On Fri, Jan 11, 2019 at 08:31:24AM -0500, Brian Foster wrote: > XFS has a bug where page writeback can end up sending data to the > wrong location due to a stale, cached file mapping. Add a test to > trigger this problem by racing background writeback with a > truncate/rewrite of the final page of the file. > > Signed-off-by: Brian Foster <bfoster@xxxxxxxxxx> > --- > > Hi all, > > This is a resend of an old post[1] that never quite made it upstream. It > wasn't a big deal at the time because we didn't really have a proper fix > for the problem. I'm resending now because there is a proposed fix[2]. Thanks for the resending! > > I've verified that this still reproduces the problem and no longer fails > with the fix applied (in hundreds of iters). Note that reproduction may > require many iterations. It took me anywhere from 5 to 30 or so on the > box I tested, which I think is reasonable for the tradeoff of a fairly > quick test. There was some discussion on the original post around making > the test run longer for a more reliable reproducer, but I'm not sure how > valuable that is given this is a targeted regression test. Thoughts > appreciated. It took me around 5 iterations to hit the corruption, I think it's fine. But a couple of things changed over the years :) > > Brian > > [1] https://marc.info/?l=fstests&m=150902929900510&w=2 > [2] https://marc.info/?l=linux-xfs&m=154721212321112&w=2 > > tests/generic/999 | 94 +++++++++++++++++++++++++++++++++++++++++++ > tests/generic/999.out | 2 + > tests/generic/group | 1 + > 3 files changed, 97 insertions(+) > create mode 100755 tests/generic/999 > create mode 100644 tests/generic/999.out > > diff --git a/tests/generic/999 b/tests/generic/999 > new file mode 100755 > index 00000000..9e56a1e0 > --- /dev/null > +++ b/tests/generic/999 > @@ -0,0 +1,94 @@ > +#! /bin/bash > +# FS QA Test 999 > +# > +# Test XFS page writeback code for races with the cached file mapping. XFS > +# caches the file -> block mapping for a full extent once it is initially looked > +# up. The cached mapping is used for all subsequent pages in the same writeback > +# cycle that cover the associated extent. Under certain conditions, it is > +# possible for concurrent operations on the file to invalidate the cached > +# mapping without the knowledge of writeback. Writeback ends up sending I/O to a > +# partly stale mapping and potentially leaving delalloc blocks in the current > +# mapping unconverted. > +# > +#----------------------------------------------------------------------- > +# Copyright (c) 2017 Red Hat, Inc. All Rights Reserved. ^^^^ 2019? > +# > +# This program is free software; you can redistribute it and/or > +# modify it under the terms of the GNU General Public License as > +# published by the Free Software Foundation. > +# > +# This program is distributed in the hope that it would be useful, > +# but WITHOUT ANY WARRANTY; without even the implied warranty of > +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > +# GNU General Public License for more details. > +# > +# You should have received a copy of the GNU General Public License > +# along with this program; if not, write the Free Software Foundation, > +# Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA > +#----------------------------------------------------------------------- And please change this to SPDX-License-Identifier. > +# > + > +seq=`basename $0` > +seqres=$RESULT_DIR/$seq > +echo "QA output created by $seq" > + > +here=`pwd` > +tmp=/tmp/$$ > +status=1 # failure is the default! > +trap "_cleanup; exit \$status" 0 1 2 3 15 > + > +_cleanup() > +{ > + cd / > + rm -f $tmp.* > +} > + > +# get standard environment, filters and checks > +. ./common/rc > + > +# remove previous $seqres.full before test > +rm -f $seqres.full > + > +# real QA test starts here > + > +# Modify as appropriate. > +_supported_fs generic > +_supported_os Linux > +_require_scratch > +_require_test_program "feature" _require_xfs_io_command "sync_range" > + > +_scratch_mkfs >> $seqres.full 2>&1 || _fail "mkfs failed" > +_scratch_mount || _fail "mount failed" _scratch_mount will _fail the test on failure now :) > + > +file=$SCRATCH_MNT/file > +filesize=$((1024 * 1024 * 32)) > +pagesize=`src/feature -s` > +truncsize=$((filesize - pagesize)) > + > +for i in $(seq 0 15); do > + # Truncate the file and fsync to persist the final size on-disk. This is > + # required so the subsequent truncate will not wait on writeback. > + $XFS_IO_PROG -fc "truncate 0" $file > + $XFS_IO_PROG -c "truncate $filesize" -c fsync $file > + > + # create a small enough delalloc extent to likely be contiguous > + $XFS_IO_PROG -c "pwrite 0 $filesize" $file >> $seqres.full 2>&1 > + > + # Start writeback and a racing truncate and rewrite of the final page. > + $XFS_IO_PROG -c "sync_range -w 0 0" $file & > + sync_pid=$! > + $XFS_IO_PROG -c "truncate $truncsize" \ > + -c "pwrite $truncsize $pagesize" $file >> $seqres.full 2>&1 > + > + # If the test fails, the most likely outcome is an sb_fdblocks mismatch > + # and/or an associated delalloc assert failure on inode reclaim. Cycle > + # the mount to trigger detection. > + wait $sync_pid > + _scratch_cycle_mount || _fail "mount failed" And _scratch_cycle_mount will exit the test on failure as well. Thanks, Eryu > +done > + > +echo Silence is golden > + > +# success, all done > +status=0 > +exit > diff --git a/tests/generic/999.out b/tests/generic/999.out > new file mode 100644 > index 00000000..3b276ca8 > --- /dev/null > +++ b/tests/generic/999.out > @@ -0,0 +1,2 @@ > +QA output created by 999 > +Silence is golden > diff --git a/tests/generic/group b/tests/generic/group > index ea5aa7aa..ce165981 100644 > --- a/tests/generic/group > +++ b/tests/generic/group > @@ -525,3 +525,4 @@ > 520 auto quick log > 521 soak long_rw > 522 soak long_rw > +999 auto quick > -- > 2.17.2 >
