Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
4 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
5 *
6 * This file is released under the GPL.
7 */
8
9#include "dm-core.h"
10#include "dm-rq.h"
11#include "dm-uevent.h"
12#include "dm-ima.h"
13
14#include <linux/bio-integrity.h>
15#include <linux/init.h>
16#include <linux/module.h>
17#include <linux/mutex.h>
18#include <linux/sched/mm.h>
19#include <linux/sched/signal.h>
20#include <linux/blkpg.h>
21#include <linux/bio.h>
22#include <linux/mempool.h>
23#include <linux/dax.h>
24#include <linux/slab.h>
25#include <linux/idr.h>
26#include <linux/uio.h>
27#include <linux/hdreg.h>
28#include <linux/delay.h>
29#include <linux/wait.h>
30#include <linux/pr.h>
31#include <linux/refcount.h>
32#include <linux/part_stat.h>
33#include <linux/blk-crypto.h>
34#include <linux/blk-crypto-profile.h>
35
36#define DM_MSG_PREFIX "core"
37
38/*
39 * Cookies are numeric values sent with CHANGE and REMOVE
40 * uevents while resuming, removing or renaming the device.
41 */
42#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
43#define DM_COOKIE_LENGTH 24
44
45/*
46 * For REQ_POLLED fs bio, this flag is set if we link mapped underlying
47 * dm_io into one list, and reuse bio->bi_private as the list head. Before
48 * ending this fs bio, we will recover its ->bi_private.
49 */
50#define REQ_DM_POLL_LIST REQ_DRV
51
52static const char *_name = DM_NAME;
53
54static unsigned int major;
55static unsigned int _major;
56
57static DEFINE_IDR(_minor_idr);
58
59static DEFINE_SPINLOCK(_minor_lock);
60
61static void do_deferred_remove(struct work_struct *w);
62
63static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
64
65static struct workqueue_struct *deferred_remove_workqueue;
66
67atomic_t dm_global_event_nr = ATOMIC_INIT(0);
68DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
69
70void dm_issue_global_event(void)
71{
72 atomic_inc(&dm_global_event_nr);
73 wake_up(&dm_global_eventq);
74}
75
76DEFINE_STATIC_KEY_FALSE(stats_enabled);
77DEFINE_STATIC_KEY_FALSE(swap_bios_enabled);
78DEFINE_STATIC_KEY_FALSE(zoned_enabled);
79
80/*
81 * One of these is allocated (on-stack) per original bio.
82 */
83struct clone_info {
84 struct dm_table *map;
85 struct bio *bio;
86 struct dm_io *io;
87 sector_t sector;
88 unsigned int sector_count;
89 bool is_abnormal_io:1;
90 bool submit_as_polled:1;
91};
92
93static inline struct dm_target_io *clone_to_tio(struct bio *clone)
94{
95 return container_of(clone, struct dm_target_io, clone);
96}
97
98void *dm_per_bio_data(struct bio *bio, size_t data_size)
99{
100 if (!dm_tio_flagged(clone_to_tio(bio), DM_TIO_INSIDE_DM_IO))
101 return (char *)bio - DM_TARGET_IO_BIO_OFFSET - data_size;
102 return (char *)bio - DM_IO_BIO_OFFSET - data_size;
103}
104EXPORT_SYMBOL_GPL(dm_per_bio_data);
105
106struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
107{
108 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
109
110 if (io->magic == DM_IO_MAGIC)
111 return (struct bio *)((char *)io + DM_IO_BIO_OFFSET);
112 BUG_ON(io->magic != DM_TIO_MAGIC);
113 return (struct bio *)((char *)io + DM_TARGET_IO_BIO_OFFSET);
114}
115EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
116
117unsigned int dm_bio_get_target_bio_nr(const struct bio *bio)
118{
119 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
120}
121EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
122
123#define MINOR_ALLOCED ((void *)-1)
124
125#define DM_NUMA_NODE NUMA_NO_NODE
126static int dm_numa_node = DM_NUMA_NODE;
127
128#define DEFAULT_SWAP_BIOS (8 * 1048576 / PAGE_SIZE)
129static int swap_bios = DEFAULT_SWAP_BIOS;
130static int get_swap_bios(void)
131{
132 int latch = READ_ONCE(swap_bios);
133
134 if (unlikely(latch <= 0))
135 latch = DEFAULT_SWAP_BIOS;
136 return latch;
137}
138
139struct table_device {
140 struct list_head list;
141 refcount_t count;
142 struct dm_dev dm_dev;
143};
144
145/*
146 * Bio-based DM's mempools' reserved IOs set by the user.
147 */
148#define RESERVED_BIO_BASED_IOS 16
149static unsigned int reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
150
151static int __dm_get_module_param_int(int *module_param, int min, int max)
152{
153 int param = READ_ONCE(*module_param);
154 int modified_param = 0;
155 bool modified = true;
156
157 if (param < min)
158 modified_param = min;
159 else if (param > max)
160 modified_param = max;
161 else
162 modified = false;
163
164 if (modified) {
165 (void)cmpxchg(module_param, param, modified_param);
166 param = modified_param;
167 }
168
169 return param;
170}
171
172unsigned int __dm_get_module_param(unsigned int *module_param, unsigned int def, unsigned int max)
173{
174 unsigned int param = READ_ONCE(*module_param);
175 unsigned int modified_param = 0;
176
177 if (!param)
178 modified_param = def;
179 else if (param > max)
180 modified_param = max;
181
182 if (modified_param) {
183 (void)cmpxchg(module_param, param, modified_param);
184 param = modified_param;
185 }
186
187 return param;
188}
189
190unsigned int dm_get_reserved_bio_based_ios(void)
191{
192 return __dm_get_module_param(&reserved_bio_based_ios,
193 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
194}
195EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
196
197static unsigned int dm_get_numa_node(void)
198{
199 return __dm_get_module_param_int(&dm_numa_node,
200 DM_NUMA_NODE, num_online_nodes() - 1);
201}
202
203static int __init local_init(void)
204{
205 int r;
206
207 r = dm_uevent_init();
208 if (r)
209 return r;
210
211 deferred_remove_workqueue = alloc_ordered_workqueue("kdmremove", 0);
212 if (!deferred_remove_workqueue) {
213 r = -ENOMEM;
214 goto out_uevent_exit;
215 }
216
217 _major = major;
218 r = register_blkdev(_major, _name);
219 if (r < 0)
220 goto out_free_workqueue;
221
222 if (!_major)
223 _major = r;
224
225 return 0;
226
227out_free_workqueue:
228 destroy_workqueue(deferred_remove_workqueue);
229out_uevent_exit:
230 dm_uevent_exit();
231
232 return r;
233}
234
235static void local_exit(void)
236{
237 destroy_workqueue(deferred_remove_workqueue);
238
239 unregister_blkdev(_major, _name);
240 dm_uevent_exit();
241
242 _major = 0;
243
244 DMINFO("cleaned up");
245}
246
247static int (*_inits[])(void) __initdata = {
248 local_init,
249 dm_target_init,
250 dm_linear_init,
251 dm_stripe_init,
252 dm_io_init,
253 dm_kcopyd_init,
254 dm_interface_init,
255 dm_statistics_init,
256};
257
258static void (*_exits[])(void) = {
259 local_exit,
260 dm_target_exit,
261 dm_linear_exit,
262 dm_stripe_exit,
263 dm_io_exit,
264 dm_kcopyd_exit,
265 dm_interface_exit,
266 dm_statistics_exit,
267};
268
269static int __init dm_init(void)
270{
271 const int count = ARRAY_SIZE(_inits);
272 int r, i;
273
274#if (IS_ENABLED(CONFIG_IMA) && !IS_ENABLED(CONFIG_IMA_DISABLE_HTABLE))
275 DMWARN("CONFIG_IMA_DISABLE_HTABLE is disabled."
276 " Duplicate IMA measurements will not be recorded in the IMA log.");
277#endif
278
279 for (i = 0; i < count; i++) {
280 r = _inits[i]();
281 if (r)
282 goto bad;
283 }
284
285 return 0;
286bad:
287 while (i--)
288 _exits[i]();
289
290 return r;
291}
292
293static void __exit dm_exit(void)
294{
295 int i = ARRAY_SIZE(_exits);
296
297 while (i--)
298 _exits[i]();
299
300 /*
301 * Should be empty by this point.
302 */
303 idr_destroy(&_minor_idr);
304}
305
306/*
307 * Block device functions
308 */
309int dm_deleting_md(struct mapped_device *md)
310{
311 return test_bit(DMF_DELETING, &md->flags);
312}
313
314static int dm_blk_open(struct gendisk *disk, blk_mode_t mode)
315{
316 struct mapped_device *md;
317
318 spin_lock(&_minor_lock);
319
320 md = disk->private_data;
321 if (!md)
322 goto out;
323
324 if (test_bit(DMF_FREEING, &md->flags) ||
325 dm_deleting_md(md)) {
326 md = NULL;
327 goto out;
328 }
329
330 dm_get(md);
331 atomic_inc(&md->open_count);
332out:
333 spin_unlock(&_minor_lock);
334
335 return md ? 0 : -ENXIO;
336}
337
338static void dm_blk_close(struct gendisk *disk)
339{
340 struct mapped_device *md;
341
342 spin_lock(&_minor_lock);
343
344 md = disk->private_data;
345 if (WARN_ON(!md))
346 goto out;
347
348 if (atomic_dec_and_test(&md->open_count) &&
349 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
350 queue_work(deferred_remove_workqueue, &deferred_remove_work);
351
352 dm_put(md);
353out:
354 spin_unlock(&_minor_lock);
355}
356
357int dm_open_count(struct mapped_device *md)
358{
359 return atomic_read(&md->open_count);
360}
361
362/*
363 * Guarantees nothing is using the device before it's deleted.
364 */
365int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
366{
367 int r = 0;
368
369 spin_lock(&_minor_lock);
370
371 if (dm_open_count(md)) {
372 r = -EBUSY;
373 if (mark_deferred)
374 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
375 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
376 r = -EEXIST;
377 else
378 set_bit(DMF_DELETING, &md->flags);
379
380 spin_unlock(&_minor_lock);
381
382 return r;
383}
384
385int dm_cancel_deferred_remove(struct mapped_device *md)
386{
387 int r = 0;
388
389 spin_lock(&_minor_lock);
390
391 if (test_bit(DMF_DELETING, &md->flags))
392 r = -EBUSY;
393 else
394 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
395
396 spin_unlock(&_minor_lock);
397
398 return r;
399}
400
401static void do_deferred_remove(struct work_struct *w)
402{
403 dm_deferred_remove();
404}
405
406static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
407{
408 struct mapped_device *md = bdev->bd_disk->private_data;
409
410 return dm_get_geometry(md, geo);
411}
412
413static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
414 struct block_device **bdev)
415{
416 struct dm_target *ti;
417 struct dm_table *map;
418 int r;
419
420retry:
421 r = -ENOTTY;
422 map = dm_get_live_table(md, srcu_idx);
423 if (!map || !dm_table_get_size(map))
424 return r;
425
426 /* We only support devices that have a single target */
427 if (map->num_targets != 1)
428 return r;
429
430 ti = dm_table_get_target(map, 0);
431 if (!ti->type->prepare_ioctl)
432 return r;
433
434 if (dm_suspended_md(md))
435 return -EAGAIN;
436
437 r = ti->type->prepare_ioctl(ti, bdev);
438 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
439 dm_put_live_table(md, *srcu_idx);
440 fsleep(10000);
441 goto retry;
442 }
443
444 return r;
445}
446
447static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
448{
449 dm_put_live_table(md, srcu_idx);
450}
451
452static int dm_blk_ioctl(struct block_device *bdev, blk_mode_t mode,
453 unsigned int cmd, unsigned long arg)
454{
455 struct mapped_device *md = bdev->bd_disk->private_data;
456 int r, srcu_idx;
457
458 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
459 if (r < 0)
460 goto out;
461
462 if (r > 0) {
463 /*
464 * Target determined this ioctl is being issued against a
465 * subset of the parent bdev; require extra privileges.
466 */
467 if (!capable(CAP_SYS_RAWIO)) {
468 DMDEBUG_LIMIT(
469 "%s: sending ioctl %x to DM device without required privilege.",
470 current->comm, cmd);
471 r = -ENOIOCTLCMD;
472 goto out;
473 }
474 }
475
476 if (!bdev->bd_disk->fops->ioctl)
477 r = -ENOTTY;
478 else
479 r = bdev->bd_disk->fops->ioctl(bdev, mode, cmd, arg);
480out:
481 dm_unprepare_ioctl(md, srcu_idx);
482 return r;
483}
484
485u64 dm_start_time_ns_from_clone(struct bio *bio)
486{
487 return jiffies_to_nsecs(clone_to_tio(bio)->io->start_time);
488}
489EXPORT_SYMBOL_GPL(dm_start_time_ns_from_clone);
490
491static inline bool bio_is_flush_with_data(struct bio *bio)
492{
493 return ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size);
494}
495
496static inline unsigned int dm_io_sectors(struct dm_io *io, struct bio *bio)
497{
498 /*
499 * If REQ_PREFLUSH set, don't account payload, it will be
500 * submitted (and accounted) after this flush completes.
501 */
502 if (bio_is_flush_with_data(bio))
503 return 0;
504 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
505 return io->sectors;
506 return bio_sectors(bio);
507}
508
509static void dm_io_acct(struct dm_io *io, bool end)
510{
511 struct bio *bio = io->orig_bio;
512
513 if (dm_io_flagged(io, DM_IO_BLK_STAT)) {
514 if (!end)
515 bdev_start_io_acct(bio->bi_bdev, bio_op(bio),
516 io->start_time);
517 else
518 bdev_end_io_acct(bio->bi_bdev, bio_op(bio),
519 dm_io_sectors(io, bio),
520 io->start_time);
521 }
522
523 if (static_branch_unlikely(&stats_enabled) &&
524 unlikely(dm_stats_used(&io->md->stats))) {
525 sector_t sector;
526
527 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
528 sector = bio_end_sector(bio) - io->sector_offset;
529 else
530 sector = bio->bi_iter.bi_sector;
531
532 dm_stats_account_io(&io->md->stats, bio_data_dir(bio),
533 sector, dm_io_sectors(io, bio),
534 end, io->start_time, &io->stats_aux);
535 }
536}
537
538static void __dm_start_io_acct(struct dm_io *io)
539{
540 dm_io_acct(io, false);
541}
542
543static void dm_start_io_acct(struct dm_io *io, struct bio *clone)
544{
545 /*
546 * Ensure IO accounting is only ever started once.
547 */
548 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
549 return;
550
551 /* Expect no possibility for race unless DM_TIO_IS_DUPLICATE_BIO. */
552 if (!clone || likely(dm_tio_is_normal(clone_to_tio(clone)))) {
553 dm_io_set_flag(io, DM_IO_ACCOUNTED);
554 } else {
555 unsigned long flags;
556 /* Can afford locking given DM_TIO_IS_DUPLICATE_BIO */
557 spin_lock_irqsave(&io->lock, flags);
558 if (dm_io_flagged(io, DM_IO_ACCOUNTED)) {
559 spin_unlock_irqrestore(&io->lock, flags);
560 return;
561 }
562 dm_io_set_flag(io, DM_IO_ACCOUNTED);
563 spin_unlock_irqrestore(&io->lock, flags);
564 }
565
566 __dm_start_io_acct(io);
567}
568
569static void dm_end_io_acct(struct dm_io *io)
570{
571 dm_io_acct(io, true);
572}
573
574static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio, gfp_t gfp_mask)
575{
576 struct dm_io *io;
577 struct dm_target_io *tio;
578 struct bio *clone;
579
580 clone = bio_alloc_clone(NULL, bio, gfp_mask, &md->mempools->io_bs);
581 if (unlikely(!clone))
582 return NULL;
583 tio = clone_to_tio(clone);
584 tio->flags = 0;
585 dm_tio_set_flag(tio, DM_TIO_INSIDE_DM_IO);
586 tio->io = NULL;
587
588 io = container_of(tio, struct dm_io, tio);
589 io->magic = DM_IO_MAGIC;
590 io->status = BLK_STS_OK;
591
592 /* one ref is for submission, the other is for completion */
593 atomic_set(&io->io_count, 2);
594 this_cpu_inc(*md->pending_io);
595 io->orig_bio = bio;
596 io->md = md;
597 spin_lock_init(&io->lock);
598 io->start_time = jiffies;
599 io->flags = 0;
600 if (blk_queue_io_stat(md->queue))
601 dm_io_set_flag(io, DM_IO_BLK_STAT);
602
603 if (static_branch_unlikely(&stats_enabled) &&
604 unlikely(dm_stats_used(&md->stats)))
605 dm_stats_record_start(&md->stats, &io->stats_aux);
606
607 return io;
608}
609
610static void free_io(struct dm_io *io)
611{
612 bio_put(&io->tio.clone);
613}
614
615static struct bio *alloc_tio(struct clone_info *ci, struct dm_target *ti,
616 unsigned int target_bio_nr, unsigned int *len, gfp_t gfp_mask)
617{
618 struct mapped_device *md = ci->io->md;
619 struct dm_target_io *tio;
620 struct bio *clone;
621
622 if (!ci->io->tio.io) {
623 /* the dm_target_io embedded in ci->io is available */
624 tio = &ci->io->tio;
625 /* alloc_io() already initialized embedded clone */
626 clone = &tio->clone;
627 } else {
628 clone = bio_alloc_clone(NULL, ci->bio, gfp_mask,
629 &md->mempools->bs);
630 if (!clone)
631 return NULL;
632
633 /* REQ_DM_POLL_LIST shouldn't be inherited */
634 clone->bi_opf &= ~REQ_DM_POLL_LIST;
635
636 tio = clone_to_tio(clone);
637 tio->flags = 0; /* also clears DM_TIO_INSIDE_DM_IO */
638 }
639
640 tio->magic = DM_TIO_MAGIC;
641 tio->io = ci->io;
642 tio->ti = ti;
643 tio->target_bio_nr = target_bio_nr;
644 tio->len_ptr = len;
645 tio->old_sector = 0;
646
647 /* Set default bdev, but target must bio_set_dev() before issuing IO */
648 clone->bi_bdev = md->disk->part0;
649 if (likely(ti != NULL) && unlikely(ti->needs_bio_set_dev))
650 bio_set_dev(clone, md->disk->part0);
651
652 if (len) {
653 clone->bi_iter.bi_size = to_bytes(*len);
654 if (bio_integrity(clone))
655 bio_integrity_trim(clone);
656 }
657
658 return clone;
659}
660
661static void free_tio(struct bio *clone)
662{
663 if (dm_tio_flagged(clone_to_tio(clone), DM_TIO_INSIDE_DM_IO))
664 return;
665 bio_put(clone);
666}
667
668/*
669 * Add the bio to the list of deferred io.
670 */
671static void queue_io(struct mapped_device *md, struct bio *bio)
672{
673 unsigned long flags;
674
675 spin_lock_irqsave(&md->deferred_lock, flags);
676 bio_list_add(&md->deferred, bio);
677 spin_unlock_irqrestore(&md->deferred_lock, flags);
678 queue_work(md->wq, &md->work);
679}
680
681/*
682 * Everyone (including functions in this file), should use this
683 * function to access the md->map field, and make sure they call
684 * dm_put_live_table() when finished.
685 */
686struct dm_table *dm_get_live_table(struct mapped_device *md,
687 int *srcu_idx) __acquires(md->io_barrier)
688{
689 *srcu_idx = srcu_read_lock(&md->io_barrier);
690
691 return srcu_dereference(md->map, &md->io_barrier);
692}
693
694void dm_put_live_table(struct mapped_device *md,
695 int srcu_idx) __releases(md->io_barrier)
696{
697 srcu_read_unlock(&md->io_barrier, srcu_idx);
698}
699
700void dm_sync_table(struct mapped_device *md)
701{
702 synchronize_srcu(&md->io_barrier);
703 synchronize_rcu_expedited();
704}
705
706/*
707 * A fast alternative to dm_get_live_table/dm_put_live_table.
708 * The caller must not block between these two functions.
709 */
710static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
711{
712 rcu_read_lock();
713 return rcu_dereference(md->map);
714}
715
716static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
717{
718 rcu_read_unlock();
719}
720
721static char *_dm_claim_ptr = "I belong to device-mapper";
722
723/*
724 * Open a table device so we can use it as a map destination.
725 */
726static struct table_device *open_table_device(struct mapped_device *md,
727 dev_t dev, blk_mode_t mode)
728{
729 struct table_device *td;
730 struct file *bdev_file;
731 struct block_device *bdev;
732 u64 part_off;
733 int r;
734
735 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
736 if (!td)
737 return ERR_PTR(-ENOMEM);
738 refcount_set(&td->count, 1);
739
740 bdev_file = bdev_file_open_by_dev(dev, mode, _dm_claim_ptr, NULL);
741 if (IS_ERR(bdev_file)) {
742 r = PTR_ERR(bdev_file);
743 goto out_free_td;
744 }
745
746 bdev = file_bdev(bdev_file);
747
748 /*
749 * We can be called before the dm disk is added. In that case we can't
750 * register the holder relation here. It will be done once add_disk was
751 * called.
752 */
753 if (md->disk->slave_dir) {
754 r = bd_link_disk_holder(bdev, md->disk);
755 if (r)
756 goto out_blkdev_put;
757 }
758
759 td->dm_dev.mode = mode;
760 td->dm_dev.bdev = bdev;
761 td->dm_dev.bdev_file = bdev_file;
762 td->dm_dev.dax_dev = fs_dax_get_by_bdev(bdev, &part_off,
763 NULL, NULL);
764 format_dev_t(td->dm_dev.name, dev);
765 list_add(&td->list, &md->table_devices);
766 return td;
767
768out_blkdev_put:
769 __fput_sync(bdev_file);
770out_free_td:
771 kfree(td);
772 return ERR_PTR(r);
773}
774
775/*
776 * Close a table device that we've been using.
777 */
778static void close_table_device(struct table_device *td, struct mapped_device *md)
779{
780 if (md->disk->slave_dir)
781 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
782
783 /* Leverage async fput() if DMF_DEFERRED_REMOVE set */
784 if (unlikely(test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
785 fput(td->dm_dev.bdev_file);
786 else
787 __fput_sync(td->dm_dev.bdev_file);
788
789 put_dax(td->dm_dev.dax_dev);
790 list_del(&td->list);
791 kfree(td);
792}
793
794static struct table_device *find_table_device(struct list_head *l, dev_t dev,
795 blk_mode_t mode)
796{
797 struct table_device *td;
798
799 list_for_each_entry(td, l, list)
800 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
801 return td;
802
803 return NULL;
804}
805
806int dm_get_table_device(struct mapped_device *md, dev_t dev, blk_mode_t mode,
807 struct dm_dev **result)
808{
809 struct table_device *td;
810
811 mutex_lock(&md->table_devices_lock);
812 td = find_table_device(&md->table_devices, dev, mode);
813 if (!td) {
814 td = open_table_device(md, dev, mode);
815 if (IS_ERR(td)) {
816 mutex_unlock(&md->table_devices_lock);
817 return PTR_ERR(td);
818 }
819 } else {
820 refcount_inc(&td->count);
821 }
822 mutex_unlock(&md->table_devices_lock);
823
824 *result = &td->dm_dev;
825 return 0;
826}
827
828void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
829{
830 struct table_device *td = container_of(d, struct table_device, dm_dev);
831
832 mutex_lock(&md->table_devices_lock);
833 if (refcount_dec_and_test(&td->count))
834 close_table_device(td, md);
835 mutex_unlock(&md->table_devices_lock);
836}
837
838/*
839 * Get the geometry associated with a dm device
840 */
841int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
842{
843 *geo = md->geometry;
844
845 return 0;
846}
847
848/*
849 * Set the geometry of a device.
850 */
851int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
852{
853 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
854
855 if (geo->start > sz) {
856 DMERR("Start sector is beyond the geometry limits.");
857 return -EINVAL;
858 }
859
860 md->geometry = *geo;
861
862 return 0;
863}
864
865static int __noflush_suspending(struct mapped_device *md)
866{
867 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
868}
869
870static void dm_requeue_add_io(struct dm_io *io, bool first_stage)
871{
872 struct mapped_device *md = io->md;
873
874 if (first_stage) {
875 struct dm_io *next = md->requeue_list;
876
877 md->requeue_list = io;
878 io->next = next;
879 } else {
880 bio_list_add_head(&md->deferred, io->orig_bio);
881 }
882}
883
884static void dm_kick_requeue(struct mapped_device *md, bool first_stage)
885{
886 if (first_stage)
887 queue_work(md->wq, &md->requeue_work);
888 else
889 queue_work(md->wq, &md->work);
890}
891
892/*
893 * Return true if the dm_io's original bio is requeued.
894 * io->status is updated with error if requeue disallowed.
895 */
896static bool dm_handle_requeue(struct dm_io *io, bool first_stage)
897{
898 struct bio *bio = io->orig_bio;
899 bool handle_requeue = (io->status == BLK_STS_DM_REQUEUE);
900 bool handle_polled_eagain = ((io->status == BLK_STS_AGAIN) &&
901 (bio->bi_opf & REQ_POLLED));
902 struct mapped_device *md = io->md;
903 bool requeued = false;
904
905 if (handle_requeue || handle_polled_eagain) {
906 unsigned long flags;
907
908 if (bio->bi_opf & REQ_POLLED) {
909 /*
910 * Upper layer won't help us poll split bio
911 * (io->orig_bio may only reflect a subset of the
912 * pre-split original) so clear REQ_POLLED.
913 */
914 bio_clear_polled(bio);
915 }
916
917 /*
918 * Target requested pushing back the I/O or
919 * polled IO hit BLK_STS_AGAIN.
920 */
921 spin_lock_irqsave(&md->deferred_lock, flags);
922 if ((__noflush_suspending(md) &&
923 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) ||
924 handle_polled_eagain || first_stage) {
925 dm_requeue_add_io(io, first_stage);
926 requeued = true;
927 } else {
928 /*
929 * noflush suspend was interrupted or this is
930 * a write to a zoned target.
931 */
932 io->status = BLK_STS_IOERR;
933 }
934 spin_unlock_irqrestore(&md->deferred_lock, flags);
935 }
936
937 if (requeued)
938 dm_kick_requeue(md, first_stage);
939
940 return requeued;
941}
942
943static void __dm_io_complete(struct dm_io *io, bool first_stage)
944{
945 struct bio *bio = io->orig_bio;
946 struct mapped_device *md = io->md;
947 blk_status_t io_error;
948 bool requeued;
949
950 requeued = dm_handle_requeue(io, first_stage);
951 if (requeued && first_stage)
952 return;
953
954 io_error = io->status;
955 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
956 dm_end_io_acct(io);
957 else if (!io_error) {
958 /*
959 * Must handle target that DM_MAPIO_SUBMITTED only to
960 * then bio_endio() rather than dm_submit_bio_remap()
961 */
962 __dm_start_io_acct(io);
963 dm_end_io_acct(io);
964 }
965 free_io(io);
966 smp_wmb();
967 this_cpu_dec(*md->pending_io);
968
969 /* nudge anyone waiting on suspend queue */
970 if (unlikely(wq_has_sleeper(&md->wait)))
971 wake_up(&md->wait);
972
973 /* Return early if the original bio was requeued */
974 if (requeued)
975 return;
976
977 if (bio_is_flush_with_data(bio)) {
978 /*
979 * Preflush done for flush with data, reissue
980 * without REQ_PREFLUSH.
981 */
982 bio->bi_opf &= ~REQ_PREFLUSH;
983 queue_io(md, bio);
984 } else {
985 /* done with normal IO or empty flush */
986 if (io_error)
987 bio->bi_status = io_error;
988 bio_endio(bio);
989 }
990}
991
992static void dm_wq_requeue_work(struct work_struct *work)
993{
994 struct mapped_device *md = container_of(work, struct mapped_device,
995 requeue_work);
996 unsigned long flags;
997 struct dm_io *io;
998
999 /* reuse deferred lock to simplify dm_handle_requeue */
1000 spin_lock_irqsave(&md->deferred_lock, flags);
1001 io = md->requeue_list;
1002 md->requeue_list = NULL;
1003 spin_unlock_irqrestore(&md->deferred_lock, flags);
1004
1005 while (io) {
1006 struct dm_io *next = io->next;
1007
1008 dm_io_rewind(io, &md->disk->bio_split);
1009
1010 io->next = NULL;
1011 __dm_io_complete(io, false);
1012 io = next;
1013 cond_resched();
1014 }
1015}
1016
1017/*
1018 * Two staged requeue:
1019 *
1020 * 1) io->orig_bio points to the real original bio, and the part mapped to
1021 * this io must be requeued, instead of other parts of the original bio.
1022 *
1023 * 2) io->orig_bio points to new cloned bio which matches the requeued dm_io.
1024 */
1025static void dm_io_complete(struct dm_io *io)
1026{
1027 bool first_requeue;
1028
1029 /*
1030 * Only dm_io that has been split needs two stage requeue, otherwise
1031 * we may run into long bio clone chain during suspend and OOM could
1032 * be triggered.
1033 *
1034 * Also flush data dm_io won't be marked as DM_IO_WAS_SPLIT, so they
1035 * also aren't handled via the first stage requeue.
1036 */
1037 if (dm_io_flagged(io, DM_IO_WAS_SPLIT))
1038 first_requeue = true;
1039 else
1040 first_requeue = false;
1041
1042 __dm_io_complete(io, first_requeue);
1043}
1044
1045/*
1046 * Decrements the number of outstanding ios that a bio has been
1047 * cloned into, completing the original io if necc.
1048 */
1049static inline void __dm_io_dec_pending(struct dm_io *io)
1050{
1051 if (atomic_dec_and_test(&io->io_count))
1052 dm_io_complete(io);
1053}
1054
1055static void dm_io_set_error(struct dm_io *io, blk_status_t error)
1056{
1057 unsigned long flags;
1058
1059 /* Push-back supersedes any I/O errors */
1060 spin_lock_irqsave(&io->lock, flags);
1061 if (!(io->status == BLK_STS_DM_REQUEUE &&
1062 __noflush_suspending(io->md))) {
1063 io->status = error;
1064 }
1065 spin_unlock_irqrestore(&io->lock, flags);
1066}
1067
1068static void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
1069{
1070 if (unlikely(error))
1071 dm_io_set_error(io, error);
1072
1073 __dm_io_dec_pending(io);
1074}
1075
1076/*
1077 * The queue_limits are only valid as long as you have a reference
1078 * count on 'md'. But _not_ imposing verification to avoid atomic_read(),
1079 */
1080static inline struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
1081{
1082 return &md->queue->limits;
1083}
1084
1085void disable_discard(struct mapped_device *md)
1086{
1087 struct queue_limits *limits = dm_get_queue_limits(md);
1088
1089 /* device doesn't really support DISCARD, disable it */
1090 limits->max_hw_discard_sectors = 0;
1091}
1092
1093void disable_write_zeroes(struct mapped_device *md)
1094{
1095 struct queue_limits *limits = dm_get_queue_limits(md);
1096
1097 /* device doesn't really support WRITE ZEROES, disable it */
1098 limits->max_write_zeroes_sectors = 0;
1099}
1100
1101static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
1102{
1103 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
1104}
1105
1106static void clone_endio(struct bio *bio)
1107{
1108 blk_status_t error = bio->bi_status;
1109 struct dm_target_io *tio = clone_to_tio(bio);
1110 struct dm_target *ti = tio->ti;
1111 dm_endio_fn endio = likely(ti != NULL) ? ti->type->end_io : NULL;
1112 struct dm_io *io = tio->io;
1113 struct mapped_device *md = io->md;
1114
1115 if (unlikely(error == BLK_STS_TARGET)) {
1116 if (bio_op(bio) == REQ_OP_DISCARD &&
1117 !bdev_max_discard_sectors(bio->bi_bdev))
1118 disable_discard(md);
1119 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
1120 !bdev_write_zeroes_sectors(bio->bi_bdev))
1121 disable_write_zeroes(md);
1122 }
1123
1124 if (static_branch_unlikely(&zoned_enabled) &&
1125 unlikely(bdev_is_zoned(bio->bi_bdev)))
1126 dm_zone_endio(io, bio);
1127
1128 if (endio) {
1129 int r = endio(ti, bio, &error);
1130
1131 switch (r) {
1132 case DM_ENDIO_REQUEUE:
1133 if (static_branch_unlikely(&zoned_enabled)) {
1134 /*
1135 * Requeuing writes to a sequential zone of a zoned
1136 * target will break the sequential write pattern:
1137 * fail such IO.
1138 */
1139 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
1140 error = BLK_STS_IOERR;
1141 else
1142 error = BLK_STS_DM_REQUEUE;
1143 } else
1144 error = BLK_STS_DM_REQUEUE;
1145 fallthrough;
1146 case DM_ENDIO_DONE:
1147 break;
1148 case DM_ENDIO_INCOMPLETE:
1149 /* The target will handle the io */
1150 return;
1151 default:
1152 DMCRIT("unimplemented target endio return value: %d", r);
1153 BUG();
1154 }
1155 }
1156
1157 if (static_branch_unlikely(&swap_bios_enabled) &&
1158 likely(ti != NULL) && unlikely(swap_bios_limit(ti, bio)))
1159 up(&md->swap_bios_semaphore);
1160
1161 free_tio(bio);
1162 dm_io_dec_pending(io, error);
1163}
1164
1165/*
1166 * Return maximum size of I/O possible at the supplied sector up to the current
1167 * target boundary.
1168 */
1169static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
1170 sector_t target_offset)
1171{
1172 return ti->len - target_offset;
1173}
1174
1175static sector_t __max_io_len(struct dm_target *ti, sector_t sector,
1176 unsigned int max_granularity,
1177 unsigned int max_sectors)
1178{
1179 sector_t target_offset = dm_target_offset(ti, sector);
1180 sector_t len = max_io_len_target_boundary(ti, target_offset);
1181
1182 /*
1183 * Does the target need to split IO even further?
1184 * - varied (per target) IO splitting is a tenet of DM; this
1185 * explains why stacked chunk_sectors based splitting via
1186 * bio_split_to_limits() isn't possible here.
1187 */
1188 if (!max_granularity)
1189 return len;
1190 return min_t(sector_t, len,
1191 min(max_sectors ? : queue_max_sectors(ti->table->md->queue),
1192 blk_boundary_sectors_left(target_offset, max_granularity)));
1193}
1194
1195static inline sector_t max_io_len(struct dm_target *ti, sector_t sector)
1196{
1197 return __max_io_len(ti, sector, ti->max_io_len, 0);
1198}
1199
1200int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1201{
1202 if (len > UINT_MAX) {
1203 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1204 (unsigned long long)len, UINT_MAX);
1205 ti->error = "Maximum size of target IO is too large";
1206 return -EINVAL;
1207 }
1208
1209 ti->max_io_len = (uint32_t) len;
1210
1211 return 0;
1212}
1213EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1214
1215static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1216 sector_t sector, int *srcu_idx)
1217 __acquires(md->io_barrier)
1218{
1219 struct dm_table *map;
1220 struct dm_target *ti;
1221
1222 map = dm_get_live_table(md, srcu_idx);
1223 if (!map)
1224 return NULL;
1225
1226 ti = dm_table_find_target(map, sector);
1227 if (!ti)
1228 return NULL;
1229
1230 return ti;
1231}
1232
1233static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1234 long nr_pages, enum dax_access_mode mode, void **kaddr,
1235 pfn_t *pfn)
1236{
1237 struct mapped_device *md = dax_get_private(dax_dev);
1238 sector_t sector = pgoff * PAGE_SECTORS;
1239 struct dm_target *ti;
1240 long len, ret = -EIO;
1241 int srcu_idx;
1242
1243 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1244
1245 if (!ti)
1246 goto out;
1247 if (!ti->type->direct_access)
1248 goto out;
1249 len = max_io_len(ti, sector) / PAGE_SECTORS;
1250 if (len < 1)
1251 goto out;
1252 nr_pages = min(len, nr_pages);
1253 ret = ti->type->direct_access(ti, pgoff, nr_pages, mode, kaddr, pfn);
1254
1255 out:
1256 dm_put_live_table(md, srcu_idx);
1257
1258 return ret;
1259}
1260
1261static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1262 size_t nr_pages)
1263{
1264 struct mapped_device *md = dax_get_private(dax_dev);
1265 sector_t sector = pgoff * PAGE_SECTORS;
1266 struct dm_target *ti;
1267 int ret = -EIO;
1268 int srcu_idx;
1269
1270 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1271
1272 if (!ti)
1273 goto out;
1274 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1275 /*
1276 * ->zero_page_range() is mandatory dax operation. If we are
1277 * here, something is wrong.
1278 */
1279 goto out;
1280 }
1281 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1282 out:
1283 dm_put_live_table(md, srcu_idx);
1284
1285 return ret;
1286}
1287
1288static size_t dm_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
1289 void *addr, size_t bytes, struct iov_iter *i)
1290{
1291 struct mapped_device *md = dax_get_private(dax_dev);
1292 sector_t sector = pgoff * PAGE_SECTORS;
1293 struct dm_target *ti;
1294 int srcu_idx;
1295 long ret = 0;
1296
1297 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1298 if (!ti || !ti->type->dax_recovery_write)
1299 goto out;
1300
1301 ret = ti->type->dax_recovery_write(ti, pgoff, addr, bytes, i);
1302out:
1303 dm_put_live_table(md, srcu_idx);
1304 return ret;
1305}
1306
1307/*
1308 * A target may call dm_accept_partial_bio only from the map routine. It is
1309 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1310 * operations, REQ_OP_ZONE_APPEND (zone append writes) and any bio serviced by
1311 * __send_duplicate_bios().
1312 *
1313 * dm_accept_partial_bio informs the dm that the target only wants to process
1314 * additional n_sectors sectors of the bio and the rest of the data should be
1315 * sent in a next bio.
1316 *
1317 * A diagram that explains the arithmetics:
1318 * +--------------------+---------------+-------+
1319 * | 1 | 2 | 3 |
1320 * +--------------------+---------------+-------+
1321 *
1322 * <-------------- *tio->len_ptr --------------->
1323 * <----- bio_sectors ----->
1324 * <-- n_sectors -->
1325 *
1326 * Region 1 was already iterated over with bio_advance or similar function.
1327 * (it may be empty if the target doesn't use bio_advance)
1328 * Region 2 is the remaining bio size that the target wants to process.
1329 * (it may be empty if region 1 is non-empty, although there is no reason
1330 * to make it empty)
1331 * The target requires that region 3 is to be sent in the next bio.
1332 *
1333 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1334 * the partially processed part (the sum of regions 1+2) must be the same for all
1335 * copies of the bio.
1336 */
1337void dm_accept_partial_bio(struct bio *bio, unsigned int n_sectors)
1338{
1339 struct dm_target_io *tio = clone_to_tio(bio);
1340 struct dm_io *io = tio->io;
1341 unsigned int bio_sectors = bio_sectors(bio);
1342
1343 BUG_ON(dm_tio_flagged(tio, DM_TIO_IS_DUPLICATE_BIO));
1344 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1345 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1346 BUG_ON(bio_sectors > *tio->len_ptr);
1347 BUG_ON(n_sectors > bio_sectors);
1348
1349 *tio->len_ptr -= bio_sectors - n_sectors;
1350 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1351
1352 /*
1353 * __split_and_process_bio() may have already saved mapped part
1354 * for accounting but it is being reduced so update accordingly.
1355 */
1356 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1357 io->sectors = n_sectors;
1358 io->sector_offset = bio_sectors(io->orig_bio);
1359}
1360EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1361
1362/*
1363 * @clone: clone bio that DM core passed to target's .map function
1364 * @tgt_clone: clone of @clone bio that target needs submitted
1365 *
1366 * Targets should use this interface to submit bios they take
1367 * ownership of when returning DM_MAPIO_SUBMITTED.
1368 *
1369 * Target should also enable ti->accounts_remapped_io
1370 */
1371void dm_submit_bio_remap(struct bio *clone, struct bio *tgt_clone)
1372{
1373 struct dm_target_io *tio = clone_to_tio(clone);
1374 struct dm_io *io = tio->io;
1375
1376 /* establish bio that will get submitted */
1377 if (!tgt_clone)
1378 tgt_clone = clone;
1379
1380 /*
1381 * Account io->origin_bio to DM dev on behalf of target
1382 * that took ownership of IO with DM_MAPIO_SUBMITTED.
1383 */
1384 dm_start_io_acct(io, clone);
1385
1386 trace_block_bio_remap(tgt_clone, disk_devt(io->md->disk),
1387 tio->old_sector);
1388 submit_bio_noacct(tgt_clone);
1389}
1390EXPORT_SYMBOL_GPL(dm_submit_bio_remap);
1391
1392static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1393{
1394 mutex_lock(&md->swap_bios_lock);
1395 while (latch < md->swap_bios) {
1396 cond_resched();
1397 down(&md->swap_bios_semaphore);
1398 md->swap_bios--;
1399 }
1400 while (latch > md->swap_bios) {
1401 cond_resched();
1402 up(&md->swap_bios_semaphore);
1403 md->swap_bios++;
1404 }
1405 mutex_unlock(&md->swap_bios_lock);
1406}
1407
1408static void __map_bio(struct bio *clone)
1409{
1410 struct dm_target_io *tio = clone_to_tio(clone);
1411 struct dm_target *ti = tio->ti;
1412 struct dm_io *io = tio->io;
1413 struct mapped_device *md = io->md;
1414 int r;
1415
1416 clone->bi_end_io = clone_endio;
1417
1418 /*
1419 * Map the clone.
1420 */
1421 tio->old_sector = clone->bi_iter.bi_sector;
1422
1423 if (static_branch_unlikely(&swap_bios_enabled) &&
1424 unlikely(swap_bios_limit(ti, clone))) {
1425 int latch = get_swap_bios();
1426
1427 if (unlikely(latch != md->swap_bios))
1428 __set_swap_bios_limit(md, latch);
1429 down(&md->swap_bios_semaphore);
1430 }
1431
1432 if (likely(ti->type->map == linear_map))
1433 r = linear_map(ti, clone);
1434 else if (ti->type->map == stripe_map)
1435 r = stripe_map(ti, clone);
1436 else
1437 r = ti->type->map(ti, clone);
1438
1439 switch (r) {
1440 case DM_MAPIO_SUBMITTED:
1441 /* target has assumed ownership of this io */
1442 if (!ti->accounts_remapped_io)
1443 dm_start_io_acct(io, clone);
1444 break;
1445 case DM_MAPIO_REMAPPED:
1446 dm_submit_bio_remap(clone, NULL);
1447 break;
1448 case DM_MAPIO_KILL:
1449 case DM_MAPIO_REQUEUE:
1450 if (static_branch_unlikely(&swap_bios_enabled) &&
1451 unlikely(swap_bios_limit(ti, clone)))
1452 up(&md->swap_bios_semaphore);
1453 free_tio(clone);
1454 if (r == DM_MAPIO_KILL)
1455 dm_io_dec_pending(io, BLK_STS_IOERR);
1456 else
1457 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1458 break;
1459 default:
1460 DMCRIT("unimplemented target map return value: %d", r);
1461 BUG();
1462 }
1463}
1464
1465static void setup_split_accounting(struct clone_info *ci, unsigned int len)
1466{
1467 struct dm_io *io = ci->io;
1468
1469 if (ci->sector_count > len) {
1470 /*
1471 * Split needed, save the mapped part for accounting.
1472 * NOTE: dm_accept_partial_bio() will update accordingly.
1473 */
1474 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1475 io->sectors = len;
1476 io->sector_offset = bio_sectors(ci->bio);
1477 }
1478}
1479
1480static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1481 struct dm_target *ti, unsigned int num_bios,
1482 unsigned *len, gfp_t gfp_flag)
1483{
1484 struct bio *bio;
1485 int try = (gfp_flag & GFP_NOWAIT) ? 0 : 1;
1486
1487 for (; try < 2; try++) {
1488 int bio_nr;
1489
1490 if (try && num_bios > 1)
1491 mutex_lock(&ci->io->md->table_devices_lock);
1492 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1493 bio = alloc_tio(ci, ti, bio_nr, len,
1494 try ? GFP_NOIO : GFP_NOWAIT);
1495 if (!bio)
1496 break;
1497
1498 bio_list_add(blist, bio);
1499 }
1500 if (try && num_bios > 1)
1501 mutex_unlock(&ci->io->md->table_devices_lock);
1502 if (bio_nr == num_bios)
1503 return;
1504
1505 while ((bio = bio_list_pop(blist)))
1506 free_tio(bio);
1507 }
1508}
1509
1510static unsigned int __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1511 unsigned int num_bios, unsigned int *len,
1512 gfp_t gfp_flag)
1513{
1514 struct bio_list blist = BIO_EMPTY_LIST;
1515 struct bio *clone;
1516 unsigned int ret = 0;
1517
1518 if (WARN_ON_ONCE(num_bios == 0)) /* num_bios = 0 is a bug in caller */
1519 return 0;
1520
1521 /* dm_accept_partial_bio() is not supported with shared tio->len_ptr */
1522 if (len)
1523 setup_split_accounting(ci, *len);
1524
1525 /*
1526 * Using alloc_multiple_bios(), even if num_bios is 1, to consistently
1527 * support allocating using GFP_NOWAIT with GFP_NOIO fallback.
1528 */
1529 alloc_multiple_bios(&blist, ci, ti, num_bios, len, gfp_flag);
1530 while ((clone = bio_list_pop(&blist))) {
1531 if (num_bios > 1)
1532 dm_tio_set_flag(clone_to_tio(clone), DM_TIO_IS_DUPLICATE_BIO);
1533 __map_bio(clone);
1534 ret += 1;
1535 }
1536
1537 return ret;
1538}
1539
1540static void __send_empty_flush(struct clone_info *ci)
1541{
1542 struct dm_table *t = ci->map;
1543 struct bio flush_bio;
1544
1545 /*
1546 * Use an on-stack bio for this, it's safe since we don't
1547 * need to reference it after submit. It's just used as
1548 * the basis for the clone(s).
1549 */
1550 bio_init(&flush_bio, ci->io->md->disk->part0, NULL, 0,
1551 REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
1552
1553 ci->bio = &flush_bio;
1554 ci->sector_count = 0;
1555 ci->io->tio.clone.bi_iter.bi_size = 0;
1556
1557 if (!t->flush_bypasses_map) {
1558 for (unsigned int i = 0; i < t->num_targets; i++) {
1559 unsigned int bios;
1560 struct dm_target *ti = dm_table_get_target(t, i);
1561
1562 if (unlikely(ti->num_flush_bios == 0))
1563 continue;
1564
1565 atomic_add(ti->num_flush_bios, &ci->io->io_count);
1566 bios = __send_duplicate_bios(ci, ti, ti->num_flush_bios,
1567 NULL, GFP_NOWAIT);
1568 atomic_sub(ti->num_flush_bios - bios, &ci->io->io_count);
1569 }
1570 } else {
1571 /*
1572 * Note that there's no need to grab t->devices_lock here
1573 * because the targets that support flush optimization don't
1574 * modify the list of devices.
1575 */
1576 struct list_head *devices = dm_table_get_devices(t);
1577 unsigned int len = 0;
1578 struct dm_dev_internal *dd;
1579 list_for_each_entry(dd, devices, list) {
1580 struct bio *clone;
1581 /*
1582 * Note that the structure dm_target_io is not
1583 * associated with any target (because the device may be
1584 * used by multiple targets), so we set tio->ti = NULL.
1585 * We must check for NULL in the I/O processing path, to
1586 * avoid NULL pointer dereference.
1587 */
1588 clone = alloc_tio(ci, NULL, 0, &len, GFP_NOIO);
1589 atomic_add(1, &ci->io->io_count);
1590 bio_set_dev(clone, dd->dm_dev->bdev);
1591 clone->bi_end_io = clone_endio;
1592 dm_submit_bio_remap(clone, NULL);
1593 }
1594 }
1595
1596 /*
1597 * alloc_io() takes one extra reference for submission, so the
1598 * reference won't reach 0 without the following subtraction
1599 */
1600 atomic_sub(1, &ci->io->io_count);
1601
1602 bio_uninit(ci->bio);
1603}
1604
1605static void __send_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1606 unsigned int num_bios, unsigned int max_granularity,
1607 unsigned int max_sectors)
1608{
1609 unsigned int len, bios;
1610
1611 len = min_t(sector_t, ci->sector_count,
1612 __max_io_len(ti, ci->sector, max_granularity, max_sectors));
1613
1614 atomic_add(num_bios, &ci->io->io_count);
1615 bios = __send_duplicate_bios(ci, ti, num_bios, &len, GFP_NOIO);
1616 /*
1617 * alloc_io() takes one extra reference for submission, so the
1618 * reference won't reach 0 without the following (+1) subtraction
1619 */
1620 atomic_sub(num_bios - bios + 1, &ci->io->io_count);
1621
1622 ci->sector += len;
1623 ci->sector_count -= len;
1624}
1625
1626static bool is_abnormal_io(struct bio *bio)
1627{
1628 switch (bio_op(bio)) {
1629 case REQ_OP_READ:
1630 case REQ_OP_WRITE:
1631 case REQ_OP_FLUSH:
1632 return false;
1633 case REQ_OP_DISCARD:
1634 case REQ_OP_SECURE_ERASE:
1635 case REQ_OP_WRITE_ZEROES:
1636 case REQ_OP_ZONE_RESET_ALL:
1637 return true;
1638 default:
1639 return false;
1640 }
1641}
1642
1643static blk_status_t __process_abnormal_io(struct clone_info *ci,
1644 struct dm_target *ti)
1645{
1646 unsigned int num_bios = 0;
1647 unsigned int max_granularity = 0;
1648 unsigned int max_sectors = 0;
1649 struct queue_limits *limits = dm_get_queue_limits(ti->table->md);
1650
1651 switch (bio_op(ci->bio)) {
1652 case REQ_OP_DISCARD:
1653 num_bios = ti->num_discard_bios;
1654 max_sectors = limits->max_discard_sectors;
1655 if (ti->max_discard_granularity)
1656 max_granularity = max_sectors;
1657 break;
1658 case REQ_OP_SECURE_ERASE:
1659 num_bios = ti->num_secure_erase_bios;
1660 max_sectors = limits->max_secure_erase_sectors;
1661 break;
1662 case REQ_OP_WRITE_ZEROES:
1663 num_bios = ti->num_write_zeroes_bios;
1664 max_sectors = limits->max_write_zeroes_sectors;
1665 break;
1666 default:
1667 break;
1668 }
1669
1670 /*
1671 * Even though the device advertised support for this type of
1672 * request, that does not mean every target supports it, and
1673 * reconfiguration might also have changed that since the
1674 * check was performed.
1675 */
1676 if (unlikely(!num_bios))
1677 return BLK_STS_NOTSUPP;
1678
1679 __send_abnormal_io(ci, ti, num_bios, max_granularity, max_sectors);
1680
1681 return BLK_STS_OK;
1682}
1683
1684/*
1685 * Reuse ->bi_private as dm_io list head for storing all dm_io instances
1686 * associated with this bio, and this bio's bi_private needs to be
1687 * stored in dm_io->data before the reuse.
1688 *
1689 * bio->bi_private is owned by fs or upper layer, so block layer won't
1690 * touch it after splitting. Meantime it won't be changed by anyone after
1691 * bio is submitted. So this reuse is safe.
1692 */
1693static inline struct dm_io **dm_poll_list_head(struct bio *bio)
1694{
1695 return (struct dm_io **)&bio->bi_private;
1696}
1697
1698static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
1699{
1700 struct dm_io **head = dm_poll_list_head(bio);
1701
1702 if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
1703 bio->bi_opf |= REQ_DM_POLL_LIST;
1704 /*
1705 * Save .bi_private into dm_io, so that we can reuse
1706 * .bi_private as dm_io list head for storing dm_io list
1707 */
1708 io->data = bio->bi_private;
1709
1710 /* tell block layer to poll for completion */
1711 bio->bi_cookie = ~BLK_QC_T_NONE;
1712
1713 io->next = NULL;
1714 } else {
1715 /*
1716 * bio recursed due to split, reuse original poll list,
1717 * and save bio->bi_private too.
1718 */
1719 io->data = (*head)->data;
1720 io->next = *head;
1721 }
1722
1723 *head = io;
1724}
1725
1726/*
1727 * Select the correct strategy for processing a non-flush bio.
1728 */
1729static blk_status_t __split_and_process_bio(struct clone_info *ci)
1730{
1731 struct bio *clone;
1732 struct dm_target *ti;
1733 unsigned int len;
1734
1735 ti = dm_table_find_target(ci->map, ci->sector);
1736 if (unlikely(!ti))
1737 return BLK_STS_IOERR;
1738
1739 if (unlikely(ci->is_abnormal_io))
1740 return __process_abnormal_io(ci, ti);
1741
1742 /*
1743 * Only support bio polling for normal IO, and the target io is
1744 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
1745 */
1746 ci->submit_as_polled = !!(ci->bio->bi_opf & REQ_POLLED);
1747
1748 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1749 setup_split_accounting(ci, len);
1750
1751 if (unlikely(ci->bio->bi_opf & REQ_NOWAIT)) {
1752 if (unlikely(!dm_target_supports_nowait(ti->type)))
1753 return BLK_STS_NOTSUPP;
1754
1755 clone = alloc_tio(ci, ti, 0, &len, GFP_NOWAIT);
1756 if (unlikely(!clone))
1757 return BLK_STS_AGAIN;
1758 } else {
1759 clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
1760 }
1761 __map_bio(clone);
1762
1763 ci->sector += len;
1764 ci->sector_count -= len;
1765
1766 return BLK_STS_OK;
1767}
1768
1769static void init_clone_info(struct clone_info *ci, struct dm_io *io,
1770 struct dm_table *map, struct bio *bio, bool is_abnormal)
1771{
1772 ci->map = map;
1773 ci->io = io;
1774 ci->bio = bio;
1775 ci->is_abnormal_io = is_abnormal;
1776 ci->submit_as_polled = false;
1777 ci->sector = bio->bi_iter.bi_sector;
1778 ci->sector_count = bio_sectors(bio);
1779
1780 /* Shouldn't happen but sector_count was being set to 0 so... */
1781 if (static_branch_unlikely(&zoned_enabled) &&
1782 WARN_ON_ONCE(op_is_zone_mgmt(bio_op(bio)) && ci->sector_count))
1783 ci->sector_count = 0;
1784}
1785
1786#ifdef CONFIG_BLK_DEV_ZONED
1787static inline bool dm_zone_bio_needs_split(struct mapped_device *md,
1788 struct bio *bio)
1789{
1790 /*
1791 * For mapped device that need zone append emulation, we must
1792 * split any large BIO that straddles zone boundaries.
1793 */
1794 return dm_emulate_zone_append(md) && bio_straddles_zones(bio) &&
1795 !bio_flagged(bio, BIO_ZONE_WRITE_PLUGGING);
1796}
1797static inline bool dm_zone_plug_bio(struct mapped_device *md, struct bio *bio)
1798{
1799 return dm_emulate_zone_append(md) && blk_zone_plug_bio(bio, 0);
1800}
1801
1802static blk_status_t __send_zone_reset_all_emulated(struct clone_info *ci,
1803 struct dm_target *ti)
1804{
1805 struct bio_list blist = BIO_EMPTY_LIST;
1806 struct mapped_device *md = ci->io->md;
1807 unsigned int zone_sectors = md->disk->queue->limits.chunk_sectors;
1808 unsigned long *need_reset;
1809 unsigned int i, nr_zones, nr_reset;
1810 unsigned int num_bios = 0;
1811 blk_status_t sts = BLK_STS_OK;
1812 sector_t sector = ti->begin;
1813 struct bio *clone;
1814 int ret;
1815
1816 nr_zones = ti->len >> ilog2(zone_sectors);
1817 need_reset = bitmap_zalloc(nr_zones, GFP_NOIO);
1818 if (!need_reset)
1819 return BLK_STS_RESOURCE;
1820
1821 ret = dm_zone_get_reset_bitmap(md, ci->map, ti->begin,
1822 nr_zones, need_reset);
1823 if (ret) {
1824 sts = BLK_STS_IOERR;
1825 goto free_bitmap;
1826 }
1827
1828 /* If we have no zone to reset, we are done. */
1829 nr_reset = bitmap_weight(need_reset, nr_zones);
1830 if (!nr_reset)
1831 goto free_bitmap;
1832
1833 atomic_add(nr_zones, &ci->io->io_count);
1834
1835 for (i = 0; i < nr_zones; i++) {
1836
1837 if (!test_bit(i, need_reset)) {
1838 sector += zone_sectors;
1839 continue;
1840 }
1841
1842 if (bio_list_empty(&blist)) {
1843 /* This may take a while, so be nice to others */
1844 if (num_bios)
1845 cond_resched();
1846
1847 /*
1848 * We may need to reset thousands of zones, so let's
1849 * not go crazy with the clone allocation.
1850 */
1851 alloc_multiple_bios(&blist, ci, ti, min(nr_reset, 32),
1852 NULL, GFP_NOIO);
1853 }
1854
1855 /* Get a clone and change it to a regular reset operation. */
1856 clone = bio_list_pop(&blist);
1857 clone->bi_opf &= ~REQ_OP_MASK;
1858 clone->bi_opf |= REQ_OP_ZONE_RESET | REQ_SYNC;
1859 clone->bi_iter.bi_sector = sector;
1860 clone->bi_iter.bi_size = 0;
1861 __map_bio(clone);
1862
1863 sector += zone_sectors;
1864 num_bios++;
1865 nr_reset--;
1866 }
1867
1868 WARN_ON_ONCE(!bio_list_empty(&blist));
1869 atomic_sub(nr_zones - num_bios, &ci->io->io_count);
1870 ci->sector_count = 0;
1871
1872free_bitmap:
1873 bitmap_free(need_reset);
1874
1875 return sts;
1876}
1877
1878static void __send_zone_reset_all_native(struct clone_info *ci,
1879 struct dm_target *ti)
1880{
1881 unsigned int bios;
1882
1883 atomic_add(1, &ci->io->io_count);
1884 bios = __send_duplicate_bios(ci, ti, 1, NULL, GFP_NOIO);
1885 atomic_sub(1 - bios, &ci->io->io_count);
1886
1887 ci->sector_count = 0;
1888}
1889
1890static blk_status_t __send_zone_reset_all(struct clone_info *ci)
1891{
1892 struct dm_table *t = ci->map;
1893 blk_status_t sts = BLK_STS_OK;
1894
1895 for (unsigned int i = 0; i < t->num_targets; i++) {
1896 struct dm_target *ti = dm_table_get_target(t, i);
1897
1898 if (ti->zone_reset_all_supported) {
1899 __send_zone_reset_all_native(ci, ti);
1900 continue;
1901 }
1902
1903 sts = __send_zone_reset_all_emulated(ci, ti);
1904 if (sts != BLK_STS_OK)
1905 break;
1906 }
1907
1908 /* Release the reference that alloc_io() took for submission. */
1909 atomic_sub(1, &ci->io->io_count);
1910
1911 return sts;
1912}
1913
1914#else
1915static inline bool dm_zone_bio_needs_split(struct mapped_device *md,
1916 struct bio *bio)
1917{
1918 return false;
1919}
1920static inline bool dm_zone_plug_bio(struct mapped_device *md, struct bio *bio)
1921{
1922 return false;
1923}
1924static blk_status_t __send_zone_reset_all(struct clone_info *ci)
1925{
1926 return BLK_STS_NOTSUPP;
1927}
1928#endif
1929
1930/*
1931 * Entry point to split a bio into clones and submit them to the targets.
1932 */
1933static void dm_split_and_process_bio(struct mapped_device *md,
1934 struct dm_table *map, struct bio *bio)
1935{
1936 struct clone_info ci;
1937 struct dm_io *io;
1938 blk_status_t error = BLK_STS_OK;
1939 bool is_abnormal, need_split;
1940
1941 is_abnormal = is_abnormal_io(bio);
1942 if (static_branch_unlikely(&zoned_enabled)) {
1943 /* Special case REQ_OP_ZONE_RESET_ALL as it cannot be split. */
1944 need_split = (bio_op(bio) != REQ_OP_ZONE_RESET_ALL) &&
1945 (is_abnormal || dm_zone_bio_needs_split(md, bio));
1946 } else {
1947 need_split = is_abnormal;
1948 }
1949
1950 if (unlikely(need_split)) {
1951 /*
1952 * Use bio_split_to_limits() for abnormal IO (e.g. discard, etc)
1953 * otherwise associated queue_limits won't be imposed.
1954 * Also split the BIO for mapped devices needing zone append
1955 * emulation to ensure that the BIO does not cross zone
1956 * boundaries.
1957 */
1958 bio = bio_split_to_limits(bio);
1959 if (!bio)
1960 return;
1961 }
1962
1963 /*
1964 * Use the block layer zone write plugging for mapped devices that
1965 * need zone append emulation (e.g. dm-crypt).
1966 */
1967 if (static_branch_unlikely(&zoned_enabled) && dm_zone_plug_bio(md, bio))
1968 return;
1969
1970 /* Only support nowait for normal IO */
1971 if (unlikely(bio->bi_opf & REQ_NOWAIT) && !is_abnormal) {
1972 io = alloc_io(md, bio, GFP_NOWAIT);
1973 if (unlikely(!io)) {
1974 /* Unable to do anything without dm_io. */
1975 bio_wouldblock_error(bio);
1976 return;
1977 }
1978 } else {
1979 io = alloc_io(md, bio, GFP_NOIO);
1980 }
1981 init_clone_info(&ci, io, map, bio, is_abnormal);
1982
1983 if (bio->bi_opf & REQ_PREFLUSH) {
1984 __send_empty_flush(&ci);
1985 /* dm_io_complete submits any data associated with flush */
1986 goto out;
1987 }
1988
1989 if (static_branch_unlikely(&zoned_enabled) &&
1990 (bio_op(bio) == REQ_OP_ZONE_RESET_ALL)) {
1991 error = __send_zone_reset_all(&ci);
1992 goto out;
1993 }
1994
1995 error = __split_and_process_bio(&ci);
1996 if (error || !ci.sector_count)
1997 goto out;
1998 /*
1999 * Remainder must be passed to submit_bio_noacct() so it gets handled
2000 * *after* bios already submitted have been completely processed.
2001 */
2002 bio_trim(bio, io->sectors, ci.sector_count);
2003 trace_block_split(bio, bio->bi_iter.bi_sector);
2004 bio_inc_remaining(bio);
2005 submit_bio_noacct(bio);
2006out:
2007 /*
2008 * Drop the extra reference count for non-POLLED bio, and hold one
2009 * reference for POLLED bio, which will be released in dm_poll_bio
2010 *
2011 * Add every dm_io instance into the dm_io list head which is stored
2012 * in bio->bi_private, so that dm_poll_bio can poll them all.
2013 */
2014 if (error || !ci.submit_as_polled) {
2015 /*
2016 * In case of submission failure, the extra reference for
2017 * submitting io isn't consumed yet
2018 */
2019 if (error)
2020 atomic_dec(&io->io_count);
2021 dm_io_dec_pending(io, error);
2022 } else
2023 dm_queue_poll_io(bio, io);
2024}
2025
2026static void dm_submit_bio(struct bio *bio)
2027{
2028 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
2029 int srcu_idx;
2030 struct dm_table *map;
2031
2032 map = dm_get_live_table(md, &srcu_idx);
2033 if (unlikely(!map)) {
2034 DMERR_LIMIT("%s: mapping table unavailable, erroring io",
2035 dm_device_name(md));
2036 bio_io_error(bio);
2037 goto out;
2038 }
2039
2040 /* If suspended, queue this IO for later */
2041 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
2042 if (bio->bi_opf & REQ_NOWAIT)
2043 bio_wouldblock_error(bio);
2044 else if (bio->bi_opf & REQ_RAHEAD)
2045 bio_io_error(bio);
2046 else
2047 queue_io(md, bio);
2048 goto out;
2049 }
2050
2051 dm_split_and_process_bio(md, map, bio);
2052out:
2053 dm_put_live_table(md, srcu_idx);
2054}
2055
2056static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
2057 unsigned int flags)
2058{
2059 WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
2060
2061 /* don't poll if the mapped io is done */
2062 if (atomic_read(&io->io_count) > 1)
2063 bio_poll(&io->tio.clone, iob, flags);
2064
2065 /* bio_poll holds the last reference */
2066 return atomic_read(&io->io_count) == 1;
2067}
2068
2069static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
2070 unsigned int flags)
2071{
2072 struct dm_io **head = dm_poll_list_head(bio);
2073 struct dm_io *list = *head;
2074 struct dm_io *tmp = NULL;
2075 struct dm_io *curr, *next;
2076
2077 /* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
2078 if (!(bio->bi_opf & REQ_DM_POLL_LIST))
2079 return 0;
2080
2081 WARN_ON_ONCE(!list);
2082
2083 /*
2084 * Restore .bi_private before possibly completing dm_io.
2085 *
2086 * bio_poll() is only possible once @bio has been completely
2087 * submitted via submit_bio_noacct()'s depth-first submission.
2088 * So there is no dm_queue_poll_io() race associated with
2089 * clearing REQ_DM_POLL_LIST here.
2090 */
2091 bio->bi_opf &= ~REQ_DM_POLL_LIST;
2092 bio->bi_private = list->data;
2093
2094 for (curr = list, next = curr->next; curr; curr = next, next =
2095 curr ? curr->next : NULL) {
2096 if (dm_poll_dm_io(curr, iob, flags)) {
2097 /*
2098 * clone_endio() has already occurred, so no
2099 * error handling is needed here.
2100 */
2101 __dm_io_dec_pending(curr);
2102 } else {
2103 curr->next = tmp;
2104 tmp = curr;
2105 }
2106 }
2107
2108 /* Not done? */
2109 if (tmp) {
2110 bio->bi_opf |= REQ_DM_POLL_LIST;
2111 /* Reset bio->bi_private to dm_io list head */
2112 *head = tmp;
2113 return 0;
2114 }
2115 return 1;
2116}
2117
2118/*
2119 *---------------------------------------------------------------
2120 * An IDR is used to keep track of allocated minor numbers.
2121 *---------------------------------------------------------------
2122 */
2123static void free_minor(int minor)
2124{
2125 spin_lock(&_minor_lock);
2126 idr_remove(&_minor_idr, minor);
2127 spin_unlock(&_minor_lock);
2128}
2129
2130/*
2131 * See if the device with a specific minor # is free.
2132 */
2133static int specific_minor(int minor)
2134{
2135 int r;
2136
2137 if (minor >= (1 << MINORBITS))
2138 return -EINVAL;
2139
2140 idr_preload(GFP_KERNEL);
2141 spin_lock(&_minor_lock);
2142
2143 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2144
2145 spin_unlock(&_minor_lock);
2146 idr_preload_end();
2147 if (r < 0)
2148 return r == -ENOSPC ? -EBUSY : r;
2149 return 0;
2150}
2151
2152static int next_free_minor(int *minor)
2153{
2154 int r;
2155
2156 idr_preload(GFP_KERNEL);
2157 spin_lock(&_minor_lock);
2158
2159 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2160
2161 spin_unlock(&_minor_lock);
2162 idr_preload_end();
2163 if (r < 0)
2164 return r;
2165 *minor = r;
2166 return 0;
2167}
2168
2169static const struct block_device_operations dm_blk_dops;
2170static const struct block_device_operations dm_rq_blk_dops;
2171static const struct dax_operations dm_dax_ops;
2172
2173static void dm_wq_work(struct work_struct *work);
2174
2175#ifdef CONFIG_BLK_INLINE_ENCRYPTION
2176static void dm_queue_destroy_crypto_profile(struct request_queue *q)
2177{
2178 dm_destroy_crypto_profile(q->crypto_profile);
2179}
2180
2181#else /* CONFIG_BLK_INLINE_ENCRYPTION */
2182
2183static inline void dm_queue_destroy_crypto_profile(struct request_queue *q)
2184{
2185}
2186#endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
2187
2188static void cleanup_mapped_device(struct mapped_device *md)
2189{
2190 if (md->wq)
2191 destroy_workqueue(md->wq);
2192 dm_free_md_mempools(md->mempools);
2193
2194 if (md->dax_dev) {
2195 dax_remove_host(md->disk);
2196 kill_dax(md->dax_dev);
2197 put_dax(md->dax_dev);
2198 md->dax_dev = NULL;
2199 }
2200
2201 if (md->disk) {
2202 spin_lock(&_minor_lock);
2203 md->disk->private_data = NULL;
2204 spin_unlock(&_minor_lock);
2205 if (dm_get_md_type(md) != DM_TYPE_NONE) {
2206 struct table_device *td;
2207
2208 dm_sysfs_exit(md);
2209 list_for_each_entry(td, &md->table_devices, list) {
2210 bd_unlink_disk_holder(td->dm_dev.bdev,
2211 md->disk);
2212 }
2213
2214 /*
2215 * Hold lock to make sure del_gendisk() won't concurrent
2216 * with open/close_table_device().
2217 */
2218 mutex_lock(&md->table_devices_lock);
2219 del_gendisk(md->disk);
2220 mutex_unlock(&md->table_devices_lock);
2221 }
2222 dm_queue_destroy_crypto_profile(md->queue);
2223 put_disk(md->disk);
2224 }
2225
2226 if (md->pending_io) {
2227 free_percpu(md->pending_io);
2228 md->pending_io = NULL;
2229 }
2230
2231 cleanup_srcu_struct(&md->io_barrier);
2232
2233 mutex_destroy(&md->suspend_lock);
2234 mutex_destroy(&md->type_lock);
2235 mutex_destroy(&md->table_devices_lock);
2236 mutex_destroy(&md->swap_bios_lock);
2237
2238 dm_mq_cleanup_mapped_device(md);
2239}
2240
2241/*
2242 * Allocate and initialise a blank device with a given minor.
2243 */
2244static struct mapped_device *alloc_dev(int minor)
2245{
2246 int r, numa_node_id = dm_get_numa_node();
2247 struct dax_device *dax_dev;
2248 struct mapped_device *md;
2249 void *old_md;
2250
2251 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
2252 if (!md) {
2253 DMERR("unable to allocate device, out of memory.");
2254 return NULL;
2255 }
2256
2257 if (!try_module_get(THIS_MODULE))
2258 goto bad_module_get;
2259
2260 /* get a minor number for the dev */
2261 if (minor == DM_ANY_MINOR)
2262 r = next_free_minor(&minor);
2263 else
2264 r = specific_minor(minor);
2265 if (r < 0)
2266 goto bad_minor;
2267
2268 r = init_srcu_struct(&md->io_barrier);
2269 if (r < 0)
2270 goto bad_io_barrier;
2271
2272 md->numa_node_id = numa_node_id;
2273 md->init_tio_pdu = false;
2274 md->type = DM_TYPE_NONE;
2275 mutex_init(&md->suspend_lock);
2276 mutex_init(&md->type_lock);
2277 mutex_init(&md->table_devices_lock);
2278 spin_lock_init(&md->deferred_lock);
2279 atomic_set(&md->holders, 1);
2280 atomic_set(&md->open_count, 0);
2281 atomic_set(&md->event_nr, 0);
2282 atomic_set(&md->uevent_seq, 0);
2283 INIT_LIST_HEAD(&md->uevent_list);
2284 INIT_LIST_HEAD(&md->table_devices);
2285 spin_lock_init(&md->uevent_lock);
2286
2287 /*
2288 * default to bio-based until DM table is loaded and md->type
2289 * established. If request-based table is loaded: blk-mq will
2290 * override accordingly.
2291 */
2292 md->disk = blk_alloc_disk(NULL, md->numa_node_id);
2293 if (IS_ERR(md->disk)) {
2294 md->disk = NULL;
2295 goto bad;
2296 }
2297 md->queue = md->disk->queue;
2298
2299 init_waitqueue_head(&md->wait);
2300 INIT_WORK(&md->work, dm_wq_work);
2301 INIT_WORK(&md->requeue_work, dm_wq_requeue_work);
2302 init_waitqueue_head(&md->eventq);
2303 init_completion(&md->kobj_holder.completion);
2304
2305 md->requeue_list = NULL;
2306 md->swap_bios = get_swap_bios();
2307 sema_init(&md->swap_bios_semaphore, md->swap_bios);
2308 mutex_init(&md->swap_bios_lock);
2309
2310 md->disk->major = _major;
2311 md->disk->first_minor = minor;
2312 md->disk->minors = 1;
2313 md->disk->flags |= GENHD_FL_NO_PART;
2314 md->disk->fops = &dm_blk_dops;
2315 md->disk->private_data = md;
2316 sprintf(md->disk->disk_name, "dm-%d", minor);
2317
2318 dax_dev = alloc_dax(md, &dm_dax_ops);
2319 if (IS_ERR(dax_dev)) {
2320 if (PTR_ERR(dax_dev) != -EOPNOTSUPP)
2321 goto bad;
2322 } else {
2323 set_dax_nocache(dax_dev);
2324 set_dax_nomc(dax_dev);
2325 md->dax_dev = dax_dev;
2326 if (dax_add_host(dax_dev, md->disk))
2327 goto bad;
2328 }
2329
2330 format_dev_t(md->name, MKDEV(_major, minor));
2331
2332 md->wq = alloc_workqueue("kdmflush/%s", WQ_MEM_RECLAIM, 0, md->name);
2333 if (!md->wq)
2334 goto bad;
2335
2336 md->pending_io = alloc_percpu(unsigned long);
2337 if (!md->pending_io)
2338 goto bad;
2339
2340 r = dm_stats_init(&md->stats);
2341 if (r < 0)
2342 goto bad;
2343
2344 /* Populate the mapping, nobody knows we exist yet */
2345 spin_lock(&_minor_lock);
2346 old_md = idr_replace(&_minor_idr, md, minor);
2347 spin_unlock(&_minor_lock);
2348
2349 BUG_ON(old_md != MINOR_ALLOCED);
2350
2351 return md;
2352
2353bad:
2354 cleanup_mapped_device(md);
2355bad_io_barrier:
2356 free_minor(minor);
2357bad_minor:
2358 module_put(THIS_MODULE);
2359bad_module_get:
2360 kvfree(md);
2361 return NULL;
2362}
2363
2364static void unlock_fs(struct mapped_device *md);
2365
2366static void free_dev(struct mapped_device *md)
2367{
2368 int minor = MINOR(disk_devt(md->disk));
2369
2370 unlock_fs(md);
2371
2372 cleanup_mapped_device(md);
2373
2374 WARN_ON_ONCE(!list_empty(&md->table_devices));
2375 dm_stats_cleanup(&md->stats);
2376 free_minor(minor);
2377
2378 module_put(THIS_MODULE);
2379 kvfree(md);
2380}
2381
2382/*
2383 * Bind a table to the device.
2384 */
2385static void event_callback(void *context)
2386{
2387 unsigned long flags;
2388 LIST_HEAD(uevents);
2389 struct mapped_device *md = context;
2390
2391 spin_lock_irqsave(&md->uevent_lock, flags);
2392 list_splice_init(&md->uevent_list, &uevents);
2393 spin_unlock_irqrestore(&md->uevent_lock, flags);
2394
2395 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2396
2397 atomic_inc(&md->event_nr);
2398 wake_up(&md->eventq);
2399 dm_issue_global_event();
2400}
2401
2402/*
2403 * Returns old map, which caller must destroy.
2404 */
2405static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2406 struct queue_limits *limits)
2407{
2408 struct dm_table *old_map;
2409 sector_t size;
2410 int ret;
2411
2412 lockdep_assert_held(&md->suspend_lock);
2413
2414 size = dm_table_get_size(t);
2415
2416 /*
2417 * Wipe any geometry if the size of the table changed.
2418 */
2419 if (size != dm_get_size(md))
2420 memset(&md->geometry, 0, sizeof(md->geometry));
2421
2422 set_capacity(md->disk, size);
2423
2424 dm_table_event_callback(t, event_callback, md);
2425
2426 if (dm_table_request_based(t)) {
2427 /*
2428 * Leverage the fact that request-based DM targets are
2429 * immutable singletons - used to optimize dm_mq_queue_rq.
2430 */
2431 md->immutable_target = dm_table_get_immutable_target(t);
2432
2433 /*
2434 * There is no need to reload with request-based dm because the
2435 * size of front_pad doesn't change.
2436 *
2437 * Note for future: If you are to reload bioset, prep-ed
2438 * requests in the queue may refer to bio from the old bioset,
2439 * so you must walk through the queue to unprep.
2440 */
2441 if (!md->mempools) {
2442 md->mempools = t->mempools;
2443 t->mempools = NULL;
2444 }
2445 } else {
2446 /*
2447 * The md may already have mempools that need changing.
2448 * If so, reload bioset because front_pad may have changed
2449 * because a different table was loaded.
2450 */
2451 dm_free_md_mempools(md->mempools);
2452 md->mempools = t->mempools;
2453 t->mempools = NULL;
2454 }
2455
2456 ret = dm_table_set_restrictions(t, md->queue, limits);
2457 if (ret) {
2458 old_map = ERR_PTR(ret);
2459 goto out;
2460 }
2461
2462 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2463 rcu_assign_pointer(md->map, (void *)t);
2464 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2465
2466 if (old_map)
2467 dm_sync_table(md);
2468out:
2469 return old_map;
2470}
2471
2472/*
2473 * Returns unbound table for the caller to free.
2474 */
2475static struct dm_table *__unbind(struct mapped_device *md)
2476{
2477 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2478
2479 if (!map)
2480 return NULL;
2481
2482 dm_table_event_callback(map, NULL, NULL);
2483 RCU_INIT_POINTER(md->map, NULL);
2484 dm_sync_table(md);
2485
2486 return map;
2487}
2488
2489/*
2490 * Constructor for a new device.
2491 */
2492int dm_create(int minor, struct mapped_device **result)
2493{
2494 struct mapped_device *md;
2495
2496 md = alloc_dev(minor);
2497 if (!md)
2498 return -ENXIO;
2499
2500 dm_ima_reset_data(md);
2501
2502 *result = md;
2503 return 0;
2504}
2505
2506/*
2507 * Functions to manage md->type.
2508 * All are required to hold md->type_lock.
2509 */
2510void dm_lock_md_type(struct mapped_device *md)
2511{
2512 mutex_lock(&md->type_lock);
2513}
2514
2515void dm_unlock_md_type(struct mapped_device *md)
2516{
2517 mutex_unlock(&md->type_lock);
2518}
2519
2520enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2521{
2522 return md->type;
2523}
2524
2525struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2526{
2527 return md->immutable_target_type;
2528}
2529
2530/*
2531 * Setup the DM device's queue based on md's type
2532 */
2533int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2534{
2535 enum dm_queue_mode type = dm_table_get_type(t);
2536 struct queue_limits limits;
2537 struct table_device *td;
2538 int r;
2539
2540 WARN_ON_ONCE(type == DM_TYPE_NONE);
2541
2542 if (type == DM_TYPE_REQUEST_BASED) {
2543 md->disk->fops = &dm_rq_blk_dops;
2544 r = dm_mq_init_request_queue(md, t);
2545 if (r) {
2546 DMERR("Cannot initialize queue for request-based dm mapped device");
2547 return r;
2548 }
2549 }
2550
2551 r = dm_calculate_queue_limits(t, &limits);
2552 if (r) {
2553 DMERR("Cannot calculate initial queue limits");
2554 return r;
2555 }
2556 r = dm_table_set_restrictions(t, md->queue, &limits);
2557 if (r)
2558 return r;
2559
2560 /*
2561 * Hold lock to make sure add_disk() and del_gendisk() won't concurrent
2562 * with open_table_device() and close_table_device().
2563 */
2564 mutex_lock(&md->table_devices_lock);
2565 r = add_disk(md->disk);
2566 mutex_unlock(&md->table_devices_lock);
2567 if (r)
2568 return r;
2569
2570 /*
2571 * Register the holder relationship for devices added before the disk
2572 * was live.
2573 */
2574 list_for_each_entry(td, &md->table_devices, list) {
2575 r = bd_link_disk_holder(td->dm_dev.bdev, md->disk);
2576 if (r)
2577 goto out_undo_holders;
2578 }
2579
2580 r = dm_sysfs_init(md);
2581 if (r)
2582 goto out_undo_holders;
2583
2584 md->type = type;
2585 return 0;
2586
2587out_undo_holders:
2588 list_for_each_entry_continue_reverse(td, &md->table_devices, list)
2589 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
2590 mutex_lock(&md->table_devices_lock);
2591 del_gendisk(md->disk);
2592 mutex_unlock(&md->table_devices_lock);
2593 return r;
2594}
2595
2596struct mapped_device *dm_get_md(dev_t dev)
2597{
2598 struct mapped_device *md;
2599 unsigned int minor = MINOR(dev);
2600
2601 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2602 return NULL;
2603
2604 spin_lock(&_minor_lock);
2605
2606 md = idr_find(&_minor_idr, minor);
2607 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2608 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2609 md = NULL;
2610 goto out;
2611 }
2612 dm_get(md);
2613out:
2614 spin_unlock(&_minor_lock);
2615
2616 return md;
2617}
2618EXPORT_SYMBOL_GPL(dm_get_md);
2619
2620void *dm_get_mdptr(struct mapped_device *md)
2621{
2622 return md->interface_ptr;
2623}
2624
2625void dm_set_mdptr(struct mapped_device *md, void *ptr)
2626{
2627 md->interface_ptr = ptr;
2628}
2629
2630void dm_get(struct mapped_device *md)
2631{
2632 atomic_inc(&md->holders);
2633 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2634}
2635
2636int dm_hold(struct mapped_device *md)
2637{
2638 spin_lock(&_minor_lock);
2639 if (test_bit(DMF_FREEING, &md->flags)) {
2640 spin_unlock(&_minor_lock);
2641 return -EBUSY;
2642 }
2643 dm_get(md);
2644 spin_unlock(&_minor_lock);
2645 return 0;
2646}
2647EXPORT_SYMBOL_GPL(dm_hold);
2648
2649const char *dm_device_name(struct mapped_device *md)
2650{
2651 return md->name;
2652}
2653EXPORT_SYMBOL_GPL(dm_device_name);
2654
2655static void __dm_destroy(struct mapped_device *md, bool wait)
2656{
2657 struct dm_table *map;
2658 int srcu_idx;
2659
2660 might_sleep();
2661
2662 spin_lock(&_minor_lock);
2663 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2664 set_bit(DMF_FREEING, &md->flags);
2665 spin_unlock(&_minor_lock);
2666
2667 blk_mark_disk_dead(md->disk);
2668
2669 /*
2670 * Take suspend_lock so that presuspend and postsuspend methods
2671 * do not race with internal suspend.
2672 */
2673 mutex_lock(&md->suspend_lock);
2674 map = dm_get_live_table(md, &srcu_idx);
2675 if (!dm_suspended_md(md)) {
2676 dm_table_presuspend_targets(map);
2677 set_bit(DMF_SUSPENDED, &md->flags);
2678 set_bit(DMF_POST_SUSPENDING, &md->flags);
2679 dm_table_postsuspend_targets(map);
2680 }
2681 /* dm_put_live_table must be before fsleep, otherwise deadlock is possible */
2682 dm_put_live_table(md, srcu_idx);
2683 mutex_unlock(&md->suspend_lock);
2684
2685 /*
2686 * Rare, but there may be I/O requests still going to complete,
2687 * for example. Wait for all references to disappear.
2688 * No one should increment the reference count of the mapped_device,
2689 * after the mapped_device state becomes DMF_FREEING.
2690 */
2691 if (wait)
2692 while (atomic_read(&md->holders))
2693 fsleep(1000);
2694 else if (atomic_read(&md->holders))
2695 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2696 dm_device_name(md), atomic_read(&md->holders));
2697
2698 dm_table_destroy(__unbind(md));
2699 free_dev(md);
2700}
2701
2702void dm_destroy(struct mapped_device *md)
2703{
2704 __dm_destroy(md, true);
2705}
2706
2707void dm_destroy_immediate(struct mapped_device *md)
2708{
2709 __dm_destroy(md, false);
2710}
2711
2712void dm_put(struct mapped_device *md)
2713{
2714 atomic_dec(&md->holders);
2715}
2716EXPORT_SYMBOL_GPL(dm_put);
2717
2718static bool dm_in_flight_bios(struct mapped_device *md)
2719{
2720 int cpu;
2721 unsigned long sum = 0;
2722
2723 for_each_possible_cpu(cpu)
2724 sum += *per_cpu_ptr(md->pending_io, cpu);
2725
2726 return sum != 0;
2727}
2728
2729static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2730{
2731 int r = 0;
2732 DEFINE_WAIT(wait);
2733
2734 while (true) {
2735 prepare_to_wait(&md->wait, &wait, task_state);
2736
2737 if (!dm_in_flight_bios(md))
2738 break;
2739
2740 if (signal_pending_state(task_state, current)) {
2741 r = -ERESTARTSYS;
2742 break;
2743 }
2744
2745 io_schedule();
2746 }
2747 finish_wait(&md->wait, &wait);
2748
2749 smp_rmb();
2750
2751 return r;
2752}
2753
2754static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2755{
2756 int r = 0;
2757
2758 if (!queue_is_mq(md->queue))
2759 return dm_wait_for_bios_completion(md, task_state);
2760
2761 while (true) {
2762 if (!blk_mq_queue_inflight(md->queue))
2763 break;
2764
2765 if (signal_pending_state(task_state, current)) {
2766 r = -ERESTARTSYS;
2767 break;
2768 }
2769
2770 fsleep(5000);
2771 }
2772
2773 return r;
2774}
2775
2776/*
2777 * Process the deferred bios
2778 */
2779static void dm_wq_work(struct work_struct *work)
2780{
2781 struct mapped_device *md = container_of(work, struct mapped_device, work);
2782 struct bio *bio;
2783
2784 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2785 spin_lock_irq(&md->deferred_lock);
2786 bio = bio_list_pop(&md->deferred);
2787 spin_unlock_irq(&md->deferred_lock);
2788
2789 if (!bio)
2790 break;
2791
2792 submit_bio_noacct(bio);
2793 cond_resched();
2794 }
2795}
2796
2797static void dm_queue_flush(struct mapped_device *md)
2798{
2799 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2800 smp_mb__after_atomic();
2801 queue_work(md->wq, &md->work);
2802}
2803
2804/*
2805 * Swap in a new table, returning the old one for the caller to destroy.
2806 */
2807struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2808{
2809 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2810 struct queue_limits limits;
2811 int r;
2812
2813 mutex_lock(&md->suspend_lock);
2814
2815 /* device must be suspended */
2816 if (!dm_suspended_md(md))
2817 goto out;
2818
2819 /*
2820 * If the new table has no data devices, retain the existing limits.
2821 * This helps multipath with queue_if_no_path if all paths disappear,
2822 * then new I/O is queued based on these limits, and then some paths
2823 * reappear.
2824 */
2825 if (dm_table_has_no_data_devices(table)) {
2826 live_map = dm_get_live_table_fast(md);
2827 if (live_map)
2828 limits = md->queue->limits;
2829 dm_put_live_table_fast(md);
2830 }
2831
2832 if (!live_map) {
2833 r = dm_calculate_queue_limits(table, &limits);
2834 if (r) {
2835 map = ERR_PTR(r);
2836 goto out;
2837 }
2838 }
2839
2840 map = __bind(md, table, &limits);
2841 dm_issue_global_event();
2842
2843out:
2844 mutex_unlock(&md->suspend_lock);
2845 return map;
2846}
2847
2848/*
2849 * Functions to lock and unlock any filesystem running on the
2850 * device.
2851 */
2852static int lock_fs(struct mapped_device *md)
2853{
2854 int r;
2855
2856 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2857
2858 r = bdev_freeze(md->disk->part0);
2859 if (!r)
2860 set_bit(DMF_FROZEN, &md->flags);
2861 return r;
2862}
2863
2864static void unlock_fs(struct mapped_device *md)
2865{
2866 if (!test_bit(DMF_FROZEN, &md->flags))
2867 return;
2868 bdev_thaw(md->disk->part0);
2869 clear_bit(DMF_FROZEN, &md->flags);
2870}
2871
2872/*
2873 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2874 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2875 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2876 *
2877 * If __dm_suspend returns 0, the device is completely quiescent
2878 * now. There is no request-processing activity. All new requests
2879 * are being added to md->deferred list.
2880 */
2881static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2882 unsigned int suspend_flags, unsigned int task_state,
2883 int dmf_suspended_flag)
2884{
2885 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2886 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2887 int r;
2888
2889 lockdep_assert_held(&md->suspend_lock);
2890
2891 /*
2892 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2893 * This flag is cleared before dm_suspend returns.
2894 */
2895 if (noflush)
2896 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2897 else
2898 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2899
2900 /*
2901 * This gets reverted if there's an error later and the targets
2902 * provide the .presuspend_undo hook.
2903 */
2904 dm_table_presuspend_targets(map);
2905
2906 /*
2907 * Flush I/O to the device.
2908 * Any I/O submitted after lock_fs() may not be flushed.
2909 * noflush takes precedence over do_lockfs.
2910 * (lock_fs() flushes I/Os and waits for them to complete.)
2911 */
2912 if (!noflush && do_lockfs) {
2913 r = lock_fs(md);
2914 if (r) {
2915 dm_table_presuspend_undo_targets(map);
2916 return r;
2917 }
2918 }
2919
2920 /*
2921 * Here we must make sure that no processes are submitting requests
2922 * to target drivers i.e. no one may be executing
2923 * dm_split_and_process_bio from dm_submit_bio.
2924 *
2925 * To get all processes out of dm_split_and_process_bio in dm_submit_bio,
2926 * we take the write lock. To prevent any process from reentering
2927 * dm_split_and_process_bio from dm_submit_bio and quiesce the thread
2928 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2929 * flush_workqueue(md->wq).
2930 */
2931 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2932 if (map)
2933 synchronize_srcu(&md->io_barrier);
2934
2935 /*
2936 * Stop md->queue before flushing md->wq in case request-based
2937 * dm defers requests to md->wq from md->queue.
2938 */
2939 if (dm_request_based(md))
2940 dm_stop_queue(md->queue);
2941
2942 flush_workqueue(md->wq);
2943
2944 /*
2945 * At this point no more requests are entering target request routines.
2946 * We call dm_wait_for_completion to wait for all existing requests
2947 * to finish.
2948 */
2949 r = dm_wait_for_completion(md, task_state);
2950 if (!r)
2951 set_bit(dmf_suspended_flag, &md->flags);
2952
2953 if (noflush)
2954 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2955 if (map)
2956 synchronize_srcu(&md->io_barrier);
2957
2958 /* were we interrupted ? */
2959 if (r < 0) {
2960 dm_queue_flush(md);
2961
2962 if (dm_request_based(md))
2963 dm_start_queue(md->queue);
2964
2965 unlock_fs(md);
2966 dm_table_presuspend_undo_targets(map);
2967 /* pushback list is already flushed, so skip flush */
2968 }
2969
2970 return r;
2971}
2972
2973/*
2974 * We need to be able to change a mapping table under a mounted
2975 * filesystem. For example we might want to move some data in
2976 * the background. Before the table can be swapped with
2977 * dm_bind_table, dm_suspend must be called to flush any in
2978 * flight bios and ensure that any further io gets deferred.
2979 */
2980/*
2981 * Suspend mechanism in request-based dm.
2982 *
2983 * 1. Flush all I/Os by lock_fs() if needed.
2984 * 2. Stop dispatching any I/O by stopping the request_queue.
2985 * 3. Wait for all in-flight I/Os to be completed or requeued.
2986 *
2987 * To abort suspend, start the request_queue.
2988 */
2989int dm_suspend(struct mapped_device *md, unsigned int suspend_flags)
2990{
2991 struct dm_table *map = NULL;
2992 int r = 0;
2993
2994retry:
2995 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2996
2997 if (dm_suspended_md(md)) {
2998 r = -EINVAL;
2999 goto out_unlock;
3000 }
3001
3002 if (dm_suspended_internally_md(md)) {
3003 /* already internally suspended, wait for internal resume */
3004 mutex_unlock(&md->suspend_lock);
3005 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3006 if (r)
3007 return r;
3008 goto retry;
3009 }
3010
3011 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3012 if (!map) {
3013 /* avoid deadlock with fs/namespace.c:do_mount() */
3014 suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
3015 }
3016
3017 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
3018 if (r)
3019 goto out_unlock;
3020
3021 set_bit(DMF_POST_SUSPENDING, &md->flags);
3022 dm_table_postsuspend_targets(map);
3023 clear_bit(DMF_POST_SUSPENDING, &md->flags);
3024
3025out_unlock:
3026 mutex_unlock(&md->suspend_lock);
3027 return r;
3028}
3029
3030static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3031{
3032 if (map) {
3033 int r = dm_table_resume_targets(map);
3034
3035 if (r)
3036 return r;
3037 }
3038
3039 dm_queue_flush(md);
3040
3041 /*
3042 * Flushing deferred I/Os must be done after targets are resumed
3043 * so that mapping of targets can work correctly.
3044 * Request-based dm is queueing the deferred I/Os in its request_queue.
3045 */
3046 if (dm_request_based(md))
3047 dm_start_queue(md->queue);
3048
3049 unlock_fs(md);
3050
3051 return 0;
3052}
3053
3054int dm_resume(struct mapped_device *md)
3055{
3056 int r;
3057 struct dm_table *map = NULL;
3058
3059retry:
3060 r = -EINVAL;
3061 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3062
3063 if (!dm_suspended_md(md))
3064 goto out;
3065
3066 if (dm_suspended_internally_md(md)) {
3067 /* already internally suspended, wait for internal resume */
3068 mutex_unlock(&md->suspend_lock);
3069 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3070 if (r)
3071 return r;
3072 goto retry;
3073 }
3074
3075 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3076 if (!map || !dm_table_get_size(map))
3077 goto out;
3078
3079 r = __dm_resume(md, map);
3080 if (r)
3081 goto out;
3082
3083 clear_bit(DMF_SUSPENDED, &md->flags);
3084out:
3085 mutex_unlock(&md->suspend_lock);
3086
3087 return r;
3088}
3089
3090/*
3091 * Internal suspend/resume works like userspace-driven suspend. It waits
3092 * until all bios finish and prevents issuing new bios to the target drivers.
3093 * It may be used only from the kernel.
3094 */
3095
3096static void __dm_internal_suspend(struct mapped_device *md, unsigned int suspend_flags)
3097{
3098 struct dm_table *map = NULL;
3099
3100 lockdep_assert_held(&md->suspend_lock);
3101
3102 if (md->internal_suspend_count++)
3103 return; /* nested internal suspend */
3104
3105 if (dm_suspended_md(md)) {
3106 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3107 return; /* nest suspend */
3108 }
3109
3110 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3111
3112 /*
3113 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3114 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3115 * would require changing .presuspend to return an error -- avoid this
3116 * until there is a need for more elaborate variants of internal suspend.
3117 */
3118 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
3119 DMF_SUSPENDED_INTERNALLY);
3120
3121 set_bit(DMF_POST_SUSPENDING, &md->flags);
3122 dm_table_postsuspend_targets(map);
3123 clear_bit(DMF_POST_SUSPENDING, &md->flags);
3124}
3125
3126static void __dm_internal_resume(struct mapped_device *md)
3127{
3128 int r;
3129 struct dm_table *map;
3130
3131 BUG_ON(!md->internal_suspend_count);
3132
3133 if (--md->internal_suspend_count)
3134 return; /* resume from nested internal suspend */
3135
3136 if (dm_suspended_md(md))
3137 goto done; /* resume from nested suspend */
3138
3139 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3140 r = __dm_resume(md, map);
3141 if (r) {
3142 /*
3143 * If a preresume method of some target failed, we are in a
3144 * tricky situation. We can't return an error to the caller. We
3145 * can't fake success because then the "resume" and
3146 * "postsuspend" methods would not be paired correctly, and it
3147 * would break various targets, for example it would cause list
3148 * corruption in the "origin" target.
3149 *
3150 * So, we fake normal suspend here, to make sure that the
3151 * "resume" and "postsuspend" methods will be paired correctly.
3152 */
3153 DMERR("Preresume method failed: %d", r);
3154 set_bit(DMF_SUSPENDED, &md->flags);
3155 }
3156done:
3157 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3158 smp_mb__after_atomic();
3159 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3160}
3161
3162void dm_internal_suspend_noflush(struct mapped_device *md)
3163{
3164 mutex_lock(&md->suspend_lock);
3165 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3166 mutex_unlock(&md->suspend_lock);
3167}
3168EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3169
3170void dm_internal_resume(struct mapped_device *md)
3171{
3172 mutex_lock(&md->suspend_lock);
3173 __dm_internal_resume(md);
3174 mutex_unlock(&md->suspend_lock);
3175}
3176EXPORT_SYMBOL_GPL(dm_internal_resume);
3177
3178/*
3179 * Fast variants of internal suspend/resume hold md->suspend_lock,
3180 * which prevents interaction with userspace-driven suspend.
3181 */
3182
3183void dm_internal_suspend_fast(struct mapped_device *md)
3184{
3185 mutex_lock(&md->suspend_lock);
3186 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3187 return;
3188
3189 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3190 synchronize_srcu(&md->io_barrier);
3191 flush_workqueue(md->wq);
3192 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3193}
3194EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3195
3196void dm_internal_resume_fast(struct mapped_device *md)
3197{
3198 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3199 goto done;
3200
3201 dm_queue_flush(md);
3202
3203done:
3204 mutex_unlock(&md->suspend_lock);
3205}
3206EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3207
3208/*
3209 *---------------------------------------------------------------
3210 * Event notification.
3211 *---------------------------------------------------------------
3212 */
3213int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3214 unsigned int cookie, bool need_resize_uevent)
3215{
3216 int r;
3217 unsigned int noio_flag;
3218 char udev_cookie[DM_COOKIE_LENGTH];
3219 char *envp[3] = { NULL, NULL, NULL };
3220 char **envpp = envp;
3221 if (cookie) {
3222 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3223 DM_COOKIE_ENV_VAR_NAME, cookie);
3224 *envpp++ = udev_cookie;
3225 }
3226 if (need_resize_uevent) {
3227 *envpp++ = "RESIZE=1";
3228 }
3229
3230 noio_flag = memalloc_noio_save();
3231
3232 r = kobject_uevent_env(&disk_to_dev(md->disk)->kobj, action, envp);
3233
3234 memalloc_noio_restore(noio_flag);
3235
3236 return r;
3237}
3238
3239uint32_t dm_next_uevent_seq(struct mapped_device *md)
3240{
3241 return atomic_add_return(1, &md->uevent_seq);
3242}
3243
3244uint32_t dm_get_event_nr(struct mapped_device *md)
3245{
3246 return atomic_read(&md->event_nr);
3247}
3248
3249int dm_wait_event(struct mapped_device *md, int event_nr)
3250{
3251 return wait_event_interruptible(md->eventq,
3252 (event_nr != atomic_read(&md->event_nr)));
3253}
3254
3255void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3256{
3257 unsigned long flags;
3258
3259 spin_lock_irqsave(&md->uevent_lock, flags);
3260 list_add(elist, &md->uevent_list);
3261 spin_unlock_irqrestore(&md->uevent_lock, flags);
3262}
3263
3264/*
3265 * The gendisk is only valid as long as you have a reference
3266 * count on 'md'.
3267 */
3268struct gendisk *dm_disk(struct mapped_device *md)
3269{
3270 return md->disk;
3271}
3272EXPORT_SYMBOL_GPL(dm_disk);
3273
3274struct kobject *dm_kobject(struct mapped_device *md)
3275{
3276 return &md->kobj_holder.kobj;
3277}
3278
3279struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3280{
3281 struct mapped_device *md;
3282
3283 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3284
3285 spin_lock(&_minor_lock);
3286 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
3287 md = NULL;
3288 goto out;
3289 }
3290 dm_get(md);
3291out:
3292 spin_unlock(&_minor_lock);
3293
3294 return md;
3295}
3296
3297int dm_suspended_md(struct mapped_device *md)
3298{
3299 return test_bit(DMF_SUSPENDED, &md->flags);
3300}
3301
3302static int dm_post_suspending_md(struct mapped_device *md)
3303{
3304 return test_bit(DMF_POST_SUSPENDING, &md->flags);
3305}
3306
3307int dm_suspended_internally_md(struct mapped_device *md)
3308{
3309 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3310}
3311
3312int dm_test_deferred_remove_flag(struct mapped_device *md)
3313{
3314 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3315}
3316
3317int dm_suspended(struct dm_target *ti)
3318{
3319 return dm_suspended_md(ti->table->md);
3320}
3321EXPORT_SYMBOL_GPL(dm_suspended);
3322
3323int dm_post_suspending(struct dm_target *ti)
3324{
3325 return dm_post_suspending_md(ti->table->md);
3326}
3327EXPORT_SYMBOL_GPL(dm_post_suspending);
3328
3329int dm_noflush_suspending(struct dm_target *ti)
3330{
3331 return __noflush_suspending(ti->table->md);
3332}
3333EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3334
3335void dm_free_md_mempools(struct dm_md_mempools *pools)
3336{
3337 if (!pools)
3338 return;
3339
3340 bioset_exit(&pools->bs);
3341 bioset_exit(&pools->io_bs);
3342
3343 kfree(pools);
3344}
3345
3346struct dm_blkdev_id {
3347 u8 *id;
3348 enum blk_unique_id type;
3349};
3350
3351static int __dm_get_unique_id(struct dm_target *ti, struct dm_dev *dev,
3352 sector_t start, sector_t len, void *data)
3353{
3354 struct dm_blkdev_id *dm_id = data;
3355 const struct block_device_operations *fops = dev->bdev->bd_disk->fops;
3356
3357 if (!fops->get_unique_id)
3358 return 0;
3359
3360 return fops->get_unique_id(dev->bdev->bd_disk, dm_id->id, dm_id->type);
3361}
3362
3363/*
3364 * Allow access to get_unique_id() for the first device returning a
3365 * non-zero result. Reasonable use expects all devices to have the
3366 * same unique id.
3367 */
3368static int dm_blk_get_unique_id(struct gendisk *disk, u8 *id,
3369 enum blk_unique_id type)
3370{
3371 struct mapped_device *md = disk->private_data;
3372 struct dm_table *table;
3373 struct dm_target *ti;
3374 int ret = 0, srcu_idx;
3375
3376 struct dm_blkdev_id dm_id = {
3377 .id = id,
3378 .type = type,
3379 };
3380
3381 table = dm_get_live_table(md, &srcu_idx);
3382 if (!table || !dm_table_get_size(table))
3383 goto out;
3384
3385 /* We only support devices that have a single target */
3386 if (table->num_targets != 1)
3387 goto out;
3388 ti = dm_table_get_target(table, 0);
3389
3390 if (!ti->type->iterate_devices)
3391 goto out;
3392
3393 ret = ti->type->iterate_devices(ti, __dm_get_unique_id, &dm_id);
3394out:
3395 dm_put_live_table(md, srcu_idx);
3396 return ret;
3397}
3398
3399struct dm_pr {
3400 u64 old_key;
3401 u64 new_key;
3402 u32 flags;
3403 bool abort;
3404 bool fail_early;
3405 int ret;
3406 enum pr_type type;
3407 struct pr_keys *read_keys;
3408 struct pr_held_reservation *rsv;
3409};
3410
3411static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3412 struct dm_pr *pr)
3413{
3414 struct mapped_device *md = bdev->bd_disk->private_data;
3415 struct dm_table *table;
3416 struct dm_target *ti;
3417 int ret = -ENOTTY, srcu_idx;
3418
3419 table = dm_get_live_table(md, &srcu_idx);
3420 if (!table || !dm_table_get_size(table))
3421 goto out;
3422
3423 /* We only support devices that have a single target */
3424 if (table->num_targets != 1)
3425 goto out;
3426 ti = dm_table_get_target(table, 0);
3427
3428 if (dm_suspended_md(md)) {
3429 ret = -EAGAIN;
3430 goto out;
3431 }
3432
3433 ret = -EINVAL;
3434 if (!ti->type->iterate_devices)
3435 goto out;
3436
3437 ti->type->iterate_devices(ti, fn, pr);
3438 ret = 0;
3439out:
3440 dm_put_live_table(md, srcu_idx);
3441 return ret;
3442}
3443
3444/*
3445 * For register / unregister we need to manually call out to every path.
3446 */
3447static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3448 sector_t start, sector_t len, void *data)
3449{
3450 struct dm_pr *pr = data;
3451 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3452 int ret;
3453
3454 if (!ops || !ops->pr_register) {
3455 pr->ret = -EOPNOTSUPP;
3456 return -1;
3457 }
3458
3459 ret = ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3460 if (!ret)
3461 return 0;
3462
3463 if (!pr->ret)
3464 pr->ret = ret;
3465
3466 if (pr->fail_early)
3467 return -1;
3468
3469 return 0;
3470}
3471
3472static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3473 u32 flags)
3474{
3475 struct dm_pr pr = {
3476 .old_key = old_key,
3477 .new_key = new_key,
3478 .flags = flags,
3479 .fail_early = true,
3480 .ret = 0,
3481 };
3482 int ret;
3483
3484 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3485 if (ret) {
3486 /* Didn't even get to register a path */
3487 return ret;
3488 }
3489
3490 if (!pr.ret)
3491 return 0;
3492 ret = pr.ret;
3493
3494 if (!new_key)
3495 return ret;
3496
3497 /* unregister all paths if we failed to register any path */
3498 pr.old_key = new_key;
3499 pr.new_key = 0;
3500 pr.flags = 0;
3501 pr.fail_early = false;
3502 (void) dm_call_pr(bdev, __dm_pr_register, &pr);
3503 return ret;
3504}
3505
3506
3507static int __dm_pr_reserve(struct dm_target *ti, struct dm_dev *dev,
3508 sector_t start, sector_t len, void *data)
3509{
3510 struct dm_pr *pr = data;
3511 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3512
3513 if (!ops || !ops->pr_reserve) {
3514 pr->ret = -EOPNOTSUPP;
3515 return -1;
3516 }
3517
3518 pr->ret = ops->pr_reserve(dev->bdev, pr->old_key, pr->type, pr->flags);
3519 if (!pr->ret)
3520 return -1;
3521
3522 return 0;
3523}
3524
3525static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3526 u32 flags)
3527{
3528 struct dm_pr pr = {
3529 .old_key = key,
3530 .flags = flags,
3531 .type = type,
3532 .fail_early = false,
3533 .ret = 0,
3534 };
3535 int ret;
3536
3537 ret = dm_call_pr(bdev, __dm_pr_reserve, &pr);
3538 if (ret)
3539 return ret;
3540
3541 return pr.ret;
3542}
3543
3544/*
3545 * If there is a non-All Registrants type of reservation, the release must be
3546 * sent down the holding path. For the cases where there is no reservation or
3547 * the path is not the holder the device will also return success, so we must
3548 * try each path to make sure we got the correct path.
3549 */
3550static int __dm_pr_release(struct dm_target *ti, struct dm_dev *dev,
3551 sector_t start, sector_t len, void *data)
3552{
3553 struct dm_pr *pr = data;
3554 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3555
3556 if (!ops || !ops->pr_release) {
3557 pr->ret = -EOPNOTSUPP;
3558 return -1;
3559 }
3560
3561 pr->ret = ops->pr_release(dev->bdev, pr->old_key, pr->type);
3562 if (pr->ret)
3563 return -1;
3564
3565 return 0;
3566}
3567
3568static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3569{
3570 struct dm_pr pr = {
3571 .old_key = key,
3572 .type = type,
3573 .fail_early = false,
3574 };
3575 int ret;
3576
3577 ret = dm_call_pr(bdev, __dm_pr_release, &pr);
3578 if (ret)
3579 return ret;
3580
3581 return pr.ret;
3582}
3583
3584static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
3585 sector_t start, sector_t len, void *data)
3586{
3587 struct dm_pr *pr = data;
3588 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3589
3590 if (!ops || !ops->pr_preempt) {
3591 pr->ret = -EOPNOTSUPP;
3592 return -1;
3593 }
3594
3595 pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
3596 pr->abort);
3597 if (!pr->ret)
3598 return -1;
3599
3600 return 0;
3601}
3602
3603static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3604 enum pr_type type, bool abort)
3605{
3606 struct dm_pr pr = {
3607 .new_key = new_key,
3608 .old_key = old_key,
3609 .type = type,
3610 .fail_early = false,
3611 };
3612 int ret;
3613
3614 ret = dm_call_pr(bdev, __dm_pr_preempt, &pr);
3615 if (ret)
3616 return ret;
3617
3618 return pr.ret;
3619}
3620
3621static int dm_pr_clear(struct block_device *bdev, u64 key)
3622{
3623 struct mapped_device *md = bdev->bd_disk->private_data;
3624 const struct pr_ops *ops;
3625 int r, srcu_idx;
3626
3627 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3628 if (r < 0)
3629 goto out;
3630
3631 ops = bdev->bd_disk->fops->pr_ops;
3632 if (ops && ops->pr_clear)
3633 r = ops->pr_clear(bdev, key);
3634 else
3635 r = -EOPNOTSUPP;
3636out:
3637 dm_unprepare_ioctl(md, srcu_idx);
3638 return r;
3639}
3640
3641static int __dm_pr_read_keys(struct dm_target *ti, struct dm_dev *dev,
3642 sector_t start, sector_t len, void *data)
3643{
3644 struct dm_pr *pr = data;
3645 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3646
3647 if (!ops || !ops->pr_read_keys) {
3648 pr->ret = -EOPNOTSUPP;
3649 return -1;
3650 }
3651
3652 pr->ret = ops->pr_read_keys(dev->bdev, pr->read_keys);
3653 if (!pr->ret)
3654 return -1;
3655
3656 return 0;
3657}
3658
3659static int dm_pr_read_keys(struct block_device *bdev, struct pr_keys *keys)
3660{
3661 struct dm_pr pr = {
3662 .read_keys = keys,
3663 };
3664 int ret;
3665
3666 ret = dm_call_pr(bdev, __dm_pr_read_keys, &pr);
3667 if (ret)
3668 return ret;
3669
3670 return pr.ret;
3671}
3672
3673static int __dm_pr_read_reservation(struct dm_target *ti, struct dm_dev *dev,
3674 sector_t start, sector_t len, void *data)
3675{
3676 struct dm_pr *pr = data;
3677 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3678
3679 if (!ops || !ops->pr_read_reservation) {
3680 pr->ret = -EOPNOTSUPP;
3681 return -1;
3682 }
3683
3684 pr->ret = ops->pr_read_reservation(dev->bdev, pr->rsv);
3685 if (!pr->ret)
3686 return -1;
3687
3688 return 0;
3689}
3690
3691static int dm_pr_read_reservation(struct block_device *bdev,
3692 struct pr_held_reservation *rsv)
3693{
3694 struct dm_pr pr = {
3695 .rsv = rsv,
3696 };
3697 int ret;
3698
3699 ret = dm_call_pr(bdev, __dm_pr_read_reservation, &pr);
3700 if (ret)
3701 return ret;
3702
3703 return pr.ret;
3704}
3705
3706static const struct pr_ops dm_pr_ops = {
3707 .pr_register = dm_pr_register,
3708 .pr_reserve = dm_pr_reserve,
3709 .pr_release = dm_pr_release,
3710 .pr_preempt = dm_pr_preempt,
3711 .pr_clear = dm_pr_clear,
3712 .pr_read_keys = dm_pr_read_keys,
3713 .pr_read_reservation = dm_pr_read_reservation,
3714};
3715
3716static const struct block_device_operations dm_blk_dops = {
3717 .submit_bio = dm_submit_bio,
3718 .poll_bio = dm_poll_bio,
3719 .open = dm_blk_open,
3720 .release = dm_blk_close,
3721 .ioctl = dm_blk_ioctl,
3722 .getgeo = dm_blk_getgeo,
3723 .report_zones = dm_blk_report_zones,
3724 .get_unique_id = dm_blk_get_unique_id,
3725 .pr_ops = &dm_pr_ops,
3726 .owner = THIS_MODULE
3727};
3728
3729static const struct block_device_operations dm_rq_blk_dops = {
3730 .open = dm_blk_open,
3731 .release = dm_blk_close,
3732 .ioctl = dm_blk_ioctl,
3733 .getgeo = dm_blk_getgeo,
3734 .get_unique_id = dm_blk_get_unique_id,
3735 .pr_ops = &dm_pr_ops,
3736 .owner = THIS_MODULE
3737};
3738
3739static const struct dax_operations dm_dax_ops = {
3740 .direct_access = dm_dax_direct_access,
3741 .zero_page_range = dm_dax_zero_page_range,
3742 .recovery_write = dm_dax_recovery_write,
3743};
3744
3745/*
3746 * module hooks
3747 */
3748module_init(dm_init);
3749module_exit(dm_exit);
3750
3751module_param(major, uint, 0);
3752MODULE_PARM_DESC(major, "The major number of the device mapper");
3753
3754module_param(reserved_bio_based_ios, uint, 0644);
3755MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3756
3757module_param(dm_numa_node, int, 0644);
3758MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3759
3760module_param(swap_bios, int, 0644);
3761MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3762
3763MODULE_DESCRIPTION(DM_NAME " driver");
3764MODULE_AUTHOR("Joe Thornber <dm-devel@lists.linux.dev>");
3765MODULE_LICENSE("GPL");
1/*
2 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm.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
23#include <trace/events/block.h>
24
25#define DM_MSG_PREFIX "core"
26
27#ifdef CONFIG_PRINTK
28/*
29 * ratelimit state to be used in DMXXX_LIMIT().
30 */
31DEFINE_RATELIMIT_STATE(dm_ratelimit_state,
32 DEFAULT_RATELIMIT_INTERVAL,
33 DEFAULT_RATELIMIT_BURST);
34EXPORT_SYMBOL(dm_ratelimit_state);
35#endif
36
37/*
38 * Cookies are numeric values sent with CHANGE and REMOVE
39 * uevents while resuming, removing or renaming the device.
40 */
41#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
42#define DM_COOKIE_LENGTH 24
43
44static const char *_name = DM_NAME;
45
46static unsigned int major = 0;
47static unsigned int _major = 0;
48
49static DEFINE_IDR(_minor_idr);
50
51static DEFINE_SPINLOCK(_minor_lock);
52/*
53 * For bio-based dm.
54 * One of these is allocated per bio.
55 */
56struct dm_io {
57 struct mapped_device *md;
58 int error;
59 atomic_t io_count;
60 struct bio *bio;
61 unsigned long start_time;
62 spinlock_t endio_lock;
63};
64
65/*
66 * For bio-based dm.
67 * One of these is allocated per target within a bio. Hopefully
68 * this will be simplified out one day.
69 */
70struct dm_target_io {
71 struct dm_io *io;
72 struct dm_target *ti;
73 union map_info info;
74};
75
76/*
77 * For request-based dm.
78 * One of these is allocated per request.
79 */
80struct dm_rq_target_io {
81 struct mapped_device *md;
82 struct dm_target *ti;
83 struct request *orig, clone;
84 int error;
85 union map_info info;
86};
87
88/*
89 * For request-based dm.
90 * One of these is allocated per bio.
91 */
92struct dm_rq_clone_bio_info {
93 struct bio *orig;
94 struct dm_rq_target_io *tio;
95};
96
97union map_info *dm_get_mapinfo(struct bio *bio)
98{
99 if (bio && bio->bi_private)
100 return &((struct dm_target_io *)bio->bi_private)->info;
101 return NULL;
102}
103
104union map_info *dm_get_rq_mapinfo(struct request *rq)
105{
106 if (rq && rq->end_io_data)
107 return &((struct dm_rq_target_io *)rq->end_io_data)->info;
108 return NULL;
109}
110EXPORT_SYMBOL_GPL(dm_get_rq_mapinfo);
111
112#define MINOR_ALLOCED ((void *)-1)
113
114/*
115 * Bits for the md->flags field.
116 */
117#define DMF_BLOCK_IO_FOR_SUSPEND 0
118#define DMF_SUSPENDED 1
119#define DMF_FROZEN 2
120#define DMF_FREEING 3
121#define DMF_DELETING 4
122#define DMF_NOFLUSH_SUSPENDING 5
123#define DMF_MERGE_IS_OPTIONAL 6
124
125/*
126 * Work processed by per-device workqueue.
127 */
128struct mapped_device {
129 struct rw_semaphore io_lock;
130 struct mutex suspend_lock;
131 rwlock_t map_lock;
132 atomic_t holders;
133 atomic_t open_count;
134
135 unsigned long flags;
136
137 struct request_queue *queue;
138 unsigned type;
139 /* Protect queue and type against concurrent access. */
140 struct mutex type_lock;
141
142 struct target_type *immutable_target_type;
143
144 struct gendisk *disk;
145 char name[16];
146
147 void *interface_ptr;
148
149 /*
150 * A list of ios that arrived while we were suspended.
151 */
152 atomic_t pending[2];
153 wait_queue_head_t wait;
154 struct work_struct work;
155 struct bio_list deferred;
156 spinlock_t deferred_lock;
157
158 /*
159 * Processing queue (flush)
160 */
161 struct workqueue_struct *wq;
162
163 /*
164 * The current mapping.
165 */
166 struct dm_table *map;
167
168 /*
169 * io objects are allocated from here.
170 */
171 mempool_t *io_pool;
172 mempool_t *tio_pool;
173
174 struct bio_set *bs;
175
176 /*
177 * Event handling.
178 */
179 atomic_t event_nr;
180 wait_queue_head_t eventq;
181 atomic_t uevent_seq;
182 struct list_head uevent_list;
183 spinlock_t uevent_lock; /* Protect access to uevent_list */
184
185 /*
186 * freeze/thaw support require holding onto a super block
187 */
188 struct super_block *frozen_sb;
189 struct block_device *bdev;
190
191 /* forced geometry settings */
192 struct hd_geometry geometry;
193
194 /* sysfs handle */
195 struct kobject kobj;
196
197 /* zero-length flush that will be cloned and submitted to targets */
198 struct bio flush_bio;
199};
200
201/*
202 * For mempools pre-allocation at the table loading time.
203 */
204struct dm_md_mempools {
205 mempool_t *io_pool;
206 mempool_t *tio_pool;
207 struct bio_set *bs;
208};
209
210#define MIN_IOS 256
211static struct kmem_cache *_io_cache;
212static struct kmem_cache *_tio_cache;
213static struct kmem_cache *_rq_tio_cache;
214static struct kmem_cache *_rq_bio_info_cache;
215
216static int __init local_init(void)
217{
218 int r = -ENOMEM;
219
220 /* allocate a slab for the dm_ios */
221 _io_cache = KMEM_CACHE(dm_io, 0);
222 if (!_io_cache)
223 return r;
224
225 /* allocate a slab for the target ios */
226 _tio_cache = KMEM_CACHE(dm_target_io, 0);
227 if (!_tio_cache)
228 goto out_free_io_cache;
229
230 _rq_tio_cache = KMEM_CACHE(dm_rq_target_io, 0);
231 if (!_rq_tio_cache)
232 goto out_free_tio_cache;
233
234 _rq_bio_info_cache = KMEM_CACHE(dm_rq_clone_bio_info, 0);
235 if (!_rq_bio_info_cache)
236 goto out_free_rq_tio_cache;
237
238 r = dm_uevent_init();
239 if (r)
240 goto out_free_rq_bio_info_cache;
241
242 _major = major;
243 r = register_blkdev(_major, _name);
244 if (r < 0)
245 goto out_uevent_exit;
246
247 if (!_major)
248 _major = r;
249
250 return 0;
251
252out_uevent_exit:
253 dm_uevent_exit();
254out_free_rq_bio_info_cache:
255 kmem_cache_destroy(_rq_bio_info_cache);
256out_free_rq_tio_cache:
257 kmem_cache_destroy(_rq_tio_cache);
258out_free_tio_cache:
259 kmem_cache_destroy(_tio_cache);
260out_free_io_cache:
261 kmem_cache_destroy(_io_cache);
262
263 return r;
264}
265
266static void local_exit(void)
267{
268 kmem_cache_destroy(_rq_bio_info_cache);
269 kmem_cache_destroy(_rq_tio_cache);
270 kmem_cache_destroy(_tio_cache);
271 kmem_cache_destroy(_io_cache);
272 unregister_blkdev(_major, _name);
273 dm_uevent_exit();
274
275 _major = 0;
276
277 DMINFO("cleaned up");
278}
279
280static int (*_inits[])(void) __initdata = {
281 local_init,
282 dm_target_init,
283 dm_linear_init,
284 dm_stripe_init,
285 dm_io_init,
286 dm_kcopyd_init,
287 dm_interface_init,
288};
289
290static void (*_exits[])(void) = {
291 local_exit,
292 dm_target_exit,
293 dm_linear_exit,
294 dm_stripe_exit,
295 dm_io_exit,
296 dm_kcopyd_exit,
297 dm_interface_exit,
298};
299
300static int __init dm_init(void)
301{
302 const int count = ARRAY_SIZE(_inits);
303
304 int r, i;
305
306 for (i = 0; i < count; i++) {
307 r = _inits[i]();
308 if (r)
309 goto bad;
310 }
311
312 return 0;
313
314 bad:
315 while (i--)
316 _exits[i]();
317
318 return r;
319}
320
321static void __exit dm_exit(void)
322{
323 int i = ARRAY_SIZE(_exits);
324
325 while (i--)
326 _exits[i]();
327
328 /*
329 * Should be empty by this point.
330 */
331 idr_remove_all(&_minor_idr);
332 idr_destroy(&_minor_idr);
333}
334
335/*
336 * Block device functions
337 */
338int dm_deleting_md(struct mapped_device *md)
339{
340 return test_bit(DMF_DELETING, &md->flags);
341}
342
343static int dm_blk_open(struct block_device *bdev, fmode_t mode)
344{
345 struct mapped_device *md;
346
347 spin_lock(&_minor_lock);
348
349 md = bdev->bd_disk->private_data;
350 if (!md)
351 goto out;
352
353 if (test_bit(DMF_FREEING, &md->flags) ||
354 dm_deleting_md(md)) {
355 md = NULL;
356 goto out;
357 }
358
359 dm_get(md);
360 atomic_inc(&md->open_count);
361
362out:
363 spin_unlock(&_minor_lock);
364
365 return md ? 0 : -ENXIO;
366}
367
368static int dm_blk_close(struct gendisk *disk, fmode_t mode)
369{
370 struct mapped_device *md = disk->private_data;
371
372 spin_lock(&_minor_lock);
373
374 atomic_dec(&md->open_count);
375 dm_put(md);
376
377 spin_unlock(&_minor_lock);
378
379 return 0;
380}
381
382int dm_open_count(struct mapped_device *md)
383{
384 return atomic_read(&md->open_count);
385}
386
387/*
388 * Guarantees nothing is using the device before it's deleted.
389 */
390int dm_lock_for_deletion(struct mapped_device *md)
391{
392 int r = 0;
393
394 spin_lock(&_minor_lock);
395
396 if (dm_open_count(md))
397 r = -EBUSY;
398 else
399 set_bit(DMF_DELETING, &md->flags);
400
401 spin_unlock(&_minor_lock);
402
403 return r;
404}
405
406static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
407{
408 struct mapped_device *md = bdev->bd_disk->private_data;
409
410 return dm_get_geometry(md, geo);
411}
412
413static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
414 unsigned int cmd, unsigned long arg)
415{
416 struct mapped_device *md = bdev->bd_disk->private_data;
417 struct dm_table *map = dm_get_live_table(md);
418 struct dm_target *tgt;
419 int r = -ENOTTY;
420
421 if (!map || !dm_table_get_size(map))
422 goto out;
423
424 /* We only support devices that have a single target */
425 if (dm_table_get_num_targets(map) != 1)
426 goto out;
427
428 tgt = dm_table_get_target(map, 0);
429
430 if (dm_suspended_md(md)) {
431 r = -EAGAIN;
432 goto out;
433 }
434
435 if (tgt->type->ioctl)
436 r = tgt->type->ioctl(tgt, cmd, arg);
437
438out:
439 dm_table_put(map);
440
441 return r;
442}
443
444static struct dm_io *alloc_io(struct mapped_device *md)
445{
446 return mempool_alloc(md->io_pool, GFP_NOIO);
447}
448
449static void free_io(struct mapped_device *md, struct dm_io *io)
450{
451 mempool_free(io, md->io_pool);
452}
453
454static void free_tio(struct mapped_device *md, struct dm_target_io *tio)
455{
456 mempool_free(tio, md->tio_pool);
457}
458
459static struct dm_rq_target_io *alloc_rq_tio(struct mapped_device *md,
460 gfp_t gfp_mask)
461{
462 return mempool_alloc(md->tio_pool, gfp_mask);
463}
464
465static void free_rq_tio(struct dm_rq_target_io *tio)
466{
467 mempool_free(tio, tio->md->tio_pool);
468}
469
470static struct dm_rq_clone_bio_info *alloc_bio_info(struct mapped_device *md)
471{
472 return mempool_alloc(md->io_pool, GFP_ATOMIC);
473}
474
475static void free_bio_info(struct dm_rq_clone_bio_info *info)
476{
477 mempool_free(info, info->tio->md->io_pool);
478}
479
480static int md_in_flight(struct mapped_device *md)
481{
482 return atomic_read(&md->pending[READ]) +
483 atomic_read(&md->pending[WRITE]);
484}
485
486static void start_io_acct(struct dm_io *io)
487{
488 struct mapped_device *md = io->md;
489 int cpu;
490 int rw = bio_data_dir(io->bio);
491
492 io->start_time = jiffies;
493
494 cpu = part_stat_lock();
495 part_round_stats(cpu, &dm_disk(md)->part0);
496 part_stat_unlock();
497 atomic_set(&dm_disk(md)->part0.in_flight[rw],
498 atomic_inc_return(&md->pending[rw]));
499}
500
501static void end_io_acct(struct dm_io *io)
502{
503 struct mapped_device *md = io->md;
504 struct bio *bio = io->bio;
505 unsigned long duration = jiffies - io->start_time;
506 int pending, cpu;
507 int rw = bio_data_dir(bio);
508
509 cpu = part_stat_lock();
510 part_round_stats(cpu, &dm_disk(md)->part0);
511 part_stat_add(cpu, &dm_disk(md)->part0, ticks[rw], duration);
512 part_stat_unlock();
513
514 /*
515 * After this is decremented the bio must not be touched if it is
516 * a flush.
517 */
518 pending = atomic_dec_return(&md->pending[rw]);
519 atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
520 pending += atomic_read(&md->pending[rw^0x1]);
521
522 /* nudge anyone waiting on suspend queue */
523 if (!pending)
524 wake_up(&md->wait);
525}
526
527/*
528 * Add the bio to the list of deferred io.
529 */
530static void queue_io(struct mapped_device *md, struct bio *bio)
531{
532 unsigned long flags;
533
534 spin_lock_irqsave(&md->deferred_lock, flags);
535 bio_list_add(&md->deferred, bio);
536 spin_unlock_irqrestore(&md->deferred_lock, flags);
537 queue_work(md->wq, &md->work);
538}
539
540/*
541 * Everyone (including functions in this file), should use this
542 * function to access the md->map field, and make sure they call
543 * dm_table_put() when finished.
544 */
545struct dm_table *dm_get_live_table(struct mapped_device *md)
546{
547 struct dm_table *t;
548 unsigned long flags;
549
550 read_lock_irqsave(&md->map_lock, flags);
551 t = md->map;
552 if (t)
553 dm_table_get(t);
554 read_unlock_irqrestore(&md->map_lock, flags);
555
556 return t;
557}
558
559/*
560 * Get the geometry associated with a dm device
561 */
562int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
563{
564 *geo = md->geometry;
565
566 return 0;
567}
568
569/*
570 * Set the geometry of a device.
571 */
572int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
573{
574 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
575
576 if (geo->start > sz) {
577 DMWARN("Start sector is beyond the geometry limits.");
578 return -EINVAL;
579 }
580
581 md->geometry = *geo;
582
583 return 0;
584}
585
586/*-----------------------------------------------------------------
587 * CRUD START:
588 * A more elegant soln is in the works that uses the queue
589 * merge fn, unfortunately there are a couple of changes to
590 * the block layer that I want to make for this. So in the
591 * interests of getting something for people to use I give
592 * you this clearly demarcated crap.
593 *---------------------------------------------------------------*/
594
595static int __noflush_suspending(struct mapped_device *md)
596{
597 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
598}
599
600/*
601 * Decrements the number of outstanding ios that a bio has been
602 * cloned into, completing the original io if necc.
603 */
604static void dec_pending(struct dm_io *io, int error)
605{
606 unsigned long flags;
607 int io_error;
608 struct bio *bio;
609 struct mapped_device *md = io->md;
610
611 /* Push-back supersedes any I/O errors */
612 if (unlikely(error)) {
613 spin_lock_irqsave(&io->endio_lock, flags);
614 if (!(io->error > 0 && __noflush_suspending(md)))
615 io->error = error;
616 spin_unlock_irqrestore(&io->endio_lock, flags);
617 }
618
619 if (atomic_dec_and_test(&io->io_count)) {
620 if (io->error == DM_ENDIO_REQUEUE) {
621 /*
622 * Target requested pushing back the I/O.
623 */
624 spin_lock_irqsave(&md->deferred_lock, flags);
625 if (__noflush_suspending(md))
626 bio_list_add_head(&md->deferred, io->bio);
627 else
628 /* noflush suspend was interrupted. */
629 io->error = -EIO;
630 spin_unlock_irqrestore(&md->deferred_lock, flags);
631 }
632
633 io_error = io->error;
634 bio = io->bio;
635 end_io_acct(io);
636 free_io(md, io);
637
638 if (io_error == DM_ENDIO_REQUEUE)
639 return;
640
641 if ((bio->bi_rw & REQ_FLUSH) && bio->bi_size) {
642 /*
643 * Preflush done for flush with data, reissue
644 * without REQ_FLUSH.
645 */
646 bio->bi_rw &= ~REQ_FLUSH;
647 queue_io(md, bio);
648 } else {
649 /* done with normal IO or empty flush */
650 trace_block_bio_complete(md->queue, bio, io_error);
651 bio_endio(bio, io_error);
652 }
653 }
654}
655
656static void clone_endio(struct bio *bio, int error)
657{
658 int r = 0;
659 struct dm_target_io *tio = bio->bi_private;
660 struct dm_io *io = tio->io;
661 struct mapped_device *md = tio->io->md;
662 dm_endio_fn endio = tio->ti->type->end_io;
663
664 if (!bio_flagged(bio, BIO_UPTODATE) && !error)
665 error = -EIO;
666
667 if (endio) {
668 r = endio(tio->ti, bio, error, &tio->info);
669 if (r < 0 || r == DM_ENDIO_REQUEUE)
670 /*
671 * error and requeue request are handled
672 * in dec_pending().
673 */
674 error = r;
675 else if (r == DM_ENDIO_INCOMPLETE)
676 /* The target will handle the io */
677 return;
678 else if (r) {
679 DMWARN("unimplemented target endio return value: %d", r);
680 BUG();
681 }
682 }
683
684 /*
685 * Store md for cleanup instead of tio which is about to get freed.
686 */
687 bio->bi_private = md->bs;
688
689 free_tio(md, tio);
690 bio_put(bio);
691 dec_pending(io, error);
692}
693
694/*
695 * Partial completion handling for request-based dm
696 */
697static void end_clone_bio(struct bio *clone, int error)
698{
699 struct dm_rq_clone_bio_info *info = clone->bi_private;
700 struct dm_rq_target_io *tio = info->tio;
701 struct bio *bio = info->orig;
702 unsigned int nr_bytes = info->orig->bi_size;
703
704 bio_put(clone);
705
706 if (tio->error)
707 /*
708 * An error has already been detected on the request.
709 * Once error occurred, just let clone->end_io() handle
710 * the remainder.
711 */
712 return;
713 else if (error) {
714 /*
715 * Don't notice the error to the upper layer yet.
716 * The error handling decision is made by the target driver,
717 * when the request is completed.
718 */
719 tio->error = error;
720 return;
721 }
722
723 /*
724 * I/O for the bio successfully completed.
725 * Notice the data completion to the upper layer.
726 */
727
728 /*
729 * bios are processed from the head of the list.
730 * So the completing bio should always be rq->bio.
731 * If it's not, something wrong is happening.
732 */
733 if (tio->orig->bio != bio)
734 DMERR("bio completion is going in the middle of the request");
735
736 /*
737 * Update the original request.
738 * Do not use blk_end_request() here, because it may complete
739 * the original request before the clone, and break the ordering.
740 */
741 blk_update_request(tio->orig, 0, nr_bytes);
742}
743
744/*
745 * Don't touch any member of the md after calling this function because
746 * the md may be freed in dm_put() at the end of this function.
747 * Or do dm_get() before calling this function and dm_put() later.
748 */
749static void rq_completed(struct mapped_device *md, int rw, int run_queue)
750{
751 atomic_dec(&md->pending[rw]);
752
753 /* nudge anyone waiting on suspend queue */
754 if (!md_in_flight(md))
755 wake_up(&md->wait);
756
757 if (run_queue)
758 blk_run_queue(md->queue);
759
760 /*
761 * dm_put() must be at the end of this function. See the comment above
762 */
763 dm_put(md);
764}
765
766static void free_rq_clone(struct request *clone)
767{
768 struct dm_rq_target_io *tio = clone->end_io_data;
769
770 blk_rq_unprep_clone(clone);
771 free_rq_tio(tio);
772}
773
774/*
775 * Complete the clone and the original request.
776 * Must be called without queue lock.
777 */
778static void dm_end_request(struct request *clone, int error)
779{
780 int rw = rq_data_dir(clone);
781 struct dm_rq_target_io *tio = clone->end_io_data;
782 struct mapped_device *md = tio->md;
783 struct request *rq = tio->orig;
784
785 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
786 rq->errors = clone->errors;
787 rq->resid_len = clone->resid_len;
788
789 if (rq->sense)
790 /*
791 * We are using the sense buffer of the original
792 * request.
793 * So setting the length of the sense data is enough.
794 */
795 rq->sense_len = clone->sense_len;
796 }
797
798 free_rq_clone(clone);
799 blk_end_request_all(rq, error);
800 rq_completed(md, rw, true);
801}
802
803static void dm_unprep_request(struct request *rq)
804{
805 struct request *clone = rq->special;
806
807 rq->special = NULL;
808 rq->cmd_flags &= ~REQ_DONTPREP;
809
810 free_rq_clone(clone);
811}
812
813/*
814 * Requeue the original request of a clone.
815 */
816void dm_requeue_unmapped_request(struct request *clone)
817{
818 int rw = rq_data_dir(clone);
819 struct dm_rq_target_io *tio = clone->end_io_data;
820 struct mapped_device *md = tio->md;
821 struct request *rq = tio->orig;
822 struct request_queue *q = rq->q;
823 unsigned long flags;
824
825 dm_unprep_request(rq);
826
827 spin_lock_irqsave(q->queue_lock, flags);
828 blk_requeue_request(q, rq);
829 spin_unlock_irqrestore(q->queue_lock, flags);
830
831 rq_completed(md, rw, 0);
832}
833EXPORT_SYMBOL_GPL(dm_requeue_unmapped_request);
834
835static void __stop_queue(struct request_queue *q)
836{
837 blk_stop_queue(q);
838}
839
840static void stop_queue(struct request_queue *q)
841{
842 unsigned long flags;
843
844 spin_lock_irqsave(q->queue_lock, flags);
845 __stop_queue(q);
846 spin_unlock_irqrestore(q->queue_lock, flags);
847}
848
849static void __start_queue(struct request_queue *q)
850{
851 if (blk_queue_stopped(q))
852 blk_start_queue(q);
853}
854
855static void start_queue(struct request_queue *q)
856{
857 unsigned long flags;
858
859 spin_lock_irqsave(q->queue_lock, flags);
860 __start_queue(q);
861 spin_unlock_irqrestore(q->queue_lock, flags);
862}
863
864static void dm_done(struct request *clone, int error, bool mapped)
865{
866 int r = error;
867 struct dm_rq_target_io *tio = clone->end_io_data;
868 dm_request_endio_fn rq_end_io = NULL;
869
870 if (tio->ti) {
871 rq_end_io = tio->ti->type->rq_end_io;
872
873 if (mapped && rq_end_io)
874 r = rq_end_io(tio->ti, clone, error, &tio->info);
875 }
876
877 if (r <= 0)
878 /* The target wants to complete the I/O */
879 dm_end_request(clone, r);
880 else if (r == DM_ENDIO_INCOMPLETE)
881 /* The target will handle the I/O */
882 return;
883 else if (r == DM_ENDIO_REQUEUE)
884 /* The target wants to requeue the I/O */
885 dm_requeue_unmapped_request(clone);
886 else {
887 DMWARN("unimplemented target endio return value: %d", r);
888 BUG();
889 }
890}
891
892/*
893 * Request completion handler for request-based dm
894 */
895static void dm_softirq_done(struct request *rq)
896{
897 bool mapped = true;
898 struct request *clone = rq->completion_data;
899 struct dm_rq_target_io *tio = clone->end_io_data;
900
901 if (rq->cmd_flags & REQ_FAILED)
902 mapped = false;
903
904 dm_done(clone, tio->error, mapped);
905}
906
907/*
908 * Complete the clone and the original request with the error status
909 * through softirq context.
910 */
911static void dm_complete_request(struct request *clone, int error)
912{
913 struct dm_rq_target_io *tio = clone->end_io_data;
914 struct request *rq = tio->orig;
915
916 tio->error = error;
917 rq->completion_data = clone;
918 blk_complete_request(rq);
919}
920
921/*
922 * Complete the not-mapped clone and the original request with the error status
923 * through softirq context.
924 * Target's rq_end_io() function isn't called.
925 * This may be used when the target's map_rq() function fails.
926 */
927void dm_kill_unmapped_request(struct request *clone, int error)
928{
929 struct dm_rq_target_io *tio = clone->end_io_data;
930 struct request *rq = tio->orig;
931
932 rq->cmd_flags |= REQ_FAILED;
933 dm_complete_request(clone, error);
934}
935EXPORT_SYMBOL_GPL(dm_kill_unmapped_request);
936
937/*
938 * Called with the queue lock held
939 */
940static void end_clone_request(struct request *clone, int error)
941{
942 /*
943 * For just cleaning up the information of the queue in which
944 * the clone was dispatched.
945 * The clone is *NOT* freed actually here because it is alloced from
946 * dm own mempool and REQ_ALLOCED isn't set in clone->cmd_flags.
947 */
948 __blk_put_request(clone->q, clone);
949
950 /*
951 * Actual request completion is done in a softirq context which doesn't
952 * hold the queue lock. Otherwise, deadlock could occur because:
953 * - another request may be submitted by the upper level driver
954 * of the stacking during the completion
955 * - the submission which requires queue lock may be done
956 * against this queue
957 */
958 dm_complete_request(clone, error);
959}
960
961/*
962 * Return maximum size of I/O possible at the supplied sector up to the current
963 * target boundary.
964 */
965static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
966{
967 sector_t target_offset = dm_target_offset(ti, sector);
968
969 return ti->len - target_offset;
970}
971
972static sector_t max_io_len(sector_t sector, struct dm_target *ti)
973{
974 sector_t len = max_io_len_target_boundary(sector, ti);
975
976 /*
977 * Does the target need to split even further ?
978 */
979 if (ti->split_io) {
980 sector_t boundary;
981 sector_t offset = dm_target_offset(ti, sector);
982 boundary = ((offset + ti->split_io) & ~(ti->split_io - 1))
983 - offset;
984 if (len > boundary)
985 len = boundary;
986 }
987
988 return len;
989}
990
991static void __map_bio(struct dm_target *ti, struct bio *clone,
992 struct dm_target_io *tio)
993{
994 int r;
995 sector_t sector;
996 struct mapped_device *md;
997
998 clone->bi_end_io = clone_endio;
999 clone->bi_private = tio;
1000
1001 /*
1002 * Map the clone. If r == 0 we don't need to do
1003 * anything, the target has assumed ownership of
1004 * this io.
1005 */
1006 atomic_inc(&tio->io->io_count);
1007 sector = clone->bi_sector;
1008 r = ti->type->map(ti, clone, &tio->info);
1009 if (r == DM_MAPIO_REMAPPED) {
1010 /* the bio has been remapped so dispatch it */
1011
1012 trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1013 tio->io->bio->bi_bdev->bd_dev, sector);
1014
1015 generic_make_request(clone);
1016 } else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1017 /* error the io and bail out, or requeue it if needed */
1018 md = tio->io->md;
1019 dec_pending(tio->io, r);
1020 /*
1021 * Store bio_set for cleanup.
1022 */
1023 clone->bi_end_io = NULL;
1024 clone->bi_private = md->bs;
1025 bio_put(clone);
1026 free_tio(md, tio);
1027 } else if (r) {
1028 DMWARN("unimplemented target map return value: %d", r);
1029 BUG();
1030 }
1031}
1032
1033struct clone_info {
1034 struct mapped_device *md;
1035 struct dm_table *map;
1036 struct bio *bio;
1037 struct dm_io *io;
1038 sector_t sector;
1039 sector_t sector_count;
1040 unsigned short idx;
1041};
1042
1043static void dm_bio_destructor(struct bio *bio)
1044{
1045 struct bio_set *bs = bio->bi_private;
1046
1047 bio_free(bio, bs);
1048}
1049
1050/*
1051 * Creates a little bio that just does part of a bvec.
1052 */
1053static struct bio *split_bvec(struct bio *bio, sector_t sector,
1054 unsigned short idx, unsigned int offset,
1055 unsigned int len, struct bio_set *bs)
1056{
1057 struct bio *clone;
1058 struct bio_vec *bv = bio->bi_io_vec + idx;
1059
1060 clone = bio_alloc_bioset(GFP_NOIO, 1, bs);
1061 clone->bi_destructor = dm_bio_destructor;
1062 *clone->bi_io_vec = *bv;
1063
1064 clone->bi_sector = sector;
1065 clone->bi_bdev = bio->bi_bdev;
1066 clone->bi_rw = bio->bi_rw;
1067 clone->bi_vcnt = 1;
1068 clone->bi_size = to_bytes(len);
1069 clone->bi_io_vec->bv_offset = offset;
1070 clone->bi_io_vec->bv_len = clone->bi_size;
1071 clone->bi_flags |= 1 << BIO_CLONED;
1072
1073 if (bio_integrity(bio)) {
1074 bio_integrity_clone(clone, bio, GFP_NOIO, bs);
1075 bio_integrity_trim(clone,
1076 bio_sector_offset(bio, idx, offset), len);
1077 }
1078
1079 return clone;
1080}
1081
1082/*
1083 * Creates a bio that consists of range of complete bvecs.
1084 */
1085static struct bio *clone_bio(struct bio *bio, sector_t sector,
1086 unsigned short idx, unsigned short bv_count,
1087 unsigned int len, struct bio_set *bs)
1088{
1089 struct bio *clone;
1090
1091 clone = bio_alloc_bioset(GFP_NOIO, bio->bi_max_vecs, bs);
1092 __bio_clone(clone, bio);
1093 clone->bi_destructor = dm_bio_destructor;
1094 clone->bi_sector = sector;
1095 clone->bi_idx = idx;
1096 clone->bi_vcnt = idx + bv_count;
1097 clone->bi_size = to_bytes(len);
1098 clone->bi_flags &= ~(1 << BIO_SEG_VALID);
1099
1100 if (bio_integrity(bio)) {
1101 bio_integrity_clone(clone, bio, GFP_NOIO, bs);
1102
1103 if (idx != bio->bi_idx || clone->bi_size < bio->bi_size)
1104 bio_integrity_trim(clone,
1105 bio_sector_offset(bio, idx, 0), len);
1106 }
1107
1108 return clone;
1109}
1110
1111static struct dm_target_io *alloc_tio(struct clone_info *ci,
1112 struct dm_target *ti)
1113{
1114 struct dm_target_io *tio = mempool_alloc(ci->md->tio_pool, GFP_NOIO);
1115
1116 tio->io = ci->io;
1117 tio->ti = ti;
1118 memset(&tio->info, 0, sizeof(tio->info));
1119
1120 return tio;
1121}
1122
1123static void __issue_target_request(struct clone_info *ci, struct dm_target *ti,
1124 unsigned request_nr, sector_t len)
1125{
1126 struct dm_target_io *tio = alloc_tio(ci, ti);
1127 struct bio *clone;
1128
1129 tio->info.target_request_nr = request_nr;
1130
1131 /*
1132 * Discard requests require the bio's inline iovecs be initialized.
1133 * ci->bio->bi_max_vecs is BIO_INLINE_VECS anyway, for both flush
1134 * and discard, so no need for concern about wasted bvec allocations.
1135 */
1136 clone = bio_alloc_bioset(GFP_NOIO, ci->bio->bi_max_vecs, ci->md->bs);
1137 __bio_clone(clone, ci->bio);
1138 clone->bi_destructor = dm_bio_destructor;
1139 if (len) {
1140 clone->bi_sector = ci->sector;
1141 clone->bi_size = to_bytes(len);
1142 }
1143
1144 __map_bio(ti, clone, tio);
1145}
1146
1147static void __issue_target_requests(struct clone_info *ci, struct dm_target *ti,
1148 unsigned num_requests, sector_t len)
1149{
1150 unsigned request_nr;
1151
1152 for (request_nr = 0; request_nr < num_requests; request_nr++)
1153 __issue_target_request(ci, ti, request_nr, len);
1154}
1155
1156static int __clone_and_map_empty_flush(struct clone_info *ci)
1157{
1158 unsigned target_nr = 0;
1159 struct dm_target *ti;
1160
1161 BUG_ON(bio_has_data(ci->bio));
1162 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1163 __issue_target_requests(ci, ti, ti->num_flush_requests, 0);
1164
1165 return 0;
1166}
1167
1168/*
1169 * Perform all io with a single clone.
1170 */
1171static void __clone_and_map_simple(struct clone_info *ci, struct dm_target *ti)
1172{
1173 struct bio *clone, *bio = ci->bio;
1174 struct dm_target_io *tio;
1175
1176 tio = alloc_tio(ci, ti);
1177 clone = clone_bio(bio, ci->sector, ci->idx,
1178 bio->bi_vcnt - ci->idx, ci->sector_count,
1179 ci->md->bs);
1180 __map_bio(ti, clone, tio);
1181 ci->sector_count = 0;
1182}
1183
1184static int __clone_and_map_discard(struct clone_info *ci)
1185{
1186 struct dm_target *ti;
1187 sector_t len;
1188
1189 do {
1190 ti = dm_table_find_target(ci->map, ci->sector);
1191 if (!dm_target_is_valid(ti))
1192 return -EIO;
1193
1194 /*
1195 * Even though the device advertised discard support,
1196 * that does not mean every target supports it, and
1197 * reconfiguration might also have changed that since the
1198 * check was performed.
1199 */
1200 if (!ti->num_discard_requests)
1201 return -EOPNOTSUPP;
1202
1203 len = min(ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1204
1205 __issue_target_requests(ci, ti, ti->num_discard_requests, len);
1206
1207 ci->sector += len;
1208 } while (ci->sector_count -= len);
1209
1210 return 0;
1211}
1212
1213static int __clone_and_map(struct clone_info *ci)
1214{
1215 struct bio *clone, *bio = ci->bio;
1216 struct dm_target *ti;
1217 sector_t len = 0, max;
1218 struct dm_target_io *tio;
1219
1220 if (unlikely(bio->bi_rw & REQ_DISCARD))
1221 return __clone_and_map_discard(ci);
1222
1223 ti = dm_table_find_target(ci->map, ci->sector);
1224 if (!dm_target_is_valid(ti))
1225 return -EIO;
1226
1227 max = max_io_len(ci->sector, ti);
1228
1229 if (ci->sector_count <= max) {
1230 /*
1231 * Optimise for the simple case where we can do all of
1232 * the remaining io with a single clone.
1233 */
1234 __clone_and_map_simple(ci, ti);
1235
1236 } else if (to_sector(bio->bi_io_vec[ci->idx].bv_len) <= max) {
1237 /*
1238 * There are some bvecs that don't span targets.
1239 * Do as many of these as possible.
1240 */
1241 int i;
1242 sector_t remaining = max;
1243 sector_t bv_len;
1244
1245 for (i = ci->idx; remaining && (i < bio->bi_vcnt); i++) {
1246 bv_len = to_sector(bio->bi_io_vec[i].bv_len);
1247
1248 if (bv_len > remaining)
1249 break;
1250
1251 remaining -= bv_len;
1252 len += bv_len;
1253 }
1254
1255 tio = alloc_tio(ci, ti);
1256 clone = clone_bio(bio, ci->sector, ci->idx, i - ci->idx, len,
1257 ci->md->bs);
1258 __map_bio(ti, clone, tio);
1259
1260 ci->sector += len;
1261 ci->sector_count -= len;
1262 ci->idx = i;
1263
1264 } else {
1265 /*
1266 * Handle a bvec that must be split between two or more targets.
1267 */
1268 struct bio_vec *bv = bio->bi_io_vec + ci->idx;
1269 sector_t remaining = to_sector(bv->bv_len);
1270 unsigned int offset = 0;
1271
1272 do {
1273 if (offset) {
1274 ti = dm_table_find_target(ci->map, ci->sector);
1275 if (!dm_target_is_valid(ti))
1276 return -EIO;
1277
1278 max = max_io_len(ci->sector, ti);
1279 }
1280
1281 len = min(remaining, max);
1282
1283 tio = alloc_tio(ci, ti);
1284 clone = split_bvec(bio, ci->sector, ci->idx,
1285 bv->bv_offset + offset, len,
1286 ci->md->bs);
1287
1288 __map_bio(ti, clone, tio);
1289
1290 ci->sector += len;
1291 ci->sector_count -= len;
1292 offset += to_bytes(len);
1293 } while (remaining -= len);
1294
1295 ci->idx++;
1296 }
1297
1298 return 0;
1299}
1300
1301/*
1302 * Split the bio into several clones and submit it to targets.
1303 */
1304static void __split_and_process_bio(struct mapped_device *md, struct bio *bio)
1305{
1306 struct clone_info ci;
1307 int error = 0;
1308
1309 ci.map = dm_get_live_table(md);
1310 if (unlikely(!ci.map)) {
1311 bio_io_error(bio);
1312 return;
1313 }
1314
1315 ci.md = md;
1316 ci.io = alloc_io(md);
1317 ci.io->error = 0;
1318 atomic_set(&ci.io->io_count, 1);
1319 ci.io->bio = bio;
1320 ci.io->md = md;
1321 spin_lock_init(&ci.io->endio_lock);
1322 ci.sector = bio->bi_sector;
1323 ci.idx = bio->bi_idx;
1324
1325 start_io_acct(ci.io);
1326 if (bio->bi_rw & REQ_FLUSH) {
1327 ci.bio = &ci.md->flush_bio;
1328 ci.sector_count = 0;
1329 error = __clone_and_map_empty_flush(&ci);
1330 /* dec_pending submits any data associated with flush */
1331 } else {
1332 ci.bio = bio;
1333 ci.sector_count = bio_sectors(bio);
1334 while (ci.sector_count && !error)
1335 error = __clone_and_map(&ci);
1336 }
1337
1338 /* drop the extra reference count */
1339 dec_pending(ci.io, error);
1340 dm_table_put(ci.map);
1341}
1342/*-----------------------------------------------------------------
1343 * CRUD END
1344 *---------------------------------------------------------------*/
1345
1346static int dm_merge_bvec(struct request_queue *q,
1347 struct bvec_merge_data *bvm,
1348 struct bio_vec *biovec)
1349{
1350 struct mapped_device *md = q->queuedata;
1351 struct dm_table *map = dm_get_live_table(md);
1352 struct dm_target *ti;
1353 sector_t max_sectors;
1354 int max_size = 0;
1355
1356 if (unlikely(!map))
1357 goto out;
1358
1359 ti = dm_table_find_target(map, bvm->bi_sector);
1360 if (!dm_target_is_valid(ti))
1361 goto out_table;
1362
1363 /*
1364 * Find maximum amount of I/O that won't need splitting
1365 */
1366 max_sectors = min(max_io_len(bvm->bi_sector, ti),
1367 (sector_t) BIO_MAX_SECTORS);
1368 max_size = (max_sectors << SECTOR_SHIFT) - bvm->bi_size;
1369 if (max_size < 0)
1370 max_size = 0;
1371
1372 /*
1373 * merge_bvec_fn() returns number of bytes
1374 * it can accept at this offset
1375 * max is precomputed maximal io size
1376 */
1377 if (max_size && ti->type->merge)
1378 max_size = ti->type->merge(ti, bvm, biovec, max_size);
1379 /*
1380 * If the target doesn't support merge method and some of the devices
1381 * provided their merge_bvec method (we know this by looking at
1382 * queue_max_hw_sectors), then we can't allow bios with multiple vector
1383 * entries. So always set max_size to 0, and the code below allows
1384 * just one page.
1385 */
1386 else if (queue_max_hw_sectors(q) <= PAGE_SIZE >> 9)
1387
1388 max_size = 0;
1389
1390out_table:
1391 dm_table_put(map);
1392
1393out:
1394 /*
1395 * Always allow an entire first page
1396 */
1397 if (max_size <= biovec->bv_len && !(bvm->bi_size >> SECTOR_SHIFT))
1398 max_size = biovec->bv_len;
1399
1400 return max_size;
1401}
1402
1403/*
1404 * The request function that just remaps the bio built up by
1405 * dm_merge_bvec.
1406 */
1407static void _dm_request(struct request_queue *q, struct bio *bio)
1408{
1409 int rw = bio_data_dir(bio);
1410 struct mapped_device *md = q->queuedata;
1411 int cpu;
1412
1413 down_read(&md->io_lock);
1414
1415 cpu = part_stat_lock();
1416 part_stat_inc(cpu, &dm_disk(md)->part0, ios[rw]);
1417 part_stat_add(cpu, &dm_disk(md)->part0, sectors[rw], bio_sectors(bio));
1418 part_stat_unlock();
1419
1420 /* if we're suspended, we have to queue this io for later */
1421 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1422 up_read(&md->io_lock);
1423
1424 if (bio_rw(bio) != READA)
1425 queue_io(md, bio);
1426 else
1427 bio_io_error(bio);
1428 return;
1429 }
1430
1431 __split_and_process_bio(md, bio);
1432 up_read(&md->io_lock);
1433 return;
1434}
1435
1436static int dm_request_based(struct mapped_device *md)
1437{
1438 return blk_queue_stackable(md->queue);
1439}
1440
1441static void dm_request(struct request_queue *q, struct bio *bio)
1442{
1443 struct mapped_device *md = q->queuedata;
1444
1445 if (dm_request_based(md))
1446 blk_queue_bio(q, bio);
1447 else
1448 _dm_request(q, bio);
1449}
1450
1451void dm_dispatch_request(struct request *rq)
1452{
1453 int r;
1454
1455 if (blk_queue_io_stat(rq->q))
1456 rq->cmd_flags |= REQ_IO_STAT;
1457
1458 rq->start_time = jiffies;
1459 r = blk_insert_cloned_request(rq->q, rq);
1460 if (r)
1461 dm_complete_request(rq, r);
1462}
1463EXPORT_SYMBOL_GPL(dm_dispatch_request);
1464
1465static void dm_rq_bio_destructor(struct bio *bio)
1466{
1467 struct dm_rq_clone_bio_info *info = bio->bi_private;
1468 struct mapped_device *md = info->tio->md;
1469
1470 free_bio_info(info);
1471 bio_free(bio, md->bs);
1472}
1473
1474static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
1475 void *data)
1476{
1477 struct dm_rq_target_io *tio = data;
1478 struct mapped_device *md = tio->md;
1479 struct dm_rq_clone_bio_info *info = alloc_bio_info(md);
1480
1481 if (!info)
1482 return -ENOMEM;
1483
1484 info->orig = bio_orig;
1485 info->tio = tio;
1486 bio->bi_end_io = end_clone_bio;
1487 bio->bi_private = info;
1488 bio->bi_destructor = dm_rq_bio_destructor;
1489
1490 return 0;
1491}
1492
1493static int setup_clone(struct request *clone, struct request *rq,
1494 struct dm_rq_target_io *tio)
1495{
1496 int r;
1497
1498 r = blk_rq_prep_clone(clone, rq, tio->md->bs, GFP_ATOMIC,
1499 dm_rq_bio_constructor, tio);
1500 if (r)
1501 return r;
1502
1503 clone->cmd = rq->cmd;
1504 clone->cmd_len = rq->cmd_len;
1505 clone->sense = rq->sense;
1506 clone->buffer = rq->buffer;
1507 clone->end_io = end_clone_request;
1508 clone->end_io_data = tio;
1509
1510 return 0;
1511}
1512
1513static struct request *clone_rq(struct request *rq, struct mapped_device *md,
1514 gfp_t gfp_mask)
1515{
1516 struct request *clone;
1517 struct dm_rq_target_io *tio;
1518
1519 tio = alloc_rq_tio(md, gfp_mask);
1520 if (!tio)
1521 return NULL;
1522
1523 tio->md = md;
1524 tio->ti = NULL;
1525 tio->orig = rq;
1526 tio->error = 0;
1527 memset(&tio->info, 0, sizeof(tio->info));
1528
1529 clone = &tio->clone;
1530 if (setup_clone(clone, rq, tio)) {
1531 /* -ENOMEM */
1532 free_rq_tio(tio);
1533 return NULL;
1534 }
1535
1536 return clone;
1537}
1538
1539/*
1540 * Called with the queue lock held.
1541 */
1542static int dm_prep_fn(struct request_queue *q, struct request *rq)
1543{
1544 struct mapped_device *md = q->queuedata;
1545 struct request *clone;
1546
1547 if (unlikely(rq->special)) {
1548 DMWARN("Already has something in rq->special.");
1549 return BLKPREP_KILL;
1550 }
1551
1552 clone = clone_rq(rq, md, GFP_ATOMIC);
1553 if (!clone)
1554 return BLKPREP_DEFER;
1555
1556 rq->special = clone;
1557 rq->cmd_flags |= REQ_DONTPREP;
1558
1559 return BLKPREP_OK;
1560}
1561
1562/*
1563 * Returns:
1564 * 0 : the request has been processed (not requeued)
1565 * !0 : the request has been requeued
1566 */
1567static int map_request(struct dm_target *ti, struct request *clone,
1568 struct mapped_device *md)
1569{
1570 int r, requeued = 0;
1571 struct dm_rq_target_io *tio = clone->end_io_data;
1572
1573 tio->ti = ti;
1574 r = ti->type->map_rq(ti, clone, &tio->info);
1575 switch (r) {
1576 case DM_MAPIO_SUBMITTED:
1577 /* The target has taken the I/O to submit by itself later */
1578 break;
1579 case DM_MAPIO_REMAPPED:
1580 /* The target has remapped the I/O so dispatch it */
1581 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
1582 blk_rq_pos(tio->orig));
1583 dm_dispatch_request(clone);
1584 break;
1585 case DM_MAPIO_REQUEUE:
1586 /* The target wants to requeue the I/O */
1587 dm_requeue_unmapped_request(clone);
1588 requeued = 1;
1589 break;
1590 default:
1591 if (r > 0) {
1592 DMWARN("unimplemented target map return value: %d", r);
1593 BUG();
1594 }
1595
1596 /* The target wants to complete the I/O */
1597 dm_kill_unmapped_request(clone, r);
1598 break;
1599 }
1600
1601 return requeued;
1602}
1603
1604static struct request *dm_start_request(struct mapped_device *md, struct request *orig)
1605{
1606 struct request *clone;
1607
1608 blk_start_request(orig);
1609 clone = orig->special;
1610 atomic_inc(&md->pending[rq_data_dir(clone)]);
1611
1612 /*
1613 * Hold the md reference here for the in-flight I/O.
1614 * We can't rely on the reference count by device opener,
1615 * because the device may be closed during the request completion
1616 * when all bios are completed.
1617 * See the comment in rq_completed() too.
1618 */
1619 dm_get(md);
1620
1621 return clone;
1622}
1623
1624/*
1625 * q->request_fn for request-based dm.
1626 * Called with the queue lock held.
1627 */
1628static void dm_request_fn(struct request_queue *q)
1629{
1630 struct mapped_device *md = q->queuedata;
1631 struct dm_table *map = dm_get_live_table(md);
1632 struct dm_target *ti;
1633 struct request *rq, *clone;
1634 sector_t pos;
1635
1636 /*
1637 * For suspend, check blk_queue_stopped() and increment
1638 * ->pending within a single queue_lock not to increment the
1639 * number of in-flight I/Os after the queue is stopped in
1640 * dm_suspend().
1641 */
1642 while (!blk_queue_stopped(q)) {
1643 rq = blk_peek_request(q);
1644 if (!rq)
1645 goto delay_and_out;
1646
1647 /* always use block 0 to find the target for flushes for now */
1648 pos = 0;
1649 if (!(rq->cmd_flags & REQ_FLUSH))
1650 pos = blk_rq_pos(rq);
1651
1652 ti = dm_table_find_target(map, pos);
1653 if (!dm_target_is_valid(ti)) {
1654 /*
1655 * Must perform setup, that dm_done() requires,
1656 * before calling dm_kill_unmapped_request
1657 */
1658 DMERR_LIMIT("request attempted access beyond the end of device");
1659 clone = dm_start_request(md, rq);
1660 dm_kill_unmapped_request(clone, -EIO);
1661 continue;
1662 }
1663
1664 if (ti->type->busy && ti->type->busy(ti))
1665 goto delay_and_out;
1666
1667 clone = dm_start_request(md, rq);
1668
1669 spin_unlock(q->queue_lock);
1670 if (map_request(ti, clone, md))
1671 goto requeued;
1672
1673 BUG_ON(!irqs_disabled());
1674 spin_lock(q->queue_lock);
1675 }
1676
1677 goto out;
1678
1679requeued:
1680 BUG_ON(!irqs_disabled());
1681 spin_lock(q->queue_lock);
1682
1683delay_and_out:
1684 blk_delay_queue(q, HZ / 10);
1685out:
1686 dm_table_put(map);
1687}
1688
1689int dm_underlying_device_busy(struct request_queue *q)
1690{
1691 return blk_lld_busy(q);
1692}
1693EXPORT_SYMBOL_GPL(dm_underlying_device_busy);
1694
1695static int dm_lld_busy(struct request_queue *q)
1696{
1697 int r;
1698 struct mapped_device *md = q->queuedata;
1699 struct dm_table *map = dm_get_live_table(md);
1700
1701 if (!map || test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))
1702 r = 1;
1703 else
1704 r = dm_table_any_busy_target(map);
1705
1706 dm_table_put(map);
1707
1708 return r;
1709}
1710
1711static int dm_any_congested(void *congested_data, int bdi_bits)
1712{
1713 int r = bdi_bits;
1714 struct mapped_device *md = congested_data;
1715 struct dm_table *map;
1716
1717 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
1718 map = dm_get_live_table(md);
1719 if (map) {
1720 /*
1721 * Request-based dm cares about only own queue for
1722 * the query about congestion status of request_queue
1723 */
1724 if (dm_request_based(md))
1725 r = md->queue->backing_dev_info.state &
1726 bdi_bits;
1727 else
1728 r = dm_table_any_congested(map, bdi_bits);
1729
1730 dm_table_put(map);
1731 }
1732 }
1733
1734 return r;
1735}
1736
1737/*-----------------------------------------------------------------
1738 * An IDR is used to keep track of allocated minor numbers.
1739 *---------------------------------------------------------------*/
1740static void free_minor(int minor)
1741{
1742 spin_lock(&_minor_lock);
1743 idr_remove(&_minor_idr, minor);
1744 spin_unlock(&_minor_lock);
1745}
1746
1747/*
1748 * See if the device with a specific minor # is free.
1749 */
1750static int specific_minor(int minor)
1751{
1752 int r, m;
1753
1754 if (minor >= (1 << MINORBITS))
1755 return -EINVAL;
1756
1757 r = idr_pre_get(&_minor_idr, GFP_KERNEL);
1758 if (!r)
1759 return -ENOMEM;
1760
1761 spin_lock(&_minor_lock);
1762
1763 if (idr_find(&_minor_idr, minor)) {
1764 r = -EBUSY;
1765 goto out;
1766 }
1767
1768 r = idr_get_new_above(&_minor_idr, MINOR_ALLOCED, minor, &m);
1769 if (r)
1770 goto out;
1771
1772 if (m != minor) {
1773 idr_remove(&_minor_idr, m);
1774 r = -EBUSY;
1775 goto out;
1776 }
1777
1778out:
1779 spin_unlock(&_minor_lock);
1780 return r;
1781}
1782
1783static int next_free_minor(int *minor)
1784{
1785 int r, m;
1786
1787 r = idr_pre_get(&_minor_idr, GFP_KERNEL);
1788 if (!r)
1789 return -ENOMEM;
1790
1791 spin_lock(&_minor_lock);
1792
1793 r = idr_get_new(&_minor_idr, MINOR_ALLOCED, &m);
1794 if (r)
1795 goto out;
1796
1797 if (m >= (1 << MINORBITS)) {
1798 idr_remove(&_minor_idr, m);
1799 r = -ENOSPC;
1800 goto out;
1801 }
1802
1803 *minor = m;
1804
1805out:
1806 spin_unlock(&_minor_lock);
1807 return r;
1808}
1809
1810static const struct block_device_operations dm_blk_dops;
1811
1812static void dm_wq_work(struct work_struct *work);
1813
1814static void dm_init_md_queue(struct mapped_device *md)
1815{
1816 /*
1817 * Request-based dm devices cannot be stacked on top of bio-based dm
1818 * devices. The type of this dm device has not been decided yet.
1819 * The type is decided at the first table loading time.
1820 * To prevent problematic device stacking, clear the queue flag
1821 * for request stacking support until then.
1822 *
1823 * This queue is new, so no concurrency on the queue_flags.
1824 */
1825 queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
1826
1827 md->queue->queuedata = md;
1828 md->queue->backing_dev_info.congested_fn = dm_any_congested;
1829 md->queue->backing_dev_info.congested_data = md;
1830 blk_queue_make_request(md->queue, dm_request);
1831 blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
1832 blk_queue_merge_bvec(md->queue, dm_merge_bvec);
1833}
1834
1835/*
1836 * Allocate and initialise a blank device with a given minor.
1837 */
1838static struct mapped_device *alloc_dev(int minor)
1839{
1840 int r;
1841 struct mapped_device *md = kzalloc(sizeof(*md), GFP_KERNEL);
1842 void *old_md;
1843
1844 if (!md) {
1845 DMWARN("unable to allocate device, out of memory.");
1846 return NULL;
1847 }
1848
1849 if (!try_module_get(THIS_MODULE))
1850 goto bad_module_get;
1851
1852 /* get a minor number for the dev */
1853 if (minor == DM_ANY_MINOR)
1854 r = next_free_minor(&minor);
1855 else
1856 r = specific_minor(minor);
1857 if (r < 0)
1858 goto bad_minor;
1859
1860 md->type = DM_TYPE_NONE;
1861 init_rwsem(&md->io_lock);
1862 mutex_init(&md->suspend_lock);
1863 mutex_init(&md->type_lock);
1864 spin_lock_init(&md->deferred_lock);
1865 rwlock_init(&md->map_lock);
1866 atomic_set(&md->holders, 1);
1867 atomic_set(&md->open_count, 0);
1868 atomic_set(&md->event_nr, 0);
1869 atomic_set(&md->uevent_seq, 0);
1870 INIT_LIST_HEAD(&md->uevent_list);
1871 spin_lock_init(&md->uevent_lock);
1872
1873 md->queue = blk_alloc_queue(GFP_KERNEL);
1874 if (!md->queue)
1875 goto bad_queue;
1876
1877 dm_init_md_queue(md);
1878
1879 md->disk = alloc_disk(1);
1880 if (!md->disk)
1881 goto bad_disk;
1882
1883 atomic_set(&md->pending[0], 0);
1884 atomic_set(&md->pending[1], 0);
1885 init_waitqueue_head(&md->wait);
1886 INIT_WORK(&md->work, dm_wq_work);
1887 init_waitqueue_head(&md->eventq);
1888
1889 md->disk->major = _major;
1890 md->disk->first_minor = minor;
1891 md->disk->fops = &dm_blk_dops;
1892 md->disk->queue = md->queue;
1893 md->disk->private_data = md;
1894 sprintf(md->disk->disk_name, "dm-%d", minor);
1895 add_disk(md->disk);
1896 format_dev_t(md->name, MKDEV(_major, minor));
1897
1898 md->wq = alloc_workqueue("kdmflush",
1899 WQ_NON_REENTRANT | WQ_MEM_RECLAIM, 0);
1900 if (!md->wq)
1901 goto bad_thread;
1902
1903 md->bdev = bdget_disk(md->disk, 0);
1904 if (!md->bdev)
1905 goto bad_bdev;
1906
1907 bio_init(&md->flush_bio);
1908 md->flush_bio.bi_bdev = md->bdev;
1909 md->flush_bio.bi_rw = WRITE_FLUSH;
1910
1911 /* Populate the mapping, nobody knows we exist yet */
1912 spin_lock(&_minor_lock);
1913 old_md = idr_replace(&_minor_idr, md, minor);
1914 spin_unlock(&_minor_lock);
1915
1916 BUG_ON(old_md != MINOR_ALLOCED);
1917
1918 return md;
1919
1920bad_bdev:
1921 destroy_workqueue(md->wq);
1922bad_thread:
1923 del_gendisk(md->disk);
1924 put_disk(md->disk);
1925bad_disk:
1926 blk_cleanup_queue(md->queue);
1927bad_queue:
1928 free_minor(minor);
1929bad_minor:
1930 module_put(THIS_MODULE);
1931bad_module_get:
1932 kfree(md);
1933 return NULL;
1934}
1935
1936static void unlock_fs(struct mapped_device *md);
1937
1938static void free_dev(struct mapped_device *md)
1939{
1940 int minor = MINOR(disk_devt(md->disk));
1941
1942 unlock_fs(md);
1943 bdput(md->bdev);
1944 destroy_workqueue(md->wq);
1945 if (md->tio_pool)
1946 mempool_destroy(md->tio_pool);
1947 if (md->io_pool)
1948 mempool_destroy(md->io_pool);
1949 if (md->bs)
1950 bioset_free(md->bs);
1951 blk_integrity_unregister(md->disk);
1952 del_gendisk(md->disk);
1953 free_minor(minor);
1954
1955 spin_lock(&_minor_lock);
1956 md->disk->private_data = NULL;
1957 spin_unlock(&_minor_lock);
1958
1959 put_disk(md->disk);
1960 blk_cleanup_queue(md->queue);
1961 module_put(THIS_MODULE);
1962 kfree(md);
1963}
1964
1965static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
1966{
1967 struct dm_md_mempools *p;
1968
1969 if (md->io_pool && md->tio_pool && md->bs)
1970 /* the md already has necessary mempools */
1971 goto out;
1972
1973 p = dm_table_get_md_mempools(t);
1974 BUG_ON(!p || md->io_pool || md->tio_pool || md->bs);
1975
1976 md->io_pool = p->io_pool;
1977 p->io_pool = NULL;
1978 md->tio_pool = p->tio_pool;
1979 p->tio_pool = NULL;
1980 md->bs = p->bs;
1981 p->bs = NULL;
1982
1983out:
1984 /* mempool bind completed, now no need any mempools in the table */
1985 dm_table_free_md_mempools(t);
1986}
1987
1988/*
1989 * Bind a table to the device.
1990 */
1991static void event_callback(void *context)
1992{
1993 unsigned long flags;
1994 LIST_HEAD(uevents);
1995 struct mapped_device *md = (struct mapped_device *) context;
1996
1997 spin_lock_irqsave(&md->uevent_lock, flags);
1998 list_splice_init(&md->uevent_list, &uevents);
1999 spin_unlock_irqrestore(&md->uevent_lock, flags);
2000
2001 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2002
2003 atomic_inc(&md->event_nr);
2004 wake_up(&md->eventq);
2005}
2006
2007/*
2008 * Protected by md->suspend_lock obtained by dm_swap_table().
2009 */
2010static void __set_size(struct mapped_device *md, sector_t size)
2011{
2012 set_capacity(md->disk, size);
2013
2014 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2015}
2016
2017/*
2018 * Return 1 if the queue has a compulsory merge_bvec_fn function.
2019 *
2020 * If this function returns 0, then the device is either a non-dm
2021 * device without a merge_bvec_fn, or it is a dm device that is
2022 * able to split any bios it receives that are too big.
2023 */
2024int dm_queue_merge_is_compulsory(struct request_queue *q)
2025{
2026 struct mapped_device *dev_md;
2027
2028 if (!q->merge_bvec_fn)
2029 return 0;
2030
2031 if (q->make_request_fn == dm_request) {
2032 dev_md = q->queuedata;
2033 if (test_bit(DMF_MERGE_IS_OPTIONAL, &dev_md->flags))
2034 return 0;
2035 }
2036
2037 return 1;
2038}
2039
2040static int dm_device_merge_is_compulsory(struct dm_target *ti,
2041 struct dm_dev *dev, sector_t start,
2042 sector_t len, void *data)
2043{
2044 struct block_device *bdev = dev->bdev;
2045 struct request_queue *q = bdev_get_queue(bdev);
2046
2047 return dm_queue_merge_is_compulsory(q);
2048}
2049
2050/*
2051 * Return 1 if it is acceptable to ignore merge_bvec_fn based
2052 * on the properties of the underlying devices.
2053 */
2054static int dm_table_merge_is_optional(struct dm_table *table)
2055{
2056 unsigned i = 0;
2057 struct dm_target *ti;
2058
2059 while (i < dm_table_get_num_targets(table)) {
2060 ti = dm_table_get_target(table, i++);
2061
2062 if (ti->type->iterate_devices &&
2063 ti->type->iterate_devices(ti, dm_device_merge_is_compulsory, NULL))
2064 return 0;
2065 }
2066
2067 return 1;
2068}
2069
2070/*
2071 * Returns old map, which caller must destroy.
2072 */
2073static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2074 struct queue_limits *limits)
2075{
2076 struct dm_table *old_map;
2077 struct request_queue *q = md->queue;
2078 sector_t size;
2079 unsigned long flags;
2080 int merge_is_optional;
2081
2082 size = dm_table_get_size(t);
2083
2084 /*
2085 * Wipe any geometry if the size of the table changed.
2086 */
2087 if (size != get_capacity(md->disk))
2088 memset(&md->geometry, 0, sizeof(md->geometry));
2089
2090 __set_size(md, size);
2091
2092 dm_table_event_callback(t, event_callback, md);
2093
2094 /*
2095 * The queue hasn't been stopped yet, if the old table type wasn't
2096 * for request-based during suspension. So stop it to prevent
2097 * I/O mapping before resume.
2098 * This must be done before setting the queue restrictions,
2099 * because request-based dm may be run just after the setting.
2100 */
2101 if (dm_table_request_based(t) && !blk_queue_stopped(q))
2102 stop_queue(q);
2103
2104 __bind_mempools(md, t);
2105
2106 merge_is_optional = dm_table_merge_is_optional(t);
2107
2108 write_lock_irqsave(&md->map_lock, flags);
2109 old_map = md->map;
2110 md->map = t;
2111 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2112
2113 dm_table_set_restrictions(t, q, limits);
2114 if (merge_is_optional)
2115 set_bit(DMF_MERGE_IS_OPTIONAL, &md->flags);
2116 else
2117 clear_bit(DMF_MERGE_IS_OPTIONAL, &md->flags);
2118 write_unlock_irqrestore(&md->map_lock, flags);
2119
2120 return old_map;
2121}
2122
2123/*
2124 * Returns unbound table for the caller to free.
2125 */
2126static struct dm_table *__unbind(struct mapped_device *md)
2127{
2128 struct dm_table *map = md->map;
2129 unsigned long flags;
2130
2131 if (!map)
2132 return NULL;
2133
2134 dm_table_event_callback(map, NULL, NULL);
2135 write_lock_irqsave(&md->map_lock, flags);
2136 md->map = NULL;
2137 write_unlock_irqrestore(&md->map_lock, flags);
2138
2139 return map;
2140}
2141
2142/*
2143 * Constructor for a new device.
2144 */
2145int dm_create(int minor, struct mapped_device **result)
2146{
2147 struct mapped_device *md;
2148
2149 md = alloc_dev(minor);
2150 if (!md)
2151 return -ENXIO;
2152
2153 dm_sysfs_init(md);
2154
2155 *result = md;
2156 return 0;
2157}
2158
2159/*
2160 * Functions to manage md->type.
2161 * All are required to hold md->type_lock.
2162 */
2163void dm_lock_md_type(struct mapped_device *md)
2164{
2165 mutex_lock(&md->type_lock);
2166}
2167
2168void dm_unlock_md_type(struct mapped_device *md)
2169{
2170 mutex_unlock(&md->type_lock);
2171}
2172
2173void dm_set_md_type(struct mapped_device *md, unsigned type)
2174{
2175 md->type = type;
2176}
2177
2178unsigned dm_get_md_type(struct mapped_device *md)
2179{
2180 return md->type;
2181}
2182
2183struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2184{
2185 return md->immutable_target_type;
2186}
2187
2188/*
2189 * Fully initialize a request-based queue (->elevator, ->request_fn, etc).
2190 */
2191static int dm_init_request_based_queue(struct mapped_device *md)
2192{
2193 struct request_queue *q = NULL;
2194
2195 if (md->queue->elevator)
2196 return 1;
2197
2198 /* Fully initialize the queue */
2199 q = blk_init_allocated_queue(md->queue, dm_request_fn, NULL);
2200 if (!q)
2201 return 0;
2202
2203 md->queue = q;
2204 dm_init_md_queue(md);
2205 blk_queue_softirq_done(md->queue, dm_softirq_done);
2206 blk_queue_prep_rq(md->queue, dm_prep_fn);
2207 blk_queue_lld_busy(md->queue, dm_lld_busy);
2208
2209 elv_register_queue(md->queue);
2210
2211 return 1;
2212}
2213
2214/*
2215 * Setup the DM device's queue based on md's type
2216 */
2217int dm_setup_md_queue(struct mapped_device *md)
2218{
2219 if ((dm_get_md_type(md) == DM_TYPE_REQUEST_BASED) &&
2220 !dm_init_request_based_queue(md)) {
2221 DMWARN("Cannot initialize queue for request-based mapped device");
2222 return -EINVAL;
2223 }
2224
2225 return 0;
2226}
2227
2228static struct mapped_device *dm_find_md(dev_t dev)
2229{
2230 struct mapped_device *md;
2231 unsigned minor = MINOR(dev);
2232
2233 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2234 return NULL;
2235
2236 spin_lock(&_minor_lock);
2237
2238 md = idr_find(&_minor_idr, minor);
2239 if (md && (md == MINOR_ALLOCED ||
2240 (MINOR(disk_devt(dm_disk(md))) != minor) ||
2241 dm_deleting_md(md) ||
2242 test_bit(DMF_FREEING, &md->flags))) {
2243 md = NULL;
2244 goto out;
2245 }
2246
2247out:
2248 spin_unlock(&_minor_lock);
2249
2250 return md;
2251}
2252
2253struct mapped_device *dm_get_md(dev_t dev)
2254{
2255 struct mapped_device *md = dm_find_md(dev);
2256
2257 if (md)
2258 dm_get(md);
2259
2260 return md;
2261}
2262EXPORT_SYMBOL_GPL(dm_get_md);
2263
2264void *dm_get_mdptr(struct mapped_device *md)
2265{
2266 return md->interface_ptr;
2267}
2268
2269void dm_set_mdptr(struct mapped_device *md, void *ptr)
2270{
2271 md->interface_ptr = ptr;
2272}
2273
2274void dm_get(struct mapped_device *md)
2275{
2276 atomic_inc(&md->holders);
2277 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2278}
2279
2280const char *dm_device_name(struct mapped_device *md)
2281{
2282 return md->name;
2283}
2284EXPORT_SYMBOL_GPL(dm_device_name);
2285
2286static void __dm_destroy(struct mapped_device *md, bool wait)
2287{
2288 struct dm_table *map;
2289
2290 might_sleep();
2291
2292 spin_lock(&_minor_lock);
2293 map = dm_get_live_table(md);
2294 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2295 set_bit(DMF_FREEING, &md->flags);
2296 spin_unlock(&_minor_lock);
2297
2298 if (!dm_suspended_md(md)) {
2299 dm_table_presuspend_targets(map);
2300 dm_table_postsuspend_targets(map);
2301 }
2302
2303 /*
2304 * Rare, but there may be I/O requests still going to complete,
2305 * for example. Wait for all references to disappear.
2306 * No one should increment the reference count of the mapped_device,
2307 * after the mapped_device state becomes DMF_FREEING.
2308 */
2309 if (wait)
2310 while (atomic_read(&md->holders))
2311 msleep(1);
2312 else if (atomic_read(&md->holders))
2313 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2314 dm_device_name(md), atomic_read(&md->holders));
2315
2316 dm_sysfs_exit(md);
2317 dm_table_put(map);
2318 dm_table_destroy(__unbind(md));
2319 free_dev(md);
2320}
2321
2322void dm_destroy(struct mapped_device *md)
2323{
2324 __dm_destroy(md, true);
2325}
2326
2327void dm_destroy_immediate(struct mapped_device *md)
2328{
2329 __dm_destroy(md, false);
2330}
2331
2332void dm_put(struct mapped_device *md)
2333{
2334 atomic_dec(&md->holders);
2335}
2336EXPORT_SYMBOL_GPL(dm_put);
2337
2338static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
2339{
2340 int r = 0;
2341 DECLARE_WAITQUEUE(wait, current);
2342
2343 add_wait_queue(&md->wait, &wait);
2344
2345 while (1) {
2346 set_current_state(interruptible);
2347
2348 if (!md_in_flight(md))
2349 break;
2350
2351 if (interruptible == TASK_INTERRUPTIBLE &&
2352 signal_pending(current)) {
2353 r = -EINTR;
2354 break;
2355 }
2356
2357 io_schedule();
2358 }
2359 set_current_state(TASK_RUNNING);
2360
2361 remove_wait_queue(&md->wait, &wait);
2362
2363 return r;
2364}
2365
2366/*
2367 * Process the deferred bios
2368 */
2369static void dm_wq_work(struct work_struct *work)
2370{
2371 struct mapped_device *md = container_of(work, struct mapped_device,
2372 work);
2373 struct bio *c;
2374
2375 down_read(&md->io_lock);
2376
2377 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2378 spin_lock_irq(&md->deferred_lock);
2379 c = bio_list_pop(&md->deferred);
2380 spin_unlock_irq(&md->deferred_lock);
2381
2382 if (!c)
2383 break;
2384
2385 up_read(&md->io_lock);
2386
2387 if (dm_request_based(md))
2388 generic_make_request(c);
2389 else
2390 __split_and_process_bio(md, c);
2391
2392 down_read(&md->io_lock);
2393 }
2394
2395 up_read(&md->io_lock);
2396}
2397
2398static void dm_queue_flush(struct mapped_device *md)
2399{
2400 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2401 smp_mb__after_clear_bit();
2402 queue_work(md->wq, &md->work);
2403}
2404
2405/*
2406 * Swap in a new table, returning the old one for the caller to destroy.
2407 */
2408struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2409{
2410 struct dm_table *map = ERR_PTR(-EINVAL);
2411 struct queue_limits limits;
2412 int r;
2413
2414 mutex_lock(&md->suspend_lock);
2415
2416 /* device must be suspended */
2417 if (!dm_suspended_md(md))
2418 goto out;
2419
2420 r = dm_calculate_queue_limits(table, &limits);
2421 if (r) {
2422 map = ERR_PTR(r);
2423 goto out;
2424 }
2425
2426 map = __bind(md, table, &limits);
2427
2428out:
2429 mutex_unlock(&md->suspend_lock);
2430 return map;
2431}
2432
2433/*
2434 * Functions to lock and unlock any filesystem running on the
2435 * device.
2436 */
2437static int lock_fs(struct mapped_device *md)
2438{
2439 int r;
2440
2441 WARN_ON(md->frozen_sb);
2442
2443 md->frozen_sb = freeze_bdev(md->bdev);
2444 if (IS_ERR(md->frozen_sb)) {
2445 r = PTR_ERR(md->frozen_sb);
2446 md->frozen_sb = NULL;
2447 return r;
2448 }
2449
2450 set_bit(DMF_FROZEN, &md->flags);
2451
2452 return 0;
2453}
2454
2455static void unlock_fs(struct mapped_device *md)
2456{
2457 if (!test_bit(DMF_FROZEN, &md->flags))
2458 return;
2459
2460 thaw_bdev(md->bdev, md->frozen_sb);
2461 md->frozen_sb = NULL;
2462 clear_bit(DMF_FROZEN, &md->flags);
2463}
2464
2465/*
2466 * We need to be able to change a mapping table under a mounted
2467 * filesystem. For example we might want to move some data in
2468 * the background. Before the table can be swapped with
2469 * dm_bind_table, dm_suspend must be called to flush any in
2470 * flight bios and ensure that any further io gets deferred.
2471 */
2472/*
2473 * Suspend mechanism in request-based dm.
2474 *
2475 * 1. Flush all I/Os by lock_fs() if needed.
2476 * 2. Stop dispatching any I/O by stopping the request_queue.
2477 * 3. Wait for all in-flight I/Os to be completed or requeued.
2478 *
2479 * To abort suspend, start the request_queue.
2480 */
2481int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
2482{
2483 struct dm_table *map = NULL;
2484 int r = 0;
2485 int do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG ? 1 : 0;
2486 int noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG ? 1 : 0;
2487
2488 mutex_lock(&md->suspend_lock);
2489
2490 if (dm_suspended_md(md)) {
2491 r = -EINVAL;
2492 goto out_unlock;
2493 }
2494
2495 map = dm_get_live_table(md);
2496
2497 /*
2498 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2499 * This flag is cleared before dm_suspend returns.
2500 */
2501 if (noflush)
2502 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2503
2504 /* This does not get reverted if there's an error later. */
2505 dm_table_presuspend_targets(map);
2506
2507 /*
2508 * Flush I/O to the device.
2509 * Any I/O submitted after lock_fs() may not be flushed.
2510 * noflush takes precedence over do_lockfs.
2511 * (lock_fs() flushes I/Os and waits for them to complete.)
2512 */
2513 if (!noflush && do_lockfs) {
2514 r = lock_fs(md);
2515 if (r)
2516 goto out;
2517 }
2518
2519 /*
2520 * Here we must make sure that no processes are submitting requests
2521 * to target drivers i.e. no one may be executing
2522 * __split_and_process_bio. This is called from dm_request and
2523 * dm_wq_work.
2524 *
2525 * To get all processes out of __split_and_process_bio in dm_request,
2526 * we take the write lock. To prevent any process from reentering
2527 * __split_and_process_bio from dm_request and quiesce the thread
2528 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
2529 * flush_workqueue(md->wq).
2530 */
2531 down_write(&md->io_lock);
2532 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2533 up_write(&md->io_lock);
2534
2535 /*
2536 * Stop md->queue before flushing md->wq in case request-based
2537 * dm defers requests to md->wq from md->queue.
2538 */
2539 if (dm_request_based(md))
2540 stop_queue(md->queue);
2541
2542 flush_workqueue(md->wq);
2543
2544 /*
2545 * At this point no more requests are entering target request routines.
2546 * We call dm_wait_for_completion to wait for all existing requests
2547 * to finish.
2548 */
2549 r = dm_wait_for_completion(md, TASK_INTERRUPTIBLE);
2550
2551 down_write(&md->io_lock);
2552 if (noflush)
2553 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2554 up_write(&md->io_lock);
2555
2556 /* were we interrupted ? */
2557 if (r < 0) {
2558 dm_queue_flush(md);
2559
2560 if (dm_request_based(md))
2561 start_queue(md->queue);
2562
2563 unlock_fs(md);
2564 goto out; /* pushback list is already flushed, so skip flush */
2565 }
2566
2567 /*
2568 * If dm_wait_for_completion returned 0, the device is completely
2569 * quiescent now. There is no request-processing activity. All new
2570 * requests are being added to md->deferred list.
2571 */
2572
2573 set_bit(DMF_SUSPENDED, &md->flags);
2574
2575 dm_table_postsuspend_targets(map);
2576
2577out:
2578 dm_table_put(map);
2579
2580out_unlock:
2581 mutex_unlock(&md->suspend_lock);
2582 return r;
2583}
2584
2585int dm_resume(struct mapped_device *md)
2586{
2587 int r = -EINVAL;
2588 struct dm_table *map = NULL;
2589
2590 mutex_lock(&md->suspend_lock);
2591 if (!dm_suspended_md(md))
2592 goto out;
2593
2594 map = dm_get_live_table(md);
2595 if (!map || !dm_table_get_size(map))
2596 goto out;
2597
2598 r = dm_table_resume_targets(map);
2599 if (r)
2600 goto out;
2601
2602 dm_queue_flush(md);
2603
2604 /*
2605 * Flushing deferred I/Os must be done after targets are resumed
2606 * so that mapping of targets can work correctly.
2607 * Request-based dm is queueing the deferred I/Os in its request_queue.
2608 */
2609 if (dm_request_based(md))
2610 start_queue(md->queue);
2611
2612 unlock_fs(md);
2613
2614 clear_bit(DMF_SUSPENDED, &md->flags);
2615
2616 r = 0;
2617out:
2618 dm_table_put(map);
2619 mutex_unlock(&md->suspend_lock);
2620
2621 return r;
2622}
2623
2624/*-----------------------------------------------------------------
2625 * Event notification.
2626 *---------------------------------------------------------------*/
2627int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
2628 unsigned cookie)
2629{
2630 char udev_cookie[DM_COOKIE_LENGTH];
2631 char *envp[] = { udev_cookie, NULL };
2632
2633 if (!cookie)
2634 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
2635 else {
2636 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
2637 DM_COOKIE_ENV_VAR_NAME, cookie);
2638 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
2639 action, envp);
2640 }
2641}
2642
2643uint32_t dm_next_uevent_seq(struct mapped_device *md)
2644{
2645 return atomic_add_return(1, &md->uevent_seq);
2646}
2647
2648uint32_t dm_get_event_nr(struct mapped_device *md)
2649{
2650 return atomic_read(&md->event_nr);
2651}
2652
2653int dm_wait_event(struct mapped_device *md, int event_nr)
2654{
2655 return wait_event_interruptible(md->eventq,
2656 (event_nr != atomic_read(&md->event_nr)));
2657}
2658
2659void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
2660{
2661 unsigned long flags;
2662
2663 spin_lock_irqsave(&md->uevent_lock, flags);
2664 list_add(elist, &md->uevent_list);
2665 spin_unlock_irqrestore(&md->uevent_lock, flags);
2666}
2667
2668/*
2669 * The gendisk is only valid as long as you have a reference
2670 * count on 'md'.
2671 */
2672struct gendisk *dm_disk(struct mapped_device *md)
2673{
2674 return md->disk;
2675}
2676
2677struct kobject *dm_kobject(struct mapped_device *md)
2678{
2679 return &md->kobj;
2680}
2681
2682/*
2683 * struct mapped_device should not be exported outside of dm.c
2684 * so use this check to verify that kobj is part of md structure
2685 */
2686struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
2687{
2688 struct mapped_device *md;
2689
2690 md = container_of(kobj, struct mapped_device, kobj);
2691 if (&md->kobj != kobj)
2692 return NULL;
2693
2694 if (test_bit(DMF_FREEING, &md->flags) ||
2695 dm_deleting_md(md))
2696 return NULL;
2697
2698 dm_get(md);
2699 return md;
2700}
2701
2702int dm_suspended_md(struct mapped_device *md)
2703{
2704 return test_bit(DMF_SUSPENDED, &md->flags);
2705}
2706
2707int dm_suspended(struct dm_target *ti)
2708{
2709 return dm_suspended_md(dm_table_get_md(ti->table));
2710}
2711EXPORT_SYMBOL_GPL(dm_suspended);
2712
2713int dm_noflush_suspending(struct dm_target *ti)
2714{
2715 return __noflush_suspending(dm_table_get_md(ti->table));
2716}
2717EXPORT_SYMBOL_GPL(dm_noflush_suspending);
2718
2719struct dm_md_mempools *dm_alloc_md_mempools(unsigned type, unsigned integrity)
2720{
2721 struct dm_md_mempools *pools = kmalloc(sizeof(*pools), GFP_KERNEL);
2722 unsigned int pool_size = (type == DM_TYPE_BIO_BASED) ? 16 : MIN_IOS;
2723
2724 if (!pools)
2725 return NULL;
2726
2727 pools->io_pool = (type == DM_TYPE_BIO_BASED) ?
2728 mempool_create_slab_pool(MIN_IOS, _io_cache) :
2729 mempool_create_slab_pool(MIN_IOS, _rq_bio_info_cache);
2730 if (!pools->io_pool)
2731 goto free_pools_and_out;
2732
2733 pools->tio_pool = (type == DM_TYPE_BIO_BASED) ?
2734 mempool_create_slab_pool(MIN_IOS, _tio_cache) :
2735 mempool_create_slab_pool(MIN_IOS, _rq_tio_cache);
2736 if (!pools->tio_pool)
2737 goto free_io_pool_and_out;
2738
2739 pools->bs = bioset_create(pool_size, 0);
2740 if (!pools->bs)
2741 goto free_tio_pool_and_out;
2742
2743 if (integrity && bioset_integrity_create(pools->bs, pool_size))
2744 goto free_bioset_and_out;
2745
2746 return pools;
2747
2748free_bioset_and_out:
2749 bioset_free(pools->bs);
2750
2751free_tio_pool_and_out:
2752 mempool_destroy(pools->tio_pool);
2753
2754free_io_pool_and_out:
2755 mempool_destroy(pools->io_pool);
2756
2757free_pools_and_out:
2758 kfree(pools);
2759
2760 return NULL;
2761}
2762
2763void dm_free_md_mempools(struct dm_md_mempools *pools)
2764{
2765 if (!pools)
2766 return;
2767
2768 if (pools->io_pool)
2769 mempool_destroy(pools->io_pool);
2770
2771 if (pools->tio_pool)
2772 mempool_destroy(pools->tio_pool);
2773
2774 if (pools->bs)
2775 bioset_free(pools->bs);
2776
2777 kfree(pools);
2778}
2779
2780static const struct block_device_operations dm_blk_dops = {
2781 .open = dm_blk_open,
2782 .release = dm_blk_close,
2783 .ioctl = dm_blk_ioctl,
2784 .getgeo = dm_blk_getgeo,
2785 .owner = THIS_MODULE
2786};
2787
2788EXPORT_SYMBOL(dm_get_mapinfo);
2789
2790/*
2791 * module hooks
2792 */
2793module_init(dm_init);
2794module_exit(dm_exit);
2795
2796module_param(major, uint, 0);
2797MODULE_PARM_DESC(major, "The major number of the device mapper");
2798MODULE_DESCRIPTION(DM_NAME " driver");
2799MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2800MODULE_LICENSE("GPL");