Re: epoll oops.

[Peter Hurley <peter@hurleysoftware.com>]

On 10/14/2013 04:57 PM, Linus Torvalds wrote:
> [ Adding Pekka to verify the SLAB_DESTROY_BY_RCU semantics, and Peter
> Hurley due to the possible tty association ]

> And I see a few worrisome cases. For example, look at "tty_poll()". It
> ends up doing something very similar, except it uses the tty instead
> of sighand. And exactly like the sighand struct, the tty allocation
> lifespan can - thanks to hangup() - be shorter than the file
> allocation lifespan.
>
> Peter? Does a tty hangup end up actually possibly freeing the tty
> struct? Looking at it, I'm starting to think that it only affects
> f_op, and the "struct tty" stays around, in which case this is all
> fine.

The tty_struct is only freed at the completion of the tty's
file_operations .release method (tty_release()). Further, it should
not be possible to advance past the tty_ldisc_release() call in
tty_release() while file operations such as tty_poll() -> poll_wait()
or a tty hangup are in-progress.

[Notwithstanding the above, if some kernel driver failed to acquire
a tty reference, either directly or via tty_port_tty_hangup(), before
hanging up, then the hangup could be racing with the .release(). But
I don't think that's what's happening here.]


On 10/15/2013 11:48 AM, Oleg Nesterov wrote:>> Hmm? There might be other cases..
 >
 > Yes.
 >
 > Dave, perhaps you have vmcore? I have no idea if this is possible or
 > not, but perhaps you can look at eventpoll_release_file's frame and
 > print file->f_op ?

I think Oleg's suggestion is the next diagnostic step.

Regards,
Peter Hurley