fix prefetching of indirect blocks while destroying #14603
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
amotin
reviewed
Mar 10, 2023
|
The zloop test failed because this PR's branch was not rebased on the most recent master, so it is missing the recently merged fix from #14583. The FreeBSD test failure is a pre-existing issue. |
When traversing a tree of block pointers (e.g. for `zfs destroy <fs>` or `zfs send`), we prefetch the indirect blocks that will be needed, in `traverse_prefetch_metadata()`. In the case of `zfs destroy <fs>`, we do a little traversing each txg, and resume the traversal the next txg. So the indirect blocks that will be needed, and thus are candidates for prefetching, does not include blocks that are before the resume point. The problem is that the logic for determining if the indirect blocks are before the resume point is incorrect, causing the (up to 1024) L1 indirect blocks that are inside the first L2 to not be prefetched. In practice, if we are able to read many more than 1024 blocks per txg, then this will be inconsequential. But if i/o latency is more than a few milliseconds, almost no L1's will be prefetched, so they will be read serially, and thus the destroying will be very slow. This can be observed as `zpool get freeing` decreasing very slowly. Specifically: When we first examine the L2 that contains the block we'll be resuming from, we have not yet resumed, so `td_resume` is nonzero. At this point, all calls to `traverse_prefetch_metadata()` will fail, even if the L1 in question is after the resume point. It isn't until the callback is issued for the resume point that we zero out `td_resume`, but by this point we've already attempted and failed to prefetch everything under this L2 indirect block. This commit addresses the issue by reusing the existing `resume_skip_check()` to determine if the L1's bookmark is before or after the resume point. To do so, this function is made non-mutating (the caller now zeros `td_resume`). Note, this bug likely predates (was not introduced by) openzfs#11803. Signed-off-by: Matthew Ahrens <[email protected]>
ahrens
force-pushed
the
prefetch_indirect
branch
from
e3cfa1f to
8e2b28c
Compare
This was referenced Mar 13, 2023
amotin
approved these changes
Mar 14, 2023
behlendorf
approved these changes
Mar 16, 2023
behlendorf
added
Status: Accepted
andrewc12
pushed a commit
to andrewc12/openzfs
that referenced
this pull request
Mar 26, 2023
When traversing a tree of block pointers (e.g. for `zfs destroy <fs>` or `zfs send`), we prefetch the indirect blocks that will be needed, in `traverse_prefetch_metadata()`. In the case of `zfs destroy <fs>`, we do a little traversing each txg, and resume the traversal the next txg. So the indirect blocks that will be needed, and thus are candidates for prefetching, does not include blocks that are before the resume point. The problem is that the logic for determining if the indirect blocks are before the resume point is incorrect, causing the (up to 1024) L1 indirect blocks that are inside the first L2 to not be prefetched. In practice, if we are able to read many more than 1024 blocks per txg, then this will be inconsequential. But if i/o latency is more than a few milliseconds, almost no L1's will be prefetched, so they will be read serially, and thus the destroying will be very slow. This can be observed as `zpool get freeing` decreasing very slowly. Specifically: When we first examine the L2 that contains the block we'll be resuming from, we have not yet resumed, so `td_resume` is nonzero. At this point, all calls to `traverse_prefetch_metadata()` will fail, even if the L1 in question is after the resume point. It isn't until the callback is issued for the resume point that we zero out `td_resume`, but by this point we've already attempted and failed to prefetch everything under this L2 indirect block. This commit addresses the issue by reusing the existing `resume_skip_check()` to determine if the L1's bookmark is before or after the resume point. To do so, this function is made non-mutating (the caller now zeros `td_resume`). Note, this bug likely predates (was not introduced by) openzfs#11803. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Matthew Ahrens <[email protected]> Closes openzfs#14603
andrewc12
pushed a commit
to andrewc12/openzfs
that referenced
this pull request
Mar 26, 2023
When traversing a tree of block pointers (e.g. for `zfs destroy <fs>` or `zfs send`), we prefetch the indirect blocks that will be needed, in `traverse_prefetch_metadata()`. In the case of `zfs destroy <fs>`, we do a little traversing each txg, and resume the traversal the next txg. So the indirect blocks that will be needed, and thus are candidates for prefetching, does not include blocks that are before the resume point. The problem is that the logic for determining if the indirect blocks are before the resume point is incorrect, causing the (up to 1024) L1 indirect blocks that are inside the first L2 to not be prefetched. In practice, if we are able to read many more than 1024 blocks per txg, then this will be inconsequential. But if i/o latency is more than a few milliseconds, almost no L1's will be prefetched, so they will be read serially, and thus the destroying will be very slow. This can be observed as `zpool get freeing` decreasing very slowly. Specifically: When we first examine the L2 that contains the block we'll be resuming from, we have not yet resumed, so `td_resume` is nonzero. At this point, all calls to `traverse_prefetch_metadata()` will fail, even if the L1 in question is after the resume point. It isn't until the callback is issued for the resume point that we zero out `td_resume`, but by this point we've already attempted and failed to prefetch everything under this L2 indirect block. This commit addresses the issue by reusing the existing `resume_skip_check()` to determine if the L1's bookmark is before or after the resume point. To do so, this function is made non-mutating (the caller now zeros `td_resume`). Note, this bug likely predates (was not introduced by) openzfs#11803. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Matthew Ahrens <[email protected]> Closes openzfs#14603
pcd1193182
pushed a commit
to pcd1193182/zfs
that referenced
this pull request
Sep 26, 2023
When traversing a tree of block pointers (e.g. for `zfs destroy <fs>` or `zfs send`), we prefetch the indirect blocks that will be needed, in `traverse_prefetch_metadata()`. In the case of `zfs destroy <fs>`, we do a little traversing each txg, and resume the traversal the next txg. So the indirect blocks that will be needed, and thus are candidates for prefetching, does not include blocks that are before the resume point. The problem is that the logic for determining if the indirect blocks are before the resume point is incorrect, causing the (up to 1024) L1 indirect blocks that are inside the first L2 to not be prefetched. In practice, if we are able to read many more than 1024 blocks per txg, then this will be inconsequential. But if i/o latency is more than a few milliseconds, almost no L1's will be prefetched, so they will be read serially, and thus the destroying will be very slow. This can be observed as `zpool get freeing` decreasing very slowly. Specifically: When we first examine the L2 that contains the block we'll be resuming from, we have not yet resumed, so `td_resume` is nonzero. At this point, all calls to `traverse_prefetch_metadata()` will fail, even if the L1 in question is after the resume point. It isn't until the callback is issued for the resume point that we zero out `td_resume`, but by this point we've already attempted and failed to prefetch everything under this L2 indirect block. This commit addresses the issue by reusing the existing `resume_skip_check()` to determine if the L1's bookmark is before or after the resume point. To do so, this function is made non-mutating (the caller now zeros `td_resume`). Note, this bug likely predates (was not introduced by) openzfs#11803. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Matthew Ahrens <[email protected]> Closes openzfs#14603
behlendorf
added a commit
that referenced
this pull request
Oct 13, 2023
New Features - Block cloning (#13392) - Linux container support (#14070, #14097, #12263) - Scrub error log (#12812, #12355) - BLAKE3 checksums (#12918) - Corrective "zfs receive" - Vdev and zpool user properties Performance - Fully adaptive ARC (#14359) - SHA2 checksums (#13741) - Edon-R checksums (#13618) - Zstd early abort (#13244) - Prefetch improvements (#14603, #14516, #14402, #14243, #13452) - General optimization (#14121, #14123, #14039, #13680, #13613, #13606, #13576, #13553, #12789, #14925, #14948) Signed-off-by: Brian Behlendorf <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Motivation and Context
When traversing a tree of block pointers (e.g. for
zfs destroy <fs>orzfs send), we prefetch the indirect blocks that will be needed, intraverse_prefetch_metadata(). In the case ofzfs destroy <fs>, we do a little traversing each txg, and resume the traversal the next txg. So the indirect blocks that will be needed, and thus are candidates for prefetching, does not include blocks that are before the resume point.The problem is that the logic for determining if the indirect blocks are before the resume point is incorrect, causing the (up to 1024) L1 indirect blocks that are inside the first L2 to not be prefetched. In practice, if we are able to read many more than 1024 blocks per txg, then this will be inconsequential. But if i/o latency is more than a few milliseconds, almost no L1's will be prefetched, so they will be read serially, and thus the destroying will be very slow. This can be observed as
zpool get freeingdecreasing very slowly.Specifically: When we first examine the L2 that contains the block we'll be resuming from, we have not yet resumed, so
td_resumeis nonzero. At this point, all calls totraverse_prefetch_metadata()will fail, even if the L1 in question is after the resume point. It isn't until the callback is issued for the resume point that we zero outtd_resume, but by this point we've already attempted and failed to prefetch everything under this L2 indirect block.Description
This commit addresses the issue by reusing the existing
resume_skip_check()to determine if the L1's bookmark is before or after the resume point. To do so, this function is made non-mutating (the caller now zerostd_resume).Note, this bug likely predates (was not introduced by) #11803.
How Has This Been Tested?
With high-latency storage, I saw a >10x improvement in space reclamation performance following a
zfs destroy <fs>.Types of changes
Checklist:
Signed-off-by.