Hi Ilya, On Tue, Oct 3, 2023 at 5:00 PM Ilya Dryomov <idryomov@xxxxxxxxx> wrote: > > On Tue, Oct 3, 2023 at 1:06 PM Venky Shankar <vshankar@xxxxxxxxxx> wrote: > > > > From: Xiubo Li <xiubli@xxxxxxxxxx> > > > > This reverts commit f7a67b463fb83a4b9b11ceaa8ec4950b8fb7f902. > > > > We have identified an issue in the MDS affecting CephFS users using > > the kernel driver. The issue was first introduced in the octopus > > release that added support for clients to perform asynchronous > > directory operations using the `nowsync` mount option. The issue > > presents itself as an MDS crash resembling (any of) the following > > crashes: > > > > https://tracker.ceph.com/issues/61009 > > https://tracker.ceph.com/issues/58489 > > > > There is no apparent data loss or corruption, but since the underlying > > cause is related to an (operation) ordering issue, the extent of the > > problem could surface in other forms - most likely MDS crashes > > involving preallocated inodes. > > > > The fix is being reviewed and is being worked on priority: > > > > https://github.com/ceph/ceph/pull/53752 > > > > As a workaround, we recommend (kernel) clients be remounted with the > > `wsync` mount option which disables asynchronous directory operations > > (depending on the kernel version being used, the default could be > > `nowsync`). > > > > This change reverts the default, so, async dirops is disabled (by default). > > Hi Venky, > > Given that the fix is now up and being reviewed on priority, does it > still make sense to change the default? > > According to Xiubo, https://tracker.ceph.com/issues/58489 which morphed > into https://tracker.ceph.com/issues/61009 isn't the only concern -- he > also brought up https://tracker.ceph.com/issues/62810. If the move to > revert (change of default) is also prompted by that issue, it should be > described in the patch. Fair enough -- I'll push out with an updated commit message. > > Thanks, > > Ilya > -- Cheers, Venky
