Re: [PATCH 2/2] fanotify: Add pidfd support to the fanotify API

[Matthew Bobrowski <repnop@google.com>]

On Mon, Apr 19, 2021 at 05:02:33PM +0200, Christian Brauner wrote:
> On Mon, Apr 19, 2021 at 03:55:50PM +0200, Jan Kara wrote:
> > On Mon 19-04-21 15:20:20, Christian Brauner wrote:
> > > On Fri, Apr 16, 2021 at 09:22:25AM +1000, Matthew Bobrowski wrote:
> > > > Introduce a new flag FAN_REPORT_PIDFD for fanotify_init(2) which
> > > > allows userspace applications to control whether a pidfd is to be
> > > > returned instead of a pid for `struct fanotify_event_metadata.pid`.
> > > > 
> > > > FAN_REPORT_PIDFD is mutually exclusive with FAN_REPORT_TID as the
> > > > pidfd API is currently restricted to only support pidfd generation for
> > > > thread-group leaders. Attempting to set them both when calling
> > > > fanotify_init(2) will result in -EINVAL being returned to the
> > > > caller. As the pidfd API evolves and support is added for tids, this
> > > > is something that could be relaxed in the future.
> > > > 
> > > > If pidfd creation fails, the pid in struct fanotify_event_metadata is
> > > > set to FAN_NOPIDFD(-1). Falling back and providing a pid instead of a
> > > > pidfd on pidfd creation failures was considered, although this could
> > > > possibly lead to confusion and unpredictability within userspace
> > > > applications as distinguishing between whether an actual pidfd or pid
> > > > was returned could be difficult, so it's best to be explicit.
> > > > 
> > > > Signed-off-by: Matthew Bobrowski <repnop@google.com>
> > > > ---
> > > >  fs/notify/fanotify/fanotify_user.c | 33 +++++++++++++++++++++++++++---
> > > >  include/linux/fanotify.h           |  2 +-
> > > >  include/uapi/linux/fanotify.h      |  2 ++
> > > >  3 files changed, 33 insertions(+), 4 deletions(-)
> > > > 
> > > > diff --git a/fs/notify/fanotify/fanotify_user.c b/fs/notify/fanotify/fanotify_user.c
> > > > index 9e0c1afac8bd..fd8ae88796a8 100644
> > > > --- a/fs/notify/fanotify/fanotify_user.c
> > > > +++ b/fs/notify/fanotify/fanotify_user.c
> > > > @@ -329,7 +329,7 @@ static ssize_t copy_event_to_user(struct fsnotify_group *group,
> > > >  	struct fanotify_info *info = fanotify_event_info(event);
> > > >  	unsigned int fid_mode = FAN_GROUP_FLAG(group, FANOTIFY_FID_BITS);
> > > >  	struct file *f = NULL;
> > > > -	int ret, fd = FAN_NOFD;
> > > > +	int ret, pidfd, fd = FAN_NOFD;
> > > >  	int info_type = 0;
> > > >  
> > > >  	pr_debug("%s: group=%p event=%p\n", __func__, group, event);
> > > > @@ -340,7 +340,25 @@ static ssize_t copy_event_to_user(struct fsnotify_group *group,
> > > >  	metadata.vers = FANOTIFY_METADATA_VERSION;
> > > >  	metadata.reserved = 0;
> > > >  	metadata.mask = event->mask & FANOTIFY_OUTGOING_EVENTS;
> > > > -	metadata.pid = pid_vnr(event->pid);
> > > > +
> > > > +	if (FAN_GROUP_FLAG(group, FAN_REPORT_PIDFD) &&
> > > > +		pid_has_task(event->pid, PIDTYPE_TGID)) {
> > > > +		/*
> > > > +		 * Given FAN_REPORT_PIDFD is to be mutually exclusive with
> > > > +		 * FAN_REPORT_TID, panic here if the mutual exclusion is ever
> > > > +		 * blindly lifted without pidfds for threads actually being
> > > > +		 * supported.
> > > > +		 */
> > > > +		WARN_ON(FAN_GROUP_FLAG(group, FAN_REPORT_TID));
> > > > +
> > > > +		pidfd = pidfd_create(event->pid, 0);
> > > > +		if (unlikely(pidfd < 0))
> > > > +			metadata.pid = FAN_NOPIDFD;
> > > > +		else
> > > > +			metadata.pid = pidfd;
> > > 
> > > I'm not a fan of overloading fields (Yes, we did this for the _legacy_
> > > clone() syscall for CLONE_PIDFD/CLONE_PARENT_SETTID but in general it's
> > > never a good idea if there are other options, imho.).
> > > Could/should we consider the possibility of adding a new pidfd field to
> > > struct fanotify_event_metadata?
> > 
> > I'm not a huge fan of overloading fields either but in this particular case
> > I'm fine with that because:
> > 
> > a) storage size & type matches
> > b) it describes exactly the same information, just in a different way
> > 
> > It is not possible to store the pidfd elsewhere in fanotify_event_metadata.
> > But it is certainly possible to use extended event info to return pidfd
> > instead - similarly to how we return e.g. handle + fsid for some
> > notification groups. It just means somewhat longer events and more
> > complicated parsing of structured events in userspace. But as I write
> > above, in this case I don't think it is worth it - only if we think that
> > returning both pid and pidfd could ever be useful.
> 
> Yeah, I don't hink users need to do that. After all they can parse the
> PID out of the /proc/self/fdinfo/<pidfd> file.

I agree. We can do this little PID extraction dance in userspace, it's
not a big deal at this point IMO.

> A general question about struct fanotify_event_metadata and its
> extensibility model:
> looking through the code it seems that this struct is read via
> fanotify_rad(). So the user is expected to supply a buffer with at least
> 
> #define FAN_EVENT_METADATA_LEN (sizeof(struct fanotify_event_metadata))
> 
> bytes. In addition you can return the info to the user about how many
> bytes the kernel has written from fanotify_read().
> 
> So afaict extending fanotify_event_metadata should be _fairly_
> straightforward, right? It would essentially the complement to
> copy_struct_from_user() which Aleksa and I added (1 or 2 years ago)
> which deals with user->kernel and you're dealing with kernel->user:
> - If the user supplied a buffer smaller than the minimum known struct
>   size -> reject.
> - If the user supplied a buffer < smaller than what the current kernel
>   supports -> copy only what userspace knows about, and return the size
>   userspace knows about.
> - If the user supplied a buffer that is larger than what the current
>   kernel knows about -> copy only what the kernel knows about, zero the
>   rest, and return the kernel size.
> 
> Extension should then be fairly straightforward (64bit aligned
> increments)?

You'd think that it's fairly straightforward, but I have a feeling
that the whole fanotify_event_metadata extensibility discussion and
the current limitation to do so revolves around whether it can be
achieved in a way which can guarantee that no userspace applications
would break. I think the answer to this is that there's no guarantee
because of <<reasons>>, so the decision to extend fanotify's feature
set was done via other means i.e. introduction of additional
structures.

/M