Folks. I suspect that this isn't actually a ZFS bug but a general memory manager problem. I would appreciate input from mm folks. I parked the detailed bug report here for now: https://github.com/zfsonlinux/zfs/issues/8396
Short story: When running a ZFS on a Linux 4.19/4.20 NFS server the clients are occasionally unsuccessful in perfectly normal page-in on mapped files. Example: executable on NFS mounted filesystem. Some executable-mapped page is referenced. The page is supposed to be retrieved on demand if not resident yet. This occasionally fails with recent Linux or ZoL code. The semantics of the error are identical to what is legitimately happening when you have a page fault in a NFS client mapped file after that file has been unlinked on the server side. Here it is happening without the unlinking. I suspect this might be a general Linux mm problem because I cross-checked with a FreeBSD server with very similar ZFS code. Although I cannot track when the error started appearing I know it is about within the last 6 months and I read all commits to ZoL that manage pages and couldn't see anything suspicious off-hand. On the other hand the errors do not appear when moving the server side file tree to ext4fs. The errors get more frequent with uptime of the server and are not impressed by drop_caches or by trying to evict ZFS' own caches with memory pressure. Details with a line of reasoning why I blame the server and all other info I am collecting: https://github.com/zfsonlinux/zfs/issues/8396 Thanks Martin -- %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Martin Cracauer <craca...@cons.org> http://www.cons.org/cracauer/
