[PATCH] 2.5.46: epoll ep_insert doesn't wake waiters if events exist

[PATCH] 2.5.46: epoll ep_insert doesn't wake waiters if events exist

[Jay Vosburgh]

[-- Attachment #1: Type: text/plain, Size: 1240 bytes --]


      The logic in fs/eventpoll.c:ep_insert()  checks to see if events are
already present when processing an EP_CTL_ADD operation, and, if any are,
adds them to the list.  It does not wake up any waiters, so, e.g., another
thread waiting in epoll_wait() will not be awakened for these events.  The
fix is to have ep_insert() do wakeups, just as ep_poll_callback() does when
it adds events.

      -J

diff -ur linux-2.5.46.orig/fs/eventpoll.c linux-2.5.46/fs/eventpoll.c
--- linux-2.5.46.orig/fs/eventpoll.c      Thu Nov  7 14:31:21 2002
+++ linux-2.5.46/fs/eventpoll.c     Thu Nov  7 14:32:16 2002
@@ -838,8 +838,13 @@
      list_add(&dpi->llink, ep_hash_entry(ep, ep_hash_index(ep, tfile)));

      /* If the file is already "ready" we drop it inside the ready list */
-     if ((revents & pfd->events) && !EP_IS_LINKED(&dpi->rdllink))
+     if ((revents & pfd->events) && !EP_IS_LINKED(&dpi->rdllink)) {
            list_add(&dpi->rdllink, &ep->rdllist);
+           if (waitqueue_active(&ep->wq))
+                 wake_up(&ep->wq);
+           if (waitqueue_active(&ep->poll_wait))
+                 wake_up(&ep->poll_wait);
+     }

      write_unlock_irqrestore(&ep->lock, flags);



(See attached file: epoll-2.5.46-insert.patch)

[-- Attachment #2: epoll-2.5.46-insert.patch --]
[-- Type: application/octet-stream, Size: 703 bytes --]

diff -ur linux-2.5.46.orig/fs/eventpoll.c linux-2.5.46/fs/eventpoll.c
--- linux-2.5.46.orig/fs/eventpoll.c	Thu Nov  7 14:31:21 2002
+++ linux-2.5.46/fs/eventpoll.c	Thu Nov  7 14:32:16 2002
@@ -838,8 +838,13 @@
 	list_add(&dpi->llink, ep_hash_entry(ep, ep_hash_index(ep, tfile)));
 
 	/* If the file is already "ready" we drop it inside the ready list */
-	if ((revents & pfd->events) && !EP_IS_LINKED(&dpi->rdllink))
+	if ((revents & pfd->events) && !EP_IS_LINKED(&dpi->rdllink)) {
 		list_add(&dpi->rdllink, &ep->rdllist);
+		if (waitqueue_active(&ep->wq))
+			wake_up(&ep->wq);
+		if (waitqueue_active(&ep->poll_wait))
+			wake_up(&ep->poll_wait);
+	}
 
 	write_unlock_irqrestore(&ep->lock, flags);
 

Re: [PATCH] 2.5.46: epoll ep_insert doesn't wake waiters if events exist

[Davide Libenzi]

On Thu, 7 Nov 2002, Jay Vosburgh wrote:

>
>       The logic in fs/eventpoll.c:ep_insert()  checks to see if events are
> already present when processing an EP_CTL_ADD operation, and, if any are,
> adds them to the list.  It does not wake up any waiters, so, e.g., another
> thread waiting in epoll_wait() will not be awakened for these events.  The
> fix is to have ep_insert() do wakeups, just as ep_poll_callback() does when
> it adds events.

Matthew Kirk already made me notice this yesterday and it's in my code ( 0.37 ).
I was ashame to make another post to Linus :) so I delayed the post.
Linus, I have a few bits. Let me know if you want them now or we will
delay to 2.5.47 ...



- Davide

Re: [PATCH] 2.5.46: epoll ep_insert doesn't wake waiters if events exist

[Mark Hamblin]

Looking at the documentation at
http://www.xmailserver.org/linux-patches/epoll_wait.txt, I noticed that the
description for epoll_wait states:  "On success, epoll_wait(2) returns the
number of file descriptors ready for the requested I/O."  My question is if
I have, for example, 1000 sockets registered with epoll and 100 of them have
received data and I call epoll_wait with max_events set to 10, will
epoll_wait return 10 or 100?  Furthermore, does the edge-triggered nature of
epoll "eat the edge" for the other 90 sockets even though they didn't get
returned?  Finally, if those 10 sockets get more data before I call
epoll_wait with max_events = 10 again, will those same 10 sockets get
returned?

- Mark

Re: [PATCH] 2.5.46: epoll ep_insert doesn't wake waiters if events exist

[Davide Libenzi]

On Fri, 8 Nov 2002, Mark Hamblin wrote:

> Looking at the documentation at
> http://www.xmailserver.org/linux-patches/epoll_wait.txt, I noticed that the
> description for epoll_wait states:  "On success, epoll_wait(2) returns the
> number of file descriptors ready for the requested I/O."  My question is if
> I have, for example, 1000 sockets registered with epoll and 100 of them have
> received data and I call epoll_wait with max_events set to 10, will
> epoll_wait return 10 or 100?  Furthermore, does the edge-triggered nature of
> epoll "eat the edge" for the other 90 sockets even though they didn't get
> returned?  Finally, if those 10 sockets get more data before I call
> epoll_wait with max_events = 10 again, will those same 10 sockets get
> returned?

epoll will return 10 and the remaining 90 will remain available for future
epoll_wait() calls. The last question will trigger a change in my code,
from list_add() to list_add_tail(). Since I don't want to stress Linus
with 24590 patches, I'll wait for 2.5.47 to merge the new code ...



- Davide