Some further information:

When the deadlock occurs, the client is prefetching
into the client cache.  There is heavy VM pressure.
When the writeback to disk is about to happen, a page
is needed and this causes kswapd to wake up to try to
free some pages.  Unfortunately, it chooses a page
in the cache file being prefetched, causing a deadlock.

It looks like a lot of locking changes were made in 1.3
in afs_GetDCache for other reasons -- I guess something
may have to change in there to fix this problem.
--
Joe Buehler

_______________________________________________
OpenAFS-devel mailing list
[EMAIL PROTECTED]
https://lists.openafs.org/mailman/listinfo/openafs-devel

Reply via email to