Hi all,
after obsessing over these for another month after xds 2010, this is the first version of my vblank timestamping patches that i'm not totally ashamed off.
1st patch modifies the drm vblank handling to maintain precise timestamps of when vblanks happen. It timestamps the end of each vblank interval (= start of active scanout of the following frame), as required by the OpenML OML_sync_control extension spec, so the DRI2 swap & sync bits should return proper timestamps to gl clients. It also takes care of keeping timestamps and vblcounts consistent across vblank irq on/off and hopefully fixes a few off-by-one vblank counts due to race conditions in those on/off bits.
There is still one small race condition between the vblank on/off bits and the gpu (updating its hardware vblank counter at the wrong moment) which can cause miscounting of vblanks across enable/disable. I have an idea how to fix that, based on these patches but will need to tinker around with it a bit more. After that we could reduce the vblankoffdelay from 5000 msecs to something smaller like 100 msecs to save more power on desktops with desktop composition enabled, where already a blinking cursor in a terminal window or a ticking clock in the menu bar can keep vblanks turned on all the time with a 5000 msecs timeout.
All the timestamping inside the drmWaitVblank ioctl() and in vblank event handling now uses the timestamps calculated in this patch.
2nd and 3rd patches add the new timestamping and scanout position functions to the i915 and radeon drivers.
Wrt. what i showed you at xds, i made the following changes:
I've moved the actual timestamping routine as a helper function into the core. kms drivers now optionally export the crtc scanout position query function in addition to the vblank timestamping function to the drm core. Whenever the core has to recalculate timestamps it goes like:
core -> kmsdriver-timestamping() -> core-timestampinghelper() -> kmsdriver-getscanoutposition()
The kms driver can implement its own version of the timestamping function, e.g., using special hardware timestamping registers in some more recent gpu's, or it can simply export a getscanoutposition() function with standardized behaviour across drivers and call the timestamping routine implemented in the core. The routine in the core works in non-interlaced, non-doublescan video modes, with or without panel scaling. It can compensate for any random delay up to 1 video refresh duration. For more robustness one would need to use a gpu specific hardware method.
I also moved the calculation of the constants needed for timetamping into the drm_crtc_helper routine, so they only get recalculated after a mode switch, not at each vblank. At modeset time i cache the adjusted_mode after validation/adjustment by the crtc and encoders inside the drm_crtc struct. Needed to account for panel scaling, and i have a few other ideas that would require that cached value.
More details in the different commit messages and the code, just some more remarks.
@Jesse: I replaced the n extra dynamically allocated spin-locks for n crtc's by one statically allocated lock shared by all crtc's and renamed it. I think i also worked all your other suggestions into the patches. Thanks a lot for your first review of the earlier version.
The intel driver uses the timestamps inside the finish pageflip function for timestamps of bufferswap completion. I've tested with an Atom Mini netbook with Intel 945 GME gpu and there it works perfect. I don't have any other intel test machines, so the finish pageflip and scanout query functions for ironlake, gen6 and g4x aren't tested, only coded based on some of the intel manuals at xorg. In the pageflip irq handlers there is this uncertainty if pageflip irq's get delivered before or after the vblank irq's for the vblank interval of swap completion. The i945 delivers pagflip irq's before handling the vblank irq's so the completion handler needs to account for that. I guessed from the code, which ones do deliver before/after vblank, but that guesses may be wrong, so this needs testing.
@Alex: Thanks for your feedback at xds. I think the radeon part should be as we discussed. The radeon patch builds onto what you committed already into drm-radeon-testing for the reclocking fixes. It slightly changes the interface between the scanout query function and radeon_pm_in_vbl(), but doesn't change the implementation.
Testing of the timestamps on the Intel 945GME, a Radeon R500 and a R600 against external measurement equipment shows that the timestamps are essentially perfect. Error wrt. to the external measurement is less than 20 microseconds (for the radeon's) and jitter is +/- 1 usec. The intel error is less than 0.5 msecs, but i can't test how much better because my photo-diode can't do better and my usec precise equipment only works with DVI, but the intel test machine only has a vga output.
Now this swapbuffers timestamping is only trustworthy for pageflipped swaps, ie. at the moment only on intel. Alex is working on pageflipping support for radeon and i can test those bits and add timestamping. For the swap by blitting case, more work needs to be done.
My current test scripts are written as matlab/octave scripts against my toolkit (http://www.psychtoolbox.org). I could prepare a little zip file with the minimum bits needed to run those under gnu/octave. I'll translate the most essential tests into c in the next couple of days if there is some kind of test suite for this?
Thanks, -mario