[PATCH] generic: add a test that stresses metadata eviction

[PATCH] generic: add a test that stresses metadata eviction

[Christoph Hellwig]

The fsverity test generic/759 could be used to reproduce a regression in
the xfs buffer hashing.  Carlos Maiolino figured out that even removing
the fsverity-specific bits would reproduce the issue.  This test is
thus a copy of generic/759 with the fsverity bits striped out, and
run 5 times as that seems to be the threshold at which the issue
reproduces reliably.

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

Not sure how to best credit Eric for the original patch and Carlos
for the extraction, any ideas?

 tests/generic/4201     | 108 +++++++++++++++++++++++++++++++++++++++++
 tests/generic/4201.out |   2 +
 2 files changed, 110 insertions(+)
 create mode 100755 tests/generic/4201
 create mode 100644 tests/generic/4201.out

diff --git a/tests/generic/4201 b/tests/generic/4201
new file mode 100755
index 000000000000..fc367beeadaf
--- /dev/null
+++ b/tests/generic/4201
@@ -0,0 +1,108 @@
+#! /bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Copyright 2019 Google LLC
+#
+# FS QA Test generic/4201
+#
+# Race dropping file systems caches vs fsstress an repeated reading of files.
+# Based on generic/579 with the fsverity bits stripped out, and an increased
+# number of runs.
+#
+. ./common/preamble
+_begin_fstest auto stress
+
+stop_run()
+{
+	# Stop all subprocesses.
+	_kill_fsstress
+	touch $tmp.done
+	wait
+}
+
+_cleanup()
+{
+	stop_run
+	rm -f $tmp.*
+}
+
+. ./common/filter
+
+fsv_file_size=10000000
+nproc_enabler=$((4 * LOAD_FACTOR))
+nproc_reader=$((6 * LOAD_FACTOR))
+nproc_stress=$((3 * LOAD_FACTOR))
+runtime=$((20 * TIME_FACTOR))
+
+run_test()
+{
+	rm -rf $tmp.done
+
+	# Create the test files and start the fs-verity enabler processes.
+	for ((proc = 0; proc < nproc_enabler; proc++)); do
+		orig_file=$SCRATCH_MNT/orig$proc
+		fsv_file=$SCRATCH_MNT/fsv$proc
+		head -c $fsv_file_size /dev/urandom > $orig_file
+		(
+			while [ ! -e $tmp.done ]; do
+				rm -f $fsv_file
+				cp $orig_file $fsv_file
+				# Give the readers some time to read from the file.
+				sleep 0.$((RANDOM % 100))
+			done
+		) &
+	done
+
+	# Start the reader processes.
+	for ((proc = 0; proc < nproc_reader; proc++)); do
+		(
+			while [ ! -e $tmp.done ]; do
+				# Choose a random file for each iteration, so that
+				# sometimes multiple processes read from the same file.
+				i=$((RANDOM % nproc_enabler))
+				orig_file=$SCRATCH_MNT/orig$i
+				fsv_file=$SCRATCH_MNT/fsv$i
+
+				# After the copy from $orig_file to $fsv_file has
+				# completed, the contents of these two files should
+				# match, regardless of whether verity has been enabled
+				# or not yet (or is currently being enabled).
+				cmp $orig_file $fsv_file |& _filter_scratch | \
+					sed -e "s/'//g" | \
+					grep -v "SCRATCH_MNT/fsv$i: No such file or directory" | \
+					grep -v "EOF on SCRATCH_MNT/fsv$i"
+				done
+		) &
+	done
+
+	# Start a process that occasionally runs 'sync && drop_caches'.  This makes more
+	# reads go through fs-verity for real, rather than just returning pagecache.
+	(
+		while [ ! -e $tmp.done ]; do
+			sleep 2.$((RANDOM % 100))
+			_scratch_sync && echo 3 > /proc/sys/vm/drop_caches
+		done
+	) &
+
+	# Start the fsstress processes.
+	_run_fsstress_bg -p $nproc_stress -l 0 -d $SCRATCH_MNT/stressdir
+
+	# Run for a while.
+	sleep $runtime
+
+	stop_run
+}
+
+for i in $(seq 1 5); do
+	echo "Starting run $i" >> $seqres.full
+
+	_scratch_mkfs &>> $seqres.full
+	_scratch_mount
+	run_test
+	_scratch_unmount
+done
+
+echo "Silence is golden"
+
+# success, all done
+status=0
+exit
diff --git a/tests/generic/4201.out b/tests/generic/4201.out
new file mode 100644
index 000000000000..2c3f3d42ff3f
--- /dev/null
+++ b/tests/generic/4201.out
@@ -0,0 +1,2 @@
+QA output created by 4201
+Silence is golden
-- 
2.53.0

Re: [PATCH] generic: add a test that stresses metadata eviction

[Carlos Maiolino]

On Mon, May 18, 2026 at 08:24:53AM +0200, Christoph Hellwig wrote:
> The fsverity test generic/759 could be used to reproduce a regression in
> the xfs buffer hashing.  Carlos Maiolino figured out that even removing
> the fsverity-specific bits would reproduce the issue.  This test is
> thus a copy of generic/759 with the fsverity bits striped out, and
> run 5 times as that seems to be the threshold at which the issue
> reproduces reliably.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> ---
> 
> Not sure how to best credit Eric for the original patch and Carlos
> for the extraction, any ideas?

To be fair, Andrey initially suggested stripping off the fs-verity
parts, IIRC because he wanted to isolate his own series from this crash.
My role was mostly reproduce it and hunt the bug down.
Risking to come with yet another tag, perhaps something like:

Based-on-generic-579-by: <Eric...>

It doesn't look ugly to me, but who knows...

> 
>  tests/generic/4201     | 108 +++++++++++++++++++++++++++++++++++++++++
>  tests/generic/4201.out |   2 +
>  2 files changed, 110 insertions(+)
>  create mode 100755 tests/generic/4201
>  create mode 100644 tests/generic/4201.out
> 
> diff --git a/tests/generic/4201 b/tests/generic/4201
> new file mode 100755
> index 000000000000..fc367beeadaf
> --- /dev/null
> +++ b/tests/generic/4201
> @@ -0,0 +1,108 @@
> +#! /bin/bash
> +# SPDX-License-Identifier: GPL-2.0
> +# Copyright 2019 Google LLC
> +#
> +# FS QA Test generic/4201
> +#
> +# Race dropping file systems caches vs fsstress an repeated reading of files.
> +# Based on generic/579 with the fsverity bits stripped out, and an increased
> +# number of runs.
> +#
> +. ./common/preamble
> +_begin_fstest auto stress
> +
> +stop_run()
> +{
> +	# Stop all subprocesses.
> +	_kill_fsstress
> +	touch $tmp.done
> +	wait
> +}
> +
> +_cleanup()
> +{
> +	stop_run
> +	rm -f $tmp.*
> +}
> +
> +. ./common/filter
> +
> +fsv_file_size=10000000
> +nproc_enabler=$((4 * LOAD_FACTOR))
> +nproc_reader=$((6 * LOAD_FACTOR))
> +nproc_stress=$((3 * LOAD_FACTOR))
> +runtime=$((20 * TIME_FACTOR))
> +
> +run_test()
> +{
> +	rm -rf $tmp.done
> +
> +	# Create the test files and start the fs-verity enabler processes.
> +	for ((proc = 0; proc < nproc_enabler; proc++)); do
> +		orig_file=$SCRATCH_MNT/orig$proc
> +		fsv_file=$SCRATCH_MNT/fsv$proc
> +		head -c $fsv_file_size /dev/urandom > $orig_file
> +		(
> +			while [ ! -e $tmp.done ]; do
> +				rm -f $fsv_file
> +				cp $orig_file $fsv_file
> +				# Give the readers some time to read from the file.
> +				sleep 0.$((RANDOM % 100))
> +			done
> +		) &
> +	done
> +
> +	# Start the reader processes.
> +	for ((proc = 0; proc < nproc_reader; proc++)); do
> +		(
> +			while [ ! -e $tmp.done ]; do
> +				# Choose a random file for each iteration, so that
> +				# sometimes multiple processes read from the same file.
> +				i=$((RANDOM % nproc_enabler))
> +				orig_file=$SCRATCH_MNT/orig$i
> +				fsv_file=$SCRATCH_MNT/fsv$i
> +
> +				# After the copy from $orig_file to $fsv_file has
> +				# completed, the contents of these two files should
> +				# match, regardless of whether verity has been enabled
> +				# or not yet (or is currently being enabled).
> +				cmp $orig_file $fsv_file |& _filter_scratch | \
> +					sed -e "s/'//g" | \
> +					grep -v "SCRATCH_MNT/fsv$i: No such file or directory" | \
> +					grep -v "EOF on SCRATCH_MNT/fsv$i"
> +				done
> +		) &
> +	done
> +
> +	# Start a process that occasionally runs 'sync && drop_caches'.  This makes more
> +	# reads go through fs-verity for real, rather than just returning pagecache.
> +	(
> +		while [ ! -e $tmp.done ]; do
> +			sleep 2.$((RANDOM % 100))
> +			_scratch_sync && echo 3 > /proc/sys/vm/drop_caches
> +		done
> +	) &
> +
> +	# Start the fsstress processes.
> +	_run_fsstress_bg -p $nproc_stress -l 0 -d $SCRATCH_MNT/stressdir
> +
> +	# Run for a while.
> +	sleep $runtime
> +
> +	stop_run
> +}
> +
> +for i in $(seq 1 5); do
> +	echo "Starting run $i" >> $seqres.full
> +
> +	_scratch_mkfs &>> $seqres.full
> +	_scratch_mount
> +	run_test
> +	_scratch_unmount
> +done
> +
> +echo "Silence is golden"
> +
> +# success, all done
> +status=0
> +exit
> diff --git a/tests/generic/4201.out b/tests/generic/4201.out
> new file mode 100644
> index 000000000000..2c3f3d42ff3f
> --- /dev/null
> +++ b/tests/generic/4201.out
> @@ -0,0 +1,2 @@
> +QA output created by 4201
> +Silence is golden
> -- 
> 2.53.0
> 
> 

Re: [PATCH] generic: add a test that stresses metadata eviction

[Zorro Lang]

On Mon, May 18, 2026 at 02:27:42PM +0200, Carlos Maiolino wrote:
> On Mon, May 18, 2026 at 08:24:53AM +0200, Christoph Hellwig wrote:
> > The fsverity test generic/759 could be used to reproduce a regression in
                              ^^^
                              579

> > the xfs buffer hashing.  Carlos Maiolino figured out that even removing
> > the fsverity-specific bits would reproduce the issue.  This test is
> > thus a copy of generic/759 with the fsverity bits striped out, and
> > run 5 times as that seems to be the threshold at which the issue
> > reproduces reliably.
> > 
> > Signed-off-by: Christoph Hellwig <hch@lst.de>
> > ---
> > 
> > Not sure how to best credit Eric for the original patch and Carlos
> > for the extraction, any ideas?
> 
> To be fair, Andrey initially suggested stripping off the fs-verity
> parts, IIRC because he wanted to isolate his own series from this crash.
> My role was mostly reproduce it and hunt the bug down.
> Risking to come with yet another tag, perhaps something like:
> 
> Based-on-generic-579-by: <Eric...>
> 
> It doesn't look ugly to me, but who knows...

Hmm... Generally, we state which case was copied/forked from in the introductory
comments of the case and in the commit log. If you want to specifically mention
the original author, I think we can add a dedicated line in the commit log, for
example:

  'This case is copied from generic/579, which was written by Eric Biggers.'

Currently, we don’t have a formal commit trailer to define this specific scenario,
but perhaps this is a good opportunity for us to discuss and agree on one. After
all, copying and modifying existing cases happens quite often in xfstests. For
example, we could specify the source case in the commit body, and then add a
commit trailer likes "Based-on-code-by", the whole things like:

   This case is copied by generic/579, I removed the fsverity-specific part to
   reproduce ....
   ...

   Based-on-code-by: Eric Biggers <ebiggers@google.com>
   Signed-off-by: Christoph Hellwig <hch@lst.de>

What do you and others think?

> 
> > 
> >  tests/generic/4201     | 108 +++++++++++++++++++++++++++++++++++++++++
> >  tests/generic/4201.out |   2 +
> >  2 files changed, 110 insertions(+)
> >  create mode 100755 tests/generic/4201
> >  create mode 100644 tests/generic/4201.out
> > 
> > diff --git a/tests/generic/4201 b/tests/generic/4201
> > new file mode 100755
> > index 000000000000..fc367beeadaf
> > --- /dev/null
> > +++ b/tests/generic/4201
> > @@ -0,0 +1,108 @@
> > +#! /bin/bash
> > +# SPDX-License-Identifier: GPL-2.0
> > +# Copyright 2019 Google LLC

I saw you tried to keep the original copyright of generic/579. I'm not sure if
we can have two Copyright lines at here, one is yours, the other is the original
one?

Thanks,
Zorro

> > +#
> > +# FS QA Test generic/4201
> > +#
> > +# Race dropping file systems caches vs fsstress an repeated reading of files.
> > +# Based on generic/579 with the fsverity bits stripped out, and an increased
> > +# number of runs.
> > +#
> > +. ./common/preamble
> > +_begin_fstest auto stress
> > +
> > +stop_run()
> > +{
> > +	# Stop all subprocesses.
> > +	_kill_fsstress
> > +	touch $tmp.done
> > +	wait
> > +}
> > +
> > +_cleanup()
> > +{
> > +	stop_run
> > +	rm -f $tmp.*
> > +}
> > +
> > +. ./common/filter
> > +
> > +fsv_file_size=10000000
> > +nproc_enabler=$((4 * LOAD_FACTOR))
> > +nproc_reader=$((6 * LOAD_FACTOR))
> > +nproc_stress=$((3 * LOAD_FACTOR))
> > +runtime=$((20 * TIME_FACTOR))
> > +
> > +run_test()
> > +{
> > +	rm -rf $tmp.done
> > +
> > +	# Create the test files and start the fs-verity enabler processes.
> > +	for ((proc = 0; proc < nproc_enabler; proc++)); do
> > +		orig_file=$SCRATCH_MNT/orig$proc
> > +		fsv_file=$SCRATCH_MNT/fsv$proc
> > +		head -c $fsv_file_size /dev/urandom > $orig_file
> > +		(
> > +			while [ ! -e $tmp.done ]; do
> > +				rm -f $fsv_file
> > +				cp $orig_file $fsv_file
> > +				# Give the readers some time to read from the file.
> > +				sleep 0.$((RANDOM % 100))
> > +			done
> > +		) &
> > +	done
> > +
> > +	# Start the reader processes.
> > +	for ((proc = 0; proc < nproc_reader; proc++)); do
> > +		(
> > +			while [ ! -e $tmp.done ]; do
> > +				# Choose a random file for each iteration, so that
> > +				# sometimes multiple processes read from the same file.
> > +				i=$((RANDOM % nproc_enabler))
> > +				orig_file=$SCRATCH_MNT/orig$i
> > +				fsv_file=$SCRATCH_MNT/fsv$i
> > +
> > +				# After the copy from $orig_file to $fsv_file has
> > +				# completed, the contents of these two files should
> > +				# match, regardless of whether verity has been enabled
> > +				# or not yet (or is currently being enabled).
> > +				cmp $orig_file $fsv_file |& _filter_scratch | \
> > +					sed -e "s/'//g" | \
> > +					grep -v "SCRATCH_MNT/fsv$i: No such file or directory" | \
> > +					grep -v "EOF on SCRATCH_MNT/fsv$i"
> > +				done
> > +		) &
> > +	done
> > +
> > +	# Start a process that occasionally runs 'sync && drop_caches'.  This makes more
> > +	# reads go through fs-verity for real, rather than just returning pagecache.
> > +	(
> > +		while [ ! -e $tmp.done ]; do
> > +			sleep 2.$((RANDOM % 100))
> > +			_scratch_sync && echo 3 > /proc/sys/vm/drop_caches
> > +		done
> > +	) &
> > +
> > +	# Start the fsstress processes.
> > +	_run_fsstress_bg -p $nproc_stress -l 0 -d $SCRATCH_MNT/stressdir
> > +
> > +	# Run for a while.
> > +	sleep $runtime
> > +
> > +	stop_run
> > +}
> > +
> > +for i in $(seq 1 5); do
> > +	echo "Starting run $i" >> $seqres.full
> > +
> > +	_scratch_mkfs &>> $seqres.full
> > +	_scratch_mount
> > +	run_test
> > +	_scratch_unmount
> > +done
> > +
> > +echo "Silence is golden"
> > +
> > +# success, all done
> > +status=0
> > +exit
> > diff --git a/tests/generic/4201.out b/tests/generic/4201.out
> > new file mode 100644
> > index 000000000000..2c3f3d42ff3f
> > --- /dev/null
> > +++ b/tests/generic/4201.out
> > @@ -0,0 +1,2 @@
> > +QA output created by 4201
> > +Silence is golden
> > -- 
> > 2.53.0
> > 
> > 

Re: [PATCH] generic: add a test that stresses metadata eviction

[Christoph Hellwig]

On Tue, May 19, 2026 at 03:03:37AM +0800, Zorro Lang wrote:
> On Mon, May 18, 2026 at 02:27:42PM +0200, Carlos Maiolino wrote:
> > On Mon, May 18, 2026 at 08:24:53AM +0200, Christoph Hellwig wrote:
> > > The fsverity test generic/759 could be used to reproduce a regression in
>                               ^^^
>                               579

Yeah.

> Hmm... Generally, we state which case was copied/forked from in the introductory
> comments of the case and in the commit log. If you want to specifically mention
> the original author, I think we can add a dedicated line in the commit log, for
> example:

If Eric is fine with that I can stick to the normal rule.

> What do you and others think?

I don't really care at all.  I just want to make sure I don't smub
everyone involved with this the wrong way :)

Re: [PATCH] generic: add a test that stresses metadata eviction

[Theodore Tso]

On Tue, May 19, 2026 at 03:03:37AM +0800, Zorro Lang wrote:
> Currently, we don’t have a formal commit trailer to define this
> specific scenario, but perhaps this is a good opportunity for us to
> discuss and agree on one.

My suggestion is that this should go in a comment of the forked test, e.g.

# Based on generic/579 with fsverity-specific bits removed

That way, if someone fixes a problem in generic/579 in the future, a
quick grep might turn up other tests that need a similar adjustment.
The commit trailer doesn't seem as useful to me since it's harder to
find.

							- Ted

Re: [PATCH] generic: add a test that stresses metadata eviction

[Eric Biggers]

On Tue, May 19, 2026 at 10:37:58AM +0200, Christoph Hellwig wrote:
> On Tue, May 19, 2026 at 03:03:37AM +0800, Zorro Lang wrote:
> > On Mon, May 18, 2026 at 02:27:42PM +0200, Carlos Maiolino wrote:
> > > On Mon, May 18, 2026 at 08:24:53AM +0200, Christoph Hellwig wrote:
> > > > The fsverity test generic/759 could be used to reproduce a regression in
> >                               ^^^
> >                               579
> 
> Yeah.
> 
> > Hmm... Generally, we state which case was copied/forked from in the introductory
> > comments of the case and in the commit log. If you want to specifically mention
> > the original author, I think we can add a dedicated line in the commit log, for
> > example:
> 
> If Eric is fine with that I can stick to the normal rule.
> 
> > What do you and others think?
> 
> I don't really care at all.  I just want to make sure I don't smub
> everyone involved with this the wrong way :)

The patch looks fine as-is (other than the typo already pointed out).
It already preserves the original copyright line and has a comment
"# Based on generic/579 with the fsverity bits stripped out".  I don't
need any additional credit.  Thanks,

- Eric

Re: [PATCH] generic: add a test that stresses metadata eviction

[Zorro Lang]

On Tue, May 19, 2026 at 08:57:04AM -0400, Theodore Tso wrote:
> On Tue, May 19, 2026 at 03:03:37AM +0800, Zorro Lang wrote:
> > Currently, we don’t have a formal commit trailer to define this
> > specific scenario, but perhaps this is a good opportunity for us to
> > discuss and agree on one.
> 
> My suggestion is that this should go in a comment of the forked test, e.g.
> 
> # Based on generic/579 with fsverity-specific bits removed
> 
> That way, if someone fixes a problem in generic/579 in the future, a
> quick grep might turn up other tests that need a similar adjustment.
> The commit trailer doesn't seem as useful to me since it's harder to
> find.

Thanks Ted, since Eric agreed with this as well and there are no other
objections, I think we can hold off on creating a new git trailer. Let's
just proceed as usual: mention the original case number in the commit log,
and also add it to the case's initial comments would make it easier to track.

Thanks,
Zorro

> 
> 							- Ted
> 

Re: [PATCH] generic: add a test that stresses metadata eviction

[Zorro Lang]

On Tue, May 19, 2026 at 10:37:58AM +0200, Christoph Hellwig wrote:
> On Tue, May 19, 2026 at 03:03:37AM +0800, Zorro Lang wrote:
> > On Mon, May 18, 2026 at 02:27:42PM +0200, Carlos Maiolino wrote:
> > > On Mon, May 18, 2026 at 08:24:53AM +0200, Christoph Hellwig wrote:
> > > > The fsverity test generic/759 could be used to reproduce a regression in
> >                               ^^^
> >                               579
> 
> Yeah.
> 
> > Hmm... Generally, we state which case was copied/forked from in the introductory
> > comments of the case and in the commit log. If you want to specifically mention
> > the original author, I think we can add a dedicated line in the commit log, for
> > example:
> 
> If Eric is fine with that I can stick to the normal rule.
> 
> > What do you and others think?
> 
> I don't really care at all.  I just want to make sure I don't smub
> everyone involved with this the wrong way :)

I've merged this patch with above typo fixes. Please feel free to check
patches-in-queue branch.

Thanks,
Zorro

> 
>