[Intel-gfx] Breakage for Ironlake due to some watermarks changes in Linux 4.0+?

[Mario Kleiner]

Hi all,

since Linux 4.0 i experience some massive display flicker problem on my 
Intel HD Ironlake mobile (2010 MacBookPro6,2) under Waylands reference 
compositor Weston.

- Only happens on Linux >= 4.0 on intel-kms with the Intel HD, not under 
nouveau-kms with the discrete NVidia gpu. Strangely on Linux 4.1-rc it 
happens all the time, whereas on Linux 4.0 it can work normally for 
quite a while, but once the problem starts only a reboot can cure it.

- Almost only happens on Weston, but only very rarely under the XServer. 
VT switching from Weston to XOrg makes the problem disappear, switching 
back to Weston and it starts again immediately.

- Only happens if a hardware cursor is displayed - hiding the cursor 
stops the flicker immediately, showing the cursor starts the flicker.

- The drm and desktop is completely idle during this - drm.debug=15 
shows no activity while this happens.

Symptom:

Up to the scanline where the cursor is located, the desktop image is 
displayed, but jumps horizontally left and right by some random number 
of pixels, maybe in the range 0 - 200 pixels with high frequency, making 
the content unreadable. Starting with the scanline where scanout of the 
cursor starts, the display goes blank, as if some display controller 
fifo would underflow and the controller blanks the display in response. 
Seems having to scanout the cursor plane in addition to the primary 
plane is just enough to push it over some limit?

I also see cpu and pch pipe a fifo underruns reported by the underflow 
irq handlers.

I saw there were many changes around Linux 4.0 in the kms driver wrt. 
watermark calculations, so this might be related?

thanks,
-mario