On Thu, Apr 18, 2024 at 07:26:00PM +0100, Conor Dooley wrote:
> On Thu, Apr 18, 2024 at 10:39:46AM -0700, Eric Biggers wrote:
> > On Thu, Apr 18, 2024 at 10:32:03AM -0700, Eric Biggers wrote:
> > > On Thu, Apr 18, 2024 at 05:53:55PM +0100, Conor Dooley wrote:
> > > > > If it would be useful to do so, we should be able to enable some of the code
> > > > > with a smaller VLEN and/or EEW once it has been tested in those configurations.
> > > > > Some of it should work, but some of it won't be able to work.  (For example, the
> > > > > SHA512 instructions require EEW==64.)
> > > > > 
> > > > > Also note that currently all the RISC-V vector crypto code only supports riscv64
> > > > > (XLEN=64).  Similarly, that could be relaxed in the future if people really need
> > > > > the vector crypto acceleration on 32-bit CPUs...  But similarly, the code would
> > > > > need to be revised and tested in that configuration.
> > > > > 
> > > > > > Eric/Jerry (although read the previous paragraph too):
> > > > > > I noticed that the sha256 glue code calls crypto_simd_usable(), and in
> > > > > > turn may_use_simd() before kernel_vector_begin(). The chacha20 glue code
> > > > > > does not call either, which seems to violate the edict in
> > > > > > kernel_vector_begin()'s kerneldoc:
> > > > > > "Must not be called unless may_use_simd() returns true."
> > > > > 
> > > > > skcipher algorithms can only be invoked in process and softirq context.  This
> > > > > differs from shash algorithms which can be invoked in any context.
> > > > > 
> > > > > My understanding is that, like arm64, RISC-V always allows non-nested
> > > > > kernel-mode vector to be used in process and softirq context -- and in fact,
> > > > > this was intentionally done in order to support use cases like this.  So that's
> > > > > why the RISC-V skcipher algorithms don't check for may_use_simd() before calling
> > > > > kernel_vector_begin().
> > > > 
> > > > I see, thanks for explaining that. I think you should probably check
> > > > somewhere if has_vector() returns true in that driver though before
> > > > using vector instructions. Only checking vlen seems to me like relying on
> > > > an implementation detail and if we set vlen for the T-Head/0.7.1 vector
> > > > it'd be fooled. That said, I don't think that any of the 0.7.1 vector
> > > > systems actually support Zvkb, but I hope you get my drift.
> > > 
> > > All the algorithms check for at least one of the vector crypto extensions being
> > > supported, for example Zvkb.  'if (riscv_isa_extension_available(NULL, ZVKB))'
> > > should return whether the ratified version of Zvkb is supported, and likewise
> > > for the other vector crypto extensions.  The ratified version of the vector
> > > crypto extensions depends on the ratified version of the vector extension.
> 
> That's great if it does require that the version of the vector extension
> must be standard. Higher quality spec than most if it does. But
> "supported" in the context of riscv_isa_extension_available() means that
> the hardware supports it (or set of harts), not that the currently
> running kernel does. The Kconfig deps that must be met for the code to be
> built at least mean the kernel is built with vector support, leaving only
> "the kernel was built with vector support and the hardware supports vector
> but for $reason the kernel refused to enable it".
> 
> I'm not sure if that final condition is actually possible with the system
> ending up in a broken state, however - I'm not sure that we ever do turn
> off access to the VPU at present (after we mark it usable), and if we do
> it doesn't get reflected in has_vector() so the kernel and userspace would
> both break, with what a crypto driver does probably being the least of
> your worries.
> 
> > > So
> > > there should be no issue.  If there is, the RISC-V core architecture code needs
> > > to be fixed to not declare that extensions are supported when they are actually
> > > incompatible non-standard versions of those extensions.  Incompatible
> > > non-standard extensions should be represented as separate extensions.
> > > 
> > 
> > It probably makes sense to check has_vector() to exclude Zve* for now, though.
> 
> I think you might actually be better served at present, given the code can
> only be built if the core vector code is, by using
> riscv_isa_extension_available(NULL, v). That way you know for sure that
> you're getting the ratified extension and nothing else.

Poor choice of wording here - I meant, of course, the "main" vector
extension, rather than the Zve* variants.

> Prior to this conversation I thought that has_vector() should return true
> if there's a standard compliant vector unit available - given all users
> Andy added only need Zve32x.
> 
> > I am just concerned about how you're suggesting that non-standard extensions
> > might be pretending to be standard ones and individual users of kernel-mode
> > vector would need to work around that.
> 
> I am absolutely not suggesting that non-standard extensions should
> masquerade as standard ones, I don't know where you got that from. What
> I said was that a non-standard vector extension could reuse riscv_v_vlen
> (and should IMO for simplicity reasons), not that any of the APIs we have
> for checking extension availability would lie and say it was standard.
> riscv_v_vlen having a value greater than 128 is not one of those APIs ;)
> 
> > I think that neither has_vector() nor
> > 'if (riscv_isa_extension_available(NULL, ZVKB))' should return true if the CPU's
> > vector extension is non-standard.
> 
> riscv_isa_extension_available(NULL, ZVKB) only checks whether the extension
> was present in DT or ACPI for all harts. It doesn't check whether or not
> the required config option for vector has been set or anything related
> to dependencies. has_vector() at least checks that the vector core has
> been enabled (and uses the alternative-patched version of the check
> given it is used in some hotter paths). That's kinda moot for code
> that's only built if the vector core stuff is enabled as I said above
> though.
> 
> We could of course make riscv_isa_extension_available() check
> extension dependencies, but I'd rather leave dt validation to the dt
> tooling (apparently ACPI tables are never wrong...). Either would allow
> you to rely on the crypto extensions present only when the standard vector
> extensions unless someone's DT/ACPI stuff is shite, but then they keep the
> pieces IMO :)
> 
> Hope that makes sense?
> Conor.