Re: fanotify permission events on virtual filesystem

[Jan Kara <jack@suse.cz>]

On Wed 20-03-19 17:02:54, Amir Goldstein wrote:
> On Wed, Mar 20, 2019 at 4:30 PM Jan Kara <jack@suse.cz> wrote:
> >
> > On Wed 20-03-19 15:46:20, Amir Goldstein wrote:
> > > On Wed, Mar 20, 2019 at 3:16 PM Jan Kara <jack@suse.cz> wrote:
> > > > recently, one of our customers has reported a deadlock with fanotify. The
> > > > analysis has shown that a process has put (likely by mistake) FAN_OPEN_PERM
> > > > mark on /proc and / filesystem. That resulted in a deadlock like follows:
> > > >
> > > > process 1:                      process 2:              process 3:
> > > > open("/proc/process 2/maps")
> > > >   - blocks waiting for response to
> > > >     FAN_OPEN_PERM event
> > > >
> > > >                                 exec(2)
> > > >                                   __do_execve_file()
> > > >                                     - grabs current->signal->cred_guard_mutex
> > > >                                     do_open_execat()
> > > >                                       - blocks waiting for response to
> > > >                                         FAN_OPEN_PERM event
> > > >
> > > >                                                         reads fanotify event
> > > >                                                         generated by process 1
> > > >                                                           create_fd()
> > > >                                                             dentry_open()
> > > >                                                               proc_maps_open()
> > > >                                                                 blocks on
> > > >                                                 cred_guard_mutex of process 2.
> > > >
> > > > Now this is not the only case where you can setup fanotify permissions
> > > > events so that your listener deadlocks but I'd argue that this case is
> > > > especially nasty and it is unrealistic to expect from userspace that it
> > > > would be able to implement fanotify listener in such a way that is
> > > > deadlock-free for proc filesystem since the lock dependencies there are
> > > > very different. So how about we just forbid placing marks with fanotify
> > > > permission events on proc? Any other virtual filesystem we should exclude?
> > > >
> > >
> > > I bet if we forbid placing marks on /proc, some apps would break.
> >
> > Well, I didn't mean all marks, just the permission ones. I'm not sure there
> > are apps that place permission events on /proc...
> >
> 
> Maybe not intentionally.
> I once tested a few fanotify based AntiVirus solutions.
> In some of them, setting an "Exclude path" on some mount point
> would cause mark to not be set on that path, but for one in particular,
> the mark was still being set on the mount so path pattern filtering was
> done after receiving the events.
> I did not check whether /proc was blacklisted out of the box or if it
> could be marked/excluded from scan.
> IMO, assuming that all AntiVirus vendors blacklist all virtual filesystems
> is an assumption that we need to validate.
> [CC Marko from F-Secure for commenting on the above.]
> 
> > > I always though that allowing O_PATH in event_f_flags can make
> > > sense for some apps.
> > >
> > > What if instead of forbiding marks of /proc, we would force those
> > > marks to use O_PATH for fd creation. Some of the functionality
> > > will remain. Apps are less likely to break. Deadlocks will be less
> > > likely, although maybe still possible.
> >
> > Yes, that's another option. But if this is automatic, it is going to be
> > confusing to potential users - report O_PATH fd if getting normal fd is
> > dangerous isn't great. And also the deadlocks are there only for permission
> > events so there's no strong reason to restrict normal ones.
> >
> 
> Of course, we can do this hack only for reading permission events
> for virtual filesystems. The question is what would be more surprising
> to existing apps:
> - marks for permission events no longer working on /proc
> - fd from permission events no longer readable from /proc
> - something completely different

I think that option 1) is at least very explicit failure and given what
Marko wrote I'm willing to give that a try. If people start complaining,
we'll probably have to revert and try either reporting O_PATH fd or even
FAN_NOFD. It's a bit of a gamble I agree but current status quo isn't good
either.

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR