Hi Daniel, On Tue, Jul 18, 2017 at 09:35:03AM +0200, Daniel Vetter wrote: > On Tue, Jul 18, 2017 at 9:07 AM, Maxime Ripard > wrote: > > On Mon, Jul 17, 2017 at 02:57:19PM +0800, Chen-Yu Tsai wrote: > >> On Mon, Jul 17, 2017 at 2:55 PM, Maxime Ripard > >> wrote: > >> > On Fri, Jul 14, 2017 at 04:56:01PM +0800, Chen-Yu Tsai wrote: > >> >> Hi, > >> >> > >> >> On Thu, Jul 13, 2017 at 10:41 PM, Maxime Ripard > >> >> wrote: > >> >> > In the earlier display engine designs, any register access while a commit > >> >> > is pending is forbidden. > >> >> > > >> >> > One of the symptoms is that reading a register will return another, random, > >> >> > register value which can lead to register corruptions if we ever do a > >> >> > read/modify/write cycle. > >> >> > >> >> Alternatively, if changes to the backend (layers) are guaranteed to happen > >> >> while the CRTC is disabled (which seems to be the case after looking at > >> >> drm_atomic_helper_commit_planes and drm_atomic_helper_commit_tail), we > >> >> could just turn on register auto-commit all the time and not deal with > >> >> this. > >> > > >> > As far as I understand, it will only be the case if we need a new > >> > modeset or we changed the active CRTC or connectors. But if you change > >> > only the format, buffers or properties it won't be the case, and we'll > >> > need to commit. > >> > >> So in other words, if someone were to use it for actual compositing and > >> moved the upper composited layer around, we would need commit support to be > >> safe. > >> > >> Sounds more or less like something a video player would do. > > > > Not only that. A change of buffer will happen every frame or so, and > > we can change the format whenever we want too (even if it's usually > > going to be in sync with a new buffer). Changing a property can happen > > any time too (like zpos for example). > > You can upgrade any property change to an atomic modeset by e.g. > setting connector->mode_changed (and then making sure to call > check_modeset() helper again perhaps). This is for cases where your hw > can't handle a property change within 1 vblank. The default is just > the solution for most common hw. > > The other way round works too, you can clear these flags in your > atomic_check callbacks. But that requires a bit more care (to make > sure you never clear it when there's something else also changing that > still needs a full modeset sequence to commit to hw). Hmm, that's good to know, but that would imply disabling the CRTC each time we change even a small property, with all the visual artifacts it might imply, right? Maxime -- Maxime Ripard, Free Electrons Embedded Linux and Kernel engineering http://free-electrons.com