Re: [GIT PULL] notification tree: directory events

[Andreas Gruenbacher <agruen@suse.de>]

On Friday 20 August 2010 05:38:17 Eric Paris wrote:
> As to when a listener dies:  You have to define 'dies'.

I meant a process that gets killed or simply exits while there are outstanding 
perm events.  Nothing in the code would wake up the processes blocked on the 
perm event; they will simply be stuck forever.  (They cannot even be killed 
with SIGKILL.)

This is easy to reproduce with a listener that is killed while processing a 
perm event: just run the fanotify example [1] with the new -s option and hit 
^C while it is sleeping.

Bonus points would be awarded to make a process blocked on a listener killable 
with SIGKILL or other lethal signals.

[1] http://git.kernel.org/?p=linux/kernel/git/agruen/fanotify-example.git

The listener will also hang forever in that case; not sure why that is the 
case.

I already told you about this before (http://lkml.org/lkml/2010/8/16/349); not 
sure why everything needs to repeated multiple times until it sinks in.

> If the process just stops respond it is supposed to lock forever.

I agree.

> I was explicitly ask to remove timeouts (even though I've already been ask
> off list to put them back)

I disagree with putting timeouts back in.

> In Linus' tree there is a race in which both processes (the
> listener and the process doing an action waiting on the listener) can
> deadlock and hang forever.  A (much less racy but not right) patch to
> fix this has already been posted.
>
> I have a better version which has worked well for me today which I will
> likely post tomorrow morning after I look over it again.  I'd love your
> feedback.

Do you have a pointer to that?

Thanks,
Andreas