Re: [PATCH] epoll: add exclusive wakeups flag
From: Michael Kerrisk (man-pages)
Date: Fri Jan 29 2016 - 03:14:48 EST
On 01/28/2016 06:57 PM, Jason Baron wrote:
> On 01/28/2016 02:16 AM, Michael Kerrisk (man-pages) wrote:
>> Hi Jason,
>> On 12/08/2015 04:23 AM, Jason Baron wrote:
>>> Re-post of an old series addressing thundering herd issues when sharing
>>> an event source fd amongst multiple epoll fds. Last posting was here
>>> for reference: https://lkml.org/lkml/2015/2/25/56
>>> The patch herein drops the core scheduler 'rotate' changes I had previously
>>> proposed as this patch seems performant without those.
>>> I was prompted to re-post this because Madars Vitolins reported some good
>>> speedups with this patch using Enduro/X application. His writeup is here:
>>> Sample epoll_clt text:
>> Thanks for the proposed text. I have some questions about points
>> that are not quite clear to me.
>>> Sets an exclusive wakeup mode for the epfd file descriptor that is
>>> being attached to the target file descriptor, fd. Thus, when an
>>> event occurs and multiple epfd file descriptors are attached to the
>>> same target file using EPOLLEXCLUSIVE, one or more epfds will receive
>>> an event with epoll_wait(2). The default in this scenario (when
>>> EPOLLEXCLUSIVE is not set) is for all epfds to receive an event.
>>> EPOLLEXLUSVIE may only be specified with the op EPOLL_CTL_ADD.
>> So, assuming an FD is present in the interest list of multiple (say 6)
>> epoll FDs, and some (say 3) of those attachments were done using
>> EPOLLEXCLUSVE. Which of the following statements are correct:
>> (a) It's guaranteed that *none* of the epoll FDs that did NOT specify
>> EPOLLEXCLUSIVE will receive an event.
>> (b) It's guaranteed that *all* of the epoll FDs that did NOT specify
>> EPOLLEXCLUSIVE will receive an event.
>> (c) From 1 to 3 of the epoll FDs that did specify EPOLLEXCLUSIVE
>> will receive an event.
>> (d) Exactly one epoll FD that did specify EPOLLEXCLUSIVE will get
>> an event, and it is indeterminate which one.
> So b and c. All the non-exclusive adds will get it and at least 1 of the
> exclusive adds will as well.
So is it fair to say that the expected use case is that all epoll sets
would use EPOLLEXCLUSIVE?
>> I suppose one point I'm trying to uncover in the above is: what is
>> the scope of EPOLLEXCLUSIVE? Is it just applicable for one process's
>> FD, or is it setting an attribute in the epoll "interest list" record
>> for that FD that affects notification behavior across all processes?
> Right - so 'EPOLLEXCLUSIVE' will affect other epoll sets that are also
> using 'EPOLLEXCLUSIVE' against the the same fd, but will have no affect
> on epoll sets connected to fd that do not specify it.
>> And then:
>> (1) What are the semantics of EPOLLEXCLUSIVE if the added FD becomes
>> disabled via EPOLLONESHOT (or explicitly via EPOLL_CTL_MOD with
>> the 'events' field set to 0)?
> In the case of EPOLLEXCLUSIVE and EPOLLONESHOT, one would have to re-arm
> at least 1 of threads that was woken up by doing EPOLL_CTL_MOD to
> guarantee further wakeups.
> And like-wise with an EPOLL_CTL_MOD with 'events' all set to 0, one
> would need to either re-arm the thread that set the 'events' field to 0
> (by setting back to non-zero), or re-arm in at least one other thread
> via EPOLL_CTL_MOD (or delete and add).
Okay -- so when an EPOLLEXCLUSIVE FD becomes disarmed it is possible
to re-enable rith EPOLL_CTL_MOD; one doesn't need to delete and re-add
>> (2) The source code contains a comment "we do not currently supported
>> nested exclusive wakeups". Could you elaborate on this point? It
>> sounds like something that should be documented.
> So I was just trying to say that we return -EINVAL if you try to do and
> EPOLL_CTL_ADD with EPOLLEXCLUSIVE and the 'fd' argument is a epoll fd
> returned via epoll_create().
Okay -- that definitely belongs in the man page.
I'll work up a text, but would like to get input about the "use case"
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/