On Thu, Aug 03, 2023 at 01:43:08PM +0200, Maxime Ripard wrote: > On Thu, Aug 03, 2023 at 12:26:03PM +0200, Daniel Vetter wrote: > > On Thu, 3 Aug 2023 at 11:22, Maxime Ripard <mripard@xxxxxxxxxx> wrote: > > > > > > On Thu, Aug 03, 2023 at 10:51:57AM +0200, Daniel Vetter wrote: > > > > On Thu, Aug 03, 2023 at 10:48:57AM +0200, Maxime Ripard wrote: > > > > > On Thu, Aug 03, 2023 at 10:11:22AM +0200, Neil Armstrong wrote: > > > > > > Hi, > > > > > > > > > > > > On 18/07/2023 17:31, Michael Riesch wrote: > > > > > > > Hi all, > > > > > > > > > > > > > > This series adds support for the partial display mode to the Sitronix > > > > > > > ST7789V panel driver. This is useful for panels that are partially > > > > > > > occluded by design, such as the Jasonic JT240MHQS-HWT-EK-E3. Support > > > > > > > for this particular panel is added as well. > > > > > > > > > > > > > > Note: This series is already based on > > > > > > > https://lore.kernel.org/lkml/20230714013756.1546769-1-sre@xxxxxxxxxx/ > > > > > > > > > > > > I understand Maxime's arguments, but by looking closely at the code, > > > > > > this doesn't look like an hack at all and uses capabilities of the > > > > > > panel controller to expose a smaller area without depending on any > > > > > > changes or hacks on the display controller side which is coherent. > > > > > > > > > > > > Following's Daniel's summary we cannot compare it to TV overscan > > > > > > because overscan is only on *some* displays, we can still get 100% > > > > > > of the picture from the signal. > > > > > > > > > > Still disagree on the fact that it only affects some display. But it's > > > > > not really relevant for that series. > > > > > > > > See my 2nd point, from a quick grep aside from i915 hdmi support, no one > > > > else sets all the required hdmi infoframes correctly. Which means on a > > > > compliant hdmi tv, you _should_ get overscan. That's how that stuff is > > > > speced. > > > > > > > > Iirc you need to at least set both the VIC and the content type, maybe > > > > even more stuff. > > > > > > > > Unless all that stuff is set I'd say it's a kms driver bug if you get > > > > overscan on a hdmi TV. > > > > > > I have no doubt that i915 works there. The source of my disagreement is > > > that if all drivers but one don't do that, then userspace will have to > > > care. You kind of said it yourself, i915 is kind of the exception there. > > > > > > The exception can be (and I'm sure it is) right, but still, it deviates > > > from the norm. > > > > The right fix for these is sending the right infoframes, _not_ trying > > to fiddle with overscan margins. Only the kernel can make sure the > > right infoframes are sent out. If you try to paper over this in > > userspace, you'll make the situation worse, not better (because > > fiddling with overscan means you get scaling, and so rescaling > > artifacts, and for hard contrasts along pixel lines that'll look like > > crap). > > > > So yeah this is a case of "most upstream hdmi drivers are broken". > > Please don't try to fix kernel bugs in userspace. > > ACK. > > > > > > I think I'll still like to have something clarified before we merge it: > > > > > if userspace forces a mode, does it contain the margins or not? I don't > > > > > have an opinion there, I just think it should be documented. > > > > > > > > The mode comes with the margins, so if userspace does something really > > > > funny then either it gets garbage (as in, part of it's crtc area isn't > > > > visible, or maybe black bars on the screen), or the driver rejects it > > > > (which I think is the case for panels, they only take their mode and > > > > nothing else). > > > > > > Panels can usually be quite flexible when it comes to the timings they > > > accept, and we could actually use that to our advantage, but even if we > > > assume that they have a single mode, I don't think we have anything that > > > enforces that, either at the framework or documentation levels? > > > > Maybe more bugs? We've been slowly filling out all kinds of atomic kms > > validation bugs in core/helper code because as a rule of thumb, > > drivers get it wrong. Developers test until things work, then call it > > good enough, and very few driver teams make a serious effort in trying > > to really validate all invalid input. Because doing that is an > > enormous amount of work. > > > > I think for clear-cut cases like drm_panel the fix is to just put more > > stricter validation into shared code (and then if we break something, > > figure out how we can be sufficiently lenient again). > > Panels are kind of weird, since they essentially don't exist at all in > the framework so it's difficult to make it handle them or their state. > > It's typically handled by encoders directly, so each and every driver > would need to make that check, and from a quick grep, none of them are > (for the reasons you said). I think the panel bridge driver infra is the right spot to put this, and then push drivers a bit more towards using that. Because yeah if they hand-roll the panel integration, they'll probably miss a lot of these details :-/ -Sima > > Just like for HDMI, even though we can commit to changing those facts, > it won't happen overnight, so to circle back to that series, I'd like a > comment in the driver when the partial mode is enabled that if userspace > ever pushes a mode different from the expected one, we'll add the margins. > > That way, if and when we come back to it, we'll know what the original > intent and semantics were. > > Maxime -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch