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

[Will Deacon]

On Thu, Nov 19, 2020 at 10:57:47PM +0800, liwei (GF) wrote:
> On 2020/10/2 18:57, Will Deacon wrote:
> > On Wed, Sep 30, 2020 at 05:31:35PM +0800, Wei Li wrote:
> >> @@ -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?
> > 
> 
> Sorry for the long delay.
> 
> These is also an array-index-out-of-bounds issue when accessing 'sys_pmsevfr_el1_mask', if
> the pmsver read in the future is bigger than what the driver supports.
> 
> So how about change to:
> 
> static u64 arm_spe_pmsevfr_mask(u16 pmsver)
> {
> 	u64 mask = 0;
> 
> 	if (pmsver >= ID_AA64DFR0_PMSVER_8_3)
> 		mask |= BIT_ULL(18) | BIT_ULL(17) | BIT_ULL(11);
> 	if (pmsver >= ID_AA64DFR0_PMSVER_8_2)
> 		mask |= GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> 			GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
> 			BIT_ULL(1);
> 	return mask;
> }
> 
> Thus, the driver can try its best to support, and the definition is clear enough to show
> the difference between versions of SPE.
> 
> Or should i still define them as what you advised and add a check of pmsver to just serve
> the versions what the driver support?

I think I'd prefer that, yes.

Will