218 return ret;
3219}
3220
3221/*
3222 * For register / unregister we need to manually call out to every path.
3223 */
3224static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3225 sector_t start, sector_t len, void *data)
3226{
3227 struct dm_pr *pr = data;
3228 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3229 int ret;
3230
3231 if (!ops || !ops->pr_register) {
3232 pr->ret = -EOPNOTSUPP;
3233 return -1;
3234 }
3235
3236 ret = ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3237 if (!ret)
3238 return 0;
3239
3240 if (!pr->ret)
3241 pr->ret = ret;
3242
3243 if (pr->fail_early)
3244 return -1;
3245
3246 return 0;
3247}
3248
3249static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3250 u32 flags)
3251{
3252 struct dm_pr pr = {
3253 .old_key = old_key,
3254 .new_key = new_key,
3255 .flags = flags,
3256 .fail_early = true,
3257 .ret = 0,
3258 };
3259 int ret;
3260
3261 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3262 if (ret) {
3263 /* Didn't even get to register a path */
3264 return ret;
3265 }
3266
3267 if (!pr.ret)
3268 return 0;
3269 ret = pr.ret;
3270
3271 if (!new_key)
3272 return ret;
3273
3274 /* unregister all paths if we failed to register any path */
3275 pr.old_key = new_key;
3276 pr.new_key = 0;
3277 pr.flags = 0;
3278 pr.fail_early = false;
3279 (void) dm_call_pr(bdev, __dm_pr_register, &pr);
3280 return ret;
3281}
3282
3283
3284static int __dm_pr_reserve(struct dm_target *ti, struct dm_dev *dev,
3285 sector_t start, sector_t len, void *data)
3286{
3287 struct dm_pr *pr = data;
3288 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3289
3290 if (!ops || !ops->pr_reserve) {
3291 pr->ret = -EOPNOTSUPP;
3292 return -1;
3293 }
3294
3295 pr->ret = ops->pr_reserve(dev->bdev, pr->old_key, pr->type, pr->flags);
3296 if (!pr->ret)
3297 return -1;
3298
3299 return 0;
3300}
3301
3302static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3303 u32 flags)
3304{
3305 struct dm_pr pr = {
3306 .old_key = key,
3307 .flags = flags,
3308 .type = type,
3309 .fail_early = false,
3310 .ret = 0,
3311 };
3312 int ret;
3313
3314 ret = dm_call_pr(bdev, __dm_pr_reserve, &pr);
3315 if (ret)
3316 return ret;
3317
3318 return pr.ret;
3319}
3320
3321/*
3322 * If there is a non-All Registrants type of reservation, the release must be
3323 * sent down the holding path. For the cases where there is no reservation or
3324 * the path is not the holder the device will also return success, so we must
3325 * try each path to make sure we got the correct path.
3326 */
3327static int __dm_pr_release(struct dm_target *ti, struct dm_dev *dev,
3328 sector_t start, sector_t len, void *data)
3329{
3330 struct dm_pr *pr = data;
3331 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3332
3333 if (!ops || !ops->pr_release) {
3334 pr->ret = -EOPNOTSUPP;
3335 return -1;
3336 }
3337
3338 pr->ret = ops->pr_release(dev->bdev, pr->old_key, pr->type);
3339 if (pr->ret)
3340 return -1;
3341
3342 return 0;
3343}
3344
3345static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3346{
3347 struct dm_pr pr = {
3348 .old_key = key,
3349 .type = type,
3350 .fail_early = false,
3351 };
3352 int ret;
3353
3354 ret = dm_call_pr(bdev, __dm_pr_release, &pr);
3355 if (ret)
3356 return ret;
3357
3358 return pr.ret;
3359}
3360
3361static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
3362 sector_t start, sector_t len, void *data)
3363{
3364 struct dm_pr *pr = data;
3365 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3366
3367 if (!ops || !ops->pr_preempt) {
3368 pr->ret = -EOPNOTSUPP;
3369 return -1;
3370 }
3371
3372 pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
3373 pr->abort);
3374 if (!pr->ret)
3375 return -1;
3376
3377 return 0;
3378}
3379
3380static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3381 enum pr_type type, bool abort)
3382{
3383 struct dm_pr pr = {
3384 .new_key = new_key,
3385 .old_key = old_key,
3386 .type = type,
3387 .fail_early = false,
3388 };
3389 int ret;
3390
3391 ret = dm_call_pr(bdev, __dm_pr_preempt, &pr);
3392 if (ret)
3393 return ret;
3394
3395 return pr.ret;
3396}
3397
3398static int dm_pr_clear(struct block_device *bdev, u64 key)
3399{
3400 struct mapped_device *md = bdev->bd_disk->private_data;
3401 const struct pr_ops *ops;
3402 int r, srcu_idx;
3403
3404 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3405 if (r < 0)
3406 goto out;
3407
3408 ops = bdev->bd_disk->fops->pr_ops;
3409 if (ops && ops->pr_clear)
3410 r = ops->pr_clear(bdev, key);
3411 else
3412 r = -EOPNOTSUPP;
3413out:
3414 dm_unprepare_ioctl(md, srcu_idx);
3415 return r;
3416}
3417
3418static int __dm_pr_read_keys(struct dm_target *ti, struct dm_dev *dev,
3419 sector_t start, sector_t len, void *data)
3420{
3421 struct dm_pr *pr = data;
3422 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3423
3424 if (!ops || !ops->pr_read_keys) {
3425 pr->ret = -EOPNOTSUPP;
3426 return -1;
3427 }
3428
3429 pr->ret = ops->pr_read_keys(dev->bdev, pr->read_keys);
3430 if (!pr->ret)
3431 return -1;
3432
3433 return 0;
3434}
3435
3436static int dm_pr_read_keys(struct block_device *bdev, struct pr_keys *keys)
3437{
3438 struct dm_pr pr = {
3439 .read_keys = keys,
3440 };
3441 int ret;
3442
3443 ret = dm_call_pr(bdev, __dm_pr_read_keys, &pr);
3444 if (ret)
3445 return ret;
3446
3447 return pr.ret;
3448}
3449
3450static int __dm_pr_read_reservation(struct dm_target *ti, struct dm_dev *dev,
3451 sector_t start, sector_t len, void *data)
3452{
3453 struct dm_pr *pr = data;
3454 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3455
3456 if (!ops || !ops->pr_read_reservation) {
3457 pr->ret = -EOPNOTSUPP;
3458 return -1;
3459 }
3460
3461 pr->ret = ops->pr_read_reservation(dev->bdev, pr->rsv);
3462 if (!pr->ret)
3463 return -1;
3464
3465 return 0;
3466}
3467
3468static int dm_pr_read_reservation(struct block_device *bdev,
3469 struct pr_held_reservation *rsv)
3470{
3471 struct dm_pr pr = {
3472 .rsv = rsv,
3473 };
3474 int ret;
3475
3476 ret = dm_call_pr(bdev, __dm_pr_read_reservation, &pr);
3477 if (ret)
3478 return ret;
3479
3480 return pr.ret;
3481}
3482
3483static const struct pr_ops dm_pr_ops = {
3484 .pr_register = dm_pr_register,
3485 .pr_reserve = dm_pr_reserve,
3486 .pr_release = dm_pr_release,
3487 .pr_preempt = dm_pr_preempt,
3488 .pr_clear = dm_pr_clear,
3489 .pr_read_keys = dm_pr_read_keys,
3490 .pr_read_reservation = dm_pr_read_reservation,
3491};
3492
3493static const struct block_device_operations dm_blk_dops = {
3494 .submit_bio = dm_submit_bio,
3495 .poll_bio = dm_poll_bio,
3496 .open = dm_blk_open,
3497 .release = dm_blk_close,
3498 .ioctl = dm_blk_ioctl,
3499 .getgeo = dm_blk_getgeo,
3500 .report_zones = dm_blk_report_zones,
3501 .pr_ops = &dm_pr_ops,
3502 .owner = THIS_MODULE
3503};
3504
3505static const struct block_device_operations dm_rq_blk_dops = {
3506 .open = dm_blk_open,
3507 .release = dm_blk_close,
3508 .ioctl = dm_blk_ioctl,
3509 .getgeo = dm_blk_getgeo,
3510 .pr_ops = &dm_pr_ops,
3511 .owner = THIS_MODULE
3512};
3513
3514static const struct dax_operations dm_dax_ops = {
3515 .direct_access = dm_dax_direct_access,
3516 .zero_page_range = dm_dax_zero_page_range,
3517 .recovery_write = dm_dax_recovery_write,
3518};
3519
3520/*
3521 * module hooks
3522 */
3523module_init(dm_init);
3524module_exit(dm_exit);
3525
3526module_param(major, uint, 0);
3527MODULE_PARM_DESC(major, "The major number of the device mapper");
3528
3529module_param(reserved_bio_based_ios, uint, 0644);
3530MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3531
3532module_param(dm_numa_node, int, 0644);
3533MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3534
3535module_param(swap_bios, int, 0644);
3536MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3537
3538MODULE_DESCRIPTION(DM_NAME " driver");
3539MODULE_AUTHOR("Joe Thornber <dm-devel@lists.linux.dev>");
3540MODULE_LICENSE("GPL");