Re: [PATCH v2 8/8] arch_topology: Add support to build llc_sibling on DT platforms

From: Sudeep Holla <sudeep.holla@arm.com>
To: Rob Herring <robh@kernel.org>
Cc: Atish Patra <atishp@atishpatra.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	Atish Patra <atishp@rivosinc.com>,
	Sudeep Holla <sudeep.holla@arm.com>,
	Vincent Guittot <vincent.guittot@linaro.org>,
	Morten Rasmussen <morten.rasmussen@arm.com>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Qing Wang <wangqing@vivo.com>,
	linux-arm-kernel <linux-arm-kernel@lists.infradead.org>,
	linux-riscv <linux-riscv@lists.infradead.org>
Subject: Re: [PATCH v2 8/8] arch_topology: Add support to build llc_sibling on DT platforms
Date: Fri, 20 May 2022 16:06:52 +0100	[thread overview]
Message-ID: <20220520150652.n6cfl7qpzbjrjh2f@bogus> (raw)
In-Reply-To: <20220520143620.GA3788938-robh@kernel.org>

On Fri, May 20, 2022 at 09:36:20AM -0500, Rob Herring wrote:
> On Fri, May 20, 2022 at 01:59:59PM +0100, Sudeep Holla wrote:
> > On Thu, May 19, 2022 at 01:10:51PM -0500, Rob Herring wrote:
> > > On Wed, May 18, 2022 at 4:34 AM Sudeep Holla <sudeep.holla@arm.com> wrote:
> > > >
> > > > ACPI PPTT provides cache identifiers and especially the last level cache
> > > > identifier is used in obtaining last level cache siblings amongst CPUs.
> > > >
> > > > While we have the cpu map representing all the CPUs sharing last level
> > > > cache in the cacheinfo driver, it is populated quite late in the boot
> > > > while the information is needed to build scheduler domains quite early.
> > >
> > > Late is because it's a device_initcall() rather than late in the cpu
> > > hotplug state machine, right?
> > 
> > Right. The expectation is to run in on each online CPU in CPU hotplug state
> > machine for some architectures. We may not need that on arm64 especially
> > since we get all info from DT or ACPI, but e.g. x86 uses cpuid which needs
> > to be executed on that CPU.
> 
> That's a separate issue. I'm not suggesting changing that part (that 
> would just be an optimization).
> 
> > > The late aspect is for sysfs presumably,but I think we could decouple that.
> > 
> > OK, not sure when this sched_domain info is actually needed. It think it
> > could be decoupled if we can wait until all the cpus are online.
> 
> No need to wait for all cpus to be online. I think you keep doing 
> it as part of cpu hotplug. The device_initcall() is used because you 
> cannot have struct device or sysfs calls before the driver core is 
> initialized. If we run the cacheinfo code earlier (I think the arch code 
> will have to call it) just like the topology code and skip the sysfs 
> parts, then you can use it.
>

Yes I was thinking something on similar lines, though I didn't think of
pushing code to arch. Let me check, must be possible.

> > > Do all the firmware cache parsing early and then populate the sysfs parts
> > > later.
> > 
> > Yes that may work on DT/ACPI based systems, as I said x86 relies on cpuid.
> 
> I'd assume using cpuid works at any time?
>

I think so, I can't recall the details since I looked at that about 5 years
ago. I have to check again anyways to explore early initialisation.

-- 
Regards,
Sudeep