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> --- Note again that this test is expected to occasionally fail on upstream XFS. The fix is undergoing development and review. Brian v2: - Fix up copyright header. - Fix up _require statements and _scratch_mount* error checks. v1: https://marc.info/?l=fstests&m=154721349921715&w=2 tests/generic/999 | 80 +++++++++++++++++++++++++++++++++++++++++++ tests/generic/999.out | 2 ++ tests/generic/group | 1 + 3 files changed, 83 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..0546125e --- /dev/null +++ b/tests/generic/999 @@ -0,0 +1,80 @@ +#! /bin/bash +# SPDX-License-Identifier: GPL-2.0 +# Copyright (c) 2017-2019 Red Hat, Inc. All Rights Reserved. +# +# 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. + +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 >> $seqres.full 2>&1 || _fail "mkfs failed" +_scratch_mount + +file=$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 +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
