Epoll is doing multiple passes over the ready set at the moment, because of
the constraints over the f_op->poll() call.  Looking at the code again, I
noticed that we already hold the epoll semaphore in read, and this
(together with other locking conditions that hold while doing an
epoll_wait()) can lead to a smarter way [1] to "ship" events to userspace
(in a single pass).

This is a stress application that can be used to test the new code.  It
spwans multiple thread and call epoll_wait() and epoll_ctl() from many
threads.  Stress tested on my dual Opteron 254 w/out any problems.

http://www.xmailserver.org/totalmess.c

This is not a benchmark, just something that tries to stress and exploit
possible problems with the new code.
Also, I made a stupid micro-benchmark:

http://www.xmailserver.org/epwbench.c

[1] Considering that epoll must be thread-safe, there are five ways we can
    be hit during an epoll_wait() transfer loop (ep_send_events()):

    1) The epoll fd going away and calling ep_free
       This just can't happen, since we did an fget() in sys_epoll_wait

    2) An epoll_ctl(EPOLL_CTL_DEL)
       This can't happen because epoll_ctl() gets ep->sem in write, and
       we're holding it in read during ep_send_events()

    3) An fd stored inside the epoll fd going away
       This can't happen because in eventpoll_release_file() we get
       ep->sem in write, and we're holding it in read during
       ep_send_events()

    4) Another epoll_wait() happening on another thread
       They both can be inside ep_send_events() at the same time, we get
       (splice) the ready-list under the spinlock, so each one will get
       its own ready list. Note that an fd cannot be at the same time
       inside more than one ready list, because ep_poll_callback() will
       not re-queue it if it sees it already linked:

       if (ep_is_linked(&epi->rdllink))
                goto is_linked;

       Another case that can happen, is two concurrent epoll_wait(),
       coming in with a userspace event buffer of size, say, ten.
       Suppose there are 50 event ready in the list. The first
       epoll_wait() will "steal" the whole list, while the second, seeing
       no events, will go to sleep. But at the end of ep_send_events() in
       the first epoll_wait(), we will re-inject surplus ready fds, and we
       will trigger the proper wake_up to the second epoll_wait().

    5) ep_poll_callback() hitting us asyncronously
       This is the tricky part. As I said above, the ep_is_linked() test
       done inside ep_poll_callback(), will guarantee us that until the
       item will result linked to a list, ep_poll_callback() will not try
       to re-queue it again (read, write data on any of its members). When
       we do a list_del() in ep_send_events(), the item will still satisfy
       the ep_is_linked() test (whatever data is written in prev/next,
       it'll never be its own pointer), so ep_poll_callback() will still
       leave us alone. It's only after the eventual smp_mb()+INIT_LIST_HEAD(&epi->rdllink)
       that it'll become visible to ep_poll_callback(), but at the point
       we're already past it.

[akpm@osdl.org: 80 cols]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>