Re: [fuse-devel] Interrupt vs. original request races

[Miklos S. <mi...@sz...>]

On Thu, May 21, 2015 at 08:51:28AM -0700, Tom Ramsdell wrote:
> Hello.
>
> I’d like to pick up on this as I’ve posted something similar.  There is an
> unfortunate user mode behavior that we’ve been fighting relating to this.
>
> The problem isn’t with the particular request as it is a problem at higher
> levels of software.  In particular, “ls” on a very large direction becomes
> uninterruptible.
>
> What we’ve determined is that our (low level) interrupt callback is invoked
> with a request that’s been completed already.  We can detect that and ignore
> it easily.  The problem we are seeing is that the interrupt is eaten in this
> case - so “ls” continues to enumerate the directory.

Fuse shouldn't "eat" interrupts, in fact taking a signal out of the queue is
something very low level, never done by filesystem code.  So there's probably
something else going on.

"ls" could block, ingore or do something different with the SIGINT signal, but
we'd need to see a strace to see if that is the case.  "ls" on my machine can be
interrupted fine.

Thanks,
Miklos