Loading...
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#include "dm-ima.h"
12
13#include <linux/init.h>
14#include <linux/module.h>
15#include <linux/mutex.h>
16#include <linux/sched/mm.h>
17#include <linux/sched/signal.h>
18#include <linux/blkpg.h>
19#include <linux/bio.h>
20#include <linux/mempool.h>
21#include <linux/dax.h>
22#include <linux/slab.h>
23#include <linux/idr.h>
24#include <linux/uio.h>
25#include <linux/hdreg.h>
26#include <linux/delay.h>
27#include <linux/wait.h>
28#include <linux/pr.h>
29#include <linux/refcount.h>
30#include <linux/part_stat.h>
31#include <linux/blk-crypto.h>
32#include <linux/blk-crypto-profile.h>
33
34#define DM_MSG_PREFIX "core"
35
36/*
37 * Cookies are numeric values sent with CHANGE and REMOVE
38 * uevents while resuming, removing or renaming the device.
39 */
40#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
41#define DM_COOKIE_LENGTH 24
42
43/*
44 * For REQ_POLLED fs bio, this flag is set if we link mapped underlying
45 * dm_io into one list, and reuse bio->bi_private as the list head. Before
46 * ending this fs bio, we will recover its ->bi_private.
47 */
48#define REQ_DM_POLL_LIST REQ_DRV
49
50static const char *_name = DM_NAME;
51
52static unsigned int major = 0;
53static unsigned int _major = 0;
54
55static DEFINE_IDR(_minor_idr);
56
57static DEFINE_SPINLOCK(_minor_lock);
58
59static void do_deferred_remove(struct work_struct *w);
60
61static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
62
63static struct workqueue_struct *deferred_remove_workqueue;
64
65atomic_t dm_global_event_nr = ATOMIC_INIT(0);
66DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
67
68void dm_issue_global_event(void)
69{
70 atomic_inc(&dm_global_event_nr);
71 wake_up(&dm_global_eventq);
72}
73
74DEFINE_STATIC_KEY_FALSE(stats_enabled);
75DEFINE_STATIC_KEY_FALSE(swap_bios_enabled);
76DEFINE_STATIC_KEY_FALSE(zoned_enabled);
77
78/*
79 * One of these is allocated (on-stack) per original bio.
80 */
81struct clone_info {
82 struct dm_table *map;
83 struct bio *bio;
84 struct dm_io *io;
85 sector_t sector;
86 unsigned sector_count;
87 bool is_abnormal_io:1;
88 bool submit_as_polled:1;
89};
90
91static inline struct dm_target_io *clone_to_tio(struct bio *clone)
92{
93 return container_of(clone, struct dm_target_io, clone);
94}
95
96void *dm_per_bio_data(struct bio *bio, size_t data_size)
97{
98 if (!dm_tio_flagged(clone_to_tio(bio), DM_TIO_INSIDE_DM_IO))
99 return (char *)bio - DM_TARGET_IO_BIO_OFFSET - data_size;
100 return (char *)bio - DM_IO_BIO_OFFSET - data_size;
101}
102EXPORT_SYMBOL_GPL(dm_per_bio_data);
103
104struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
105{
106 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
107 if (io->magic == DM_IO_MAGIC)
108 return (struct bio *)((char *)io + DM_IO_BIO_OFFSET);
109 BUG_ON(io->magic != DM_TIO_MAGIC);
110 return (struct bio *)((char *)io + DM_TARGET_IO_BIO_OFFSET);
111}
112EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
113
114unsigned dm_bio_get_target_bio_nr(const struct bio *bio)
115{
116 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
117}
118EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
119
120#define MINOR_ALLOCED ((void *)-1)
121
122#define DM_NUMA_NODE NUMA_NO_NODE
123static int dm_numa_node = DM_NUMA_NODE;
124
125#define DEFAULT_SWAP_BIOS (8 * 1048576 / PAGE_SIZE)
126static int swap_bios = DEFAULT_SWAP_BIOS;
127static int get_swap_bios(void)
128{
129 int latch = READ_ONCE(swap_bios);
130 if (unlikely(latch <= 0))
131 latch = DEFAULT_SWAP_BIOS;
132 return latch;
133}
134
135struct table_device {
136 struct list_head list;
137 refcount_t count;
138 struct dm_dev dm_dev;
139};
140
141/*
142 * Bio-based DM's mempools' reserved IOs set by the user.
143 */
144#define RESERVED_BIO_BASED_IOS 16
145static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
146
147static int __dm_get_module_param_int(int *module_param, int min, int max)
148{
149 int param = READ_ONCE(*module_param);
150 int modified_param = 0;
151 bool modified = true;
152
153 if (param < min)
154 modified_param = min;
155 else if (param > max)
156 modified_param = max;
157 else
158 modified = false;
159
160 if (modified) {
161 (void)cmpxchg(module_param, param, modified_param);
162 param = modified_param;
163 }
164
165 return param;
166}
167
168unsigned __dm_get_module_param(unsigned *module_param,
169 unsigned def, unsigned max)
170{
171 unsigned param = READ_ONCE(*module_param);
172 unsigned modified_param = 0;
173
174 if (!param)
175 modified_param = def;
176 else if (param > max)
177 modified_param = max;
178
179 if (modified_param) {
180 (void)cmpxchg(module_param, param, modified_param);
181 param = modified_param;
182 }
183
184 return param;
185}
186
187unsigned dm_get_reserved_bio_based_ios(void)
188{
189 return __dm_get_module_param(&reserved_bio_based_ios,
190 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
191}
192EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
193
194static unsigned dm_get_numa_node(void)
195{
196 return __dm_get_module_param_int(&dm_numa_node,
197 DM_NUMA_NODE, num_online_nodes() - 1);
198}
199
200static int __init local_init(void)
201{
202 int r;
203
204 r = dm_uevent_init();
205 if (r)
206 return r;
207
208 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
209 if (!deferred_remove_workqueue) {
210 r = -ENOMEM;
211 goto out_uevent_exit;
212 }
213
214 _major = major;
215 r = register_blkdev(_major, _name);
216 if (r < 0)
217 goto out_free_workqueue;
218
219 if (!_major)
220 _major = r;
221
222 return 0;
223
224out_free_workqueue:
225 destroy_workqueue(deferred_remove_workqueue);
226out_uevent_exit:
227 dm_uevent_exit();
228
229 return r;
230}
231
232static void local_exit(void)
233{
234 flush_scheduled_work();
235 destroy_workqueue(deferred_remove_workqueue);
236
237 unregister_blkdev(_major, _name);
238 dm_uevent_exit();
239
240 _major = 0;
241
242 DMINFO("cleaned up");
243}
244
245static int (*_inits[])(void) __initdata = {
246 local_init,
247 dm_target_init,
248 dm_linear_init,
249 dm_stripe_init,
250 dm_io_init,
251 dm_kcopyd_init,
252 dm_interface_init,
253 dm_statistics_init,
254};
255
256static void (*_exits[])(void) = {
257 local_exit,
258 dm_target_exit,
259 dm_linear_exit,
260 dm_stripe_exit,
261 dm_io_exit,
262 dm_kcopyd_exit,
263 dm_interface_exit,
264 dm_statistics_exit,
265};
266
267static int __init dm_init(void)
268{
269 const int count = ARRAY_SIZE(_inits);
270 int r, i;
271
272#if (IS_ENABLED(CONFIG_IMA) && !IS_ENABLED(CONFIG_IMA_DISABLE_HTABLE))
273 DMWARN("CONFIG_IMA_DISABLE_HTABLE is disabled."
274 " Duplicate IMA measurements will not be recorded in the IMA log.");
275#endif
276
277 for (i = 0; i < count; i++) {
278 r = _inits[i]();
279 if (r)
280 goto bad;
281 }
282
283 return 0;
284bad:
285 while (i--)
286 _exits[i]();
287
288 return r;
289}
290
291static void __exit dm_exit(void)
292{
293 int i = ARRAY_SIZE(_exits);
294
295 while (i--)
296 _exits[i]();
297
298 /*
299 * Should be empty by this point.
300 */
301 idr_destroy(&_minor_idr);
302}
303
304/*
305 * Block device functions
306 */
307int dm_deleting_md(struct mapped_device *md)
308{
309 return test_bit(DMF_DELETING, &md->flags);
310}
311
312static int dm_blk_open(struct block_device *bdev, fmode_t mode)
313{
314 struct mapped_device *md;
315
316 spin_lock(&_minor_lock);
317
318 md = bdev->bd_disk->private_data;
319 if (!md)
320 goto out;
321
322 if (test_bit(DMF_FREEING, &md->flags) ||
323 dm_deleting_md(md)) {
324 md = NULL;
325 goto out;
326 }
327
328 dm_get(md);
329 atomic_inc(&md->open_count);
330out:
331 spin_unlock(&_minor_lock);
332
333 return md ? 0 : -ENXIO;
334}
335
336static void dm_blk_close(struct gendisk *disk, fmode_t mode)
337{
338 struct mapped_device *md;
339
340 spin_lock(&_minor_lock);
341
342 md = disk->private_data;
343 if (WARN_ON(!md))
344 goto out;
345
346 if (atomic_dec_and_test(&md->open_count) &&
347 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
348 queue_work(deferred_remove_workqueue, &deferred_remove_work);
349
350 dm_put(md);
351out:
352 spin_unlock(&_minor_lock);
353}
354
355int dm_open_count(struct mapped_device *md)
356{
357 return atomic_read(&md->open_count);
358}
359
360/*
361 * Guarantees nothing is using the device before it's deleted.
362 */
363int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
364{
365 int r = 0;
366
367 spin_lock(&_minor_lock);
368
369 if (dm_open_count(md)) {
370 r = -EBUSY;
371 if (mark_deferred)
372 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
373 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
374 r = -EEXIST;
375 else
376 set_bit(DMF_DELETING, &md->flags);
377
378 spin_unlock(&_minor_lock);
379
380 return r;
381}
382
383int dm_cancel_deferred_remove(struct mapped_device *md)
384{
385 int r = 0;
386
387 spin_lock(&_minor_lock);
388
389 if (test_bit(DMF_DELETING, &md->flags))
390 r = -EBUSY;
391 else
392 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
393
394 spin_unlock(&_minor_lock);
395
396 return r;
397}
398
399static void do_deferred_remove(struct work_struct *w)
400{
401 dm_deferred_remove();
402}
403
404static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
405{
406 struct mapped_device *md = bdev->bd_disk->private_data;
407
408 return dm_get_geometry(md, geo);
409}
410
411static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
412 struct block_device **bdev)
413{
414 struct dm_target *ti;
415 struct dm_table *map;
416 int r;
417
418retry:
419 r = -ENOTTY;
420 map = dm_get_live_table(md, srcu_idx);
421 if (!map || !dm_table_get_size(map))
422 return r;
423
424 /* We only support devices that have a single target */
425 if (map->num_targets != 1)
426 return r;
427
428 ti = dm_table_get_target(map, 0);
429 if (!ti->type->prepare_ioctl)
430 return r;
431
432 if (dm_suspended_md(md))
433 return -EAGAIN;
434
435 r = ti->type->prepare_ioctl(ti, bdev);
436 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
437 dm_put_live_table(md, *srcu_idx);
438 msleep(10);
439 goto retry;
440 }
441
442 return r;
443}
444
445static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
446{
447 dm_put_live_table(md, srcu_idx);
448}
449
450static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
451 unsigned int cmd, unsigned long arg)
452{
453 struct mapped_device *md = bdev->bd_disk->private_data;
454 int r, srcu_idx;
455
456 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
457 if (r < 0)
458 goto out;
459
460 if (r > 0) {
461 /*
462 * Target determined this ioctl is being issued against a
463 * subset of the parent bdev; require extra privileges.
464 */
465 if (!capable(CAP_SYS_RAWIO)) {
466 DMDEBUG_LIMIT(
467 "%s: sending ioctl %x to DM device without required privilege.",
468 current->comm, cmd);
469 r = -ENOIOCTLCMD;
470 goto out;
471 }
472 }
473
474 if (!bdev->bd_disk->fops->ioctl)
475 r = -ENOTTY;
476 else
477 r = bdev->bd_disk->fops->ioctl(bdev, mode, cmd, arg);
478out:
479 dm_unprepare_ioctl(md, srcu_idx);
480 return r;
481}
482
483u64 dm_start_time_ns_from_clone(struct bio *bio)
484{
485 return jiffies_to_nsecs(clone_to_tio(bio)->io->start_time);
486}
487EXPORT_SYMBOL_GPL(dm_start_time_ns_from_clone);
488
489static bool bio_is_flush_with_data(struct bio *bio)
490{
491 return ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size);
492}
493
494static void dm_io_acct(struct dm_io *io, bool end)
495{
496 struct dm_stats_aux *stats_aux = &io->stats_aux;
497 unsigned long start_time = io->start_time;
498 struct mapped_device *md = io->md;
499 struct bio *bio = io->orig_bio;
500 unsigned int sectors;
501
502 /*
503 * If REQ_PREFLUSH set, don't account payload, it will be
504 * submitted (and accounted) after this flush completes.
505 */
506 if (bio_is_flush_with_data(bio))
507 sectors = 0;
508 else if (likely(!(dm_io_flagged(io, DM_IO_WAS_SPLIT))))
509 sectors = bio_sectors(bio);
510 else
511 sectors = io->sectors;
512
513 if (!end)
514 bdev_start_io_acct(bio->bi_bdev, sectors, bio_op(bio),
515 start_time);
516 else
517 bdev_end_io_acct(bio->bi_bdev, bio_op(bio), start_time);
518
519 if (static_branch_unlikely(&stats_enabled) &&
520 unlikely(dm_stats_used(&md->stats))) {
521 sector_t sector;
522
523 if (likely(!dm_io_flagged(io, DM_IO_WAS_SPLIT)))
524 sector = bio->bi_iter.bi_sector;
525 else
526 sector = bio_end_sector(bio) - io->sector_offset;
527
528 dm_stats_account_io(&md->stats, bio_data_dir(bio),
529 sector, sectors,
530 end, start_time, stats_aux);
531 }
532}
533
534static void __dm_start_io_acct(struct dm_io *io)
535{
536 dm_io_acct(io, false);
537}
538
539static void dm_start_io_acct(struct dm_io *io, struct bio *clone)
540{
541 /*
542 * Ensure IO accounting is only ever started once.
543 */
544 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
545 return;
546
547 /* Expect no possibility for race unless DM_TIO_IS_DUPLICATE_BIO. */
548 if (!clone || likely(dm_tio_is_normal(clone_to_tio(clone)))) {
549 dm_io_set_flag(io, DM_IO_ACCOUNTED);
550 } else {
551 unsigned long flags;
552 /* Can afford locking given DM_TIO_IS_DUPLICATE_BIO */
553 spin_lock_irqsave(&io->lock, flags);
554 if (dm_io_flagged(io, DM_IO_ACCOUNTED)) {
555 spin_unlock_irqrestore(&io->lock, flags);
556 return;
557 }
558 dm_io_set_flag(io, DM_IO_ACCOUNTED);
559 spin_unlock_irqrestore(&io->lock, flags);
560 }
561
562 __dm_start_io_acct(io);
563}
564
565static void dm_end_io_acct(struct dm_io *io)
566{
567 dm_io_acct(io, true);
568}
569
570static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio)
571{
572 struct dm_io *io;
573 struct dm_target_io *tio;
574 struct bio *clone;
575
576 clone = bio_alloc_clone(NULL, bio, GFP_NOIO, &md->mempools->io_bs);
577 tio = clone_to_tio(clone);
578 tio->flags = 0;
579 dm_tio_set_flag(tio, DM_TIO_INSIDE_DM_IO);
580 tio->io = NULL;
581
582 io = container_of(tio, struct dm_io, tio);
583 io->magic = DM_IO_MAGIC;
584 io->status = BLK_STS_OK;
585
586 /* one ref is for submission, the other is for completion */
587 atomic_set(&io->io_count, 2);
588 this_cpu_inc(*md->pending_io);
589 io->orig_bio = bio;
590 io->md = md;
591 spin_lock_init(&io->lock);
592 io->start_time = jiffies;
593 io->flags = 0;
594
595 if (static_branch_unlikely(&stats_enabled))
596 dm_stats_record_start(&md->stats, &io->stats_aux);
597
598 return io;
599}
600
601static void free_io(struct dm_io *io)
602{
603 bio_put(&io->tio.clone);
604}
605
606static struct bio *alloc_tio(struct clone_info *ci, struct dm_target *ti,
607 unsigned target_bio_nr, unsigned *len, gfp_t gfp_mask)
608{
609 struct mapped_device *md = ci->io->md;
610 struct dm_target_io *tio;
611 struct bio *clone;
612
613 if (!ci->io->tio.io) {
614 /* the dm_target_io embedded in ci->io is available */
615 tio = &ci->io->tio;
616 /* alloc_io() already initialized embedded clone */
617 clone = &tio->clone;
618 } else {
619 clone = bio_alloc_clone(NULL, ci->bio, gfp_mask,
620 &md->mempools->bs);
621 if (!clone)
622 return NULL;
623
624 /* REQ_DM_POLL_LIST shouldn't be inherited */
625 clone->bi_opf &= ~REQ_DM_POLL_LIST;
626
627 tio = clone_to_tio(clone);
628 tio->flags = 0; /* also clears DM_TIO_INSIDE_DM_IO */
629 }
630
631 tio->magic = DM_TIO_MAGIC;
632 tio->io = ci->io;
633 tio->ti = ti;
634 tio->target_bio_nr = target_bio_nr;
635 tio->len_ptr = len;
636 tio->old_sector = 0;
637
638 /* Set default bdev, but target must bio_set_dev() before issuing IO */
639 clone->bi_bdev = md->disk->part0;
640 if (unlikely(ti->needs_bio_set_dev))
641 bio_set_dev(clone, md->disk->part0);
642
643 if (len) {
644 clone->bi_iter.bi_size = to_bytes(*len);
645 if (bio_integrity(clone))
646 bio_integrity_trim(clone);
647 }
648
649 return clone;
650}
651
652static void free_tio(struct bio *clone)
653{
654 if (dm_tio_flagged(clone_to_tio(clone), DM_TIO_INSIDE_DM_IO))
655 return;
656 bio_put(clone);
657}
658
659/*
660 * Add the bio to the list of deferred io.
661 */
662static void queue_io(struct mapped_device *md, struct bio *bio)
663{
664 unsigned long flags;
665
666 spin_lock_irqsave(&md->deferred_lock, flags);
667 bio_list_add(&md->deferred, bio);
668 spin_unlock_irqrestore(&md->deferred_lock, flags);
669 queue_work(md->wq, &md->work);
670}
671
672/*
673 * Everyone (including functions in this file), should use this
674 * function to access the md->map field, and make sure they call
675 * dm_put_live_table() when finished.
676 */
677struct dm_table *dm_get_live_table(struct mapped_device *md,
678 int *srcu_idx) __acquires(md->io_barrier)
679{
680 *srcu_idx = srcu_read_lock(&md->io_barrier);
681
682 return srcu_dereference(md->map, &md->io_barrier);
683}
684
685void dm_put_live_table(struct mapped_device *md,
686 int srcu_idx) __releases(md->io_barrier)
687{
688 srcu_read_unlock(&md->io_barrier, srcu_idx);
689}
690
691void dm_sync_table(struct mapped_device *md)
692{
693 synchronize_srcu(&md->io_barrier);
694 synchronize_rcu_expedited();
695}
696
697/*
698 * A fast alternative to dm_get_live_table/dm_put_live_table.
699 * The caller must not block between these two functions.
700 */
701static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
702{
703 rcu_read_lock();
704 return rcu_dereference(md->map);
705}
706
707static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
708{
709 rcu_read_unlock();
710}
711
712static inline struct dm_table *dm_get_live_table_bio(struct mapped_device *md,
713 int *srcu_idx, blk_opf_t bio_opf)
714{
715 if (bio_opf & REQ_NOWAIT)
716 return dm_get_live_table_fast(md);
717 else
718 return dm_get_live_table(md, srcu_idx);
719}
720
721static inline void dm_put_live_table_bio(struct mapped_device *md, int srcu_idx,
722 blk_opf_t bio_opf)
723{
724 if (bio_opf & REQ_NOWAIT)
725 dm_put_live_table_fast(md);
726 else
727 dm_put_live_table(md, srcu_idx);
728}
729
730static char *_dm_claim_ptr = "I belong to device-mapper";
731
732/*
733 * Open a table device so we can use it as a map destination.
734 */
735static struct table_device *open_table_device(struct mapped_device *md,
736 dev_t dev, fmode_t mode)
737{
738 struct table_device *td;
739 struct block_device *bdev;
740 u64 part_off;
741 int r;
742
743 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
744 if (!td)
745 return ERR_PTR(-ENOMEM);
746 refcount_set(&td->count, 1);
747
748 bdev = blkdev_get_by_dev(dev, mode | FMODE_EXCL, _dm_claim_ptr);
749 if (IS_ERR(bdev)) {
750 r = PTR_ERR(bdev);
751 goto out_free_td;
752 }
753
754 /*
755 * We can be called before the dm disk is added. In that case we can't
756 * register the holder relation here. It will be done once add_disk was
757 * called.
758 */
759 if (md->disk->slave_dir) {
760 r = bd_link_disk_holder(bdev, md->disk);
761 if (r)
762 goto out_blkdev_put;
763 }
764
765 td->dm_dev.mode = mode;
766 td->dm_dev.bdev = bdev;
767 td->dm_dev.dax_dev = fs_dax_get_by_bdev(bdev, &part_off, NULL, NULL);
768 format_dev_t(td->dm_dev.name, dev);
769 list_add(&td->list, &md->table_devices);
770 return td;
771
772out_blkdev_put:
773 blkdev_put(bdev, mode | FMODE_EXCL);
774out_free_td:
775 kfree(td);
776 return ERR_PTR(r);
777}
778
779/*
780 * Close a table device that we've been using.
781 */
782static void close_table_device(struct table_device *td, struct mapped_device *md)
783{
784 if (md->disk->slave_dir)
785 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
786 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
787 put_dax(td->dm_dev.dax_dev);
788 list_del(&td->list);
789 kfree(td);
790}
791
792static struct table_device *find_table_device(struct list_head *l, dev_t dev,
793 fmode_t mode)
794{
795 struct table_device *td;
796
797 list_for_each_entry(td, l, list)
798 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
799 return td;
800
801 return NULL;
802}
803
804int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
805 struct dm_dev **result)
806{
807 struct table_device *td;
808
809 mutex_lock(&md->table_devices_lock);
810 td = find_table_device(&md->table_devices, dev, mode);
811 if (!td) {
812 td = open_table_device(md, dev, mode);
813 if (IS_ERR(td)) {
814 mutex_unlock(&md->table_devices_lock);
815 return PTR_ERR(td);
816 }
817 } else {
818 refcount_inc(&td->count);
819 }
820 mutex_unlock(&md->table_devices_lock);
821
822 *result = &td->dm_dev;
823 return 0;
824}
825
826void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
827{
828 struct table_device *td = container_of(d, struct table_device, dm_dev);
829
830 mutex_lock(&md->table_devices_lock);
831 if (refcount_dec_and_test(&td->count))
832 close_table_device(td, md);
833 mutex_unlock(&md->table_devices_lock);
834}
835
836/*
837 * Get the geometry associated with a dm device
838 */
839int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
840{
841 *geo = md->geometry;
842
843 return 0;
844}
845
846/*
847 * Set the geometry of a device.
848 */
849int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
850{
851 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
852
853 if (geo->start > sz) {
854 DMERR("Start sector is beyond the geometry limits.");
855 return -EINVAL;
856 }
857
858 md->geometry = *geo;
859
860 return 0;
861}
862
863static int __noflush_suspending(struct mapped_device *md)
864{
865 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
866}
867
868static void dm_requeue_add_io(struct dm_io *io, bool first_stage)
869{
870 struct mapped_device *md = io->md;
871
872 if (first_stage) {
873 struct dm_io *next = md->requeue_list;
874
875 md->requeue_list = io;
876 io->next = next;
877 } else {
878 bio_list_add_head(&md->deferred, io->orig_bio);
879 }
880}
881
882static void dm_kick_requeue(struct mapped_device *md, bool first_stage)
883{
884 if (first_stage)
885 queue_work(md->wq, &md->requeue_work);
886 else
887 queue_work(md->wq, &md->work);
888}
889
890/*
891 * Return true if the dm_io's original bio is requeued.
892 * io->status is updated with error if requeue disallowed.
893 */
894static bool dm_handle_requeue(struct dm_io *io, bool first_stage)
895{
896 struct bio *bio = io->orig_bio;
897 bool handle_requeue = (io->status == BLK_STS_DM_REQUEUE);
898 bool handle_polled_eagain = ((io->status == BLK_STS_AGAIN) &&
899 (bio->bi_opf & REQ_POLLED));
900 struct mapped_device *md = io->md;
901 bool requeued = false;
902
903 if (handle_requeue || handle_polled_eagain) {
904 unsigned long flags;
905
906 if (bio->bi_opf & REQ_POLLED) {
907 /*
908 * Upper layer won't help us poll split bio
909 * (io->orig_bio may only reflect a subset of the
910 * pre-split original) so clear REQ_POLLED.
911 */
912 bio_clear_polled(bio);
913 }
914
915 /*
916 * Target requested pushing back the I/O or
917 * polled IO hit BLK_STS_AGAIN.
918 */
919 spin_lock_irqsave(&md->deferred_lock, flags);
920 if ((__noflush_suspending(md) &&
921 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) ||
922 handle_polled_eagain || first_stage) {
923 dm_requeue_add_io(io, first_stage);
924 requeued = true;
925 } else {
926 /*
927 * noflush suspend was interrupted or this is
928 * a write to a zoned target.
929 */
930 io->status = BLK_STS_IOERR;
931 }
932 spin_unlock_irqrestore(&md->deferred_lock, flags);
933 }
934
935 if (requeued)
936 dm_kick_requeue(md, first_stage);
937
938 return requeued;
939}
940
941static void __dm_io_complete(struct dm_io *io, bool first_stage)
942{
943 struct bio *bio = io->orig_bio;
944 struct mapped_device *md = io->md;
945 blk_status_t io_error;
946 bool requeued;
947
948 requeued = dm_handle_requeue(io, first_stage);
949 if (requeued && first_stage)
950 return;
951
952 io_error = io->status;
953 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
954 dm_end_io_acct(io);
955 else if (!io_error) {
956 /*
957 * Must handle target that DM_MAPIO_SUBMITTED only to
958 * then bio_endio() rather than dm_submit_bio_remap()
959 */
960 __dm_start_io_acct(io);
961 dm_end_io_acct(io);
962 }
963 free_io(io);
964 smp_wmb();
965 this_cpu_dec(*md->pending_io);
966
967 /* nudge anyone waiting on suspend queue */
968 if (unlikely(wq_has_sleeper(&md->wait)))
969 wake_up(&md->wait);
970
971 /* Return early if the original bio was requeued */
972 if (requeued)
973 return;
974
975 if (bio_is_flush_with_data(bio)) {
976 /*
977 * Preflush done for flush with data, reissue
978 * without REQ_PREFLUSH.
979 */
980 bio->bi_opf &= ~REQ_PREFLUSH;
981 queue_io(md, bio);
982 } else {
983 /* done with normal IO or empty flush */
984 if (io_error)
985 bio->bi_status = io_error;
986 bio_endio(bio);
987 }
988}
989
990static void dm_wq_requeue_work(struct work_struct *work)
991{
992 struct mapped_device *md = container_of(work, struct mapped_device,
993 requeue_work);
994 unsigned long flags;
995 struct dm_io *io;
996
997 /* reuse deferred lock to simplify dm_handle_requeue */
998 spin_lock_irqsave(&md->deferred_lock, flags);
999 io = md->requeue_list;
1000 md->requeue_list = NULL;
1001 spin_unlock_irqrestore(&md->deferred_lock, flags);
1002
1003 while (io) {
1004 struct dm_io *next = io->next;
1005
1006 dm_io_rewind(io, &md->disk->bio_split);
1007
1008 io->next = NULL;
1009 __dm_io_complete(io, false);
1010 io = next;
1011 }
1012}
1013
1014/*
1015 * Two staged requeue:
1016 *
1017 * 1) io->orig_bio points to the real original bio, and the part mapped to
1018 * this io must be requeued, instead of other parts of the original bio.
1019 *
1020 * 2) io->orig_bio points to new cloned bio which matches the requeued dm_io.
1021 */
1022static void dm_io_complete(struct dm_io *io)
1023{
1024 bool first_requeue;
1025
1026 /*
1027 * Only dm_io that has been split needs two stage requeue, otherwise
1028 * we may run into long bio clone chain during suspend and OOM could
1029 * be triggered.
1030 *
1031 * Also flush data dm_io won't be marked as DM_IO_WAS_SPLIT, so they
1032 * also aren't handled via the first stage requeue.
1033 */
1034 if (dm_io_flagged(io, DM_IO_WAS_SPLIT))
1035 first_requeue = true;
1036 else
1037 first_requeue = false;
1038
1039 __dm_io_complete(io, first_requeue);
1040}
1041
1042/*
1043 * Decrements the number of outstanding ios that a bio has been
1044 * cloned into, completing the original io if necc.
1045 */
1046static inline void __dm_io_dec_pending(struct dm_io *io)
1047{
1048 if (atomic_dec_and_test(&io->io_count))
1049 dm_io_complete(io);
1050}
1051
1052static void dm_io_set_error(struct dm_io *io, blk_status_t error)
1053{
1054 unsigned long flags;
1055
1056 /* Push-back supersedes any I/O errors */
1057 spin_lock_irqsave(&io->lock, flags);
1058 if (!(io->status == BLK_STS_DM_REQUEUE &&
1059 __noflush_suspending(io->md))) {
1060 io->status = error;
1061 }
1062 spin_unlock_irqrestore(&io->lock, flags);
1063}
1064
1065static void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
1066{
1067 if (unlikely(error))
1068 dm_io_set_error(io, error);
1069
1070 __dm_io_dec_pending(io);
1071}
1072
1073void disable_discard(struct mapped_device *md)
1074{
1075 struct queue_limits *limits = dm_get_queue_limits(md);
1076
1077 /* device doesn't really support DISCARD, disable it */
1078 limits->max_discard_sectors = 0;
1079}
1080
1081void disable_write_zeroes(struct mapped_device *md)
1082{
1083 struct queue_limits *limits = dm_get_queue_limits(md);
1084
1085 /* device doesn't really support WRITE ZEROES, disable it */
1086 limits->max_write_zeroes_sectors = 0;
1087}
1088
1089static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
1090{
1091 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
1092}
1093
1094static void clone_endio(struct bio *bio)
1095{
1096 blk_status_t error = bio->bi_status;
1097 struct dm_target_io *tio = clone_to_tio(bio);
1098 struct dm_target *ti = tio->ti;
1099 dm_endio_fn endio = ti->type->end_io;
1100 struct dm_io *io = tio->io;
1101 struct mapped_device *md = io->md;
1102
1103 if (unlikely(error == BLK_STS_TARGET)) {
1104 if (bio_op(bio) == REQ_OP_DISCARD &&
1105 !bdev_max_discard_sectors(bio->bi_bdev))
1106 disable_discard(md);
1107 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
1108 !bdev_write_zeroes_sectors(bio->bi_bdev))
1109 disable_write_zeroes(md);
1110 }
1111
1112 if (static_branch_unlikely(&zoned_enabled) &&
1113 unlikely(bdev_is_zoned(bio->bi_bdev)))
1114 dm_zone_endio(io, bio);
1115
1116 if (endio) {
1117 int r = endio(ti, bio, &error);
1118 switch (r) {
1119 case DM_ENDIO_REQUEUE:
1120 if (static_branch_unlikely(&zoned_enabled)) {
1121 /*
1122 * Requeuing writes to a sequential zone of a zoned
1123 * target will break the sequential write pattern:
1124 * fail such IO.
1125 */
1126 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
1127 error = BLK_STS_IOERR;
1128 else
1129 error = BLK_STS_DM_REQUEUE;
1130 } else
1131 error = BLK_STS_DM_REQUEUE;
1132 fallthrough;
1133 case DM_ENDIO_DONE:
1134 break;
1135 case DM_ENDIO_INCOMPLETE:
1136 /* The target will handle the io */
1137 return;
1138 default:
1139 DMCRIT("unimplemented target endio return value: %d", r);
1140 BUG();
1141 }
1142 }
1143
1144 if (static_branch_unlikely(&swap_bios_enabled) &&
1145 unlikely(swap_bios_limit(ti, bio)))
1146 up(&md->swap_bios_semaphore);
1147
1148 free_tio(bio);
1149 dm_io_dec_pending(io, error);
1150}
1151
1152/*
1153 * Return maximum size of I/O possible at the supplied sector up to the current
1154 * target boundary.
1155 */
1156static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
1157 sector_t target_offset)
1158{
1159 return ti->len - target_offset;
1160}
1161
1162static sector_t max_io_len(struct dm_target *ti, sector_t sector)
1163{
1164 sector_t target_offset = dm_target_offset(ti, sector);
1165 sector_t len = max_io_len_target_boundary(ti, target_offset);
1166
1167 /*
1168 * Does the target need to split IO even further?
1169 * - varied (per target) IO splitting is a tenet of DM; this
1170 * explains why stacked chunk_sectors based splitting via
1171 * bio_split_to_limits() isn't possible here.
1172 */
1173 if (!ti->max_io_len)
1174 return len;
1175 return min_t(sector_t, len,
1176 min(queue_max_sectors(ti->table->md->queue),
1177 blk_chunk_sectors_left(target_offset, ti->max_io_len)));
1178}
1179
1180int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1181{
1182 if (len > UINT_MAX) {
1183 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1184 (unsigned long long)len, UINT_MAX);
1185 ti->error = "Maximum size of target IO is too large";
1186 return -EINVAL;
1187 }
1188
1189 ti->max_io_len = (uint32_t) len;
1190
1191 return 0;
1192}
1193EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1194
1195static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1196 sector_t sector, int *srcu_idx)
1197 __acquires(md->io_barrier)
1198{
1199 struct dm_table *map;
1200 struct dm_target *ti;
1201
1202 map = dm_get_live_table(md, srcu_idx);
1203 if (!map)
1204 return NULL;
1205
1206 ti = dm_table_find_target(map, sector);
1207 if (!ti)
1208 return NULL;
1209
1210 return ti;
1211}
1212
1213static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1214 long nr_pages, enum dax_access_mode mode, void **kaddr,
1215 pfn_t *pfn)
1216{
1217 struct mapped_device *md = dax_get_private(dax_dev);
1218 sector_t sector = pgoff * PAGE_SECTORS;
1219 struct dm_target *ti;
1220 long len, ret = -EIO;
1221 int srcu_idx;
1222
1223 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1224
1225 if (!ti)
1226 goto out;
1227 if (!ti->type->direct_access)
1228 goto out;
1229 len = max_io_len(ti, sector) / PAGE_SECTORS;
1230 if (len < 1)
1231 goto out;
1232 nr_pages = min(len, nr_pages);
1233 ret = ti->type->direct_access(ti, pgoff, nr_pages, mode, kaddr, pfn);
1234
1235 out:
1236 dm_put_live_table(md, srcu_idx);
1237
1238 return ret;
1239}
1240
1241static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1242 size_t nr_pages)
1243{
1244 struct mapped_device *md = dax_get_private(dax_dev);
1245 sector_t sector = pgoff * PAGE_SECTORS;
1246 struct dm_target *ti;
1247 int ret = -EIO;
1248 int srcu_idx;
1249
1250 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1251
1252 if (!ti)
1253 goto out;
1254 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1255 /*
1256 * ->zero_page_range() is mandatory dax operation. If we are
1257 * here, something is wrong.
1258 */
1259 goto out;
1260 }
1261 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1262 out:
1263 dm_put_live_table(md, srcu_idx);
1264
1265 return ret;
1266}
1267
1268static size_t dm_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
1269 void *addr, size_t bytes, struct iov_iter *i)
1270{
1271 struct mapped_device *md = dax_get_private(dax_dev);
1272 sector_t sector = pgoff * PAGE_SECTORS;
1273 struct dm_target *ti;
1274 int srcu_idx;
1275 long ret = 0;
1276
1277 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1278 if (!ti || !ti->type->dax_recovery_write)
1279 goto out;
1280
1281 ret = ti->type->dax_recovery_write(ti, pgoff, addr, bytes, i);
1282out:
1283 dm_put_live_table(md, srcu_idx);
1284 return ret;
1285}
1286
1287/*
1288 * A target may call dm_accept_partial_bio only from the map routine. It is
1289 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1290 * operations, REQ_OP_ZONE_APPEND (zone append writes) and any bio serviced by
1291 * __send_duplicate_bios().
1292 *
1293 * dm_accept_partial_bio informs the dm that the target only wants to process
1294 * additional n_sectors sectors of the bio and the rest of the data should be
1295 * sent in a next bio.
1296 *
1297 * A diagram that explains the arithmetics:
1298 * +--------------------+---------------+-------+
1299 * | 1 | 2 | 3 |
1300 * +--------------------+---------------+-------+
1301 *
1302 * <-------------- *tio->len_ptr --------------->
1303 * <----- bio_sectors ----->
1304 * <-- n_sectors -->
1305 *
1306 * Region 1 was already iterated over with bio_advance or similar function.
1307 * (it may be empty if the target doesn't use bio_advance)
1308 * Region 2 is the remaining bio size that the target wants to process.
1309 * (it may be empty if region 1 is non-empty, although there is no reason
1310 * to make it empty)
1311 * The target requires that region 3 is to be sent in the next bio.
1312 *
1313 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1314 * the partially processed part (the sum of regions 1+2) must be the same for all
1315 * copies of the bio.
1316 */
1317void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1318{
1319 struct dm_target_io *tio = clone_to_tio(bio);
1320 struct dm_io *io = tio->io;
1321 unsigned bio_sectors = bio_sectors(bio);
1322
1323 BUG_ON(dm_tio_flagged(tio, DM_TIO_IS_DUPLICATE_BIO));
1324 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1325 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1326 BUG_ON(bio_sectors > *tio->len_ptr);
1327 BUG_ON(n_sectors > bio_sectors);
1328
1329 *tio->len_ptr -= bio_sectors - n_sectors;
1330 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1331
1332 /*
1333 * __split_and_process_bio() may have already saved mapped part
1334 * for accounting but it is being reduced so update accordingly.
1335 */
1336 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1337 io->sectors = n_sectors;
1338 io->sector_offset = bio_sectors(io->orig_bio);
1339}
1340EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1341
1342/*
1343 * @clone: clone bio that DM core passed to target's .map function
1344 * @tgt_clone: clone of @clone bio that target needs submitted
1345 *
1346 * Targets should use this interface to submit bios they take
1347 * ownership of when returning DM_MAPIO_SUBMITTED.
1348 *
1349 * Target should also enable ti->accounts_remapped_io
1350 */
1351void dm_submit_bio_remap(struct bio *clone, struct bio *tgt_clone)
1352{
1353 struct dm_target_io *tio = clone_to_tio(clone);
1354 struct dm_io *io = tio->io;
1355
1356 /* establish bio that will get submitted */
1357 if (!tgt_clone)
1358 tgt_clone = clone;
1359
1360 /*
1361 * Account io->origin_bio to DM dev on behalf of target
1362 * that took ownership of IO with DM_MAPIO_SUBMITTED.
1363 */
1364 dm_start_io_acct(io, clone);
1365
1366 trace_block_bio_remap(tgt_clone, disk_devt(io->md->disk),
1367 tio->old_sector);
1368 submit_bio_noacct(tgt_clone);
1369}
1370EXPORT_SYMBOL_GPL(dm_submit_bio_remap);
1371
1372static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1373{
1374 mutex_lock(&md->swap_bios_lock);
1375 while (latch < md->swap_bios) {
1376 cond_resched();
1377 down(&md->swap_bios_semaphore);
1378 md->swap_bios--;
1379 }
1380 while (latch > md->swap_bios) {
1381 cond_resched();
1382 up(&md->swap_bios_semaphore);
1383 md->swap_bios++;
1384 }
1385 mutex_unlock(&md->swap_bios_lock);
1386}
1387
1388static void __map_bio(struct bio *clone)
1389{
1390 struct dm_target_io *tio = clone_to_tio(clone);
1391 struct dm_target *ti = tio->ti;
1392 struct dm_io *io = tio->io;
1393 struct mapped_device *md = io->md;
1394 int r;
1395
1396 clone->bi_end_io = clone_endio;
1397
1398 /*
1399 * Map the clone.
1400 */
1401 tio->old_sector = clone->bi_iter.bi_sector;
1402
1403 if (static_branch_unlikely(&swap_bios_enabled) &&
1404 unlikely(swap_bios_limit(ti, clone))) {
1405 int latch = get_swap_bios();
1406 if (unlikely(latch != md->swap_bios))
1407 __set_swap_bios_limit(md, latch);
1408 down(&md->swap_bios_semaphore);
1409 }
1410
1411 if (static_branch_unlikely(&zoned_enabled)) {
1412 /*
1413 * Check if the IO needs a special mapping due to zone append
1414 * emulation on zoned target. In this case, dm_zone_map_bio()
1415 * calls the target map operation.
1416 */
1417 if (unlikely(dm_emulate_zone_append(md)))
1418 r = dm_zone_map_bio(tio);
1419 else
1420 r = ti->type->map(ti, clone);
1421 } else
1422 r = ti->type->map(ti, clone);
1423
1424 switch (r) {
1425 case DM_MAPIO_SUBMITTED:
1426 /* target has assumed ownership of this io */
1427 if (!ti->accounts_remapped_io)
1428 dm_start_io_acct(io, clone);
1429 break;
1430 case DM_MAPIO_REMAPPED:
1431 dm_submit_bio_remap(clone, NULL);
1432 break;
1433 case DM_MAPIO_KILL:
1434 case DM_MAPIO_REQUEUE:
1435 if (static_branch_unlikely(&swap_bios_enabled) &&
1436 unlikely(swap_bios_limit(ti, clone)))
1437 up(&md->swap_bios_semaphore);
1438 free_tio(clone);
1439 if (r == DM_MAPIO_KILL)
1440 dm_io_dec_pending(io, BLK_STS_IOERR);
1441 else
1442 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1443 break;
1444 default:
1445 DMCRIT("unimplemented target map return value: %d", r);
1446 BUG();
1447 }
1448}
1449
1450static void setup_split_accounting(struct clone_info *ci, unsigned len)
1451{
1452 struct dm_io *io = ci->io;
1453
1454 if (ci->sector_count > len) {
1455 /*
1456 * Split needed, save the mapped part for accounting.
1457 * NOTE: dm_accept_partial_bio() will update accordingly.
1458 */
1459 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1460 io->sectors = len;
1461 io->sector_offset = bio_sectors(ci->bio);
1462 }
1463}
1464
1465static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1466 struct dm_target *ti, unsigned num_bios)
1467{
1468 struct bio *bio;
1469 int try;
1470
1471 for (try = 0; try < 2; try++) {
1472 int bio_nr;
1473
1474 if (try)
1475 mutex_lock(&ci->io->md->table_devices_lock);
1476 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1477 bio = alloc_tio(ci, ti, bio_nr, NULL,
1478 try ? GFP_NOIO : GFP_NOWAIT);
1479 if (!bio)
1480 break;
1481
1482 bio_list_add(blist, bio);
1483 }
1484 if (try)
1485 mutex_unlock(&ci->io->md->table_devices_lock);
1486 if (bio_nr == num_bios)
1487 return;
1488
1489 while ((bio = bio_list_pop(blist)))
1490 free_tio(bio);
1491 }
1492}
1493
1494static int __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1495 unsigned int num_bios, unsigned *len)
1496{
1497 struct bio_list blist = BIO_EMPTY_LIST;
1498 struct bio *clone;
1499 unsigned int ret = 0;
1500
1501 switch (num_bios) {
1502 case 0:
1503 break;
1504 case 1:
1505 if (len)
1506 setup_split_accounting(ci, *len);
1507 clone = alloc_tio(ci, ti, 0, len, GFP_NOIO);
1508 __map_bio(clone);
1509 ret = 1;
1510 break;
1511 default:
1512 /* dm_accept_partial_bio() is not supported with shared tio->len_ptr */
1513 alloc_multiple_bios(&blist, ci, ti, num_bios);
1514 while ((clone = bio_list_pop(&blist))) {
1515 dm_tio_set_flag(clone_to_tio(clone), DM_TIO_IS_DUPLICATE_BIO);
1516 __map_bio(clone);
1517 ret += 1;
1518 }
1519 break;
1520 }
1521
1522 return ret;
1523}
1524
1525static void __send_empty_flush(struct clone_info *ci)
1526{
1527 struct dm_table *t = ci->map;
1528 struct bio flush_bio;
1529
1530 /*
1531 * Use an on-stack bio for this, it's safe since we don't
1532 * need to reference it after submit. It's just used as
1533 * the basis for the clone(s).
1534 */
1535 bio_init(&flush_bio, ci->io->md->disk->part0, NULL, 0,
1536 REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
1537
1538 ci->bio = &flush_bio;
1539 ci->sector_count = 0;
1540 ci->io->tio.clone.bi_iter.bi_size = 0;
1541
1542 for (unsigned int i = 0; i < t->num_targets; i++) {
1543 unsigned int bios;
1544 struct dm_target *ti = dm_table_get_target(t, i);
1545
1546 atomic_add(ti->num_flush_bios, &ci->io->io_count);
1547 bios = __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1548 atomic_sub(ti->num_flush_bios - bios, &ci->io->io_count);
1549 }
1550
1551 /*
1552 * alloc_io() takes one extra reference for submission, so the
1553 * reference won't reach 0 without the following subtraction
1554 */
1555 atomic_sub(1, &ci->io->io_count);
1556
1557 bio_uninit(ci->bio);
1558}
1559
1560static void __send_changing_extent_only(struct clone_info *ci, struct dm_target *ti,
1561 unsigned num_bios)
1562{
1563 unsigned len;
1564 unsigned int bios;
1565
1566 len = min_t(sector_t, ci->sector_count,
1567 max_io_len_target_boundary(ti, dm_target_offset(ti, ci->sector)));
1568
1569 atomic_add(num_bios, &ci->io->io_count);
1570 bios = __send_duplicate_bios(ci, ti, num_bios, &len);
1571 /*
1572 * alloc_io() takes one extra reference for submission, so the
1573 * reference won't reach 0 without the following (+1) subtraction
1574 */
1575 atomic_sub(num_bios - bios + 1, &ci->io->io_count);
1576
1577 ci->sector += len;
1578 ci->sector_count -= len;
1579}
1580
1581static bool is_abnormal_io(struct bio *bio)
1582{
1583 enum req_op op = bio_op(bio);
1584
1585 if (op != REQ_OP_READ && op != REQ_OP_WRITE && op != REQ_OP_FLUSH) {
1586 switch (op) {
1587 case REQ_OP_DISCARD:
1588 case REQ_OP_SECURE_ERASE:
1589 case REQ_OP_WRITE_ZEROES:
1590 return true;
1591 default:
1592 break;
1593 }
1594 }
1595
1596 return false;
1597}
1598
1599static blk_status_t __process_abnormal_io(struct clone_info *ci,
1600 struct dm_target *ti)
1601{
1602 unsigned num_bios = 0;
1603
1604 switch (bio_op(ci->bio)) {
1605 case REQ_OP_DISCARD:
1606 num_bios = ti->num_discard_bios;
1607 break;
1608 case REQ_OP_SECURE_ERASE:
1609 num_bios = ti->num_secure_erase_bios;
1610 break;
1611 case REQ_OP_WRITE_ZEROES:
1612 num_bios = ti->num_write_zeroes_bios;
1613 break;
1614 default:
1615 break;
1616 }
1617
1618 /*
1619 * Even though the device advertised support for this type of
1620 * request, that does not mean every target supports it, and
1621 * reconfiguration might also have changed that since the
1622 * check was performed.
1623 */
1624 if (unlikely(!num_bios))
1625 return BLK_STS_NOTSUPP;
1626
1627 __send_changing_extent_only(ci, ti, num_bios);
1628 return BLK_STS_OK;
1629}
1630
1631/*
1632 * Reuse ->bi_private as dm_io list head for storing all dm_io instances
1633 * associated with this bio, and this bio's bi_private needs to be
1634 * stored in dm_io->data before the reuse.
1635 *
1636 * bio->bi_private is owned by fs or upper layer, so block layer won't
1637 * touch it after splitting. Meantime it won't be changed by anyone after
1638 * bio is submitted. So this reuse is safe.
1639 */
1640static inline struct dm_io **dm_poll_list_head(struct bio *bio)
1641{
1642 return (struct dm_io **)&bio->bi_private;
1643}
1644
1645static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
1646{
1647 struct dm_io **head = dm_poll_list_head(bio);
1648
1649 if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
1650 bio->bi_opf |= REQ_DM_POLL_LIST;
1651 /*
1652 * Save .bi_private into dm_io, so that we can reuse
1653 * .bi_private as dm_io list head for storing dm_io list
1654 */
1655 io->data = bio->bi_private;
1656
1657 /* tell block layer to poll for completion */
1658 bio->bi_cookie = ~BLK_QC_T_NONE;
1659
1660 io->next = NULL;
1661 } else {
1662 /*
1663 * bio recursed due to split, reuse original poll list,
1664 * and save bio->bi_private too.
1665 */
1666 io->data = (*head)->data;
1667 io->next = *head;
1668 }
1669
1670 *head = io;
1671}
1672
1673/*
1674 * Select the correct strategy for processing a non-flush bio.
1675 */
1676static blk_status_t __split_and_process_bio(struct clone_info *ci)
1677{
1678 struct bio *clone;
1679 struct dm_target *ti;
1680 unsigned len;
1681
1682 ti = dm_table_find_target(ci->map, ci->sector);
1683 if (unlikely(!ti))
1684 return BLK_STS_IOERR;
1685
1686 if (unlikely((ci->bio->bi_opf & REQ_NOWAIT) != 0) &&
1687 unlikely(!dm_target_supports_nowait(ti->type)))
1688 return BLK_STS_NOTSUPP;
1689
1690 if (unlikely(ci->is_abnormal_io))
1691 return __process_abnormal_io(ci, ti);
1692
1693 /*
1694 * Only support bio polling for normal IO, and the target io is
1695 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
1696 */
1697 ci->submit_as_polled = !!(ci->bio->bi_opf & REQ_POLLED);
1698
1699 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1700 setup_split_accounting(ci, len);
1701 clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
1702 __map_bio(clone);
1703
1704 ci->sector += len;
1705 ci->sector_count -= len;
1706
1707 return BLK_STS_OK;
1708}
1709
1710static void init_clone_info(struct clone_info *ci, struct mapped_device *md,
1711 struct dm_table *map, struct bio *bio, bool is_abnormal)
1712{
1713 ci->map = map;
1714 ci->io = alloc_io(md, bio);
1715 ci->bio = bio;
1716 ci->is_abnormal_io = is_abnormal;
1717 ci->submit_as_polled = false;
1718 ci->sector = bio->bi_iter.bi_sector;
1719 ci->sector_count = bio_sectors(bio);
1720
1721 /* Shouldn't happen but sector_count was being set to 0 so... */
1722 if (static_branch_unlikely(&zoned_enabled) &&
1723 WARN_ON_ONCE(op_is_zone_mgmt(bio_op(bio)) && ci->sector_count))
1724 ci->sector_count = 0;
1725}
1726
1727/*
1728 * Entry point to split a bio into clones and submit them to the targets.
1729 */
1730static void dm_split_and_process_bio(struct mapped_device *md,
1731 struct dm_table *map, struct bio *bio)
1732{
1733 struct clone_info ci;
1734 struct dm_io *io;
1735 blk_status_t error = BLK_STS_OK;
1736 bool is_abnormal;
1737
1738 is_abnormal = is_abnormal_io(bio);
1739 if (unlikely(is_abnormal)) {
1740 /*
1741 * Use bio_split_to_limits() for abnormal IO (e.g. discard, etc)
1742 * otherwise associated queue_limits won't be imposed.
1743 */
1744 bio = bio_split_to_limits(bio);
1745 if (!bio)
1746 return;
1747 }
1748
1749 init_clone_info(&ci, md, map, bio, is_abnormal);
1750 io = ci.io;
1751
1752 if (bio->bi_opf & REQ_PREFLUSH) {
1753 __send_empty_flush(&ci);
1754 /* dm_io_complete submits any data associated with flush */
1755 goto out;
1756 }
1757
1758 error = __split_and_process_bio(&ci);
1759 if (error || !ci.sector_count)
1760 goto out;
1761 /*
1762 * Remainder must be passed to submit_bio_noacct() so it gets handled
1763 * *after* bios already submitted have been completely processed.
1764 */
1765 bio_trim(bio, io->sectors, ci.sector_count);
1766 trace_block_split(bio, bio->bi_iter.bi_sector);
1767 bio_inc_remaining(bio);
1768 submit_bio_noacct(bio);
1769out:
1770 /*
1771 * Drop the extra reference count for non-POLLED bio, and hold one
1772 * reference for POLLED bio, which will be released in dm_poll_bio
1773 *
1774 * Add every dm_io instance into the dm_io list head which is stored
1775 * in bio->bi_private, so that dm_poll_bio can poll them all.
1776 */
1777 if (error || !ci.submit_as_polled) {
1778 /*
1779 * In case of submission failure, the extra reference for
1780 * submitting io isn't consumed yet
1781 */
1782 if (error)
1783 atomic_dec(&io->io_count);
1784 dm_io_dec_pending(io, error);
1785 } else
1786 dm_queue_poll_io(bio, io);
1787}
1788
1789static void dm_submit_bio(struct bio *bio)
1790{
1791 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
1792 int srcu_idx;
1793 struct dm_table *map;
1794 blk_opf_t bio_opf = bio->bi_opf;
1795
1796 map = dm_get_live_table_bio(md, &srcu_idx, bio_opf);
1797
1798 /* If suspended, or map not yet available, queue this IO for later */
1799 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) ||
1800 unlikely(!map)) {
1801 if (bio->bi_opf & REQ_NOWAIT)
1802 bio_wouldblock_error(bio);
1803 else if (bio->bi_opf & REQ_RAHEAD)
1804 bio_io_error(bio);
1805 else
1806 queue_io(md, bio);
1807 goto out;
1808 }
1809
1810 dm_split_and_process_bio(md, map, bio);
1811out:
1812 dm_put_live_table_bio(md, srcu_idx, bio_opf);
1813}
1814
1815static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
1816 unsigned int flags)
1817{
1818 WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
1819
1820 /* don't poll if the mapped io is done */
1821 if (atomic_read(&io->io_count) > 1)
1822 bio_poll(&io->tio.clone, iob, flags);
1823
1824 /* bio_poll holds the last reference */
1825 return atomic_read(&io->io_count) == 1;
1826}
1827
1828static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
1829 unsigned int flags)
1830{
1831 struct dm_io **head = dm_poll_list_head(bio);
1832 struct dm_io *list = *head;
1833 struct dm_io *tmp = NULL;
1834 struct dm_io *curr, *next;
1835
1836 /* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
1837 if (!(bio->bi_opf & REQ_DM_POLL_LIST))
1838 return 0;
1839
1840 WARN_ON_ONCE(!list);
1841
1842 /*
1843 * Restore .bi_private before possibly completing dm_io.
1844 *
1845 * bio_poll() is only possible once @bio has been completely
1846 * submitted via submit_bio_noacct()'s depth-first submission.
1847 * So there is no dm_queue_poll_io() race associated with
1848 * clearing REQ_DM_POLL_LIST here.
1849 */
1850 bio->bi_opf &= ~REQ_DM_POLL_LIST;
1851 bio->bi_private = list->data;
1852
1853 for (curr = list, next = curr->next; curr; curr = next, next =
1854 curr ? curr->next : NULL) {
1855 if (dm_poll_dm_io(curr, iob, flags)) {
1856 /*
1857 * clone_endio() has already occurred, so no
1858 * error handling is needed here.
1859 */
1860 __dm_io_dec_pending(curr);
1861 } else {
1862 curr->next = tmp;
1863 tmp = curr;
1864 }
1865 }
1866
1867 /* Not done? */
1868 if (tmp) {
1869 bio->bi_opf |= REQ_DM_POLL_LIST;
1870 /* Reset bio->bi_private to dm_io list head */
1871 *head = tmp;
1872 return 0;
1873 }
1874 return 1;
1875}
1876
1877/*-----------------------------------------------------------------
1878 * An IDR is used to keep track of allocated minor numbers.
1879 *---------------------------------------------------------------*/
1880static void free_minor(int minor)
1881{
1882 spin_lock(&_minor_lock);
1883 idr_remove(&_minor_idr, minor);
1884 spin_unlock(&_minor_lock);
1885}
1886
1887/*
1888 * See if the device with a specific minor # is free.
1889 */
1890static int specific_minor(int minor)
1891{
1892 int r;
1893
1894 if (minor >= (1 << MINORBITS))
1895 return -EINVAL;
1896
1897 idr_preload(GFP_KERNEL);
1898 spin_lock(&_minor_lock);
1899
1900 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
1901
1902 spin_unlock(&_minor_lock);
1903 idr_preload_end();
1904 if (r < 0)
1905 return r == -ENOSPC ? -EBUSY : r;
1906 return 0;
1907}
1908
1909static int next_free_minor(int *minor)
1910{
1911 int r;
1912
1913 idr_preload(GFP_KERNEL);
1914 spin_lock(&_minor_lock);
1915
1916 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
1917
1918 spin_unlock(&_minor_lock);
1919 idr_preload_end();
1920 if (r < 0)
1921 return r;
1922 *minor = r;
1923 return 0;
1924}
1925
1926static const struct block_device_operations dm_blk_dops;
1927static const struct block_device_operations dm_rq_blk_dops;
1928static const struct dax_operations dm_dax_ops;
1929
1930static void dm_wq_work(struct work_struct *work);
1931
1932#ifdef CONFIG_BLK_INLINE_ENCRYPTION
1933static void dm_queue_destroy_crypto_profile(struct request_queue *q)
1934{
1935 dm_destroy_crypto_profile(q->crypto_profile);
1936}
1937
1938#else /* CONFIG_BLK_INLINE_ENCRYPTION */
1939
1940static inline void dm_queue_destroy_crypto_profile(struct request_queue *q)
1941{
1942}
1943#endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
1944
1945static void cleanup_mapped_device(struct mapped_device *md)
1946{
1947 if (md->wq)
1948 destroy_workqueue(md->wq);
1949 dm_free_md_mempools(md->mempools);
1950
1951 if (md->dax_dev) {
1952 dax_remove_host(md->disk);
1953 kill_dax(md->dax_dev);
1954 put_dax(md->dax_dev);
1955 md->dax_dev = NULL;
1956 }
1957
1958 dm_cleanup_zoned_dev(md);
1959 if (md->disk) {
1960 spin_lock(&_minor_lock);
1961 md->disk->private_data = NULL;
1962 spin_unlock(&_minor_lock);
1963 if (dm_get_md_type(md) != DM_TYPE_NONE) {
1964 struct table_device *td;
1965
1966 dm_sysfs_exit(md);
1967 list_for_each_entry(td, &md->table_devices, list) {
1968 bd_unlink_disk_holder(td->dm_dev.bdev,
1969 md->disk);
1970 }
1971
1972 /*
1973 * Hold lock to make sure del_gendisk() won't concurrent
1974 * with open/close_table_device().
1975 */
1976 mutex_lock(&md->table_devices_lock);
1977 del_gendisk(md->disk);
1978 mutex_unlock(&md->table_devices_lock);
1979 }
1980 dm_queue_destroy_crypto_profile(md->queue);
1981 put_disk(md->disk);
1982 }
1983
1984 if (md->pending_io) {
1985 free_percpu(md->pending_io);
1986 md->pending_io = NULL;
1987 }
1988
1989 cleanup_srcu_struct(&md->io_barrier);
1990
1991 mutex_destroy(&md->suspend_lock);
1992 mutex_destroy(&md->type_lock);
1993 mutex_destroy(&md->table_devices_lock);
1994 mutex_destroy(&md->swap_bios_lock);
1995
1996 dm_mq_cleanup_mapped_device(md);
1997}
1998
1999/*
2000 * Allocate and initialise a blank device with a given minor.
2001 */
2002static struct mapped_device *alloc_dev(int minor)
2003{
2004 int r, numa_node_id = dm_get_numa_node();
2005 struct mapped_device *md;
2006 void *old_md;
2007
2008 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
2009 if (!md) {
2010 DMERR("unable to allocate device, out of memory.");
2011 return NULL;
2012 }
2013
2014 if (!try_module_get(THIS_MODULE))
2015 goto bad_module_get;
2016
2017 /* get a minor number for the dev */
2018 if (minor == DM_ANY_MINOR)
2019 r = next_free_minor(&minor);
2020 else
2021 r = specific_minor(minor);
2022 if (r < 0)
2023 goto bad_minor;
2024
2025 r = init_srcu_struct(&md->io_barrier);
2026 if (r < 0)
2027 goto bad_io_barrier;
2028
2029 md->numa_node_id = numa_node_id;
2030 md->init_tio_pdu = false;
2031 md->type = DM_TYPE_NONE;
2032 mutex_init(&md->suspend_lock);
2033 mutex_init(&md->type_lock);
2034 mutex_init(&md->table_devices_lock);
2035 spin_lock_init(&md->deferred_lock);
2036 atomic_set(&md->holders, 1);
2037 atomic_set(&md->open_count, 0);
2038 atomic_set(&md->event_nr, 0);
2039 atomic_set(&md->uevent_seq, 0);
2040 INIT_LIST_HEAD(&md->uevent_list);
2041 INIT_LIST_HEAD(&md->table_devices);
2042 spin_lock_init(&md->uevent_lock);
2043
2044 /*
2045 * default to bio-based until DM table is loaded and md->type
2046 * established. If request-based table is loaded: blk-mq will
2047 * override accordingly.
2048 */
2049 md->disk = blk_alloc_disk(md->numa_node_id);
2050 if (!md->disk)
2051 goto bad;
2052 md->queue = md->disk->queue;
2053
2054 init_waitqueue_head(&md->wait);
2055 INIT_WORK(&md->work, dm_wq_work);
2056 INIT_WORK(&md->requeue_work, dm_wq_requeue_work);
2057 init_waitqueue_head(&md->eventq);
2058 init_completion(&md->kobj_holder.completion);
2059
2060 md->requeue_list = NULL;
2061 md->swap_bios = get_swap_bios();
2062 sema_init(&md->swap_bios_semaphore, md->swap_bios);
2063 mutex_init(&md->swap_bios_lock);
2064
2065 md->disk->major = _major;
2066 md->disk->first_minor = minor;
2067 md->disk->minors = 1;
2068 md->disk->flags |= GENHD_FL_NO_PART;
2069 md->disk->fops = &dm_blk_dops;
2070 md->disk->private_data = md;
2071 sprintf(md->disk->disk_name, "dm-%d", minor);
2072
2073 if (IS_ENABLED(CONFIG_FS_DAX)) {
2074 md->dax_dev = alloc_dax(md, &dm_dax_ops);
2075 if (IS_ERR(md->dax_dev)) {
2076 md->dax_dev = NULL;
2077 goto bad;
2078 }
2079 set_dax_nocache(md->dax_dev);
2080 set_dax_nomc(md->dax_dev);
2081 if (dax_add_host(md->dax_dev, md->disk))
2082 goto bad;
2083 }
2084
2085 format_dev_t(md->name, MKDEV(_major, minor));
2086
2087 md->wq = alloc_workqueue("kdmflush/%s", WQ_MEM_RECLAIM, 0, md->name);
2088 if (!md->wq)
2089 goto bad;
2090
2091 md->pending_io = alloc_percpu(unsigned long);
2092 if (!md->pending_io)
2093 goto bad;
2094
2095 dm_stats_init(&md->stats);
2096
2097 /* Populate the mapping, nobody knows we exist yet */
2098 spin_lock(&_minor_lock);
2099 old_md = idr_replace(&_minor_idr, md, minor);
2100 spin_unlock(&_minor_lock);
2101
2102 BUG_ON(old_md != MINOR_ALLOCED);
2103
2104 return md;
2105
2106bad:
2107 cleanup_mapped_device(md);
2108bad_io_barrier:
2109 free_minor(minor);
2110bad_minor:
2111 module_put(THIS_MODULE);
2112bad_module_get:
2113 kvfree(md);
2114 return NULL;
2115}
2116
2117static void unlock_fs(struct mapped_device *md);
2118
2119static void free_dev(struct mapped_device *md)
2120{
2121 int minor = MINOR(disk_devt(md->disk));
2122
2123 unlock_fs(md);
2124
2125 cleanup_mapped_device(md);
2126
2127 WARN_ON_ONCE(!list_empty(&md->table_devices));
2128 dm_stats_cleanup(&md->stats);
2129 free_minor(minor);
2130
2131 module_put(THIS_MODULE);
2132 kvfree(md);
2133}
2134
2135/*
2136 * Bind a table to the device.
2137 */
2138static void event_callback(void *context)
2139{
2140 unsigned long flags;
2141 LIST_HEAD(uevents);
2142 struct mapped_device *md = (struct mapped_device *) context;
2143
2144 spin_lock_irqsave(&md->uevent_lock, flags);
2145 list_splice_init(&md->uevent_list, &uevents);
2146 spin_unlock_irqrestore(&md->uevent_lock, flags);
2147
2148 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2149
2150 atomic_inc(&md->event_nr);
2151 wake_up(&md->eventq);
2152 dm_issue_global_event();
2153}
2154
2155/*
2156 * Returns old map, which caller must destroy.
2157 */
2158static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2159 struct queue_limits *limits)
2160{
2161 struct dm_table *old_map;
2162 sector_t size;
2163 int ret;
2164
2165 lockdep_assert_held(&md->suspend_lock);
2166
2167 size = dm_table_get_size(t);
2168
2169 /*
2170 * Wipe any geometry if the size of the table changed.
2171 */
2172 if (size != dm_get_size(md))
2173 memset(&md->geometry, 0, sizeof(md->geometry));
2174
2175 if (!get_capacity(md->disk))
2176 set_capacity(md->disk, size);
2177 else
2178 set_capacity_and_notify(md->disk, size);
2179
2180 dm_table_event_callback(t, event_callback, md);
2181
2182 if (dm_table_request_based(t)) {
2183 /*
2184 * Leverage the fact that request-based DM targets are
2185 * immutable singletons - used to optimize dm_mq_queue_rq.
2186 */
2187 md->immutable_target = dm_table_get_immutable_target(t);
2188
2189 /*
2190 * There is no need to reload with request-based dm because the
2191 * size of front_pad doesn't change.
2192 *
2193 * Note for future: If you are to reload bioset, prep-ed
2194 * requests in the queue may refer to bio from the old bioset,
2195 * so you must walk through the queue to unprep.
2196 */
2197 if (!md->mempools) {
2198 md->mempools = t->mempools;
2199 t->mempools = NULL;
2200 }
2201 } else {
2202 /*
2203 * The md may already have mempools that need changing.
2204 * If so, reload bioset because front_pad may have changed
2205 * because a different table was loaded.
2206 */
2207 dm_free_md_mempools(md->mempools);
2208 md->mempools = t->mempools;
2209 t->mempools = NULL;
2210 }
2211
2212 ret = dm_table_set_restrictions(t, md->queue, limits);
2213 if (ret) {
2214 old_map = ERR_PTR(ret);
2215 goto out;
2216 }
2217
2218 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2219 rcu_assign_pointer(md->map, (void *)t);
2220 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2221
2222 if (old_map)
2223 dm_sync_table(md);
2224out:
2225 return old_map;
2226}
2227
2228/*
2229 * Returns unbound table for the caller to free.
2230 */
2231static struct dm_table *__unbind(struct mapped_device *md)
2232{
2233 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2234
2235 if (!map)
2236 return NULL;
2237
2238 dm_table_event_callback(map, NULL, NULL);
2239 RCU_INIT_POINTER(md->map, NULL);
2240 dm_sync_table(md);
2241
2242 return map;
2243}
2244
2245/*
2246 * Constructor for a new device.
2247 */
2248int dm_create(int minor, struct mapped_device **result)
2249{
2250 struct mapped_device *md;
2251
2252 md = alloc_dev(minor);
2253 if (!md)
2254 return -ENXIO;
2255
2256 dm_ima_reset_data(md);
2257
2258 *result = md;
2259 return 0;
2260}
2261
2262/*
2263 * Functions to manage md->type.
2264 * All are required to hold md->type_lock.
2265 */
2266void dm_lock_md_type(struct mapped_device *md)
2267{
2268 mutex_lock(&md->type_lock);
2269}
2270
2271void dm_unlock_md_type(struct mapped_device *md)
2272{
2273 mutex_unlock(&md->type_lock);
2274}
2275
2276void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2277{
2278 BUG_ON(!mutex_is_locked(&md->type_lock));
2279 md->type = type;
2280}
2281
2282enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2283{
2284 return md->type;
2285}
2286
2287struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2288{
2289 return md->immutable_target_type;
2290}
2291
2292/*
2293 * The queue_limits are only valid as long as you have a reference
2294 * count on 'md'.
2295 */
2296struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2297{
2298 BUG_ON(!atomic_read(&md->holders));
2299 return &md->queue->limits;
2300}
2301EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2302
2303/*
2304 * Setup the DM device's queue based on md's type
2305 */
2306int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2307{
2308 enum dm_queue_mode type = dm_table_get_type(t);
2309 struct queue_limits limits;
2310 struct table_device *td;
2311 int r;
2312
2313 switch (type) {
2314 case DM_TYPE_REQUEST_BASED:
2315 md->disk->fops = &dm_rq_blk_dops;
2316 r = dm_mq_init_request_queue(md, t);
2317 if (r) {
2318 DMERR("Cannot initialize queue for request-based dm mapped device");
2319 return r;
2320 }
2321 break;
2322 case DM_TYPE_BIO_BASED:
2323 case DM_TYPE_DAX_BIO_BASED:
2324 break;
2325 case DM_TYPE_NONE:
2326 WARN_ON_ONCE(true);
2327 break;
2328 }
2329
2330 r = dm_calculate_queue_limits(t, &limits);
2331 if (r) {
2332 DMERR("Cannot calculate initial queue limits");
2333 return r;
2334 }
2335 r = dm_table_set_restrictions(t, md->queue, &limits);
2336 if (r)
2337 return r;
2338
2339 /*
2340 * Hold lock to make sure add_disk() and del_gendisk() won't concurrent
2341 * with open_table_device() and close_table_device().
2342 */
2343 mutex_lock(&md->table_devices_lock);
2344 r = add_disk(md->disk);
2345 mutex_unlock(&md->table_devices_lock);
2346 if (r)
2347 return r;
2348
2349 /*
2350 * Register the holder relationship for devices added before the disk
2351 * was live.
2352 */
2353 list_for_each_entry(td, &md->table_devices, list) {
2354 r = bd_link_disk_holder(td->dm_dev.bdev, md->disk);
2355 if (r)
2356 goto out_undo_holders;
2357 }
2358
2359 r = dm_sysfs_init(md);
2360 if (r)
2361 goto out_undo_holders;
2362
2363 md->type = type;
2364 return 0;
2365
2366out_undo_holders:
2367 list_for_each_entry_continue_reverse(td, &md->table_devices, list)
2368 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
2369 mutex_lock(&md->table_devices_lock);
2370 del_gendisk(md->disk);
2371 mutex_unlock(&md->table_devices_lock);
2372 return r;
2373}
2374
2375struct mapped_device *dm_get_md(dev_t dev)
2376{
2377 struct mapped_device *md;
2378 unsigned minor = MINOR(dev);
2379
2380 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2381 return NULL;
2382
2383 spin_lock(&_minor_lock);
2384
2385 md = idr_find(&_minor_idr, minor);
2386 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2387 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2388 md = NULL;
2389 goto out;
2390 }
2391 dm_get(md);
2392out:
2393 spin_unlock(&_minor_lock);
2394
2395 return md;
2396}
2397EXPORT_SYMBOL_GPL(dm_get_md);
2398
2399void *dm_get_mdptr(struct mapped_device *md)
2400{
2401 return md->interface_ptr;
2402}
2403
2404void dm_set_mdptr(struct mapped_device *md, void *ptr)
2405{
2406 md->interface_ptr = ptr;
2407}
2408
2409void dm_get(struct mapped_device *md)
2410{
2411 atomic_inc(&md->holders);
2412 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2413}
2414
2415int dm_hold(struct mapped_device *md)
2416{
2417 spin_lock(&_minor_lock);
2418 if (test_bit(DMF_FREEING, &md->flags)) {
2419 spin_unlock(&_minor_lock);
2420 return -EBUSY;
2421 }
2422 dm_get(md);
2423 spin_unlock(&_minor_lock);
2424 return 0;
2425}
2426EXPORT_SYMBOL_GPL(dm_hold);
2427
2428const char *dm_device_name(struct mapped_device *md)
2429{
2430 return md->name;
2431}
2432EXPORT_SYMBOL_GPL(dm_device_name);
2433
2434static void __dm_destroy(struct mapped_device *md, bool wait)
2435{
2436 struct dm_table *map;
2437 int srcu_idx;
2438
2439 might_sleep();
2440
2441 spin_lock(&_minor_lock);
2442 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2443 set_bit(DMF_FREEING, &md->flags);
2444 spin_unlock(&_minor_lock);
2445
2446 blk_mark_disk_dead(md->disk);
2447
2448 /*
2449 * Take suspend_lock so that presuspend and postsuspend methods
2450 * do not race with internal suspend.
2451 */
2452 mutex_lock(&md->suspend_lock);
2453 map = dm_get_live_table(md, &srcu_idx);
2454 if (!dm_suspended_md(md)) {
2455 dm_table_presuspend_targets(map);
2456 set_bit(DMF_SUSPENDED, &md->flags);
2457 set_bit(DMF_POST_SUSPENDING, &md->flags);
2458 dm_table_postsuspend_targets(map);
2459 }
2460 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2461 dm_put_live_table(md, srcu_idx);
2462 mutex_unlock(&md->suspend_lock);
2463
2464 /*
2465 * Rare, but there may be I/O requests still going to complete,
2466 * for example. Wait for all references to disappear.
2467 * No one should increment the reference count of the mapped_device,
2468 * after the mapped_device state becomes DMF_FREEING.
2469 */
2470 if (wait)
2471 while (atomic_read(&md->holders))
2472 msleep(1);
2473 else if (atomic_read(&md->holders))
2474 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2475 dm_device_name(md), atomic_read(&md->holders));
2476
2477 dm_table_destroy(__unbind(md));
2478 free_dev(md);
2479}
2480
2481void dm_destroy(struct mapped_device *md)
2482{
2483 __dm_destroy(md, true);
2484}
2485
2486void dm_destroy_immediate(struct mapped_device *md)
2487{
2488 __dm_destroy(md, false);
2489}
2490
2491void dm_put(struct mapped_device *md)
2492{
2493 atomic_dec(&md->holders);
2494}
2495EXPORT_SYMBOL_GPL(dm_put);
2496
2497static bool dm_in_flight_bios(struct mapped_device *md)
2498{
2499 int cpu;
2500 unsigned long sum = 0;
2501
2502 for_each_possible_cpu(cpu)
2503 sum += *per_cpu_ptr(md->pending_io, cpu);
2504
2505 return sum != 0;
2506}
2507
2508static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2509{
2510 int r = 0;
2511 DEFINE_WAIT(wait);
2512
2513 while (true) {
2514 prepare_to_wait(&md->wait, &wait, task_state);
2515
2516 if (!dm_in_flight_bios(md))
2517 break;
2518
2519 if (signal_pending_state(task_state, current)) {
2520 r = -EINTR;
2521 break;
2522 }
2523
2524 io_schedule();
2525 }
2526 finish_wait(&md->wait, &wait);
2527
2528 smp_rmb();
2529
2530 return r;
2531}
2532
2533static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2534{
2535 int r = 0;
2536
2537 if (!queue_is_mq(md->queue))
2538 return dm_wait_for_bios_completion(md, task_state);
2539
2540 while (true) {
2541 if (!blk_mq_queue_inflight(md->queue))
2542 break;
2543
2544 if (signal_pending_state(task_state, current)) {
2545 r = -EINTR;
2546 break;
2547 }
2548
2549 msleep(5);
2550 }
2551
2552 return r;
2553}
2554
2555/*
2556 * Process the deferred bios
2557 */
2558static void dm_wq_work(struct work_struct *work)
2559{
2560 struct mapped_device *md = container_of(work, struct mapped_device, work);
2561 struct bio *bio;
2562
2563 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2564 spin_lock_irq(&md->deferred_lock);
2565 bio = bio_list_pop(&md->deferred);
2566 spin_unlock_irq(&md->deferred_lock);
2567
2568 if (!bio)
2569 break;
2570
2571 submit_bio_noacct(bio);
2572 }
2573}
2574
2575static void dm_queue_flush(struct mapped_device *md)
2576{
2577 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2578 smp_mb__after_atomic();
2579 queue_work(md->wq, &md->work);
2580}
2581
2582/*
2583 * Swap in a new table, returning the old one for the caller to destroy.
2584 */
2585struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2586{
2587 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2588 struct queue_limits limits;
2589 int r;
2590
2591 mutex_lock(&md->suspend_lock);
2592
2593 /* device must be suspended */
2594 if (!dm_suspended_md(md))
2595 goto out;
2596
2597 /*
2598 * If the new table has no data devices, retain the existing limits.
2599 * This helps multipath with queue_if_no_path if all paths disappear,
2600 * then new I/O is queued based on these limits, and then some paths
2601 * reappear.
2602 */
2603 if (dm_table_has_no_data_devices(table)) {
2604 live_map = dm_get_live_table_fast(md);
2605 if (live_map)
2606 limits = md->queue->limits;
2607 dm_put_live_table_fast(md);
2608 }
2609
2610 if (!live_map) {
2611 r = dm_calculate_queue_limits(table, &limits);
2612 if (r) {
2613 map = ERR_PTR(r);
2614 goto out;
2615 }
2616 }
2617
2618 map = __bind(md, table, &limits);
2619 dm_issue_global_event();
2620
2621out:
2622 mutex_unlock(&md->suspend_lock);
2623 return map;
2624}
2625
2626/*
2627 * Functions to lock and unlock any filesystem running on the
2628 * device.
2629 */
2630static int lock_fs(struct mapped_device *md)
2631{
2632 int r;
2633
2634 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2635
2636 r = freeze_bdev(md->disk->part0);
2637 if (!r)
2638 set_bit(DMF_FROZEN, &md->flags);
2639 return r;
2640}
2641
2642static void unlock_fs(struct mapped_device *md)
2643{
2644 if (!test_bit(DMF_FROZEN, &md->flags))
2645 return;
2646 thaw_bdev(md->disk->part0);
2647 clear_bit(DMF_FROZEN, &md->flags);
2648}
2649
2650/*
2651 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2652 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2653 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2654 *
2655 * If __dm_suspend returns 0, the device is completely quiescent
2656 * now. There is no request-processing activity. All new requests
2657 * are being added to md->deferred list.
2658 */
2659static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2660 unsigned suspend_flags, unsigned int task_state,
2661 int dmf_suspended_flag)
2662{
2663 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2664 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2665 int r;
2666
2667 lockdep_assert_held(&md->suspend_lock);
2668
2669 /*
2670 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2671 * This flag is cleared before dm_suspend returns.
2672 */
2673 if (noflush)
2674 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2675 else
2676 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2677
2678 /*
2679 * This gets reverted if there's an error later and the targets
2680 * provide the .presuspend_undo hook.
2681 */
2682 dm_table_presuspend_targets(map);
2683
2684 /*
2685 * Flush I/O to the device.
2686 * Any I/O submitted after lock_fs() may not be flushed.
2687 * noflush takes precedence over do_lockfs.
2688 * (lock_fs() flushes I/Os and waits for them to complete.)
2689 */
2690 if (!noflush && do_lockfs) {
2691 r = lock_fs(md);
2692 if (r) {
2693 dm_table_presuspend_undo_targets(map);
2694 return r;
2695 }
2696 }
2697
2698 /*
2699 * Here we must make sure that no processes are submitting requests
2700 * to target drivers i.e. no one may be executing
2701 * dm_split_and_process_bio from dm_submit_bio.
2702 *
2703 * To get all processes out of dm_split_and_process_bio in dm_submit_bio,
2704 * we take the write lock. To prevent any process from reentering
2705 * dm_split_and_process_bio from dm_submit_bio and quiesce the thread
2706 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2707 * flush_workqueue(md->wq).
2708 */
2709 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2710 if (map)
2711 synchronize_srcu(&md->io_barrier);
2712
2713 /*
2714 * Stop md->queue before flushing md->wq in case request-based
2715 * dm defers requests to md->wq from md->queue.
2716 */
2717 if (dm_request_based(md))
2718 dm_stop_queue(md->queue);
2719
2720 flush_workqueue(md->wq);
2721
2722 /*
2723 * At this point no more requests are entering target request routines.
2724 * We call dm_wait_for_completion to wait for all existing requests
2725 * to finish.
2726 */
2727 r = dm_wait_for_completion(md, task_state);
2728 if (!r)
2729 set_bit(dmf_suspended_flag, &md->flags);
2730
2731 if (noflush)
2732 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2733 if (map)
2734 synchronize_srcu(&md->io_barrier);
2735
2736 /* were we interrupted ? */
2737 if (r < 0) {
2738 dm_queue_flush(md);
2739
2740 if (dm_request_based(md))
2741 dm_start_queue(md->queue);
2742
2743 unlock_fs(md);
2744 dm_table_presuspend_undo_targets(map);
2745 /* pushback list is already flushed, so skip flush */
2746 }
2747
2748 return r;
2749}
2750
2751/*
2752 * We need to be able to change a mapping table under a mounted
2753 * filesystem. For example we might want to move some data in
2754 * the background. Before the table can be swapped with
2755 * dm_bind_table, dm_suspend must be called to flush any in
2756 * flight bios and ensure that any further io gets deferred.
2757 */
2758/*
2759 * Suspend mechanism in request-based dm.
2760 *
2761 * 1. Flush all I/Os by lock_fs() if needed.
2762 * 2. Stop dispatching any I/O by stopping the request_queue.
2763 * 3. Wait for all in-flight I/Os to be completed or requeued.
2764 *
2765 * To abort suspend, start the request_queue.
2766 */
2767int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
2768{
2769 struct dm_table *map = NULL;
2770 int r = 0;
2771
2772retry:
2773 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2774
2775 if (dm_suspended_md(md)) {
2776 r = -EINVAL;
2777 goto out_unlock;
2778 }
2779
2780 if (dm_suspended_internally_md(md)) {
2781 /* already internally suspended, wait for internal resume */
2782 mutex_unlock(&md->suspend_lock);
2783 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2784 if (r)
2785 return r;
2786 goto retry;
2787 }
2788
2789 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2790
2791 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2792 if (r)
2793 goto out_unlock;
2794
2795 set_bit(DMF_POST_SUSPENDING, &md->flags);
2796 dm_table_postsuspend_targets(map);
2797 clear_bit(DMF_POST_SUSPENDING, &md->flags);
2798
2799out_unlock:
2800 mutex_unlock(&md->suspend_lock);
2801 return r;
2802}
2803
2804static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2805{
2806 if (map) {
2807 int r = dm_table_resume_targets(map);
2808 if (r)
2809 return r;
2810 }
2811
2812 dm_queue_flush(md);
2813
2814 /*
2815 * Flushing deferred I/Os must be done after targets are resumed
2816 * so that mapping of targets can work correctly.
2817 * Request-based dm is queueing the deferred I/Os in its request_queue.
2818 */
2819 if (dm_request_based(md))
2820 dm_start_queue(md->queue);
2821
2822 unlock_fs(md);
2823
2824 return 0;
2825}
2826
2827int dm_resume(struct mapped_device *md)
2828{
2829 int r;
2830 struct dm_table *map = NULL;
2831
2832retry:
2833 r = -EINVAL;
2834 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2835
2836 if (!dm_suspended_md(md))
2837 goto out;
2838
2839 if (dm_suspended_internally_md(md)) {
2840 /* already internally suspended, wait for internal resume */
2841 mutex_unlock(&md->suspend_lock);
2842 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2843 if (r)
2844 return r;
2845 goto retry;
2846 }
2847
2848 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2849 if (!map || !dm_table_get_size(map))
2850 goto out;
2851
2852 r = __dm_resume(md, map);
2853 if (r)
2854 goto out;
2855
2856 clear_bit(DMF_SUSPENDED, &md->flags);
2857out:
2858 mutex_unlock(&md->suspend_lock);
2859
2860 return r;
2861}
2862
2863/*
2864 * Internal suspend/resume works like userspace-driven suspend. It waits
2865 * until all bios finish and prevents issuing new bios to the target drivers.
2866 * It may be used only from the kernel.
2867 */
2868
2869static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
2870{
2871 struct dm_table *map = NULL;
2872
2873 lockdep_assert_held(&md->suspend_lock);
2874
2875 if (md->internal_suspend_count++)
2876 return; /* nested internal suspend */
2877
2878 if (dm_suspended_md(md)) {
2879 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2880 return; /* nest suspend */
2881 }
2882
2883 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2884
2885 /*
2886 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2887 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2888 * would require changing .presuspend to return an error -- avoid this
2889 * until there is a need for more elaborate variants of internal suspend.
2890 */
2891 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2892 DMF_SUSPENDED_INTERNALLY);
2893
2894 set_bit(DMF_POST_SUSPENDING, &md->flags);
2895 dm_table_postsuspend_targets(map);
2896 clear_bit(DMF_POST_SUSPENDING, &md->flags);
2897}
2898
2899static void __dm_internal_resume(struct mapped_device *md)
2900{
2901 BUG_ON(!md->internal_suspend_count);
2902
2903 if (--md->internal_suspend_count)
2904 return; /* resume from nested internal suspend */
2905
2906 if (dm_suspended_md(md))
2907 goto done; /* resume from nested suspend */
2908
2909 /*
2910 * NOTE: existing callers don't need to call dm_table_resume_targets
2911 * (which may fail -- so best to avoid it for now by passing NULL map)
2912 */
2913 (void) __dm_resume(md, NULL);
2914
2915done:
2916 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2917 smp_mb__after_atomic();
2918 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
2919}
2920
2921void dm_internal_suspend_noflush(struct mapped_device *md)
2922{
2923 mutex_lock(&md->suspend_lock);
2924 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2925 mutex_unlock(&md->suspend_lock);
2926}
2927EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2928
2929void dm_internal_resume(struct mapped_device *md)
2930{
2931 mutex_lock(&md->suspend_lock);
2932 __dm_internal_resume(md);
2933 mutex_unlock(&md->suspend_lock);
2934}
2935EXPORT_SYMBOL_GPL(dm_internal_resume);
2936
2937/*
2938 * Fast variants of internal suspend/resume hold md->suspend_lock,
2939 * which prevents interaction with userspace-driven suspend.
2940 */
2941
2942void dm_internal_suspend_fast(struct mapped_device *md)
2943{
2944 mutex_lock(&md->suspend_lock);
2945 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2946 return;
2947
2948 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2949 synchronize_srcu(&md->io_barrier);
2950 flush_workqueue(md->wq);
2951 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
2952}
2953EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
2954
2955void dm_internal_resume_fast(struct mapped_device *md)
2956{
2957 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2958 goto done;
2959
2960 dm_queue_flush(md);
2961
2962done:
2963 mutex_unlock(&md->suspend_lock);
2964}
2965EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
2966
2967/*-----------------------------------------------------------------
2968 * Event notification.
2969 *---------------------------------------------------------------*/
2970int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
2971 unsigned cookie)
2972{
2973 int r;
2974 unsigned noio_flag;
2975 char udev_cookie[DM_COOKIE_LENGTH];
2976 char *envp[] = { udev_cookie, NULL };
2977
2978 noio_flag = memalloc_noio_save();
2979
2980 if (!cookie)
2981 r = kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
2982 else {
2983 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
2984 DM_COOKIE_ENV_VAR_NAME, cookie);
2985 r = kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
2986 action, envp);
2987 }
2988
2989 memalloc_noio_restore(noio_flag);
2990
2991 return r;
2992}
2993
2994uint32_t dm_next_uevent_seq(struct mapped_device *md)
2995{
2996 return atomic_add_return(1, &md->uevent_seq);
2997}
2998
2999uint32_t dm_get_event_nr(struct mapped_device *md)
3000{
3001 return atomic_read(&md->event_nr);
3002}
3003
3004int dm_wait_event(struct mapped_device *md, int event_nr)
3005{
3006 return wait_event_interruptible(md->eventq,
3007 (event_nr != atomic_read(&md->event_nr)));
3008}
3009
3010void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3011{
3012 unsigned long flags;
3013
3014 spin_lock_irqsave(&md->uevent_lock, flags);
3015 list_add(elist, &md->uevent_list);
3016 spin_unlock_irqrestore(&md->uevent_lock, flags);
3017}
3018
3019/*
3020 * The gendisk is only valid as long as you have a reference
3021 * count on 'md'.
3022 */
3023struct gendisk *dm_disk(struct mapped_device *md)
3024{
3025 return md->disk;
3026}
3027EXPORT_SYMBOL_GPL(dm_disk);
3028
3029struct kobject *dm_kobject(struct mapped_device *md)
3030{
3031 return &md->kobj_holder.kobj;
3032}
3033
3034struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3035{
3036 struct mapped_device *md;
3037
3038 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3039
3040 spin_lock(&_minor_lock);
3041 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
3042 md = NULL;
3043 goto out;
3044 }
3045 dm_get(md);
3046out:
3047 spin_unlock(&_minor_lock);
3048
3049 return md;
3050}
3051
3052int dm_suspended_md(struct mapped_device *md)
3053{
3054 return test_bit(DMF_SUSPENDED, &md->flags);
3055}
3056
3057static int dm_post_suspending_md(struct mapped_device *md)
3058{
3059 return test_bit(DMF_POST_SUSPENDING, &md->flags);
3060}
3061
3062int dm_suspended_internally_md(struct mapped_device *md)
3063{
3064 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3065}
3066
3067int dm_test_deferred_remove_flag(struct mapped_device *md)
3068{
3069 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3070}
3071
3072int dm_suspended(struct dm_target *ti)
3073{
3074 return dm_suspended_md(ti->table->md);
3075}
3076EXPORT_SYMBOL_GPL(dm_suspended);
3077
3078int dm_post_suspending(struct dm_target *ti)
3079{
3080 return dm_post_suspending_md(ti->table->md);
3081}
3082EXPORT_SYMBOL_GPL(dm_post_suspending);
3083
3084int dm_noflush_suspending(struct dm_target *ti)
3085{
3086 return __noflush_suspending(ti->table->md);
3087}
3088EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3089
3090void dm_free_md_mempools(struct dm_md_mempools *pools)
3091{
3092 if (!pools)
3093 return;
3094
3095 bioset_exit(&pools->bs);
3096 bioset_exit(&pools->io_bs);
3097
3098 kfree(pools);
3099}
3100
3101struct dm_pr {
3102 u64 old_key;
3103 u64 new_key;
3104 u32 flags;
3105 bool abort;
3106 bool fail_early;
3107 int ret;
3108 enum pr_type type;
3109};
3110
3111static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3112 struct dm_pr *pr)
3113{
3114 struct mapped_device *md = bdev->bd_disk->private_data;
3115 struct dm_table *table;
3116 struct dm_target *ti;
3117 int ret = -ENOTTY, srcu_idx;
3118
3119 table = dm_get_live_table(md, &srcu_idx);
3120 if (!table || !dm_table_get_size(table))
3121 goto out;
3122
3123 /* We only support devices that have a single target */
3124 if (table->num_targets != 1)
3125 goto out;
3126 ti = dm_table_get_target(table, 0);
3127
3128 if (dm_suspended_md(md)) {
3129 ret = -EAGAIN;
3130 goto out;
3131 }
3132
3133 ret = -EINVAL;
3134 if (!ti->type->iterate_devices)
3135 goto out;
3136
3137 ti->type->iterate_devices(ti, fn, pr);
3138 ret = 0;
3139out:
3140 dm_put_live_table(md, srcu_idx);
3141 return ret;
3142}
3143
3144/*
3145 * For register / unregister we need to manually call out to every path.
3146 */
3147static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3148 sector_t start, sector_t len, void *data)
3149{
3150 struct dm_pr *pr = data;
3151 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3152 int ret;
3153
3154 if (!ops || !ops->pr_register) {
3155 pr->ret = -EOPNOTSUPP;
3156 return -1;
3157 }
3158
3159 ret = ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3160 if (!ret)
3161 return 0;
3162
3163 if (!pr->ret)
3164 pr->ret = ret;
3165
3166 if (pr->fail_early)
3167 return -1;
3168
3169 return 0;
3170}
3171
3172static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3173 u32 flags)
3174{
3175 struct dm_pr pr = {
3176 .old_key = old_key,
3177 .new_key = new_key,
3178 .flags = flags,
3179 .fail_early = true,
3180 .ret = 0,
3181 };
3182 int ret;
3183
3184 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3185 if (ret) {
3186 /* Didn't even get to register a path */
3187 return ret;
3188 }
3189
3190 if (!pr.ret)
3191 return 0;
3192 ret = pr.ret;
3193
3194 if (!new_key)
3195 return ret;
3196
3197 /* unregister all paths if we failed to register any path */
3198 pr.old_key = new_key;
3199 pr.new_key = 0;
3200 pr.flags = 0;
3201 pr.fail_early = false;
3202 (void) dm_call_pr(bdev, __dm_pr_register, &pr);
3203 return ret;
3204}
3205
3206
3207static int __dm_pr_reserve(struct dm_target *ti, struct dm_dev *dev,
3208 sector_t start, sector_t len, void *data)
3209{
3210 struct dm_pr *pr = data;
3211 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3212
3213 if (!ops || !ops->pr_reserve) {
3214 pr->ret = -EOPNOTSUPP;
3215 return -1;
3216 }
3217
3218 pr->ret = ops->pr_reserve(dev->bdev, pr->old_key, pr->type, pr->flags);
3219 if (!pr->ret)
3220 return -1;
3221
3222 return 0;
3223}
3224
3225static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3226 u32 flags)
3227{
3228 struct dm_pr pr = {
3229 .old_key = key,
3230 .flags = flags,
3231 .type = type,
3232 .fail_early = false,
3233 .ret = 0,
3234 };
3235 int ret;
3236
3237 ret = dm_call_pr(bdev, __dm_pr_reserve, &pr);
3238 if (ret)
3239 return ret;
3240
3241 return pr.ret;
3242}
3243
3244/*
3245 * If there is a non-All Registrants type of reservation, the release must be
3246 * sent down the holding path. For the cases where there is no reservation or
3247 * the path is not the holder the device will also return success, so we must
3248 * try each path to make sure we got the correct path.
3249 */
3250static int __dm_pr_release(struct dm_target *ti, struct dm_dev *dev,
3251 sector_t start, sector_t len, void *data)
3252{
3253 struct dm_pr *pr = data;
3254 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3255
3256 if (!ops || !ops->pr_release) {
3257 pr->ret = -EOPNOTSUPP;
3258 return -1;
3259 }
3260
3261 pr->ret = ops->pr_release(dev->bdev, pr->old_key, pr->type);
3262 if (pr->ret)
3263 return -1;
3264
3265 return 0;
3266}
3267
3268static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3269{
3270 struct dm_pr pr = {
3271 .old_key = key,
3272 .type = type,
3273 .fail_early = false,
3274 };
3275 int ret;
3276
3277 ret = dm_call_pr(bdev, __dm_pr_release, &pr);
3278 if (ret)
3279 return ret;
3280
3281 return pr.ret;
3282}
3283
3284static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
3285 sector_t start, sector_t len, void *data)
3286{
3287 struct dm_pr *pr = data;
3288 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3289
3290 if (!ops || !ops->pr_preempt) {
3291 pr->ret = -EOPNOTSUPP;
3292 return -1;
3293 }
3294
3295 pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
3296 pr->abort);
3297 if (!pr->ret)
3298 return -1;
3299
3300 return 0;
3301}
3302
3303static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3304 enum pr_type type, bool abort)
3305{
3306 struct dm_pr pr = {
3307 .new_key = new_key,
3308 .old_key = old_key,
3309 .type = type,
3310 .fail_early = false,
3311 };
3312 int ret;
3313
3314 ret = dm_call_pr(bdev, __dm_pr_preempt, &pr);
3315 if (ret)
3316 return ret;
3317
3318 return pr.ret;
3319}
3320
3321static int dm_pr_clear(struct block_device *bdev, u64 key)
3322{
3323 struct mapped_device *md = bdev->bd_disk->private_data;
3324 const struct pr_ops *ops;
3325 int r, srcu_idx;
3326
3327 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3328 if (r < 0)
3329 goto out;
3330
3331 ops = bdev->bd_disk->fops->pr_ops;
3332 if (ops && ops->pr_clear)
3333 r = ops->pr_clear(bdev, key);
3334 else
3335 r = -EOPNOTSUPP;
3336out:
3337 dm_unprepare_ioctl(md, srcu_idx);
3338 return r;
3339}
3340
3341static const struct pr_ops dm_pr_ops = {
3342 .pr_register = dm_pr_register,
3343 .pr_reserve = dm_pr_reserve,
3344 .pr_release = dm_pr_release,
3345 .pr_preempt = dm_pr_preempt,
3346 .pr_clear = dm_pr_clear,
3347};
3348
3349static const struct block_device_operations dm_blk_dops = {
3350 .submit_bio = dm_submit_bio,
3351 .poll_bio = dm_poll_bio,
3352 .open = dm_blk_open,
3353 .release = dm_blk_close,
3354 .ioctl = dm_blk_ioctl,
3355 .getgeo = dm_blk_getgeo,
3356 .report_zones = dm_blk_report_zones,
3357 .pr_ops = &dm_pr_ops,
3358 .owner = THIS_MODULE
3359};
3360
3361static const struct block_device_operations dm_rq_blk_dops = {
3362 .open = dm_blk_open,
3363 .release = dm_blk_close,
3364 .ioctl = dm_blk_ioctl,
3365 .getgeo = dm_blk_getgeo,
3366 .pr_ops = &dm_pr_ops,
3367 .owner = THIS_MODULE
3368};
3369
3370static const struct dax_operations dm_dax_ops = {
3371 .direct_access = dm_dax_direct_access,
3372 .zero_page_range = dm_dax_zero_page_range,
3373 .recovery_write = dm_dax_recovery_write,
3374};
3375
3376/*
3377 * module hooks
3378 */
3379module_init(dm_init);
3380module_exit(dm_exit);
3381
3382module_param(major, uint, 0);
3383MODULE_PARM_DESC(major, "The major number of the device mapper");
3384
3385module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3386MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3387
3388module_param(dm_numa_node, int, S_IRUGO | S_IWUSR);
3389MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3390
3391module_param(swap_bios, int, S_IRUGO | S_IWUSR);
3392MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3393
3394MODULE_DESCRIPTION(DM_NAME " driver");
3395MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3396MODULE_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.h"
9#include "dm-uevent.h"
10
11#include <linux/init.h>
12#include <linux/module.h>
13#include <linux/mutex.h>
14#include <linux/moduleparam.h>
15#include <linux/blkpg.h>
16#include <linux/bio.h>
17#include <linux/mempool.h>
18#include <linux/slab.h>
19#include <linux/idr.h>
20#include <linux/hdreg.h>
21#include <linux/delay.h>
22#include <linux/wait.h>
23#include <linux/kthread.h>
24#include <linux/ktime.h>
25#include <linux/elevator.h> /* for rq_end_sector() */
26#include <linux/blk-mq.h>
27#include <linux/pr.h>
28
29#include <trace/events/block.h>
30
31#define DM_MSG_PREFIX "core"
32
33#ifdef CONFIG_PRINTK
34/*
35 * ratelimit state to be used in DMXXX_LIMIT().
36 */
37DEFINE_RATELIMIT_STATE(dm_ratelimit_state,
38 DEFAULT_RATELIMIT_INTERVAL,
39 DEFAULT_RATELIMIT_BURST);
40EXPORT_SYMBOL(dm_ratelimit_state);
41#endif
42
43/*
44 * Cookies are numeric values sent with CHANGE and REMOVE
45 * uevents while resuming, removing or renaming the device.
46 */
47#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
48#define DM_COOKIE_LENGTH 24
49
50static const char *_name = DM_NAME;
51
52static unsigned int major = 0;
53static unsigned int _major = 0;
54
55static DEFINE_IDR(_minor_idr);
56
57static DEFINE_SPINLOCK(_minor_lock);
58
59static void do_deferred_remove(struct work_struct *w);
60
61static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
62
63static struct workqueue_struct *deferred_remove_workqueue;
64
65/*
66 * For bio-based dm.
67 * One of these is allocated per bio.
68 */
69struct dm_io {
70 struct mapped_device *md;
71 int error;
72 atomic_t io_count;
73 struct bio *bio;
74 unsigned long start_time;
75 spinlock_t endio_lock;
76 struct dm_stats_aux stats_aux;
77};
78
79/*
80 * For request-based dm.
81 * One of these is allocated per request.
82 */
83struct dm_rq_target_io {
84 struct mapped_device *md;
85 struct dm_target *ti;
86 struct request *orig, *clone;
87 struct kthread_work work;
88 int error;
89 union map_info info;
90 struct dm_stats_aux stats_aux;
91 unsigned long duration_jiffies;
92 unsigned n_sectors;
93};
94
95/*
96 * For request-based dm - the bio clones we allocate are embedded in these
97 * structs.
98 *
99 * We allocate these with bio_alloc_bioset, using the front_pad parameter when
100 * the bioset is created - this means the bio has to come at the end of the
101 * struct.
102 */
103struct dm_rq_clone_bio_info {
104 struct bio *orig;
105 struct dm_rq_target_io *tio;
106 struct bio clone;
107};
108
109#define MINOR_ALLOCED ((void *)-1)
110
111/*
112 * Bits for the md->flags field.
113 */
114#define DMF_BLOCK_IO_FOR_SUSPEND 0
115#define DMF_SUSPENDED 1
116#define DMF_FROZEN 2
117#define DMF_FREEING 3
118#define DMF_DELETING 4
119#define DMF_NOFLUSH_SUSPENDING 5
120#define DMF_DEFERRED_REMOVE 6
121#define DMF_SUSPENDED_INTERNALLY 7
122
123/*
124 * Work processed by per-device workqueue.
125 */
126struct mapped_device {
127 struct srcu_struct io_barrier;
128 struct mutex suspend_lock;
129
130 /*
131 * The current mapping (struct dm_table *).
132 * Use dm_get_live_table{_fast} or take suspend_lock for
133 * dereference.
134 */
135 void __rcu *map;
136
137 struct list_head table_devices;
138 struct mutex table_devices_lock;
139
140 unsigned long flags;
141
142 struct request_queue *queue;
143 int numa_node_id;
144
145 unsigned type;
146 /* Protect queue and type against concurrent access. */
147 struct mutex type_lock;
148
149 atomic_t holders;
150 atomic_t open_count;
151
152 struct dm_target *immutable_target;
153 struct target_type *immutable_target_type;
154
155 struct gendisk *disk;
156 char name[16];
157
158 void *interface_ptr;
159
160 /*
161 * A list of ios that arrived while we were suspended.
162 */
163 atomic_t pending[2];
164 wait_queue_head_t wait;
165 struct work_struct work;
166 spinlock_t deferred_lock;
167 struct bio_list deferred;
168
169 /*
170 * Event handling.
171 */
172 wait_queue_head_t eventq;
173 atomic_t event_nr;
174 atomic_t uevent_seq;
175 struct list_head uevent_list;
176 spinlock_t uevent_lock; /* Protect access to uevent_list */
177
178 /* the number of internal suspends */
179 unsigned internal_suspend_count;
180
181 /*
182 * Processing queue (flush)
183 */
184 struct workqueue_struct *wq;
185
186 /*
187 * io objects are allocated from here.
188 */
189 mempool_t *io_pool;
190 mempool_t *rq_pool;
191
192 struct bio_set *bs;
193
194 /*
195 * freeze/thaw support require holding onto a super block
196 */
197 struct super_block *frozen_sb;
198
199 /* forced geometry settings */
200 struct hd_geometry geometry;
201
202 struct block_device *bdev;
203
204 /* kobject and completion */
205 struct dm_kobject_holder kobj_holder;
206
207 /* zero-length flush that will be cloned and submitted to targets */
208 struct bio flush_bio;
209
210 struct dm_stats stats;
211
212 struct kthread_worker kworker;
213 struct task_struct *kworker_task;
214
215 /* for request-based merge heuristic in dm_request_fn() */
216 unsigned seq_rq_merge_deadline_usecs;
217 int last_rq_rw;
218 sector_t last_rq_pos;
219 ktime_t last_rq_start_time;
220
221 /* for blk-mq request-based DM support */
222 struct blk_mq_tag_set *tag_set;
223 bool use_blk_mq:1;
224 bool init_tio_pdu:1;
225};
226
227#ifdef CONFIG_DM_MQ_DEFAULT
228static bool use_blk_mq = true;
229#else
230static bool use_blk_mq = false;
231#endif
232
233#define DM_MQ_NR_HW_QUEUES 1
234#define DM_MQ_QUEUE_DEPTH 2048
235#define DM_NUMA_NODE NUMA_NO_NODE
236
237static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
238static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
239static int dm_numa_node = DM_NUMA_NODE;
240
241bool dm_use_blk_mq(struct mapped_device *md)
242{
243 return md->use_blk_mq;
244}
245EXPORT_SYMBOL_GPL(dm_use_blk_mq);
246
247/*
248 * For mempools pre-allocation at the table loading time.
249 */
250struct dm_md_mempools {
251 mempool_t *io_pool;
252 mempool_t *rq_pool;
253 struct bio_set *bs;
254};
255
256struct table_device {
257 struct list_head list;
258 atomic_t count;
259 struct dm_dev dm_dev;
260};
261
262#define RESERVED_BIO_BASED_IOS 16
263#define RESERVED_REQUEST_BASED_IOS 256
264#define RESERVED_MAX_IOS 1024
265static struct kmem_cache *_io_cache;
266static struct kmem_cache *_rq_tio_cache;
267static struct kmem_cache *_rq_cache;
268
269/*
270 * Bio-based DM's mempools' reserved IOs set by the user.
271 */
272static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
273
274/*
275 * Request-based DM's mempools' reserved IOs set by the user.
276 */
277static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
278
279static int __dm_get_module_param_int(int *module_param, int min, int max)
280{
281 int param = ACCESS_ONCE(*module_param);
282 int modified_param = 0;
283 bool modified = true;
284
285 if (param < min)
286 modified_param = min;
287 else if (param > max)
288 modified_param = max;
289 else
290 modified = false;
291
292 if (modified) {
293 (void)cmpxchg(module_param, param, modified_param);
294 param = modified_param;
295 }
296
297 return param;
298}
299
300static unsigned __dm_get_module_param(unsigned *module_param,
301 unsigned def, unsigned max)
302{
303 unsigned param = ACCESS_ONCE(*module_param);
304 unsigned modified_param = 0;
305
306 if (!param)
307 modified_param = def;
308 else if (param > max)
309 modified_param = max;
310
311 if (modified_param) {
312 (void)cmpxchg(module_param, param, modified_param);
313 param = modified_param;
314 }
315
316 return param;
317}
318
319unsigned dm_get_reserved_bio_based_ios(void)
320{
321 return __dm_get_module_param(&reserved_bio_based_ios,
322 RESERVED_BIO_BASED_IOS, RESERVED_MAX_IOS);
323}
324EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
325
326unsigned dm_get_reserved_rq_based_ios(void)
327{
328 return __dm_get_module_param(&reserved_rq_based_ios,
329 RESERVED_REQUEST_BASED_IOS, RESERVED_MAX_IOS);
330}
331EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
332
333static unsigned dm_get_blk_mq_nr_hw_queues(void)
334{
335 return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
336}
337
338static unsigned dm_get_blk_mq_queue_depth(void)
339{
340 return __dm_get_module_param(&dm_mq_queue_depth,
341 DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
342}
343
344static unsigned dm_get_numa_node(void)
345{
346 return __dm_get_module_param_int(&dm_numa_node,
347 DM_NUMA_NODE, num_online_nodes() - 1);
348}
349
350static int __init local_init(void)
351{
352 int r = -ENOMEM;
353
354 /* allocate a slab for the dm_ios */
355 _io_cache = KMEM_CACHE(dm_io, 0);
356 if (!_io_cache)
357 return r;
358
359 _rq_tio_cache = KMEM_CACHE(dm_rq_target_io, 0);
360 if (!_rq_tio_cache)
361 goto out_free_io_cache;
362
363 _rq_cache = kmem_cache_create("dm_old_clone_request", sizeof(struct request),
364 __alignof__(struct request), 0, NULL);
365 if (!_rq_cache)
366 goto out_free_rq_tio_cache;
367
368 r = dm_uevent_init();
369 if (r)
370 goto out_free_rq_cache;
371
372 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
373 if (!deferred_remove_workqueue) {
374 r = -ENOMEM;
375 goto out_uevent_exit;
376 }
377
378 _major = major;
379 r = register_blkdev(_major, _name);
380 if (r < 0)
381 goto out_free_workqueue;
382
383 if (!_major)
384 _major = r;
385
386 return 0;
387
388out_free_workqueue:
389 destroy_workqueue(deferred_remove_workqueue);
390out_uevent_exit:
391 dm_uevent_exit();
392out_free_rq_cache:
393 kmem_cache_destroy(_rq_cache);
394out_free_rq_tio_cache:
395 kmem_cache_destroy(_rq_tio_cache);
396out_free_io_cache:
397 kmem_cache_destroy(_io_cache);
398
399 return r;
400}
401
402static void local_exit(void)
403{
404 flush_scheduled_work();
405 destroy_workqueue(deferred_remove_workqueue);
406
407 kmem_cache_destroy(_rq_cache);
408 kmem_cache_destroy(_rq_tio_cache);
409 kmem_cache_destroy(_io_cache);
410 unregister_blkdev(_major, _name);
411 dm_uevent_exit();
412
413 _major = 0;
414
415 DMINFO("cleaned up");
416}
417
418static int (*_inits[])(void) __initdata = {
419 local_init,
420 dm_target_init,
421 dm_linear_init,
422 dm_stripe_init,
423 dm_io_init,
424 dm_kcopyd_init,
425 dm_interface_init,
426 dm_statistics_init,
427};
428
429static void (*_exits[])(void) = {
430 local_exit,
431 dm_target_exit,
432 dm_linear_exit,
433 dm_stripe_exit,
434 dm_io_exit,
435 dm_kcopyd_exit,
436 dm_interface_exit,
437 dm_statistics_exit,
438};
439
440static int __init dm_init(void)
441{
442 const int count = ARRAY_SIZE(_inits);
443
444 int r, i;
445
446 for (i = 0; i < count; i++) {
447 r = _inits[i]();
448 if (r)
449 goto bad;
450 }
451
452 return 0;
453
454 bad:
455 while (i--)
456 _exits[i]();
457
458 return r;
459}
460
461static void __exit dm_exit(void)
462{
463 int i = ARRAY_SIZE(_exits);
464
465 while (i--)
466 _exits[i]();
467
468 /*
469 * Should be empty by this point.
470 */
471 idr_destroy(&_minor_idr);
472}
473
474/*
475 * Block device functions
476 */
477int dm_deleting_md(struct mapped_device *md)
478{
479 return test_bit(DMF_DELETING, &md->flags);
480}
481
482static int dm_blk_open(struct block_device *bdev, fmode_t mode)
483{
484 struct mapped_device *md;
485
486 spin_lock(&_minor_lock);
487
488 md = bdev->bd_disk->private_data;
489 if (!md)
490 goto out;
491
492 if (test_bit(DMF_FREEING, &md->flags) ||
493 dm_deleting_md(md)) {
494 md = NULL;
495 goto out;
496 }
497
498 dm_get(md);
499 atomic_inc(&md->open_count);
500out:
501 spin_unlock(&_minor_lock);
502
503 return md ? 0 : -ENXIO;
504}
505
506static void dm_blk_close(struct gendisk *disk, fmode_t mode)
507{
508 struct mapped_device *md;
509
510 spin_lock(&_minor_lock);
511
512 md = disk->private_data;
513 if (WARN_ON(!md))
514 goto out;
515
516 if (atomic_dec_and_test(&md->open_count) &&
517 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
518 queue_work(deferred_remove_workqueue, &deferred_remove_work);
519
520 dm_put(md);
521out:
522 spin_unlock(&_minor_lock);
523}
524
525int dm_open_count(struct mapped_device *md)
526{
527 return atomic_read(&md->open_count);
528}
529
530/*
531 * Guarantees nothing is using the device before it's deleted.
532 */
533int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
534{
535 int r = 0;
536
537 spin_lock(&_minor_lock);
538
539 if (dm_open_count(md)) {
540 r = -EBUSY;
541 if (mark_deferred)
542 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
543 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
544 r = -EEXIST;
545 else
546 set_bit(DMF_DELETING, &md->flags);
547
548 spin_unlock(&_minor_lock);
549
550 return r;
551}
552
553int dm_cancel_deferred_remove(struct mapped_device *md)
554{
555 int r = 0;
556
557 spin_lock(&_minor_lock);
558
559 if (test_bit(DMF_DELETING, &md->flags))
560 r = -EBUSY;
561 else
562 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
563
564 spin_unlock(&_minor_lock);
565
566 return r;
567}
568
569static void do_deferred_remove(struct work_struct *w)
570{
571 dm_deferred_remove();
572}
573
574sector_t dm_get_size(struct mapped_device *md)
575{
576 return get_capacity(md->disk);
577}
578
579struct request_queue *dm_get_md_queue(struct mapped_device *md)
580{
581 return md->queue;
582}
583
584struct dm_stats *dm_get_stats(struct mapped_device *md)
585{
586 return &md->stats;
587}
588
589static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
590{
591 struct mapped_device *md = bdev->bd_disk->private_data;
592
593 return dm_get_geometry(md, geo);
594}
595
596static int dm_grab_bdev_for_ioctl(struct mapped_device *md,
597 struct block_device **bdev,
598 fmode_t *mode)
599{
600 struct dm_target *tgt;
601 struct dm_table *map;
602 int srcu_idx, r;
603
604retry:
605 r = -ENOTTY;
606 map = dm_get_live_table(md, &srcu_idx);
607 if (!map || !dm_table_get_size(map))
608 goto out;
609
610 /* We only support devices that have a single target */
611 if (dm_table_get_num_targets(map) != 1)
612 goto out;
613
614 tgt = dm_table_get_target(map, 0);
615 if (!tgt->type->prepare_ioctl)
616 goto out;
617
618 if (dm_suspended_md(md)) {
619 r = -EAGAIN;
620 goto out;
621 }
622
623 r = tgt->type->prepare_ioctl(tgt, bdev, mode);
624 if (r < 0)
625 goto out;
626
627 bdgrab(*bdev);
628 dm_put_live_table(md, srcu_idx);
629 return r;
630
631out:
632 dm_put_live_table(md, srcu_idx);
633 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
634 msleep(10);
635 goto retry;
636 }
637 return r;
638}
639
640static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
641 unsigned int cmd, unsigned long arg)
642{
643 struct mapped_device *md = bdev->bd_disk->private_data;
644 int r;
645
646 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
647 if (r < 0)
648 return r;
649
650 if (r > 0) {
651 /*
652 * Target determined this ioctl is being issued against
653 * a logical partition of the parent bdev; so extra
654 * validation is needed.
655 */
656 r = scsi_verify_blk_ioctl(NULL, cmd);
657 if (r)
658 goto out;
659 }
660
661 r = __blkdev_driver_ioctl(bdev, mode, cmd, arg);
662out:
663 bdput(bdev);
664 return r;
665}
666
667static struct dm_io *alloc_io(struct mapped_device *md)
668{
669 return mempool_alloc(md->io_pool, GFP_NOIO);
670}
671
672static void free_io(struct mapped_device *md, struct dm_io *io)
673{
674 mempool_free(io, md->io_pool);
675}
676
677static void free_tio(struct mapped_device *md, struct dm_target_io *tio)
678{
679 bio_put(&tio->clone);
680}
681
682static struct dm_rq_target_io *alloc_old_rq_tio(struct mapped_device *md,
683 gfp_t gfp_mask)
684{
685 return mempool_alloc(md->io_pool, gfp_mask);
686}
687
688static void free_old_rq_tio(struct dm_rq_target_io *tio)
689{
690 mempool_free(tio, tio->md->io_pool);
691}
692
693static struct request *alloc_old_clone_request(struct mapped_device *md,
694 gfp_t gfp_mask)
695{
696 return mempool_alloc(md->rq_pool, gfp_mask);
697}
698
699static void free_old_clone_request(struct mapped_device *md, struct request *rq)
700{
701 mempool_free(rq, md->rq_pool);
702}
703
704static int md_in_flight(struct mapped_device *md)
705{
706 return atomic_read(&md->pending[READ]) +
707 atomic_read(&md->pending[WRITE]);
708}
709
710static void start_io_acct(struct dm_io *io)
711{
712 struct mapped_device *md = io->md;
713 struct bio *bio = io->bio;
714 int cpu;
715 int rw = bio_data_dir(bio);
716
717 io->start_time = jiffies;
718
719 cpu = part_stat_lock();
720 part_round_stats(cpu, &dm_disk(md)->part0);
721 part_stat_unlock();
722 atomic_set(&dm_disk(md)->part0.in_flight[rw],
723 atomic_inc_return(&md->pending[rw]));
724
725 if (unlikely(dm_stats_used(&md->stats)))
726 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
727 bio_sectors(bio), false, 0, &io->stats_aux);
728}
729
730static void end_io_acct(struct dm_io *io)
731{
732 struct mapped_device *md = io->md;
733 struct bio *bio = io->bio;
734 unsigned long duration = jiffies - io->start_time;
735 int pending;
736 int rw = bio_data_dir(bio);
737
738 generic_end_io_acct(rw, &dm_disk(md)->part0, io->start_time);
739
740 if (unlikely(dm_stats_used(&md->stats)))
741 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
742 bio_sectors(bio), true, duration, &io->stats_aux);
743
744 /*
745 * After this is decremented the bio must not be touched if it is
746 * a flush.
747 */
748 pending = atomic_dec_return(&md->pending[rw]);
749 atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
750 pending += atomic_read(&md->pending[rw^0x1]);
751
752 /* nudge anyone waiting on suspend queue */
753 if (!pending)
754 wake_up(&md->wait);
755}
756
757/*
758 * Add the bio to the list of deferred io.
759 */
760static void queue_io(struct mapped_device *md, struct bio *bio)
761{
762 unsigned long flags;
763
764 spin_lock_irqsave(&md->deferred_lock, flags);
765 bio_list_add(&md->deferred, bio);
766 spin_unlock_irqrestore(&md->deferred_lock, flags);
767 queue_work(md->wq, &md->work);
768}
769
770/*
771 * Everyone (including functions in this file), should use this
772 * function to access the md->map field, and make sure they call
773 * dm_put_live_table() when finished.
774 */
775struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
776{
777 *srcu_idx = srcu_read_lock(&md->io_barrier);
778
779 return srcu_dereference(md->map, &md->io_barrier);
780}
781
782void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
783{
784 srcu_read_unlock(&md->io_barrier, srcu_idx);
785}
786
787void dm_sync_table(struct mapped_device *md)
788{
789 synchronize_srcu(&md->io_barrier);
790 synchronize_rcu_expedited();
791}
792
793/*
794 * A fast alternative to dm_get_live_table/dm_put_live_table.
795 * The caller must not block between these two functions.
796 */
797static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
798{
799 rcu_read_lock();
800 return rcu_dereference(md->map);
801}
802
803static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
804{
805 rcu_read_unlock();
806}
807
808/*
809 * Open a table device so we can use it as a map destination.
810 */
811static int open_table_device(struct table_device *td, dev_t dev,
812 struct mapped_device *md)
813{
814 static char *_claim_ptr = "I belong to device-mapper";
815 struct block_device *bdev;
816
817 int r;
818
819 BUG_ON(td->dm_dev.bdev);
820
821 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _claim_ptr);
822 if (IS_ERR(bdev))
823 return PTR_ERR(bdev);
824
825 r = bd_link_disk_holder(bdev, dm_disk(md));
826 if (r) {
827 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
828 return r;
829 }
830
831 td->dm_dev.bdev = bdev;
832 return 0;
833}
834
835/*
836 * Close a table device that we've been using.
837 */
838static void close_table_device(struct table_device *td, struct mapped_device *md)
839{
840 if (!td->dm_dev.bdev)
841 return;
842
843 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
844 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
845 td->dm_dev.bdev = NULL;
846}
847
848static struct table_device *find_table_device(struct list_head *l, dev_t dev,
849 fmode_t mode) {
850 struct table_device *td;
851
852 list_for_each_entry(td, l, list)
853 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
854 return td;
855
856 return NULL;
857}
858
859int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
860 struct dm_dev **result) {
861 int r;
862 struct table_device *td;
863
864 mutex_lock(&md->table_devices_lock);
865 td = find_table_device(&md->table_devices, dev, mode);
866 if (!td) {
867 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
868 if (!td) {
869 mutex_unlock(&md->table_devices_lock);
870 return -ENOMEM;
871 }
872
873 td->dm_dev.mode = mode;
874 td->dm_dev.bdev = NULL;
875
876 if ((r = open_table_device(td, dev, md))) {
877 mutex_unlock(&md->table_devices_lock);
878 kfree(td);
879 return r;
880 }
881
882 format_dev_t(td->dm_dev.name, dev);
883
884 atomic_set(&td->count, 0);
885 list_add(&td->list, &md->table_devices);
886 }
887 atomic_inc(&td->count);
888 mutex_unlock(&md->table_devices_lock);
889
890 *result = &td->dm_dev;
891 return 0;
892}
893EXPORT_SYMBOL_GPL(dm_get_table_device);
894
895void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
896{
897 struct table_device *td = container_of(d, struct table_device, dm_dev);
898
899 mutex_lock(&md->table_devices_lock);
900 if (atomic_dec_and_test(&td->count)) {
901 close_table_device(td, md);
902 list_del(&td->list);
903 kfree(td);
904 }
905 mutex_unlock(&md->table_devices_lock);
906}
907EXPORT_SYMBOL(dm_put_table_device);
908
909static void free_table_devices(struct list_head *devices)
910{
911 struct list_head *tmp, *next;
912
913 list_for_each_safe(tmp, next, devices) {
914 struct table_device *td = list_entry(tmp, struct table_device, list);
915
916 DMWARN("dm_destroy: %s still exists with %d references",
917 td->dm_dev.name, atomic_read(&td->count));
918 kfree(td);
919 }
920}
921
922/*
923 * Get the geometry associated with a dm device
924 */
925int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
926{
927 *geo = md->geometry;
928
929 return 0;
930}
931
932/*
933 * Set the geometry of a device.
934 */
935int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
936{
937 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
938
939 if (geo->start > sz) {
940 DMWARN("Start sector is beyond the geometry limits.");
941 return -EINVAL;
942 }
943
944 md->geometry = *geo;
945
946 return 0;
947}
948
949/*-----------------------------------------------------------------
950 * CRUD START:
951 * A more elegant soln is in the works that uses the queue
952 * merge fn, unfortunately there are a couple of changes to
953 * the block layer that I want to make for this. So in the
954 * interests of getting something for people to use I give
955 * you this clearly demarcated crap.
956 *---------------------------------------------------------------*/
957
958static int __noflush_suspending(struct mapped_device *md)
959{
960 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
961}
962
963/*
964 * Decrements the number of outstanding ios that a bio has been
965 * cloned into, completing the original io if necc.
966 */
967static void dec_pending(struct dm_io *io, int error)
968{
969 unsigned long flags;
970 int io_error;
971 struct bio *bio;
972 struct mapped_device *md = io->md;
973
974 /* Push-back supersedes any I/O errors */
975 if (unlikely(error)) {
976 spin_lock_irqsave(&io->endio_lock, flags);
977 if (!(io->error > 0 && __noflush_suspending(md)))
978 io->error = error;
979 spin_unlock_irqrestore(&io->endio_lock, flags);
980 }
981
982 if (atomic_dec_and_test(&io->io_count)) {
983 if (io->error == DM_ENDIO_REQUEUE) {
984 /*
985 * Target requested pushing back the I/O.
986 */
987 spin_lock_irqsave(&md->deferred_lock, flags);
988 if (__noflush_suspending(md))
989 bio_list_add_head(&md->deferred, io->bio);
990 else
991 /* noflush suspend was interrupted. */
992 io->error = -EIO;
993 spin_unlock_irqrestore(&md->deferred_lock, flags);
994 }
995
996 io_error = io->error;
997 bio = io->bio;
998 end_io_acct(io);
999 free_io(md, io);
1000
1001 if (io_error == DM_ENDIO_REQUEUE)
1002 return;
1003
1004 if ((bio->bi_rw & REQ_FLUSH) && bio->bi_iter.bi_size) {
1005 /*
1006 * Preflush done for flush with data, reissue
1007 * without REQ_FLUSH.
1008 */
1009 bio->bi_rw &= ~REQ_FLUSH;
1010 queue_io(md, bio);
1011 } else {
1012 /* done with normal IO or empty flush */
1013 trace_block_bio_complete(md->queue, bio, io_error);
1014 bio->bi_error = io_error;
1015 bio_endio(bio);
1016 }
1017 }
1018}
1019
1020static void disable_write_same(struct mapped_device *md)
1021{
1022 struct queue_limits *limits = dm_get_queue_limits(md);
1023
1024 /* device doesn't really support WRITE SAME, disable it */
1025 limits->max_write_same_sectors = 0;
1026}
1027
1028static void clone_endio(struct bio *bio)
1029{
1030 int error = bio->bi_error;
1031 int r = error;
1032 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1033 struct dm_io *io = tio->io;
1034 struct mapped_device *md = tio->io->md;
1035 dm_endio_fn endio = tio->ti->type->end_io;
1036
1037 if (endio) {
1038 r = endio(tio->ti, bio, error);
1039 if (r < 0 || r == DM_ENDIO_REQUEUE)
1040 /*
1041 * error and requeue request are handled
1042 * in dec_pending().
1043 */
1044 error = r;
1045 else if (r == DM_ENDIO_INCOMPLETE)
1046 /* The target will handle the io */
1047 return;
1048 else if (r) {
1049 DMWARN("unimplemented target endio return value: %d", r);
1050 BUG();
1051 }
1052 }
1053
1054 if (unlikely(r == -EREMOTEIO && (bio->bi_rw & REQ_WRITE_SAME) &&
1055 !bdev_get_queue(bio->bi_bdev)->limits.max_write_same_sectors))
1056 disable_write_same(md);
1057
1058 free_tio(md, tio);
1059 dec_pending(io, error);
1060}
1061
1062/*
1063 * Partial completion handling for request-based dm
1064 */
1065static void end_clone_bio(struct bio *clone)
1066{
1067 struct dm_rq_clone_bio_info *info =
1068 container_of(clone, struct dm_rq_clone_bio_info, clone);
1069 struct dm_rq_target_io *tio = info->tio;
1070 struct bio *bio = info->orig;
1071 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
1072 int error = clone->bi_error;
1073
1074 bio_put(clone);
1075
1076 if (tio->error)
1077 /*
1078 * An error has already been detected on the request.
1079 * Once error occurred, just let clone->end_io() handle
1080 * the remainder.
1081 */
1082 return;
1083 else if (error) {
1084 /*
1085 * Don't notice the error to the upper layer yet.
1086 * The error handling decision is made by the target driver,
1087 * when the request is completed.
1088 */
1089 tio->error = error;
1090 return;
1091 }
1092
1093 /*
1094 * I/O for the bio successfully completed.
1095 * Notice the data completion to the upper layer.
1096 */
1097
1098 /*
1099 * bios are processed from the head of the list.
1100 * So the completing bio should always be rq->bio.
1101 * If it's not, something wrong is happening.
1102 */
1103 if (tio->orig->bio != bio)
1104 DMERR("bio completion is going in the middle of the request");
1105
1106 /*
1107 * Update the original request.
1108 * Do not use blk_end_request() here, because it may complete
1109 * the original request before the clone, and break the ordering.
1110 */
1111 blk_update_request(tio->orig, 0, nr_bytes);
1112}
1113
1114static struct dm_rq_target_io *tio_from_request(struct request *rq)
1115{
1116 return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
1117}
1118
1119static void rq_end_stats(struct mapped_device *md, struct request *orig)
1120{
1121 if (unlikely(dm_stats_used(&md->stats))) {
1122 struct dm_rq_target_io *tio = tio_from_request(orig);
1123 tio->duration_jiffies = jiffies - tio->duration_jiffies;
1124 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
1125 tio->n_sectors, true, tio->duration_jiffies,
1126 &tio->stats_aux);
1127 }
1128}
1129
1130/*
1131 * Don't touch any member of the md after calling this function because
1132 * the md may be freed in dm_put() at the end of this function.
1133 * Or do dm_get() before calling this function and dm_put() later.
1134 */
1135static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
1136{
1137 atomic_dec(&md->pending[rw]);
1138
1139 /* nudge anyone waiting on suspend queue */
1140 if (!md_in_flight(md))
1141 wake_up(&md->wait);
1142
1143 /*
1144 * Run this off this callpath, as drivers could invoke end_io while
1145 * inside their request_fn (and holding the queue lock). Calling
1146 * back into ->request_fn() could deadlock attempting to grab the
1147 * queue lock again.
1148 */
1149 if (!md->queue->mq_ops && run_queue)
1150 blk_run_queue_async(md->queue);
1151
1152 /*
1153 * dm_put() must be at the end of this function. See the comment above
1154 */
1155 dm_put(md);
1156}
1157
1158static void free_rq_clone(struct request *clone)
1159{
1160 struct dm_rq_target_io *tio = clone->end_io_data;
1161 struct mapped_device *md = tio->md;
1162
1163 blk_rq_unprep_clone(clone);
1164
1165 if (md->type == DM_TYPE_MQ_REQUEST_BASED)
1166 /* stacked on blk-mq queue(s) */
1167 tio->ti->type->release_clone_rq(clone);
1168 else if (!md->queue->mq_ops)
1169 /* request_fn queue stacked on request_fn queue(s) */
1170 free_old_clone_request(md, clone);
1171
1172 if (!md->queue->mq_ops)
1173 free_old_rq_tio(tio);
1174}
1175
1176/*
1177 * Complete the clone and the original request.
1178 * Must be called without clone's queue lock held,
1179 * see end_clone_request() for more details.
1180 */
1181static void dm_end_request(struct request *clone, int error)
1182{
1183 int rw = rq_data_dir(clone);
1184 struct dm_rq_target_io *tio = clone->end_io_data;
1185 struct mapped_device *md = tio->md;
1186 struct request *rq = tio->orig;
1187
1188 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
1189 rq->errors = clone->errors;
1190 rq->resid_len = clone->resid_len;
1191
1192 if (rq->sense)
1193 /*
1194 * We are using the sense buffer of the original
1195 * request.
1196 * So setting the length of the sense data is enough.
1197 */
1198 rq->sense_len = clone->sense_len;
1199 }
1200
1201 free_rq_clone(clone);
1202 rq_end_stats(md, rq);
1203 if (!rq->q->mq_ops)
1204 blk_end_request_all(rq, error);
1205 else
1206 blk_mq_end_request(rq, error);
1207 rq_completed(md, rw, true);
1208}
1209
1210static void dm_unprep_request(struct request *rq)
1211{
1212 struct dm_rq_target_io *tio = tio_from_request(rq);
1213 struct request *clone = tio->clone;
1214
1215 if (!rq->q->mq_ops) {
1216 rq->special = NULL;
1217 rq->cmd_flags &= ~REQ_DONTPREP;
1218 }
1219
1220 if (clone)
1221 free_rq_clone(clone);
1222 else if (!tio->md->queue->mq_ops)
1223 free_old_rq_tio(tio);
1224}
1225
1226/*
1227 * Requeue the original request of a clone.
1228 */
1229static void dm_old_requeue_request(struct request *rq)
1230{
1231 struct request_queue *q = rq->q;
1232 unsigned long flags;
1233
1234 spin_lock_irqsave(q->queue_lock, flags);
1235 blk_requeue_request(q, rq);
1236 blk_run_queue_async(q);
1237 spin_unlock_irqrestore(q->queue_lock, flags);
1238}
1239
1240static void dm_mq_requeue_request(struct request *rq)
1241{
1242 struct request_queue *q = rq->q;
1243 unsigned long flags;
1244
1245 blk_mq_requeue_request(rq);
1246 spin_lock_irqsave(q->queue_lock, flags);
1247 if (!blk_queue_stopped(q))
1248 blk_mq_kick_requeue_list(q);
1249 spin_unlock_irqrestore(q->queue_lock, flags);
1250}
1251
1252static void dm_requeue_original_request(struct mapped_device *md,
1253 struct request *rq)
1254{
1255 int rw = rq_data_dir(rq);
1256
1257 rq_end_stats(md, rq);
1258 dm_unprep_request(rq);
1259
1260 if (!rq->q->mq_ops)
1261 dm_old_requeue_request(rq);
1262 else
1263 dm_mq_requeue_request(rq);
1264
1265 rq_completed(md, rw, false);
1266}
1267
1268static void dm_old_stop_queue(struct request_queue *q)
1269{
1270 unsigned long flags;
1271
1272 spin_lock_irqsave(q->queue_lock, flags);
1273 if (blk_queue_stopped(q)) {
1274 spin_unlock_irqrestore(q->queue_lock, flags);
1275 return;
1276 }
1277
1278 blk_stop_queue(q);
1279 spin_unlock_irqrestore(q->queue_lock, flags);
1280}
1281
1282static void dm_stop_queue(struct request_queue *q)
1283{
1284 if (!q->mq_ops)
1285 dm_old_stop_queue(q);
1286 else
1287 blk_mq_stop_hw_queues(q);
1288}
1289
1290static void dm_old_start_queue(struct request_queue *q)
1291{
1292 unsigned long flags;
1293
1294 spin_lock_irqsave(q->queue_lock, flags);
1295 if (blk_queue_stopped(q))
1296 blk_start_queue(q);
1297 spin_unlock_irqrestore(q->queue_lock, flags);
1298}
1299
1300static void dm_start_queue(struct request_queue *q)
1301{
1302 if (!q->mq_ops)
1303 dm_old_start_queue(q);
1304 else {
1305 blk_mq_start_stopped_hw_queues(q, true);
1306 blk_mq_kick_requeue_list(q);
1307 }
1308}
1309
1310static void dm_done(struct request *clone, int error, bool mapped)
1311{
1312 int r = error;
1313 struct dm_rq_target_io *tio = clone->end_io_data;
1314 dm_request_endio_fn rq_end_io = NULL;
1315
1316 if (tio->ti) {
1317 rq_end_io = tio->ti->type->rq_end_io;
1318
1319 if (mapped && rq_end_io)
1320 r = rq_end_io(tio->ti, clone, error, &tio->info);
1321 }
1322
1323 if (unlikely(r == -EREMOTEIO && (clone->cmd_flags & REQ_WRITE_SAME) &&
1324 !clone->q->limits.max_write_same_sectors))
1325 disable_write_same(tio->md);
1326
1327 if (r <= 0)
1328 /* The target wants to complete the I/O */
1329 dm_end_request(clone, r);
1330 else if (r == DM_ENDIO_INCOMPLETE)
1331 /* The target will handle the I/O */
1332 return;
1333 else if (r == DM_ENDIO_REQUEUE)
1334 /* The target wants to requeue the I/O */
1335 dm_requeue_original_request(tio->md, tio->orig);
1336 else {
1337 DMWARN("unimplemented target endio return value: %d", r);
1338 BUG();
1339 }
1340}
1341
1342/*
1343 * Request completion handler for request-based dm
1344 */
1345static void dm_softirq_done(struct request *rq)
1346{
1347 bool mapped = true;
1348 struct dm_rq_target_io *tio = tio_from_request(rq);
1349 struct request *clone = tio->clone;
1350 int rw;
1351
1352 if (!clone) {
1353 rq_end_stats(tio->md, rq);
1354 rw = rq_data_dir(rq);
1355 if (!rq->q->mq_ops) {
1356 blk_end_request_all(rq, tio->error);
1357 rq_completed(tio->md, rw, false);
1358 free_old_rq_tio(tio);
1359 } else {
1360 blk_mq_end_request(rq, tio->error);
1361 rq_completed(tio->md, rw, false);
1362 }
1363 return;
1364 }
1365
1366 if (rq->cmd_flags & REQ_FAILED)
1367 mapped = false;
1368
1369 dm_done(clone, tio->error, mapped);
1370}
1371
1372/*
1373 * Complete the clone and the original request with the error status
1374 * through softirq context.
1375 */
1376static void dm_complete_request(struct request *rq, int error)
1377{
1378 struct dm_rq_target_io *tio = tio_from_request(rq);
1379
1380 tio->error = error;
1381 if (!rq->q->mq_ops)
1382 blk_complete_request(rq);
1383 else
1384 blk_mq_complete_request(rq, error);
1385}
1386
1387/*
1388 * Complete the not-mapped clone and the original request with the error status
1389 * through softirq context.
1390 * Target's rq_end_io() function isn't called.
1391 * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
1392 */
1393static void dm_kill_unmapped_request(struct request *rq, int error)
1394{
1395 rq->cmd_flags |= REQ_FAILED;
1396 dm_complete_request(rq, error);
1397}
1398
1399/*
1400 * Called with the clone's queue lock held (in the case of .request_fn)
1401 */
1402static void end_clone_request(struct request *clone, int error)
1403{
1404 struct dm_rq_target_io *tio = clone->end_io_data;
1405
1406 if (!clone->q->mq_ops) {
1407 /*
1408 * For just cleaning up the information of the queue in which
1409 * the clone was dispatched.
1410 * The clone is *NOT* freed actually here because it is alloced
1411 * from dm own mempool (REQ_ALLOCED isn't set).
1412 */
1413 __blk_put_request(clone->q, clone);
1414 }
1415
1416 /*
1417 * Actual request completion is done in a softirq context which doesn't
1418 * hold the clone's queue lock. Otherwise, deadlock could occur because:
1419 * - another request may be submitted by the upper level driver
1420 * of the stacking during the completion
1421 * - the submission which requires queue lock may be done
1422 * against this clone's queue
1423 */
1424 dm_complete_request(tio->orig, error);
1425}
1426
1427/*
1428 * Return maximum size of I/O possible at the supplied sector up to the current
1429 * target boundary.
1430 */
1431static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1432{
1433 sector_t target_offset = dm_target_offset(ti, sector);
1434
1435 return ti->len - target_offset;
1436}
1437
1438static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1439{
1440 sector_t len = max_io_len_target_boundary(sector, ti);
1441 sector_t offset, max_len;
1442
1443 /*
1444 * Does the target need to split even further?
1445 */
1446 if (ti->max_io_len) {
1447 offset = dm_target_offset(ti, sector);
1448 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1449 max_len = sector_div(offset, ti->max_io_len);
1450 else
1451 max_len = offset & (ti->max_io_len - 1);
1452 max_len = ti->max_io_len - max_len;
1453
1454 if (len > max_len)
1455 len = max_len;
1456 }
1457
1458 return len;
1459}
1460
1461int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1462{
1463 if (len > UINT_MAX) {
1464 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1465 (unsigned long long)len, UINT_MAX);
1466 ti->error = "Maximum size of target IO is too large";
1467 return -EINVAL;
1468 }
1469
1470 ti->max_io_len = (uint32_t) len;
1471
1472 return 0;
1473}
1474EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1475
1476/*
1477 * A target may call dm_accept_partial_bio only from the map routine. It is
1478 * allowed for all bio types except REQ_FLUSH.
1479 *
1480 * dm_accept_partial_bio informs the dm that the target only wants to process
1481 * additional n_sectors sectors of the bio and the rest of the data should be
1482 * sent in a next bio.
1483 *
1484 * A diagram that explains the arithmetics:
1485 * +--------------------+---------------+-------+
1486 * | 1 | 2 | 3 |
1487 * +--------------------+---------------+-------+
1488 *
1489 * <-------------- *tio->len_ptr --------------->
1490 * <------- bi_size ------->
1491 * <-- n_sectors -->
1492 *
1493 * Region 1 was already iterated over with bio_advance or similar function.
1494 * (it may be empty if the target doesn't use bio_advance)
1495 * Region 2 is the remaining bio size that the target wants to process.
1496 * (it may be empty if region 1 is non-empty, although there is no reason
1497 * to make it empty)
1498 * The target requires that region 3 is to be sent in the next bio.
1499 *
1500 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1501 * the partially processed part (the sum of regions 1+2) must be the same for all
1502 * copies of the bio.
1503 */
1504void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1505{
1506 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1507 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1508 BUG_ON(bio->bi_rw & REQ_FLUSH);
1509 BUG_ON(bi_size > *tio->len_ptr);
1510 BUG_ON(n_sectors > bi_size);
1511 *tio->len_ptr -= bi_size - n_sectors;
1512 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1513}
1514EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1515
1516static void __map_bio(struct dm_target_io *tio)
1517{
1518 int r;
1519 sector_t sector;
1520 struct mapped_device *md;
1521 struct bio *clone = &tio->clone;
1522 struct dm_target *ti = tio->ti;
1523
1524 clone->bi_end_io = clone_endio;
1525
1526 /*
1527 * Map the clone. If r == 0 we don't need to do
1528 * anything, the target has assumed ownership of
1529 * this io.
1530 */
1531 atomic_inc(&tio->io->io_count);
1532 sector = clone->bi_iter.bi_sector;
1533 r = ti->type->map(ti, clone);
1534 if (r == DM_MAPIO_REMAPPED) {
1535 /* the bio has been remapped so dispatch it */
1536
1537 trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1538 tio->io->bio->bi_bdev->bd_dev, sector);
1539
1540 generic_make_request(clone);
1541 } else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1542 /* error the io and bail out, or requeue it if needed */
1543 md = tio->io->md;
1544 dec_pending(tio->io, r);
1545 free_tio(md, tio);
1546 } else if (r != DM_MAPIO_SUBMITTED) {
1547 DMWARN("unimplemented target map return value: %d", r);
1548 BUG();
1549 }
1550}
1551
1552struct clone_info {
1553 struct mapped_device *md;
1554 struct dm_table *map;
1555 struct bio *bio;
1556 struct dm_io *io;
1557 sector_t sector;
1558 unsigned sector_count;
1559};
1560
1561static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1562{
1563 bio->bi_iter.bi_sector = sector;
1564 bio->bi_iter.bi_size = to_bytes(len);
1565}
1566
1567/*
1568 * Creates a bio that consists of range of complete bvecs.
1569 */
1570static int clone_bio(struct dm_target_io *tio, struct bio *bio,
1571 sector_t sector, unsigned len)
1572{
1573 struct bio *clone = &tio->clone;
1574
1575 __bio_clone_fast(clone, bio);
1576
1577 if (bio_integrity(bio)) {
1578 int r = bio_integrity_clone(clone, bio, GFP_NOIO);
1579 if (r < 0)
1580 return r;
1581 }
1582
1583 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1584 clone->bi_iter.bi_size = to_bytes(len);
1585
1586 if (bio_integrity(bio))
1587 bio_integrity_trim(clone, 0, len);
1588
1589 return 0;
1590}
1591
1592static struct dm_target_io *alloc_tio(struct clone_info *ci,
1593 struct dm_target *ti,
1594 unsigned target_bio_nr)
1595{
1596 struct dm_target_io *tio;
1597 struct bio *clone;
1598
1599 clone = bio_alloc_bioset(GFP_NOIO, 0, ci->md->bs);
1600 tio = container_of(clone, struct dm_target_io, clone);
1601
1602 tio->io = ci->io;
1603 tio->ti = ti;
1604 tio->target_bio_nr = target_bio_nr;
1605
1606 return tio;
1607}
1608
1609static void __clone_and_map_simple_bio(struct clone_info *ci,
1610 struct dm_target *ti,
1611 unsigned target_bio_nr, unsigned *len)
1612{
1613 struct dm_target_io *tio = alloc_tio(ci, ti, target_bio_nr);
1614 struct bio *clone = &tio->clone;
1615
1616 tio->len_ptr = len;
1617
1618 __bio_clone_fast(clone, ci->bio);
1619 if (len)
1620 bio_setup_sector(clone, ci->sector, *len);
1621
1622 __map_bio(tio);
1623}
1624
1625static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1626 unsigned num_bios, unsigned *len)
1627{
1628 unsigned target_bio_nr;
1629
1630 for (target_bio_nr = 0; target_bio_nr < num_bios; target_bio_nr++)
1631 __clone_and_map_simple_bio(ci, ti, target_bio_nr, len);
1632}
1633
1634static int __send_empty_flush(struct clone_info *ci)
1635{
1636 unsigned target_nr = 0;
1637 struct dm_target *ti;
1638
1639 BUG_ON(bio_has_data(ci->bio));
1640 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1641 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1642
1643 return 0;
1644}
1645
1646static int __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1647 sector_t sector, unsigned *len)
1648{
1649 struct bio *bio = ci->bio;
1650 struct dm_target_io *tio;
1651 unsigned target_bio_nr;
1652 unsigned num_target_bios = 1;
1653 int r = 0;
1654
1655 /*
1656 * Does the target want to receive duplicate copies of the bio?
1657 */
1658 if (bio_data_dir(bio) == WRITE && ti->num_write_bios)
1659 num_target_bios = ti->num_write_bios(ti, bio);
1660
1661 for (target_bio_nr = 0; target_bio_nr < num_target_bios; target_bio_nr++) {
1662 tio = alloc_tio(ci, ti, target_bio_nr);
1663 tio->len_ptr = len;
1664 r = clone_bio(tio, bio, sector, *len);
1665 if (r < 0) {
1666 free_tio(ci->md, tio);
1667 break;
1668 }
1669 __map_bio(tio);
1670 }
1671
1672 return r;
1673}
1674
1675typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1676
1677static unsigned get_num_discard_bios(struct dm_target *ti)
1678{
1679 return ti->num_discard_bios;
1680}
1681
1682static unsigned get_num_write_same_bios(struct dm_target *ti)
1683{
1684 return ti->num_write_same_bios;
1685}
1686
1687typedef bool (*is_split_required_fn)(struct dm_target *ti);
1688
1689static bool is_split_required_for_discard(struct dm_target *ti)
1690{
1691 return ti->split_discard_bios;
1692}
1693
1694static int __send_changing_extent_only(struct clone_info *ci,
1695 get_num_bios_fn get_num_bios,
1696 is_split_required_fn is_split_required)
1697{
1698 struct dm_target *ti;
1699 unsigned len;
1700 unsigned num_bios;
1701
1702 do {
1703 ti = dm_table_find_target(ci->map, ci->sector);
1704 if (!dm_target_is_valid(ti))
1705 return -EIO;
1706
1707 /*
1708 * Even though the device advertised support for this type of
1709 * request, that does not mean every target supports it, and
1710 * reconfiguration might also have changed that since the
1711 * check was performed.
1712 */
1713 num_bios = get_num_bios ? get_num_bios(ti) : 0;
1714 if (!num_bios)
1715 return -EOPNOTSUPP;
1716
1717 if (is_split_required && !is_split_required(ti))
1718 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1719 else
1720 len = min((sector_t)ci->sector_count, max_io_len(ci->sector, ti));
1721
1722 __send_duplicate_bios(ci, ti, num_bios, &len);
1723
1724 ci->sector += len;
1725 } while (ci->sector_count -= len);
1726
1727 return 0;
1728}
1729
1730static int __send_discard(struct clone_info *ci)
1731{
1732 return __send_changing_extent_only(ci, get_num_discard_bios,
1733 is_split_required_for_discard);
1734}
1735
1736static int __send_write_same(struct clone_info *ci)
1737{
1738 return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
1739}
1740
1741/*
1742 * Select the correct strategy for processing a non-flush bio.
1743 */
1744static int __split_and_process_non_flush(struct clone_info *ci)
1745{
1746 struct bio *bio = ci->bio;
1747 struct dm_target *ti;
1748 unsigned len;
1749 int r;
1750
1751 if (unlikely(bio->bi_rw & REQ_DISCARD))
1752 return __send_discard(ci);
1753 else if (unlikely(bio->bi_rw & REQ_WRITE_SAME))
1754 return __send_write_same(ci);
1755
1756 ti = dm_table_find_target(ci->map, ci->sector);
1757 if (!dm_target_is_valid(ti))
1758 return -EIO;
1759
1760 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1761
1762 r = __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1763 if (r < 0)
1764 return r;
1765
1766 ci->sector += len;
1767 ci->sector_count -= len;
1768
1769 return 0;
1770}
1771
1772/*
1773 * Entry point to split a bio into clones and submit them to the targets.
1774 */
1775static void __split_and_process_bio(struct mapped_device *md,
1776 struct dm_table *map, struct bio *bio)
1777{
1778 struct clone_info ci;
1779 int error = 0;
1780
1781 if (unlikely(!map)) {
1782 bio_io_error(bio);
1783 return;
1784 }
1785
1786 ci.map = map;
1787 ci.md = md;
1788 ci.io = alloc_io(md);
1789 ci.io->error = 0;
1790 atomic_set(&ci.io->io_count, 1);
1791 ci.io->bio = bio;
1792 ci.io->md = md;
1793 spin_lock_init(&ci.io->endio_lock);
1794 ci.sector = bio->bi_iter.bi_sector;
1795
1796 start_io_acct(ci.io);
1797
1798 if (bio->bi_rw & REQ_FLUSH) {
1799 ci.bio = &ci.md->flush_bio;
1800 ci.sector_count = 0;
1801 error = __send_empty_flush(&ci);
1802 /* dec_pending submits any data associated with flush */
1803 } else {
1804 ci.bio = bio;
1805 ci.sector_count = bio_sectors(bio);
1806 while (ci.sector_count && !error)
1807 error = __split_and_process_non_flush(&ci);
1808 }
1809
1810 /* drop the extra reference count */
1811 dec_pending(ci.io, error);
1812}
1813/*-----------------------------------------------------------------
1814 * CRUD END
1815 *---------------------------------------------------------------*/
1816
1817/*
1818 * The request function that just remaps the bio built up by
1819 * dm_merge_bvec.
1820 */
1821static blk_qc_t dm_make_request(struct request_queue *q, struct bio *bio)
1822{
1823 int rw = bio_data_dir(bio);
1824 struct mapped_device *md = q->queuedata;
1825 int srcu_idx;
1826 struct dm_table *map;
1827
1828 map = dm_get_live_table(md, &srcu_idx);
1829
1830 generic_start_io_acct(rw, bio_sectors(bio), &dm_disk(md)->part0);
1831
1832 /* if we're suspended, we have to queue this io for later */
1833 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1834 dm_put_live_table(md, srcu_idx);
1835
1836 if (bio_rw(bio) != READA)
1837 queue_io(md, bio);
1838 else
1839 bio_io_error(bio);
1840 return BLK_QC_T_NONE;
1841 }
1842
1843 __split_and_process_bio(md, map, bio);
1844 dm_put_live_table(md, srcu_idx);
1845 return BLK_QC_T_NONE;
1846}
1847
1848int dm_request_based(struct mapped_device *md)
1849{
1850 return blk_queue_stackable(md->queue);
1851}
1852
1853static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
1854{
1855 int r;
1856
1857 if (blk_queue_io_stat(clone->q))
1858 clone->cmd_flags |= REQ_IO_STAT;
1859
1860 clone->start_time = jiffies;
1861 r = blk_insert_cloned_request(clone->q, clone);
1862 if (r)
1863 /* must complete clone in terms of original request */
1864 dm_complete_request(rq, r);
1865}
1866
1867static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
1868 void *data)
1869{
1870 struct dm_rq_target_io *tio = data;
1871 struct dm_rq_clone_bio_info *info =
1872 container_of(bio, struct dm_rq_clone_bio_info, clone);
1873
1874 info->orig = bio_orig;
1875 info->tio = tio;
1876 bio->bi_end_io = end_clone_bio;
1877
1878 return 0;
1879}
1880
1881static int setup_clone(struct request *clone, struct request *rq,
1882 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1883{
1884 int r;
1885
1886 r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
1887 dm_rq_bio_constructor, tio);
1888 if (r)
1889 return r;
1890
1891 clone->cmd = rq->cmd;
1892 clone->cmd_len = rq->cmd_len;
1893 clone->sense = rq->sense;
1894 clone->end_io = end_clone_request;
1895 clone->end_io_data = tio;
1896
1897 tio->clone = clone;
1898
1899 return 0;
1900}
1901
1902static struct request *clone_old_rq(struct request *rq, struct mapped_device *md,
1903 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1904{
1905 /*
1906 * Create clone for use with .request_fn request_queue
1907 */
1908 struct request *clone;
1909
1910 clone = alloc_old_clone_request(md, gfp_mask);
1911 if (!clone)
1912 return NULL;
1913
1914 blk_rq_init(NULL, clone);
1915 if (setup_clone(clone, rq, tio, gfp_mask)) {
1916 /* -ENOMEM */
1917 free_old_clone_request(md, clone);
1918 return NULL;
1919 }
1920
1921 return clone;
1922}
1923
1924static void map_tio_request(struct kthread_work *work);
1925
1926static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
1927 struct mapped_device *md)
1928{
1929 tio->md = md;
1930 tio->ti = NULL;
1931 tio->clone = NULL;
1932 tio->orig = rq;
1933 tio->error = 0;
1934 /*
1935 * Avoid initializing info for blk-mq; it passes
1936 * target-specific data through info.ptr
1937 * (see: dm_mq_init_request)
1938 */
1939 if (!md->init_tio_pdu)
1940 memset(&tio->info, 0, sizeof(tio->info));
1941 if (md->kworker_task)
1942 init_kthread_work(&tio->work, map_tio_request);
1943}
1944
1945static struct dm_rq_target_io *dm_old_prep_tio(struct request *rq,
1946 struct mapped_device *md,
1947 gfp_t gfp_mask)
1948{
1949 struct dm_rq_target_io *tio;
1950 int srcu_idx;
1951 struct dm_table *table;
1952
1953 tio = alloc_old_rq_tio(md, gfp_mask);
1954 if (!tio)
1955 return NULL;
1956
1957 init_tio(tio, rq, md);
1958
1959 table = dm_get_live_table(md, &srcu_idx);
1960 /*
1961 * Must clone a request if this .request_fn DM device
1962 * is stacked on .request_fn device(s).
1963 */
1964 if (!dm_table_mq_request_based(table)) {
1965 if (!clone_old_rq(rq, md, tio, gfp_mask)) {
1966 dm_put_live_table(md, srcu_idx);
1967 free_old_rq_tio(tio);
1968 return NULL;
1969 }
1970 }
1971 dm_put_live_table(md, srcu_idx);
1972
1973 return tio;
1974}
1975
1976/*
1977 * Called with the queue lock held.
1978 */
1979static int dm_old_prep_fn(struct request_queue *q, struct request *rq)
1980{
1981 struct mapped_device *md = q->queuedata;
1982 struct dm_rq_target_io *tio;
1983
1984 if (unlikely(rq->special)) {
1985 DMWARN("Already has something in rq->special.");
1986 return BLKPREP_KILL;
1987 }
1988
1989 tio = dm_old_prep_tio(rq, md, GFP_ATOMIC);
1990 if (!tio)
1991 return BLKPREP_DEFER;
1992
1993 rq->special = tio;
1994 rq->cmd_flags |= REQ_DONTPREP;
1995
1996 return BLKPREP_OK;
1997}
1998
1999/*
2000 * Returns:
2001 * 0 : the request has been processed
2002 * DM_MAPIO_REQUEUE : the original request needs to be requeued
2003 * < 0 : the request was completed due to failure
2004 */
2005static int map_request(struct dm_rq_target_io *tio, struct request *rq,
2006 struct mapped_device *md)
2007{
2008 int r;
2009 struct dm_target *ti = tio->ti;
2010 struct request *clone = NULL;
2011
2012 if (tio->clone) {
2013 clone = tio->clone;
2014 r = ti->type->map_rq(ti, clone, &tio->info);
2015 } else {
2016 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
2017 if (r < 0) {
2018 /* The target wants to complete the I/O */
2019 dm_kill_unmapped_request(rq, r);
2020 return r;
2021 }
2022 if (r != DM_MAPIO_REMAPPED)
2023 return r;
2024 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
2025 /* -ENOMEM */
2026 ti->type->release_clone_rq(clone);
2027 return DM_MAPIO_REQUEUE;
2028 }
2029 }
2030
2031 switch (r) {
2032 case DM_MAPIO_SUBMITTED:
2033 /* The target has taken the I/O to submit by itself later */
2034 break;
2035 case DM_MAPIO_REMAPPED:
2036 /* The target has remapped the I/O so dispatch it */
2037 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
2038 blk_rq_pos(rq));
2039 dm_dispatch_clone_request(clone, rq);
2040 break;
2041 case DM_MAPIO_REQUEUE:
2042 /* The target wants to requeue the I/O */
2043 dm_requeue_original_request(md, tio->orig);
2044 break;
2045 default:
2046 if (r > 0) {
2047 DMWARN("unimplemented target map return value: %d", r);
2048 BUG();
2049 }
2050
2051 /* The target wants to complete the I/O */
2052 dm_kill_unmapped_request(rq, r);
2053 return r;
2054 }
2055
2056 return 0;
2057}
2058
2059static void map_tio_request(struct kthread_work *work)
2060{
2061 struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
2062 struct request *rq = tio->orig;
2063 struct mapped_device *md = tio->md;
2064
2065 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE)
2066 dm_requeue_original_request(md, rq);
2067}
2068
2069static void dm_start_request(struct mapped_device *md, struct request *orig)
2070{
2071 if (!orig->q->mq_ops)
2072 blk_start_request(orig);
2073 else
2074 blk_mq_start_request(orig);
2075 atomic_inc(&md->pending[rq_data_dir(orig)]);
2076
2077 if (md->seq_rq_merge_deadline_usecs) {
2078 md->last_rq_pos = rq_end_sector(orig);
2079 md->last_rq_rw = rq_data_dir(orig);
2080 md->last_rq_start_time = ktime_get();
2081 }
2082
2083 if (unlikely(dm_stats_used(&md->stats))) {
2084 struct dm_rq_target_io *tio = tio_from_request(orig);
2085 tio->duration_jiffies = jiffies;
2086 tio->n_sectors = blk_rq_sectors(orig);
2087 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
2088 tio->n_sectors, false, 0, &tio->stats_aux);
2089 }
2090
2091 /*
2092 * Hold the md reference here for the in-flight I/O.
2093 * We can't rely on the reference count by device opener,
2094 * because the device may be closed during the request completion
2095 * when all bios are completed.
2096 * See the comment in rq_completed() too.
2097 */
2098 dm_get(md);
2099}
2100
2101#define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
2102
2103ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
2104{
2105 return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
2106}
2107
2108ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
2109 const char *buf, size_t count)
2110{
2111 unsigned deadline;
2112
2113 if (!dm_request_based(md) || md->use_blk_mq)
2114 return count;
2115
2116 if (kstrtouint(buf, 10, &deadline))
2117 return -EINVAL;
2118
2119 if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
2120 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
2121
2122 md->seq_rq_merge_deadline_usecs = deadline;
2123
2124 return count;
2125}
2126
2127static bool dm_request_peeked_before_merge_deadline(struct mapped_device *md)
2128{
2129 ktime_t kt_deadline;
2130
2131 if (!md->seq_rq_merge_deadline_usecs)
2132 return false;
2133
2134 kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
2135 kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
2136
2137 return !ktime_after(ktime_get(), kt_deadline);
2138}
2139
2140/*
2141 * q->request_fn for request-based dm.
2142 * Called with the queue lock held.
2143 */
2144static void dm_request_fn(struct request_queue *q)
2145{
2146 struct mapped_device *md = q->queuedata;
2147 struct dm_target *ti = md->immutable_target;
2148 struct request *rq;
2149 struct dm_rq_target_io *tio;
2150 sector_t pos = 0;
2151
2152 if (unlikely(!ti)) {
2153 int srcu_idx;
2154 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2155
2156 ti = dm_table_find_target(map, pos);
2157 dm_put_live_table(md, srcu_idx);
2158 }
2159
2160 /*
2161 * For suspend, check blk_queue_stopped() and increment
2162 * ->pending within a single queue_lock not to increment the
2163 * number of in-flight I/Os after the queue is stopped in
2164 * dm_suspend().
2165 */
2166 while (!blk_queue_stopped(q)) {
2167 rq = blk_peek_request(q);
2168 if (!rq)
2169 return;
2170
2171 /* always use block 0 to find the target for flushes for now */
2172 pos = 0;
2173 if (!(rq->cmd_flags & REQ_FLUSH))
2174 pos = blk_rq_pos(rq);
2175
2176 if ((dm_request_peeked_before_merge_deadline(md) &&
2177 md_in_flight(md) && rq->bio && rq->bio->bi_vcnt == 1 &&
2178 md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
2179 (ti->type->busy && ti->type->busy(ti))) {
2180 blk_delay_queue(q, HZ / 100);
2181 return;
2182 }
2183
2184 dm_start_request(md, rq);
2185
2186 tio = tio_from_request(rq);
2187 /* Establish tio->ti before queuing work (map_tio_request) */
2188 tio->ti = ti;
2189 queue_kthread_work(&md->kworker, &tio->work);
2190 BUG_ON(!irqs_disabled());
2191 }
2192}
2193
2194static int dm_any_congested(void *congested_data, int bdi_bits)
2195{
2196 int r = bdi_bits;
2197 struct mapped_device *md = congested_data;
2198 struct dm_table *map;
2199
2200 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2201 if (dm_request_based(md)) {
2202 /*
2203 * With request-based DM we only need to check the
2204 * top-level queue for congestion.
2205 */
2206 r = md->queue->backing_dev_info.wb.state & bdi_bits;
2207 } else {
2208 map = dm_get_live_table_fast(md);
2209 if (map)
2210 r = dm_table_any_congested(map, bdi_bits);
2211 dm_put_live_table_fast(md);
2212 }
2213 }
2214
2215 return r;
2216}
2217
2218/*-----------------------------------------------------------------
2219 * An IDR is used to keep track of allocated minor numbers.
2220 *---------------------------------------------------------------*/
2221static void free_minor(int minor)
2222{
2223 spin_lock(&_minor_lock);
2224 idr_remove(&_minor_idr, minor);
2225 spin_unlock(&_minor_lock);
2226}
2227
2228/*
2229 * See if the device with a specific minor # is free.
2230 */
2231static int specific_minor(int minor)
2232{
2233 int r;
2234
2235 if (minor >= (1 << MINORBITS))
2236 return -EINVAL;
2237
2238 idr_preload(GFP_KERNEL);
2239 spin_lock(&_minor_lock);
2240
2241 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2242
2243 spin_unlock(&_minor_lock);
2244 idr_preload_end();
2245 if (r < 0)
2246 return r == -ENOSPC ? -EBUSY : r;
2247 return 0;
2248}
2249
2250static int next_free_minor(int *minor)
2251{
2252 int r;
2253
2254 idr_preload(GFP_KERNEL);
2255 spin_lock(&_minor_lock);
2256
2257 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2258
2259 spin_unlock(&_minor_lock);
2260 idr_preload_end();
2261 if (r < 0)
2262 return r;
2263 *minor = r;
2264 return 0;
2265}
2266
2267static const struct block_device_operations dm_blk_dops;
2268
2269static void dm_wq_work(struct work_struct *work);
2270
2271static void dm_init_md_queue(struct mapped_device *md)
2272{
2273 /*
2274 * Request-based dm devices cannot be stacked on top of bio-based dm
2275 * devices. The type of this dm device may not have been decided yet.
2276 * The type is decided at the first table loading time.
2277 * To prevent problematic device stacking, clear the queue flag
2278 * for request stacking support until then.
2279 *
2280 * This queue is new, so no concurrency on the queue_flags.
2281 */
2282 queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
2283
2284 /*
2285 * Initialize data that will only be used by a non-blk-mq DM queue
2286 * - must do so here (in alloc_dev callchain) before queue is used
2287 */
2288 md->queue->queuedata = md;
2289 md->queue->backing_dev_info.congested_data = md;
2290}
2291
2292static void dm_init_normal_md_queue(struct mapped_device *md)
2293{
2294 md->use_blk_mq = false;
2295 dm_init_md_queue(md);
2296
2297 /*
2298 * Initialize aspects of queue that aren't relevant for blk-mq
2299 */
2300 md->queue->backing_dev_info.congested_fn = dm_any_congested;
2301 blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
2302}
2303
2304static void cleanup_mapped_device(struct mapped_device *md)
2305{
2306 if (md->wq)
2307 destroy_workqueue(md->wq);
2308 if (md->kworker_task)
2309 kthread_stop(md->kworker_task);
2310 mempool_destroy(md->io_pool);
2311 mempool_destroy(md->rq_pool);
2312 if (md->bs)
2313 bioset_free(md->bs);
2314
2315 cleanup_srcu_struct(&md->io_barrier);
2316
2317 if (md->disk) {
2318 spin_lock(&_minor_lock);
2319 md->disk->private_data = NULL;
2320 spin_unlock(&_minor_lock);
2321 del_gendisk(md->disk);
2322 put_disk(md->disk);
2323 }
2324
2325 if (md->queue)
2326 blk_cleanup_queue(md->queue);
2327
2328 if (md->bdev) {
2329 bdput(md->bdev);
2330 md->bdev = NULL;
2331 }
2332}
2333
2334/*
2335 * Allocate and initialise a blank device with a given minor.
2336 */
2337static struct mapped_device *alloc_dev(int minor)
2338{
2339 int r, numa_node_id = dm_get_numa_node();
2340 struct mapped_device *md;
2341 void *old_md;
2342
2343 md = kzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
2344 if (!md) {
2345 DMWARN("unable to allocate device, out of memory.");
2346 return NULL;
2347 }
2348
2349 if (!try_module_get(THIS_MODULE))
2350 goto bad_module_get;
2351
2352 /* get a minor number for the dev */
2353 if (minor == DM_ANY_MINOR)
2354 r = next_free_minor(&minor);
2355 else
2356 r = specific_minor(minor);
2357 if (r < 0)
2358 goto bad_minor;
2359
2360 r = init_srcu_struct(&md->io_barrier);
2361 if (r < 0)
2362 goto bad_io_barrier;
2363
2364 md->numa_node_id = numa_node_id;
2365 md->use_blk_mq = use_blk_mq;
2366 md->init_tio_pdu = false;
2367 md->type = DM_TYPE_NONE;
2368 mutex_init(&md->suspend_lock);
2369 mutex_init(&md->type_lock);
2370 mutex_init(&md->table_devices_lock);
2371 spin_lock_init(&md->deferred_lock);
2372 atomic_set(&md->holders, 1);
2373 atomic_set(&md->open_count, 0);
2374 atomic_set(&md->event_nr, 0);
2375 atomic_set(&md->uevent_seq, 0);
2376 INIT_LIST_HEAD(&md->uevent_list);
2377 INIT_LIST_HEAD(&md->table_devices);
2378 spin_lock_init(&md->uevent_lock);
2379
2380 md->queue = blk_alloc_queue_node(GFP_KERNEL, numa_node_id);
2381 if (!md->queue)
2382 goto bad;
2383
2384 dm_init_md_queue(md);
2385
2386 md->disk = alloc_disk_node(1, numa_node_id);
2387 if (!md->disk)
2388 goto bad;
2389
2390 atomic_set(&md->pending[0], 0);
2391 atomic_set(&md->pending[1], 0);
2392 init_waitqueue_head(&md->wait);
2393 INIT_WORK(&md->work, dm_wq_work);
2394 init_waitqueue_head(&md->eventq);
2395 init_completion(&md->kobj_holder.completion);
2396 md->kworker_task = NULL;
2397
2398 md->disk->major = _major;
2399 md->disk->first_minor = minor;
2400 md->disk->fops = &dm_blk_dops;
2401 md->disk->queue = md->queue;
2402 md->disk->private_data = md;
2403 sprintf(md->disk->disk_name, "dm-%d", minor);
2404 add_disk(md->disk);
2405 format_dev_t(md->name, MKDEV(_major, minor));
2406
2407 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
2408 if (!md->wq)
2409 goto bad;
2410
2411 md->bdev = bdget_disk(md->disk, 0);
2412 if (!md->bdev)
2413 goto bad;
2414
2415 bio_init(&md->flush_bio);
2416 md->flush_bio.bi_bdev = md->bdev;
2417 md->flush_bio.bi_rw = WRITE_FLUSH;
2418
2419 dm_stats_init(&md->stats);
2420
2421 /* Populate the mapping, nobody knows we exist yet */
2422 spin_lock(&_minor_lock);
2423 old_md = idr_replace(&_minor_idr, md, minor);
2424 spin_unlock(&_minor_lock);
2425
2426 BUG_ON(old_md != MINOR_ALLOCED);
2427
2428 return md;
2429
2430bad:
2431 cleanup_mapped_device(md);
2432bad_io_barrier:
2433 free_minor(minor);
2434bad_minor:
2435 module_put(THIS_MODULE);
2436bad_module_get:
2437 kfree(md);
2438 return NULL;
2439}
2440
2441static void unlock_fs(struct mapped_device *md);
2442
2443static void free_dev(struct mapped_device *md)
2444{
2445 int minor = MINOR(disk_devt(md->disk));
2446
2447 unlock_fs(md);
2448
2449 cleanup_mapped_device(md);
2450 if (md->tag_set) {
2451 blk_mq_free_tag_set(md->tag_set);
2452 kfree(md->tag_set);
2453 }
2454
2455 free_table_devices(&md->table_devices);
2456 dm_stats_cleanup(&md->stats);
2457 free_minor(minor);
2458
2459 module_put(THIS_MODULE);
2460 kfree(md);
2461}
2462
2463static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
2464{
2465 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
2466
2467 if (md->bs) {
2468 /* The md already has necessary mempools. */
2469 if (dm_table_get_type(t) == DM_TYPE_BIO_BASED) {
2470 /*
2471 * Reload bioset because front_pad may have changed
2472 * because a different table was loaded.
2473 */
2474 bioset_free(md->bs);
2475 md->bs = p->bs;
2476 p->bs = NULL;
2477 }
2478 /*
2479 * There's no need to reload with request-based dm
2480 * because the size of front_pad doesn't change.
2481 * Note for future: If you are to reload bioset,
2482 * prep-ed requests in the queue may refer
2483 * to bio from the old bioset, so you must walk
2484 * through the queue to unprep.
2485 */
2486 goto out;
2487 }
2488
2489 BUG_ON(!p || md->io_pool || md->rq_pool || md->bs);
2490
2491 md->io_pool = p->io_pool;
2492 p->io_pool = NULL;
2493 md->rq_pool = p->rq_pool;
2494 p->rq_pool = NULL;
2495 md->bs = p->bs;
2496 p->bs = NULL;
2497
2498out:
2499 /* mempool bind completed, no longer need any mempools in the table */
2500 dm_table_free_md_mempools(t);
2501}
2502
2503/*
2504 * Bind a table to the device.
2505 */
2506static void event_callback(void *context)
2507{
2508 unsigned long flags;
2509 LIST_HEAD(uevents);
2510 struct mapped_device *md = (struct mapped_device *) context;
2511
2512 spin_lock_irqsave(&md->uevent_lock, flags);
2513 list_splice_init(&md->uevent_list, &uevents);
2514 spin_unlock_irqrestore(&md->uevent_lock, flags);
2515
2516 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2517
2518 atomic_inc(&md->event_nr);
2519 wake_up(&md->eventq);
2520}
2521
2522/*
2523 * Protected by md->suspend_lock obtained by dm_swap_table().
2524 */
2525static void __set_size(struct mapped_device *md, sector_t size)
2526{
2527 set_capacity(md->disk, size);
2528
2529 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2530}
2531
2532/*
2533 * Returns old map, which caller must destroy.
2534 */
2535static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2536 struct queue_limits *limits)
2537{
2538 struct dm_table *old_map;
2539 struct request_queue *q = md->queue;
2540 sector_t size;
2541
2542 size = dm_table_get_size(t);
2543
2544 /*
2545 * Wipe any geometry if the size of the table changed.
2546 */
2547 if (size != dm_get_size(md))
2548 memset(&md->geometry, 0, sizeof(md->geometry));
2549
2550 __set_size(md, size);
2551
2552 dm_table_event_callback(t, event_callback, md);
2553
2554 /*
2555 * The queue hasn't been stopped yet, if the old table type wasn't
2556 * for request-based during suspension. So stop it to prevent
2557 * I/O mapping before resume.
2558 * This must be done before setting the queue restrictions,
2559 * because request-based dm may be run just after the setting.
2560 */
2561 if (dm_table_request_based(t)) {
2562 dm_stop_queue(q);
2563 /*
2564 * Leverage the fact that request-based DM targets are
2565 * immutable singletons and establish md->immutable_target
2566 * - used to optimize both dm_request_fn and dm_mq_queue_rq
2567 */
2568 md->immutable_target = dm_table_get_immutable_target(t);
2569 }
2570
2571 __bind_mempools(md, t);
2572
2573 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2574 rcu_assign_pointer(md->map, (void *)t);
2575 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2576
2577 dm_table_set_restrictions(t, q, limits);
2578 if (old_map)
2579 dm_sync_table(md);
2580
2581 return old_map;
2582}
2583
2584/*
2585 * Returns unbound table for the caller to free.
2586 */
2587static struct dm_table *__unbind(struct mapped_device *md)
2588{
2589 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2590
2591 if (!map)
2592 return NULL;
2593
2594 dm_table_event_callback(map, NULL, NULL);
2595 RCU_INIT_POINTER(md->map, NULL);
2596 dm_sync_table(md);
2597
2598 return map;
2599}
2600
2601/*
2602 * Constructor for a new device.
2603 */
2604int dm_create(int minor, struct mapped_device **result)
2605{
2606 struct mapped_device *md;
2607
2608 md = alloc_dev(minor);
2609 if (!md)
2610 return -ENXIO;
2611
2612 dm_sysfs_init(md);
2613
2614 *result = md;
2615 return 0;
2616}
2617
2618/*
2619 * Functions to manage md->type.
2620 * All are required to hold md->type_lock.
2621 */
2622void dm_lock_md_type(struct mapped_device *md)
2623{
2624 mutex_lock(&md->type_lock);
2625}
2626
2627void dm_unlock_md_type(struct mapped_device *md)
2628{
2629 mutex_unlock(&md->type_lock);
2630}
2631
2632void dm_set_md_type(struct mapped_device *md, unsigned type)
2633{
2634 BUG_ON(!mutex_is_locked(&md->type_lock));
2635 md->type = type;
2636}
2637
2638unsigned dm_get_md_type(struct mapped_device *md)
2639{
2640 return md->type;
2641}
2642
2643struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2644{
2645 return md->immutable_target_type;
2646}
2647
2648/*
2649 * The queue_limits are only valid as long as you have a reference
2650 * count on 'md'.
2651 */
2652struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2653{
2654 BUG_ON(!atomic_read(&md->holders));
2655 return &md->queue->limits;
2656}
2657EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2658
2659static void dm_old_init_rq_based_worker_thread(struct mapped_device *md)
2660{
2661 /* Initialize the request-based DM worker thread */
2662 init_kthread_worker(&md->kworker);
2663 md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
2664 "kdmwork-%s", dm_device_name(md));
2665}
2666
2667/*
2668 * Fully initialize a .request_fn request-based queue.
2669 */
2670static int dm_old_init_request_queue(struct mapped_device *md)
2671{
2672 /* Fully initialize the queue */
2673 if (!blk_init_allocated_queue(md->queue, dm_request_fn, NULL))
2674 return -EINVAL;
2675
2676 /* disable dm_request_fn's merge heuristic by default */
2677 md->seq_rq_merge_deadline_usecs = 0;
2678
2679 dm_init_normal_md_queue(md);
2680 blk_queue_softirq_done(md->queue, dm_softirq_done);
2681 blk_queue_prep_rq(md->queue, dm_old_prep_fn);
2682
2683 dm_old_init_rq_based_worker_thread(md);
2684
2685 elv_register_queue(md->queue);
2686
2687 return 0;
2688}
2689
2690static int dm_mq_init_request(void *data, struct request *rq,
2691 unsigned int hctx_idx, unsigned int request_idx,
2692 unsigned int numa_node)
2693{
2694 struct mapped_device *md = data;
2695 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2696
2697 /*
2698 * Must initialize md member of tio, otherwise it won't
2699 * be available in dm_mq_queue_rq.
2700 */
2701 tio->md = md;
2702
2703 if (md->init_tio_pdu) {
2704 /* target-specific per-io data is immediately after the tio */
2705 tio->info.ptr = tio + 1;
2706 }
2707
2708 return 0;
2709}
2710
2711static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
2712 const struct blk_mq_queue_data *bd)
2713{
2714 struct request *rq = bd->rq;
2715 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2716 struct mapped_device *md = tio->md;
2717 struct dm_target *ti = md->immutable_target;
2718
2719 if (unlikely(!ti)) {
2720 int srcu_idx;
2721 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2722
2723 ti = dm_table_find_target(map, 0);
2724 dm_put_live_table(md, srcu_idx);
2725 }
2726
2727 if (ti->type->busy && ti->type->busy(ti))
2728 return BLK_MQ_RQ_QUEUE_BUSY;
2729
2730 dm_start_request(md, rq);
2731
2732 /* Init tio using md established in .init_request */
2733 init_tio(tio, rq, md);
2734
2735 /*
2736 * Establish tio->ti before queuing work (map_tio_request)
2737 * or making direct call to map_request().
2738 */
2739 tio->ti = ti;
2740
2741 /* Direct call is fine since .queue_rq allows allocations */
2742 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
2743 /* Undo dm_start_request() before requeuing */
2744 rq_end_stats(md, rq);
2745 rq_completed(md, rq_data_dir(rq), false);
2746 return BLK_MQ_RQ_QUEUE_BUSY;
2747 }
2748
2749 return BLK_MQ_RQ_QUEUE_OK;
2750}
2751
2752static struct blk_mq_ops dm_mq_ops = {
2753 .queue_rq = dm_mq_queue_rq,
2754 .map_queue = blk_mq_map_queue,
2755 .complete = dm_softirq_done,
2756 .init_request = dm_mq_init_request,
2757};
2758
2759static int dm_mq_init_request_queue(struct mapped_device *md,
2760 struct dm_target *immutable_tgt)
2761{
2762 struct request_queue *q;
2763 int err;
2764
2765 if (dm_get_md_type(md) == DM_TYPE_REQUEST_BASED) {
2766 DMERR("request-based dm-mq may only be stacked on blk-mq device(s)");
2767 return -EINVAL;
2768 }
2769
2770 md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
2771 if (!md->tag_set)
2772 return -ENOMEM;
2773
2774 md->tag_set->ops = &dm_mq_ops;
2775 md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
2776 md->tag_set->numa_node = md->numa_node_id;
2777 md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2778 md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
2779 md->tag_set->driver_data = md;
2780
2781 md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
2782 if (immutable_tgt && immutable_tgt->per_io_data_size) {
2783 /* any target-specific per-io data is immediately after the tio */
2784 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
2785 md->init_tio_pdu = true;
2786 }
2787
2788 err = blk_mq_alloc_tag_set(md->tag_set);
2789 if (err)
2790 goto out_kfree_tag_set;
2791
2792 q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
2793 if (IS_ERR(q)) {
2794 err = PTR_ERR(q);
2795 goto out_tag_set;
2796 }
2797 dm_init_md_queue(md);
2798
2799 /* backfill 'mq' sysfs registration normally done in blk_register_queue */
2800 blk_mq_register_disk(md->disk);
2801
2802 return 0;
2803
2804out_tag_set:
2805 blk_mq_free_tag_set(md->tag_set);
2806out_kfree_tag_set:
2807 kfree(md->tag_set);
2808
2809 return err;
2810}
2811
2812static unsigned filter_md_type(unsigned type, struct mapped_device *md)
2813{
2814 if (type == DM_TYPE_BIO_BASED)
2815 return type;
2816
2817 return !md->use_blk_mq ? DM_TYPE_REQUEST_BASED : DM_TYPE_MQ_REQUEST_BASED;
2818}
2819
2820/*
2821 * Setup the DM device's queue based on md's type
2822 */
2823int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2824{
2825 int r;
2826 unsigned md_type = filter_md_type(dm_get_md_type(md), md);
2827
2828 switch (md_type) {
2829 case DM_TYPE_REQUEST_BASED:
2830 r = dm_old_init_request_queue(md);
2831 if (r) {
2832 DMERR("Cannot initialize queue for request-based mapped device");
2833 return r;
2834 }
2835 break;
2836 case DM_TYPE_MQ_REQUEST_BASED:
2837 r = dm_mq_init_request_queue(md, dm_table_get_immutable_target(t));
2838 if (r) {
2839 DMERR("Cannot initialize queue for request-based dm-mq mapped device");
2840 return r;
2841 }
2842 break;
2843 case DM_TYPE_BIO_BASED:
2844 dm_init_normal_md_queue(md);
2845 blk_queue_make_request(md->queue, dm_make_request);
2846 /*
2847 * DM handles splitting bios as needed. Free the bio_split bioset
2848 * since it won't be used (saves 1 process per bio-based DM device).
2849 */
2850 bioset_free(md->queue->bio_split);
2851 md->queue->bio_split = NULL;
2852 break;
2853 }
2854
2855 return 0;
2856}
2857
2858struct mapped_device *dm_get_md(dev_t dev)
2859{
2860 struct mapped_device *md;
2861 unsigned minor = MINOR(dev);
2862
2863 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2864 return NULL;
2865
2866 spin_lock(&_minor_lock);
2867
2868 md = idr_find(&_minor_idr, minor);
2869 if (md) {
2870 if ((md == MINOR_ALLOCED ||
2871 (MINOR(disk_devt(dm_disk(md))) != minor) ||
2872 dm_deleting_md(md) ||
2873 test_bit(DMF_FREEING, &md->flags))) {
2874 md = NULL;
2875 goto out;
2876 }
2877 dm_get(md);
2878 }
2879
2880out:
2881 spin_unlock(&_minor_lock);
2882
2883 return md;
2884}
2885EXPORT_SYMBOL_GPL(dm_get_md);
2886
2887void *dm_get_mdptr(struct mapped_device *md)
2888{
2889 return md->interface_ptr;
2890}
2891
2892void dm_set_mdptr(struct mapped_device *md, void *ptr)
2893{
2894 md->interface_ptr = ptr;
2895}
2896
2897void dm_get(struct mapped_device *md)
2898{
2899 atomic_inc(&md->holders);
2900 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2901}
2902
2903int dm_hold(struct mapped_device *md)
2904{
2905 spin_lock(&_minor_lock);
2906 if (test_bit(DMF_FREEING, &md->flags)) {
2907 spin_unlock(&_minor_lock);
2908 return -EBUSY;
2909 }
2910 dm_get(md);
2911 spin_unlock(&_minor_lock);
2912 return 0;
2913}
2914EXPORT_SYMBOL_GPL(dm_hold);
2915
2916const char *dm_device_name(struct mapped_device *md)
2917{
2918 return md->name;
2919}
2920EXPORT_SYMBOL_GPL(dm_device_name);
2921
2922static void __dm_destroy(struct mapped_device *md, bool wait)
2923{
2924 struct dm_table *map;
2925 int srcu_idx;
2926
2927 might_sleep();
2928
2929 spin_lock(&_minor_lock);
2930 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2931 set_bit(DMF_FREEING, &md->flags);
2932 spin_unlock(&_minor_lock);
2933
2934 if (dm_request_based(md) && md->kworker_task)
2935 flush_kthread_worker(&md->kworker);
2936
2937 /*
2938 * Take suspend_lock so that presuspend and postsuspend methods
2939 * do not race with internal suspend.
2940 */
2941 mutex_lock(&md->suspend_lock);
2942 map = dm_get_live_table(md, &srcu_idx);
2943 if (!dm_suspended_md(md)) {
2944 dm_table_presuspend_targets(map);
2945 dm_table_postsuspend_targets(map);
2946 }
2947 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2948 dm_put_live_table(md, srcu_idx);
2949 mutex_unlock(&md->suspend_lock);
2950
2951 /*
2952 * Rare, but there may be I/O requests still going to complete,
2953 * for example. Wait for all references to disappear.
2954 * No one should increment the reference count of the mapped_device,
2955 * after the mapped_device state becomes DMF_FREEING.
2956 */
2957 if (wait)
2958 while (atomic_read(&md->holders))
2959 msleep(1);
2960 else if (atomic_read(&md->holders))
2961 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2962 dm_device_name(md), atomic_read(&md->holders));
2963
2964 dm_sysfs_exit(md);
2965 dm_table_destroy(__unbind(md));
2966 free_dev(md);
2967}
2968
2969void dm_destroy(struct mapped_device *md)
2970{
2971 __dm_destroy(md, true);
2972}
2973
2974void dm_destroy_immediate(struct mapped_device *md)
2975{
2976 __dm_destroy(md, false);
2977}
2978
2979void dm_put(struct mapped_device *md)
2980{
2981 atomic_dec(&md->holders);
2982}
2983EXPORT_SYMBOL_GPL(dm_put);
2984
2985static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
2986{
2987 int r = 0;
2988 DECLARE_WAITQUEUE(wait, current);
2989
2990 add_wait_queue(&md->wait, &wait);
2991
2992 while (1) {
2993 set_current_state(interruptible);
2994
2995 if (!md_in_flight(md))
2996 break;
2997
2998 if (interruptible == TASK_INTERRUPTIBLE &&
2999 signal_pending(current)) {
3000 r = -EINTR;
3001 break;
3002 }
3003
3004 io_schedule();
3005 }
3006 set_current_state(TASK_RUNNING);
3007
3008 remove_wait_queue(&md->wait, &wait);
3009
3010 return r;
3011}
3012
3013/*
3014 * Process the deferred bios
3015 */
3016static void dm_wq_work(struct work_struct *work)
3017{
3018 struct mapped_device *md = container_of(work, struct mapped_device,
3019 work);
3020 struct bio *c;
3021 int srcu_idx;
3022 struct dm_table *map;
3023
3024 map = dm_get_live_table(md, &srcu_idx);
3025
3026 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
3027 spin_lock_irq(&md->deferred_lock);
3028 c = bio_list_pop(&md->deferred);
3029 spin_unlock_irq(&md->deferred_lock);
3030
3031 if (!c)
3032 break;
3033
3034 if (dm_request_based(md))
3035 generic_make_request(c);
3036 else
3037 __split_and_process_bio(md, map, c);
3038 }
3039
3040 dm_put_live_table(md, srcu_idx);
3041}
3042
3043static void dm_queue_flush(struct mapped_device *md)
3044{
3045 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3046 smp_mb__after_atomic();
3047 queue_work(md->wq, &md->work);
3048}
3049
3050/*
3051 * Swap in a new table, returning the old one for the caller to destroy.
3052 */
3053struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
3054{
3055 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
3056 struct queue_limits limits;
3057 int r;
3058
3059 mutex_lock(&md->suspend_lock);
3060
3061 /* device must be suspended */
3062 if (!dm_suspended_md(md))
3063 goto out;
3064
3065 /*
3066 * If the new table has no data devices, retain the existing limits.
3067 * This helps multipath with queue_if_no_path if all paths disappear,
3068 * then new I/O is queued based on these limits, and then some paths
3069 * reappear.
3070 */
3071 if (dm_table_has_no_data_devices(table)) {
3072 live_map = dm_get_live_table_fast(md);
3073 if (live_map)
3074 limits = md->queue->limits;
3075 dm_put_live_table_fast(md);
3076 }
3077
3078 if (!live_map) {
3079 r = dm_calculate_queue_limits(table, &limits);
3080 if (r) {
3081 map = ERR_PTR(r);
3082 goto out;
3083 }
3084 }
3085
3086 map = __bind(md, table, &limits);
3087
3088out:
3089 mutex_unlock(&md->suspend_lock);
3090 return map;
3091}
3092
3093/*
3094 * Functions to lock and unlock any filesystem running on the
3095 * device.
3096 */
3097static int lock_fs(struct mapped_device *md)
3098{
3099 int r;
3100
3101 WARN_ON(md->frozen_sb);
3102
3103 md->frozen_sb = freeze_bdev(md->bdev);
3104 if (IS_ERR(md->frozen_sb)) {
3105 r = PTR_ERR(md->frozen_sb);
3106 md->frozen_sb = NULL;
3107 return r;
3108 }
3109
3110 set_bit(DMF_FROZEN, &md->flags);
3111
3112 return 0;
3113}
3114
3115static void unlock_fs(struct mapped_device *md)
3116{
3117 if (!test_bit(DMF_FROZEN, &md->flags))
3118 return;
3119
3120 thaw_bdev(md->bdev, md->frozen_sb);
3121 md->frozen_sb = NULL;
3122 clear_bit(DMF_FROZEN, &md->flags);
3123}
3124
3125/*
3126 * If __dm_suspend returns 0, the device is completely quiescent
3127 * now. There is no request-processing activity. All new requests
3128 * are being added to md->deferred list.
3129 *
3130 * Caller must hold md->suspend_lock
3131 */
3132static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
3133 unsigned suspend_flags, int interruptible)
3134{
3135 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
3136 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
3137 int r;
3138
3139 /*
3140 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
3141 * This flag is cleared before dm_suspend returns.
3142 */
3143 if (noflush)
3144 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3145
3146 /*
3147 * This gets reverted if there's an error later and the targets
3148 * provide the .presuspend_undo hook.
3149 */
3150 dm_table_presuspend_targets(map);
3151
3152 /*
3153 * Flush I/O to the device.
3154 * Any I/O submitted after lock_fs() may not be flushed.
3155 * noflush takes precedence over do_lockfs.
3156 * (lock_fs() flushes I/Os and waits for them to complete.)
3157 */
3158 if (!noflush && do_lockfs) {
3159 r = lock_fs(md);
3160 if (r) {
3161 dm_table_presuspend_undo_targets(map);
3162 return r;
3163 }
3164 }
3165
3166 /*
3167 * Here we must make sure that no processes are submitting requests
3168 * to target drivers i.e. no one may be executing
3169 * __split_and_process_bio. This is called from dm_request and
3170 * dm_wq_work.
3171 *
3172 * To get all processes out of __split_and_process_bio in dm_request,
3173 * we take the write lock. To prevent any process from reentering
3174 * __split_and_process_bio from dm_request and quiesce the thread
3175 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
3176 * flush_workqueue(md->wq).
3177 */
3178 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3179 if (map)
3180 synchronize_srcu(&md->io_barrier);
3181
3182 /*
3183 * Stop md->queue before flushing md->wq in case request-based
3184 * dm defers requests to md->wq from md->queue.
3185 */
3186 if (dm_request_based(md)) {
3187 dm_stop_queue(md->queue);
3188 if (md->kworker_task)
3189 flush_kthread_worker(&md->kworker);
3190 }
3191
3192 flush_workqueue(md->wq);
3193
3194 /*
3195 * At this point no more requests are entering target request routines.
3196 * We call dm_wait_for_completion to wait for all existing requests
3197 * to finish.
3198 */
3199 r = dm_wait_for_completion(md, interruptible);
3200
3201 if (noflush)
3202 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3203 if (map)
3204 synchronize_srcu(&md->io_barrier);
3205
3206 /* were we interrupted ? */
3207 if (r < 0) {
3208 dm_queue_flush(md);
3209
3210 if (dm_request_based(md))
3211 dm_start_queue(md->queue);
3212
3213 unlock_fs(md);
3214 dm_table_presuspend_undo_targets(map);
3215 /* pushback list is already flushed, so skip flush */
3216 }
3217
3218 return r;
3219}
3220
3221/*
3222 * We need to be able to change a mapping table under a mounted
3223 * filesystem. For example we might want to move some data in
3224 * the background. Before the table can be swapped with
3225 * dm_bind_table, dm_suspend must be called to flush any in
3226 * flight bios and ensure that any further io gets deferred.
3227 */
3228/*
3229 * Suspend mechanism in request-based dm.
3230 *
3231 * 1. Flush all I/Os by lock_fs() if needed.
3232 * 2. Stop dispatching any I/O by stopping the request_queue.
3233 * 3. Wait for all in-flight I/Os to be completed or requeued.
3234 *
3235 * To abort suspend, start the request_queue.
3236 */
3237int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
3238{
3239 struct dm_table *map = NULL;
3240 int r = 0;
3241
3242retry:
3243 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3244
3245 if (dm_suspended_md(md)) {
3246 r = -EINVAL;
3247 goto out_unlock;
3248 }
3249
3250 if (dm_suspended_internally_md(md)) {
3251 /* already internally suspended, wait for internal resume */
3252 mutex_unlock(&md->suspend_lock);
3253 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3254 if (r)
3255 return r;
3256 goto retry;
3257 }
3258
3259 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3260
3261 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE);
3262 if (r)
3263 goto out_unlock;
3264
3265 set_bit(DMF_SUSPENDED, &md->flags);
3266
3267 dm_table_postsuspend_targets(map);
3268
3269out_unlock:
3270 mutex_unlock(&md->suspend_lock);
3271 return r;
3272}
3273
3274static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3275{
3276 if (map) {
3277 int r = dm_table_resume_targets(map);
3278 if (r)
3279 return r;
3280 }
3281
3282 dm_queue_flush(md);
3283
3284 /*
3285 * Flushing deferred I/Os must be done after targets are resumed
3286 * so that mapping of targets can work correctly.
3287 * Request-based dm is queueing the deferred I/Os in its request_queue.
3288 */
3289 if (dm_request_based(md))
3290 dm_start_queue(md->queue);
3291
3292 unlock_fs(md);
3293
3294 return 0;
3295}
3296
3297int dm_resume(struct mapped_device *md)
3298{
3299 int r = -EINVAL;
3300 struct dm_table *map = NULL;
3301
3302retry:
3303 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3304
3305 if (!dm_suspended_md(md))
3306 goto out;
3307
3308 if (dm_suspended_internally_md(md)) {
3309 /* already internally suspended, wait for internal resume */
3310 mutex_unlock(&md->suspend_lock);
3311 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3312 if (r)
3313 return r;
3314 goto retry;
3315 }
3316
3317 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3318 if (!map || !dm_table_get_size(map))
3319 goto out;
3320
3321 r = __dm_resume(md, map);
3322 if (r)
3323 goto out;
3324
3325 clear_bit(DMF_SUSPENDED, &md->flags);
3326
3327 r = 0;
3328out:
3329 mutex_unlock(&md->suspend_lock);
3330
3331 return r;
3332}
3333
3334/*
3335 * Internal suspend/resume works like userspace-driven suspend. It waits
3336 * until all bios finish and prevents issuing new bios to the target drivers.
3337 * It may be used only from the kernel.
3338 */
3339
3340static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
3341{
3342 struct dm_table *map = NULL;
3343
3344 if (md->internal_suspend_count++)
3345 return; /* nested internal suspend */
3346
3347 if (dm_suspended_md(md)) {
3348 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3349 return; /* nest suspend */
3350 }
3351
3352 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3353
3354 /*
3355 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3356 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3357 * would require changing .presuspend to return an error -- avoid this
3358 * until there is a need for more elaborate variants of internal suspend.
3359 */
3360 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE);
3361
3362 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3363
3364 dm_table_postsuspend_targets(map);
3365}
3366
3367static void __dm_internal_resume(struct mapped_device *md)
3368{
3369 BUG_ON(!md->internal_suspend_count);
3370
3371 if (--md->internal_suspend_count)
3372 return; /* resume from nested internal suspend */
3373
3374 if (dm_suspended_md(md))
3375 goto done; /* resume from nested suspend */
3376
3377 /*
3378 * NOTE: existing callers don't need to call dm_table_resume_targets
3379 * (which may fail -- so best to avoid it for now by passing NULL map)
3380 */
3381 (void) __dm_resume(md, NULL);
3382
3383done:
3384 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3385 smp_mb__after_atomic();
3386 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3387}
3388
3389void dm_internal_suspend_noflush(struct mapped_device *md)
3390{
3391 mutex_lock(&md->suspend_lock);
3392 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3393 mutex_unlock(&md->suspend_lock);
3394}
3395EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3396
3397void dm_internal_resume(struct mapped_device *md)
3398{
3399 mutex_lock(&md->suspend_lock);
3400 __dm_internal_resume(md);
3401 mutex_unlock(&md->suspend_lock);
3402}
3403EXPORT_SYMBOL_GPL(dm_internal_resume);
3404
3405/*
3406 * Fast variants of internal suspend/resume hold md->suspend_lock,
3407 * which prevents interaction with userspace-driven suspend.
3408 */
3409
3410void dm_internal_suspend_fast(struct mapped_device *md)
3411{
3412 mutex_lock(&md->suspend_lock);
3413 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3414 return;
3415
3416 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3417 synchronize_srcu(&md->io_barrier);
3418 flush_workqueue(md->wq);
3419 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3420}
3421EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3422
3423void dm_internal_resume_fast(struct mapped_device *md)
3424{
3425 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3426 goto done;
3427
3428 dm_queue_flush(md);
3429
3430done:
3431 mutex_unlock(&md->suspend_lock);
3432}
3433EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3434
3435/*-----------------------------------------------------------------
3436 * Event notification.
3437 *---------------------------------------------------------------*/
3438int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3439 unsigned cookie)
3440{
3441 char udev_cookie[DM_COOKIE_LENGTH];
3442 char *envp[] = { udev_cookie, NULL };
3443
3444 if (!cookie)
3445 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
3446 else {
3447 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3448 DM_COOKIE_ENV_VAR_NAME, cookie);
3449 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
3450 action, envp);
3451 }
3452}
3453
3454uint32_t dm_next_uevent_seq(struct mapped_device *md)
3455{
3456 return atomic_add_return(1, &md->uevent_seq);
3457}
3458
3459uint32_t dm_get_event_nr(struct mapped_device *md)
3460{
3461 return atomic_read(&md->event_nr);
3462}
3463
3464int dm_wait_event(struct mapped_device *md, int event_nr)
3465{
3466 return wait_event_interruptible(md->eventq,
3467 (event_nr != atomic_read(&md->event_nr)));
3468}
3469
3470void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3471{
3472 unsigned long flags;
3473
3474 spin_lock_irqsave(&md->uevent_lock, flags);
3475 list_add(elist, &md->uevent_list);
3476 spin_unlock_irqrestore(&md->uevent_lock, flags);
3477}
3478
3479/*
3480 * The gendisk is only valid as long as you have a reference
3481 * count on 'md'.
3482 */
3483struct gendisk *dm_disk(struct mapped_device *md)
3484{
3485 return md->disk;
3486}
3487EXPORT_SYMBOL_GPL(dm_disk);
3488
3489struct kobject *dm_kobject(struct mapped_device *md)
3490{
3491 return &md->kobj_holder.kobj;
3492}
3493
3494struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3495{
3496 struct mapped_device *md;
3497
3498 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3499
3500 if (test_bit(DMF_FREEING, &md->flags) ||
3501 dm_deleting_md(md))
3502 return NULL;
3503
3504 dm_get(md);
3505 return md;
3506}
3507
3508int dm_suspended_md(struct mapped_device *md)
3509{
3510 return test_bit(DMF_SUSPENDED, &md->flags);
3511}
3512
3513int dm_suspended_internally_md(struct mapped_device *md)
3514{
3515 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3516}
3517
3518int dm_test_deferred_remove_flag(struct mapped_device *md)
3519{
3520 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3521}
3522
3523int dm_suspended(struct dm_target *ti)
3524{
3525 return dm_suspended_md(dm_table_get_md(ti->table));
3526}
3527EXPORT_SYMBOL_GPL(dm_suspended);
3528
3529int dm_noflush_suspending(struct dm_target *ti)
3530{
3531 return __noflush_suspending(dm_table_get_md(ti->table));
3532}
3533EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3534
3535struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
3536 unsigned integrity, unsigned per_io_data_size)
3537{
3538 struct dm_md_mempools *pools = kzalloc_node(sizeof(*pools), GFP_KERNEL, md->numa_node_id);
3539 struct kmem_cache *cachep = NULL;
3540 unsigned int pool_size = 0;
3541 unsigned int front_pad;
3542
3543 if (!pools)
3544 return NULL;
3545
3546 type = filter_md_type(type, md);
3547
3548 switch (type) {
3549 case DM_TYPE_BIO_BASED:
3550 cachep = _io_cache;
3551 pool_size = dm_get_reserved_bio_based_ios();
3552 front_pad = roundup(per_io_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
3553 break;
3554 case DM_TYPE_REQUEST_BASED:
3555 cachep = _rq_tio_cache;
3556 pool_size = dm_get_reserved_rq_based_ios();
3557 pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
3558 if (!pools->rq_pool)
3559 goto out;
3560 /* fall through to setup remaining rq-based pools */
3561 case DM_TYPE_MQ_REQUEST_BASED:
3562 if (!pool_size)
3563 pool_size = dm_get_reserved_rq_based_ios();
3564 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
3565 /* per_io_data_size is used for blk-mq pdu at queue allocation */
3566 break;
3567 default:
3568 BUG();
3569 }
3570
3571 if (cachep) {
3572 pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
3573 if (!pools->io_pool)
3574 goto out;
3575 }
3576
3577 pools->bs = bioset_create_nobvec(pool_size, front_pad);
3578 if (!pools->bs)
3579 goto out;
3580
3581 if (integrity && bioset_integrity_create(pools->bs, pool_size))
3582 goto out;
3583
3584 return pools;
3585
3586out:
3587 dm_free_md_mempools(pools);
3588
3589 return NULL;
3590}
3591
3592void dm_free_md_mempools(struct dm_md_mempools *pools)
3593{
3594 if (!pools)
3595 return;
3596
3597 mempool_destroy(pools->io_pool);
3598 mempool_destroy(pools->rq_pool);
3599
3600 if (pools->bs)
3601 bioset_free(pools->bs);
3602
3603 kfree(pools);
3604}
3605
3606static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3607 u32 flags)
3608{
3609 struct mapped_device *md = bdev->bd_disk->private_data;
3610 const struct pr_ops *ops;
3611 fmode_t mode;
3612 int r;
3613
3614 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3615 if (r < 0)
3616 return r;
3617
3618 ops = bdev->bd_disk->fops->pr_ops;
3619 if (ops && ops->pr_register)
3620 r = ops->pr_register(bdev, old_key, new_key, flags);
3621 else
3622 r = -EOPNOTSUPP;
3623
3624 bdput(bdev);
3625 return r;
3626}
3627
3628static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3629 u32 flags)
3630{
3631 struct mapped_device *md = bdev->bd_disk->private_data;
3632 const struct pr_ops *ops;
3633 fmode_t mode;
3634 int r;
3635
3636 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3637 if (r < 0)
3638 return r;
3639
3640 ops = bdev->bd_disk->fops->pr_ops;
3641 if (ops && ops->pr_reserve)
3642 r = ops->pr_reserve(bdev, key, type, flags);
3643 else
3644 r = -EOPNOTSUPP;
3645
3646 bdput(bdev);
3647 return r;
3648}
3649
3650static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3651{
3652 struct mapped_device *md = bdev->bd_disk->private_data;
3653 const struct pr_ops *ops;
3654 fmode_t mode;
3655 int r;
3656
3657 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3658 if (r < 0)
3659 return r;
3660
3661 ops = bdev->bd_disk->fops->pr_ops;
3662 if (ops && ops->pr_release)
3663 r = ops->pr_release(bdev, key, type);
3664 else
3665 r = -EOPNOTSUPP;
3666
3667 bdput(bdev);
3668 return r;
3669}
3670
3671static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3672 enum pr_type type, bool abort)
3673{
3674 struct mapped_device *md = bdev->bd_disk->private_data;
3675 const struct pr_ops *ops;
3676 fmode_t mode;
3677 int r;
3678
3679 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3680 if (r < 0)
3681 return r;
3682
3683 ops = bdev->bd_disk->fops->pr_ops;
3684 if (ops && ops->pr_preempt)
3685 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
3686 else
3687 r = -EOPNOTSUPP;
3688
3689 bdput(bdev);
3690 return r;
3691}
3692
3693static int dm_pr_clear(struct block_device *bdev, u64 key)
3694{
3695 struct mapped_device *md = bdev->bd_disk->private_data;
3696 const struct pr_ops *ops;
3697 fmode_t mode;
3698 int r;
3699
3700 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3701 if (r < 0)
3702 return r;
3703
3704 ops = bdev->bd_disk->fops->pr_ops;
3705 if (ops && ops->pr_clear)
3706 r = ops->pr_clear(bdev, key);
3707 else
3708 r = -EOPNOTSUPP;
3709
3710 bdput(bdev);
3711 return r;
3712}
3713
3714static const struct pr_ops dm_pr_ops = {
3715 .pr_register = dm_pr_register,
3716 .pr_reserve = dm_pr_reserve,
3717 .pr_release = dm_pr_release,
3718 .pr_preempt = dm_pr_preempt,
3719 .pr_clear = dm_pr_clear,
3720};
3721
3722static const struct block_device_operations dm_blk_dops = {
3723 .open = dm_blk_open,
3724 .release = dm_blk_close,
3725 .ioctl = dm_blk_ioctl,
3726 .getgeo = dm_blk_getgeo,
3727 .pr_ops = &dm_pr_ops,
3728 .owner = THIS_MODULE
3729};
3730
3731/*
3732 * module hooks
3733 */
3734module_init(dm_init);
3735module_exit(dm_exit);
3736
3737module_param(major, uint, 0);
3738MODULE_PARM_DESC(major, "The major number of the device mapper");
3739
3740module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3741MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3742
3743module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
3744MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
3745
3746module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
3747MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
3748
3749module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
3750MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
3751
3752module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
3753MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");
3754
3755module_param(dm_numa_node, int, S_IRUGO | S_IWUSR);
3756MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3757
3758MODULE_DESCRIPTION(DM_NAME " driver");
3759MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3760MODULE_LICENSE("GPL");