Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:1 as permitted sender) client-ip=2620:137:e000::3:1;
Message-ID: <3f383b96-87ae-4580-a23d-0196bbd8ec91@rivosinc.com>
Date:   Mon, 23 Oct 2023 09:24:30 +0200
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: [PATCH v2 05/19] riscv: add ISA extension parsing for vector
 crypto extensions
Content-Language: en-US
To:     Evan Green <evan@rivosinc.com>, Conor Dooley <conor@kernel.org>
Cc:     Jerry Shih <jerry.shih@sifive.com>,
        linux-riscv@lists.infradead.org, devicetree@vger.kernel.org,
        linux-kernel@vger.kernel.org, linux-doc@vger.kernel.org,
        Palmer Dabbelt <palmer@rivosinc.com>,
        Paul Walmsley <paul.walmsley@sifive.com>,
        Rob Herring <robh+dt@kernel.org>,
        Krzysztof Kozlowski <krzysztof.kozlowski+dt@linaro.org>,
        Albert Ou <aou@eecs.berkeley.edu>,
        Jonathan Corbet <corbet@lwn.net>,
        Andrew Jones <ajones@ventanamicro.com>,
        Samuel Ortiz <sameo@rivosinc.com>
References: <20231017131456.2053396-1-cleger@rivosinc.com>
 <20231017131456.2053396-6-cleger@rivosinc.com>
 <DA8B4610-D514-4733-B875-C247FFCCC7AA@sifive.com>
 <af785f0f-9de7-4548-9cdb-f392cde1cc2b@rivosinc.com>
 <CALs-HstEBt-ntCcETa9YwS6On3nGyoEc2p7R-gaBLG9+aFJL5w@mail.gmail.com>
 <7626f978-e9ea-4f8f-b814-aeac02bd3712@rivosinc.com>
 <20231019-nuclear-vista-ef3e0b9bef71@spud>
 <CALs-HssWZyhnbo=ErH+LEfMi4m04i39Cw3PGXP1WhCFfSP8G=A@mail.gmail.com>
From:   =?UTF-8?B?Q2zDqW1lbnQgTMOpZ2Vy?= <cleger@rivosinc.com>
In-Reply-To: <CALs-HssWZyhnbo=ErH+LEfMi4m04i39Cw3PGXP1WhCFfSP8G=A@mail.gmail.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Precedence: bulk


On 19/10/2023 18:19, Evan Green wrote:
> On Thu, Oct 19, 2023 at 8:33 AM Conor Dooley <conor@kernel.org> wrote:
>>
>> On Thu, Oct 19, 2023 at 11:35:59AM +0200, Clément Léger wrote:
>>>
>>>
>>> On 18/10/2023 19:26, Evan Green wrote:
>>>> On Wed, Oct 18, 2023 at 5:53 AM Clément Léger <cleger@rivosinc.com> wrote:
>>>>>
>>>>>
>>>>>
>>>>> On 18/10/2023 03:45, Jerry Shih wrote:
>>>>>> On Oct 17, 2023, at 21:14, Clément Léger <cleger@rivosinc.com> wrote:
>>>>>>> @@ -221,6 +261,22 @@ const struct riscv_isa_ext_data riscv_isa_ext[] = {
>>>>>>>      __RISCV_ISA_EXT_DATA(zkt, RISCV_ISA_EXT_ZKT),
>>>>>>>      __RISCV_ISA_EXT_DATA(zksed, RISCV_ISA_EXT_ZKSED),
>>>>>>>      __RISCV_ISA_EXT_DATA(zksh, RISCV_ISA_EXT_ZKSH),
>>>>>>> +    __RISCV_ISA_EXT_DATA(zvbb, RISCV_ISA_EXT_ZVBB),
>>>>>>> +    __RISCV_ISA_EXT_DATA(zvbc, RISCV_ISA_EXT_ZVBC),
>>>>>>> +    __RISCV_ISA_EXT_DATA(zvkb, RISCV_ISA_EXT_ZVKB),
>>>>>>
>>>>>> The `Zvkb` is the subset of `Zvbb`[1]. So, the `Zvkb` should be bundled with `Zvbb`.
>>>>>
>>>>> Hi Jerry,
>>>>>
>>>>> Thanks for catching this, I think some other extensions will fall in
>>>>> this category as well then (Zvknha/Zvknhb). I will verify that.
>>>>
>>>> The bundling mechanism works well when an extension is a pure lasso
>>>> around other extensions. We'd have to tweak that code if we wanted to
>>>> support cases like this, where the extension is a superset of others,
>>>> but also contains loose change not present anywhere else (and
>>>> therefore also needs to stand as a separate bit).
>>>
>>> For Zvbb and Zvknhb, I used the following code:
>>>
>>> static const unsigned int riscv_zvbb_bundled_exts[] = {
>>>       RISCV_ISA_EXT_ZVKB,
>>>       RISCV_ISA_EXT_ZVBB
>>> };
>>>
>>> static const unsigned int riscv_zvknhb_bundled_exts[] = {
>>>       RISCV_ISA_EXT_ZVKNHA,
>>>       RISCV_ISA_EXT_ZVKNHB
>>> };
>>>
>>> Which correctly results in both extension (superset + base set) being
>>> enabled when only one is set. Is there something that I'm missing ?
>>>
>>>>
>>>> IMO, decomposing "pure" bundles makes sense since otherwise usermode
>>>> would have to query multiple distinct bitmaps that meant the same
>>>> thing (eg check the Zk bit, or maybe check the Zkn/Zkr/Zkt bits, or
>>>> maybe check the Zbkb/Zbkc... bits, and they're all equivalent). But
>>>> when an extension is a superset that also contains loose change, there
>>>> really aren't two equivalent bitmasks, each bit adds something new.
>>>
>>> Agreed but if a system only report ZVBB for instance and the user wants
>>> ZVKB, then it is clear that ZVKB should be reported as well I guess. So
>>> in the end, it works much like "bundle" extension, just that the bundle
>>> is actually a "real" ISA extension by itself.
>>>
>>> Clément
>>>
>>>>
>>>> There's an argument to be made for still turning on the containing
>>>> extensions to cover for silly ISA strings (eg ISA strings that
>>>> advertise the superset but fail to advertise the containing
>>>> extensions). We can decide if we want to work that hard to cover
>>>> hypothetical broken ISA strings now, or wait until they show up.
>>>> Personally I would wait until something broken shows up. But others
>>>> may feel differently.
>>
>> I'm not really sure that those are "silly" ISA strings. People are going
>> to do it that way because it is much easier than spelling out 5 dozen
>> sub-components, and it is pretty inevitable that subsets will be
>> introduced in the future for extensions we currently have.
>>
>> IMO, it's perfectly valid to say you have the supersets and not spell
>> out all the subcomponents.
> 
> Hm, ok. If ISA strings are likely to be written that way, then I agree
> having the kernel flip on all the contained extensions is a good idea.
> We can tweak patch 2 to support the parsing of struct
> riscv_isa_ext_data with both .id and .bundle_size set (instead of only
> one or the other as it is now). Looking back at that patch, it looks
> quite doable. Alright!

Hey Evan,

do you have anything against using this code:

static const unsigned int riscv_zvbb_bundled_exts[] = {
	RISCV_ISA_EXT_ZVKB,
	RISCV_ISA_EXT_ZVBB
};

...

Then declaring zvbb like that:

__RISCV_ISA_EXT_BUNDLE(zvbb, riscv_zvbb_bundled_exts),

I agree that it is *not* a bundled extension but it actually already
works with Conor's code. Not sure that adding more code is needed to
handle that case.

Clément


> 
> -Evan