kernel.git/fs/btrfs/raid56.c, branch linux-6.5.y

btrfs: scrub: avoid unnecessary csum tree search preparing stripes

2023-09-19T10:30:24Z

commit 3c771c194402ffe20d4de68d9fc21e703179a9ce upstream. One of the bottleneck of the new scrub code is the extra csum tree search. The old code would only do the csum tree search for each scrub bio, which can be as large as 512KiB, thus they can afford to allocate a new path each time. But the new scrub code is doing csum tree search for each stripe, which is only 64KiB, this means we'd better re-use the same csum path during each search. This patch would introduce a per-sctx path for csum tree search, as we don't need to re-allocate the path every time we need to do a csum tree search. With this change we can further improve the queue depth and improve the scrub read performance: Before (with regression and cached extent tree path): Device r/s rkB/s rrqm/s %rrqm r_await rareq-sz aqu-sz %util nvme0n1p3 15875.00 1013328.00 12.00 0.08 0.08 63.83 1.35 100.00 After (with both cached extent/csum tree path): nvme0n1p3 17759.00 1133280.00 10.00 0.06 0.08 63.81 1.50 100.00 Fixes: e02ee89baa66 ("btrfs: scrub: switch scrub_simple_mirror() to scrub_stripe infrastructure") CC: stable@vger.kernel.org # 6.4+ Signed-off-by: Qu Wenruo Reviewed-by: David Sterba Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman

btrfs: raid56: always verify the P/Q contents for scrub

2023-07-18T01:13:15Z

[REGRESSION] Commit 75b470332965 ("btrfs: raid56: migrate recovery and scrub recovery path to use error_bitmap") changed the behavior of scrub_rbio(). Initially if we have no error reading the raid bio, we will assign @need_check to true, then finish_parity_scrub() would later verify the content of P/Q stripes before writeback. But after that commit we never verify the content of P/Q stripes and just writeback them. This can lead to unrepaired P/Q stripes during scrub, or already corrupted P/Q copied to the dev-replace target. [FIX] The situation is more complex than the regression, in fact the initial behavior is not 100% correct either. If we have the following rare case, it can still lead to the same problem using the old behavior: 0 16K 32K 48K 64K Data 1: |IIIIIII| | Data 2: | | Parity: | |CCCCCCC| | Where "I" means IO error, "C" means corruption. In the above case, we're scrubbing the parity stripe, then read out all the contents of Data 1, Data 2, Parity stripes. But found IO error in Data 1, which leads to rebuild using Data 2 and Parity and got the correct data. In that case, we would not verify if the Parity is correct for range [16K, 32K). So here we have to always verify the content of Parity no matter if we did recovery or not. This patch would remove the @need_check parameter of finish_parity_scrub() completely, and would always do the P/Q verification before writeback. Fixes: 75b470332965 ("btrfs: raid56: migrate recovery and scrub recovery path to use error_bitmap") CC: stable@vger.kernel.org # 6.2+ Signed-off-by: Qu Wenruo Signed-off-by: David Sterba

btrfs: scrub: use recovered data stripes as cache to avoid unnecessary read

2023-06-19T11:59:24Z

For P/Q stripe scrub, we have quite some duplicated read IO: - Data stripes read for verification This is triggered by the scrub_submit_initial_read() inside scrub_raid56_parity_stripe(). - Data stripes read (again) for P/Q stripe verification This is triggered by scrub_assemble_read_bios() from scrub_rbio(). Although we can have hit rbio cache and avoid unnecessary read, the chance is very low, as scrub would easily flush the whole rbio cache. This means, even we're just scrubbing a single P/Q stripe, we would read the data stripes twice for the best case scenario. If we need to recover some data stripes, it would cause more reads on the same data stripes, again and again. However before we call raid56_parity_submit_scrub_rbio() we already have all data stripes repaired and their contents ready to use. But RAID56 cache is unaware about the scrub cache, thus RAID56 layer itself still needs to re-read the data stripes. To avoid such cache miss, this patch would: - Introduce a new helper, raid56_parity_cache_data_pages() This function would grab the pages from an array, and copy the content to the rbio, marking all the involved sectors uptodate. The page copy is unavoidable because of the cache pages of rbio are all self managed, thus can not utilize outside pages without screwing up the lifespan. - Use the repaired data stripes as cache inside scrub_raid56_parity_stripe() By this, we ensure all the data sectors of the scrub rbio are already uptodate, and no need to read them again from disk. Signed-off-by: Qu Wenruo Signed-off-by: David Sterba

btrfs: use SECTOR_SHIFT to convert LBA to physical offset

2023-06-19T11:59:23Z

Using SECTOR_SHIFT to convert LBA to physical address makes it more readable. Signed-off-by: Anand Jain Reviewed-by: David Sterba Signed-off-by: David Sterba

btrfs: use SECTOR_SHIFT to convert physical offset to LBA

2023-06-19T11:59:23Z

Use SECTOR_SHIFT while converting a physical address to an LBA, makes it more readable. Signed-off-by: Anand Jain Reviewed-by: David Sterba Signed-off-by: David Sterba

btrfs: remove unused raid56 functions which were dedicated for scrub

2023-04-17T17:52:18Z

Since the scrub rework, the following RAID56 functions are no longer called: - raid56_add_scrub_pages() - raid56_alloc_missing_rbio() - raid56_submit_missing_rbio() Those functions are all utilized by scrub to handle missing device cases for RAID56. However the new scrub code handle them in a completely different way: - If it's data stripe, go recovery path through btrfs_submit_bio() - If it's P/Q stripe, it would be handled through raid56_parity_submit_scrub_rbio() And that function would handle dev-replace and repair properly. Thus we can safely remove those functions. Signed-off-by: Qu Wenruo Reviewed-by: David Sterba Signed-off-by: David Sterba

btrfs: scrub: introduce helper to find and fill sector info for a scrub_stripe

2023-04-17T16:01:23Z

The new helper will search the extent tree to find the first extent of a logical range, then fill the sectors array by two loops: - Loop 1 to fill common bits and metadata generation - Loop 2 to fill csum data (only for data bgs) This loop will use the new btrfs_lookup_csums_bitmap() to fill the full csum buffer, and set scrub_sector_verification::csum. With all the needed info filled by this function, later we only need to submit and verify the stripe. Here we temporarily export the helper to avoid warning on unused static function. Signed-off-by: Qu Wenruo Reviewed-by: David Sterba Signed-off-by: David Sterba

btrfs: use __bio_add_page to add single a page in rbio_add_io_sector

2023-04-17T16:01:20Z

The btrfs raid56 sector submission code uses bio_add_page() to add a page to a newly created bio. bio_add_page() can fail, but the return value is never checked. Use __bio_add_page() as adding a single page to a newly created bio is guaranteed to succeed. This brings us a step closer to marking bio_add_page() as __must_check. Reviewed-by: Damien Le Moal Signed-off-by: Johannes Thumshirn Reviewed-by: David Sterba Signed-off-by: David Sterba

btrfs: replace btrfs_io_context::raid_map with a fixed u64 value

2023-04-17T16:01:14Z

In btrfs_io_context structure, we have a pointer raid_map, which indicates the logical bytenr for each stripe. But considering we always call sort_parity_stripes(), the result raid_map[] is always sorted, thus raid_map[0] is always the logical bytenr of the full stripe. So why we waste the space and time (for sorting) for raid_map? This patch will replace btrfs_io_context::raid_map with a single u64 number, full_stripe_start, by: - Replace btrfs_io_context::raid_map with full_stripe_start - Replace call sites using raid_map[0] to use full_stripe_start - Replace call sites using raid_map[i] to compare with nr_data_stripes. The benefits are: - Less memory wasted on raid_map It's sizeof(u64) * num_stripes vs sizeof(u64). It'll always save at least one u64, and the benefit grows larger with num_stripes. - No more weird alloc_btrfs_io_context() behavior As there is only one fixed size + one variable length array. Signed-off-by: Qu Wenruo Signed-off-by: David Sterba

btrfs: use an efficient way to represent source of duplicated stripes

2023-04-17T16:01:14Z

For btrfs dev-replace, we have to duplicate writes to the source device into the target device. For non-RAID56, all writes into the same mapped ranges are sharing the same content, thus they don't really need to bother anything. (E.g. in btrfs_submit_bio() for non-RAID56 range we just submit the same write to all involved devices). But for RAID56, all stripes contain different content, thus we must have a clear mapping of which stripe is duplicated from which original stripe. Currently we use a complex way using tgtdev_map[] array, e.g: num_tgtdevs = 1 tgtdev_map[0] = 0 <- Means stripes[0] is not involved in replace. tgtdev_map[1] = 3 <- Means stripes[1] is involved in replace, and it's duplicated to stripes[3]. tgtdev_map[2] = 0 <- Means stripes[2] is not involved in replace. But this is wasting some space, and ignores one important thing for dev-replace, there is at most one running replace. Thus we can change it to a fixed array to represent the mapping: replace_nr_stripes = 1 replace_stripe_src = 1 <- Means stripes[1] is involved in replace. thus the extra stripe is a copy of stripes[1] By this we can save some space for bioc on RAID56 chunks with many devices. And we get rid of one variable sized array from bioc. Thus the patch involves the following changes: - Replace @num_tgtdevs and @tgtdev_map[] with @replace_nr_stripes and @replace_stripe_src. @num_tgtdevs is just renamed to @replace_nr_stripes. While the mapping is completely changed. - Add extra ASSERT()s for RAID56 code - Only add two more extra stripes for dev-replace cases. As we have an upper limit on how many dev-replace stripes we can have. - Unify the behavior of handle_ops_on_dev_replace() Previously handle_ops_on_dev_replace() go two different paths for WRITE and GET_READ_MIRRORS. Now unify them by always going the WRITE path first (with at most 2 replace stripes), then if we're doing GET_READ_MIRRORS and we have 2 extra stripes, just drop one stripe. - Remove the @real_stripes argument from alloc_btrfs_io_context() As we don't need the old variable length array any more. Signed-off-by: Qu Wenruo Signed-off-by: David Sterba