Thread (5 messages) 5 messages, 2 authors, 2024-07-31

Re: [PATCH v3] epoll.7: clarify the event distribution under edge-triggered mode

From: Alejandro Colomar <alx@kernel.org>
Date: 2024-07-31 10:39:40
Also in: linux-api

Hi Andy,

On Tue, Jul 30, 2024 at 10:48:28AM GMT, Andy Pan via B4 Relay wrote:
quoted hunk ↗ jump to hunk
From: Andy Pan <redacted>

For the moment, the edge-triggered epoll generates an event for each
receipt of a chunk of data, that is to say, epoll_wait() will return
and tell us a monitored file descriptor is ready whenever there is a
new activity on that FD since we were last informed about that FD.
This is not a real _edge_ implementation for epoll, but it's been
working this way for years and plenty of projects are relying on it
to eliminate the overhead of one system call of read(2) per wakeup event.

There are several renowned open-source projects relying on this feature
for notification function (with eventfd): register eventfd with EPOLLET
and avoid calling read(2) on the eventfd when there is wakeup event (eventfd being written).
Examples: nginx [1], netty [2], tokio [3], libevent [4], ect. [5]
These projects are widely used in today's Internet infrastructures.
Thus, changing this behavior of epoll ET will fundamentally break them
and cause a significant negative impact.
Linux has changed it for pipe before [6], breaking some Android libraries,
which had got "reverted" somehow. [7] [8]

Nevertheless, the paragraph in the manual pages describing this
characteristic of epoll ET seems ambiguous, I think a more explict
sentence should be used to clarify it. We're improving the notification
mechanism for libuv recently by exploiting this feature with eventfd,
which brings us a significant performance boost. [9]

Therefore, we (as well as the maintainers of nginx, netty, tokio, etc.)
would have a sense of security to build an enhanced notification function
based on this feature if there is a guarantee of retaining this implementation
of epoll ET for the backward compatibility in the man pages.

[1]: https://github.com/nginx/nginx/blob/efc6a217b92985a1ee211b6bb7337cd2f62deb90/src/event/modules/ngx_epoll_module.c#L386-L457
[2]: https://github.com/netty/netty/pull/9192
[3]: https://github.com/tokio-rs/mio/blob/309daae21ecb1d46203a7dbc0cf4c80310240cba/src/sys/unix/waker.rs#L111-L143
[4]: https://github.com/libevent/libevent/blob/525f5d0a14c9c103be750f2ca175328c25505ea4/event.c#L2597-L2614
[5]: https://github.com/libuv/libuv/pull/4400#issuecomment-2123798748
[6]: https://lkml.iu.edu/hypermail/linux/kernel/2010.1/04363.html
[7]: https://github.com/torvalds/linux/commit/3a34b13a88caeb2800ab44a4918f230041b37dd9
[8]: https://github.com/torvalds/linux/commit/3b844826b6c6affa80755254da322b017358a2f4
[9]: https://github.com/libuv/libuv/pull/4400#issuecomment-2103232402

Signed-off-by: Andy Pan <redacted>
---
Changes in v3:
- Updated the git commit description
- Link to v2: https://lore.kernel.org/r/20240727-epoll-et-desc-v2-1-c99b2ac66775@andypan.me (local)

Changes in v2:
- Added the git commit description based on feedback
- Link to v1: https://lore.kernel.org/r/20240727-epoll-et-desc-v1-1-390bafc678b9@andypan.me (local)
---
 man/man7/epoll.7 | 1 +
 1 file changed, 1 insertion(+)
diff --git a/man/man7/epoll.7 b/man/man7/epoll.7
index 951500131..361d9db99 100644
--- a/man/man7/epoll.7
+++ b/man/man7/epoll.7
@@ -172,6 +172,7 @@ .SS Level-triggered and edge-triggered
 Since even with edge-triggered
 .BR epoll ,
This paragraph is not only about EPOLLET.
 multiple events can be generated upon receipt of multiple chunks of data,
+that is, an event will be generated upon each receipt of a chunk of data,
Is this strictly true in all modes?  The commit message seems to say
that this is the behavior of EPOLLET, but this paragraph is not only
about EPOLLET.  Is this correct here, or should it go somewhere else?

Have a lovely day!
Alex
 the caller has the option to specify the
 .B EPOLLONESHOT
 flag, to tell

---
base-commit: cbc0a111e4dceea2037c51098de33e6bc8c16a5c
change-id: 20240727-epoll-et-desc-04ea9a590f3b

Best regards,
-- 
Andy Pan [off-list ref]
-- 
<https://www.alejandro-colomar.es/>

Attachments

Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help