2019-05-08 19:16:52

by Salman Qazi

[permalink] [raw]
Subject: icache_is_aliasing and big.LITTLE

Hi,

What is the intention behind icache_is_aliasing on big.LITTLE systems
where some icaches are VIPT and others are PIPT? Is it meant to be
conservative in some sense or should it be made per-CPU?

Thanks for your help,

Salman


2019-05-09 08:51:15

by Catalin Marinas

[permalink] [raw]
Subject: Re: icache_is_aliasing and big.LITTLE

Hi,

On Wed, May 08, 2019 at 11:45:03AM -0700, Salman Qazi wrote:
> What is the intention behind icache_is_aliasing on big.LITTLE systems
> where some icaches are VIPT and others are PIPT? Is it meant to be
> conservative in some sense or should it be made per-CPU?

It needs to cover the worst case scenario across all CPUs, i.e. aliasing
VIPT if one of the CPUs has this. We can't make it per-CPU because a
thread performing cache maintenance might be migrated to another CPU
with different cache policy (e.g. sync_icache_aliases()).

--
Catalin

2019-05-09 18:15:57

by Salman Qazi

[permalink] [raw]
Subject: Re: icache_is_aliasing and big.LITTLE

Thank you.

On Thu, May 9, 2019 at 1:50 AM Catalin Marinas <[email protected]> wrote:
>
> Hi,
>
> On Wed, May 08, 2019 at 11:45:03AM -0700, Salman Qazi wrote:
> > What is the intention behind icache_is_aliasing on big.LITTLE systems
> > where some icaches are VIPT and others are PIPT? Is it meant to be
> > conservative in some sense or should it be made per-CPU?
>
> It needs to cover the worst case scenario across all CPUs, i.e. aliasing
> VIPT if one of the CPUs has this. We can't make it per-CPU because a
> thread performing cache maintenance might be migrated to another CPU
> with different cache policy (e.g. sync_icache_aliases()).
>
> --
> Catalin

2019-05-10 11:13:49

by Mark Rutland

[permalink] [raw]
Subject: Re: icache_is_aliasing and big.LITTLE

On Thu, May 09, 2019 at 09:50:04AM +0100, Catalin Marinas wrote:
> Hi,
>
> On Wed, May 08, 2019 at 11:45:03AM -0700, Salman Qazi wrote:
> > What is the intention behind icache_is_aliasing on big.LITTLE systems
> > where some icaches are VIPT and others are PIPT? Is it meant to be
> > conservative in some sense or should it be made per-CPU?
>
> It needs to cover the worst case scenario across all CPUs, i.e. aliasing
> VIPT if one of the CPUs has this. We can't make it per-CPU because a
> thread performing cache maintenance might be migrated to another CPU
> with different cache policy (e.g. sync_icache_aliases()).

It's slightly more subtle than that -- for broadcast maintenance the
policy of the CPU receiving the broadcast matters.

So even if all i-cache maintenance were performed on a thread pinned to
a CPU with PIPT caches, to correctly affect any VIPT i-caches in the
system it would be necessary to perform maintenance as-if the CPU
performing the maintenance had VIPT i-caches.

Thanks,
Mark.