[PATCH V13 - RESEND 06/10] arm64/perf: Enable branch stack events via FEAT_BRBE

[Anshuman Khandual]

Hello Yang,

On 7/25/23 12:42, Yang Shen wrote:
>> +    if (!(branch_type & PERF_SAMPLE_BRANCH_NO_CYCLES))
>> +        brbcr |= BRBCR_EL1_CC;
> 
> Hi Anshuman,
> 
> Here is problem about enable CYCLES_COUNT. The SPEC defines that the CYCLES_COUNT is only
> 
> valid when the BRECR_EL1.CC & BRBCR_EL2.CC is true. And here the SPEC also defines that
> 
> when PSTATE.EL == EL2 and HCR_EL2.E2h == '1', 'MSR BRBCR_EL1, <Xt>' means writing to
> 
> BRBCR_EL2 actually. So 'armv8pmu_branch_enable' can only set the BRBCR_EL2.CC, while the
> 
> BRECR_EL1.CC is still 0. The CYCLES_COUNT will be always 0 in records.


Agreed, this is a valid problem i.e BRBCR_EL1.CC and BRBCR_EL2.CC both needs to be set
for valid cycle count information regardless if the kernel runs in EL1 or EL2. A simple
hack in the current code setting BRBCR_EL12.C, which in turn sets BRBCR_EL1.CC when the
kernel runs in EL2 solves the problem.

> 
> As a solution, maybe BRBCR_EL12 should be added for driver according to the registers definition.

Right, will add the definition for BRBCR_EL12 in arch/arm64/tools/sysreg

> 
> Or, do you have a more standard solution?

Right, there are some nuances involved here.

Kernel could boot
	
a. Directly into EL2 and stays in EL2 for good
b. Directly into EL2 but switches into EL1
c. Directly into EL1 without ever going into EL2

In all the above cases BRBCR_EL1.CC and BRBCR_EL2.CC needs to be set when cycle count
is requested in the perf event interface (event->attr.branch_sample_type) via clearing
PERF_SAMPLE_BRANCH_NO_CYCLES.


- For the case as in (c) where kernel boots into EL1 directly and hence cannot ever set
  EL2 register, BRBCR_EL2.CC would be a booting requirement - updated in booting.rst

- For the cases as in (a) and (b) kernel boots via EL2, hence there is an opportunity
  to set both BRBCR_EL1.CC (via accessed BRBCR_EL12.CC) and BRBCR_EL2.CC. Depending on
  where the kernel lands up eventually, either BRBCR_EL1.CC or BRBCR_EL2.CC will be the
  toggle switch to ON or OFF cycle counting in the driver via branch_type_to_brbcr().
  So a new macro __init_el2_brbe is required which will get called from init_el2_state
  setting both the register fields as explained earlier.

I am working on these changes, will post the code soon.