Skip to contentSkip to conversation

btrfs: make read repair work in synchronous mode

RFC PATCH Superseded Qu Wenruo 1 participant 2 messages Apr 27, 2022 – Apr 28, 2022 Btrfs file system

lore
Share & tools
lore

Qu Wenruo

author
#0 lore raw

v2 cover btrfs: make read repair work in synchronous mode

[CHANGELOG]
RFC v1 -> RFC v2:
- Assemble a bio list for read/write bios and submit them in one go This allows less submit bio hooks, while still allow us to wait for them all to finish.

- Completely remove io_failure_tree infrastructure
Now we don't need to remember which mirror we hit error. At end_bio_extent_readpage() we either get good data and done the repair already, or we there aren't enough mirrors for us to recover all data.

This is mostly trading on-stack memory of end_bio_extent_readpage() with btrfs_inode::io_failure_tree.
The latter tree has a much longer lifespan, thus I think it's still a win overall

[RFC POINTS]
- How to improve read_repair_get_sector()?
Currently we always iterate the whole bio to grab the target page/pgoff.

Is there any better cached method to avoid such iteration?

- Is this new code logically more reader-friendly?
It's more for sure straight-forward, but I doubt if it's any easier to read compared to the old code.

- btrfs/157 failure
Need extra check to find out why btrfs/157 failed.
In theory, we should just iterate through all mirrors, I guess it's we have no way to exhaust all combinations, thus the extra 2 "mirrors" can gave us wrong result for RAID6.

[BEFORE]
For corrupted sectors, we just record the logical bytenr and mirror number into io_failure_tree, then re-queue the same block with different mirror number and call it a day.

The re-queued read will trigger enter the same endio function, with extra failrec handling to either continue re-queue (csum mismatch/read failure), or clear the current failrec and submit a write to fix the corrupted mirror (read succeeded and csum match/no csum).

This is harder to read, as we need to enter the same river twice or even more.

[AFTER]
For corrupted sectors, we record the following things into an on-stack structure in end_bio_extent_readpage():

- The original bio

- The original file offset of the bio
This is for direct IO case, as we can not grab file offset just using page_offset()

- Offset inside the bio of the corrupted sector

- Corrupted mirror

Then in the new btrfs_read_repair_ctrl structure, we hold those info like:

Original bio logical = X, file_offset = Y, inode=(R/I)

Offset inside bio: 0 4k 8K 12K 16K
cur_bad_bitmap | X| X| | X|

Each set bit will indicate we have a corrupted sector inside the original bio.

During endio function, we only populate the cur_bad_bitmap.

After we have iterated all sectors of the original bio, then we call btrfs_read_repair_finish() to do the real repair by:

- Build a list of bios for cur_bad_bitmap
For above case, bio offset [0, 8K) will be inside one bio, while another bio for bio offset [12K, 16K).

And the page/pgoff will be extracted from the orignial bio.

This is a little different from the old behavior, as old behavior will submit a new bio for each sector.
The new behavior will save some btrfs_map_bio() calls.

- Submit all the bios in the bio list and wait them to finish

- Re-verify the read result

- Submit write for the corrupted mirror
Currently the write is still submitted for each sector and we will wait for each sector to finish.
This needs some optimization.

And for repaired sectors, remove them from @cur_bad_bitmap.

- Do the same loop until either 1) we tried all mirrors, or 2) no more corrupted sectors

- Handle the remaining corrupted sectors
Either mark them error for buffered read, or just return an error for direct IO.

By this we can:
- Remove the re-entry behavior of endio function Now everything is handled inside end_bio_extent_readpage().

- Remove the io_failure_tree completely
As we don't need to record which mirror has failed.

- Slightly reduced overhead on read repair
Now we won't call btrfs_map_bio() for each corrupted sector, as we can merge the sectors into a much larger bio.

Qu Wenruo (12):
btrfs: introduce a pure data checksum checking helper
btrfs: always save bio::bi_iter into btrfs_bio::iter before submitting
btrfs: remove duplicated parameters from submit_data_read_repair()
btrfs: add btrfs_read_repair_ctrl to record corrupted sectors
btrfs: add a helper to queue a corrupted sector for read repair
btrfs: introduce a helper to repair from one mirror
btrfs: allow btrfs read repair to submit all writes in one go
btrfs: switch buffered read to the new btrfs_read_repair_* based
repair routine
btrfs: switch direct IO routine to use btrfs_read_repair_ctrl
btrfs: cleanup btrfs_repair_one_sector()
btrfs: remove io_failure_record infrastructure completely
btrfs: remove btrfs_inode::io_failure_tree

 fs/btrfs/btrfs_inode.h       |   5 -
 fs/btrfs/compression.c       |  12 +-
 fs/btrfs/ctree.h             |   2 +
 fs/btrfs/extent-io-tree.h    |  15 -
 fs/btrfs/extent_io.c         | 744 ++++++++++++++++++-----------------
 fs/btrfs/extent_io.h         |  89 +++--
 fs/btrfs/inode.c             | 108 +++--
 include/trace/events/btrfs.h |   1 -
 8 files changed, 518 insertions(+), 458 deletions(-)
Signature · 1 line
2.36.0
author
#1 lore raw
Quoted 28 lines · Qu Wenruo

[CHANGELOG]
RFC v1 -> RFC v2:
- Assemble a bio list for read/write bios and submit them in one go This allows less submit bio hooks, while still allow us to wait for them all to finish.

- Completely remove io_failure_tree infrastructure
Now we don't need to remember which mirror we hit error. At end_bio_extent_readpage() we either get good data and done the repair already, or we there aren't enough mirrors for us to recover all data.

This is mostly trading on-stack memory of end_bio_extent_readpage() with btrfs_inode::io_failure_tree.
The latter tree has a much longer lifespan, thus I think it's still a win overall

[RFC POINTS]
- How to improve read_repair_get_sector()?
Currently we always iterate the whole bio to grab the target page/pgoff.

Is there any better cached method to avoid such iteration?

- Is this new code logically more reader-friendly?
It's more for sure straight-forward, but I doubt if it's any easier to read compared to the old code.

Quoted 5 lines

- btrfs/157 failure
Need extra check to find out why btrfs/157 failed.
In theory, we should just iterate through all mirrors, I guess it's we have no way to exhaust all combinations, thus the extra 2 "mirrors" can gave us wrong result for RAID6.

This is related to the writeback behavior for bad mirrors.

For RAID56, the mirror_num is not really indicating a mirror, but an hint on how to rebuild the data.

For RAID6 it can be as large as the number of stripes. The data stripe where our read is, is always corrupted (or we won't need to rebuild). The mirror number can be used to iterate through all combination of the next possible corrupted stripe.

When we got the correct data using a specific mirror number, we should not write back the correct data for RAID56.

As it will trigger RMW, and RMW always read data stripes from disk (including the other corrupted data in RAID6), thus the write back will in fact cause the other corrupted data to be eternal.

In that case, we just need to skip the unnecessary writeback for RAID56.

This will not write the correct full stripe back to disk, but it's way better than corrupting the data furthermore.

The fix is already updated in my github repo.

Thanks,
Qu

Quoted 109 lines

[BEFORE]
For corrupted sectors, we just record the logical bytenr and mirror number into io_failure_tree, then re-queue the same block with different mirror number and call it a day.

The re-queued read will trigger enter the same endio function, with extra failrec handling to either continue re-queue (csum mismatch/read failure), or clear the current failrec and submit a write to fix the corrupted mirror (read succeeded and csum match/no csum).

This is harder to read, as we need to enter the same river twice or even more.

[AFTER]
For corrupted sectors, we record the following things into an on-stack structure in end_bio_extent_readpage():

- The original bio

- The original file offset of the bio
This is for direct IO case, as we can not grab file offset just using page_offset()

- Offset inside the bio of the corrupted sector

- Corrupted mirror

Then in the new btrfs_read_repair_ctrl structure, we hold those info like:

Original bio logical = X, file_offset = Y, inode=(R/I)

Offset inside bio: 0 4k 8K 12K 16K
cur_bad_bitmap | X| X| | X|

Each set bit will indicate we have a corrupted sector inside the original bio.

During endio function, we only populate the cur_bad_bitmap.

After we have iterated all sectors of the original bio, then we call btrfs_read_repair_finish() to do the real repair by:

- Build a list of bios for cur_bad_bitmap
For above case, bio offset [0, 8K) will be inside one bio, while another bio for bio offset [12K, 16K).

And the page/pgoff will be extracted from the orignial bio.

This is a little different from the old behavior, as old behavior will submit a new bio for each sector.
The new behavior will save some btrfs_map_bio() calls.

- Submit all the bios in the bio list and wait them to finish

- Re-verify the read result

- Submit write for the corrupted mirror
Currently the write is still submitted for each sector and we will wait for each sector to finish.
This needs some optimization.

And for repaired sectors, remove them from @cur_bad_bitmap.

- Do the same loop until either 1) we tried all mirrors, or 2) no more corrupted sectors

- Handle the remaining corrupted sectors
Either mark them error for buffered read, or just return an error for direct IO.

By this we can:
- Remove the re-entry behavior of endio function
Now everything is handled inside end_bio_extent_readpage().

- Remove the io_failure_tree completely
As we don't need to record which mirror has failed.

- Slightly reduced overhead on read repair
Now we won't call btrfs_map_bio() for each corrupted sector, as we can merge the sectors into a much larger bio.

Qu Wenruo (12):
btrfs: introduce a pure data checksum checking helper
btrfs: always save bio::bi_iter into btrfs_bio::iter before submitting
btrfs: remove duplicated parameters from submit_data_read_repair()
btrfs: add btrfs_read_repair_ctrl to record corrupted sectors
btrfs: add a helper to queue a corrupted sector for read repair
btrfs: introduce a helper to repair from one mirror
btrfs: allow btrfs read repair to submit all writes in one go
btrfs: switch buffered read to the new btrfs_read_repair_* based
repair routine
btrfs: switch direct IO routine to use btrfs_read_repair_ctrl
btrfs: cleanup btrfs_repair_one_sector()
btrfs: remove io_failure_record infrastructure completely
btrfs: remove btrfs_inode::io_failure_tree

  fs/btrfs/btrfs_inode.h       |   5 -
  fs/btrfs/compression.c       |  12 +-
  fs/btrfs/ctree.h             |   2 +
  fs/btrfs/extent-io-tree.h    |  15 -
  fs/btrfs/extent_io.c         | 744 ++++++++++++++++++-----------------
  fs/btrfs/extent_io.h         |  89 +++--
  fs/btrfs/inode.c             | 108 +++--
  include/trace/events/btrfs.h |   1 -
  8 files changed, 518 insertions(+), 458 deletions(-)