If you are comfortable rebuilding the kernel (actually simply rebuilding
the module cifs.ko) - why not simply put
socket->sk->sk_rcvtimeo = 7 * HZ;
socket->sk->sk_sndtimeo = 5 * HZ;
in fs/cifs/connect.c in ipv4_connect (and ipv6_connect) before the call to
rc = (*csocket)->ops->connect(*csocket, (struct sockaddr *)
psin_server,
sizeof(struct sockaddr_in), 0);
and see if that is good enough
On Tue, Nov 27, 2012 at 11:27 AM, Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> On Tue, 27 Nov 2012 08:58:41 -0500
> "Matthew M. DeLoera" <mdeloera@xxxxxxxxx> wrote:
>
>> I'm testing in an older kernel - 2.6.15-55-386. While mount sits and waits, I see cifsoplockd and cifsnotifyd, but no cifsd. And no /proc/PID/stack for either. When I kill mount -a, both of those cifs daemons remain running (both sleeping).
>>
>> I then tried later 2.6.24-24-generic kernel and got same behavior.
>>
>> Either way, for the record, cifsnotifyd was already running. Once I started mount, then cifsoplockd started.
>>
>> Now in an even newer kernel, 2.6.32-15.generic, I get much quicker timeout - just over 30 seconds.
>>
>> I still have to work with the older kernels I mentioned, though I'm glad to see the behavior in 32-15. And FYI, I never saw cifsd running in 32-15.
>>
>> Agreed about kill -9. It seems like cifs timeout behavior must have changed along the way, but my kernels span about 6 years, so that doesn't surprise me. But since I am still stuck with the older kernels I mentioned, is there something better that I could do, or just settle for kill -9 and allow enough time for newer kernels to timeout? Is there a way that I might be able to detect what timeout value I might expect?
>>
>> Cheers,
>> - Matthew
>>
>>
>
> Strange...cifsd is the demultiplex thread that handles the receive
> path. It usually also handles connecting the socket, but I don't recall
> if that was always the case as far back as 2.6.15...
The socket connection is made in the thread executing the mount.
Basically we called ipv4_connect or ipv6_connect, which created the socket
then connected to the ip address (which presumably hangs in your case)
then set the rcvtimeout to 7 seconds
(then set the send and rcvbuf sizes) then sent the negotiate, then
returned and launched cifsd which listens and processes responses
from the server
cifsnotifyd and cifsoplockd are not going to be there in 2.6.32,
but are there in 2.6.24 (and don't really matter for this
> In any case, I think your choices boil down to backporting patches to
> lower the timeouts, or somehow working around the problem in userspace.
For RHEL, I thought that the backported fixes were reasonably large for 2.6.15
Wonder if you backported the setting of the socket sndtimeout (and setting
before the socket connect call)
--
Thanks,
Steve
--
To unsubscribe from this list: send the line "unsubscribe linux-cifs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
