On 1/27/21 8:04 PM, Jason Wang wrote: > > On 2021/1/27 下午5:11, Yongji Xie wrote: >> On Wed, Jan 27, 2021 at 11:38 AM Jason Wang <jasowang@xxxxxxxxxx> wrote: >>> >>> On 2021/1/20 下午2:52, Yongji Xie wrote: >>>> On Wed, Jan 20, 2021 at 12:24 PM Jason Wang <jasowang@xxxxxxxxxx> wrote: >>>>> On 2021/1/19 下午12:59, Xie Yongji wrote: >>>>>> Now we have a global percpu counter to limit the recursion depth >>>>>> of eventfd_signal(). This can avoid deadlock or stack overflow. >>>>>> But in stack overflow case, it should be OK to increase the >>>>>> recursion depth if needed. So we add a percpu counter in eventfd_ctx >>>>>> to limit the recursion depth for deadlock case. Then it could be >>>>>> fine to increase the global percpu counter later. >>>>> I wonder whether or not it's worth to introduce percpu for each eventfd. >>>>> >>>>> How about simply check if eventfd_signal_count() is greater than 2? >>>>> >>>> It can't avoid deadlock in this way. >>> >>> I may miss something but the count is to avoid recursive eventfd call. >>> So for VDUSE what we suffers is e.g the interrupt injection path: >>> >>> userspace write IRQFD -> vq->cb() -> another IRQFD. >>> >>> It looks like increasing EVENTFD_WAKEUP_DEPTH should be sufficient? >>> >> Actually I mean the deadlock described in commit f0b493e ("io_uring: >> prevent potential eventfd recursion on poll"). It can break this bug >> fix if we just increase EVENTFD_WAKEUP_DEPTH. > > > Ok, so can wait do something similar in that commit? (using async stuffs > like wq). io_uring should be fine in current kernels, but aio would still be affected by this. But just in terms of recursion, bumping it one more should probably still be fine. -- Jens Axboe