1724420 - An "SCTP timer" thread remains in every content process that used a peer connection, and uses CPU

Status: RESOLVED FIXED

(Core :: WebRTC: Networking, defect)

Description

[Florian Quèze [:florian]]

I have two content processes that have used peer connections (but afaik these connections no longer exist). Each of them has an "SCTP timer" thread that according to about:processes (with toolkit.aboutProcesses.showThreads set to true in about:config) uses 0.2% of a CPU core each.

Comment 1

[Florian Quèze [:florian]]

I don't know what this thread does with the CPU, it is named at https://searchfox.org/mozilla-central/rev/67637695876e27ddbfb6ff99202bb1c951de847a/netwerk/sctp/src/netinet/sctp_callout.c#198 but doesn't seem to be registered with the Gecko Profiler.

Comment 2

[Valentin Gosu [:valentin] (he/him)]

I believe sctp is only used for webrtc. Maybe Nils knows if we can do something about it.

Comment 3

[Florian Quèze [:florian]]

This thread seems to remain active forever in any content process that ever used a data channel (I verified using the demo at https://tomashubelbauer.github.io/webrtc-data-channel-demo/src/index.html). I see this in Slack tabs, my bank's website, air mozilla. https://searchfox.org/mozilla-central/rev/67637695876e27ddbfb6ff99202bb1c951de847a/netwerk/sctp/src/netinet/sctp_callout.c#193,201,206,209 causes a CPU wakeup every 10ms, which probably wastes a lot of power.

So questions:

1. Is this 10ms repeating timer really needed in the first place? Could the same result be achieved in a more power efficient way?
2. Can we ensure this thread is stopped when it's no longer useful?
3. Would it be possible to register this thead with the Gecko Profiler?

Comment 4

[Byron Campen [:bwc]]

(In reply to Florian Quèze [:florian] from comment #3)

This thread seems to remain active forever in any content process that ever used a data channel (I verified using the demo at https://tomashubelbauer.github.io/webrtc-data-channel-demo/src/index.html). I see this in Slack tabs, my bank's website, air mozilla. https://searchfox.org/mozilla-central/rev/67637695876e27ddbfb6ff99202bb1c951de847a/netwerk/sctp/src/netinet/sctp_callout.c#193,201,206,209 causes a CPU wakeup every 10ms, which probably wastes a lot of power.

So questions:

1. Is this 10ms repeating timer really needed in the first place? Could the same result be achieved in a more power efficient way?

Unsure, but this thread is internal to libusrsctp, so our control here is limited.

2. Can we ensure this thread is stopped when it's no longer useful?

Currently, no. See bug 1646716. We're waiting on a TSan fix from upstream, and once we have that, we can then evaluate whether it is safe to try.

3. Would it be possible to register this thead with the Gecko Profiler?

So this thread is internal to libusrsctp, but maybe we could wait for the first callback we get from that thread and register at that time?

Comment 5

[Byron Campen [:bwc]]

Attachment: Bug 1724420: Init/de-init libusrsctp as needed. r?ng

Comment 6

[Byron Campen [:bwc]]

https://treeherder.mozilla.org/#/jobs?repo=try&revision=f56f8d6eefbed72c29f79cd22c5b74838a44c382

https://treeherder.mozilla.org/#/jobs?repo=try&revision=f3cd9dd065ae52a6cb509bc24ee4fdc2280f056f

Comment 7

[Byron Campen [:bwc]]

https://treeherder.mozilla.org/#/jobs?repo=try&revision=3a61b5e00d2a13d869f9a25c0e17f265fce26060

https://treeherder.mozilla.org/#/jobs?repo=try&revision=8d46dc1aaf7873cb2c96e4ae2190abb49dd55a3d

Comment 8

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/4236c8dfe678
Init/de-init libusrsctp as needed. r=ng

Comment 9

[Cristian Tuns]

https://hg.mozilla.org/mozilla-central/rev/4236c8dfe678