[PATCH v4] RISC-V: Show accurate per-hart isa in /proc/cpuinfo

[Andrew Jones]

On Thu, Aug 24, 2023 at 03:06:53PM -0700, Evan Green wrote:
> On Thu, Aug 24, 2023 at 10:29 AM Conor Dooley <conor at kernel.org> wrote:
...
> > Do you want to have this new thing in cpuinfo tell the user "this hart
> > has xyz extensions that are supported by a kernel, but maybe not this
> > kernel" or to tell the user "this hart has xyz extensions that are
> > supported by this kernel"? Your text above says "understood by the
> > kernel", but I think that's a poor definition that needs to be improved
> > to spell out exactly what you mean. IOW does "understood" mean the
> > kernel will parse them into a structure, or does it mean "yes you can
> > use this extension on this particular hart".
> 
> I'm imagining /proc/cpuinfo being closer to "the CPU has it and the
> kernel at least vaguely understands it, but may not have full support
> for it enabled". I'm assuming /proc/cpuinfo is mostly used by 1)
> humans wanting to know if they have hardware support for a feature,
> and 2) administrators collecting telemetry to manage fleets (ie do I
> have any hardware deployed that supports X).

Is (2) a special case of (1)? (I want to make sure I understand all the
cases.)

> Programmers looking to
> see "is the kernel support for this feature ready right now" would
> ideally not be parsing /proc/cpuinfo text, as more direct mechanisms
> like specific hwprobe bits for "am I fully ready to go" would be
> easier to work with. Feel free to yell at me if this overall vision
> seems flawed.
> 
> I tried to look to see if there was consensus among the other
> architectures. Aarch64 seems to go with "supported and fully enabled",
> as their cpu_has_feature() directly tests elf_hwcap, and elements in
> arm64_elf_hwcaps[] are Kconfig gated. X86 is complicated, but IIRC is
> more along the lines of "hardware has it". They have two macros,
> cpu_has() for "raw capability" and cpu_feature_enabled() for "kernel
> can do it too", and they use cpu_has() for /proc/cpuinfo flags.
>

I'd lean more towards the way AArch64 goes, because, unless /proc/cpuinfo
is just a blind regurgitation of an isa string from DT / ACPI, then the
kernel must at least know something about it. Advertising a feature which
is known, but also known not to work, seems odd to me. So my vote is that
only features which are present and enabled in the kernel or present and
not necessary to be enabled in the kernel in order for userspace or
virtual machines to use be advertised in /proc/cpuinfo.

We still have SMBIOS (dmidecode) to blindly dump what the hardware
supports for cases (1) and (2) above.

Thanks,
drew