Peter Tribble wrote: > I know that locally mounting using nfs is problematic. I'm trying > to understand why (and what symptoms you're likely to see). > > I think the relevant bug is 4498652, but the content is somewhat > non-existent. > > Could someone explain the problem and likely result, or actually > open up the comments that presumably contains the useful details > in this case? > Hi Peter,
Below's my explanation of the problem. On systems under extreme stress you'll run into a deadlock between NFS/UFS/VM threads which depending on a scenario you hit, can either make the particular client mount/server share inaccessible or deadlock the entire system. The root cause for this really boils down to interaction between UFS and NFS modules through the segmap driver and NFS client doing synchronous commits. There can be several different scenarios which can trigger this. One scenario could go something like this - all the nfs server threads are stuck in rfs3_write() trying to obtain the rwlock for the vnode. The one thread which owns this lock is actually stuck trying to free the pages which happen to be nfs pages. To free the pages, the NFS client needs a commit, which means you need to talk back to the server where you deadlock trying to get the rwlock. ie. UFS write tries to acquire a segmap slot previously used by the NFS client mount: ufs_write() -> segmap_getmapflt() -> segmap_pagefree() -> page_release() -> VN_DISPOSE() and if the page was written but not commited, then the client needs to send a commit to the server. VN_DISPOSE() -> nfs3_dispose() -> nfs3_commit() -> synchronous commit to the nfs server Now if the server happens to be the same system and data that needs to be flushed happens to be associated with the same file, then all commits/writes will stall. nfs-server -> rfs3_commit -> VOP_PUTPAGE() -> ufs_putpage() -> stall. There are several different variations to this and different locks involved in either nfs/ufs/vm modules (so it is not necessary that you deadlock on the same client). This deadlock issue appears to be specific to segmap based filesystems like UFS and should not be an issue with ZFS. In future we may change our supported configuration to enable support for loopback mounting ZFS based filesystems. However until that stage loopback mounting on NFS filesystems is not a supported configuration. Thanks, Mahesh > Thanks, > >
