Surfaceflinger takes more that 50% of the CPU usage on the latest Android builds

[Zach Pfeffer]

ARM found this while profiling Snowball on DS-5. We suspect all platforms have this problem.

[Bernhard Rosenkraenzer (berolinux)]

Off the top of my head, I'd say this is triggered by the latest kernels: EARLYSUSPEND stuff is gone in current kernels, and surfaceflinger used to spam the logs with around 10 warnings about that per second. I'm not sure that has been fixed...

Can you confirm/deny that theory?

[Amit Pundir (pundiramit)]

We suspect the same EARLYSUSPEND issue here. This is supposed to be fixed in AOSP's 3.4 tree but not sure if it fixed there or not.

[Yann BENIGOT (yann-benigot-w)]

An quick fix to this is to add a usleep() call in the DisplayHardwareBase::DisplayEventThread::threadLoop() function in case FB_EARLYSUSPEND is missing so that the thread stops using all CPU.
With the attached patch, the threadLoop() function will only be called six times per second. Perhaps it would be possible to suspend the thread entirely when FB_EARLYSUSPEND is missing since the thread does not seem to be doing anything.

It reduced surfaceflinger's CPU usage to 5-7% with linaro-wallpaper, 0-1% without.

[Yann BENIGOT (yann-benigot-w)]

Seems to be a duplicate of https://bugs.launchpad.net/linaro-android/+bug/1023985

[John Stultz (jstultz)]

So what version of userland is being used here?

Since with AOSP 4.2, the suspend/display managment logic has been moved from surfaceflinger to the PowerManagerService.

See commit message here for details: https://android.googlesource.com/platform/frameworks/base/+/9630704ed3b265f008a8f64ec60a33cf9dcd3345