Re: [PATCH v1 2/2] perf auxtrace: Optimize barriers with load-acquire and store-release

From: Adrian Hunter
Date: Tue Jun 01 2021 - 05:45:24 EST


On 1/06/21 12:17 pm, Peter Zijlstra wrote:
> On Tue, Jun 01, 2021 at 12:07:31PM +0300, Adrian Hunter wrote:
>> __sync_xxx_compare_and_swap is out-of-date now. This page:
>>
>> https://gcc.gnu.org/onlinedocs/gcc/_005f_005fsync-Builtins.html#g_t_005f_005fsync-Builtins
>>
>> recommends '__atomic' builtins instead.
>
> perf doesn't seem to use that.

I guess we could drop support for the compat case; add validation:

"Error, 32-bit perf cannot record AUX area traces from a 64-bit kernel.
Please use a 64-bit version of perf instead."

>
>> Since atomics are needed only for the "compat" case (i.e. 32-bit perf with 64-bit kernel)
>> you could try to find an elegant way to check for a 64-bit kernel, and avoid the atomics
>> for a 32-bit perf with 32-bit kernel.
>
> Most 32bit archs cannot do 64bit atomics. I suppose the only reason this
> doesn't explode is because the aux stuff isn't supported on many
> architectures?
>

Yes but presumably the race itself is unlikely since the upper byte changes only once every 4GiB.