On Fri, May 21, 2021 at 12:40:56PM +0200, Jan Kara wrote:

On Fri 21-05-21 20:15:35, Matthew Bobrowski wrote:

quoted

On Thu, May 20, 2021 at 03:55:27PM +0200, Jan Kara wrote:
There's one thing that I'd like to mention, and it's something in
regards to the overall approach we've taken that I'm not particularly
happy about and I'd like to hear all your thoughts. Basically, with
this approach the pidfd creation is done only once an event has been
queued and the notification worker wakes up and picks up the event
from the queue processes it. There's a subtle latency introduced when
taking such an approach which at times leads to pidfd creation
failures. As in, by the time pidfd_create() is called the struct pid
has already been reaped, which then results in FAN_NOPIDFD being
returned in the pidfd info record.

Having said that, I'm wondering what the thoughts are on doing pidfd
creation earlier on i.e. in the event allocation stages? This way, the
struct pid is pinned earlier on and rather than FAN_NOPIDFD being
returned in the pidfd info record because the struct pid has been
already reaped, userspace application will atleast receive a valid
pidfd which can be used to check whether the process still exists or
not. I think it'll just set the expectation better from an API
perspective.

Yes, there's this race. OTOH if FAN_NOPIDFD is returned, the listener can
be sure the original process doesn't exist anymore. So is it useful to
still receive pidfd of the dead process?

Well, you're absolutely right. However, FWIW I was approaching this
from two different angles:

1) I wanted to keep the pattern in which the listener checks for the
   existence/recycling of the process consistent. As in, the listener
   would receive the pidfd, then send the pidfd a signal via
   pidfd_send_signal() and check for -ESRCH which clearly indicates
   that the target process has terminated.

2) I didn't want to mask failed pidfd creation because of early
   process termination and other possible failures behind a single
   FAN_NOPIDFD. IOW, if we take the -ESRCH approach above, the
   listener can take clear corrective branches as what's to be done
   next if a race is to have been detected, whereas simply returning
   FAN_NOPIDFD at this stage can mean multiple things.

Now that I've written the above and keeping in mind that we'd like to
refrain from doing anything in the event allocation stages, perhaps we
could introduce a different error code for detecting early process
termination while attempting to construct the info record. WDYT?

Also opening pidfd in the context of event generation is problematic
for two reasons:

1) Technically, the context under which events are generated can be rather
constrained (various locks held etc.). Adding relatively complex operations
such as pidfd creation is going to introduce strange lock dependencies,
possibly deadlocks.

2) Doing pidfd generation in the context of the process generating event is
problematic - you don't know in which fd_table the fd will live. Also that
process is unfairly penalized (performance wise) because someone else is
listening. We try to keep overhead of event generation as low as possible
for this reason.

Fair points, thanks for sharing your view. :)

/M