[Intel-gfx] [PATCH] drm: Fix page flip ioctl format check

[Daniel Vetter]

On Mon, May 11, 2020 at 2:37 PM Ville Syrjälä
<ville.syrjala at linux.intel.com> wrote:
>
> On Sat, May 09, 2020 at 12:13:02PM +0200, Daniel Vetter wrote:
> > On Fri, May 8, 2020 at 7:09 PM Rodrigo Vivi <rodrigo.vivi at intel.com> wrote:
> > >
> > > On Fri, Apr 17, 2020 at 09:28:34PM +0300, Ville Syrjälä wrote:
> > > > On Fri, Apr 17, 2020 at 08:10:26PM +0200, Daniel Vetter wrote:
> > > > > On Fri, Apr 17, 2020 at 5:43 PM Ville Syrjälä
> > > > > <ville.syrjala at linux.intel.com> wrote:
> > > > > >
> > > > > > On Fri, Apr 17, 2020 at 05:23:10PM +0200, Daniel Vetter wrote:
> > > > > > > On Thu, Apr 16, 2020 at 08:04:20PM +0300, Ville Syrjala wrote:
> > > > > > > > From: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > > > > > > >
> > > > > > > > Revert back to comparing fb->format->format instead fb->format for the
> > > > > > > > page flip ioctl. This check was originally only here to disallow pixel
> > > > > > > > format changes, but when we changed it to do the pointer comparison
> > > > > > > > we potentially started to reject some (but definitely not all) modifier
> > > > > > > > changes as well. In fact the current behaviour depends on whether the
> > > > > > > > driver overrides the format info for a specific format+modifier combo.
> > > > > > > > Eg. on i915 this now rejects compression vs. no compression changes but
> > > > > > > > does not reject any other tiling changes. That's just inconsistent
> > > > > > > > nonsense.
> > > > > > > >
> > > > > > > > The main reason we have to go back to the old behaviour is to fix page
> > > > > > > > flipping with Xorg. At some point Xorg got its atomic rights taken away
> > > > > > > > and since then we can't page flip between compressed and non-compressed
> > > > > > > > fbs on i915. Currently we get no page flipping for any games pretty much
> > > > > > > > since Mesa likes to use compressed buffers. Not sure how compositors are
> > > > > > > > working around this (don't use one myself). I guess they must be doing
> > > > > > > > something to get non-compressed buffers instead. Either that or
> > > > > > > > somehow no one noticed the tearing from the blit fallback.
> > > > > > >
> > > > > > > Mesa only uses compressed buffers if you enable modifiers, and there's a
> > > > > > > _loooooooooooot_ more that needs to be fixed in Xorg to enable that for
> > > > > > > real. Like real atomic support.
> > > > > >
> > > > > > Why would you need atomic for modifiers? Xorg doesn't even have
> > > > > > any sensible framework for atomic and I suspect it never will.
> > > > >
> > > > > Frankly if no one cares about atomic in X I don't think we should do
> > > > > work-arounds for lack of atomic in X.
> > > > >
> > > > > > > Without modifiers all you get is X tiling,
> > > > > > > and that works just fine.
> > > > > > >
> > > > > > > Which would also fix this issue here you're papering over.
> > > > > > >
> > > > > > > So if this is the entire reason for this, I'm inclined to not do this.
> > > > > > > Current Xorg is toast wrt modifiers, that's not news.
> > > > > >
> > > > > > Works just fine. Also pretty sure modifiers are even enabled by
> > > > > > default now in modesetting.
> > > > >
> > > > > Y/CSS is harder to scan out, you need to verify with TEST_ONLY whether
> > > > > it works. Otherwise good chances for some oddball black screens on
> > > > > configurations that worked before. Which is why all non-atomic
> > > > > compositors reverted modifiers by default again.
> > > >
> > > > Y alone is hard to scanout also, and yet we do nothing to reject that.
> > > > It's just an inconsistent mess.
> > > >
> > > > If we really want to keep this check then we should rewrite it
> > > > to be explicit:
> > > >
> > > > if (old_fb->format->format != new_fb->format->format ||
> > > >     is_ccs(old_fb->modifier) != is_ccs(new_fb->modifier))
> > > >     return -EINVAL;
> > > >
> > > > Now it's just a random thing that may even stop doing what it's
> > > > currently doing if anyone touches their .get_format_info()
> > > > implementation.
> > > >
> > > > >
> > > > > > And as stated the current check doesn't have consistent behaviour
> > > > > > anyway. You can still flip between different modifiers as long a the
> > > > > > driver doesn't override .get_format_info() for one of them. The *only*
> > > > > > case where that happens is CCS on i915. There is no valid reason to
> > > > > > special case that one.
> > > > >
> > > > > The thing is, you need atomic to make CCS work reliably enough for
> > > > > compositors and distros to dare enabling it by default.
> > > >
> > > > If it's not enabled by default then there is no harm in letting people
> > > > explicitly enable it and get better performance.
> > > >
> > > > > CCS flipping
> > > > > works with atomic. I really see no point in baking this in with as
> > > > > uapi.
> > > >
> > > > It's just going back to the original intention of the check.
> > > > Heck, the debug message doesn't even match what it's doing now.
> > > >
> > > > > Just fix Xorg.
> > > >
> > > > Be serious. No one is going to rewrite all the randr code to be atomic.
> > >
> > > I fully understand Daniel's concern here, but I also believe this won't be
> > > done so soon at least. Meanwhile would it be acceptable to have a comment
> > > with the code /* XXX: Xorg blah... */ or /* FIXME: After Xorg blah.. */
> > > ?
> >
> > Here's a few numbers:
> >
> > - skl shipped in Aug 2015, so about 5 years. Since then would we like
> > to have modifiers enabled for intel, because it costs us quite a bit
> > of performance. This isn't new at all.
> > - the last Xorg release is from May 2018, so two years. Meanwhile even
> > patches to fix some of the atomic mixups in -modesetting landed, but
> > they never shipped so not useful.
> > - I spent a few hours (which really is nothing) reading Xorg code
> > yesterday, and I concur with Daniel Stone's napkin estimate that this
> > will take about half to one year to fix properly. It's not happening,
> > no one is working on that.
> >
> > Conclusion: No one cares about modifiers on Xorg-modesetting. I don't
> > see why the kernel should bend over for that.
> >
> > Once that has changed (I'm not betting on that) and there's clear
> > effort behind modifiers for Xorg-modesetting I guess we can look into
> > stop-gap measures, but meanwhile the best imo is to not disturb the
> > dead.
>
> The alternative interpretation is that the current kernel code is
> just nonsense, and since no one is depending on the current nonsense
> behaviour we can safely change it it back to make sense.
>
> Would allow people to at least test modifier plumbing via dri3/etc.
> Also those of us who know what they're doing and want to actually
> play games on Intel GPUs can flip it on for a a bit extra performance.
> In the meantime I'll just have to keep carrying this patch in my own
> kernels.

You can also carry a one-liner for -modesetting to re-enable atomic on
master (it's fixed up there, simply never released, why we've had to
take it away). And then you can also play with modifiers.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch