Re: [PATCH] driver: perf: arm_pmuv3: Read PMMIR_EL1 unconditionally

[Mark Rutland]

On Thu, Oct 12, 2023 at 04:28:28PM +0530, Anshuman Khandual wrote:
> 
> 
> On 10/12/23 14:54, Mark Rutland wrote:
> > On Mon, Oct 09, 2023 at 09:59:19AM +0100, James Clark wrote:
> >> On 09/10/2023 08:56, Anshuman Khandual wrote:
> >>> PMMIR_EL1 needs to be captured in 'armpmu->reg_pmmir', for all appropriate
> >>> PMU version implementations where the register is available and reading it
> >>> is valid . Hence checking for bus slot event presence is redundant and can
> >>> be dropped.
> >>>
> >>> Cc: Will Deacon <will@xxxxxxxxxx>
> >>> Cc: Mark Rutland <mark.rutland@xxxxxxx>
> >>> Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
> >>> Cc: linux-kernel@xxxxxxxxxxxxxxx
> >>> Signed-off-by: Anshuman Khandual <anshuman.khandual@xxxxxxx>
> >>> ---
> >>> This applies on v6.6-rc5.
> >>>  
> >>>  drivers/perf/arm_pmuv3.c | 2 +-
> >>>  1 file changed, 1 insertion(+), 1 deletion(-)
> >>>
> >>> diff --git a/drivers/perf/arm_pmuv3.c b/drivers/perf/arm_pmuv3.c
> >>> index 1e72b486c033..9fc1b6da5106 100644
> >>> --- a/drivers/perf/arm_pmuv3.c
> >>> +++ b/drivers/perf/arm_pmuv3.c
> >>> @@ -1129,7 +1129,7 @@ static void __armv8pmu_probe_pmu(void *info)
> >>>  			     pmceid, ARMV8_PMUV3_MAX_COMMON_EVENTS);
> >>>  
> >>>  	/* store PMMIR register for sysfs */
> >>> -	if (is_pmuv3p4(pmuver) && (pmceid_raw[1] & BIT(31)))
> >>> +	if (is_pmuv3p4(pmuver))
> >>>  		cpu_pmu->reg_pmmir = read_pmmir();
> >>>  	else
> >>>  		cpu_pmu->reg_pmmir = 0;
> >>
> >>
> >> This does have the side effect of showing non-zero values in caps/slots
> >> even when the STALL_SLOT event isn't implemented. I think that's the
> >> scenario that the original commit (f5be3a61fd) was trying to avoid:
> >>
> >>   /sys/bus/event_source/devices/armv8_pmuv3_0/caps/slots is exposed
> >>   under sysfs. [If] Both ARMv8.4-PMU and STALL_SLOT event are
> >>   implemented, it returns the slots from PMMIR_EL1, otherwise it will
> >>   return 0.
> > 
> > We check for the STALL_SLOT event becuase (at the time) the ARM ARM said:
> > 
> > | If STALL_SLOT is not implemented, it is IMPLEMENTATION DEFINED whether the
> > | PMMIR System registers are implemented.
> > 
> > ... and this was necessary to avoid triggering an UNDEFINED exception if we
> > attempted to read PMMIR on a CPU which didn't actually implement it.
> > 
> > See: 
> > 
> >   https://lore.kernel.org/linux-arm-kernel/20200720101518.GA11516@willie-the-truck/
> >   https://lore.kernel.org/linux-arm-kernel/20200720105019.GA54220@C02TD0UTHF1T.local/
> >   https://lore.kernel.org/linux-arm-kernel/20200720105410.GD11516@willie-the-truck/
> > 
> > As I promised in that thread, I did raise that with our architects. According
> > to the bug I filed against the architecture, this was tightened such that
> > ARMv8.4-PMU gauaranteed the presence of PMMIR, and that should have changed
> > between the G.a and G.b releases of the ARM ARM.
> > 
> > Anshuman, can you go and check that the wording did chaange between G.a and G.b?
> 
> G.a was the last ARM ARM version to have that STALL_SLOT event dependency for
> PMMIR_EL1 system register which got dropped in G.b version ARM ARM.
> 
> > 
> > Assuming it did (and the wording in the latest J.a release is also fine),
> > please update the commit message to describe the history above.
> 
> Sure, will update the commit message as follows.
> 
>     driver: perf: arm_pmuv3: Read PMMIR_EL1 unconditionally
>     
>     PMMIR_EL1 needs to be captured in 'armpmu->reg_pmmir', for all appropriate
>     PMU version implementations i.e FEAT_PMUv3p4 onward, where the register is
>     available and reading it is valid. However STALL_SLOT event dependency was
>     included earlier as per previous version ARM ARM wordings [1]. But later
>     i.e ARM ARM version G.a onward, STALL_SLOT event dependency for PMMIR_EL1
>     register has been dropped. Checking for that dependency is now redundant,
>     and should be dropped.
>     
>     [1] https://lore.kernel.org/linux-arm-kernel/20200720105410.GD11516@willie-the-truck/

This is quite painful to read, and makes it the reader's responsibility to go
and find information that should really be in the commit message itself, e.g.
the exact woridng from the ARM ARM, which is earlier in the email context.

This should say something like:

    drivers/perf: arm_pmuv3: Read PMMIR_EL1 unconditionally

    Currently the PMUv3 driver only reads PMMIR_EL1 if the PMU implements
    FEAT_PMUv3p4 and the STALL_SLOT event, but the check for STALL_SLOT isn't
    necessary and can be removed.

    The check for STALL_SLOT was introduced with the read of PMMIR_EL1 in
    commit:

      f5be3a61fdb5dd11 ("arm64: perf: Add support caps under sysfs:)
   
    When this logic was written, the ARM ARM said:

    | If STALL_SLOT is not implemented, it is IMPLEMENTATION DEFINED whether
    | the PMMIR System registers are implemented.

    ... and thus the driver had to check for STALL_SLOT to verify that
    PMMIR_EL1 was implemented and accesses to PMMIR_EL1 would not be UNDEFINED.

    Subsequently, the architecture was retrospectively tightened to require
    that any FEAT_PMUv3p4 implementation implements PMMIR_EL1. Since the G.b
    release of the ARM ARM, the wording regarding STALL_SLOT has been removed,
    and the description of PMMIR_EL1 says:

    | This register is present only when FEAT_PMUv3p4 is implemented.

    Drop the unnecessary check for STALL_SLOT when reading PMMIR_EL1.

... which gives the reader *all* the information they need, and explains why
the change is safe.

Mark.