While trying to figure out why we keep getting TCP NFSv3 client mounts in
CLOSE_WAIT state, I notice that the socket closes and the host becomes
remountable if I kill all D-state processes blocking on it. However, some
Apache processes are often uninterruptible:
# cat /proc/8959/stack
[<ffffffff810d6b69>] sleep_on_page+0x9/0x10
[<ffffffff810d6b54>] __lock_page+0x64/0x70
[<ffffffff81147f5d>] __generic_file_splice_read+0x2dd/0x500
[<ffffffff811481cd>] generic_file_splice_read+0x4d/0x90
[<ffffffff811f70b5>] nfs_file_splice_read+0x85/0xd0
[<ffffffff81146522>] do_splice_to+0x72/0xa0
[<ffffffff81146d54>] splice_direct_to_actor+0xc4/0x1d0
[<ffffffff81146eb2>] do_splice_direct+0x52/0x70
[<ffffffff8111c1fe>] do_sendfile+0x16e/0x1e0
[<ffffffff8111c2f5>] sys_sendfile64+0x85/0xb0
[<ffffffff816e1392>] system_call_fastpath+0x16/0x1b
[<ffffffffffffffff>] 0xffffffffffffffff
However, from loking in generic_file_splice_read, it looks like not
something easy to make interruptible....
What causes the sockets to close, anyway? I notice there seems to be some
sort of idle timeout. Is this configured somewhere, so I can set it
really low and reproduce this more easily? I suspect we must have some
race with the socket closing and something trying to use it again...
Simon-