Created attachment 342386 [details, diff] Patch to revert commits Some changes to the memory management code during the Linux 3.6 development cycle can cause a deadlock regression in ZFS. The URL field has a link to the upstream issue that describes precisely why the deadlock occurs. I am working on a fix, but it will take 3 to 4 weeks at the minimum before it is ready. Would it be possible to revert the following commits in gentoo-sources and hardened-sources until I have a fix ready for the tree? https://github.com/torvalds/linux/commit/c3b94f44fcb0725471ecebb701c077a0ed67bd07 https://github.com/torvalds/linux/commit/e62e384e9da8d9a0c599795464a7e76fd490931c These patches only benefit kernels built with CONFIG_MEMCG (default N in Kconfig) when memory cgroups are actually used, are given small quantities of RAM and have significant disk IO occur in them. As far as I can tell, nothing in the main tree uses this and the one user on the forums that appears to use memory cgroups does not use them under conditions that would trigger the issue that these patches are meant to solve.
Created attachment 342410 [details, diff] Patch to revert commits There was a minor typo in the previous patch's comment that broke the upstream URL. I have attached a revised version that resolves it.
+ 17 Mar 2013; Tom Wijsman <TomWij@gentoo.org> +gentoo-sources-3.2.40-r1.ebuild, + +gentoo-sources-3.6.11-r2.ebuild, +gentoo-sources-3.7.10-r1.ebuild, + +gentoo-sources-3.8.3-r1.ebuild, -gentoo-sources-3.0.67.ebuild, + -gentoo-sources-3.0.68.ebuild, -gentoo-sources-3.2.39.ebuild, + -gentoo-sources-3.2.40.ebuild, -gentoo-sources-3.4.35.ebuild, + -gentoo-sources-3.8.1.ebuild, -gentoo-sources-3.8.2.ebuild, + -gentoo-sources-3.8.3.ebuild: + Revision bumps for 3.2.40, 3.6.11, 3.7.10 and 3.8.3 and removal of + intermediate versions: Revert memcg patches that prevent OOM with too many + dirty pages which caused ZFS to flush dirty pages to deadlock, see bug + #462066; in 3.6, 3.7 and 3.8. Backported fixes from 3.8 to 3.7 that affect + some of our Gentoo users. Backported one of these patches to 3.2 as well since + there was no 3.2 bump last week.
This was resolved in sys-fs/zfs-kmod-0.6.4.