Re: [PATCH v2] PCI: Mark LSI FW643 to avoid bus reset

[Lukas Wunner <lukas@wunner.de>]

On Fri, Mar 29, 2024 at 05:12:19PM +0900, Takashi Sakamoto wrote:
> On Fri, Mar 29, 2024 at 05:41:16AM +0100, Lukas Wunner wrote:
> > Just checked the ACPI tables and there's an FPEN method below the
> > FRWR device which toggles GPIO 48 on the PCH.  Checked the schematics
> > as well and GPIO 48 is marked FW_PWR_EN.  The GPIO controls load
> > switches which cut power to the FW643 chip when nothing is connected.
> > 
> > Also, FW_PWR_EN feeds into an SLG4AP016V chip where it seems to
> > internally gate FW_CLKREQ_L.
> > 
> > I'm guessing the driver may need to call the FPEN ACPI method after
> > issuing a SBR to force the chip on (or perhaps first off, then on)
> > and thereby re-enable Clock Request.
> > 
> > It's a pity the ohci.c driver doesn't seem to support runtime PM.
> > That would allow cutting power to the chip when nothing is connected
> > and thus increase battery life.  The ACPI tables indicate that the
> > platform sends a notification when something is plugged in, so all
> > the necessary ingredients are there but we're not taking advantage
> > of them.
> 
> Yup. In both PCI drivers and unit drivers belonging to Linux FireWire
> subsystem, any type of runtime PM is not supported. If I integrate 1394
> OHCI driver, I should implement all of the above in any callback of
> runtime PM, or the part of the above is already supported by any driver
> in parent PCI layer?

The power management method Apple uses to cut power to the FireWire
controller, Thunderbolt controller and discrete GPU is nonstandard.
It's *implemented* in ACPI, but doesn't *conform* to ACPI:  There are
no Power Resources described in the ACPI tables, just custom methods.

This can be made to work on Linux by assigning a dev_pm_domain to the
Root Port above the FireWire controller.  The dev_pm_domain callbacks
cut power to the FireWire controller on ->runtime_suspend() and
reinstate it on ->runtime_resume().  The reason this needs to be done
at the Root Port level is that the PCI core assumes the FireWire
controller is powered on when it calls pci_pm_runtime_resume() for it.
Normally that function would reinstate power through ACPI via
pci_power_up(), but that doesn't work due to the nonstandard nature
of Apple's ACPI tables.

I've implemented this 8 years ago for Thunderbolt but unfortunately
got sidetracked and thus haven't been able to finish upstreaming it yet:

https://github.com/l1k/linux/commit/a53d44439d42

I'm not as familiar with ohci.c as I am (or was) with thunderbolt.ko.
If you could amend ohci.c to call pm_runtime_get() when something is
attached and call pm_runtime_put() when something is detached, I could
look into bringing up the ACPI stuff.  You could acquire one runtime PM
ref for each attached device or just acquire a single ref if *anything*
is connected at all.  Doesn't matter.  But acquiring one ref per attached
device might be simpler.

I would also need a function to perform a bus scan upon runtime resume
which looks for new devices and acquires refs as necessary.  Once that
infrastructure exists, adding the Apple-specific ACPI stuff wouldn't
be too hard for me to do as I could just adapt what I did for Thunderbolt.

Thanks,

Lukas