Re: [PATCH RFC 2/2] soc: Add a basic ACPI generic driver

[Olof Johansson]

On Tue, Jan 28, 2020 at 10:22 AM John Garry <john.garry@xxxxxxxxxx> wrote:
>
> On 28/01/2020 17:51, Olof Johansson wrote:
> > Hi,
> >
> > On Tue, Jan 28, 2020 at 3:18 AM John Garry <john.garry@xxxxxxxxxx> wrote:
> >>
>
> Hi Olof,
>
> >> Add a generic driver for platforms which populate their ACPI PPTT
> >> processor package ID Type Structure according to suggestion in the ACPI
> >> spec - see ACPI 6.2, section 5.2.29.3 ID structure Type 2.
> >>
> >> The soc_id is from member LEVEL_2_ID.
> >>
> >> For this, we need to use a whitelist of platforms which are known to
> >> populate the structure as suggested.
> >>
> >> For now, only the vendor and soc_id fields are exposed.
> >>
> >> Signed-off-by: John Garry <john.garry@xxxxxxxxxx>
> >> ---
> >>   drivers/soc/Makefile       |   1 +
> >>   drivers/soc/acpi_generic.c | 102 +++++++++++++++++++++++++++++++++++++
> >>   2 files changed, 103 insertions(+)
> >>   create mode 100644 drivers/soc/acpi_generic.c
> >>
> >> diff --git a/drivers/soc/Makefile b/drivers/soc/Makefile
> >> index 8b49d782a1ab..2a59a30a22cd 100644
> >> --- a/drivers/soc/Makefile
> >> +++ b/drivers/soc/Makefile
> >> @@ -3,6 +3,7 @@
> >>   # Makefile for the Linux Kernel SOC specific device drivers.
> >>   #
> >>
> >> +obj-$(CONFIG_ACPI_PPTT)                += acpi_generic.o
> >>   obj-$(CONFIG_ARCH_ACTIONS)     += actions/
> >>   obj-$(CONFIG_SOC_ASPEED)       += aspeed/
> >>   obj-$(CONFIG_ARCH_AT91)                += atmel/
> >
> > Based on everything I've seen so far, this should go under drivers/acpi instead.
>
> soc drivers seem to live in drivers/soc (non-arm32, anyway), so I
> decided on this location. But drivers/acpi would also seem reasonable now.

We don't want drivers/soc to be too much of a catch-all -- it is meant
for some of the glue pieces that don't have good homes elsewhere.
Unfortunately, the slope is slippery and we've already gone down it a
bit, but I think we can fairly clearly declare that this kind of
cross-soc material is likely not the right home for it -- especially
when drivers/acpi is a good fit in this case.

> >> diff --git a/drivers/soc/acpi_generic.c b/drivers/soc/acpi_generic.c
> >> new file mode 100644
> >> index 000000000000..34a1f5f8e063
> >> --- /dev/null
> >> +++ b/drivers/soc/acpi_generic.c
> >> @@ -0,0 +1,102 @@
> >> +// SPDX-License-Identifier: GPL-2.0
> >> +/*
> >> + * Copyright (c) John Garry, john.garry@xxxxxxxxxx
> >> + */
> >> +
> >> +#define pr_fmt(fmt) "SOC ACPI GENERIC: " fmt
> >> +
> >> +#include <linux/acpi.h>
> >> +#include <linux/sys_soc.h>
> >> +
> >> +/*
> >> + * Known platforms that fill in PPTT package ID structures according to
> >> + * ACPI spec examples, that being:
> >> + * - Custom driver attribute is in ID Type Structure VENDOR_ID member
> >> + * - SoC id is in ID Type Structure LEVEL_2_ID member
> >> + *    See ACPI SPEC 6.2 Table 5-154 for PPTT ID Type Structure
> >> + */
> >> +static struct acpi_platform_list plat_list[] = {
> >> +       {"HISI  ", "HIP08   ", 0, ACPI_SIG_PPTT, all_versions},
> >> +       { } /* End */
> >> +};
> >
> > As others have said, this will become a mess over time, and will
> > require changes for every new platform. Which, unfortunately, is
> > exactly what ACPI is supposed to provide relief from by making
> > standardized platforms... standardized.
> >
>
> Right, and I think that it can be dropped. As discussed with Sudeep, I
> was concerned how this PPTT ID structure could be interpreted, and had a
> whitelist as a conservative approach.

[...]

> >
> > Hmm, this doesn't look like much of a driver to me. This looks like
> > the export of an attribute to userspace, and should probably be done
> > by ACPI core instead of creating an empty driver for it.
>
> OK, but I'm thinking that having a soc driver can be useful as it is
> common to DT, and so userspace only has to check a single location. And
> the soc driver can also cover multiple-chip systems without have to
> reinvent that code for ACPI core. And it saves adding a new ABI.

While having a single location could be convenient, the actual data
read/written would be different (I'm guessing).

We also already have a supposed standard way of figuring out what SoC
we're on (toplevel compatible for the DT). So no matter what, I think
userspace will need to handle two ways of probing this.


-Olof