2 * Copyright (C) 2014 Facebook. All rights reserved.
4 * This file is released under the GPL.
7 #include <linux/device-mapper.h>
9 #include <linux/module.h>
10 #include <linux/init.h>
11 #include <linux/blkdev.h>
12 #include <linux/bio.h>
13 #include <linux/slab.h>
14 #include <linux/kthread.h>
15 #include <linux/freezer.h>
17 #define DM_MSG_PREFIX "log-writes"
20 * This target will sequentially log all writes to the target device onto the
21 * log device. This is helpful for replaying writes to check for fs consistency
22 * at all times. This target provides a mechanism to mark specific events to
23 * check data at a later time. So for example you would:
27 * dmsetup message /dev/whatever mark mymark
30 * Then replay the log up to mymark and check the contents of the replay to
31 * verify it matches what was written.
33 * We log writes only after they have been flushed, this makes the log describe
34 * close to the order in which the data hits the actual disk, not its cache. So
35 * for example the following sequence (W means write, C means complete)
37 * Wa,Wb,Wc,Cc,Ca,FLUSH,FUAd,Cb,CFLUSH,CFUAd
39 * Would result in the log looking like this:
41 * c,a,flush,fuad,b,<other writes>,<next flush>
43 * This is meant to help expose problems where file systems do not properly wait
44 * on data being written before invoking a FLUSH. FUA bypasses cache so once it
45 * completes it is added to the log as it should be on disk.
47 * We treat DISCARDs as if they don't bypass cache so that they are logged in
48 * order of completion along with the normal writes. If we didn't do it this
49 * way we would process all the discards first and then write all the data, when
50 * in fact we want to do the data and the discard in the order that they
53 #define LOG_FLUSH_FLAG (1 << 0)
54 #define LOG_FUA_FLAG (1 << 1)
55 #define LOG_DISCARD_FLAG (1 << 2)
56 #define LOG_MARK_FLAG (1 << 3)
58 #define WRITE_LOG_VERSION 1ULL
59 #define WRITE_LOG_MAGIC 0x6a736677736872ULL
62 * The disk format for this is braindead simple.
64 * At byte 0 we have our super, followed by the following sequence for
67 * [ 1 sector ][ entry->nr_sectors ]
68 * [log_write_entry][ data written ]
70 * The log_write_entry takes up a full sector so we can have arbitrary length
71 * marks and it leaves us room for extra content in the future.
75 * Basic info about the log for userspace.
77 struct log_write_super {
85 * sector - the sector we wrote.
86 * nr_sectors - the number of sectors we wrote.
87 * flags - flags for this log entry.
88 * data_len - the size of the data in this log entry, this is for private log
89 * entry stuff, the MARK data provided by userspace for example.
91 struct log_write_entry {
100 struct dm_dev *logdev;
104 atomic_t pending_blocks;
105 sector_t next_sector;
107 bool logging_enabled;
108 bool device_supports_discard;
109 spinlock_t blocks_lock;
110 struct list_head unflushed_blocks;
111 struct list_head logging_blocks;
112 wait_queue_head_t wait;
113 struct task_struct *log_kthread;
116 struct pending_block {
123 struct list_head list;
124 struct bio_vec vecs[0];
127 struct per_bio_data {
128 struct pending_block *block;
131 static void put_pending_block(struct log_writes_c *lc)
133 if (atomic_dec_and_test(&lc->pending_blocks)) {
134 smp_mb__after_atomic();
135 if (waitqueue_active(&lc->wait))
140 static void put_io_block(struct log_writes_c *lc)
142 if (atomic_dec_and_test(&lc->io_blocks)) {
143 smp_mb__after_atomic();
144 if (waitqueue_active(&lc->wait))
149 static void log_end_io(struct bio *bio)
151 struct log_writes_c *lc = bio->bi_private;
152 struct bio_vec *bvec;
158 DMERR("Error writing log block, error=%d", bio->bi_error);
159 spin_lock_irqsave(&lc->blocks_lock, flags);
160 lc->logging_enabled = false;
161 spin_unlock_irqrestore(&lc->blocks_lock, flags);
164 bio_for_each_segment_all(bvec, bio, i)
165 __free_page(bvec->bv_page);
172 * Meant to be called if there is an error, it will free all the pages
173 * associated with the block.
175 static void free_pending_block(struct log_writes_c *lc,
176 struct pending_block *block)
180 for (i = 0; i < block->vec_cnt; i++) {
181 if (block->vecs[i].bv_page)
182 __free_page(block->vecs[i].bv_page);
186 put_pending_block(lc);
189 static int write_metadata(struct log_writes_c *lc, void *entry,
190 size_t entrylen, void *data, size_t datalen,
198 bio = bio_alloc(GFP_KERNEL, 1);
200 DMERR("Couldn't alloc log bio");
203 bio->bi_iter.bi_size = 0;
204 bio->bi_iter.bi_sector = sector;
205 bio->bi_bdev = lc->logdev->bdev;
206 bio->bi_end_io = log_end_io;
207 bio->bi_private = lc;
208 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
210 page = alloc_page(GFP_KERNEL);
212 DMERR("Couldn't alloc log page");
217 ptr = kmap_atomic(page);
218 memcpy(ptr, entry, entrylen);
220 memcpy(ptr + entrylen, data, datalen);
221 memset(ptr + entrylen + datalen, 0,
222 lc->sectorsize - entrylen - datalen);
225 ret = bio_add_page(bio, page, lc->sectorsize, 0);
226 if (ret != lc->sectorsize) {
227 DMERR("Couldn't add page to the log block");
240 static int log_one_block(struct log_writes_c *lc,
241 struct pending_block *block, sector_t sector)
244 struct log_write_entry entry;
248 entry.sector = cpu_to_le64(block->sector);
249 entry.nr_sectors = cpu_to_le64(block->nr_sectors);
250 entry.flags = cpu_to_le64(block->flags);
251 entry.data_len = cpu_to_le64(block->datalen);
252 if (write_metadata(lc, &entry, sizeof(entry), block->data,
253 block->datalen, sector)) {
254 free_pending_block(lc, block);
262 bio = bio_alloc(GFP_KERNEL, block->vec_cnt);
264 DMERR("Couldn't alloc log bio");
267 atomic_inc(&lc->io_blocks);
268 bio->bi_iter.bi_size = 0;
269 bio->bi_iter.bi_sector = sector;
270 bio->bi_bdev = lc->logdev->bdev;
271 bio->bi_end_io = log_end_io;
272 bio->bi_private = lc;
273 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
275 for (i = 0; i < block->vec_cnt; i++) {
277 * The page offset is always 0 because we allocate a new page
278 * for every bvec in the original bio for simplicity sake.
280 ret = bio_add_page(bio, block->vecs[i].bv_page,
281 block->vecs[i].bv_len, 0);
282 if (ret != block->vecs[i].bv_len) {
283 atomic_inc(&lc->io_blocks);
285 bio = bio_alloc(GFP_KERNEL, block->vec_cnt - i);
287 DMERR("Couldn't alloc log bio");
290 bio->bi_iter.bi_size = 0;
291 bio->bi_iter.bi_sector = sector;
292 bio->bi_bdev = lc->logdev->bdev;
293 bio->bi_end_io = log_end_io;
294 bio->bi_private = lc;
295 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
297 ret = bio_add_page(bio, block->vecs[i].bv_page,
298 block->vecs[i].bv_len, 0);
299 if (ret != block->vecs[i].bv_len) {
300 DMERR("Couldn't add page on new bio?");
305 sector += block->vecs[i].bv_len >> SECTOR_SHIFT;
311 put_pending_block(lc);
314 free_pending_block(lc, block);
319 static int log_super(struct log_writes_c *lc)
321 struct log_write_super super;
323 super.magic = cpu_to_le64(WRITE_LOG_MAGIC);
324 super.version = cpu_to_le64(WRITE_LOG_VERSION);
325 super.nr_entries = cpu_to_le64(lc->logged_entries);
326 super.sectorsize = cpu_to_le32(lc->sectorsize);
328 if (write_metadata(lc, &super, sizeof(super), NULL, 0, 0)) {
329 DMERR("Couldn't write super");
336 static inline sector_t logdev_last_sector(struct log_writes_c *lc)
338 return i_size_read(lc->logdev->bdev->bd_inode) >> SECTOR_SHIFT;
341 static int log_writes_kthread(void *arg)
343 struct log_writes_c *lc = (struct log_writes_c *)arg;
346 while (!kthread_should_stop()) {
348 bool logging_enabled;
349 struct pending_block *block = NULL;
352 spin_lock_irq(&lc->blocks_lock);
353 if (!list_empty(&lc->logging_blocks)) {
354 block = list_first_entry(&lc->logging_blocks,
355 struct pending_block, list);
356 list_del_init(&block->list);
357 if (!lc->logging_enabled)
360 sector = lc->next_sector;
361 if (block->flags & LOG_DISCARD_FLAG)
364 lc->next_sector += block->nr_sectors + 1;
367 * Apparently the size of the device may not be known
368 * right away, so handle this properly.
371 lc->end_sector = logdev_last_sector(lc);
372 if (lc->end_sector &&
373 lc->next_sector >= lc->end_sector) {
374 DMERR("Ran out of space on the logdev");
375 lc->logging_enabled = false;
378 lc->logged_entries++;
379 atomic_inc(&lc->io_blocks);
381 super = (block->flags & (LOG_FUA_FLAG | LOG_MARK_FLAG));
383 atomic_inc(&lc->io_blocks);
386 logging_enabled = lc->logging_enabled;
387 spin_unlock_irq(&lc->blocks_lock);
389 if (logging_enabled) {
390 ret = log_one_block(lc, block, sector);
394 spin_lock_irq(&lc->blocks_lock);
395 lc->logging_enabled = false;
396 spin_unlock_irq(&lc->blocks_lock);
399 free_pending_block(lc, block);
403 if (!try_to_freeze()) {
404 set_current_state(TASK_INTERRUPTIBLE);
405 if (!kthread_should_stop() &&
406 !atomic_read(&lc->pending_blocks))
408 __set_current_state(TASK_RUNNING);
415 * Construct a log-writes mapping:
416 * log-writes <dev_path> <log_dev_path>
418 static int log_writes_ctr(struct dm_target *ti, unsigned int argc, char **argv)
420 struct log_writes_c *lc;
421 struct dm_arg_set as;
422 const char *devname, *logdevname;
429 ti->error = "Invalid argument count";
433 lc = kzalloc(sizeof(struct log_writes_c), GFP_KERNEL);
435 ti->error = "Cannot allocate context";
438 spin_lock_init(&lc->blocks_lock);
439 INIT_LIST_HEAD(&lc->unflushed_blocks);
440 INIT_LIST_HEAD(&lc->logging_blocks);
441 init_waitqueue_head(&lc->wait);
442 lc->sectorsize = 1 << SECTOR_SHIFT;
443 atomic_set(&lc->io_blocks, 0);
444 atomic_set(&lc->pending_blocks, 0);
446 devname = dm_shift_arg(&as);
447 ret = dm_get_device(ti, devname, dm_table_get_mode(ti->table), &lc->dev);
449 ti->error = "Device lookup failed";
453 logdevname = dm_shift_arg(&as);
454 ret = dm_get_device(ti, logdevname, dm_table_get_mode(ti->table),
457 ti->error = "Log device lookup failed";
458 dm_put_device(ti, lc->dev);
463 lc->log_kthread = kthread_run(log_writes_kthread, lc, "log-write");
464 if (!lc->log_kthread) {
465 ti->error = "Couldn't alloc kthread";
466 dm_put_device(ti, lc->dev);
467 dm_put_device(ti, lc->logdev);
471 /* We put the super at sector 0, start logging at sector 1 */
473 lc->logging_enabled = true;
474 lc->end_sector = logdev_last_sector(lc);
475 lc->device_supports_discard = true;
477 ti->num_flush_bios = 1;
478 ti->flush_supported = true;
479 ti->num_discard_bios = 1;
480 ti->discards_supported = true;
481 ti->per_io_data_size = sizeof(struct per_bio_data);
490 static int log_mark(struct log_writes_c *lc, char *data)
492 struct pending_block *block;
493 size_t maxsize = lc->sectorsize - sizeof(struct log_write_entry);
495 block = kzalloc(sizeof(struct pending_block), GFP_KERNEL);
497 DMERR("Error allocating pending block");
501 block->data = kstrndup(data, maxsize, GFP_KERNEL);
503 DMERR("Error copying mark data");
507 atomic_inc(&lc->pending_blocks);
508 block->datalen = strlen(block->data);
509 block->flags |= LOG_MARK_FLAG;
510 spin_lock_irq(&lc->blocks_lock);
511 list_add_tail(&block->list, &lc->logging_blocks);
512 spin_unlock_irq(&lc->blocks_lock);
513 wake_up_process(lc->log_kthread);
517 static void log_writes_dtr(struct dm_target *ti)
519 struct log_writes_c *lc = ti->private;
521 spin_lock_irq(&lc->blocks_lock);
522 list_splice_init(&lc->unflushed_blocks, &lc->logging_blocks);
523 spin_unlock_irq(&lc->blocks_lock);
526 * This is just nice to have since it'll update the super to include the
527 * unflushed blocks, if it fails we don't really care.
529 log_mark(lc, "dm-log-writes-end");
530 wake_up_process(lc->log_kthread);
531 wait_event(lc->wait, !atomic_read(&lc->io_blocks) &&
532 !atomic_read(&lc->pending_blocks));
533 kthread_stop(lc->log_kthread);
535 WARN_ON(!list_empty(&lc->logging_blocks));
536 WARN_ON(!list_empty(&lc->unflushed_blocks));
537 dm_put_device(ti, lc->dev);
538 dm_put_device(ti, lc->logdev);
542 static void normal_map_bio(struct dm_target *ti, struct bio *bio)
544 struct log_writes_c *lc = ti->private;
546 bio->bi_bdev = lc->dev->bdev;
549 static int log_writes_map(struct dm_target *ti, struct bio *bio)
551 struct log_writes_c *lc = ti->private;
552 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
553 struct pending_block *block;
554 struct bvec_iter iter;
558 bool flush_bio = (bio->bi_opf & REQ_PREFLUSH);
559 bool fua_bio = (bio->bi_opf & REQ_FUA);
560 bool discard_bio = (bio_op(bio) == REQ_OP_DISCARD);
564 /* Don't bother doing anything if logging has been disabled */
565 if (!lc->logging_enabled)
569 * Map reads as normal.
571 if (bio_data_dir(bio) == READ)
574 /* No sectors and not a flush? Don't care */
575 if (!bio_sectors(bio) && !flush_bio)
579 * Discards will have bi_size set but there's no actual data, so just
580 * allocate the size of the pending block.
583 alloc_size = sizeof(struct pending_block);
585 alloc_size = sizeof(struct pending_block) + sizeof(struct bio_vec) * bio_segments(bio);
587 block = kzalloc(alloc_size, GFP_NOIO);
589 DMERR("Error allocating pending block");
590 spin_lock_irq(&lc->blocks_lock);
591 lc->logging_enabled = false;
592 spin_unlock_irq(&lc->blocks_lock);
595 INIT_LIST_HEAD(&block->list);
597 atomic_inc(&lc->pending_blocks);
600 block->flags |= LOG_FLUSH_FLAG;
602 block->flags |= LOG_FUA_FLAG;
604 block->flags |= LOG_DISCARD_FLAG;
606 block->sector = bio->bi_iter.bi_sector;
607 block->nr_sectors = bio_sectors(bio);
609 /* We don't need the data, just submit */
611 WARN_ON(flush_bio || fua_bio);
612 if (lc->device_supports_discard)
615 return DM_MAPIO_SUBMITTED;
618 /* Flush bio, splice the unflushed blocks onto this list and submit */
619 if (flush_bio && !bio_sectors(bio)) {
620 spin_lock_irq(&lc->blocks_lock);
621 list_splice_init(&lc->unflushed_blocks, &block->list);
622 spin_unlock_irq(&lc->blocks_lock);
627 * We will write this bio somewhere else way later so we need to copy
628 * the actual contents into new pages so we know the data will always be
631 * We do this because this could be a bio from O_DIRECT in which case we
632 * can't just hold onto the page until some later point, we have to
633 * manually copy the contents.
635 bio_for_each_segment(bv, bio, iter) {
639 page = alloc_page(GFP_NOIO);
641 DMERR("Error allocing page");
642 free_pending_block(lc, block);
643 spin_lock_irq(&lc->blocks_lock);
644 lc->logging_enabled = false;
645 spin_unlock_irq(&lc->blocks_lock);
649 src = kmap_atomic(bv.bv_page);
650 dst = kmap_atomic(page);
651 memcpy(dst, src + bv.bv_offset, bv.bv_len);
654 block->vecs[i].bv_page = page;
655 block->vecs[i].bv_len = bv.bv_len;
660 /* Had a flush with data in it, weird */
662 spin_lock_irq(&lc->blocks_lock);
663 list_splice_init(&lc->unflushed_blocks, &block->list);
664 spin_unlock_irq(&lc->blocks_lock);
667 normal_map_bio(ti, bio);
668 return DM_MAPIO_REMAPPED;
671 static int normal_end_io(struct dm_target *ti, struct bio *bio, int error)
673 struct log_writes_c *lc = ti->private;
674 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
676 if (bio_data_dir(bio) == WRITE && pb->block) {
677 struct pending_block *block = pb->block;
680 spin_lock_irqsave(&lc->blocks_lock, flags);
681 if (block->flags & LOG_FLUSH_FLAG) {
682 list_splice_tail_init(&block->list, &lc->logging_blocks);
683 list_add_tail(&block->list, &lc->logging_blocks);
684 wake_up_process(lc->log_kthread);
685 } else if (block->flags & LOG_FUA_FLAG) {
686 list_add_tail(&block->list, &lc->logging_blocks);
687 wake_up_process(lc->log_kthread);
689 list_add_tail(&block->list, &lc->unflushed_blocks);
690 spin_unlock_irqrestore(&lc->blocks_lock, flags);
697 * INFO format: <logged entries> <highest allocated sector>
699 static void log_writes_status(struct dm_target *ti, status_type_t type,
700 unsigned status_flags, char *result,
704 struct log_writes_c *lc = ti->private;
707 case STATUSTYPE_INFO:
708 DMEMIT("%llu %llu", lc->logged_entries,
709 (unsigned long long)lc->next_sector - 1);
710 if (!lc->logging_enabled)
711 DMEMIT(" logging_disabled");
714 case STATUSTYPE_TABLE:
715 DMEMIT("%s %s", lc->dev->name, lc->logdev->name);
720 static int log_writes_prepare_ioctl(struct dm_target *ti,
721 struct block_device **bdev, fmode_t *mode)
723 struct log_writes_c *lc = ti->private;
724 struct dm_dev *dev = lc->dev;
728 * Only pass ioctls through if the device sizes match exactly.
730 if (ti->len != i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT)
735 static int log_writes_iterate_devices(struct dm_target *ti,
736 iterate_devices_callout_fn fn,
739 struct log_writes_c *lc = ti->private;
741 return fn(ti, lc->dev, 0, ti->len, data);
745 * Messages supported:
746 * mark <mark data> - specify the marked data.
748 static int log_writes_message(struct dm_target *ti, unsigned argc, char **argv)
751 struct log_writes_c *lc = ti->private;
754 DMWARN("Invalid log-writes message arguments, expect 2 arguments, got %d", argc);
758 if (!strcasecmp(argv[0], "mark"))
759 r = log_mark(lc, argv[1]);
761 DMWARN("Unrecognised log writes target message received: %s", argv[0]);
766 static void log_writes_io_hints(struct dm_target *ti, struct queue_limits *limits)
768 struct log_writes_c *lc = ti->private;
769 struct request_queue *q = bdev_get_queue(lc->dev->bdev);
771 if (!q || !blk_queue_discard(q)) {
772 lc->device_supports_discard = false;
773 limits->discard_granularity = 1 << SECTOR_SHIFT;
774 limits->max_discard_sectors = (UINT_MAX >> SECTOR_SHIFT);
778 static struct target_type log_writes_target = {
779 .name = "log-writes",
780 .version = {1, 0, 0},
781 .module = THIS_MODULE,
782 .ctr = log_writes_ctr,
783 .dtr = log_writes_dtr,
784 .map = log_writes_map,
785 .end_io = normal_end_io,
786 .status = log_writes_status,
787 .prepare_ioctl = log_writes_prepare_ioctl,
788 .message = log_writes_message,
789 .iterate_devices = log_writes_iterate_devices,
790 .io_hints = log_writes_io_hints,
793 static int __init dm_log_writes_init(void)
795 int r = dm_register_target(&log_writes_target);
798 DMERR("register failed %d", r);
803 static void __exit dm_log_writes_exit(void)
805 dm_unregister_target(&log_writes_target);
808 module_init(dm_log_writes_init);
809 module_exit(dm_log_writes_exit);
811 MODULE_DESCRIPTION(DM_NAME " log writes target");
812 MODULE_AUTHOR("Josef Bacik <jbacik@fb.com>");
813 MODULE_LICENSE("GPL");