Re: [RFC PATCH 00/20] x86/intel_rdt: Start abstraction for a second arch

From: Yury Norov <ynorov@caviumnetworks.com>
To: James Morse <james.morse@arm.com>
Cc: "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"x86@kernel.org" <x86@kernel.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	Fenghua Yu <fenghua.yu@intel.com>,
	Tony Luck <tony.luck@intel.com>, Ingo Molnar <mingo@redhat.com>,
	H Peter Anvin <hpa@zytor.com>,
	Reinette Chatre <reinette.chatre@intel.com>,
	Vikas Shivappa <vikas.shivappa@linux.intel.com>,
	"Norov, Yuri" <Yuri.Norov@cavium.com>
Subject: Re: [RFC PATCH 00/20] x86/intel_rdt: Start abstraction for a second arch
Date: Tue, 27 Nov 2018 12:33:30 +0000	[thread overview]
Message-ID: <20181127123319.GA24166@yury-thinkpad> (raw)
In-Reply-To: <20180824104519.11203-1-james.morse@arm.com>

Hi James,

On Fri, Aug 24, 2018 at 11:44:59AM +0100, James Morse wrote:
> Hi folks,
> 
> ARM have some upcoming CPU features that are similar to Intel RDT. Resctrl
> is the defacto ABI for this sort of thing, but it lives under arch/x86.
> 
> To get existing software working, we need to make resctrl work with arm64.
> This series is the first chunk of that. The aim is to move the filesystem/ABI
> parts into /fs/resctrl, and implement a second arch backend.
> 
> 
> What are the ARM features?
> Future ARM SoCs may have a feature called MPAM: Memory Partitioning and
> Monitoring. This is an umbrella term like RDT, and covers a range of controls
> (like CAT) and monitors (like MBM, CMT).
> 
> This series is almost all about CDP. MPAM has equivalent functionality, but
> it doesn't need enabling, and doesn't affect the available closids. (I'll
> try and use Intel terms). MPAM expects the equivalent to IA32_PRQ_MSR to
> be configured with an Instruction closid and a Data closid. These are the
> same for no-CDP, and different otherwise. There is no need for them to be
> adjacent.
> 
> To avoid emulating CDP in arm64's arch code, this series moves all the ABI
> parts of the CDP behaviour, (half the closid-space, each having two
> configurations) into the filesystem parts of resctrl. These will eventually
> be moved to /fs/.
> 
> MPAMs control and monitor configuration is all memory mapped, the base
> addresses are discovered via firmware tables, so we won't have a table of
> possible resources that just need alloc_enabling.
> 
> Is this it? No... there are another two series of a similar size that
> abstract the MBM/CMT overflow threads and avoid 'fs' code accessing things
> that have moved into the 'hw' arch specific struct.
> 
> 
> I'm after feedback on the general approach taken here, bugs, as there are
> certainly subtleties I've missed, and any strong-opinions on what should be
> arch-specific, and what shouldn't.
> 
> This series is based on v4.18, and can be retrieved from:
> git://linux-arm.org/linux-jm.git -b mpam/resctrl_rework/rfc_1

Thank you a lot for this work on cache allocation.

We are very interested in enabling CAT on Cavium / Marvell devices.
Could you please share another two series you mentioned above?

Do you have a working ARM64 CAT driver? It will much help us in our
experimenting.

Thanks in advance,
Yury