get_file() unsafe under epoll (was Re: [syzbot] [fs?] [io-uring?] general protection fault in __ep_remove)

[Al Viro]

On Fri, May 03, 2024 at 01:28:37PM -0700, Kees Cook wrote:
> 
> Is this the right approach? It still feels to me like get_file() needs
> to happen much earlier...

I don't believe it needs to happen at all.  The problem is not that
->release() can be called during ->poll() - it can't and it doesn't.
It's that this instance of ->poll() is trying to extend the lifetime
of that struct file, when it might very well be past the point of no
return.

What we need is
	* promise that ep_item_poll() won't happen after eventpoll_release_file().
AFAICS, we do have that.
	* ->poll() not playing silly buggers.

As it is, dma_buf ->poll() is very suspicious regardless of that
mess - it can grab reference to file for unspecified interval.
Have that happen shortly before reboot and you are asking for failing
umount.

->poll() must be refcount-neutral wrt file passed to it.  I'm seriously
tempted to make ->poll() take const struct file * and see if there's
anything else that would fall out.