Add EV_EXCLUSIVE to kevent.

[dchagin]

second iteration, it's works, but I'm not sure.

[pauamma_gundo.com]

Manual page change is probably small enough to be fixed while committing if nothing else needs revision. (As always, I can't attest to source change or consistency between it and manual page.)

lib/libc/sys/kqueue.2

289–290

[dchagin]

In D35155#799022, @pauamma_gundo.com wrote:

Manual page change is probably small enough to be fixed while committing if nothing else needs revision. (As always, I can't attest to source change or consistency between it and manual page.)

Thank you! Since I don't really understand the role of influx in kevent() I won't commit until Kostya criticizes kqueue_sccan changes

[editor_callfortesting.org]

Is this Review still valid, valuable, and ready to land?

[dchagin]

In D35155#944705, @editor_callfortesting.org wrote:

Is this Review still valid, valuable, and ready to land?

Please, see my previous comment, also it would be nice to test in the real environment

[markj]

Reading the description of EPOLLEXCLUSIVE from the epoll_ctl manual page, this looks like it provides the wrong semantics. The use-case for EPOLLEXCLUSIVE seems to be that you have multiple epoll descriptors being used to monitor events on a single object; when an event occurs, only one of the epoll descriptors will receive the event. In the Linuxulator there is a 1-1 correspondence between kqueues and epoll descriptors. You are translating EPOLLEXCLUSIVE to a per-knote flag, and if N kqueues contain an event for a given resource, there will be N independent knotes. The tests work because they use a single kqueue, but I believe that epoll semantics require EV_EXCLUSIVE to work across multiple kqueues. I believe it is unusual to have multiple threads sharing a kqueue anyway; see https://www.usenix.org/system/files/atc21-zhao.pdf

So, I would expect an implementation of EPOLLEXCLUSIVE to, at least, modify knote() such that we avoid activating more than one knote with EV_EXCLUSIVE set.

lib/libc/sys/kqueue.2

282

The flag applies to the knote, not the entire queue.

sys/kern/kern_event.c

2074

There is no need to add a separate case, you can change line 2061 to if ((kn->kn_flags & EV_CLEAR | EV_DISPATCH | EV_EXCLUSIVE)) != 0) { and the behaviour will be unchanged.

[kevans]

Someone was asking about something like this on Discord; thinking just in terms of providing a flag that matches EPOLLEXCLUSIVE semantics and ignoring the Linuxolator aspect for a minute: I think the bulk of the implementation would actually be in knote() instead, effectively only KNOTE_ACTIVATE'ing some subset of what we touched if some of them are marked EXCLUSIVE. I don't know that it's quite that straightforward, though.