Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
4 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
5 *
6 * This file is released under the GPL.
7 */
8
9#include "dm-core.h"
10#include "dm-rq.h"
11#include "dm-uevent.h"
12#include "dm-ima.h"
13
14#include <linux/bio-integrity.h>
15#include <linux/init.h>
16#include <linux/module.h>
17#include <linux/mutex.h>
18#include <linux/sched/mm.h>
19#include <linux/sched/signal.h>
20#include <linux/blkpg.h>
21#include <linux/bio.h>
22#include <linux/mempool.h>
23#include <linux/dax.h>
24#include <linux/slab.h>
25#include <linux/idr.h>
26#include <linux/uio.h>
27#include <linux/hdreg.h>
28#include <linux/delay.h>
29#include <linux/wait.h>
30#include <linux/pr.h>
31#include <linux/refcount.h>
32#include <linux/part_stat.h>
33#include <linux/blk-crypto.h>
34#include <linux/blk-crypto-profile.h>
35
36#define DM_MSG_PREFIX "core"
37
38/*
39 * Cookies are numeric values sent with CHANGE and REMOVE
40 * uevents while resuming, removing or renaming the device.
41 */
42#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
43#define DM_COOKIE_LENGTH 24
44
45/*
46 * For REQ_POLLED fs bio, this flag is set if we link mapped underlying
47 * dm_io into one list, and reuse bio->bi_private as the list head. Before
48 * ending this fs bio, we will recover its ->bi_private.
49 */
50#define REQ_DM_POLL_LIST REQ_DRV
51
52static const char *_name = DM_NAME;
53
54static unsigned int major;
55static unsigned int _major;
56
57static DEFINE_IDR(_minor_idr);
58
59static DEFINE_SPINLOCK(_minor_lock);
60
61static void do_deferred_remove(struct work_struct *w);
62
63static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
64
65static struct workqueue_struct *deferred_remove_workqueue;
66
67atomic_t dm_global_event_nr = ATOMIC_INIT(0);
68DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
69
70void dm_issue_global_event(void)
71{
72 atomic_inc(&dm_global_event_nr);
73 wake_up(&dm_global_eventq);
74}
75
76DEFINE_STATIC_KEY_FALSE(stats_enabled);
77DEFINE_STATIC_KEY_FALSE(swap_bios_enabled);
78DEFINE_STATIC_KEY_FALSE(zoned_enabled);
79
80/*
81 * One of these is allocated (on-stack) per original bio.
82 */
83struct clone_info {
84 struct dm_table *map;
85 struct bio *bio;
86 struct dm_io *io;
87 sector_t sector;
88 unsigned int sector_count;
89 bool is_abnormal_io:1;
90 bool submit_as_polled:1;
91};
92
93static inline struct dm_target_io *clone_to_tio(struct bio *clone)
94{
95 return container_of(clone, struct dm_target_io, clone);
96}
97
98void *dm_per_bio_data(struct bio *bio, size_t data_size)
99{
100 if (!dm_tio_flagged(clone_to_tio(bio), DM_TIO_INSIDE_DM_IO))
101 return (char *)bio - DM_TARGET_IO_BIO_OFFSET - data_size;
102 return (char *)bio - DM_IO_BIO_OFFSET - data_size;
103}
104EXPORT_SYMBOL_GPL(dm_per_bio_data);
105
106struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
107{
108 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
109
110 if (io->magic == DM_IO_MAGIC)
111 return (struct bio *)((char *)io + DM_IO_BIO_OFFSET);
112 BUG_ON(io->magic != DM_TIO_MAGIC);
113 return (struct bio *)((char *)io + DM_TARGET_IO_BIO_OFFSET);
114}
115EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
116
117unsigned int dm_bio_get_target_bio_nr(const struct bio *bio)
118{
119 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
120}
121EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
122
123#define MINOR_ALLOCED ((void *)-1)
124
125#define DM_NUMA_NODE NUMA_NO_NODE
126static int dm_numa_node = DM_NUMA_NODE;
127
128#define DEFAULT_SWAP_BIOS (8 * 1048576 / PAGE_SIZE)
129static int swap_bios = DEFAULT_SWAP_BIOS;
130static int get_swap_bios(void)
131{
132 int latch = READ_ONCE(swap_bios);
133
134 if (unlikely(latch <= 0))
135 latch = DEFAULT_SWAP_BIOS;
136 return latch;
137}
138
139struct table_device {
140 struct list_head list;
141 refcount_t count;
142 struct dm_dev dm_dev;
143};
144
145/*
146 * Bio-based DM's mempools' reserved IOs set by the user.
147 */
148#define RESERVED_BIO_BASED_IOS 16
149static unsigned int reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
150
151static int __dm_get_module_param_int(int *module_param, int min, int max)
152{
153 int param = READ_ONCE(*module_param);
154 int modified_param = 0;
155 bool modified = true;
156
157 if (param < min)
158 modified_param = min;
159 else if (param > max)
160 modified_param = max;
161 else
162 modified = false;
163
164 if (modified) {
165 (void)cmpxchg(module_param, param, modified_param);
166 param = modified_param;
167 }
168
169 return param;
170}
171
172unsigned int __dm_get_module_param(unsigned int *module_param, unsigned int def, unsigned int max)
173{
174 unsigned int param = READ_ONCE(*module_param);
175 unsigned int modified_param = 0;
176
177 if (!param)
178 modified_param = def;
179 else if (param > max)
180 modified_param = max;
181
182 if (modified_param) {
183 (void)cmpxchg(module_param, param, modified_param);
184 param = modified_param;
185 }
186
187 return param;
188}
189
190unsigned int dm_get_reserved_bio_based_ios(void)
191{
192 return __dm_get_module_param(&reserved_bio_based_ios,
193 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
194}
195EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
196
197static unsigned int dm_get_numa_node(void)
198{
199 return __dm_get_module_param_int(&dm_numa_node,
200 DM_NUMA_NODE, num_online_nodes() - 1);
201}
202
203static int __init local_init(void)
204{
205 int r;
206
207 r = dm_uevent_init();
208 if (r)
209 return r;
210
211 deferred_remove_workqueue = alloc_ordered_workqueue("kdmremove", 0);
212 if (!deferred_remove_workqueue) {
213 r = -ENOMEM;
214 goto out_uevent_exit;
215 }
216
217 _major = major;
218 r = register_blkdev(_major, _name);
219 if (r < 0)
220 goto out_free_workqueue;
221
222 if (!_major)
223 _major = r;
224
225 return 0;
226
227out_free_workqueue:
228 destroy_workqueue(deferred_remove_workqueue);
229out_uevent_exit:
230 dm_uevent_exit();
231
232 return r;
233}
234
235static void local_exit(void)
236{
237 destroy_workqueue(deferred_remove_workqueue);
238
239 unregister_blkdev(_major, _name);
240 dm_uevent_exit();
241
242 _major = 0;
243
244 DMINFO("cleaned up");
245}
246
247static int (*_inits[])(void) __initdata = {
248 local_init,
249 dm_target_init,
250 dm_linear_init,
251 dm_stripe_init,
252 dm_io_init,
253 dm_kcopyd_init,
254 dm_interface_init,
255 dm_statistics_init,
256};
257
258static void (*_exits[])(void) = {
259 local_exit,
260 dm_target_exit,
261 dm_linear_exit,
262 dm_stripe_exit,
263 dm_io_exit,
264 dm_kcopyd_exit,
265 dm_interface_exit,
266 dm_statistics_exit,
267};
268
269static int __init dm_init(void)
270{
271 const int count = ARRAY_SIZE(_inits);
272 int r, i;
273
274#if (IS_ENABLED(CONFIG_IMA) && !IS_ENABLED(CONFIG_IMA_DISABLE_HTABLE))
275 DMWARN("CONFIG_IMA_DISABLE_HTABLE is disabled."
276 " Duplicate IMA measurements will not be recorded in the IMA log.");
277#endif
278
279 for (i = 0; i < count; i++) {
280 r = _inits[i]();
281 if (r)
282 goto bad;
283 }
284
285 return 0;
286bad:
287 while (i--)
288 _exits[i]();
289
290 return r;
291}
292
293static void __exit dm_exit(void)
294{
295 int i = ARRAY_SIZE(_exits);
296
297 while (i--)
298 _exits[i]();
299
300 /*
301 * Should be empty by this point.
302 */
303 idr_destroy(&_minor_idr);
304}
305
306/*
307 * Block device functions
308 */
309int dm_deleting_md(struct mapped_device *md)
310{
311 return test_bit(DMF_DELETING, &md->flags);
312}
313
314static int dm_blk_open(struct gendisk *disk, blk_mode_t mode)
315{
316 struct mapped_device *md;
317
318 spin_lock(&_minor_lock);
319
320 md = disk->private_data;
321 if (!md)
322 goto out;
323
324 if (test_bit(DMF_FREEING, &md->flags) ||
325 dm_deleting_md(md)) {
326 md = NULL;
327 goto out;
328 }
329
330 dm_get(md);
331 atomic_inc(&md->open_count);
332out:
333 spin_unlock(&_minor_lock);
334
335 return md ? 0 : -ENXIO;
336}
337
338static void dm_blk_close(struct gendisk *disk)
339{
340 struct mapped_device *md;
341
342 spin_lock(&_minor_lock);
343
344 md = disk->private_data;
345 if (WARN_ON(!md))
346 goto out;
347
348 if (atomic_dec_and_test(&md->open_count) &&
349 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
350 queue_work(deferred_remove_workqueue, &deferred_remove_work);
351
352 dm_put(md);
353out:
354 spin_unlock(&_minor_lock);
355}
356
357int dm_open_count(struct mapped_device *md)
358{
359 return atomic_read(&md->open_count);
360}
361
362/*
363 * Guarantees nothing is using the device before it's deleted.
364 */
365int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
366{
367 int r = 0;
368
369 spin_lock(&_minor_lock);
370
371 if (dm_open_count(md)) {
372 r = -EBUSY;
373 if (mark_deferred)
374 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
375 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
376 r = -EEXIST;
377 else
378 set_bit(DMF_DELETING, &md->flags);
379
380 spin_unlock(&_minor_lock);
381
382 return r;
383}
384
385int dm_cancel_deferred_remove(struct mapped_device *md)
386{
387 int r = 0;
388
389 spin_lock(&_minor_lock);
390
391 if (test_bit(DMF_DELETING, &md->flags))
392 r = -EBUSY;
393 else
394 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
395
396 spin_unlock(&_minor_lock);
397
398 return r;
399}
400
401static void do_deferred_remove(struct work_struct *w)
402{
403 dm_deferred_remove();
404}
405
406static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
407{
408 struct mapped_device *md = bdev->bd_disk->private_data;
409
410 return dm_get_geometry(md, geo);
411}
412
413static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
414 struct block_device **bdev)
415{
416 struct dm_target *ti;
417 struct dm_table *map;
418 int r;
419
420retry:
421 r = -ENOTTY;
422 map = dm_get_live_table(md, srcu_idx);
423 if (!map || !dm_table_get_size(map))
424 return r;
425
426 /* We only support devices that have a single target */
427 if (map->num_targets != 1)
428 return r;
429
430 ti = dm_table_get_target(map, 0);
431 if (!ti->type->prepare_ioctl)
432 return r;
433
434 if (dm_suspended_md(md))
435 return -EAGAIN;
436
437 r = ti->type->prepare_ioctl(ti, bdev);
438 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
439 dm_put_live_table(md, *srcu_idx);
440 fsleep(10000);
441 goto retry;
442 }
443
444 return r;
445}
446
447static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
448{
449 dm_put_live_table(md, srcu_idx);
450}
451
452static int dm_blk_ioctl(struct block_device *bdev, blk_mode_t mode,
453 unsigned int cmd, unsigned long arg)
454{
455 struct mapped_device *md = bdev->bd_disk->private_data;
456 int r, srcu_idx;
457
458 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
459 if (r < 0)
460 goto out;
461
462 if (r > 0) {
463 /*
464 * Target determined this ioctl is being issued against a
465 * subset of the parent bdev; require extra privileges.
466 */
467 if (!capable(CAP_SYS_RAWIO)) {
468 DMDEBUG_LIMIT(
469 "%s: sending ioctl %x to DM device without required privilege.",
470 current->comm, cmd);
471 r = -ENOIOCTLCMD;
472 goto out;
473 }
474 }
475
476 if (!bdev->bd_disk->fops->ioctl)
477 r = -ENOTTY;
478 else
479 r = bdev->bd_disk->fops->ioctl(bdev, mode, cmd, arg);
480out:
481 dm_unprepare_ioctl(md, srcu_idx);
482 return r;
483}
484
485u64 dm_start_time_ns_from_clone(struct bio *bio)
486{
487 return jiffies_to_nsecs(clone_to_tio(bio)->io->start_time);
488}
489EXPORT_SYMBOL_GPL(dm_start_time_ns_from_clone);
490
491static inline bool bio_is_flush_with_data(struct bio *bio)
492{
493 return ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size);
494}
495
496static inline unsigned int dm_io_sectors(struct dm_io *io, struct bio *bio)
497{
498 /*
499 * If REQ_PREFLUSH set, don't account payload, it will be
500 * submitted (and accounted) after this flush completes.
501 */
502 if (bio_is_flush_with_data(bio))
503 return 0;
504 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
505 return io->sectors;
506 return bio_sectors(bio);
507}
508
509static void dm_io_acct(struct dm_io *io, bool end)
510{
511 struct bio *bio = io->orig_bio;
512
513 if (dm_io_flagged(io, DM_IO_BLK_STAT)) {
514 if (!end)
515 bdev_start_io_acct(bio->bi_bdev, bio_op(bio),
516 io->start_time);
517 else
518 bdev_end_io_acct(bio->bi_bdev, bio_op(bio),
519 dm_io_sectors(io, bio),
520 io->start_time);
521 }
522
523 if (static_branch_unlikely(&stats_enabled) &&
524 unlikely(dm_stats_used(&io->md->stats))) {
525 sector_t sector;
526
527 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
528 sector = bio_end_sector(bio) - io->sector_offset;
529 else
530 sector = bio->bi_iter.bi_sector;
531
532 dm_stats_account_io(&io->md->stats, bio_data_dir(bio),
533 sector, dm_io_sectors(io, bio),
534 end, io->start_time, &io->stats_aux);
535 }
536}
537
538static void __dm_start_io_acct(struct dm_io *io)
539{
540 dm_io_acct(io, false);
541}
542
543static void dm_start_io_acct(struct dm_io *io, struct bio *clone)
544{
545 /*
546 * Ensure IO accounting is only ever started once.
547 */
548 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
549 return;
550
551 /* Expect no possibility for race unless DM_TIO_IS_DUPLICATE_BIO. */
552 if (!clone || likely(dm_tio_is_normal(clone_to_tio(clone)))) {
553 dm_io_set_flag(io, DM_IO_ACCOUNTED);
554 } else {
555 unsigned long flags;
556 /* Can afford locking given DM_TIO_IS_DUPLICATE_BIO */
557 spin_lock_irqsave(&io->lock, flags);
558 if (dm_io_flagged(io, DM_IO_ACCOUNTED)) {
559 spin_unlock_irqrestore(&io->lock, flags);
560 return;
561 }
562 dm_io_set_flag(io, DM_IO_ACCOUNTED);
563 spin_unlock_irqrestore(&io->lock, flags);
564 }
565
566 __dm_start_io_acct(io);
567}
568
569static void dm_end_io_acct(struct dm_io *io)
570{
571 dm_io_acct(io, true);
572}
573
574static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio, gfp_t gfp_mask)
575{
576 struct dm_io *io;
577 struct dm_target_io *tio;
578 struct bio *clone;
579
580 clone = bio_alloc_clone(NULL, bio, gfp_mask, &md->mempools->io_bs);
581 if (unlikely(!clone))
582 return NULL;
583 tio = clone_to_tio(clone);
584 tio->flags = 0;
585 dm_tio_set_flag(tio, DM_TIO_INSIDE_DM_IO);
586 tio->io = NULL;
587
588 io = container_of(tio, struct dm_io, tio);
589 io->magic = DM_IO_MAGIC;
590 io->status = BLK_STS_OK;
591
592 /* one ref is for submission, the other is for completion */
593 atomic_set(&io->io_count, 2);
594 this_cpu_inc(*md->pending_io);
595 io->orig_bio = bio;
596 io->md = md;
597 spin_lock_init(&io->lock);
598 io->start_time = jiffies;
599 io->flags = 0;
600 if (blk_queue_io_stat(md->queue))
601 dm_io_set_flag(io, DM_IO_BLK_STAT);
602
603 if (static_branch_unlikely(&stats_enabled) &&
604 unlikely(dm_stats_used(&md->stats)))
605 dm_stats_record_start(&md->stats, &io->stats_aux);
606
607 return io;
608}
609
610static void free_io(struct dm_io *io)
611{
612 bio_put(&io->tio.clone);
613}
614
615static struct bio *alloc_tio(struct clone_info *ci, struct dm_target *ti,
616 unsigned int target_bio_nr, unsigned int *len, gfp_t gfp_mask)
617{
618 struct mapped_device *md = ci->io->md;
619 struct dm_target_io *tio;
620 struct bio *clone;
621
622 if (!ci->io->tio.io) {
623 /* the dm_target_io embedded in ci->io is available */
624 tio = &ci->io->tio;
625 /* alloc_io() already initialized embedded clone */
626 clone = &tio->clone;
627 } else {
628 clone = bio_alloc_clone(NULL, ci->bio, gfp_mask,
629 &md->mempools->bs);
630 if (!clone)
631 return NULL;
632
633 /* REQ_DM_POLL_LIST shouldn't be inherited */
634 clone->bi_opf &= ~REQ_DM_POLL_LIST;
635
636 tio = clone_to_tio(clone);
637 tio->flags = 0; /* also clears DM_TIO_INSIDE_DM_IO */
638 }
639
640 tio->magic = DM_TIO_MAGIC;
641 tio->io = ci->io;
642 tio->ti = ti;
643 tio->target_bio_nr = target_bio_nr;
644 tio->len_ptr = len;
645 tio->old_sector = 0;
646
647 /* Set default bdev, but target must bio_set_dev() before issuing IO */
648 clone->bi_bdev = md->disk->part0;
649 if (likely(ti != NULL) && unlikely(ti->needs_bio_set_dev))
650 bio_set_dev(clone, md->disk->part0);
651
652 if (len) {
653 clone->bi_iter.bi_size = to_bytes(*len);
654 if (bio_integrity(clone))
655 bio_integrity_trim(clone);
656 }
657
658 return clone;
659}
660
661static void free_tio(struct bio *clone)
662{
663 if (dm_tio_flagged(clone_to_tio(clone), DM_TIO_INSIDE_DM_IO))
664 return;
665 bio_put(clone);
666}
667
668/*
669 * Add the bio to the list of deferred io.
670 */
671static void queue_io(struct mapped_device *md, struct bio *bio)
672{
673 unsigned long flags;
674
675 spin_lock_irqsave(&md->deferred_lock, flags);
676 bio_list_add(&md->deferred, bio);
677 spin_unlock_irqrestore(&md->deferred_lock, flags);
678 queue_work(md->wq, &md->work);
679}
680
681/*
682 * Everyone (including functions in this file), should use this
683 * function to access the md->map field, and make sure they call
684 * dm_put_live_table() when finished.
685 */
686struct dm_table *dm_get_live_table(struct mapped_device *md,
687 int *srcu_idx) __acquires(md->io_barrier)
688{
689 *srcu_idx = srcu_read_lock(&md->io_barrier);
690
691 return srcu_dereference(md->map, &md->io_barrier);
692}
693
694void dm_put_live_table(struct mapped_device *md,
695 int srcu_idx) __releases(md->io_barrier)
696{
697 srcu_read_unlock(&md->io_barrier, srcu_idx);
698}
699
700void dm_sync_table(struct mapped_device *md)
701{
702 synchronize_srcu(&md->io_barrier);
703 synchronize_rcu_expedited();
704}
705
706/*
707 * A fast alternative to dm_get_live_table/dm_put_live_table.
708 * The caller must not block between these two functions.
709 */
710static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
711{
712 rcu_read_lock();
713 return rcu_dereference(md->map);
714}
715
716static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
717{
718 rcu_read_unlock();
719}
720
721static char *_dm_claim_ptr = "I belong to device-mapper";
722
723/*
724 * Open a table device so we can use it as a map destination.
725 */
726static struct table_device *open_table_device(struct mapped_device *md,
727 dev_t dev, blk_mode_t mode)
728{
729 struct table_device *td;
730 struct file *bdev_file;
731 struct block_device *bdev;
732 u64 part_off;
733 int r;
734
735 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
736 if (!td)
737 return ERR_PTR(-ENOMEM);
738 refcount_set(&td->count, 1);
739
740 bdev_file = bdev_file_open_by_dev(dev, mode, _dm_claim_ptr, NULL);
741 if (IS_ERR(bdev_file)) {
742 r = PTR_ERR(bdev_file);
743 goto out_free_td;
744 }
745
746 bdev = file_bdev(bdev_file);
747
748 /*
749 * We can be called before the dm disk is added. In that case we can't
750 * register the holder relation here. It will be done once add_disk was
751 * called.
752 */
753 if (md->disk->slave_dir) {
754 r = bd_link_disk_holder(bdev, md->disk);
755 if (r)
756 goto out_blkdev_put;
757 }
758
759 td->dm_dev.mode = mode;
760 td->dm_dev.bdev = bdev;
761 td->dm_dev.bdev_file = bdev_file;
762 td->dm_dev.dax_dev = fs_dax_get_by_bdev(bdev, &part_off,
763 NULL, NULL);
764 format_dev_t(td->dm_dev.name, dev);
765 list_add(&td->list, &md->table_devices);
766 return td;
767
768out_blkdev_put:
769 __fput_sync(bdev_file);
770out_free_td:
771 kfree(td);
772 return ERR_PTR(r);
773}
774
775/*
776 * Close a table device that we've been using.
777 */
778static void close_table_device(struct table_device *td, struct mapped_device *md)
779{
780 if (md->disk->slave_dir)
781 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
782
783 /* Leverage async fput() if DMF_DEFERRED_REMOVE set */
784 if (unlikely(test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
785 fput(td->dm_dev.bdev_file);
786 else
787 __fput_sync(td->dm_dev.bdev_file);
788
789 put_dax(td->dm_dev.dax_dev);
790 list_del(&td->list);
791 kfree(td);
792}
793
794static struct table_device *find_table_device(struct list_head *l, dev_t dev,
795 blk_mode_t mode)
796{
797 struct table_device *td;
798
799 list_for_each_entry(td, l, list)
800 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
801 return td;
802
803 return NULL;
804}
805
806int dm_get_table_device(struct mapped_device *md, dev_t dev, blk_mode_t mode,
807 struct dm_dev **result)
808{
809 struct table_device *td;
810
811 mutex_lock(&md->table_devices_lock);
812 td = find_table_device(&md->table_devices, dev, mode);
813 if (!td) {
814 td = open_table_device(md, dev, mode);
815 if (IS_ERR(td)) {
816 mutex_unlock(&md->table_devices_lock);
817 return PTR_ERR(td);
818 }
819 } else {
820 refcount_inc(&td->count);
821 }
822 mutex_unlock(&md->table_devices_lock);
823
824 *result = &td->dm_dev;
825 return 0;
826}
827
828void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
829{
830 struct table_device *td = container_of(d, struct table_device, dm_dev);
831
832 mutex_lock(&md->table_devices_lock);
833 if (refcount_dec_and_test(&td->count))
834 close_table_device(td, md);
835 mutex_unlock(&md->table_devices_lock);
836}
837
838/*
839 * Get the geometry associated with a dm device
840 */
841int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
842{
843 *geo = md->geometry;
844
845 return 0;
846}
847
848/*
849 * Set the geometry of a device.
850 */
851int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
852{
853 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
854
855 if (geo->start > sz) {
856 DMERR("Start sector is beyond the geometry limits.");
857 return -EINVAL;
858 }
859
860 md->geometry = *geo;
861
862 return 0;
863}
864
865static int __noflush_suspending(struct mapped_device *md)
866{
867 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
868}
869
870static void dm_requeue_add_io(struct dm_io *io, bool first_stage)
871{
872 struct mapped_device *md = io->md;
873
874 if (first_stage) {
875 struct dm_io *next = md->requeue_list;
876
877 md->requeue_list = io;
878 io->next = next;
879 } else {
880 bio_list_add_head(&md->deferred, io->orig_bio);
881 }
882}
883
884static void dm_kick_requeue(struct mapped_device *md, bool first_stage)
885{
886 if (first_stage)
887 queue_work(md->wq, &md->requeue_work);
888 else
889 queue_work(md->wq, &md->work);
890}
891
892/*
893 * Return true if the dm_io's original bio is requeued.
894 * io->status is updated with error if requeue disallowed.
895 */
896static bool dm_handle_requeue(struct dm_io *io, bool first_stage)
897{
898 struct bio *bio = io->orig_bio;
899 bool handle_requeue = (io->status == BLK_STS_DM_REQUEUE);
900 bool handle_polled_eagain = ((io->status == BLK_STS_AGAIN) &&
901 (bio->bi_opf & REQ_POLLED));
902 struct mapped_device *md = io->md;
903 bool requeued = false;
904
905 if (handle_requeue || handle_polled_eagain) {
906 unsigned long flags;
907
908 if (bio->bi_opf & REQ_POLLED) {
909 /*
910 * Upper layer won't help us poll split bio
911 * (io->orig_bio may only reflect a subset of the
912 * pre-split original) so clear REQ_POLLED.
913 */
914 bio_clear_polled(bio);
915 }
916
917 /*
918 * Target requested pushing back the I/O or
919 * polled IO hit BLK_STS_AGAIN.
920 */
921 spin_lock_irqsave(&md->deferred_lock, flags);
922 if ((__noflush_suspending(md) &&
923 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) ||
924 handle_polled_eagain || first_stage) {
925 dm_requeue_add_io(io, first_stage);
926 requeued = true;
927 } else {
928 /*
929 * noflush suspend was interrupted or this is
930 * a write to a zoned target.
931 */
932 io->status = BLK_STS_IOERR;
933 }
934 spin_unlock_irqrestore(&md->deferred_lock, flags);
935 }
936
937 if (requeued)
938 dm_kick_requeue(md, first_stage);
939
940 return requeued;
941}
942
943static void __dm_io_complete(struct dm_io *io, bool first_stage)
944{
945 struct bio *bio = io->orig_bio;
946 struct mapped_device *md = io->md;
947 blk_status_t io_error;
948 bool requeued;
949
950 requeued = dm_handle_requeue(io, first_stage);
951 if (requeued && first_stage)
952 return;
953
954 io_error = io->status;
955 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
956 dm_end_io_acct(io);
957 else if (!io_error) {
958 /*
959 * Must handle target that DM_MAPIO_SUBMITTED only to
960 * then bio_endio() rather than dm_submit_bio_remap()
961 */
962 __dm_start_io_acct(io);
963 dm_end_io_acct(io);
964 }
965 free_io(io);
966 smp_wmb();
967 this_cpu_dec(*md->pending_io);
968
969 /* nudge anyone waiting on suspend queue */
970 if (unlikely(wq_has_sleeper(&md->wait)))
971 wake_up(&md->wait);
972
973 /* Return early if the original bio was requeued */
974 if (requeued)
975 return;
976
977 if (bio_is_flush_with_data(bio)) {
978 /*
979 * Preflush done for flush with data, reissue
980 * without REQ_PREFLUSH.
981 */
982 bio->bi_opf &= ~REQ_PREFLUSH;
983 queue_io(md, bio);
984 } else {
985 /* done with normal IO or empty flush */
986 if (io_error)
987 bio->bi_status = io_error;
988 bio_endio(bio);
989 }
990}
991
992static void dm_wq_requeue_work(struct work_struct *work)
993{
994 struct mapped_device *md = container_of(work, struct mapped_device,
995 requeue_work);
996 unsigned long flags;
997 struct dm_io *io;
998
999 /* reuse deferred lock to simplify dm_handle_requeue */
1000 spin_lock_irqsave(&md->deferred_lock, flags);
1001 io = md->requeue_list;
1002 md->requeue_list = NULL;
1003 spin_unlock_irqrestore(&md->deferred_lock, flags);
1004
1005 while (io) {
1006 struct dm_io *next = io->next;
1007
1008 dm_io_rewind(io, &md->disk->bio_split);
1009
1010 io->next = NULL;
1011 __dm_io_complete(io, false);
1012 io = next;
1013 cond_resched();
1014 }
1015}
1016
1017/*
1018 * Two staged requeue:
1019 *
1020 * 1) io->orig_bio points to the real original bio, and the part mapped to
1021 * this io must be requeued, instead of other parts of the original bio.
1022 *
1023 * 2) io->orig_bio points to new cloned bio which matches the requeued dm_io.
1024 */
1025static void dm_io_complete(struct dm_io *io)
1026{
1027 bool first_requeue;
1028
1029 /*
1030 * Only dm_io that has been split needs two stage requeue, otherwise
1031 * we may run into long bio clone chain during suspend and OOM could
1032 * be triggered.
1033 *
1034 * Also flush data dm_io won't be marked as DM_IO_WAS_SPLIT, so they
1035 * also aren't handled via the first stage requeue.
1036 */
1037 if (dm_io_flagged(io, DM_IO_WAS_SPLIT))
1038 first_requeue = true;
1039 else
1040 first_requeue = false;
1041
1042 __dm_io_complete(io, first_requeue);
1043}
1044
1045/*
1046 * Decrements the number of outstanding ios that a bio has been
1047 * cloned into, completing the original io if necc.
1048 */
1049static inline void __dm_io_dec_pending(struct dm_io *io)
1050{
1051 if (atomic_dec_and_test(&io->io_count))
1052 dm_io_complete(io);
1053}
1054
1055static void dm_io_set_error(struct dm_io *io, blk_status_t error)
1056{
1057 unsigned long flags;
1058
1059 /* Push-back supersedes any I/O errors */
1060 spin_lock_irqsave(&io->lock, flags);
1061 if (!(io->status == BLK_STS_DM_REQUEUE &&
1062 __noflush_suspending(io->md))) {
1063 io->status = error;
1064 }
1065 spin_unlock_irqrestore(&io->lock, flags);
1066}
1067
1068static void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
1069{
1070 if (unlikely(error))
1071 dm_io_set_error(io, error);
1072
1073 __dm_io_dec_pending(io);
1074}
1075
1076/*
1077 * The queue_limits are only valid as long as you have a reference
1078 * count on 'md'. But _not_ imposing verification to avoid atomic_read(),
1079 */
1080static inline struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
1081{
1082 return &md->queue->limits;
1083}
1084
1085void disable_discard(struct mapped_device *md)
1086{
1087 struct queue_limits *limits = dm_get_queue_limits(md);
1088
1089 /* device doesn't really support DISCARD, disable it */
1090 limits->max_hw_discard_sectors = 0;
1091}
1092
1093void disable_write_zeroes(struct mapped_device *md)
1094{
1095 struct queue_limits *limits = dm_get_queue_limits(md);
1096
1097 /* device doesn't really support WRITE ZEROES, disable it */
1098 limits->max_write_zeroes_sectors = 0;
1099}
1100
1101static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
1102{
1103 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
1104}
1105
1106static void clone_endio(struct bio *bio)
1107{
1108 blk_status_t error = bio->bi_status;
1109 struct dm_target_io *tio = clone_to_tio(bio);
1110 struct dm_target *ti = tio->ti;
1111 dm_endio_fn endio = likely(ti != NULL) ? ti->type->end_io : NULL;
1112 struct dm_io *io = tio->io;
1113 struct mapped_device *md = io->md;
1114
1115 if (unlikely(error == BLK_STS_TARGET)) {
1116 if (bio_op(bio) == REQ_OP_DISCARD &&
1117 !bdev_max_discard_sectors(bio->bi_bdev))
1118 disable_discard(md);
1119 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
1120 !bdev_write_zeroes_sectors(bio->bi_bdev))
1121 disable_write_zeroes(md);
1122 }
1123
1124 if (static_branch_unlikely(&zoned_enabled) &&
1125 unlikely(bdev_is_zoned(bio->bi_bdev)))
1126 dm_zone_endio(io, bio);
1127
1128 if (endio) {
1129 int r = endio(ti, bio, &error);
1130
1131 switch (r) {
1132 case DM_ENDIO_REQUEUE:
1133 if (static_branch_unlikely(&zoned_enabled)) {
1134 /*
1135 * Requeuing writes to a sequential zone of a zoned
1136 * target will break the sequential write pattern:
1137 * fail such IO.
1138 */
1139 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
1140 error = BLK_STS_IOERR;
1141 else
1142 error = BLK_STS_DM_REQUEUE;
1143 } else
1144 error = BLK_STS_DM_REQUEUE;
1145 fallthrough;
1146 case DM_ENDIO_DONE:
1147 break;
1148 case DM_ENDIO_INCOMPLETE:
1149 /* The target will handle the io */
1150 return;
1151 default:
1152 DMCRIT("unimplemented target endio return value: %d", r);
1153 BUG();
1154 }
1155 }
1156
1157 if (static_branch_unlikely(&swap_bios_enabled) &&
1158 likely(ti != NULL) && unlikely(swap_bios_limit(ti, bio)))
1159 up(&md->swap_bios_semaphore);
1160
1161 free_tio(bio);
1162 dm_io_dec_pending(io, error);
1163}
1164
1165/*
1166 * Return maximum size of I/O possible at the supplied sector up to the current
1167 * target boundary.
1168 */
1169static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
1170 sector_t target_offset)
1171{
1172 return ti->len - target_offset;
1173}
1174
1175static sector_t __max_io_len(struct dm_target *ti, sector_t sector,
1176 unsigned int max_granularity,
1177 unsigned int max_sectors)
1178{
1179 sector_t target_offset = dm_target_offset(ti, sector);
1180 sector_t len = max_io_len_target_boundary(ti, target_offset);
1181
1182 /*
1183 * Does the target need to split IO even further?
1184 * - varied (per target) IO splitting is a tenet of DM; this
1185 * explains why stacked chunk_sectors based splitting via
1186 * bio_split_to_limits() isn't possible here.
1187 */
1188 if (!max_granularity)
1189 return len;
1190 return min_t(sector_t, len,
1191 min(max_sectors ? : queue_max_sectors(ti->table->md->queue),
1192 blk_boundary_sectors_left(target_offset, max_granularity)));
1193}
1194
1195static inline sector_t max_io_len(struct dm_target *ti, sector_t sector)
1196{
1197 return __max_io_len(ti, sector, ti->max_io_len, 0);
1198}
1199
1200int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1201{
1202 if (len > UINT_MAX) {
1203 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1204 (unsigned long long)len, UINT_MAX);
1205 ti->error = "Maximum size of target IO is too large";
1206 return -EINVAL;
1207 }
1208
1209 ti->max_io_len = (uint32_t) len;
1210
1211 return 0;
1212}
1213EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1214
1215static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1216 sector_t sector, int *srcu_idx)
1217 __acquires(md->io_barrier)
1218{
1219 struct dm_table *map;
1220 struct dm_target *ti;
1221
1222 map = dm_get_live_table(md, srcu_idx);
1223 if (!map)
1224 return NULL;
1225
1226 ti = dm_table_find_target(map, sector);
1227 if (!ti)
1228 return NULL;
1229
1230 return ti;
1231}
1232
1233static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1234 long nr_pages, enum dax_access_mode mode, void **kaddr,
1235 pfn_t *pfn)
1236{
1237 struct mapped_device *md = dax_get_private(dax_dev);
1238 sector_t sector = pgoff * PAGE_SECTORS;
1239 struct dm_target *ti;
1240 long len, ret = -EIO;
1241 int srcu_idx;
1242
1243 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1244
1245 if (!ti)
1246 goto out;
1247 if (!ti->type->direct_access)
1248 goto out;
1249 len = max_io_len(ti, sector) / PAGE_SECTORS;
1250 if (len < 1)
1251 goto out;
1252 nr_pages = min(len, nr_pages);
1253 ret = ti->type->direct_access(ti, pgoff, nr_pages, mode, kaddr, pfn);
1254
1255 out:
1256 dm_put_live_table(md, srcu_idx);
1257
1258 return ret;
1259}
1260
1261static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1262 size_t nr_pages)
1263{
1264 struct mapped_device *md = dax_get_private(dax_dev);
1265 sector_t sector = pgoff * PAGE_SECTORS;
1266 struct dm_target *ti;
1267 int ret = -EIO;
1268 int srcu_idx;
1269
1270 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1271
1272 if (!ti)
1273 goto out;
1274 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1275 /*
1276 * ->zero_page_range() is mandatory dax operation. If we are
1277 * here, something is wrong.
1278 */
1279 goto out;
1280 }
1281 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1282 out:
1283 dm_put_live_table(md, srcu_idx);
1284
1285 return ret;
1286}
1287
1288static size_t dm_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
1289 void *addr, size_t bytes, struct iov_iter *i)
1290{
1291 struct mapped_device *md = dax_get_private(dax_dev);
1292 sector_t sector = pgoff * PAGE_SECTORS;
1293 struct dm_target *ti;
1294 int srcu_idx;
1295 long ret = 0;
1296
1297 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1298 if (!ti || !ti->type->dax_recovery_write)
1299 goto out;
1300
1301 ret = ti->type->dax_recovery_write(ti, pgoff, addr, bytes, i);
1302out:
1303 dm_put_live_table(md, srcu_idx);
1304 return ret;
1305}
1306
1307/*
1308 * A target may call dm_accept_partial_bio only from the map routine. It is
1309 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1310 * operations, REQ_OP_ZONE_APPEND (zone append writes) and any bio serviced by
1311 * __send_duplicate_bios().
1312 *
1313 * dm_accept_partial_bio informs the dm that the target only wants to process
1314 * additional n_sectors sectors of the bio and the rest of the data should be
1315 * sent in a next bio.
1316 *
1317 * A diagram that explains the arithmetics:
1318 * +--------------------+---------------+-------+
1319 * | 1 | 2 | 3 |
1320 * +--------------------+---------------+-------+
1321 *
1322 * <-------------- *tio->len_ptr --------------->
1323 * <----- bio_sectors ----->
1324 * <-- n_sectors -->
1325 *
1326 * Region 1 was already iterated over with bio_advance or similar function.
1327 * (it may be empty if the target doesn't use bio_advance)
1328 * Region 2 is the remaining bio size that the target wants to process.
1329 * (it may be empty if region 1 is non-empty, although there is no reason
1330 * to make it empty)
1331 * The target requires that region 3 is to be sent in the next bio.
1332 *
1333 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1334 * the partially processed part (the sum of regions 1+2) must be the same for all
1335 * copies of the bio.
1336 */
1337void dm_accept_partial_bio(struct bio *bio, unsigned int n_sectors)
1338{
1339 struct dm_target_io *tio = clone_to_tio(bio);
1340 struct dm_io *io = tio->io;
1341 unsigned int bio_sectors = bio_sectors(bio);
1342
1343 BUG_ON(dm_tio_flagged(tio, DM_TIO_IS_DUPLICATE_BIO));
1344 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1345 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1346 BUG_ON(bio_sectors > *tio->len_ptr);
1347 BUG_ON(n_sectors > bio_sectors);
1348
1349 *tio->len_ptr -= bio_sectors - n_sectors;
1350 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1351
1352 /*
1353 * __split_and_process_bio() may have already saved mapped part
1354 * for accounting but it is being reduced so update accordingly.
1355 */
1356 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1357 io->sectors = n_sectors;
1358 io->sector_offset = bio_sectors(io->orig_bio);
1359}
1360EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1361
1362/*
1363 * @clone: clone bio that DM core passed to target's .map function
1364 * @tgt_clone: clone of @clone bio that target needs submitted
1365 *
1366 * Targets should use this interface to submit bios they take
1367 * ownership of when returning DM_MAPIO_SUBMITTED.
1368 *
1369 * Target should also enable ti->accounts_remapped_io
1370 */
1371void dm_submit_bio_remap(struct bio *clone, struct bio *tgt_clone)
1372{
1373 struct dm_target_io *tio = clone_to_tio(clone);
1374 struct dm_io *io = tio->io;
1375
1376 /* establish bio that will get submitted */
1377 if (!tgt_clone)
1378 tgt_clone = clone;
1379
1380 /*
1381 * Account io->origin_bio to DM dev on behalf of target
1382 * that took ownership of IO with DM_MAPIO_SUBMITTED.
1383 */
1384 dm_start_io_acct(io, clone);
1385
1386 trace_block_bio_remap(tgt_clone, disk_devt(io->md->disk),
1387 tio->old_sector);
1388 submit_bio_noacct(tgt_clone);
1389}
1390EXPORT_SYMBOL_GPL(dm_submit_bio_remap);
1391
1392static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1393{
1394 mutex_lock(&md->swap_bios_lock);
1395 while (latch < md->swap_bios) {
1396 cond_resched();
1397 down(&md->swap_bios_semaphore);
1398 md->swap_bios--;
1399 }
1400 while (latch > md->swap_bios) {
1401 cond_resched();
1402 up(&md->swap_bios_semaphore);
1403 md->swap_bios++;
1404 }
1405 mutex_unlock(&md->swap_bios_lock);
1406}
1407
1408static void __map_bio(struct bio *clone)
1409{
1410 struct dm_target_io *tio = clone_to_tio(clone);
1411 struct dm_target *ti = tio->ti;
1412 struct dm_io *io = tio->io;
1413 struct mapped_device *md = io->md;
1414 int r;
1415
1416 clone->bi_end_io = clone_endio;
1417
1418 /*
1419 * Map the clone.
1420 */
1421 tio->old_sector = clone->bi_iter.bi_sector;
1422
1423 if (static_branch_unlikely(&swap_bios_enabled) &&
1424 unlikely(swap_bios_limit(ti, clone))) {
1425 int latch = get_swap_bios();
1426
1427 if (unlikely(latch != md->swap_bios))
1428 __set_swap_bios_limit(md, latch);
1429 down(&md->swap_bios_semaphore);
1430 }
1431
1432 if (likely(ti->type->map == linear_map))
1433 r = linear_map(ti, clone);
1434 else if (ti->type->map == stripe_map)
1435 r = stripe_map(ti, clone);
1436 else
1437 r = ti->type->map(ti, clone);
1438
1439 switch (r) {
1440 case DM_MAPIO_SUBMITTED:
1441 /* target has assumed ownership of this io */
1442 if (!ti->accounts_remapped_io)
1443 dm_start_io_acct(io, clone);
1444 break;
1445 case DM_MAPIO_REMAPPED:
1446 dm_submit_bio_remap(clone, NULL);
1447 break;
1448 case DM_MAPIO_KILL:
1449 case DM_MAPIO_REQUEUE:
1450 if (static_branch_unlikely(&swap_bios_enabled) &&
1451 unlikely(swap_bios_limit(ti, clone)))
1452 up(&md->swap_bios_semaphore);
1453 free_tio(clone);
1454 if (r == DM_MAPIO_KILL)
1455 dm_io_dec_pending(io, BLK_STS_IOERR);
1456 else
1457 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1458 break;
1459 default:
1460 DMCRIT("unimplemented target map return value: %d", r);
1461 BUG();
1462 }
1463}
1464
1465static void setup_split_accounting(struct clone_info *ci, unsigned int len)
1466{
1467 struct dm_io *io = ci->io;
1468
1469 if (ci->sector_count > len) {
1470 /*
1471 * Split needed, save the mapped part for accounting.
1472 * NOTE: dm_accept_partial_bio() will update accordingly.
1473 */
1474 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1475 io->sectors = len;
1476 io->sector_offset = bio_sectors(ci->bio);
1477 }
1478}
1479
1480static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1481 struct dm_target *ti, unsigned int num_bios,
1482 unsigned *len, gfp_t gfp_flag)
1483{
1484 struct bio *bio;
1485 int try = (gfp_flag & GFP_NOWAIT) ? 0 : 1;
1486
1487 for (; try < 2; try++) {
1488 int bio_nr;
1489
1490 if (try && num_bios > 1)
1491 mutex_lock(&ci->io->md->table_devices_lock);
1492 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1493 bio = alloc_tio(ci, ti, bio_nr, len,
1494 try ? GFP_NOIO : GFP_NOWAIT);
1495 if (!bio)
1496 break;
1497
1498 bio_list_add(blist, bio);
1499 }
1500 if (try && num_bios > 1)
1501 mutex_unlock(&ci->io->md->table_devices_lock);
1502 if (bio_nr == num_bios)
1503 return;
1504
1505 while ((bio = bio_list_pop(blist)))
1506 free_tio(bio);
1507 }
1508}
1509
1510static unsigned int __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1511 unsigned int num_bios, unsigned int *len,
1512 gfp_t gfp_flag)
1513{
1514 struct bio_list blist = BIO_EMPTY_LIST;
1515 struct bio *clone;
1516 unsigned int ret = 0;
1517
1518 if (WARN_ON_ONCE(num_bios == 0)) /* num_bios = 0 is a bug in caller */
1519 return 0;
1520
1521 /* dm_accept_partial_bio() is not supported with shared tio->len_ptr */
1522 if (len)
1523 setup_split_accounting(ci, *len);
1524
1525 /*
1526 * Using alloc_multiple_bios(), even if num_bios is 1, to consistently
1527 * support allocating using GFP_NOWAIT with GFP_NOIO fallback.
1528 */
1529 alloc_multiple_bios(&blist, ci, ti, num_bios, len, gfp_flag);
1530 while ((clone = bio_list_pop(&blist))) {
1531 if (num_bios > 1)
1532 dm_tio_set_flag(clone_to_tio(clone), DM_TIO_IS_DUPLICATE_BIO);
1533 __map_bio(clone);
1534 ret += 1;
1535 }
1536
1537 return ret;
1538}
1539
1540static void __send_empty_flush(struct clone_info *ci)
1541{
1542 struct dm_table *t = ci->map;
1543 struct bio flush_bio;
1544
1545 /*
1546 * Use an on-stack bio for this, it's safe since we don't
1547 * need to reference it after submit. It's just used as
1548 * the basis for the clone(s).
1549 */
1550 bio_init(&flush_bio, ci->io->md->disk->part0, NULL, 0,
1551 REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
1552
1553 ci->bio = &flush_bio;
1554 ci->sector_count = 0;
1555 ci->io->tio.clone.bi_iter.bi_size = 0;
1556
1557 if (!t->flush_bypasses_map) {
1558 for (unsigned int i = 0; i < t->num_targets; i++) {
1559 unsigned int bios;
1560 struct dm_target *ti = dm_table_get_target(t, i);
1561
1562 if (unlikely(ti->num_flush_bios == 0))
1563 continue;
1564
1565 atomic_add(ti->num_flush_bios, &ci->io->io_count);
1566 bios = __send_duplicate_bios(ci, ti, ti->num_flush_bios,
1567 NULL, GFP_NOWAIT);
1568 atomic_sub(ti->num_flush_bios - bios, &ci->io->io_count);
1569 }
1570 } else {
1571 /*
1572 * Note that there's no need to grab t->devices_lock here
1573 * because the targets that support flush optimization don't
1574 * modify the list of devices.
1575 */
1576 struct list_head *devices = dm_table_get_devices(t);
1577 unsigned int len = 0;
1578 struct dm_dev_internal *dd;
1579 list_for_each_entry(dd, devices, list) {
1580 struct bio *clone;
1581 /*
1582 * Note that the structure dm_target_io is not
1583 * associated with any target (because the device may be
1584 * used by multiple targets), so we set tio->ti = NULL.
1585 * We must check for NULL in the I/O processing path, to
1586 * avoid NULL pointer dereference.
1587 */
1588 clone = alloc_tio(ci, NULL, 0, &len, GFP_NOIO);
1589 atomic_add(1, &ci->io->io_count);
1590 bio_set_dev(clone, dd->dm_dev->bdev);
1591 clone->bi_end_io = clone_endio;
1592 dm_submit_bio_remap(clone, NULL);
1593 }
1594 }
1595
1596 /*
1597 * alloc_io() takes one extra reference for submission, so the
1598 * reference won't reach 0 without the following subtraction
1599 */
1600 atomic_sub(1, &ci->io->io_count);
1601
1602 bio_uninit(ci->bio);
1603}
1604
1605static void __send_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1606 unsigned int num_bios, unsigned int max_granularity,
1607 unsigned int max_sectors)
1608{
1609 unsigned int len, bios;
1610
1611 len = min_t(sector_t, ci->sector_count,
1612 __max_io_len(ti, ci->sector, max_granularity, max_sectors));
1613
1614 atomic_add(num_bios, &ci->io->io_count);
1615 bios = __send_duplicate_bios(ci, ti, num_bios, &len, GFP_NOIO);
1616 /*
1617 * alloc_io() takes one extra reference for submission, so the
1618 * reference won't reach 0 without the following (+1) subtraction
1619 */
1620 atomic_sub(num_bios - bios + 1, &ci->io->io_count);
1621
1622 ci->sector += len;
1623 ci->sector_count -= len;
1624}
1625
1626static bool is_abnormal_io(struct bio *bio)
1627{
1628 switch (bio_op(bio)) {
1629 case REQ_OP_READ:
1630 case REQ_OP_WRITE:
1631 case REQ_OP_FLUSH:
1632 return false;
1633 case REQ_OP_DISCARD:
1634 case REQ_OP_SECURE_ERASE:
1635 case REQ_OP_WRITE_ZEROES:
1636 case REQ_OP_ZONE_RESET_ALL:
1637 return true;
1638 default:
1639 return false;
1640 }
1641}
1642
1643static blk_status_t __process_abnormal_io(struct clone_info *ci,
1644 struct dm_target *ti)
1645{
1646 unsigned int num_bios = 0;
1647 unsigned int max_granularity = 0;
1648 unsigned int max_sectors = 0;
1649 struct queue_limits *limits = dm_get_queue_limits(ti->table->md);
1650
1651 switch (bio_op(ci->bio)) {
1652 case REQ_OP_DISCARD:
1653 num_bios = ti->num_discard_bios;
1654 max_sectors = limits->max_discard_sectors;
1655 if (ti->max_discard_granularity)
1656 max_granularity = max_sectors;
1657 break;
1658 case REQ_OP_SECURE_ERASE:
1659 num_bios = ti->num_secure_erase_bios;
1660 max_sectors = limits->max_secure_erase_sectors;
1661 break;
1662 case REQ_OP_WRITE_ZEROES:
1663 num_bios = ti->num_write_zeroes_bios;
1664 max_sectors = limits->max_write_zeroes_sectors;
1665 break;
1666 default:
1667 break;
1668 }
1669
1670 /*
1671 * Even though the device advertised support for this type of
1672 * request, that does not mean every target supports it, and
1673 * reconfiguration might also have changed that since the
1674 * check was performed.
1675 */
1676 if (unlikely(!num_bios))
1677 return BLK_STS_NOTSUPP;
1678
1679 __send_abnormal_io(ci, ti, num_bios, max_granularity, max_sectors);
1680
1681 return BLK_STS_OK;
1682}
1683
1684/*
1685 * Reuse ->bi_private as dm_io list head for storing all dm_io instances
1686 * associated with this bio, and this bio's bi_private needs to be
1687 * stored in dm_io->data before the reuse.
1688 *
1689 * bio->bi_private is owned by fs or upper layer, so block layer won't
1690 * touch it after splitting. Meantime it won't be changed by anyone after
1691 * bio is submitted. So this reuse is safe.
1692 */
1693static inline struct dm_io **dm_poll_list_head(struct bio *bio)
1694{
1695 return (struct dm_io **)&bio->bi_private;
1696}
1697
1698static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
1699{
1700 struct dm_io **head = dm_poll_list_head(bio);
1701
1702 if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
1703 bio->bi_opf |= REQ_DM_POLL_LIST;
1704 /*
1705 * Save .bi_private into dm_io, so that we can reuse
1706 * .bi_private as dm_io list head for storing dm_io list
1707 */
1708 io->data = bio->bi_private;
1709
1710 /* tell block layer to poll for completion */
1711 bio->bi_cookie = ~BLK_QC_T_NONE;
1712
1713 io->next = NULL;
1714 } else {
1715 /*
1716 * bio recursed due to split, reuse original poll list,
1717 * and save bio->bi_private too.
1718 */
1719 io->data = (*head)->data;
1720 io->next = *head;
1721 }
1722
1723 *head = io;
1724}
1725
1726/*
1727 * Select the correct strategy for processing a non-flush bio.
1728 */
1729static blk_status_t __split_and_process_bio(struct clone_info *ci)
1730{
1731 struct bio *clone;
1732 struct dm_target *ti;
1733 unsigned int len;
1734
1735 ti = dm_table_find_target(ci->map, ci->sector);
1736 if (unlikely(!ti))
1737 return BLK_STS_IOERR;
1738
1739 if (unlikely(ci->is_abnormal_io))
1740 return __process_abnormal_io(ci, ti);
1741
1742 /*
1743 * Only support bio polling for normal IO, and the target io is
1744 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
1745 */
1746 ci->submit_as_polled = !!(ci->bio->bi_opf & REQ_POLLED);
1747
1748 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1749 setup_split_accounting(ci, len);
1750
1751 if (unlikely(ci->bio->bi_opf & REQ_NOWAIT)) {
1752 if (unlikely(!dm_target_supports_nowait(ti->type)))
1753 return BLK_STS_NOTSUPP;
1754
1755 clone = alloc_tio(ci, ti, 0, &len, GFP_NOWAIT);
1756 if (unlikely(!clone))
1757 return BLK_STS_AGAIN;
1758 } else {
1759 clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
1760 }
1761 __map_bio(clone);
1762
1763 ci->sector += len;
1764 ci->sector_count -= len;
1765
1766 return BLK_STS_OK;
1767}
1768
1769static void init_clone_info(struct clone_info *ci, struct dm_io *io,
1770 struct dm_table *map, struct bio *bio, bool is_abnormal)
1771{
1772 ci->map = map;
1773 ci->io = io;
1774 ci->bio = bio;
1775 ci->is_abnormal_io = is_abnormal;
1776 ci->submit_as_polled = false;
1777 ci->sector = bio->bi_iter.bi_sector;
1778 ci->sector_count = bio_sectors(bio);
1779
1780 /* Shouldn't happen but sector_count was being set to 0 so... */
1781 if (static_branch_unlikely(&zoned_enabled) &&
1782 WARN_ON_ONCE(op_is_zone_mgmt(bio_op(bio)) && ci->sector_count))
1783 ci->sector_count = 0;
1784}
1785
1786#ifdef CONFIG_BLK_DEV_ZONED
1787static inline bool dm_zone_bio_needs_split(struct mapped_device *md,
1788 struct bio *bio)
1789{
1790 /*
1791 * For mapped device that need zone append emulation, we must
1792 * split any large BIO that straddles zone boundaries.
1793 */
1794 return dm_emulate_zone_append(md) && bio_straddles_zones(bio) &&
1795 !bio_flagged(bio, BIO_ZONE_WRITE_PLUGGING);
1796}
1797static inline bool dm_zone_plug_bio(struct mapped_device *md, struct bio *bio)
1798{
1799 return dm_emulate_zone_append(md) && blk_zone_plug_bio(bio, 0);
1800}
1801
1802static blk_status_t __send_zone_reset_all_emulated(struct clone_info *ci,
1803 struct dm_target *ti)
1804{
1805 struct bio_list blist = BIO_EMPTY_LIST;
1806 struct mapped_device *md = ci->io->md;
1807 unsigned int zone_sectors = md->disk->queue->limits.chunk_sectors;
1808 unsigned long *need_reset;
1809 unsigned int i, nr_zones, nr_reset;
1810 unsigned int num_bios = 0;
1811 blk_status_t sts = BLK_STS_OK;
1812 sector_t sector = ti->begin;
1813 struct bio *clone;
1814 int ret;
1815
1816 nr_zones = ti->len >> ilog2(zone_sectors);
1817 need_reset = bitmap_zalloc(nr_zones, GFP_NOIO);
1818 if (!need_reset)
1819 return BLK_STS_RESOURCE;
1820
1821 ret = dm_zone_get_reset_bitmap(md, ci->map, ti->begin,
1822 nr_zones, need_reset);
1823 if (ret) {
1824 sts = BLK_STS_IOERR;
1825 goto free_bitmap;
1826 }
1827
1828 /* If we have no zone to reset, we are done. */
1829 nr_reset = bitmap_weight(need_reset, nr_zones);
1830 if (!nr_reset)
1831 goto free_bitmap;
1832
1833 atomic_add(nr_zones, &ci->io->io_count);
1834
1835 for (i = 0; i < nr_zones; i++) {
1836
1837 if (!test_bit(i, need_reset)) {
1838 sector += zone_sectors;
1839 continue;
1840 }
1841
1842 if (bio_list_empty(&blist)) {
1843 /* This may take a while, so be nice to others */
1844 if (num_bios)
1845 cond_resched();
1846
1847 /*
1848 * We may need to reset thousands of zones, so let's
1849 * not go crazy with the clone allocation.
1850 */
1851 alloc_multiple_bios(&blist, ci, ti, min(nr_reset, 32),
1852 NULL, GFP_NOIO);
1853 }
1854
1855 /* Get a clone and change it to a regular reset operation. */
1856 clone = bio_list_pop(&blist);
1857 clone->bi_opf &= ~REQ_OP_MASK;
1858 clone->bi_opf |= REQ_OP_ZONE_RESET | REQ_SYNC;
1859 clone->bi_iter.bi_sector = sector;
1860 clone->bi_iter.bi_size = 0;
1861 __map_bio(clone);
1862
1863 sector += zone_sectors;
1864 num_bios++;
1865 nr_reset--;
1866 }
1867
1868 WARN_ON_ONCE(!bio_list_empty(&blist));
1869 atomic_sub(nr_zones - num_bios, &ci->io->io_count);
1870 ci->sector_count = 0;
1871
1872free_bitmap:
1873 bitmap_free(need_reset);
1874
1875 return sts;
1876}
1877
1878static void __send_zone_reset_all_native(struct clone_info *ci,
1879 struct dm_target *ti)
1880{
1881 unsigned int bios;
1882
1883 atomic_add(1, &ci->io->io_count);
1884 bios = __send_duplicate_bios(ci, ti, 1, NULL, GFP_NOIO);
1885 atomic_sub(1 - bios, &ci->io->io_count);
1886
1887 ci->sector_count = 0;
1888}
1889
1890static blk_status_t __send_zone_reset_all(struct clone_info *ci)
1891{
1892 struct dm_table *t = ci->map;
1893 blk_status_t sts = BLK_STS_OK;
1894
1895 for (unsigned int i = 0; i < t->num_targets; i++) {
1896 struct dm_target *ti = dm_table_get_target(t, i);
1897
1898 if (ti->zone_reset_all_supported) {
1899 __send_zone_reset_all_native(ci, ti);
1900 continue;
1901 }
1902
1903 sts = __send_zone_reset_all_emulated(ci, ti);
1904 if (sts != BLK_STS_OK)
1905 break;
1906 }
1907
1908 /* Release the reference that alloc_io() took for submission. */
1909 atomic_sub(1, &ci->io->io_count);
1910
1911 return sts;
1912}
1913
1914#else
1915static inline bool dm_zone_bio_needs_split(struct mapped_device *md,
1916 struct bio *bio)
1917{
1918 return false;
1919}
1920static inline bool dm_zone_plug_bio(struct mapped_device *md, struct bio *bio)
1921{
1922 return false;
1923}
1924static blk_status_t __send_zone_reset_all(struct clone_info *ci)
1925{
1926 return BLK_STS_NOTSUPP;
1927}
1928#endif
1929
1930/*
1931 * Entry point to split a bio into clones and submit them to the targets.
1932 */
1933static void dm_split_and_process_bio(struct mapped_device *md,
1934 struct dm_table *map, struct bio *bio)
1935{
1936 struct clone_info ci;
1937 struct dm_io *io;
1938 blk_status_t error = BLK_STS_OK;
1939 bool is_abnormal, need_split;
1940
1941 is_abnormal = is_abnormal_io(bio);
1942 if (static_branch_unlikely(&zoned_enabled)) {
1943 /* Special case REQ_OP_ZONE_RESET_ALL as it cannot be split. */
1944 need_split = (bio_op(bio) != REQ_OP_ZONE_RESET_ALL) &&
1945 (is_abnormal || dm_zone_bio_needs_split(md, bio));
1946 } else {
1947 need_split = is_abnormal;
1948 }
1949
1950 if (unlikely(need_split)) {
1951 /*
1952 * Use bio_split_to_limits() for abnormal IO (e.g. discard, etc)
1953 * otherwise associated queue_limits won't be imposed.
1954 * Also split the BIO for mapped devices needing zone append
1955 * emulation to ensure that the BIO does not cross zone
1956 * boundaries.
1957 */
1958 bio = bio_split_to_limits(bio);
1959 if (!bio)
1960 return;
1961 }
1962
1963 /*
1964 * Use the block layer zone write plugging for mapped devices that
1965 * need zone append emulation (e.g. dm-crypt).
1966 */
1967 if (static_branch_unlikely(&zoned_enabled) && dm_zone_plug_bio(md, bio))
1968 return;
1969
1970 /* Only support nowait for normal IO */
1971 if (unlikely(bio->bi_opf & REQ_NOWAIT) && !is_abnormal) {
1972 io = alloc_io(md, bio, GFP_NOWAIT);
1973 if (unlikely(!io)) {
1974 /* Unable to do anything without dm_io. */
1975 bio_wouldblock_error(bio);
1976 return;
1977 }
1978 } else {
1979 io = alloc_io(md, bio, GFP_NOIO);
1980 }
1981 init_clone_info(&ci, io, map, bio, is_abnormal);
1982
1983 if (bio->bi_opf & REQ_PREFLUSH) {
1984 __send_empty_flush(&ci);
1985 /* dm_io_complete submits any data associated with flush */
1986 goto out;
1987 }
1988
1989 if (static_branch_unlikely(&zoned_enabled) &&
1990 (bio_op(bio) == REQ_OP_ZONE_RESET_ALL)) {
1991 error = __send_zone_reset_all(&ci);
1992 goto out;
1993 }
1994
1995 error = __split_and_process_bio(&ci);
1996 if (error || !ci.sector_count)
1997 goto out;
1998 /*
1999 * Remainder must be passed to submit_bio_noacct() so it gets handled
2000 * *after* bios already submitted have been completely processed.
2001 */
2002 bio_trim(bio, io->sectors, ci.sector_count);
2003 trace_block_split(bio, bio->bi_iter.bi_sector);
2004 bio_inc_remaining(bio);
2005 submit_bio_noacct(bio);
2006out:
2007 /*
2008 * Drop the extra reference count for non-POLLED bio, and hold one
2009 * reference for POLLED bio, which will be released in dm_poll_bio
2010 *
2011 * Add every dm_io instance into the dm_io list head which is stored
2012 * in bio->bi_private, so that dm_poll_bio can poll them all.
2013 */
2014 if (error || !ci.submit_as_polled) {
2015 /*
2016 * In case of submission failure, the extra reference for
2017 * submitting io isn't consumed yet
2018 */
2019 if (error)
2020 atomic_dec(&io->io_count);
2021 dm_io_dec_pending(io, error);
2022 } else
2023 dm_queue_poll_io(bio, io);
2024}
2025
2026static void dm_submit_bio(struct bio *bio)
2027{
2028 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
2029 int srcu_idx;
2030 struct dm_table *map;
2031
2032 map = dm_get_live_table(md, &srcu_idx);
2033 if (unlikely(!map)) {
2034 DMERR_LIMIT("%s: mapping table unavailable, erroring io",
2035 dm_device_name(md));
2036 bio_io_error(bio);
2037 goto out;
2038 }
2039
2040 /* If suspended, queue this IO for later */
2041 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
2042 if (bio->bi_opf & REQ_NOWAIT)
2043 bio_wouldblock_error(bio);
2044 else if (bio->bi_opf & REQ_RAHEAD)
2045 bio_io_error(bio);
2046 else
2047 queue_io(md, bio);
2048 goto out;
2049 }
2050
2051 dm_split_and_process_bio(md, map, bio);
2052out:
2053 dm_put_live_table(md, srcu_idx);
2054}
2055
2056static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
2057 unsigned int flags)
2058{
2059 WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
2060
2061 /* don't poll if the mapped io is done */
2062 if (atomic_read(&io->io_count) > 1)
2063 bio_poll(&io->tio.clone, iob, flags);
2064
2065 /* bio_poll holds the last reference */
2066 return atomic_read(&io->io_count) == 1;
2067}
2068
2069static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
2070 unsigned int flags)
2071{
2072 struct dm_io **head = dm_poll_list_head(bio);
2073 struct dm_io *list = *head;
2074 struct dm_io *tmp = NULL;
2075 struct dm_io *curr, *next;
2076
2077 /* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
2078 if (!(bio->bi_opf & REQ_DM_POLL_LIST))
2079 return 0;
2080
2081 WARN_ON_ONCE(!list);
2082
2083 /*
2084 * Restore .bi_private before possibly completing dm_io.
2085 *
2086 * bio_poll() is only possible once @bio has been completely
2087 * submitted via submit_bio_noacct()'s depth-first submission.
2088 * So there is no dm_queue_poll_io() race associated with
2089 * clearing REQ_DM_POLL_LIST here.
2090 */
2091 bio->bi_opf &= ~REQ_DM_POLL_LIST;
2092 bio->bi_private = list->data;
2093
2094 for (curr = list, next = curr->next; curr; curr = next, next =
2095 curr ? curr->next : NULL) {
2096 if (dm_poll_dm_io(curr, iob, flags)) {
2097 /*
2098 * clone_endio() has already occurred, so no
2099 * error handling is needed here.
2100 */
2101 __dm_io_dec_pending(curr);
2102 } else {
2103 curr->next = tmp;
2104 tmp = curr;
2105 }
2106 }
2107
2108 /* Not done? */
2109 if (tmp) {
2110 bio->bi_opf |= REQ_DM_POLL_LIST;
2111 /* Reset bio->bi_private to dm_io list head */
2112 *head = tmp;
2113 return 0;
2114 }
2115 return 1;
2116}
2117
2118/*
2119 *---------------------------------------------------------------
2120 * An IDR is used to keep track of allocated minor numbers.
2121 *---------------------------------------------------------------
2122 */
2123static void free_minor(int minor)
2124{
2125 spin_lock(&_minor_lock);
2126 idr_remove(&_minor_idr, minor);
2127 spin_unlock(&_minor_lock);
2128}
2129
2130/*
2131 * See if the device with a specific minor # is free.
2132 */
2133static int specific_minor(int minor)
2134{
2135 int r;
2136
2137 if (minor >= (1 << MINORBITS))
2138 return -EINVAL;
2139
2140 idr_preload(GFP_KERNEL);
2141 spin_lock(&_minor_lock);
2142
2143 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2144
2145 spin_unlock(&_minor_lock);
2146 idr_preload_end();
2147 if (r < 0)
2148 return r == -ENOSPC ? -EBUSY : r;
2149 return 0;
2150}
2151
2152static int next_free_minor(int *minor)
2153{
2154 int r;
2155
2156 idr_preload(GFP_KERNEL);
2157 spin_lock(&_minor_lock);
2158
2159 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2160
2161 spin_unlock(&_minor_lock);
2162 idr_preload_end();
2163 if (r < 0)
2164 return r;
2165 *minor = r;
2166 return 0;
2167}
2168
2169static const struct block_device_operations dm_blk_dops;
2170static const struct block_device_operations dm_rq_blk_dops;
2171static const struct dax_operations dm_dax_ops;
2172
2173static void dm_wq_work(struct work_struct *work);
2174
2175#ifdef CONFIG_BLK_INLINE_ENCRYPTION
2176static void dm_queue_destroy_crypto_profile(struct request_queue *q)
2177{
2178 dm_destroy_crypto_profile(q->crypto_profile);
2179}
2180
2181#else /* CONFIG_BLK_INLINE_ENCRYPTION */
2182
2183static inline void dm_queue_destroy_crypto_profile(struct request_queue *q)
2184{
2185}
2186#endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
2187
2188static void cleanup_mapped_device(struct mapped_device *md)
2189{
2190 if (md->wq)
2191 destroy_workqueue(md->wq);
2192 dm_free_md_mempools(md->mempools);
2193
2194 if (md->dax_dev) {
2195 dax_remove_host(md->disk);
2196 kill_dax(md->dax_dev);
2197 put_dax(md->dax_dev);
2198 md->dax_dev = NULL;
2199 }
2200
2201 if (md->disk) {
2202 spin_lock(&_minor_lock);
2203 md->disk->private_data = NULL;
2204 spin_unlock(&_minor_lock);
2205 if (dm_get_md_type(md) != DM_TYPE_NONE) {
2206 struct table_device *td;
2207
2208 dm_sysfs_exit(md);
2209 list_for_each_entry(td, &md->table_devices, list) {
2210 bd_unlink_disk_holder(td->dm_dev.bdev,
2211 md->disk);
2212 }
2213
2214 /*
2215 * Hold lock to make sure del_gendisk() won't concurrent
2216 * with open/close_table_device().
2217 */
2218 mutex_lock(&md->table_devices_lock);
2219 del_gendisk(md->disk);
2220 mutex_unlock(&md->table_devices_lock);
2221 }
2222 dm_queue_destroy_crypto_profile(md->queue);
2223 put_disk(md->disk);
2224 }
2225
2226 if (md->pending_io) {
2227 free_percpu(md->pending_io);
2228 md->pending_io = NULL;
2229 }
2230
2231 cleanup_srcu_struct(&md->io_barrier);
2232
2233 mutex_destroy(&md->suspend_lock);
2234 mutex_destroy(&md->type_lock);
2235 mutex_destroy(&md->table_devices_lock);
2236 mutex_destroy(&md->swap_bios_lock);
2237
2238 dm_mq_cleanup_mapped_device(md);
2239}
2240
2241/*
2242 * Allocate and initialise a blank device with a given minor.
2243 */
2244static struct mapped_device *alloc_dev(int minor)
2245{
2246 int r, numa_node_id = dm_get_numa_node();
2247 struct dax_device *dax_dev;
2248 struct mapped_device *md;
2249 void *old_md;
2250
2251 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
2252 if (!md) {
2253 DMERR("unable to allocate device, out of memory.");
2254 return NULL;
2255 }
2256
2257 if (!try_module_get(THIS_MODULE))
2258 goto bad_module_get;
2259
2260 /* get a minor number for the dev */
2261 if (minor == DM_ANY_MINOR)
2262 r = next_free_minor(&minor);
2263 else
2264 r = specific_minor(minor);
2265 if (r < 0)
2266 goto bad_minor;
2267
2268 r = init_srcu_struct(&md->io_barrier);
2269 if (r < 0)
2270 goto bad_io_barrier;
2271
2272 md->numa_node_id = numa_node_id;
2273 md->init_tio_pdu = false;
2274 md->type = DM_TYPE_NONE;
2275 mutex_init(&md->suspend_lock);
2276 mutex_init(&md->type_lock);
2277 mutex_init(&md->table_devices_lock);
2278 spin_lock_init(&md->deferred_lock);
2279 atomic_set(&md->holders, 1);
2280 atomic_set(&md->open_count, 0);
2281 atomic_set(&md->event_nr, 0);
2282 atomic_set(&md->uevent_seq, 0);
2283 INIT_LIST_HEAD(&md->uevent_list);
2284 INIT_LIST_HEAD(&md->table_devices);
2285 spin_lock_init(&md->uevent_lock);
2286
2287 /*
2288 * default to bio-based until DM table is loaded and md->type
2289 * established. If request-based table is loaded: blk-mq will
2290 * override accordingly.
2291 */
2292 md->disk = blk_alloc_disk(NULL, md->numa_node_id);
2293 if (IS_ERR(md->disk)) {
2294 md->disk = NULL;
2295 goto bad;
2296 }
2297 md->queue = md->disk->queue;
2298
2299 init_waitqueue_head(&md->wait);
2300 INIT_WORK(&md->work, dm_wq_work);
2301 INIT_WORK(&md->requeue_work, dm_wq_requeue_work);
2302 init_waitqueue_head(&md->eventq);
2303 init_completion(&md->kobj_holder.completion);
2304
2305 md->requeue_list = NULL;
2306 md->swap_bios = get_swap_bios();
2307 sema_init(&md->swap_bios_semaphore, md->swap_bios);
2308 mutex_init(&md->swap_bios_lock);
2309
2310 md->disk->major = _major;
2311 md->disk->first_minor = minor;
2312 md->disk->minors = 1;
2313 md->disk->flags |= GENHD_FL_NO_PART;
2314 md->disk->fops = &dm_blk_dops;
2315 md->disk->private_data = md;
2316 sprintf(md->disk->disk_name, "dm-%d", minor);
2317
2318 dax_dev = alloc_dax(md, &dm_dax_ops);
2319 if (IS_ERR(dax_dev)) {
2320 if (PTR_ERR(dax_dev) != -EOPNOTSUPP)
2321 goto bad;
2322 } else {
2323 set_dax_nocache(dax_dev);
2324 set_dax_nomc(dax_dev);
2325 md->dax_dev = dax_dev;
2326 if (dax_add_host(dax_dev, md->disk))
2327 goto bad;
2328 }
2329
2330 format_dev_t(md->name, MKDEV(_major, minor));
2331
2332 md->wq = alloc_workqueue("kdmflush/%s", WQ_MEM_RECLAIM, 0, md->name);
2333 if (!md->wq)
2334 goto bad;
2335
2336 md->pending_io = alloc_percpu(unsigned long);
2337 if (!md->pending_io)
2338 goto bad;
2339
2340 r = dm_stats_init(&md->stats);
2341 if (r < 0)
2342 goto bad;
2343
2344 /* Populate the mapping, nobody knows we exist yet */
2345 spin_lock(&_minor_lock);
2346 old_md = idr_replace(&_minor_idr, md, minor);
2347 spin_unlock(&_minor_lock);
2348
2349 BUG_ON(old_md != MINOR_ALLOCED);
2350
2351 return md;
2352
2353bad:
2354 cleanup_mapped_device(md);
2355bad_io_barrier:
2356 free_minor(minor);
2357bad_minor:
2358 module_put(THIS_MODULE);
2359bad_module_get:
2360 kvfree(md);
2361 return NULL;
2362}
2363
2364static void unlock_fs(struct mapped_device *md);
2365
2366static void free_dev(struct mapped_device *md)
2367{
2368 int minor = MINOR(disk_devt(md->disk));
2369
2370 unlock_fs(md);
2371
2372 cleanup_mapped_device(md);
2373
2374 WARN_ON_ONCE(!list_empty(&md->table_devices));
2375 dm_stats_cleanup(&md->stats);
2376 free_minor(minor);
2377
2378 module_put(THIS_MODULE);
2379 kvfree(md);
2380}
2381
2382/*
2383 * Bind a table to the device.
2384 */
2385static void event_callback(void *context)
2386{
2387 unsigned long flags;
2388 LIST_HEAD(uevents);
2389 struct mapped_device *md = context;
2390
2391 spin_lock_irqsave(&md->uevent_lock, flags);
2392 list_splice_init(&md->uevent_list, &uevents);
2393 spin_unlock_irqrestore(&md->uevent_lock, flags);
2394
2395 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2396
2397 atomic_inc(&md->event_nr);
2398 wake_up(&md->eventq);
2399 dm_issue_global_event();
2400}
2401
2402/*
2403 * Returns old map, which caller must destroy.
2404 */
2405static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2406 struct queue_limits *limits)
2407{
2408 struct dm_table *old_map;
2409 sector_t size;
2410 int ret;
2411
2412 lockdep_assert_held(&md->suspend_lock);
2413
2414 size = dm_table_get_size(t);
2415
2416 /*
2417 * Wipe any geometry if the size of the table changed.
2418 */
2419 if (size != dm_get_size(md))
2420 memset(&md->geometry, 0, sizeof(md->geometry));
2421
2422 set_capacity(md->disk, size);
2423
2424 dm_table_event_callback(t, event_callback, md);
2425
2426 if (dm_table_request_based(t)) {
2427 /*
2428 * Leverage the fact that request-based DM targets are
2429 * immutable singletons - used to optimize dm_mq_queue_rq.
2430 */
2431 md->immutable_target = dm_table_get_immutable_target(t);
2432
2433 /*
2434 * There is no need to reload with request-based dm because the
2435 * size of front_pad doesn't change.
2436 *
2437 * Note for future: If you are to reload bioset, prep-ed
2438 * requests in the queue may refer to bio from the old bioset,
2439 * so you must walk through the queue to unprep.
2440 */
2441 if (!md->mempools) {
2442 md->mempools = t->mempools;
2443 t->mempools = NULL;
2444 }
2445 } else {
2446 /*
2447 * The md may already have mempools that need changing.
2448 * If so, reload bioset because front_pad may have changed
2449 * because a different table was loaded.
2450 */
2451 dm_free_md_mempools(md->mempools);
2452 md->mempools = t->mempools;
2453 t->mempools = NULL;
2454 }
2455
2456 ret = dm_table_set_restrictions(t, md->queue, limits);
2457 if (ret) {
2458 old_map = ERR_PTR(ret);
2459 goto out;
2460 }
2461
2462 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2463 rcu_assign_pointer(md->map, (void *)t);
2464 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2465
2466 if (old_map)
2467 dm_sync_table(md);
2468out:
2469 return old_map;
2470}
2471
2472/*
2473 * Returns unbound table for the caller to free.
2474 */
2475static struct dm_table *__unbind(struct mapped_device *md)
2476{
2477 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2478
2479 if (!map)
2480 return NULL;
2481
2482 dm_table_event_callback(map, NULL, NULL);
2483 RCU_INIT_POINTER(md->map, NULL);
2484 dm_sync_table(md);
2485
2486 return map;
2487}
2488
2489/*
2490 * Constructor for a new device.
2491 */
2492int dm_create(int minor, struct mapped_device **result)
2493{
2494 struct mapped_device *md;
2495
2496 md = alloc_dev(minor);
2497 if (!md)
2498 return -ENXIO;
2499
2500 dm_ima_reset_data(md);
2501
2502 *result = md;
2503 return 0;
2504}
2505
2506/*
2507 * Functions to manage md->type.
2508 * All are required to hold md->type_lock.
2509 */
2510void dm_lock_md_type(struct mapped_device *md)
2511{
2512 mutex_lock(&md->type_lock);
2513}
2514
2515void dm_unlock_md_type(struct mapped_device *md)
2516{
2517 mutex_unlock(&md->type_lock);
2518}
2519
2520enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2521{
2522 return md->type;
2523}
2524
2525struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2526{
2527 return md->immutable_target_type;
2528}
2529
2530/*
2531 * Setup the DM device's queue based on md's type
2532 */
2533int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2534{
2535 enum dm_queue_mode type = dm_table_get_type(t);
2536 struct queue_limits limits;
2537 struct table_device *td;
2538 int r;
2539
2540 WARN_ON_ONCE(type == DM_TYPE_NONE);
2541
2542 if (type == DM_TYPE_REQUEST_BASED) {
2543 md->disk->fops = &dm_rq_blk_dops;
2544 r = dm_mq_init_request_queue(md, t);
2545 if (r) {
2546 DMERR("Cannot initialize queue for request-based dm mapped device");
2547 return r;
2548 }
2549 }
2550
2551 r = dm_calculate_queue_limits(t, &limits);
2552 if (r) {
2553 DMERR("Cannot calculate initial queue limits");
2554 return r;
2555 }
2556 r = dm_table_set_restrictions(t, md->queue, &limits);
2557 if (r)
2558 return r;
2559
2560 /*
2561 * Hold lock to make sure add_disk() and del_gendisk() won't concurrent
2562 * with open_table_device() and close_table_device().
2563 */
2564 mutex_lock(&md->table_devices_lock);
2565 r = add_disk(md->disk);
2566 mutex_unlock(&md->table_devices_lock);
2567 if (r)
2568 return r;
2569
2570 /*
2571 * Register the holder relationship for devices added before the disk
2572 * was live.
2573 */
2574 list_for_each_entry(td, &md->table_devices, list) {
2575 r = bd_link_disk_holder(td->dm_dev.bdev, md->disk);
2576 if (r)
2577 goto out_undo_holders;
2578 }
2579
2580 r = dm_sysfs_init(md);
2581 if (r)
2582 goto out_undo_holders;
2583
2584 md->type = type;
2585 return 0;
2586
2587out_undo_holders:
2588 list_for_each_entry_continue_reverse(td, &md->table_devices, list)
2589 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
2590 mutex_lock(&md->table_devices_lock);
2591 del_gendisk(md->disk);
2592 mutex_unlock(&md->table_devices_lock);
2593 return r;
2594}
2595
2596struct mapped_device *dm_get_md(dev_t dev)
2597{
2598 struct mapped_device *md;
2599 unsigned int minor = MINOR(dev);
2600
2601 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2602 return NULL;
2603
2604 spin_lock(&_minor_lock);
2605
2606 md = idr_find(&_minor_idr, minor);
2607 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2608 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2609 md = NULL;
2610 goto out;
2611 }
2612 dm_get(md);
2613out:
2614 spin_unlock(&_minor_lock);
2615
2616 return md;
2617}
2618EXPORT_SYMBOL_GPL(dm_get_md);
2619
2620void *dm_get_mdptr(struct mapped_device *md)
2621{
2622 return md->interface_ptr;
2623}
2624
2625void dm_set_mdptr(struct mapped_device *md, void *ptr)
2626{
2627 md->interface_ptr = ptr;
2628}
2629
2630void dm_get(struct mapped_device *md)
2631{
2632 atomic_inc(&md->holders);
2633 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2634}
2635
2636int dm_hold(struct mapped_device *md)
2637{
2638 spin_lock(&_minor_lock);
2639 if (test_bit(DMF_FREEING, &md->flags)) {
2640 spin_unlock(&_minor_lock);
2641 return -EBUSY;
2642 }
2643 dm_get(md);
2644 spin_unlock(&_minor_lock);
2645 return 0;
2646}
2647EXPORT_SYMBOL_GPL(dm_hold);
2648
2649const char *dm_device_name(struct mapped_device *md)
2650{
2651 return md->name;
2652}
2653EXPORT_SYMBOL_GPL(dm_device_name);
2654
2655static void __dm_destroy(struct mapped_device *md, bool wait)
2656{
2657 struct dm_table *map;
2658 int srcu_idx;
2659
2660 might_sleep();
2661
2662 spin_lock(&_minor_lock);
2663 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2664 set_bit(DMF_FREEING, &md->flags);
2665 spin_unlock(&_minor_lock);
2666
2667 blk_mark_disk_dead(md->disk);
2668
2669 /*
2670 * Take suspend_lock so that presuspend and postsuspend methods
2671 * do not race with internal suspend.
2672 */
2673 mutex_lock(&md->suspend_lock);
2674 map = dm_get_live_table(md, &srcu_idx);
2675 if (!dm_suspended_md(md)) {
2676 dm_table_presuspend_targets(map);
2677 set_bit(DMF_SUSPENDED, &md->flags);
2678 set_bit(DMF_POST_SUSPENDING, &md->flags);
2679 dm_table_postsuspend_targets(map);
2680 }
2681 /* dm_put_live_table must be before fsleep, otherwise deadlock is possible */
2682 dm_put_live_table(md, srcu_idx);
2683 mutex_unlock(&md->suspend_lock);
2684
2685 /*
2686 * Rare, but there may be I/O requests still going to complete,
2687 * for example. Wait for all references to disappear.
2688 * No one should increment the reference count of the mapped_device,
2689 * after the mapped_device state becomes DMF_FREEING.
2690 */
2691 if (wait)
2692 while (atomic_read(&md->holders))
2693 fsleep(1000);
2694 else if (atomic_read(&md->holders))
2695 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2696 dm_device_name(md), atomic_read(&md->holders));
2697
2698 dm_table_destroy(__unbind(md));
2699 free_dev(md);
2700}
2701
2702void dm_destroy(struct mapped_device *md)
2703{
2704 __dm_destroy(md, true);
2705}
2706
2707void dm_destroy_immediate(struct mapped_device *md)
2708{
2709 __dm_destroy(md, false);
2710}
2711
2712void dm_put(struct mapped_device *md)
2713{
2714 atomic_dec(&md->holders);
2715}
2716EXPORT_SYMBOL_GPL(dm_put);
2717
2718static bool dm_in_flight_bios(struct mapped_device *md)
2719{
2720 int cpu;
2721 unsigned long sum = 0;
2722
2723 for_each_possible_cpu(cpu)
2724 sum += *per_cpu_ptr(md->pending_io, cpu);
2725
2726 return sum != 0;
2727}
2728
2729static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2730{
2731 int r = 0;
2732 DEFINE_WAIT(wait);
2733
2734 while (true) {
2735 prepare_to_wait(&md->wait, &wait, task_state);
2736
2737 if (!dm_in_flight_bios(md))
2738 break;
2739
2740 if (signal_pending_state(task_state, current)) {
2741 r = -ERESTARTSYS;
2742 break;
2743 }
2744
2745 io_schedule();
2746 }
2747 finish_wait(&md->wait, &wait);
2748
2749 smp_rmb();
2750
2751 return r;
2752}
2753
2754static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2755{
2756 int r = 0;
2757
2758 if (!queue_is_mq(md->queue))
2759 return dm_wait_for_bios_completion(md, task_state);
2760
2761 while (true) {
2762 if (!blk_mq_queue_inflight(md->queue))
2763 break;
2764
2765 if (signal_pending_state(task_state, current)) {
2766 r = -ERESTARTSYS;
2767 break;
2768 }
2769
2770 fsleep(5000);
2771 }
2772
2773 return r;
2774}
2775
2776/*
2777 * Process the deferred bios
2778 */
2779static void dm_wq_work(struct work_struct *work)
2780{
2781 struct mapped_device *md = container_of(work, struct mapped_device, work);
2782 struct bio *bio;
2783
2784 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2785 spin_lock_irq(&md->deferred_lock);
2786 bio = bio_list_pop(&md->deferred);
2787 spin_unlock_irq(&md->deferred_lock);
2788
2789 if (!bio)
2790 break;
2791
2792 submit_bio_noacct(bio);
2793 cond_resched();
2794 }
2795}
2796
2797static void dm_queue_flush(struct mapped_device *md)
2798{
2799 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2800 smp_mb__after_atomic();
2801 queue_work(md->wq, &md->work);
2802}
2803
2804/*
2805 * Swap in a new table, returning the old one for the caller to destroy.
2806 */
2807struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2808{
2809 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2810 struct queue_limits limits;
2811 int r;
2812
2813 mutex_lock(&md->suspend_lock);
2814
2815 /* device must be suspended */
2816 if (!dm_suspended_md(md))
2817 goto out;
2818
2819 /*
2820 * If the new table has no data devices, retain the existing limits.
2821 * This helps multipath with queue_if_no_path if all paths disappear,
2822 * then new I/O is queued based on these limits, and then some paths
2823 * reappear.
2824 */
2825 if (dm_table_has_no_data_devices(table)) {
2826 live_map = dm_get_live_table_fast(md);
2827 if (live_map)
2828 limits = md->queue->limits;
2829 dm_put_live_table_fast(md);
2830 }
2831
2832 if (!live_map) {
2833 r = dm_calculate_queue_limits(table, &limits);
2834 if (r) {
2835 map = ERR_PTR(r);
2836 goto out;
2837 }
2838 }
2839
2840 map = __bind(md, table, &limits);
2841 dm_issue_global_event();
2842
2843out:
2844 mutex_unlock(&md->suspend_lock);
2845 return map;
2846}
2847
2848/*
2849 * Functions to lock and unlock any filesystem running on the
2850 * device.
2851 */
2852static int lock_fs(struct mapped_device *md)
2853{
2854 int r;
2855
2856 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2857
2858 r = bdev_freeze(md->disk->part0);
2859 if (!r)
2860 set_bit(DMF_FROZEN, &md->flags);
2861 return r;
2862}
2863
2864static void unlock_fs(struct mapped_device *md)
2865{
2866 if (!test_bit(DMF_FROZEN, &md->flags))
2867 return;
2868 bdev_thaw(md->disk->part0);
2869 clear_bit(DMF_FROZEN, &md->flags);
2870}
2871
2872/*
2873 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2874 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2875 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2876 *
2877 * If __dm_suspend returns 0, the device is completely quiescent
2878 * now. There is no request-processing activity. All new requests
2879 * are being added to md->deferred list.
2880 */
2881static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2882 unsigned int suspend_flags, unsigned int task_state,
2883 int dmf_suspended_flag)
2884{
2885 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2886 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2887 int r;
2888
2889 lockdep_assert_held(&md->suspend_lock);
2890
2891 /*
2892 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2893 * This flag is cleared before dm_suspend returns.
2894 */
2895 if (noflush)
2896 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2897 else
2898 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2899
2900 /*
2901 * This gets reverted if there's an error later and the targets
2902 * provide the .presuspend_undo hook.
2903 */
2904 dm_table_presuspend_targets(map);
2905
2906 /*
2907 * Flush I/O to the device.
2908 * Any I/O submitted after lock_fs() may not be flushed.
2909 * noflush takes precedence over do_lockfs.
2910 * (lock_fs() flushes I/Os and waits for them to complete.)
2911 */
2912 if (!noflush && do_lockfs) {
2913 r = lock_fs(md);
2914 if (r) {
2915 dm_table_presuspend_undo_targets(map);
2916 return r;
2917 }
2918 }
2919
2920 /*
2921 * Here we must make sure that no processes are submitting requests
2922 * to target drivers i.e. no one may be executing
2923 * dm_split_and_process_bio from dm_submit_bio.
2924 *
2925 * To get all processes out of dm_split_and_process_bio in dm_submit_bio,
2926 * we take the write lock. To prevent any process from reentering
2927 * dm_split_and_process_bio from dm_submit_bio and quiesce the thread
2928 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2929 * flush_workqueue(md->wq).
2930 */
2931 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2932 if (map)
2933 synchronize_srcu(&md->io_barrier);
2934
2935 /*
2936 * Stop md->queue before flushing md->wq in case request-based
2937 * dm defers requests to md->wq from md->queue.
2938 */
2939 if (dm_request_based(md))
2940 dm_stop_queue(md->queue);
2941
2942 flush_workqueue(md->wq);
2943
2944 /*
2945 * At this point no more requests are entering target request routines.
2946 * We call dm_wait_for_completion to wait for all existing requests
2947 * to finish.
2948 */
2949 r = dm_wait_for_completion(md, task_state);
2950 if (!r)
2951 set_bit(dmf_suspended_flag, &md->flags);
2952
2953 if (noflush)
2954 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2955 if (map)
2956 synchronize_srcu(&md->io_barrier);
2957
2958 /* were we interrupted ? */
2959 if (r < 0) {
2960 dm_queue_flush(md);
2961
2962 if (dm_request_based(md))
2963 dm_start_queue(md->queue);
2964
2965 unlock_fs(md);
2966 dm_table_presuspend_undo_targets(map);
2967 /* pushback list is already flushed, so skip flush */
2968 }
2969
2970 return r;
2971}
2972
2973/*
2974 * We need to be able to change a mapping table under a mounted
2975 * filesystem. For example we might want to move some data in
2976 * the background. Before the table can be swapped with
2977 * dm_bind_table, dm_suspend must be called to flush any in
2978 * flight bios and ensure that any further io gets deferred.
2979 */
2980/*
2981 * Suspend mechanism in request-based dm.
2982 *
2983 * 1. Flush all I/Os by lock_fs() if needed.
2984 * 2. Stop dispatching any I/O by stopping the request_queue.
2985 * 3. Wait for all in-flight I/Os to be completed or requeued.
2986 *
2987 * To abort suspend, start the request_queue.
2988 */
2989int dm_suspend(struct mapped_device *md, unsigned int suspend_flags)
2990{
2991 struct dm_table *map = NULL;
2992 int r = 0;
2993
2994retry:
2995 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2996
2997 if (dm_suspended_md(md)) {
2998 r = -EINVAL;
2999 goto out_unlock;
3000 }
3001
3002 if (dm_suspended_internally_md(md)) {
3003 /* already internally suspended, wait for internal resume */
3004 mutex_unlock(&md->suspend_lock);
3005 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3006 if (r)
3007 return r;
3008 goto retry;
3009 }
3010
3011 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3012 if (!map) {
3013 /* avoid deadlock with fs/namespace.c:do_mount() */
3014 suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
3015 }
3016
3017 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
3018 if (r)
3019 goto out_unlock;
3020
3021 set_bit(DMF_POST_SUSPENDING, &md->flags);
3022 dm_table_postsuspend_targets(map);
3023 clear_bit(DMF_POST_SUSPENDING, &md->flags);
3024
3025out_unlock:
3026 mutex_unlock(&md->suspend_lock);
3027 return r;
3028}
3029
3030static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3031{
3032 if (map) {
3033 int r = dm_table_resume_targets(map);
3034
3035 if (r)
3036 return r;
3037 }
3038
3039 dm_queue_flush(md);
3040
3041 /*
3042 * Flushing deferred I/Os must be done after targets are resumed
3043 * so that mapping of targets can work correctly.
3044 * Request-based dm is queueing the deferred I/Os in its request_queue.
3045 */
3046 if (dm_request_based(md))
3047 dm_start_queue(md->queue);
3048
3049 unlock_fs(md);
3050
3051 return 0;
3052}
3053
3054int dm_resume(struct mapped_device *md)
3055{
3056 int r;
3057 struct dm_table *map = NULL;
3058
3059retry:
3060 r = -EINVAL;
3061 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3062
3063 if (!dm_suspended_md(md))
3064 goto out;
3065
3066 if (dm_suspended_internally_md(md)) {
3067 /* already internally suspended, wait for internal resume */
3068 mutex_unlock(&md->suspend_lock);
3069 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3070 if (r)
3071 return r;
3072 goto retry;
3073 }
3074
3075 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3076 if (!map || !dm_table_get_size(map))
3077 goto out;
3078
3079 r = __dm_resume(md, map);
3080 if (r)
3081 goto out;
3082
3083 clear_bit(DMF_SUSPENDED, &md->flags);
3084out:
3085 mutex_unlock(&md->suspend_lock);
3086
3087 return r;
3088}
3089
3090/*
3091 * Internal suspend/resume works like userspace-driven suspend. It waits
3092 * until all bios finish and prevents issuing new bios to the target drivers.
3093 * It may be used only from the kernel.
3094 */
3095
3096static void __dm_internal_suspend(struct mapped_device *md, unsigned int suspend_flags)
3097{
3098 struct dm_table *map = NULL;
3099
3100 lockdep_assert_held(&md->suspend_lock);
3101
3102 if (md->internal_suspend_count++)
3103 return; /* nested internal suspend */
3104
3105 if (dm_suspended_md(md)) {
3106 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3107 return; /* nest suspend */
3108 }
3109
3110 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3111
3112 /*
3113 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3114 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3115 * would require changing .presuspend to return an error -- avoid this
3116 * until there is a need for more elaborate variants of internal suspend.
3117 */
3118 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
3119 DMF_SUSPENDED_INTERNALLY);
3120
3121 set_bit(DMF_POST_SUSPENDING, &md->flags);
3122 dm_table_postsuspend_targets(map);
3123 clear_bit(DMF_POST_SUSPENDING, &md->flags);
3124}
3125
3126static void __dm_internal_resume(struct mapped_device *md)
3127{
3128 int r;
3129 struct dm_table *map;
3130
3131 BUG_ON(!md->internal_suspend_count);
3132
3133 if (--md->internal_suspend_count)
3134 return; /* resume from nested internal suspend */
3135
3136 if (dm_suspended_md(md))
3137 goto done; /* resume from nested suspend */
3138
3139 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3140 r = __dm_resume(md, map);
3141 if (r) {
3142 /*
3143 * If a preresume method of some target failed, we are in a
3144 * tricky situation. We can't return an error to the caller. We
3145 * can't fake success because then the "resume" and
3146 * "postsuspend" methods would not be paired correctly, and it
3147 * would break various targets, for example it would cause list
3148 * corruption in the "origin" target.
3149 *
3150 * So, we fake normal suspend here, to make sure that the
3151 * "resume" and "postsuspend" methods will be paired correctly.
3152 */
3153 DMERR("Preresume method failed: %d", r);
3154 set_bit(DMF_SUSPENDED, &md->flags);
3155 }
3156done:
3157 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3158 smp_mb__after_atomic();
3159 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3160}
3161
3162void dm_internal_suspend_noflush(struct mapped_device *md)
3163{
3164 mutex_lock(&md->suspend_lock);
3165 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3166 mutex_unlock(&md->suspend_lock);
3167}
3168EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3169
3170void dm_internal_resume(struct mapped_device *md)
3171{
3172 mutex_lock(&md->suspend_lock);
3173 __dm_internal_resume(md);
3174 mutex_unlock(&md->suspend_lock);
3175}
3176EXPORT_SYMBOL_GPL(dm_internal_resume);
3177
3178/*
3179 * Fast variants of internal suspend/resume hold md->suspend_lock,
3180 * which prevents interaction with userspace-driven suspend.
3181 */
3182
3183void dm_internal_suspend_fast(struct mapped_device *md)
3184{
3185 mutex_lock(&md->suspend_lock);
3186 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3187 return;
3188
3189 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3190 synchronize_srcu(&md->io_barrier);
3191 flush_workqueue(md->wq);
3192 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3193}
3194EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3195
3196void dm_internal_resume_fast(struct mapped_device *md)
3197{
3198 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3199 goto done;
3200
3201 dm_queue_flush(md);
3202
3203done:
3204 mutex_unlock(&md->suspend_lock);
3205}
3206EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3207
3208/*
3209 *---------------------------------------------------------------
3210 * Event notification.
3211 *---------------------------------------------------------------
3212 */
3213int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3214 unsigned int cookie, bool need_resize_uevent)
3215{
3216 int r;
3217 unsigned int noio_flag;
3218 char udev_cookie[DM_COOKIE_LENGTH];
3219 char *envp[3] = { NULL, NULL, NULL };
3220 char **envpp = envp;
3221 if (cookie) {
3222 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3223 DM_COOKIE_ENV_VAR_NAME, cookie);
3224 *envpp++ = udev_cookie;
3225 }
3226 if (need_resize_uevent) {
3227 *envpp++ = "RESIZE=1";
3228 }
3229
3230 noio_flag = memalloc_noio_save();
3231
3232 r = kobject_uevent_env(&disk_to_dev(md->disk)->kobj, action, envp);
3233
3234 memalloc_noio_restore(noio_flag);
3235
3236 return r;
3237}
3238
3239uint32_t dm_next_uevent_seq(struct mapped_device *md)
3240{
3241 return atomic_add_return(1, &md->uevent_seq);
3242}
3243
3244uint32_t dm_get_event_nr(struct mapped_device *md)
3245{
3246 return atomic_read(&md->event_nr);
3247}
3248
3249int dm_wait_event(struct mapped_device *md, int event_nr)
3250{
3251 return wait_event_interruptible(md->eventq,
3252 (event_nr != atomic_read(&md->event_nr)));
3253}
3254
3255void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3256{
3257 unsigned long flags;
3258
3259 spin_lock_irqsave(&md->uevent_lock, flags);
3260 list_add(elist, &md->uevent_list);
3261 spin_unlock_irqrestore(&md->uevent_lock, flags);
3262}
3263
3264/*
3265 * The gendisk is only valid as long as you have a reference
3266 * count on 'md'.
3267 */
3268struct gendisk *dm_disk(struct mapped_device *md)
3269{
3270 return md->disk;
3271}
3272EXPORT_SYMBOL_GPL(dm_disk);
3273
3274struct kobject *dm_kobject(struct mapped_device *md)
3275{
3276 return &md->kobj_holder.kobj;
3277}
3278
3279struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3280{
3281 struct mapped_device *md;
3282
3283 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3284
3285 spin_lock(&_minor_lock);
3286 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
3287 md = NULL;
3288 goto out;
3289 }
3290 dm_get(md);
3291out:
3292 spin_unlock(&_minor_lock);
3293
3294 return md;
3295}
3296
3297int dm_suspended_md(struct mapped_device *md)
3298{
3299 return test_bit(DMF_SUSPENDED, &md->flags);
3300}
3301
3302static int dm_post_suspending_md(struct mapped_device *md)
3303{
3304 return test_bit(DMF_POST_SUSPENDING, &md->flags);
3305}
3306
3307int dm_suspended_internally_md(struct mapped_device *md)
3308{
3309 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3310}
3311
3312int dm_test_deferred_remove_flag(struct mapped_device *md)
3313{
3314 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3315}
3316
3317int dm_suspended(struct dm_target *ti)
3318{
3319 return dm_suspended_md(ti->table->md);
3320}
3321EXPORT_SYMBOL_GPL(dm_suspended);
3322
3323int dm_post_suspending(struct dm_target *ti)
3324{
3325 return dm_post_suspending_md(ti->table->md);
3326}
3327EXPORT_SYMBOL_GPL(dm_post_suspending);
3328
3329int dm_noflush_suspending(struct dm_target *ti)
3330{
3331 return __noflush_suspending(ti->table->md);
3332}
3333EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3334
3335void dm_free_md_mempools(struct dm_md_mempools *pools)
3336{
3337 if (!pools)
3338 return;
3339
3340 bioset_exit(&pools->bs);
3341 bioset_exit(&pools->io_bs);
3342
3343 kfree(pools);
3344}
3345
3346struct dm_blkdev_id {
3347 u8 *id;
3348 enum blk_unique_id type;
3349};
3350
3351static int __dm_get_unique_id(struct dm_target *ti, struct dm_dev *dev,
3352 sector_t start, sector_t len, void *data)
3353{
3354 struct dm_blkdev_id *dm_id = data;
3355 const struct block_device_operations *fops = dev->bdev->bd_disk->fops;
3356
3357 if (!fops->get_unique_id)
3358 return 0;
3359
3360 return fops->get_unique_id(dev->bdev->bd_disk, dm_id->id, dm_id->type);
3361}
3362
3363/*
3364 * Allow access to get_unique_id() for the first device returning a
3365 * non-zero result. Reasonable use expects all devices to have the
3366 * same unique id.
3367 */
3368static int dm_blk_get_unique_id(struct gendisk *disk, u8 *id,
3369 enum blk_unique_id type)
3370{
3371 struct mapped_device *md = disk->private_data;
3372 struct dm_table *table;
3373 struct dm_target *ti;
3374 int ret = 0, srcu_idx;
3375
3376 struct dm_blkdev_id dm_id = {
3377 .id = id,
3378 .type = type,
3379 };
3380
3381 table = dm_get_live_table(md, &srcu_idx);
3382 if (!table || !dm_table_get_size(table))
3383 goto out;
3384
3385 /* We only support devices that have a single target */
3386 if (table->num_targets != 1)
3387 goto out;
3388 ti = dm_table_get_target(table, 0);
3389
3390 if (!ti->type->iterate_devices)
3391 goto out;
3392
3393 ret = ti->type->iterate_devices(ti, __dm_get_unique_id, &dm_id);
3394out:
3395 dm_put_live_table(md, srcu_idx);
3396 return ret;
3397}
3398
3399struct dm_pr {
3400 u64 old_key;
3401 u64 new_key;
3402 u32 flags;
3403 bool abort;
3404 bool fail_early;
3405 int ret;
3406 enum pr_type type;
3407 struct pr_keys *read_keys;
3408 struct pr_held_reservation *rsv;
3409};
3410
3411static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3412 struct dm_pr *pr)
3413{
3414 struct mapped_device *md = bdev->bd_disk->private_data;
3415 struct dm_table *table;
3416 struct dm_target *ti;
3417 int ret = -ENOTTY, srcu_idx;
3418
3419 table = dm_get_live_table(md, &srcu_idx);
3420 if (!table || !dm_table_get_size(table))
3421 goto out;
3422
3423 /* We only support devices that have a single target */
3424 if (table->num_targets != 1)
3425 goto out;
3426 ti = dm_table_get_target(table, 0);
3427
3428 if (dm_suspended_md(md)) {
3429 ret = -EAGAIN;
3430 goto out;
3431 }
3432
3433 ret = -EINVAL;
3434 if (!ti->type->iterate_devices)
3435 goto out;
3436
3437 ti->type->iterate_devices(ti, fn, pr);
3438 ret = 0;
3439out:
3440 dm_put_live_table(md, srcu_idx);
3441 return ret;
3442}
3443
3444/*
3445 * For register / unregister we need to manually call out to every path.
3446 */
3447static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3448 sector_t start, sector_t len, void *data)
3449{
3450 struct dm_pr *pr = data;
3451 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3452 int ret;
3453
3454 if (!ops || !ops->pr_register) {
3455 pr->ret = -EOPNOTSUPP;
3456 return -1;
3457 }
3458
3459 ret = ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3460 if (!ret)
3461 return 0;
3462
3463 if (!pr->ret)
3464 pr->ret = ret;
3465
3466 if (pr->fail_early)
3467 return -1;
3468
3469 return 0;
3470}
3471
3472static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3473 u32 flags)
3474{
3475 struct dm_pr pr = {
3476 .old_key = old_key,
3477 .new_key = new_key,
3478 .flags = flags,
3479 .fail_early = true,
3480 .ret = 0,
3481 };
3482 int ret;
3483
3484 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3485 if (ret) {
3486 /* Didn't even get to register a path */
3487 return ret;
3488 }
3489
3490 if (!pr.ret)
3491 return 0;
3492 ret = pr.ret;
3493
3494 if (!new_key)
3495 return ret;
3496
3497 /* unregister all paths if we failed to register any path */
3498 pr.old_key = new_key;
3499 pr.new_key = 0;
3500 pr.flags = 0;
3501 pr.fail_early = false;
3502 (void) dm_call_pr(bdev, __dm_pr_register, &pr);
3503 return ret;
3504}
3505
3506
3507static int __dm_pr_reserve(struct dm_target *ti, struct dm_dev *dev,
3508 sector_t start, sector_t len, void *data)
3509{
3510 struct dm_pr *pr = data;
3511 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3512
3513 if (!ops || !ops->pr_reserve) {
3514 pr->ret = -EOPNOTSUPP;
3515 return -1;
3516 }
3517
3518 pr->ret = ops->pr_reserve(dev->bdev, pr->old_key, pr->type, pr->flags);
3519 if (!pr->ret)
3520 return -1;
3521
3522 return 0;
3523}
3524
3525static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3526 u32 flags)
3527{
3528 struct dm_pr pr = {
3529 .old_key = key,
3530 .flags = flags,
3531 .type = type,
3532 .fail_early = false,
3533 .ret = 0,
3534 };
3535 int ret;
3536
3537 ret = dm_call_pr(bdev, __dm_pr_reserve, &pr);
3538 if (ret)
3539 return ret;
3540
3541 return pr.ret;
3542}
3543
3544/*
3545 * If there is a non-All Registrants type of reservation, the release must be
3546 * sent down the holding path. For the cases where there is no reservation or
3547 * the path is not the holder the device will also return success, so we must
3548 * try each path to make sure we got the correct path.
3549 */
3550static int __dm_pr_release(struct dm_target *ti, struct dm_dev *dev,
3551 sector_t start, sector_t len, void *data)
3552{
3553 struct dm_pr *pr = data;
3554 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3555
3556 if (!ops || !ops->pr_release) {
3557 pr->ret = -EOPNOTSUPP;
3558 return -1;
3559 }
3560
3561 pr->ret = ops->pr_release(dev->bdev, pr->old_key, pr->type);
3562 if (pr->ret)
3563 return -1;
3564
3565 return 0;
3566}
3567
3568static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3569{
3570 struct dm_pr pr = {
3571 .old_key = key,
3572 .type = type,
3573 .fail_early = false,
3574 };
3575 int ret;
3576
3577 ret = dm_call_pr(bdev, __dm_pr_release, &pr);
3578 if (ret)
3579 return ret;
3580
3581 return pr.ret;
3582}
3583
3584static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
3585 sector_t start, sector_t len, void *data)
3586{
3587 struct dm_pr *pr = data;
3588 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3589
3590 if (!ops || !ops->pr_preempt) {
3591 pr->ret = -EOPNOTSUPP;
3592 return -1;
3593 }
3594
3595 pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
3596 pr->abort);
3597 if (!pr->ret)
3598 return -1;
3599
3600 return 0;
3601}
3602
3603static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3604 enum pr_type type, bool abort)
3605{
3606 struct dm_pr pr = {
3607 .new_key = new_key,
3608 .old_key = old_key,
3609 .type = type,
3610 .fail_early = false,
3611 };
3612 int ret;
3613
3614 ret = dm_call_pr(bdev, __dm_pr_preempt, &pr);
3615 if (ret)
3616 return ret;
3617
3618 return pr.ret;
3619}
3620
3621static int dm_pr_clear(struct block_device *bdev, u64 key)
3622{
3623 struct mapped_device *md = bdev->bd_disk->private_data;
3624 const struct pr_ops *ops;
3625 int r, srcu_idx;
3626
3627 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3628 if (r < 0)
3629 goto out;
3630
3631 ops = bdev->bd_disk->fops->pr_ops;
3632 if (ops && ops->pr_clear)
3633 r = ops->pr_clear(bdev, key);
3634 else
3635 r = -EOPNOTSUPP;
3636out:
3637 dm_unprepare_ioctl(md, srcu_idx);
3638 return r;
3639}
3640
3641static int __dm_pr_read_keys(struct dm_target *ti, struct dm_dev *dev,
3642 sector_t start, sector_t len, void *data)
3643{
3644 struct dm_pr *pr = data;
3645 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3646
3647 if (!ops || !ops->pr_read_keys) {
3648 pr->ret = -EOPNOTSUPP;
3649 return -1;
3650 }
3651
3652 pr->ret = ops->pr_read_keys(dev->bdev, pr->read_keys);
3653 if (!pr->ret)
3654 return -1;
3655
3656 return 0;
3657}
3658
3659static int dm_pr_read_keys(struct block_device *bdev, struct pr_keys *keys)
3660{
3661 struct dm_pr pr = {
3662 .read_keys = keys,
3663 };
3664 int ret;
3665
3666 ret = dm_call_pr(bdev, __dm_pr_read_keys, &pr);
3667 if (ret)
3668 return ret;
3669
3670 return pr.ret;
3671}
3672
3673static int __dm_pr_read_reservation(struct dm_target *ti, struct dm_dev *dev,
3674 sector_t start, sector_t len, void *data)
3675{
3676 struct dm_pr *pr = data;
3677 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3678
3679 if (!ops || !ops->pr_read_reservation) {
3680 pr->ret = -EOPNOTSUPP;
3681 return -1;
3682 }
3683
3684 pr->ret = ops->pr_read_reservation(dev->bdev, pr->rsv);
3685 if (!pr->ret)
3686 return -1;
3687
3688 return 0;
3689}
3690
3691static int dm_pr_read_reservation(struct block_device *bdev,
3692 struct pr_held_reservation *rsv)
3693{
3694 struct dm_pr pr = {
3695 .rsv = rsv,
3696 };
3697 int ret;
3698
3699 ret = dm_call_pr(bdev, __dm_pr_read_reservation, &pr);
3700 if (ret)
3701 return ret;
3702
3703 return pr.ret;
3704}
3705
3706static const struct pr_ops dm_pr_ops = {
3707 .pr_register = dm_pr_register,
3708 .pr_reserve = dm_pr_reserve,
3709 .pr_release = dm_pr_release,
3710 .pr_preempt = dm_pr_preempt,
3711 .pr_clear = dm_pr_clear,
3712 .pr_read_keys = dm_pr_read_keys,
3713 .pr_read_reservation = dm_pr_read_reservation,
3714};
3715
3716static const struct block_device_operations dm_blk_dops = {
3717 .submit_bio = dm_submit_bio,
3718 .poll_bio = dm_poll_bio,
3719 .open = dm_blk_open,
3720 .release = dm_blk_close,
3721 .ioctl = dm_blk_ioctl,
3722 .getgeo = dm_blk_getgeo,
3723 .report_zones = dm_blk_report_zones,
3724 .get_unique_id = dm_blk_get_unique_id,
3725 .pr_ops = &dm_pr_ops,
3726 .owner = THIS_MODULE
3727};
3728
3729static const struct block_device_operations dm_rq_blk_dops = {
3730 .open = dm_blk_open,
3731 .release = dm_blk_close,
3732 .ioctl = dm_blk_ioctl,
3733 .getgeo = dm_blk_getgeo,
3734 .get_unique_id = dm_blk_get_unique_id,
3735 .pr_ops = &dm_pr_ops,
3736 .owner = THIS_MODULE
3737};
3738
3739static const struct dax_operations dm_dax_ops = {
3740 .direct_access = dm_dax_direct_access,
3741 .zero_page_range = dm_dax_zero_page_range,
3742 .recovery_write = dm_dax_recovery_write,
3743};
3744
3745/*
3746 * module hooks
3747 */
3748module_init(dm_init);
3749module_exit(dm_exit);
3750
3751module_param(major, uint, 0);
3752MODULE_PARM_DESC(major, "The major number of the device mapper");
3753
3754module_param(reserved_bio_based_ios, uint, 0644);
3755MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3756
3757module_param(dm_numa_node, int, 0644);
3758MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3759
3760module_param(swap_bios, int, 0644);
3761MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3762
3763MODULE_DESCRIPTION(DM_NAME " driver");
3764MODULE_AUTHOR("Joe Thornber <dm-devel@lists.linux.dev>");
3765MODULE_LICENSE("GPL");
1/*
2 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm-core.h"
9#include "dm-rq.h"
10#include "dm-uevent.h"
11
12#include <linux/init.h>
13#include <linux/module.h>
14#include <linux/mutex.h>
15#include <linux/sched/signal.h>
16#include <linux/blkpg.h>
17#include <linux/bio.h>
18#include <linux/mempool.h>
19#include <linux/dax.h>
20#include <linux/slab.h>
21#include <linux/idr.h>
22#include <linux/uio.h>
23#include <linux/hdreg.h>
24#include <linux/delay.h>
25#include <linux/wait.h>
26#include <linux/pr.h>
27#include <linux/refcount.h>
28
29#define DM_MSG_PREFIX "core"
30
31/*
32 * Cookies are numeric values sent with CHANGE and REMOVE
33 * uevents while resuming, removing or renaming the device.
34 */
35#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
36#define DM_COOKIE_LENGTH 24
37
38static const char *_name = DM_NAME;
39
40static unsigned int major = 0;
41static unsigned int _major = 0;
42
43static DEFINE_IDR(_minor_idr);
44
45static DEFINE_SPINLOCK(_minor_lock);
46
47static void do_deferred_remove(struct work_struct *w);
48
49static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
50
51static struct workqueue_struct *deferred_remove_workqueue;
52
53atomic_t dm_global_event_nr = ATOMIC_INIT(0);
54DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
55
56void dm_issue_global_event(void)
57{
58 atomic_inc(&dm_global_event_nr);
59 wake_up(&dm_global_eventq);
60}
61
62/*
63 * One of these is allocated (on-stack) per original bio.
64 */
65struct clone_info {
66 struct dm_table *map;
67 struct bio *bio;
68 struct dm_io *io;
69 sector_t sector;
70 unsigned sector_count;
71};
72
73/*
74 * One of these is allocated per clone bio.
75 */
76#define DM_TIO_MAGIC 7282014
77struct dm_target_io {
78 unsigned magic;
79 struct dm_io *io;
80 struct dm_target *ti;
81 unsigned target_bio_nr;
82 unsigned *len_ptr;
83 bool inside_dm_io;
84 struct bio clone;
85};
86
87/*
88 * One of these is allocated per original bio.
89 * It contains the first clone used for that original.
90 */
91#define DM_IO_MAGIC 5191977
92struct dm_io {
93 unsigned magic;
94 struct mapped_device *md;
95 blk_status_t status;
96 atomic_t io_count;
97 struct bio *orig_bio;
98 unsigned long start_time;
99 spinlock_t endio_lock;
100 struct dm_stats_aux stats_aux;
101 /* last member of dm_target_io is 'struct bio' */
102 struct dm_target_io tio;
103};
104
105void *dm_per_bio_data(struct bio *bio, size_t data_size)
106{
107 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
108 if (!tio->inside_dm_io)
109 return (char *)bio - offsetof(struct dm_target_io, clone) - data_size;
110 return (char *)bio - offsetof(struct dm_target_io, clone) - offsetof(struct dm_io, tio) - data_size;
111}
112EXPORT_SYMBOL_GPL(dm_per_bio_data);
113
114struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
115{
116 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
117 if (io->magic == DM_IO_MAGIC)
118 return (struct bio *)((char *)io + offsetof(struct dm_io, tio) + offsetof(struct dm_target_io, clone));
119 BUG_ON(io->magic != DM_TIO_MAGIC);
120 return (struct bio *)((char *)io + offsetof(struct dm_target_io, clone));
121}
122EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
123
124unsigned dm_bio_get_target_bio_nr(const struct bio *bio)
125{
126 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
127}
128EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
129
130#define MINOR_ALLOCED ((void *)-1)
131
132/*
133 * Bits for the md->flags field.
134 */
135#define DMF_BLOCK_IO_FOR_SUSPEND 0
136#define DMF_SUSPENDED 1
137#define DMF_FROZEN 2
138#define DMF_FREEING 3
139#define DMF_DELETING 4
140#define DMF_NOFLUSH_SUSPENDING 5
141#define DMF_DEFERRED_REMOVE 6
142#define DMF_SUSPENDED_INTERNALLY 7
143
144#define DM_NUMA_NODE NUMA_NO_NODE
145static int dm_numa_node = DM_NUMA_NODE;
146
147/*
148 * For mempools pre-allocation at the table loading time.
149 */
150struct dm_md_mempools {
151 struct bio_set bs;
152 struct bio_set io_bs;
153};
154
155struct table_device {
156 struct list_head list;
157 refcount_t count;
158 struct dm_dev dm_dev;
159};
160
161/*
162 * Bio-based DM's mempools' reserved IOs set by the user.
163 */
164#define RESERVED_BIO_BASED_IOS 16
165static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
166
167static int __dm_get_module_param_int(int *module_param, int min, int max)
168{
169 int param = READ_ONCE(*module_param);
170 int modified_param = 0;
171 bool modified = true;
172
173 if (param < min)
174 modified_param = min;
175 else if (param > max)
176 modified_param = max;
177 else
178 modified = false;
179
180 if (modified) {
181 (void)cmpxchg(module_param, param, modified_param);
182 param = modified_param;
183 }
184
185 return param;
186}
187
188unsigned __dm_get_module_param(unsigned *module_param,
189 unsigned def, unsigned max)
190{
191 unsigned param = READ_ONCE(*module_param);
192 unsigned modified_param = 0;
193
194 if (!param)
195 modified_param = def;
196 else if (param > max)
197 modified_param = max;
198
199 if (modified_param) {
200 (void)cmpxchg(module_param, param, modified_param);
201 param = modified_param;
202 }
203
204 return param;
205}
206
207unsigned dm_get_reserved_bio_based_ios(void)
208{
209 return __dm_get_module_param(&reserved_bio_based_ios,
210 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
211}
212EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
213
214static unsigned dm_get_numa_node(void)
215{
216 return __dm_get_module_param_int(&dm_numa_node,
217 DM_NUMA_NODE, num_online_nodes() - 1);
218}
219
220static int __init local_init(void)
221{
222 int r;
223
224 r = dm_uevent_init();
225 if (r)
226 return r;
227
228 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
229 if (!deferred_remove_workqueue) {
230 r = -ENOMEM;
231 goto out_uevent_exit;
232 }
233
234 _major = major;
235 r = register_blkdev(_major, _name);
236 if (r < 0)
237 goto out_free_workqueue;
238
239 if (!_major)
240 _major = r;
241
242 return 0;
243
244out_free_workqueue:
245 destroy_workqueue(deferred_remove_workqueue);
246out_uevent_exit:
247 dm_uevent_exit();
248
249 return r;
250}
251
252static void local_exit(void)
253{
254 flush_scheduled_work();
255 destroy_workqueue(deferred_remove_workqueue);
256
257 unregister_blkdev(_major, _name);
258 dm_uevent_exit();
259
260 _major = 0;
261
262 DMINFO("cleaned up");
263}
264
265static int (*_inits[])(void) __initdata = {
266 local_init,
267 dm_target_init,
268 dm_linear_init,
269 dm_stripe_init,
270 dm_io_init,
271 dm_kcopyd_init,
272 dm_interface_init,
273 dm_statistics_init,
274};
275
276static void (*_exits[])(void) = {
277 local_exit,
278 dm_target_exit,
279 dm_linear_exit,
280 dm_stripe_exit,
281 dm_io_exit,
282 dm_kcopyd_exit,
283 dm_interface_exit,
284 dm_statistics_exit,
285};
286
287static int __init dm_init(void)
288{
289 const int count = ARRAY_SIZE(_inits);
290
291 int r, i;
292
293 for (i = 0; i < count; i++) {
294 r = _inits[i]();
295 if (r)
296 goto bad;
297 }
298
299 return 0;
300
301 bad:
302 while (i--)
303 _exits[i]();
304
305 return r;
306}
307
308static void __exit dm_exit(void)
309{
310 int i = ARRAY_SIZE(_exits);
311
312 while (i--)
313 _exits[i]();
314
315 /*
316 * Should be empty by this point.
317 */
318 idr_destroy(&_minor_idr);
319}
320
321/*
322 * Block device functions
323 */
324int dm_deleting_md(struct mapped_device *md)
325{
326 return test_bit(DMF_DELETING, &md->flags);
327}
328
329static int dm_blk_open(struct block_device *bdev, fmode_t mode)
330{
331 struct mapped_device *md;
332
333 spin_lock(&_minor_lock);
334
335 md = bdev->bd_disk->private_data;
336 if (!md)
337 goto out;
338
339 if (test_bit(DMF_FREEING, &md->flags) ||
340 dm_deleting_md(md)) {
341 md = NULL;
342 goto out;
343 }
344
345 dm_get(md);
346 atomic_inc(&md->open_count);
347out:
348 spin_unlock(&_minor_lock);
349
350 return md ? 0 : -ENXIO;
351}
352
353static void dm_blk_close(struct gendisk *disk, fmode_t mode)
354{
355 struct mapped_device *md;
356
357 spin_lock(&_minor_lock);
358
359 md = disk->private_data;
360 if (WARN_ON(!md))
361 goto out;
362
363 if (atomic_dec_and_test(&md->open_count) &&
364 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
365 queue_work(deferred_remove_workqueue, &deferred_remove_work);
366
367 dm_put(md);
368out:
369 spin_unlock(&_minor_lock);
370}
371
372int dm_open_count(struct mapped_device *md)
373{
374 return atomic_read(&md->open_count);
375}
376
377/*
378 * Guarantees nothing is using the device before it's deleted.
379 */
380int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
381{
382 int r = 0;
383
384 spin_lock(&_minor_lock);
385
386 if (dm_open_count(md)) {
387 r = -EBUSY;
388 if (mark_deferred)
389 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
390 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
391 r = -EEXIST;
392 else
393 set_bit(DMF_DELETING, &md->flags);
394
395 spin_unlock(&_minor_lock);
396
397 return r;
398}
399
400int dm_cancel_deferred_remove(struct mapped_device *md)
401{
402 int r = 0;
403
404 spin_lock(&_minor_lock);
405
406 if (test_bit(DMF_DELETING, &md->flags))
407 r = -EBUSY;
408 else
409 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
410
411 spin_unlock(&_minor_lock);
412
413 return r;
414}
415
416static void do_deferred_remove(struct work_struct *w)
417{
418 dm_deferred_remove();
419}
420
421sector_t dm_get_size(struct mapped_device *md)
422{
423 return get_capacity(md->disk);
424}
425
426struct request_queue *dm_get_md_queue(struct mapped_device *md)
427{
428 return md->queue;
429}
430
431struct dm_stats *dm_get_stats(struct mapped_device *md)
432{
433 return &md->stats;
434}
435
436static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
437{
438 struct mapped_device *md = bdev->bd_disk->private_data;
439
440 return dm_get_geometry(md, geo);
441}
442
443static int dm_blk_report_zones(struct gendisk *disk, sector_t sector,
444 struct blk_zone *zones, unsigned int *nr_zones)
445{
446#ifdef CONFIG_BLK_DEV_ZONED
447 struct mapped_device *md = disk->private_data;
448 struct dm_target *tgt;
449 struct dm_table *map;
450 int srcu_idx, ret;
451
452 if (dm_suspended_md(md))
453 return -EAGAIN;
454
455 map = dm_get_live_table(md, &srcu_idx);
456 if (!map)
457 return -EIO;
458
459 tgt = dm_table_find_target(map, sector);
460 if (!tgt) {
461 ret = -EIO;
462 goto out;
463 }
464
465 /*
466 * If we are executing this, we already know that the block device
467 * is a zoned device and so each target should have support for that
468 * type of drive. A missing report_zones method means that the target
469 * driver has a problem.
470 */
471 if (WARN_ON(!tgt->type->report_zones)) {
472 ret = -EIO;
473 goto out;
474 }
475
476 /*
477 * blkdev_report_zones() will loop and call this again to cover all the
478 * zones of the target, eventually moving on to the next target.
479 * So there is no need to loop here trying to fill the entire array
480 * of zones.
481 */
482 ret = tgt->type->report_zones(tgt, sector, zones, nr_zones);
483
484out:
485 dm_put_live_table(md, srcu_idx);
486 return ret;
487#else
488 return -ENOTSUPP;
489#endif
490}
491
492static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
493 struct block_device **bdev)
494 __acquires(md->io_barrier)
495{
496 struct dm_target *tgt;
497 struct dm_table *map;
498 int r;
499
500retry:
501 r = -ENOTTY;
502 map = dm_get_live_table(md, srcu_idx);
503 if (!map || !dm_table_get_size(map))
504 return r;
505
506 /* We only support devices that have a single target */
507 if (dm_table_get_num_targets(map) != 1)
508 return r;
509
510 tgt = dm_table_get_target(map, 0);
511 if (!tgt->type->prepare_ioctl)
512 return r;
513
514 if (dm_suspended_md(md))
515 return -EAGAIN;
516
517 r = tgt->type->prepare_ioctl(tgt, bdev);
518 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
519 dm_put_live_table(md, *srcu_idx);
520 msleep(10);
521 goto retry;
522 }
523
524 return r;
525}
526
527static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
528 __releases(md->io_barrier)
529{
530 dm_put_live_table(md, srcu_idx);
531}
532
533static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
534 unsigned int cmd, unsigned long arg)
535{
536 struct mapped_device *md = bdev->bd_disk->private_data;
537 int r, srcu_idx;
538
539 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
540 if (r < 0)
541 goto out;
542
543 if (r > 0) {
544 /*
545 * Target determined this ioctl is being issued against a
546 * subset of the parent bdev; require extra privileges.
547 */
548 if (!capable(CAP_SYS_RAWIO)) {
549 DMWARN_LIMIT(
550 "%s: sending ioctl %x to DM device without required privilege.",
551 current->comm, cmd);
552 r = -ENOIOCTLCMD;
553 goto out;
554 }
555 }
556
557 r = __blkdev_driver_ioctl(bdev, mode, cmd, arg);
558out:
559 dm_unprepare_ioctl(md, srcu_idx);
560 return r;
561}
562
563static void start_io_acct(struct dm_io *io);
564
565static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio)
566{
567 struct dm_io *io;
568 struct dm_target_io *tio;
569 struct bio *clone;
570
571 clone = bio_alloc_bioset(GFP_NOIO, 0, &md->io_bs);
572 if (!clone)
573 return NULL;
574
575 tio = container_of(clone, struct dm_target_io, clone);
576 tio->inside_dm_io = true;
577 tio->io = NULL;
578
579 io = container_of(tio, struct dm_io, tio);
580 io->magic = DM_IO_MAGIC;
581 io->status = 0;
582 atomic_set(&io->io_count, 1);
583 io->orig_bio = bio;
584 io->md = md;
585 spin_lock_init(&io->endio_lock);
586
587 start_io_acct(io);
588
589 return io;
590}
591
592static void free_io(struct mapped_device *md, struct dm_io *io)
593{
594 bio_put(&io->tio.clone);
595}
596
597static struct dm_target_io *alloc_tio(struct clone_info *ci, struct dm_target *ti,
598 unsigned target_bio_nr, gfp_t gfp_mask)
599{
600 struct dm_target_io *tio;
601
602 if (!ci->io->tio.io) {
603 /* the dm_target_io embedded in ci->io is available */
604 tio = &ci->io->tio;
605 } else {
606 struct bio *clone = bio_alloc_bioset(gfp_mask, 0, &ci->io->md->bs);
607 if (!clone)
608 return NULL;
609
610 tio = container_of(clone, struct dm_target_io, clone);
611 tio->inside_dm_io = false;
612 }
613
614 tio->magic = DM_TIO_MAGIC;
615 tio->io = ci->io;
616 tio->ti = ti;
617 tio->target_bio_nr = target_bio_nr;
618
619 return tio;
620}
621
622static void free_tio(struct dm_target_io *tio)
623{
624 if (tio->inside_dm_io)
625 return;
626 bio_put(&tio->clone);
627}
628
629static bool md_in_flight_bios(struct mapped_device *md)
630{
631 int cpu;
632 struct hd_struct *part = &dm_disk(md)->part0;
633 long sum = 0;
634
635 for_each_possible_cpu(cpu) {
636 sum += part_stat_local_read_cpu(part, in_flight[0], cpu);
637 sum += part_stat_local_read_cpu(part, in_flight[1], cpu);
638 }
639
640 return sum != 0;
641}
642
643static bool md_in_flight(struct mapped_device *md)
644{
645 if (queue_is_mq(md->queue))
646 return blk_mq_queue_inflight(md->queue);
647 else
648 return md_in_flight_bios(md);
649}
650
651static void start_io_acct(struct dm_io *io)
652{
653 struct mapped_device *md = io->md;
654 struct bio *bio = io->orig_bio;
655
656 io->start_time = jiffies;
657
658 generic_start_io_acct(md->queue, bio_op(bio), bio_sectors(bio),
659 &dm_disk(md)->part0);
660
661 if (unlikely(dm_stats_used(&md->stats)))
662 dm_stats_account_io(&md->stats, bio_data_dir(bio),
663 bio->bi_iter.bi_sector, bio_sectors(bio),
664 false, 0, &io->stats_aux);
665}
666
667static void end_io_acct(struct dm_io *io)
668{
669 struct mapped_device *md = io->md;
670 struct bio *bio = io->orig_bio;
671 unsigned long duration = jiffies - io->start_time;
672
673 generic_end_io_acct(md->queue, bio_op(bio), &dm_disk(md)->part0,
674 io->start_time);
675
676 if (unlikely(dm_stats_used(&md->stats)))
677 dm_stats_account_io(&md->stats, bio_data_dir(bio),
678 bio->bi_iter.bi_sector, bio_sectors(bio),
679 true, duration, &io->stats_aux);
680
681 /* nudge anyone waiting on suspend queue */
682 if (unlikely(wq_has_sleeper(&md->wait)))
683 wake_up(&md->wait);
684}
685
686/*
687 * Add the bio to the list of deferred io.
688 */
689static void queue_io(struct mapped_device *md, struct bio *bio)
690{
691 unsigned long flags;
692
693 spin_lock_irqsave(&md->deferred_lock, flags);
694 bio_list_add(&md->deferred, bio);
695 spin_unlock_irqrestore(&md->deferred_lock, flags);
696 queue_work(md->wq, &md->work);
697}
698
699/*
700 * Everyone (including functions in this file), should use this
701 * function to access the md->map field, and make sure they call
702 * dm_put_live_table() when finished.
703 */
704struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
705{
706 *srcu_idx = srcu_read_lock(&md->io_barrier);
707
708 return srcu_dereference(md->map, &md->io_barrier);
709}
710
711void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
712{
713 srcu_read_unlock(&md->io_barrier, srcu_idx);
714}
715
716void dm_sync_table(struct mapped_device *md)
717{
718 synchronize_srcu(&md->io_barrier);
719 synchronize_rcu_expedited();
720}
721
722/*
723 * A fast alternative to dm_get_live_table/dm_put_live_table.
724 * The caller must not block between these two functions.
725 */
726static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
727{
728 rcu_read_lock();
729 return rcu_dereference(md->map);
730}
731
732static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
733{
734 rcu_read_unlock();
735}
736
737static char *_dm_claim_ptr = "I belong to device-mapper";
738
739/*
740 * Open a table device so we can use it as a map destination.
741 */
742static int open_table_device(struct table_device *td, dev_t dev,
743 struct mapped_device *md)
744{
745 struct block_device *bdev;
746
747 int r;
748
749 BUG_ON(td->dm_dev.bdev);
750
751 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _dm_claim_ptr);
752 if (IS_ERR(bdev))
753 return PTR_ERR(bdev);
754
755 r = bd_link_disk_holder(bdev, dm_disk(md));
756 if (r) {
757 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
758 return r;
759 }
760
761 td->dm_dev.bdev = bdev;
762 td->dm_dev.dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
763 return 0;
764}
765
766/*
767 * Close a table device that we've been using.
768 */
769static void close_table_device(struct table_device *td, struct mapped_device *md)
770{
771 if (!td->dm_dev.bdev)
772 return;
773
774 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
775 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
776 put_dax(td->dm_dev.dax_dev);
777 td->dm_dev.bdev = NULL;
778 td->dm_dev.dax_dev = NULL;
779}
780
781static struct table_device *find_table_device(struct list_head *l, dev_t dev,
782 fmode_t mode)
783{
784 struct table_device *td;
785
786 list_for_each_entry(td, l, list)
787 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
788 return td;
789
790 return NULL;
791}
792
793int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
794 struct dm_dev **result)
795{
796 int r;
797 struct table_device *td;
798
799 mutex_lock(&md->table_devices_lock);
800 td = find_table_device(&md->table_devices, dev, mode);
801 if (!td) {
802 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
803 if (!td) {
804 mutex_unlock(&md->table_devices_lock);
805 return -ENOMEM;
806 }
807
808 td->dm_dev.mode = mode;
809 td->dm_dev.bdev = NULL;
810
811 if ((r = open_table_device(td, dev, md))) {
812 mutex_unlock(&md->table_devices_lock);
813 kfree(td);
814 return r;
815 }
816
817 format_dev_t(td->dm_dev.name, dev);
818
819 refcount_set(&td->count, 1);
820 list_add(&td->list, &md->table_devices);
821 } else {
822 refcount_inc(&td->count);
823 }
824 mutex_unlock(&md->table_devices_lock);
825
826 *result = &td->dm_dev;
827 return 0;
828}
829EXPORT_SYMBOL_GPL(dm_get_table_device);
830
831void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
832{
833 struct table_device *td = container_of(d, struct table_device, dm_dev);
834
835 mutex_lock(&md->table_devices_lock);
836 if (refcount_dec_and_test(&td->count)) {
837 close_table_device(td, md);
838 list_del(&td->list);
839 kfree(td);
840 }
841 mutex_unlock(&md->table_devices_lock);
842}
843EXPORT_SYMBOL(dm_put_table_device);
844
845static void free_table_devices(struct list_head *devices)
846{
847 struct list_head *tmp, *next;
848
849 list_for_each_safe(tmp, next, devices) {
850 struct table_device *td = list_entry(tmp, struct table_device, list);
851
852 DMWARN("dm_destroy: %s still exists with %d references",
853 td->dm_dev.name, refcount_read(&td->count));
854 kfree(td);
855 }
856}
857
858/*
859 * Get the geometry associated with a dm device
860 */
861int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
862{
863 *geo = md->geometry;
864
865 return 0;
866}
867
868/*
869 * Set the geometry of a device.
870 */
871int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
872{
873 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
874
875 if (geo->start > sz) {
876 DMWARN("Start sector is beyond the geometry limits.");
877 return -EINVAL;
878 }
879
880 md->geometry = *geo;
881
882 return 0;
883}
884
885static int __noflush_suspending(struct mapped_device *md)
886{
887 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
888}
889
890/*
891 * Decrements the number of outstanding ios that a bio has been
892 * cloned into, completing the original io if necc.
893 */
894static void dec_pending(struct dm_io *io, blk_status_t error)
895{
896 unsigned long flags;
897 blk_status_t io_error;
898 struct bio *bio;
899 struct mapped_device *md = io->md;
900
901 /* Push-back supersedes any I/O errors */
902 if (unlikely(error)) {
903 spin_lock_irqsave(&io->endio_lock, flags);
904 if (!(io->status == BLK_STS_DM_REQUEUE && __noflush_suspending(md)))
905 io->status = error;
906 spin_unlock_irqrestore(&io->endio_lock, flags);
907 }
908
909 if (atomic_dec_and_test(&io->io_count)) {
910 if (io->status == BLK_STS_DM_REQUEUE) {
911 /*
912 * Target requested pushing back the I/O.
913 */
914 spin_lock_irqsave(&md->deferred_lock, flags);
915 if (__noflush_suspending(md))
916 /* NOTE early return due to BLK_STS_DM_REQUEUE below */
917 bio_list_add_head(&md->deferred, io->orig_bio);
918 else
919 /* noflush suspend was interrupted. */
920 io->status = BLK_STS_IOERR;
921 spin_unlock_irqrestore(&md->deferred_lock, flags);
922 }
923
924 io_error = io->status;
925 bio = io->orig_bio;
926 end_io_acct(io);
927 free_io(md, io);
928
929 if (io_error == BLK_STS_DM_REQUEUE)
930 return;
931
932 if ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size) {
933 /*
934 * Preflush done for flush with data, reissue
935 * without REQ_PREFLUSH.
936 */
937 bio->bi_opf &= ~REQ_PREFLUSH;
938 queue_io(md, bio);
939 } else {
940 /* done with normal IO or empty flush */
941 if (io_error)
942 bio->bi_status = io_error;
943 bio_endio(bio);
944 }
945 }
946}
947
948void disable_discard(struct mapped_device *md)
949{
950 struct queue_limits *limits = dm_get_queue_limits(md);
951
952 /* device doesn't really support DISCARD, disable it */
953 limits->max_discard_sectors = 0;
954 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, md->queue);
955}
956
957void disable_write_same(struct mapped_device *md)
958{
959 struct queue_limits *limits = dm_get_queue_limits(md);
960
961 /* device doesn't really support WRITE SAME, disable it */
962 limits->max_write_same_sectors = 0;
963}
964
965void disable_write_zeroes(struct mapped_device *md)
966{
967 struct queue_limits *limits = dm_get_queue_limits(md);
968
969 /* device doesn't really support WRITE ZEROES, disable it */
970 limits->max_write_zeroes_sectors = 0;
971}
972
973static void clone_endio(struct bio *bio)
974{
975 blk_status_t error = bio->bi_status;
976 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
977 struct dm_io *io = tio->io;
978 struct mapped_device *md = tio->io->md;
979 dm_endio_fn endio = tio->ti->type->end_io;
980
981 if (unlikely(error == BLK_STS_TARGET) && md->type != DM_TYPE_NVME_BIO_BASED) {
982 if (bio_op(bio) == REQ_OP_DISCARD &&
983 !bio->bi_disk->queue->limits.max_discard_sectors)
984 disable_discard(md);
985 else if (bio_op(bio) == REQ_OP_WRITE_SAME &&
986 !bio->bi_disk->queue->limits.max_write_same_sectors)
987 disable_write_same(md);
988 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
989 !bio->bi_disk->queue->limits.max_write_zeroes_sectors)
990 disable_write_zeroes(md);
991 }
992
993 if (endio) {
994 int r = endio(tio->ti, bio, &error);
995 switch (r) {
996 case DM_ENDIO_REQUEUE:
997 error = BLK_STS_DM_REQUEUE;
998 /*FALLTHRU*/
999 case DM_ENDIO_DONE:
1000 break;
1001 case DM_ENDIO_INCOMPLETE:
1002 /* The target will handle the io */
1003 return;
1004 default:
1005 DMWARN("unimplemented target endio return value: %d", r);
1006 BUG();
1007 }
1008 }
1009
1010 free_tio(tio);
1011 dec_pending(io, error);
1012}
1013
1014/*
1015 * Return maximum size of I/O possible at the supplied sector up to the current
1016 * target boundary.
1017 */
1018static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1019{
1020 sector_t target_offset = dm_target_offset(ti, sector);
1021
1022 return ti->len - target_offset;
1023}
1024
1025static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1026{
1027 sector_t len = max_io_len_target_boundary(sector, ti);
1028 sector_t offset, max_len;
1029
1030 /*
1031 * Does the target need to split even further?
1032 */
1033 if (ti->max_io_len) {
1034 offset = dm_target_offset(ti, sector);
1035 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1036 max_len = sector_div(offset, ti->max_io_len);
1037 else
1038 max_len = offset & (ti->max_io_len - 1);
1039 max_len = ti->max_io_len - max_len;
1040
1041 if (len > max_len)
1042 len = max_len;
1043 }
1044
1045 return len;
1046}
1047
1048int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1049{
1050 if (len > UINT_MAX) {
1051 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1052 (unsigned long long)len, UINT_MAX);
1053 ti->error = "Maximum size of target IO is too large";
1054 return -EINVAL;
1055 }
1056
1057 ti->max_io_len = (uint32_t) len;
1058
1059 return 0;
1060}
1061EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1062
1063static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1064 sector_t sector, int *srcu_idx)
1065 __acquires(md->io_barrier)
1066{
1067 struct dm_table *map;
1068 struct dm_target *ti;
1069
1070 map = dm_get_live_table(md, srcu_idx);
1071 if (!map)
1072 return NULL;
1073
1074 ti = dm_table_find_target(map, sector);
1075 if (!ti)
1076 return NULL;
1077
1078 return ti;
1079}
1080
1081static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1082 long nr_pages, void **kaddr, pfn_t *pfn)
1083{
1084 struct mapped_device *md = dax_get_private(dax_dev);
1085 sector_t sector = pgoff * PAGE_SECTORS;
1086 struct dm_target *ti;
1087 long len, ret = -EIO;
1088 int srcu_idx;
1089
1090 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1091
1092 if (!ti)
1093 goto out;
1094 if (!ti->type->direct_access)
1095 goto out;
1096 len = max_io_len(sector, ti) / PAGE_SECTORS;
1097 if (len < 1)
1098 goto out;
1099 nr_pages = min(len, nr_pages);
1100 ret = ti->type->direct_access(ti, pgoff, nr_pages, kaddr, pfn);
1101
1102 out:
1103 dm_put_live_table(md, srcu_idx);
1104
1105 return ret;
1106}
1107
1108static bool dm_dax_supported(struct dax_device *dax_dev, struct block_device *bdev,
1109 int blocksize, sector_t start, sector_t len)
1110{
1111 struct mapped_device *md = dax_get_private(dax_dev);
1112 struct dm_table *map;
1113 int srcu_idx;
1114 bool ret;
1115
1116 map = dm_get_live_table(md, &srcu_idx);
1117 if (!map)
1118 return false;
1119
1120 ret = dm_table_supports_dax(map, device_supports_dax, &blocksize);
1121
1122 dm_put_live_table(md, srcu_idx);
1123
1124 return ret;
1125}
1126
1127static size_t dm_dax_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
1128 void *addr, size_t bytes, struct iov_iter *i)
1129{
1130 struct mapped_device *md = dax_get_private(dax_dev);
1131 sector_t sector = pgoff * PAGE_SECTORS;
1132 struct dm_target *ti;
1133 long ret = 0;
1134 int srcu_idx;
1135
1136 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1137
1138 if (!ti)
1139 goto out;
1140 if (!ti->type->dax_copy_from_iter) {
1141 ret = copy_from_iter(addr, bytes, i);
1142 goto out;
1143 }
1144 ret = ti->type->dax_copy_from_iter(ti, pgoff, addr, bytes, i);
1145 out:
1146 dm_put_live_table(md, srcu_idx);
1147
1148 return ret;
1149}
1150
1151static size_t dm_dax_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
1152 void *addr, size_t bytes, struct iov_iter *i)
1153{
1154 struct mapped_device *md = dax_get_private(dax_dev);
1155 sector_t sector = pgoff * PAGE_SECTORS;
1156 struct dm_target *ti;
1157 long ret = 0;
1158 int srcu_idx;
1159
1160 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1161
1162 if (!ti)
1163 goto out;
1164 if (!ti->type->dax_copy_to_iter) {
1165 ret = copy_to_iter(addr, bytes, i);
1166 goto out;
1167 }
1168 ret = ti->type->dax_copy_to_iter(ti, pgoff, addr, bytes, i);
1169 out:
1170 dm_put_live_table(md, srcu_idx);
1171
1172 return ret;
1173}
1174
1175/*
1176 * A target may call dm_accept_partial_bio only from the map routine. It is
1177 * allowed for all bio types except REQ_PREFLUSH and REQ_OP_ZONE_RESET.
1178 *
1179 * dm_accept_partial_bio informs the dm that the target only wants to process
1180 * additional n_sectors sectors of the bio and the rest of the data should be
1181 * sent in a next bio.
1182 *
1183 * A diagram that explains the arithmetics:
1184 * +--------------------+---------------+-------+
1185 * | 1 | 2 | 3 |
1186 * +--------------------+---------------+-------+
1187 *
1188 * <-------------- *tio->len_ptr --------------->
1189 * <------- bi_size ------->
1190 * <-- n_sectors -->
1191 *
1192 * Region 1 was already iterated over with bio_advance or similar function.
1193 * (it may be empty if the target doesn't use bio_advance)
1194 * Region 2 is the remaining bio size that the target wants to process.
1195 * (it may be empty if region 1 is non-empty, although there is no reason
1196 * to make it empty)
1197 * The target requires that region 3 is to be sent in the next bio.
1198 *
1199 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1200 * the partially processed part (the sum of regions 1+2) must be the same for all
1201 * copies of the bio.
1202 */
1203void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1204{
1205 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1206 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1207 BUG_ON(bio->bi_opf & REQ_PREFLUSH);
1208 BUG_ON(bi_size > *tio->len_ptr);
1209 BUG_ON(n_sectors > bi_size);
1210 *tio->len_ptr -= bi_size - n_sectors;
1211 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1212}
1213EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1214
1215/*
1216 * The zone descriptors obtained with a zone report indicate
1217 * zone positions within the underlying device of the target. The zone
1218 * descriptors must be remapped to match their position within the dm device.
1219 * The caller target should obtain the zones information using
1220 * blkdev_report_zones() to ensure that remapping for partition offset is
1221 * already handled.
1222 */
1223void dm_remap_zone_report(struct dm_target *ti, sector_t start,
1224 struct blk_zone *zones, unsigned int *nr_zones)
1225{
1226#ifdef CONFIG_BLK_DEV_ZONED
1227 struct blk_zone *zone;
1228 unsigned int nrz = *nr_zones;
1229 int i;
1230
1231 /*
1232 * Remap the start sector and write pointer position of the zones in
1233 * the array. Since we may have obtained from the target underlying
1234 * device more zones that the target size, also adjust the number
1235 * of zones.
1236 */
1237 for (i = 0; i < nrz; i++) {
1238 zone = zones + i;
1239 if (zone->start >= start + ti->len) {
1240 memset(zone, 0, sizeof(struct blk_zone) * (nrz - i));
1241 break;
1242 }
1243
1244 zone->start = zone->start + ti->begin - start;
1245 if (zone->type == BLK_ZONE_TYPE_CONVENTIONAL)
1246 continue;
1247
1248 if (zone->cond == BLK_ZONE_COND_FULL)
1249 zone->wp = zone->start + zone->len;
1250 else if (zone->cond == BLK_ZONE_COND_EMPTY)
1251 zone->wp = zone->start;
1252 else
1253 zone->wp = zone->wp + ti->begin - start;
1254 }
1255
1256 *nr_zones = i;
1257#else /* !CONFIG_BLK_DEV_ZONED */
1258 *nr_zones = 0;
1259#endif
1260}
1261EXPORT_SYMBOL_GPL(dm_remap_zone_report);
1262
1263static blk_qc_t __map_bio(struct dm_target_io *tio)
1264{
1265 int r;
1266 sector_t sector;
1267 struct bio *clone = &tio->clone;
1268 struct dm_io *io = tio->io;
1269 struct mapped_device *md = io->md;
1270 struct dm_target *ti = tio->ti;
1271 blk_qc_t ret = BLK_QC_T_NONE;
1272
1273 clone->bi_end_io = clone_endio;
1274
1275 /*
1276 * Map the clone. If r == 0 we don't need to do
1277 * anything, the target has assumed ownership of
1278 * this io.
1279 */
1280 atomic_inc(&io->io_count);
1281 sector = clone->bi_iter.bi_sector;
1282
1283 r = ti->type->map(ti, clone);
1284 switch (r) {
1285 case DM_MAPIO_SUBMITTED:
1286 break;
1287 case DM_MAPIO_REMAPPED:
1288 /* the bio has been remapped so dispatch it */
1289 trace_block_bio_remap(clone->bi_disk->queue, clone,
1290 bio_dev(io->orig_bio), sector);
1291 if (md->type == DM_TYPE_NVME_BIO_BASED)
1292 ret = direct_make_request(clone);
1293 else
1294 ret = generic_make_request(clone);
1295 break;
1296 case DM_MAPIO_KILL:
1297 free_tio(tio);
1298 dec_pending(io, BLK_STS_IOERR);
1299 break;
1300 case DM_MAPIO_REQUEUE:
1301 free_tio(tio);
1302 dec_pending(io, BLK_STS_DM_REQUEUE);
1303 break;
1304 default:
1305 DMWARN("unimplemented target map return value: %d", r);
1306 BUG();
1307 }
1308
1309 return ret;
1310}
1311
1312static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1313{
1314 bio->bi_iter.bi_sector = sector;
1315 bio->bi_iter.bi_size = to_bytes(len);
1316}
1317
1318/*
1319 * Creates a bio that consists of range of complete bvecs.
1320 */
1321static int clone_bio(struct dm_target_io *tio, struct bio *bio,
1322 sector_t sector, unsigned len)
1323{
1324 struct bio *clone = &tio->clone;
1325
1326 __bio_clone_fast(clone, bio);
1327
1328 if (bio_integrity(bio)) {
1329 int r;
1330
1331 if (unlikely(!dm_target_has_integrity(tio->ti->type) &&
1332 !dm_target_passes_integrity(tio->ti->type))) {
1333 DMWARN("%s: the target %s doesn't support integrity data.",
1334 dm_device_name(tio->io->md),
1335 tio->ti->type->name);
1336 return -EIO;
1337 }
1338
1339 r = bio_integrity_clone(clone, bio, GFP_NOIO);
1340 if (r < 0)
1341 return r;
1342 }
1343
1344 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1345 clone->bi_iter.bi_size = to_bytes(len);
1346
1347 if (bio_integrity(bio))
1348 bio_integrity_trim(clone);
1349
1350 return 0;
1351}
1352
1353static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1354 struct dm_target *ti, unsigned num_bios)
1355{
1356 struct dm_target_io *tio;
1357 int try;
1358
1359 if (!num_bios)
1360 return;
1361
1362 if (num_bios == 1) {
1363 tio = alloc_tio(ci, ti, 0, GFP_NOIO);
1364 bio_list_add(blist, &tio->clone);
1365 return;
1366 }
1367
1368 for (try = 0; try < 2; try++) {
1369 int bio_nr;
1370 struct bio *bio;
1371
1372 if (try)
1373 mutex_lock(&ci->io->md->table_devices_lock);
1374 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1375 tio = alloc_tio(ci, ti, bio_nr, try ? GFP_NOIO : GFP_NOWAIT);
1376 if (!tio)
1377 break;
1378
1379 bio_list_add(blist, &tio->clone);
1380 }
1381 if (try)
1382 mutex_unlock(&ci->io->md->table_devices_lock);
1383 if (bio_nr == num_bios)
1384 return;
1385
1386 while ((bio = bio_list_pop(blist))) {
1387 tio = container_of(bio, struct dm_target_io, clone);
1388 free_tio(tio);
1389 }
1390 }
1391}
1392
1393static blk_qc_t __clone_and_map_simple_bio(struct clone_info *ci,
1394 struct dm_target_io *tio, unsigned *len)
1395{
1396 struct bio *clone = &tio->clone;
1397
1398 tio->len_ptr = len;
1399
1400 __bio_clone_fast(clone, ci->bio);
1401 if (len)
1402 bio_setup_sector(clone, ci->sector, *len);
1403
1404 return __map_bio(tio);
1405}
1406
1407static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1408 unsigned num_bios, unsigned *len)
1409{
1410 struct bio_list blist = BIO_EMPTY_LIST;
1411 struct bio *bio;
1412 struct dm_target_io *tio;
1413
1414 alloc_multiple_bios(&blist, ci, ti, num_bios);
1415
1416 while ((bio = bio_list_pop(&blist))) {
1417 tio = container_of(bio, struct dm_target_io, clone);
1418 (void) __clone_and_map_simple_bio(ci, tio, len);
1419 }
1420}
1421
1422static int __send_empty_flush(struct clone_info *ci)
1423{
1424 unsigned target_nr = 0;
1425 struct dm_target *ti;
1426
1427 /*
1428 * Empty flush uses a statically initialized bio, as the base for
1429 * cloning. However, blkg association requires that a bdev is
1430 * associated with a gendisk, which doesn't happen until the bdev is
1431 * opened. So, blkg association is done at issue time of the flush
1432 * rather than when the device is created in alloc_dev().
1433 */
1434 bio_set_dev(ci->bio, ci->io->md->bdev);
1435
1436 BUG_ON(bio_has_data(ci->bio));
1437 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1438 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1439
1440 bio_disassociate_blkg(ci->bio);
1441
1442 return 0;
1443}
1444
1445static int __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1446 sector_t sector, unsigned *len)
1447{
1448 struct bio *bio = ci->bio;
1449 struct dm_target_io *tio;
1450 int r;
1451
1452 tio = alloc_tio(ci, ti, 0, GFP_NOIO);
1453 tio->len_ptr = len;
1454 r = clone_bio(tio, bio, sector, *len);
1455 if (r < 0) {
1456 free_tio(tio);
1457 return r;
1458 }
1459 (void) __map_bio(tio);
1460
1461 return 0;
1462}
1463
1464typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1465
1466static unsigned get_num_discard_bios(struct dm_target *ti)
1467{
1468 return ti->num_discard_bios;
1469}
1470
1471static unsigned get_num_secure_erase_bios(struct dm_target *ti)
1472{
1473 return ti->num_secure_erase_bios;
1474}
1475
1476static unsigned get_num_write_same_bios(struct dm_target *ti)
1477{
1478 return ti->num_write_same_bios;
1479}
1480
1481static unsigned get_num_write_zeroes_bios(struct dm_target *ti)
1482{
1483 return ti->num_write_zeroes_bios;
1484}
1485
1486static int __send_changing_extent_only(struct clone_info *ci, struct dm_target *ti,
1487 unsigned num_bios)
1488{
1489 unsigned len;
1490
1491 /*
1492 * Even though the device advertised support for this type of
1493 * request, that does not mean every target supports it, and
1494 * reconfiguration might also have changed that since the
1495 * check was performed.
1496 */
1497 if (!num_bios)
1498 return -EOPNOTSUPP;
1499
1500 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1501
1502 __send_duplicate_bios(ci, ti, num_bios, &len);
1503
1504 ci->sector += len;
1505 ci->sector_count -= len;
1506
1507 return 0;
1508}
1509
1510static int __send_discard(struct clone_info *ci, struct dm_target *ti)
1511{
1512 return __send_changing_extent_only(ci, ti, get_num_discard_bios(ti));
1513}
1514
1515static int __send_secure_erase(struct clone_info *ci, struct dm_target *ti)
1516{
1517 return __send_changing_extent_only(ci, ti, get_num_secure_erase_bios(ti));
1518}
1519
1520static int __send_write_same(struct clone_info *ci, struct dm_target *ti)
1521{
1522 return __send_changing_extent_only(ci, ti, get_num_write_same_bios(ti));
1523}
1524
1525static int __send_write_zeroes(struct clone_info *ci, struct dm_target *ti)
1526{
1527 return __send_changing_extent_only(ci, ti, get_num_write_zeroes_bios(ti));
1528}
1529
1530static bool is_abnormal_io(struct bio *bio)
1531{
1532 bool r = false;
1533
1534 switch (bio_op(bio)) {
1535 case REQ_OP_DISCARD:
1536 case REQ_OP_SECURE_ERASE:
1537 case REQ_OP_WRITE_SAME:
1538 case REQ_OP_WRITE_ZEROES:
1539 r = true;
1540 break;
1541 }
1542
1543 return r;
1544}
1545
1546static bool __process_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1547 int *result)
1548{
1549 struct bio *bio = ci->bio;
1550
1551 if (bio_op(bio) == REQ_OP_DISCARD)
1552 *result = __send_discard(ci, ti);
1553 else if (bio_op(bio) == REQ_OP_SECURE_ERASE)
1554 *result = __send_secure_erase(ci, ti);
1555 else if (bio_op(bio) == REQ_OP_WRITE_SAME)
1556 *result = __send_write_same(ci, ti);
1557 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
1558 *result = __send_write_zeroes(ci, ti);
1559 else
1560 return false;
1561
1562 return true;
1563}
1564
1565/*
1566 * Select the correct strategy for processing a non-flush bio.
1567 */
1568static int __split_and_process_non_flush(struct clone_info *ci)
1569{
1570 struct dm_target *ti;
1571 unsigned len;
1572 int r;
1573
1574 ti = dm_table_find_target(ci->map, ci->sector);
1575 if (!ti)
1576 return -EIO;
1577
1578 if (__process_abnormal_io(ci, ti, &r))
1579 return r;
1580
1581 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1582
1583 r = __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1584 if (r < 0)
1585 return r;
1586
1587 ci->sector += len;
1588 ci->sector_count -= len;
1589
1590 return 0;
1591}
1592
1593static void init_clone_info(struct clone_info *ci, struct mapped_device *md,
1594 struct dm_table *map, struct bio *bio)
1595{
1596 ci->map = map;
1597 ci->io = alloc_io(md, bio);
1598 ci->sector = bio->bi_iter.bi_sector;
1599}
1600
1601#define __dm_part_stat_sub(part, field, subnd) \
1602 (part_stat_get(part, field) -= (subnd))
1603
1604/*
1605 * Entry point to split a bio into clones and submit them to the targets.
1606 */
1607static blk_qc_t __split_and_process_bio(struct mapped_device *md,
1608 struct dm_table *map, struct bio *bio)
1609{
1610 struct clone_info ci;
1611 blk_qc_t ret = BLK_QC_T_NONE;
1612 int error = 0;
1613
1614 init_clone_info(&ci, md, map, bio);
1615
1616 if (bio->bi_opf & REQ_PREFLUSH) {
1617 struct bio flush_bio;
1618
1619 /*
1620 * Use an on-stack bio for this, it's safe since we don't
1621 * need to reference it after submit. It's just used as
1622 * the basis for the clone(s).
1623 */
1624 bio_init(&flush_bio, NULL, 0);
1625 flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC;
1626 ci.bio = &flush_bio;
1627 ci.sector_count = 0;
1628 error = __send_empty_flush(&ci);
1629 /* dec_pending submits any data associated with flush */
1630 } else if (bio_op(bio) == REQ_OP_ZONE_RESET) {
1631 ci.bio = bio;
1632 ci.sector_count = 0;
1633 error = __split_and_process_non_flush(&ci);
1634 } else {
1635 ci.bio = bio;
1636 ci.sector_count = bio_sectors(bio);
1637 while (ci.sector_count && !error) {
1638 error = __split_and_process_non_flush(&ci);
1639 if (current->bio_list && ci.sector_count && !error) {
1640 /*
1641 * Remainder must be passed to generic_make_request()
1642 * so that it gets handled *after* bios already submitted
1643 * have been completely processed.
1644 * We take a clone of the original to store in
1645 * ci.io->orig_bio to be used by end_io_acct() and
1646 * for dec_pending to use for completion handling.
1647 */
1648 struct bio *b = bio_split(bio, bio_sectors(bio) - ci.sector_count,
1649 GFP_NOIO, &md->queue->bio_split);
1650 ci.io->orig_bio = b;
1651
1652 /*
1653 * Adjust IO stats for each split, otherwise upon queue
1654 * reentry there will be redundant IO accounting.
1655 * NOTE: this is a stop-gap fix, a proper fix involves
1656 * significant refactoring of DM core's bio splitting
1657 * (by eliminating DM's splitting and just using bio_split)
1658 */
1659 part_stat_lock();
1660 __dm_part_stat_sub(&dm_disk(md)->part0,
1661 sectors[op_stat_group(bio_op(bio))], ci.sector_count);
1662 part_stat_unlock();
1663
1664 bio_chain(b, bio);
1665 trace_block_split(md->queue, b, bio->bi_iter.bi_sector);
1666 ret = generic_make_request(bio);
1667 break;
1668 }
1669 }
1670 }
1671
1672 /* drop the extra reference count */
1673 dec_pending(ci.io, errno_to_blk_status(error));
1674 return ret;
1675}
1676
1677/*
1678 * Optimized variant of __split_and_process_bio that leverages the
1679 * fact that targets that use it do _not_ have a need to split bios.
1680 */
1681static blk_qc_t __process_bio(struct mapped_device *md, struct dm_table *map,
1682 struct bio *bio, struct dm_target *ti)
1683{
1684 struct clone_info ci;
1685 blk_qc_t ret = BLK_QC_T_NONE;
1686 int error = 0;
1687
1688 init_clone_info(&ci, md, map, bio);
1689
1690 if (bio->bi_opf & REQ_PREFLUSH) {
1691 struct bio flush_bio;
1692
1693 /*
1694 * Use an on-stack bio for this, it's safe since we don't
1695 * need to reference it after submit. It's just used as
1696 * the basis for the clone(s).
1697 */
1698 bio_init(&flush_bio, NULL, 0);
1699 flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC;
1700 ci.bio = &flush_bio;
1701 ci.sector_count = 0;
1702 error = __send_empty_flush(&ci);
1703 /* dec_pending submits any data associated with flush */
1704 } else {
1705 struct dm_target_io *tio;
1706
1707 ci.bio = bio;
1708 ci.sector_count = bio_sectors(bio);
1709 if (__process_abnormal_io(&ci, ti, &error))
1710 goto out;
1711
1712 tio = alloc_tio(&ci, ti, 0, GFP_NOIO);
1713 ret = __clone_and_map_simple_bio(&ci, tio, NULL);
1714 }
1715out:
1716 /* drop the extra reference count */
1717 dec_pending(ci.io, errno_to_blk_status(error));
1718 return ret;
1719}
1720
1721static void dm_queue_split(struct mapped_device *md, struct dm_target *ti, struct bio **bio)
1722{
1723 unsigned len, sector_count;
1724
1725 sector_count = bio_sectors(*bio);
1726 len = min_t(sector_t, max_io_len((*bio)->bi_iter.bi_sector, ti), sector_count);
1727
1728 if (sector_count > len) {
1729 struct bio *split = bio_split(*bio, len, GFP_NOIO, &md->queue->bio_split);
1730
1731 bio_chain(split, *bio);
1732 trace_block_split(md->queue, split, (*bio)->bi_iter.bi_sector);
1733 generic_make_request(*bio);
1734 *bio = split;
1735 }
1736}
1737
1738static blk_qc_t dm_process_bio(struct mapped_device *md,
1739 struct dm_table *map, struct bio *bio)
1740{
1741 blk_qc_t ret = BLK_QC_T_NONE;
1742 struct dm_target *ti = md->immutable_target;
1743
1744 if (unlikely(!map)) {
1745 bio_io_error(bio);
1746 return ret;
1747 }
1748
1749 if (!ti) {
1750 ti = dm_table_find_target(map, bio->bi_iter.bi_sector);
1751 if (unlikely(!ti)) {
1752 bio_io_error(bio);
1753 return ret;
1754 }
1755 }
1756
1757 /*
1758 * If in ->make_request_fn we need to use blk_queue_split(), otherwise
1759 * queue_limits for abnormal requests (e.g. discard, writesame, etc)
1760 * won't be imposed.
1761 */
1762 if (current->bio_list) {
1763 blk_queue_split(md->queue, &bio);
1764 if (!is_abnormal_io(bio))
1765 dm_queue_split(md, ti, &bio);
1766 }
1767
1768 if (dm_get_md_type(md) == DM_TYPE_NVME_BIO_BASED)
1769 return __process_bio(md, map, bio, ti);
1770 else
1771 return __split_and_process_bio(md, map, bio);
1772}
1773
1774static blk_qc_t dm_make_request(struct request_queue *q, struct bio *bio)
1775{
1776 struct mapped_device *md = q->queuedata;
1777 blk_qc_t ret = BLK_QC_T_NONE;
1778 int srcu_idx;
1779 struct dm_table *map;
1780
1781 map = dm_get_live_table(md, &srcu_idx);
1782
1783 /* if we're suspended, we have to queue this io for later */
1784 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1785 dm_put_live_table(md, srcu_idx);
1786
1787 if (!(bio->bi_opf & REQ_RAHEAD))
1788 queue_io(md, bio);
1789 else
1790 bio_io_error(bio);
1791 return ret;
1792 }
1793
1794 ret = dm_process_bio(md, map, bio);
1795
1796 dm_put_live_table(md, srcu_idx);
1797 return ret;
1798}
1799
1800static int dm_any_congested(void *congested_data, int bdi_bits)
1801{
1802 int r = bdi_bits;
1803 struct mapped_device *md = congested_data;
1804 struct dm_table *map;
1805
1806 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
1807 if (dm_request_based(md)) {
1808 /*
1809 * With request-based DM we only need to check the
1810 * top-level queue for congestion.
1811 */
1812 r = md->queue->backing_dev_info->wb.state & bdi_bits;
1813 } else {
1814 map = dm_get_live_table_fast(md);
1815 if (map)
1816 r = dm_table_any_congested(map, bdi_bits);
1817 dm_put_live_table_fast(md);
1818 }
1819 }
1820
1821 return r;
1822}
1823
1824/*-----------------------------------------------------------------
1825 * An IDR is used to keep track of allocated minor numbers.
1826 *---------------------------------------------------------------*/
1827static void free_minor(int minor)
1828{
1829 spin_lock(&_minor_lock);
1830 idr_remove(&_minor_idr, minor);
1831 spin_unlock(&_minor_lock);
1832}
1833
1834/*
1835 * See if the device with a specific minor # is free.
1836 */
1837static int specific_minor(int minor)
1838{
1839 int r;
1840
1841 if (minor >= (1 << MINORBITS))
1842 return -EINVAL;
1843
1844 idr_preload(GFP_KERNEL);
1845 spin_lock(&_minor_lock);
1846
1847 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
1848
1849 spin_unlock(&_minor_lock);
1850 idr_preload_end();
1851 if (r < 0)
1852 return r == -ENOSPC ? -EBUSY : r;
1853 return 0;
1854}
1855
1856static int next_free_minor(int *minor)
1857{
1858 int r;
1859
1860 idr_preload(GFP_KERNEL);
1861 spin_lock(&_minor_lock);
1862
1863 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
1864
1865 spin_unlock(&_minor_lock);
1866 idr_preload_end();
1867 if (r < 0)
1868 return r;
1869 *minor = r;
1870 return 0;
1871}
1872
1873static const struct block_device_operations dm_blk_dops;
1874static const struct dax_operations dm_dax_ops;
1875
1876static void dm_wq_work(struct work_struct *work);
1877
1878static void dm_init_normal_md_queue(struct mapped_device *md)
1879{
1880 /*
1881 * Initialize aspects of queue that aren't relevant for blk-mq
1882 */
1883 md->queue->backing_dev_info->congested_fn = dm_any_congested;
1884}
1885
1886static void cleanup_mapped_device(struct mapped_device *md)
1887{
1888 if (md->wq)
1889 destroy_workqueue(md->wq);
1890 bioset_exit(&md->bs);
1891 bioset_exit(&md->io_bs);
1892
1893 if (md->dax_dev) {
1894 kill_dax(md->dax_dev);
1895 put_dax(md->dax_dev);
1896 md->dax_dev = NULL;
1897 }
1898
1899 if (md->disk) {
1900 spin_lock(&_minor_lock);
1901 md->disk->private_data = NULL;
1902 spin_unlock(&_minor_lock);
1903 del_gendisk(md->disk);
1904 put_disk(md->disk);
1905 }
1906
1907 if (md->queue)
1908 blk_cleanup_queue(md->queue);
1909
1910 cleanup_srcu_struct(&md->io_barrier);
1911
1912 if (md->bdev) {
1913 bdput(md->bdev);
1914 md->bdev = NULL;
1915 }
1916
1917 mutex_destroy(&md->suspend_lock);
1918 mutex_destroy(&md->type_lock);
1919 mutex_destroy(&md->table_devices_lock);
1920
1921 dm_mq_cleanup_mapped_device(md);
1922}
1923
1924/*
1925 * Allocate and initialise a blank device with a given minor.
1926 */
1927static struct mapped_device *alloc_dev(int minor)
1928{
1929 int r, numa_node_id = dm_get_numa_node();
1930 struct mapped_device *md;
1931 void *old_md;
1932
1933 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
1934 if (!md) {
1935 DMWARN("unable to allocate device, out of memory.");
1936 return NULL;
1937 }
1938
1939 if (!try_module_get(THIS_MODULE))
1940 goto bad_module_get;
1941
1942 /* get a minor number for the dev */
1943 if (minor == DM_ANY_MINOR)
1944 r = next_free_minor(&minor);
1945 else
1946 r = specific_minor(minor);
1947 if (r < 0)
1948 goto bad_minor;
1949
1950 r = init_srcu_struct(&md->io_barrier);
1951 if (r < 0)
1952 goto bad_io_barrier;
1953
1954 md->numa_node_id = numa_node_id;
1955 md->init_tio_pdu = false;
1956 md->type = DM_TYPE_NONE;
1957 mutex_init(&md->suspend_lock);
1958 mutex_init(&md->type_lock);
1959 mutex_init(&md->table_devices_lock);
1960 spin_lock_init(&md->deferred_lock);
1961 atomic_set(&md->holders, 1);
1962 atomic_set(&md->open_count, 0);
1963 atomic_set(&md->event_nr, 0);
1964 atomic_set(&md->uevent_seq, 0);
1965 INIT_LIST_HEAD(&md->uevent_list);
1966 INIT_LIST_HEAD(&md->table_devices);
1967 spin_lock_init(&md->uevent_lock);
1968
1969 md->queue = blk_alloc_queue_node(GFP_KERNEL, numa_node_id);
1970 if (!md->queue)
1971 goto bad;
1972 md->queue->queuedata = md;
1973 md->queue->backing_dev_info->congested_data = md;
1974
1975 md->disk = alloc_disk_node(1, md->numa_node_id);
1976 if (!md->disk)
1977 goto bad;
1978
1979 init_waitqueue_head(&md->wait);
1980 INIT_WORK(&md->work, dm_wq_work);
1981 init_waitqueue_head(&md->eventq);
1982 init_completion(&md->kobj_holder.completion);
1983
1984 md->disk->major = _major;
1985 md->disk->first_minor = minor;
1986 md->disk->fops = &dm_blk_dops;
1987 md->disk->queue = md->queue;
1988 md->disk->private_data = md;
1989 sprintf(md->disk->disk_name, "dm-%d", minor);
1990
1991 if (IS_ENABLED(CONFIG_DAX_DRIVER)) {
1992 md->dax_dev = alloc_dax(md, md->disk->disk_name,
1993 &dm_dax_ops, 0);
1994 if (!md->dax_dev)
1995 goto bad;
1996 }
1997
1998 add_disk_no_queue_reg(md->disk);
1999 format_dev_t(md->name, MKDEV(_major, minor));
2000
2001 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
2002 if (!md->wq)
2003 goto bad;
2004
2005 md->bdev = bdget_disk(md->disk, 0);
2006 if (!md->bdev)
2007 goto bad;
2008
2009 dm_stats_init(&md->stats);
2010
2011 /* Populate the mapping, nobody knows we exist yet */
2012 spin_lock(&_minor_lock);
2013 old_md = idr_replace(&_minor_idr, md, minor);
2014 spin_unlock(&_minor_lock);
2015
2016 BUG_ON(old_md != MINOR_ALLOCED);
2017
2018 return md;
2019
2020bad:
2021 cleanup_mapped_device(md);
2022bad_io_barrier:
2023 free_minor(minor);
2024bad_minor:
2025 module_put(THIS_MODULE);
2026bad_module_get:
2027 kvfree(md);
2028 return NULL;
2029}
2030
2031static void unlock_fs(struct mapped_device *md);
2032
2033static void free_dev(struct mapped_device *md)
2034{
2035 int minor = MINOR(disk_devt(md->disk));
2036
2037 unlock_fs(md);
2038
2039 cleanup_mapped_device(md);
2040
2041 free_table_devices(&md->table_devices);
2042 dm_stats_cleanup(&md->stats);
2043 free_minor(minor);
2044
2045 module_put(THIS_MODULE);
2046 kvfree(md);
2047}
2048
2049static int __bind_mempools(struct mapped_device *md, struct dm_table *t)
2050{
2051 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
2052 int ret = 0;
2053
2054 if (dm_table_bio_based(t)) {
2055 /*
2056 * The md may already have mempools that need changing.
2057 * If so, reload bioset because front_pad may have changed
2058 * because a different table was loaded.
2059 */
2060 bioset_exit(&md->bs);
2061 bioset_exit(&md->io_bs);
2062
2063 } else if (bioset_initialized(&md->bs)) {
2064 /*
2065 * There's no need to reload with request-based dm
2066 * because the size of front_pad doesn't change.
2067 * Note for future: If you are to reload bioset,
2068 * prep-ed requests in the queue may refer
2069 * to bio from the old bioset, so you must walk
2070 * through the queue to unprep.
2071 */
2072 goto out;
2073 }
2074
2075 BUG_ON(!p ||
2076 bioset_initialized(&md->bs) ||
2077 bioset_initialized(&md->io_bs));
2078
2079 ret = bioset_init_from_src(&md->bs, &p->bs);
2080 if (ret)
2081 goto out;
2082 ret = bioset_init_from_src(&md->io_bs, &p->io_bs);
2083 if (ret)
2084 bioset_exit(&md->bs);
2085out:
2086 /* mempool bind completed, no longer need any mempools in the table */
2087 dm_table_free_md_mempools(t);
2088 return ret;
2089}
2090
2091/*
2092 * Bind a table to the device.
2093 */
2094static void event_callback(void *context)
2095{
2096 unsigned long flags;
2097 LIST_HEAD(uevents);
2098 struct mapped_device *md = (struct mapped_device *) context;
2099
2100 spin_lock_irqsave(&md->uevent_lock, flags);
2101 list_splice_init(&md->uevent_list, &uevents);
2102 spin_unlock_irqrestore(&md->uevent_lock, flags);
2103
2104 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2105
2106 atomic_inc(&md->event_nr);
2107 wake_up(&md->eventq);
2108 dm_issue_global_event();
2109}
2110
2111/*
2112 * Protected by md->suspend_lock obtained by dm_swap_table().
2113 */
2114static void __set_size(struct mapped_device *md, sector_t size)
2115{
2116 lockdep_assert_held(&md->suspend_lock);
2117
2118 set_capacity(md->disk, size);
2119
2120 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2121}
2122
2123/*
2124 * Returns old map, which caller must destroy.
2125 */
2126static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2127 struct queue_limits *limits)
2128{
2129 struct dm_table *old_map;
2130 struct request_queue *q = md->queue;
2131 bool request_based = dm_table_request_based(t);
2132 sector_t size;
2133 int ret;
2134
2135 lockdep_assert_held(&md->suspend_lock);
2136
2137 size = dm_table_get_size(t);
2138
2139 /*
2140 * Wipe any geometry if the size of the table changed.
2141 */
2142 if (size != dm_get_size(md))
2143 memset(&md->geometry, 0, sizeof(md->geometry));
2144
2145 __set_size(md, size);
2146
2147 dm_table_event_callback(t, event_callback, md);
2148
2149 /*
2150 * The queue hasn't been stopped yet, if the old table type wasn't
2151 * for request-based during suspension. So stop it to prevent
2152 * I/O mapping before resume.
2153 * This must be done before setting the queue restrictions,
2154 * because request-based dm may be run just after the setting.
2155 */
2156 if (request_based)
2157 dm_stop_queue(q);
2158
2159 if (request_based || md->type == DM_TYPE_NVME_BIO_BASED) {
2160 /*
2161 * Leverage the fact that request-based DM targets and
2162 * NVMe bio based targets are immutable singletons
2163 * - used to optimize both dm_request_fn and dm_mq_queue_rq;
2164 * and __process_bio.
2165 */
2166 md->immutable_target = dm_table_get_immutable_target(t);
2167 }
2168
2169 ret = __bind_mempools(md, t);
2170 if (ret) {
2171 old_map = ERR_PTR(ret);
2172 goto out;
2173 }
2174
2175 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2176 rcu_assign_pointer(md->map, (void *)t);
2177 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2178
2179 dm_table_set_restrictions(t, q, limits);
2180 if (old_map)
2181 dm_sync_table(md);
2182
2183out:
2184 return old_map;
2185}
2186
2187/*
2188 * Returns unbound table for the caller to free.
2189 */
2190static struct dm_table *__unbind(struct mapped_device *md)
2191{
2192 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2193
2194 if (!map)
2195 return NULL;
2196
2197 dm_table_event_callback(map, NULL, NULL);
2198 RCU_INIT_POINTER(md->map, NULL);
2199 dm_sync_table(md);
2200
2201 return map;
2202}
2203
2204/*
2205 * Constructor for a new device.
2206 */
2207int dm_create(int minor, struct mapped_device **result)
2208{
2209 int r;
2210 struct mapped_device *md;
2211
2212 md = alloc_dev(minor);
2213 if (!md)
2214 return -ENXIO;
2215
2216 r = dm_sysfs_init(md);
2217 if (r) {
2218 free_dev(md);
2219 return r;
2220 }
2221
2222 *result = md;
2223 return 0;
2224}
2225
2226/*
2227 * Functions to manage md->type.
2228 * All are required to hold md->type_lock.
2229 */
2230void dm_lock_md_type(struct mapped_device *md)
2231{
2232 mutex_lock(&md->type_lock);
2233}
2234
2235void dm_unlock_md_type(struct mapped_device *md)
2236{
2237 mutex_unlock(&md->type_lock);
2238}
2239
2240void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2241{
2242 BUG_ON(!mutex_is_locked(&md->type_lock));
2243 md->type = type;
2244}
2245
2246enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2247{
2248 return md->type;
2249}
2250
2251struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2252{
2253 return md->immutable_target_type;
2254}
2255
2256/*
2257 * The queue_limits are only valid as long as you have a reference
2258 * count on 'md'.
2259 */
2260struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2261{
2262 BUG_ON(!atomic_read(&md->holders));
2263 return &md->queue->limits;
2264}
2265EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2266
2267/*
2268 * Setup the DM device's queue based on md's type
2269 */
2270int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2271{
2272 int r;
2273 struct queue_limits limits;
2274 enum dm_queue_mode type = dm_get_md_type(md);
2275
2276 switch (type) {
2277 case DM_TYPE_REQUEST_BASED:
2278 r = dm_mq_init_request_queue(md, t);
2279 if (r) {
2280 DMERR("Cannot initialize queue for request-based dm-mq mapped device");
2281 return r;
2282 }
2283 break;
2284 case DM_TYPE_BIO_BASED:
2285 case DM_TYPE_DAX_BIO_BASED:
2286 case DM_TYPE_NVME_BIO_BASED:
2287 dm_init_normal_md_queue(md);
2288 blk_queue_make_request(md->queue, dm_make_request);
2289 break;
2290 case DM_TYPE_NONE:
2291 WARN_ON_ONCE(true);
2292 break;
2293 }
2294
2295 r = dm_calculate_queue_limits(t, &limits);
2296 if (r) {
2297 DMERR("Cannot calculate initial queue limits");
2298 return r;
2299 }
2300 dm_table_set_restrictions(t, md->queue, &limits);
2301 blk_register_queue(md->disk);
2302
2303 return 0;
2304}
2305
2306struct mapped_device *dm_get_md(dev_t dev)
2307{
2308 struct mapped_device *md;
2309 unsigned minor = MINOR(dev);
2310
2311 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2312 return NULL;
2313
2314 spin_lock(&_minor_lock);
2315
2316 md = idr_find(&_minor_idr, minor);
2317 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2318 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2319 md = NULL;
2320 goto out;
2321 }
2322 dm_get(md);
2323out:
2324 spin_unlock(&_minor_lock);
2325
2326 return md;
2327}
2328EXPORT_SYMBOL_GPL(dm_get_md);
2329
2330void *dm_get_mdptr(struct mapped_device *md)
2331{
2332 return md->interface_ptr;
2333}
2334
2335void dm_set_mdptr(struct mapped_device *md, void *ptr)
2336{
2337 md->interface_ptr = ptr;
2338}
2339
2340void dm_get(struct mapped_device *md)
2341{
2342 atomic_inc(&md->holders);
2343 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2344}
2345
2346int dm_hold(struct mapped_device *md)
2347{
2348 spin_lock(&_minor_lock);
2349 if (test_bit(DMF_FREEING, &md->flags)) {
2350 spin_unlock(&_minor_lock);
2351 return -EBUSY;
2352 }
2353 dm_get(md);
2354 spin_unlock(&_minor_lock);
2355 return 0;
2356}
2357EXPORT_SYMBOL_GPL(dm_hold);
2358
2359const char *dm_device_name(struct mapped_device *md)
2360{
2361 return md->name;
2362}
2363EXPORT_SYMBOL_GPL(dm_device_name);
2364
2365static void __dm_destroy(struct mapped_device *md, bool wait)
2366{
2367 struct dm_table *map;
2368 int srcu_idx;
2369
2370 might_sleep();
2371
2372 spin_lock(&_minor_lock);
2373 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2374 set_bit(DMF_FREEING, &md->flags);
2375 spin_unlock(&_minor_lock);
2376
2377 blk_set_queue_dying(md->queue);
2378
2379 /*
2380 * Take suspend_lock so that presuspend and postsuspend methods
2381 * do not race with internal suspend.
2382 */
2383 mutex_lock(&md->suspend_lock);
2384 map = dm_get_live_table(md, &srcu_idx);
2385 if (!dm_suspended_md(md)) {
2386 dm_table_presuspend_targets(map);
2387 dm_table_postsuspend_targets(map);
2388 }
2389 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2390 dm_put_live_table(md, srcu_idx);
2391 mutex_unlock(&md->suspend_lock);
2392
2393 /*
2394 * Rare, but there may be I/O requests still going to complete,
2395 * for example. Wait for all references to disappear.
2396 * No one should increment the reference count of the mapped_device,
2397 * after the mapped_device state becomes DMF_FREEING.
2398 */
2399 if (wait)
2400 while (atomic_read(&md->holders))
2401 msleep(1);
2402 else if (atomic_read(&md->holders))
2403 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2404 dm_device_name(md), atomic_read(&md->holders));
2405
2406 dm_sysfs_exit(md);
2407 dm_table_destroy(__unbind(md));
2408 free_dev(md);
2409}
2410
2411void dm_destroy(struct mapped_device *md)
2412{
2413 __dm_destroy(md, true);
2414}
2415
2416void dm_destroy_immediate(struct mapped_device *md)
2417{
2418 __dm_destroy(md, false);
2419}
2420
2421void dm_put(struct mapped_device *md)
2422{
2423 atomic_dec(&md->holders);
2424}
2425EXPORT_SYMBOL_GPL(dm_put);
2426
2427static int dm_wait_for_completion(struct mapped_device *md, long task_state)
2428{
2429 int r = 0;
2430 DEFINE_WAIT(wait);
2431
2432 while (1) {
2433 prepare_to_wait(&md->wait, &wait, task_state);
2434
2435 if (!md_in_flight(md))
2436 break;
2437
2438 if (signal_pending_state(task_state, current)) {
2439 r = -EINTR;
2440 break;
2441 }
2442
2443 io_schedule();
2444 }
2445 finish_wait(&md->wait, &wait);
2446
2447 return r;
2448}
2449
2450/*
2451 * Process the deferred bios
2452 */
2453static void dm_wq_work(struct work_struct *work)
2454{
2455 struct mapped_device *md = container_of(work, struct mapped_device,
2456 work);
2457 struct bio *c;
2458 int srcu_idx;
2459 struct dm_table *map;
2460
2461 map = dm_get_live_table(md, &srcu_idx);
2462
2463 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2464 spin_lock_irq(&md->deferred_lock);
2465 c = bio_list_pop(&md->deferred);
2466 spin_unlock_irq(&md->deferred_lock);
2467
2468 if (!c)
2469 break;
2470
2471 if (dm_request_based(md))
2472 (void) generic_make_request(c);
2473 else
2474 (void) dm_process_bio(md, map, c);
2475 }
2476
2477 dm_put_live_table(md, srcu_idx);
2478}
2479
2480static void dm_queue_flush(struct mapped_device *md)
2481{
2482 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2483 smp_mb__after_atomic();
2484 queue_work(md->wq, &md->work);
2485}
2486
2487/*
2488 * Swap in a new table, returning the old one for the caller to destroy.
2489 */
2490struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2491{
2492 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2493 struct queue_limits limits;
2494 int r;
2495
2496 mutex_lock(&md->suspend_lock);
2497
2498 /* device must be suspended */
2499 if (!dm_suspended_md(md))
2500 goto out;
2501
2502 /*
2503 * If the new table has no data devices, retain the existing limits.
2504 * This helps multipath with queue_if_no_path if all paths disappear,
2505 * then new I/O is queued based on these limits, and then some paths
2506 * reappear.
2507 */
2508 if (dm_table_has_no_data_devices(table)) {
2509 live_map = dm_get_live_table_fast(md);
2510 if (live_map)
2511 limits = md->queue->limits;
2512 dm_put_live_table_fast(md);
2513 }
2514
2515 if (!live_map) {
2516 r = dm_calculate_queue_limits(table, &limits);
2517 if (r) {
2518 map = ERR_PTR(r);
2519 goto out;
2520 }
2521 }
2522
2523 map = __bind(md, table, &limits);
2524 dm_issue_global_event();
2525
2526out:
2527 mutex_unlock(&md->suspend_lock);
2528 return map;
2529}
2530
2531/*
2532 * Functions to lock and unlock any filesystem running on the
2533 * device.
2534 */
2535static int lock_fs(struct mapped_device *md)
2536{
2537 int r;
2538
2539 WARN_ON(md->frozen_sb);
2540
2541 md->frozen_sb = freeze_bdev(md->bdev);
2542 if (IS_ERR(md->frozen_sb)) {
2543 r = PTR_ERR(md->frozen_sb);
2544 md->frozen_sb = NULL;
2545 return r;
2546 }
2547
2548 set_bit(DMF_FROZEN, &md->flags);
2549
2550 return 0;
2551}
2552
2553static void unlock_fs(struct mapped_device *md)
2554{
2555 if (!test_bit(DMF_FROZEN, &md->flags))
2556 return;
2557
2558 thaw_bdev(md->bdev, md->frozen_sb);
2559 md->frozen_sb = NULL;
2560 clear_bit(DMF_FROZEN, &md->flags);
2561}
2562
2563/*
2564 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2565 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2566 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2567 *
2568 * If __dm_suspend returns 0, the device is completely quiescent
2569 * now. There is no request-processing activity. All new requests
2570 * are being added to md->deferred list.
2571 */
2572static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2573 unsigned suspend_flags, long task_state,
2574 int dmf_suspended_flag)
2575{
2576 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2577 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2578 int r;
2579
2580 lockdep_assert_held(&md->suspend_lock);
2581
2582 /*
2583 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2584 * This flag is cleared before dm_suspend returns.
2585 */
2586 if (noflush)
2587 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2588 else
2589 pr_debug("%s: suspending with flush\n", dm_device_name(md));
2590
2591 /*
2592 * This gets reverted if there's an error later and the targets
2593 * provide the .presuspend_undo hook.
2594 */
2595 dm_table_presuspend_targets(map);
2596
2597 /*
2598 * Flush I/O to the device.
2599 * Any I/O submitted after lock_fs() may not be flushed.
2600 * noflush takes precedence over do_lockfs.
2601 * (lock_fs() flushes I/Os and waits for them to complete.)
2602 */
2603 if (!noflush && do_lockfs) {
2604 r = lock_fs(md);
2605 if (r) {
2606 dm_table_presuspend_undo_targets(map);
2607 return r;
2608 }
2609 }
2610
2611 /*
2612 * Here we must make sure that no processes are submitting requests
2613 * to target drivers i.e. no one may be executing
2614 * __split_and_process_bio. This is called from dm_request and
2615 * dm_wq_work.
2616 *
2617 * To get all processes out of __split_and_process_bio in dm_request,
2618 * we take the write lock. To prevent any process from reentering
2619 * __split_and_process_bio from dm_request and quiesce the thread
2620 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
2621 * flush_workqueue(md->wq).
2622 */
2623 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2624 if (map)
2625 synchronize_srcu(&md->io_barrier);
2626
2627 /*
2628 * Stop md->queue before flushing md->wq in case request-based
2629 * dm defers requests to md->wq from md->queue.
2630 */
2631 if (dm_request_based(md))
2632 dm_stop_queue(md->queue);
2633
2634 flush_workqueue(md->wq);
2635
2636 /*
2637 * At this point no more requests are entering target request routines.
2638 * We call dm_wait_for_completion to wait for all existing requests
2639 * to finish.
2640 */
2641 r = dm_wait_for_completion(md, task_state);
2642 if (!r)
2643 set_bit(dmf_suspended_flag, &md->flags);
2644
2645 if (noflush)
2646 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2647 if (map)
2648 synchronize_srcu(&md->io_barrier);
2649
2650 /* were we interrupted ? */
2651 if (r < 0) {
2652 dm_queue_flush(md);
2653
2654 if (dm_request_based(md))
2655 dm_start_queue(md->queue);
2656
2657 unlock_fs(md);
2658 dm_table_presuspend_undo_targets(map);
2659 /* pushback list is already flushed, so skip flush */
2660 }
2661
2662 return r;
2663}
2664
2665/*
2666 * We need to be able to change a mapping table under a mounted
2667 * filesystem. For example we might want to move some data in
2668 * the background. Before the table can be swapped with
2669 * dm_bind_table, dm_suspend must be called to flush any in
2670 * flight bios and ensure that any further io gets deferred.
2671 */
2672/*
2673 * Suspend mechanism in request-based dm.
2674 *
2675 * 1. Flush all I/Os by lock_fs() if needed.
2676 * 2. Stop dispatching any I/O by stopping the request_queue.
2677 * 3. Wait for all in-flight I/Os to be completed or requeued.
2678 *
2679 * To abort suspend, start the request_queue.
2680 */
2681int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
2682{
2683 struct dm_table *map = NULL;
2684 int r = 0;
2685
2686retry:
2687 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2688
2689 if (dm_suspended_md(md)) {
2690 r = -EINVAL;
2691 goto out_unlock;
2692 }
2693
2694 if (dm_suspended_internally_md(md)) {
2695 /* already internally suspended, wait for internal resume */
2696 mutex_unlock(&md->suspend_lock);
2697 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2698 if (r)
2699 return r;
2700 goto retry;
2701 }
2702
2703 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2704
2705 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2706 if (r)
2707 goto out_unlock;
2708
2709 dm_table_postsuspend_targets(map);
2710
2711out_unlock:
2712 mutex_unlock(&md->suspend_lock);
2713 return r;
2714}
2715
2716static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2717{
2718 if (map) {
2719 int r = dm_table_resume_targets(map);
2720 if (r)
2721 return r;
2722 }
2723
2724 dm_queue_flush(md);
2725
2726 /*
2727 * Flushing deferred I/Os must be done after targets are resumed
2728 * so that mapping of targets can work correctly.
2729 * Request-based dm is queueing the deferred I/Os in its request_queue.
2730 */
2731 if (dm_request_based(md))
2732 dm_start_queue(md->queue);
2733
2734 unlock_fs(md);
2735
2736 return 0;
2737}
2738
2739int dm_resume(struct mapped_device *md)
2740{
2741 int r;
2742 struct dm_table *map = NULL;
2743
2744retry:
2745 r = -EINVAL;
2746 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2747
2748 if (!dm_suspended_md(md))
2749 goto out;
2750
2751 if (dm_suspended_internally_md(md)) {
2752 /* already internally suspended, wait for internal resume */
2753 mutex_unlock(&md->suspend_lock);
2754 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2755 if (r)
2756 return r;
2757 goto retry;
2758 }
2759
2760 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2761 if (!map || !dm_table_get_size(map))
2762 goto out;
2763
2764 r = __dm_resume(md, map);
2765 if (r)
2766 goto out;
2767
2768 clear_bit(DMF_SUSPENDED, &md->flags);
2769out:
2770 mutex_unlock(&md->suspend_lock);
2771
2772 return r;
2773}
2774
2775/*
2776 * Internal suspend/resume works like userspace-driven suspend. It waits
2777 * until all bios finish and prevents issuing new bios to the target drivers.
2778 * It may be used only from the kernel.
2779 */
2780
2781static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
2782{
2783 struct dm_table *map = NULL;
2784
2785 lockdep_assert_held(&md->suspend_lock);
2786
2787 if (md->internal_suspend_count++)
2788 return; /* nested internal suspend */
2789
2790 if (dm_suspended_md(md)) {
2791 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2792 return; /* nest suspend */
2793 }
2794
2795 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2796
2797 /*
2798 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2799 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2800 * would require changing .presuspend to return an error -- avoid this
2801 * until there is a need for more elaborate variants of internal suspend.
2802 */
2803 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2804 DMF_SUSPENDED_INTERNALLY);
2805
2806 dm_table_postsuspend_targets(map);
2807}
2808
2809static void __dm_internal_resume(struct mapped_device *md)
2810{
2811 BUG_ON(!md->internal_suspend_count);
2812
2813 if (--md->internal_suspend_count)
2814 return; /* resume from nested internal suspend */
2815
2816 if (dm_suspended_md(md))
2817 goto done; /* resume from nested suspend */
2818
2819 /*
2820 * NOTE: existing callers don't need to call dm_table_resume_targets
2821 * (which may fail -- so best to avoid it for now by passing NULL map)
2822 */
2823 (void) __dm_resume(md, NULL);
2824
2825done:
2826 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2827 smp_mb__after_atomic();
2828 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
2829}
2830
2831void dm_internal_suspend_noflush(struct mapped_device *md)
2832{
2833 mutex_lock(&md->suspend_lock);
2834 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2835 mutex_unlock(&md->suspend_lock);
2836}
2837EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2838
2839void dm_internal_resume(struct mapped_device *md)
2840{
2841 mutex_lock(&md->suspend_lock);
2842 __dm_internal_resume(md);
2843 mutex_unlock(&md->suspend_lock);
2844}
2845EXPORT_SYMBOL_GPL(dm_internal_resume);
2846
2847/*
2848 * Fast variants of internal suspend/resume hold md->suspend_lock,
2849 * which prevents interaction with userspace-driven suspend.
2850 */
2851
2852void dm_internal_suspend_fast(struct mapped_device *md)
2853{
2854 mutex_lock(&md->suspend_lock);
2855 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2856 return;
2857
2858 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2859 synchronize_srcu(&md->io_barrier);
2860 flush_workqueue(md->wq);
2861 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
2862}
2863EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
2864
2865void dm_internal_resume_fast(struct mapped_device *md)
2866{
2867 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2868 goto done;
2869
2870 dm_queue_flush(md);
2871
2872done:
2873 mutex_unlock(&md->suspend_lock);
2874}
2875EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
2876
2877/*-----------------------------------------------------------------
2878 * Event notification.
2879 *---------------------------------------------------------------*/
2880int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
2881 unsigned cookie)
2882{
2883 char udev_cookie[DM_COOKIE_LENGTH];
2884 char *envp[] = { udev_cookie, NULL };
2885
2886 if (!cookie)
2887 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
2888 else {
2889 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
2890 DM_COOKIE_ENV_VAR_NAME, cookie);
2891 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
2892 action, envp);
2893 }
2894}
2895
2896uint32_t dm_next_uevent_seq(struct mapped_device *md)
2897{
2898 return atomic_add_return(1, &md->uevent_seq);
2899}
2900
2901uint32_t dm_get_event_nr(struct mapped_device *md)
2902{
2903 return atomic_read(&md->event_nr);
2904}
2905
2906int dm_wait_event(struct mapped_device *md, int event_nr)
2907{
2908 return wait_event_interruptible(md->eventq,
2909 (event_nr != atomic_read(&md->event_nr)));
2910}
2911
2912void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
2913{
2914 unsigned long flags;
2915
2916 spin_lock_irqsave(&md->uevent_lock, flags);
2917 list_add(elist, &md->uevent_list);
2918 spin_unlock_irqrestore(&md->uevent_lock, flags);
2919}
2920
2921/*
2922 * The gendisk is only valid as long as you have a reference
2923 * count on 'md'.
2924 */
2925struct gendisk *dm_disk(struct mapped_device *md)
2926{
2927 return md->disk;
2928}
2929EXPORT_SYMBOL_GPL(dm_disk);
2930
2931struct kobject *dm_kobject(struct mapped_device *md)
2932{
2933 return &md->kobj_holder.kobj;
2934}
2935
2936struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
2937{
2938 struct mapped_device *md;
2939
2940 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
2941
2942 spin_lock(&_minor_lock);
2943 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2944 md = NULL;
2945 goto out;
2946 }
2947 dm_get(md);
2948out:
2949 spin_unlock(&_minor_lock);
2950
2951 return md;
2952}
2953
2954int dm_suspended_md(struct mapped_device *md)
2955{
2956 return test_bit(DMF_SUSPENDED, &md->flags);
2957}
2958
2959int dm_suspended_internally_md(struct mapped_device *md)
2960{
2961 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2962}
2963
2964int dm_test_deferred_remove_flag(struct mapped_device *md)
2965{
2966 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
2967}
2968
2969int dm_suspended(struct dm_target *ti)
2970{
2971 return dm_suspended_md(dm_table_get_md(ti->table));
2972}
2973EXPORT_SYMBOL_GPL(dm_suspended);
2974
2975int dm_noflush_suspending(struct dm_target *ti)
2976{
2977 return __noflush_suspending(dm_table_get_md(ti->table));
2978}
2979EXPORT_SYMBOL_GPL(dm_noflush_suspending);
2980
2981struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, enum dm_queue_mode type,
2982 unsigned integrity, unsigned per_io_data_size,
2983 unsigned min_pool_size)
2984{
2985 struct dm_md_mempools *pools = kzalloc_node(sizeof(*pools), GFP_KERNEL, md->numa_node_id);
2986 unsigned int pool_size = 0;
2987 unsigned int front_pad, io_front_pad;
2988 int ret;
2989
2990 if (!pools)
2991 return NULL;
2992
2993 switch (type) {
2994 case DM_TYPE_BIO_BASED:
2995 case DM_TYPE_DAX_BIO_BASED:
2996 case DM_TYPE_NVME_BIO_BASED:
2997 pool_size = max(dm_get_reserved_bio_based_ios(), min_pool_size);
2998 front_pad = roundup(per_io_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
2999 io_front_pad = roundup(front_pad, __alignof__(struct dm_io)) + offsetof(struct dm_io, tio);
3000 ret = bioset_init(&pools->io_bs, pool_size, io_front_pad, 0);
3001 if (ret)
3002 goto out;
3003 if (integrity && bioset_integrity_create(&pools->io_bs, pool_size))
3004 goto out;
3005 break;
3006 case DM_TYPE_REQUEST_BASED:
3007 pool_size = max(dm_get_reserved_rq_based_ios(), min_pool_size);
3008 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
3009 /* per_io_data_size is used for blk-mq pdu at queue allocation */
3010 break;
3011 default:
3012 BUG();
3013 }
3014
3015 ret = bioset_init(&pools->bs, pool_size, front_pad, 0);
3016 if (ret)
3017 goto out;
3018
3019 if (integrity && bioset_integrity_create(&pools->bs, pool_size))
3020 goto out;
3021
3022 return pools;
3023
3024out:
3025 dm_free_md_mempools(pools);
3026
3027 return NULL;
3028}
3029
3030void dm_free_md_mempools(struct dm_md_mempools *pools)
3031{
3032 if (!pools)
3033 return;
3034
3035 bioset_exit(&pools->bs);
3036 bioset_exit(&pools->io_bs);
3037
3038 kfree(pools);
3039}
3040
3041struct dm_pr {
3042 u64 old_key;
3043 u64 new_key;
3044 u32 flags;
3045 bool fail_early;
3046};
3047
3048static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3049 void *data)
3050{
3051 struct mapped_device *md = bdev->bd_disk->private_data;
3052 struct dm_table *table;
3053 struct dm_target *ti;
3054 int ret = -ENOTTY, srcu_idx;
3055
3056 table = dm_get_live_table(md, &srcu_idx);
3057 if (!table || !dm_table_get_size(table))
3058 goto out;
3059
3060 /* We only support devices that have a single target */
3061 if (dm_table_get_num_targets(table) != 1)
3062 goto out;
3063 ti = dm_table_get_target(table, 0);
3064
3065 ret = -EINVAL;
3066 if (!ti->type->iterate_devices)
3067 goto out;
3068
3069 ret = ti->type->iterate_devices(ti, fn, data);
3070out:
3071 dm_put_live_table(md, srcu_idx);
3072 return ret;
3073}
3074
3075/*
3076 * For register / unregister we need to manually call out to every path.
3077 */
3078static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3079 sector_t start, sector_t len, void *data)
3080{
3081 struct dm_pr *pr = data;
3082 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3083
3084 if (!ops || !ops->pr_register)
3085 return -EOPNOTSUPP;
3086 return ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3087}
3088
3089static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3090 u32 flags)
3091{
3092 struct dm_pr pr = {
3093 .old_key = old_key,
3094 .new_key = new_key,
3095 .flags = flags,
3096 .fail_early = true,
3097 };
3098 int ret;
3099
3100 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3101 if (ret && new_key) {
3102 /* unregister all paths if we failed to register any path */
3103 pr.old_key = new_key;
3104 pr.new_key = 0;
3105 pr.flags = 0;
3106 pr.fail_early = false;
3107 dm_call_pr(bdev, __dm_pr_register, &pr);
3108 }
3109
3110 return ret;
3111}
3112
3113static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3114 u32 flags)
3115{
3116 struct mapped_device *md = bdev->bd_disk->private_data;
3117 const struct pr_ops *ops;
3118 int r, srcu_idx;
3119
3120 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3121 if (r < 0)
3122 goto out;
3123
3124 ops = bdev->bd_disk->fops->pr_ops;
3125 if (ops && ops->pr_reserve)
3126 r = ops->pr_reserve(bdev, key, type, flags);
3127 else
3128 r = -EOPNOTSUPP;
3129out:
3130 dm_unprepare_ioctl(md, srcu_idx);
3131 return r;
3132}
3133
3134static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3135{
3136 struct mapped_device *md = bdev->bd_disk->private_data;
3137 const struct pr_ops *ops;
3138 int r, srcu_idx;
3139
3140 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3141 if (r < 0)
3142 goto out;
3143
3144 ops = bdev->bd_disk->fops->pr_ops;
3145 if (ops && ops->pr_release)
3146 r = ops->pr_release(bdev, key, type);
3147 else
3148 r = -EOPNOTSUPP;
3149out:
3150 dm_unprepare_ioctl(md, srcu_idx);
3151 return r;
3152}
3153
3154static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3155 enum pr_type type, bool abort)
3156{
3157 struct mapped_device *md = bdev->bd_disk->private_data;
3158 const struct pr_ops *ops;
3159 int r, srcu_idx;
3160
3161 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3162 if (r < 0)
3163 goto out;
3164
3165 ops = bdev->bd_disk->fops->pr_ops;
3166 if (ops && ops->pr_preempt)
3167 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
3168 else
3169 r = -EOPNOTSUPP;
3170out:
3171 dm_unprepare_ioctl(md, srcu_idx);
3172 return r;
3173}
3174
3175static int dm_pr_clear(struct block_device *bdev, u64 key)
3176{
3177 struct mapped_device *md = bdev->bd_disk->private_data;
3178 const struct pr_ops *ops;
3179 int r, srcu_idx;
3180
3181 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3182 if (r < 0)
3183 goto out;
3184
3185 ops = bdev->bd_disk->fops->pr_ops;
3186 if (ops && ops->pr_clear)
3187 r = ops->pr_clear(bdev, key);
3188 else
3189 r = -EOPNOTSUPP;
3190out:
3191 dm_unprepare_ioctl(md, srcu_idx);
3192 return r;
3193}
3194
3195static const struct pr_ops dm_pr_ops = {
3196 .pr_register = dm_pr_register,
3197 .pr_reserve = dm_pr_reserve,
3198 .pr_release = dm_pr_release,
3199 .pr_preempt = dm_pr_preempt,
3200 .pr_clear = dm_pr_clear,
3201};
3202
3203static const struct block_device_operations dm_blk_dops = {
3204 .open = dm_blk_open,
3205 .release = dm_blk_close,
3206 .ioctl = dm_blk_ioctl,
3207 .getgeo = dm_blk_getgeo,
3208 .report_zones = dm_blk_report_zones,
3209 .pr_ops = &dm_pr_ops,
3210 .owner = THIS_MODULE
3211};
3212
3213static const struct dax_operations dm_dax_ops = {
3214 .direct_access = dm_dax_direct_access,
3215 .dax_supported = dm_dax_supported,
3216 .copy_from_iter = dm_dax_copy_from_iter,
3217 .copy_to_iter = dm_dax_copy_to_iter,
3218};
3219
3220/*
3221 * module hooks
3222 */
3223module_init(dm_init);
3224module_exit(dm_exit);
3225
3226module_param(major, uint, 0);
3227MODULE_PARM_DESC(major, "The major number of the device mapper");
3228
3229module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3230MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3231
3232module_param(dm_numa_node, int, S_IRUGO | S_IWUSR);
3233MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3234
3235MODULE_DESCRIPTION(DM_NAME " driver");
3236MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3237MODULE_LICENSE("GPL");