lkml.org 
[lkml]   [2021]   [Oct]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH V2 1/2] riscv: Add RISC-V svpbmt extension
From
On Tue, 28 Sep 2021 07:58:51 PDT (-0700), alexandre.ghiti@canonical.com wrote:
> On Tue, Sep 28, 2021 at 3:48 PM Philipp Tomsich
> <philipp.tomsich@vrull.eu> wrote:
>>
>> Nick,
>>
>> On Tue, 28 Sept 2021 at 15:19, Nick Kossifidis <mick@ics.forth.gr> wrote:
>> >
>> > On 9/28/21 7:26 AM, Atish Patra wrote:
>> > > On Mon, Sep 27, 2021 at 8:50 PM Anup Patel <anup@brainfault.org> wrote:
>> > >>
>> > >> On Tue, Sep 28, 2021 at 6:32 AM Nick Kossifidis <mick@ics.forth.gr> wrote:
>> > >>>
>> > >>> Στις 2021-09-27 23:13, Atish Patra έγραψε:
>> > >>>>> We need to decide whether we should support the upstream kernel for
>> > >>>>> D1. Few things to consider.
>> > >>>>> – Can it be considered as an errata ?
>> > >>>
>> > >>> It's one thing to follow the spec and have an error in the
>> > >>> implementation, and another to not follow the spec.
>> > >>>
>> > >>>>> – Does it set a bad precedent and open can of worms in future ?
>> > >>>
>> > >>> IMHO yes, I'm thinking of Kendryte 210 devs for example coming up and
>> > >>> asking for MMU support, they 've also shipped many chips already. I can
>> > >>> also imagine other vendors in the future coming up with implementations
>> > >>> that violate the spec in which case handling the standard stuff will
>> > >>> become messy and complex, and hurt performance/security. We'll end up
>> > >>> filling the code with exceptions and tweaks all over the place. We need
>> > >>> to be strict about what is "riscv" and what's "draft riscv" or "riscv
>> > >>> inspired", and what we are willing to support upstream. I can understand
>> > >>> supporting vendor extensions upstream but they need to fit within the
>> > >>> standard spec, we can't have for example extensions that use encoding
>> > >>> space/csrs/fields etc reserved for standard use, they may only use
>> > >>> what's reserved for custom/vendor use. At least let's agree on that.
>> > >>
>> > >> Totally agree with Nick here. It's a slippery slope.
>> > >>
>> > >> Including D1 PTE bits (or Kendryte K210 MMU) part of the Linux RISC-V
>> > >> means future hardware which intentionally violates specs will also have to
>> > >> be merged and the RISC-V patch acceptance policy will have no significance.
>> > >>
>> > >>>
>> > >>>>> – Can we just ignore D1 given the mass volume ?
>> > >>>>>
>> > >>>
>> > >>> IMHO no, we need to find a way to support it upstream but I believe
>> > >>> there is another question to answer:
>> > >>>
>> > >>> Do we also guarantee "one image to rule them all" approach, required by
>> > >>> binary distros, for implementations that violate the spec ? Are we ok
>> > >>> for example to support Allwinner D1 upstream but require a custom
>> > >>> configuration/build instead of supporting it with the "generic" image ?
>> > >>> In one case we need to handle the violation at runtime and introduce
>> > >>> overhead for everyone (like looking up __riscv_svpbmt every time we set
>> > >>> a PTE in this case), in the other it's an #ifdef.
>> > >>
>> > >> At least, we should not have hardware violating specs as part of the
>> > >> unified kernel image instead have these intentional deviations/violations
>> > >> under separate kconfig which will not be enabled by default. This means
>> > >> vendors (of such hardware) and distros will have to explicitly enable
>> > >> support for such violations/deviations.
>> > >>
>> > >
>> > > If we merge the code and are not enabled by default, it would be a
>> > > maintenance nightmare in future.
>> > > These part of the kernel will not be regularly tested but we have to
>> > > carry the changes for a long time.
>> >
>> > I don't see a difference between having these features as part of the
>> > generic image vs having them as custom configs/builds. The code will get
>> > executed only on boards that support the custom/non-compliant
>> > implementation anyway. To the contrary we'll have more code to test if
>> > we are doing things at runtime vs at compile time.
>> >
>> > > Similar changes will only grow over time causing a lot of custom
>> > > configs that are not enabled by default.
>> > >
>> >
>> > We'll have a lot of custom configs that will only get used on boards
>> > that use them, vs runtime code that will run for no reason on every
>> > board and choose the default/standard-compliant implementation most of
>> > the time. In the end the code will only get tested on specific hardware
>> > anyway.
>> >
>> > > IMHO, if we want to support this board in upstream, we should just
>> > > clearly state that it is one time special exception
>> > > for this board only because of the following reasons
>> > >
>> > > 1. The board design predates the patch acceptance policy.
>> > > 2. We don't have enough affordable Linux compatible platforms today.
>> > > 3. Allowing running an upstream kernel on D1 helps the RISC-V software
>> > > ecosystem to grow.
>> > >
>> >
>> > The same can be said for Kendryte as well, are we willing to also
>> > support their MMU implementation on the generic image if a patch comes
>> > in? To be clear I'm not saying we shouldn't support D1 or Kendryte
>> > upstream, I'm just saying that we shouldn't sacrifice the complexity and
>> > performance of the code path for standard-compliant implementations, to
>> > support non-compliant implementations, and instead support non-compliant
>> > implementations with custom kernel builds using compile time options. It
>>
>> For priming the pump on the software effort, having a solution that is enabled
>> on distro-builds is clearly preferable — that leads to the solution that Palmer
>> had outlined at LPC, which is to have a KCONFIG option that enables the
>> alternate code paths and can be turned off for embedded use-cases.
>>
>> > still counts as upstream support, they won't have to maintain their own
>> > forks. It'll also allow custom implementations to have more flexibility
>> > on what they can do since they will be able to use completely
>> > different/custom code paths, instead of trying to fit in the standard
>> > code path (which will become a mess over time). I think this approach is
>> > much more flexible and will allow more customizations to be supported
>> > upstream in the future.
>>
>> The important detail will be the ground rules: changes have to be sufficiently
>> quarantined that (a) they can be turned off, (b) can be reverted easily (in case
>> that vendors fail to perform their maintenance obligations),
>
> Can we really remove support once it is in and widely used?

We'll follow the standard deprecation policies for anything I have any
say over, which in the kernel I've always heard described as
forever-ish. Since this is pretty coupled to a specific chip one could
imagine deprecating it when we can convince ourselves those chips have
all had their smoke let out, but that's a decade timescale sort of
thing.

>> and (c) they don't
>> affect the performance and complexity of the standard code paths.
>>
>> Cheers,
>> Philipp.
>>
>> _______________________________________________
>> linux-riscv mailing list
>> linux-riscv@lists.infradead.org
>> http://lists.infradead.org/mailman/listinfo/linux-riscv

\
 
 \ /
  Last update: 2021-10-05 02:44    [W:0.055 / U:0.072 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site