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 bdev_handle *bdev_handle;
730 u64 part_off;
731 int r;
732
733 td = kmalloc_node(size: sizeof(*td), GFP_KERNEL, node: md->numa_node_id);
734 if (!td)
735 return ERR_PTR(error: -ENOMEM);
736 refcount_set(r: &td->count, n: 1);
737
738 bdev_handle = bdev_open_by_dev(dev, mode, holder: _dm_claim_ptr, NULL);
739 if (IS_ERR(ptr: bdev_handle)) {
740 r = PTR_ERR(ptr: bdev_handle);
741 goto out_free_td;
742 }
743
744 /*
745 * We can be called before the dm disk is added. In that case we can't
746 * register the holder relation here. It will be done once add_disk was
747 * called.
748 */
749 if (md->disk->slave_dir) {
750 r = bd_link_disk_holder(bdev: bdev_handle->bdev, disk: md->disk);
751 if (r)
752 goto out_blkdev_put;
753 }
754
755 td->dm_dev.mode = mode;
756 td->dm_dev.bdev = bdev_handle->bdev;
757 td->dm_dev.bdev_handle = bdev_handle;
758 td->dm_dev.dax_dev = fs_dax_get_by_bdev(bdev: bdev_handle->bdev, start_off: &part_off,
759 NULL, NULL);
760 format_dev_t(td->dm_dev.name, dev);
761 list_add(new: &td->list, head: &md->table_devices);
762 return td;
763
764out_blkdev_put:
765 bdev_release(handle: bdev_handle);
766out_free_td:
767 kfree(objp: td);
768 return ERR_PTR(error: r);
769}
770
771/*
772 * Close a table device that we've been using.
773 */
774static void close_table_device(struct table_device *td, struct mapped_device *md)
775{
776 if (md->disk->slave_dir)
777 bd_unlink_disk_holder(bdev: td->dm_dev.bdev, disk: md->disk);
778 bdev_release(handle: td->dm_dev.bdev_handle);
779 put_dax(dax_dev: td->dm_dev.dax_dev);
780 list_del(entry: &td->list);
781 kfree(objp: td);
782}
783
784static struct table_device *find_table_device(struct list_head *l, dev_t dev,
785 blk_mode_t mode)
786{
787 struct table_device *td;
788
789 list_for_each_entry(td, l, list)
790 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
791 return td;
792
793 return NULL;
794}
795
796int dm_get_table_device(struct mapped_device *md, dev_t dev, blk_mode_t mode,
797 struct dm_dev **result)
798{
799 struct table_device *td;
800
801 mutex_lock(&md->table_devices_lock);
802 td = find_table_device(l: &md->table_devices, dev, mode);
803 if (!td) {
804 td = open_table_device(md, dev, mode);
805 if (IS_ERR(ptr: td)) {
806 mutex_unlock(lock: &md->table_devices_lock);
807 return PTR_ERR(ptr: td);
808 }
809 } else {
810 refcount_inc(r: &td->count);
811 }
812 mutex_unlock(lock: &md->table_devices_lock);
813
814 *result = &td->dm_dev;
815 return 0;
816}
817
818void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
819{
820 struct table_device *td = container_of(d, struct table_device, dm_dev);
821
822 mutex_lock(&md->table_devices_lock);
823 if (refcount_dec_and_test(r: &td->count))
824 close_table_device(td, md);
825 mutex_unlock(lock: &md->table_devices_lock);
826}
827
828/*
829 * Get the geometry associated with a dm device
830 */
831int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
832{
833 *geo = md->geometry;
834
835 return 0;
836}
837
838/*
839 * Set the geometry of a device.
840 */
841int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
842{
843 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
844
845 if (geo->start > sz) {
846 DMERR("Start sector is beyond the geometry limits.");
847 return -EINVAL;
848 }
849
850 md->geometry = *geo;
851
852 return 0;
853}
854
855static int __noflush_suspending(struct mapped_device *md)
856{
857 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
858}
859
860static void dm_requeue_add_io(struct dm_io *io, bool first_stage)
861{
862 struct mapped_device *md = io->md;
863
864 if (first_stage) {
865 struct dm_io *next = md->requeue_list;
866
867 md->requeue_list = io;
868 io->next = next;
869 } else {
870 bio_list_add_head(bl: &md->deferred, bio: io->orig_bio);
871 }
872}
873
874static void dm_kick_requeue(struct mapped_device *md, bool first_stage)
875{
876 if (first_stage)
877 queue_work(wq: md->wq, work: &md->requeue_work);
878 else
879 queue_work(wq: md->wq, work: &md->work);
880}
881
882/*
883 * Return true if the dm_io's original bio is requeued.
884 * io->status is updated with error if requeue disallowed.
885 */
886static bool dm_handle_requeue(struct dm_io *io, bool first_stage)
887{
888 struct bio *bio = io->orig_bio;
889 bool handle_requeue = (io->status == BLK_STS_DM_REQUEUE);
890 bool handle_polled_eagain = ((io->status == BLK_STS_AGAIN) &&
891 (bio->bi_opf & REQ_POLLED));
892 struct mapped_device *md = io->md;
893 bool requeued = false;
894
895 if (handle_requeue || handle_polled_eagain) {
896 unsigned long flags;
897
898 if (bio->bi_opf & REQ_POLLED) {
899 /*
900 * Upper layer won't help us poll split bio
901 * (io->orig_bio may only reflect a subset of the
902 * pre-split original) so clear REQ_POLLED.
903 */
904 bio_clear_polled(bio);
905 }
906
907 /*
908 * Target requested pushing back the I/O or
909 * polled IO hit BLK_STS_AGAIN.
910 */
911 spin_lock_irqsave(&md->deferred_lock, flags);
912 if ((__noflush_suspending(md) &&
913 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) ||
914 handle_polled_eagain || first_stage) {
915 dm_requeue_add_io(io, first_stage);
916 requeued = true;
917 } else {
918 /*
919 * noflush suspend was interrupted or this is
920 * a write to a zoned target.
921 */
922 io->status = BLK_STS_IOERR;
923 }
924 spin_unlock_irqrestore(lock: &md->deferred_lock, flags);
925 }
926
927 if (requeued)
928 dm_kick_requeue(md, first_stage);
929
930 return requeued;
931}
932
933static void __dm_io_complete(struct dm_io *io, bool first_stage)
934{
935 struct bio *bio = io->orig_bio;
936 struct mapped_device *md = io->md;
937 blk_status_t io_error;
938 bool requeued;
939
940 requeued = dm_handle_requeue(io, first_stage);
941 if (requeued && first_stage)
942 return;
943
944 io_error = io->status;
945 if (dm_io_flagged(io, bit: DM_IO_ACCOUNTED))
946 dm_end_io_acct(io);
947 else if (!io_error) {
948 /*
949 * Must handle target that DM_MAPIO_SUBMITTED only to
950 * then bio_endio() rather than dm_submit_bio_remap()
951 */
952 __dm_start_io_acct(io);
953 dm_end_io_acct(io);
954 }
955 free_io(io);
956 smp_wmb();
957 this_cpu_dec(*md->pending_io);
958
959 /* nudge anyone waiting on suspend queue */
960 if (unlikely(wq_has_sleeper(&md->wait)))
961 wake_up(&md->wait);
962
963 /* Return early if the original bio was requeued */
964 if (requeued)
965 return;
966
967 if (bio_is_flush_with_data(bio)) {
968 /*
969 * Preflush done for flush with data, reissue
970 * without REQ_PREFLUSH.
971 */
972 bio->bi_opf &= ~REQ_PREFLUSH;
973 queue_io(md, bio);
974 } else {
975 /* done with normal IO or empty flush */
976 if (io_error)
977 bio->bi_status = io_error;
978 bio_endio(bio);
979 }
980}
981
982static void dm_wq_requeue_work(struct work_struct *work)
983{
984 struct mapped_device *md = container_of(work, struct mapped_device,
985 requeue_work);
986 unsigned long flags;
987 struct dm_io *io;
988
989 /* reuse deferred lock to simplify dm_handle_requeue */
990 spin_lock_irqsave(&md->deferred_lock, flags);
991 io = md->requeue_list;
992 md->requeue_list = NULL;
993 spin_unlock_irqrestore(lock: &md->deferred_lock, flags);
994
995 while (io) {
996 struct dm_io *next = io->next;
997
998 dm_io_rewind(io, bs: &md->disk->bio_split);
999
1000 io->next = NULL;
1001 __dm_io_complete(io, first_stage: false);
1002 io = next;
1003 cond_resched();
1004 }
1005}
1006
1007/*
1008 * Two staged requeue:
1009 *
1010 * 1) io->orig_bio points to the real original bio, and the part mapped to
1011 * this io must be requeued, instead of other parts of the original bio.
1012 *
1013 * 2) io->orig_bio points to new cloned bio which matches the requeued dm_io.
1014 */
1015static void dm_io_complete(struct dm_io *io)
1016{
1017 bool first_requeue;
1018
1019 /*
1020 * Only dm_io that has been split needs two stage requeue, otherwise
1021 * we may run into long bio clone chain during suspend and OOM could
1022 * be triggered.
1023 *
1024 * Also flush data dm_io won't be marked as DM_IO_WAS_SPLIT, so they
1025 * also aren't handled via the first stage requeue.
1026 */
1027 if (dm_io_flagged(io, bit: DM_IO_WAS_SPLIT))
1028 first_requeue = true;
1029 else
1030 first_requeue = false;
1031
1032 __dm_io_complete(io, first_stage: first_requeue);
1033}
1034
1035/*
1036 * Decrements the number of outstanding ios that a bio has been
1037 * cloned into, completing the original io if necc.
1038 */
1039static inline void __dm_io_dec_pending(struct dm_io *io)
1040{
1041 if (atomic_dec_and_test(v: &io->io_count))
1042 dm_io_complete(io);
1043}
1044
1045static void dm_io_set_error(struct dm_io *io, blk_status_t error)
1046{
1047 unsigned long flags;
1048
1049 /* Push-back supersedes any I/O errors */
1050 spin_lock_irqsave(&io->lock, flags);
1051 if (!(io->status == BLK_STS_DM_REQUEUE &&
1052 __noflush_suspending(md: io->md))) {
1053 io->status = error;
1054 }
1055 spin_unlock_irqrestore(lock: &io->lock, flags);
1056}
1057
1058static void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
1059{
1060 if (unlikely(error))
1061 dm_io_set_error(io, error);
1062
1063 __dm_io_dec_pending(io);
1064}
1065
1066/*
1067 * The queue_limits are only valid as long as you have a reference
1068 * count on 'md'. But _not_ imposing verification to avoid atomic_read(),
1069 */
1070static inline struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
1071{
1072 return &md->queue->limits;
1073}
1074
1075void disable_discard(struct mapped_device *md)
1076{
1077 struct queue_limits *limits = dm_get_queue_limits(md);
1078
1079 /* device doesn't really support DISCARD, disable it */
1080 limits->max_discard_sectors = 0;
1081}
1082
1083void disable_write_zeroes(struct mapped_device *md)
1084{
1085 struct queue_limits *limits = dm_get_queue_limits(md);
1086
1087 /* device doesn't really support WRITE ZEROES, disable it */
1088 limits->max_write_zeroes_sectors = 0;
1089}
1090
1091static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
1092{
1093 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
1094}
1095
1096static void clone_endio(struct bio *bio)
1097{
1098 blk_status_t error = bio->bi_status;
1099 struct dm_target_io *tio = clone_to_tio(clone: bio);
1100 struct dm_target *ti = tio->ti;
1101 dm_endio_fn endio = ti->type->end_io;
1102 struct dm_io *io = tio->io;
1103 struct mapped_device *md = io->md;
1104
1105 if (unlikely(error == BLK_STS_TARGET)) {
1106 if (bio_op(bio) == REQ_OP_DISCARD &&
1107 !bdev_max_discard_sectors(bdev: bio->bi_bdev))
1108 disable_discard(md);
1109 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
1110 !bdev_write_zeroes_sectors(bdev: bio->bi_bdev))
1111 disable_write_zeroes(md);
1112 }
1113
1114 if (static_branch_unlikely(&zoned_enabled) &&
1115 unlikely(bdev_is_zoned(bio->bi_bdev)))
1116 dm_zone_endio(io, clone: bio);
1117
1118 if (endio) {
1119 int r = endio(ti, bio, &error);
1120
1121 switch (r) {
1122 case DM_ENDIO_REQUEUE:
1123 if (static_branch_unlikely(&zoned_enabled)) {
1124 /*
1125 * Requeuing writes to a sequential zone of a zoned
1126 * target will break the sequential write pattern:
1127 * fail such IO.
1128 */
1129 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
1130 error = BLK_STS_IOERR;
1131 else
1132 error = BLK_STS_DM_REQUEUE;
1133 } else
1134 error = BLK_STS_DM_REQUEUE;
1135 fallthrough;
1136 case DM_ENDIO_DONE:
1137 break;
1138 case DM_ENDIO_INCOMPLETE:
1139 /* The target will handle the io */
1140 return;
1141 default:
1142 DMCRIT("unimplemented target endio return value: %d", r);
1143 BUG();
1144 }
1145 }
1146
1147 if (static_branch_unlikely(&swap_bios_enabled) &&
1148 unlikely(swap_bios_limit(ti, bio)))
1149 up(sem: &md->swap_bios_semaphore);
1150
1151 free_tio(clone: bio);
1152 dm_io_dec_pending(io, error);
1153}
1154
1155/*
1156 * Return maximum size of I/O possible at the supplied sector up to the current
1157 * target boundary.
1158 */
1159static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
1160 sector_t target_offset)
1161{
1162 return ti->len - target_offset;
1163}
1164
1165static sector_t __max_io_len(struct dm_target *ti, sector_t sector,
1166 unsigned int max_granularity,
1167 unsigned int max_sectors)
1168{
1169 sector_t target_offset = dm_target_offset(ti, sector);
1170 sector_t len = max_io_len_target_boundary(ti, target_offset);
1171
1172 /*
1173 * Does the target need to split IO even further?
1174 * - varied (per target) IO splitting is a tenet of DM; this
1175 * explains why stacked chunk_sectors based splitting via
1176 * bio_split_to_limits() isn't possible here.
1177 */
1178 if (!max_granularity)
1179 return len;
1180 return min_t(sector_t, len,
1181 min(max_sectors ? : queue_max_sectors(ti->table->md->queue),
1182 blk_chunk_sectors_left(target_offset, max_granularity)));
1183}
1184
1185static inline sector_t max_io_len(struct dm_target *ti, sector_t sector)
1186{
1187 return __max_io_len(ti, sector, max_granularity: ti->max_io_len, max_sectors: 0);
1188}
1189
1190int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1191{
1192 if (len > UINT_MAX) {
1193 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1194 (unsigned long long)len, UINT_MAX);
1195 ti->error = "Maximum size of target IO is too large";
1196 return -EINVAL;
1197 }
1198
1199 ti->max_io_len = (uint32_t) len;
1200
1201 return 0;
1202}
1203EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1204
1205static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1206 sector_t sector, int *srcu_idx)
1207 __acquires(md->io_barrier)
1208{
1209 struct dm_table *map;
1210 struct dm_target *ti;
1211
1212 map = dm_get_live_table(md, srcu_idx);
1213 if (!map)
1214 return NULL;
1215
1216 ti = dm_table_find_target(t: map, sector);
1217 if (!ti)
1218 return NULL;
1219
1220 return ti;
1221}
1222
1223static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1224 long nr_pages, enum dax_access_mode mode, void **kaddr,
1225 pfn_t *pfn)
1226{
1227 struct mapped_device *md = dax_get_private(dax_dev);
1228 sector_t sector = pgoff * PAGE_SECTORS;
1229 struct dm_target *ti;
1230 long len, ret = -EIO;
1231 int srcu_idx;
1232
1233 ti = dm_dax_get_live_target(md, sector, srcu_idx: &srcu_idx);
1234
1235 if (!ti)
1236 goto out;
1237 if (!ti->type->direct_access)
1238 goto out;
1239 len = max_io_len(ti, sector) / PAGE_SECTORS;
1240 if (len < 1)
1241 goto out;
1242 nr_pages = min(len, nr_pages);
1243 ret = ti->type->direct_access(ti, pgoff, nr_pages, mode, kaddr, pfn);
1244
1245 out:
1246 dm_put_live_table(md, srcu_idx);
1247
1248 return ret;
1249}
1250
1251static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1252 size_t nr_pages)
1253{
1254 struct mapped_device *md = dax_get_private(dax_dev);
1255 sector_t sector = pgoff * PAGE_SECTORS;
1256 struct dm_target *ti;
1257 int ret = -EIO;
1258 int srcu_idx;
1259
1260 ti = dm_dax_get_live_target(md, sector, srcu_idx: &srcu_idx);
1261
1262 if (!ti)
1263 goto out;
1264 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1265 /*
1266 * ->zero_page_range() is mandatory dax operation. If we are
1267 * here, something is wrong.
1268 */
1269 goto out;
1270 }
1271 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1272 out:
1273 dm_put_live_table(md, srcu_idx);
1274
1275 return ret;
1276}
1277
1278static size_t dm_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
1279 void *addr, size_t bytes, struct iov_iter *i)
1280{
1281 struct mapped_device *md = dax_get_private(dax_dev);
1282 sector_t sector = pgoff * PAGE_SECTORS;
1283 struct dm_target *ti;
1284 int srcu_idx;
1285 long ret = 0;
1286
1287 ti = dm_dax_get_live_target(md, sector, srcu_idx: &srcu_idx);
1288 if (!ti || !ti->type->dax_recovery_write)
1289 goto out;
1290
1291 ret = ti->type->dax_recovery_write(ti, pgoff, addr, bytes, i);
1292out:
1293 dm_put_live_table(md, srcu_idx);
1294 return ret;
1295}
1296
1297/*
1298 * A target may call dm_accept_partial_bio only from the map routine. It is
1299 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1300 * operations, REQ_OP_ZONE_APPEND (zone append writes) and any bio serviced by
1301 * __send_duplicate_bios().
1302 *
1303 * dm_accept_partial_bio informs the dm that the target only wants to process
1304 * additional n_sectors sectors of the bio and the rest of the data should be
1305 * sent in a next bio.
1306 *
1307 * A diagram that explains the arithmetics:
1308 * +--------------------+---------------+-------+
1309 * | 1 | 2 | 3 |
1310 * +--------------------+---------------+-------+
1311 *
1312 * <-------------- *tio->len_ptr --------------->
1313 * <----- bio_sectors ----->
1314 * <-- n_sectors -->
1315 *
1316 * Region 1 was already iterated over with bio_advance or similar function.
1317 * (it may be empty if the target doesn't use bio_advance)
1318 * Region 2 is the remaining bio size that the target wants to process.
1319 * (it may be empty if region 1 is non-empty, although there is no reason
1320 * to make it empty)
1321 * The target requires that region 3 is to be sent in the next bio.
1322 *
1323 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1324 * the partially processed part (the sum of regions 1+2) must be the same for all
1325 * copies of the bio.
1326 */
1327void dm_accept_partial_bio(struct bio *bio, unsigned int n_sectors)
1328{
1329 struct dm_target_io *tio = clone_to_tio(clone: bio);
1330 struct dm_io *io = tio->io;
1331 unsigned int bio_sectors = bio_sectors(bio);
1332
1333 BUG_ON(dm_tio_flagged(tio, DM_TIO_IS_DUPLICATE_BIO));
1334 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1335 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1336 BUG_ON(bio_sectors > *tio->len_ptr);
1337 BUG_ON(n_sectors > bio_sectors);
1338
1339 *tio->len_ptr -= bio_sectors - n_sectors;
1340 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1341
1342 /*
1343 * __split_and_process_bio() may have already saved mapped part
1344 * for accounting but it is being reduced so update accordingly.
1345 */
1346 dm_io_set_flag(io, bit: DM_IO_WAS_SPLIT);
1347 io->sectors = n_sectors;
1348 io->sector_offset = bio_sectors(io->orig_bio);
1349}
1350EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1351
1352/*
1353 * @clone: clone bio that DM core passed to target's .map function
1354 * @tgt_clone: clone of @clone bio that target needs submitted
1355 *
1356 * Targets should use this interface to submit bios they take
1357 * ownership of when returning DM_MAPIO_SUBMITTED.
1358 *
1359 * Target should also enable ti->accounts_remapped_io
1360 */
1361void dm_submit_bio_remap(struct bio *clone, struct bio *tgt_clone)
1362{
1363 struct dm_target_io *tio = clone_to_tio(clone);
1364 struct dm_io *io = tio->io;
1365
1366 /* establish bio that will get submitted */
1367 if (!tgt_clone)
1368 tgt_clone = clone;
1369
1370 /*
1371 * Account io->origin_bio to DM dev on behalf of target
1372 * that took ownership of IO with DM_MAPIO_SUBMITTED.
1373 */
1374 dm_start_io_acct(io, clone);
1375
1376 trace_block_bio_remap(bio: tgt_clone, dev: disk_devt(disk: io->md->disk),
1377 from: tio->old_sector);
1378 submit_bio_noacct(bio: tgt_clone);
1379}
1380EXPORT_SYMBOL_GPL(dm_submit_bio_remap);
1381
1382static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1383{
1384 mutex_lock(&md->swap_bios_lock);
1385 while (latch < md->swap_bios) {
1386 cond_resched();
1387 down(sem: &md->swap_bios_semaphore);
1388 md->swap_bios--;
1389 }
1390 while (latch > md->swap_bios) {
1391 cond_resched();
1392 up(sem: &md->swap_bios_semaphore);
1393 md->swap_bios++;
1394 }
1395 mutex_unlock(lock: &md->swap_bios_lock);
1396}
1397
1398static void __map_bio(struct bio *clone)
1399{
1400 struct dm_target_io *tio = clone_to_tio(clone);
1401 struct dm_target *ti = tio->ti;
1402 struct dm_io *io = tio->io;
1403 struct mapped_device *md = io->md;
1404 int r;
1405
1406 clone->bi_end_io = clone_endio;
1407
1408 /*
1409 * Map the clone.
1410 */
1411 tio->old_sector = clone->bi_iter.bi_sector;
1412
1413 if (static_branch_unlikely(&swap_bios_enabled) &&
1414 unlikely(swap_bios_limit(ti, clone))) {
1415 int latch = get_swap_bios();
1416
1417 if (unlikely(latch != md->swap_bios))
1418 __set_swap_bios_limit(md, latch);
1419 down(sem: &md->swap_bios_semaphore);
1420 }
1421
1422 if (static_branch_unlikely(&zoned_enabled)) {
1423 /*
1424 * Check if the IO needs a special mapping due to zone append
1425 * emulation on zoned target. In this case, dm_zone_map_bio()
1426 * calls the target map operation.
1427 */
1428 if (unlikely(dm_emulate_zone_append(md)))
1429 r = dm_zone_map_bio(io: tio);
1430 else
1431 goto do_map;
1432 } else {
1433do_map:
1434 if (likely(ti->type->map == linear_map))
1435 r = linear_map(ti, bio: clone);
1436 else if (ti->type->map == stripe_map)
1437 r = stripe_map(ti, bio: clone);
1438 else
1439 r = ti->type->map(ti, clone);
1440 }
1441
1442 switch (r) {
1443 case DM_MAPIO_SUBMITTED:
1444 /* target has assumed ownership of this io */
1445 if (!ti->accounts_remapped_io)
1446 dm_start_io_acct(io, clone);
1447 break;
1448 case DM_MAPIO_REMAPPED:
1449 dm_submit_bio_remap(clone, NULL);
1450 break;
1451 case DM_MAPIO_KILL:
1452 case DM_MAPIO_REQUEUE:
1453 if (static_branch_unlikely(&swap_bios_enabled) &&
1454 unlikely(swap_bios_limit(ti, clone)))
1455 up(sem: &md->swap_bios_semaphore);
1456 free_tio(clone);
1457 if (r == DM_MAPIO_KILL)
1458 dm_io_dec_pending(io, BLK_STS_IOERR);
1459 else
1460 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1461 break;
1462 default:
1463 DMCRIT("unimplemented target map return value: %d", r);
1464 BUG();
1465 }
1466}
1467
1468static void setup_split_accounting(struct clone_info *ci, unsigned int len)
1469{
1470 struct dm_io *io = ci->io;
1471
1472 if (ci->sector_count > len) {
1473 /*
1474 * Split needed, save the mapped part for accounting.
1475 * NOTE: dm_accept_partial_bio() will update accordingly.
1476 */
1477 dm_io_set_flag(io, bit: DM_IO_WAS_SPLIT);
1478 io->sectors = len;
1479 io->sector_offset = bio_sectors(ci->bio);
1480 }
1481}
1482
1483static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1484 struct dm_target *ti, unsigned int num_bios,
1485 unsigned *len, gfp_t gfp_flag)
1486{
1487 struct bio *bio;
1488 int try = (gfp_flag & GFP_NOWAIT) ? 0 : 1;
1489
1490 for (; try < 2; try++) {
1491 int bio_nr;
1492
1493 if (try && num_bios > 1)
1494 mutex_lock(&ci->io->md->table_devices_lock);
1495 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1496 bio = alloc_tio(ci, ti, target_bio_nr: bio_nr, len,
1497 gfp_mask: try ? GFP_NOIO : GFP_NOWAIT);
1498 if (!bio)
1499 break;
1500
1501 bio_list_add(bl: blist, bio);
1502 }
1503 if (try && num_bios > 1)
1504 mutex_unlock(lock: &ci->io->md->table_devices_lock);
1505 if (bio_nr == num_bios)
1506 return;
1507
1508 while ((bio = bio_list_pop(bl: blist)))
1509 free_tio(clone: bio);
1510 }
1511}
1512
1513static unsigned int __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1514 unsigned int num_bios, unsigned int *len,
1515 gfp_t gfp_flag)
1516{
1517 struct bio_list blist = BIO_EMPTY_LIST;
1518 struct bio *clone;
1519 unsigned int ret = 0;
1520
1521 if (WARN_ON_ONCE(num_bios == 0)) /* num_bios = 0 is a bug in caller */
1522 return 0;
1523
1524 /* dm_accept_partial_bio() is not supported with shared tio->len_ptr */
1525 if (len)
1526 setup_split_accounting(ci, len: *len);
1527
1528 /*
1529 * Using alloc_multiple_bios(), even if num_bios is 1, to consistently
1530 * support allocating using GFP_NOWAIT with GFP_NOIO fallback.
1531 */
1532 alloc_multiple_bios(blist: &blist, ci, ti, num_bios, len, gfp_flag);
1533 while ((clone = bio_list_pop(bl: &blist))) {
1534 if (num_bios > 1)
1535 dm_tio_set_flag(tio: clone_to_tio(clone), bit: DM_TIO_IS_DUPLICATE_BIO);
1536 __map_bio(clone);
1537 ret += 1;
1538 }
1539
1540 return ret;
1541}
1542
1543static void __send_empty_flush(struct clone_info *ci)
1544{
1545 struct dm_table *t = ci->map;
1546 struct bio flush_bio;
1547
1548 /*
1549 * Use an on-stack bio for this, it's safe since we don't
1550 * need to reference it after submit. It's just used as
1551 * the basis for the clone(s).
1552 */
1553 bio_init(bio: &flush_bio, bdev: ci->io->md->disk->part0, NULL, max_vecs: 0,
1554 opf: REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
1555
1556 ci->bio = &flush_bio;
1557 ci->sector_count = 0;
1558 ci->io->tio.clone.bi_iter.bi_size = 0;
1559
1560 for (unsigned int i = 0; i < t->num_targets; i++) {
1561 unsigned int bios;
1562 struct dm_target *ti = dm_table_get_target(t, index: i);
1563
1564 if (unlikely(ti->num_flush_bios == 0))
1565 continue;
1566
1567 atomic_add(i: ti->num_flush_bios, v: &ci->io->io_count);
1568 bios = __send_duplicate_bios(ci, ti, num_bios: ti->num_flush_bios,
1569 NULL, GFP_NOWAIT);
1570 atomic_sub(i: ti->num_flush_bios - bios, v: &ci->io->io_count);
1571 }
1572
1573 /*
1574 * alloc_io() takes one extra reference for submission, so the
1575 * reference won't reach 0 without the following subtraction
1576 */
1577 atomic_sub(i: 1, v: &ci->io->io_count);
1578
1579 bio_uninit(ci->bio);
1580}
1581
1582static void __send_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1583 unsigned int num_bios, unsigned int max_granularity,
1584 unsigned int max_sectors)
1585{
1586 unsigned int len, bios;
1587
1588 len = min_t(sector_t, ci->sector_count,
1589 __max_io_len(ti, ci->sector, max_granularity, max_sectors));
1590
1591 atomic_add(i: num_bios, v: &ci->io->io_count);
1592 bios = __send_duplicate_bios(ci, ti, num_bios, len: &len, GFP_NOIO);
1593 /*
1594 * alloc_io() takes one extra reference for submission, so the
1595 * reference won't reach 0 without the following (+1) subtraction
1596 */
1597 atomic_sub(i: num_bios - bios + 1, v: &ci->io->io_count);
1598
1599 ci->sector += len;
1600 ci->sector_count -= len;
1601}
1602
1603static bool is_abnormal_io(struct bio *bio)
1604{
1605 enum req_op op = bio_op(bio);
1606
1607 if (op != REQ_OP_READ && op != REQ_OP_WRITE && op != REQ_OP_FLUSH) {
1608 switch (op) {
1609 case REQ_OP_DISCARD:
1610 case REQ_OP_SECURE_ERASE:
1611 case REQ_OP_WRITE_ZEROES:
1612 return true;
1613 default:
1614 break;
1615 }
1616 }
1617
1618 return false;
1619}
1620
1621static blk_status_t __process_abnormal_io(struct clone_info *ci,
1622 struct dm_target *ti)
1623{
1624 unsigned int num_bios = 0;
1625 unsigned int max_granularity = 0;
1626 unsigned int max_sectors = 0;
1627 struct queue_limits *limits = dm_get_queue_limits(md: ti->table->md);
1628
1629 switch (bio_op(bio: ci->bio)) {
1630 case REQ_OP_DISCARD:
1631 num_bios = ti->num_discard_bios;
1632 max_sectors = limits->max_discard_sectors;
1633 if (ti->max_discard_granularity)
1634 max_granularity = max_sectors;
1635 break;
1636 case REQ_OP_SECURE_ERASE:
1637 num_bios = ti->num_secure_erase_bios;
1638 max_sectors = limits->max_secure_erase_sectors;
1639 if (ti->max_secure_erase_granularity)
1640 max_granularity = max_sectors;
1641 break;
1642 case REQ_OP_WRITE_ZEROES:
1643 num_bios = ti->num_write_zeroes_bios;
1644 max_sectors = limits->max_write_zeroes_sectors;
1645 if (ti->max_write_zeroes_granularity)
1646 max_granularity = max_sectors;
1647 break;
1648 default:
1649 break;
1650 }
1651
1652 /*
1653 * Even though the device advertised support for this type of
1654 * request, that does not mean every target supports it, and
1655 * reconfiguration might also have changed that since the
1656 * check was performed.
1657 */
1658 if (unlikely(!num_bios))
1659 return BLK_STS_NOTSUPP;
1660
1661 __send_abnormal_io(ci, ti, num_bios, max_granularity, max_sectors);
1662
1663 return BLK_STS_OK;
1664}
1665
1666/*
1667 * Reuse ->bi_private as dm_io list head for storing all dm_io instances
1668 * associated with this bio, and this bio's bi_private needs to be
1669 * stored in dm_io->data before the reuse.
1670 *
1671 * bio->bi_private is owned by fs or upper layer, so block layer won't
1672 * touch it after splitting. Meantime it won't be changed by anyone after
1673 * bio is submitted. So this reuse is safe.
1674 */
1675static inline struct dm_io **dm_poll_list_head(struct bio *bio)
1676{
1677 return (struct dm_io **)&bio->bi_private;
1678}
1679
1680static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
1681{
1682 struct dm_io **head = dm_poll_list_head(bio);
1683
1684 if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
1685 bio->bi_opf |= REQ_DM_POLL_LIST;
1686 /*
1687 * Save .bi_private into dm_io, so that we can reuse
1688 * .bi_private as dm_io list head for storing dm_io list
1689 */
1690 io->data = bio->bi_private;
1691
1692 /* tell block layer to poll for completion */
1693 bio->bi_cookie = ~BLK_QC_T_NONE;
1694
1695 io->next = NULL;
1696 } else {
1697 /*
1698 * bio recursed due to split, reuse original poll list,
1699 * and save bio->bi_private too.
1700 */
1701 io->data = (*head)->data;
1702 io->next = *head;
1703 }
1704
1705 *head = io;
1706}
1707
1708/*
1709 * Select the correct strategy for processing a non-flush bio.
1710 */
1711static blk_status_t __split_and_process_bio(struct clone_info *ci)
1712{
1713 struct bio *clone;
1714 struct dm_target *ti;
1715 unsigned int len;
1716
1717 ti = dm_table_find_target(t: ci->map, sector: ci->sector);
1718 if (unlikely(!ti))
1719 return BLK_STS_IOERR;
1720
1721 if (unlikely(ci->is_abnormal_io))
1722 return __process_abnormal_io(ci, ti);
1723
1724 /*
1725 * Only support bio polling for normal IO, and the target io is
1726 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
1727 */
1728 ci->submit_as_polled = !!(ci->bio->bi_opf & REQ_POLLED);
1729
1730 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1731 setup_split_accounting(ci, len);
1732
1733 if (unlikely(ci->bio->bi_opf & REQ_NOWAIT)) {
1734 if (unlikely(!dm_target_supports_nowait(ti->type)))
1735 return BLK_STS_NOTSUPP;
1736
1737 clone = alloc_tio(ci, ti, target_bio_nr: 0, len: &len, GFP_NOWAIT);
1738 if (unlikely(!clone))
1739 return BLK_STS_AGAIN;
1740 } else {
1741 clone = alloc_tio(ci, ti, target_bio_nr: 0, len: &len, GFP_NOIO);
1742 }
1743 __map_bio(clone);
1744
1745 ci->sector += len;
1746 ci->sector_count -= len;
1747
1748 return BLK_STS_OK;
1749}
1750
1751static void init_clone_info(struct clone_info *ci, struct dm_io *io,
1752 struct dm_table *map, struct bio *bio, bool is_abnormal)
1753{
1754 ci->map = map;
1755 ci->io = io;
1756 ci->bio = bio;
1757 ci->is_abnormal_io = is_abnormal;
1758 ci->submit_as_polled = false;
1759 ci->sector = bio->bi_iter.bi_sector;
1760 ci->sector_count = bio_sectors(bio);
1761
1762 /* Shouldn't happen but sector_count was being set to 0 so... */
1763 if (static_branch_unlikely(&zoned_enabled) &&
1764 WARN_ON_ONCE(op_is_zone_mgmt(bio_op(bio)) && ci->sector_count))
1765 ci->sector_count = 0;
1766}
1767
1768/*
1769 * Entry point to split a bio into clones and submit them to the targets.
1770 */
1771static void dm_split_and_process_bio(struct mapped_device *md,
1772 struct dm_table *map, struct bio *bio)
1773{
1774 struct clone_info ci;
1775 struct dm_io *io;
1776 blk_status_t error = BLK_STS_OK;
1777 bool is_abnormal;
1778
1779 is_abnormal = is_abnormal_io(bio);
1780 if (unlikely(is_abnormal)) {
1781 /*
1782 * Use bio_split_to_limits() for abnormal IO (e.g. discard, etc)
1783 * otherwise associated queue_limits won't be imposed.
1784 */
1785 bio = bio_split_to_limits(bio);
1786 if (!bio)
1787 return;
1788 }
1789
1790 /* Only support nowait for normal IO */
1791 if (unlikely(bio->bi_opf & REQ_NOWAIT) && !is_abnormal) {
1792 io = alloc_io(md, bio, GFP_NOWAIT);
1793 if (unlikely(!io)) {
1794 /* Unable to do anything without dm_io. */
1795 bio_wouldblock_error(bio);
1796 return;
1797 }
1798 } else {
1799 io = alloc_io(md, bio, GFP_NOIO);
1800 }
1801 init_clone_info(ci: &ci, io, map, bio, is_abnormal);
1802
1803 if (bio->bi_opf & REQ_PREFLUSH) {
1804 __send_empty_flush(ci: &ci);
1805 /* dm_io_complete submits any data associated with flush */
1806 goto out;
1807 }
1808
1809 error = __split_and_process_bio(ci: &ci);
1810 if (error || !ci.sector_count)
1811 goto out;
1812 /*
1813 * Remainder must be passed to submit_bio_noacct() so it gets handled
1814 * *after* bios already submitted have been completely processed.
1815 */
1816 bio_trim(bio, offset: io->sectors, size: ci.sector_count);
1817 trace_block_split(bio, new_sector: bio->bi_iter.bi_sector);
1818 bio_inc_remaining(bio);
1819 submit_bio_noacct(bio);
1820out:
1821 /*
1822 * Drop the extra reference count for non-POLLED bio, and hold one
1823 * reference for POLLED bio, which will be released in dm_poll_bio
1824 *
1825 * Add every dm_io instance into the dm_io list head which is stored
1826 * in bio->bi_private, so that dm_poll_bio can poll them all.
1827 */
1828 if (error || !ci.submit_as_polled) {
1829 /*
1830 * In case of submission failure, the extra reference for
1831 * submitting io isn't consumed yet
1832 */
1833 if (error)
1834 atomic_dec(v: &io->io_count);
1835 dm_io_dec_pending(io, error);
1836 } else
1837 dm_queue_poll_io(bio, io);
1838}
1839
1840static void dm_submit_bio(struct bio *bio)
1841{
1842 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
1843 int srcu_idx;
1844 struct dm_table *map;
1845
1846 map = dm_get_live_table(md, srcu_idx: &srcu_idx);
1847
1848 /* If suspended, or map not yet available, queue this IO for later */
1849 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) ||
1850 unlikely(!map)) {
1851 if (bio->bi_opf & REQ_NOWAIT)
1852 bio_wouldblock_error(bio);
1853 else if (bio->bi_opf & REQ_RAHEAD)
1854 bio_io_error(bio);
1855 else
1856 queue_io(md, bio);
1857 goto out;
1858 }
1859
1860 dm_split_and_process_bio(md, map, bio);
1861out:
1862 dm_put_live_table(md, srcu_idx);
1863}
1864
1865static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
1866 unsigned int flags)
1867{
1868 WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
1869
1870 /* don't poll if the mapped io is done */
1871 if (atomic_read(v: &io->io_count) > 1)
1872 bio_poll(bio: &io->tio.clone, iob, flags);
1873
1874 /* bio_poll holds the last reference */
1875 return atomic_read(v: &io->io_count) == 1;
1876}
1877
1878static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
1879 unsigned int flags)
1880{
1881 struct dm_io **head = dm_poll_list_head(bio);
1882 struct dm_io *list = *head;
1883 struct dm_io *tmp = NULL;
1884 struct dm_io *curr, *next;
1885
1886 /* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
1887 if (!(bio->bi_opf & REQ_DM_POLL_LIST))
1888 return 0;
1889
1890 WARN_ON_ONCE(!list);
1891
1892 /*
1893 * Restore .bi_private before possibly completing dm_io.
1894 *
1895 * bio_poll() is only possible once @bio has been completely
1896 * submitted via submit_bio_noacct()'s depth-first submission.
1897 * So there is no dm_queue_poll_io() race associated with
1898 * clearing REQ_DM_POLL_LIST here.
1899 */
1900 bio->bi_opf &= ~REQ_DM_POLL_LIST;
1901 bio->bi_private = list->data;
1902
1903 for (curr = list, next = curr->next; curr; curr = next, next =
1904 curr ? curr->next : NULL) {
1905 if (dm_poll_dm_io(io: curr, iob, flags)) {
1906 /*
1907 * clone_endio() has already occurred, so no
1908 * error handling is needed here.
1909 */
1910 __dm_io_dec_pending(io: curr);
1911 } else {
1912 curr->next = tmp;
1913 tmp = curr;
1914 }
1915 }
1916
1917 /* Not done? */
1918 if (tmp) {
1919 bio->bi_opf |= REQ_DM_POLL_LIST;
1920 /* Reset bio->bi_private to dm_io list head */
1921 *head = tmp;
1922 return 0;
1923 }
1924 return 1;
1925}
1926
1927/*
1928 *---------------------------------------------------------------
1929 * An IDR is used to keep track of allocated minor numbers.
1930 *---------------------------------------------------------------
1931 */
1932static void free_minor(int minor)
1933{
1934 spin_lock(lock: &_minor_lock);
1935 idr_remove(&_minor_idr, id: minor);
1936 spin_unlock(lock: &_minor_lock);
1937}
1938
1939/*
1940 * See if the device with a specific minor # is free.
1941 */
1942static int specific_minor(int minor)
1943{
1944 int r;
1945
1946 if (minor >= (1 << MINORBITS))
1947 return -EINVAL;
1948
1949 idr_preload(GFP_KERNEL);
1950 spin_lock(lock: &_minor_lock);
1951
1952 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, start: minor, end: minor + 1, GFP_NOWAIT);
1953
1954 spin_unlock(lock: &_minor_lock);
1955 idr_preload_end();
1956 if (r < 0)
1957 return r == -ENOSPC ? -EBUSY : r;
1958 return 0;
1959}
1960
1961static int next_free_minor(int *minor)
1962{
1963 int r;
1964
1965 idr_preload(GFP_KERNEL);
1966 spin_lock(lock: &_minor_lock);
1967
1968 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, start: 0, end: 1 << MINORBITS, GFP_NOWAIT);
1969
1970 spin_unlock(lock: &_minor_lock);
1971 idr_preload_end();
1972 if (r < 0)
1973 return r;
1974 *minor = r;
1975 return 0;
1976}
1977
1978static const struct block_device_operations dm_blk_dops;
1979static const struct block_device_operations dm_rq_blk_dops;
1980static const struct dax_operations dm_dax_ops;
1981
1982static void dm_wq_work(struct work_struct *work);
1983
1984#ifdef CONFIG_BLK_INLINE_ENCRYPTION
1985static void dm_queue_destroy_crypto_profile(struct request_queue *q)
1986{
1987 dm_destroy_crypto_profile(profile: q->crypto_profile);
1988}
1989
1990#else /* CONFIG_BLK_INLINE_ENCRYPTION */
1991
1992static inline void dm_queue_destroy_crypto_profile(struct request_queue *q)
1993{
1994}
1995#endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
1996
1997static void cleanup_mapped_device(struct mapped_device *md)
1998{
1999 if (md->wq)
2000 destroy_workqueue(wq: md->wq);
2001 dm_free_md_mempools(pools: md->mempools);
2002
2003 if (md->dax_dev) {
2004 dax_remove_host(disk: md->disk);
2005 kill_dax(dax_dev: md->dax_dev);
2006 put_dax(dax_dev: md->dax_dev);
2007 md->dax_dev = NULL;
2008 }
2009
2010 dm_cleanup_zoned_dev(md);
2011 if (md->disk) {
2012 spin_lock(lock: &_minor_lock);
2013 md->disk->private_data = NULL;
2014 spin_unlock(lock: &_minor_lock);
2015 if (dm_get_md_type(md) != DM_TYPE_NONE) {
2016 struct table_device *td;
2017
2018 dm_sysfs_exit(md);
2019 list_for_each_entry(td, &md->table_devices, list) {
2020 bd_unlink_disk_holder(bdev: td->dm_dev.bdev,
2021 disk: md->disk);
2022 }
2023
2024 /*
2025 * Hold lock to make sure del_gendisk() won't concurrent
2026 * with open/close_table_device().
2027 */
2028 mutex_lock(&md->table_devices_lock);
2029 del_gendisk(gp: md->disk);
2030 mutex_unlock(lock: &md->table_devices_lock);
2031 }
2032 dm_queue_destroy_crypto_profile(q: md->queue);
2033 put_disk(disk: md->disk);
2034 }
2035
2036 if (md->pending_io) {
2037 free_percpu(pdata: md->pending_io);
2038 md->pending_io = NULL;
2039 }
2040
2041 cleanup_srcu_struct(ssp: &md->io_barrier);
2042
2043 mutex_destroy(lock: &md->suspend_lock);
2044 mutex_destroy(lock: &md->type_lock);
2045 mutex_destroy(lock: &md->table_devices_lock);
2046 mutex_destroy(lock: &md->swap_bios_lock);
2047
2048 dm_mq_cleanup_mapped_device(md);
2049}
2050
2051/*
2052 * Allocate and initialise a blank device with a given minor.
2053 */
2054static struct mapped_device *alloc_dev(int minor)
2055{
2056 int r, numa_node_id = dm_get_numa_node();
2057 struct mapped_device *md;
2058 void *old_md;
2059
2060 md = kvzalloc_node(size: sizeof(*md), GFP_KERNEL, node: numa_node_id);
2061 if (!md) {
2062 DMERR("unable to allocate device, out of memory.");
2063 return NULL;
2064 }
2065
2066 if (!try_module_get(THIS_MODULE))
2067 goto bad_module_get;
2068
2069 /* get a minor number for the dev */
2070 if (minor == DM_ANY_MINOR)
2071 r = next_free_minor(minor: &minor);
2072 else
2073 r = specific_minor(minor);
2074 if (r < 0)
2075 goto bad_minor;
2076
2077 r = init_srcu_struct(&md->io_barrier);
2078 if (r < 0)
2079 goto bad_io_barrier;
2080
2081 md->numa_node_id = numa_node_id;
2082 md->init_tio_pdu = false;
2083 md->type = DM_TYPE_NONE;
2084 mutex_init(&md->suspend_lock);
2085 mutex_init(&md->type_lock);
2086 mutex_init(&md->table_devices_lock);
2087 spin_lock_init(&md->deferred_lock);
2088 atomic_set(v: &md->holders, i: 1);
2089 atomic_set(v: &md->open_count, i: 0);
2090 atomic_set(v: &md->event_nr, i: 0);
2091 atomic_set(v: &md->uevent_seq, i: 0);
2092 INIT_LIST_HEAD(list: &md->uevent_list);
2093 INIT_LIST_HEAD(list: &md->table_devices);
2094 spin_lock_init(&md->uevent_lock);
2095
2096 /*
2097 * default to bio-based until DM table is loaded and md->type
2098 * established. If request-based table is loaded: blk-mq will
2099 * override accordingly.
2100 */
2101 md->disk = blk_alloc_disk(md->numa_node_id);
2102 if (!md->disk)
2103 goto bad;
2104 md->queue = md->disk->queue;
2105
2106 init_waitqueue_head(&md->wait);
2107 INIT_WORK(&md->work, dm_wq_work);
2108 INIT_WORK(&md->requeue_work, dm_wq_requeue_work);
2109 init_waitqueue_head(&md->eventq);
2110 init_completion(x: &md->kobj_holder.completion);
2111
2112 md->requeue_list = NULL;
2113 md->swap_bios = get_swap_bios();
2114 sema_init(sem: &md->swap_bios_semaphore, val: md->swap_bios);
2115 mutex_init(&md->swap_bios_lock);
2116
2117 md->disk->major = _major;
2118 md->disk->first_minor = minor;
2119 md->disk->minors = 1;
2120 md->disk->flags |= GENHD_FL_NO_PART;
2121 md->disk->fops = &dm_blk_dops;
2122 md->disk->private_data = md;
2123 sprintf(buf: md->disk->disk_name, fmt: "dm-%d", minor);
2124
2125 if (IS_ENABLED(CONFIG_FS_DAX)) {
2126 md->dax_dev = alloc_dax(private: md, ops: &dm_dax_ops);
2127 if (IS_ERR(ptr: md->dax_dev)) {
2128 md->dax_dev = NULL;
2129 goto bad;
2130 }
2131 set_dax_nocache(md->dax_dev);
2132 set_dax_nomc(md->dax_dev);
2133 if (dax_add_host(dax_dev: md->dax_dev, disk: md->disk))
2134 goto bad;
2135 }
2136
2137 format_dev_t(md->name, MKDEV(_major, minor));
2138
2139 md->wq = alloc_workqueue(fmt: "kdmflush/%s", flags: WQ_MEM_RECLAIM, max_active: 0, md->name);
2140 if (!md->wq)
2141 goto bad;
2142
2143 md->pending_io = alloc_percpu(unsigned long);
2144 if (!md->pending_io)
2145 goto bad;
2146
2147 r = dm_stats_init(st: &md->stats);
2148 if (r < 0)
2149 goto bad;
2150
2151 /* Populate the mapping, nobody knows we exist yet */
2152 spin_lock(lock: &_minor_lock);
2153 old_md = idr_replace(&_minor_idr, md, id: minor);
2154 spin_unlock(lock: &_minor_lock);
2155
2156 BUG_ON(old_md != MINOR_ALLOCED);
2157
2158 return md;
2159
2160bad:
2161 cleanup_mapped_device(md);
2162bad_io_barrier:
2163 free_minor(minor);
2164bad_minor:
2165 module_put(THIS_MODULE);
2166bad_module_get:
2167 kvfree(addr: md);
2168 return NULL;
2169}
2170
2171static void unlock_fs(struct mapped_device *md);
2172
2173static void free_dev(struct mapped_device *md)
2174{
2175 int minor = MINOR(disk_devt(md->disk));
2176
2177 unlock_fs(md);
2178
2179 cleanup_mapped_device(md);
2180
2181 WARN_ON_ONCE(!list_empty(&md->table_devices));
2182 dm_stats_cleanup(st: &md->stats);
2183 free_minor(minor);
2184
2185 module_put(THIS_MODULE);
2186 kvfree(addr: md);
2187}
2188
2189/*
2190 * Bind a table to the device.
2191 */
2192static void event_callback(void *context)
2193{
2194 unsigned long flags;
2195 LIST_HEAD(uevents);
2196 struct mapped_device *md = context;
2197
2198 spin_lock_irqsave(&md->uevent_lock, flags);
2199 list_splice_init(list: &md->uevent_list, head: &uevents);
2200 spin_unlock_irqrestore(lock: &md->uevent_lock, flags);
2201
2202 dm_send_uevents(events: &uevents, kobj: &disk_to_dev(md->disk)->kobj);
2203
2204 atomic_inc(v: &md->event_nr);
2205 wake_up(&md->eventq);
2206 dm_issue_global_event();
2207}
2208
2209/*
2210 * Returns old map, which caller must destroy.
2211 */
2212static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2213 struct queue_limits *limits)
2214{
2215 struct dm_table *old_map;
2216 sector_t size;
2217 int ret;
2218
2219 lockdep_assert_held(&md->suspend_lock);
2220
2221 size = dm_table_get_size(t);
2222
2223 /*
2224 * Wipe any geometry if the size of the table changed.
2225 */
2226 if (size != dm_get_size(md))
2227 memset(&md->geometry, 0, sizeof(md->geometry));
2228
2229 set_capacity(disk: md->disk, size);
2230
2231 dm_table_event_callback(t, fn: event_callback, context: md);
2232
2233 if (dm_table_request_based(t)) {
2234 /*
2235 * Leverage the fact that request-based DM targets are
2236 * immutable singletons - used to optimize dm_mq_queue_rq.
2237 */
2238 md->immutable_target = dm_table_get_immutable_target(t);
2239
2240 /*
2241 * There is no need to reload with request-based dm because the
2242 * size of front_pad doesn't change.
2243 *
2244 * Note for future: If you are to reload bioset, prep-ed
2245 * requests in the queue may refer to bio from the old bioset,
2246 * so you must walk through the queue to unprep.
2247 */
2248 if (!md->mempools) {
2249 md->mempools = t->mempools;
2250 t->mempools = NULL;
2251 }
2252 } else {
2253 /*
2254 * The md may already have mempools that need changing.
2255 * If so, reload bioset because front_pad may have changed
2256 * because a different table was loaded.
2257 */
2258 dm_free_md_mempools(pools: md->mempools);
2259 md->mempools = t->mempools;
2260 t->mempools = NULL;
2261 }
2262
2263 ret = dm_table_set_restrictions(t, q: md->queue, limits);
2264 if (ret) {
2265 old_map = ERR_PTR(error: ret);
2266 goto out;
2267 }
2268
2269 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2270 rcu_assign_pointer(md->map, (void *)t);
2271 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2272
2273 if (old_map)
2274 dm_sync_table(md);
2275out:
2276 return old_map;
2277}
2278
2279/*
2280 * Returns unbound table for the caller to free.
2281 */
2282static struct dm_table *__unbind(struct mapped_device *md)
2283{
2284 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2285
2286 if (!map)
2287 return NULL;
2288
2289 dm_table_event_callback(t: map, NULL, NULL);
2290 RCU_INIT_POINTER(md->map, NULL);
2291 dm_sync_table(md);
2292
2293 return map;
2294}
2295
2296/*
2297 * Constructor for a new device.
2298 */
2299int dm_create(int minor, struct mapped_device **result)
2300{
2301 struct mapped_device *md;
2302
2303 md = alloc_dev(minor);
2304 if (!md)
2305 return -ENXIO;
2306
2307 dm_ima_reset_data(md);
2308
2309 *result = md;
2310 return 0;
2311}
2312
2313/*
2314 * Functions to manage md->type.
2315 * All are required to hold md->type_lock.
2316 */
2317void dm_lock_md_type(struct mapped_device *md)
2318{
2319 mutex_lock(&md->type_lock);
2320}
2321
2322void dm_unlock_md_type(struct mapped_device *md)
2323{
2324 mutex_unlock(lock: &md->type_lock);
2325}
2326
2327void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2328{
2329 BUG_ON(!mutex_is_locked(&md->type_lock));
2330 md->type = type;
2331}
2332
2333enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2334{
2335 return md->type;
2336}
2337
2338struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2339{
2340 return md->immutable_target_type;
2341}
2342
2343/*
2344 * Setup the DM device's queue based on md's type
2345 */
2346int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2347{
2348 enum dm_queue_mode type = dm_table_get_type(t);
2349 struct queue_limits limits;
2350 struct table_device *td;
2351 int r;
2352
2353 switch (type) {
2354 case DM_TYPE_REQUEST_BASED:
2355 md->disk->fops = &dm_rq_blk_dops;
2356 r = dm_mq_init_request_queue(md, t);
2357 if (r) {
2358 DMERR("Cannot initialize queue for request-based dm mapped device");
2359 return r;
2360 }
2361 break;
2362 case DM_TYPE_BIO_BASED:
2363 case DM_TYPE_DAX_BIO_BASED:
2364 blk_queue_flag_set(QUEUE_FLAG_IO_STAT, q: md->queue);
2365 break;
2366 case DM_TYPE_NONE:
2367 WARN_ON_ONCE(true);
2368 break;
2369 }
2370
2371 r = dm_calculate_queue_limits(table: t, limits: &limits);
2372 if (r) {
2373 DMERR("Cannot calculate initial queue limits");
2374 return r;
2375 }
2376 r = dm_table_set_restrictions(t, q: md->queue, limits: &limits);
2377 if (r)
2378 return r;
2379
2380 /*
2381 * Hold lock to make sure add_disk() and del_gendisk() won't concurrent
2382 * with open_table_device() and close_table_device().
2383 */
2384 mutex_lock(&md->table_devices_lock);
2385 r = add_disk(disk: md->disk);
2386 mutex_unlock(lock: &md->table_devices_lock);
2387 if (r)
2388 return r;
2389
2390 /*
2391 * Register the holder relationship for devices added before the disk
2392 * was live.
2393 */
2394 list_for_each_entry(td, &md->table_devices, list) {
2395 r = bd_link_disk_holder(bdev: td->dm_dev.bdev, disk: md->disk);
2396 if (r)
2397 goto out_undo_holders;
2398 }
2399
2400 r = dm_sysfs_init(md);
2401 if (r)
2402 goto out_undo_holders;
2403
2404 md->type = type;
2405 return 0;
2406
2407out_undo_holders:
2408 list_for_each_entry_continue_reverse(td, &md->table_devices, list)
2409 bd_unlink_disk_holder(bdev: td->dm_dev.bdev, disk: md->disk);
2410 mutex_lock(&md->table_devices_lock);
2411 del_gendisk(gp: md->disk);
2412 mutex_unlock(lock: &md->table_devices_lock);
2413 return r;
2414}
2415
2416struct mapped_device *dm_get_md(dev_t dev)
2417{
2418 struct mapped_device *md;
2419 unsigned int minor = MINOR(dev);
2420
2421 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2422 return NULL;
2423
2424 spin_lock(lock: &_minor_lock);
2425
2426 md = idr_find(&_minor_idr, id: minor);
2427 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2428 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2429 md = NULL;
2430 goto out;
2431 }
2432 dm_get(md);
2433out:
2434 spin_unlock(lock: &_minor_lock);
2435
2436 return md;
2437}
2438EXPORT_SYMBOL_GPL(dm_get_md);
2439
2440void *dm_get_mdptr(struct mapped_device *md)
2441{
2442 return md->interface_ptr;
2443}
2444
2445void dm_set_mdptr(struct mapped_device *md, void *ptr)
2446{
2447 md->interface_ptr = ptr;
2448}
2449
2450void dm_get(struct mapped_device *md)
2451{
2452 atomic_inc(v: &md->holders);
2453 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2454}
2455
2456int dm_hold(struct mapped_device *md)
2457{
2458 spin_lock(lock: &_minor_lock);
2459 if (test_bit(DMF_FREEING, &md->flags)) {
2460 spin_unlock(lock: &_minor_lock);
2461 return -EBUSY;
2462 }
2463 dm_get(md);
2464 spin_unlock(lock: &_minor_lock);
2465 return 0;
2466}
2467EXPORT_SYMBOL_GPL(dm_hold);
2468
2469const char *dm_device_name(struct mapped_device *md)
2470{
2471 return md->name;
2472}
2473EXPORT_SYMBOL_GPL(dm_device_name);
2474
2475static void __dm_destroy(struct mapped_device *md, bool wait)
2476{
2477 struct dm_table *map;
2478 int srcu_idx;
2479
2480 might_sleep();
2481
2482 spin_lock(lock: &_minor_lock);
2483 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2484 set_bit(DMF_FREEING, addr: &md->flags);
2485 spin_unlock(lock: &_minor_lock);
2486
2487 blk_mark_disk_dead(disk: md->disk);
2488
2489 /*
2490 * Take suspend_lock so that presuspend and postsuspend methods
2491 * do not race with internal suspend.
2492 */
2493 mutex_lock(&md->suspend_lock);
2494 map = dm_get_live_table(md, srcu_idx: &srcu_idx);
2495 if (!dm_suspended_md(md)) {
2496 dm_table_presuspend_targets(t: map);
2497 set_bit(DMF_SUSPENDED, addr: &md->flags);
2498 set_bit(DMF_POST_SUSPENDING, addr: &md->flags);
2499 dm_table_postsuspend_targets(t: map);
2500 }
2501 /* dm_put_live_table must be before fsleep, otherwise deadlock is possible */
2502 dm_put_live_table(md, srcu_idx);
2503 mutex_unlock(lock: &md->suspend_lock);
2504
2505 /*
2506 * Rare, but there may be I/O requests still going to complete,
2507 * for example. Wait for all references to disappear.
2508 * No one should increment the reference count of the mapped_device,
2509 * after the mapped_device state becomes DMF_FREEING.
2510 */
2511 if (wait)
2512 while (atomic_read(v: &md->holders))
2513 fsleep(usecs: 1000);
2514 else if (atomic_read(v: &md->holders))
2515 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2516 dm_device_name(md), atomic_read(&md->holders));
2517
2518 dm_table_destroy(t: __unbind(md));
2519 free_dev(md);
2520}
2521
2522void dm_destroy(struct mapped_device *md)
2523{
2524 __dm_destroy(md, wait: true);
2525}
2526
2527void dm_destroy_immediate(struct mapped_device *md)
2528{
2529 __dm_destroy(md, wait: false);
2530}
2531
2532void dm_put(struct mapped_device *md)
2533{
2534 atomic_dec(v: &md->holders);
2535}
2536EXPORT_SYMBOL_GPL(dm_put);
2537
2538static bool dm_in_flight_bios(struct mapped_device *md)
2539{
2540 int cpu;
2541 unsigned long sum = 0;
2542
2543 for_each_possible_cpu(cpu)
2544 sum += *per_cpu_ptr(md->pending_io, cpu);
2545
2546 return sum != 0;
2547}
2548
2549static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2550{
2551 int r = 0;
2552 DEFINE_WAIT(wait);
2553
2554 while (true) {
2555 prepare_to_wait(wq_head: &md->wait, wq_entry: &wait, state: task_state);
2556
2557 if (!dm_in_flight_bios(md))
2558 break;
2559
2560 if (signal_pending_state(state: task_state, current)) {
2561 r = -EINTR;
2562 break;
2563 }
2564
2565 io_schedule();
2566 }
2567 finish_wait(wq_head: &md->wait, wq_entry: &wait);
2568
2569 smp_rmb();
2570
2571 return r;
2572}
2573
2574static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2575{
2576 int r = 0;
2577
2578 if (!queue_is_mq(q: md->queue))
2579 return dm_wait_for_bios_completion(md, task_state);
2580
2581 while (true) {
2582 if (!blk_mq_queue_inflight(q: md->queue))
2583 break;
2584
2585 if (signal_pending_state(state: task_state, current)) {
2586 r = -EINTR;
2587 break;
2588 }
2589
2590 fsleep(usecs: 5000);
2591 }
2592
2593 return r;
2594}
2595
2596/*
2597 * Process the deferred bios
2598 */
2599static void dm_wq_work(struct work_struct *work)
2600{
2601 struct mapped_device *md = container_of(work, struct mapped_device, work);
2602 struct bio *bio;
2603
2604 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2605 spin_lock_irq(lock: &md->deferred_lock);
2606 bio = bio_list_pop(bl: &md->deferred);
2607 spin_unlock_irq(lock: &md->deferred_lock);
2608
2609 if (!bio)
2610 break;
2611
2612 submit_bio_noacct(bio);
2613 cond_resched();
2614 }
2615}
2616
2617static void dm_queue_flush(struct mapped_device *md)
2618{
2619 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, addr: &md->flags);
2620 smp_mb__after_atomic();
2621 queue_work(wq: md->wq, work: &md->work);
2622}
2623
2624/*
2625 * Swap in a new table, returning the old one for the caller to destroy.
2626 */
2627struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2628{
2629 struct dm_table *live_map = NULL, *map = ERR_PTR(error: -EINVAL);
2630 struct queue_limits limits;
2631 int r;
2632
2633 mutex_lock(&md->suspend_lock);
2634
2635 /* device must be suspended */
2636 if (!dm_suspended_md(md))
2637 goto out;
2638
2639 /*
2640 * If the new table has no data devices, retain the existing limits.
2641 * This helps multipath with queue_if_no_path if all paths disappear,
2642 * then new I/O is queued based on these limits, and then some paths
2643 * reappear.
2644 */
2645 if (dm_table_has_no_data_devices(table)) {
2646 live_map = dm_get_live_table_fast(md);
2647 if (live_map)
2648 limits = md->queue->limits;
2649 dm_put_live_table_fast(md);
2650 }
2651
2652 if (!live_map) {
2653 r = dm_calculate_queue_limits(table, limits: &limits);
2654 if (r) {
2655 map = ERR_PTR(error: r);
2656 goto out;
2657 }
2658 }
2659
2660 map = __bind(md, t: table, limits: &limits);
2661 dm_issue_global_event();
2662
2663out:
2664 mutex_unlock(lock: &md->suspend_lock);
2665 return map;
2666}
2667
2668/*
2669 * Functions to lock and unlock any filesystem running on the
2670 * device.
2671 */
2672static int lock_fs(struct mapped_device *md)
2673{
2674 int r;
2675
2676 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2677
2678 r = freeze_bdev(bdev: md->disk->part0);
2679 if (!r)
2680 set_bit(DMF_FROZEN, addr: &md->flags);
2681 return r;
2682}
2683
2684static void unlock_fs(struct mapped_device *md)
2685{
2686 if (!test_bit(DMF_FROZEN, &md->flags))
2687 return;
2688 thaw_bdev(bdev: md->disk->part0);
2689 clear_bit(DMF_FROZEN, addr: &md->flags);
2690}
2691
2692/*
2693 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2694 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2695 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2696 *
2697 * If __dm_suspend returns 0, the device is completely quiescent
2698 * now. There is no request-processing activity. All new requests
2699 * are being added to md->deferred list.
2700 */
2701static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2702 unsigned int suspend_flags, unsigned int task_state,
2703 int dmf_suspended_flag)
2704{
2705 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2706 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2707 int r;
2708
2709 lockdep_assert_held(&md->suspend_lock);
2710
2711 /*
2712 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2713 * This flag is cleared before dm_suspend returns.
2714 */
2715 if (noflush)
2716 set_bit(DMF_NOFLUSH_SUSPENDING, addr: &md->flags);
2717 else
2718 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2719
2720 /*
2721 * This gets reverted if there's an error later and the targets
2722 * provide the .presuspend_undo hook.
2723 */
2724 dm_table_presuspend_targets(t: map);
2725
2726 /*
2727 * Flush I/O to the device.
2728 * Any I/O submitted after lock_fs() may not be flushed.
2729 * noflush takes precedence over do_lockfs.
2730 * (lock_fs() flushes I/Os and waits for them to complete.)
2731 */
2732 if (!noflush && do_lockfs) {
2733 r = lock_fs(md);
2734 if (r) {
2735 dm_table_presuspend_undo_targets(t: map);
2736 return r;
2737 }
2738 }
2739
2740 /*
2741 * Here we must make sure that no processes are submitting requests
2742 * to target drivers i.e. no one may be executing
2743 * dm_split_and_process_bio from dm_submit_bio.
2744 *
2745 * To get all processes out of dm_split_and_process_bio in dm_submit_bio,
2746 * we take the write lock. To prevent any process from reentering
2747 * dm_split_and_process_bio from dm_submit_bio and quiesce the thread
2748 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2749 * flush_workqueue(md->wq).
2750 */
2751 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, addr: &md->flags);
2752 if (map)
2753 synchronize_srcu(ssp: &md->io_barrier);
2754
2755 /*
2756 * Stop md->queue before flushing md->wq in case request-based
2757 * dm defers requests to md->wq from md->queue.
2758 */
2759 if (dm_request_based(md))
2760 dm_stop_queue(q: md->queue);
2761
2762 flush_workqueue(md->wq);
2763
2764 /*
2765 * At this point no more requests are entering target request routines.
2766 * We call dm_wait_for_completion to wait for all existing requests
2767 * to finish.
2768 */
2769 r = dm_wait_for_completion(md, task_state);
2770 if (!r)
2771 set_bit(nr: dmf_suspended_flag, addr: &md->flags);
2772
2773 if (noflush)
2774 clear_bit(DMF_NOFLUSH_SUSPENDING, addr: &md->flags);
2775 if (map)
2776 synchronize_srcu(ssp: &md->io_barrier);
2777
2778 /* were we interrupted ? */
2779 if (r < 0) {
2780 dm_queue_flush(md);
2781
2782 if (dm_request_based(md))
2783 dm_start_queue(q: md->queue);
2784
2785 unlock_fs(md);
2786 dm_table_presuspend_undo_targets(t: map);
2787 /* pushback list is already flushed, so skip flush */
2788 }
2789
2790 return r;
2791}
2792
2793/*
2794 * We need to be able to change a mapping table under a mounted
2795 * filesystem. For example we might want to move some data in
2796 * the background. Before the table can be swapped with
2797 * dm_bind_table, dm_suspend must be called to flush any in
2798 * flight bios and ensure that any further io gets deferred.
2799 */
2800/*
2801 * Suspend mechanism in request-based dm.
2802 *
2803 * 1. Flush all I/Os by lock_fs() if needed.
2804 * 2. Stop dispatching any I/O by stopping the request_queue.
2805 * 3. Wait for all in-flight I/Os to be completed or requeued.
2806 *
2807 * To abort suspend, start the request_queue.
2808 */
2809int dm_suspend(struct mapped_device *md, unsigned int suspend_flags)
2810{
2811 struct dm_table *map = NULL;
2812 int r = 0;
2813
2814retry:
2815 mutex_lock_nested(lock: &md->suspend_lock, SINGLE_DEPTH_NESTING);
2816
2817 if (dm_suspended_md(md)) {
2818 r = -EINVAL;
2819 goto out_unlock;
2820 }
2821
2822 if (dm_suspended_internally_md(md)) {
2823 /* already internally suspended, wait for internal resume */
2824 mutex_unlock(lock: &md->suspend_lock);
2825 r = wait_on_bit(word: &md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2826 if (r)
2827 return r;
2828 goto retry;
2829 }
2830
2831 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2832 if (!map) {
2833 /* avoid deadlock with fs/namespace.c:do_mount() */
2834 suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
2835 }
2836
2837 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2838 if (r)
2839 goto out_unlock;
2840
2841 set_bit(DMF_POST_SUSPENDING, addr: &md->flags);
2842 dm_table_postsuspend_targets(t: map);
2843 clear_bit(DMF_POST_SUSPENDING, addr: &md->flags);
2844
2845out_unlock:
2846 mutex_unlock(lock: &md->suspend_lock);
2847 return r;
2848}
2849
2850static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2851{
2852 if (map) {
2853 int r = dm_table_resume_targets(t: map);
2854
2855 if (r)
2856 return r;
2857 }
2858
2859 dm_queue_flush(md);
2860
2861 /*
2862 * Flushing deferred I/Os must be done after targets are resumed
2863 * so that mapping of targets can work correctly.
2864 * Request-based dm is queueing the deferred I/Os in its request_queue.
2865 */
2866 if (dm_request_based(md))
2867 dm_start_queue(q: md->queue);
2868
2869 unlock_fs(md);
2870
2871 return 0;
2872}
2873
2874int dm_resume(struct mapped_device *md)
2875{
2876 int r;
2877 struct dm_table *map = NULL;
2878
2879retry:
2880 r = -EINVAL;
2881 mutex_lock_nested(lock: &md->suspend_lock, SINGLE_DEPTH_NESTING);
2882
2883 if (!dm_suspended_md(md))
2884 goto out;
2885
2886 if (dm_suspended_internally_md(md)) {
2887 /* already internally suspended, wait for internal resume */
2888 mutex_unlock(lock: &md->suspend_lock);
2889 r = wait_on_bit(word: &md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2890 if (r)
2891 return r;
2892 goto retry;
2893 }
2894
2895 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2896 if (!map || !dm_table_get_size(t: map))
2897 goto out;
2898
2899 r = __dm_resume(md, map);
2900 if (r)
2901 goto out;
2902
2903 clear_bit(DMF_SUSPENDED, addr: &md->flags);
2904out:
2905 mutex_unlock(lock: &md->suspend_lock);
2906
2907 return r;
2908}
2909
2910/*
2911 * Internal suspend/resume works like userspace-driven suspend. It waits
2912 * until all bios finish and prevents issuing new bios to the target drivers.
2913 * It may be used only from the kernel.
2914 */
2915
2916static void __dm_internal_suspend(struct mapped_device *md, unsigned int suspend_flags)
2917{
2918 struct dm_table *map = NULL;
2919
2920 lockdep_assert_held(&md->suspend_lock);
2921
2922 if (md->internal_suspend_count++)
2923 return; /* nested internal suspend */
2924
2925 if (dm_suspended_md(md)) {
2926 set_bit(DMF_SUSPENDED_INTERNALLY, addr: &md->flags);
2927 return; /* nest suspend */
2928 }
2929
2930 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2931
2932 /*
2933 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2934 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2935 * would require changing .presuspend to return an error -- avoid this
2936 * until there is a need for more elaborate variants of internal suspend.
2937 */
2938 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2939 DMF_SUSPENDED_INTERNALLY);
2940
2941 set_bit(DMF_POST_SUSPENDING, addr: &md->flags);
2942 dm_table_postsuspend_targets(t: map);
2943 clear_bit(DMF_POST_SUSPENDING, addr: &md->flags);
2944}
2945
2946static void __dm_internal_resume(struct mapped_device *md)
2947{
2948 BUG_ON(!md->internal_suspend_count);
2949
2950 if (--md->internal_suspend_count)
2951 return; /* resume from nested internal suspend */
2952
2953 if (dm_suspended_md(md))
2954 goto done; /* resume from nested suspend */
2955
2956 /*
2957 * NOTE: existing callers don't need to call dm_table_resume_targets
2958 * (which may fail -- so best to avoid it for now by passing NULL map)
2959 */
2960 (void) __dm_resume(md, NULL);
2961
2962done:
2963 clear_bit(DMF_SUSPENDED_INTERNALLY, addr: &md->flags);
2964 smp_mb__after_atomic();
2965 wake_up_bit(word: &md->flags, DMF_SUSPENDED_INTERNALLY);
2966}
2967
2968void dm_internal_suspend_noflush(struct mapped_device *md)
2969{
2970 mutex_lock(&md->suspend_lock);
2971 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2972 mutex_unlock(lock: &md->suspend_lock);
2973}
2974EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2975
2976void dm_internal_resume(struct mapped_device *md)
2977{
2978 mutex_lock(&md->suspend_lock);
2979 __dm_internal_resume(md);
2980 mutex_unlock(lock: &md->suspend_lock);
2981}
2982EXPORT_SYMBOL_GPL(dm_internal_resume);
2983
2984/*
2985 * Fast variants of internal suspend/resume hold md->suspend_lock,
2986 * which prevents interaction with userspace-driven suspend.
2987 */
2988
2989void dm_internal_suspend_fast(struct mapped_device *md)
2990{
2991 mutex_lock(&md->suspend_lock);
2992 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2993 return;
2994
2995 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, addr: &md->flags);
2996 synchronize_srcu(ssp: &md->io_barrier);
2997 flush_workqueue(md->wq);
2998 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
2999}
3000EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3001
3002void dm_internal_resume_fast(struct mapped_device *md)
3003{
3004 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3005 goto done;
3006
3007 dm_queue_flush(md);
3008
3009done:
3010 mutex_unlock(lock: &md->suspend_lock);
3011}
3012EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3013
3014/*
3015 *---------------------------------------------------------------
3016 * Event notification.
3017 *---------------------------------------------------------------
3018 */
3019int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3020 unsigned int cookie, bool need_resize_uevent)
3021{
3022 int r;
3023 unsigned int noio_flag;
3024 char udev_cookie[DM_COOKIE_LENGTH];
3025 char *envp[3] = { NULL, NULL, NULL };
3026 char **envpp = envp;
3027 if (cookie) {
3028 snprintf(buf: udev_cookie, DM_COOKIE_LENGTH, fmt: "%s=%u",
3029 DM_COOKIE_ENV_VAR_NAME, cookie);
3030 *envpp++ = udev_cookie;
3031 }
3032 if (need_resize_uevent) {
3033 *envpp++ = "RESIZE=1";
3034 }
3035
3036 noio_flag = memalloc_noio_save();
3037
3038 r = kobject_uevent_env(kobj: &disk_to_dev(md->disk)->kobj, action, envp);
3039
3040 memalloc_noio_restore(flags: noio_flag);
3041
3042 return r;
3043}
3044
3045uint32_t dm_next_uevent_seq(struct mapped_device *md)
3046{
3047 return atomic_add_return(i: 1, v: &md->uevent_seq);
3048}
3049
3050uint32_t dm_get_event_nr(struct mapped_device *md)
3051{
3052 return atomic_read(v: &md->event_nr);
3053}
3054
3055int dm_wait_event(struct mapped_device *md, int event_nr)
3056{
3057 return wait_event_interruptible(md->eventq,
3058 (event_nr != atomic_read(&md->event_nr)));
3059}
3060
3061void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3062{
3063 unsigned long flags;
3064
3065 spin_lock_irqsave(&md->uevent_lock, flags);
3066 list_add(new: elist, head: &md->uevent_list);
3067 spin_unlock_irqrestore(lock: &md->uevent_lock, flags);
3068}
3069
3070/*
3071 * The gendisk is only valid as long as you have a reference
3072 * count on 'md'.
3073 */
3074struct gendisk *dm_disk(struct mapped_device *md)
3075{
3076 return md->disk;
3077}
3078EXPORT_SYMBOL_GPL(dm_disk);
3079
3080struct kobject *dm_kobject(struct mapped_device *md)
3081{
3082 return &md->kobj_holder.kobj;
3083}
3084
3085struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3086{
3087 struct mapped_device *md;
3088
3089 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3090
3091 spin_lock(lock: &_minor_lock);
3092 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
3093 md = NULL;
3094 goto out;
3095 }
3096 dm_get(md);
3097out:
3098 spin_unlock(lock: &_minor_lock);
3099
3100 return md;
3101}
3102
3103int dm_suspended_md(struct mapped_device *md)
3104{
3105 return test_bit(DMF_SUSPENDED, &md->flags);
3106}
3107
3108static int dm_post_suspending_md(struct mapped_device *md)
3109{
3110 return test_bit(DMF_POST_SUSPENDING, &md->flags);
3111}
3112
3113int dm_suspended_internally_md(struct mapped_device *md)
3114{
3115 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3116}
3117
3118int dm_test_deferred_remove_flag(struct mapped_device *md)
3119{
3120 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3121}
3122
3123int dm_suspended(struct dm_target *ti)
3124{
3125 return dm_suspended_md(md: ti->table->md);
3126}
3127EXPORT_SYMBOL_GPL(dm_suspended);
3128
3129int dm_post_suspending(struct dm_target *ti)
3130{
3131 return dm_post_suspending_md(md: ti->table->md);
3132}
3133EXPORT_SYMBOL_GPL(dm_post_suspending);
3134
3135int dm_noflush_suspending(struct dm_target *ti)
3136{
3137 return __noflush_suspending(md: ti->table->md);
3138}
3139EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3140
3141void dm_free_md_mempools(struct dm_md_mempools *pools)
3142{
3143 if (!pools)
3144 return;
3145
3146 bioset_exit(&pools->bs);
3147 bioset_exit(&pools->io_bs);
3148
3149 kfree(objp: pools);
3150}
3151
3152struct dm_pr {
3153 u64 old_key;
3154 u64 new_key;
3155 u32 flags;
3156 bool abort;
3157 bool fail_early;
3158 int ret;
3159 enum pr_type type;
3160 struct pr_keys *read_keys;
3161 struct pr_held_reservation *rsv;
3162};
3163
3164static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3165 struct dm_pr *pr)
3166{
3167 struct mapped_device *md = bdev->bd_disk->private_data;
3168 struct dm_table *table;
3169 struct dm_target *ti;
3170 int ret = -ENOTTY, srcu_idx;
3171
3172 table = dm_get_live_table(md, srcu_idx: &srcu_idx);
3173 if (!table || !dm_table_get_size(t: table))
3174 goto out;
3175
3176 /* We only support devices that have a single target */
3177 if (table->num_targets != 1)
3178 goto out;
3179 ti = dm_table_get_target(t: table, index: 0);
3180
3181 if (dm_suspended_md(md)) {
3182 ret = -EAGAIN;
3183 goto out;
3184 }
3185
3186 ret = -EINVAL;
3187 if (!ti->type->iterate_devices)
3188 goto out;
3189
3190 ti->type->iterate_devices(ti, fn, pr);
3191 ret = 0;
3192out:
3193 dm_put_live_table(md, srcu_idx);
3194 return ret;
3195}
3196
3197/*
3198 * For register / unregister we need to manually call out to every path.
3199 */
3200static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3201 sector_t start, sector_t len, void *data)
3202{
3203 struct dm_pr *pr = data;
3204 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3205 int ret;
3206
3207 if (!ops || !ops->pr_register) {
3208 pr->ret = -EOPNOTSUPP;
3209 return -1;
3210 }
3211
3212 ret = ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3213 if (!ret)
3214 return 0;
3215
3216 if (!pr->ret)
3217 pr->ret = ret;
3218
3219 if (pr->fail_early)
3220 return -1;
3221
3222 return 0;
3223}
3224
3225static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3226 u32 flags)
3227{
3228 struct dm_pr pr = {
3229 .old_key = old_key,
3230 .new_key = new_key,
3231 .flags = flags,
3232 .fail_early = true,
3233 .ret = 0,
3234 };
3235 int ret;
3236
3237 ret = dm_call_pr(bdev, fn: __dm_pr_register, pr: &pr);
3238 if (ret) {
3239 /* Didn't even get to register a path */
3240 return ret;
3241 }
3242
3243 if (!pr.ret)
3244 return 0;
3245 ret = pr.ret;
3246
3247 if (!new_key)
3248 return ret;
3249
3250 /* unregister all paths if we failed to register any path */
3251 pr.old_key = new_key;
3252 pr.new_key = 0;
3253 pr.flags = 0;
3254 pr.fail_early = false;
3255 (void) dm_call_pr(bdev, fn: __dm_pr_register, pr: &pr);
3256 return ret;
3257}
3258
3259
3260static int __dm_pr_reserve(struct dm_target *ti, struct dm_dev *dev,
3261 sector_t start, sector_t len, void *data)
3262{
3263 struct dm_pr *pr = data;
3264 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3265
3266 if (!ops || !ops->pr_reserve) {
3267 pr->ret = -EOPNOTSUPP;
3268 return -1;
3269 }
3270
3271 pr->ret = ops->pr_reserve(dev->bdev, pr->old_key, pr->type, pr->flags);
3272 if (!pr->ret)
3273 return -1;
3274
3275 return 0;
3276}
3277
3278static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3279 u32 flags)
3280{
3281 struct dm_pr pr = {
3282 .old_key = key,
3283 .flags = flags,
3284 .type = type,
3285 .fail_early = false,
3286 .ret = 0,
3287 };
3288 int ret;
3289
3290 ret = dm_call_pr(bdev, fn: __dm_pr_reserve, pr: &pr);
3291 if (ret)
3292 return ret;
3293
3294 return pr.ret;
3295}
3296
3297/*
3298 * If there is a non-All Registrants type of reservation, the release must be
3299 * sent down the holding path. For the cases where there is no reservation or
3300 * the path is not the holder the device will also return success, so we must
3301 * try each path to make sure we got the correct path.
3302 */
3303static int __dm_pr_release(struct dm_target *ti, struct dm_dev *dev,
3304 sector_t start, sector_t len, void *data)
3305{
3306 struct dm_pr *pr = data;
3307 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3308
3309 if (!ops || !ops->pr_release) {
3310 pr->ret = -EOPNOTSUPP;
3311 return -1;
3312 }
3313
3314 pr->ret = ops->pr_release(dev->bdev, pr->old_key, pr->type);
3315 if (pr->ret)
3316 return -1;
3317
3318 return 0;
3319}
3320
3321static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3322{
3323 struct dm_pr pr = {
3324 .old_key = key,
3325 .type = type,
3326 .fail_early = false,
3327 };
3328 int ret;
3329
3330 ret = dm_call_pr(bdev, fn: __dm_pr_release, pr: &pr);
3331 if (ret)
3332 return ret;
3333
3334 return pr.ret;
3335}
3336
3337static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
3338 sector_t start, sector_t len, void *data)
3339{
3340 struct dm_pr *pr = data;
3341 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3342
3343 if (!ops || !ops->pr_preempt) {
3344 pr->ret = -EOPNOTSUPP;
3345 return -1;
3346 }
3347
3348 pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
3349 pr->abort);
3350 if (!pr->ret)
3351 return -1;
3352
3353 return 0;
3354}
3355
3356static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3357 enum pr_type type, bool abort)
3358{
3359 struct dm_pr pr = {
3360 .new_key = new_key,
3361 .old_key = old_key,
3362 .type = type,
3363 .fail_early = false,
3364 };
3365 int ret;
3366
3367 ret = dm_call_pr(bdev, fn: __dm_pr_preempt, pr: &pr);
3368 if (ret)
3369 return ret;
3370
3371 return pr.ret;
3372}
3373
3374static int dm_pr_clear(struct block_device *bdev, u64 key)
3375{
3376 struct mapped_device *md = bdev->bd_disk->private_data;
3377 const struct pr_ops *ops;
3378 int r, srcu_idx;
3379
3380 r = dm_prepare_ioctl(md, srcu_idx: &srcu_idx, bdev: &bdev);
3381 if (r < 0)
3382 goto out;
3383
3384 ops = bdev->bd_disk->fops->pr_ops;
3385 if (ops && ops->pr_clear)
3386 r = ops->pr_clear(bdev, key);
3387 else
3388 r = -EOPNOTSUPP;
3389out:
3390 dm_unprepare_ioctl(md, srcu_idx);
3391 return r;
3392}
3393
3394static int __dm_pr_read_keys(struct dm_target *ti, struct dm_dev *dev,
3395 sector_t start, sector_t len, void *data)
3396{
3397 struct dm_pr *pr = data;
3398 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3399
3400 if (!ops || !ops->pr_read_keys) {
3401 pr->ret = -EOPNOTSUPP;
3402 return -1;
3403 }
3404
3405 pr->ret = ops->pr_read_keys(dev->bdev, pr->read_keys);
3406 if (!pr->ret)
3407 return -1;
3408
3409 return 0;
3410}
3411
3412static int dm_pr_read_keys(struct block_device *bdev, struct pr_keys *keys)
3413{
3414 struct dm_pr pr = {
3415 .read_keys = keys,
3416 };
3417 int ret;
3418
3419 ret = dm_call_pr(bdev, fn: __dm_pr_read_keys, pr: &pr);
3420 if (ret)
3421 return ret;
3422
3423 return pr.ret;
3424}
3425
3426static int __dm_pr_read_reservation(struct dm_target *ti, struct dm_dev *dev,
3427 sector_t start, sector_t len, void *data)
3428{
3429 struct dm_pr *pr = data;
3430 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3431
3432 if (!ops || !ops->pr_read_reservation) {
3433 pr->ret = -EOPNOTSUPP;
3434 return -1;
3435 }
3436
3437 pr->ret = ops->pr_read_reservation(dev->bdev, pr->rsv);
3438 if (!pr->ret)
3439 return -1;
3440
3441 return 0;
3442}
3443
3444static int dm_pr_read_reservation(struct block_device *bdev,
3445 struct pr_held_reservation *rsv)
3446{
3447 struct dm_pr pr = {
3448 .rsv = rsv,
3449 };
3450 int ret;
3451
3452 ret = dm_call_pr(bdev, fn: __dm_pr_read_reservation, pr: &pr);
3453 if (ret)
3454 return ret;
3455
3456 return pr.ret;
3457}
3458
3459static const struct pr_ops dm_pr_ops = {
3460 .pr_register = dm_pr_register,
3461 .pr_reserve = dm_pr_reserve,
3462 .pr_release = dm_pr_release,
3463 .pr_preempt = dm_pr_preempt,
3464 .pr_clear = dm_pr_clear,
3465 .pr_read_keys = dm_pr_read_keys,
3466 .pr_read_reservation = dm_pr_read_reservation,
3467};
3468
3469static const struct block_device_operations dm_blk_dops = {
3470 .submit_bio = dm_submit_bio,
3471 .poll_bio = dm_poll_bio,
3472 .open = dm_blk_open,
3473 .release = dm_blk_close,
3474 .ioctl = dm_blk_ioctl,
3475 .getgeo = dm_blk_getgeo,
3476 .report_zones = dm_blk_report_zones,
3477 .pr_ops = &dm_pr_ops,
3478 .owner = THIS_MODULE
3479};
3480
3481static const struct block_device_operations dm_rq_blk_dops = {
3482 .open = dm_blk_open,
3483 .release = dm_blk_close,
3484 .ioctl = dm_blk_ioctl,
3485 .getgeo = dm_blk_getgeo,
3486 .pr_ops = &dm_pr_ops,
3487 .owner = THIS_MODULE
3488};
3489
3490static const struct dax_operations dm_dax_ops = {
3491 .direct_access = dm_dax_direct_access,
3492 .zero_page_range = dm_dax_zero_page_range,
3493 .recovery_write = dm_dax_recovery_write,
3494};
3495
3496/*
3497 * module hooks
3498 */
3499module_init(dm_init);
3500module_exit(dm_exit);
3501
3502module_param(major, uint, 0);
3503MODULE_PARM_DESC(major, "The major number of the device mapper");
3504
3505module_param(reserved_bio_based_ios, uint, 0644);
3506MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3507
3508module_param(dm_numa_node, int, 0644);
3509MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3510
3511module_param(swap_bios, int, 0644);
3512MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3513
3514MODULE_DESCRIPTION(DM_NAME " driver");
3515MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3516MODULE_LICENSE("GPL");
3517

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