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(v: &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(tio: clone_to_tio(clone: bio), bit: 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(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(module_param: &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(wq: 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(wq: deferred_remove_workqueue);
237
238 unregister_blkdev(major: _major, name: _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(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(v: &md->open_count);
331out:
332 spin_unlock(lock: &_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(lock: &_minor_lock);
342
343 md = disk->private_data;
344 if (WARN_ON(!md))
345 goto out;
346
347 if (atomic_dec_and_test(v: &md->open_count) &&
348 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
349 queue_work(wq: deferred_remove_workqueue, work: &deferred_remove_work);
350
351 dm_put(md);
352out:
353 spin_unlock(lock: &_minor_lock);
354}
355
356int dm_open_count(struct mapped_device *md)
357{
358 return atomic_read(v: &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(lock: &_minor_lock);
369
370 if (dm_open_count(md)) {
371 r = -EBUSY;
372 if (mark_deferred)
373 set_bit(DMF_DEFERRED_REMOVE, addr: &md->flags);
374 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
375 r = -EEXIST;
376 else
377 set_bit(DMF_DELETING, addr: &md->flags);
378
379 spin_unlock(lock: &_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(lock: &_minor_lock);
389
390 if (test_bit(DMF_DELETING, &md->flags))
391 r = -EBUSY;
392 else
393 clear_bit(DMF_DEFERRED_REMOVE, addr: &md->flags);
394
395 spin_unlock(lock: &_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(t: 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(t: map, index: 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: *srcu_idx);
439 fsleep(usecs: 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: &srcu_idx, bdev: &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(j: clone_to_tio(clone: 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, bit: DM_IO_BLK_STAT)) {
513 if (!end)
514 bdev_start_io_acct(bdev: bio->bi_bdev, op: bio_op(bio),
515 start_time: io->start_time);
516 else
517 bdev_end_io_acct(bdev: bio->bi_bdev, op: bio_op(bio),
518 sectors: dm_io_sectors(io, bio),
519 start_time: 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(stats: &io->md->stats, bio_data_dir(bio),
532 bi_sector: sector, bi_sectors: dm_io_sectors(io, bio),
533 end, start_time: io->start_time, aux: &io->stats_aux);
534 }
535}
536
537static void __dm_start_io_acct(struct dm_io *io)
538{
539 dm_io_acct(io, end: 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, bit: 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, bit: 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, bit: DM_IO_ACCOUNTED)) {
558 spin_unlock_irqrestore(lock: &io->lock, flags);
559 return;
560 }
561 dm_io_set_flag(io, bit: DM_IO_ACCOUNTED);
562 spin_unlock_irqrestore(lock: &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, end: 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_src: bio, gfp: gfp_mask, bs: &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, bit: 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(v: &io->io_count, i: 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, bit: DM_IO_BLK_STAT);
601
602 if (static_branch_unlikely(&stats_enabled) &&
603 unlikely(dm_stats_used(&md->stats)))
604 dm_stats_record_start(stats: &md->stats, aux: &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, bio_src: ci->bio, gfp: gfp_mask,
628 bs: &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(bio: clone, bdev: md->disk->part0);
650
651 if (len) {
652 clone->bi_iter.bi_size = to_bytes(n: *len);
653 if (bio_integrity(bio: 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(tio: clone_to_tio(clone), bit: 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(bl: &md->deferred, bio);
676 spin_unlock_irqrestore(lock: &md->deferred_lock, flags);
677 queue_work(wq: md->wq, work: &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(ssp: &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(ssp: &md->io_barrier, idx: srcu_idx);
697}
698
699void dm_sync_table(struct mapped_device *md)
700{
701 synchronize_srcu(ssp: &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(size: sizeof(*td), GFP_KERNEL, node: md->numa_node_id);
735 if (!td)
736 return ERR_PTR(error: -ENOMEM);
737 refcount_set(r: &td->count, n: 1);
738
739 bdev_file = bdev_file_open_by_dev(dev, mode, holder: _dm_claim_ptr, NULL);
740 if (IS_ERR(ptr: bdev_file)) {
741 r = PTR_ERR(ptr: 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, disk: 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, start_off: &part_off,
762 NULL, NULL);
763 format_dev_t(td->dm_dev.name, dev);
764 list_add(new: &td->list, head: &md->table_devices);
765 return td;
766
767out_blkdev_put:
768 fput(bdev_file);
769out_free_td:
770 kfree(objp: td);
771 return ERR_PTR(error: 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(bdev: td->dm_dev.bdev, disk: md->disk);
781 fput(td->dm_dev.bdev_file);
782 put_dax(dax_dev: td->dm_dev.dax_dev);
783 list_del(entry: &td->list);
784 kfree(objp: td);
785}
786
787static struct table_device *find_table_device(struct list_head *l, dev_t dev,
788 blk_mode_t mode)
789{
790 struct table_device *td;
791
792 list_for_each_entry(td, l, list)
793 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
794 return td;
795
796 return NULL;
797}
798
799int dm_get_table_device(struct mapped_device *md, dev_t dev, blk_mode_t mode,
800 struct dm_dev **result)
801{
802 struct table_device *td;
803
804 mutex_lock(&md->table_devices_lock);
805 td = find_table_device(l: &md->table_devices, dev, mode);
806 if (!td) {
807 td = open_table_device(md, dev, mode);
808 if (IS_ERR(ptr: td)) {
809 mutex_unlock(lock: &md->table_devices_lock);
810 return PTR_ERR(ptr: td);
811 }
812 } else {
813 refcount_inc(r: &td->count);
814 }
815 mutex_unlock(lock: &md->table_devices_lock);
816
817 *result = &td->dm_dev;
818 return 0;
819}
820
821void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
822{
823 struct table_device *td = container_of(d, struct table_device, dm_dev);
824
825 mutex_lock(&md->table_devices_lock);
826 if (refcount_dec_and_test(r: &td->count))
827 close_table_device(td, md);
828 mutex_unlock(lock: &md->table_devices_lock);
829}
830
831/*
832 * Get the geometry associated with a dm device
833 */
834int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
835{
836 *geo = md->geometry;
837
838 return 0;
839}
840
841/*
842 * Set the geometry of a device.
843 */
844int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
845{
846 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
847
848 if (geo->start > sz) {
849 DMERR("Start sector is beyond the geometry limits.");
850 return -EINVAL;
851 }
852
853 md->geometry = *geo;
854
855 return 0;
856}
857
858static int __noflush_suspending(struct mapped_device *md)
859{
860 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
861}
862
863static void dm_requeue_add_io(struct dm_io *io, bool first_stage)
864{
865 struct mapped_device *md = io->md;
866
867 if (first_stage) {
868 struct dm_io *next = md->requeue_list;
869
870 md->requeue_list = io;
871 io->next = next;
872 } else {
873 bio_list_add_head(bl: &md->deferred, bio: io->orig_bio);
874 }
875}
876
877static void dm_kick_requeue(struct mapped_device *md, bool first_stage)
878{
879 if (first_stage)
880 queue_work(wq: md->wq, work: &md->requeue_work);
881 else
882 queue_work(wq: md->wq, work: &md->work);
883}
884
885/*
886 * Return true if the dm_io's original bio is requeued.
887 * io->status is updated with error if requeue disallowed.
888 */
889static bool dm_handle_requeue(struct dm_io *io, bool first_stage)
890{
891 struct bio *bio = io->orig_bio;
892 bool handle_requeue = (io->status == BLK_STS_DM_REQUEUE);
893 bool handle_polled_eagain = ((io->status == BLK_STS_AGAIN) &&
894 (bio->bi_opf & REQ_POLLED));
895 struct mapped_device *md = io->md;
896 bool requeued = false;
897
898 if (handle_requeue || handle_polled_eagain) {
899 unsigned long flags;
900
901 if (bio->bi_opf & REQ_POLLED) {
902 /*
903 * Upper layer won't help us poll split bio
904 * (io->orig_bio may only reflect a subset of the
905 * pre-split original) so clear REQ_POLLED.
906 */
907 bio_clear_polled(bio);
908 }
909
910 /*
911 * Target requested pushing back the I/O or
912 * polled IO hit BLK_STS_AGAIN.
913 */
914 spin_lock_irqsave(&md->deferred_lock, flags);
915 if ((__noflush_suspending(md) &&
916 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) ||
917 handle_polled_eagain || first_stage) {
918 dm_requeue_add_io(io, first_stage);
919 requeued = true;
920 } else {
921 /*
922 * noflush suspend was interrupted or this is
923 * a write to a zoned target.
924 */
925 io->status = BLK_STS_IOERR;
926 }
927 spin_unlock_irqrestore(lock: &md->deferred_lock, flags);
928 }
929
930 if (requeued)
931 dm_kick_requeue(md, first_stage);
932
933 return requeued;
934}
935
936static void __dm_io_complete(struct dm_io *io, bool first_stage)
937{
938 struct bio *bio = io->orig_bio;
939 struct mapped_device *md = io->md;
940 blk_status_t io_error;
941 bool requeued;
942
943 requeued = dm_handle_requeue(io, first_stage);
944 if (requeued && first_stage)
945 return;
946
947 io_error = io->status;
948 if (dm_io_flagged(io, bit: DM_IO_ACCOUNTED))
949 dm_end_io_acct(io);
950 else if (!io_error) {
951 /*
952 * Must handle target that DM_MAPIO_SUBMITTED only to
953 * then bio_endio() rather than dm_submit_bio_remap()
954 */
955 __dm_start_io_acct(io);
956 dm_end_io_acct(io);
957 }
958 free_io(io);
959 smp_wmb();
960 this_cpu_dec(*md->pending_io);
961
962 /* nudge anyone waiting on suspend queue */
963 if (unlikely(wq_has_sleeper(&md->wait)))
964 wake_up(&md->wait);
965
966 /* Return early if the original bio was requeued */
967 if (requeued)
968 return;
969
970 if (bio_is_flush_with_data(bio)) {
971 /*
972 * Preflush done for flush with data, reissue
973 * without REQ_PREFLUSH.
974 */
975 bio->bi_opf &= ~REQ_PREFLUSH;
976 queue_io(md, bio);
977 } else {
978 /* done with normal IO or empty flush */
979 if (io_error)
980 bio->bi_status = io_error;
981 bio_endio(bio);
982 }
983}
984
985static void dm_wq_requeue_work(struct work_struct *work)
986{
987 struct mapped_device *md = container_of(work, struct mapped_device,
988 requeue_work);
989 unsigned long flags;
990 struct dm_io *io;
991
992 /* reuse deferred lock to simplify dm_handle_requeue */
993 spin_lock_irqsave(&md->deferred_lock, flags);
994 io = md->requeue_list;
995 md->requeue_list = NULL;
996 spin_unlock_irqrestore(lock: &md->deferred_lock, flags);
997
998 while (io) {
999 struct dm_io *next = io->next;
1000
1001 dm_io_rewind(io, bs: &md->disk->bio_split);
1002
1003 io->next = NULL;
1004 __dm_io_complete(io, first_stage: false);
1005 io = next;
1006 cond_resched();
1007 }
1008}
1009
1010/*
1011 * Two staged requeue:
1012 *
1013 * 1) io->orig_bio points to the real original bio, and the part mapped to
1014 * this io must be requeued, instead of other parts of the original bio.
1015 *
1016 * 2) io->orig_bio points to new cloned bio which matches the requeued dm_io.
1017 */
1018static void dm_io_complete(struct dm_io *io)
1019{
1020 bool first_requeue;
1021
1022 /*
1023 * Only dm_io that has been split needs two stage requeue, otherwise
1024 * we may run into long bio clone chain during suspend and OOM could
1025 * be triggered.
1026 *
1027 * Also flush data dm_io won't be marked as DM_IO_WAS_SPLIT, so they
1028 * also aren't handled via the first stage requeue.
1029 */
1030 if (dm_io_flagged(io, bit: DM_IO_WAS_SPLIT))
1031 first_requeue = true;
1032 else
1033 first_requeue = false;
1034
1035 __dm_io_complete(io, first_stage: first_requeue);
1036}
1037
1038/*
1039 * Decrements the number of outstanding ios that a bio has been
1040 * cloned into, completing the original io if necc.
1041 */
1042static inline void __dm_io_dec_pending(struct dm_io *io)
1043{
1044 if (atomic_dec_and_test(v: &io->io_count))
1045 dm_io_complete(io);
1046}
1047
1048static void dm_io_set_error(struct dm_io *io, blk_status_t error)
1049{
1050 unsigned long flags;
1051
1052 /* Push-back supersedes any I/O errors */
1053 spin_lock_irqsave(&io->lock, flags);
1054 if (!(io->status == BLK_STS_DM_REQUEUE &&
1055 __noflush_suspending(md: io->md))) {
1056 io->status = error;
1057 }
1058 spin_unlock_irqrestore(lock: &io->lock, flags);
1059}
1060
1061static void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
1062{
1063 if (unlikely(error))
1064 dm_io_set_error(io, error);
1065
1066 __dm_io_dec_pending(io);
1067}
1068
1069/*
1070 * The queue_limits are only valid as long as you have a reference
1071 * count on 'md'. But _not_ imposing verification to avoid atomic_read(),
1072 */
1073static inline struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
1074{
1075 return &md->queue->limits;
1076}
1077
1078void disable_discard(struct mapped_device *md)
1079{
1080 struct queue_limits *limits = dm_get_queue_limits(md);
1081
1082 /* device doesn't really support DISCARD, disable it */
1083 limits->max_discard_sectors = 0;
1084}
1085
1086void disable_write_zeroes(struct mapped_device *md)
1087{
1088 struct queue_limits *limits = dm_get_queue_limits(md);
1089
1090 /* device doesn't really support WRITE ZEROES, disable it */
1091 limits->max_write_zeroes_sectors = 0;
1092}
1093
1094static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
1095{
1096 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
1097}
1098
1099static void clone_endio(struct bio *bio)
1100{
1101 blk_status_t error = bio->bi_status;
1102 struct dm_target_io *tio = clone_to_tio(clone: bio);
1103 struct dm_target *ti = tio->ti;
1104 dm_endio_fn endio = ti->type->end_io;
1105 struct dm_io *io = tio->io;
1106 struct mapped_device *md = io->md;
1107
1108 if (unlikely(error == BLK_STS_TARGET)) {
1109 if (bio_op(bio) == REQ_OP_DISCARD &&
1110 !bdev_max_discard_sectors(bdev: bio->bi_bdev))
1111 disable_discard(md);
1112 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
1113 !bdev_write_zeroes_sectors(bdev: bio->bi_bdev))
1114 disable_write_zeroes(md);
1115 }
1116
1117 if (static_branch_unlikely(&zoned_enabled) &&
1118 unlikely(bdev_is_zoned(bio->bi_bdev)))
1119 dm_zone_endio(io, clone: bio);
1120
1121 if (endio) {
1122 int r = endio(ti, bio, &error);
1123
1124 switch (r) {
1125 case DM_ENDIO_REQUEUE:
1126 if (static_branch_unlikely(&zoned_enabled)) {
1127 /*
1128 * Requeuing writes to a sequential zone of a zoned
1129 * target will break the sequential write pattern:
1130 * fail such IO.
1131 */
1132 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
1133 error = BLK_STS_IOERR;
1134 else
1135 error = BLK_STS_DM_REQUEUE;
1136 } else
1137 error = BLK_STS_DM_REQUEUE;
1138 fallthrough;
1139 case DM_ENDIO_DONE:
1140 break;
1141 case DM_ENDIO_INCOMPLETE:
1142 /* The target will handle the io */
1143 return;
1144 default:
1145 DMCRIT("unimplemented target endio return value: %d", r);
1146 BUG();
1147 }
1148 }
1149
1150 if (static_branch_unlikely(&swap_bios_enabled) &&
1151 unlikely(swap_bios_limit(ti, bio)))
1152 up(sem: &md->swap_bios_semaphore);
1153
1154 free_tio(clone: bio);
1155 dm_io_dec_pending(io, error);
1156}
1157
1158/*
1159 * Return maximum size of I/O possible at the supplied sector up to the current
1160 * target boundary.
1161 */
1162static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
1163 sector_t target_offset)
1164{
1165 return ti->len - target_offset;
1166}
1167
1168static sector_t __max_io_len(struct dm_target *ti, sector_t sector,
1169 unsigned int max_granularity,
1170 unsigned int max_sectors)
1171{
1172 sector_t target_offset = dm_target_offset(ti, sector);
1173 sector_t len = max_io_len_target_boundary(ti, target_offset);
1174
1175 /*
1176 * Does the target need to split IO even further?
1177 * - varied (per target) IO splitting is a tenet of DM; this
1178 * explains why stacked chunk_sectors based splitting via
1179 * bio_split_to_limits() isn't possible here.
1180 */
1181 if (!max_granularity)
1182 return len;
1183 return min_t(sector_t, len,
1184 min(max_sectors ? : queue_max_sectors(ti->table->md->queue),
1185 blk_chunk_sectors_left(target_offset, max_granularity)));
1186}
1187
1188static inline sector_t max_io_len(struct dm_target *ti, sector_t sector)
1189{
1190 return __max_io_len(ti, sector, max_granularity: ti->max_io_len, max_sectors: 0);
1191}
1192
1193int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1194{
1195 if (len > UINT_MAX) {
1196 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1197 (unsigned long long)len, UINT_MAX);
1198 ti->error = "Maximum size of target IO is too large";
1199 return -EINVAL;
1200 }
1201
1202 ti->max_io_len = (uint32_t) len;
1203
1204 return 0;
1205}
1206EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1207
1208static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1209 sector_t sector, int *srcu_idx)
1210 __acquires(md->io_barrier)
1211{
1212 struct dm_table *map;
1213 struct dm_target *ti;
1214
1215 map = dm_get_live_table(md, srcu_idx);
1216 if (!map)
1217 return NULL;
1218
1219 ti = dm_table_find_target(t: map, sector);
1220 if (!ti)
1221 return NULL;
1222
1223 return ti;
1224}
1225
1226static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1227 long nr_pages, enum dax_access_mode mode, void **kaddr,
1228 pfn_t *pfn)
1229{
1230 struct mapped_device *md = dax_get_private(dax_dev);
1231 sector_t sector = pgoff * PAGE_SECTORS;
1232 struct dm_target *ti;
1233 long len, ret = -EIO;
1234 int srcu_idx;
1235
1236 ti = dm_dax_get_live_target(md, sector, srcu_idx: &srcu_idx);
1237
1238 if (!ti)
1239 goto out;
1240 if (!ti->type->direct_access)
1241 goto out;
1242 len = max_io_len(ti, sector) / PAGE_SECTORS;
1243 if (len < 1)
1244 goto out;
1245 nr_pages = min(len, nr_pages);
1246 ret = ti->type->direct_access(ti, pgoff, nr_pages, mode, kaddr, pfn);
1247
1248 out:
1249 dm_put_live_table(md, srcu_idx);
1250
1251 return ret;
1252}
1253
1254static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1255 size_t nr_pages)
1256{
1257 struct mapped_device *md = dax_get_private(dax_dev);
1258 sector_t sector = pgoff * PAGE_SECTORS;
1259 struct dm_target *ti;
1260 int ret = -EIO;
1261 int srcu_idx;
1262
1263 ti = dm_dax_get_live_target(md, sector, srcu_idx: &srcu_idx);
1264
1265 if (!ti)
1266 goto out;
1267 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1268 /*
1269 * ->zero_page_range() is mandatory dax operation. If we are
1270 * here, something is wrong.
1271 */
1272 goto out;
1273 }
1274 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1275 out:
1276 dm_put_live_table(md, srcu_idx);
1277
1278 return ret;
1279}
1280
1281static size_t dm_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
1282 void *addr, size_t bytes, struct iov_iter *i)
1283{
1284 struct mapped_device *md = dax_get_private(dax_dev);
1285 sector_t sector = pgoff * PAGE_SECTORS;
1286 struct dm_target *ti;
1287 int srcu_idx;
1288 long ret = 0;
1289
1290 ti = dm_dax_get_live_target(md, sector, srcu_idx: &srcu_idx);
1291 if (!ti || !ti->type->dax_recovery_write)
1292 goto out;
1293
1294 ret = ti->type->dax_recovery_write(ti, pgoff, addr, bytes, i);
1295out:
1296 dm_put_live_table(md, srcu_idx);
1297 return ret;
1298}
1299
1300/*
1301 * A target may call dm_accept_partial_bio only from the map routine. It is
1302 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1303 * operations, REQ_OP_ZONE_APPEND (zone append writes) and any bio serviced by
1304 * __send_duplicate_bios().
1305 *
1306 * dm_accept_partial_bio informs the dm that the target only wants to process
1307 * additional n_sectors sectors of the bio and the rest of the data should be
1308 * sent in a next bio.
1309 *
1310 * A diagram that explains the arithmetics:
1311 * +--------------------+---------------+-------+
1312 * | 1 | 2 | 3 |
1313 * +--------------------+---------------+-------+
1314 *
1315 * <-------------- *tio->len_ptr --------------->
1316 * <----- bio_sectors ----->
1317 * <-- n_sectors -->
1318 *
1319 * Region 1 was already iterated over with bio_advance or similar function.
1320 * (it may be empty if the target doesn't use bio_advance)
1321 * Region 2 is the remaining bio size that the target wants to process.
1322 * (it may be empty if region 1 is non-empty, although there is no reason
1323 * to make it empty)
1324 * The target requires that region 3 is to be sent in the next bio.
1325 *
1326 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1327 * the partially processed part (the sum of regions 1+2) must be the same for all
1328 * copies of the bio.
1329 */
1330void dm_accept_partial_bio(struct bio *bio, unsigned int n_sectors)
1331{
1332 struct dm_target_io *tio = clone_to_tio(clone: bio);
1333 struct dm_io *io = tio->io;
1334 unsigned int bio_sectors = bio_sectors(bio);
1335
1336 BUG_ON(dm_tio_flagged(tio, DM_TIO_IS_DUPLICATE_BIO));
1337 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1338 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1339 BUG_ON(bio_sectors > *tio->len_ptr);
1340 BUG_ON(n_sectors > bio_sectors);
1341
1342 *tio->len_ptr -= bio_sectors - n_sectors;
1343 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1344
1345 /*
1346 * __split_and_process_bio() may have already saved mapped part
1347 * for accounting but it is being reduced so update accordingly.
1348 */
1349 dm_io_set_flag(io, bit: DM_IO_WAS_SPLIT);
1350 io->sectors = n_sectors;
1351 io->sector_offset = bio_sectors(io->orig_bio);
1352}
1353EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1354
1355/*
1356 * @clone: clone bio that DM core passed to target's .map function
1357 * @tgt_clone: clone of @clone bio that target needs submitted
1358 *
1359 * Targets should use this interface to submit bios they take
1360 * ownership of when returning DM_MAPIO_SUBMITTED.
1361 *
1362 * Target should also enable ti->accounts_remapped_io
1363 */
1364void dm_submit_bio_remap(struct bio *clone, struct bio *tgt_clone)
1365{
1366 struct dm_target_io *tio = clone_to_tio(clone);
1367 struct dm_io *io = tio->io;
1368
1369 /* establish bio that will get submitted */
1370 if (!tgt_clone)
1371 tgt_clone = clone;
1372
1373 /*
1374 * Account io->origin_bio to DM dev on behalf of target
1375 * that took ownership of IO with DM_MAPIO_SUBMITTED.
1376 */
1377 dm_start_io_acct(io, clone);
1378
1379 trace_block_bio_remap(bio: tgt_clone, dev: disk_devt(disk: io->md->disk),
1380 from: tio->old_sector);
1381 submit_bio_noacct(bio: tgt_clone);
1382}
1383EXPORT_SYMBOL_GPL(dm_submit_bio_remap);
1384
1385static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1386{
1387 mutex_lock(&md->swap_bios_lock);
1388 while (latch < md->swap_bios) {
1389 cond_resched();
1390 down(sem: &md->swap_bios_semaphore);
1391 md->swap_bios--;
1392 }
1393 while (latch > md->swap_bios) {
1394 cond_resched();
1395 up(sem: &md->swap_bios_semaphore);
1396 md->swap_bios++;
1397 }
1398 mutex_unlock(lock: &md->swap_bios_lock);
1399}
1400
1401static void __map_bio(struct bio *clone)
1402{
1403 struct dm_target_io *tio = clone_to_tio(clone);
1404 struct dm_target *ti = tio->ti;
1405 struct dm_io *io = tio->io;
1406 struct mapped_device *md = io->md;
1407 int r;
1408
1409 clone->bi_end_io = clone_endio;
1410
1411 /*
1412 * Map the clone.
1413 */
1414 tio->old_sector = clone->bi_iter.bi_sector;
1415
1416 if (static_branch_unlikely(&swap_bios_enabled) &&
1417 unlikely(swap_bios_limit(ti, clone))) {
1418 int latch = get_swap_bios();
1419
1420 if (unlikely(latch != md->swap_bios))
1421 __set_swap_bios_limit(md, latch);
1422 down(sem: &md->swap_bios_semaphore);
1423 }
1424
1425 if (static_branch_unlikely(&zoned_enabled)) {
1426 /*
1427 * Check if the IO needs a special mapping due to zone append
1428 * emulation on zoned target. In this case, dm_zone_map_bio()
1429 * calls the target map operation.
1430 */
1431 if (unlikely(dm_emulate_zone_append(md)))
1432 r = dm_zone_map_bio(io: tio);
1433 else
1434 goto do_map;
1435 } else {
1436do_map:
1437 if (likely(ti->type->map == linear_map))
1438 r = linear_map(ti, bio: clone);
1439 else if (ti->type->map == stripe_map)
1440 r = stripe_map(ti, bio: clone);
1441 else
1442 r = ti->type->map(ti, clone);
1443 }
1444
1445 switch (r) {
1446 case DM_MAPIO_SUBMITTED:
1447 /* target has assumed ownership of this io */
1448 if (!ti->accounts_remapped_io)
1449 dm_start_io_acct(io, clone);
1450 break;
1451 case DM_MAPIO_REMAPPED:
1452 dm_submit_bio_remap(clone, NULL);
1453 break;
1454 case DM_MAPIO_KILL:
1455 case DM_MAPIO_REQUEUE:
1456 if (static_branch_unlikely(&swap_bios_enabled) &&
1457 unlikely(swap_bios_limit(ti, clone)))
1458 up(sem: &md->swap_bios_semaphore);
1459 free_tio(clone);
1460 if (r == DM_MAPIO_KILL)
1461 dm_io_dec_pending(io, BLK_STS_IOERR);
1462 else
1463 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1464 break;
1465 default:
1466 DMCRIT("unimplemented target map return value: %d", r);
1467 BUG();
1468 }
1469}
1470
1471static void setup_split_accounting(struct clone_info *ci, unsigned int len)
1472{
1473 struct dm_io *io = ci->io;
1474
1475 if (ci->sector_count > len) {
1476 /*
1477 * Split needed, save the mapped part for accounting.
1478 * NOTE: dm_accept_partial_bio() will update accordingly.
1479 */
1480 dm_io_set_flag(io, bit: DM_IO_WAS_SPLIT);
1481 io->sectors = len;
1482 io->sector_offset = bio_sectors(ci->bio);
1483 }
1484}
1485
1486static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1487 struct dm_target *ti, unsigned int num_bios,
1488 unsigned *len, gfp_t gfp_flag)
1489{
1490 struct bio *bio;
1491 int try = (gfp_flag & GFP_NOWAIT) ? 0 : 1;
1492
1493 for (; try < 2; try++) {
1494 int bio_nr;
1495
1496 if (try && num_bios > 1)
1497 mutex_lock(&ci->io->md->table_devices_lock);
1498 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1499 bio = alloc_tio(ci, ti, target_bio_nr: bio_nr, len,
1500 gfp_mask: try ? GFP_NOIO : GFP_NOWAIT);
1501 if (!bio)
1502 break;
1503
1504 bio_list_add(bl: blist, bio);
1505 }
1506 if (try && num_bios > 1)
1507 mutex_unlock(lock: &ci->io->md->table_devices_lock);
1508 if (bio_nr == num_bios)
1509 return;
1510
1511 while ((bio = bio_list_pop(bl: blist)))
1512 free_tio(clone: bio);
1513 }
1514}
1515
1516static unsigned int __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1517 unsigned int num_bios, unsigned int *len,
1518 gfp_t gfp_flag)
1519{
1520 struct bio_list blist = BIO_EMPTY_LIST;
1521 struct bio *clone;
1522 unsigned int ret = 0;
1523
1524 if (WARN_ON_ONCE(num_bios == 0)) /* num_bios = 0 is a bug in caller */
1525 return 0;
1526
1527 /* dm_accept_partial_bio() is not supported with shared tio->len_ptr */
1528 if (len)
1529 setup_split_accounting(ci, len: *len);
1530
1531 /*
1532 * Using alloc_multiple_bios(), even if num_bios is 1, to consistently
1533 * support allocating using GFP_NOWAIT with GFP_NOIO fallback.
1534 */
1535 alloc_multiple_bios(blist: &blist, ci, ti, num_bios, len, gfp_flag);
1536 while ((clone = bio_list_pop(bl: &blist))) {
1537 if (num_bios > 1)
1538 dm_tio_set_flag(tio: clone_to_tio(clone), bit: DM_TIO_IS_DUPLICATE_BIO);
1539 __map_bio(clone);
1540 ret += 1;
1541 }
1542
1543 return ret;
1544}
1545
1546static void __send_empty_flush(struct clone_info *ci)
1547{
1548 struct dm_table *t = ci->map;
1549 struct bio flush_bio;
1550
1551 /*
1552 * Use an on-stack bio for this, it's safe since we don't
1553 * need to reference it after submit. It's just used as
1554 * the basis for the clone(s).
1555 */
1556 bio_init(bio: &flush_bio, bdev: ci->io->md->disk->part0, NULL, max_vecs: 0,
1557 opf: REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
1558
1559 ci->bio = &flush_bio;
1560 ci->sector_count = 0;
1561 ci->io->tio.clone.bi_iter.bi_size = 0;
1562
1563 for (unsigned int i = 0; i < t->num_targets; i++) {
1564 unsigned int bios;
1565 struct dm_target *ti = dm_table_get_target(t, index: i);
1566
1567 if (unlikely(ti->num_flush_bios == 0))
1568 continue;
1569
1570 atomic_add(i: ti->num_flush_bios, v: &ci->io->io_count);
1571 bios = __send_duplicate_bios(ci, ti, num_bios: ti->num_flush_bios,
1572 NULL, GFP_NOWAIT);
1573 atomic_sub(i: ti->num_flush_bios - bios, v: &ci->io->io_count);
1574 }
1575
1576 /*
1577 * alloc_io() takes one extra reference for submission, so the
1578 * reference won't reach 0 without the following subtraction
1579 */
1580 atomic_sub(i: 1, v: &ci->io->io_count);
1581
1582 bio_uninit(ci->bio);
1583}
1584
1585static void __send_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1586 unsigned int num_bios, unsigned int max_granularity,
1587 unsigned int max_sectors)
1588{
1589 unsigned int len, bios;
1590
1591 len = min_t(sector_t, ci->sector_count,
1592 __max_io_len(ti, ci->sector, max_granularity, max_sectors));
1593
1594 atomic_add(i: num_bios, v: &ci->io->io_count);
1595 bios = __send_duplicate_bios(ci, ti, num_bios, len: &len, GFP_NOIO);
1596 /*
1597 * alloc_io() takes one extra reference for submission, so the
1598 * reference won't reach 0 without the following (+1) subtraction
1599 */
1600 atomic_sub(i: num_bios - bios + 1, v: &ci->io->io_count);
1601
1602 ci->sector += len;
1603 ci->sector_count -= len;
1604}
1605
1606static bool is_abnormal_io(struct bio *bio)
1607{
1608 enum req_op op = bio_op(bio);
1609
1610 if (op != REQ_OP_READ && op != REQ_OP_WRITE && op != REQ_OP_FLUSH) {
1611 switch (op) {
1612 case REQ_OP_DISCARD:
1613 case REQ_OP_SECURE_ERASE:
1614 case REQ_OP_WRITE_ZEROES:
1615 return true;
1616 default:
1617 break;
1618 }
1619 }
1620
1621 return false;
1622}
1623
1624static blk_status_t __process_abnormal_io(struct clone_info *ci,
1625 struct dm_target *ti)
1626{
1627 unsigned int num_bios = 0;
1628 unsigned int max_granularity = 0;
1629 unsigned int max_sectors = 0;
1630 struct queue_limits *limits = dm_get_queue_limits(md: ti->table->md);
1631
1632 switch (bio_op(bio: ci->bio)) {
1633 case REQ_OP_DISCARD:
1634 num_bios = ti->num_discard_bios;
1635 max_sectors = limits->max_discard_sectors;
1636 if (ti->max_discard_granularity)
1637 max_granularity = max_sectors;
1638 break;
1639 case REQ_OP_SECURE_ERASE:
1640 num_bios = ti->num_secure_erase_bios;
1641 max_sectors = limits->max_secure_erase_sectors;
1642 if (ti->max_secure_erase_granularity)
1643 max_granularity = max_sectors;
1644 break;
1645 case REQ_OP_WRITE_ZEROES:
1646 num_bios = ti->num_write_zeroes_bios;
1647 max_sectors = limits->max_write_zeroes_sectors;
1648 if (ti->max_write_zeroes_granularity)
1649 max_granularity = max_sectors;
1650 break;
1651 default:
1652 break;
1653 }
1654
1655 /*
1656 * Even though the device advertised support for this type of
1657 * request, that does not mean every target supports it, and
1658 * reconfiguration might also have changed that since the
1659 * check was performed.
1660 */
1661 if (unlikely(!num_bios))
1662 return BLK_STS_NOTSUPP;
1663
1664 __send_abnormal_io(ci, ti, num_bios, max_granularity, max_sectors);
1665
1666 return BLK_STS_OK;
1667}
1668
1669/*
1670 * Reuse ->bi_private as dm_io list head for storing all dm_io instances
1671 * associated with this bio, and this bio's bi_private needs to be
1672 * stored in dm_io->data before the reuse.
1673 *
1674 * bio->bi_private is owned by fs or upper layer, so block layer won't
1675 * touch it after splitting. Meantime it won't be changed by anyone after
1676 * bio is submitted. So this reuse is safe.
1677 */
1678static inline struct dm_io **dm_poll_list_head(struct bio *bio)
1679{
1680 return (struct dm_io **)&bio->bi_private;
1681}
1682
1683static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
1684{
1685 struct dm_io **head = dm_poll_list_head(bio);
1686
1687 if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
1688 bio->bi_opf |= REQ_DM_POLL_LIST;
1689 /*
1690 * Save .bi_private into dm_io, so that we can reuse
1691 * .bi_private as dm_io list head for storing dm_io list
1692 */
1693 io->data = bio->bi_private;
1694
1695 /* tell block layer to poll for completion */
1696 bio->bi_cookie = ~BLK_QC_T_NONE;
1697
1698 io->next = NULL;
1699 } else {
1700 /*
1701 * bio recursed due to split, reuse original poll list,
1702 * and save bio->bi_private too.
1703 */
1704 io->data = (*head)->data;
1705 io->next = *head;
1706 }
1707
1708 *head = io;
1709}
1710
1711/*
1712 * Select the correct strategy for processing a non-flush bio.
1713 */
1714static blk_status_t __split_and_process_bio(struct clone_info *ci)
1715{
1716 struct bio *clone;
1717 struct dm_target *ti;
1718 unsigned int len;
1719
1720 ti = dm_table_find_target(t: ci->map, sector: ci->sector);
1721 if (unlikely(!ti))
1722 return BLK_STS_IOERR;
1723
1724 if (unlikely(ci->is_abnormal_io))
1725 return __process_abnormal_io(ci, ti);
1726
1727 /*
1728 * Only support bio polling for normal IO, and the target io is
1729 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
1730 */
1731 ci->submit_as_polled = !!(ci->bio->bi_opf & REQ_POLLED);
1732
1733 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1734 setup_split_accounting(ci, len);
1735
1736 if (unlikely(ci->bio->bi_opf & REQ_NOWAIT)) {
1737 if (unlikely(!dm_target_supports_nowait(ti->type)))
1738 return BLK_STS_NOTSUPP;
1739
1740 clone = alloc_tio(ci, ti, target_bio_nr: 0, len: &len, GFP_NOWAIT);
1741 if (unlikely(!clone))
1742 return BLK_STS_AGAIN;
1743 } else {
1744 clone = alloc_tio(ci, ti, target_bio_nr: 0, len: &len, GFP_NOIO);
1745 }
1746 __map_bio(clone);
1747
1748 ci->sector += len;
1749 ci->sector_count -= len;
1750
1751 return BLK_STS_OK;
1752}
1753
1754static void init_clone_info(struct clone_info *ci, struct dm_io *io,
1755 struct dm_table *map, struct bio *bio, bool is_abnormal)
1756{
1757 ci->map = map;
1758 ci->io = io;
1759 ci->bio = bio;
1760 ci->is_abnormal_io = is_abnormal;
1761 ci->submit_as_polled = false;
1762 ci->sector = bio->bi_iter.bi_sector;
1763 ci->sector_count = bio_sectors(bio);
1764
1765 /* Shouldn't happen but sector_count was being set to 0 so... */
1766 if (static_branch_unlikely(&zoned_enabled) &&
1767 WARN_ON_ONCE(op_is_zone_mgmt(bio_op(bio)) && ci->sector_count))
1768 ci->sector_count = 0;
1769}
1770
1771/*
1772 * Entry point to split a bio into clones and submit them to the targets.
1773 */
1774static void dm_split_and_process_bio(struct mapped_device *md,
1775 struct dm_table *map, struct bio *bio)
1776{
1777 struct clone_info ci;
1778 struct dm_io *io;
1779 blk_status_t error = BLK_STS_OK;
1780 bool is_abnormal;
1781
1782 is_abnormal = is_abnormal_io(bio);
1783 if (unlikely(is_abnormal)) {
1784 /*
1785 * Use bio_split_to_limits() for abnormal IO (e.g. discard, etc)
1786 * otherwise associated queue_limits won't be imposed.
1787 */
1788 bio = bio_split_to_limits(bio);
1789 if (!bio)
1790 return;
1791 }
1792
1793 /* Only support nowait for normal IO */
1794 if (unlikely(bio->bi_opf & REQ_NOWAIT) && !is_abnormal) {
1795 io = alloc_io(md, bio, GFP_NOWAIT);
1796 if (unlikely(!io)) {
1797 /* Unable to do anything without dm_io. */
1798 bio_wouldblock_error(bio);
1799 return;
1800 }
1801 } else {
1802 io = alloc_io(md, bio, GFP_NOIO);
1803 }
1804 init_clone_info(ci: &ci, io, map, bio, is_abnormal);
1805
1806 if (bio->bi_opf & REQ_PREFLUSH) {
1807 __send_empty_flush(ci: &ci);
1808 /* dm_io_complete submits any data associated with flush */
1809 goto out;
1810 }
1811
1812 error = __split_and_process_bio(ci: &ci);
1813 if (error || !ci.sector_count)
1814 goto out;
1815 /*
1816 * Remainder must be passed to submit_bio_noacct() so it gets handled
1817 * *after* bios already submitted have been completely processed.
1818 */
1819 bio_trim(bio, offset: io->sectors, size: ci.sector_count);
1820 trace_block_split(bio, new_sector: bio->bi_iter.bi_sector);
1821 bio_inc_remaining(bio);
1822 submit_bio_noacct(bio);
1823out:
1824 /*
1825 * Drop the extra reference count for non-POLLED bio, and hold one
1826 * reference for POLLED bio, which will be released in dm_poll_bio
1827 *
1828 * Add every dm_io instance into the dm_io list head which is stored
1829 * in bio->bi_private, so that dm_poll_bio can poll them all.
1830 */
1831 if (error || !ci.submit_as_polled) {
1832 /*
1833 * In case of submission failure, the extra reference for
1834 * submitting io isn't consumed yet
1835 */
1836 if (error)
1837 atomic_dec(v: &io->io_count);
1838 dm_io_dec_pending(io, error);
1839 } else
1840 dm_queue_poll_io(bio, io);
1841}
1842
1843static void dm_submit_bio(struct bio *bio)
1844{
1845 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
1846 int srcu_idx;
1847 struct dm_table *map;
1848
1849 map = dm_get_live_table(md, srcu_idx: &srcu_idx);
1850
1851 /* If suspended, or map not yet available, queue this IO for later */
1852 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) ||
1853 unlikely(!map)) {
1854 if (bio->bi_opf & REQ_NOWAIT)
1855 bio_wouldblock_error(bio);
1856 else if (bio->bi_opf & REQ_RAHEAD)
1857 bio_io_error(bio);
1858 else
1859 queue_io(md, bio);
1860 goto out;
1861 }
1862
1863 dm_split_and_process_bio(md, map, bio);
1864out:
1865 dm_put_live_table(md, srcu_idx);
1866}
1867
1868static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
1869 unsigned int flags)
1870{
1871 WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
1872
1873 /* don't poll if the mapped io is done */
1874 if (atomic_read(v: &io->io_count) > 1)
1875 bio_poll(bio: &io->tio.clone, iob, flags);
1876
1877 /* bio_poll holds the last reference */
1878 return atomic_read(v: &io->io_count) == 1;
1879}
1880
1881static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
1882 unsigned int flags)
1883{
1884 struct dm_io **head = dm_poll_list_head(bio);
1885 struct dm_io *list = *head;
1886 struct dm_io *tmp = NULL;
1887 struct dm_io *curr, *next;
1888
1889 /* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
1890 if (!(bio->bi_opf & REQ_DM_POLL_LIST))
1891 return 0;
1892
1893 WARN_ON_ONCE(!list);
1894
1895 /*
1896 * Restore .bi_private before possibly completing dm_io.
1897 *
1898 * bio_poll() is only possible once @bio has been completely
1899 * submitted via submit_bio_noacct()'s depth-first submission.
1900 * So there is no dm_queue_poll_io() race associated with
1901 * clearing REQ_DM_POLL_LIST here.
1902 */
1903 bio->bi_opf &= ~REQ_DM_POLL_LIST;
1904 bio->bi_private = list->data;
1905
1906 for (curr = list, next = curr->next; curr; curr = next, next =
1907 curr ? curr->next : NULL) {
1908 if (dm_poll_dm_io(io: curr, iob, flags)) {
1909 /*
1910 * clone_endio() has already occurred, so no
1911 * error handling is needed here.
1912 */
1913 __dm_io_dec_pending(io: curr);
1914 } else {
1915 curr->next = tmp;
1916 tmp = curr;
1917 }
1918 }
1919
1920 /* Not done? */
1921 if (tmp) {
1922 bio->bi_opf |= REQ_DM_POLL_LIST;
1923 /* Reset bio->bi_private to dm_io list head */
1924 *head = tmp;
1925 return 0;
1926 }
1927 return 1;
1928}
1929
1930/*
1931 *---------------------------------------------------------------
1932 * An IDR is used to keep track of allocated minor numbers.
1933 *---------------------------------------------------------------
1934 */
1935static void free_minor(int minor)
1936{
1937 spin_lock(lock: &_minor_lock);
1938 idr_remove(&_minor_idr, id: minor);
1939 spin_unlock(lock: &_minor_lock);
1940}
1941
1942/*
1943 * See if the device with a specific minor # is free.
1944 */
1945static int specific_minor(int minor)
1946{
1947 int r;
1948
1949 if (minor >= (1 << MINORBITS))
1950 return -EINVAL;
1951
1952 idr_preload(GFP_KERNEL);
1953 spin_lock(lock: &_minor_lock);
1954
1955 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, start: minor, end: minor + 1, GFP_NOWAIT);
1956
1957 spin_unlock(lock: &_minor_lock);
1958 idr_preload_end();
1959 if (r < 0)
1960 return r == -ENOSPC ? -EBUSY : r;
1961 return 0;
1962}
1963
1964static int next_free_minor(int *minor)
1965{
1966 int r;
1967
1968 idr_preload(GFP_KERNEL);
1969 spin_lock(lock: &_minor_lock);
1970
1971 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, start: 0, end: 1 << MINORBITS, GFP_NOWAIT);
1972
1973 spin_unlock(lock: &_minor_lock);
1974 idr_preload_end();
1975 if (r < 0)
1976 return r;
1977 *minor = r;
1978 return 0;
1979}
1980
1981static const struct block_device_operations dm_blk_dops;
1982static const struct block_device_operations dm_rq_blk_dops;
1983static const struct dax_operations dm_dax_ops;
1984
1985static void dm_wq_work(struct work_struct *work);
1986
1987#ifdef CONFIG_BLK_INLINE_ENCRYPTION
1988static void dm_queue_destroy_crypto_profile(struct request_queue *q)
1989{
1990 dm_destroy_crypto_profile(profile: q->crypto_profile);
1991}
1992
1993#else /* CONFIG_BLK_INLINE_ENCRYPTION */
1994
1995static inline void dm_queue_destroy_crypto_profile(struct request_queue *q)
1996{
1997}
1998#endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
1999
2000static void cleanup_mapped_device(struct mapped_device *md)
2001{
2002 if (md->wq)
2003 destroy_workqueue(wq: md->wq);
2004 dm_free_md_mempools(pools: md->mempools);
2005
2006 if (md->dax_dev) {
2007 dax_remove_host(disk: md->disk);
2008 kill_dax(dax_dev: md->dax_dev);
2009 put_dax(dax_dev: md->dax_dev);
2010 md->dax_dev = NULL;
2011 }
2012
2013 dm_cleanup_zoned_dev(md);
2014 if (md->disk) {
2015 spin_lock(lock: &_minor_lock);
2016 md->disk->private_data = NULL;
2017 spin_unlock(lock: &_minor_lock);
2018 if (dm_get_md_type(md) != DM_TYPE_NONE) {
2019 struct table_device *td;
2020
2021 dm_sysfs_exit(md);
2022 list_for_each_entry(td, &md->table_devices, list) {
2023 bd_unlink_disk_holder(bdev: td->dm_dev.bdev,
2024 disk: md->disk);
2025 }
2026
2027 /*
2028 * Hold lock to make sure del_gendisk() won't concurrent
2029 * with open/close_table_device().
2030 */
2031 mutex_lock(&md->table_devices_lock);
2032 del_gendisk(gp: md->disk);
2033 mutex_unlock(lock: &md->table_devices_lock);
2034 }
2035 dm_queue_destroy_crypto_profile(q: md->queue);
2036 put_disk(disk: md->disk);
2037 }
2038
2039 if (md->pending_io) {
2040 free_percpu(pdata: md->pending_io);
2041 md->pending_io = NULL;
2042 }
2043
2044 cleanup_srcu_struct(ssp: &md->io_barrier);
2045
2046 mutex_destroy(lock: &md->suspend_lock);
2047 mutex_destroy(lock: &md->type_lock);
2048 mutex_destroy(lock: &md->table_devices_lock);
2049 mutex_destroy(lock: &md->swap_bios_lock);
2050
2051 dm_mq_cleanup_mapped_device(md);
2052}
2053
2054/*
2055 * Allocate and initialise a blank device with a given minor.
2056 */
2057static struct mapped_device *alloc_dev(int minor)
2058{
2059 int r, numa_node_id = dm_get_numa_node();
2060 struct dax_device *dax_dev;
2061 struct mapped_device *md;
2062 void *old_md;
2063
2064 md = kvzalloc_node(size: sizeof(*md), GFP_KERNEL, node: numa_node_id);
2065 if (!md) {
2066 DMERR("unable to allocate device, out of memory.");
2067 return NULL;
2068 }
2069
2070 if (!try_module_get(THIS_MODULE))
2071 goto bad_module_get;
2072
2073 /* get a minor number for the dev */
2074 if (minor == DM_ANY_MINOR)
2075 r = next_free_minor(minor: &minor);
2076 else
2077 r = specific_minor(minor);
2078 if (r < 0)
2079 goto bad_minor;
2080
2081 r = init_srcu_struct(&md->io_barrier);
2082 if (r < 0)
2083 goto bad_io_barrier;
2084
2085 md->numa_node_id = numa_node_id;
2086 md->init_tio_pdu = false;
2087 md->type = DM_TYPE_NONE;
2088 mutex_init(&md->suspend_lock);
2089 mutex_init(&md->type_lock);
2090 mutex_init(&md->table_devices_lock);
2091 spin_lock_init(&md->deferred_lock);
2092 atomic_set(v: &md->holders, i: 1);
2093 atomic_set(v: &md->open_count, i: 0);
2094 atomic_set(v: &md->event_nr, i: 0);
2095 atomic_set(v: &md->uevent_seq, i: 0);
2096 INIT_LIST_HEAD(list: &md->uevent_list);
2097 INIT_LIST_HEAD(list: &md->table_devices);
2098 spin_lock_init(&md->uevent_lock);
2099
2100 /*
2101 * default to bio-based until DM table is loaded and md->type
2102 * established. If request-based table is loaded: blk-mq will
2103 * override accordingly.
2104 */
2105 md->disk = blk_alloc_disk(NULL, md->numa_node_id);
2106 if (IS_ERR(ptr: md->disk))
2107 goto bad;
2108 md->queue = md->disk->queue;
2109
2110 init_waitqueue_head(&md->wait);
2111 INIT_WORK(&md->work, dm_wq_work);
2112 INIT_WORK(&md->requeue_work, dm_wq_requeue_work);
2113 init_waitqueue_head(&md->eventq);
2114 init_completion(x: &md->kobj_holder.completion);
2115
2116 md->requeue_list = NULL;
2117 md->swap_bios = get_swap_bios();
2118 sema_init(sem: &md->swap_bios_semaphore, val: md->swap_bios);
2119 mutex_init(&md->swap_bios_lock);
2120
2121 md->disk->major = _major;
2122 md->disk->first_minor = minor;
2123 md->disk->minors = 1;
2124 md->disk->flags |= GENHD_FL_NO_PART;
2125 md->disk->fops = &dm_blk_dops;
2126 md->disk->private_data = md;
2127 sprintf(buf: md->disk->disk_name, fmt: "dm-%d", minor);
2128
2129 dax_dev = alloc_dax(private: md, ops: &dm_dax_ops);
2130 if (IS_ERR(ptr: dax_dev)) {
2131 if (PTR_ERR(ptr: dax_dev) != -EOPNOTSUPP)
2132 goto bad;
2133 } else {
2134 set_dax_nocache(dax_dev);
2135 set_dax_nomc(dax_dev);
2136 md->dax_dev = dax_dev;
2137 if (dax_add_host(dax_dev, disk: md->disk))
2138 goto bad;
2139 }
2140
2141 format_dev_t(md->name, MKDEV(_major, minor));
2142
2143 md->wq = alloc_workqueue(fmt: "kdmflush/%s", flags: WQ_MEM_RECLAIM, max_active: 0, md->name);
2144 if (!md->wq)
2145 goto bad;
2146
2147 md->pending_io = alloc_percpu(unsigned long);
2148 if (!md->pending_io)
2149 goto bad;
2150
2151 r = dm_stats_init(st: &md->stats);
2152 if (r < 0)
2153 goto bad;
2154
2155 /* Populate the mapping, nobody knows we exist yet */
2156 spin_lock(lock: &_minor_lock);
2157 old_md = idr_replace(&_minor_idr, md, id: minor);
2158 spin_unlock(lock: &_minor_lock);
2159
2160 BUG_ON(old_md != MINOR_ALLOCED);
2161
2162 return md;
2163
2164bad:
2165 cleanup_mapped_device(md);
2166bad_io_barrier:
2167 free_minor(minor);
2168bad_minor:
2169 module_put(THIS_MODULE);
2170bad_module_get:
2171 kvfree(addr: md);
2172 return NULL;
2173}
2174
2175static void unlock_fs(struct mapped_device *md);
2176
2177static void free_dev(struct mapped_device *md)
2178{
2179 int minor = MINOR(disk_devt(md->disk));
2180
2181 unlock_fs(md);
2182
2183 cleanup_mapped_device(md);
2184
2185 WARN_ON_ONCE(!list_empty(&md->table_devices));
2186 dm_stats_cleanup(st: &md->stats);
2187 free_minor(minor);
2188
2189 module_put(THIS_MODULE);
2190 kvfree(addr: md);
2191}
2192
2193/*
2194 * Bind a table to the device.
2195 */
2196static void event_callback(void *context)
2197{
2198 unsigned long flags;
2199 LIST_HEAD(uevents);
2200 struct mapped_device *md = context;
2201
2202 spin_lock_irqsave(&md->uevent_lock, flags);
2203 list_splice_init(list: &md->uevent_list, head: &uevents);
2204 spin_unlock_irqrestore(lock: &md->uevent_lock, flags);
2205
2206 dm_send_uevents(events: &uevents, kobj: &disk_to_dev(md->disk)->kobj);
2207
2208 atomic_inc(v: &md->event_nr);
2209 wake_up(&md->eventq);
2210 dm_issue_global_event();
2211}
2212
2213/*
2214 * Returns old map, which caller must destroy.
2215 */
2216static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2217 struct queue_limits *limits)
2218{
2219 struct dm_table *old_map;
2220 sector_t size;
2221 int ret;
2222
2223 lockdep_assert_held(&md->suspend_lock);
2224
2225 size = dm_table_get_size(t);
2226
2227 /*
2228 * Wipe any geometry if the size of the table changed.
2229 */
2230 if (size != dm_get_size(md))
2231 memset(&md->geometry, 0, sizeof(md->geometry));
2232
2233 set_capacity(disk: md->disk, size);
2234
2235 dm_table_event_callback(t, fn: event_callback, context: md);
2236
2237 if (dm_table_request_based(t)) {
2238 /*
2239 * Leverage the fact that request-based DM targets are
2240 * immutable singletons - used to optimize dm_mq_queue_rq.
2241 */
2242 md->immutable_target = dm_table_get_immutable_target(t);
2243
2244 /*
2245 * There is no need to reload with request-based dm because the
2246 * size of front_pad doesn't change.
2247 *
2248 * Note for future: If you are to reload bioset, prep-ed
2249 * requests in the queue may refer to bio from the old bioset,
2250 * so you must walk through the queue to unprep.
2251 */
2252 if (!md->mempools) {
2253 md->mempools = t->mempools;
2254 t->mempools = NULL;
2255 }
2256 } else {
2257 /*
2258 * The md may already have mempools that need changing.
2259 * If so, reload bioset because front_pad may have changed
2260 * because a different table was loaded.
2261 */
2262 dm_free_md_mempools(pools: md->mempools);
2263 md->mempools = t->mempools;
2264 t->mempools = NULL;
2265 }
2266
2267 ret = dm_table_set_restrictions(t, q: md->queue, limits);
2268 if (ret) {
2269 old_map = ERR_PTR(error: ret);
2270 goto out;
2271 }
2272
2273 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2274 rcu_assign_pointer(md->map, (void *)t);
2275 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2276
2277 if (old_map)
2278 dm_sync_table(md);
2279out:
2280 return old_map;
2281}
2282
2283/*
2284 * Returns unbound table for the caller to free.
2285 */
2286static struct dm_table *__unbind(struct mapped_device *md)
2287{
2288 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2289
2290 if (!map)
2291 return NULL;
2292
2293 dm_table_event_callback(t: map, NULL, NULL);
2294 RCU_INIT_POINTER(md->map, NULL);
2295 dm_sync_table(md);
2296
2297 return map;
2298}
2299
2300/*
2301 * Constructor for a new device.
2302 */
2303int dm_create(int minor, struct mapped_device **result)
2304{
2305 struct mapped_device *md;
2306
2307 md = alloc_dev(minor);
2308 if (!md)
2309 return -ENXIO;
2310
2311 dm_ima_reset_data(md);
2312
2313 *result = md;
2314 return 0;
2315}
2316
2317/*
2318 * Functions to manage md->type.
2319 * All are required to hold md->type_lock.
2320 */
2321void dm_lock_md_type(struct mapped_device *md)
2322{
2323 mutex_lock(&md->type_lock);
2324}
2325
2326void dm_unlock_md_type(struct mapped_device *md)
2327{
2328 mutex_unlock(lock: &md->type_lock);
2329}
2330
2331void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2332{
2333 BUG_ON(!mutex_is_locked(&md->type_lock));
2334 md->type = type;
2335}
2336
2337enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2338{
2339 return md->type;
2340}
2341
2342struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2343{
2344 return md->immutable_target_type;
2345}
2346
2347/*
2348 * Setup the DM device's queue based on md's type
2349 */
2350int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2351{
2352 enum dm_queue_mode type = dm_table_get_type(t);
2353 struct queue_limits limits;
2354 struct table_device *td;
2355 int r;
2356
2357 switch (type) {
2358 case DM_TYPE_REQUEST_BASED:
2359 md->disk->fops = &dm_rq_blk_dops;
2360 r = dm_mq_init_request_queue(md, t);
2361 if (r) {
2362 DMERR("Cannot initialize queue for request-based dm mapped device");
2363 return r;
2364 }
2365 break;
2366 case DM_TYPE_BIO_BASED:
2367 case DM_TYPE_DAX_BIO_BASED:
2368 blk_queue_flag_set(QUEUE_FLAG_IO_STAT, q: md->queue);
2369 break;
2370 case DM_TYPE_NONE:
2371 WARN_ON_ONCE(true);
2372 break;
2373 }
2374
2375 r = dm_calculate_queue_limits(table: t, limits: &limits);
2376 if (r) {
2377 DMERR("Cannot calculate initial queue limits");
2378 return r;
2379 }
2380 r = dm_table_set_restrictions(t, q: md->queue, limits: &limits);
2381 if (r)
2382 return r;
2383
2384 /*
2385 * Hold lock to make sure add_disk() and del_gendisk() won't concurrent
2386 * with open_table_device() and close_table_device().
2387 */
2388 mutex_lock(&md->table_devices_lock);
2389 r = add_disk(disk: md->disk);
2390 mutex_unlock(lock: &md->table_devices_lock);
2391 if (r)
2392 return r;
2393
2394 /*
2395 * Register the holder relationship for devices added before the disk
2396 * was live.
2397 */
2398 list_for_each_entry(td, &md->table_devices, list) {
2399 r = bd_link_disk_holder(bdev: td->dm_dev.bdev, disk: md->disk);
2400 if (r)
2401 goto out_undo_holders;
2402 }
2403
2404 r = dm_sysfs_init(md);
2405 if (r)
2406 goto out_undo_holders;
2407
2408 md->type = type;
2409 return 0;
2410
2411out_undo_holders:
2412 list_for_each_entry_continue_reverse(td, &md->table_devices, list)
2413 bd_unlink_disk_holder(bdev: td->dm_dev.bdev, disk: md->disk);
2414 mutex_lock(&md->table_devices_lock);
2415 del_gendisk(gp: md->disk);
2416 mutex_unlock(lock: &md->table_devices_lock);
2417 return r;
2418}
2419
2420struct mapped_device *dm_get_md(dev_t dev)
2421{
2422 struct mapped_device *md;
2423 unsigned int minor = MINOR(dev);
2424
2425 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2426 return NULL;
2427
2428 spin_lock(lock: &_minor_lock);
2429
2430 md = idr_find(&_minor_idr, id: minor);
2431 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2432 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2433 md = NULL;
2434 goto out;
2435 }
2436 dm_get(md);
2437out:
2438 spin_unlock(lock: &_minor_lock);
2439
2440 return md;
2441}
2442EXPORT_SYMBOL_GPL(dm_get_md);
2443
2444void *dm_get_mdptr(struct mapped_device *md)
2445{
2446 return md->interface_ptr;
2447}
2448
2449void dm_set_mdptr(struct mapped_device *md, void *ptr)
2450{
2451 md->interface_ptr = ptr;
2452}
2453
2454void dm_get(struct mapped_device *md)
2455{
2456 atomic_inc(v: &md->holders);
2457 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2458}
2459
2460int dm_hold(struct mapped_device *md)
2461{
2462 spin_lock(lock: &_minor_lock);
2463 if (test_bit(DMF_FREEING, &md->flags)) {
2464 spin_unlock(lock: &_minor_lock);
2465 return -EBUSY;
2466 }
2467 dm_get(md);
2468 spin_unlock(lock: &_minor_lock);
2469 return 0;
2470}
2471EXPORT_SYMBOL_GPL(dm_hold);
2472
2473const char *dm_device_name(struct mapped_device *md)
2474{
2475 return md->name;
2476}
2477EXPORT_SYMBOL_GPL(dm_device_name);
2478
2479static void __dm_destroy(struct mapped_device *md, bool wait)
2480{
2481 struct dm_table *map;
2482 int srcu_idx;
2483
2484 might_sleep();
2485
2486 spin_lock(lock: &_minor_lock);
2487 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2488 set_bit(DMF_FREEING, addr: &md->flags);
2489 spin_unlock(lock: &_minor_lock);
2490
2491 blk_mark_disk_dead(disk: md->disk);
2492
2493 /*
2494 * Take suspend_lock so that presuspend and postsuspend methods
2495 * do not race with internal suspend.
2496 */
2497 mutex_lock(&md->suspend_lock);
2498 map = dm_get_live_table(md, srcu_idx: &srcu_idx);
2499 if (!dm_suspended_md(md)) {
2500 dm_table_presuspend_targets(t: map);
2501 set_bit(DMF_SUSPENDED, addr: &md->flags);
2502 set_bit(DMF_POST_SUSPENDING, addr: &md->flags);
2503 dm_table_postsuspend_targets(t: map);
2504 }
2505 /* dm_put_live_table must be before fsleep, otherwise deadlock is possible */
2506 dm_put_live_table(md, srcu_idx);
2507 mutex_unlock(lock: &md->suspend_lock);
2508
2509 /*
2510 * Rare, but there may be I/O requests still going to complete,
2511 * for example. Wait for all references to disappear.
2512 * No one should increment the reference count of the mapped_device,
2513 * after the mapped_device state becomes DMF_FREEING.
2514 */
2515 if (wait)
2516 while (atomic_read(v: &md->holders))
2517 fsleep(usecs: 1000);
2518 else if (atomic_read(v: &md->holders))
2519 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2520 dm_device_name(md), atomic_read(&md->holders));
2521
2522 dm_table_destroy(t: __unbind(md));
2523 free_dev(md);
2524}
2525
2526void dm_destroy(struct mapped_device *md)
2527{
2528 __dm_destroy(md, wait: true);
2529}
2530
2531void dm_destroy_immediate(struct mapped_device *md)
2532{
2533 __dm_destroy(md, wait: false);
2534}
2535
2536void dm_put(struct mapped_device *md)
2537{
2538 atomic_dec(v: &md->holders);
2539}
2540EXPORT_SYMBOL_GPL(dm_put);
2541
2542static bool dm_in_flight_bios(struct mapped_device *md)
2543{
2544 int cpu;
2545 unsigned long sum = 0;
2546
2547 for_each_possible_cpu(cpu)
2548 sum += *per_cpu_ptr(md->pending_io, cpu);
2549
2550 return sum != 0;
2551}
2552
2553static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2554{
2555 int r = 0;
2556 DEFINE_WAIT(wait);
2557
2558 while (true) {
2559 prepare_to_wait(wq_head: &md->wait, wq_entry: &wait, state: task_state);
2560
2561 if (!dm_in_flight_bios(md))
2562 break;
2563
2564 if (signal_pending_state(state: task_state, current)) {
2565 r = -EINTR;
2566 break;
2567 }
2568
2569 io_schedule();
2570 }
2571 finish_wait(wq_head: &md->wait, wq_entry: &wait);
2572
2573 smp_rmb();
2574
2575 return r;
2576}
2577
2578static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2579{
2580 int r = 0;
2581
2582 if (!queue_is_mq(q: md->queue))
2583 return dm_wait_for_bios_completion(md, task_state);
2584
2585 while (true) {
2586 if (!blk_mq_queue_inflight(q: md->queue))
2587 break;
2588
2589 if (signal_pending_state(state: task_state, current)) {
2590 r = -EINTR;
2591 break;
2592 }
2593
2594 fsleep(usecs: 5000);
2595 }
2596
2597 return r;
2598}
2599
2600/*
2601 * Process the deferred bios
2602 */
2603static void dm_wq_work(struct work_struct *work)
2604{
2605 struct mapped_device *md = container_of(work, struct mapped_device, work);
2606 struct bio *bio;
2607
2608 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2609 spin_lock_irq(lock: &md->deferred_lock);
2610 bio = bio_list_pop(bl: &md->deferred);
2611 spin_unlock_irq(lock: &md->deferred_lock);
2612
2613 if (!bio)
2614 break;
2615
2616 submit_bio_noacct(bio);
2617 cond_resched();
2618 }
2619}
2620
2621static void dm_queue_flush(struct mapped_device *md)
2622{
2623 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, addr: &md->flags);
2624 smp_mb__after_atomic();
2625 queue_work(wq: md->wq, work: &md->work);
2626}
2627
2628/*
2629 * Swap in a new table, returning the old one for the caller to destroy.
2630 */
2631struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2632{
2633 struct dm_table *live_map = NULL, *map = ERR_PTR(error: -EINVAL);
2634 struct queue_limits limits;
2635 int r;
2636
2637 mutex_lock(&md->suspend_lock);
2638
2639 /* device must be suspended */
2640 if (!dm_suspended_md(md))
2641 goto out;
2642
2643 /*
2644 * If the new table has no data devices, retain the existing limits.
2645 * This helps multipath with queue_if_no_path if all paths disappear,
2646 * then new I/O is queued based on these limits, and then some paths
2647 * reappear.
2648 */
2649 if (dm_table_has_no_data_devices(table)) {
2650 live_map = dm_get_live_table_fast(md);
2651 if (live_map)
2652 limits = md->queue->limits;
2653 dm_put_live_table_fast(md);
2654 }
2655
2656 if (!live_map) {
2657 r = dm_calculate_queue_limits(table, limits: &limits);
2658 if (r) {
2659 map = ERR_PTR(error: r);
2660 goto out;
2661 }
2662 }
2663
2664 map = __bind(md, t: table, limits: &limits);
2665 dm_issue_global_event();
2666
2667out:
2668 mutex_unlock(lock: &md->suspend_lock);
2669 return map;
2670}
2671
2672/*
2673 * Functions to lock and unlock any filesystem running on the
2674 * device.
2675 */
2676static int lock_fs(struct mapped_device *md)
2677{
2678 int r;
2679
2680 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2681
2682 r = bdev_freeze(bdev: md->disk->part0);
2683 if (!r)
2684 set_bit(DMF_FROZEN, addr: &md->flags);
2685 return r;
2686}
2687
2688static void unlock_fs(struct mapped_device *md)
2689{
2690 if (!test_bit(DMF_FROZEN, &md->flags))
2691 return;
2692 bdev_thaw(bdev: md->disk->part0);
2693 clear_bit(DMF_FROZEN, addr: &md->flags);
2694}
2695
2696/*
2697 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2698 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2699 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2700 *
2701 * If __dm_suspend returns 0, the device is completely quiescent
2702 * now. There is no request-processing activity. All new requests
2703 * are being added to md->deferred list.
2704 */
2705static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2706 unsigned int suspend_flags, unsigned int task_state,
2707 int dmf_suspended_flag)
2708{
2709 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2710 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2711 int r;
2712
2713 lockdep_assert_held(&md->suspend_lock);
2714
2715 /*
2716 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2717 * This flag is cleared before dm_suspend returns.
2718 */
2719 if (noflush)
2720 set_bit(DMF_NOFLUSH_SUSPENDING, addr: &md->flags);
2721 else
2722 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2723
2724 /*
2725 * This gets reverted if there's an error later and the targets
2726 * provide the .presuspend_undo hook.
2727 */
2728 dm_table_presuspend_targets(t: map);
2729
2730 /*
2731 * Flush I/O to the device.
2732 * Any I/O submitted after lock_fs() may not be flushed.
2733 * noflush takes precedence over do_lockfs.
2734 * (lock_fs() flushes I/Os and waits for them to complete.)
2735 */
2736 if (!noflush && do_lockfs) {
2737 r = lock_fs(md);
2738 if (r) {
2739 dm_table_presuspend_undo_targets(t: map);
2740 return r;
2741 }
2742 }
2743
2744 /*
2745 * Here we must make sure that no processes are submitting requests
2746 * to target drivers i.e. no one may be executing
2747 * dm_split_and_process_bio from dm_submit_bio.
2748 *
2749 * To get all processes out of dm_split_and_process_bio in dm_submit_bio,
2750 * we take the write lock. To prevent any process from reentering
2751 * dm_split_and_process_bio from dm_submit_bio and quiesce the thread
2752 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2753 * flush_workqueue(md->wq).
2754 */
2755 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, addr: &md->flags);
2756 if (map)
2757 synchronize_srcu(ssp: &md->io_barrier);
2758
2759 /*
2760 * Stop md->queue before flushing md->wq in case request-based
2761 * dm defers requests to md->wq from md->queue.
2762 */
2763 if (dm_request_based(md))
2764 dm_stop_queue(q: md->queue);
2765
2766 flush_workqueue(md->wq);
2767
2768 /*
2769 * At this point no more requests are entering target request routines.
2770 * We call dm_wait_for_completion to wait for all existing requests
2771 * to finish.
2772 */
2773 r = dm_wait_for_completion(md, task_state);
2774 if (!r)
2775 set_bit(nr: dmf_suspended_flag, addr: &md->flags);
2776
2777 if (noflush)
2778 clear_bit(DMF_NOFLUSH_SUSPENDING, addr: &md->flags);
2779 if (map)
2780 synchronize_srcu(ssp: &md->io_barrier);
2781
2782 /* were we interrupted ? */
2783 if (r < 0) {
2784 dm_queue_flush(md);
2785
2786 if (dm_request_based(md))
2787 dm_start_queue(q: md->queue);
2788
2789 unlock_fs(md);
2790 dm_table_presuspend_undo_targets(t: map);
2791 /* pushback list is already flushed, so skip flush */
2792 }
2793
2794 return r;
2795}
2796
2797/*
2798 * We need to be able to change a mapping table under a mounted
2799 * filesystem. For example we might want to move some data in
2800 * the background. Before the table can be swapped with
2801 * dm_bind_table, dm_suspend must be called to flush any in
2802 * flight bios and ensure that any further io gets deferred.
2803 */
2804/*
2805 * Suspend mechanism in request-based dm.
2806 *
2807 * 1. Flush all I/Os by lock_fs() if needed.
2808 * 2. Stop dispatching any I/O by stopping the request_queue.
2809 * 3. Wait for all in-flight I/Os to be completed or requeued.
2810 *
2811 * To abort suspend, start the request_queue.
2812 */
2813int dm_suspend(struct mapped_device *md, unsigned int suspend_flags)
2814{
2815 struct dm_table *map = NULL;
2816 int r = 0;
2817
2818retry:
2819 mutex_lock_nested(lock: &md->suspend_lock, SINGLE_DEPTH_NESTING);
2820
2821 if (dm_suspended_md(md)) {
2822 r = -EINVAL;
2823 goto out_unlock;
2824 }
2825
2826 if (dm_suspended_internally_md(md)) {
2827 /* already internally suspended, wait for internal resume */
2828 mutex_unlock(lock: &md->suspend_lock);
2829 r = wait_on_bit(word: &md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2830 if (r)
2831 return r;
2832 goto retry;
2833 }
2834
2835 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2836 if (!map) {
2837 /* avoid deadlock with fs/namespace.c:do_mount() */
2838 suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
2839 }
2840
2841 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2842 if (r)
2843 goto out_unlock;
2844
2845 set_bit(DMF_POST_SUSPENDING, addr: &md->flags);
2846 dm_table_postsuspend_targets(t: map);
2847 clear_bit(DMF_POST_SUSPENDING, addr: &md->flags);
2848
2849out_unlock:
2850 mutex_unlock(lock: &md->suspend_lock);
2851 return r;
2852}
2853
2854static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2855{
2856 if (map) {
2857 int r = dm_table_resume_targets(t: map);
2858
2859 if (r)
2860 return r;
2861 }
2862
2863 dm_queue_flush(md);
2864
2865 /*
2866 * Flushing deferred I/Os must be done after targets are resumed
2867 * so that mapping of targets can work correctly.
2868 * Request-based dm is queueing the deferred I/Os in its request_queue.
2869 */
2870 if (dm_request_based(md))
2871 dm_start_queue(q: md->queue);
2872
2873 unlock_fs(md);
2874
2875 return 0;
2876}
2877
2878int dm_resume(struct mapped_device *md)
2879{
2880 int r;
2881 struct dm_table *map = NULL;
2882
2883retry:
2884 r = -EINVAL;
2885 mutex_lock_nested(lock: &md->suspend_lock, SINGLE_DEPTH_NESTING);
2886
2887 if (!dm_suspended_md(md))
2888 goto out;
2889
2890 if (dm_suspended_internally_md(md)) {
2891 /* already internally suspended, wait for internal resume */
2892 mutex_unlock(lock: &md->suspend_lock);
2893 r = wait_on_bit(word: &md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2894 if (r)
2895 return r;
2896 goto retry;
2897 }
2898
2899 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2900 if (!map || !dm_table_get_size(t: map))
2901 goto out;
2902
2903 r = __dm_resume(md, map);
2904 if (r)
2905 goto out;
2906
2907 clear_bit(DMF_SUSPENDED, addr: &md->flags);
2908out:
2909 mutex_unlock(lock: &md->suspend_lock);
2910
2911 return r;
2912}
2913
2914/*
2915 * Internal suspend/resume works like userspace-driven suspend. It waits
2916 * until all bios finish and prevents issuing new bios to the target drivers.
2917 * It may be used only from the kernel.
2918 */
2919
2920static void __dm_internal_suspend(struct mapped_device *md, unsigned int suspend_flags)
2921{
2922 struct dm_table *map = NULL;
2923
2924 lockdep_assert_held(&md->suspend_lock);
2925
2926 if (md->internal_suspend_count++)
2927 return; /* nested internal suspend */
2928
2929 if (dm_suspended_md(md)) {
2930 set_bit(DMF_SUSPENDED_INTERNALLY, addr: &md->flags);
2931 return; /* nest suspend */
2932 }
2933
2934 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2935
2936 /*
2937 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2938 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2939 * would require changing .presuspend to return an error -- avoid this
2940 * until there is a need for more elaborate variants of internal suspend.
2941 */
2942 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2943 DMF_SUSPENDED_INTERNALLY);
2944
2945 set_bit(DMF_POST_SUSPENDING, addr: &md->flags);
2946 dm_table_postsuspend_targets(t: map);
2947 clear_bit(DMF_POST_SUSPENDING, addr: &md->flags);
2948}
2949
2950static void __dm_internal_resume(struct mapped_device *md)
2951{
2952 int r;
2953 struct dm_table *map;
2954
2955 BUG_ON(!md->internal_suspend_count);
2956
2957 if (--md->internal_suspend_count)
2958 return; /* resume from nested internal suspend */
2959
2960 if (dm_suspended_md(md))
2961 goto done; /* resume from nested suspend */
2962
2963 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2964 r = __dm_resume(md, map);
2965 if (r) {
2966 /*
2967 * If a preresume method of some target failed, we are in a
2968 * tricky situation. We can't return an error to the caller. We
2969 * can't fake success because then the "resume" and
2970 * "postsuspend" methods would not be paired correctly, and it
2971 * would break various targets, for example it would cause list
2972 * corruption in the "origin" target.
2973 *
2974 * So, we fake normal suspend here, to make sure that the
2975 * "resume" and "postsuspend" methods will be paired correctly.
2976 */
2977 DMERR("Preresume method failed: %d", r);
2978 set_bit(DMF_SUSPENDED, addr: &md->flags);
2979 }
2980done:
2981 clear_bit(DMF_SUSPENDED_INTERNALLY, addr: &md->flags);
2982 smp_mb__after_atomic();
2983 wake_up_bit(word: &md->flags, DMF_SUSPENDED_INTERNALLY);
2984}
2985
2986void dm_internal_suspend_noflush(struct mapped_device *md)
2987{
2988 mutex_lock(&md->suspend_lock);
2989 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2990 mutex_unlock(lock: &md->suspend_lock);
2991}
2992EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2993
2994void dm_internal_resume(struct mapped_device *md)
2995{
2996 mutex_lock(&md->suspend_lock);
2997 __dm_internal_resume(md);
2998 mutex_unlock(lock: &md->suspend_lock);
2999}
3000EXPORT_SYMBOL_GPL(dm_internal_resume);
3001
3002/*
3003 * Fast variants of internal suspend/resume hold md->suspend_lock,
3004 * which prevents interaction with userspace-driven suspend.
3005 */
3006
3007void dm_internal_suspend_fast(struct mapped_device *md)
3008{
3009 mutex_lock(&md->suspend_lock);
3010 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3011 return;
3012
3013 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, addr: &md->flags);
3014 synchronize_srcu(ssp: &md->io_barrier);
3015 flush_workqueue(md->wq);
3016 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3017}
3018EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3019
3020void dm_internal_resume_fast(struct mapped_device *md)
3021{
3022 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3023 goto done;
3024
3025 dm_queue_flush(md);
3026
3027done:
3028 mutex_unlock(lock: &md->suspend_lock);
3029}
3030EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3031
3032/*
3033 *---------------------------------------------------------------
3034 * Event notification.
3035 *---------------------------------------------------------------
3036 */
3037int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3038 unsigned int cookie, bool need_resize_uevent)
3039{
3040 int r;
3041 unsigned int noio_flag;
3042 char udev_cookie[DM_COOKIE_LENGTH];
3043 char *envp[3] = { NULL, NULL, NULL };
3044 char **envpp = envp;
3045 if (cookie) {
3046 snprintf(buf: udev_cookie, DM_COOKIE_LENGTH, fmt: "%s=%u",
3047 DM_COOKIE_ENV_VAR_NAME, cookie);
3048 *envpp++ = udev_cookie;
3049 }
3050 if (need_resize_uevent) {
3051 *envpp++ = "RESIZE=1";
3052 }
3053
3054 noio_flag = memalloc_noio_save();
3055
3056 r = kobject_uevent_env(kobj: &disk_to_dev(md->disk)->kobj, action, envp);
3057
3058 memalloc_noio_restore(flags: noio_flag);
3059
3060 return r;
3061}
3062
3063uint32_t dm_next_uevent_seq(struct mapped_device *md)
3064{
3065 return atomic_add_return(i: 1, v: &md->uevent_seq);
3066}
3067
3068uint32_t dm_get_event_nr(struct mapped_device *md)
3069{
3070 return atomic_read(v: &md->event_nr);
3071}
3072
3073int dm_wait_event(struct mapped_device *md, int event_nr)
3074{
3075 return wait_event_interruptible(md->eventq,
3076 (event_nr != atomic_read(&md->event_nr)));
3077}
3078
3079void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3080{
3081 unsigned long flags;
3082
3083 spin_lock_irqsave(&md->uevent_lock, flags);
3084 list_add(new: elist, head: &md->uevent_list);
3085 spin_unlock_irqrestore(lock: &md->uevent_lock, flags);
3086}
3087
3088/*
3089 * The gendisk is only valid as long as you have a reference
3090 * count on 'md'.
3091 */
3092struct gendisk *dm_disk(struct mapped_device *md)
3093{
3094 return md->disk;
3095}
3096EXPORT_SYMBOL_GPL(dm_disk);
3097
3098struct kobject *dm_kobject(struct mapped_device *md)
3099{
3100 return &md->kobj_holder.kobj;
3101}
3102
3103struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3104{
3105 struct mapped_device *md;
3106
3107 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3108
3109 spin_lock(lock: &_minor_lock);
3110 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
3111 md = NULL;
3112 goto out;
3113 }
3114 dm_get(md);
3115out:
3116 spin_unlock(lock: &_minor_lock);
3117
3118 return md;
3119}
3120
3121int dm_suspended_md(struct mapped_device *md)
3122{
3123 return test_bit(DMF_SUSPENDED, &md->flags);
3124}
3125
3126static int dm_post_suspending_md(struct mapped_device *md)
3127{
3128 return test_bit(DMF_POST_SUSPENDING, &md->flags);
3129}
3130
3131int dm_suspended_internally_md(struct mapped_device *md)
3132{
3133 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3134}
3135
3136int dm_test_deferred_remove_flag(struct mapped_device *md)
3137{
3138 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3139}
3140
3141int dm_suspended(struct dm_target *ti)
3142{
3143 return dm_suspended_md(md: ti->table->md);
3144}
3145EXPORT_SYMBOL_GPL(dm_suspended);
3146
3147int dm_post_suspending(struct dm_target *ti)
3148{
3149 return dm_post_suspending_md(md: ti->table->md);
3150}
3151EXPORT_SYMBOL_GPL(dm_post_suspending);
3152
3153int dm_noflush_suspending(struct dm_target *ti)
3154{
3155 return __noflush_suspending(md: ti->table->md);
3156}
3157EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3158
3159void dm_free_md_mempools(struct dm_md_mempools *pools)
3160{
3161 if (!pools)
3162 return;
3163
3164 bioset_exit(&pools->bs);
3165 bioset_exit(&pools->io_bs);
3166
3167 kfree(objp: pools);
3168}
3169
3170struct dm_pr {
3171 u64 old_key;
3172 u64 new_key;
3173 u32 flags;
3174 bool abort;
3175 bool fail_early;
3176 int ret;
3177 enum pr_type type;
3178 struct pr_keys *read_keys;
3179 struct pr_held_reservation *rsv;
3180};
3181
3182static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3183 struct dm_pr *pr)
3184{
3185 struct mapped_device *md = bdev->bd_disk->private_data;
3186 struct dm_table *table;
3187 struct dm_target *ti;
3188 int ret = -ENOTTY, srcu_idx;
3189
3190 table = dm_get_live_table(md, srcu_idx: &srcu_idx);
3191 if (!table || !dm_table_get_size(t: table))
3192 goto out;
3193
3194 /* We only support devices that have a single target */
3195 if (table->num_targets != 1)
3196 goto out;
3197 ti = dm_table_get_target(t: table, index: 0);
3198
3199 if (dm_suspended_md(md)) {
3200 ret = -EAGAIN;
3201 goto out;
3202 }
3203
3204 ret = -EINVAL;
3205 if (!ti->type->iterate_devices)
3206 goto out;
3207
3208 ti->type->iterate_devices(ti, fn, pr);
3209 ret = 0;
3210out:
3211 dm_put_live_table(md, srcu_idx);
3212 return ret;
3213}
3214
3215/*
3216 * For register / unregister we need to manually call out to every path.
3217 */
3218static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3219 sector_t start, sector_t len, void *data)
3220{
3221 struct dm_pr *pr = data;
3222 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3223 int ret;
3224
3225 if (!ops || !ops->pr_register) {
3226 pr->ret = -EOPNOTSUPP;
3227 return -1;
3228 }
3229
3230 ret = ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3231 if (!ret)
3232 return 0;
3233
3234 if (!pr->ret)
3235 pr->ret = ret;
3236
3237 if (pr->fail_early)
3238 return -1;
3239
3240 return 0;
3241}
3242
3243static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3244 u32 flags)
3245{
3246 struct dm_pr pr = {
3247 .old_key = old_key,
3248 .new_key = new_key,
3249 .flags = flags,
3250 .fail_early = true,
3251 .ret = 0,
3252 };
3253 int ret;
3254
3255 ret = dm_call_pr(bdev, fn: __dm_pr_register, pr: &pr);
3256 if (ret) {
3257 /* Didn't even get to register a path */
3258 return ret;
3259 }
3260
3261 if (!pr.ret)
3262 return 0;
3263 ret = pr.ret;
3264
3265 if (!new_key)
3266 return ret;
3267
3268 /* unregister all paths if we failed to register any path */
3269 pr.old_key = new_key;
3270 pr.new_key = 0;
3271 pr.flags = 0;
3272 pr.fail_early = false;
3273 (void) dm_call_pr(bdev, fn: __dm_pr_register, pr: &pr);
3274 return ret;
3275}
3276
3277
3278static int __dm_pr_reserve(struct dm_target *ti, struct dm_dev *dev,
3279 sector_t start, sector_t len, void *data)
3280{
3281 struct dm_pr *pr = data;
3282 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3283
3284 if (!ops || !ops->pr_reserve) {
3285 pr->ret = -EOPNOTSUPP;
3286 return -1;
3287 }
3288
3289 pr->ret = ops->pr_reserve(dev->bdev, pr->old_key, pr->type, pr->flags);
3290 if (!pr->ret)
3291 return -1;
3292
3293 return 0;
3294}
3295
3296static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3297 u32 flags)
3298{
3299 struct dm_pr pr = {
3300 .old_key = key,
3301 .flags = flags,
3302 .type = type,
3303 .fail_early = false,
3304 .ret = 0,
3305 };
3306 int ret;
3307
3308 ret = dm_call_pr(bdev, fn: __dm_pr_reserve, pr: &pr);
3309 if (ret)
3310 return ret;
3311
3312 return pr.ret;
3313}
3314
3315/*
3316 * If there is a non-All Registrants type of reservation, the release must be
3317 * sent down the holding path. For the cases where there is no reservation or
3318 * the path is not the holder the device will also return success, so we must
3319 * try each path to make sure we got the correct path.
3320 */
3321static int __dm_pr_release(struct dm_target *ti, struct dm_dev *dev,
3322 sector_t start, sector_t len, void *data)
3323{
3324 struct dm_pr *pr = data;
3325 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3326
3327 if (!ops || !ops->pr_release) {
3328 pr->ret = -EOPNOTSUPP;
3329 return -1;
3330 }
3331
3332 pr->ret = ops->pr_release(dev->bdev, pr->old_key, pr->type);
3333 if (pr->ret)
3334 return -1;
3335
3336 return 0;
3337}
3338
3339static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3340{
3341 struct dm_pr pr = {
3342 .old_key = key,
3343 .type = type,
3344 .fail_early = false,
3345 };
3346 int ret;
3347
3348 ret = dm_call_pr(bdev, fn: __dm_pr_release, pr: &pr);
3349 if (ret)
3350 return ret;
3351
3352 return pr.ret;
3353}
3354
3355static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
3356 sector_t start, sector_t len, void *data)
3357{
3358 struct dm_pr *pr = data;
3359 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3360
3361 if (!ops || !ops->pr_preempt) {
3362 pr->ret = -EOPNOTSUPP;
3363 return -1;
3364 }
3365
3366 pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
3367 pr->abort);
3368 if (!pr->ret)
3369 return -1;
3370
3371 return 0;
3372}
3373
3374static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3375 enum pr_type type, bool abort)
3376{
3377 struct dm_pr pr = {
3378 .new_key = new_key,
3379 .old_key = old_key,
3380 .type = type,
3381 .fail_early = false,
3382 };
3383 int ret;
3384
3385 ret = dm_call_pr(bdev, fn: __dm_pr_preempt, pr: &pr);
3386 if (ret)
3387 return ret;
3388
3389 return pr.ret;
3390}
3391
3392static int dm_pr_clear(struct block_device *bdev, u64 key)
3393{
3394 struct mapped_device *md = bdev->bd_disk->private_data;
3395 const struct pr_ops *ops;
3396 int r, srcu_idx;
3397
3398 r = dm_prepare_ioctl(md, srcu_idx: &srcu_idx, bdev: &bdev);
3399 if (r < 0)
3400 goto out;
3401
3402 ops = bdev->bd_disk->fops->pr_ops;
3403 if (ops && ops->pr_clear)
3404 r = ops->pr_clear(bdev, key);
3405 else
3406 r = -EOPNOTSUPP;
3407out:
3408 dm_unprepare_ioctl(md, srcu_idx);
3409 return r;
3410}
3411
3412static int __dm_pr_read_keys(struct dm_target *ti, struct dm_dev *dev,
3413 sector_t start, sector_t len, void *data)
3414{
3415 struct dm_pr *pr = data;
3416 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3417
3418 if (!ops || !ops->pr_read_keys) {
3419 pr->ret = -EOPNOTSUPP;
3420 return -1;
3421 }
3422
3423 pr->ret = ops->pr_read_keys(dev->bdev, pr->read_keys);
3424 if (!pr->ret)
3425 return -1;
3426
3427 return 0;
3428}
3429
3430static int dm_pr_read_keys(struct block_device *bdev, struct pr_keys *keys)
3431{
3432 struct dm_pr pr = {
3433 .read_keys = keys,
3434 };
3435 int ret;
3436
3437 ret = dm_call_pr(bdev, fn: __dm_pr_read_keys, pr: &pr);
3438 if (ret)
3439 return ret;
3440
3441 return pr.ret;
3442}
3443
3444static int __dm_pr_read_reservation(struct dm_target *ti, struct dm_dev *dev,
3445 sector_t start, sector_t len, void *data)
3446{
3447 struct dm_pr *pr = data;
3448 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3449
3450 if (!ops || !ops->pr_read_reservation) {
3451 pr->ret = -EOPNOTSUPP;
3452 return -1;
3453 }
3454
3455 pr->ret = ops->pr_read_reservation(dev->bdev, pr->rsv);
3456 if (!pr->ret)
3457 return -1;
3458
3459 return 0;
3460}
3461
3462static int dm_pr_read_reservation(struct block_device *bdev,
3463 struct pr_held_reservation *rsv)
3464{
3465 struct dm_pr pr = {
3466 .rsv = rsv,
3467 };
3468 int ret;
3469
3470 ret = dm_call_pr(bdev, fn: __dm_pr_read_reservation, pr: &pr);
3471 if (ret)
3472 return ret;
3473
3474 return pr.ret;
3475}
3476
3477static const struct pr_ops dm_pr_ops = {
3478 .pr_register = dm_pr_register,
3479 .pr_reserve = dm_pr_reserve,
3480 .pr_release = dm_pr_release,
3481 .pr_preempt = dm_pr_preempt,
3482 .pr_clear = dm_pr_clear,
3483 .pr_read_keys = dm_pr_read_keys,
3484 .pr_read_reservation = dm_pr_read_reservation,
3485};
3486
3487static const struct block_device_operations dm_blk_dops = {
3488 .submit_bio = dm_submit_bio,
3489 .poll_bio = dm_poll_bio,
3490 .open = dm_blk_open,
3491 .release = dm_blk_close,
3492 .ioctl = dm_blk_ioctl,
3493 .getgeo = dm_blk_getgeo,
3494 .report_zones = dm_blk_report_zones,
3495 .pr_ops = &dm_pr_ops,
3496 .owner = THIS_MODULE
3497};
3498
3499static const struct block_device_operations dm_rq_blk_dops = {
3500 .open = dm_blk_open,
3501 .release = dm_blk_close,
3502 .ioctl = dm_blk_ioctl,
3503 .getgeo = dm_blk_getgeo,
3504 .pr_ops = &dm_pr_ops,
3505 .owner = THIS_MODULE
3506};
3507
3508static const struct dax_operations dm_dax_ops = {
3509 .direct_access = dm_dax_direct_access,
3510 .zero_page_range = dm_dax_zero_page_range,
3511 .recovery_write = dm_dax_recovery_write,
3512};
3513
3514/*
3515 * module hooks
3516 */
3517module_init(dm_init);
3518module_exit(dm_exit);
3519
3520module_param(major, uint, 0);
3521MODULE_PARM_DESC(major, "The major number of the device mapper");
3522
3523module_param(reserved_bio_based_ios, uint, 0644);
3524MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3525
3526module_param(dm_numa_node, int, 0644);
3527MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3528
3529module_param(swap_bios, int, 0644);
3530MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3531
3532MODULE_DESCRIPTION(DM_NAME " driver");
3533MODULE_AUTHOR("Joe Thornber <dm-devel@lists.linux.dev>");
3534MODULE_LICENSE("GPL");
3535

source code of linux/drivers/md/dm.c