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;
105 atomic_t pending_blocks;
106 sector_t next_sector;
108 bool logging_enabled;
109 bool device_supports_discard;
110 spinlock_t blocks_lock;
111 struct list_head unflushed_blocks;
112 struct list_head logging_blocks;
113 wait_queue_head_t wait;
114 struct task_struct *log_kthread;
117 struct pending_block {
124 struct list_head list;
125 struct bio_vec vecs[0];
128 struct per_bio_data {
129 struct pending_block *block;
132 static inline sector_t bio_to_dev_sectors(struct log_writes_c *lc,
135 return sectors >> (lc->sectorshift - SECTOR_SHIFT);
138 static inline sector_t dev_to_bio_sectors(struct log_writes_c *lc,
141 return sectors << (lc->sectorshift - SECTOR_SHIFT);
144 static void put_pending_block(struct log_writes_c *lc)
146 if (atomic_dec_and_test(&lc->pending_blocks)) {
147 smp_mb__after_atomic();
148 if (waitqueue_active(&lc->wait))
153 static void put_io_block(struct log_writes_c *lc)
155 if (atomic_dec_and_test(&lc->io_blocks)) {
156 smp_mb__after_atomic();
157 if (waitqueue_active(&lc->wait))
162 static void log_end_io(struct bio *bio)
164 struct log_writes_c *lc = bio->bi_private;
166 if (bio->bi_status) {
169 DMERR("Error writing log block, error=%d", bio->bi_status);
170 spin_lock_irqsave(&lc->blocks_lock, flags);
171 lc->logging_enabled = false;
172 spin_unlock_irqrestore(&lc->blocks_lock, flags);
181 * Meant to be called if there is an error, it will free all the pages
182 * associated with the block.
184 static void free_pending_block(struct log_writes_c *lc,
185 struct pending_block *block)
189 for (i = 0; i < block->vec_cnt; i++) {
190 if (block->vecs[i].bv_page)
191 __free_page(block->vecs[i].bv_page);
195 put_pending_block(lc);
198 static int write_metadata(struct log_writes_c *lc, void *entry,
199 size_t entrylen, void *data, size_t datalen,
207 bio = bio_alloc(GFP_KERNEL, 1);
209 DMERR("Couldn't alloc log bio");
212 bio->bi_iter.bi_size = 0;
213 bio->bi_iter.bi_sector = sector;
214 bio_set_dev(bio, lc->logdev->bdev);
215 bio->bi_end_io = log_end_io;
216 bio->bi_private = lc;
217 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
219 page = alloc_page(GFP_KERNEL);
221 DMERR("Couldn't alloc log page");
226 ptr = kmap_atomic(page);
227 memcpy(ptr, entry, entrylen);
229 memcpy(ptr + entrylen, data, datalen);
230 memset(ptr + entrylen + datalen, 0,
231 lc->sectorsize - entrylen - datalen);
234 ret = bio_add_page(bio, page, lc->sectorsize, 0);
235 if (ret != lc->sectorsize) {
236 DMERR("Couldn't add page to the log block");
249 static int log_one_block(struct log_writes_c *lc,
250 struct pending_block *block, sector_t sector)
253 struct log_write_entry entry;
257 entry.sector = cpu_to_le64(block->sector);
258 entry.nr_sectors = cpu_to_le64(block->nr_sectors);
259 entry.flags = cpu_to_le64(block->flags);
260 entry.data_len = cpu_to_le64(block->datalen);
261 if (write_metadata(lc, &entry, sizeof(entry), block->data,
262 block->datalen, sector)) {
263 free_pending_block(lc, block);
269 sector += dev_to_bio_sectors(lc, 1);
271 atomic_inc(&lc->io_blocks);
272 bio = bio_alloc(GFP_KERNEL, min(block->vec_cnt, BIO_MAX_PAGES));
274 DMERR("Couldn't alloc log bio");
277 bio->bi_iter.bi_size = 0;
278 bio->bi_iter.bi_sector = sector;
279 bio_set_dev(bio, lc->logdev->bdev);
280 bio->bi_end_io = log_end_io;
281 bio->bi_private = lc;
282 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
284 for (i = 0; i < block->vec_cnt; i++) {
286 * The page offset is always 0 because we allocate a new page
287 * for every bvec in the original bio for simplicity sake.
289 ret = bio_add_page(bio, block->vecs[i].bv_page,
290 block->vecs[i].bv_len, 0);
291 if (ret != block->vecs[i].bv_len) {
292 atomic_inc(&lc->io_blocks);
294 bio = bio_alloc(GFP_KERNEL, min(block->vec_cnt - i, BIO_MAX_PAGES));
296 DMERR("Couldn't alloc log bio");
299 bio->bi_iter.bi_size = 0;
300 bio->bi_iter.bi_sector = sector;
301 bio_set_dev(bio, lc->logdev->bdev);
302 bio->bi_end_io = log_end_io;
303 bio->bi_private = lc;
304 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
306 ret = bio_add_page(bio, block->vecs[i].bv_page,
307 block->vecs[i].bv_len, 0);
308 if (ret != block->vecs[i].bv_len) {
309 DMERR("Couldn't add page on new bio?");
314 sector += block->vecs[i].bv_len >> SECTOR_SHIFT;
320 put_pending_block(lc);
323 free_pending_block(lc, block);
328 static int log_super(struct log_writes_c *lc)
330 struct log_write_super super;
332 super.magic = cpu_to_le64(WRITE_LOG_MAGIC);
333 super.version = cpu_to_le64(WRITE_LOG_VERSION);
334 super.nr_entries = cpu_to_le64(lc->logged_entries);
335 super.sectorsize = cpu_to_le32(lc->sectorsize);
337 if (write_metadata(lc, &super, sizeof(super), NULL, 0, 0)) {
338 DMERR("Couldn't write super");
345 static inline sector_t logdev_last_sector(struct log_writes_c *lc)
347 return i_size_read(lc->logdev->bdev->bd_inode) >> SECTOR_SHIFT;
350 static int log_writes_kthread(void *arg)
352 struct log_writes_c *lc = (struct log_writes_c *)arg;
355 while (!kthread_should_stop()) {
357 bool logging_enabled;
358 struct pending_block *block = NULL;
361 spin_lock_irq(&lc->blocks_lock);
362 if (!list_empty(&lc->logging_blocks)) {
363 block = list_first_entry(&lc->logging_blocks,
364 struct pending_block, list);
365 list_del_init(&block->list);
366 if (!lc->logging_enabled)
369 sector = lc->next_sector;
370 if (!(block->flags & LOG_DISCARD_FLAG))
371 lc->next_sector += dev_to_bio_sectors(lc, block->nr_sectors);
372 lc->next_sector += dev_to_bio_sectors(lc, 1);
375 * Apparently the size of the device may not be known
376 * right away, so handle this properly.
379 lc->end_sector = logdev_last_sector(lc);
380 if (lc->end_sector &&
381 lc->next_sector >= lc->end_sector) {
382 DMERR("Ran out of space on the logdev");
383 lc->logging_enabled = false;
386 lc->logged_entries++;
387 atomic_inc(&lc->io_blocks);
389 super = (block->flags & (LOG_FUA_FLAG | LOG_MARK_FLAG));
391 atomic_inc(&lc->io_blocks);
394 logging_enabled = lc->logging_enabled;
395 spin_unlock_irq(&lc->blocks_lock);
397 if (logging_enabled) {
398 ret = log_one_block(lc, block, sector);
402 spin_lock_irq(&lc->blocks_lock);
403 lc->logging_enabled = false;
404 spin_unlock_irq(&lc->blocks_lock);
407 free_pending_block(lc, block);
411 if (!try_to_freeze()) {
412 set_current_state(TASK_INTERRUPTIBLE);
413 if (!kthread_should_stop() &&
414 list_empty(&lc->logging_blocks))
416 __set_current_state(TASK_RUNNING);
423 * Construct a log-writes mapping:
424 * log-writes <dev_path> <log_dev_path>
426 static int log_writes_ctr(struct dm_target *ti, unsigned int argc, char **argv)
428 struct log_writes_c *lc;
429 struct dm_arg_set as;
430 const char *devname, *logdevname;
437 ti->error = "Invalid argument count";
441 lc = kzalloc(sizeof(struct log_writes_c), GFP_KERNEL);
443 ti->error = "Cannot allocate context";
446 spin_lock_init(&lc->blocks_lock);
447 INIT_LIST_HEAD(&lc->unflushed_blocks);
448 INIT_LIST_HEAD(&lc->logging_blocks);
449 init_waitqueue_head(&lc->wait);
450 atomic_set(&lc->io_blocks, 0);
451 atomic_set(&lc->pending_blocks, 0);
453 devname = dm_shift_arg(&as);
454 ret = dm_get_device(ti, devname, dm_table_get_mode(ti->table), &lc->dev);
456 ti->error = "Device lookup failed";
460 logdevname = dm_shift_arg(&as);
461 ret = dm_get_device(ti, logdevname, dm_table_get_mode(ti->table),
464 ti->error = "Log device lookup failed";
465 dm_put_device(ti, lc->dev);
469 lc->sectorsize = bdev_logical_block_size(lc->dev->bdev);
470 lc->sectorshift = ilog2(lc->sectorsize);
471 lc->log_kthread = kthread_run(log_writes_kthread, lc, "log-write");
472 if (IS_ERR(lc->log_kthread)) {
473 ret = PTR_ERR(lc->log_kthread);
474 ti->error = "Couldn't alloc kthread";
475 dm_put_device(ti, lc->dev);
476 dm_put_device(ti, lc->logdev);
481 * next_sector is in 512b sectors to correspond to what bi_sector expects.
482 * The super starts at sector 0, and the next_sector is the next logical
483 * one based on the sectorsize of the device.
485 lc->next_sector = lc->sectorsize >> SECTOR_SHIFT;
486 lc->logging_enabled = true;
487 lc->end_sector = logdev_last_sector(lc);
488 lc->device_supports_discard = true;
490 ti->num_flush_bios = 1;
491 ti->flush_supported = true;
492 ti->num_discard_bios = 1;
493 ti->discards_supported = true;
494 ti->per_io_data_size = sizeof(struct per_bio_data);
503 static int log_mark(struct log_writes_c *lc, char *data)
505 struct pending_block *block;
506 size_t maxsize = lc->sectorsize - sizeof(struct log_write_entry);
508 block = kzalloc(sizeof(struct pending_block), GFP_KERNEL);
510 DMERR("Error allocating pending block");
514 block->data = kstrndup(data, maxsize, GFP_KERNEL);
516 DMERR("Error copying mark data");
520 atomic_inc(&lc->pending_blocks);
521 block->datalen = strlen(block->data);
522 block->flags |= LOG_MARK_FLAG;
523 spin_lock_irq(&lc->blocks_lock);
524 list_add_tail(&block->list, &lc->logging_blocks);
525 spin_unlock_irq(&lc->blocks_lock);
526 wake_up_process(lc->log_kthread);
530 static void log_writes_dtr(struct dm_target *ti)
532 struct log_writes_c *lc = ti->private;
534 spin_lock_irq(&lc->blocks_lock);
535 list_splice_init(&lc->unflushed_blocks, &lc->logging_blocks);
536 spin_unlock_irq(&lc->blocks_lock);
539 * This is just nice to have since it'll update the super to include the
540 * unflushed blocks, if it fails we don't really care.
542 log_mark(lc, "dm-log-writes-end");
543 wake_up_process(lc->log_kthread);
544 wait_event(lc->wait, !atomic_read(&lc->io_blocks) &&
545 !atomic_read(&lc->pending_blocks));
546 kthread_stop(lc->log_kthread);
548 WARN_ON(!list_empty(&lc->logging_blocks));
549 WARN_ON(!list_empty(&lc->unflushed_blocks));
550 dm_put_device(ti, lc->dev);
551 dm_put_device(ti, lc->logdev);
555 static void normal_map_bio(struct dm_target *ti, struct bio *bio)
557 struct log_writes_c *lc = ti->private;
559 bio_set_dev(bio, lc->dev->bdev);
562 static int log_writes_map(struct dm_target *ti, struct bio *bio)
564 struct log_writes_c *lc = ti->private;
565 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
566 struct pending_block *block;
567 struct bvec_iter iter;
571 bool flush_bio = (bio->bi_opf & REQ_PREFLUSH);
572 bool fua_bio = (bio->bi_opf & REQ_FUA);
573 bool discard_bio = (bio_op(bio) == REQ_OP_DISCARD);
577 /* Don't bother doing anything if logging has been disabled */
578 if (!lc->logging_enabled)
582 * Map reads as normal.
584 if (bio_data_dir(bio) == READ)
587 /* No sectors and not a flush? Don't care */
588 if (!bio_sectors(bio) && !flush_bio)
592 * Discards will have bi_size set but there's no actual data, so just
593 * allocate the size of the pending block.
596 alloc_size = sizeof(struct pending_block);
598 alloc_size = sizeof(struct pending_block) + sizeof(struct bio_vec) * bio_segments(bio);
600 block = kzalloc(alloc_size, GFP_NOIO);
602 DMERR("Error allocating pending block");
603 spin_lock_irq(&lc->blocks_lock);
604 lc->logging_enabled = false;
605 spin_unlock_irq(&lc->blocks_lock);
606 return DM_MAPIO_KILL;
608 INIT_LIST_HEAD(&block->list);
610 atomic_inc(&lc->pending_blocks);
613 block->flags |= LOG_FLUSH_FLAG;
615 block->flags |= LOG_FUA_FLAG;
617 block->flags |= LOG_DISCARD_FLAG;
619 block->sector = bio_to_dev_sectors(lc, bio->bi_iter.bi_sector);
620 block->nr_sectors = bio_to_dev_sectors(lc, bio_sectors(bio));
622 /* We don't need the data, just submit */
624 WARN_ON(flush_bio || fua_bio);
625 if (lc->device_supports_discard)
628 return DM_MAPIO_SUBMITTED;
631 /* Flush bio, splice the unflushed blocks onto this list and submit */
632 if (flush_bio && !bio_sectors(bio)) {
633 spin_lock_irq(&lc->blocks_lock);
634 list_splice_init(&lc->unflushed_blocks, &block->list);
635 spin_unlock_irq(&lc->blocks_lock);
640 * We will write this bio somewhere else way later so we need to copy
641 * the actual contents into new pages so we know the data will always be
644 * We do this because this could be a bio from O_DIRECT in which case we
645 * can't just hold onto the page until some later point, we have to
646 * manually copy the contents.
648 bio_for_each_segment(bv, bio, iter) {
652 page = alloc_page(GFP_NOIO);
654 DMERR("Error allocing page");
655 free_pending_block(lc, block);
656 spin_lock_irq(&lc->blocks_lock);
657 lc->logging_enabled = false;
658 spin_unlock_irq(&lc->blocks_lock);
659 return DM_MAPIO_KILL;
662 src = kmap_atomic(bv.bv_page);
663 dst = kmap_atomic(page);
664 memcpy(dst, src + bv.bv_offset, bv.bv_len);
667 block->vecs[i].bv_page = page;
668 block->vecs[i].bv_len = bv.bv_len;
673 /* Had a flush with data in it, weird */
675 spin_lock_irq(&lc->blocks_lock);
676 list_splice_init(&lc->unflushed_blocks, &block->list);
677 spin_unlock_irq(&lc->blocks_lock);
680 normal_map_bio(ti, bio);
681 return DM_MAPIO_REMAPPED;
684 static int normal_end_io(struct dm_target *ti, struct bio *bio,
687 struct log_writes_c *lc = ti->private;
688 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
690 if (bio_data_dir(bio) == WRITE && pb->block) {
691 struct pending_block *block = pb->block;
694 spin_lock_irqsave(&lc->blocks_lock, flags);
695 if (block->flags & LOG_FLUSH_FLAG) {
696 list_splice_tail_init(&block->list, &lc->logging_blocks);
697 list_add_tail(&block->list, &lc->logging_blocks);
698 wake_up_process(lc->log_kthread);
699 } else if (block->flags & LOG_FUA_FLAG) {
700 list_add_tail(&block->list, &lc->logging_blocks);
701 wake_up_process(lc->log_kthread);
703 list_add_tail(&block->list, &lc->unflushed_blocks);
704 spin_unlock_irqrestore(&lc->blocks_lock, flags);
707 return DM_ENDIO_DONE;
711 * INFO format: <logged entries> <highest allocated sector>
713 static void log_writes_status(struct dm_target *ti, status_type_t type,
714 unsigned status_flags, char *result,
718 struct log_writes_c *lc = ti->private;
721 case STATUSTYPE_INFO:
722 DMEMIT("%llu %llu", lc->logged_entries,
723 (unsigned long long)lc->next_sector - 1);
724 if (!lc->logging_enabled)
725 DMEMIT(" logging_disabled");
728 case STATUSTYPE_TABLE:
729 DMEMIT("%s %s", lc->dev->name, lc->logdev->name);
734 static int log_writes_prepare_ioctl(struct dm_target *ti,
735 struct block_device **bdev, fmode_t *mode)
737 struct log_writes_c *lc = ti->private;
738 struct dm_dev *dev = lc->dev;
742 * Only pass ioctls through if the device sizes match exactly.
744 if (ti->len != i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT)
749 static int log_writes_iterate_devices(struct dm_target *ti,
750 iterate_devices_callout_fn fn,
753 struct log_writes_c *lc = ti->private;
755 return fn(ti, lc->dev, 0, ti->len, data);
759 * Messages supported:
760 * mark <mark data> - specify the marked data.
762 static int log_writes_message(struct dm_target *ti, unsigned argc, char **argv)
765 struct log_writes_c *lc = ti->private;
768 DMWARN("Invalid log-writes message arguments, expect 2 arguments, got %d", argc);
772 if (!strcasecmp(argv[0], "mark"))
773 r = log_mark(lc, argv[1]);
775 DMWARN("Unrecognised log writes target message received: %s", argv[0]);
780 static void log_writes_io_hints(struct dm_target *ti, struct queue_limits *limits)
782 struct log_writes_c *lc = ti->private;
783 struct request_queue *q = bdev_get_queue(lc->dev->bdev);
785 if (!q || !blk_queue_discard(q)) {
786 lc->device_supports_discard = false;
787 limits->discard_granularity = lc->sectorsize;
788 limits->max_discard_sectors = (UINT_MAX >> SECTOR_SHIFT);
790 limits->logical_block_size = bdev_logical_block_size(lc->dev->bdev);
791 limits->physical_block_size = bdev_physical_block_size(lc->dev->bdev);
792 limits->io_min = limits->physical_block_size;
795 static struct target_type log_writes_target = {
796 .name = "log-writes",
797 .version = {1, 0, 0},
798 .module = THIS_MODULE,
799 .ctr = log_writes_ctr,
800 .dtr = log_writes_dtr,
801 .map = log_writes_map,
802 .end_io = normal_end_io,
803 .status = log_writes_status,
804 .prepare_ioctl = log_writes_prepare_ioctl,
805 .message = log_writes_message,
806 .iterate_devices = log_writes_iterate_devices,
807 .io_hints = log_writes_io_hints,
810 static int __init dm_log_writes_init(void)
812 int r = dm_register_target(&log_writes_target);
815 DMERR("register failed %d", r);
820 static void __exit dm_log_writes_exit(void)
822 dm_unregister_target(&log_writes_target);
825 module_init(dm_log_writes_init);
826 module_exit(dm_log_writes_exit);
828 MODULE_DESCRIPTION(DM_NAME " log writes target");
829 MODULE_AUTHOR("Josef Bacik <jbacik@fb.com>");
830 MODULE_LICENSE("GPL");