Re: [PATCH v3 05/13] riscv: cpufeature: extend riscv_cpufeature_patch_func to all ISA extensions

[Conor Dooley]

Hey!

I guess here is the right place to follow up on all of this stuff...

On Sat, Jan 14, 2023 at 08:32:37PM +0000, Conor Dooley wrote:
> On Fri, Jan 13, 2023 at 03:18:59PM +0000, Conor Dooley wrote:
> > On Thu, Jan 12, 2023 at 10:21:36AM +0100, Andrew Jones wrote:
> > > On Thu, Jan 12, 2023 at 12:29:57AM +0100, Heiko Stübner wrote:
> > > > Am Mittwoch, 11. Januar 2023, 18:10:19 CET schrieb Jisheng Zhang:
> > > > > riscv_cpufeature_patch_func() currently only scans a limited set of
> > > > > cpufeatures, explicitly defined with macros. Extend it to probe for all
> > > > > ISA extensions.
> > > > > 
> > > > > Signed-off-by: Jisheng Zhang <jszhang@xxxxxxxxxx>
> > > > > Reviewed-by: Andrew Jones <ajones@xxxxxxxxxxxxxxxx>
> > > > > Reviewed-by: Heiko Stuebner <heiko@xxxxxxxxx>
> > > > > ---
> > > > >  arch/riscv/include/asm/errata_list.h |  9 ++--
> > > > >  arch/riscv/kernel/cpufeature.c       | 63 ++++------------------------
> > > > >  2 files changed, 11 insertions(+), 61 deletions(-)
> > > > 
> > > > hmmm ... I do see a somewhat big caveat for this.
> > > > and would like to take back my Reviewed-by for now
> > > > 
> > > > 
> > > > With this change we would limit the patchable cpufeatures to actual
> > > > riscv extensions. But cpufeatures can also be soft features like
> > > > how performant the core handles unaligned accesses.
> > > 
> > > I agree that this needs to be addressed and Jisheng also raised this
> > > yesterday here [*]. It seems we need the concept of cpufeatures, which
> > > may be extensions or non-extensions.
> > > 
> > > [*] https://lore.kernel.org/all/Y77xyNPNqnFQUqAx@xhacker/
> > > 
> > > > See Palmer's series [0].
> > > > 
> > > > 
> > > > Also this essentially codifies that each ALTERNATIVE can only ever
> > > > be attached to exactly one extension.
> > > > 
> > > > But contrary to vendor-errata, it is very likely that we will need
> > > > combinations of different extensions for some alternatives in the future.
> > > 
> > > One possible approach may be to combine extensions/non-extensions at boot
> > > time into pseudo-cpufeatures. Then, alternatives can continue attaching to
> > > a single "feature". (I'm not saying that's a better approach than the
> > > bitmap, I'm just suggesting it as something else to consider.)
> > 
> > 
> > > >         ALTERNATIVE_2("nop",
> > > >                       "j strcmp_zbb_unaligned", 0, CPUFEATURE_ZBB | CPUFEATURE_FAST_UNALIGNED, 0, CONFIG_RISCV_ISA_ZBB,
> > > >                       "j variant_zbb", 0, CPUFEATURE_ZBB, CPUFEATURE_FAST_UNALIGNED, CONFIG_RISCV_ISA_ZBB)
> > > > 
> > > > [the additional field there models a "not" component]
> > 
> > Since we're discussing theoretical implementations, and it's a little hard
> > to picture all that they entail in my head, I might be making a fool of
> > myself here with assumptions...
> > 
> > Heiko's suggestion sounded along the lines of: keep probing individual
> > "features" as we are now. Features in this case being the presence of
> > the extension or non-extension capability. And then in the alternative,
> > make all of the decisions about which to apply.
> > 
> > Drew's suggestion would have significantly more defined CPUFEATURE_FOOs,
> > but would offload the decision making about which extensions or non-
> > extension capabilities constitute a feature to regular old cpufeature
> > code. However, the order of precedence would remain in the alt macro, as
> > it does now.
> > 
> > I think I am just a wee bit biased, but adding the complexity somewhere
> > other than alternative macros seems a wise choice, especially as we are
> > likely to find that complexity increases over time?
> > 
> > The other thing that came to mind, and maybe this is just looking for
> > holes where they don't exist (or are not worth addressing), is that
> > order of precedence.
> > I can imagine that, in some cases, the order of precedence is not a
> > constant per psuedo-cpufeature, but determined by implementation of
> > the capabilities that comprise it?
> 
> Having spent longer than I maybe should've looking at your patches
> Heiko, given it's a Saturday evening, the precedence stuff is still
> sticking out to me..
> 
> For Zbb & fast unaligned, the order may be non-controversial, but in
> the general case I don't see how it can be true that the order of
> precedence for variants is a constant.
> 
> Creating pseudo cpufeatures as Drew suggested does seem like a way to
> extract complexity from the alternatives themselves (which I think is a
> good thing) but at the expense of eating up cpu_req_feature bits...
> By itself, it doesn't help with precedence, but it may better allow us
> to deal with some of the precedence in cpufeature.c, since the
> alternative would operate based on the pseudo cpufeature rather than on
> the individual capabilities that the pseudo cpufeature depends on.
> 
> I tried to come up with a suggestion for what to do about precedence,
> but everything I thought up felt a bit horrific tbh.
> The thing that fits the current model best is just allowing cpu vendors
> to add, yet more, "errata" that pick the variant that works best for
> their implementation... Although I'd be worried about ballooning some of
> these ALT_FOO macros out to a massive degree with that sort of approach.
> 
> > If my assumption/understanding holds, moving decision making out of the
> > alternative seems like it would better provision for scenarios like
> > that? I dunno, maybe that is whatever the corollary of "premature
> > optimisation" is for this discussion.
> > 
> > That's my unsolicited € 0.02, hopefully I wasn't off-base with the
> > assumptions I made.
> 
> The order in which an alternative is added to the macro does matter,
> right? At least, that's how I thought it worked and hope I've not had
> an incorrect interpretation there all along... I wasn't until I started
> reading your patch and couldn't understand why you had a construct that
> looked like
> 
> if (zbb && !fast_unaligned)
> ...
> else if (zbb && fast_unaligned)
> ...
> 
> rather than just inverting the order and dropping the !fast_unaligned
> that I realised I might have a gap in my understanding after all..
> 
> > Heiko, I figure you've got some sort of WIP stuff for this anyway since
> > you're interested in the fast unaligned? How close are you to posting any
> > of that?
> > 
> > While I think of it, w.r.t. extension versus (pseudo)cpufeature etc
> > naming, it may make sense to call the functions that this series adds
> > in patch 6 has_cpufeature_{un,}likely(), no matter what decision gets
> > made here?
> > IMO using cpufeature seems to make more sense for a general use API that
> > may be used later on for the likes of unaligned access, even if
> > initially it is not used for anything other than extensions. 

Today at [1] we talked a bit about the various bits going on here.
I'll attempt to summarise what I remember, but I meant to do this
several hours ago and am likely to make a hames of it.

For Zbb/unaligned stuff, the sentiment was along the lines of there
needing to be a performance benefit to justify the inclusion.
Some of us have HW that is (allegedly) capable of Zbb, and, if that's the
case, will give it a go.
I think it was similar for unaligned, since there is nothing yet that
supports this behaviour, we should wait until a benefit is demonstrable.

On the subject of grouping extension/non-extension capabilities into a
single cpufeature, Palmer mentioned that GCC does something similar,
for the likes of the Ventana vendor extensions, that are unlikely to be
present in isolation.
Those are (or were?) probed as a group of extensions rather than
individually.
I think it was said it'd make sense for us to unify extensions that will
only ever appear together single psuedo cpufeature too.

For the bitfield approach versus creating pseudo cpufeatures discussion
& how to deal with that in alternatives etc, I'm a bit less sure what the
outcome was.
IIRC, nothing concrete was said about either approach, but maybe it was
implied that we should do as GCC does, only grouping things that won't
ever been seen apart.
Figuring that out seems to have been punted down the road, as the
inclusion of our only current example of this (Zbb + unaligned) is
dependant on hardware showing up that actually benefits from it.

The plan then seemed to be press ahead with this series & test the
benefits of the Zbb str* functions in Zbb capable hardware before making
a decision there.

Hopefully I wasn't too far off with that summary...

Thanks,
Conor.

1 - https://lore.kernel.org/linux-riscv/mhng-775d4068-6c1e-48a4-a1dc-b4a76ff26bb3@palmer-ri-x1c9a/

Attachment: signature.asc
Description: PGP signature