[lustre-announce] Defect Alert - Lustre 2.10.5 (ZFS)

[Peter Jones]

Impact:
Page-unaligned RPC writes might cause data corruption of the unaligned portion.

Scope:
This issue affects deployments running Lustre 2.10.5 with a ZFS backend only (ldiskfs not affected). There are two known triggers:
1) Application doing page-unaligned writes (typically to a 4KB boundary) to an OST that is either low on free space or user being low on quota
2)  When page-unaligned writes are generated significantly faster than OST can absorb them resulting in "write grant" dropping to 0 while userspace is performing page-unaligned writes. If the "zero grant" condition did not appear, page-unaligned writes are fine (even synchronous) because they are converted into full-page RPC writes on the wire.

Recommendations:
While the triggering conditions may only rarely occur, we still recommend that affected sites should immediately apply the patch for LU-11663 - https://review.whamcloud.com/#/c/33748/ - and then perform an audit to see whether any files created or written to within the time since the upgrade have been affected. This fix will also be included in the upcoming 2.10.6 release.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-announce-lustre.org/attachments/20181201/3cd6a08e/attachment.html>