On 4/15/22 23:53, Jens Axboe wrote:
On 4/15/22 4:41 PM, Pavel Begunkov wrote:
On 4/15/22 23:03, Jens Axboe wrote:
On 4/15/22 3:05 PM, Pavel Begunkov wrote:
On 4/12/22 17:46, Jens Axboe wrote:
On 4/12/22 10:41 AM, Jens Axboe wrote:
On 4/12/22 10:24 AM, Pavel Begunkov wrote:
If all completed requests in io_do_iopoll() were marked with
REQ_F_CQE_SKIP, we'll not only skip CQE posting but also
io_free_batch_list() leaking memory and resources.
Move @nr_events increment before REQ_F_CQE_SKIP check. We'll potentially
return the value greater than the real one, but iopolling will deal with
it and the userspace will re-iopoll if needed. In anyway, I don't think
there are many use cases for REQ_F_CQE_SKIP + IOPOLL.
Ah good catch - yes probably not much practical concern, as the lack of
ordering for file IO means that CQE_SKIP isn't really useful for that
scenario.
One potential snag is with the change we're now doing
io_cqring_ev_posted_iopoll() even if didn't post an event. Again
probably not a practical concern, but it is theoretically a violation
if an eventfd is used.
Looks this didn't get applied. Are you concerned about eventfd?
Yep, was hoping to get a reply back, so just deferred it for now.
Is there any good reason why the userspace can't tolerate spurious
eventfd events? Because I don't think we should care this case
I always forget the details on that, but we've had cases like this in
the past where some applications assume that if they got N eventfd
events, then are are also N events in the ring. Which granted is a bit
odd, but it does also make some sense. Why would you have more eventfd
events posted than events?
For the same reason why it can get less eventfd events than there are
CQEs, as for me it's only a communication channel but not a
replacement for completion events.
That part is inherently racy in that we might get some CQEs while we
respond to the initial eventfd notifications. But I'm totally agreeing
with you, and it doesn't seem like a big deal to me.
Ok, we don't want to break old applications, but it's a new most
probably not widely used feature, and we can say that the userspace
has to handle spurious eventfd.
If I were to guess, I'd say it's probably epoll + eventfd conversions.
But it should just be made explicit. Since events reaped and checked
Didn't get it, what should be made explicit? Do you mean documenting
that there might be spurious eventfd events or something else?
happen differently anyway, it seems like a bad assumption to make that
eventfd notifications == events available.
--
Pavel Begunkov
