locked vnode / nfs... requires kill -9 in ddb

[David Xu]

On Saturday 21 October 2006 11:57, Kostik Belousov wrote:
> On Sat, Oct 21, 2006 at 08:25:00AM +0800, David Xu wrote:
> > On Thursday 19 October 2006 18:04, Kostik Belousov wrote:
> > > The nfs_reply is sleeping with the PCATCH set. The question is why
> > > SIGTSTP does not cause msleep to return with EINTR.
> >
> > I have not been tracking the thread. but if the thread is sleeping with
> > PCATCH, the SIGTSTP should cause the process to stop unless the signal
> > is masked by sigprocmask or the signal has an action handler been set,
> > this is  a correct behavior.
>
> David,
> as I understand the report, the following happens. The nfs mount point with
> intr option issued the request and waits for the reply. Some vnode locks
> are held while waiting. Code needs to catch the signals to abort the
> operation on user request. It uses msleep with PCATCH. The thread in
> question has td_locks > 0.
>
> The SIGTSTP is delivered, and thread is stopped, while holding vnode lock.
> How this situation shall be handled ? Namely, how to sleep while having the
> ability to safely clean up on attempt of stopping ? Masking SIGTSTP is not
> the option, due to SIGSTOP having the same results and not being blockable.
>
> [Would it be right to stop the threads only on returning from kernel to
> user mode ?]

I know in the case, you want signal to interrupt the thread but don't want a
job control signal to suspend the thread, but a PCATCH flag is not enough to
tell the case. I think we are trying to fix the history problem of RELENG_4 or
earlier.

David Xu