Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
References: <20200331093241.3728-1-tesheng@andestech.com> <CAK8P3a3LokurC0n9XiwtPQh9ZgQcswMKY4b+TEsQh1VgYDNeWA@mail.gmail.com>
 <20200408035118.GA1451@andestech.com>
In-Reply-To: <20200408035118.GA1451@andestech.com>
From:   Arnd Bergmann <arnd@arndb.de>
Date:   Wed, 8 Apr 2020 16:40:17 +0200
Message-ID: <CAK8P3a1JS3_2fWrhNTZx0eTWjJa-GTb4AscTPqydpSP5EB15Yw@mail.gmail.com>
Subject: Re: [PATCH 0/3] Highmem support for 32-bit RISC-V
To:     Alan Kao <alankao@andestech.com>
Cc:     Eric Lin <tesheng@andestech.com>, Gary Guo <gary@garyguo.net>,
        alex@ghiti.fr, David Abdurachmanov <david.abdurachmanov@gmail.com>,
        Anup Patel <Anup.Patel@wdc.com>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        Steven Price <steven.price@arm.com>, atish.patra@wdc.com,
        yash.shah@sifive.com, Albert Ou <aou@eecs.berkeley.edu>,
        Palmer Dabbelt <palmer@dabbelt.com>,
        Greentime Hu <green.hu@gmail.com>, zong.li@sifive.com,
        Paul Walmsley <paul.walmsley@sifive.com>,
        Mike Rapoport <rppt@linux.ibm.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Borislav Petkov <bp@suse.de>,
        Logan Gunthorpe <logang@deltatee.com>,
        linux-riscv@lists.infradead.org
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Wed, Apr 8, 2020 at 5:52 AM Alan Kao <alankao@andestech.com> wrote:
> On Thu, Apr 02, 2020 at 11:31:37AM +0200, Arnd Bergmann wrote:
> > On Tue, Mar 31, 2020 at 11:34 AM Eric Lin <tesheng@andestech.com> wrote:
> > For the arm32 architecture, we are thinking about implementing a
> > VMPLIT_4G_4G option to replace highmem in the long run. The most
> > likely way this would turn out at the moment looks like:
> >
>
> Thanks for sharing the status from ARM32.  Is there any available branch
> already?  It would be good to have a reference implementation.

No code yet, so far not much more than the ideas that I listed. We
are currently looking for someone interested in doing the work
or maybe sponsoring it if they have a strong interest.

If someone does it for RISC-V first, that would of course also help on ARM ;-)

> > - have a 256MB region for vmalloc space at the top of the 4GB address
> >   space, containing vmlinux, module, mmio mappings and vmalloc
> >   allocations
> >
> > - have 3.75GB starting at address zero for either user space or the
> >   linear map.
> >
> > - reserve one address space ID for kernel mappings to avoid tlb flushes
> >   during normal context switches
> >
> > - On any kernel entry, switch the page table to the one with the linear
> >   mapping, and back to the user page table before returning to user space
> >
>
> After some survey I found previous disccusion
> (https://lkml.org/lkml/2019/4/24/2110). The 5.2-based patch ended up not
> being merged.  But at least we will have something to start if we want to.

Ah, I see. What is the current requirement for ASIDs in hardware
implementations? If support for more than one address space is
optional, that would make the VMSPLIT_4G support fairly expensive
as it requires a full TLB flush for each context switch.

> Also interestingly, there was a PR for privileged spec that separates
> addressing modes (https://github.com/riscv/riscv-isa-manual/pull/128) as
> Sdas extension, but there was no progress afterwards.

Right, this sounds like the ideal implementation. This is what is done
in arch/s390 and probably a few of the others.

> Not very related to this thread, but there were some discussion about
> ASID design in RISC-V (https://github.com/riscv/riscv-isa-manual/issues/348).
> It is now in ratified 1.11 privileged spec.

Ok, so I suppose that would apply to about half the 32-bit implementations
and most of the future ones, but not the ones implementing the 1.10 spec
or earlier, right?

> It seems to me that VMSPLIT_4G_4G is quite different from other VMSPLITs,
> because it requires much more changes.
>
> Thanks for showing the stance of kernel community against HIGHMEM support.
> The cited discussion thread is comprehensive and clear.  Despite that RV32
> users cannot get upstream support for their large memory, mechnisms like
> VMSPLIT_4G_4G seems to be a promising way to go.  That being said, to
> support the theoretical 16G physical memory, eventually kmap* will still
> be needed.

I had not realized that Sv32 supports more than 4GB physical address
space at all. I agree that if someone puts that much RAM into a machine,
there are few alternatives to highmem (in theory one could use the
extra RAM for zswap/zram, but that's not a good replacement).

OTOH actually using more than 1GB or 2GB of physical memory on a
32-bit core is something that I expect to become completely obscure
in the future, as this is where using 32-bit cores tends to get
uneconomical. The situation that I observe across the currently supported
32-bit architectures in the kernel is that:

- There is an incentive to run 32-bit on machines with 1GB of RAM  or less
  if you have the choice, because of higher memory  consumption and
  cache utilization on 64-bit code. On systems  with 2GB or more, the
  cost of managing that memory using 32-bit  code usually outweighs
  the benefits and you should run at least a 64-bit kernel.

- The high end 32-bit cores (Arm Cortex-A15/A17, MIPS P5600,
  PowerPC 750, Intel Pentium 4, Andes A15/D15, ...) are all obsolete
  after the follow-on products use 64-bit cores on a smaller process
  node, which end up being more capable, faster *and* cheaper.

- The 32-bit cores that do survive are based on simpler in-order
  pipelines that are cheaper and can still beat the 64-bit cores in
  terms of cost (mostly chip area, sometimes royalties), but not
  performance. This includes Arm Cortex-A7, MIPS 24k and typical
  RV32 cores.

- On an SoC with a cheap and simple CPU core, there is no point
  in spending a lot of money/area/complexity on a high-end memory
  controller. On single-core 32-bit SoCs, you usually end up with single
  16 or 32-bit wide DDR2 memory controller, on an SMP system like
  most quad-Cortex-A7, you have a 32-bit wide DDR3 controller, but no
  DDR4 or LP-DDR3/4.

- The largest economical memory configuration on a 32-bit DDR3
  controller is to have two 256Mx16 chips for a total of 1GB. You can
  get 2GB with four chips using dual-channel controllers or 512Mx8
  memory, but anything beyond that is much more expensive than
  upgrading to a 64-bit SoC with LP-DDR4.

This is unlikely to change over time as 64-bit chips are also getting
cheaper and may replace more of the 32-bit chips we see today.
In particular, I expect to see multi-core chips moving to mostly
64-bit cores over time, while 32-bit chips keep using one or
occasionally two cores, further reducing the need for large and/or
fast memory.

        Arnd