Loading...
1/*
2 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm.h"
9#include "dm-uevent.h"
10
11#include <linux/init.h>
12#include <linux/module.h>
13#include <linux/mutex.h>
14#include <linux/moduleparam.h>
15#include <linux/blkpg.h>
16#include <linux/bio.h>
17#include <linux/mempool.h>
18#include <linux/slab.h>
19#include <linux/idr.h>
20#include <linux/hdreg.h>
21#include <linux/delay.h>
22#include <linux/wait.h>
23#include <linux/kthread.h>
24#include <linux/ktime.h>
25#include <linux/elevator.h> /* for rq_end_sector() */
26#include <linux/blk-mq.h>
27#include <linux/pr.h>
28
29#include <trace/events/block.h>
30
31#define DM_MSG_PREFIX "core"
32
33#ifdef CONFIG_PRINTK
34/*
35 * ratelimit state to be used in DMXXX_LIMIT().
36 */
37DEFINE_RATELIMIT_STATE(dm_ratelimit_state,
38 DEFAULT_RATELIMIT_INTERVAL,
39 DEFAULT_RATELIMIT_BURST);
40EXPORT_SYMBOL(dm_ratelimit_state);
41#endif
42
43/*
44 * Cookies are numeric values sent with CHANGE and REMOVE
45 * uevents while resuming, removing or renaming the device.
46 */
47#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
48#define DM_COOKIE_LENGTH 24
49
50static const char *_name = DM_NAME;
51
52static unsigned int major = 0;
53static unsigned int _major = 0;
54
55static DEFINE_IDR(_minor_idr);
56
57static DEFINE_SPINLOCK(_minor_lock);
58
59static void do_deferred_remove(struct work_struct *w);
60
61static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
62
63static struct workqueue_struct *deferred_remove_workqueue;
64
65/*
66 * For bio-based dm.
67 * One of these is allocated per bio.
68 */
69struct dm_io {
70 struct mapped_device *md;
71 int error;
72 atomic_t io_count;
73 struct bio *bio;
74 unsigned long start_time;
75 spinlock_t endio_lock;
76 struct dm_stats_aux stats_aux;
77};
78
79/*
80 * For request-based dm.
81 * One of these is allocated per request.
82 */
83struct dm_rq_target_io {
84 struct mapped_device *md;
85 struct dm_target *ti;
86 struct request *orig, *clone;
87 struct kthread_work work;
88 int error;
89 union map_info info;
90 struct dm_stats_aux stats_aux;
91 unsigned long duration_jiffies;
92 unsigned n_sectors;
93};
94
95/*
96 * For request-based dm - the bio clones we allocate are embedded in these
97 * structs.
98 *
99 * We allocate these with bio_alloc_bioset, using the front_pad parameter when
100 * the bioset is created - this means the bio has to come at the end of the
101 * struct.
102 */
103struct dm_rq_clone_bio_info {
104 struct bio *orig;
105 struct dm_rq_target_io *tio;
106 struct bio clone;
107};
108
109#define MINOR_ALLOCED ((void *)-1)
110
111/*
112 * Bits for the md->flags field.
113 */
114#define DMF_BLOCK_IO_FOR_SUSPEND 0
115#define DMF_SUSPENDED 1
116#define DMF_FROZEN 2
117#define DMF_FREEING 3
118#define DMF_DELETING 4
119#define DMF_NOFLUSH_SUSPENDING 5
120#define DMF_DEFERRED_REMOVE 6
121#define DMF_SUSPENDED_INTERNALLY 7
122
123/*
124 * Work processed by per-device workqueue.
125 */
126struct mapped_device {
127 struct srcu_struct io_barrier;
128 struct mutex suspend_lock;
129
130 /*
131 * The current mapping (struct dm_table *).
132 * Use dm_get_live_table{_fast} or take suspend_lock for
133 * dereference.
134 */
135 void __rcu *map;
136
137 struct list_head table_devices;
138 struct mutex table_devices_lock;
139
140 unsigned long flags;
141
142 struct request_queue *queue;
143 int numa_node_id;
144
145 unsigned type;
146 /* Protect queue and type against concurrent access. */
147 struct mutex type_lock;
148
149 atomic_t holders;
150 atomic_t open_count;
151
152 struct dm_target *immutable_target;
153 struct target_type *immutable_target_type;
154
155 struct gendisk *disk;
156 char name[16];
157
158 void *interface_ptr;
159
160 /*
161 * A list of ios that arrived while we were suspended.
162 */
163 atomic_t pending[2];
164 wait_queue_head_t wait;
165 struct work_struct work;
166 spinlock_t deferred_lock;
167 struct bio_list deferred;
168
169 /*
170 * Event handling.
171 */
172 wait_queue_head_t eventq;
173 atomic_t event_nr;
174 atomic_t uevent_seq;
175 struct list_head uevent_list;
176 spinlock_t uevent_lock; /* Protect access to uevent_list */
177
178 /* the number of internal suspends */
179 unsigned internal_suspend_count;
180
181 /*
182 * Processing queue (flush)
183 */
184 struct workqueue_struct *wq;
185
186 /*
187 * io objects are allocated from here.
188 */
189 mempool_t *io_pool;
190 mempool_t *rq_pool;
191
192 struct bio_set *bs;
193
194 /*
195 * freeze/thaw support require holding onto a super block
196 */
197 struct super_block *frozen_sb;
198
199 /* forced geometry settings */
200 struct hd_geometry geometry;
201
202 struct block_device *bdev;
203
204 /* kobject and completion */
205 struct dm_kobject_holder kobj_holder;
206
207 /* zero-length flush that will be cloned and submitted to targets */
208 struct bio flush_bio;
209
210 struct dm_stats stats;
211
212 struct kthread_worker kworker;
213 struct task_struct *kworker_task;
214
215 /* for request-based merge heuristic in dm_request_fn() */
216 unsigned seq_rq_merge_deadline_usecs;
217 int last_rq_rw;
218 sector_t last_rq_pos;
219 ktime_t last_rq_start_time;
220
221 /* for blk-mq request-based DM support */
222 struct blk_mq_tag_set *tag_set;
223 bool use_blk_mq:1;
224 bool init_tio_pdu:1;
225};
226
227#ifdef CONFIG_DM_MQ_DEFAULT
228static bool use_blk_mq = true;
229#else
230static bool use_blk_mq = false;
231#endif
232
233#define DM_MQ_NR_HW_QUEUES 1
234#define DM_MQ_QUEUE_DEPTH 2048
235#define DM_NUMA_NODE NUMA_NO_NODE
236
237static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
238static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
239static int dm_numa_node = DM_NUMA_NODE;
240
241bool dm_use_blk_mq(struct mapped_device *md)
242{
243 return md->use_blk_mq;
244}
245EXPORT_SYMBOL_GPL(dm_use_blk_mq);
246
247/*
248 * For mempools pre-allocation at the table loading time.
249 */
250struct dm_md_mempools {
251 mempool_t *io_pool;
252 mempool_t *rq_pool;
253 struct bio_set *bs;
254};
255
256struct table_device {
257 struct list_head list;
258 atomic_t count;
259 struct dm_dev dm_dev;
260};
261
262#define RESERVED_BIO_BASED_IOS 16
263#define RESERVED_REQUEST_BASED_IOS 256
264#define RESERVED_MAX_IOS 1024
265static struct kmem_cache *_io_cache;
266static struct kmem_cache *_rq_tio_cache;
267static struct kmem_cache *_rq_cache;
268
269/*
270 * Bio-based DM's mempools' reserved IOs set by the user.
271 */
272static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
273
274/*
275 * Request-based DM's mempools' reserved IOs set by the user.
276 */
277static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
278
279static int __dm_get_module_param_int(int *module_param, int min, int max)
280{
281 int param = ACCESS_ONCE(*module_param);
282 int modified_param = 0;
283 bool modified = true;
284
285 if (param < min)
286 modified_param = min;
287 else if (param > max)
288 modified_param = max;
289 else
290 modified = false;
291
292 if (modified) {
293 (void)cmpxchg(module_param, param, modified_param);
294 param = modified_param;
295 }
296
297 return param;
298}
299
300static unsigned __dm_get_module_param(unsigned *module_param,
301 unsigned def, unsigned max)
302{
303 unsigned param = ACCESS_ONCE(*module_param);
304 unsigned modified_param = 0;
305
306 if (!param)
307 modified_param = def;
308 else if (param > max)
309 modified_param = max;
310
311 if (modified_param) {
312 (void)cmpxchg(module_param, param, modified_param);
313 param = modified_param;
314 }
315
316 return param;
317}
318
319unsigned dm_get_reserved_bio_based_ios(void)
320{
321 return __dm_get_module_param(&reserved_bio_based_ios,
322 RESERVED_BIO_BASED_IOS, RESERVED_MAX_IOS);
323}
324EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
325
326unsigned dm_get_reserved_rq_based_ios(void)
327{
328 return __dm_get_module_param(&reserved_rq_based_ios,
329 RESERVED_REQUEST_BASED_IOS, RESERVED_MAX_IOS);
330}
331EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
332
333static unsigned dm_get_blk_mq_nr_hw_queues(void)
334{
335 return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
336}
337
338static unsigned dm_get_blk_mq_queue_depth(void)
339{
340 return __dm_get_module_param(&dm_mq_queue_depth,
341 DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
342}
343
344static unsigned dm_get_numa_node(void)
345{
346 return __dm_get_module_param_int(&dm_numa_node,
347 DM_NUMA_NODE, num_online_nodes() - 1);
348}
349
350static int __init local_init(void)
351{
352 int r = -ENOMEM;
353
354 /* allocate a slab for the dm_ios */
355 _io_cache = KMEM_CACHE(dm_io, 0);
356 if (!_io_cache)
357 return r;
358
359 _rq_tio_cache = KMEM_CACHE(dm_rq_target_io, 0);
360 if (!_rq_tio_cache)
361 goto out_free_io_cache;
362
363 _rq_cache = kmem_cache_create("dm_old_clone_request", sizeof(struct request),
364 __alignof__(struct request), 0, NULL);
365 if (!_rq_cache)
366 goto out_free_rq_tio_cache;
367
368 r = dm_uevent_init();
369 if (r)
370 goto out_free_rq_cache;
371
372 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
373 if (!deferred_remove_workqueue) {
374 r = -ENOMEM;
375 goto out_uevent_exit;
376 }
377
378 _major = major;
379 r = register_blkdev(_major, _name);
380 if (r < 0)
381 goto out_free_workqueue;
382
383 if (!_major)
384 _major = r;
385
386 return 0;
387
388out_free_workqueue:
389 destroy_workqueue(deferred_remove_workqueue);
390out_uevent_exit:
391 dm_uevent_exit();
392out_free_rq_cache:
393 kmem_cache_destroy(_rq_cache);
394out_free_rq_tio_cache:
395 kmem_cache_destroy(_rq_tio_cache);
396out_free_io_cache:
397 kmem_cache_destroy(_io_cache);
398
399 return r;
400}
401
402static void local_exit(void)
403{
404 flush_scheduled_work();
405 destroy_workqueue(deferred_remove_workqueue);
406
407 kmem_cache_destroy(_rq_cache);
408 kmem_cache_destroy(_rq_tio_cache);
409 kmem_cache_destroy(_io_cache);
410 unregister_blkdev(_major, _name);
411 dm_uevent_exit();
412
413 _major = 0;
414
415 DMINFO("cleaned up");
416}
417
418static int (*_inits[])(void) __initdata = {
419 local_init,
420 dm_target_init,
421 dm_linear_init,
422 dm_stripe_init,
423 dm_io_init,
424 dm_kcopyd_init,
425 dm_interface_init,
426 dm_statistics_init,
427};
428
429static void (*_exits[])(void) = {
430 local_exit,
431 dm_target_exit,
432 dm_linear_exit,
433 dm_stripe_exit,
434 dm_io_exit,
435 dm_kcopyd_exit,
436 dm_interface_exit,
437 dm_statistics_exit,
438};
439
440static int __init dm_init(void)
441{
442 const int count = ARRAY_SIZE(_inits);
443
444 int r, i;
445
446 for (i = 0; i < count; i++) {
447 r = _inits[i]();
448 if (r)
449 goto bad;
450 }
451
452 return 0;
453
454 bad:
455 while (i--)
456 _exits[i]();
457
458 return r;
459}
460
461static void __exit dm_exit(void)
462{
463 int i = ARRAY_SIZE(_exits);
464
465 while (i--)
466 _exits[i]();
467
468 /*
469 * Should be empty by this point.
470 */
471 idr_destroy(&_minor_idr);
472}
473
474/*
475 * Block device functions
476 */
477int dm_deleting_md(struct mapped_device *md)
478{
479 return test_bit(DMF_DELETING, &md->flags);
480}
481
482static int dm_blk_open(struct block_device *bdev, fmode_t mode)
483{
484 struct mapped_device *md;
485
486 spin_lock(&_minor_lock);
487
488 md = bdev->bd_disk->private_data;
489 if (!md)
490 goto out;
491
492 if (test_bit(DMF_FREEING, &md->flags) ||
493 dm_deleting_md(md)) {
494 md = NULL;
495 goto out;
496 }
497
498 dm_get(md);
499 atomic_inc(&md->open_count);
500out:
501 spin_unlock(&_minor_lock);
502
503 return md ? 0 : -ENXIO;
504}
505
506static void dm_blk_close(struct gendisk *disk, fmode_t mode)
507{
508 struct mapped_device *md;
509
510 spin_lock(&_minor_lock);
511
512 md = disk->private_data;
513 if (WARN_ON(!md))
514 goto out;
515
516 if (atomic_dec_and_test(&md->open_count) &&
517 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
518 queue_work(deferred_remove_workqueue, &deferred_remove_work);
519
520 dm_put(md);
521out:
522 spin_unlock(&_minor_lock);
523}
524
525int dm_open_count(struct mapped_device *md)
526{
527 return atomic_read(&md->open_count);
528}
529
530/*
531 * Guarantees nothing is using the device before it's deleted.
532 */
533int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
534{
535 int r = 0;
536
537 spin_lock(&_minor_lock);
538
539 if (dm_open_count(md)) {
540 r = -EBUSY;
541 if (mark_deferred)
542 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
543 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
544 r = -EEXIST;
545 else
546 set_bit(DMF_DELETING, &md->flags);
547
548 spin_unlock(&_minor_lock);
549
550 return r;
551}
552
553int dm_cancel_deferred_remove(struct mapped_device *md)
554{
555 int r = 0;
556
557 spin_lock(&_minor_lock);
558
559 if (test_bit(DMF_DELETING, &md->flags))
560 r = -EBUSY;
561 else
562 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
563
564 spin_unlock(&_minor_lock);
565
566 return r;
567}
568
569static void do_deferred_remove(struct work_struct *w)
570{
571 dm_deferred_remove();
572}
573
574sector_t dm_get_size(struct mapped_device *md)
575{
576 return get_capacity(md->disk);
577}
578
579struct request_queue *dm_get_md_queue(struct mapped_device *md)
580{
581 return md->queue;
582}
583
584struct dm_stats *dm_get_stats(struct mapped_device *md)
585{
586 return &md->stats;
587}
588
589static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
590{
591 struct mapped_device *md = bdev->bd_disk->private_data;
592
593 return dm_get_geometry(md, geo);
594}
595
596static int dm_grab_bdev_for_ioctl(struct mapped_device *md,
597 struct block_device **bdev,
598 fmode_t *mode)
599{
600 struct dm_target *tgt;
601 struct dm_table *map;
602 int srcu_idx, r;
603
604retry:
605 r = -ENOTTY;
606 map = dm_get_live_table(md, &srcu_idx);
607 if (!map || !dm_table_get_size(map))
608 goto out;
609
610 /* We only support devices that have a single target */
611 if (dm_table_get_num_targets(map) != 1)
612 goto out;
613
614 tgt = dm_table_get_target(map, 0);
615 if (!tgt->type->prepare_ioctl)
616 goto out;
617
618 if (dm_suspended_md(md)) {
619 r = -EAGAIN;
620 goto out;
621 }
622
623 r = tgt->type->prepare_ioctl(tgt, bdev, mode);
624 if (r < 0)
625 goto out;
626
627 bdgrab(*bdev);
628 dm_put_live_table(md, srcu_idx);
629 return r;
630
631out:
632 dm_put_live_table(md, srcu_idx);
633 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
634 msleep(10);
635 goto retry;
636 }
637 return r;
638}
639
640static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
641 unsigned int cmd, unsigned long arg)
642{
643 struct mapped_device *md = bdev->bd_disk->private_data;
644 int r;
645
646 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
647 if (r < 0)
648 return r;
649
650 if (r > 0) {
651 /*
652 * Target determined this ioctl is being issued against
653 * a logical partition of the parent bdev; so extra
654 * validation is needed.
655 */
656 r = scsi_verify_blk_ioctl(NULL, cmd);
657 if (r)
658 goto out;
659 }
660
661 r = __blkdev_driver_ioctl(bdev, mode, cmd, arg);
662out:
663 bdput(bdev);
664 return r;
665}
666
667static struct dm_io *alloc_io(struct mapped_device *md)
668{
669 return mempool_alloc(md->io_pool, GFP_NOIO);
670}
671
672static void free_io(struct mapped_device *md, struct dm_io *io)
673{
674 mempool_free(io, md->io_pool);
675}
676
677static void free_tio(struct mapped_device *md, struct dm_target_io *tio)
678{
679 bio_put(&tio->clone);
680}
681
682static struct dm_rq_target_io *alloc_old_rq_tio(struct mapped_device *md,
683 gfp_t gfp_mask)
684{
685 return mempool_alloc(md->io_pool, gfp_mask);
686}
687
688static void free_old_rq_tio(struct dm_rq_target_io *tio)
689{
690 mempool_free(tio, tio->md->io_pool);
691}
692
693static struct request *alloc_old_clone_request(struct mapped_device *md,
694 gfp_t gfp_mask)
695{
696 return mempool_alloc(md->rq_pool, gfp_mask);
697}
698
699static void free_old_clone_request(struct mapped_device *md, struct request *rq)
700{
701 mempool_free(rq, md->rq_pool);
702}
703
704static int md_in_flight(struct mapped_device *md)
705{
706 return atomic_read(&md->pending[READ]) +
707 atomic_read(&md->pending[WRITE]);
708}
709
710static void start_io_acct(struct dm_io *io)
711{
712 struct mapped_device *md = io->md;
713 struct bio *bio = io->bio;
714 int cpu;
715 int rw = bio_data_dir(bio);
716
717 io->start_time = jiffies;
718
719 cpu = part_stat_lock();
720 part_round_stats(cpu, &dm_disk(md)->part0);
721 part_stat_unlock();
722 atomic_set(&dm_disk(md)->part0.in_flight[rw],
723 atomic_inc_return(&md->pending[rw]));
724
725 if (unlikely(dm_stats_used(&md->stats)))
726 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
727 bio_sectors(bio), false, 0, &io->stats_aux);
728}
729
730static void end_io_acct(struct dm_io *io)
731{
732 struct mapped_device *md = io->md;
733 struct bio *bio = io->bio;
734 unsigned long duration = jiffies - io->start_time;
735 int pending;
736 int rw = bio_data_dir(bio);
737
738 generic_end_io_acct(rw, &dm_disk(md)->part0, io->start_time);
739
740 if (unlikely(dm_stats_used(&md->stats)))
741 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
742 bio_sectors(bio), true, duration, &io->stats_aux);
743
744 /*
745 * After this is decremented the bio must not be touched if it is
746 * a flush.
747 */
748 pending = atomic_dec_return(&md->pending[rw]);
749 atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
750 pending += atomic_read(&md->pending[rw^0x1]);
751
752 /* nudge anyone waiting on suspend queue */
753 if (!pending)
754 wake_up(&md->wait);
755}
756
757/*
758 * Add the bio to the list of deferred io.
759 */
760static void queue_io(struct mapped_device *md, struct bio *bio)
761{
762 unsigned long flags;
763
764 spin_lock_irqsave(&md->deferred_lock, flags);
765 bio_list_add(&md->deferred, bio);
766 spin_unlock_irqrestore(&md->deferred_lock, flags);
767 queue_work(md->wq, &md->work);
768}
769
770/*
771 * Everyone (including functions in this file), should use this
772 * function to access the md->map field, and make sure they call
773 * dm_put_live_table() when finished.
774 */
775struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
776{
777 *srcu_idx = srcu_read_lock(&md->io_barrier);
778
779 return srcu_dereference(md->map, &md->io_barrier);
780}
781
782void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
783{
784 srcu_read_unlock(&md->io_barrier, srcu_idx);
785}
786
787void dm_sync_table(struct mapped_device *md)
788{
789 synchronize_srcu(&md->io_barrier);
790 synchronize_rcu_expedited();
791}
792
793/*
794 * A fast alternative to dm_get_live_table/dm_put_live_table.
795 * The caller must not block between these two functions.
796 */
797static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
798{
799 rcu_read_lock();
800 return rcu_dereference(md->map);
801}
802
803static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
804{
805 rcu_read_unlock();
806}
807
808/*
809 * Open a table device so we can use it as a map destination.
810 */
811static int open_table_device(struct table_device *td, dev_t dev,
812 struct mapped_device *md)
813{
814 static char *_claim_ptr = "I belong to device-mapper";
815 struct block_device *bdev;
816
817 int r;
818
819 BUG_ON(td->dm_dev.bdev);
820
821 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _claim_ptr);
822 if (IS_ERR(bdev))
823 return PTR_ERR(bdev);
824
825 r = bd_link_disk_holder(bdev, dm_disk(md));
826 if (r) {
827 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
828 return r;
829 }
830
831 td->dm_dev.bdev = bdev;
832 return 0;
833}
834
835/*
836 * Close a table device that we've been using.
837 */
838static void close_table_device(struct table_device *td, struct mapped_device *md)
839{
840 if (!td->dm_dev.bdev)
841 return;
842
843 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
844 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
845 td->dm_dev.bdev = NULL;
846}
847
848static struct table_device *find_table_device(struct list_head *l, dev_t dev,
849 fmode_t mode) {
850 struct table_device *td;
851
852 list_for_each_entry(td, l, list)
853 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
854 return td;
855
856 return NULL;
857}
858
859int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
860 struct dm_dev **result) {
861 int r;
862 struct table_device *td;
863
864 mutex_lock(&md->table_devices_lock);
865 td = find_table_device(&md->table_devices, dev, mode);
866 if (!td) {
867 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
868 if (!td) {
869 mutex_unlock(&md->table_devices_lock);
870 return -ENOMEM;
871 }
872
873 td->dm_dev.mode = mode;
874 td->dm_dev.bdev = NULL;
875
876 if ((r = open_table_device(td, dev, md))) {
877 mutex_unlock(&md->table_devices_lock);
878 kfree(td);
879 return r;
880 }
881
882 format_dev_t(td->dm_dev.name, dev);
883
884 atomic_set(&td->count, 0);
885 list_add(&td->list, &md->table_devices);
886 }
887 atomic_inc(&td->count);
888 mutex_unlock(&md->table_devices_lock);
889
890 *result = &td->dm_dev;
891 return 0;
892}
893EXPORT_SYMBOL_GPL(dm_get_table_device);
894
895void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
896{
897 struct table_device *td = container_of(d, struct table_device, dm_dev);
898
899 mutex_lock(&md->table_devices_lock);
900 if (atomic_dec_and_test(&td->count)) {
901 close_table_device(td, md);
902 list_del(&td->list);
903 kfree(td);
904 }
905 mutex_unlock(&md->table_devices_lock);
906}
907EXPORT_SYMBOL(dm_put_table_device);
908
909static void free_table_devices(struct list_head *devices)
910{
911 struct list_head *tmp, *next;
912
913 list_for_each_safe(tmp, next, devices) {
914 struct table_device *td = list_entry(tmp, struct table_device, list);
915
916 DMWARN("dm_destroy: %s still exists with %d references",
917 td->dm_dev.name, atomic_read(&td->count));
918 kfree(td);
919 }
920}
921
922/*
923 * Get the geometry associated with a dm device
924 */
925int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
926{
927 *geo = md->geometry;
928
929 return 0;
930}
931
932/*
933 * Set the geometry of a device.
934 */
935int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
936{
937 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
938
939 if (geo->start > sz) {
940 DMWARN("Start sector is beyond the geometry limits.");
941 return -EINVAL;
942 }
943
944 md->geometry = *geo;
945
946 return 0;
947}
948
949/*-----------------------------------------------------------------
950 * CRUD START:
951 * A more elegant soln is in the works that uses the queue
952 * merge fn, unfortunately there are a couple of changes to
953 * the block layer that I want to make for this. So in the
954 * interests of getting something for people to use I give
955 * you this clearly demarcated crap.
956 *---------------------------------------------------------------*/
957
958static int __noflush_suspending(struct mapped_device *md)
959{
960 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
961}
962
963/*
964 * Decrements the number of outstanding ios that a bio has been
965 * cloned into, completing the original io if necc.
966 */
967static void dec_pending(struct dm_io *io, int error)
968{
969 unsigned long flags;
970 int io_error;
971 struct bio *bio;
972 struct mapped_device *md = io->md;
973
974 /* Push-back supersedes any I/O errors */
975 if (unlikely(error)) {
976 spin_lock_irqsave(&io->endio_lock, flags);
977 if (!(io->error > 0 && __noflush_suspending(md)))
978 io->error = error;
979 spin_unlock_irqrestore(&io->endio_lock, flags);
980 }
981
982 if (atomic_dec_and_test(&io->io_count)) {
983 if (io->error == DM_ENDIO_REQUEUE) {
984 /*
985 * Target requested pushing back the I/O.
986 */
987 spin_lock_irqsave(&md->deferred_lock, flags);
988 if (__noflush_suspending(md))
989 bio_list_add_head(&md->deferred, io->bio);
990 else
991 /* noflush suspend was interrupted. */
992 io->error = -EIO;
993 spin_unlock_irqrestore(&md->deferred_lock, flags);
994 }
995
996 io_error = io->error;
997 bio = io->bio;
998 end_io_acct(io);
999 free_io(md, io);
1000
1001 if (io_error == DM_ENDIO_REQUEUE)
1002 return;
1003
1004 if ((bio->bi_rw & REQ_FLUSH) && bio->bi_iter.bi_size) {
1005 /*
1006 * Preflush done for flush with data, reissue
1007 * without REQ_FLUSH.
1008 */
1009 bio->bi_rw &= ~REQ_FLUSH;
1010 queue_io(md, bio);
1011 } else {
1012 /* done with normal IO or empty flush */
1013 trace_block_bio_complete(md->queue, bio, io_error);
1014 bio->bi_error = io_error;
1015 bio_endio(bio);
1016 }
1017 }
1018}
1019
1020static void disable_write_same(struct mapped_device *md)
1021{
1022 struct queue_limits *limits = dm_get_queue_limits(md);
1023
1024 /* device doesn't really support WRITE SAME, disable it */
1025 limits->max_write_same_sectors = 0;
1026}
1027
1028static void clone_endio(struct bio *bio)
1029{
1030 int error = bio->bi_error;
1031 int r = error;
1032 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1033 struct dm_io *io = tio->io;
1034 struct mapped_device *md = tio->io->md;
1035 dm_endio_fn endio = tio->ti->type->end_io;
1036
1037 if (endio) {
1038 r = endio(tio->ti, bio, error);
1039 if (r < 0 || r == DM_ENDIO_REQUEUE)
1040 /*
1041 * error and requeue request are handled
1042 * in dec_pending().
1043 */
1044 error = r;
1045 else if (r == DM_ENDIO_INCOMPLETE)
1046 /* The target will handle the io */
1047 return;
1048 else if (r) {
1049 DMWARN("unimplemented target endio return value: %d", r);
1050 BUG();
1051 }
1052 }
1053
1054 if (unlikely(r == -EREMOTEIO && (bio->bi_rw & REQ_WRITE_SAME) &&
1055 !bdev_get_queue(bio->bi_bdev)->limits.max_write_same_sectors))
1056 disable_write_same(md);
1057
1058 free_tio(md, tio);
1059 dec_pending(io, error);
1060}
1061
1062/*
1063 * Partial completion handling for request-based dm
1064 */
1065static void end_clone_bio(struct bio *clone)
1066{
1067 struct dm_rq_clone_bio_info *info =
1068 container_of(clone, struct dm_rq_clone_bio_info, clone);
1069 struct dm_rq_target_io *tio = info->tio;
1070 struct bio *bio = info->orig;
1071 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
1072 int error = clone->bi_error;
1073
1074 bio_put(clone);
1075
1076 if (tio->error)
1077 /*
1078 * An error has already been detected on the request.
1079 * Once error occurred, just let clone->end_io() handle
1080 * the remainder.
1081 */
1082 return;
1083 else if (error) {
1084 /*
1085 * Don't notice the error to the upper layer yet.
1086 * The error handling decision is made by the target driver,
1087 * when the request is completed.
1088 */
1089 tio->error = error;
1090 return;
1091 }
1092
1093 /*
1094 * I/O for the bio successfully completed.
1095 * Notice the data completion to the upper layer.
1096 */
1097
1098 /*
1099 * bios are processed from the head of the list.
1100 * So the completing bio should always be rq->bio.
1101 * If it's not, something wrong is happening.
1102 */
1103 if (tio->orig->bio != bio)
1104 DMERR("bio completion is going in the middle of the request");
1105
1106 /*
1107 * Update the original request.
1108 * Do not use blk_end_request() here, because it may complete
1109 * the original request before the clone, and break the ordering.
1110 */
1111 blk_update_request(tio->orig, 0, nr_bytes);
1112}
1113
1114static struct dm_rq_target_io *tio_from_request(struct request *rq)
1115{
1116 return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
1117}
1118
1119static void rq_end_stats(struct mapped_device *md, struct request *orig)
1120{
1121 if (unlikely(dm_stats_used(&md->stats))) {
1122 struct dm_rq_target_io *tio = tio_from_request(orig);
1123 tio->duration_jiffies = jiffies - tio->duration_jiffies;
1124 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
1125 tio->n_sectors, true, tio->duration_jiffies,
1126 &tio->stats_aux);
1127 }
1128}
1129
1130/*
1131 * Don't touch any member of the md after calling this function because
1132 * the md may be freed in dm_put() at the end of this function.
1133 * Or do dm_get() before calling this function and dm_put() later.
1134 */
1135static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
1136{
1137 atomic_dec(&md->pending[rw]);
1138
1139 /* nudge anyone waiting on suspend queue */
1140 if (!md_in_flight(md))
1141 wake_up(&md->wait);
1142
1143 /*
1144 * Run this off this callpath, as drivers could invoke end_io while
1145 * inside their request_fn (and holding the queue lock). Calling
1146 * back into ->request_fn() could deadlock attempting to grab the
1147 * queue lock again.
1148 */
1149 if (!md->queue->mq_ops && run_queue)
1150 blk_run_queue_async(md->queue);
1151
1152 /*
1153 * dm_put() must be at the end of this function. See the comment above
1154 */
1155 dm_put(md);
1156}
1157
1158static void free_rq_clone(struct request *clone)
1159{
1160 struct dm_rq_target_io *tio = clone->end_io_data;
1161 struct mapped_device *md = tio->md;
1162
1163 blk_rq_unprep_clone(clone);
1164
1165 if (md->type == DM_TYPE_MQ_REQUEST_BASED)
1166 /* stacked on blk-mq queue(s) */
1167 tio->ti->type->release_clone_rq(clone);
1168 else if (!md->queue->mq_ops)
1169 /* request_fn queue stacked on request_fn queue(s) */
1170 free_old_clone_request(md, clone);
1171
1172 if (!md->queue->mq_ops)
1173 free_old_rq_tio(tio);
1174}
1175
1176/*
1177 * Complete the clone and the original request.
1178 * Must be called without clone's queue lock held,
1179 * see end_clone_request() for more details.
1180 */
1181static void dm_end_request(struct request *clone, int error)
1182{
1183 int rw = rq_data_dir(clone);
1184 struct dm_rq_target_io *tio = clone->end_io_data;
1185 struct mapped_device *md = tio->md;
1186 struct request *rq = tio->orig;
1187
1188 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
1189 rq->errors = clone->errors;
1190 rq->resid_len = clone->resid_len;
1191
1192 if (rq->sense)
1193 /*
1194 * We are using the sense buffer of the original
1195 * request.
1196 * So setting the length of the sense data is enough.
1197 */
1198 rq->sense_len = clone->sense_len;
1199 }
1200
1201 free_rq_clone(clone);
1202 rq_end_stats(md, rq);
1203 if (!rq->q->mq_ops)
1204 blk_end_request_all(rq, error);
1205 else
1206 blk_mq_end_request(rq, error);
1207 rq_completed(md, rw, true);
1208}
1209
1210static void dm_unprep_request(struct request *rq)
1211{
1212 struct dm_rq_target_io *tio = tio_from_request(rq);
1213 struct request *clone = tio->clone;
1214
1215 if (!rq->q->mq_ops) {
1216 rq->special = NULL;
1217 rq->cmd_flags &= ~REQ_DONTPREP;
1218 }
1219
1220 if (clone)
1221 free_rq_clone(clone);
1222 else if (!tio->md->queue->mq_ops)
1223 free_old_rq_tio(tio);
1224}
1225
1226/*
1227 * Requeue the original request of a clone.
1228 */
1229static void dm_old_requeue_request(struct request *rq)
1230{
1231 struct request_queue *q = rq->q;
1232 unsigned long flags;
1233
1234 spin_lock_irqsave(q->queue_lock, flags);
1235 blk_requeue_request(q, rq);
1236 blk_run_queue_async(q);
1237 spin_unlock_irqrestore(q->queue_lock, flags);
1238}
1239
1240static void dm_mq_requeue_request(struct request *rq)
1241{
1242 struct request_queue *q = rq->q;
1243 unsigned long flags;
1244
1245 blk_mq_requeue_request(rq);
1246 spin_lock_irqsave(q->queue_lock, flags);
1247 if (!blk_queue_stopped(q))
1248 blk_mq_kick_requeue_list(q);
1249 spin_unlock_irqrestore(q->queue_lock, flags);
1250}
1251
1252static void dm_requeue_original_request(struct mapped_device *md,
1253 struct request *rq)
1254{
1255 int rw = rq_data_dir(rq);
1256
1257 rq_end_stats(md, rq);
1258 dm_unprep_request(rq);
1259
1260 if (!rq->q->mq_ops)
1261 dm_old_requeue_request(rq);
1262 else
1263 dm_mq_requeue_request(rq);
1264
1265 rq_completed(md, rw, false);
1266}
1267
1268static void dm_old_stop_queue(struct request_queue *q)
1269{
1270 unsigned long flags;
1271
1272 spin_lock_irqsave(q->queue_lock, flags);
1273 if (blk_queue_stopped(q)) {
1274 spin_unlock_irqrestore(q->queue_lock, flags);
1275 return;
1276 }
1277
1278 blk_stop_queue(q);
1279 spin_unlock_irqrestore(q->queue_lock, flags);
1280}
1281
1282static void dm_stop_queue(struct request_queue *q)
1283{
1284 if (!q->mq_ops)
1285 dm_old_stop_queue(q);
1286 else
1287 blk_mq_stop_hw_queues(q);
1288}
1289
1290static void dm_old_start_queue(struct request_queue *q)
1291{
1292 unsigned long flags;
1293
1294 spin_lock_irqsave(q->queue_lock, flags);
1295 if (blk_queue_stopped(q))
1296 blk_start_queue(q);
1297 spin_unlock_irqrestore(q->queue_lock, flags);
1298}
1299
1300static void dm_start_queue(struct request_queue *q)
1301{
1302 if (!q->mq_ops)
1303 dm_old_start_queue(q);
1304 else {
1305 blk_mq_start_stopped_hw_queues(q, true);
1306 blk_mq_kick_requeue_list(q);
1307 }
1308}
1309
1310static void dm_done(struct request *clone, int error, bool mapped)
1311{
1312 int r = error;
1313 struct dm_rq_target_io *tio = clone->end_io_data;
1314 dm_request_endio_fn rq_end_io = NULL;
1315
1316 if (tio->ti) {
1317 rq_end_io = tio->ti->type->rq_end_io;
1318
1319 if (mapped && rq_end_io)
1320 r = rq_end_io(tio->ti, clone, error, &tio->info);
1321 }
1322
1323 if (unlikely(r == -EREMOTEIO && (clone->cmd_flags & REQ_WRITE_SAME) &&
1324 !clone->q->limits.max_write_same_sectors))
1325 disable_write_same(tio->md);
1326
1327 if (r <= 0)
1328 /* The target wants to complete the I/O */
1329 dm_end_request(clone, r);
1330 else if (r == DM_ENDIO_INCOMPLETE)
1331 /* The target will handle the I/O */
1332 return;
1333 else if (r == DM_ENDIO_REQUEUE)
1334 /* The target wants to requeue the I/O */
1335 dm_requeue_original_request(tio->md, tio->orig);
1336 else {
1337 DMWARN("unimplemented target endio return value: %d", r);
1338 BUG();
1339 }
1340}
1341
1342/*
1343 * Request completion handler for request-based dm
1344 */
1345static void dm_softirq_done(struct request *rq)
1346{
1347 bool mapped = true;
1348 struct dm_rq_target_io *tio = tio_from_request(rq);
1349 struct request *clone = tio->clone;
1350 int rw;
1351
1352 if (!clone) {
1353 rq_end_stats(tio->md, rq);
1354 rw = rq_data_dir(rq);
1355 if (!rq->q->mq_ops) {
1356 blk_end_request_all(rq, tio->error);
1357 rq_completed(tio->md, rw, false);
1358 free_old_rq_tio(tio);
1359 } else {
1360 blk_mq_end_request(rq, tio->error);
1361 rq_completed(tio->md, rw, false);
1362 }
1363 return;
1364 }
1365
1366 if (rq->cmd_flags & REQ_FAILED)
1367 mapped = false;
1368
1369 dm_done(clone, tio->error, mapped);
1370}
1371
1372/*
1373 * Complete the clone and the original request with the error status
1374 * through softirq context.
1375 */
1376static void dm_complete_request(struct request *rq, int error)
1377{
1378 struct dm_rq_target_io *tio = tio_from_request(rq);
1379
1380 tio->error = error;
1381 if (!rq->q->mq_ops)
1382 blk_complete_request(rq);
1383 else
1384 blk_mq_complete_request(rq, error);
1385}
1386
1387/*
1388 * Complete the not-mapped clone and the original request with the error status
1389 * through softirq context.
1390 * Target's rq_end_io() function isn't called.
1391 * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
1392 */
1393static void dm_kill_unmapped_request(struct request *rq, int error)
1394{
1395 rq->cmd_flags |= REQ_FAILED;
1396 dm_complete_request(rq, error);
1397}
1398
1399/*
1400 * Called with the clone's queue lock held (in the case of .request_fn)
1401 */
1402static void end_clone_request(struct request *clone, int error)
1403{
1404 struct dm_rq_target_io *tio = clone->end_io_data;
1405
1406 if (!clone->q->mq_ops) {
1407 /*
1408 * For just cleaning up the information of the queue in which
1409 * the clone was dispatched.
1410 * The clone is *NOT* freed actually here because it is alloced
1411 * from dm own mempool (REQ_ALLOCED isn't set).
1412 */
1413 __blk_put_request(clone->q, clone);
1414 }
1415
1416 /*
1417 * Actual request completion is done in a softirq context which doesn't
1418 * hold the clone's queue lock. Otherwise, deadlock could occur because:
1419 * - another request may be submitted by the upper level driver
1420 * of the stacking during the completion
1421 * - the submission which requires queue lock may be done
1422 * against this clone's queue
1423 */
1424 dm_complete_request(tio->orig, error);
1425}
1426
1427/*
1428 * Return maximum size of I/O possible at the supplied sector up to the current
1429 * target boundary.
1430 */
1431static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1432{
1433 sector_t target_offset = dm_target_offset(ti, sector);
1434
1435 return ti->len - target_offset;
1436}
1437
1438static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1439{
1440 sector_t len = max_io_len_target_boundary(sector, ti);
1441 sector_t offset, max_len;
1442
1443 /*
1444 * Does the target need to split even further?
1445 */
1446 if (ti->max_io_len) {
1447 offset = dm_target_offset(ti, sector);
1448 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1449 max_len = sector_div(offset, ti->max_io_len);
1450 else
1451 max_len = offset & (ti->max_io_len - 1);
1452 max_len = ti->max_io_len - max_len;
1453
1454 if (len > max_len)
1455 len = max_len;
1456 }
1457
1458 return len;
1459}
1460
1461int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1462{
1463 if (len > UINT_MAX) {
1464 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1465 (unsigned long long)len, UINT_MAX);
1466 ti->error = "Maximum size of target IO is too large";
1467 return -EINVAL;
1468 }
1469
1470 ti->max_io_len = (uint32_t) len;
1471
1472 return 0;
1473}
1474EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1475
1476/*
1477 * A target may call dm_accept_partial_bio only from the map routine. It is
1478 * allowed for all bio types except REQ_FLUSH.
1479 *
1480 * dm_accept_partial_bio informs the dm that the target only wants to process
1481 * additional n_sectors sectors of the bio and the rest of the data should be
1482 * sent in a next bio.
1483 *
1484 * A diagram that explains the arithmetics:
1485 * +--------------------+---------------+-------+
1486 * | 1 | 2 | 3 |
1487 * +--------------------+---------------+-------+
1488 *
1489 * <-------------- *tio->len_ptr --------------->
1490 * <------- bi_size ------->
1491 * <-- n_sectors -->
1492 *
1493 * Region 1 was already iterated over with bio_advance or similar function.
1494 * (it may be empty if the target doesn't use bio_advance)
1495 * Region 2 is the remaining bio size that the target wants to process.
1496 * (it may be empty if region 1 is non-empty, although there is no reason
1497 * to make it empty)
1498 * The target requires that region 3 is to be sent in the next bio.
1499 *
1500 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1501 * the partially processed part (the sum of regions 1+2) must be the same for all
1502 * copies of the bio.
1503 */
1504void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1505{
1506 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1507 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1508 BUG_ON(bio->bi_rw & REQ_FLUSH);
1509 BUG_ON(bi_size > *tio->len_ptr);
1510 BUG_ON(n_sectors > bi_size);
1511 *tio->len_ptr -= bi_size - n_sectors;
1512 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1513}
1514EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1515
1516static void __map_bio(struct dm_target_io *tio)
1517{
1518 int r;
1519 sector_t sector;
1520 struct mapped_device *md;
1521 struct bio *clone = &tio->clone;
1522 struct dm_target *ti = tio->ti;
1523
1524 clone->bi_end_io = clone_endio;
1525
1526 /*
1527 * Map the clone. If r == 0 we don't need to do
1528 * anything, the target has assumed ownership of
1529 * this io.
1530 */
1531 atomic_inc(&tio->io->io_count);
1532 sector = clone->bi_iter.bi_sector;
1533 r = ti->type->map(ti, clone);
1534 if (r == DM_MAPIO_REMAPPED) {
1535 /* the bio has been remapped so dispatch it */
1536
1537 trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1538 tio->io->bio->bi_bdev->bd_dev, sector);
1539
1540 generic_make_request(clone);
1541 } else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1542 /* error the io and bail out, or requeue it if needed */
1543 md = tio->io->md;
1544 dec_pending(tio->io, r);
1545 free_tio(md, tio);
1546 } else if (r != DM_MAPIO_SUBMITTED) {
1547 DMWARN("unimplemented target map return value: %d", r);
1548 BUG();
1549 }
1550}
1551
1552struct clone_info {
1553 struct mapped_device *md;
1554 struct dm_table *map;
1555 struct bio *bio;
1556 struct dm_io *io;
1557 sector_t sector;
1558 unsigned sector_count;
1559};
1560
1561static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1562{
1563 bio->bi_iter.bi_sector = sector;
1564 bio->bi_iter.bi_size = to_bytes(len);
1565}
1566
1567/*
1568 * Creates a bio that consists of range of complete bvecs.
1569 */
1570static int clone_bio(struct dm_target_io *tio, struct bio *bio,
1571 sector_t sector, unsigned len)
1572{
1573 struct bio *clone = &tio->clone;
1574
1575 __bio_clone_fast(clone, bio);
1576
1577 if (bio_integrity(bio)) {
1578 int r = bio_integrity_clone(clone, bio, GFP_NOIO);
1579 if (r < 0)
1580 return r;
1581 }
1582
1583 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1584 clone->bi_iter.bi_size = to_bytes(len);
1585
1586 if (bio_integrity(bio))
1587 bio_integrity_trim(clone, 0, len);
1588
1589 return 0;
1590}
1591
1592static struct dm_target_io *alloc_tio(struct clone_info *ci,
1593 struct dm_target *ti,
1594 unsigned target_bio_nr)
1595{
1596 struct dm_target_io *tio;
1597 struct bio *clone;
1598
1599 clone = bio_alloc_bioset(GFP_NOIO, 0, ci->md->bs);
1600 tio = container_of(clone, struct dm_target_io, clone);
1601
1602 tio->io = ci->io;
1603 tio->ti = ti;
1604 tio->target_bio_nr = target_bio_nr;
1605
1606 return tio;
1607}
1608
1609static void __clone_and_map_simple_bio(struct clone_info *ci,
1610 struct dm_target *ti,
1611 unsigned target_bio_nr, unsigned *len)
1612{
1613 struct dm_target_io *tio = alloc_tio(ci, ti, target_bio_nr);
1614 struct bio *clone = &tio->clone;
1615
1616 tio->len_ptr = len;
1617
1618 __bio_clone_fast(clone, ci->bio);
1619 if (len)
1620 bio_setup_sector(clone, ci->sector, *len);
1621
1622 __map_bio(tio);
1623}
1624
1625static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1626 unsigned num_bios, unsigned *len)
1627{
1628 unsigned target_bio_nr;
1629
1630 for (target_bio_nr = 0; target_bio_nr < num_bios; target_bio_nr++)
1631 __clone_and_map_simple_bio(ci, ti, target_bio_nr, len);
1632}
1633
1634static int __send_empty_flush(struct clone_info *ci)
1635{
1636 unsigned target_nr = 0;
1637 struct dm_target *ti;
1638
1639 BUG_ON(bio_has_data(ci->bio));
1640 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1641 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1642
1643 return 0;
1644}
1645
1646static int __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1647 sector_t sector, unsigned *len)
1648{
1649 struct bio *bio = ci->bio;
1650 struct dm_target_io *tio;
1651 unsigned target_bio_nr;
1652 unsigned num_target_bios = 1;
1653 int r = 0;
1654
1655 /*
1656 * Does the target want to receive duplicate copies of the bio?
1657 */
1658 if (bio_data_dir(bio) == WRITE && ti->num_write_bios)
1659 num_target_bios = ti->num_write_bios(ti, bio);
1660
1661 for (target_bio_nr = 0; target_bio_nr < num_target_bios; target_bio_nr++) {
1662 tio = alloc_tio(ci, ti, target_bio_nr);
1663 tio->len_ptr = len;
1664 r = clone_bio(tio, bio, sector, *len);
1665 if (r < 0) {
1666 free_tio(ci->md, tio);
1667 break;
1668 }
1669 __map_bio(tio);
1670 }
1671
1672 return r;
1673}
1674
1675typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1676
1677static unsigned get_num_discard_bios(struct dm_target *ti)
1678{
1679 return ti->num_discard_bios;
1680}
1681
1682static unsigned get_num_write_same_bios(struct dm_target *ti)
1683{
1684 return ti->num_write_same_bios;
1685}
1686
1687typedef bool (*is_split_required_fn)(struct dm_target *ti);
1688
1689static bool is_split_required_for_discard(struct dm_target *ti)
1690{
1691 return ti->split_discard_bios;
1692}
1693
1694static int __send_changing_extent_only(struct clone_info *ci,
1695 get_num_bios_fn get_num_bios,
1696 is_split_required_fn is_split_required)
1697{
1698 struct dm_target *ti;
1699 unsigned len;
1700 unsigned num_bios;
1701
1702 do {
1703 ti = dm_table_find_target(ci->map, ci->sector);
1704 if (!dm_target_is_valid(ti))
1705 return -EIO;
1706
1707 /*
1708 * Even though the device advertised support for this type of
1709 * request, that does not mean every target supports it, and
1710 * reconfiguration might also have changed that since the
1711 * check was performed.
1712 */
1713 num_bios = get_num_bios ? get_num_bios(ti) : 0;
1714 if (!num_bios)
1715 return -EOPNOTSUPP;
1716
1717 if (is_split_required && !is_split_required(ti))
1718 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1719 else
1720 len = min((sector_t)ci->sector_count, max_io_len(ci->sector, ti));
1721
1722 __send_duplicate_bios(ci, ti, num_bios, &len);
1723
1724 ci->sector += len;
1725 } while (ci->sector_count -= len);
1726
1727 return 0;
1728}
1729
1730static int __send_discard(struct clone_info *ci)
1731{
1732 return __send_changing_extent_only(ci, get_num_discard_bios,
1733 is_split_required_for_discard);
1734}
1735
1736static int __send_write_same(struct clone_info *ci)
1737{
1738 return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
1739}
1740
1741/*
1742 * Select the correct strategy for processing a non-flush bio.
1743 */
1744static int __split_and_process_non_flush(struct clone_info *ci)
1745{
1746 struct bio *bio = ci->bio;
1747 struct dm_target *ti;
1748 unsigned len;
1749 int r;
1750
1751 if (unlikely(bio->bi_rw & REQ_DISCARD))
1752 return __send_discard(ci);
1753 else if (unlikely(bio->bi_rw & REQ_WRITE_SAME))
1754 return __send_write_same(ci);
1755
1756 ti = dm_table_find_target(ci->map, ci->sector);
1757 if (!dm_target_is_valid(ti))
1758 return -EIO;
1759
1760 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1761
1762 r = __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1763 if (r < 0)
1764 return r;
1765
1766 ci->sector += len;
1767 ci->sector_count -= len;
1768
1769 return 0;
1770}
1771
1772/*
1773 * Entry point to split a bio into clones and submit them to the targets.
1774 */
1775static void __split_and_process_bio(struct mapped_device *md,
1776 struct dm_table *map, struct bio *bio)
1777{
1778 struct clone_info ci;
1779 int error = 0;
1780
1781 if (unlikely(!map)) {
1782 bio_io_error(bio);
1783 return;
1784 }
1785
1786 ci.map = map;
1787 ci.md = md;
1788 ci.io = alloc_io(md);
1789 ci.io->error = 0;
1790 atomic_set(&ci.io->io_count, 1);
1791 ci.io->bio = bio;
1792 ci.io->md = md;
1793 spin_lock_init(&ci.io->endio_lock);
1794 ci.sector = bio->bi_iter.bi_sector;
1795
1796 start_io_acct(ci.io);
1797
1798 if (bio->bi_rw & REQ_FLUSH) {
1799 ci.bio = &ci.md->flush_bio;
1800 ci.sector_count = 0;
1801 error = __send_empty_flush(&ci);
1802 /* dec_pending submits any data associated with flush */
1803 } else {
1804 ci.bio = bio;
1805 ci.sector_count = bio_sectors(bio);
1806 while (ci.sector_count && !error)
1807 error = __split_and_process_non_flush(&ci);
1808 }
1809
1810 /* drop the extra reference count */
1811 dec_pending(ci.io, error);
1812}
1813/*-----------------------------------------------------------------
1814 * CRUD END
1815 *---------------------------------------------------------------*/
1816
1817/*
1818 * The request function that just remaps the bio built up by
1819 * dm_merge_bvec.
1820 */
1821static blk_qc_t dm_make_request(struct request_queue *q, struct bio *bio)
1822{
1823 int rw = bio_data_dir(bio);
1824 struct mapped_device *md = q->queuedata;
1825 int srcu_idx;
1826 struct dm_table *map;
1827
1828 map = dm_get_live_table(md, &srcu_idx);
1829
1830 generic_start_io_acct(rw, bio_sectors(bio), &dm_disk(md)->part0);
1831
1832 /* if we're suspended, we have to queue this io for later */
1833 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1834 dm_put_live_table(md, srcu_idx);
1835
1836 if (bio_rw(bio) != READA)
1837 queue_io(md, bio);
1838 else
1839 bio_io_error(bio);
1840 return BLK_QC_T_NONE;
1841 }
1842
1843 __split_and_process_bio(md, map, bio);
1844 dm_put_live_table(md, srcu_idx);
1845 return BLK_QC_T_NONE;
1846}
1847
1848int dm_request_based(struct mapped_device *md)
1849{
1850 return blk_queue_stackable(md->queue);
1851}
1852
1853static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
1854{
1855 int r;
1856
1857 if (blk_queue_io_stat(clone->q))
1858 clone->cmd_flags |= REQ_IO_STAT;
1859
1860 clone->start_time = jiffies;
1861 r = blk_insert_cloned_request(clone->q, clone);
1862 if (r)
1863 /* must complete clone in terms of original request */
1864 dm_complete_request(rq, r);
1865}
1866
1867static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
1868 void *data)
1869{
1870 struct dm_rq_target_io *tio = data;
1871 struct dm_rq_clone_bio_info *info =
1872 container_of(bio, struct dm_rq_clone_bio_info, clone);
1873
1874 info->orig = bio_orig;
1875 info->tio = tio;
1876 bio->bi_end_io = end_clone_bio;
1877
1878 return 0;
1879}
1880
1881static int setup_clone(struct request *clone, struct request *rq,
1882 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1883{
1884 int r;
1885
1886 r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
1887 dm_rq_bio_constructor, tio);
1888 if (r)
1889 return r;
1890
1891 clone->cmd = rq->cmd;
1892 clone->cmd_len = rq->cmd_len;
1893 clone->sense = rq->sense;
1894 clone->end_io = end_clone_request;
1895 clone->end_io_data = tio;
1896
1897 tio->clone = clone;
1898
1899 return 0;
1900}
1901
1902static struct request *clone_old_rq(struct request *rq, struct mapped_device *md,
1903 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1904{
1905 /*
1906 * Create clone for use with .request_fn request_queue
1907 */
1908 struct request *clone;
1909
1910 clone = alloc_old_clone_request(md, gfp_mask);
1911 if (!clone)
1912 return NULL;
1913
1914 blk_rq_init(NULL, clone);
1915 if (setup_clone(clone, rq, tio, gfp_mask)) {
1916 /* -ENOMEM */
1917 free_old_clone_request(md, clone);
1918 return NULL;
1919 }
1920
1921 return clone;
1922}
1923
1924static void map_tio_request(struct kthread_work *work);
1925
1926static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
1927 struct mapped_device *md)
1928{
1929 tio->md = md;
1930 tio->ti = NULL;
1931 tio->clone = NULL;
1932 tio->orig = rq;
1933 tio->error = 0;
1934 /*
1935 * Avoid initializing info for blk-mq; it passes
1936 * target-specific data through info.ptr
1937 * (see: dm_mq_init_request)
1938 */
1939 if (!md->init_tio_pdu)
1940 memset(&tio->info, 0, sizeof(tio->info));
1941 if (md->kworker_task)
1942 init_kthread_work(&tio->work, map_tio_request);
1943}
1944
1945static struct dm_rq_target_io *dm_old_prep_tio(struct request *rq,
1946 struct mapped_device *md,
1947 gfp_t gfp_mask)
1948{
1949 struct dm_rq_target_io *tio;
1950 int srcu_idx;
1951 struct dm_table *table;
1952
1953 tio = alloc_old_rq_tio(md, gfp_mask);
1954 if (!tio)
1955 return NULL;
1956
1957 init_tio(tio, rq, md);
1958
1959 table = dm_get_live_table(md, &srcu_idx);
1960 /*
1961 * Must clone a request if this .request_fn DM device
1962 * is stacked on .request_fn device(s).
1963 */
1964 if (!dm_table_mq_request_based(table)) {
1965 if (!clone_old_rq(rq, md, tio, gfp_mask)) {
1966 dm_put_live_table(md, srcu_idx);
1967 free_old_rq_tio(tio);
1968 return NULL;
1969 }
1970 }
1971 dm_put_live_table(md, srcu_idx);
1972
1973 return tio;
1974}
1975
1976/*
1977 * Called with the queue lock held.
1978 */
1979static int dm_old_prep_fn(struct request_queue *q, struct request *rq)
1980{
1981 struct mapped_device *md = q->queuedata;
1982 struct dm_rq_target_io *tio;
1983
1984 if (unlikely(rq->special)) {
1985 DMWARN("Already has something in rq->special.");
1986 return BLKPREP_KILL;
1987 }
1988
1989 tio = dm_old_prep_tio(rq, md, GFP_ATOMIC);
1990 if (!tio)
1991 return BLKPREP_DEFER;
1992
1993 rq->special = tio;
1994 rq->cmd_flags |= REQ_DONTPREP;
1995
1996 return BLKPREP_OK;
1997}
1998
1999/*
2000 * Returns:
2001 * 0 : the request has been processed
2002 * DM_MAPIO_REQUEUE : the original request needs to be requeued
2003 * < 0 : the request was completed due to failure
2004 */
2005static int map_request(struct dm_rq_target_io *tio, struct request *rq,
2006 struct mapped_device *md)
2007{
2008 int r;
2009 struct dm_target *ti = tio->ti;
2010 struct request *clone = NULL;
2011
2012 if (tio->clone) {
2013 clone = tio->clone;
2014 r = ti->type->map_rq(ti, clone, &tio->info);
2015 } else {
2016 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
2017 if (r < 0) {
2018 /* The target wants to complete the I/O */
2019 dm_kill_unmapped_request(rq, r);
2020 return r;
2021 }
2022 if (r != DM_MAPIO_REMAPPED)
2023 return r;
2024 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
2025 /* -ENOMEM */
2026 ti->type->release_clone_rq(clone);
2027 return DM_MAPIO_REQUEUE;
2028 }
2029 }
2030
2031 switch (r) {
2032 case DM_MAPIO_SUBMITTED:
2033 /* The target has taken the I/O to submit by itself later */
2034 break;
2035 case DM_MAPIO_REMAPPED:
2036 /* The target has remapped the I/O so dispatch it */
2037 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
2038 blk_rq_pos(rq));
2039 dm_dispatch_clone_request(clone, rq);
2040 break;
2041 case DM_MAPIO_REQUEUE:
2042 /* The target wants to requeue the I/O */
2043 dm_requeue_original_request(md, tio->orig);
2044 break;
2045 default:
2046 if (r > 0) {
2047 DMWARN("unimplemented target map return value: %d", r);
2048 BUG();
2049 }
2050
2051 /* The target wants to complete the I/O */
2052 dm_kill_unmapped_request(rq, r);
2053 return r;
2054 }
2055
2056 return 0;
2057}
2058
2059static void map_tio_request(struct kthread_work *work)
2060{
2061 struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
2062 struct request *rq = tio->orig;
2063 struct mapped_device *md = tio->md;
2064
2065 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE)
2066 dm_requeue_original_request(md, rq);
2067}
2068
2069static void dm_start_request(struct mapped_device *md, struct request *orig)
2070{
2071 if (!orig->q->mq_ops)
2072 blk_start_request(orig);
2073 else
2074 blk_mq_start_request(orig);
2075 atomic_inc(&md->pending[rq_data_dir(orig)]);
2076
2077 if (md->seq_rq_merge_deadline_usecs) {
2078 md->last_rq_pos = rq_end_sector(orig);
2079 md->last_rq_rw = rq_data_dir(orig);
2080 md->last_rq_start_time = ktime_get();
2081 }
2082
2083 if (unlikely(dm_stats_used(&md->stats))) {
2084 struct dm_rq_target_io *tio = tio_from_request(orig);
2085 tio->duration_jiffies = jiffies;
2086 tio->n_sectors = blk_rq_sectors(orig);
2087 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
2088 tio->n_sectors, false, 0, &tio->stats_aux);
2089 }
2090
2091 /*
2092 * Hold the md reference here for the in-flight I/O.
2093 * We can't rely on the reference count by device opener,
2094 * because the device may be closed during the request completion
2095 * when all bios are completed.
2096 * See the comment in rq_completed() too.
2097 */
2098 dm_get(md);
2099}
2100
2101#define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
2102
2103ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
2104{
2105 return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
2106}
2107
2108ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
2109 const char *buf, size_t count)
2110{
2111 unsigned deadline;
2112
2113 if (!dm_request_based(md) || md->use_blk_mq)
2114 return count;
2115
2116 if (kstrtouint(buf, 10, &deadline))
2117 return -EINVAL;
2118
2119 if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
2120 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
2121
2122 md->seq_rq_merge_deadline_usecs = deadline;
2123
2124 return count;
2125}
2126
2127static bool dm_request_peeked_before_merge_deadline(struct mapped_device *md)
2128{
2129 ktime_t kt_deadline;
2130
2131 if (!md->seq_rq_merge_deadline_usecs)
2132 return false;
2133
2134 kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
2135 kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
2136
2137 return !ktime_after(ktime_get(), kt_deadline);
2138}
2139
2140/*
2141 * q->request_fn for request-based dm.
2142 * Called with the queue lock held.
2143 */
2144static void dm_request_fn(struct request_queue *q)
2145{
2146 struct mapped_device *md = q->queuedata;
2147 struct dm_target *ti = md->immutable_target;
2148 struct request *rq;
2149 struct dm_rq_target_io *tio;
2150 sector_t pos = 0;
2151
2152 if (unlikely(!ti)) {
2153 int srcu_idx;
2154 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2155
2156 ti = dm_table_find_target(map, pos);
2157 dm_put_live_table(md, srcu_idx);
2158 }
2159
2160 /*
2161 * For suspend, check blk_queue_stopped() and increment
2162 * ->pending within a single queue_lock not to increment the
2163 * number of in-flight I/Os after the queue is stopped in
2164 * dm_suspend().
2165 */
2166 while (!blk_queue_stopped(q)) {
2167 rq = blk_peek_request(q);
2168 if (!rq)
2169 return;
2170
2171 /* always use block 0 to find the target for flushes for now */
2172 pos = 0;
2173 if (!(rq->cmd_flags & REQ_FLUSH))
2174 pos = blk_rq_pos(rq);
2175
2176 if ((dm_request_peeked_before_merge_deadline(md) &&
2177 md_in_flight(md) && rq->bio && rq->bio->bi_vcnt == 1 &&
2178 md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
2179 (ti->type->busy && ti->type->busy(ti))) {
2180 blk_delay_queue(q, HZ / 100);
2181 return;
2182 }
2183
2184 dm_start_request(md, rq);
2185
2186 tio = tio_from_request(rq);
2187 /* Establish tio->ti before queuing work (map_tio_request) */
2188 tio->ti = ti;
2189 queue_kthread_work(&md->kworker, &tio->work);
2190 BUG_ON(!irqs_disabled());
2191 }
2192}
2193
2194static int dm_any_congested(void *congested_data, int bdi_bits)
2195{
2196 int r = bdi_bits;
2197 struct mapped_device *md = congested_data;
2198 struct dm_table *map;
2199
2200 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2201 if (dm_request_based(md)) {
2202 /*
2203 * With request-based DM we only need to check the
2204 * top-level queue for congestion.
2205 */
2206 r = md->queue->backing_dev_info.wb.state & bdi_bits;
2207 } else {
2208 map = dm_get_live_table_fast(md);
2209 if (map)
2210 r = dm_table_any_congested(map, bdi_bits);
2211 dm_put_live_table_fast(md);
2212 }
2213 }
2214
2215 return r;
2216}
2217
2218/*-----------------------------------------------------------------
2219 * An IDR is used to keep track of allocated minor numbers.
2220 *---------------------------------------------------------------*/
2221static void free_minor(int minor)
2222{
2223 spin_lock(&_minor_lock);
2224 idr_remove(&_minor_idr, minor);
2225 spin_unlock(&_minor_lock);
2226}
2227
2228/*
2229 * See if the device with a specific minor # is free.
2230 */
2231static int specific_minor(int minor)
2232{
2233 int r;
2234
2235 if (minor >= (1 << MINORBITS))
2236 return -EINVAL;
2237
2238 idr_preload(GFP_KERNEL);
2239 spin_lock(&_minor_lock);
2240
2241 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2242
2243 spin_unlock(&_minor_lock);
2244 idr_preload_end();
2245 if (r < 0)
2246 return r == -ENOSPC ? -EBUSY : r;
2247 return 0;
2248}
2249
2250static int next_free_minor(int *minor)
2251{
2252 int r;
2253
2254 idr_preload(GFP_KERNEL);
2255 spin_lock(&_minor_lock);
2256
2257 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2258
2259 spin_unlock(&_minor_lock);
2260 idr_preload_end();
2261 if (r < 0)
2262 return r;
2263 *minor = r;
2264 return 0;
2265}
2266
2267static const struct block_device_operations dm_blk_dops;
2268
2269static void dm_wq_work(struct work_struct *work);
2270
2271static void dm_init_md_queue(struct mapped_device *md)
2272{
2273 /*
2274 * Request-based dm devices cannot be stacked on top of bio-based dm
2275 * devices. The type of this dm device may not have been decided yet.
2276 * The type is decided at the first table loading time.
2277 * To prevent problematic device stacking, clear the queue flag
2278 * for request stacking support until then.
2279 *
2280 * This queue is new, so no concurrency on the queue_flags.
2281 */
2282 queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
2283
2284 /*
2285 * Initialize data that will only be used by a non-blk-mq DM queue
2286 * - must do so here (in alloc_dev callchain) before queue is used
2287 */
2288 md->queue->queuedata = md;
2289 md->queue->backing_dev_info.congested_data = md;
2290}
2291
2292static void dm_init_normal_md_queue(struct mapped_device *md)
2293{
2294 md->use_blk_mq = false;
2295 dm_init_md_queue(md);
2296
2297 /*
2298 * Initialize aspects of queue that aren't relevant for blk-mq
2299 */
2300 md->queue->backing_dev_info.congested_fn = dm_any_congested;
2301 blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
2302}
2303
2304static void cleanup_mapped_device(struct mapped_device *md)
2305{
2306 if (md->wq)
2307 destroy_workqueue(md->wq);
2308 if (md->kworker_task)
2309 kthread_stop(md->kworker_task);
2310 mempool_destroy(md->io_pool);
2311 mempool_destroy(md->rq_pool);
2312 if (md->bs)
2313 bioset_free(md->bs);
2314
2315 cleanup_srcu_struct(&md->io_barrier);
2316
2317 if (md->disk) {
2318 spin_lock(&_minor_lock);
2319 md->disk->private_data = NULL;
2320 spin_unlock(&_minor_lock);
2321 del_gendisk(md->disk);
2322 put_disk(md->disk);
2323 }
2324
2325 if (md->queue)
2326 blk_cleanup_queue(md->queue);
2327
2328 if (md->bdev) {
2329 bdput(md->bdev);
2330 md->bdev = NULL;
2331 }
2332}
2333
2334/*
2335 * Allocate and initialise a blank device with a given minor.
2336 */
2337static struct mapped_device *alloc_dev(int minor)
2338{
2339 int r, numa_node_id = dm_get_numa_node();
2340 struct mapped_device *md;
2341 void *old_md;
2342
2343 md = kzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
2344 if (!md) {
2345 DMWARN("unable to allocate device, out of memory.");
2346 return NULL;
2347 }
2348
2349 if (!try_module_get(THIS_MODULE))
2350 goto bad_module_get;
2351
2352 /* get a minor number for the dev */
2353 if (minor == DM_ANY_MINOR)
2354 r = next_free_minor(&minor);
2355 else
2356 r = specific_minor(minor);
2357 if (r < 0)
2358 goto bad_minor;
2359
2360 r = init_srcu_struct(&md->io_barrier);
2361 if (r < 0)
2362 goto bad_io_barrier;
2363
2364 md->numa_node_id = numa_node_id;
2365 md->use_blk_mq = use_blk_mq;
2366 md->init_tio_pdu = false;
2367 md->type = DM_TYPE_NONE;
2368 mutex_init(&md->suspend_lock);
2369 mutex_init(&md->type_lock);
2370 mutex_init(&md->table_devices_lock);
2371 spin_lock_init(&md->deferred_lock);
2372 atomic_set(&md->holders, 1);
2373 atomic_set(&md->open_count, 0);
2374 atomic_set(&md->event_nr, 0);
2375 atomic_set(&md->uevent_seq, 0);
2376 INIT_LIST_HEAD(&md->uevent_list);
2377 INIT_LIST_HEAD(&md->table_devices);
2378 spin_lock_init(&md->uevent_lock);
2379
2380 md->queue = blk_alloc_queue_node(GFP_KERNEL, numa_node_id);
2381 if (!md->queue)
2382 goto bad;
2383
2384 dm_init_md_queue(md);
2385
2386 md->disk = alloc_disk_node(1, numa_node_id);
2387 if (!md->disk)
2388 goto bad;
2389
2390 atomic_set(&md->pending[0], 0);
2391 atomic_set(&md->pending[1], 0);
2392 init_waitqueue_head(&md->wait);
2393 INIT_WORK(&md->work, dm_wq_work);
2394 init_waitqueue_head(&md->eventq);
2395 init_completion(&md->kobj_holder.completion);
2396 md->kworker_task = NULL;
2397
2398 md->disk->major = _major;
2399 md->disk->first_minor = minor;
2400 md->disk->fops = &dm_blk_dops;
2401 md->disk->queue = md->queue;
2402 md->disk->private_data = md;
2403 sprintf(md->disk->disk_name, "dm-%d", minor);
2404 add_disk(md->disk);
2405 format_dev_t(md->name, MKDEV(_major, minor));
2406
2407 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
2408 if (!md->wq)
2409 goto bad;
2410
2411 md->bdev = bdget_disk(md->disk, 0);
2412 if (!md->bdev)
2413 goto bad;
2414
2415 bio_init(&md->flush_bio);
2416 md->flush_bio.bi_bdev = md->bdev;
2417 md->flush_bio.bi_rw = WRITE_FLUSH;
2418
2419 dm_stats_init(&md->stats);
2420
2421 /* Populate the mapping, nobody knows we exist yet */
2422 spin_lock(&_minor_lock);
2423 old_md = idr_replace(&_minor_idr, md, minor);
2424 spin_unlock(&_minor_lock);
2425
2426 BUG_ON(old_md != MINOR_ALLOCED);
2427
2428 return md;
2429
2430bad:
2431 cleanup_mapped_device(md);
2432bad_io_barrier:
2433 free_minor(minor);
2434bad_minor:
2435 module_put(THIS_MODULE);
2436bad_module_get:
2437 kfree(md);
2438 return NULL;
2439}
2440
2441static void unlock_fs(struct mapped_device *md);
2442
2443static void free_dev(struct mapped_device *md)
2444{
2445 int minor = MINOR(disk_devt(md->disk));
2446
2447 unlock_fs(md);
2448
2449 cleanup_mapped_device(md);
2450 if (md->tag_set) {
2451 blk_mq_free_tag_set(md->tag_set);
2452 kfree(md->tag_set);
2453 }
2454
2455 free_table_devices(&md->table_devices);
2456 dm_stats_cleanup(&md->stats);
2457 free_minor(minor);
2458
2459 module_put(THIS_MODULE);
2460 kfree(md);
2461}
2462
2463static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
2464{
2465 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
2466
2467 if (md->bs) {
2468 /* The md already has necessary mempools. */
2469 if (dm_table_get_type(t) == DM_TYPE_BIO_BASED) {
2470 /*
2471 * Reload bioset because front_pad may have changed
2472 * because a different table was loaded.
2473 */
2474 bioset_free(md->bs);
2475 md->bs = p->bs;
2476 p->bs = NULL;
2477 }
2478 /*
2479 * There's no need to reload with request-based dm
2480 * because the size of front_pad doesn't change.
2481 * Note for future: If you are to reload bioset,
2482 * prep-ed requests in the queue may refer
2483 * to bio from the old bioset, so you must walk
2484 * through the queue to unprep.
2485 */
2486 goto out;
2487 }
2488
2489 BUG_ON(!p || md->io_pool || md->rq_pool || md->bs);
2490
2491 md->io_pool = p->io_pool;
2492 p->io_pool = NULL;
2493 md->rq_pool = p->rq_pool;
2494 p->rq_pool = NULL;
2495 md->bs = p->bs;
2496 p->bs = NULL;
2497
2498out:
2499 /* mempool bind completed, no longer need any mempools in the table */
2500 dm_table_free_md_mempools(t);
2501}
2502
2503/*
2504 * Bind a table to the device.
2505 */
2506static void event_callback(void *context)
2507{
2508 unsigned long flags;
2509 LIST_HEAD(uevents);
2510 struct mapped_device *md = (struct mapped_device *) context;
2511
2512 spin_lock_irqsave(&md->uevent_lock, flags);
2513 list_splice_init(&md->uevent_list, &uevents);
2514 spin_unlock_irqrestore(&md->uevent_lock, flags);
2515
2516 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2517
2518 atomic_inc(&md->event_nr);
2519 wake_up(&md->eventq);
2520}
2521
2522/*
2523 * Protected by md->suspend_lock obtained by dm_swap_table().
2524 */
2525static void __set_size(struct mapped_device *md, sector_t size)
2526{
2527 set_capacity(md->disk, size);
2528
2529 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2530}
2531
2532/*
2533 * Returns old map, which caller must destroy.
2534 */
2535static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2536 struct queue_limits *limits)
2537{
2538 struct dm_table *old_map;
2539 struct request_queue *q = md->queue;
2540 sector_t size;
2541
2542 size = dm_table_get_size(t);
2543
2544 /*
2545 * Wipe any geometry if the size of the table changed.
2546 */
2547 if (size != dm_get_size(md))
2548 memset(&md->geometry, 0, sizeof(md->geometry));
2549
2550 __set_size(md, size);
2551
2552 dm_table_event_callback(t, event_callback, md);
2553
2554 /*
2555 * The queue hasn't been stopped yet, if the old table type wasn't
2556 * for request-based during suspension. So stop it to prevent
2557 * I/O mapping before resume.
2558 * This must be done before setting the queue restrictions,
2559 * because request-based dm may be run just after the setting.
2560 */
2561 if (dm_table_request_based(t)) {
2562 dm_stop_queue(q);
2563 /*
2564 * Leverage the fact that request-based DM targets are
2565 * immutable singletons and establish md->immutable_target
2566 * - used to optimize both dm_request_fn and dm_mq_queue_rq
2567 */
2568 md->immutable_target = dm_table_get_immutable_target(t);
2569 }
2570
2571 __bind_mempools(md, t);
2572
2573 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2574 rcu_assign_pointer(md->map, (void *)t);
2575 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2576
2577 dm_table_set_restrictions(t, q, limits);
2578 if (old_map)
2579 dm_sync_table(md);
2580
2581 return old_map;
2582}
2583
2584/*
2585 * Returns unbound table for the caller to free.
2586 */
2587static struct dm_table *__unbind(struct mapped_device *md)
2588{
2589 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2590
2591 if (!map)
2592 return NULL;
2593
2594 dm_table_event_callback(map, NULL, NULL);
2595 RCU_INIT_POINTER(md->map, NULL);
2596 dm_sync_table(md);
2597
2598 return map;
2599}
2600
2601/*
2602 * Constructor for a new device.
2603 */
2604int dm_create(int minor, struct mapped_device **result)
2605{
2606 struct mapped_device *md;
2607
2608 md = alloc_dev(minor);
2609 if (!md)
2610 return -ENXIO;
2611
2612 dm_sysfs_init(md);
2613
2614 *result = md;
2615 return 0;
2616}
2617
2618/*
2619 * Functions to manage md->type.
2620 * All are required to hold md->type_lock.
2621 */
2622void dm_lock_md_type(struct mapped_device *md)
2623{
2624 mutex_lock(&md->type_lock);
2625}
2626
2627void dm_unlock_md_type(struct mapped_device *md)
2628{
2629 mutex_unlock(&md->type_lock);
2630}
2631
2632void dm_set_md_type(struct mapped_device *md, unsigned type)
2633{
2634 BUG_ON(!mutex_is_locked(&md->type_lock));
2635 md->type = type;
2636}
2637
2638unsigned dm_get_md_type(struct mapped_device *md)
2639{
2640 return md->type;
2641}
2642
2643struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2644{
2645 return md->immutable_target_type;
2646}
2647
2648/*
2649 * The queue_limits are only valid as long as you have a reference
2650 * count on 'md'.
2651 */
2652struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2653{
2654 BUG_ON(!atomic_read(&md->holders));
2655 return &md->queue->limits;
2656}
2657EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2658
2659static void dm_old_init_rq_based_worker_thread(struct mapped_device *md)
2660{
2661 /* Initialize the request-based DM worker thread */
2662 init_kthread_worker(&md->kworker);
2663 md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
2664 "kdmwork-%s", dm_device_name(md));
2665}
2666
2667/*
2668 * Fully initialize a .request_fn request-based queue.
2669 */
2670static int dm_old_init_request_queue(struct mapped_device *md)
2671{
2672 /* Fully initialize the queue */
2673 if (!blk_init_allocated_queue(md->queue, dm_request_fn, NULL))
2674 return -EINVAL;
2675
2676 /* disable dm_request_fn's merge heuristic by default */
2677 md->seq_rq_merge_deadline_usecs = 0;
2678
2679 dm_init_normal_md_queue(md);
2680 blk_queue_softirq_done(md->queue, dm_softirq_done);
2681 blk_queue_prep_rq(md->queue, dm_old_prep_fn);
2682
2683 dm_old_init_rq_based_worker_thread(md);
2684
2685 elv_register_queue(md->queue);
2686
2687 return 0;
2688}
2689
2690static int dm_mq_init_request(void *data, struct request *rq,
2691 unsigned int hctx_idx, unsigned int request_idx,
2692 unsigned int numa_node)
2693{
2694 struct mapped_device *md = data;
2695 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2696
2697 /*
2698 * Must initialize md member of tio, otherwise it won't
2699 * be available in dm_mq_queue_rq.
2700 */
2701 tio->md = md;
2702
2703 if (md->init_tio_pdu) {
2704 /* target-specific per-io data is immediately after the tio */
2705 tio->info.ptr = tio + 1;
2706 }
2707
2708 return 0;
2709}
2710
2711static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
2712 const struct blk_mq_queue_data *bd)
2713{
2714 struct request *rq = bd->rq;
2715 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2716 struct mapped_device *md = tio->md;
2717 struct dm_target *ti = md->immutable_target;
2718
2719 if (unlikely(!ti)) {
2720 int srcu_idx;
2721 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2722
2723 ti = dm_table_find_target(map, 0);
2724 dm_put_live_table(md, srcu_idx);
2725 }
2726
2727 if (ti->type->busy && ti->type->busy(ti))
2728 return BLK_MQ_RQ_QUEUE_BUSY;
2729
2730 dm_start_request(md, rq);
2731
2732 /* Init tio using md established in .init_request */
2733 init_tio(tio, rq, md);
2734
2735 /*
2736 * Establish tio->ti before queuing work (map_tio_request)
2737 * or making direct call to map_request().
2738 */
2739 tio->ti = ti;
2740
2741 /* Direct call is fine since .queue_rq allows allocations */
2742 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
2743 /* Undo dm_start_request() before requeuing */
2744 rq_end_stats(md, rq);
2745 rq_completed(md, rq_data_dir(rq), false);
2746 return BLK_MQ_RQ_QUEUE_BUSY;
2747 }
2748
2749 return BLK_MQ_RQ_QUEUE_OK;
2750}
2751
2752static struct blk_mq_ops dm_mq_ops = {
2753 .queue_rq = dm_mq_queue_rq,
2754 .map_queue = blk_mq_map_queue,
2755 .complete = dm_softirq_done,
2756 .init_request = dm_mq_init_request,
2757};
2758
2759static int dm_mq_init_request_queue(struct mapped_device *md,
2760 struct dm_target *immutable_tgt)
2761{
2762 struct request_queue *q;
2763 int err;
2764
2765 if (dm_get_md_type(md) == DM_TYPE_REQUEST_BASED) {
2766 DMERR("request-based dm-mq may only be stacked on blk-mq device(s)");
2767 return -EINVAL;
2768 }
2769
2770 md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
2771 if (!md->tag_set)
2772 return -ENOMEM;
2773
2774 md->tag_set->ops = &dm_mq_ops;
2775 md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
2776 md->tag_set->numa_node = md->numa_node_id;
2777 md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2778 md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
2779 md->tag_set->driver_data = md;
2780
2781 md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
2782 if (immutable_tgt && immutable_tgt->per_io_data_size) {
2783 /* any target-specific per-io data is immediately after the tio */
2784 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
2785 md->init_tio_pdu = true;
2786 }
2787
2788 err = blk_mq_alloc_tag_set(md->tag_set);
2789 if (err)
2790 goto out_kfree_tag_set;
2791
2792 q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
2793 if (IS_ERR(q)) {
2794 err = PTR_ERR(q);
2795 goto out_tag_set;
2796 }
2797 dm_init_md_queue(md);
2798
2799 /* backfill 'mq' sysfs registration normally done in blk_register_queue */
2800 blk_mq_register_disk(md->disk);
2801
2802 return 0;
2803
2804out_tag_set:
2805 blk_mq_free_tag_set(md->tag_set);
2806out_kfree_tag_set:
2807 kfree(md->tag_set);
2808
2809 return err;
2810}
2811
2812static unsigned filter_md_type(unsigned type, struct mapped_device *md)
2813{
2814 if (type == DM_TYPE_BIO_BASED)
2815 return type;
2816
2817 return !md->use_blk_mq ? DM_TYPE_REQUEST_BASED : DM_TYPE_MQ_REQUEST_BASED;
2818}
2819
2820/*
2821 * Setup the DM device's queue based on md's type
2822 */
2823int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2824{
2825 int r;
2826 unsigned md_type = filter_md_type(dm_get_md_type(md), md);
2827
2828 switch (md_type) {
2829 case DM_TYPE_REQUEST_BASED:
2830 r = dm_old_init_request_queue(md);
2831 if (r) {
2832 DMERR("Cannot initialize queue for request-based mapped device");
2833 return r;
2834 }
2835 break;
2836 case DM_TYPE_MQ_REQUEST_BASED:
2837 r = dm_mq_init_request_queue(md, dm_table_get_immutable_target(t));
2838 if (r) {
2839 DMERR("Cannot initialize queue for request-based dm-mq mapped device");
2840 return r;
2841 }
2842 break;
2843 case DM_TYPE_BIO_BASED:
2844 dm_init_normal_md_queue(md);
2845 blk_queue_make_request(md->queue, dm_make_request);
2846 /*
2847 * DM handles splitting bios as needed. Free the bio_split bioset
2848 * since it won't be used (saves 1 process per bio-based DM device).
2849 */
2850 bioset_free(md->queue->bio_split);
2851 md->queue->bio_split = NULL;
2852 break;
2853 }
2854
2855 return 0;
2856}
2857
2858struct mapped_device *dm_get_md(dev_t dev)
2859{
2860 struct mapped_device *md;
2861 unsigned minor = MINOR(dev);
2862
2863 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2864 return NULL;
2865
2866 spin_lock(&_minor_lock);
2867
2868 md = idr_find(&_minor_idr, minor);
2869 if (md) {
2870 if ((md == MINOR_ALLOCED ||
2871 (MINOR(disk_devt(dm_disk(md))) != minor) ||
2872 dm_deleting_md(md) ||
2873 test_bit(DMF_FREEING, &md->flags))) {
2874 md = NULL;
2875 goto out;
2876 }
2877 dm_get(md);
2878 }
2879
2880out:
2881 spin_unlock(&_minor_lock);
2882
2883 return md;
2884}
2885EXPORT_SYMBOL_GPL(dm_get_md);
2886
2887void *dm_get_mdptr(struct mapped_device *md)
2888{
2889 return md->interface_ptr;
2890}
2891
2892void dm_set_mdptr(struct mapped_device *md, void *ptr)
2893{
2894 md->interface_ptr = ptr;
2895}
2896
2897void dm_get(struct mapped_device *md)
2898{
2899 atomic_inc(&md->holders);
2900 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2901}
2902
2903int dm_hold(struct mapped_device *md)
2904{
2905 spin_lock(&_minor_lock);
2906 if (test_bit(DMF_FREEING, &md->flags)) {
2907 spin_unlock(&_minor_lock);
2908 return -EBUSY;
2909 }
2910 dm_get(md);
2911 spin_unlock(&_minor_lock);
2912 return 0;
2913}
2914EXPORT_SYMBOL_GPL(dm_hold);
2915
2916const char *dm_device_name(struct mapped_device *md)
2917{
2918 return md->name;
2919}
2920EXPORT_SYMBOL_GPL(dm_device_name);
2921
2922static void __dm_destroy(struct mapped_device *md, bool wait)
2923{
2924 struct dm_table *map;
2925 int srcu_idx;
2926
2927 might_sleep();
2928
2929 spin_lock(&_minor_lock);
2930 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2931 set_bit(DMF_FREEING, &md->flags);
2932 spin_unlock(&_minor_lock);
2933
2934 if (dm_request_based(md) && md->kworker_task)
2935 flush_kthread_worker(&md->kworker);
2936
2937 /*
2938 * Take suspend_lock so that presuspend and postsuspend methods
2939 * do not race with internal suspend.
2940 */
2941 mutex_lock(&md->suspend_lock);
2942 map = dm_get_live_table(md, &srcu_idx);
2943 if (!dm_suspended_md(md)) {
2944 dm_table_presuspend_targets(map);
2945 dm_table_postsuspend_targets(map);
2946 }
2947 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2948 dm_put_live_table(md, srcu_idx);
2949 mutex_unlock(&md->suspend_lock);
2950
2951 /*
2952 * Rare, but there may be I/O requests still going to complete,
2953 * for example. Wait for all references to disappear.
2954 * No one should increment the reference count of the mapped_device,
2955 * after the mapped_device state becomes DMF_FREEING.
2956 */
2957 if (wait)
2958 while (atomic_read(&md->holders))
2959 msleep(1);
2960 else if (atomic_read(&md->holders))
2961 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2962 dm_device_name(md), atomic_read(&md->holders));
2963
2964 dm_sysfs_exit(md);
2965 dm_table_destroy(__unbind(md));
2966 free_dev(md);
2967}
2968
2969void dm_destroy(struct mapped_device *md)
2970{
2971 __dm_destroy(md, true);
2972}
2973
2974void dm_destroy_immediate(struct mapped_device *md)
2975{
2976 __dm_destroy(md, false);
2977}
2978
2979void dm_put(struct mapped_device *md)
2980{
2981 atomic_dec(&md->holders);
2982}
2983EXPORT_SYMBOL_GPL(dm_put);
2984
2985static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
2986{
2987 int r = 0;
2988 DECLARE_WAITQUEUE(wait, current);
2989
2990 add_wait_queue(&md->wait, &wait);
2991
2992 while (1) {
2993 set_current_state(interruptible);
2994
2995 if (!md_in_flight(md))
2996 break;
2997
2998 if (interruptible == TASK_INTERRUPTIBLE &&
2999 signal_pending(current)) {
3000 r = -EINTR;
3001 break;
3002 }
3003
3004 io_schedule();
3005 }
3006 set_current_state(TASK_RUNNING);
3007
3008 remove_wait_queue(&md->wait, &wait);
3009
3010 return r;
3011}
3012
3013/*
3014 * Process the deferred bios
3015 */
3016static void dm_wq_work(struct work_struct *work)
3017{
3018 struct mapped_device *md = container_of(work, struct mapped_device,
3019 work);
3020 struct bio *c;
3021 int srcu_idx;
3022 struct dm_table *map;
3023
3024 map = dm_get_live_table(md, &srcu_idx);
3025
3026 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
3027 spin_lock_irq(&md->deferred_lock);
3028 c = bio_list_pop(&md->deferred);
3029 spin_unlock_irq(&md->deferred_lock);
3030
3031 if (!c)
3032 break;
3033
3034 if (dm_request_based(md))
3035 generic_make_request(c);
3036 else
3037 __split_and_process_bio(md, map, c);
3038 }
3039
3040 dm_put_live_table(md, srcu_idx);
3041}
3042
3043static void dm_queue_flush(struct mapped_device *md)
3044{
3045 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3046 smp_mb__after_atomic();
3047 queue_work(md->wq, &md->work);
3048}
3049
3050/*
3051 * Swap in a new table, returning the old one for the caller to destroy.
3052 */
3053struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
3054{
3055 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
3056 struct queue_limits limits;
3057 int r;
3058
3059 mutex_lock(&md->suspend_lock);
3060
3061 /* device must be suspended */
3062 if (!dm_suspended_md(md))
3063 goto out;
3064
3065 /*
3066 * If the new table has no data devices, retain the existing limits.
3067 * This helps multipath with queue_if_no_path if all paths disappear,
3068 * then new I/O is queued based on these limits, and then some paths
3069 * reappear.
3070 */
3071 if (dm_table_has_no_data_devices(table)) {
3072 live_map = dm_get_live_table_fast(md);
3073 if (live_map)
3074 limits = md->queue->limits;
3075 dm_put_live_table_fast(md);
3076 }
3077
3078 if (!live_map) {
3079 r = dm_calculate_queue_limits(table, &limits);
3080 if (r) {
3081 map = ERR_PTR(r);
3082 goto out;
3083 }
3084 }
3085
3086 map = __bind(md, table, &limits);
3087
3088out:
3089 mutex_unlock(&md->suspend_lock);
3090 return map;
3091}
3092
3093/*
3094 * Functions to lock and unlock any filesystem running on the
3095 * device.
3096 */
3097static int lock_fs(struct mapped_device *md)
3098{
3099 int r;
3100
3101 WARN_ON(md->frozen_sb);
3102
3103 md->frozen_sb = freeze_bdev(md->bdev);
3104 if (IS_ERR(md->frozen_sb)) {
3105 r = PTR_ERR(md->frozen_sb);
3106 md->frozen_sb = NULL;
3107 return r;
3108 }
3109
3110 set_bit(DMF_FROZEN, &md->flags);
3111
3112 return 0;
3113}
3114
3115static void unlock_fs(struct mapped_device *md)
3116{
3117 if (!test_bit(DMF_FROZEN, &md->flags))
3118 return;
3119
3120 thaw_bdev(md->bdev, md->frozen_sb);
3121 md->frozen_sb = NULL;
3122 clear_bit(DMF_FROZEN, &md->flags);
3123}
3124
3125/*
3126 * If __dm_suspend returns 0, the device is completely quiescent
3127 * now. There is no request-processing activity. All new requests
3128 * are being added to md->deferred list.
3129 *
3130 * Caller must hold md->suspend_lock
3131 */
3132static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
3133 unsigned suspend_flags, int interruptible)
3134{
3135 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
3136 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
3137 int r;
3138
3139 /*
3140 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
3141 * This flag is cleared before dm_suspend returns.
3142 */
3143 if (noflush)
3144 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3145
3146 /*
3147 * This gets reverted if there's an error later and the targets
3148 * provide the .presuspend_undo hook.
3149 */
3150 dm_table_presuspend_targets(map);
3151
3152 /*
3153 * Flush I/O to the device.
3154 * Any I/O submitted after lock_fs() may not be flushed.
3155 * noflush takes precedence over do_lockfs.
3156 * (lock_fs() flushes I/Os and waits for them to complete.)
3157 */
3158 if (!noflush && do_lockfs) {
3159 r = lock_fs(md);
3160 if (r) {
3161 dm_table_presuspend_undo_targets(map);
3162 return r;
3163 }
3164 }
3165
3166 /*
3167 * Here we must make sure that no processes are submitting requests
3168 * to target drivers i.e. no one may be executing
3169 * __split_and_process_bio. This is called from dm_request and
3170 * dm_wq_work.
3171 *
3172 * To get all processes out of __split_and_process_bio in dm_request,
3173 * we take the write lock. To prevent any process from reentering
3174 * __split_and_process_bio from dm_request and quiesce the thread
3175 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
3176 * flush_workqueue(md->wq).
3177 */
3178 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3179 if (map)
3180 synchronize_srcu(&md->io_barrier);
3181
3182 /*
3183 * Stop md->queue before flushing md->wq in case request-based
3184 * dm defers requests to md->wq from md->queue.
3185 */
3186 if (dm_request_based(md)) {
3187 dm_stop_queue(md->queue);
3188 if (md->kworker_task)
3189 flush_kthread_worker(&md->kworker);
3190 }
3191
3192 flush_workqueue(md->wq);
3193
3194 /*
3195 * At this point no more requests are entering target request routines.
3196 * We call dm_wait_for_completion to wait for all existing requests
3197 * to finish.
3198 */
3199 r = dm_wait_for_completion(md, interruptible);
3200
3201 if (noflush)
3202 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3203 if (map)
3204 synchronize_srcu(&md->io_barrier);
3205
3206 /* were we interrupted ? */
3207 if (r < 0) {
3208 dm_queue_flush(md);
3209
3210 if (dm_request_based(md))
3211 dm_start_queue(md->queue);
3212
3213 unlock_fs(md);
3214 dm_table_presuspend_undo_targets(map);
3215 /* pushback list is already flushed, so skip flush */
3216 }
3217
3218 return r;
3219}
3220
3221/*
3222 * We need to be able to change a mapping table under a mounted
3223 * filesystem. For example we might want to move some data in
3224 * the background. Before the table can be swapped with
3225 * dm_bind_table, dm_suspend must be called to flush any in
3226 * flight bios and ensure that any further io gets deferred.
3227 */
3228/*
3229 * Suspend mechanism in request-based dm.
3230 *
3231 * 1. Flush all I/Os by lock_fs() if needed.
3232 * 2. Stop dispatching any I/O by stopping the request_queue.
3233 * 3. Wait for all in-flight I/Os to be completed or requeued.
3234 *
3235 * To abort suspend, start the request_queue.
3236 */
3237int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
3238{
3239 struct dm_table *map = NULL;
3240 int r = 0;
3241
3242retry:
3243 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3244
3245 if (dm_suspended_md(md)) {
3246 r = -EINVAL;
3247 goto out_unlock;
3248 }
3249
3250 if (dm_suspended_internally_md(md)) {
3251 /* already internally suspended, wait for internal resume */
3252 mutex_unlock(&md->suspend_lock);
3253 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3254 if (r)
3255 return r;
3256 goto retry;
3257 }
3258
3259 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3260
3261 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE);
3262 if (r)
3263 goto out_unlock;
3264
3265 set_bit(DMF_SUSPENDED, &md->flags);
3266
3267 dm_table_postsuspend_targets(map);
3268
3269out_unlock:
3270 mutex_unlock(&md->suspend_lock);
3271 return r;
3272}
3273
3274static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3275{
3276 if (map) {
3277 int r = dm_table_resume_targets(map);
3278 if (r)
3279 return r;
3280 }
3281
3282 dm_queue_flush(md);
3283
3284 /*
3285 * Flushing deferred I/Os must be done after targets are resumed
3286 * so that mapping of targets can work correctly.
3287 * Request-based dm is queueing the deferred I/Os in its request_queue.
3288 */
3289 if (dm_request_based(md))
3290 dm_start_queue(md->queue);
3291
3292 unlock_fs(md);
3293
3294 return 0;
3295}
3296
3297int dm_resume(struct mapped_device *md)
3298{
3299 int r = -EINVAL;
3300 struct dm_table *map = NULL;
3301
3302retry:
3303 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3304
3305 if (!dm_suspended_md(md))
3306 goto out;
3307
3308 if (dm_suspended_internally_md(md)) {
3309 /* already internally suspended, wait for internal resume */
3310 mutex_unlock(&md->suspend_lock);
3311 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3312 if (r)
3313 return r;
3314 goto retry;
3315 }
3316
3317 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3318 if (!map || !dm_table_get_size(map))
3319 goto out;
3320
3321 r = __dm_resume(md, map);
3322 if (r)
3323 goto out;
3324
3325 clear_bit(DMF_SUSPENDED, &md->flags);
3326
3327 r = 0;
3328out:
3329 mutex_unlock(&md->suspend_lock);
3330
3331 return r;
3332}
3333
3334/*
3335 * Internal suspend/resume works like userspace-driven suspend. It waits
3336 * until all bios finish and prevents issuing new bios to the target drivers.
3337 * It may be used only from the kernel.
3338 */
3339
3340static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
3341{
3342 struct dm_table *map = NULL;
3343
3344 if (md->internal_suspend_count++)
3345 return; /* nested internal suspend */
3346
3347 if (dm_suspended_md(md)) {
3348 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3349 return; /* nest suspend */
3350 }
3351
3352 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3353
3354 /*
3355 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3356 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3357 * would require changing .presuspend to return an error -- avoid this
3358 * until there is a need for more elaborate variants of internal suspend.
3359 */
3360 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE);
3361
3362 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3363
3364 dm_table_postsuspend_targets(map);
3365}
3366
3367static void __dm_internal_resume(struct mapped_device *md)
3368{
3369 BUG_ON(!md->internal_suspend_count);
3370
3371 if (--md->internal_suspend_count)
3372 return; /* resume from nested internal suspend */
3373
3374 if (dm_suspended_md(md))
3375 goto done; /* resume from nested suspend */
3376
3377 /*
3378 * NOTE: existing callers don't need to call dm_table_resume_targets
3379 * (which may fail -- so best to avoid it for now by passing NULL map)
3380 */
3381 (void) __dm_resume(md, NULL);
3382
3383done:
3384 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3385 smp_mb__after_atomic();
3386 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3387}
3388
3389void dm_internal_suspend_noflush(struct mapped_device *md)
3390{
3391 mutex_lock(&md->suspend_lock);
3392 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3393 mutex_unlock(&md->suspend_lock);
3394}
3395EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3396
3397void dm_internal_resume(struct mapped_device *md)
3398{
3399 mutex_lock(&md->suspend_lock);
3400 __dm_internal_resume(md);
3401 mutex_unlock(&md->suspend_lock);
3402}
3403EXPORT_SYMBOL_GPL(dm_internal_resume);
3404
3405/*
3406 * Fast variants of internal suspend/resume hold md->suspend_lock,
3407 * which prevents interaction with userspace-driven suspend.
3408 */
3409
3410void dm_internal_suspend_fast(struct mapped_device *md)
3411{
3412 mutex_lock(&md->suspend_lock);
3413 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3414 return;
3415
3416 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3417 synchronize_srcu(&md->io_barrier);
3418 flush_workqueue(md->wq);
3419 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3420}
3421EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3422
3423void dm_internal_resume_fast(struct mapped_device *md)
3424{
3425 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3426 goto done;
3427
3428 dm_queue_flush(md);
3429
3430done:
3431 mutex_unlock(&md->suspend_lock);
3432}
3433EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3434
3435/*-----------------------------------------------------------------
3436 * Event notification.
3437 *---------------------------------------------------------------*/
3438int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3439 unsigned cookie)
3440{
3441 char udev_cookie[DM_COOKIE_LENGTH];
3442 char *envp[] = { udev_cookie, NULL };
3443
3444 if (!cookie)
3445 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
3446 else {
3447 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3448 DM_COOKIE_ENV_VAR_NAME, cookie);
3449 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
3450 action, envp);
3451 }
3452}
3453
3454uint32_t dm_next_uevent_seq(struct mapped_device *md)
3455{
3456 return atomic_add_return(1, &md->uevent_seq);
3457}
3458
3459uint32_t dm_get_event_nr(struct mapped_device *md)
3460{
3461 return atomic_read(&md->event_nr);
3462}
3463
3464int dm_wait_event(struct mapped_device *md, int event_nr)
3465{
3466 return wait_event_interruptible(md->eventq,
3467 (event_nr != atomic_read(&md->event_nr)));
3468}
3469
3470void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3471{
3472 unsigned long flags;
3473
3474 spin_lock_irqsave(&md->uevent_lock, flags);
3475 list_add(elist, &md->uevent_list);
3476 spin_unlock_irqrestore(&md->uevent_lock, flags);
3477}
3478
3479/*
3480 * The gendisk is only valid as long as you have a reference
3481 * count on 'md'.
3482 */
3483struct gendisk *dm_disk(struct mapped_device *md)
3484{
3485 return md->disk;
3486}
3487EXPORT_SYMBOL_GPL(dm_disk);
3488
3489struct kobject *dm_kobject(struct mapped_device *md)
3490{
3491 return &md->kobj_holder.kobj;
3492}
3493
3494struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3495{
3496 struct mapped_device *md;
3497
3498 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3499
3500 if (test_bit(DMF_FREEING, &md->flags) ||
3501 dm_deleting_md(md))
3502 return NULL;
3503
3504 dm_get(md);
3505 return md;
3506}
3507
3508int dm_suspended_md(struct mapped_device *md)
3509{
3510 return test_bit(DMF_SUSPENDED, &md->flags);
3511}
3512
3513int dm_suspended_internally_md(struct mapped_device *md)
3514{
3515 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3516}
3517
3518int dm_test_deferred_remove_flag(struct mapped_device *md)
3519{
3520 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3521}
3522
3523int dm_suspended(struct dm_target *ti)
3524{
3525 return dm_suspended_md(dm_table_get_md(ti->table));
3526}
3527EXPORT_SYMBOL_GPL(dm_suspended);
3528
3529int dm_noflush_suspending(struct dm_target *ti)
3530{
3531 return __noflush_suspending(dm_table_get_md(ti->table));
3532}
3533EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3534
3535struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
3536 unsigned integrity, unsigned per_io_data_size)
3537{
3538 struct dm_md_mempools *pools = kzalloc_node(sizeof(*pools), GFP_KERNEL, md->numa_node_id);
3539 struct kmem_cache *cachep = NULL;
3540 unsigned int pool_size = 0;
3541 unsigned int front_pad;
3542
3543 if (!pools)
3544 return NULL;
3545
3546 type = filter_md_type(type, md);
3547
3548 switch (type) {
3549 case DM_TYPE_BIO_BASED:
3550 cachep = _io_cache;
3551 pool_size = dm_get_reserved_bio_based_ios();
3552 front_pad = roundup(per_io_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
3553 break;
3554 case DM_TYPE_REQUEST_BASED:
3555 cachep = _rq_tio_cache;
3556 pool_size = dm_get_reserved_rq_based_ios();
3557 pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
3558 if (!pools->rq_pool)
3559 goto out;
3560 /* fall through to setup remaining rq-based pools */
3561 case DM_TYPE_MQ_REQUEST_BASED:
3562 if (!pool_size)
3563 pool_size = dm_get_reserved_rq_based_ios();
3564 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
3565 /* per_io_data_size is used for blk-mq pdu at queue allocation */
3566 break;
3567 default:
3568 BUG();
3569 }
3570
3571 if (cachep) {
3572 pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
3573 if (!pools->io_pool)
3574 goto out;
3575 }
3576
3577 pools->bs = bioset_create_nobvec(pool_size, front_pad);
3578 if (!pools->bs)
3579 goto out;
3580
3581 if (integrity && bioset_integrity_create(pools->bs, pool_size))
3582 goto out;
3583
3584 return pools;
3585
3586out:
3587 dm_free_md_mempools(pools);
3588
3589 return NULL;
3590}
3591
3592void dm_free_md_mempools(struct dm_md_mempools *pools)
3593{
3594 if (!pools)
3595 return;
3596
3597 mempool_destroy(pools->io_pool);
3598 mempool_destroy(pools->rq_pool);
3599
3600 if (pools->bs)
3601 bioset_free(pools->bs);
3602
3603 kfree(pools);
3604}
3605
3606static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3607 u32 flags)
3608{
3609 struct mapped_device *md = bdev->bd_disk->private_data;
3610 const struct pr_ops *ops;
3611 fmode_t mode;
3612 int r;
3613
3614 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3615 if (r < 0)
3616 return r;
3617
3618 ops = bdev->bd_disk->fops->pr_ops;
3619 if (ops && ops->pr_register)
3620 r = ops->pr_register(bdev, old_key, new_key, flags);
3621 else
3622 r = -EOPNOTSUPP;
3623
3624 bdput(bdev);
3625 return r;
3626}
3627
3628static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3629 u32 flags)
3630{
3631 struct mapped_device *md = bdev->bd_disk->private_data;
3632 const struct pr_ops *ops;
3633 fmode_t mode;
3634 int r;
3635
3636 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3637 if (r < 0)
3638 return r;
3639
3640 ops = bdev->bd_disk->fops->pr_ops;
3641 if (ops && ops->pr_reserve)
3642 r = ops->pr_reserve(bdev, key, type, flags);
3643 else
3644 r = -EOPNOTSUPP;
3645
3646 bdput(bdev);
3647 return r;
3648}
3649
3650static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3651{
3652 struct mapped_device *md = bdev->bd_disk->private_data;
3653 const struct pr_ops *ops;
3654 fmode_t mode;
3655 int r;
3656
3657 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3658 if (r < 0)
3659 return r;
3660
3661 ops = bdev->bd_disk->fops->pr_ops;
3662 if (ops && ops->pr_release)
3663 r = ops->pr_release(bdev, key, type);
3664 else
3665 r = -EOPNOTSUPP;
3666
3667 bdput(bdev);
3668 return r;
3669}
3670
3671static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3672 enum pr_type type, bool abort)
3673{
3674 struct mapped_device *md = bdev->bd_disk->private_data;
3675 const struct pr_ops *ops;
3676 fmode_t mode;
3677 int r;
3678
3679 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3680 if (r < 0)
3681 return r;
3682
3683 ops = bdev->bd_disk->fops->pr_ops;
3684 if (ops && ops->pr_preempt)
3685 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
3686 else
3687 r = -EOPNOTSUPP;
3688
3689 bdput(bdev);
3690 return r;
3691}
3692
3693static int dm_pr_clear(struct block_device *bdev, u64 key)
3694{
3695 struct mapped_device *md = bdev->bd_disk->private_data;
3696 const struct pr_ops *ops;
3697 fmode_t mode;
3698 int r;
3699
3700 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3701 if (r < 0)
3702 return r;
3703
3704 ops = bdev->bd_disk->fops->pr_ops;
3705 if (ops && ops->pr_clear)
3706 r = ops->pr_clear(bdev, key);
3707 else
3708 r = -EOPNOTSUPP;
3709
3710 bdput(bdev);
3711 return r;
3712}
3713
3714static const struct pr_ops dm_pr_ops = {
3715 .pr_register = dm_pr_register,
3716 .pr_reserve = dm_pr_reserve,
3717 .pr_release = dm_pr_release,
3718 .pr_preempt = dm_pr_preempt,
3719 .pr_clear = dm_pr_clear,
3720};
3721
3722static const struct block_device_operations dm_blk_dops = {
3723 .open = dm_blk_open,
3724 .release = dm_blk_close,
3725 .ioctl = dm_blk_ioctl,
3726 .getgeo = dm_blk_getgeo,
3727 .pr_ops = &dm_pr_ops,
3728 .owner = THIS_MODULE
3729};
3730
3731/*
3732 * module hooks
3733 */
3734module_init(dm_init);
3735module_exit(dm_exit);
3736
3737module_param(major, uint, 0);
3738MODULE_PARM_DESC(major, "The major number of the device mapper");
3739
3740module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3741MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3742
3743module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
3744MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
3745
3746module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
3747MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
3748
3749module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
3750MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
3751
3752module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
3753MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");
3754
3755module_param(dm_numa_node, int, S_IRUGO | S_IWUSR);
3756MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3757
3758MODULE_DESCRIPTION(DM_NAME " driver");
3759MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3760MODULE_LICENSE("GPL");
1/*
2 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm-core.h"
9#include "dm-rq.h"
10#include "dm-uevent.h"
11
12#include <linux/init.h>
13#include <linux/module.h>
14#include <linux/mutex.h>
15#include <linux/sched/signal.h>
16#include <linux/blkpg.h>
17#include <linux/bio.h>
18#include <linux/mempool.h>
19#include <linux/dax.h>
20#include <linux/slab.h>
21#include <linux/idr.h>
22#include <linux/uio.h>
23#include <linux/hdreg.h>
24#include <linux/delay.h>
25#include <linux/wait.h>
26#include <linux/pr.h>
27#include <linux/refcount.h>
28
29#define DM_MSG_PREFIX "core"
30
31/*
32 * Cookies are numeric values sent with CHANGE and REMOVE
33 * uevents while resuming, removing or renaming the device.
34 */
35#define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
36#define DM_COOKIE_LENGTH 24
37
38static const char *_name = DM_NAME;
39
40static unsigned int major = 0;
41static unsigned int _major = 0;
42
43static DEFINE_IDR(_minor_idr);
44
45static DEFINE_SPINLOCK(_minor_lock);
46
47static void do_deferred_remove(struct work_struct *w);
48
49static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
50
51static struct workqueue_struct *deferred_remove_workqueue;
52
53atomic_t dm_global_event_nr = ATOMIC_INIT(0);
54DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
55
56void dm_issue_global_event(void)
57{
58 atomic_inc(&dm_global_event_nr);
59 wake_up(&dm_global_eventq);
60}
61
62/*
63 * One of these is allocated (on-stack) per original bio.
64 */
65struct clone_info {
66 struct dm_table *map;
67 struct bio *bio;
68 struct dm_io *io;
69 sector_t sector;
70 unsigned sector_count;
71};
72
73/*
74 * One of these is allocated per clone bio.
75 */
76#define DM_TIO_MAGIC 7282014
77struct dm_target_io {
78 unsigned magic;
79 struct dm_io *io;
80 struct dm_target *ti;
81 unsigned target_bio_nr;
82 unsigned *len_ptr;
83 bool inside_dm_io;
84 struct bio clone;
85};
86
87/*
88 * One of these is allocated per original bio.
89 * It contains the first clone used for that original.
90 */
91#define DM_IO_MAGIC 5191977
92struct dm_io {
93 unsigned magic;
94 struct mapped_device *md;
95 blk_status_t status;
96 atomic_t io_count;
97 struct bio *orig_bio;
98 unsigned long start_time;
99 spinlock_t endio_lock;
100 struct dm_stats_aux stats_aux;
101 /* last member of dm_target_io is 'struct bio' */
102 struct dm_target_io tio;
103};
104
105void *dm_per_bio_data(struct bio *bio, size_t data_size)
106{
107 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
108 if (!tio->inside_dm_io)
109 return (char *)bio - offsetof(struct dm_target_io, clone) - data_size;
110 return (char *)bio - offsetof(struct dm_target_io, clone) - offsetof(struct dm_io, tio) - data_size;
111}
112EXPORT_SYMBOL_GPL(dm_per_bio_data);
113
114struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
115{
116 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
117 if (io->magic == DM_IO_MAGIC)
118 return (struct bio *)((char *)io + offsetof(struct dm_io, tio) + offsetof(struct dm_target_io, clone));
119 BUG_ON(io->magic != DM_TIO_MAGIC);
120 return (struct bio *)((char *)io + offsetof(struct dm_target_io, clone));
121}
122EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
123
124unsigned dm_bio_get_target_bio_nr(const struct bio *bio)
125{
126 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
127}
128EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
129
130#define MINOR_ALLOCED ((void *)-1)
131
132/*
133 * Bits for the md->flags field.
134 */
135#define DMF_BLOCK_IO_FOR_SUSPEND 0
136#define DMF_SUSPENDED 1
137#define DMF_FROZEN 2
138#define DMF_FREEING 3
139#define DMF_DELETING 4
140#define DMF_NOFLUSH_SUSPENDING 5
141#define DMF_DEFERRED_REMOVE 6
142#define DMF_SUSPENDED_INTERNALLY 7
143
144#define DM_NUMA_NODE NUMA_NO_NODE
145static int dm_numa_node = DM_NUMA_NODE;
146
147/*
148 * For mempools pre-allocation at the table loading time.
149 */
150struct dm_md_mempools {
151 struct bio_set bs;
152 struct bio_set io_bs;
153};
154
155struct table_device {
156 struct list_head list;
157 refcount_t count;
158 struct dm_dev dm_dev;
159};
160
161/*
162 * Bio-based DM's mempools' reserved IOs set by the user.
163 */
164#define RESERVED_BIO_BASED_IOS 16
165static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
166
167static int __dm_get_module_param_int(int *module_param, int min, int max)
168{
169 int param = READ_ONCE(*module_param);
170 int modified_param = 0;
171 bool modified = true;
172
173 if (param < min)
174 modified_param = min;
175 else if (param > max)
176 modified_param = max;
177 else
178 modified = false;
179
180 if (modified) {
181 (void)cmpxchg(module_param, param, modified_param);
182 param = modified_param;
183 }
184
185 return param;
186}
187
188unsigned __dm_get_module_param(unsigned *module_param,
189 unsigned def, unsigned max)
190{
191 unsigned param = READ_ONCE(*module_param);
192 unsigned modified_param = 0;
193
194 if (!param)
195 modified_param = def;
196 else if (param > max)
197 modified_param = max;
198
199 if (modified_param) {
200 (void)cmpxchg(module_param, param, modified_param);
201 param = modified_param;
202 }
203
204 return param;
205}
206
207unsigned dm_get_reserved_bio_based_ios(void)
208{
209 return __dm_get_module_param(&reserved_bio_based_ios,
210 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
211}
212EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
213
214static unsigned dm_get_numa_node(void)
215{
216 return __dm_get_module_param_int(&dm_numa_node,
217 DM_NUMA_NODE, num_online_nodes() - 1);
218}
219
220static int __init local_init(void)
221{
222 int r;
223
224 r = dm_uevent_init();
225 if (r)
226 return r;
227
228 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
229 if (!deferred_remove_workqueue) {
230 r = -ENOMEM;
231 goto out_uevent_exit;
232 }
233
234 _major = major;
235 r = register_blkdev(_major, _name);
236 if (r < 0)
237 goto out_free_workqueue;
238
239 if (!_major)
240 _major = r;
241
242 return 0;
243
244out_free_workqueue:
245 destroy_workqueue(deferred_remove_workqueue);
246out_uevent_exit:
247 dm_uevent_exit();
248
249 return r;
250}
251
252static void local_exit(void)
253{
254 flush_scheduled_work();
255 destroy_workqueue(deferred_remove_workqueue);
256
257 unregister_blkdev(_major, _name);
258 dm_uevent_exit();
259
260 _major = 0;
261
262 DMINFO("cleaned up");
263}
264
265static int (*_inits[])(void) __initdata = {
266 local_init,
267 dm_target_init,
268 dm_linear_init,
269 dm_stripe_init,
270 dm_io_init,
271 dm_kcopyd_init,
272 dm_interface_init,
273 dm_statistics_init,
274};
275
276static void (*_exits[])(void) = {
277 local_exit,
278 dm_target_exit,
279 dm_linear_exit,
280 dm_stripe_exit,
281 dm_io_exit,
282 dm_kcopyd_exit,
283 dm_interface_exit,
284 dm_statistics_exit,
285};
286
287static int __init dm_init(void)
288{
289 const int count = ARRAY_SIZE(_inits);
290
291 int r, i;
292
293 for (i = 0; i < count; i++) {
294 r = _inits[i]();
295 if (r)
296 goto bad;
297 }
298
299 return 0;
300
301 bad:
302 while (i--)
303 _exits[i]();
304
305 return r;
306}
307
308static void __exit dm_exit(void)
309{
310 int i = ARRAY_SIZE(_exits);
311
312 while (i--)
313 _exits[i]();
314
315 /*
316 * Should be empty by this point.
317 */
318 idr_destroy(&_minor_idr);
319}
320
321/*
322 * Block device functions
323 */
324int dm_deleting_md(struct mapped_device *md)
325{
326 return test_bit(DMF_DELETING, &md->flags);
327}
328
329static int dm_blk_open(struct block_device *bdev, fmode_t mode)
330{
331 struct mapped_device *md;
332
333 spin_lock(&_minor_lock);
334
335 md = bdev->bd_disk->private_data;
336 if (!md)
337 goto out;
338
339 if (test_bit(DMF_FREEING, &md->flags) ||
340 dm_deleting_md(md)) {
341 md = NULL;
342 goto out;
343 }
344
345 dm_get(md);
346 atomic_inc(&md->open_count);
347out:
348 spin_unlock(&_minor_lock);
349
350 return md ? 0 : -ENXIO;
351}
352
353static void dm_blk_close(struct gendisk *disk, fmode_t mode)
354{
355 struct mapped_device *md;
356
357 spin_lock(&_minor_lock);
358
359 md = disk->private_data;
360 if (WARN_ON(!md))
361 goto out;
362
363 if (atomic_dec_and_test(&md->open_count) &&
364 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
365 queue_work(deferred_remove_workqueue, &deferred_remove_work);
366
367 dm_put(md);
368out:
369 spin_unlock(&_minor_lock);
370}
371
372int dm_open_count(struct mapped_device *md)
373{
374 return atomic_read(&md->open_count);
375}
376
377/*
378 * Guarantees nothing is using the device before it's deleted.
379 */
380int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
381{
382 int r = 0;
383
384 spin_lock(&_minor_lock);
385
386 if (dm_open_count(md)) {
387 r = -EBUSY;
388 if (mark_deferred)
389 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
390 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
391 r = -EEXIST;
392 else
393 set_bit(DMF_DELETING, &md->flags);
394
395 spin_unlock(&_minor_lock);
396
397 return r;
398}
399
400int dm_cancel_deferred_remove(struct mapped_device *md)
401{
402 int r = 0;
403
404 spin_lock(&_minor_lock);
405
406 if (test_bit(DMF_DELETING, &md->flags))
407 r = -EBUSY;
408 else
409 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
410
411 spin_unlock(&_minor_lock);
412
413 return r;
414}
415
416static void do_deferred_remove(struct work_struct *w)
417{
418 dm_deferred_remove();
419}
420
421sector_t dm_get_size(struct mapped_device *md)
422{
423 return get_capacity(md->disk);
424}
425
426struct request_queue *dm_get_md_queue(struct mapped_device *md)
427{
428 return md->queue;
429}
430
431struct dm_stats *dm_get_stats(struct mapped_device *md)
432{
433 return &md->stats;
434}
435
436static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
437{
438 struct mapped_device *md = bdev->bd_disk->private_data;
439
440 return dm_get_geometry(md, geo);
441}
442
443static int dm_blk_report_zones(struct gendisk *disk, sector_t sector,
444 struct blk_zone *zones, unsigned int *nr_zones)
445{
446#ifdef CONFIG_BLK_DEV_ZONED
447 struct mapped_device *md = disk->private_data;
448 struct dm_target *tgt;
449 struct dm_table *map;
450 int srcu_idx, ret;
451
452 if (dm_suspended_md(md))
453 return -EAGAIN;
454
455 map = dm_get_live_table(md, &srcu_idx);
456 if (!map)
457 return -EIO;
458
459 tgt = dm_table_find_target(map, sector);
460 if (!tgt) {
461 ret = -EIO;
462 goto out;
463 }
464
465 /*
466 * If we are executing this, we already know that the block device
467 * is a zoned device and so each target should have support for that
468 * type of drive. A missing report_zones method means that the target
469 * driver has a problem.
470 */
471 if (WARN_ON(!tgt->type->report_zones)) {
472 ret = -EIO;
473 goto out;
474 }
475
476 /*
477 * blkdev_report_zones() will loop and call this again to cover all the
478 * zones of the target, eventually moving on to the next target.
479 * So there is no need to loop here trying to fill the entire array
480 * of zones.
481 */
482 ret = tgt->type->report_zones(tgt, sector, zones, nr_zones);
483
484out:
485 dm_put_live_table(md, srcu_idx);
486 return ret;
487#else
488 return -ENOTSUPP;
489#endif
490}
491
492static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
493 struct block_device **bdev)
494 __acquires(md->io_barrier)
495{
496 struct dm_target *tgt;
497 struct dm_table *map;
498 int r;
499
500retry:
501 r = -ENOTTY;
502 map = dm_get_live_table(md, srcu_idx);
503 if (!map || !dm_table_get_size(map))
504 return r;
505
506 /* We only support devices that have a single target */
507 if (dm_table_get_num_targets(map) != 1)
508 return r;
509
510 tgt = dm_table_get_target(map, 0);
511 if (!tgt->type->prepare_ioctl)
512 return r;
513
514 if (dm_suspended_md(md))
515 return -EAGAIN;
516
517 r = tgt->type->prepare_ioctl(tgt, bdev);
518 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
519 dm_put_live_table(md, *srcu_idx);
520 msleep(10);
521 goto retry;
522 }
523
524 return r;
525}
526
527static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
528 __releases(md->io_barrier)
529{
530 dm_put_live_table(md, srcu_idx);
531}
532
533static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
534 unsigned int cmd, unsigned long arg)
535{
536 struct mapped_device *md = bdev->bd_disk->private_data;
537 int r, srcu_idx;
538
539 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
540 if (r < 0)
541 goto out;
542
543 if (r > 0) {
544 /*
545 * Target determined this ioctl is being issued against a
546 * subset of the parent bdev; require extra privileges.
547 */
548 if (!capable(CAP_SYS_RAWIO)) {
549 DMWARN_LIMIT(
550 "%s: sending ioctl %x to DM device without required privilege.",
551 current->comm, cmd);
552 r = -ENOIOCTLCMD;
553 goto out;
554 }
555 }
556
557 r = __blkdev_driver_ioctl(bdev, mode, cmd, arg);
558out:
559 dm_unprepare_ioctl(md, srcu_idx);
560 return r;
561}
562
563static void start_io_acct(struct dm_io *io);
564
565static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio)
566{
567 struct dm_io *io;
568 struct dm_target_io *tio;
569 struct bio *clone;
570
571 clone = bio_alloc_bioset(GFP_NOIO, 0, &md->io_bs);
572 if (!clone)
573 return NULL;
574
575 tio = container_of(clone, struct dm_target_io, clone);
576 tio->inside_dm_io = true;
577 tio->io = NULL;
578
579 io = container_of(tio, struct dm_io, tio);
580 io->magic = DM_IO_MAGIC;
581 io->status = 0;
582 atomic_set(&io->io_count, 1);
583 io->orig_bio = bio;
584 io->md = md;
585 spin_lock_init(&io->endio_lock);
586
587 start_io_acct(io);
588
589 return io;
590}
591
592static void free_io(struct mapped_device *md, struct dm_io *io)
593{
594 bio_put(&io->tio.clone);
595}
596
597static struct dm_target_io *alloc_tio(struct clone_info *ci, struct dm_target *ti,
598 unsigned target_bio_nr, gfp_t gfp_mask)
599{
600 struct dm_target_io *tio;
601
602 if (!ci->io->tio.io) {
603 /* the dm_target_io embedded in ci->io is available */
604 tio = &ci->io->tio;
605 } else {
606 struct bio *clone = bio_alloc_bioset(gfp_mask, 0, &ci->io->md->bs);
607 if (!clone)
608 return NULL;
609
610 tio = container_of(clone, struct dm_target_io, clone);
611 tio->inside_dm_io = false;
612 }
613
614 tio->magic = DM_TIO_MAGIC;
615 tio->io = ci->io;
616 tio->ti = ti;
617 tio->target_bio_nr = target_bio_nr;
618
619 return tio;
620}
621
622static void free_tio(struct dm_target_io *tio)
623{
624 if (tio->inside_dm_io)
625 return;
626 bio_put(&tio->clone);
627}
628
629static bool md_in_flight_bios(struct mapped_device *md)
630{
631 int cpu;
632 struct hd_struct *part = &dm_disk(md)->part0;
633 long sum = 0;
634
635 for_each_possible_cpu(cpu) {
636 sum += part_stat_local_read_cpu(part, in_flight[0], cpu);
637 sum += part_stat_local_read_cpu(part, in_flight[1], cpu);
638 }
639
640 return sum != 0;
641}
642
643static bool md_in_flight(struct mapped_device *md)
644{
645 if (queue_is_mq(md->queue))
646 return blk_mq_queue_inflight(md->queue);
647 else
648 return md_in_flight_bios(md);
649}
650
651static void start_io_acct(struct dm_io *io)
652{
653 struct mapped_device *md = io->md;
654 struct bio *bio = io->orig_bio;
655
656 io->start_time = jiffies;
657
658 generic_start_io_acct(md->queue, bio_op(bio), bio_sectors(bio),
659 &dm_disk(md)->part0);
660
661 if (unlikely(dm_stats_used(&md->stats)))
662 dm_stats_account_io(&md->stats, bio_data_dir(bio),
663 bio->bi_iter.bi_sector, bio_sectors(bio),
664 false, 0, &io->stats_aux);
665}
666
667static void end_io_acct(struct dm_io *io)
668{
669 struct mapped_device *md = io->md;
670 struct bio *bio = io->orig_bio;
671 unsigned long duration = jiffies - io->start_time;
672
673 generic_end_io_acct(md->queue, bio_op(bio), &dm_disk(md)->part0,
674 io->start_time);
675
676 if (unlikely(dm_stats_used(&md->stats)))
677 dm_stats_account_io(&md->stats, bio_data_dir(bio),
678 bio->bi_iter.bi_sector, bio_sectors(bio),
679 true, duration, &io->stats_aux);
680
681 /* nudge anyone waiting on suspend queue */
682 if (unlikely(wq_has_sleeper(&md->wait)))
683 wake_up(&md->wait);
684}
685
686/*
687 * Add the bio to the list of deferred io.
688 */
689static void queue_io(struct mapped_device *md, struct bio *bio)
690{
691 unsigned long flags;
692
693 spin_lock_irqsave(&md->deferred_lock, flags);
694 bio_list_add(&md->deferred, bio);
695 spin_unlock_irqrestore(&md->deferred_lock, flags);
696 queue_work(md->wq, &md->work);
697}
698
699/*
700 * Everyone (including functions in this file), should use this
701 * function to access the md->map field, and make sure they call
702 * dm_put_live_table() when finished.
703 */
704struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
705{
706 *srcu_idx = srcu_read_lock(&md->io_barrier);
707
708 return srcu_dereference(md->map, &md->io_barrier);
709}
710
711void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
712{
713 srcu_read_unlock(&md->io_barrier, srcu_idx);
714}
715
716void dm_sync_table(struct mapped_device *md)
717{
718 synchronize_srcu(&md->io_barrier);
719 synchronize_rcu_expedited();
720}
721
722/*
723 * A fast alternative to dm_get_live_table/dm_put_live_table.
724 * The caller must not block between these two functions.
725 */
726static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
727{
728 rcu_read_lock();
729 return rcu_dereference(md->map);
730}
731
732static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
733{
734 rcu_read_unlock();
735}
736
737static char *_dm_claim_ptr = "I belong to device-mapper";
738
739/*
740 * Open a table device so we can use it as a map destination.
741 */
742static int open_table_device(struct table_device *td, dev_t dev,
743 struct mapped_device *md)
744{
745 struct block_device *bdev;
746
747 int r;
748
749 BUG_ON(td->dm_dev.bdev);
750
751 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _dm_claim_ptr);
752 if (IS_ERR(bdev))
753 return PTR_ERR(bdev);
754
755 r = bd_link_disk_holder(bdev, dm_disk(md));
756 if (r) {
757 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
758 return r;
759 }
760
761 td->dm_dev.bdev = bdev;
762 td->dm_dev.dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
763 return 0;
764}
765
766/*
767 * Close a table device that we've been using.
768 */
769static void close_table_device(struct table_device *td, struct mapped_device *md)
770{
771 if (!td->dm_dev.bdev)
772 return;
773
774 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
775 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
776 put_dax(td->dm_dev.dax_dev);
777 td->dm_dev.bdev = NULL;
778 td->dm_dev.dax_dev = NULL;
779}
780
781static struct table_device *find_table_device(struct list_head *l, dev_t dev,
782 fmode_t mode)
783{
784 struct table_device *td;
785
786 list_for_each_entry(td, l, list)
787 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
788 return td;
789
790 return NULL;
791}
792
793int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
794 struct dm_dev **result)
795{
796 int r;
797 struct table_device *td;
798
799 mutex_lock(&md->table_devices_lock);
800 td = find_table_device(&md->table_devices, dev, mode);
801 if (!td) {
802 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
803 if (!td) {
804 mutex_unlock(&md->table_devices_lock);
805 return -ENOMEM;
806 }
807
808 td->dm_dev.mode = mode;
809 td->dm_dev.bdev = NULL;
810
811 if ((r = open_table_device(td, dev, md))) {
812 mutex_unlock(&md->table_devices_lock);
813 kfree(td);
814 return r;
815 }
816
817 format_dev_t(td->dm_dev.name, dev);
818
819 refcount_set(&td->count, 1);
820 list_add(&td->list, &md->table_devices);
821 } else {
822 refcount_inc(&td->count);
823 }
824 mutex_unlock(&md->table_devices_lock);
825
826 *result = &td->dm_dev;
827 return 0;
828}
829EXPORT_SYMBOL_GPL(dm_get_table_device);
830
831void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
832{
833 struct table_device *td = container_of(d, struct table_device, dm_dev);
834
835 mutex_lock(&md->table_devices_lock);
836 if (refcount_dec_and_test(&td->count)) {
837 close_table_device(td, md);
838 list_del(&td->list);
839 kfree(td);
840 }
841 mutex_unlock(&md->table_devices_lock);
842}
843EXPORT_SYMBOL(dm_put_table_device);
844
845static void free_table_devices(struct list_head *devices)
846{
847 struct list_head *tmp, *next;
848
849 list_for_each_safe(tmp, next, devices) {
850 struct table_device *td = list_entry(tmp, struct table_device, list);
851
852 DMWARN("dm_destroy: %s still exists with %d references",
853 td->dm_dev.name, refcount_read(&td->count));
854 kfree(td);
855 }
856}
857
858/*
859 * Get the geometry associated with a dm device
860 */
861int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
862{
863 *geo = md->geometry;
864
865 return 0;
866}
867
868/*
869 * Set the geometry of a device.
870 */
871int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
872{
873 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
874
875 if (geo->start > sz) {
876 DMWARN("Start sector is beyond the geometry limits.");
877 return -EINVAL;
878 }
879
880 md->geometry = *geo;
881
882 return 0;
883}
884
885static int __noflush_suspending(struct mapped_device *md)
886{
887 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
888}
889
890/*
891 * Decrements the number of outstanding ios that a bio has been
892 * cloned into, completing the original io if necc.
893 */
894static void dec_pending(struct dm_io *io, blk_status_t error)
895{
896 unsigned long flags;
897 blk_status_t io_error;
898 struct bio *bio;
899 struct mapped_device *md = io->md;
900
901 /* Push-back supersedes any I/O errors */
902 if (unlikely(error)) {
903 spin_lock_irqsave(&io->endio_lock, flags);
904 if (!(io->status == BLK_STS_DM_REQUEUE && __noflush_suspending(md)))
905 io->status = error;
906 spin_unlock_irqrestore(&io->endio_lock, flags);
907 }
908
909 if (atomic_dec_and_test(&io->io_count)) {
910 if (io->status == BLK_STS_DM_REQUEUE) {
911 /*
912 * Target requested pushing back the I/O.
913 */
914 spin_lock_irqsave(&md->deferred_lock, flags);
915 if (__noflush_suspending(md))
916 /* NOTE early return due to BLK_STS_DM_REQUEUE below */
917 bio_list_add_head(&md->deferred, io->orig_bio);
918 else
919 /* noflush suspend was interrupted. */
920 io->status = BLK_STS_IOERR;
921 spin_unlock_irqrestore(&md->deferred_lock, flags);
922 }
923
924 io_error = io->status;
925 bio = io->orig_bio;
926 end_io_acct(io);
927 free_io(md, io);
928
929 if (io_error == BLK_STS_DM_REQUEUE)
930 return;
931
932 if ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size) {
933 /*
934 * Preflush done for flush with data, reissue
935 * without REQ_PREFLUSH.
936 */
937 bio->bi_opf &= ~REQ_PREFLUSH;
938 queue_io(md, bio);
939 } else {
940 /* done with normal IO or empty flush */
941 if (io_error)
942 bio->bi_status = io_error;
943 bio_endio(bio);
944 }
945 }
946}
947
948void disable_discard(struct mapped_device *md)
949{
950 struct queue_limits *limits = dm_get_queue_limits(md);
951
952 /* device doesn't really support DISCARD, disable it */
953 limits->max_discard_sectors = 0;
954 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, md->queue);
955}
956
957void disable_write_same(struct mapped_device *md)
958{
959 struct queue_limits *limits = dm_get_queue_limits(md);
960
961 /* device doesn't really support WRITE SAME, disable it */
962 limits->max_write_same_sectors = 0;
963}
964
965void disable_write_zeroes(struct mapped_device *md)
966{
967 struct queue_limits *limits = dm_get_queue_limits(md);
968
969 /* device doesn't really support WRITE ZEROES, disable it */
970 limits->max_write_zeroes_sectors = 0;
971}
972
973static void clone_endio(struct bio *bio)
974{
975 blk_status_t error = bio->bi_status;
976 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
977 struct dm_io *io = tio->io;
978 struct mapped_device *md = tio->io->md;
979 dm_endio_fn endio = tio->ti->type->end_io;
980
981 if (unlikely(error == BLK_STS_TARGET) && md->type != DM_TYPE_NVME_BIO_BASED) {
982 if (bio_op(bio) == REQ_OP_DISCARD &&
983 !bio->bi_disk->queue->limits.max_discard_sectors)
984 disable_discard(md);
985 else if (bio_op(bio) == REQ_OP_WRITE_SAME &&
986 !bio->bi_disk->queue->limits.max_write_same_sectors)
987 disable_write_same(md);
988 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
989 !bio->bi_disk->queue->limits.max_write_zeroes_sectors)
990 disable_write_zeroes(md);
991 }
992
993 if (endio) {
994 int r = endio(tio->ti, bio, &error);
995 switch (r) {
996 case DM_ENDIO_REQUEUE:
997 error = BLK_STS_DM_REQUEUE;
998 /*FALLTHRU*/
999 case DM_ENDIO_DONE:
1000 break;
1001 case DM_ENDIO_INCOMPLETE:
1002 /* The target will handle the io */
1003 return;
1004 default:
1005 DMWARN("unimplemented target endio return value: %d", r);
1006 BUG();
1007 }
1008 }
1009
1010 free_tio(tio);
1011 dec_pending(io, error);
1012}
1013
1014/*
1015 * Return maximum size of I/O possible at the supplied sector up to the current
1016 * target boundary.
1017 */
1018static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1019{
1020 sector_t target_offset = dm_target_offset(ti, sector);
1021
1022 return ti->len - target_offset;
1023}
1024
1025static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1026{
1027 sector_t len = max_io_len_target_boundary(sector, ti);
1028 sector_t offset, max_len;
1029
1030 /*
1031 * Does the target need to split even further?
1032 */
1033 if (ti->max_io_len) {
1034 offset = dm_target_offset(ti, sector);
1035 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1036 max_len = sector_div(offset, ti->max_io_len);
1037 else
1038 max_len = offset & (ti->max_io_len - 1);
1039 max_len = ti->max_io_len - max_len;
1040
1041 if (len > max_len)
1042 len = max_len;
1043 }
1044
1045 return len;
1046}
1047
1048int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1049{
1050 if (len > UINT_MAX) {
1051 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1052 (unsigned long long)len, UINT_MAX);
1053 ti->error = "Maximum size of target IO is too large";
1054 return -EINVAL;
1055 }
1056
1057 ti->max_io_len = (uint32_t) len;
1058
1059 return 0;
1060}
1061EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1062
1063static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1064 sector_t sector, int *srcu_idx)
1065 __acquires(md->io_barrier)
1066{
1067 struct dm_table *map;
1068 struct dm_target *ti;
1069
1070 map = dm_get_live_table(md, srcu_idx);
1071 if (!map)
1072 return NULL;
1073
1074 ti = dm_table_find_target(map, sector);
1075 if (!ti)
1076 return NULL;
1077
1078 return ti;
1079}
1080
1081static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1082 long nr_pages, void **kaddr, pfn_t *pfn)
1083{
1084 struct mapped_device *md = dax_get_private(dax_dev);
1085 sector_t sector = pgoff * PAGE_SECTORS;
1086 struct dm_target *ti;
1087 long len, ret = -EIO;
1088 int srcu_idx;
1089
1090 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1091
1092 if (!ti)
1093 goto out;
1094 if (!ti->type->direct_access)
1095 goto out;
1096 len = max_io_len(sector, ti) / PAGE_SECTORS;
1097 if (len < 1)
1098 goto out;
1099 nr_pages = min(len, nr_pages);
1100 ret = ti->type->direct_access(ti, pgoff, nr_pages, kaddr, pfn);
1101
1102 out:
1103 dm_put_live_table(md, srcu_idx);
1104
1105 return ret;
1106}
1107
1108static bool dm_dax_supported(struct dax_device *dax_dev, struct block_device *bdev,
1109 int blocksize, sector_t start, sector_t len)
1110{
1111 struct mapped_device *md = dax_get_private(dax_dev);
1112 struct dm_table *map;
1113 int srcu_idx;
1114 bool ret;
1115
1116 map = dm_get_live_table(md, &srcu_idx);
1117 if (!map)
1118 return false;
1119
1120 ret = dm_table_supports_dax(map, device_supports_dax, &blocksize);
1121
1122 dm_put_live_table(md, srcu_idx);
1123
1124 return ret;
1125}
1126
1127static size_t dm_dax_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
1128 void *addr, size_t bytes, struct iov_iter *i)
1129{
1130 struct mapped_device *md = dax_get_private(dax_dev);
1131 sector_t sector = pgoff * PAGE_SECTORS;
1132 struct dm_target *ti;
1133 long ret = 0;
1134 int srcu_idx;
1135
1136 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1137
1138 if (!ti)
1139 goto out;
1140 if (!ti->type->dax_copy_from_iter) {
1141 ret = copy_from_iter(addr, bytes, i);
1142 goto out;
1143 }
1144 ret = ti->type->dax_copy_from_iter(ti, pgoff, addr, bytes, i);
1145 out:
1146 dm_put_live_table(md, srcu_idx);
1147
1148 return ret;
1149}
1150
1151static size_t dm_dax_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
1152 void *addr, size_t bytes, struct iov_iter *i)
1153{
1154 struct mapped_device *md = dax_get_private(dax_dev);
1155 sector_t sector = pgoff * PAGE_SECTORS;
1156 struct dm_target *ti;
1157 long ret = 0;
1158 int srcu_idx;
1159
1160 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1161
1162 if (!ti)
1163 goto out;
1164 if (!ti->type->dax_copy_to_iter) {
1165 ret = copy_to_iter(addr, bytes, i);
1166 goto out;
1167 }
1168 ret = ti->type->dax_copy_to_iter(ti, pgoff, addr, bytes, i);
1169 out:
1170 dm_put_live_table(md, srcu_idx);
1171
1172 return ret;
1173}
1174
1175/*
1176 * A target may call dm_accept_partial_bio only from the map routine. It is
1177 * allowed for all bio types except REQ_PREFLUSH and REQ_OP_ZONE_RESET.
1178 *
1179 * dm_accept_partial_bio informs the dm that the target only wants to process
1180 * additional n_sectors sectors of the bio and the rest of the data should be
1181 * sent in a next bio.
1182 *
1183 * A diagram that explains the arithmetics:
1184 * +--------------------+---------------+-------+
1185 * | 1 | 2 | 3 |
1186 * +--------------------+---------------+-------+
1187 *
1188 * <-------------- *tio->len_ptr --------------->
1189 * <------- bi_size ------->
1190 * <-- n_sectors -->
1191 *
1192 * Region 1 was already iterated over with bio_advance or similar function.
1193 * (it may be empty if the target doesn't use bio_advance)
1194 * Region 2 is the remaining bio size that the target wants to process.
1195 * (it may be empty if region 1 is non-empty, although there is no reason
1196 * to make it empty)
1197 * The target requires that region 3 is to be sent in the next bio.
1198 *
1199 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1200 * the partially processed part (the sum of regions 1+2) must be the same for all
1201 * copies of the bio.
1202 */
1203void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1204{
1205 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1206 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1207 BUG_ON(bio->bi_opf & REQ_PREFLUSH);
1208 BUG_ON(bi_size > *tio->len_ptr);
1209 BUG_ON(n_sectors > bi_size);
1210 *tio->len_ptr -= bi_size - n_sectors;
1211 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1212}
1213EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1214
1215/*
1216 * The zone descriptors obtained with a zone report indicate
1217 * zone positions within the underlying device of the target. The zone
1218 * descriptors must be remapped to match their position within the dm device.
1219 * The caller target should obtain the zones information using
1220 * blkdev_report_zones() to ensure that remapping for partition offset is
1221 * already handled.
1222 */
1223void dm_remap_zone_report(struct dm_target *ti, sector_t start,
1224 struct blk_zone *zones, unsigned int *nr_zones)
1225{
1226#ifdef CONFIG_BLK_DEV_ZONED
1227 struct blk_zone *zone;
1228 unsigned int nrz = *nr_zones;
1229 int i;
1230
1231 /*
1232 * Remap the start sector and write pointer position of the zones in
1233 * the array. Since we may have obtained from the target underlying
1234 * device more zones that the target size, also adjust the number
1235 * of zones.
1236 */
1237 for (i = 0; i < nrz; i++) {
1238 zone = zones + i;
1239 if (zone->start >= start + ti->len) {
1240 memset(zone, 0, sizeof(struct blk_zone) * (nrz - i));
1241 break;
1242 }
1243
1244 zone->start = zone->start + ti->begin - start;
1245 if (zone->type == BLK_ZONE_TYPE_CONVENTIONAL)
1246 continue;
1247
1248 if (zone->cond == BLK_ZONE_COND_FULL)
1249 zone->wp = zone->start + zone->len;
1250 else if (zone->cond == BLK_ZONE_COND_EMPTY)
1251 zone->wp = zone->start;
1252 else
1253 zone->wp = zone->wp + ti->begin - start;
1254 }
1255
1256 *nr_zones = i;
1257#else /* !CONFIG_BLK_DEV_ZONED */
1258 *nr_zones = 0;
1259#endif
1260}
1261EXPORT_SYMBOL_GPL(dm_remap_zone_report);
1262
1263static blk_qc_t __map_bio(struct dm_target_io *tio)
1264{
1265 int r;
1266 sector_t sector;
1267 struct bio *clone = &tio->clone;
1268 struct dm_io *io = tio->io;
1269 struct mapped_device *md = io->md;
1270 struct dm_target *ti = tio->ti;
1271 blk_qc_t ret = BLK_QC_T_NONE;
1272
1273 clone->bi_end_io = clone_endio;
1274
1275 /*
1276 * Map the clone. If r == 0 we don't need to do
1277 * anything, the target has assumed ownership of
1278 * this io.
1279 */
1280 atomic_inc(&io->io_count);
1281 sector = clone->bi_iter.bi_sector;
1282
1283 r = ti->type->map(ti, clone);
1284 switch (r) {
1285 case DM_MAPIO_SUBMITTED:
1286 break;
1287 case DM_MAPIO_REMAPPED:
1288 /* the bio has been remapped so dispatch it */
1289 trace_block_bio_remap(clone->bi_disk->queue, clone,
1290 bio_dev(io->orig_bio), sector);
1291 if (md->type == DM_TYPE_NVME_BIO_BASED)
1292 ret = direct_make_request(clone);
1293 else
1294 ret = generic_make_request(clone);
1295 break;
1296 case DM_MAPIO_KILL:
1297 free_tio(tio);
1298 dec_pending(io, BLK_STS_IOERR);
1299 break;
1300 case DM_MAPIO_REQUEUE:
1301 free_tio(tio);
1302 dec_pending(io, BLK_STS_DM_REQUEUE);
1303 break;
1304 default:
1305 DMWARN("unimplemented target map return value: %d", r);
1306 BUG();
1307 }
1308
1309 return ret;
1310}
1311
1312static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1313{
1314 bio->bi_iter.bi_sector = sector;
1315 bio->bi_iter.bi_size = to_bytes(len);
1316}
1317
1318/*
1319 * Creates a bio that consists of range of complete bvecs.
1320 */
1321static int clone_bio(struct dm_target_io *tio, struct bio *bio,
1322 sector_t sector, unsigned len)
1323{
1324 struct bio *clone = &tio->clone;
1325
1326 __bio_clone_fast(clone, bio);
1327
1328 if (bio_integrity(bio)) {
1329 int r;
1330
1331 if (unlikely(!dm_target_has_integrity(tio->ti->type) &&
1332 !dm_target_passes_integrity(tio->ti->type))) {
1333 DMWARN("%s: the target %s doesn't support integrity data.",
1334 dm_device_name(tio->io->md),
1335 tio->ti->type->name);
1336 return -EIO;
1337 }
1338
1339 r = bio_integrity_clone(clone, bio, GFP_NOIO);
1340 if (r < 0)
1341 return r;
1342 }
1343
1344 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1345 clone->bi_iter.bi_size = to_bytes(len);
1346
1347 if (bio_integrity(bio))
1348 bio_integrity_trim(clone);
1349
1350 return 0;
1351}
1352
1353static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1354 struct dm_target *ti, unsigned num_bios)
1355{
1356 struct dm_target_io *tio;
1357 int try;
1358
1359 if (!num_bios)
1360 return;
1361
1362 if (num_bios == 1) {
1363 tio = alloc_tio(ci, ti, 0, GFP_NOIO);
1364 bio_list_add(blist, &tio->clone);
1365 return;
1366 }
1367
1368 for (try = 0; try < 2; try++) {
1369 int bio_nr;
1370 struct bio *bio;
1371
1372 if (try)
1373 mutex_lock(&ci->io->md->table_devices_lock);
1374 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1375 tio = alloc_tio(ci, ti, bio_nr, try ? GFP_NOIO : GFP_NOWAIT);
1376 if (!tio)
1377 break;
1378
1379 bio_list_add(blist, &tio->clone);
1380 }
1381 if (try)
1382 mutex_unlock(&ci->io->md->table_devices_lock);
1383 if (bio_nr == num_bios)
1384 return;
1385
1386 while ((bio = bio_list_pop(blist))) {
1387 tio = container_of(bio, struct dm_target_io, clone);
1388 free_tio(tio);
1389 }
1390 }
1391}
1392
1393static blk_qc_t __clone_and_map_simple_bio(struct clone_info *ci,
1394 struct dm_target_io *tio, unsigned *len)
1395{
1396 struct bio *clone = &tio->clone;
1397
1398 tio->len_ptr = len;
1399
1400 __bio_clone_fast(clone, ci->bio);
1401 if (len)
1402 bio_setup_sector(clone, ci->sector, *len);
1403
1404 return __map_bio(tio);
1405}
1406
1407static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1408 unsigned num_bios, unsigned *len)
1409{
1410 struct bio_list blist = BIO_EMPTY_LIST;
1411 struct bio *bio;
1412 struct dm_target_io *tio;
1413
1414 alloc_multiple_bios(&blist, ci, ti, num_bios);
1415
1416 while ((bio = bio_list_pop(&blist))) {
1417 tio = container_of(bio, struct dm_target_io, clone);
1418 (void) __clone_and_map_simple_bio(ci, tio, len);
1419 }
1420}
1421
1422static int __send_empty_flush(struct clone_info *ci)
1423{
1424 unsigned target_nr = 0;
1425 struct dm_target *ti;
1426
1427 /*
1428 * Empty flush uses a statically initialized bio, as the base for
1429 * cloning. However, blkg association requires that a bdev is
1430 * associated with a gendisk, which doesn't happen until the bdev is
1431 * opened. So, blkg association is done at issue time of the flush
1432 * rather than when the device is created in alloc_dev().
1433 */
1434 bio_set_dev(ci->bio, ci->io->md->bdev);
1435
1436 BUG_ON(bio_has_data(ci->bio));
1437 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1438 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1439
1440 bio_disassociate_blkg(ci->bio);
1441
1442 return 0;
1443}
1444
1445static int __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1446 sector_t sector, unsigned *len)
1447{
1448 struct bio *bio = ci->bio;
1449 struct dm_target_io *tio;
1450 int r;
1451
1452 tio = alloc_tio(ci, ti, 0, GFP_NOIO);
1453 tio->len_ptr = len;
1454 r = clone_bio(tio, bio, sector, *len);
1455 if (r < 0) {
1456 free_tio(tio);
1457 return r;
1458 }
1459 (void) __map_bio(tio);
1460
1461 return 0;
1462}
1463
1464typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1465
1466static unsigned get_num_discard_bios(struct dm_target *ti)
1467{
1468 return ti->num_discard_bios;
1469}
1470
1471static unsigned get_num_secure_erase_bios(struct dm_target *ti)
1472{
1473 return ti->num_secure_erase_bios;
1474}
1475
1476static unsigned get_num_write_same_bios(struct dm_target *ti)
1477{
1478 return ti->num_write_same_bios;
1479}
1480
1481static unsigned get_num_write_zeroes_bios(struct dm_target *ti)
1482{
1483 return ti->num_write_zeroes_bios;
1484}
1485
1486static int __send_changing_extent_only(struct clone_info *ci, struct dm_target *ti,
1487 unsigned num_bios)
1488{
1489 unsigned len;
1490
1491 /*
1492 * Even though the device advertised support for this type of
1493 * request, that does not mean every target supports it, and
1494 * reconfiguration might also have changed that since the
1495 * check was performed.
1496 */
1497 if (!num_bios)
1498 return -EOPNOTSUPP;
1499
1500 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1501
1502 __send_duplicate_bios(ci, ti, num_bios, &len);
1503
1504 ci->sector += len;
1505 ci->sector_count -= len;
1506
1507 return 0;
1508}
1509
1510static int __send_discard(struct clone_info *ci, struct dm_target *ti)
1511{
1512 return __send_changing_extent_only(ci, ti, get_num_discard_bios(ti));
1513}
1514
1515static int __send_secure_erase(struct clone_info *ci, struct dm_target *ti)
1516{
1517 return __send_changing_extent_only(ci, ti, get_num_secure_erase_bios(ti));
1518}
1519
1520static int __send_write_same(struct clone_info *ci, struct dm_target *ti)
1521{
1522 return __send_changing_extent_only(ci, ti, get_num_write_same_bios(ti));
1523}
1524
1525static int __send_write_zeroes(struct clone_info *ci, struct dm_target *ti)
1526{
1527 return __send_changing_extent_only(ci, ti, get_num_write_zeroes_bios(ti));
1528}
1529
1530static bool is_abnormal_io(struct bio *bio)
1531{
1532 bool r = false;
1533
1534 switch (bio_op(bio)) {
1535 case REQ_OP_DISCARD:
1536 case REQ_OP_SECURE_ERASE:
1537 case REQ_OP_WRITE_SAME:
1538 case REQ_OP_WRITE_ZEROES:
1539 r = true;
1540 break;
1541 }
1542
1543 return r;
1544}
1545
1546static bool __process_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1547 int *result)
1548{
1549 struct bio *bio = ci->bio;
1550
1551 if (bio_op(bio) == REQ_OP_DISCARD)
1552 *result = __send_discard(ci, ti);
1553 else if (bio_op(bio) == REQ_OP_SECURE_ERASE)
1554 *result = __send_secure_erase(ci, ti);
1555 else if (bio_op(bio) == REQ_OP_WRITE_SAME)
1556 *result = __send_write_same(ci, ti);
1557 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
1558 *result = __send_write_zeroes(ci, ti);
1559 else
1560 return false;
1561
1562 return true;
1563}
1564
1565/*
1566 * Select the correct strategy for processing a non-flush bio.
1567 */
1568static int __split_and_process_non_flush(struct clone_info *ci)
1569{
1570 struct dm_target *ti;
1571 unsigned len;
1572 int r;
1573
1574 ti = dm_table_find_target(ci->map, ci->sector);
1575 if (!ti)
1576 return -EIO;
1577
1578 if (__process_abnormal_io(ci, ti, &r))
1579 return r;
1580
1581 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1582
1583 r = __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1584 if (r < 0)
1585 return r;
1586
1587 ci->sector += len;
1588 ci->sector_count -= len;
1589
1590 return 0;
1591}
1592
1593static void init_clone_info(struct clone_info *ci, struct mapped_device *md,
1594 struct dm_table *map, struct bio *bio)
1595{
1596 ci->map = map;
1597 ci->io = alloc_io(md, bio);
1598 ci->sector = bio->bi_iter.bi_sector;
1599}
1600
1601#define __dm_part_stat_sub(part, field, subnd) \
1602 (part_stat_get(part, field) -= (subnd))
1603
1604/*
1605 * Entry point to split a bio into clones and submit them to the targets.
1606 */
1607static blk_qc_t __split_and_process_bio(struct mapped_device *md,
1608 struct dm_table *map, struct bio *bio)
1609{
1610 struct clone_info ci;
1611 blk_qc_t ret = BLK_QC_T_NONE;
1612 int error = 0;
1613
1614 init_clone_info(&ci, md, map, bio);
1615
1616 if (bio->bi_opf & REQ_PREFLUSH) {
1617 struct bio flush_bio;
1618
1619 /*
1620 * Use an on-stack bio for this, it's safe since we don't
1621 * need to reference it after submit. It's just used as
1622 * the basis for the clone(s).
1623 */
1624 bio_init(&flush_bio, NULL, 0);
1625 flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC;
1626 ci.bio = &flush_bio;
1627 ci.sector_count = 0;
1628 error = __send_empty_flush(&ci);
1629 /* dec_pending submits any data associated with flush */
1630 } else if (bio_op(bio) == REQ_OP_ZONE_RESET) {
1631 ci.bio = bio;
1632 ci.sector_count = 0;
1633 error = __split_and_process_non_flush(&ci);
1634 } else {
1635 ci.bio = bio;
1636 ci.sector_count = bio_sectors(bio);
1637 while (ci.sector_count && !error) {
1638 error = __split_and_process_non_flush(&ci);
1639 if (current->bio_list && ci.sector_count && !error) {
1640 /*
1641 * Remainder must be passed to generic_make_request()
1642 * so that it gets handled *after* bios already submitted
1643 * have been completely processed.
1644 * We take a clone of the original to store in
1645 * ci.io->orig_bio to be used by end_io_acct() and
1646 * for dec_pending to use for completion handling.
1647 */
1648 struct bio *b = bio_split(bio, bio_sectors(bio) - ci.sector_count,
1649 GFP_NOIO, &md->queue->bio_split);
1650 ci.io->orig_bio = b;
1651
1652 /*
1653 * Adjust IO stats for each split, otherwise upon queue
1654 * reentry there will be redundant IO accounting.
1655 * NOTE: this is a stop-gap fix, a proper fix involves
1656 * significant refactoring of DM core's bio splitting
1657 * (by eliminating DM's splitting and just using bio_split)
1658 */
1659 part_stat_lock();
1660 __dm_part_stat_sub(&dm_disk(md)->part0,
1661 sectors[op_stat_group(bio_op(bio))], ci.sector_count);
1662 part_stat_unlock();
1663
1664 bio_chain(b, bio);
1665 trace_block_split(md->queue, b, bio->bi_iter.bi_sector);
1666 ret = generic_make_request(bio);
1667 break;
1668 }
1669 }
1670 }
1671
1672 /* drop the extra reference count */
1673 dec_pending(ci.io, errno_to_blk_status(error));
1674 return ret;
1675}
1676
1677/*
1678 * Optimized variant of __split_and_process_bio that leverages the
1679 * fact that targets that use it do _not_ have a need to split bios.
1680 */
1681static blk_qc_t __process_bio(struct mapped_device *md, struct dm_table *map,
1682 struct bio *bio, struct dm_target *ti)
1683{
1684 struct clone_info ci;
1685 blk_qc_t ret = BLK_QC_T_NONE;
1686 int error = 0;
1687
1688 init_clone_info(&ci, md, map, bio);
1689
1690 if (bio->bi_opf & REQ_PREFLUSH) {
1691 struct bio flush_bio;
1692
1693 /*
1694 * Use an on-stack bio for this, it's safe since we don't
1695 * need to reference it after submit. It's just used as
1696 * the basis for the clone(s).
1697 */
1698 bio_init(&flush_bio, NULL, 0);
1699 flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC;
1700 ci.bio = &flush_bio;
1701 ci.sector_count = 0;
1702 error = __send_empty_flush(&ci);
1703 /* dec_pending submits any data associated with flush */
1704 } else {
1705 struct dm_target_io *tio;
1706
1707 ci.bio = bio;
1708 ci.sector_count = bio_sectors(bio);
1709 if (__process_abnormal_io(&ci, ti, &error))
1710 goto out;
1711
1712 tio = alloc_tio(&ci, ti, 0, GFP_NOIO);
1713 ret = __clone_and_map_simple_bio(&ci, tio, NULL);
1714 }
1715out:
1716 /* drop the extra reference count */
1717 dec_pending(ci.io, errno_to_blk_status(error));
1718 return ret;
1719}
1720
1721static void dm_queue_split(struct mapped_device *md, struct dm_target *ti, struct bio **bio)
1722{
1723 unsigned len, sector_count;
1724
1725 sector_count = bio_sectors(*bio);
1726 len = min_t(sector_t, max_io_len((*bio)->bi_iter.bi_sector, ti), sector_count);
1727
1728 if (sector_count > len) {
1729 struct bio *split = bio_split(*bio, len, GFP_NOIO, &md->queue->bio_split);
1730
1731 bio_chain(split, *bio);
1732 trace_block_split(md->queue, split, (*bio)->bi_iter.bi_sector);
1733 generic_make_request(*bio);
1734 *bio = split;
1735 }
1736}
1737
1738static blk_qc_t dm_process_bio(struct mapped_device *md,
1739 struct dm_table *map, struct bio *bio)
1740{
1741 blk_qc_t ret = BLK_QC_T_NONE;
1742 struct dm_target *ti = md->immutable_target;
1743
1744 if (unlikely(!map)) {
1745 bio_io_error(bio);
1746 return ret;
1747 }
1748
1749 if (!ti) {
1750 ti = dm_table_find_target(map, bio->bi_iter.bi_sector);
1751 if (unlikely(!ti)) {
1752 bio_io_error(bio);
1753 return ret;
1754 }
1755 }
1756
1757 /*
1758 * If in ->make_request_fn we need to use blk_queue_split(), otherwise
1759 * queue_limits for abnormal requests (e.g. discard, writesame, etc)
1760 * won't be imposed.
1761 */
1762 if (current->bio_list) {
1763 blk_queue_split(md->queue, &bio);
1764 if (!is_abnormal_io(bio))
1765 dm_queue_split(md, ti, &bio);
1766 }
1767
1768 if (dm_get_md_type(md) == DM_TYPE_NVME_BIO_BASED)
1769 return __process_bio(md, map, bio, ti);
1770 else
1771 return __split_and_process_bio(md, map, bio);
1772}
1773
1774static blk_qc_t dm_make_request(struct request_queue *q, struct bio *bio)
1775{
1776 struct mapped_device *md = q->queuedata;
1777 blk_qc_t ret = BLK_QC_T_NONE;
1778 int srcu_idx;
1779 struct dm_table *map;
1780
1781 map = dm_get_live_table(md, &srcu_idx);
1782
1783 /* if we're suspended, we have to queue this io for later */
1784 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1785 dm_put_live_table(md, srcu_idx);
1786
1787 if (!(bio->bi_opf & REQ_RAHEAD))
1788 queue_io(md, bio);
1789 else
1790 bio_io_error(bio);
1791 return ret;
1792 }
1793
1794 ret = dm_process_bio(md, map, bio);
1795
1796 dm_put_live_table(md, srcu_idx);
1797 return ret;
1798}
1799
1800static int dm_any_congested(void *congested_data, int bdi_bits)
1801{
1802 int r = bdi_bits;
1803 struct mapped_device *md = congested_data;
1804 struct dm_table *map;
1805
1806 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
1807 if (dm_request_based(md)) {
1808 /*
1809 * With request-based DM we only need to check the
1810 * top-level queue for congestion.
1811 */
1812 r = md->queue->backing_dev_info->wb.state & bdi_bits;
1813 } else {
1814 map = dm_get_live_table_fast(md);
1815 if (map)
1816 r = dm_table_any_congested(map, bdi_bits);
1817 dm_put_live_table_fast(md);
1818 }
1819 }
1820
1821 return r;
1822}
1823
1824/*-----------------------------------------------------------------
1825 * An IDR is used to keep track of allocated minor numbers.
1826 *---------------------------------------------------------------*/
1827static void free_minor(int minor)
1828{
1829 spin_lock(&_minor_lock);
1830 idr_remove(&_minor_idr, minor);
1831 spin_unlock(&_minor_lock);
1832}
1833
1834/*
1835 * See if the device with a specific minor # is free.
1836 */
1837static int specific_minor(int minor)
1838{
1839 int r;
1840
1841 if (minor >= (1 << MINORBITS))
1842 return -EINVAL;
1843
1844 idr_preload(GFP_KERNEL);
1845 spin_lock(&_minor_lock);
1846
1847 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
1848
1849 spin_unlock(&_minor_lock);
1850 idr_preload_end();
1851 if (r < 0)
1852 return r == -ENOSPC ? -EBUSY : r;
1853 return 0;
1854}
1855
1856static int next_free_minor(int *minor)
1857{
1858 int r;
1859
1860 idr_preload(GFP_KERNEL);
1861 spin_lock(&_minor_lock);
1862
1863 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
1864
1865 spin_unlock(&_minor_lock);
1866 idr_preload_end();
1867 if (r < 0)
1868 return r;
1869 *minor = r;
1870 return 0;
1871}
1872
1873static const struct block_device_operations dm_blk_dops;
1874static const struct dax_operations dm_dax_ops;
1875
1876static void dm_wq_work(struct work_struct *work);
1877
1878static void dm_init_normal_md_queue(struct mapped_device *md)
1879{
1880 /*
1881 * Initialize aspects of queue that aren't relevant for blk-mq
1882 */
1883 md->queue->backing_dev_info->congested_fn = dm_any_congested;
1884}
1885
1886static void cleanup_mapped_device(struct mapped_device *md)
1887{
1888 if (md->wq)
1889 destroy_workqueue(md->wq);
1890 bioset_exit(&md->bs);
1891 bioset_exit(&md->io_bs);
1892
1893 if (md->dax_dev) {
1894 kill_dax(md->dax_dev);
1895 put_dax(md->dax_dev);
1896 md->dax_dev = NULL;
1897 }
1898
1899 if (md->disk) {
1900 spin_lock(&_minor_lock);
1901 md->disk->private_data = NULL;
1902 spin_unlock(&_minor_lock);
1903 del_gendisk(md->disk);
1904 put_disk(md->disk);
1905 }
1906
1907 if (md->queue)
1908 blk_cleanup_queue(md->queue);
1909
1910 cleanup_srcu_struct(&md->io_barrier);
1911
1912 if (md->bdev) {
1913 bdput(md->bdev);
1914 md->bdev = NULL;
1915 }
1916
1917 mutex_destroy(&md->suspend_lock);
1918 mutex_destroy(&md->type_lock);
1919 mutex_destroy(&md->table_devices_lock);
1920
1921 dm_mq_cleanup_mapped_device(md);
1922}
1923
1924/*
1925 * Allocate and initialise a blank device with a given minor.
1926 */
1927static struct mapped_device *alloc_dev(int minor)
1928{
1929 int r, numa_node_id = dm_get_numa_node();
1930 struct mapped_device *md;
1931 void *old_md;
1932
1933 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
1934 if (!md) {
1935 DMWARN("unable to allocate device, out of memory.");
1936 return NULL;
1937 }
1938
1939 if (!try_module_get(THIS_MODULE))
1940 goto bad_module_get;
1941
1942 /* get a minor number for the dev */
1943 if (minor == DM_ANY_MINOR)
1944 r = next_free_minor(&minor);
1945 else
1946 r = specific_minor(minor);
1947 if (r < 0)
1948 goto bad_minor;
1949
1950 r = init_srcu_struct(&md->io_barrier);
1951 if (r < 0)
1952 goto bad_io_barrier;
1953
1954 md->numa_node_id = numa_node_id;
1955 md->init_tio_pdu = false;
1956 md->type = DM_TYPE_NONE;
1957 mutex_init(&md->suspend_lock);
1958 mutex_init(&md->type_lock);
1959 mutex_init(&md->table_devices_lock);
1960 spin_lock_init(&md->deferred_lock);
1961 atomic_set(&md->holders, 1);
1962 atomic_set(&md->open_count, 0);
1963 atomic_set(&md->event_nr, 0);
1964 atomic_set(&md->uevent_seq, 0);
1965 INIT_LIST_HEAD(&md->uevent_list);
1966 INIT_LIST_HEAD(&md->table_devices);
1967 spin_lock_init(&md->uevent_lock);
1968
1969 md->queue = blk_alloc_queue_node(GFP_KERNEL, numa_node_id);
1970 if (!md->queue)
1971 goto bad;
1972 md->queue->queuedata = md;
1973 md->queue->backing_dev_info->congested_data = md;
1974
1975 md->disk = alloc_disk_node(1, md->numa_node_id);
1976 if (!md->disk)
1977 goto bad;
1978
1979 init_waitqueue_head(&md->wait);
1980 INIT_WORK(&md->work, dm_wq_work);
1981 init_waitqueue_head(&md->eventq);
1982 init_completion(&md->kobj_holder.completion);
1983
1984 md->disk->major = _major;
1985 md->disk->first_minor = minor;
1986 md->disk->fops = &dm_blk_dops;
1987 md->disk->queue = md->queue;
1988 md->disk->private_data = md;
1989 sprintf(md->disk->disk_name, "dm-%d", minor);
1990
1991 if (IS_ENABLED(CONFIG_DAX_DRIVER)) {
1992 md->dax_dev = alloc_dax(md, md->disk->disk_name,
1993 &dm_dax_ops, 0);
1994 if (!md->dax_dev)
1995 goto bad;
1996 }
1997
1998 add_disk_no_queue_reg(md->disk);
1999 format_dev_t(md->name, MKDEV(_major, minor));
2000
2001 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
2002 if (!md->wq)
2003 goto bad;
2004
2005 md->bdev = bdget_disk(md->disk, 0);
2006 if (!md->bdev)
2007 goto bad;
2008
2009 dm_stats_init(&md->stats);
2010
2011 /* Populate the mapping, nobody knows we exist yet */
2012 spin_lock(&_minor_lock);
2013 old_md = idr_replace(&_minor_idr, md, minor);
2014 spin_unlock(&_minor_lock);
2015
2016 BUG_ON(old_md != MINOR_ALLOCED);
2017
2018 return md;
2019
2020bad:
2021 cleanup_mapped_device(md);
2022bad_io_barrier:
2023 free_minor(minor);
2024bad_minor:
2025 module_put(THIS_MODULE);
2026bad_module_get:
2027 kvfree(md);
2028 return NULL;
2029}
2030
2031static void unlock_fs(struct mapped_device *md);
2032
2033static void free_dev(struct mapped_device *md)
2034{
2035 int minor = MINOR(disk_devt(md->disk));
2036
2037 unlock_fs(md);
2038
2039 cleanup_mapped_device(md);
2040
2041 free_table_devices(&md->table_devices);
2042 dm_stats_cleanup(&md->stats);
2043 free_minor(minor);
2044
2045 module_put(THIS_MODULE);
2046 kvfree(md);
2047}
2048
2049static int __bind_mempools(struct mapped_device *md, struct dm_table *t)
2050{
2051 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
2052 int ret = 0;
2053
2054 if (dm_table_bio_based(t)) {
2055 /*
2056 * The md may already have mempools that need changing.
2057 * If so, reload bioset because front_pad may have changed
2058 * because a different table was loaded.
2059 */
2060 bioset_exit(&md->bs);
2061 bioset_exit(&md->io_bs);
2062
2063 } else if (bioset_initialized(&md->bs)) {
2064 /*
2065 * There's no need to reload with request-based dm
2066 * because the size of front_pad doesn't change.
2067 * Note for future: If you are to reload bioset,
2068 * prep-ed requests in the queue may refer
2069 * to bio from the old bioset, so you must walk
2070 * through the queue to unprep.
2071 */
2072 goto out;
2073 }
2074
2075 BUG_ON(!p ||
2076 bioset_initialized(&md->bs) ||
2077 bioset_initialized(&md->io_bs));
2078
2079 ret = bioset_init_from_src(&md->bs, &p->bs);
2080 if (ret)
2081 goto out;
2082 ret = bioset_init_from_src(&md->io_bs, &p->io_bs);
2083 if (ret)
2084 bioset_exit(&md->bs);
2085out:
2086 /* mempool bind completed, no longer need any mempools in the table */
2087 dm_table_free_md_mempools(t);
2088 return ret;
2089}
2090
2091/*
2092 * Bind a table to the device.
2093 */
2094static void event_callback(void *context)
2095{
2096 unsigned long flags;
2097 LIST_HEAD(uevents);
2098 struct mapped_device *md = (struct mapped_device *) context;
2099
2100 spin_lock_irqsave(&md->uevent_lock, flags);
2101 list_splice_init(&md->uevent_list, &uevents);
2102 spin_unlock_irqrestore(&md->uevent_lock, flags);
2103
2104 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2105
2106 atomic_inc(&md->event_nr);
2107 wake_up(&md->eventq);
2108 dm_issue_global_event();
2109}
2110
2111/*
2112 * Protected by md->suspend_lock obtained by dm_swap_table().
2113 */
2114static void __set_size(struct mapped_device *md, sector_t size)
2115{
2116 lockdep_assert_held(&md->suspend_lock);
2117
2118 set_capacity(md->disk, size);
2119
2120 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2121}
2122
2123/*
2124 * Returns old map, which caller must destroy.
2125 */
2126static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2127 struct queue_limits *limits)
2128{
2129 struct dm_table *old_map;
2130 struct request_queue *q = md->queue;
2131 bool request_based = dm_table_request_based(t);
2132 sector_t size;
2133 int ret;
2134
2135 lockdep_assert_held(&md->suspend_lock);
2136
2137 size = dm_table_get_size(t);
2138
2139 /*
2140 * Wipe any geometry if the size of the table changed.
2141 */
2142 if (size != dm_get_size(md))
2143 memset(&md->geometry, 0, sizeof(md->geometry));
2144
2145 __set_size(md, size);
2146
2147 dm_table_event_callback(t, event_callback, md);
2148
2149 /*
2150 * The queue hasn't been stopped yet, if the old table type wasn't
2151 * for request-based during suspension. So stop it to prevent
2152 * I/O mapping before resume.
2153 * This must be done before setting the queue restrictions,
2154 * because request-based dm may be run just after the setting.
2155 */
2156 if (request_based)
2157 dm_stop_queue(q);
2158
2159 if (request_based || md->type == DM_TYPE_NVME_BIO_BASED) {
2160 /*
2161 * Leverage the fact that request-based DM targets and
2162 * NVMe bio based targets are immutable singletons
2163 * - used to optimize both dm_request_fn and dm_mq_queue_rq;
2164 * and __process_bio.
2165 */
2166 md->immutable_target = dm_table_get_immutable_target(t);
2167 }
2168
2169 ret = __bind_mempools(md, t);
2170 if (ret) {
2171 old_map = ERR_PTR(ret);
2172 goto out;
2173 }
2174
2175 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2176 rcu_assign_pointer(md->map, (void *)t);
2177 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2178
2179 dm_table_set_restrictions(t, q, limits);
2180 if (old_map)
2181 dm_sync_table(md);
2182
2183out:
2184 return old_map;
2185}
2186
2187/*
2188 * Returns unbound table for the caller to free.
2189 */
2190static struct dm_table *__unbind(struct mapped_device *md)
2191{
2192 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2193
2194 if (!map)
2195 return NULL;
2196
2197 dm_table_event_callback(map, NULL, NULL);
2198 RCU_INIT_POINTER(md->map, NULL);
2199 dm_sync_table(md);
2200
2201 return map;
2202}
2203
2204/*
2205 * Constructor for a new device.
2206 */
2207int dm_create(int minor, struct mapped_device **result)
2208{
2209 int r;
2210 struct mapped_device *md;
2211
2212 md = alloc_dev(minor);
2213 if (!md)
2214 return -ENXIO;
2215
2216 r = dm_sysfs_init(md);
2217 if (r) {
2218 free_dev(md);
2219 return r;
2220 }
2221
2222 *result = md;
2223 return 0;
2224}
2225
2226/*
2227 * Functions to manage md->type.
2228 * All are required to hold md->type_lock.
2229 */
2230void dm_lock_md_type(struct mapped_device *md)
2231{
2232 mutex_lock(&md->type_lock);
2233}
2234
2235void dm_unlock_md_type(struct mapped_device *md)
2236{
2237 mutex_unlock(&md->type_lock);
2238}
2239
2240void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2241{
2242 BUG_ON(!mutex_is_locked(&md->type_lock));
2243 md->type = type;
2244}
2245
2246enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2247{
2248 return md->type;
2249}
2250
2251struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2252{
2253 return md->immutable_target_type;
2254}
2255
2256/*
2257 * The queue_limits are only valid as long as you have a reference
2258 * count on 'md'.
2259 */
2260struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2261{
2262 BUG_ON(!atomic_read(&md->holders));
2263 return &md->queue->limits;
2264}
2265EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2266
2267/*
2268 * Setup the DM device's queue based on md's type
2269 */
2270int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2271{
2272 int r;
2273 struct queue_limits limits;
2274 enum dm_queue_mode type = dm_get_md_type(md);
2275
2276 switch (type) {
2277 case DM_TYPE_REQUEST_BASED:
2278 r = dm_mq_init_request_queue(md, t);
2279 if (r) {
2280 DMERR("Cannot initialize queue for request-based dm-mq mapped device");
2281 return r;
2282 }
2283 break;
2284 case DM_TYPE_BIO_BASED:
2285 case DM_TYPE_DAX_BIO_BASED:
2286 case DM_TYPE_NVME_BIO_BASED:
2287 dm_init_normal_md_queue(md);
2288 blk_queue_make_request(md->queue, dm_make_request);
2289 break;
2290 case DM_TYPE_NONE:
2291 WARN_ON_ONCE(true);
2292 break;
2293 }
2294
2295 r = dm_calculate_queue_limits(t, &limits);
2296 if (r) {
2297 DMERR("Cannot calculate initial queue limits");
2298 return r;
2299 }
2300 dm_table_set_restrictions(t, md->queue, &limits);
2301 blk_register_queue(md->disk);
2302
2303 return 0;
2304}
2305
2306struct mapped_device *dm_get_md(dev_t dev)
2307{
2308 struct mapped_device *md;
2309 unsigned minor = MINOR(dev);
2310
2311 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2312 return NULL;
2313
2314 spin_lock(&_minor_lock);
2315
2316 md = idr_find(&_minor_idr, minor);
2317 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2318 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2319 md = NULL;
2320 goto out;
2321 }
2322 dm_get(md);
2323out:
2324 spin_unlock(&_minor_lock);
2325
2326 return md;
2327}
2328EXPORT_SYMBOL_GPL(dm_get_md);
2329
2330void *dm_get_mdptr(struct mapped_device *md)
2331{
2332 return md->interface_ptr;
2333}
2334
2335void dm_set_mdptr(struct mapped_device *md, void *ptr)
2336{
2337 md->interface_ptr = ptr;
2338}
2339
2340void dm_get(struct mapped_device *md)
2341{
2342 atomic_inc(&md->holders);
2343 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2344}
2345
2346int dm_hold(struct mapped_device *md)
2347{
2348 spin_lock(&_minor_lock);
2349 if (test_bit(DMF_FREEING, &md->flags)) {
2350 spin_unlock(&_minor_lock);
2351 return -EBUSY;
2352 }
2353 dm_get(md);
2354 spin_unlock(&_minor_lock);
2355 return 0;
2356}
2357EXPORT_SYMBOL_GPL(dm_hold);
2358
2359const char *dm_device_name(struct mapped_device *md)
2360{
2361 return md->name;
2362}
2363EXPORT_SYMBOL_GPL(dm_device_name);
2364
2365static void __dm_destroy(struct mapped_device *md, bool wait)
2366{
2367 struct dm_table *map;
2368 int srcu_idx;
2369
2370 might_sleep();
2371
2372 spin_lock(&_minor_lock);
2373 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2374 set_bit(DMF_FREEING, &md->flags);
2375 spin_unlock(&_minor_lock);
2376
2377 blk_set_queue_dying(md->queue);
2378
2379 /*
2380 * Take suspend_lock so that presuspend and postsuspend methods
2381 * do not race with internal suspend.
2382 */
2383 mutex_lock(&md->suspend_lock);
2384 map = dm_get_live_table(md, &srcu_idx);
2385 if (!dm_suspended_md(md)) {
2386 dm_table_presuspend_targets(map);
2387 dm_table_postsuspend_targets(map);
2388 }
2389 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2390 dm_put_live_table(md, srcu_idx);
2391 mutex_unlock(&md->suspend_lock);
2392
2393 /*
2394 * Rare, but there may be I/O requests still going to complete,
2395 * for example. Wait for all references to disappear.
2396 * No one should increment the reference count of the mapped_device,
2397 * after the mapped_device state becomes DMF_FREEING.
2398 */
2399 if (wait)
2400 while (atomic_read(&md->holders))
2401 msleep(1);
2402 else if (atomic_read(&md->holders))
2403 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2404 dm_device_name(md), atomic_read(&md->holders));
2405
2406 dm_sysfs_exit(md);
2407 dm_table_destroy(__unbind(md));
2408 free_dev(md);
2409}
2410
2411void dm_destroy(struct mapped_device *md)
2412{
2413 __dm_destroy(md, true);
2414}
2415
2416void dm_destroy_immediate(struct mapped_device *md)
2417{
2418 __dm_destroy(md, false);
2419}
2420
2421void dm_put(struct mapped_device *md)
2422{
2423 atomic_dec(&md->holders);
2424}
2425EXPORT_SYMBOL_GPL(dm_put);
2426
2427static int dm_wait_for_completion(struct mapped_device *md, long task_state)
2428{
2429 int r = 0;
2430 DEFINE_WAIT(wait);
2431
2432 while (1) {
2433 prepare_to_wait(&md->wait, &wait, task_state);
2434
2435 if (!md_in_flight(md))
2436 break;
2437
2438 if (signal_pending_state(task_state, current)) {
2439 r = -EINTR;
2440 break;
2441 }
2442
2443 io_schedule();
2444 }
2445 finish_wait(&md->wait, &wait);
2446
2447 return r;
2448}
2449
2450/*
2451 * Process the deferred bios
2452 */
2453static void dm_wq_work(struct work_struct *work)
2454{
2455 struct mapped_device *md = container_of(work, struct mapped_device,
2456 work);
2457 struct bio *c;
2458 int srcu_idx;
2459 struct dm_table *map;
2460
2461 map = dm_get_live_table(md, &srcu_idx);
2462
2463 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2464 spin_lock_irq(&md->deferred_lock);
2465 c = bio_list_pop(&md->deferred);
2466 spin_unlock_irq(&md->deferred_lock);
2467
2468 if (!c)
2469 break;
2470
2471 if (dm_request_based(md))
2472 (void) generic_make_request(c);
2473 else
2474 (void) dm_process_bio(md, map, c);
2475 }
2476
2477 dm_put_live_table(md, srcu_idx);
2478}
2479
2480static void dm_queue_flush(struct mapped_device *md)
2481{
2482 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2483 smp_mb__after_atomic();
2484 queue_work(md->wq, &md->work);
2485}
2486
2487/*
2488 * Swap in a new table, returning the old one for the caller to destroy.
2489 */
2490struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2491{
2492 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2493 struct queue_limits limits;
2494 int r;
2495
2496 mutex_lock(&md->suspend_lock);
2497
2498 /* device must be suspended */
2499 if (!dm_suspended_md(md))
2500 goto out;
2501
2502 /*
2503 * If the new table has no data devices, retain the existing limits.
2504 * This helps multipath with queue_if_no_path if all paths disappear,
2505 * then new I/O is queued based on these limits, and then some paths
2506 * reappear.
2507 */
2508 if (dm_table_has_no_data_devices(table)) {
2509 live_map = dm_get_live_table_fast(md);
2510 if (live_map)
2511 limits = md->queue->limits;
2512 dm_put_live_table_fast(md);
2513 }
2514
2515 if (!live_map) {
2516 r = dm_calculate_queue_limits(table, &limits);
2517 if (r) {
2518 map = ERR_PTR(r);
2519 goto out;
2520 }
2521 }
2522
2523 map = __bind(md, table, &limits);
2524 dm_issue_global_event();
2525
2526out:
2527 mutex_unlock(&md->suspend_lock);
2528 return map;
2529}
2530
2531/*
2532 * Functions to lock and unlock any filesystem running on the
2533 * device.
2534 */
2535static int lock_fs(struct mapped_device *md)
2536{
2537 int r;
2538
2539 WARN_ON(md->frozen_sb);
2540
2541 md->frozen_sb = freeze_bdev(md->bdev);
2542 if (IS_ERR(md->frozen_sb)) {
2543 r = PTR_ERR(md->frozen_sb);
2544 md->frozen_sb = NULL;
2545 return r;
2546 }
2547
2548 set_bit(DMF_FROZEN, &md->flags);
2549
2550 return 0;
2551}
2552
2553static void unlock_fs(struct mapped_device *md)
2554{
2555 if (!test_bit(DMF_FROZEN, &md->flags))
2556 return;
2557
2558 thaw_bdev(md->bdev, md->frozen_sb);
2559 md->frozen_sb = NULL;
2560 clear_bit(DMF_FROZEN, &md->flags);
2561}
2562
2563/*
2564 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2565 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2566 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2567 *
2568 * If __dm_suspend returns 0, the device is completely quiescent
2569 * now. There is no request-processing activity. All new requests
2570 * are being added to md->deferred list.
2571 */
2572static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2573 unsigned suspend_flags, long task_state,
2574 int dmf_suspended_flag)
2575{
2576 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2577 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2578 int r;
2579
2580 lockdep_assert_held(&md->suspend_lock);
2581
2582 /*
2583 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2584 * This flag is cleared before dm_suspend returns.
2585 */
2586 if (noflush)
2587 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2588 else
2589 pr_debug("%s: suspending with flush\n", dm_device_name(md));
2590
2591 /*
2592 * This gets reverted if there's an error later and the targets
2593 * provide the .presuspend_undo hook.
2594 */
2595 dm_table_presuspend_targets(map);
2596
2597 /*
2598 * Flush I/O to the device.
2599 * Any I/O submitted after lock_fs() may not be flushed.
2600 * noflush takes precedence over do_lockfs.
2601 * (lock_fs() flushes I/Os and waits for them to complete.)
2602 */
2603 if (!noflush && do_lockfs) {
2604 r = lock_fs(md);
2605 if (r) {
2606 dm_table_presuspend_undo_targets(map);
2607 return r;
2608 }
2609 }
2610
2611 /*
2612 * Here we must make sure that no processes are submitting requests
2613 * to target drivers i.e. no one may be executing
2614 * __split_and_process_bio. This is called from dm_request and
2615 * dm_wq_work.
2616 *
2617 * To get all processes out of __split_and_process_bio in dm_request,
2618 * we take the write lock. To prevent any process from reentering
2619 * __split_and_process_bio from dm_request and quiesce the thread
2620 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
2621 * flush_workqueue(md->wq).
2622 */
2623 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2624 if (map)
2625 synchronize_srcu(&md->io_barrier);
2626
2627 /*
2628 * Stop md->queue before flushing md->wq in case request-based
2629 * dm defers requests to md->wq from md->queue.
2630 */
2631 if (dm_request_based(md))
2632 dm_stop_queue(md->queue);
2633
2634 flush_workqueue(md->wq);
2635
2636 /*
2637 * At this point no more requests are entering target request routines.
2638 * We call dm_wait_for_completion to wait for all existing requests
2639 * to finish.
2640 */
2641 r = dm_wait_for_completion(md, task_state);
2642 if (!r)
2643 set_bit(dmf_suspended_flag, &md->flags);
2644
2645 if (noflush)
2646 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2647 if (map)
2648 synchronize_srcu(&md->io_barrier);
2649
2650 /* were we interrupted ? */
2651 if (r < 0) {
2652 dm_queue_flush(md);
2653
2654 if (dm_request_based(md))
2655 dm_start_queue(md->queue);
2656
2657 unlock_fs(md);
2658 dm_table_presuspend_undo_targets(map);
2659 /* pushback list is already flushed, so skip flush */
2660 }
2661
2662 return r;
2663}
2664
2665/*
2666 * We need to be able to change a mapping table under a mounted
2667 * filesystem. For example we might want to move some data in
2668 * the background. Before the table can be swapped with
2669 * dm_bind_table, dm_suspend must be called to flush any in
2670 * flight bios and ensure that any further io gets deferred.
2671 */
2672/*
2673 * Suspend mechanism in request-based dm.
2674 *
2675 * 1. Flush all I/Os by lock_fs() if needed.
2676 * 2. Stop dispatching any I/O by stopping the request_queue.
2677 * 3. Wait for all in-flight I/Os to be completed or requeued.
2678 *
2679 * To abort suspend, start the request_queue.
2680 */
2681int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
2682{
2683 struct dm_table *map = NULL;
2684 int r = 0;
2685
2686retry:
2687 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2688
2689 if (dm_suspended_md(md)) {
2690 r = -EINVAL;
2691 goto out_unlock;
2692 }
2693
2694 if (dm_suspended_internally_md(md)) {
2695 /* already internally suspended, wait for internal resume */
2696 mutex_unlock(&md->suspend_lock);
2697 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2698 if (r)
2699 return r;
2700 goto retry;
2701 }
2702
2703 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2704
2705 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2706 if (r)
2707 goto out_unlock;
2708
2709 dm_table_postsuspend_targets(map);
2710
2711out_unlock:
2712 mutex_unlock(&md->suspend_lock);
2713 return r;
2714}
2715
2716static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2717{
2718 if (map) {
2719 int r = dm_table_resume_targets(map);
2720 if (r)
2721 return r;
2722 }
2723
2724 dm_queue_flush(md);
2725
2726 /*
2727 * Flushing deferred I/Os must be done after targets are resumed
2728 * so that mapping of targets can work correctly.
2729 * Request-based dm is queueing the deferred I/Os in its request_queue.
2730 */
2731 if (dm_request_based(md))
2732 dm_start_queue(md->queue);
2733
2734 unlock_fs(md);
2735
2736 return 0;
2737}
2738
2739int dm_resume(struct mapped_device *md)
2740{
2741 int r;
2742 struct dm_table *map = NULL;
2743
2744retry:
2745 r = -EINVAL;
2746 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2747
2748 if (!dm_suspended_md(md))
2749 goto out;
2750
2751 if (dm_suspended_internally_md(md)) {
2752 /* already internally suspended, wait for internal resume */
2753 mutex_unlock(&md->suspend_lock);
2754 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2755 if (r)
2756 return r;
2757 goto retry;
2758 }
2759
2760 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2761 if (!map || !dm_table_get_size(map))
2762 goto out;
2763
2764 r = __dm_resume(md, map);
2765 if (r)
2766 goto out;
2767
2768 clear_bit(DMF_SUSPENDED, &md->flags);
2769out:
2770 mutex_unlock(&md->suspend_lock);
2771
2772 return r;
2773}
2774
2775/*
2776 * Internal suspend/resume works like userspace-driven suspend. It waits
2777 * until all bios finish and prevents issuing new bios to the target drivers.
2778 * It may be used only from the kernel.
2779 */
2780
2781static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
2782{
2783 struct dm_table *map = NULL;
2784
2785 lockdep_assert_held(&md->suspend_lock);
2786
2787 if (md->internal_suspend_count++)
2788 return; /* nested internal suspend */
2789
2790 if (dm_suspended_md(md)) {
2791 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2792 return; /* nest suspend */
2793 }
2794
2795 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2796
2797 /*
2798 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2799 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2800 * would require changing .presuspend to return an error -- avoid this
2801 * until there is a need for more elaborate variants of internal suspend.
2802 */
2803 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2804 DMF_SUSPENDED_INTERNALLY);
2805
2806 dm_table_postsuspend_targets(map);
2807}
2808
2809static void __dm_internal_resume(struct mapped_device *md)
2810{
2811 BUG_ON(!md->internal_suspend_count);
2812
2813 if (--md->internal_suspend_count)
2814 return; /* resume from nested internal suspend */
2815
2816 if (dm_suspended_md(md))
2817 goto done; /* resume from nested suspend */
2818
2819 /*
2820 * NOTE: existing callers don't need to call dm_table_resume_targets
2821 * (which may fail -- so best to avoid it for now by passing NULL map)
2822 */
2823 (void) __dm_resume(md, NULL);
2824
2825done:
2826 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2827 smp_mb__after_atomic();
2828 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
2829}
2830
2831void dm_internal_suspend_noflush(struct mapped_device *md)
2832{
2833 mutex_lock(&md->suspend_lock);
2834 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2835 mutex_unlock(&md->suspend_lock);
2836}
2837EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2838
2839void dm_internal_resume(struct mapped_device *md)
2840{
2841 mutex_lock(&md->suspend_lock);
2842 __dm_internal_resume(md);
2843 mutex_unlock(&md->suspend_lock);
2844}
2845EXPORT_SYMBOL_GPL(dm_internal_resume);
2846
2847/*
2848 * Fast variants of internal suspend/resume hold md->suspend_lock,
2849 * which prevents interaction with userspace-driven suspend.
2850 */
2851
2852void dm_internal_suspend_fast(struct mapped_device *md)
2853{
2854 mutex_lock(&md->suspend_lock);
2855 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2856 return;
2857
2858 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2859 synchronize_srcu(&md->io_barrier);
2860 flush_workqueue(md->wq);
2861 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
2862}
2863EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
2864
2865void dm_internal_resume_fast(struct mapped_device *md)
2866{
2867 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
2868 goto done;
2869
2870 dm_queue_flush(md);
2871
2872done:
2873 mutex_unlock(&md->suspend_lock);
2874}
2875EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
2876
2877/*-----------------------------------------------------------------
2878 * Event notification.
2879 *---------------------------------------------------------------*/
2880int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
2881 unsigned cookie)
2882{
2883 char udev_cookie[DM_COOKIE_LENGTH];
2884 char *envp[] = { udev_cookie, NULL };
2885
2886 if (!cookie)
2887 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
2888 else {
2889 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
2890 DM_COOKIE_ENV_VAR_NAME, cookie);
2891 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
2892 action, envp);
2893 }
2894}
2895
2896uint32_t dm_next_uevent_seq(struct mapped_device *md)
2897{
2898 return atomic_add_return(1, &md->uevent_seq);
2899}
2900
2901uint32_t dm_get_event_nr(struct mapped_device *md)
2902{
2903 return atomic_read(&md->event_nr);
2904}
2905
2906int dm_wait_event(struct mapped_device *md, int event_nr)
2907{
2908 return wait_event_interruptible(md->eventq,
2909 (event_nr != atomic_read(&md->event_nr)));
2910}
2911
2912void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
2913{
2914 unsigned long flags;
2915
2916 spin_lock_irqsave(&md->uevent_lock, flags);
2917 list_add(elist, &md->uevent_list);
2918 spin_unlock_irqrestore(&md->uevent_lock, flags);
2919}
2920
2921/*
2922 * The gendisk is only valid as long as you have a reference
2923 * count on 'md'.
2924 */
2925struct gendisk *dm_disk(struct mapped_device *md)
2926{
2927 return md->disk;
2928}
2929EXPORT_SYMBOL_GPL(dm_disk);
2930
2931struct kobject *dm_kobject(struct mapped_device *md)
2932{
2933 return &md->kobj_holder.kobj;
2934}
2935
2936struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
2937{
2938 struct mapped_device *md;
2939
2940 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
2941
2942 spin_lock(&_minor_lock);
2943 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2944 md = NULL;
2945 goto out;
2946 }
2947 dm_get(md);
2948out:
2949 spin_unlock(&_minor_lock);
2950
2951 return md;
2952}
2953
2954int dm_suspended_md(struct mapped_device *md)
2955{
2956 return test_bit(DMF_SUSPENDED, &md->flags);
2957}
2958
2959int dm_suspended_internally_md(struct mapped_device *md)
2960{
2961 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2962}
2963
2964int dm_test_deferred_remove_flag(struct mapped_device *md)
2965{
2966 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
2967}
2968
2969int dm_suspended(struct dm_target *ti)
2970{
2971 return dm_suspended_md(dm_table_get_md(ti->table));
2972}
2973EXPORT_SYMBOL_GPL(dm_suspended);
2974
2975int dm_noflush_suspending(struct dm_target *ti)
2976{
2977 return __noflush_suspending(dm_table_get_md(ti->table));
2978}
2979EXPORT_SYMBOL_GPL(dm_noflush_suspending);
2980
2981struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, enum dm_queue_mode type,
2982 unsigned integrity, unsigned per_io_data_size,
2983 unsigned min_pool_size)
2984{
2985 struct dm_md_mempools *pools = kzalloc_node(sizeof(*pools), GFP_KERNEL, md->numa_node_id);
2986 unsigned int pool_size = 0;
2987 unsigned int front_pad, io_front_pad;
2988 int ret;
2989
2990 if (!pools)
2991 return NULL;
2992
2993 switch (type) {
2994 case DM_TYPE_BIO_BASED:
2995 case DM_TYPE_DAX_BIO_BASED:
2996 case DM_TYPE_NVME_BIO_BASED:
2997 pool_size = max(dm_get_reserved_bio_based_ios(), min_pool_size);
2998 front_pad = roundup(per_io_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
2999 io_front_pad = roundup(front_pad, __alignof__(struct dm_io)) + offsetof(struct dm_io, tio);
3000 ret = bioset_init(&pools->io_bs, pool_size, io_front_pad, 0);
3001 if (ret)
3002 goto out;
3003 if (integrity && bioset_integrity_create(&pools->io_bs, pool_size))
3004 goto out;
3005 break;
3006 case DM_TYPE_REQUEST_BASED:
3007 pool_size = max(dm_get_reserved_rq_based_ios(), min_pool_size);
3008 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
3009 /* per_io_data_size is used for blk-mq pdu at queue allocation */
3010 break;
3011 default:
3012 BUG();
3013 }
3014
3015 ret = bioset_init(&pools->bs, pool_size, front_pad, 0);
3016 if (ret)
3017 goto out;
3018
3019 if (integrity && bioset_integrity_create(&pools->bs, pool_size))
3020 goto out;
3021
3022 return pools;
3023
3024out:
3025 dm_free_md_mempools(pools);
3026
3027 return NULL;
3028}
3029
3030void dm_free_md_mempools(struct dm_md_mempools *pools)
3031{
3032 if (!pools)
3033 return;
3034
3035 bioset_exit(&pools->bs);
3036 bioset_exit(&pools->io_bs);
3037
3038 kfree(pools);
3039}
3040
3041struct dm_pr {
3042 u64 old_key;
3043 u64 new_key;
3044 u32 flags;
3045 bool fail_early;
3046};
3047
3048static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3049 void *data)
3050{
3051 struct mapped_device *md = bdev->bd_disk->private_data;
3052 struct dm_table *table;
3053 struct dm_target *ti;
3054 int ret = -ENOTTY, srcu_idx;
3055
3056 table = dm_get_live_table(md, &srcu_idx);
3057 if (!table || !dm_table_get_size(table))
3058 goto out;
3059
3060 /* We only support devices that have a single target */
3061 if (dm_table_get_num_targets(table) != 1)
3062 goto out;
3063 ti = dm_table_get_target(table, 0);
3064
3065 ret = -EINVAL;
3066 if (!ti->type->iterate_devices)
3067 goto out;
3068
3069 ret = ti->type->iterate_devices(ti, fn, data);
3070out:
3071 dm_put_live_table(md, srcu_idx);
3072 return ret;
3073}
3074
3075/*
3076 * For register / unregister we need to manually call out to every path.
3077 */
3078static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3079 sector_t start, sector_t len, void *data)
3080{
3081 struct dm_pr *pr = data;
3082 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3083
3084 if (!ops || !ops->pr_register)
3085 return -EOPNOTSUPP;
3086 return ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3087}
3088
3089static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3090 u32 flags)
3091{
3092 struct dm_pr pr = {
3093 .old_key = old_key,
3094 .new_key = new_key,
3095 .flags = flags,
3096 .fail_early = true,
3097 };
3098 int ret;
3099
3100 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3101 if (ret && new_key) {
3102 /* unregister all paths if we failed to register any path */
3103 pr.old_key = new_key;
3104 pr.new_key = 0;
3105 pr.flags = 0;
3106 pr.fail_early = false;
3107 dm_call_pr(bdev, __dm_pr_register, &pr);
3108 }
3109
3110 return ret;
3111}
3112
3113static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3114 u32 flags)
3115{
3116 struct mapped_device *md = bdev->bd_disk->private_data;
3117 const struct pr_ops *ops;
3118 int r, srcu_idx;
3119
3120 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3121 if (r < 0)
3122 goto out;
3123
3124 ops = bdev->bd_disk->fops->pr_ops;
3125 if (ops && ops->pr_reserve)
3126 r = ops->pr_reserve(bdev, key, type, flags);
3127 else
3128 r = -EOPNOTSUPP;
3129out:
3130 dm_unprepare_ioctl(md, srcu_idx);
3131 return r;
3132}
3133
3134static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3135{
3136 struct mapped_device *md = bdev->bd_disk->private_data;
3137 const struct pr_ops *ops;
3138 int r, srcu_idx;
3139
3140 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3141 if (r < 0)
3142 goto out;
3143
3144 ops = bdev->bd_disk->fops->pr_ops;
3145 if (ops && ops->pr_release)
3146 r = ops->pr_release(bdev, key, type);
3147 else
3148 r = -EOPNOTSUPP;
3149out:
3150 dm_unprepare_ioctl(md, srcu_idx);
3151 return r;
3152}
3153
3154static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3155 enum pr_type type, bool abort)
3156{
3157 struct mapped_device *md = bdev->bd_disk->private_data;
3158 const struct pr_ops *ops;
3159 int r, srcu_idx;
3160
3161 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3162 if (r < 0)
3163 goto out;
3164
3165 ops = bdev->bd_disk->fops->pr_ops;
3166 if (ops && ops->pr_preempt)
3167 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
3168 else
3169 r = -EOPNOTSUPP;
3170out:
3171 dm_unprepare_ioctl(md, srcu_idx);
3172 return r;
3173}
3174
3175static int dm_pr_clear(struct block_device *bdev, u64 key)
3176{
3177 struct mapped_device *md = bdev->bd_disk->private_data;
3178 const struct pr_ops *ops;
3179 int r, srcu_idx;
3180
3181 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3182 if (r < 0)
3183 goto out;
3184
3185 ops = bdev->bd_disk->fops->pr_ops;
3186 if (ops && ops->pr_clear)
3187 r = ops->pr_clear(bdev, key);
3188 else
3189 r = -EOPNOTSUPP;
3190out:
3191 dm_unprepare_ioctl(md, srcu_idx);
3192 return r;
3193}
3194
3195static const struct pr_ops dm_pr_ops = {
3196 .pr_register = dm_pr_register,
3197 .pr_reserve = dm_pr_reserve,
3198 .pr_release = dm_pr_release,
3199 .pr_preempt = dm_pr_preempt,
3200 .pr_clear = dm_pr_clear,
3201};
3202
3203static const struct block_device_operations dm_blk_dops = {
3204 .open = dm_blk_open,
3205 .release = dm_blk_close,
3206 .ioctl = dm_blk_ioctl,
3207 .getgeo = dm_blk_getgeo,
3208 .report_zones = dm_blk_report_zones,
3209 .pr_ops = &dm_pr_ops,
3210 .owner = THIS_MODULE
3211};
3212
3213static const struct dax_operations dm_dax_ops = {
3214 .direct_access = dm_dax_direct_access,
3215 .dax_supported = dm_dax_supported,
3216 .copy_from_iter = dm_dax_copy_from_iter,
3217 .copy_to_iter = dm_dax_copy_to_iter,
3218};
3219
3220/*
3221 * module hooks
3222 */
3223module_init(dm_init);
3224module_exit(dm_exit);
3225
3226module_param(major, uint, 0);
3227MODULE_PARM_DESC(major, "The major number of the device mapper");
3228
3229module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3230MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3231
3232module_param(dm_numa_node, int, S_IRUGO | S_IWUSR);
3233MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3234
3235MODULE_DESCRIPTION(DM_NAME " driver");
3236MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3237MODULE_LICENSE("GPL");