Re: [PATCH 4/4] perf: arm_spe: Enable ACPI/Platform automatic module loading

[Jeremy Linton]

Hi,

On 4/16/19 8:50 AM, Will Deacon wrote:
> On Thu, Apr 04, 2019 at 12:24:38PM -0500, Jeremy Linton wrote:
> > On 4/4/19 12:04 PM, Will Deacon wrote:
> > > On Tue, Mar 26, 2019 at 05:39:38PM -0500, Jeremy Linton wrote:
> > > > Lets add the MODULE_TABLE and platform id_table entries so that
> > > > the SPE driver can attach to the ACPI platform device created by
> > > > the core pmu code.
> > > >
> > > > Signed-off-by: Jeremy Linton <jeremy.linton@xxxxxxx>
> > > > Reviewed-by: Sudeep Holla <sudeep.holla@xxxxxxx>
> > > > ---
> > > >    drivers/perf/arm_spe_pmu.c | 11 +++++++++--
> > > >    1 file changed, 9 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
> > > > index 7cb766dafe85..ffa2c76c08bb 100644
> > > > --- a/drivers/perf/arm_spe_pmu.c
> > > > +++ b/drivers/perf/arm_spe_pmu.c
> > > > @@ -1176,7 +1176,13 @@ static const struct of_device_id arm_spe_pmu_of_match[] = {
> > > >    };
> > > >    MODULE_DEVICE_TABLE(of, arm_spe_pmu_of_match);
> > > > -static int arm_spe_pmu_device_dt_probe(struct platform_device *pdev)
> > > > +static const struct platform_device_id arm_spe_match[] = {
> > > > +	{ "arm,spe-v1", 0},
> > >
> > > It would be nice if we could avoid duplicating this string from the ACPI
> > > parsing code.
> >
> > Ok sure, I just need to find a good common place for it.

There doesn't appear to be a good common place for this, so maybe 
arm_pmu.h, which can then be included in the spe driver is the right thing.


> > > > +	{ }
> > > > +};
> > > > +MODULE_DEVICE_TABLE(platform, arm_spe_match);
> > > > +
> > > > +static int arm_spe_pmu_device_probe(struct platform_device *pdev)
> > > >    {
> > > >    	int ret;
> > > >    	struct arm_spe_pmu *spe_pmu;
> > > > @@ -1236,11 +1242,12 @@ static int arm_spe_pmu_device_remove(struct platform_device *pdev)
> > > >    }
> > > >    static struct platform_driver arm_spe_pmu_driver = {
> > > > +	.id_table = arm_spe_match,
> > > >    	.driver	= {
> > > >    		.name		= DRVNAME,
> > > >    		.of_match_table	= of_match_ptr(arm_spe_pmu_of_match),
> > >
> > > Hmm, so some other drivers don't hook .id_table like you do, but instead
> > > hook .acpi_match_table in the driver structure. Is that not better?
> >
> > This isn't actually an ACPI device, (aka not defined in the namespace), so
> > its missing much of the ACPI functionality. I think that also means its
> > needs to be declared this way.
>
> Looking at platform_match(), I'd really like to avoid having both an
> .id_table and an .of_match_table field.


> acpi_of_match_device() will actually use the .of_match_table, but it relies
> on ACPI_COMPANION returning a valid acpi_device. If we don't have one of

Right, via the fwnode it can cause an acpi DSDT defined device with a 
_DSD "compatible" property to match an entry in the of_match_table 
compatible string. I don't think this is us...

> those, perhaps we can use the .id_table exclusively and drop the
> .of_match_table instead?

This definitely made me do my homework, the following is AFAIK:

Its possible to match on just a .id_table, but this requires matching 
the OF device name against the id_table name rather than against the OF 
compatible string (*). This doesn't seem like a good idea, despite 
platform_device_id entries being significantly smaller than the 
of_device_id ones. Plus, I think we end up with two duplicate tables 
because we still need the MODULE_TABLE(of,xxx) to assure that userspace 
can associate the modalias with the module.

OTOH, it seems possible to match on module name directly 
('arm_spe_pmu'), but this limits us to only a single device type for all 
ACPI device variations unless we put platform checks in the module 
itself (ick!). I suspect in the future if a spe.v2 were to come out this 
would be a problem unless a separate module were created. Then there is 
the fact this still needs a platform_device_id table, as the modalias 
will read "platform:arm_spe_pmu". Which will cause people to question 
why its not just assigned and matched against the .id_table.


*(interestingly trivia: There doesn't appear to be a single arm64 module 
which matches on a MODULE_TABLE OF name. They only match type or 
compatible. Out of the 3534 modules on my machine only three do any OF 
table type matching, ipmi_si and two drivers for freescale networking 
fsl_pq_mdio and gianfar_driver. In those cases, i'm not even sure its 
actually necessary.)