13bd94003SHeinz Mauelshagen // SPDX-License-Identifier: GPL-2.0-only
20e9cebe7SJosef Bacik /*
30e9cebe7SJosef Bacik * Copyright (C) 2014 Facebook. All rights reserved.
40e9cebe7SJosef Bacik *
50e9cebe7SJosef Bacik * This file is released under the GPL.
60e9cebe7SJosef Bacik */
70e9cebe7SJosef Bacik
80e9cebe7SJosef Bacik #include <linux/device-mapper.h>
90e9cebe7SJosef Bacik
100e9cebe7SJosef Bacik #include <linux/module.h>
110e9cebe7SJosef Bacik #include <linux/init.h>
120e9cebe7SJosef Bacik #include <linux/blkdev.h>
130e9cebe7SJosef Bacik #include <linux/bio.h>
1498d82f48SRoss Zwisler #include <linux/dax.h>
150e9cebe7SJosef Bacik #include <linux/slab.h>
160e9cebe7SJosef Bacik #include <linux/kthread.h>
170e9cebe7SJosef Bacik #include <linux/freezer.h>
1898d82f48SRoss Zwisler #include <linux/uio.h>
190e9cebe7SJosef Bacik
200e9cebe7SJosef Bacik #define DM_MSG_PREFIX "log-writes"
210e9cebe7SJosef Bacik
220e9cebe7SJosef Bacik /*
230e9cebe7SJosef Bacik * This target will sequentially log all writes to the target device onto the
240e9cebe7SJosef Bacik * log device. This is helpful for replaying writes to check for fs consistency
250e9cebe7SJosef Bacik * at all times. This target provides a mechanism to mark specific events to
260e9cebe7SJosef Bacik * check data at a later time. So for example you would:
270e9cebe7SJosef Bacik *
280e9cebe7SJosef Bacik * write data
290e9cebe7SJosef Bacik * fsync
300e9cebe7SJosef Bacik * dmsetup message /dev/whatever mark mymark
310e9cebe7SJosef Bacik * unmount /mnt/test
320e9cebe7SJosef Bacik *
330e9cebe7SJosef Bacik * Then replay the log up to mymark and check the contents of the replay to
340e9cebe7SJosef Bacik * verify it matches what was written.
350e9cebe7SJosef Bacik *
360e9cebe7SJosef Bacik * We log writes only after they have been flushed, this makes the log describe
370e9cebe7SJosef Bacik * close to the order in which the data hits the actual disk, not its cache. So
380e9cebe7SJosef Bacik * for example the following sequence (W means write, C means complete)
390e9cebe7SJosef Bacik *
400e9cebe7SJosef Bacik * Wa,Wb,Wc,Cc,Ca,FLUSH,FUAd,Cb,CFLUSH,CFUAd
410e9cebe7SJosef Bacik *
420e9cebe7SJosef Bacik * Would result in the log looking like this:
430e9cebe7SJosef Bacik *
447537dad7SQu Wenruo * c,a,b,flush,fuad,<other writes>,<next flush>
450e9cebe7SJosef Bacik *
460e9cebe7SJosef Bacik * This is meant to help expose problems where file systems do not properly wait
470e9cebe7SJosef Bacik * on data being written before invoking a FLUSH. FUA bypasses cache so once it
480e9cebe7SJosef Bacik * completes it is added to the log as it should be on disk.
490e9cebe7SJosef Bacik *
500e9cebe7SJosef Bacik * We treat DISCARDs as if they don't bypass cache so that they are logged in
510e9cebe7SJosef Bacik * order of completion along with the normal writes. If we didn't do it this
520e9cebe7SJosef Bacik * way we would process all the discards first and then write all the data, when
530e9cebe7SJosef Bacik * in fact we want to do the data and the discard in the order that they
540e9cebe7SJosef Bacik * completed.
550e9cebe7SJosef Bacik */
560e9cebe7SJosef Bacik #define LOG_FLUSH_FLAG (1 << 0)
570e9cebe7SJosef Bacik #define LOG_FUA_FLAG (1 << 1)
580e9cebe7SJosef Bacik #define LOG_DISCARD_FLAG (1 << 2)
590e9cebe7SJosef Bacik #define LOG_MARK_FLAG (1 << 3)
60e5c4cb9bSQu Wenruo #define LOG_METADATA_FLAG (1 << 4)
610e9cebe7SJosef Bacik
62f4ad317aSGeert Uytterhoeven #define WRITE_LOG_VERSION 1ULL
63f4ad317aSGeert Uytterhoeven #define WRITE_LOG_MAGIC 0x6a736677736872ULL
64211ad4b7Szhangyi (F) #define WRITE_LOG_SUPER_SECTOR 0
650e9cebe7SJosef Bacik
660e9cebe7SJosef Bacik /*
670e9cebe7SJosef Bacik * The disk format for this is braindead simple.
680e9cebe7SJosef Bacik *
690e9cebe7SJosef Bacik * At byte 0 we have our super, followed by the following sequence for
700e9cebe7SJosef Bacik * nr_entries:
710e9cebe7SJosef Bacik *
720e9cebe7SJosef Bacik * [ 1 sector ][ entry->nr_sectors ]
730e9cebe7SJosef Bacik * [log_write_entry][ data written ]
740e9cebe7SJosef Bacik *
750e9cebe7SJosef Bacik * The log_write_entry takes up a full sector so we can have arbitrary length
760e9cebe7SJosef Bacik * marks and it leaves us room for extra content in the future.
770e9cebe7SJosef Bacik */
780e9cebe7SJosef Bacik
790e9cebe7SJosef Bacik /*
800e9cebe7SJosef Bacik * Basic info about the log for userspace.
810e9cebe7SJosef Bacik */
820e9cebe7SJosef Bacik struct log_write_super {
830e9cebe7SJosef Bacik __le64 magic;
840e9cebe7SJosef Bacik __le64 version;
850e9cebe7SJosef Bacik __le64 nr_entries;
860e9cebe7SJosef Bacik __le32 sectorsize;
870e9cebe7SJosef Bacik };
880e9cebe7SJosef Bacik
890e9cebe7SJosef Bacik /*
900e9cebe7SJosef Bacik * sector - the sector we wrote.
910e9cebe7SJosef Bacik * nr_sectors - the number of sectors we wrote.
920e9cebe7SJosef Bacik * flags - flags for this log entry.
930e9cebe7SJosef Bacik * data_len - the size of the data in this log entry, this is for private log
940e9cebe7SJosef Bacik * entry stuff, the MARK data provided by userspace for example.
950e9cebe7SJosef Bacik */
960e9cebe7SJosef Bacik struct log_write_entry {
970e9cebe7SJosef Bacik __le64 sector;
980e9cebe7SJosef Bacik __le64 nr_sectors;
990e9cebe7SJosef Bacik __le64 flags;
1000e9cebe7SJosef Bacik __le64 data_len;
1010e9cebe7SJosef Bacik };
1020e9cebe7SJosef Bacik
1030e9cebe7SJosef Bacik struct log_writes_c {
1040e9cebe7SJosef Bacik struct dm_dev *dev;
1050e9cebe7SJosef Bacik struct dm_dev *logdev;
1060e9cebe7SJosef Bacik u64 logged_entries;
1070e9cebe7SJosef Bacik u32 sectorsize;
108228bb5b2SJosef Bacik u32 sectorshift;
1090e9cebe7SJosef Bacik atomic_t io_blocks;
1100e9cebe7SJosef Bacik atomic_t pending_blocks;
1110e9cebe7SJosef Bacik sector_t next_sector;
1120e9cebe7SJosef Bacik sector_t end_sector;
1130e9cebe7SJosef Bacik bool logging_enabled;
1140e9cebe7SJosef Bacik bool device_supports_discard;
1150e9cebe7SJosef Bacik spinlock_t blocks_lock;
1160e9cebe7SJosef Bacik struct list_head unflushed_blocks;
1170e9cebe7SJosef Bacik struct list_head logging_blocks;
1180e9cebe7SJosef Bacik wait_queue_head_t wait;
1190e9cebe7SJosef Bacik struct task_struct *log_kthread;
120211ad4b7Szhangyi (F) struct completion super_done;
1210e9cebe7SJosef Bacik };
1220e9cebe7SJosef Bacik
1230e9cebe7SJosef Bacik struct pending_block {
1240e9cebe7SJosef Bacik int vec_cnt;
1250e9cebe7SJosef Bacik u64 flags;
1260e9cebe7SJosef Bacik sector_t sector;
1270e9cebe7SJosef Bacik sector_t nr_sectors;
1280e9cebe7SJosef Bacik char *data;
1290e9cebe7SJosef Bacik u32 datalen;
1300e9cebe7SJosef Bacik struct list_head list;
131b18ae8ddSGustavo A. R. Silva struct bio_vec vecs[];
1320e9cebe7SJosef Bacik };
1330e9cebe7SJosef Bacik
1340e9cebe7SJosef Bacik struct per_bio_data {
1350e9cebe7SJosef Bacik struct pending_block *block;
1360e9cebe7SJosef Bacik };
1370e9cebe7SJosef Bacik
bio_to_dev_sectors(struct log_writes_c * lc,sector_t sectors)138228bb5b2SJosef Bacik static inline sector_t bio_to_dev_sectors(struct log_writes_c *lc,
139228bb5b2SJosef Bacik sector_t sectors)
140228bb5b2SJosef Bacik {
141228bb5b2SJosef Bacik return sectors >> (lc->sectorshift - SECTOR_SHIFT);
142228bb5b2SJosef Bacik }
143228bb5b2SJosef Bacik
dev_to_bio_sectors(struct log_writes_c * lc,sector_t sectors)144228bb5b2SJosef Bacik static inline sector_t dev_to_bio_sectors(struct log_writes_c *lc,
145228bb5b2SJosef Bacik sector_t sectors)
146228bb5b2SJosef Bacik {
147228bb5b2SJosef Bacik return sectors << (lc->sectorshift - SECTOR_SHIFT);
148228bb5b2SJosef Bacik }
149228bb5b2SJosef Bacik
put_pending_block(struct log_writes_c * lc)1500e9cebe7SJosef Bacik static void put_pending_block(struct log_writes_c *lc)
1510e9cebe7SJosef Bacik {
1520e9cebe7SJosef Bacik if (atomic_dec_and_test(&lc->pending_blocks)) {
1530e9cebe7SJosef Bacik smp_mb__after_atomic();
1540e9cebe7SJosef Bacik if (waitqueue_active(&lc->wait))
1550e9cebe7SJosef Bacik wake_up(&lc->wait);
1560e9cebe7SJosef Bacik }
1570e9cebe7SJosef Bacik }
1580e9cebe7SJosef Bacik
put_io_block(struct log_writes_c * lc)1590e9cebe7SJosef Bacik static void put_io_block(struct log_writes_c *lc)
1600e9cebe7SJosef Bacik {
1610e9cebe7SJosef Bacik if (atomic_dec_and_test(&lc->io_blocks)) {
1620e9cebe7SJosef Bacik smp_mb__after_atomic();
1630e9cebe7SJosef Bacik if (waitqueue_active(&lc->wait))
1640e9cebe7SJosef Bacik wake_up(&lc->wait);
1650e9cebe7SJosef Bacik }
1660e9cebe7SJosef Bacik }
1670e9cebe7SJosef Bacik
log_end_io(struct bio * bio)1684246a0b6SChristoph Hellwig static void log_end_io(struct bio *bio)
1690e9cebe7SJosef Bacik {
1700e9cebe7SJosef Bacik struct log_writes_c *lc = bio->bi_private;
1710e9cebe7SJosef Bacik
1724e4cbee9SChristoph Hellwig if (bio->bi_status) {
1730e9cebe7SJosef Bacik unsigned long flags;
1740e9cebe7SJosef Bacik
1754e4cbee9SChristoph Hellwig DMERR("Error writing log block, error=%d", bio->bi_status);
1760e9cebe7SJosef Bacik spin_lock_irqsave(&lc->blocks_lock, flags);
1770e9cebe7SJosef Bacik lc->logging_enabled = false;
1780e9cebe7SJosef Bacik spin_unlock_irqrestore(&lc->blocks_lock, flags);
1790e9cebe7SJosef Bacik }
1800e9cebe7SJosef Bacik
181491221f8SGuoqing Jiang bio_free_pages(bio);
1820e9cebe7SJosef Bacik put_io_block(lc);
1830e9cebe7SJosef Bacik bio_put(bio);
1840e9cebe7SJosef Bacik }
1850e9cebe7SJosef Bacik
log_end_super(struct bio * bio)186211ad4b7Szhangyi (F) static void log_end_super(struct bio *bio)
187211ad4b7Szhangyi (F) {
188211ad4b7Szhangyi (F) struct log_writes_c *lc = bio->bi_private;
189211ad4b7Szhangyi (F)
190211ad4b7Szhangyi (F) complete(&lc->super_done);
191211ad4b7Szhangyi (F) log_end_io(bio);
192211ad4b7Szhangyi (F) }
193211ad4b7Szhangyi (F)
1940e9cebe7SJosef Bacik /*
1950e9cebe7SJosef Bacik * Meant to be called if there is an error, it will free all the pages
1960e9cebe7SJosef Bacik * associated with the block.
1970e9cebe7SJosef Bacik */
free_pending_block(struct log_writes_c * lc,struct pending_block * block)1980e9cebe7SJosef Bacik static void free_pending_block(struct log_writes_c *lc,
1990e9cebe7SJosef Bacik struct pending_block *block)
2000e9cebe7SJosef Bacik {
2010e9cebe7SJosef Bacik int i;
2020e9cebe7SJosef Bacik
2030e9cebe7SJosef Bacik for (i = 0; i < block->vec_cnt; i++) {
2040e9cebe7SJosef Bacik if (block->vecs[i].bv_page)
2050e9cebe7SJosef Bacik __free_page(block->vecs[i].bv_page);
2060e9cebe7SJosef Bacik }
2070e9cebe7SJosef Bacik kfree(block->data);
2080e9cebe7SJosef Bacik kfree(block);
2090e9cebe7SJosef Bacik put_pending_block(lc);
2100e9cebe7SJosef Bacik }
2110e9cebe7SJosef Bacik
write_metadata(struct log_writes_c * lc,void * entry,size_t entrylen,void * data,size_t datalen,sector_t sector)2120e9cebe7SJosef Bacik static int write_metadata(struct log_writes_c *lc, void *entry,
2130e9cebe7SJosef Bacik size_t entrylen, void *data, size_t datalen,
2140e9cebe7SJosef Bacik sector_t sector)
2150e9cebe7SJosef Bacik {
2160e9cebe7SJosef Bacik struct bio *bio;
2170e9cebe7SJosef Bacik struct page *page;
2180e9cebe7SJosef Bacik void *ptr;
2190e9cebe7SJosef Bacik size_t ret;
2200e9cebe7SJosef Bacik
22107888c66SChristoph Hellwig bio = bio_alloc(lc->logdev->bdev, 1, REQ_OP_WRITE, GFP_KERNEL);
2220e9cebe7SJosef Bacik bio->bi_iter.bi_size = 0;
2230e9cebe7SJosef Bacik bio->bi_iter.bi_sector = sector;
224211ad4b7Szhangyi (F) bio->bi_end_io = (sector == WRITE_LOG_SUPER_SECTOR) ?
225211ad4b7Szhangyi (F) log_end_super : log_end_io;
2260e9cebe7SJosef Bacik bio->bi_private = lc;
2270e9cebe7SJosef Bacik
2280e9cebe7SJosef Bacik page = alloc_page(GFP_KERNEL);
2290e9cebe7SJosef Bacik if (!page) {
2300e9cebe7SJosef Bacik DMERR("Couldn't alloc log page");
2310e9cebe7SJosef Bacik bio_put(bio);
2320e9cebe7SJosef Bacik goto error;
2330e9cebe7SJosef Bacik }
2340e9cebe7SJosef Bacik
2350d78954aSHeinz Mauelshagen ptr = kmap_local_page(page);
2360e9cebe7SJosef Bacik memcpy(ptr, entry, entrylen);
2370e9cebe7SJosef Bacik if (datalen)
2380e9cebe7SJosef Bacik memcpy(ptr + entrylen, data, datalen);
2390e9cebe7SJosef Bacik memset(ptr + entrylen + datalen, 0,
2400e9cebe7SJosef Bacik lc->sectorsize - entrylen - datalen);
2410d78954aSHeinz Mauelshagen kunmap_local(ptr);
2420e9cebe7SJosef Bacik
2430e9cebe7SJosef Bacik ret = bio_add_page(bio, page, lc->sectorsize, 0);
2440e9cebe7SJosef Bacik if (ret != lc->sectorsize) {
2450e9cebe7SJosef Bacik DMERR("Couldn't add page to the log block");
2460e9cebe7SJosef Bacik goto error_bio;
2470e9cebe7SJosef Bacik }
2484e49ea4aSMike Christie submit_bio(bio);
2490e9cebe7SJosef Bacik return 0;
2500e9cebe7SJosef Bacik error_bio:
2510e9cebe7SJosef Bacik bio_put(bio);
2520e9cebe7SJosef Bacik __free_page(page);
2530e9cebe7SJosef Bacik error:
2540e9cebe7SJosef Bacik put_io_block(lc);
2550e9cebe7SJosef Bacik return -1;
2560e9cebe7SJosef Bacik }
2570e9cebe7SJosef Bacik
write_inline_data(struct log_writes_c * lc,void * entry,size_t entrylen,void * data,size_t datalen,sector_t sector)258e5a20660SRoss Zwisler static int write_inline_data(struct log_writes_c *lc, void *entry,
259e5a20660SRoss Zwisler size_t entrylen, void *data, size_t datalen,
260e5a20660SRoss Zwisler sector_t sector)
261e5a20660SRoss Zwisler {
2625f7136dbSMatthew Wilcox (Oracle) int bio_pages, pg_datalen, pg_sectorlen, i;
263e5a20660SRoss Zwisler struct page *page;
264e5a20660SRoss Zwisler struct bio *bio;
265e5a20660SRoss Zwisler size_t ret;
266e5a20660SRoss Zwisler void *ptr;
267e5a20660SRoss Zwisler
268e5a20660SRoss Zwisler while (datalen) {
2695f7136dbSMatthew Wilcox (Oracle) bio_pages = bio_max_segs(DIV_ROUND_UP(datalen, PAGE_SIZE));
270e5a20660SRoss Zwisler
271e5a20660SRoss Zwisler atomic_inc(&lc->io_blocks);
272e5a20660SRoss Zwisler
27307888c66SChristoph Hellwig bio = bio_alloc(lc->logdev->bdev, bio_pages, REQ_OP_WRITE,
27407888c66SChristoph Hellwig GFP_KERNEL);
275e5a20660SRoss Zwisler bio->bi_iter.bi_size = 0;
276e5a20660SRoss Zwisler bio->bi_iter.bi_sector = sector;
277e5a20660SRoss Zwisler bio->bi_end_io = log_end_io;
278e5a20660SRoss Zwisler bio->bi_private = lc;
279e5a20660SRoss Zwisler
280e5a20660SRoss Zwisler for (i = 0; i < bio_pages; i++) {
281e5a20660SRoss Zwisler pg_datalen = min_t(int, datalen, PAGE_SIZE);
282e5a20660SRoss Zwisler pg_sectorlen = ALIGN(pg_datalen, lc->sectorsize);
283e5a20660SRoss Zwisler
284e5a20660SRoss Zwisler page = alloc_page(GFP_KERNEL);
285e5a20660SRoss Zwisler if (!page) {
286e5a20660SRoss Zwisler DMERR("Couldn't alloc inline data page");
287e5a20660SRoss Zwisler goto error_bio;
288e5a20660SRoss Zwisler }
289e5a20660SRoss Zwisler
2900d78954aSHeinz Mauelshagen ptr = kmap_local_page(page);
291e5a20660SRoss Zwisler memcpy(ptr, data, pg_datalen);
292e5a20660SRoss Zwisler if (pg_sectorlen > pg_datalen)
293e5a20660SRoss Zwisler memset(ptr + pg_datalen, 0, pg_sectorlen - pg_datalen);
2940d78954aSHeinz Mauelshagen kunmap_local(ptr);
295e5a20660SRoss Zwisler
296e5a20660SRoss Zwisler ret = bio_add_page(bio, page, pg_sectorlen, 0);
297e5a20660SRoss Zwisler if (ret != pg_sectorlen) {
298e5a20660SRoss Zwisler DMERR("Couldn't add page of inline data");
299e5a20660SRoss Zwisler __free_page(page);
300e5a20660SRoss Zwisler goto error_bio;
301e5a20660SRoss Zwisler }
302e5a20660SRoss Zwisler
303e5a20660SRoss Zwisler datalen -= pg_datalen;
304e5a20660SRoss Zwisler data += pg_datalen;
305e5a20660SRoss Zwisler }
306e5a20660SRoss Zwisler submit_bio(bio);
307e5a20660SRoss Zwisler
308e5a20660SRoss Zwisler sector += bio_pages * PAGE_SECTORS;
309e5a20660SRoss Zwisler }
310e5a20660SRoss Zwisler return 0;
311e5a20660SRoss Zwisler error_bio:
312e5a20660SRoss Zwisler bio_free_pages(bio);
313e5a20660SRoss Zwisler bio_put(bio);
314e5a20660SRoss Zwisler put_io_block(lc);
315e5a20660SRoss Zwisler return -1;
316e5a20660SRoss Zwisler }
317e5a20660SRoss Zwisler
log_one_block(struct log_writes_c * lc,struct pending_block * block,sector_t sector)3180e9cebe7SJosef Bacik static int log_one_block(struct log_writes_c *lc,
3190e9cebe7SJosef Bacik struct pending_block *block, sector_t sector)
3200e9cebe7SJosef Bacik {
3210e9cebe7SJosef Bacik struct bio *bio;
3220e9cebe7SJosef Bacik struct log_write_entry entry;
323e5a20660SRoss Zwisler size_t metadatalen, ret;
3240e9cebe7SJosef Bacik int i;
3250e9cebe7SJosef Bacik
3260e9cebe7SJosef Bacik entry.sector = cpu_to_le64(block->sector);
3270e9cebe7SJosef Bacik entry.nr_sectors = cpu_to_le64(block->nr_sectors);
3280e9cebe7SJosef Bacik entry.flags = cpu_to_le64(block->flags);
3290e9cebe7SJosef Bacik entry.data_len = cpu_to_le64(block->datalen);
330e5a20660SRoss Zwisler
331e5a20660SRoss Zwisler metadatalen = (block->flags & LOG_MARK_FLAG) ? block->datalen : 0;
3320e9cebe7SJosef Bacik if (write_metadata(lc, &entry, sizeof(entry), block->data,
333e5a20660SRoss Zwisler metadatalen, sector)) {
3340e9cebe7SJosef Bacik free_pending_block(lc, block);
3350e9cebe7SJosef Bacik return -1;
3360e9cebe7SJosef Bacik }
3370e9cebe7SJosef Bacik
338e5a20660SRoss Zwisler sector += dev_to_bio_sectors(lc, 1);
339e5a20660SRoss Zwisler
340e5a20660SRoss Zwisler if (block->datalen && metadatalen == 0) {
341e5a20660SRoss Zwisler if (write_inline_data(lc, &entry, sizeof(entry), block->data,
342e5a20660SRoss Zwisler block->datalen, sector)) {
343e5a20660SRoss Zwisler free_pending_block(lc, block);
344e5a20660SRoss Zwisler return -1;
345e5a20660SRoss Zwisler }
346e5a20660SRoss Zwisler /* we don't support both inline data & bio data */
347e5a20660SRoss Zwisler goto out;
348e5a20660SRoss Zwisler }
349e5a20660SRoss Zwisler
3500e9cebe7SJosef Bacik if (!block->vec_cnt)
3510e9cebe7SJosef Bacik goto out;
3520e9cebe7SJosef Bacik
353a5d60783SMikulas Patocka atomic_inc(&lc->io_blocks);
35407888c66SChristoph Hellwig bio = bio_alloc(lc->logdev->bdev, bio_max_segs(block->vec_cnt),
35507888c66SChristoph Hellwig REQ_OP_WRITE, GFP_KERNEL);
3560e9cebe7SJosef Bacik bio->bi_iter.bi_size = 0;
3570e9cebe7SJosef Bacik bio->bi_iter.bi_sector = sector;
3580e9cebe7SJosef Bacik bio->bi_end_io = log_end_io;
3590e9cebe7SJosef Bacik bio->bi_private = lc;
3600e9cebe7SJosef Bacik
3610e9cebe7SJosef Bacik for (i = 0; i < block->vec_cnt; i++) {
3620e9cebe7SJosef Bacik /*
3630e9cebe7SJosef Bacik * The page offset is always 0 because we allocate a new page
3640e9cebe7SJosef Bacik * for every bvec in the original bio for simplicity sake.
3650e9cebe7SJosef Bacik */
3660e9cebe7SJosef Bacik ret = bio_add_page(bio, block->vecs[i].bv_page,
3670e9cebe7SJosef Bacik block->vecs[i].bv_len, 0);
3680e9cebe7SJosef Bacik if (ret != block->vecs[i].bv_len) {
3690e9cebe7SJosef Bacik atomic_inc(&lc->io_blocks);
3704e49ea4aSMike Christie submit_bio(bio);
37107888c66SChristoph Hellwig bio = bio_alloc(lc->logdev->bdev,
37207888c66SChristoph Hellwig bio_max_segs(block->vec_cnt - i),
37307888c66SChristoph Hellwig REQ_OP_WRITE, GFP_KERNEL);
3740e9cebe7SJosef Bacik bio->bi_iter.bi_size = 0;
3750e9cebe7SJosef Bacik bio->bi_iter.bi_sector = sector;
3760e9cebe7SJosef Bacik bio->bi_end_io = log_end_io;
3770e9cebe7SJosef Bacik bio->bi_private = lc;
3780e9cebe7SJosef Bacik
3790e9cebe7SJosef Bacik ret = bio_add_page(bio, block->vecs[i].bv_page,
3800e9cebe7SJosef Bacik block->vecs[i].bv_len, 0);
3810e9cebe7SJosef Bacik if (ret != block->vecs[i].bv_len) {
3820e9cebe7SJosef Bacik DMERR("Couldn't add page on new bio?");
3830e9cebe7SJosef Bacik bio_put(bio);
3840e9cebe7SJosef Bacik goto error;
3850e9cebe7SJosef Bacik }
3860e9cebe7SJosef Bacik }
3870e9cebe7SJosef Bacik sector += block->vecs[i].bv_len >> SECTOR_SHIFT;
3880e9cebe7SJosef Bacik }
3894e49ea4aSMike Christie submit_bio(bio);
3900e9cebe7SJosef Bacik out:
3910e9cebe7SJosef Bacik kfree(block->data);
3920e9cebe7SJosef Bacik kfree(block);
3930e9cebe7SJosef Bacik put_pending_block(lc);
3940e9cebe7SJosef Bacik return 0;
3950e9cebe7SJosef Bacik error:
3960e9cebe7SJosef Bacik free_pending_block(lc, block);
3970e9cebe7SJosef Bacik put_io_block(lc);
3980e9cebe7SJosef Bacik return -1;
3990e9cebe7SJosef Bacik }
4000e9cebe7SJosef Bacik
log_super(struct log_writes_c * lc)4010e9cebe7SJosef Bacik static int log_super(struct log_writes_c *lc)
4020e9cebe7SJosef Bacik {
4030e9cebe7SJosef Bacik struct log_write_super super;
4040e9cebe7SJosef Bacik
4050e9cebe7SJosef Bacik super.magic = cpu_to_le64(WRITE_LOG_MAGIC);
4060e9cebe7SJosef Bacik super.version = cpu_to_le64(WRITE_LOG_VERSION);
4070e9cebe7SJosef Bacik super.nr_entries = cpu_to_le64(lc->logged_entries);
4080e9cebe7SJosef Bacik super.sectorsize = cpu_to_le32(lc->sectorsize);
4090e9cebe7SJosef Bacik
410211ad4b7Szhangyi (F) if (write_metadata(lc, &super, sizeof(super), NULL, 0,
411211ad4b7Szhangyi (F) WRITE_LOG_SUPER_SECTOR)) {
4120e9cebe7SJosef Bacik DMERR("Couldn't write super");
4130e9cebe7SJosef Bacik return -1;
4140e9cebe7SJosef Bacik }
4150e9cebe7SJosef Bacik
416211ad4b7Szhangyi (F) /*
417211ad4b7Szhangyi (F) * Super sector should be writen in-order, otherwise the
418211ad4b7Szhangyi (F) * nr_entries could be rewritten incorrectly by an old bio.
419211ad4b7Szhangyi (F) */
420211ad4b7Szhangyi (F) wait_for_completion_io(&lc->super_done);
421211ad4b7Szhangyi (F)
4220e9cebe7SJosef Bacik return 0;
4230e9cebe7SJosef Bacik }
4240e9cebe7SJosef Bacik
logdev_last_sector(struct log_writes_c * lc)4250e9cebe7SJosef Bacik static inline sector_t logdev_last_sector(struct log_writes_c *lc)
4260e9cebe7SJosef Bacik {
4276dcbb52cSChristoph Hellwig return bdev_nr_sectors(lc->logdev->bdev);
4280e9cebe7SJosef Bacik }
4290e9cebe7SJosef Bacik
log_writes_kthread(void * arg)4300e9cebe7SJosef Bacik static int log_writes_kthread(void *arg)
4310e9cebe7SJosef Bacik {
43226cb62a2SYu Zhe struct log_writes_c *lc = arg;
4330e9cebe7SJosef Bacik sector_t sector = 0;
4340e9cebe7SJosef Bacik
4350e9cebe7SJosef Bacik while (!kthread_should_stop()) {
4360e9cebe7SJosef Bacik bool super = false;
4370e9cebe7SJosef Bacik bool logging_enabled;
4380e9cebe7SJosef Bacik struct pending_block *block = NULL;
4390e9cebe7SJosef Bacik int ret;
4400e9cebe7SJosef Bacik
4410e9cebe7SJosef Bacik spin_lock_irq(&lc->blocks_lock);
4420e9cebe7SJosef Bacik if (!list_empty(&lc->logging_blocks)) {
4430e9cebe7SJosef Bacik block = list_first_entry(&lc->logging_blocks,
4440e9cebe7SJosef Bacik struct pending_block, list);
4450e9cebe7SJosef Bacik list_del_init(&block->list);
4460e9cebe7SJosef Bacik if (!lc->logging_enabled)
4470e9cebe7SJosef Bacik goto next;
4480e9cebe7SJosef Bacik
4490e9cebe7SJosef Bacik sector = lc->next_sector;
450228bb5b2SJosef Bacik if (!(block->flags & LOG_DISCARD_FLAG))
451228bb5b2SJosef Bacik lc->next_sector += dev_to_bio_sectors(lc, block->nr_sectors);
452228bb5b2SJosef Bacik lc->next_sector += dev_to_bio_sectors(lc, 1);
4530e9cebe7SJosef Bacik
4540e9cebe7SJosef Bacik /*
4550e9cebe7SJosef Bacik * Apparently the size of the device may not be known
4560e9cebe7SJosef Bacik * right away, so handle this properly.
4570e9cebe7SJosef Bacik */
4580e9cebe7SJosef Bacik if (!lc->end_sector)
4590e9cebe7SJosef Bacik lc->end_sector = logdev_last_sector(lc);
4600e9cebe7SJosef Bacik if (lc->end_sector &&
4610e9cebe7SJosef Bacik lc->next_sector >= lc->end_sector) {
4620e9cebe7SJosef Bacik DMERR("Ran out of space on the logdev");
4630e9cebe7SJosef Bacik lc->logging_enabled = false;
4640e9cebe7SJosef Bacik goto next;
4650e9cebe7SJosef Bacik }
4660e9cebe7SJosef Bacik lc->logged_entries++;
4670e9cebe7SJosef Bacik atomic_inc(&lc->io_blocks);
4680e9cebe7SJosef Bacik
4690e9cebe7SJosef Bacik super = (block->flags & (LOG_FUA_FLAG | LOG_MARK_FLAG));
4700e9cebe7SJosef Bacik if (super)
4710e9cebe7SJosef Bacik atomic_inc(&lc->io_blocks);
4720e9cebe7SJosef Bacik }
4730e9cebe7SJosef Bacik next:
4740e9cebe7SJosef Bacik logging_enabled = lc->logging_enabled;
4750e9cebe7SJosef Bacik spin_unlock_irq(&lc->blocks_lock);
4760e9cebe7SJosef Bacik if (block) {
4770e9cebe7SJosef Bacik if (logging_enabled) {
4780e9cebe7SJosef Bacik ret = log_one_block(lc, block, sector);
4790e9cebe7SJosef Bacik if (!ret && super)
4800e9cebe7SJosef Bacik ret = log_super(lc);
4810e9cebe7SJosef Bacik if (ret) {
4820e9cebe7SJosef Bacik spin_lock_irq(&lc->blocks_lock);
4830e9cebe7SJosef Bacik lc->logging_enabled = false;
4840e9cebe7SJosef Bacik spin_unlock_irq(&lc->blocks_lock);
4850e9cebe7SJosef Bacik }
4860e9cebe7SJosef Bacik } else
4870e9cebe7SJosef Bacik free_pending_block(lc, block);
4880e9cebe7SJosef Bacik continue;
4890e9cebe7SJosef Bacik }
4900e9cebe7SJosef Bacik
4910e9cebe7SJosef Bacik if (!try_to_freeze()) {
4920e9cebe7SJosef Bacik set_current_state(TASK_INTERRUPTIBLE);
4930e9cebe7SJosef Bacik if (!kthread_should_stop() &&
4940c79c620SJosef Bacik list_empty(&lc->logging_blocks))
4950e9cebe7SJosef Bacik schedule();
4960e9cebe7SJosef Bacik __set_current_state(TASK_RUNNING);
4970e9cebe7SJosef Bacik }
4980e9cebe7SJosef Bacik }
4990e9cebe7SJosef Bacik return 0;
5000e9cebe7SJosef Bacik }
5010e9cebe7SJosef Bacik
5020e9cebe7SJosef Bacik /*
5030e9cebe7SJosef Bacik * Construct a log-writes mapping:
5040e9cebe7SJosef Bacik * log-writes <dev_path> <log_dev_path>
5050e9cebe7SJosef Bacik */
log_writes_ctr(struct dm_target * ti,unsigned int argc,char ** argv)5060e9cebe7SJosef Bacik static int log_writes_ctr(struct dm_target *ti, unsigned int argc, char **argv)
5070e9cebe7SJosef Bacik {
5080e9cebe7SJosef Bacik struct log_writes_c *lc;
5090e9cebe7SJosef Bacik struct dm_arg_set as;
5100e9cebe7SJosef Bacik const char *devname, *logdevname;
511e80d1c80SVivek Goyal int ret;
5120e9cebe7SJosef Bacik
5130e9cebe7SJosef Bacik as.argc = argc;
5140e9cebe7SJosef Bacik as.argv = argv;
5150e9cebe7SJosef Bacik
5160e9cebe7SJosef Bacik if (argc < 2) {
5170e9cebe7SJosef Bacik ti->error = "Invalid argument count";
5180e9cebe7SJosef Bacik return -EINVAL;
5190e9cebe7SJosef Bacik }
5200e9cebe7SJosef Bacik
5210e9cebe7SJosef Bacik lc = kzalloc(sizeof(struct log_writes_c), GFP_KERNEL);
5220e9cebe7SJosef Bacik if (!lc) {
5230e9cebe7SJosef Bacik ti->error = "Cannot allocate context";
5240e9cebe7SJosef Bacik return -ENOMEM;
5250e9cebe7SJosef Bacik }
5260e9cebe7SJosef Bacik spin_lock_init(&lc->blocks_lock);
5270e9cebe7SJosef Bacik INIT_LIST_HEAD(&lc->unflushed_blocks);
5280e9cebe7SJosef Bacik INIT_LIST_HEAD(&lc->logging_blocks);
5290e9cebe7SJosef Bacik init_waitqueue_head(&lc->wait);
530211ad4b7Szhangyi (F) init_completion(&lc->super_done);
5310e9cebe7SJosef Bacik atomic_set(&lc->io_blocks, 0);
5320e9cebe7SJosef Bacik atomic_set(&lc->pending_blocks, 0);
5330e9cebe7SJosef Bacik
5340e9cebe7SJosef Bacik devname = dm_shift_arg(&as);
535e80d1c80SVivek Goyal ret = dm_get_device(ti, devname, dm_table_get_mode(ti->table), &lc->dev);
536e80d1c80SVivek Goyal if (ret) {
5370e9cebe7SJosef Bacik ti->error = "Device lookup failed";
5380e9cebe7SJosef Bacik goto bad;
5390e9cebe7SJosef Bacik }
5400e9cebe7SJosef Bacik
5410e9cebe7SJosef Bacik logdevname = dm_shift_arg(&as);
542e80d1c80SVivek Goyal ret = dm_get_device(ti, logdevname, dm_table_get_mode(ti->table),
543e80d1c80SVivek Goyal &lc->logdev);
544e80d1c80SVivek Goyal if (ret) {
5450e9cebe7SJosef Bacik ti->error = "Log device lookup failed";
5460e9cebe7SJosef Bacik dm_put_device(ti, lc->dev);
5470e9cebe7SJosef Bacik goto bad;
5480e9cebe7SJosef Bacik }
5490e9cebe7SJosef Bacik
550228bb5b2SJosef Bacik lc->sectorsize = bdev_logical_block_size(lc->dev->bdev);
551228bb5b2SJosef Bacik lc->sectorshift = ilog2(lc->sectorsize);
5520e9cebe7SJosef Bacik lc->log_kthread = kthread_run(log_writes_kthread, lc, "log-write");
55391e630d9SVladimir Zapolskiy if (IS_ERR(lc->log_kthread)) {
55491e630d9SVladimir Zapolskiy ret = PTR_ERR(lc->log_kthread);
5550e9cebe7SJosef Bacik ti->error = "Couldn't alloc kthread";
5560e9cebe7SJosef Bacik dm_put_device(ti, lc->dev);
5570e9cebe7SJosef Bacik dm_put_device(ti, lc->logdev);
5580e9cebe7SJosef Bacik goto bad;
5590e9cebe7SJosef Bacik }
5600e9cebe7SJosef Bacik
561228bb5b2SJosef Bacik /*
562228bb5b2SJosef Bacik * next_sector is in 512b sectors to correspond to what bi_sector expects.
563228bb5b2SJosef Bacik * The super starts at sector 0, and the next_sector is the next logical
564228bb5b2SJosef Bacik * one based on the sectorsize of the device.
565228bb5b2SJosef Bacik */
566228bb5b2SJosef Bacik lc->next_sector = lc->sectorsize >> SECTOR_SHIFT;
5670e9cebe7SJosef Bacik lc->logging_enabled = true;
5680e9cebe7SJosef Bacik lc->end_sector = logdev_last_sector(lc);
5690e9cebe7SJosef Bacik lc->device_supports_discard = true;
5700e9cebe7SJosef Bacik
5710e9cebe7SJosef Bacik ti->num_flush_bios = 1;
5720e9cebe7SJosef Bacik ti->flush_supported = true;
5730e9cebe7SJosef Bacik ti->num_discard_bios = 1;
5740e9cebe7SJosef Bacik ti->discards_supported = true;
57530187e1dSMike Snitzer ti->per_io_data_size = sizeof(struct per_bio_data);
5760e9cebe7SJosef Bacik ti->private = lc;
5770e9cebe7SJosef Bacik return 0;
5780e9cebe7SJosef Bacik
5790e9cebe7SJosef Bacik bad:
5800e9cebe7SJosef Bacik kfree(lc);
581e80d1c80SVivek Goyal return ret;
5820e9cebe7SJosef Bacik }
5830e9cebe7SJosef Bacik
log_mark(struct log_writes_c * lc,char * data)5840e9cebe7SJosef Bacik static int log_mark(struct log_writes_c *lc, char *data)
5850e9cebe7SJosef Bacik {
5860e9cebe7SJosef Bacik struct pending_block *block;
5870e9cebe7SJosef Bacik size_t maxsize = lc->sectorsize - sizeof(struct log_write_entry);
5880e9cebe7SJosef Bacik
5890e9cebe7SJosef Bacik block = kzalloc(sizeof(struct pending_block), GFP_KERNEL);
5900e9cebe7SJosef Bacik if (!block) {
5910e9cebe7SJosef Bacik DMERR("Error allocating pending block");
5920e9cebe7SJosef Bacik return -ENOMEM;
5930e9cebe7SJosef Bacik }
5940e9cebe7SJosef Bacik
5954b259fc4SMa Shimiao block->data = kstrndup(data, maxsize - 1, GFP_KERNEL);
5960e9cebe7SJosef Bacik if (!block->data) {
5970e9cebe7SJosef Bacik DMERR("Error copying mark data");
5980e9cebe7SJosef Bacik kfree(block);
5990e9cebe7SJosef Bacik return -ENOMEM;
6000e9cebe7SJosef Bacik }
6010e9cebe7SJosef Bacik atomic_inc(&lc->pending_blocks);
6020e9cebe7SJosef Bacik block->datalen = strlen(block->data);
6030e9cebe7SJosef Bacik block->flags |= LOG_MARK_FLAG;
6040e9cebe7SJosef Bacik spin_lock_irq(&lc->blocks_lock);
6050e9cebe7SJosef Bacik list_add_tail(&block->list, &lc->logging_blocks);
6060e9cebe7SJosef Bacik spin_unlock_irq(&lc->blocks_lock);
6070e9cebe7SJosef Bacik wake_up_process(lc->log_kthread);
6080e9cebe7SJosef Bacik return 0;
6090e9cebe7SJosef Bacik }
6100e9cebe7SJosef Bacik
log_writes_dtr(struct dm_target * ti)6110e9cebe7SJosef Bacik static void log_writes_dtr(struct dm_target *ti)
6120e9cebe7SJosef Bacik {
6130e9cebe7SJosef Bacik struct log_writes_c *lc = ti->private;
6140e9cebe7SJosef Bacik
6150e9cebe7SJosef Bacik spin_lock_irq(&lc->blocks_lock);
6160e9cebe7SJosef Bacik list_splice_init(&lc->unflushed_blocks, &lc->logging_blocks);
6170e9cebe7SJosef Bacik spin_unlock_irq(&lc->blocks_lock);
6180e9cebe7SJosef Bacik
6190e9cebe7SJosef Bacik /*
6200e9cebe7SJosef Bacik * This is just nice to have since it'll update the super to include the
6210e9cebe7SJosef Bacik * unflushed blocks, if it fails we don't really care.
6220e9cebe7SJosef Bacik */
6230e9cebe7SJosef Bacik log_mark(lc, "dm-log-writes-end");
6240e9cebe7SJosef Bacik wake_up_process(lc->log_kthread);
6250e9cebe7SJosef Bacik wait_event(lc->wait, !atomic_read(&lc->io_blocks) &&
6260e9cebe7SJosef Bacik !atomic_read(&lc->pending_blocks));
6270e9cebe7SJosef Bacik kthread_stop(lc->log_kthread);
6280e9cebe7SJosef Bacik
6290e9cebe7SJosef Bacik WARN_ON(!list_empty(&lc->logging_blocks));
6300e9cebe7SJosef Bacik WARN_ON(!list_empty(&lc->unflushed_blocks));
6310e9cebe7SJosef Bacik dm_put_device(ti, lc->dev);
6320e9cebe7SJosef Bacik dm_put_device(ti, lc->logdev);
6330e9cebe7SJosef Bacik kfree(lc);
6340e9cebe7SJosef Bacik }
6350e9cebe7SJosef Bacik
normal_map_bio(struct dm_target * ti,struct bio * bio)6360e9cebe7SJosef Bacik static void normal_map_bio(struct dm_target *ti, struct bio *bio)
6370e9cebe7SJosef Bacik {
6380e9cebe7SJosef Bacik struct log_writes_c *lc = ti->private;
6390e9cebe7SJosef Bacik
64074d46992SChristoph Hellwig bio_set_dev(bio, lc->dev->bdev);
6410e9cebe7SJosef Bacik }
6420e9cebe7SJosef Bacik
log_writes_map(struct dm_target * ti,struct bio * bio)6430e9cebe7SJosef Bacik static int log_writes_map(struct dm_target *ti, struct bio *bio)
6440e9cebe7SJosef Bacik {
6450e9cebe7SJosef Bacik struct log_writes_c *lc = ti->private;
6460e9cebe7SJosef Bacik struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
6470e9cebe7SJosef Bacik struct pending_block *block;
6480e9cebe7SJosef Bacik struct bvec_iter iter;
6490e9cebe7SJosef Bacik struct bio_vec bv;
6500e9cebe7SJosef Bacik size_t alloc_size;
6510e9cebe7SJosef Bacik int i = 0;
6521eff9d32SJens Axboe bool flush_bio = (bio->bi_opf & REQ_PREFLUSH);
6531eff9d32SJens Axboe bool fua_bio = (bio->bi_opf & REQ_FUA);
654e6047149SMike Christie bool discard_bio = (bio_op(bio) == REQ_OP_DISCARD);
655e5c4cb9bSQu Wenruo bool meta_bio = (bio->bi_opf & REQ_META);
6560e9cebe7SJosef Bacik
6570e9cebe7SJosef Bacik pb->block = NULL;
6580e9cebe7SJosef Bacik
6590e9cebe7SJosef Bacik /* Don't bother doing anything if logging has been disabled */
6600e9cebe7SJosef Bacik if (!lc->logging_enabled)
6610e9cebe7SJosef Bacik goto map_bio;
6620e9cebe7SJosef Bacik
6630e9cebe7SJosef Bacik /*
6640e9cebe7SJosef Bacik * Map reads as normal.
6650e9cebe7SJosef Bacik */
6660e9cebe7SJosef Bacik if (bio_data_dir(bio) == READ)
6670e9cebe7SJosef Bacik goto map_bio;
6680e9cebe7SJosef Bacik
6690e9cebe7SJosef Bacik /* No sectors and not a flush? Don't care */
6700e9cebe7SJosef Bacik if (!bio_sectors(bio) && !flush_bio)
6710e9cebe7SJosef Bacik goto map_bio;
6720e9cebe7SJosef Bacik
6730e9cebe7SJosef Bacik /*
6740e9cebe7SJosef Bacik * Discards will have bi_size set but there's no actual data, so just
6750e9cebe7SJosef Bacik * allocate the size of the pending block.
6760e9cebe7SJosef Bacik */
6770e9cebe7SJosef Bacik if (discard_bio)
6780e9cebe7SJosef Bacik alloc_size = sizeof(struct pending_block);
6790e9cebe7SJosef Bacik else
680d4e6e836SZhengyuan Liu alloc_size = struct_size(block, vecs, bio_segments(bio));
6810e9cebe7SJosef Bacik
6820e9cebe7SJosef Bacik block = kzalloc(alloc_size, GFP_NOIO);
6830e9cebe7SJosef Bacik if (!block) {
6840e9cebe7SJosef Bacik DMERR("Error allocating pending block");
6850e9cebe7SJosef Bacik spin_lock_irq(&lc->blocks_lock);
6860e9cebe7SJosef Bacik lc->logging_enabled = false;
6870e9cebe7SJosef Bacik spin_unlock_irq(&lc->blocks_lock);
688846785e6SChristoph Hellwig return DM_MAPIO_KILL;
6890e9cebe7SJosef Bacik }
6900e9cebe7SJosef Bacik INIT_LIST_HEAD(&block->list);
6910e9cebe7SJosef Bacik pb->block = block;
6920e9cebe7SJosef Bacik atomic_inc(&lc->pending_blocks);
6930e9cebe7SJosef Bacik
6940e9cebe7SJosef Bacik if (flush_bio)
6950e9cebe7SJosef Bacik block->flags |= LOG_FLUSH_FLAG;
6960e9cebe7SJosef Bacik if (fua_bio)
6970e9cebe7SJosef Bacik block->flags |= LOG_FUA_FLAG;
6980e9cebe7SJosef Bacik if (discard_bio)
6990e9cebe7SJosef Bacik block->flags |= LOG_DISCARD_FLAG;
700e5c4cb9bSQu Wenruo if (meta_bio)
701e5c4cb9bSQu Wenruo block->flags |= LOG_METADATA_FLAG;
7020e9cebe7SJosef Bacik
703228bb5b2SJosef Bacik block->sector = bio_to_dev_sectors(lc, bio->bi_iter.bi_sector);
704228bb5b2SJosef Bacik block->nr_sectors = bio_to_dev_sectors(lc, bio_sectors(bio));
7050e9cebe7SJosef Bacik
7060e9cebe7SJosef Bacik /* We don't need the data, just submit */
7070e9cebe7SJosef Bacik if (discard_bio) {
7080e9cebe7SJosef Bacik WARN_ON(flush_bio || fua_bio);
7090e9cebe7SJosef Bacik if (lc->device_supports_discard)
7100e9cebe7SJosef Bacik goto map_bio;
7114246a0b6SChristoph Hellwig bio_endio(bio);
7120e9cebe7SJosef Bacik return DM_MAPIO_SUBMITTED;
7130e9cebe7SJosef Bacik }
7140e9cebe7SJosef Bacik
7150e9cebe7SJosef Bacik /* Flush bio, splice the unflushed blocks onto this list and submit */
7160e9cebe7SJosef Bacik if (flush_bio && !bio_sectors(bio)) {
7170e9cebe7SJosef Bacik spin_lock_irq(&lc->blocks_lock);
7180e9cebe7SJosef Bacik list_splice_init(&lc->unflushed_blocks, &block->list);
7190e9cebe7SJosef Bacik spin_unlock_irq(&lc->blocks_lock);
7200e9cebe7SJosef Bacik goto map_bio;
7210e9cebe7SJosef Bacik }
7220e9cebe7SJosef Bacik
7230e9cebe7SJosef Bacik /*
7240e9cebe7SJosef Bacik * We will write this bio somewhere else way later so we need to copy
7250e9cebe7SJosef Bacik * the actual contents into new pages so we know the data will always be
7260e9cebe7SJosef Bacik * there.
7270e9cebe7SJosef Bacik *
7280e9cebe7SJosef Bacik * We do this because this could be a bio from O_DIRECT in which case we
7290e9cebe7SJosef Bacik * can't just hold onto the page until some later point, we have to
7300e9cebe7SJosef Bacik * manually copy the contents.
7310e9cebe7SJosef Bacik */
7320e9cebe7SJosef Bacik bio_for_each_segment(bv, bio, iter) {
7330e9cebe7SJosef Bacik struct page *page;
73427db2717SChristoph Hellwig void *dst;
7350e9cebe7SJosef Bacik
7360e9cebe7SJosef Bacik page = alloc_page(GFP_NOIO);
7370e9cebe7SJosef Bacik if (!page) {
7380e9cebe7SJosef Bacik DMERR("Error allocing page");
7390e9cebe7SJosef Bacik free_pending_block(lc, block);
7400e9cebe7SJosef Bacik spin_lock_irq(&lc->blocks_lock);
7410e9cebe7SJosef Bacik lc->logging_enabled = false;
7420e9cebe7SJosef Bacik spin_unlock_irq(&lc->blocks_lock);
743846785e6SChristoph Hellwig return DM_MAPIO_KILL;
7440e9cebe7SJosef Bacik }
7450e9cebe7SJosef Bacik
7460d78954aSHeinz Mauelshagen dst = kmap_local_page(page);
74727db2717SChristoph Hellwig memcpy_from_bvec(dst, &bv);
7480d78954aSHeinz Mauelshagen kunmap_local(dst);
7490e9cebe7SJosef Bacik block->vecs[i].bv_page = page;
7500e9cebe7SJosef Bacik block->vecs[i].bv_len = bv.bv_len;
7510e9cebe7SJosef Bacik block->vec_cnt++;
7520e9cebe7SJosef Bacik i++;
7530e9cebe7SJosef Bacik }
7540e9cebe7SJosef Bacik
7550e9cebe7SJosef Bacik /* Had a flush with data in it, weird */
7560e9cebe7SJosef Bacik if (flush_bio) {
7570e9cebe7SJosef Bacik spin_lock_irq(&lc->blocks_lock);
7580e9cebe7SJosef Bacik list_splice_init(&lc->unflushed_blocks, &block->list);
7590e9cebe7SJosef Bacik spin_unlock_irq(&lc->blocks_lock);
7600e9cebe7SJosef Bacik }
7610e9cebe7SJosef Bacik map_bio:
7620e9cebe7SJosef Bacik normal_map_bio(ti, bio);
7630e9cebe7SJosef Bacik return DM_MAPIO_REMAPPED;
7640e9cebe7SJosef Bacik }
7650e9cebe7SJosef Bacik
normal_end_io(struct dm_target * ti,struct bio * bio,blk_status_t * error)7664e4cbee9SChristoph Hellwig static int normal_end_io(struct dm_target *ti, struct bio *bio,
7674e4cbee9SChristoph Hellwig blk_status_t *error)
7680e9cebe7SJosef Bacik {
7690e9cebe7SJosef Bacik struct log_writes_c *lc = ti->private;
7700e9cebe7SJosef Bacik struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
7710e9cebe7SJosef Bacik
7720e9cebe7SJosef Bacik if (bio_data_dir(bio) == WRITE && pb->block) {
7730e9cebe7SJosef Bacik struct pending_block *block = pb->block;
7740e9cebe7SJosef Bacik unsigned long flags;
7750e9cebe7SJosef Bacik
7760e9cebe7SJosef Bacik spin_lock_irqsave(&lc->blocks_lock, flags);
7770e9cebe7SJosef Bacik if (block->flags & LOG_FLUSH_FLAG) {
7780e9cebe7SJosef Bacik list_splice_tail_init(&block->list, &lc->logging_blocks);
7790e9cebe7SJosef Bacik list_add_tail(&block->list, &lc->logging_blocks);
7800e9cebe7SJosef Bacik wake_up_process(lc->log_kthread);
7810e9cebe7SJosef Bacik } else if (block->flags & LOG_FUA_FLAG) {
7820e9cebe7SJosef Bacik list_add_tail(&block->list, &lc->logging_blocks);
7830e9cebe7SJosef Bacik wake_up_process(lc->log_kthread);
7840e9cebe7SJosef Bacik } else
7850e9cebe7SJosef Bacik list_add_tail(&block->list, &lc->unflushed_blocks);
7860e9cebe7SJosef Bacik spin_unlock_irqrestore(&lc->blocks_lock, flags);
7870e9cebe7SJosef Bacik }
7880e9cebe7SJosef Bacik
7891be56909SChristoph Hellwig return DM_ENDIO_DONE;
7900e9cebe7SJosef Bacik }
7910e9cebe7SJosef Bacik
7920e9cebe7SJosef Bacik /*
7930e9cebe7SJosef Bacik * INFO format: <logged entries> <highest allocated sector>
7940e9cebe7SJosef Bacik */
log_writes_status(struct dm_target * ti,status_type_t type,unsigned int status_flags,char * result,unsigned int maxlen)7950e9cebe7SJosef Bacik static void log_writes_status(struct dm_target *ti, status_type_t type,
79686a3238cSHeinz Mauelshagen unsigned int status_flags, char *result,
79786a3238cSHeinz Mauelshagen unsigned int maxlen)
7980e9cebe7SJosef Bacik {
79986a3238cSHeinz Mauelshagen unsigned int sz = 0;
8000e9cebe7SJosef Bacik struct log_writes_c *lc = ti->private;
8010e9cebe7SJosef Bacik
8020e9cebe7SJosef Bacik switch (type) {
8030e9cebe7SJosef Bacik case STATUSTYPE_INFO:
8040e9cebe7SJosef Bacik DMEMIT("%llu %llu", lc->logged_entries,
8050e9cebe7SJosef Bacik (unsigned long long)lc->next_sector - 1);
8060e9cebe7SJosef Bacik if (!lc->logging_enabled)
8070e9cebe7SJosef Bacik DMEMIT(" logging_disabled");
8080e9cebe7SJosef Bacik break;
8090e9cebe7SJosef Bacik
8100e9cebe7SJosef Bacik case STATUSTYPE_TABLE:
8110e9cebe7SJosef Bacik DMEMIT("%s %s", lc->dev->name, lc->logdev->name);
8120e9cebe7SJosef Bacik break;
8138ec45662STushar Sugandhi
8148ec45662STushar Sugandhi case STATUSTYPE_IMA:
8158ec45662STushar Sugandhi *result = '\0';
8168ec45662STushar Sugandhi break;
8170e9cebe7SJosef Bacik }
8180e9cebe7SJosef Bacik }
8190e9cebe7SJosef Bacik
log_writes_prepare_ioctl(struct dm_target * ti,struct block_device ** bdev)820e56f81e0SChristoph Hellwig static int log_writes_prepare_ioctl(struct dm_target *ti,
8215bd5e8d8SMike Snitzer struct block_device **bdev)
8220e9cebe7SJosef Bacik {
8230e9cebe7SJosef Bacik struct log_writes_c *lc = ti->private;
8240e9cebe7SJosef Bacik struct dm_dev *dev = lc->dev;
8250e9cebe7SJosef Bacik
826e56f81e0SChristoph Hellwig *bdev = dev->bdev;
8270e9cebe7SJosef Bacik /*
8280e9cebe7SJosef Bacik * Only pass ioctls through if the device sizes match exactly.
8290e9cebe7SJosef Bacik */
8306dcbb52cSChristoph Hellwig if (ti->len != bdev_nr_sectors(dev->bdev))
831e56f81e0SChristoph Hellwig return 1;
832e56f81e0SChristoph Hellwig return 0;
8330e9cebe7SJosef Bacik }
8340e9cebe7SJosef Bacik
log_writes_iterate_devices(struct dm_target * ti,iterate_devices_callout_fn fn,void * data)8350e9cebe7SJosef Bacik static int log_writes_iterate_devices(struct dm_target *ti,
8360e9cebe7SJosef Bacik iterate_devices_callout_fn fn,
8370e9cebe7SJosef Bacik void *data)
8380e9cebe7SJosef Bacik {
8390e9cebe7SJosef Bacik struct log_writes_c *lc = ti->private;
8400e9cebe7SJosef Bacik
8410e9cebe7SJosef Bacik return fn(ti, lc->dev, 0, ti->len, data);
8420e9cebe7SJosef Bacik }
8430e9cebe7SJosef Bacik
8440e9cebe7SJosef Bacik /*
8450e9cebe7SJosef Bacik * Messages supported:
8460e9cebe7SJosef Bacik * mark <mark data> - specify the marked data.
8470e9cebe7SJosef Bacik */
log_writes_message(struct dm_target * ti,unsigned int argc,char ** argv,char * result,unsigned int maxlen)84886a3238cSHeinz Mauelshagen static int log_writes_message(struct dm_target *ti, unsigned int argc, char **argv,
84986a3238cSHeinz Mauelshagen char *result, unsigned int maxlen)
8500e9cebe7SJosef Bacik {
8510e9cebe7SJosef Bacik int r = -EINVAL;
8520e9cebe7SJosef Bacik struct log_writes_c *lc = ti->private;
8530e9cebe7SJosef Bacik
8540e9cebe7SJosef Bacik if (argc != 2) {
8550e9cebe7SJosef Bacik DMWARN("Invalid log-writes message arguments, expect 2 arguments, got %d", argc);
8560e9cebe7SJosef Bacik return r;
8570e9cebe7SJosef Bacik }
8580e9cebe7SJosef Bacik
8590e9cebe7SJosef Bacik if (!strcasecmp(argv[0], "mark"))
8600e9cebe7SJosef Bacik r = log_mark(lc, argv[1]);
8610e9cebe7SJosef Bacik else
8620e9cebe7SJosef Bacik DMWARN("Unrecognised log writes target message received: %s", argv[0]);
8630e9cebe7SJosef Bacik
8640e9cebe7SJosef Bacik return r;
8650e9cebe7SJosef Bacik }
8660e9cebe7SJosef Bacik
log_writes_io_hints(struct dm_target * ti,struct queue_limits * limits)8670e9cebe7SJosef Bacik static void log_writes_io_hints(struct dm_target *ti, struct queue_limits *limits)
8680e9cebe7SJosef Bacik {
8690e9cebe7SJosef Bacik struct log_writes_c *lc = ti->private;
8700e9cebe7SJosef Bacik
87170200574SChristoph Hellwig if (!bdev_max_discard_sectors(lc->dev->bdev)) {
8720e9cebe7SJosef Bacik lc->device_supports_discard = false;
873228bb5b2SJosef Bacik limits->discard_granularity = lc->sectorsize;
8740e9cebe7SJosef Bacik limits->max_discard_sectors = (UINT_MAX >> SECTOR_SHIFT);
8750e9cebe7SJosef Bacik }
876228bb5b2SJosef Bacik limits->logical_block_size = bdev_logical_block_size(lc->dev->bdev);
877228bb5b2SJosef Bacik limits->physical_block_size = bdev_physical_block_size(lc->dev->bdev);
878228bb5b2SJosef Bacik limits->io_min = limits->physical_block_size;
87950a89335SKeith Busch limits->dma_alignment = limits->logical_block_size - 1;
8800e9cebe7SJosef Bacik }
8810e9cebe7SJosef Bacik
8825d2a228bSChristoph Hellwig #if IS_ENABLED(CONFIG_FS_DAX)
log_writes_dax_pgoff(struct dm_target * ti,pgoff_t * pgoff)883d19bd675SChristoph Hellwig static struct dax_device *log_writes_dax_pgoff(struct dm_target *ti,
884d19bd675SChristoph Hellwig pgoff_t *pgoff)
885d19bd675SChristoph Hellwig {
886d19bd675SChristoph Hellwig struct log_writes_c *lc = ti->private;
887d19bd675SChristoph Hellwig
888d19bd675SChristoph Hellwig *pgoff += (get_start_sect(lc->dev->bdev) >> PAGE_SECTORS_SHIFT);
889d19bd675SChristoph Hellwig return lc->dev->dax_dev;
890d19bd675SChristoph Hellwig }
891d19bd675SChristoph Hellwig
log_writes_dax_direct_access(struct dm_target * ti,pgoff_t pgoff,long nr_pages,enum dax_access_mode mode,void ** kaddr,pfn_t * pfn)89298d82f48SRoss Zwisler static long log_writes_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
893e511c4a3SJane Chu long nr_pages, enum dax_access_mode mode, void **kaddr,
894e511c4a3SJane Chu pfn_t *pfn)
89598d82f48SRoss Zwisler {
896d19bd675SChristoph Hellwig struct dax_device *dax_dev = log_writes_dax_pgoff(ti, &pgoff);
89798d82f48SRoss Zwisler
898e511c4a3SJane Chu return dax_direct_access(dax_dev, pgoff, nr_pages, mode, kaddr, pfn);
89998d82f48SRoss Zwisler }
90098d82f48SRoss Zwisler
log_writes_dax_zero_page_range(struct dm_target * ti,pgoff_t pgoff,size_t nr_pages)901cdf6cdcdSVivek Goyal static int log_writes_dax_zero_page_range(struct dm_target *ti, pgoff_t pgoff,
902cdf6cdcdSVivek Goyal size_t nr_pages)
903cdf6cdcdSVivek Goyal {
904d19bd675SChristoph Hellwig struct dax_device *dax_dev = log_writes_dax_pgoff(ti, &pgoff);
905cdf6cdcdSVivek Goyal
906d19bd675SChristoph Hellwig return dax_zero_page_range(dax_dev, pgoff, nr_pages << PAGE_SHIFT);
907cdf6cdcdSVivek Goyal }
908cdf6cdcdSVivek Goyal
log_writes_dax_recovery_write(struct dm_target * ti,pgoff_t pgoff,void * addr,size_t bytes,struct iov_iter * i)909047218ecSJane Chu static size_t log_writes_dax_recovery_write(struct dm_target *ti,
910047218ecSJane Chu pgoff_t pgoff, void *addr, size_t bytes, struct iov_iter *i)
911047218ecSJane Chu {
912047218ecSJane Chu struct dax_device *dax_dev = log_writes_dax_pgoff(ti, &pgoff);
913047218ecSJane Chu
914047218ecSJane Chu return dax_recovery_write(dax_dev, pgoff, addr, bytes, i);
915047218ecSJane Chu }
916047218ecSJane Chu
917976431b0SDan Williams #else
918976431b0SDan Williams #define log_writes_dax_direct_access NULL
919cdf6cdcdSVivek Goyal #define log_writes_dax_zero_page_range NULL
920047218ecSJane Chu #define log_writes_dax_recovery_write NULL
921976431b0SDan Williams #endif
92298d82f48SRoss Zwisler
9230e9cebe7SJosef Bacik static struct target_type log_writes_target = {
9240e9cebe7SJosef Bacik .name = "log-writes",
92598d82f48SRoss Zwisler .version = {1, 1, 0},
9260e9cebe7SJosef Bacik .module = THIS_MODULE,
9270e9cebe7SJosef Bacik .ctr = log_writes_ctr,
9280e9cebe7SJosef Bacik .dtr = log_writes_dtr,
9290e9cebe7SJosef Bacik .map = log_writes_map,
9300e9cebe7SJosef Bacik .end_io = normal_end_io,
9310e9cebe7SJosef Bacik .status = log_writes_status,
932e56f81e0SChristoph Hellwig .prepare_ioctl = log_writes_prepare_ioctl,
9330e9cebe7SJosef Bacik .message = log_writes_message,
9340e9cebe7SJosef Bacik .iterate_devices = log_writes_iterate_devices,
9350e9cebe7SJosef Bacik .io_hints = log_writes_io_hints,
93698d82f48SRoss Zwisler .direct_access = log_writes_dax_direct_access,
937cdf6cdcdSVivek Goyal .dax_zero_page_range = log_writes_dax_zero_page_range,
938047218ecSJane Chu .dax_recovery_write = log_writes_dax_recovery_write,
9390e9cebe7SJosef Bacik };
940*3664ff82SYangtao Li module_dm(log_writes);
9410e9cebe7SJosef Bacik
9420e9cebe7SJosef Bacik MODULE_DESCRIPTION(DM_NAME " log writes target");
9430e9cebe7SJosef Bacik MODULE_AUTHOR("Josef Bacik <jbacik@fb.com>");
9440e9cebe7SJosef Bacik MODULE_LICENSE("GPL");
945