| Message ID | 162561727244.543423.13321546742830675478.stgit@locust (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | fstests: random fixes | expand |
On 7/6/21 5:21 PM, Darrick J. Wong wrote: > From: Darrick J. Wong <djwong@kernel.org> > > This test tries to exploit an interaction between delayed allocation and > writeback on full filesystems to see if it can trip up the filestreams > allocator. The behaviors do not present if the filesystem allocates > space at write time, so disable it under these scenarios. Ok, makes sense Reviewed-by: Allison Henderson <allison.henderson@oracle.com> > > Signed-off-by: Darrick J. Wong <djwong@kernel.org> > --- > tests/xfs/172 | 30 +++++++++++++++++++++++++++++- > 1 file changed, 29 insertions(+), 1 deletion(-) > > > diff --git a/tests/xfs/172 b/tests/xfs/172 > index 0d1b441e..c0495305 100755 > --- a/tests/xfs/172 > +++ b/tests/xfs/172 > @@ -16,9 +16,37 @@ _begin_fstest rw filestreams > > # real QA test starts here > _supported_fs xfs > - > +_require_command "$FILEFRAG_PROG" filefrag > _require_scratch > > +# The first _test_streams call sets up the filestreams allocator to fail and > +# then checks that it actually failed. It does this by creating a very small > +# filesystem, writing a lot of data in parallel to separate streams, and then > +# flushes the dirty data, also in parallel. To trip the allocator, the test > +# relies on writeback combining adjacent dirty ranges into large allocation > +# requests which eventually bleed across AGs. This happens either because the > +# big writes are slow enough that filestreams contexts expire between > +# allocation requests, or because the AGs are so full at allocation time that > +# the bmapi allocator decides to scan for a less full AG. Either way, stream > +# directories share AGs, which is what the test considers a success. > +# > +# However, this only happens if writes use the delayed allocation code paths. > +# If the kernel allocates small amounts of space at the time of each write() > +# call, the successive small allocations never trip the bmapi allocator's > +# rescan thresholds and will keep pushing out the expiration time, with the > +# result that the filestreams allocator succeeds in maintaining the streams. > +# The test considers this a failure. > +# > +# Make sure that a regular buffered write produces delalloc reservations. > +# This effectively disables the test for files with extent size hints or DAX > +# mode set. > +_scratch_mkfs > $seqres.full > +_scratch_mount > +$XFS_IO_PROG -f -c 'pwrite 0 64k' $SCRATCH_MNT/testy &> /dev/null > +$FILEFRAG_PROG -v $SCRATCH_MNT/testy 2>&1 | grep -q delalloc || \ > + _notrun "test requires delayed allocation buffered writes" > +_scratch_unmount > + > _check_filestreams_support || _notrun "filestreams not available" > > # test reaper works by setting timeout low. Expected to fail >
diff --git a/tests/xfs/172 b/tests/xfs/172 index 0d1b441e..c0495305 100755 --- a/tests/xfs/172 +++ b/tests/xfs/172 @@ -16,9 +16,37 @@ _begin_fstest rw filestreams # real QA test starts here _supported_fs xfs - +_require_command "$FILEFRAG_PROG" filefrag _require_scratch +# The first _test_streams call sets up the filestreams allocator to fail and +# then checks that it actually failed. It does this by creating a very small +# filesystem, writing a lot of data in parallel to separate streams, and then +# flushes the dirty data, also in parallel. To trip the allocator, the test +# relies on writeback combining adjacent dirty ranges into large allocation +# requests which eventually bleed across AGs. This happens either because the +# big writes are slow enough that filestreams contexts expire between +# allocation requests, or because the AGs are so full at allocation time that +# the bmapi allocator decides to scan for a less full AG. Either way, stream +# directories share AGs, which is what the test considers a success. +# +# However, this only happens if writes use the delayed allocation code paths. +# If the kernel allocates small amounts of space at the time of each write() +# call, the successive small allocations never trip the bmapi allocator's +# rescan thresholds and will keep pushing out the expiration time, with the +# result that the filestreams allocator succeeds in maintaining the streams. +# The test considers this a failure. +# +# Make sure that a regular buffered write produces delalloc reservations. +# This effectively disables the test for files with extent size hints or DAX +# mode set. +_scratch_mkfs > $seqres.full +_scratch_mount +$XFS_IO_PROG -f -c 'pwrite 0 64k' $SCRATCH_MNT/testy &> /dev/null +$FILEFRAG_PROG -v $SCRATCH_MNT/testy 2>&1 | grep -q delalloc || \ + _notrun "test requires delayed allocation buffered writes" +_scratch_unmount + _check_filestreams_support || _notrun "filestreams not available" # test reaper works by setting timeout low. Expected to fail