[Intel-gfx] [PATCH v8 07/17] pwm: lpss: Always update state and set update bit

[Thierry Reding]

On Mon, Aug 31, 2020 at 01:26:46PM +0200, Hans de Goede wrote:
> Hi,
> 
> On 8/31/20 1:13 PM, Thierry Reding wrote:
> > On Sun, Aug 30, 2020 at 02:57:43PM +0200, Hans de Goede wrote:
> > > This commit removes a check where we would skip writing the ctrl register
> > > and then setting the update bit in case the ctrl register already contains
> > > the correct values.
> > > 
> > > In a perfect world skipping the update should be fine in these cases, but
> > > on Cherry Trail devices the AML code in the GFX0 devices' PS0 and PS3
> > > methods messes with the PWM controller.
> > > 
> > > The "ACPI / LPSS: Resume Cherry Trail PWM controller in no-irq phase" patch
> > > earlier in this series stops the GFX0._PS0 method from messing with the PWM
> > > controller and on the DSDT-s inspected sofar the _PS3 method only reads
> > > from the PWM controller (and turns it off before we get a change to do so):
> > > 
> > >      {
> > >          PWMB = PWMC /* \_SB_.PCI0.GFX0.PWMC */
> > >          PSAT |= 0x03
> > >          Local0 = PSAT /* \_SB_.PCI0.GFX0.PSAT */
> > >      }
> > > 
> > > The PWM controller getting turning off before we do this ourselves is
> > > a bit annoying but not really an issue.
> > > 
> > > The problem this patch fixes comes from a new variant of the GFX0._PS3 code
> > > messing with the PWM controller found on the Acer One 10 S1003 (1):
> > > 
> > >      {
> > >          PWMB = PWMC /* \_SB_.PCI0.GFX0.PWMC */
> > >          PWMT = PWMC /* \_SB_.PCI0.GFX0.PWMC */
> > >          PWMT &= 0xFF0000FF
> > >          PWMT |= 0xC0000000
> > >          PWMC = PWMT /* \_SB_.PCI0.GFX0.PWMT */
> > >          PWMT = PWMC /* \_SB_.PCI0.GFX0.PWMC */
> > >          Sleep (0x64)
> > >          PWMB &= 0x3FFFFFFF
> > >          PWMC = PWMB /* \_SB_.PCI0.GFX0.PWMB */
> > >          PSAT |= 0x03
> > >          Local0 = PSAT /* \_SB_.PCI0.GFX0.PSAT */
> > >      }
> > > 
> > > This "beautiful" piece of code clears the base-unit part of the ctrl-reg,
> > > which effectively disables the controller, and it sets the update flag
> > > to apply this change. Then after this it restores the original ctrl-reg
> > > value, so we do not see it has mucked with the controller.
> > > 
> > > *But* it does not set the update flag when restoring the original value.
> > > So the check to see if we can skip writing the ctrl register succeeds
> > > but since the update flag was not set, the old base-unit value of 0 is
> > > still in use and the PWM controller is effectively disabled.
> > > 
> > > IOW this PWM controller poking means that we cannot trust the base-unit /
> > > on-time-div value we read back from the PWM controller since it may not
> > > have been applied/committed. Thus we must always update the ctrl-register
> > > and set the update bit.
> > 
> > Doesn't this now make patch 6/17 obsolete?
> 
> No, there is no guarantee we will get any changes soon after resume,
> so we must restore the state properly on resume, before 5.17
> we were just blindly restoring the old ctrl reg state, but
> if either the freq-div or the duty-cycle changes, we should
> also set the update bit in that case to apply the new freq-div/
> duty-cycle.

Hm... I didn't realize the driver was already saving/restoring context
before this. And from a quick look through the subsystem it looks like
I've done a pretty poor job of enforcing the "no context save/restore
from PWM drivers" rule. There are some cases that have had this support
since before we realized that this is problematic, but I think at least
pwm-img is newer than that and should never have had that code either.

> This actually also helps with that case since patch 6/17 uses
> pwm_lpss_prepare and this makes pwm_lpss_prepare set the
> update but unconditionally.
> 
> Also on resume we most do the set the enable bit vs set
> the update bit in the right order, depending on the
> generation of the SoC in which the PWM controller is
> embedded. 6/17 fixes all this by resume, by treating
> resume as a special case of apply() taking all the
> steps apply does.

As I mentioned earlier this works only under the assumption that the
suspend/resume order is correct. And that's possibly true for LPSS. It
won't work in the general case, though, because a backlight could end up
suspending/resuming completely out of sync with the rest of the display
pipeline and that's not something that we want.

I would expect that on i915 you also do have a controlled call sequence
that LPSS is part of, so I would expect that some consumer would
eventually call pwm_apply_state() and then any new settings would get
applied. Yes, that may perhaps be not immediately at the point where the
LPSS resumes, but it should be exactly at the point where the consumer
wants to enable it and therefore the only point where you can expect it
to make sense to enable the PWM.

Anyway, if this really turns out to be the only way to make this work I
can't object to it. But if you do rely on this, perhaps just make a
mental note that this can lead to sequencing problems that you may
potentially run into at some point.

Thierry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/intel-gfx/attachments/20200831/55906fbc/attachment.sig>