On 2/4/20 2:38 AM, Daurnimator wrote: > On Fri, 31 Jan 2020 at 16:25, Jens Axboe <axboe@xxxxxxxxx> wrote: >> >> If we register an eventfd with io_uring for completion notification, >> and then subsequently does a poll for that very descriptor, then we >> can trigger a deadlock scenario. Once a request completes and signals >> the eventfd context, that will in turn trigger the poll command to >> complete. When that poll request completes, it'll try trigger another >> event on the eventfd, but this time off the path led us to complete >> the poll in the first place. The result is a deadlock in eventfd, >> as it tries to ctx->wqh.lock in a nested fashion. >> >> Check if the file in question for the poll request is our eventfd >> context, and if it is, don't trigger a nested event for the poll >> completion. > > Could this deadlock/loop also happen via an epoll fd? No, only through io_uring and aio with the way they handle the notification through a (potentially nested) waitqueue wakeup handler. -- Jens Axboe
