Re: [PATCH v6] fs: Improve eventpoll logging to stop indicting timerfd

[Isaac Manjarres <isaacmanjarres@google.com>]

On Tue, Jul 09, 2024 at 02:04:43PM -0700, Isaac Manjarres wrote:
> On Thu, Jul 04, 2024 at 04:03:59PM +0200, Christian Brauner wrote:
> > On Wed, Jul 03, 2024 at 02:43:14PM GMT, Isaac J. Manjarres wrote:
> > > From: Manish Varma <varmam@google.com>
> > > 
> > > We'll often see aborted suspend operations that look like:
> > > 
> > >  PM: suspend entry 2024-07-03 15:55:15.372419634 UTC
> > >  PM: PM: Pending Wakeup Sources: [timerfd]
> > >  Abort: Pending Wakeup Sources: [timerfd]
> > >  PM: suspend exit 2024-07-03 15:55:15.445281857 UTC
> > > 
> > > From this, it seems a timerfd caused the abort, but that can be
> > > confusing, as timerfds don't create wakeup sources. However,
> > > eventpoll can, and when it does, it names them after the underlying
> > > file descriptor. Unfortunately, all the file descriptors are called
> > > "[timerfd]", and a system may have many timerfds, so this isn't very
> > > useful to debug what's going on to cause the suspend to abort.
> > > 
> > > To improve this, change the way eventpoll wakeup sources are named:
> > > 
> > > 1) The top-level per-process eventpoll wakeup source is now named
> > > "epollN:P" (instead of just "eventpoll"), where N is a unique ID token,
> > > and P is the PID of the creating process.
> > > 
> > > 2) Individual eventpoll item wakeup sources are now named
> > > "epollitemN:P.F", where N is a unique ID token, P is PID of the creating
> > > process, and F is the name of the underlying file descriptor.
> > 
> > Fyi, that PID is meaningless or even actively misleading in the face of
> > pid namespaces. And since such wakeups seem to be registered in sysfs
> > globally they are visible to all containers. That means a container will
> > now see some timerfd wakeup source with a PID that might just accidently
> > correspond to a process inside the container. Which in turn also means
> Thanks for your feedback on this, Christian. With regards to this
> scenario: would it be useful to use a namespace ID, along with the PID,
> to uniquely identify the process? If not, do you have a suggestion for
> this?
> 
> I understand that the proposed naming scheme has a chance of causing
> collisions, however, it is still an improvement over the existing
> naming scheme in terms of being able to attribute wakeups to a
> particular application.
> 
> > you're leaking the info about the creating process into the container.
> > IOW, if PID 1 ends up registering some wakeup source the container gets
> > to know about it.
> Is there a general security concern about this? If not, can you please
> elaborate why this is a problem?
> 
Hey Christian,

I just wanted to follow-up to see if you had a chance to go through my
questions above?

Thanks,
Isaac