Re: [PATCH v2] drivers/perf: Add support for ARMv8.3-SPE

[Will Deacon <will@kernel.org>]

On Wed, Sep 30, 2020 at 05:31:35PM +0800, Wei Li wrote:
> Armv8.3 extends the SPE by adding:
> - Alignment field in the Events packet, and filtering on this event
>   using PMSEVFR_EL1.
> - Support for the Scalable Vector Extension (SVE).
> 
> The main additions for SVE are:
> - Recording the vector length for SVE operations in the Operation Type
>   packet. It is not possible to filter on vector length.
> - Incomplete predicate and empty predicate fields in the Events packet,
>   and filtering on these events using PMSEVFR_EL1.
> 
> Update the check of pmsevfr for empty/partial predicated SVE and
> alignment event in SPE driver. For adaption by the version of SPE,
> expose 'pmsver' as cap attribute to userspace.
> 
> Signed-off-by: Wei Li <liwei391@huawei.com>
> ---
> v1 -> v2:
>  - Rename 'pmuver' to 'pmsver', change it's type to 'u16' from 'int'.
>    (Suggested by Will and Leo.)
>  - Expose 'pmsver' as cap attribute through sysfs, instead of printing.
>    (Suggested by Will.)
> ---
>  arch/arm64/include/asm/sysreg.h |  4 +++-
>  drivers/perf/arm_spe_pmu.c      | 18 ++++++++++++++++--
>  2 files changed, 19 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
> index 554a7e8ecb07..f4f9c1fc6398 100644
> --- a/arch/arm64/include/asm/sysreg.h
> +++ b/arch/arm64/include/asm/sysreg.h
> @@ -281,7 +281,6 @@
>  #define SYS_PMSFCR_EL1_ST_SHIFT		18
>  
>  #define SYS_PMSEVFR_EL1			sys_reg(3, 0, 9, 9, 5)
> -#define SYS_PMSEVFR_EL1_RES0		0x0000ffff00ff0f55UL
>  
>  #define SYS_PMSLATFR_EL1		sys_reg(3, 0, 9, 9, 6)
>  #define SYS_PMSLATFR_EL1_MINLAT_SHIFT	0
> @@ -787,6 +786,9 @@
>  #define ID_AA64DFR0_PMUVER_8_5		0x6
>  #define ID_AA64DFR0_PMUVER_IMP_DEF	0xf
>  
> +#define ID_AA64DFR0_PMSVER_8_2		0x1
> +#define ID_AA64DFR0_PMSVER_8_3		0x2
> +
>  #define ID_DFR0_PERFMON_SHIFT		24
>  
>  #define ID_DFR0_PERFMON_8_1		0x4
> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
> index cc00915ad6d1..52e7869f5621 100644
> --- a/drivers/perf/arm_spe_pmu.c
> +++ b/drivers/perf/arm_spe_pmu.c
> @@ -54,7 +54,7 @@ struct arm_spe_pmu {
>  	struct hlist_node			hotplug_node;
>  
>  	int					irq; /* PPI */
> -
> +	u16					pmsver;
>  	u16					min_period;
>  	u16					counter_sz;
>  
> @@ -80,6 +80,15 @@ struct arm_spe_pmu {
>  /* Keep track of our dynamic hotplug state */
>  static enum cpuhp_state arm_spe_pmu_online;
>  
> +static u64 sys_pmsevfr_el1_mask[] = {
> +	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> +		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
> +		BIT_ULL(1),
> +	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> +		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
> +		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
> +};

Ok, so I finally figured out what I don't like about this: it's the fact
that the RES0 mask only ever reduces, but we have no way of ensuring that
by construction with this approach. In other words, it's a bit brittle to
keep all of these things defined entirely separately from one another.

How about a small change so that we define things like:

#define SYS_PMSEVFR_EL1_RES0_8_2	SYS_PMSEVFR_EL1_RES0 &
					~(...)

#define SYS_PMSEVFR_EL1_RES0_8_3	SYS_PMSEVFR_EL1_RES0_8_2 &
					~(...)

where the '...' parts identify the bits that are no longer RES0 for that
version of the architecture?

What do you think?

Will