On Sat 22-05-21 09:32:36, Matthew Bobrowski wrote: > On Fri, May 21, 2021 at 12:40:56PM +0200, Jan Kara wrote: > > On Fri 21-05-21 20:15:35, Matthew Bobrowski wrote: > > > 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? Sure, I wouldn't like to overengineer it but having one special fd value for "process doesn't exist anymore" and another for general "creating pidfd failed" looks OK to me. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR