[Intel-gfx] Fwd: [PATCH] drm/i915: Avoid vblank counter for gen9+

[Patrik Jakobsson]

On Wed, Mar 02, 2016 at 07:13:07PM +0200, Imre Deak wrote:
> On Fri, 2016-02-26 at 10:02 -0800, Rodrigo Vivi wrote:
> > [...]
> > Well, I have this tree:
> > https://cgit.freedesktop.org/~vivijim/drm-intel/log/?h=rpm-domains-psr-vblank-counter-full
> > with mainly:
> > 1 - vblank domain on pre-enable post-disable vblanks hooks as Ville
> > had suggested
> > 2 - psr domain so we just enable dc state when screen is really in
> > idle.
> > 3 - restore counter on vblank enable.
> > 
> > From what I understood so far of this problem, only the patch 1
> > should
> > be enough, but with only this one I don't get the screen frozen but
> > the typying is so slow that is visible that we have something
> > wrong.... Maybe dc state transition with mutexes there are slow?
> 
> I'm not aware of any big latencies caused by toggling DC states alone.

According to bspec, any MMIO access when in DC6 have a considerably higher
latency. The recommendation from bspec is to disable DC states around longer
sequences of MMIO. I don't know if this is the case here but going from DC6 ->
DC3 is a quite heavy operation (enable PG0, PG1, CDCLK and do CSR for these). I
guess we could measure it to see how long it actually takes?

> 
> > Patch 2 by iitself also doesn't solve this and I still have frozen
> > screens, but when combined to  patch 1 everything works really
> > well...
> > In the point that I believe we really don't need patch 3.
> 
> I think something like 1 and 3 is a good idea (and both are needed).
> About 2, it's strange that you have to disable DC states when enabling
> the panel. Since the pipe is active it should prevent DC5 (and hence
> DC6). We wouldn't waste any power with your changes, since you re-
> enable DC states before entering PSR, but imo we should find out why
> exactly this is needed.
> 
> Some notes/ideas about the patches:
> - intel_display_power_get() is called from page_flip_completed(), which
>  is bad since we can be in interrupt context.
> - There is a drm_crtc_vblank_get() in intel_crtc_page_flip(), but there
> is no corresponding intel_display_power_get() for it.
> - The same goes for the FBC, CRC code, couldn't you just call the new
> vblank hooks from drm_vblank_get/put()?
> - PIPE_FLIPCOUNT_G4X is also read-only, so it could get corrupted the
> same way as the frame counter register. page_flip_finished() depends on
> PIPE_FLIPCOUNT_G4X, so isn't this a problem?
> 
> I haven't checked this in detail, but it could be that we need to exit
> PSR explicitly when waiting for a vblank or doing a flip. In PSR mode
> the pipe may not be running, so I'm not sure how the vblank and flip
> interrupts would be delivered.
> 
> --Imre
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
---
Intel Sweden AB Registered Office: Knarrarnasgatan 15, 164 40 Kista, Stockholm, Sweden Registration Number: 556189-6027