[PATCH] xfs/013: fix ENOSPC handling

[PATCH] xfs/013: fix ENOSPC handling

[Shin'ichiro Kawasaki]

Commit 000813899afb ("fstests: scale some tests for high CPU count
sanity") replaced "touch" with "echo -n >" to speed up file creation by
using a shell builtin redirection instead of forking and execing an
external binary. However, this broke the ENOSPC filtering the test
relies on.

When the scratch filesystem fills up, the shell redirection "> $dir/$i"
fails instead of the "touch" command. The shell applies redirections
left-to-right, so "> $dir/$i" is attempted before "2>&1". Because
opening the file fails immediately with ENOSPC, the command never runs
and the shell writes the error message "No space left on device" to the
stderr instead of the pipe. Then the error message is not passed to
"filter_enospc" and leaks into the test output, causing the test case
failure. The failure is recreated by running the test case with rather
small size of block devices, such as 128MiB null_blk:

xfs/013            - output mismatch (see /home/test/xfstests/results//xfs/013.out.bad)
    --- tests/xfs/013.out	2026-05-28 16:52:16.881159200 +0900
    +++ /home/test/xfstests/results//xfs/013.out.bad	2026-06-11 17:20:36.269003467 +0900
    @@ -5,3 +5,12305 @@
     naming   =VERN bsize=XXX
     log      =LDEV bsize=XXX blocks=XXX
     realtime =RDEV extsz=XXX blocks=XXX, rtextents=XXX
    +/home/test/xfstests/tests/xfs/013: line 45: /var/scratch/dir2/19013: No space left on device
    +/home/test/xfstests/tests/xfs/013: line 45: /var/scratch/dir2/6452: No space left on device
    +/home/test/xfstests/tests/xfs/013: line 45: /var/scratch/dir2/14964: No space left on device
    +/home/test/xfstests/tests/xfs/013: line 45: /var/scratch/dir2/10051: No space left on device
    ...
    (Run 'diff -u /home/test/xfstests/tests/xfs/013.out /home/test/xfstests/results//xfs/013.out.bad'  to see the entire diff)

To fix the ENOSPC handling, wrap the redirection in a brace group so
that "2>&1" is applied to the group before the inner redirection is
attempted. The redirection failure is then written to the group's
stderr, which is piped into "filter_enospc". This keeps the faster
builtin file creation than the "touch" command as the trigger commit
intended.

Fixes: 000813899afb ("fstests: scale some tests for high CPU count sanity")
Signed-off-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@wdc.com>
---
 tests/xfs/013 | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tests/xfs/013 b/tests/xfs/013
index fd011445..2e9b9850 100755
--- a/tests/xfs/013
+++ b/tests/xfs/013
@@ -28,7 +28,7 @@ _create()
 	mkdir -p $dir
 	for i in $(seq 0 $count)
 	do
-		echo -n > $dir/$i 2>&1 | filter_enospc
+		{ echo -n > $dir/$i ; } 2>&1 | filter_enospc
 	done
 }
 
@@ -42,7 +42,7 @@ _rand_replace()
 	do
 		file=$((RANDOM % count))
 		rm -f $dir/$file
-		echo -n > $dir/$file 2>&1 | filter_enospc
+		{ echo -n > $dir/$file ; } 2>&1 | filter_enospc
 	done
 }
 
-- 
2.54.0

Re: [PATCH] xfs/013: fix ENOSPC handling

[Christoph Hellwig]

On Mon, Jun 15, 2026 at 09:12:57PM +0900, Shin'ichiro Kawasaki wrote:
> Commit 000813899afb ("fstests: scale some tests for high CPU count
> sanity") replaced "touch" with "echo -n >" to speed up file creation by
> using a shell builtin redirection instead of forking and execing an
> external binary. However, this broke the ENOSPC filtering the test
> relies on.
> 
> When the scratch filesystem fills up, the shell redirection "> $dir/$i"
> fails instead of the "touch" command. The shell applies redirections
> left-to-right, so "> $dir/$i" is attempted before "2>&1". Because
> opening the file fails immediately with ENOSPC, the command never runs
> and the shell writes the error message "No space left on device" to the
> stderr instead of the pipe. Then the error message is not passed to
> "filter_enospc" and leaks into the test output, causing the test case
> failure. The failure is recreated by running the test case with rather
> small size of block devices, such as 128MiB null_blk:

Ewww.  I've seen the same errors sporadically too in zoned testing,
but never got to the ground of it.  The fix looks good:

Reviewed-by: Christoph Hellwig <hch@lst.de>