Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Fri, 11 Feb 2022 00:01:42 +0800
From:   Jisheng Zhang <jszhang@kernel.org>
To:     Heiko =?utf-8?Q?St=C3=BCbner?= <heiko@sntech.de>
Cc:     palmer@dabbelt.com, paul.walmsley@sifive.com,
        aou@eecs.berkeley.edu, linux-riscv@lists.infradead.org,
        devicetree@vger.kernel.org, linux-kernel@vger.kernel.org,
        robh+dt@kernel.org, wefu@redhat.com, liush@allwinnertech.com,
        guoren@kernel.org, atishp@atishpatra.org, anup@brainfault.org,
        drew@beagleboard.org, hch@lst.de, arnd@arndb.de, wens@csie.org,
        maxime@cerno.tech, gfavor@ventanamicro.com,
        andrea.mondelli@huawei.com, behrensj@mit.edu, xinhaoqu@huawei.com,
        huffman@cadence.com, mick@ics.forth.gr,
        allen.baum@esperantotech.com, jscheid@ventanamicro.com,
        rtrauben@gmail.com, samuel@sholland.org, cmuellner@linux.com,
        philipp.tomsich@vrull.eu
Subject: Re: [PATCH v6 00/14] riscv: support for Svpbmt and D1 memory types
Message-ID: <YgU25sWTf8EXTina@xhacker>
References: <20220209123800.269774-1-heiko@sntech.de>
 <YgP+n5OMhQPSbICV@xhacker>
 <14426959.46CLvVMboC@diego>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <14426959.46CLvVMboC@diego>
Precedence: bulk

On Thu, Feb 10, 2022 at 12:44:04AM +0100, Heiko Stübner wrote:
> Hi,
> 
> Am Mittwoch, 9. Februar 2022, 18:49:19 CET schrieb Jisheng Zhang:
> > On Wed, Feb 09, 2022 at 01:37:46PM +0100, Heiko Stuebner wrote:
> > > Svpbmt is an extension defining "Supervisor-mode: page-based memory types"
> > > for things like non-cacheable pages or I/O memory pages.
> > > 
> > > 
> > > So this is my 2nd try at implementing Svpbmt (and the diverging D1 memory
> > > types) using the alternatives framework.
> > > 
> > > This includes a number of changes to the alternatives mechanism itself.
> > > The biggest one being the move to a more central location, as I expect
> > > in the future, nearly every chip needing some sort of patching, be it
> > > either for erratas or for optional features (svpbmt or others).
> > > 
> > > The dt-binding for svpbmt itself is of course not finished and is still
> > > using the binding introduced in previous versions, as where to put
> > > a svpbmt-property in the devicetree is still under dicussion.
> > > Atish seems to be working on a framework for extensions [0],
> > > 
> > > The series also introduces support for the memory types of the D1
> > > which are implemented differently to svpbmt. But when patching anyway
> > > it's pretty clean to add the D1 variant via ALTERNATIVE_2 to the same
> > > location.
> > > 
> > > The only slightly bigger difference is that the "normal" type is not 0
> > > as with svpbmt, so kernel patches for this PMA type need to be applied
> > > even before the MMU is brought up, so the series introduces a separate
> > > stage for that.
> > > 
> > > 
> > > In theory this series is 3 parts:
> > > - sbi cache-flush / null-ptr
> > > - alternatives improvements
> > > - svpbmt+d1
> > > 
> > > So expecially patches from the first 2 areas could be applied when
> > > deemed ready, I just thought to keep it together to show-case where
> > > the end-goal is and not requiring jumping between different series.
> > > 
> > > 
> > > The sbi cache-flush patch is based on Atish's sparse-hartid patch [1],
> > > as it touches a similar area in mm/cacheflush.c
> > > 
> > > 
> > > I picked the recipient list from the previous version, hopefully
> > > I didn't forget anybody.
> > > 
> > > changes in v6:
> > > - rebase onto 5.17-rc1
> > > - handle sbi null-ptr differently
> > > - improve commit messages
> > > - use riscv,mmu as property name
> > > 
> > > changes in v5:
> > > - move to use alternatives for runtime-patching
> > 
> > another choice is using static key mechanism. Pros: no need to coding
> > in asm, all in c.
> > 
> > To support new arch features, I see other arch sometimes use static
> > key, sometimes use alternative mechanism, so one question here would
> > be which mechanism is better? Any guide?
> 
> For me it's also a bit of a learn-as-you-go experience, but I do see some

I hope old hands can give some suggestions here about static key VS.
alternative ;). When to use which mechanism, and why.

> advantages in using alternatives:
> 
> - Static keys need the jump-label infrastructure, which the RiscV kernel
>   only seems to provide on non-XIP kernels [0]

I think you found one bug here.
I believe alternative mechanism also doesn't work for XIP kernel. I will
submit a patch for this case.

> - the amount of asm here is somewhat minimal for the core no-cache and io
>   types (load immediate + shift)
> - using the static key mechanism still does incur more overhead for the
>   conditional

do you mean the icache overhead due to the other disabled branch of
static key? It deserves a check.

> - and if we want to support the strange family-members like the D1,
>   (and it seems we do want that) this would create more conditionals

Maybe implement the standard svpbmt via. static key and cope with D1 as
errata alternative.
>   as we have to test for svpbmt, d1 and maybe future special cases,

From Documentation/riscv/patch-acceptance.rst, the "Submit Checklist
Addendum" section, "we'll only to accept patches for extensions that
have been officially frozen or ratified by the RISC-V Foundation."
This rule hasn't been changed.
Per my understanding of history of the svpbmt patch set, no future
special cases any more.

>   where alternatives-patching on the other hand simply replaces the
>   relevant code with the appropriate variant.
>