We're kind of getting deep into implementation details, but here goes. The routine in question, for those of you following along with source, is afs_FlushActiveVcaches. This does two things. One, it keeps file locks alive. Two, it stores core files and files exported via NFS back to the server (because they don't go through the normal open/close handling). >Could you adopt this sort of incremental scan of the hash table? That's one of the possibilities I had in mind. Another shortcoming of this code is that it will stall on files which have RPCs being performed on them, which may be fixable. A third possible problem I'm still researching. If you can periodically check the kernel variable afs_brsdaemons, especially during the observed hangs, see if it ever goes to zero, or spends a lot of time near 1. If so, you would benefit from kicking up the value you use for "afsd -daemons". Also, during the hang, run cmdebug against the client. That might prove something.
