add support for busypolling #3462

sleepybishop · 2025-01-30T20:14:42Z

This pr add busypolling support to h2o. It is linux only.

For notes on busy polling and its internals, please see the following:

The current implementation make several assumptions that may need further thought

It is assumed that num-threads will be equivalent to the number of cpus configured in the busypoll map.
all interfaces will be declared in busypoll map using mode to specify busypoll behavior. Eg. non-busypolling interfaces will be included but marked with mode: OFF in the options subsection.
loopback interface will be assigned to 1 cpu

lib/common/busypoll.c

lib/handler/configurator/busypoll.c

include/h2o/busypoll.h

…e being passed into open_inet_listener and adjust the logic in create_per_thread_nic_map_listeners to call finish_open_listener for all listeners

src/main.c

…ate iface being passed into open_inet_listener and adjust the logic in create_per_thread_nic_map_listeners to call finish_open_listener for all listeners" This reverts commit daf55c7.

include/h2o/busypoll.h

src/main.c

lib/common/busypoll.c

t/90busypoll.t

kazuho

Thank you for the changes! I like how the changes are becoming isolated.

CMakeLists.txt

src/main.c

include/h2o/busypoll.h

Co-authored-by: Kazuho Oku <kazuhooku@gmail.com>

kazuho

Thank you for all the changes. I really like what we have now, all my comments below are nitpicks.

I might try myself to see if we can further reduce the changes to src/main.c, but I am starting to guess I'd fail, as it already looks quite polished up.

include/h2o/socket/evloop.h

include/h2o/busypoll.h

lib/common/socket/evloop/epoll.c.h

src/main.c

CMakeLists.txt

Co-authored-by: Kazuho Oku <kazuhooku@gmail.com>

jdamato-fsly · 2025-05-01T19:03:03Z

there's a bug in the loopback setup with this branch, issuing two requests to loopback causes it to fail:

thread 0 NAPI id: 0, iface: lo, listener fd: 4
found matching iface in conf map: lo
thread idx 0, has taken cpu: 0, for NIC: lo, via napi: 0 via listener 4

thread 0 NAPI id: 0, iface: lo, listener fd: 4
found matching iface in conf map: lo
no CPU found for thread with NAPI from NIC: lo:0
busy-poll-cpu-map configuration or queue setup is buggy -- exiting h2o now.

which shouldn't happen -- the right listener fd is being awoken, so either:

multiple threads have the FD for that listener in their epoll loop (they shouldn't)
or the code doesn't claim a CPU / thread after the first accept (seems unlikely, since the second request can find no available CPU, meaning the first request triggered the listener to take a CPU and pin itself to that CPU)

possibly worth adding a test which makes multiple requests to both loopback and non-loopback interfaces?

sleepybishop · 2025-05-01T20:10:34Z

possibly worth adding a test which makes multiple requests to both loopback and non-loopback interfaces?

in t/90busypoll.tright before sending large requests to the netdevsim interfaces theres a block that sends requests to loopback

    for (1..5) {
        my $resp = `ip netns exec nssv curl -ksi http://127.0.0.1:$port`;
    }

the same logic runs after the sends to the netdevsim interfaces as well, im not sure why it's not triggering the behavior you are seeing.

jdamato-fsly · 2025-05-01T22:03:55Z

the same logic runs after the sends to the netdevsim interfaces as well, im not sure why it's not triggering the behavior you are seeing.

Because the test is incorrect.

If you change the test configuration according to the following diff:

diff --git a/t/90busypoll.t b/t/90busypoll.t
index 87bad8346..3388450ef 100644
--- a/t/90busypoll.t
+++ b/t/90busypoll.t
@@ -185,7 +185,8 @@ busy-poll-map:
         suspend-timeout: $suspend_timeout
         mode: $mode
     - ifindex: 1
-      cpus: [$total_threads]
+      cpus:
+        - 0
       options:
         mode: OFF
 EOT

the problem reproduces:

  3.544553   % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
  3.544583                                  Dload  Upload   Total   Spent    Left  Speed
  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0thread 4 NAPI id: 0, iface: lo, listener fd: 11
  3.544789 found matching iface in conf map: lo
  3.544855 thread idx 4, has taken cpu: 0, for NIC: lo, via napi: 0 via listener 11
100   177  100   177    0     0   331k      0 --:--:-- --:--:-- --:--:--  172k
  3.551733   % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
  3.551763                                  Dload  Upload   Total   Spent    Left  Speed
  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0thread 4 NAPI id: 0, iface: lo, listener fd: 11
  3.551931 found matching iface in conf map: lo
  3.552273 no CPU found for thread with NAPI from NIC: lo:0
  3.552277 busy-poll-cpu-map configuration or queue setup is buggy -- exiting h2o now.
  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0
  3.553185 curl: (56) Recv failure: Connection reset by peer
  3.559614   % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
  3.559631                                  Dload  Upload   Total   Spent    Left  Speed
  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0
  3.559809 curl: (7) Failed to connect to localhost port 32773 after 0 ms: Couldn't connect to server
  3.565928   % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
  3.565981                                  Dload  Upload   Total   Spent    Left  Speed
  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0
  3.566234 curl: (7) Failed to connect to localhost port 32773 after 0 ms: Couldn't connect to server
  3.572766   % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
  3.572789                                  Dload  Upload   Total   Spent    Left  Speed
  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0
  3.572957 curl: (7) Failed to connect to localhost port 32773 after 0 ms: Couldn't connect to server

sleepybishop · 2025-05-01T22:47:55Z

If you change the test configuration according to the following diff:

This originated with the change to add the serving cpu index to as a logging directive, the cpu_claimed thread local changed from being a boolean to holding the index of the pinned cpu, that introduced an invalid assumption that the cpu indexes started at one. This has been addressed in 3f3a09b

add support for busypolling

35ca0f7

jdamato-fsly reviewed Jan 30, 2025

View reviewed changes

lib/common/busypoll.c Outdated Show resolved Hide resolved

jdamato-fsly reviewed Jan 30, 2025

View reviewed changes

lib/common/busypoll.c Show resolved Hide resolved

jdamato-fsly reviewed Jan 30, 2025

View reviewed changes

lib/common/busypoll.c Outdated Show resolved Hide resolved

jdamato-fsly reviewed Jan 30, 2025

View reviewed changes

lib/handler/configurator/busypoll.c Outdated Show resolved Hide resolved

include/h2o/busypoll.h Show resolved Hide resolved

sleepybishop added 2 commits January 31, 2025 08:52

both busypoll modes require libynl

bf340c7

move h2o_busypoll_bind_device to finish_open_listener, eliminate ifac…

daf55c7

…e being passed into open_inet_listener and adjust the logic in create_per_thread_nic_map_listeners to call finish_open_listener for all listeners

sleepybishop force-pushed the jcalderon/busypoller branch 2 times, most recently from 3feaff9 to 4a6a7ab Compare January 31, 2025 19:46

jdamato-fsly reviewed Jan 31, 2025

View reviewed changes

src/main.c Outdated Show resolved Hide resolved

sleepybishop added 2 commits January 31, 2025 11:51

use epoll_pwait2 directly

aca50c5

plumbing for epoll.prefer_busy_poll

fff1b13

sleepybishop force-pushed the jcalderon/busypoller branch from 4a6a7ab to fff1b13 Compare January 31, 2025 19:51

mark busypoll options experimental

df876a0

sleepybishop force-pushed the jcalderon/busypoller branch from cfec1b9 to 9be384e Compare January 31, 2025 21:09

Revert "move h2o_busypoll_bind_device to finish_open_listener, elimin…

a53af73

…ate iface being passed into open_inet_listener and adjust the logic in create_per_thread_nic_map_listeners to call finish_open_listener for all listeners" This reverts commit daf55c7.

sleepybishop force-pushed the jcalderon/busypoller branch from 9be384e to a53af73 Compare February 1, 2025 00:03

sleepybishop added 2 commits February 1, 2025 13:26

use epoll_pwait2 if H2O_HAS_YNL_H is set

ea7b934

remove fixme

77a1544

sleepybishop force-pushed the jcalderon/busypoller branch 2 times, most recently from febb4fe to e55be3c Compare February 5, 2025 00:25

set bp.prefer_busypoll for nics in suspend mode at assign time

2bed894

sleepybishop force-pushed the jcalderon/busypoller branch from ef09964 to 1f8387d Compare February 6, 2025 23:35

derive, save configure napi_ids for configured interfaces

29542d8

sleepybishop force-pushed the jcalderon/busypoller branch 2 times, most recently from 70c5bd5 to 426b9d8 Compare February 13, 2025 00:54

update ci to ubuntu2404 (noble)

e321105

sleepybishop force-pushed the jcalderon/busypoller branch from 426b9d8 to e321105 Compare February 13, 2025 01:48

sleepybishop added 2 commits February 13, 2025 14:12

standardize error handling

e4d01f4

restore get_napi_id

e120987

busypoll off should still map nic queues to cpus

20ea51b