[linux-cifs-client] Re: [PATCH] CIFS: make cifsd (more) signal-safe

[Christoph Hellwig]

On Tue, Jun 05, 2007 at 03:23:40PM -0400, Jeff Layton wrote:
> I recently sent a similar, smaller patch for this problem. After some
> discussion with Steve and Shaggy, I think I better understand why cifsd
> allows signals through, and I realize that my earlier patch wasn't
> comprehensive enough
> 
> The mount and unmount calls will send a KILL signal to cifsd to wake it
> up if it happens to be blocked in kernel_recvmsg. The problem is that
> it doesn't distinguish between "legitimate" signals sent for this
> reason and spurious signals sent by a userspace process (for instance).
> While this is definitely a "don't do that" sort of situation, we might
> as well try to have cifsd be as signal-safe as possible.
> 
> The following patch does this by making sure that we set tcpStatus to
> CifsExiting before sending cifsd a signal, and having cifsd check for
> that when it sees that it's been signalled. If the tcpStatus is not set
> correctly, it ignores it, flushes signals and moves on.
> 
> I've tested a similar backported version of this on an earlier kernel,
> but have not tested this particular patch as of yet.

The right way to fix this is to stop sending signals at all and have
a kernel-internal way to get out of kernel_recvmsg.  Uses of signals by
kernel thread generally are bugs.