Re: [3.1-rc4] NFSv3 client hang

[Trond Myklebust]

On Fri, 2011-09-09 at 12:45 -0700, Simon Kirby wrote: 
> The 3.1-rc4 NFSv3 client hung on another box (separate from the other one
> which Oopsed in vfs_rmdir() with similar workload). This build was also
> of 9e79e3e9dd9672b37ac9412e9a926714306551fe (slightly past 3.1-rc4), and
> "git log 9e79e3e9dd96.. fs/nfs net/sunrpc" is empty.
> 
> All mounts to one server IP have hung, while all other mounts work fine.
> I ran "cd /proc/sys/sunrpc; echo 255 > rpc_debug; echo 255 > nfs_debug"
> for a while, then kill -9'd all D-state processes to simplify the
> debugging, and was left with one that was not interruptible:
> 
> 28612 D    /usr/local/apache2/bin/http sleep_on_page
> # cat /proc/28612/stack
> [<ffffffff810bdf49>] sleep_on_page+0x9/0x10
> [<ffffffff810bdf34>] __lock_page+0x64/0x70
> [<ffffffff8112a9e5>] __generic_file_splice_read+0x2d5/0x500
> [<ffffffff8112ac5a>] generic_file_splice_read+0x4a/0x90
> [<ffffffff812030e5>] nfs_file_splice_read+0x85/0xd0
> [<ffffffff81128fb2>] do_splice_to+0x72/0xa0
> [<ffffffff811297e4>] splice_direct_to_actor+0xc4/0x1d0
> [<ffffffff81129942>] do_splice_direct+0x52/0x70
> [<ffffffff81100096>] do_sendfile+0x166/0x1d0
> [<ffffffff81100185>] sys_sendfile64+0x85/0xb0
> [<ffffffff816af57b>] system_call_fastpath+0x16/0x1b
> [<ffffffffffffffff>] 0xffffffffffffffff
> 
> echo 1 > /proc/sys/sunrpc/rpc_debug emits:
> 
> -pid- flgs status -client- --rqstp- -timeout ---ops--
> 37163 0001    -11 ffff8802251bca00   (null)        0 ffffffff816e4110 nfsv3 READ a:call_reserveresult q:xprt_sending
> 
> tcpdump to this server shows absolutely no packets to the server IP for
> several minutes. netstat shows the socket in CLOSE_WAIT:
> 
> # netstat -tan|grep 2049
> tcp        0      0 10.10.52.50:806         10.10.52.230:2049       CLOSE_WAIT
> 
> This is the only port-2049 socket that still exists.
> rpcinfo -p 10.10.52.230, -t 10.10.52.230 lockmgr, etc., all show the
> server seems fine. rpciod is sleeping in rescuer_thread, and nothing
> else is in D state.
> 
> mount opts were "rw,hard,intr,tcp,timeo=300,retrans=2,vers=3"
> 
> Running another "df" on the mountpoint with rpc_debug = 255 shows:
> 
> -pid- flgs status -client- --rqstp- -timeout ---ops--
> 37163 0001    -11 ffff8802251bca00   (null)        0 ffffffff816e4110 nfsv3 READ a:call_reserveresult q:xprt_sending
> RPC:       looking up Generic cred
> NFS call  access
> RPC:       new task initialized, procpid 30679
> RPC:       allocated task ffff880030c17a00
> RPC: 37133 __rpc_execute flags=0x80
> RPC: 37133 call_start nfs3 proc ACCESS (sync)
> RPC: 37133 call_reserve (status 0)
> RPC: 37133 failed to lock transport ffff880223d0a000
> RPC: 37133 sleep_on(queue "xprt_sending" time 4489651610)
> RPC: 37133 added to queue ffff880223d0a178 "xprt_sending"
> RPC: 37133 sync task going to sleep
> 
> So something is not closing the old transport socket here?

Hi Simon,

I've stared at this for some time now, and I'm having trouble seeing
what could be going wrong. Can you therefore please apply the following
patch, and see if you can reproduce the problem?

Cheers
  Trond
8<-------------------------------------------------------------------------