Re: [PATCH v3 5/5] fanotify: add pidfd support to the fanotify API

[Matthew Bobrowski <repnop@google.com>]

On Wed, Aug 04, 2021 at 02:39:40PM +0200, Jan Kara wrote:
> On Wed 04-08-21 13:46:05, Matthew Bobrowski wrote:
> > On Tue, Aug 03, 2021 at 04:04:21PM +0200, Jan Kara wrote:
> > > On Tue 03-08-21 13:07:57, Amir Goldstein wrote:
> > > > On Tue, Aug 3, 2021 at 12:37 PM Christian Brauner
> > > > <christian.brauner@ubuntu.com> wrote:
> > > > >
> > > > > On Mon, Aug 02, 2021 at 05:38:20PM +0300, Amir Goldstein wrote:
> > > > > > On Mon, Aug 2, 2021 at 3:34 PM Jan Kara <jack@suse.cz> wrote:
> > > > > > >
> > > > > > > On Fri 30-07-21 08:03:01, Amir Goldstein wrote:
> > > > > > > > On Thu, Jul 29, 2021 at 6:13 PM Amir Goldstein <amir73il@gmail.com> wrote:
> > > > > > > > > On Thu, Jul 29, 2021 at 4:39 PM Jan Kara <jack@suse.cz> wrote:
> > > > > > > > > > Well, but pidfd also makes sure that /proc/<pid>/ keeps belonging to the
> > > > > > > > > > same process while you read various data from it. And you cannot achieve
> > > > > > > > > > that with pid+generation thing you've suggested. Plus the additional
> > > > > > > > > > concept and its complexity is non-trivial So I tend to agree with
> > > > > > > > > > Christian that we really want to return pidfd.
> > > > > > > > > >
> > > > > > > > > > Given returning pidfd is CAP_SYS_ADMIN priviledged operation I'm undecided
> > > > > > > > > > whether it is worth the trouble to come up with some other mechanism how to
> > > > > > > > > > return pidfd with the event. We could return some cookie which could be
> > > > > > > > > > then (by some ioctl or so) either transformed into real pidfd or released
> > > > > > > > > > (so that we can release pid handle in the kernel) but it looks ugly and
> > > > > > > > > > complicates things for everybody without bringing significant security
> > > > > > > > > > improvement (we already can pass fd with the event). So I'm pondering
> > > > > > > > > > whether there's some other way how we could make the interface safer - e.g.
> > > > > > > > > > so that the process receiving the event (not the one creating the group)
> > > > > > > > > > would also need to opt in for getting fds created in its file table.
> > > > > > > > > >
> > > > > > > > > > But so far nothing bright has come to my mind. :-|
> > > > > > > > > >
> > > > > > > > >
> > > > > > > > > There is a way, it is not bright, but it is pretty simple -
> > > > > > > > > store an optional pid in group->fanotify_data.fd_reader.
> > > > > > > > >
> > > > > > > > > With flag FAN_REPORT_PIDFD, both pidfd and event->fd reporting
> > > > > > > > > will be disabled to any process other than fd_reader.
> > > > > > > > > Without FAN_REPORT_PIDFD, event->fd reporting will be disabled
> > > > > > > > > if fd_reaader is set to a process other than the reader.
> > > > > > > > >
> > > > > > > > > A process can call ioctl START_FD_READER to set fd_reader to itself.
> > > > > > > > > With FAN_REPORT_PIDFD, if reaader_fd is NULL and the reader
> > > > > > > > > process has CAP_SYS_ADMIN, read() sets fd_reader to itself.
> > > > > > > > >
> > > > > > > > > Permission wise, START_FD_READER is allowed with
> > > > > > > > > CAP_SYS_ADMIN or if fd_reader is not owned by another process.
> > > > > > > > > We may consider YIELD_FD_READER ioctl if needed.
> > > > > > > > >
> > > > > > > > > I think that this is a pretty cheap price for implementation
> > > > > > > > > and maybe acceptable overhead for complicating the API?
> > > > > > > > > Note that without passing fd, there is no need for any ioctl.
> > > > > > > > >
> > > > > > > > > An added security benefit is that the ioctl adds is a way for the
> > > > > > > > > caller of fanotify_init() to make sure that even if the fanotify_fd is
> > > > > > > > > leaked, that event->fd will not be leaked, regardless of flag
> > > > > > > > > FAN_REPORT_PIDFD.
> > > > > > > > >
> > > > > > > > > So the START_FD_READER ioctl feature could be implemented
> > > > > > > > > and documented first.
> > > > > > > > > And then FAN_REPORT_PIDFD could use the feature with a
> > > > > > > > > very minor API difference:
> > > > > > > > > - Without the flag, other processes can read fds by default and
> > > > > > > > >   group initiator can opt-out
> > > > > > > > > - With the flag, other processes cannot read fds by default and
> > > > > > > > >   need to opt-in
> > > > > > > >
> > > > > > > > Or maybe something even simpler... fanotify_init() flag
> > > > > > > > FAN_PRIVATE (or FAN_PROTECTED) that limits event reading
> > > > > > > > to the initiator process (not only fd reading).
> > > > > > > >
> > > > > > > > FAN_REPORT_PIDFD requires FAN_PRIVATE.
> > > > > > > > If we do not know there is a use case for passing fanotify_fd
> > > > > > > > that reports pidfds to another process why implement the ioctl.
> > > > > > > > We can always implement it later if the need arises.
> > > > > > > > If we contemplate this future change, though, maybe the name
> > > > > > > > FAN_PROTECTED is better to start with.
> > > > > > >
> > > > > > > Good ideas. I think we are fine with returning pidfd only to the process
> > > > > > > creating the fanotify group. Later we can add an ioctl which would indicate
> > > > > > > that the process is also prepared to have fds created in its file table.
> > > > > > > But I have still some open questions:
> > > > > > > Do we want threads of the same process to still be able to receive fds?
> > > > > >
> > > > > > I don't see why not.
> > > > > > They will be bloating the same fd table as the thread that called
> > > > > > fanotify_init().
> > > > > >
> > > > > > > Also pids can be recycled so they are probably not completely reliable
> > > > > > > identifiers?
> > > > > >
> > > > > > Not sure I follow. The group hold a refcount on struct pid of the process that
> > > > > > called fanotify_init() - I think that can used to check if reader process is
> > > > > > the same process, but not sure. Maybe there is another way (Christian?).
> > > > >
> > > > > If the fanotify group hold's a reference to struct pid it won't get
> > > > > recycled. And it can be used to check if the reader thread is the same
> > > > > thread with some care. You also have to be specific what exactly you
> > > > > want to know.  If you're asking if the reading process is the same as
> > > > > the fanotify_init() process you can be asking one of two things.
> > > > >
> > > > > You can be asking if the reader is a thread in the same thread-group as
> > > > > the thread that called fanotify_init(). In that case you might need to
> > > > > do something like
> > > > >
> > > > > rcu_read_lock();
> > > > > struct task_struct *fanotify_init_task_struct = pid_task(stashed_struct_pid, PIDTYPE_PID);
> > > > > if (!fanotify_init_task_struct) {
> > > > >         /* The thread which called fanotify_init() has died already. */
> > > > >         return -ESRCH;
> > > > > }
> > > > > if (same_thread_group(fanotify_init_task_struct, current))
> > > > > rcu_read_unlock();
> > > > >
> > > > > though thinking about it makes me realise that there's a corner case. If
> > > > > the thread that called fanotify_init() is a thread in a non-empty
> > > > > thread-group it can already have died and been reaped. This would mean,
> > > > > pid_task(..., PIDTYPE_PID) will return NULL but there are still other
> > > > > threads alive in the thread-group. Handling that case might be a bit
> > > > > complicated.
> > > > >
> > > > > If you're asking whether the reading thread is really the same as the
> > > > > thread that created the fanotify instance then you might need to do sm
> > > > > like
> > > > >
> > > > > rcu_read_lock();
> > > > > if (pid_task(stashed_struct_pid, PIDTYPE_PID) == current)
> > > > > rcu_read_unlock();
> > > > >
> > > > > Just for completeness if I remember all of this right: there's a corner
> > > > > case because of how de_thread() works.
> > > > > During exec the thread that is execing will assume the struct pid of the
> > > > > old thread-group leader. (All other threads in the same thread-group
> > > > > will get killed.)
> > > > > Assume the thread that created the fanotify instance is not the
> > > > > thread-group leader in its non-empty thread-group. And further assume it
> > > > > exec's. Then it will assume the struct pid of the old thread-group
> > > > > leader during de_thread().
> > > > > Assume the thread inherits the fanotify fd across the exec. Now, when it
> > > > > tries to read a new event after the exec then pid_task() will return
> > > > > NULL.
> > > > > However, if the thread was already the thread-group leader before the
> > > > > exec then pid_task() will return the same task struct as before after
> > > > > the exec (because no struct pid swapping needed to take place).
> > > > >
> > > > > I hope this causes more clarity ?then confusion. :)
> > > > 
> > > > I'm afraid it's the latter :D
> > > > 
> > > > Sigh! We must simplify.
> > > > 
> > > > Thinking out loud, instead of sealing the possibility of another
> > > > process reading pidfd, maybe just avoid the most obvious unintentional
> > > > leak of fanotify_fd to another process by mandating  FAN_CLOEXEC?
> > >
> > > Well, I don't think we need any protection from leaking fanotify_fd. It is
> > > special fd with special priviledges as any other. If you leak it, well, bad
> > > luck but that's how Unix priviledge model works.
> > > 
> > > The threat IMO is that you have a process X, that process expects to
> > > receive fd to work with from process Y. Now process Y is malicious (or
> > > taken over by an attacker) and passes to X fanotify_fd. X reads from
> > > fanotify_fd to get data to process, it performs all kinds of validity
> > > checks on untrusted input but it does not expect that the read has side
> > > effects on X's file_table and in the worst case can lead to some compromise
> > > of X or easily to DoS on X by exhausting its file_table space.
> > >
> > > Currently this attack vector is moot because you have to have CAP_SYS_ADMIN
> > > to get to fanotify_fd and then you can certainly do worse things. But OTOH
> > > I can see why Jann was uneasy about this.
> > 
> > As I have breifly expressed in my previous emails, the cause for concern
> > here is flakey IMO. If there's sensible something that I'm clearly missing,
> > then please explain.
> 
> No, I think your understanding is correct.
> 
> > From my perspective, the only sensible attack vector that's maybe worth
> > worrying about here is the possibility of exhausting the fdtable of a given
> > process, which yes, can be considered as a form of DoS. However, in any
> > case, there are other defensive protections/measures that a programmer
> > could employ in their application code which could prevent such from ever
> > happening.
> > 
> > The whole passing of file descriptors between process Y and process X and
> > the leaking of a file descriptor thing simply goes back to what you've
> > mentioned above Jan. I consider it a very weak argument. When enabling
> > FAN_REPORT_PIDFD, the process requires CAP_SYS_ADMIN. If that process ever
> > has its execution flow hijacked by an attacker, then I'm sorry, I think
> > there's other larger causes for concern at that point rather then worrying
> > about the state of some other child processes fdtable.
> > 
> > In general cases, I get that passing a file descriptor between process Y
> > and process X and then having process X's fdtable modified as result of
> > calling functions like read() is considered undesired. But, for
> > applications that makes use of fanotify is there ever a case where we pass
> > the fanotify file descriptor to a random/unexpected process and have it
> > process events? I don't think so. So, I suppose what I'm trying to say is
> > that, if an application chooses to opt-in and use a flag like
> > FAN_REPORT_PIDFD or any other future file descriptor generating variant,
> > the expectation is that which ever process is created and event processing
> > is passed to that process, then it should always expect to have its fdtable
> > modified when reading events.
> 
> Yes, I was thinking about this some more and at this point, given the lack
> of convenient options for the hardening, I think the best option is to keep
> the interface as originally planned. Because I'm afraid the hardening options
> we were able to come up with would only cause confusion (and from confusion
> bugs easily arise) for little security gain.

OK, in that case are you happy for me to post hopefully the last iteration
of this series with the minor nits addressed?

/M