[systemd-devel] [PATCH] sd-event: we do not support EPOLLONESHOT correctly

[Shawn Landden]

On Fri, Dec 13, 2013 at 8:04 PM, Shawn Landden <shawn at churchofgit.com> wrote:
> On Fri, Dec 13, 2013 at 7:17 PM, Lennart Poettering
> <lennart at poettering.net> wrote:
>> On Fri, 13.12.13 13:08, Shawn Landden (shawn at churchofgit.com) wrote:
>>
>>> If this event is not the highest priority, then it will not be
>>> dispatched when the epoll triggers, and since we will not get any more
>>> wakeups (due to the way EPOLLONESHOT works) will never be dispatched.
>>
>> Hmm, so the way this is supposed to work is that as soon as epoll told
>> us about some event once, we will set the event source to "pending", and
>> it stays "pending" until we have dispatched it, regardless if we ever
>> get the event from epoll again or not. However, we currently will
>> overwrite the "revents" field should we ever get it again with the new
>> one. It probably makes more sense to OR it in, so that we don't forget
>> readability if writability happens at a later point or vice versa, if
>> you follow what I mean. I'll prep a fix for that.
> good
>>
>> With that fixed is there something else to still fix?
> Oh I missed how the pending code worked. Because of that I think I can
> use EPOLLET
> with sd-event, and will send a Distribute= patch that uses it. (already have it,
> just wasn't sure if it was only working cause my test setup has near-0 load)
With either of these, can't we end up in a situation where we have
events pending
but are still blocking on epoll_wait()?

shouldn't we do something like

p = have_pending()

epoll_wait(foo, foo, foo, p ? timeout : 0);

do prioq stuff()
mark_pending()
dispatch_highest_prio()

to avoid this issue?
>>
>>> Since we only handle one event per epoll_wait() wakeup,
>>> and we dequeue a ONE_SHOT event when we dispatch it, we
>>> do not have the problem described in epoll(7) of
>>> "multiple events can be generated upon receipt of multiple
>>> chunks of data".
>>
>> Not sure I follow here? That part appears to be about EPOLLET?
> That part explains why EPOLLET cannot be used to implement EPOLLONESHOT.
> The reason is that even with EPOLLET, epoll_wait can generate multiple
> events for the
> same watch, if more than one connection came in.
>>
>> Lennart
>>
>> --
>> Lennart Poettering, Red Hat
>> _______________________________________________
>> systemd-devel mailing list
>> systemd-devel at lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/systemd-devel