Perfect -- thank you! As it turns out, this was -- as you surmised -- a bug in our epoll implementation. I have pushed the fix:
https://github.com/joyent/illumos-joyent/commit/471f1f3ef02e25a84d7c36c857ed260ae2b7ac99 Thanks for taking the time for the clean test case -- and apologies for the bug! - Bryan On Thu, Jul 31, 2014 at 7:28 AM, Youzhong Yang <[email protected]> wrote: > Thanks Bryan. > > I take back what I said. epoll_* does not ignore epoll_event.data passed > in. > > I managed to reproduce the issue with a program rewritten from yours: > > # gcc myepoll.c > # ./a.out > EPOLL_CTL_ADD: fd = 4, ptr = fffffd7fffdffba0 > writing to pipe fd 5 > written 6 bytes to fd 5 > Got an event on fd 4, events = 1, ptr = fffffd7fffdffba0 > got 6 bytes from fd 5 > EPOLL_CTL_DEL: fd = 4, ptr = fffffd7fffdffba0 > EPOLL_CTL_ADD: fd = 4, ptr = fffffd7fffdffb80 > writing to pipe fd 5 > written 6 bytes to fd 5 > Got an event on fd 4, events = 1, ptr = fffffd7fffdffba0 > got 6 bytes from fd 5 > > As it indicates, the issue here is that after EPOLL_CTL_DEL, and then > EPOLL_CTL_ADD of the same file descriptor with a new data.ptr, epoll_wait > still returns an old pointer. > > C code is attached for your reference. > > Thanks very much, > > -Youzhong > > > On Wed, Jul 30, 2014 at 5:38 PM, Bryan Cantrill <[email protected]> > wrote: > >> >> Seems like a bug, but not sure I entirely understand: there may be some >> edge conditions where data specified in the "data" field of the epoll_event >> structure via epoll_ctl() is not properly returned, but it's not true that >> SmartOS simply ignores it. Take, for example, the following program: >> >> https://us-east.manta.joyent.com/bcantrill/public/epoll_add.c >> >> (It's essentially the most basic possible use of epoll(5).) Compiling >> that program and running it yields (for me, anyway) the same result on >> SmartOS as it does on Linux: >> >> $ ./epoll_add >> writing to pipe >> adding event... >> parent: got an event on fd 4 >> >> I certainly don't doubt that there's a bug here -- I just don't think >> it's quite as broad as you're implying; could you point me to the core file >> for Samba and/or instructions as to how to reproduce the issue? >> >> - Bryan >> >> >> >> On Wed, Jul 30, 2014 at 2:08 PM, Youzhong Yang via smartos-discuss < >> [email protected]> wrote: >> >>> Hi All, >>> >>> The man page of epoll_ctl says: >>> >>> The data field specifies the datum to be associated with the >>> event and will be returned via epoll_wait(3C). The events field >>> denotes both the desired events (when specified via epoll_ctl()) >>> and the events that have occurred (when returned via >>> epoll_wait(3C)). >>> >>> This does not seem to be the case and is not compatible with Linux's >>> implementation. Recently I tried to compile Samba 4.1.10 on smartos and >>> 'net ads join' just crashes. Samba uses epoll_event.data.ptr to keep track >>> of its tevent_fd data but epoll_ctl() on smartos simply ignores it and does >>> not keep a copy of the 'data' structure and returns it when epoll_wait is >>> called. >>> >>> Can I assume this is an implementation specific bug? >>> >>> Thanks, >>> >>> -Youzhong >>> *smartos-discuss* | Archives >>> <https://www.listbox.com/member/archive/184463/=now> >>> <https://www.listbox.com/member/archive/rss/184463/21493597-819305a5> | >>> Modify >>> <https://www.listbox.com/member/?&;> >>> Your Subscription <http://www.listbox.com> >>> >> >> > ------------------------------------------- smartos-discuss Archives: https://www.listbox.com/member/archive/184463/=now RSS Feed: https://www.listbox.com/member/archive/rss/184463/25769125-55cfbc00 Modify Your Subscription: https://www.listbox.com/member/?member_id=25769125&id_secret=25769125-7688e9fb Powered by Listbox: http://www.listbox.com
