| Message ID | YYXhNip3PctJAaDY@mit.edu (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | xfs/076 takes a long long time testing with a realtime volume | expand |
On Fri, Nov 05, 2021 at 09:58:14PM -0400, Theodore Ts'o wrote: > After committing some exclusions into my test runner framework (see > below), I tested a potential fix to xfs/076 which disables the > real-time volume when creating the scratch volume. Should I send it > as a formal patch to fstests? Does adding: _xfs_force_bdev data $SCRATCH_MNT right after _scratch_mount make the performance problem go away? Sparse inodes and realtime are a supported configuration. --D > diff --git a/tests/xfs/076 b/tests/xfs/076 > index eac7410e..5628c08f 100755 > --- a/tests/xfs/076 > +++ b/tests/xfs/076 > @@ -60,6 +60,7 @@ _require_xfs_io_command "falloc" > _require_xfs_io_command "fpunch" > _require_xfs_sparse_inodes > > +unset SCRATCH_RTDEV > _scratch_mkfs "-d size=50m -m crc=1 -i sparse" | > _filter_mkfs > /dev/null 2> $tmp.mkfs > . $tmp.mkfs # for isize > > - Ted > > For why this is needed, see the commit description below: > > commit c41ae1cc0b21eafd2858541c0bc195f951c0726c > Author: Theodore Ts'o <tytso@mit.edu> > Date: Fri Nov 5 20:46:19 2021 -0400 > > test-appliance: exclude xfs/076 from the realtime configs > > The xfs/076 test takes two minutes on a normal xfs file system (e.g., > a normal 4k block size file system). However, when there is a > real-time volume attached, this test takes over 80 minutes. The > reason for this seems to be because the test is spending a lot more > time failing to create files due to missing directories. Compare: > > root@xfstests-2:~# ls -sh /results/xfs/results-4k/xfs/076.full > 48K /results/xfs/results-4k/xfs/076.full > root@xfstests-2:~# ls -sh /tmp/realtime-076.full > 25M /tmp/realtime-076.full > > and: > > root@xfstests-2:~# grep "cannot touch" /results/xfs/results-4k/xfs/076.full | wc -l > 656 > root@xfstests-2:~# grep "cannot touch" /tmp/realtime-076.full | wc -l > 327664 > > The failures from 076.full look like this: > > touch: cannot touch '/xt-vdc/offset.21473722368/25659': No space left on device > touch: cannot touch '/xt-vdc/offset.21473656832/0': No such file or directory > touch: cannot touch '/xt-vdc/offset.21473591296/0': No such file or directory > ... > touch: cannot touch '/xt-vdc/offset.196608/0': No such file or directory > touch: cannot touch '/xt-vdc/offset.131072/0': No such file or directory > touch: cannot touch '/xt-vdc/offset.65536/0': No such file or directory > touch: cannot touch '/xt-vdc/offset.0/0': No such file or directory > > What seems to be going on is that xfs/076 tries to create a small > scratch file system --- but when we attach a real-time volume this > balloons the available size of the file system. Of course, that space > can't be used for normal files. As a result, xfs/076 is incorrectly > estimating how many files it needs to create to fill the file system. > > I'm not sure what's the best way to fix this in the test; perhaps the > test should forcibly unset SCRATCH_RTDEV environment variable before > running _scratch_mkfs? Anyway, for now, we'll just skip running > xfs/076 for the xfs/realtime* configs. > > Signed-off-by: Theodore Ts'o <tytso@mit.edu> > > diff --git a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime.exclude b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime.exclude > index a9acba9c..bafce552 100644 > --- a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime.exclude > +++ b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime.exclude > @@ -1,2 +1,7 @@ > # Normal configurations don't support dax > -g dax > + > +# The xfs/076 test takes well over an hour (80 minutes using 100GB GCE > +# PD/SSD) when run with an external realtime device, which triggers > +# the ltm "test is stalled" failsafe which aborts the VM. > +xfs/076 > diff --git a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_28k_logdev.exclude b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_28k_logdev.exclude > index a9acba9c..bafce552 100644 > --- a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_28k_logdev.exclude > +++ b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_28k_logdev.exclude > @@ -1,2 +1,7 @@ > # Normal configurations don't support dax > -g dax > + > +# The xfs/076 test takes well over an hour (80 minutes using 100GB GCE > +# PD/SSD) when run with an external realtime device, which triggers > +# the ltm "test is stalled" failsafe which aborts the VM. > +xfs/076 > diff --git a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_logdev.exclude b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_logdev.exclude > index a9acba9c..bafce552 100644 > --- a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_logdev.exclude > +++ b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_logdev.exclude > @@ -1,2 +1,7 @@ > # Normal configurations don't support dax > -g dax > + > +# The xfs/076 test takes well over an hour (80 minutes using 100GB GCE > +# PD/SSD) when run with an external realtime device, which triggers > +# the ltm "test is stalled" failsafe which aborts the VM. > +xfs/076
On Fri, Nov 05, 2021 at 07:08:04PM -0700, Darrick J. Wong wrote: > On Fri, Nov 05, 2021 at 09:58:14PM -0400, Theodore Ts'o wrote: > > After committing some exclusions into my test runner framework (see > > below), I tested a potential fix to xfs/076 which disables the > > real-time volume when creating the scratch volume. Should I send it > > as a formal patch to fstests? > > Does adding: > > _xfs_force_bdev data $SCRATCH_MNT > > right after _scratch_mount make the performance problem go away? Sparse > inodes and realtime are a supported configuration. The test fails with an "fpunch failed" in 076.out.bad, and nothing enlightening in 076.full. But it does complete in roughly two minutes. - Ted
diff --git a/tests/xfs/076 b/tests/xfs/076 index eac7410e..5628c08f 100755 --- a/tests/xfs/076 +++ b/tests/xfs/076 @@ -60,6 +60,7 @@ _require_xfs_io_command "falloc" _require_xfs_io_command "fpunch" _require_xfs_sparse_inodes +unset SCRATCH_RTDEV _scratch_mkfs "-d size=50m -m crc=1 -i sparse" | _filter_mkfs > /dev/null 2> $tmp.mkfs . $tmp.mkfs # for isize - Ted For why this is needed, see the commit description below: commit c41ae1cc0b21eafd2858541c0bc195f951c0726c Author: Theodore Ts'o <tytso@mit.edu> Date: Fri Nov 5 20:46:19 2021 -0400 test-appliance: exclude xfs/076 from the realtime configs The xfs/076 test takes two minutes on a normal xfs file system (e.g., a normal 4k block size file system). However, when there is a real-time volume attached, this test takes over 80 minutes. The reason for this seems to be because the test is spending a lot more time failing to create files due to missing directories. Compare: root@xfstests-2:~# ls -sh /results/xfs/results-4k/xfs/076.full 48K /results/xfs/results-4k/xfs/076.full root@xfstests-2:~# ls -sh /tmp/realtime-076.full 25M /tmp/realtime-076.full and: root@xfstests-2:~# grep "cannot touch" /results/xfs/results-4k/xfs/076.full | wc -l 656 root@xfstests-2:~# grep "cannot touch" /tmp/realtime-076.full | wc -l 327664 The failures from 076.full look like this: touch: cannot touch '/xt-vdc/offset.21473722368/25659': No space left on device touch: cannot touch '/xt-vdc/offset.21473656832/0': No such file or directory touch: cannot touch '/xt-vdc/offset.21473591296/0': No such file or directory ... touch: cannot touch '/xt-vdc/offset.196608/0': No such file or directory touch: cannot touch '/xt-vdc/offset.131072/0': No such file or directory touch: cannot touch '/xt-vdc/offset.65536/0': No such file or directory touch: cannot touch '/xt-vdc/offset.0/0': No such file or directory What seems to be going on is that xfs/076 tries to create a small scratch file system --- but when we attach a real-time volume this balloons the available size of the file system. Of course, that space can't be used for normal files. As a result, xfs/076 is incorrectly estimating how many files it needs to create to fill the file system. I'm not sure what's the best way to fix this in the test; perhaps the test should forcibly unset SCRATCH_RTDEV environment variable before running _scratch_mkfs? Anyway, for now, we'll just skip running xfs/076 for the xfs/realtime* configs. Signed-off-by: Theodore Ts'o <tytso@mit.edu> diff --git a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime.exclude b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime.exclude index a9acba9c..bafce552 100644 --- a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime.exclude +++ b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime.exclude @@ -1,2 +1,7 @@ # Normal configurations don't support dax -g dax + +# The xfs/076 test takes well over an hour (80 minutes using 100GB GCE +# PD/SSD) when run with an external realtime device, which triggers +# the ltm "test is stalled" failsafe which aborts the VM. +xfs/076 diff --git a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_28k_logdev.exclude b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_28k_logdev.exclude index a9acba9c..bafce552 100644 --- a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_28k_logdev.exclude +++ b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_28k_logdev.exclude @@ -1,2 +1,7 @@ # Normal configurations don't support dax -g dax + +# The xfs/076 test takes well over an hour (80 minutes using 100GB GCE +# PD/SSD) when run with an external realtime device, which triggers +# the ltm "test is stalled" failsafe which aborts the VM. +xfs/076 diff --git a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_logdev.exclude b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_logdev.exclude index a9acba9c..bafce552 100644 --- a/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_logdev.exclude +++ b/kvm-xfstests/test-appliance/files/root/fs/xfs/cfg/realtime_logdev.exclude @@ -1,2 +1,7 @@ # Normal configurations don't support dax -g dax + +# The xfs/076 test takes well over an hour (80 minutes using 100GB GCE +# PD/SSD) when run with an external realtime device, which triggers +# the ltm "test is stalled" failsafe which aborts the VM. +xfs/076