2020-03-19 15:01:16

by Eric Whitney

[permalink] [raw]
Subject: [PATCH] ext4: disable dioread_nolock whenever delayed allocation is disabled

The patch "ext4: make dioread_nolock the default" (244adf6426ee) causes
generic/422 to fail when run in kvm-xfstests' ext3conv test case. This
applies both the dioread_nolock and nodelalloc mount options, a
combination not previously tested by kvm-xfstests. The failure occurs
because the dioread_nolock code path splits a previously fallocated
multiblock extent into a series of single block extents when overwriting
a portion of that extent. That causes allocation of an extent tree leaf
node and a reshuffling of extents. Once writeback is completed, the
individual extents are recombined into a single extent, the extent is
moved again, and the leaf node is deleted. The difference in block
utilization before and after writeback due to the leaf node triggers the
failure.

The original reason for this behavior was to avoid ENOSPC when handling
I/O completions during writeback in the dioread_nolock code paths when
delayed allocation is disabled. It may no longer be necessary, because
code was added in the past to reserve extra space to solve this problem
when delayed allocation is enabled, and this code may also apply when
delayed allocation is disabled. Until this can be verified, don't use
the dioread_nolock code paths if delayed allocation is disabled.

Signed-off-by: Eric Whitney <[email protected]>
---
fs/ext4/ext4_jbd2.h | 3 +++
1 file changed, 3 insertions(+)

diff --git a/fs/ext4/ext4_jbd2.h b/fs/ext4/ext4_jbd2.h
index 7ea4f6fa173b..4b9002f0e84c 100644
--- a/fs/ext4/ext4_jbd2.h
+++ b/fs/ext4/ext4_jbd2.h
@@ -512,6 +512,9 @@ static inline int ext4_should_dioread_nolock(struct inode *inode)
return 0;
if (ext4_should_journal_data(inode))
return 0;
+ /* temporary fix to prevent generic/422 test failures */
+ if (!test_opt(inode->i_sb, DELALLOC))
+ return 0;
return 1;
}

--
2.11.0


2020-03-26 14:58:55

by Theodore Ts'o

[permalink] [raw]
Subject: Re: [PATCH] ext4: disable dioread_nolock whenever delayed allocation is disabled

On Thu, Mar 19, 2020 at 11:00:28AM -0400, Eric Whitney wrote:
> The patch "ext4: make dioread_nolock the default" (244adf6426ee) causes
> generic/422 to fail when run in kvm-xfstests' ext3conv test case. This
> applies both the dioread_nolock and nodelalloc mount options, a
> combination not previously tested by kvm-xfstests. The failure occurs
> because the dioread_nolock code path splits a previously fallocated
> multiblock extent into a series of single block extents when overwriting
> a portion of that extent. That causes allocation of an extent tree leaf
> node and a reshuffling of extents. Once writeback is completed, the
> individual extents are recombined into a single extent, the extent is
> moved again, and the leaf node is deleted. The difference in block
> utilization before and after writeback due to the leaf node triggers the
> failure.
>
> The original reason for this behavior was to avoid ENOSPC when handling
> I/O completions during writeback in the dioread_nolock code paths when
> delayed allocation is disabled. It may no longer be necessary, because
> code was added in the past to reserve extra space to solve this problem
> when delayed allocation is enabled, and this code may also apply when
> delayed allocation is disabled. Until this can be verified, don't use
> the dioread_nolock code paths if delayed allocation is disabled.
>
> Signed-off-by: Eric Whitney <[email protected]>

Applied, thanks.

- Ted