On 2018年03月16日 16:19, Eryu Guan wrote: > On Fri, Mar 16, 2018 at 01:17:07PM +0800, Qu Wenruo wrote: >> >> >> On 2018年03月16日 12:01, Eryu Guan wrote: >>> On Wed, Mar 14, 2018 at 05:02:30PM +0800, Qu Wenruo wrote: >>>> Basic test case which triggers fsstress with dm-log-writes, and then >>>> check the fs after each FUA writes. >>>> With needed infrastructure and special handlers for journal based fs. >>> >>> It's not clear to me why the existing infrastructure is not sufficient >>> for the test. It'd be great if you could provide more information and/or >>> background in commit log. >> >> The main problem of current infrastructure is we don't have the >> following things: >> >> 1) Way to take full advantage of dm-log-writes >> The main thing is, we don't have test cases to check each FUA (this >> patch) and flush (later patch after clearing all the RFC comments). >> >> We have some dm-flakey test cases to emulate power loss, but they are >> mostly for fsync. >> Here we are not only testing fsync, but also every superblock update. >> Which should be a super set of dm-flakey tests. >> >> 2) Workaround for journal replay >> In fact, if we only test btrfs, we don't even need such complicated >> work, just 'replay-log --fsck "btrfs check" --check fua' will be >> enough. As btrfs check doesn't report dirty journal (log tree) as >> problem. >> But for journal based fses, their fsck all report dirty journal as >> error, which needs current snapshot works to replay them before >> running fsck. > > And replay-to-fua doesn't guarantee a consistent filesystem state, > that's why we need to mount/umount the target device to replay the > filesystem journal, and to avoid replaying already-replayed-log over and > over again, we create a snapshot of the target device and mount cycle & > fsck the snapshot, right? > > I'm wondering if the overhead of repeatly create & destroy snapshots is > larger than replaying log from start. Maybe snapshots take more time?
This is the interesting part. Repeatedly creating snapshot will become slower and slower, and in fact for ext4/xfs it's pretty slow, so I use the small nops for fsstress (512), as for later snapshot creation it's taking quite a long time. But if we switch back to no snapshot, but pure replay-log method, replay-log itself will become slow too. But at least it can go much further (nops=2048) and still get a somewhat acceptable speed. Either way, we will get slower and slower when the number of operation grow. However personally speaking, the pure replay-log one looks a little better, as I don't need to use LVM to create snapshot. :) > >> >> I would add them in next version if there is no extra comment on this. >> >>> >>>> >>>> Signed-off-by: Qu Wenruo <w...@suse.com> >>>> --- >>>> In my test, xfs and btrfs survives while ext4 would report error during >>>> fsck. >>>> >>>> My current biggest concern is, we abuse $TEST_DEV and mkfs on it all by >>>> ourselves. Not sure if it's allowed. >>> >>> As Amir already replied, that's not allowed, any destructive operations >>> should be done on $SCRATCH_DEV. >> >> Yep, I'm looking for similar case who uses $SCRATCH_DEV as LVM pv do get >> extra device. >> >> Or can we reuse the scratch_dev_pool even for ext4/xfs? > > I think so, IMO pool devices are not limited to btrfs. But I think we > could use a loop device reside on $TEST_DIR? Or if snapshots take longer > time, then we don't need this extra device at all :) That would be much better! Both common/dmlogwrites and the test case will be much simpler. Thanks, Qu > > I have some other comments, will reply to the RFC patch in another > email. > > Thanks, > Eryu > -- > To unsubscribe from this list: send the line "unsubscribe fstests" in > the body of a message to majord...@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html >
signature.asc
Description: OpenPGP digital signature