Re: [RFC PATCH 08/12] perf/x86: Add APX in extended regs

[Liang, Kan]

On 2025-06-13 12:02 p.m., Dave Hansen wrote:
> On 6/13/25 06:49, kan.liang@xxxxxxxxxxxxxxx wrote:
>> +#define __x86_pmu_get_regs(_mask, _regs, _size)		\
>> +do {							\
>> +	if (mask & _mask && xcomp_bv & _mask) {		\
>> +		_regs = xsave;				\
>> +		xsave += _size;				\
>> +	}						\
>> +} while (0)
> 
> Ewww.
> 
> First of all, this doesn't work generally because of the previously
> mentioned alignment. Second, it's using xcomp_bv which doesn't tell you
> if XSAVES wrote the data.
> 
> Last, this attempts to reimplement get_xsave_addr().
> 
> I'd do something like this:
> 
> 	for (xfeature_nr in mask) {
> 		void *src = get_xsave_addr(xsave, xfeature_nr);
> 		void *dst = ... a function to map XFEATURE_MASK_APX
> 				to perf_regs->apx_regs,
> 		int size  = xstate_sizes(xfeature_nr);
> 
> 		if (!src)
> 			continue;
> 
> 		memcpy(dst, src, size);

The data will eventually be copied to a buffer which shared with the
user-space tool. I probably have to avoid the memcpy and do it later
when outputting the sample.
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/events/core.c#n7389

> 	}
> 
> That should handle *all* of the nastiness. The alignment, the init
> optimization. *Please* use get_xsave_addr() or one of the other helpers.

Thanks, I will try to reuse the existing fpu functions as much as I can.

Thanks,
Kan