Re: [PATCH v2 4/5] apple-gmux: support MMIO gmux on T2 Macs

From: Orlando Chamberlain
Date: Sat Feb 18 2023 - 07:52:38 EST


On Sat, 18 Feb 2023 11:49:52 +0100
Hans de Goede <hdegoede@xxxxxxxxxx> wrote:

> Hi,
>
> On 2/17/23 13:02, Orlando Chamberlain wrote:
> > On Fri, 17 Feb 2023 11:05:31 +1100
> > Orlando Chamberlain <orlandoch.dev@xxxxxxxxx> wrote:
> >>>
> >>> Question are we not worried about MacBooks with an "APP000B"
> >>> ACPI device (with a value IORSOURCE_MEM entry) but which do not
> >>> actually have a gmux, because they are iGPU only ?
> >>
> >> It looks like iMac20,1, iMac20,2, and iMacPro1,1 have APP000B:
> >>
> >> apple_gmux: Failed to find gmux I/O resource
> >>
> >> iMac20,2: https://linux-hardware.org/?probe=ec2af584b3&log=dmesg
> >> iMac20,1: https://linux-hardware.org/?probe=fee7644b9c&log=dmesg
> >> iMacPro1,1: https://linux-hardware.org/?probe=6c26c9ff8c&log=dmesg
> >>
> >> But I'm not sure if they actually have it or not. I'll see if I can
> >> get people with those models to test if it's a real gmux. There
> >> does seem to be a pattern in that those three all have AMD GPU's.
> >
> > Kerem Karabay managed to find the acpi tables and macOS's ioreg
> > from and iMacPro1,1:
> >
> > https://github.com/khronokernel/DarwinDumped/blob/master/iMacPro/iMacPro1%2C1/Darwin%20Dumper/DarwinDumper_3.0.4_30.12_15.30.40_iMacPro1%2C1_Apple_X64_High%20Sierra_17C2120_apple/ACPI%20Tables/DSL/DSDT.dsl#L10423
> > https://github.com/khronokernel/DarwinDumped/blob/master/iMacPro/iMacPro1%2C1/Darwin%20Dumper/DarwinDumper_3.0.4_30.12_15.30.40_iMacPro1%2C1_Apple_X64_High%20Sierra_17C2120_apple/IORegistry/IOReg.txt#L5096
> >
> > The DSDT table has the same APP000B device as MacBooks with actual
> > gmux, while the ioreg has no mention of Apple's driver
> > AppleMuxControl2 being used for that device.
> >
> > I think that confirms Apple has not fixed the issue of putting
> > APP000B's where they don't need to.
> >
> > Solutions to this I can think of are:
> >
> > - Use DMI matching to ignore product_names "iMacPro1,1" "iMac20,1",
> > "iMac20,2"
> > - Maybe check if the MMIO region for gmux is filled with 0xff*
> >
> > *I don't know if this would work or not as I don't have a machine to
> > check with. On my machine everything surrounding the 16 bytes used
> > for gmux is 0xff:
> >
> > # hexdump -n48 -C -s 0xfe0b01f0 /dev/mem
> > fe0b01f0 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff
> > |................| fe0b0200 00 00 3e 4f 00 00 00 00 00 00 00 00
> > 00 00 14 00 |..>O............| fe0b0210 ff ff ff ff ff ff ff ff
> > ff ff ff ff ff ff ff ff |................|
> >
> > so maybe on the iMacPro and iMac's, this would all be 0xff.
>
> Yes checking for a regular ioread32 returning 0xffffffff sounds
> like it should work. Can you add a check for that in the next version
> please ? Note this means we still need to do an iomap + unmap as
> you pointed out in another email, but I see no way around that.

I'll check that GMUX_MMIO_COMMAND_SEND (16th byte) is not 0xff, as if
the gmux is present it will reset that to 0x00, unless a command isn't
finished yet, in which case it will be one of 0x1, 0x4, 0x41, or 0x44.

>
> Regards,
>
> Hans
>
>
>
>
> >
> >>
> >> I've looked at dmesg or at least lsmod on all the models with the
> >> T2 chip and there wasn't evidence of any other models having that
> >> error or having apple-gmux loaded on any models that shouldn't
> >> have a gmux, other than the three mentioned above. Of course I
> >> don't know if its possible for there to be firmware versions where
> >> this isn't the case.
> >>>
> >>> I have learned the hard way (through backlight control regressions
> >>> in 6.1) that at least some older model MacBooks with an IO
> >>> resource have an APP000B ACPI device without them actually having
> >>> a gmux, these get caught by the version check and then do not
> >>> pass the indexed check so that apple_gmux_detect() properly
> >>> returns false.
> >>>
> >>> Maybe make gmux_mmio_read32() a static inline inside
> >>> include/linux/apple-gmux.h and try to read the version here ?
> >>
> >> For that would we need to ioremap() and iounmap()?
> >>>
> >>> Has this been tested on iGPU only T2 Macs?
> >>
> >> I don't think so.
> >>
> >>>
> >>> Regards,
> >>>
> >>> Hans
> >>>
> >>>
> >>
> >
>