Re: [PATCH v3 7/8] riscv: Add parameter for skipping access speed tests

[Alexandre Ghiti]

On 18/03/2025 13:45, Andrew Jones wrote:
> On Tue, Mar 18, 2025 at 01:13:18PM +0100, Alexandre Ghiti wrote:
> > On 18/03/2025 09:48, Andrew Jones wrote:
> > > On Mon, Mar 17, 2025 at 03:39:01PM +0100, Alexandre Ghiti wrote:
> > > > Hi Drew,
> > > >
> > > > On 04/03/2025 13:00, Andrew Jones wrote:
> > > > > Allow skipping scalar and vector unaligned access speed tests. This
> > > > > is useful for testing alternative code paths and to skip the tests in
> > > > > environments where they run too slowly. All CPUs must have the same
> > > > > unaligned access speed.
> > > > I'm not a big fan of the command line parameter, this is not where we should
> > > > push uarch decisions because there could be many other in the future, the
> > > > best solution to me should be in DT/ACPI and since the DT folks, according
> > > > to Palmer, shut down this solution, it remains using an extension.
> > > >
> > > > I have been reading a bit about unaligned accesses. Zicclsm was described as
> > > > "Even though mandated, misaligned loads and stores might execute extremely
> > > > slowly. Standard software distributions should assume their existence only
> > > > for correctness, not for performance." in rva20/22 but *not* in rva23. So
> > > > what about using this "hole" and consider that a platform that *advertises*
> > > > Zicclsm means its unaligned accesses are fast? After internal discussion, It
> > > > actually does not make sense to advertise Zicclsm if the platform accesses
> > > > are slow right?
> > > This topic pops up every so often, including in yesterday's server
> > > platform TG call. In that call, and, afaict, every other time it has
> > > popped up, the result is to reiterate that ISA extensions never say
> > > anything about performance. So, Zicclsm will never mean fast and we
> > > won't likely be able to add any extension that does.
> >
> > Ok, I should not say "fast". Usually, when an extension is advertised by a
> > platform, we don't question its speed (zicboz, zicbom...etc), we simply use
> > it and it's up to the vendor to benchmark its implementation and act
> > accordingly (i.e. do not set it in the isa string).
> >
> >
> > > > arm64 for example considers that armv8 has fast unaligned accesses and can
> > > > then enable HAVE_EFFICIENT_ALIGNED_ACCESS in the kernel, even though some
> > > > uarchs are slow. Distros will very likely use rva23 as baseline so they will
> > > > enable Zicclsm which would allow us to take advantage of this too, without
> > > > this, we lose a lot of perf improvement in the kernel, see
> > > > https://lore.kernel.org/lkml/20231225044207.3821-1-jszhang@xxxxxxxxxx/.
> > > >
> > > > Or we could have a new named feature for this, even though it's weird to
> > > > have a named feature which would basically  mean "Zicclsm is fast". We don't
> > > > have, for example, a named feature to say "Zicboz is fast" but given the
> > > > vague wording in the profile spec, maybe we can ask for one in that case?
> > > >
> > > > Sorry for the late review and for triggering this debate...
> > > No problem, let's try to pick the best option. I'll try listing all the
> > > options and there pros/cons.
> > >
> > > 1. Leave as is, which is to always probe
> > >      pro: Nothing to do
> > >      con: Not ideal in all environments
> > >
> > > 2. New DT/ACPI description
> > >      pro: Describing whether or not misaligned accesses are implemented in
> > >           HW (which presumably means fast) is something that should be done
> > > 	in HW descriptions
> > >      con: We'll need to live with probing until we can get the descriptions
> > >           defined, which may be never if there's too much opposition
> > >
> > > 3. Command line
> > >      pro: Easy and serves its purpose, which is to skip probing in the
> > >           environments where probing is not desired
> > >      con: Yet another command line option (which we may want to deprecate
> > >           someday)
> > >
> > > 4. New ISA extension
> > >      pro: Easy to add to HW descriptions
> > >      con: Not likely to get it through ratification
> > >
> > > 5. New SBI FWFT feature
> > >      pro: Probably easier to get through ratification than an ISA extension
> > >      con: Instead of probing, kernel would have to ask SBI -- would that
> > >           even be faster? Will all the environments that want to skip
> > > 	probing even have a complete SBI?
> > >
> > > 6. ??
> >
> > So what about:
> >
> > 7. New enum value describing the performance as "FORCED" or "HW" (or
> > anything better)
> >      pro: We only use the existing Zicclsm
> >      con: It's not clear that the accesses are fast but it basically says to
> > SW "don't think too much, I'm telling you that you can use it", up to us to
> > describe this correctly for users to understand.
> But Zicclsm doesn't mean misaligned accesses are in HW, it just means
> they're not going to explode.


They never explode since if they are not supported by the HW, we rely on 
S-mode emulation already.


> We'd still need the probing to find out
> if the accesses are emulated (slow) or hw (fast). We at least want to
> know the answer to that question because we advertise it to userspace
> through hwprobe.
>
> (BTW, another pro of the command line is that it can be used to test
> both slow and fast paths without recompiling.)
>
> Thanks,
> drew