On 4/1/24 8:36 AM, Jens Axboe wrote: > timerfd is utterly buggy wrt nonblocking reads - regardless of whether > or not there's data available, it returns -EAGAIN. This is incompatible > with how nonblocking reads should work. If there's data available, it > should be returned. > > Convert it to use fops->read_iter() so it can handle both nonblocking > fds and IOCB_NOWAIT, mark it as FMODE_NOWAIT to signify that it's > compatible with nonblocking reads, and finally have timerfd_read_iter() > properly check for data availability in nonblocking mode. > > Cc: stable@xxxxxxxxxxxxxxx > Signed-off-by: Jens Axboe <axboe@xxxxxxxxx> > > --- > > Can't believe it's this broken... Patch has been tested with a test case > that was reported via io_uring, and I also ran the ltp timerfd test > cases and it passes all of those too. OK pre-coffee email, it's not actually that broken. Which makes me feel better, because I'm not used to finding stuff like this in core bits, mostly just the odd drivers. I'll send a revised version of this with a fixed up changelog, as supporting IOCB_NOWAIT is still very useful. Won't change the actual patch, as that's really all it does. -- Jens Axboe
