On 26/09/2019 21:08, Kevin Hilman wrote:

Neil Armstrong [off-list ref] writes:

quoted

On 25/09/2019 23:35, Kevin Hilman wrote:

quoted

From: Kevin Hilman <khilman@baylibre.com>

During init, ensure that the driver on/off state as well as clock and
reset state matches the hardware state.  Do this by always calling the
drivers 'on' function, and then callling the 'off' function if the
HW state was initially detected as off.

[...]

quoted

I don't see what you are trying to solve except simplifying the code.

Simplifying the code is a worthwhile goal on its own, but that's not the
only thing I'm tring to accomplish.

I still find it ugly to power_on a domain to power it off right afterwards.
The issue is with the CCF enable handling which is not in sync with the
HW, if you boot with an already enabled clock, it won't be marked enabled
in CCF, and it's clearly bad when you want to have a fine-tuned gate state
handling.

Part of the goal is make the init less VPU specific and thus make it
more understandable/maintainable.  But the more important part is to
ensure that the driver state and HW state is consistent for all the
domains (not just VPU.)  I've had to debug lots of power-domain issues,
and inconsistiences between HW state and driver state tend to be the
primary cause of problems.

Also I'm generally not a fan of the "don't mess with bootloader state"
approach as that leads to subtle dependencies on specific bootloader
versions that are also difficult to debug.

Antother equally important goal is to actually be able to power-down the
VPU when it's not used.  Right now, we'll never power off the VPU if the
bootloader enabled it, and that seems a bit extreme so I'm looking to
improve that and be able to power off the VPU when not used.

quoted

And the case is more that "matching the clock state" here, the
pm_domain_always_on_gov was is a real case when booting from the Amlogic
U-boot.

I'm not understanding what you mean about vendor uboot.  I've done
testing with vendor uboot too:

I tried on g12a-u200, g12a-x96-max, and sm1-khadas-vim3l all of which
have vendor uboot, and I tried before and after $SUBJECT patch.

In all cases, I see the vendor uboot splash screen, and I see the
framebuffer console from linux after kernel boot.  I see the same
behavior before and after my patch.

I also tried on g12b-odroid-n2 (vendor uboot), and there is _no_ uboot
spash screen, but I see the linux framebuffer console come up both
before and after my patch.

Thanks for testing all these cases

What's the specific case you're worried about with vendor uboot?

It's an issue I got when bringing up mainline uboot and the vpu power controller
driver, I had regressions on GXBB and GXL boards.
But it seems it's no more the case on G12A/G12B, we'll see this when
GX support will be added in this driver.

Also...  note something interesting I noticed on sm1-khadas-vim3l:
before my patch, the framebuffer console appears, but the background is
a bluish/green color.  After my patch, the background is black (as
expected.)  

Yes it's an issue I have on my infinite todo list, it's a different
init done by the latest vendor uboot. I have a fix but it seems it
breaks display when booting other boards.

quoted

The display power domain is complex and as been half solved by using
"simple-framebuffer" on gx and is missing on g12a/g12b/sm1.

For example, Debian installer runs without the modules, but will use
the EFIfb set by U-Boot, but in this precise case :
- the DRM driver isn't loaded
- we can't hook this power domain with EFIfb

OK, so I agree that this case where we want the display to continue to
work but linux DRM drivers never loaded is a special case.

Is there a way to check if efifb is enabled?  Is the the linux driver
(drivers/video/fbdev/efifb.c) or something else only done by uboot?

If we can detect efifb from the kernel (not just whether the domain is
already on), then a simple addition to my patch like this:

 	if (is_off)
 		meson_ee_pwrc_off(&dom->base);
+	else if (efifb_is_enabled)
+		dom->base.flags |= GENPD_FLAG_ALWAYS_ON;

would force the domain always-on in the case of efifb.  

In fact, now that I think of it, simply adding:

 	if (is_off)
 		meson_ee_pwrc_off(&dom->base);
+	else
+		dom->base.flags |= GENPD_FLAG_ALWAYS_ON;

to my current patch would get back to the same behavior of the existing
driver.  But I still don't like the idea that we can *never* power off
the VPU if the bootloader enables it. :( I'd rather see patches
conditionally adding that flag with detailed explanations as to why it's
needed.

quoted

When *not* in EFIfb, we use simple-framebuffer on GX, using this
power domain, but it hasn't been copied to G12A.

I don't quite understand what problem simple-framebuffer is
solving. Could you explain that in more detail.

simple-framebuffer has the power domain hooked in it's node, so
when u-boot will boot linux with HDMI enabled it will enable
this node and until the DRM driver loads the simple-framebuffer
will live and keep the power domain enabled.

Assuming it is solving something, why can't it be used on g12[ab]/sm1 ?

It will, but I need to push the patches.

quoted

Personally I'll leave this code until we really tested and checked all
uses cases, 

Right, I don't want to break anything on purpose, but I think the
current state of this driver is fragile and difficult to
understand/maintain, so I would be grateful for any help in
understanding the corner cases better, as well as testing them (or
explaining to me how to reproduce them.)

Generally, with long-term maintenance in mind, if I'm forced to choose
between understandable/maintainable code and "covers 100% of corner
cases" I will most often chose the former.

quoted

not only on the sei510/sei610 using mainline u-boot.

See above about all the other boards with vendor uboots also tested.

Let's leave apart the vendor uboot issue for g12.

Since display support for G12A will land soon in mainline U-boot, let me
push the DT patches for simple-framebuffer to we have a fallback in
case the DRM driver isn't loaded/built.

We can consider that in case of EFIfb, the simple-framebuffer node would have
been enabled and the power domain hooked and enabled.

Neil

Kevin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel