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