Loading...
1/*
2 * Copyright (C) 2014 Facebook. All rights reserved.
3 *
4 * This file is released under the GPL.
5 */
6
7#include <linux/device-mapper.h>
8
9#include <linux/module.h>
10#include <linux/init.h>
11#include <linux/blkdev.h>
12#include <linux/bio.h>
13#include <linux/dax.h>
14#include <linux/slab.h>
15#include <linux/kthread.h>
16#include <linux/freezer.h>
17#include <linux/uio.h>
18
19#define DM_MSG_PREFIX "log-writes"
20
21/*
22 * This target will sequentially log all writes to the target device onto the
23 * log device. This is helpful for replaying writes to check for fs consistency
24 * at all times. This target provides a mechanism to mark specific events to
25 * check data at a later time. So for example you would:
26 *
27 * write data
28 * fsync
29 * dmsetup message /dev/whatever mark mymark
30 * unmount /mnt/test
31 *
32 * Then replay the log up to mymark and check the contents of the replay to
33 * verify it matches what was written.
34 *
35 * We log writes only after they have been flushed, this makes the log describe
36 * close to the order in which the data hits the actual disk, not its cache. So
37 * for example the following sequence (W means write, C means complete)
38 *
39 * Wa,Wb,Wc,Cc,Ca,FLUSH,FUAd,Cb,CFLUSH,CFUAd
40 *
41 * Would result in the log looking like this:
42 *
43 * c,a,b,flush,fuad,<other writes>,<next flush>
44 *
45 * This is meant to help expose problems where file systems do not properly wait
46 * on data being written before invoking a FLUSH. FUA bypasses cache so once it
47 * completes it is added to the log as it should be on disk.
48 *
49 * We treat DISCARDs as if they don't bypass cache so that they are logged in
50 * order of completion along with the normal writes. If we didn't do it this
51 * way we would process all the discards first and then write all the data, when
52 * in fact we want to do the data and the discard in the order that they
53 * completed.
54 */
55#define LOG_FLUSH_FLAG (1 << 0)
56#define LOG_FUA_FLAG (1 << 1)
57#define LOG_DISCARD_FLAG (1 << 2)
58#define LOG_MARK_FLAG (1 << 3)
59#define LOG_METADATA_FLAG (1 << 4)
60
61#define WRITE_LOG_VERSION 1ULL
62#define WRITE_LOG_MAGIC 0x6a736677736872ULL
63#define WRITE_LOG_SUPER_SECTOR 0
64
65/*
66 * The disk format for this is braindead simple.
67 *
68 * At byte 0 we have our super, followed by the following sequence for
69 * nr_entries:
70 *
71 * [ 1 sector ][ entry->nr_sectors ]
72 * [log_write_entry][ data written ]
73 *
74 * The log_write_entry takes up a full sector so we can have arbitrary length
75 * marks and it leaves us room for extra content in the future.
76 */
77
78/*
79 * Basic info about the log for userspace.
80 */
81struct log_write_super {
82 __le64 magic;
83 __le64 version;
84 __le64 nr_entries;
85 __le32 sectorsize;
86};
87
88/*
89 * sector - the sector we wrote.
90 * nr_sectors - the number of sectors we wrote.
91 * flags - flags for this log entry.
92 * data_len - the size of the data in this log entry, this is for private log
93 * entry stuff, the MARK data provided by userspace for example.
94 */
95struct log_write_entry {
96 __le64 sector;
97 __le64 nr_sectors;
98 __le64 flags;
99 __le64 data_len;
100};
101
102struct log_writes_c {
103 struct dm_dev *dev;
104 struct dm_dev *logdev;
105 u64 logged_entries;
106 u32 sectorsize;
107 u32 sectorshift;
108 atomic_t io_blocks;
109 atomic_t pending_blocks;
110 sector_t next_sector;
111 sector_t end_sector;
112 bool logging_enabled;
113 bool device_supports_discard;
114 spinlock_t blocks_lock;
115 struct list_head unflushed_blocks;
116 struct list_head logging_blocks;
117 wait_queue_head_t wait;
118 struct task_struct *log_kthread;
119 struct completion super_done;
120};
121
122struct pending_block {
123 int vec_cnt;
124 u64 flags;
125 sector_t sector;
126 sector_t nr_sectors;
127 char *data;
128 u32 datalen;
129 struct list_head list;
130 struct bio_vec vecs[];
131};
132
133struct per_bio_data {
134 struct pending_block *block;
135};
136
137static inline sector_t bio_to_dev_sectors(struct log_writes_c *lc,
138 sector_t sectors)
139{
140 return sectors >> (lc->sectorshift - SECTOR_SHIFT);
141}
142
143static inline sector_t dev_to_bio_sectors(struct log_writes_c *lc,
144 sector_t sectors)
145{
146 return sectors << (lc->sectorshift - SECTOR_SHIFT);
147}
148
149static void put_pending_block(struct log_writes_c *lc)
150{
151 if (atomic_dec_and_test(&lc->pending_blocks)) {
152 smp_mb__after_atomic();
153 if (waitqueue_active(&lc->wait))
154 wake_up(&lc->wait);
155 }
156}
157
158static void put_io_block(struct log_writes_c *lc)
159{
160 if (atomic_dec_and_test(&lc->io_blocks)) {
161 smp_mb__after_atomic();
162 if (waitqueue_active(&lc->wait))
163 wake_up(&lc->wait);
164 }
165}
166
167static void log_end_io(struct bio *bio)
168{
169 struct log_writes_c *lc = bio->bi_private;
170
171 if (bio->bi_status) {
172 unsigned long flags;
173
174 DMERR("Error writing log block, error=%d", bio->bi_status);
175 spin_lock_irqsave(&lc->blocks_lock, flags);
176 lc->logging_enabled = false;
177 spin_unlock_irqrestore(&lc->blocks_lock, flags);
178 }
179
180 bio_free_pages(bio);
181 put_io_block(lc);
182 bio_put(bio);
183}
184
185static void log_end_super(struct bio *bio)
186{
187 struct log_writes_c *lc = bio->bi_private;
188
189 complete(&lc->super_done);
190 log_end_io(bio);
191}
192
193/*
194 * Meant to be called if there is an error, it will free all the pages
195 * associated with the block.
196 */
197static void free_pending_block(struct log_writes_c *lc,
198 struct pending_block *block)
199{
200 int i;
201
202 for (i = 0; i < block->vec_cnt; i++) {
203 if (block->vecs[i].bv_page)
204 __free_page(block->vecs[i].bv_page);
205 }
206 kfree(block->data);
207 kfree(block);
208 put_pending_block(lc);
209}
210
211static int write_metadata(struct log_writes_c *lc, void *entry,
212 size_t entrylen, void *data, size_t datalen,
213 sector_t sector)
214{
215 struct bio *bio;
216 struct page *page;
217 void *ptr;
218 size_t ret;
219
220 bio = bio_alloc(lc->logdev->bdev, 1, REQ_OP_WRITE, GFP_KERNEL);
221 bio->bi_iter.bi_size = 0;
222 bio->bi_iter.bi_sector = sector;
223 bio->bi_end_io = (sector == WRITE_LOG_SUPER_SECTOR) ?
224 log_end_super : log_end_io;
225 bio->bi_private = lc;
226
227 page = alloc_page(GFP_KERNEL);
228 if (!page) {
229 DMERR("Couldn't alloc log page");
230 bio_put(bio);
231 goto error;
232 }
233
234 ptr = kmap_atomic(page);
235 memcpy(ptr, entry, entrylen);
236 if (datalen)
237 memcpy(ptr + entrylen, data, datalen);
238 memset(ptr + entrylen + datalen, 0,
239 lc->sectorsize - entrylen - datalen);
240 kunmap_atomic(ptr);
241
242 ret = bio_add_page(bio, page, lc->sectorsize, 0);
243 if (ret != lc->sectorsize) {
244 DMERR("Couldn't add page to the log block");
245 goto error_bio;
246 }
247 submit_bio(bio);
248 return 0;
249error_bio:
250 bio_put(bio);
251 __free_page(page);
252error:
253 put_io_block(lc);
254 return -1;
255}
256
257static int write_inline_data(struct log_writes_c *lc, void *entry,
258 size_t entrylen, void *data, size_t datalen,
259 sector_t sector)
260{
261 int bio_pages, pg_datalen, pg_sectorlen, i;
262 struct page *page;
263 struct bio *bio;
264 size_t ret;
265 void *ptr;
266
267 while (datalen) {
268 bio_pages = bio_max_segs(DIV_ROUND_UP(datalen, PAGE_SIZE));
269
270 atomic_inc(&lc->io_blocks);
271
272 bio = bio_alloc(lc->logdev->bdev, bio_pages, REQ_OP_WRITE,
273 GFP_KERNEL);
274 bio->bi_iter.bi_size = 0;
275 bio->bi_iter.bi_sector = sector;
276 bio->bi_end_io = log_end_io;
277 bio->bi_private = lc;
278
279 for (i = 0; i < bio_pages; i++) {
280 pg_datalen = min_t(int, datalen, PAGE_SIZE);
281 pg_sectorlen = ALIGN(pg_datalen, lc->sectorsize);
282
283 page = alloc_page(GFP_KERNEL);
284 if (!page) {
285 DMERR("Couldn't alloc inline data page");
286 goto error_bio;
287 }
288
289 ptr = kmap_atomic(page);
290 memcpy(ptr, data, pg_datalen);
291 if (pg_sectorlen > pg_datalen)
292 memset(ptr + pg_datalen, 0, pg_sectorlen - pg_datalen);
293 kunmap_atomic(ptr);
294
295 ret = bio_add_page(bio, page, pg_sectorlen, 0);
296 if (ret != pg_sectorlen) {
297 DMERR("Couldn't add page of inline data");
298 __free_page(page);
299 goto error_bio;
300 }
301
302 datalen -= pg_datalen;
303 data += pg_datalen;
304 }
305 submit_bio(bio);
306
307 sector += bio_pages * PAGE_SECTORS;
308 }
309 return 0;
310error_bio:
311 bio_free_pages(bio);
312 bio_put(bio);
313 put_io_block(lc);
314 return -1;
315}
316
317static int log_one_block(struct log_writes_c *lc,
318 struct pending_block *block, sector_t sector)
319{
320 struct bio *bio;
321 struct log_write_entry entry;
322 size_t metadatalen, ret;
323 int i;
324
325 entry.sector = cpu_to_le64(block->sector);
326 entry.nr_sectors = cpu_to_le64(block->nr_sectors);
327 entry.flags = cpu_to_le64(block->flags);
328 entry.data_len = cpu_to_le64(block->datalen);
329
330 metadatalen = (block->flags & LOG_MARK_FLAG) ? block->datalen : 0;
331 if (write_metadata(lc, &entry, sizeof(entry), block->data,
332 metadatalen, sector)) {
333 free_pending_block(lc, block);
334 return -1;
335 }
336
337 sector += dev_to_bio_sectors(lc, 1);
338
339 if (block->datalen && metadatalen == 0) {
340 if (write_inline_data(lc, &entry, sizeof(entry), block->data,
341 block->datalen, sector)) {
342 free_pending_block(lc, block);
343 return -1;
344 }
345 /* we don't support both inline data & bio data */
346 goto out;
347 }
348
349 if (!block->vec_cnt)
350 goto out;
351
352 atomic_inc(&lc->io_blocks);
353 bio = bio_alloc(lc->logdev->bdev, bio_max_segs(block->vec_cnt),
354 REQ_OP_WRITE, GFP_KERNEL);
355 bio->bi_iter.bi_size = 0;
356 bio->bi_iter.bi_sector = sector;
357 bio->bi_end_io = log_end_io;
358 bio->bi_private = lc;
359
360 for (i = 0; i < block->vec_cnt; i++) {
361 /*
362 * The page offset is always 0 because we allocate a new page
363 * for every bvec in the original bio for simplicity sake.
364 */
365 ret = bio_add_page(bio, block->vecs[i].bv_page,
366 block->vecs[i].bv_len, 0);
367 if (ret != block->vecs[i].bv_len) {
368 atomic_inc(&lc->io_blocks);
369 submit_bio(bio);
370 bio = bio_alloc(lc->logdev->bdev,
371 bio_max_segs(block->vec_cnt - i),
372 REQ_OP_WRITE, GFP_KERNEL);
373 bio->bi_iter.bi_size = 0;
374 bio->bi_iter.bi_sector = sector;
375 bio->bi_end_io = log_end_io;
376 bio->bi_private = lc;
377
378 ret = bio_add_page(bio, block->vecs[i].bv_page,
379 block->vecs[i].bv_len, 0);
380 if (ret != block->vecs[i].bv_len) {
381 DMERR("Couldn't add page on new bio?");
382 bio_put(bio);
383 goto error;
384 }
385 }
386 sector += block->vecs[i].bv_len >> SECTOR_SHIFT;
387 }
388 submit_bio(bio);
389out:
390 kfree(block->data);
391 kfree(block);
392 put_pending_block(lc);
393 return 0;
394error:
395 free_pending_block(lc, block);
396 put_io_block(lc);
397 return -1;
398}
399
400static int log_super(struct log_writes_c *lc)
401{
402 struct log_write_super super;
403
404 super.magic = cpu_to_le64(WRITE_LOG_MAGIC);
405 super.version = cpu_to_le64(WRITE_LOG_VERSION);
406 super.nr_entries = cpu_to_le64(lc->logged_entries);
407 super.sectorsize = cpu_to_le32(lc->sectorsize);
408
409 if (write_metadata(lc, &super, sizeof(super), NULL, 0,
410 WRITE_LOG_SUPER_SECTOR)) {
411 DMERR("Couldn't write super");
412 return -1;
413 }
414
415 /*
416 * Super sector should be writen in-order, otherwise the
417 * nr_entries could be rewritten incorrectly by an old bio.
418 */
419 wait_for_completion_io(&lc->super_done);
420
421 return 0;
422}
423
424static inline sector_t logdev_last_sector(struct log_writes_c *lc)
425{
426 return bdev_nr_sectors(lc->logdev->bdev);
427}
428
429static int log_writes_kthread(void *arg)
430{
431 struct log_writes_c *lc = (struct log_writes_c *)arg;
432 sector_t sector = 0;
433
434 while (!kthread_should_stop()) {
435 bool super = false;
436 bool logging_enabled;
437 struct pending_block *block = NULL;
438 int ret;
439
440 spin_lock_irq(&lc->blocks_lock);
441 if (!list_empty(&lc->logging_blocks)) {
442 block = list_first_entry(&lc->logging_blocks,
443 struct pending_block, list);
444 list_del_init(&block->list);
445 if (!lc->logging_enabled)
446 goto next;
447
448 sector = lc->next_sector;
449 if (!(block->flags & LOG_DISCARD_FLAG))
450 lc->next_sector += dev_to_bio_sectors(lc, block->nr_sectors);
451 lc->next_sector += dev_to_bio_sectors(lc, 1);
452
453 /*
454 * Apparently the size of the device may not be known
455 * right away, so handle this properly.
456 */
457 if (!lc->end_sector)
458 lc->end_sector = logdev_last_sector(lc);
459 if (lc->end_sector &&
460 lc->next_sector >= lc->end_sector) {
461 DMERR("Ran out of space on the logdev");
462 lc->logging_enabled = false;
463 goto next;
464 }
465 lc->logged_entries++;
466 atomic_inc(&lc->io_blocks);
467
468 super = (block->flags & (LOG_FUA_FLAG | LOG_MARK_FLAG));
469 if (super)
470 atomic_inc(&lc->io_blocks);
471 }
472next:
473 logging_enabled = lc->logging_enabled;
474 spin_unlock_irq(&lc->blocks_lock);
475 if (block) {
476 if (logging_enabled) {
477 ret = log_one_block(lc, block, sector);
478 if (!ret && super)
479 ret = log_super(lc);
480 if (ret) {
481 spin_lock_irq(&lc->blocks_lock);
482 lc->logging_enabled = false;
483 spin_unlock_irq(&lc->blocks_lock);
484 }
485 } else
486 free_pending_block(lc, block);
487 continue;
488 }
489
490 if (!try_to_freeze()) {
491 set_current_state(TASK_INTERRUPTIBLE);
492 if (!kthread_should_stop() &&
493 list_empty(&lc->logging_blocks))
494 schedule();
495 __set_current_state(TASK_RUNNING);
496 }
497 }
498 return 0;
499}
500
501/*
502 * Construct a log-writes mapping:
503 * log-writes <dev_path> <log_dev_path>
504 */
505static int log_writes_ctr(struct dm_target *ti, unsigned int argc, char **argv)
506{
507 struct log_writes_c *lc;
508 struct dm_arg_set as;
509 const char *devname, *logdevname;
510 int ret;
511
512 as.argc = argc;
513 as.argv = argv;
514
515 if (argc < 2) {
516 ti->error = "Invalid argument count";
517 return -EINVAL;
518 }
519
520 lc = kzalloc(sizeof(struct log_writes_c), GFP_KERNEL);
521 if (!lc) {
522 ti->error = "Cannot allocate context";
523 return -ENOMEM;
524 }
525 spin_lock_init(&lc->blocks_lock);
526 INIT_LIST_HEAD(&lc->unflushed_blocks);
527 INIT_LIST_HEAD(&lc->logging_blocks);
528 init_waitqueue_head(&lc->wait);
529 init_completion(&lc->super_done);
530 atomic_set(&lc->io_blocks, 0);
531 atomic_set(&lc->pending_blocks, 0);
532
533 devname = dm_shift_arg(&as);
534 ret = dm_get_device(ti, devname, dm_table_get_mode(ti->table), &lc->dev);
535 if (ret) {
536 ti->error = "Device lookup failed";
537 goto bad;
538 }
539
540 logdevname = dm_shift_arg(&as);
541 ret = dm_get_device(ti, logdevname, dm_table_get_mode(ti->table),
542 &lc->logdev);
543 if (ret) {
544 ti->error = "Log device lookup failed";
545 dm_put_device(ti, lc->dev);
546 goto bad;
547 }
548
549 lc->sectorsize = bdev_logical_block_size(lc->dev->bdev);
550 lc->sectorshift = ilog2(lc->sectorsize);
551 lc->log_kthread = kthread_run(log_writes_kthread, lc, "log-write");
552 if (IS_ERR(lc->log_kthread)) {
553 ret = PTR_ERR(lc->log_kthread);
554 ti->error = "Couldn't alloc kthread";
555 dm_put_device(ti, lc->dev);
556 dm_put_device(ti, lc->logdev);
557 goto bad;
558 }
559
560 /*
561 * next_sector is in 512b sectors to correspond to what bi_sector expects.
562 * The super starts at sector 0, and the next_sector is the next logical
563 * one based on the sectorsize of the device.
564 */
565 lc->next_sector = lc->sectorsize >> SECTOR_SHIFT;
566 lc->logging_enabled = true;
567 lc->end_sector = logdev_last_sector(lc);
568 lc->device_supports_discard = true;
569
570 ti->num_flush_bios = 1;
571 ti->flush_supported = true;
572 ti->num_discard_bios = 1;
573 ti->discards_supported = true;
574 ti->per_io_data_size = sizeof(struct per_bio_data);
575 ti->private = lc;
576 return 0;
577
578bad:
579 kfree(lc);
580 return ret;
581}
582
583static int log_mark(struct log_writes_c *lc, char *data)
584{
585 struct pending_block *block;
586 size_t maxsize = lc->sectorsize - sizeof(struct log_write_entry);
587
588 block = kzalloc(sizeof(struct pending_block), GFP_KERNEL);
589 if (!block) {
590 DMERR("Error allocating pending block");
591 return -ENOMEM;
592 }
593
594 block->data = kstrndup(data, maxsize - 1, GFP_KERNEL);
595 if (!block->data) {
596 DMERR("Error copying mark data");
597 kfree(block);
598 return -ENOMEM;
599 }
600 atomic_inc(&lc->pending_blocks);
601 block->datalen = strlen(block->data);
602 block->flags |= LOG_MARK_FLAG;
603 spin_lock_irq(&lc->blocks_lock);
604 list_add_tail(&block->list, &lc->logging_blocks);
605 spin_unlock_irq(&lc->blocks_lock);
606 wake_up_process(lc->log_kthread);
607 return 0;
608}
609
610static void log_writes_dtr(struct dm_target *ti)
611{
612 struct log_writes_c *lc = ti->private;
613
614 spin_lock_irq(&lc->blocks_lock);
615 list_splice_init(&lc->unflushed_blocks, &lc->logging_blocks);
616 spin_unlock_irq(&lc->blocks_lock);
617
618 /*
619 * This is just nice to have since it'll update the super to include the
620 * unflushed blocks, if it fails we don't really care.
621 */
622 log_mark(lc, "dm-log-writes-end");
623 wake_up_process(lc->log_kthread);
624 wait_event(lc->wait, !atomic_read(&lc->io_blocks) &&
625 !atomic_read(&lc->pending_blocks));
626 kthread_stop(lc->log_kthread);
627
628 WARN_ON(!list_empty(&lc->logging_blocks));
629 WARN_ON(!list_empty(&lc->unflushed_blocks));
630 dm_put_device(ti, lc->dev);
631 dm_put_device(ti, lc->logdev);
632 kfree(lc);
633}
634
635static void normal_map_bio(struct dm_target *ti, struct bio *bio)
636{
637 struct log_writes_c *lc = ti->private;
638
639 bio_set_dev(bio, lc->dev->bdev);
640}
641
642static int log_writes_map(struct dm_target *ti, struct bio *bio)
643{
644 struct log_writes_c *lc = ti->private;
645 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
646 struct pending_block *block;
647 struct bvec_iter iter;
648 struct bio_vec bv;
649 size_t alloc_size;
650 int i = 0;
651 bool flush_bio = (bio->bi_opf & REQ_PREFLUSH);
652 bool fua_bio = (bio->bi_opf & REQ_FUA);
653 bool discard_bio = (bio_op(bio) == REQ_OP_DISCARD);
654 bool meta_bio = (bio->bi_opf & REQ_META);
655
656 pb->block = NULL;
657
658 /* Don't bother doing anything if logging has been disabled */
659 if (!lc->logging_enabled)
660 goto map_bio;
661
662 /*
663 * Map reads as normal.
664 */
665 if (bio_data_dir(bio) == READ)
666 goto map_bio;
667
668 /* No sectors and not a flush? Don't care */
669 if (!bio_sectors(bio) && !flush_bio)
670 goto map_bio;
671
672 /*
673 * Discards will have bi_size set but there's no actual data, so just
674 * allocate the size of the pending block.
675 */
676 if (discard_bio)
677 alloc_size = sizeof(struct pending_block);
678 else
679 alloc_size = struct_size(block, vecs, bio_segments(bio));
680
681 block = kzalloc(alloc_size, GFP_NOIO);
682 if (!block) {
683 DMERR("Error allocating pending block");
684 spin_lock_irq(&lc->blocks_lock);
685 lc->logging_enabled = false;
686 spin_unlock_irq(&lc->blocks_lock);
687 return DM_MAPIO_KILL;
688 }
689 INIT_LIST_HEAD(&block->list);
690 pb->block = block;
691 atomic_inc(&lc->pending_blocks);
692
693 if (flush_bio)
694 block->flags |= LOG_FLUSH_FLAG;
695 if (fua_bio)
696 block->flags |= LOG_FUA_FLAG;
697 if (discard_bio)
698 block->flags |= LOG_DISCARD_FLAG;
699 if (meta_bio)
700 block->flags |= LOG_METADATA_FLAG;
701
702 block->sector = bio_to_dev_sectors(lc, bio->bi_iter.bi_sector);
703 block->nr_sectors = bio_to_dev_sectors(lc, bio_sectors(bio));
704
705 /* We don't need the data, just submit */
706 if (discard_bio) {
707 WARN_ON(flush_bio || fua_bio);
708 if (lc->device_supports_discard)
709 goto map_bio;
710 bio_endio(bio);
711 return DM_MAPIO_SUBMITTED;
712 }
713
714 /* Flush bio, splice the unflushed blocks onto this list and submit */
715 if (flush_bio && !bio_sectors(bio)) {
716 spin_lock_irq(&lc->blocks_lock);
717 list_splice_init(&lc->unflushed_blocks, &block->list);
718 spin_unlock_irq(&lc->blocks_lock);
719 goto map_bio;
720 }
721
722 /*
723 * We will write this bio somewhere else way later so we need to copy
724 * the actual contents into new pages so we know the data will always be
725 * there.
726 *
727 * We do this because this could be a bio from O_DIRECT in which case we
728 * can't just hold onto the page until some later point, we have to
729 * manually copy the contents.
730 */
731 bio_for_each_segment(bv, bio, iter) {
732 struct page *page;
733 void *dst;
734
735 page = alloc_page(GFP_NOIO);
736 if (!page) {
737 DMERR("Error allocing page");
738 free_pending_block(lc, block);
739 spin_lock_irq(&lc->blocks_lock);
740 lc->logging_enabled = false;
741 spin_unlock_irq(&lc->blocks_lock);
742 return DM_MAPIO_KILL;
743 }
744
745 dst = kmap_atomic(page);
746 memcpy_from_bvec(dst, &bv);
747 kunmap_atomic(dst);
748 block->vecs[i].bv_page = page;
749 block->vecs[i].bv_len = bv.bv_len;
750 block->vec_cnt++;
751 i++;
752 }
753
754 /* Had a flush with data in it, weird */
755 if (flush_bio) {
756 spin_lock_irq(&lc->blocks_lock);
757 list_splice_init(&lc->unflushed_blocks, &block->list);
758 spin_unlock_irq(&lc->blocks_lock);
759 }
760map_bio:
761 normal_map_bio(ti, bio);
762 return DM_MAPIO_REMAPPED;
763}
764
765static int normal_end_io(struct dm_target *ti, struct bio *bio,
766 blk_status_t *error)
767{
768 struct log_writes_c *lc = ti->private;
769 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
770
771 if (bio_data_dir(bio) == WRITE && pb->block) {
772 struct pending_block *block = pb->block;
773 unsigned long flags;
774
775 spin_lock_irqsave(&lc->blocks_lock, flags);
776 if (block->flags & LOG_FLUSH_FLAG) {
777 list_splice_tail_init(&block->list, &lc->logging_blocks);
778 list_add_tail(&block->list, &lc->logging_blocks);
779 wake_up_process(lc->log_kthread);
780 } else if (block->flags & LOG_FUA_FLAG) {
781 list_add_tail(&block->list, &lc->logging_blocks);
782 wake_up_process(lc->log_kthread);
783 } else
784 list_add_tail(&block->list, &lc->unflushed_blocks);
785 spin_unlock_irqrestore(&lc->blocks_lock, flags);
786 }
787
788 return DM_ENDIO_DONE;
789}
790
791/*
792 * INFO format: <logged entries> <highest allocated sector>
793 */
794static void log_writes_status(struct dm_target *ti, status_type_t type,
795 unsigned status_flags, char *result,
796 unsigned maxlen)
797{
798 unsigned sz = 0;
799 struct log_writes_c *lc = ti->private;
800
801 switch (type) {
802 case STATUSTYPE_INFO:
803 DMEMIT("%llu %llu", lc->logged_entries,
804 (unsigned long long)lc->next_sector - 1);
805 if (!lc->logging_enabled)
806 DMEMIT(" logging_disabled");
807 break;
808
809 case STATUSTYPE_TABLE:
810 DMEMIT("%s %s", lc->dev->name, lc->logdev->name);
811 break;
812
813 case STATUSTYPE_IMA:
814 *result = '\0';
815 break;
816 }
817}
818
819static int log_writes_prepare_ioctl(struct dm_target *ti,
820 struct block_device **bdev)
821{
822 struct log_writes_c *lc = ti->private;
823 struct dm_dev *dev = lc->dev;
824
825 *bdev = dev->bdev;
826 /*
827 * Only pass ioctls through if the device sizes match exactly.
828 */
829 if (ti->len != bdev_nr_sectors(dev->bdev))
830 return 1;
831 return 0;
832}
833
834static int log_writes_iterate_devices(struct dm_target *ti,
835 iterate_devices_callout_fn fn,
836 void *data)
837{
838 struct log_writes_c *lc = ti->private;
839
840 return fn(ti, lc->dev, 0, ti->len, data);
841}
842
843/*
844 * Messages supported:
845 * mark <mark data> - specify the marked data.
846 */
847static int log_writes_message(struct dm_target *ti, unsigned argc, char **argv,
848 char *result, unsigned maxlen)
849{
850 int r = -EINVAL;
851 struct log_writes_c *lc = ti->private;
852
853 if (argc != 2) {
854 DMWARN("Invalid log-writes message arguments, expect 2 arguments, got %d", argc);
855 return r;
856 }
857
858 if (!strcasecmp(argv[0], "mark"))
859 r = log_mark(lc, argv[1]);
860 else
861 DMWARN("Unrecognised log writes target message received: %s", argv[0]);
862
863 return r;
864}
865
866static void log_writes_io_hints(struct dm_target *ti, struct queue_limits *limits)
867{
868 struct log_writes_c *lc = ti->private;
869
870 if (!bdev_max_discard_sectors(lc->dev->bdev)) {
871 lc->device_supports_discard = false;
872 limits->discard_granularity = lc->sectorsize;
873 limits->max_discard_sectors = (UINT_MAX >> SECTOR_SHIFT);
874 }
875 limits->logical_block_size = bdev_logical_block_size(lc->dev->bdev);
876 limits->physical_block_size = bdev_physical_block_size(lc->dev->bdev);
877 limits->io_min = limits->physical_block_size;
878 limits->dma_alignment = limits->logical_block_size - 1;
879}
880
881#if IS_ENABLED(CONFIG_FS_DAX)
882static struct dax_device *log_writes_dax_pgoff(struct dm_target *ti,
883 pgoff_t *pgoff)
884{
885 struct log_writes_c *lc = ti->private;
886
887 *pgoff += (get_start_sect(lc->dev->bdev) >> PAGE_SECTORS_SHIFT);
888 return lc->dev->dax_dev;
889}
890
891static long log_writes_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
892 long nr_pages, enum dax_access_mode mode, void **kaddr,
893 pfn_t *pfn)
894{
895 struct dax_device *dax_dev = log_writes_dax_pgoff(ti, &pgoff);
896
897 return dax_direct_access(dax_dev, pgoff, nr_pages, mode, kaddr, pfn);
898}
899
900static int log_writes_dax_zero_page_range(struct dm_target *ti, pgoff_t pgoff,
901 size_t nr_pages)
902{
903 struct dax_device *dax_dev = log_writes_dax_pgoff(ti, &pgoff);
904
905 return dax_zero_page_range(dax_dev, pgoff, nr_pages << PAGE_SHIFT);
906}
907
908static size_t log_writes_dax_recovery_write(struct dm_target *ti,
909 pgoff_t pgoff, void *addr, size_t bytes, struct iov_iter *i)
910{
911 struct dax_device *dax_dev = log_writes_dax_pgoff(ti, &pgoff);
912
913 return dax_recovery_write(dax_dev, pgoff, addr, bytes, i);
914}
915
916#else
917#define log_writes_dax_direct_access NULL
918#define log_writes_dax_zero_page_range NULL
919#define log_writes_dax_recovery_write NULL
920#endif
921
922static struct target_type log_writes_target = {
923 .name = "log-writes",
924 .version = {1, 1, 0},
925 .module = THIS_MODULE,
926 .ctr = log_writes_ctr,
927 .dtr = log_writes_dtr,
928 .map = log_writes_map,
929 .end_io = normal_end_io,
930 .status = log_writes_status,
931 .prepare_ioctl = log_writes_prepare_ioctl,
932 .message = log_writes_message,
933 .iterate_devices = log_writes_iterate_devices,
934 .io_hints = log_writes_io_hints,
935 .direct_access = log_writes_dax_direct_access,
936 .dax_zero_page_range = log_writes_dax_zero_page_range,
937 .dax_recovery_write = log_writes_dax_recovery_write,
938};
939
940static int __init dm_log_writes_init(void)
941{
942 int r = dm_register_target(&log_writes_target);
943
944 if (r < 0)
945 DMERR("register failed %d", r);
946
947 return r;
948}
949
950static void __exit dm_log_writes_exit(void)
951{
952 dm_unregister_target(&log_writes_target);
953}
954
955module_init(dm_log_writes_init);
956module_exit(dm_log_writes_exit);
957
958MODULE_DESCRIPTION(DM_NAME " log writes target");
959MODULE_AUTHOR("Josef Bacik <jbacik@fb.com>");
960MODULE_LICENSE("GPL");
1/*
2 * Copyright (C) 2014 Facebook. All rights reserved.
3 *
4 * This file is released under the GPL.
5 */
6
7#include <linux/device-mapper.h>
8
9#include <linux/module.h>
10#include <linux/init.h>
11#include <linux/blkdev.h>
12#include <linux/bio.h>
13#include <linux/dax.h>
14#include <linux/slab.h>
15#include <linux/kthread.h>
16#include <linux/freezer.h>
17#include <linux/uio.h>
18
19#define DM_MSG_PREFIX "log-writes"
20
21/*
22 * This target will sequentially log all writes to the target device onto the
23 * log device. This is helpful for replaying writes to check for fs consistency
24 * at all times. This target provides a mechanism to mark specific events to
25 * check data at a later time. So for example you would:
26 *
27 * write data
28 * fsync
29 * dmsetup message /dev/whatever mark mymark
30 * unmount /mnt/test
31 *
32 * Then replay the log up to mymark and check the contents of the replay to
33 * verify it matches what was written.
34 *
35 * We log writes only after they have been flushed, this makes the log describe
36 * close to the order in which the data hits the actual disk, not its cache. So
37 * for example the following sequence (W means write, C means complete)
38 *
39 * Wa,Wb,Wc,Cc,Ca,FLUSH,FUAd,Cb,CFLUSH,CFUAd
40 *
41 * Would result in the log looking like this:
42 *
43 * c,a,b,flush,fuad,<other writes>,<next flush>
44 *
45 * This is meant to help expose problems where file systems do not properly wait
46 * on data being written before invoking a FLUSH. FUA bypasses cache so once it
47 * completes it is added to the log as it should be on disk.
48 *
49 * We treat DISCARDs as if they don't bypass cache so that they are logged in
50 * order of completion along with the normal writes. If we didn't do it this
51 * way we would process all the discards first and then write all the data, when
52 * in fact we want to do the data and the discard in the order that they
53 * completed.
54 */
55#define LOG_FLUSH_FLAG (1 << 0)
56#define LOG_FUA_FLAG (1 << 1)
57#define LOG_DISCARD_FLAG (1 << 2)
58#define LOG_MARK_FLAG (1 << 3)
59#define LOG_METADATA_FLAG (1 << 4)
60
61#define WRITE_LOG_VERSION 1ULL
62#define WRITE_LOG_MAGIC 0x6a736677736872ULL
63#define WRITE_LOG_SUPER_SECTOR 0
64
65/*
66 * The disk format for this is braindead simple.
67 *
68 * At byte 0 we have our super, followed by the following sequence for
69 * nr_entries:
70 *
71 * [ 1 sector ][ entry->nr_sectors ]
72 * [log_write_entry][ data written ]
73 *
74 * The log_write_entry takes up a full sector so we can have arbitrary length
75 * marks and it leaves us room for extra content in the future.
76 */
77
78/*
79 * Basic info about the log for userspace.
80 */
81struct log_write_super {
82 __le64 magic;
83 __le64 version;
84 __le64 nr_entries;
85 __le32 sectorsize;
86};
87
88/*
89 * sector - the sector we wrote.
90 * nr_sectors - the number of sectors we wrote.
91 * flags - flags for this log entry.
92 * data_len - the size of the data in this log entry, this is for private log
93 * entry stuff, the MARK data provided by userspace for example.
94 */
95struct log_write_entry {
96 __le64 sector;
97 __le64 nr_sectors;
98 __le64 flags;
99 __le64 data_len;
100};
101
102struct log_writes_c {
103 struct dm_dev *dev;
104 struct dm_dev *logdev;
105 u64 logged_entries;
106 u32 sectorsize;
107 u32 sectorshift;
108 atomic_t io_blocks;
109 atomic_t pending_blocks;
110 sector_t next_sector;
111 sector_t end_sector;
112 bool logging_enabled;
113 bool device_supports_discard;
114 spinlock_t blocks_lock;
115 struct list_head unflushed_blocks;
116 struct list_head logging_blocks;
117 wait_queue_head_t wait;
118 struct task_struct *log_kthread;
119 struct completion super_done;
120};
121
122struct pending_block {
123 int vec_cnt;
124 u64 flags;
125 sector_t sector;
126 sector_t nr_sectors;
127 char *data;
128 u32 datalen;
129 struct list_head list;
130 struct bio_vec vecs[];
131};
132
133struct per_bio_data {
134 struct pending_block *block;
135};
136
137static inline sector_t bio_to_dev_sectors(struct log_writes_c *lc,
138 sector_t sectors)
139{
140 return sectors >> (lc->sectorshift - SECTOR_SHIFT);
141}
142
143static inline sector_t dev_to_bio_sectors(struct log_writes_c *lc,
144 sector_t sectors)
145{
146 return sectors << (lc->sectorshift - SECTOR_SHIFT);
147}
148
149static void put_pending_block(struct log_writes_c *lc)
150{
151 if (atomic_dec_and_test(&lc->pending_blocks)) {
152 smp_mb__after_atomic();
153 if (waitqueue_active(&lc->wait))
154 wake_up(&lc->wait);
155 }
156}
157
158static void put_io_block(struct log_writes_c *lc)
159{
160 if (atomic_dec_and_test(&lc->io_blocks)) {
161 smp_mb__after_atomic();
162 if (waitqueue_active(&lc->wait))
163 wake_up(&lc->wait);
164 }
165}
166
167static void log_end_io(struct bio *bio)
168{
169 struct log_writes_c *lc = bio->bi_private;
170
171 if (bio->bi_status) {
172 unsigned long flags;
173
174 DMERR("Error writing log block, error=%d", bio->bi_status);
175 spin_lock_irqsave(&lc->blocks_lock, flags);
176 lc->logging_enabled = false;
177 spin_unlock_irqrestore(&lc->blocks_lock, flags);
178 }
179
180 bio_free_pages(bio);
181 put_io_block(lc);
182 bio_put(bio);
183}
184
185static void log_end_super(struct bio *bio)
186{
187 struct log_writes_c *lc = bio->bi_private;
188
189 complete(&lc->super_done);
190 log_end_io(bio);
191}
192
193/*
194 * Meant to be called if there is an error, it will free all the pages
195 * associated with the block.
196 */
197static void free_pending_block(struct log_writes_c *lc,
198 struct pending_block *block)
199{
200 int i;
201
202 for (i = 0; i < block->vec_cnt; i++) {
203 if (block->vecs[i].bv_page)
204 __free_page(block->vecs[i].bv_page);
205 }
206 kfree(block->data);
207 kfree(block);
208 put_pending_block(lc);
209}
210
211static int write_metadata(struct log_writes_c *lc, void *entry,
212 size_t entrylen, void *data, size_t datalen,
213 sector_t sector)
214{
215 struct bio *bio;
216 struct page *page;
217 void *ptr;
218 size_t ret;
219
220 bio = bio_alloc(GFP_KERNEL, 1);
221 if (!bio) {
222 DMERR("Couldn't alloc log bio");
223 goto error;
224 }
225 bio->bi_iter.bi_size = 0;
226 bio->bi_iter.bi_sector = sector;
227 bio_set_dev(bio, lc->logdev->bdev);
228 bio->bi_end_io = (sector == WRITE_LOG_SUPER_SECTOR) ?
229 log_end_super : log_end_io;
230 bio->bi_private = lc;
231 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
232
233 page = alloc_page(GFP_KERNEL);
234 if (!page) {
235 DMERR("Couldn't alloc log page");
236 bio_put(bio);
237 goto error;
238 }
239
240 ptr = kmap_atomic(page);
241 memcpy(ptr, entry, entrylen);
242 if (datalen)
243 memcpy(ptr + entrylen, data, datalen);
244 memset(ptr + entrylen + datalen, 0,
245 lc->sectorsize - entrylen - datalen);
246 kunmap_atomic(ptr);
247
248 ret = bio_add_page(bio, page, lc->sectorsize, 0);
249 if (ret != lc->sectorsize) {
250 DMERR("Couldn't add page to the log block");
251 goto error_bio;
252 }
253 submit_bio(bio);
254 return 0;
255error_bio:
256 bio_put(bio);
257 __free_page(page);
258error:
259 put_io_block(lc);
260 return -1;
261}
262
263static int write_inline_data(struct log_writes_c *lc, void *entry,
264 size_t entrylen, void *data, size_t datalen,
265 sector_t sector)
266{
267 int num_pages, bio_pages, pg_datalen, pg_sectorlen, i;
268 struct page *page;
269 struct bio *bio;
270 size_t ret;
271 void *ptr;
272
273 while (datalen) {
274 num_pages = ALIGN(datalen, PAGE_SIZE) >> PAGE_SHIFT;
275 bio_pages = min(num_pages, BIO_MAX_PAGES);
276
277 atomic_inc(&lc->io_blocks);
278
279 bio = bio_alloc(GFP_KERNEL, bio_pages);
280 if (!bio) {
281 DMERR("Couldn't alloc inline data bio");
282 goto error;
283 }
284
285 bio->bi_iter.bi_size = 0;
286 bio->bi_iter.bi_sector = sector;
287 bio_set_dev(bio, lc->logdev->bdev);
288 bio->bi_end_io = log_end_io;
289 bio->bi_private = lc;
290 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
291
292 for (i = 0; i < bio_pages; i++) {
293 pg_datalen = min_t(int, datalen, PAGE_SIZE);
294 pg_sectorlen = ALIGN(pg_datalen, lc->sectorsize);
295
296 page = alloc_page(GFP_KERNEL);
297 if (!page) {
298 DMERR("Couldn't alloc inline data page");
299 goto error_bio;
300 }
301
302 ptr = kmap_atomic(page);
303 memcpy(ptr, data, pg_datalen);
304 if (pg_sectorlen > pg_datalen)
305 memset(ptr + pg_datalen, 0, pg_sectorlen - pg_datalen);
306 kunmap_atomic(ptr);
307
308 ret = bio_add_page(bio, page, pg_sectorlen, 0);
309 if (ret != pg_sectorlen) {
310 DMERR("Couldn't add page of inline data");
311 __free_page(page);
312 goto error_bio;
313 }
314
315 datalen -= pg_datalen;
316 data += pg_datalen;
317 }
318 submit_bio(bio);
319
320 sector += bio_pages * PAGE_SECTORS;
321 }
322 return 0;
323error_bio:
324 bio_free_pages(bio);
325 bio_put(bio);
326error:
327 put_io_block(lc);
328 return -1;
329}
330
331static int log_one_block(struct log_writes_c *lc,
332 struct pending_block *block, sector_t sector)
333{
334 struct bio *bio;
335 struct log_write_entry entry;
336 size_t metadatalen, ret;
337 int i;
338
339 entry.sector = cpu_to_le64(block->sector);
340 entry.nr_sectors = cpu_to_le64(block->nr_sectors);
341 entry.flags = cpu_to_le64(block->flags);
342 entry.data_len = cpu_to_le64(block->datalen);
343
344 metadatalen = (block->flags & LOG_MARK_FLAG) ? block->datalen : 0;
345 if (write_metadata(lc, &entry, sizeof(entry), block->data,
346 metadatalen, sector)) {
347 free_pending_block(lc, block);
348 return -1;
349 }
350
351 sector += dev_to_bio_sectors(lc, 1);
352
353 if (block->datalen && metadatalen == 0) {
354 if (write_inline_data(lc, &entry, sizeof(entry), block->data,
355 block->datalen, sector)) {
356 free_pending_block(lc, block);
357 return -1;
358 }
359 /* we don't support both inline data & bio data */
360 goto out;
361 }
362
363 if (!block->vec_cnt)
364 goto out;
365
366 atomic_inc(&lc->io_blocks);
367 bio = bio_alloc(GFP_KERNEL, min(block->vec_cnt, BIO_MAX_PAGES));
368 if (!bio) {
369 DMERR("Couldn't alloc log bio");
370 goto error;
371 }
372 bio->bi_iter.bi_size = 0;
373 bio->bi_iter.bi_sector = sector;
374 bio_set_dev(bio, lc->logdev->bdev);
375 bio->bi_end_io = log_end_io;
376 bio->bi_private = lc;
377 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
378
379 for (i = 0; i < block->vec_cnt; i++) {
380 /*
381 * The page offset is always 0 because we allocate a new page
382 * for every bvec in the original bio for simplicity sake.
383 */
384 ret = bio_add_page(bio, block->vecs[i].bv_page,
385 block->vecs[i].bv_len, 0);
386 if (ret != block->vecs[i].bv_len) {
387 atomic_inc(&lc->io_blocks);
388 submit_bio(bio);
389 bio = bio_alloc(GFP_KERNEL, min(block->vec_cnt - i, BIO_MAX_PAGES));
390 if (!bio) {
391 DMERR("Couldn't alloc log bio");
392 goto error;
393 }
394 bio->bi_iter.bi_size = 0;
395 bio->bi_iter.bi_sector = sector;
396 bio_set_dev(bio, lc->logdev->bdev);
397 bio->bi_end_io = log_end_io;
398 bio->bi_private = lc;
399 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
400
401 ret = bio_add_page(bio, block->vecs[i].bv_page,
402 block->vecs[i].bv_len, 0);
403 if (ret != block->vecs[i].bv_len) {
404 DMERR("Couldn't add page on new bio?");
405 bio_put(bio);
406 goto error;
407 }
408 }
409 sector += block->vecs[i].bv_len >> SECTOR_SHIFT;
410 }
411 submit_bio(bio);
412out:
413 kfree(block->data);
414 kfree(block);
415 put_pending_block(lc);
416 return 0;
417error:
418 free_pending_block(lc, block);
419 put_io_block(lc);
420 return -1;
421}
422
423static int log_super(struct log_writes_c *lc)
424{
425 struct log_write_super super;
426
427 super.magic = cpu_to_le64(WRITE_LOG_MAGIC);
428 super.version = cpu_to_le64(WRITE_LOG_VERSION);
429 super.nr_entries = cpu_to_le64(lc->logged_entries);
430 super.sectorsize = cpu_to_le32(lc->sectorsize);
431
432 if (write_metadata(lc, &super, sizeof(super), NULL, 0,
433 WRITE_LOG_SUPER_SECTOR)) {
434 DMERR("Couldn't write super");
435 return -1;
436 }
437
438 /*
439 * Super sector should be writen in-order, otherwise the
440 * nr_entries could be rewritten incorrectly by an old bio.
441 */
442 wait_for_completion_io(&lc->super_done);
443
444 return 0;
445}
446
447static inline sector_t logdev_last_sector(struct log_writes_c *lc)
448{
449 return i_size_read(lc->logdev->bdev->bd_inode) >> SECTOR_SHIFT;
450}
451
452static int log_writes_kthread(void *arg)
453{
454 struct log_writes_c *lc = (struct log_writes_c *)arg;
455 sector_t sector = 0;
456
457 while (!kthread_should_stop()) {
458 bool super = false;
459 bool logging_enabled;
460 struct pending_block *block = NULL;
461 int ret;
462
463 spin_lock_irq(&lc->blocks_lock);
464 if (!list_empty(&lc->logging_blocks)) {
465 block = list_first_entry(&lc->logging_blocks,
466 struct pending_block, list);
467 list_del_init(&block->list);
468 if (!lc->logging_enabled)
469 goto next;
470
471 sector = lc->next_sector;
472 if (!(block->flags & LOG_DISCARD_FLAG))
473 lc->next_sector += dev_to_bio_sectors(lc, block->nr_sectors);
474 lc->next_sector += dev_to_bio_sectors(lc, 1);
475
476 /*
477 * Apparently the size of the device may not be known
478 * right away, so handle this properly.
479 */
480 if (!lc->end_sector)
481 lc->end_sector = logdev_last_sector(lc);
482 if (lc->end_sector &&
483 lc->next_sector >= lc->end_sector) {
484 DMERR("Ran out of space on the logdev");
485 lc->logging_enabled = false;
486 goto next;
487 }
488 lc->logged_entries++;
489 atomic_inc(&lc->io_blocks);
490
491 super = (block->flags & (LOG_FUA_FLAG | LOG_MARK_FLAG));
492 if (super)
493 atomic_inc(&lc->io_blocks);
494 }
495next:
496 logging_enabled = lc->logging_enabled;
497 spin_unlock_irq(&lc->blocks_lock);
498 if (block) {
499 if (logging_enabled) {
500 ret = log_one_block(lc, block, sector);
501 if (!ret && super)
502 ret = log_super(lc);
503 if (ret) {
504 spin_lock_irq(&lc->blocks_lock);
505 lc->logging_enabled = false;
506 spin_unlock_irq(&lc->blocks_lock);
507 }
508 } else
509 free_pending_block(lc, block);
510 continue;
511 }
512
513 if (!try_to_freeze()) {
514 set_current_state(TASK_INTERRUPTIBLE);
515 if (!kthread_should_stop() &&
516 list_empty(&lc->logging_blocks))
517 schedule();
518 __set_current_state(TASK_RUNNING);
519 }
520 }
521 return 0;
522}
523
524/*
525 * Construct a log-writes mapping:
526 * log-writes <dev_path> <log_dev_path>
527 */
528static int log_writes_ctr(struct dm_target *ti, unsigned int argc, char **argv)
529{
530 struct log_writes_c *lc;
531 struct dm_arg_set as;
532 const char *devname, *logdevname;
533 int ret;
534
535 as.argc = argc;
536 as.argv = argv;
537
538 if (argc < 2) {
539 ti->error = "Invalid argument count";
540 return -EINVAL;
541 }
542
543 lc = kzalloc(sizeof(struct log_writes_c), GFP_KERNEL);
544 if (!lc) {
545 ti->error = "Cannot allocate context";
546 return -ENOMEM;
547 }
548 spin_lock_init(&lc->blocks_lock);
549 INIT_LIST_HEAD(&lc->unflushed_blocks);
550 INIT_LIST_HEAD(&lc->logging_blocks);
551 init_waitqueue_head(&lc->wait);
552 init_completion(&lc->super_done);
553 atomic_set(&lc->io_blocks, 0);
554 atomic_set(&lc->pending_blocks, 0);
555
556 devname = dm_shift_arg(&as);
557 ret = dm_get_device(ti, devname, dm_table_get_mode(ti->table), &lc->dev);
558 if (ret) {
559 ti->error = "Device lookup failed";
560 goto bad;
561 }
562
563 logdevname = dm_shift_arg(&as);
564 ret = dm_get_device(ti, logdevname, dm_table_get_mode(ti->table),
565 &lc->logdev);
566 if (ret) {
567 ti->error = "Log device lookup failed";
568 dm_put_device(ti, lc->dev);
569 goto bad;
570 }
571
572 lc->sectorsize = bdev_logical_block_size(lc->dev->bdev);
573 lc->sectorshift = ilog2(lc->sectorsize);
574 lc->log_kthread = kthread_run(log_writes_kthread, lc, "log-write");
575 if (IS_ERR(lc->log_kthread)) {
576 ret = PTR_ERR(lc->log_kthread);
577 ti->error = "Couldn't alloc kthread";
578 dm_put_device(ti, lc->dev);
579 dm_put_device(ti, lc->logdev);
580 goto bad;
581 }
582
583 /*
584 * next_sector is in 512b sectors to correspond to what bi_sector expects.
585 * The super starts at sector 0, and the next_sector is the next logical
586 * one based on the sectorsize of the device.
587 */
588 lc->next_sector = lc->sectorsize >> SECTOR_SHIFT;
589 lc->logging_enabled = true;
590 lc->end_sector = logdev_last_sector(lc);
591 lc->device_supports_discard = true;
592
593 ti->num_flush_bios = 1;
594 ti->flush_supported = true;
595 ti->num_discard_bios = 1;
596 ti->discards_supported = true;
597 ti->per_io_data_size = sizeof(struct per_bio_data);
598 ti->private = lc;
599 return 0;
600
601bad:
602 kfree(lc);
603 return ret;
604}
605
606static int log_mark(struct log_writes_c *lc, char *data)
607{
608 struct pending_block *block;
609 size_t maxsize = lc->sectorsize - sizeof(struct log_write_entry);
610
611 block = kzalloc(sizeof(struct pending_block), GFP_KERNEL);
612 if (!block) {
613 DMERR("Error allocating pending block");
614 return -ENOMEM;
615 }
616
617 block->data = kstrndup(data, maxsize - 1, GFP_KERNEL);
618 if (!block->data) {
619 DMERR("Error copying mark data");
620 kfree(block);
621 return -ENOMEM;
622 }
623 atomic_inc(&lc->pending_blocks);
624 block->datalen = strlen(block->data);
625 block->flags |= LOG_MARK_FLAG;
626 spin_lock_irq(&lc->blocks_lock);
627 list_add_tail(&block->list, &lc->logging_blocks);
628 spin_unlock_irq(&lc->blocks_lock);
629 wake_up_process(lc->log_kthread);
630 return 0;
631}
632
633static void log_writes_dtr(struct dm_target *ti)
634{
635 struct log_writes_c *lc = ti->private;
636
637 spin_lock_irq(&lc->blocks_lock);
638 list_splice_init(&lc->unflushed_blocks, &lc->logging_blocks);
639 spin_unlock_irq(&lc->blocks_lock);
640
641 /*
642 * This is just nice to have since it'll update the super to include the
643 * unflushed blocks, if it fails we don't really care.
644 */
645 log_mark(lc, "dm-log-writes-end");
646 wake_up_process(lc->log_kthread);
647 wait_event(lc->wait, !atomic_read(&lc->io_blocks) &&
648 !atomic_read(&lc->pending_blocks));
649 kthread_stop(lc->log_kthread);
650
651 WARN_ON(!list_empty(&lc->logging_blocks));
652 WARN_ON(!list_empty(&lc->unflushed_blocks));
653 dm_put_device(ti, lc->dev);
654 dm_put_device(ti, lc->logdev);
655 kfree(lc);
656}
657
658static void normal_map_bio(struct dm_target *ti, struct bio *bio)
659{
660 struct log_writes_c *lc = ti->private;
661
662 bio_set_dev(bio, lc->dev->bdev);
663}
664
665static int log_writes_map(struct dm_target *ti, struct bio *bio)
666{
667 struct log_writes_c *lc = ti->private;
668 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
669 struct pending_block *block;
670 struct bvec_iter iter;
671 struct bio_vec bv;
672 size_t alloc_size;
673 int i = 0;
674 bool flush_bio = (bio->bi_opf & REQ_PREFLUSH);
675 bool fua_bio = (bio->bi_opf & REQ_FUA);
676 bool discard_bio = (bio_op(bio) == REQ_OP_DISCARD);
677 bool meta_bio = (bio->bi_opf & REQ_META);
678
679 pb->block = NULL;
680
681 /* Don't bother doing anything if logging has been disabled */
682 if (!lc->logging_enabled)
683 goto map_bio;
684
685 /*
686 * Map reads as normal.
687 */
688 if (bio_data_dir(bio) == READ)
689 goto map_bio;
690
691 /* No sectors and not a flush? Don't care */
692 if (!bio_sectors(bio) && !flush_bio)
693 goto map_bio;
694
695 /*
696 * Discards will have bi_size set but there's no actual data, so just
697 * allocate the size of the pending block.
698 */
699 if (discard_bio)
700 alloc_size = sizeof(struct pending_block);
701 else
702 alloc_size = struct_size(block, vecs, bio_segments(bio));
703
704 block = kzalloc(alloc_size, GFP_NOIO);
705 if (!block) {
706 DMERR("Error allocating pending block");
707 spin_lock_irq(&lc->blocks_lock);
708 lc->logging_enabled = false;
709 spin_unlock_irq(&lc->blocks_lock);
710 return DM_MAPIO_KILL;
711 }
712 INIT_LIST_HEAD(&block->list);
713 pb->block = block;
714 atomic_inc(&lc->pending_blocks);
715
716 if (flush_bio)
717 block->flags |= LOG_FLUSH_FLAG;
718 if (fua_bio)
719 block->flags |= LOG_FUA_FLAG;
720 if (discard_bio)
721 block->flags |= LOG_DISCARD_FLAG;
722 if (meta_bio)
723 block->flags |= LOG_METADATA_FLAG;
724
725 block->sector = bio_to_dev_sectors(lc, bio->bi_iter.bi_sector);
726 block->nr_sectors = bio_to_dev_sectors(lc, bio_sectors(bio));
727
728 /* We don't need the data, just submit */
729 if (discard_bio) {
730 WARN_ON(flush_bio || fua_bio);
731 if (lc->device_supports_discard)
732 goto map_bio;
733 bio_endio(bio);
734 return DM_MAPIO_SUBMITTED;
735 }
736
737 /* Flush bio, splice the unflushed blocks onto this list and submit */
738 if (flush_bio && !bio_sectors(bio)) {
739 spin_lock_irq(&lc->blocks_lock);
740 list_splice_init(&lc->unflushed_blocks, &block->list);
741 spin_unlock_irq(&lc->blocks_lock);
742 goto map_bio;
743 }
744
745 /*
746 * We will write this bio somewhere else way later so we need to copy
747 * the actual contents into new pages so we know the data will always be
748 * there.
749 *
750 * We do this because this could be a bio from O_DIRECT in which case we
751 * can't just hold onto the page until some later point, we have to
752 * manually copy the contents.
753 */
754 bio_for_each_segment(bv, bio, iter) {
755 struct page *page;
756 void *src, *dst;
757
758 page = alloc_page(GFP_NOIO);
759 if (!page) {
760 DMERR("Error allocing page");
761 free_pending_block(lc, block);
762 spin_lock_irq(&lc->blocks_lock);
763 lc->logging_enabled = false;
764 spin_unlock_irq(&lc->blocks_lock);
765 return DM_MAPIO_KILL;
766 }
767
768 src = kmap_atomic(bv.bv_page);
769 dst = kmap_atomic(page);
770 memcpy(dst, src + bv.bv_offset, bv.bv_len);
771 kunmap_atomic(dst);
772 kunmap_atomic(src);
773 block->vecs[i].bv_page = page;
774 block->vecs[i].bv_len = bv.bv_len;
775 block->vec_cnt++;
776 i++;
777 }
778
779 /* Had a flush with data in it, weird */
780 if (flush_bio) {
781 spin_lock_irq(&lc->blocks_lock);
782 list_splice_init(&lc->unflushed_blocks, &block->list);
783 spin_unlock_irq(&lc->blocks_lock);
784 }
785map_bio:
786 normal_map_bio(ti, bio);
787 return DM_MAPIO_REMAPPED;
788}
789
790static int normal_end_io(struct dm_target *ti, struct bio *bio,
791 blk_status_t *error)
792{
793 struct log_writes_c *lc = ti->private;
794 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
795
796 if (bio_data_dir(bio) == WRITE && pb->block) {
797 struct pending_block *block = pb->block;
798 unsigned long flags;
799
800 spin_lock_irqsave(&lc->blocks_lock, flags);
801 if (block->flags & LOG_FLUSH_FLAG) {
802 list_splice_tail_init(&block->list, &lc->logging_blocks);
803 list_add_tail(&block->list, &lc->logging_blocks);
804 wake_up_process(lc->log_kthread);
805 } else if (block->flags & LOG_FUA_FLAG) {
806 list_add_tail(&block->list, &lc->logging_blocks);
807 wake_up_process(lc->log_kthread);
808 } else
809 list_add_tail(&block->list, &lc->unflushed_blocks);
810 spin_unlock_irqrestore(&lc->blocks_lock, flags);
811 }
812
813 return DM_ENDIO_DONE;
814}
815
816/*
817 * INFO format: <logged entries> <highest allocated sector>
818 */
819static void log_writes_status(struct dm_target *ti, status_type_t type,
820 unsigned status_flags, char *result,
821 unsigned maxlen)
822{
823 unsigned sz = 0;
824 struct log_writes_c *lc = ti->private;
825
826 switch (type) {
827 case STATUSTYPE_INFO:
828 DMEMIT("%llu %llu", lc->logged_entries,
829 (unsigned long long)lc->next_sector - 1);
830 if (!lc->logging_enabled)
831 DMEMIT(" logging_disabled");
832 break;
833
834 case STATUSTYPE_TABLE:
835 DMEMIT("%s %s", lc->dev->name, lc->logdev->name);
836 break;
837 }
838}
839
840static int log_writes_prepare_ioctl(struct dm_target *ti,
841 struct block_device **bdev)
842{
843 struct log_writes_c *lc = ti->private;
844 struct dm_dev *dev = lc->dev;
845
846 *bdev = dev->bdev;
847 /*
848 * Only pass ioctls through if the device sizes match exactly.
849 */
850 if (ti->len != i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT)
851 return 1;
852 return 0;
853}
854
855static int log_writes_iterate_devices(struct dm_target *ti,
856 iterate_devices_callout_fn fn,
857 void *data)
858{
859 struct log_writes_c *lc = ti->private;
860
861 return fn(ti, lc->dev, 0, ti->len, data);
862}
863
864/*
865 * Messages supported:
866 * mark <mark data> - specify the marked data.
867 */
868static int log_writes_message(struct dm_target *ti, unsigned argc, char **argv,
869 char *result, unsigned maxlen)
870{
871 int r = -EINVAL;
872 struct log_writes_c *lc = ti->private;
873
874 if (argc != 2) {
875 DMWARN("Invalid log-writes message arguments, expect 2 arguments, got %d", argc);
876 return r;
877 }
878
879 if (!strcasecmp(argv[0], "mark"))
880 r = log_mark(lc, argv[1]);
881 else
882 DMWARN("Unrecognised log writes target message received: %s", argv[0]);
883
884 return r;
885}
886
887static void log_writes_io_hints(struct dm_target *ti, struct queue_limits *limits)
888{
889 struct log_writes_c *lc = ti->private;
890 struct request_queue *q = bdev_get_queue(lc->dev->bdev);
891
892 if (!q || !blk_queue_discard(q)) {
893 lc->device_supports_discard = false;
894 limits->discard_granularity = lc->sectorsize;
895 limits->max_discard_sectors = (UINT_MAX >> SECTOR_SHIFT);
896 }
897 limits->logical_block_size = bdev_logical_block_size(lc->dev->bdev);
898 limits->physical_block_size = bdev_physical_block_size(lc->dev->bdev);
899 limits->io_min = limits->physical_block_size;
900}
901
902#if IS_ENABLED(CONFIG_DAX_DRIVER)
903static int log_dax(struct log_writes_c *lc, sector_t sector, size_t bytes,
904 struct iov_iter *i)
905{
906 struct pending_block *block;
907
908 if (!bytes)
909 return 0;
910
911 block = kzalloc(sizeof(struct pending_block), GFP_KERNEL);
912 if (!block) {
913 DMERR("Error allocating dax pending block");
914 return -ENOMEM;
915 }
916
917 block->data = kzalloc(bytes, GFP_KERNEL);
918 if (!block->data) {
919 DMERR("Error allocating dax data space");
920 kfree(block);
921 return -ENOMEM;
922 }
923
924 /* write data provided via the iterator */
925 if (!copy_from_iter(block->data, bytes, i)) {
926 DMERR("Error copying dax data");
927 kfree(block->data);
928 kfree(block);
929 return -EIO;
930 }
931
932 /* rewind the iterator so that the block driver can use it */
933 iov_iter_revert(i, bytes);
934
935 block->datalen = bytes;
936 block->sector = bio_to_dev_sectors(lc, sector);
937 block->nr_sectors = ALIGN(bytes, lc->sectorsize) >> lc->sectorshift;
938
939 atomic_inc(&lc->pending_blocks);
940 spin_lock_irq(&lc->blocks_lock);
941 list_add_tail(&block->list, &lc->unflushed_blocks);
942 spin_unlock_irq(&lc->blocks_lock);
943 wake_up_process(lc->log_kthread);
944
945 return 0;
946}
947
948static long log_writes_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
949 long nr_pages, void **kaddr, pfn_t *pfn)
950{
951 struct log_writes_c *lc = ti->private;
952 sector_t sector = pgoff * PAGE_SECTORS;
953 int ret;
954
955 ret = bdev_dax_pgoff(lc->dev->bdev, sector, nr_pages * PAGE_SIZE, &pgoff);
956 if (ret)
957 return ret;
958 return dax_direct_access(lc->dev->dax_dev, pgoff, nr_pages, kaddr, pfn);
959}
960
961static size_t log_writes_dax_copy_from_iter(struct dm_target *ti,
962 pgoff_t pgoff, void *addr, size_t bytes,
963 struct iov_iter *i)
964{
965 struct log_writes_c *lc = ti->private;
966 sector_t sector = pgoff * PAGE_SECTORS;
967 int err;
968
969 if (bdev_dax_pgoff(lc->dev->bdev, sector, ALIGN(bytes, PAGE_SIZE), &pgoff))
970 return 0;
971
972 /* Don't bother doing anything if logging has been disabled */
973 if (!lc->logging_enabled)
974 goto dax_copy;
975
976 err = log_dax(lc, sector, bytes, i);
977 if (err) {
978 DMWARN("Error %d logging DAX write", err);
979 return 0;
980 }
981dax_copy:
982 return dax_copy_from_iter(lc->dev->dax_dev, pgoff, addr, bytes, i);
983}
984
985static size_t log_writes_dax_copy_to_iter(struct dm_target *ti,
986 pgoff_t pgoff, void *addr, size_t bytes,
987 struct iov_iter *i)
988{
989 struct log_writes_c *lc = ti->private;
990 sector_t sector = pgoff * PAGE_SECTORS;
991
992 if (bdev_dax_pgoff(lc->dev->bdev, sector, ALIGN(bytes, PAGE_SIZE), &pgoff))
993 return 0;
994 return dax_copy_to_iter(lc->dev->dax_dev, pgoff, addr, bytes, i);
995}
996
997static int log_writes_dax_zero_page_range(struct dm_target *ti, pgoff_t pgoff,
998 size_t nr_pages)
999{
1000 int ret;
1001 struct log_writes_c *lc = ti->private;
1002 sector_t sector = pgoff * PAGE_SECTORS;
1003
1004 ret = bdev_dax_pgoff(lc->dev->bdev, sector, nr_pages << PAGE_SHIFT,
1005 &pgoff);
1006 if (ret)
1007 return ret;
1008 return dax_zero_page_range(lc->dev->dax_dev, pgoff,
1009 nr_pages << PAGE_SHIFT);
1010}
1011
1012#else
1013#define log_writes_dax_direct_access NULL
1014#define log_writes_dax_copy_from_iter NULL
1015#define log_writes_dax_copy_to_iter NULL
1016#define log_writes_dax_zero_page_range NULL
1017#endif
1018
1019static struct target_type log_writes_target = {
1020 .name = "log-writes",
1021 .version = {1, 1, 0},
1022 .module = THIS_MODULE,
1023 .ctr = log_writes_ctr,
1024 .dtr = log_writes_dtr,
1025 .map = log_writes_map,
1026 .end_io = normal_end_io,
1027 .status = log_writes_status,
1028 .prepare_ioctl = log_writes_prepare_ioctl,
1029 .message = log_writes_message,
1030 .iterate_devices = log_writes_iterate_devices,
1031 .io_hints = log_writes_io_hints,
1032 .direct_access = log_writes_dax_direct_access,
1033 .dax_copy_from_iter = log_writes_dax_copy_from_iter,
1034 .dax_copy_to_iter = log_writes_dax_copy_to_iter,
1035 .dax_zero_page_range = log_writes_dax_zero_page_range,
1036};
1037
1038static int __init dm_log_writes_init(void)
1039{
1040 int r = dm_register_target(&log_writes_target);
1041
1042 if (r < 0)
1043 DMERR("register failed %d", r);
1044
1045 return r;
1046}
1047
1048static void __exit dm_log_writes_exit(void)
1049{
1050 dm_unregister_target(&log_writes_target);
1051}
1052
1053module_init(dm_log_writes_init);
1054module_exit(dm_log_writes_exit);
1055
1056MODULE_DESCRIPTION(DM_NAME " log writes target");
1057MODULE_AUTHOR("Josef Bacik <jbacik@fb.com>");
1058MODULE_LICENSE("GPL");