On Fri, Nov 30, 2018 at 2:13 PM Zorro Lang <zlang@xxxxxxxxxx> wrote: > > On Fri, Nov 30, 2018 at 12:12:53AM -0500, Murphy Zhou wrote: > > Hi, > > > > Hit a xfs regression issue by generic/095 on overlayfs. > > > > It's easy to reproduce. Tests processes hang there and never return. > > There are some warning in the dmesg. > > > > SIGINT (Ctrl+C) can't kill the tests, neither does SIGTERM (kill). > > However, SIGKILL (kill -9) can clean them up. > > > > This happens when testing on v4 or v5 or reflink, with the same behaviour. > > -m crc=0 > > -m crc=1,finobt=1,rmapbt=0,reflink=0 -i sparse=1 > > -m crc=1,finobt=1,rmapbt=1,reflink=1 -i sparse=1 > > > > This does not happen if ext4 as base fs for overlayfs. > > > > Bisecting points to this commit: > > 4721a601 iomap: dio data corruption and spurious errors when pipes fill > > > > Test pass soon after this commit reverted. > > > > # ps axjf > > 1839 1862 S+ 0:00 | \_ /bin/bash ./bin/single -o -t generic/095 > > 1862 2005 S+ 0:00 | \_ /bin/bash ./check -T -overlay generic/095 > > 2005 2292 S+ 0:00 | \_ /bin/bash ./tests/generic/095 > > 2292 2516 Sl+ 0:01 | \_ /usr/bin/fio /tmp/2292.fio > > 2516 2565 Rs 6:02 | \_ /usr/bin/fio /tmp/2292.fio > > 2516 2566 Rs 6:02 | \_ /usr/bin/fio /tmp/2292.fio > > 2516 2567 Rs 6:02 | \_ /usr/bin/fio /tmp/2292.fio > > 2516 2568 Rs 6:02 | \_ /usr/bin/fio /tmp/2292.fio > > 2516 2569 Rs 6:02 | \_ /usr/bin/fio /tmp/2292.fio > > Nice catch! I didn't tried overlayfs on XFS regression test this time. I never hit > issue on XFS singly, even on glusterfs(XFS underlying). > > I just gave it a try. This bug is reproducible on xfs-linux git tree > xfs-4.20-fixes-2 HEAD. And by strace all fio processes, they all > keep outputing [1]. More details as [2]. Only overlayfs on XFS can > reproduce this bug, overlayfs on Ext4 can't. > Darrick, This is my cue to insert a rant. You already know what I am going to rant about. I cannot force you to add a check -overlay xfstests run to your pull request validation routine. I can offer assistance in any questions you may have and I can offer support for check -overlay infrastructure if it breaks or if it needs improvements. I have checked on several recent point releases that check -overlay does not regress any tests compared to xfs reflink configuration, and when I found a regression (mostly in overlayfs code, but also in xfs code sometimes) I reported and/or fixed it. But I do not have the resources to validate every xfs merge and certainly not xfs-next. There is a large group of tests that is expected to notrun, which makes running the -overlay test a lot faster than any given xfs configuration and IMO running just a single xfs reflink config with -overlay would give a pretty good test coverage. So what do you say?... Thanks, Amir.
