| Message ID | 20250210-arm64-fix-2024-dpisa-v1-1-d77f43c7b253@kernel.org (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | arm64: Add missing registrations of SF8MMx hwcaps | expand |
On Mon, Feb 10, 2025 at 05:49:33PM +0000, Mark Brown wrote: > Commit 819935464cb2 ("arm64/hwcap: Describe 2024 dpISA extensions to > userspace") added definitions for HWCAP_F8MM8 and HWCAP_F8MM4 but did > not include the crucial registration in arm64_elf_hwcaps. Add it. Looking again, it looks like KERNEL_HWCAP_FPRCVT also isn't wired up? I did a quick check for suspects in that commit with: | [mark@lakrids:~/src/linux]% git show 819935464cb2 | grep '^+\s' | grep -o 'KERNEL_HWCAP_[A-Z0-9_]\+' | sort | uniq -c | sort -n | 1 KERNEL_HWCAP_F8MM4 | 1 KERNEL_HWCAP_F8MM8 | 1 KERNEL_HWCAP_FPRCVT | 1 KERNEL_HWCAP_SME_SF8MM4 | 1 KERNEL_HWCAP_SME_SF8MM8 | 2 KERNEL_HWCAP_CMPBR | 2 KERNEL_HWCAP_SME2P2 | 2 KERNEL_HWCAP_SME_AES | 2 KERNEL_HWCAP_SME_SBITPERM | 2 KERNEL_HWCAP_SME_SFEXPA | 2 KERNEL_HWCAP_SME_SMOP4 | 2 KERNEL_HWCAP_SME_STMOP | 2 KERNEL_HWCAP_SVE2P2 | 2 KERNEL_HWCAP_SVE_AES2 | 2 KERNEL_HWCAP_SVE_BFSCALE | 2 KERNEL_HWCAP_SVE_ELTPERM | 2 KERNEL_HWCAP_SVE_F16MM ... and while KERNEL_HWCAP_SME_SF8MM4 and KERNEL_HWCAP_SME_SF8MM8 are gone now, KERNEL_HWCAP_FPRCVT is still around AFAICT. Checking against a tree based on v6.14-rc1, it superficially looks like we're not missing any other KERNEL_HWCAP_* bits per a coarse search, which shows three missing entries in cpufeature.c: | [mark@lakrids:~/src/linux]% git grep -oh 'KERNEL_HWCAP_[A-Z0-9_]\+' -- arch/arm64/include/asm/hwcap.h | wc -l | 112 | [mark@lakrids:~/src/linux]% git grep -oh 'KERNEL_HWCAP_[A-Z0-9_]\+' -- arch/arm64/include/asm/hwcap.h | sort | uniq | wc -l | 112 | [mark@lakrids:~/src/linux]% git grep -oh 'KERNEL_HWCAP_[A-Z0-9_]\+' -- arch/arm64/kernel/cpuinfo.c | sort | uniq | wc -l | 112 | [mark@lakrids:~/src/linux]% git grep -oh 'KERNEL_HWCAP_[A-Z0-9_]\+' -- arch/arm64/kernel/cpufeature.c | sort | uniq | wc -l | 109 ... but maybe that misses something, and I haven't checked COMPAT_HWCAP(*) or COMPAT_HWCAP2(*). Could you please go check if there's anything else missing more generally? Mark. > Fixes: 819935464cb2 ("arm64/hwcap: Describe 2024 dpISA extensions to userspace") > Reported-by: Mark Rutland <mark.rutland@arm.com> > Signed-off-by: Mark Brown <broonie@kernel.org> > --- > arch/arm64/kernel/cpufeature.c | 2 ++ > 1 file changed, 2 insertions(+) > > diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c > index f0910f20fbf8c18fbeb63bcee18abf13371b1d5e..2cde1e6e309a1b840e0e8dc3a5a55f07f0aadd82 100644 > --- a/arch/arm64/kernel/cpufeature.c > +++ b/arch/arm64/kernel/cpufeature.c > @@ -3190,6 +3190,8 @@ static const struct arm64_cpu_capabilities arm64_elf_hwcaps[] = { > HWCAP_CAP(ID_AA64FPFR0_EL1, F8FMA, IMP, CAP_HWCAP, KERNEL_HWCAP_F8FMA), > HWCAP_CAP(ID_AA64FPFR0_EL1, F8DP4, IMP, CAP_HWCAP, KERNEL_HWCAP_F8DP4), > HWCAP_CAP(ID_AA64FPFR0_EL1, F8DP2, IMP, CAP_HWCAP, KERNEL_HWCAP_F8DP2), > + HWCAP_CAP(ID_AA64FPFR0_EL1, F8MM8, IMP, CAP_HWCAP, KERNEL_HWCAP_F8MM8), > + HWCAP_CAP(ID_AA64FPFR0_EL1, F8MM4, IMP, CAP_HWCAP, KERNEL_HWCAP_F8MM4), > HWCAP_CAP(ID_AA64FPFR0_EL1, F8E4M3, IMP, CAP_HWCAP, KERNEL_HWCAP_F8E4M3), > HWCAP_CAP(ID_AA64FPFR0_EL1, F8E5M2, IMP, CAP_HWCAP, KERNEL_HWCAP_F8E5M2), > #ifdef CONFIG_ARM64_POE > > --- > base-commit: 21fed7c223e20e694b91dbf25936d922a50c8b19 > change-id: 20250207-arm64-fix-2024-dpisa-635a5961609c > > Best regards, > -- > Mark Brown <broonie@kernel.org> >
On Mon, Feb 10, 2025 at 06:41:09PM +0000, Mark Rutland wrote: > On Mon, Feb 10, 2025 at 05:49:33PM +0000, Mark Brown wrote: > > Commit 819935464cb2 ("arm64/hwcap: Describe 2024 dpISA extensions to > > userspace") added definitions for HWCAP_F8MM8 and HWCAP_F8MM4 but did > > not include the crucial registration in arm64_elf_hwcaps. Add it. > Looking again, it looks like KERNEL_HWCAP_FPRCVT also isn't wired up? Indeed. I'll add that separately. > I did a quick check for suspects in that commit with: > | [mark@lakrids:~/src/linux]% git show 819935464cb2 | grep '^+\s' | grep -o 'KERNEL_HWCAP_[A-Z0-9_]\+' | sort | uniq -c | sort -n > Checking against a tree based on v6.14-rc1, it superficially looks like > we're not missing any other KERNEL_HWCAP_* bits per a coarse search, > which shows three missing entries in cpufeature.c: > | [mark@lakrids:~/src/linux]% git grep -oh 'KERNEL_HWCAP_[A-Z0-9_]\+' -- arch/arm64/kernel/cpuinfo.c | sort | uniq | wc -l > | 112 > | [mark@lakrids:~/src/linux]% git grep -oh 'KERNEL_HWCAP_[A-Z0-9_]\+' -- arch/arm64/kernel/cpufeature.c | sort | uniq | wc -l > | 109 > ... but maybe that misses something, and I haven't checked > COMPAT_HWCAP(*) or COMPAT_HWCAP2(*). > Could you please go check if there's anything else missing more generally? The other two (on top of FPRCVT) are CPUID and EVTSTRM. Both of these are set using cpu_set_named_feature() which doesn't use match your grep for KERNEL_HWCAP_ in cpufeature.c but instead generates the constant name via token pasting.
On Mon, Feb 10, 2025 at 07:14:58PM +0000, Mark Brown wrote: > On Mon, Feb 10, 2025 at 06:41:09PM +0000, Mark Rutland wrote: > > On Mon, Feb 10, 2025 at 05:49:33PM +0000, Mark Brown wrote: > > > Commit 819935464cb2 ("arm64/hwcap: Describe 2024 dpISA extensions to > > > userspace") added definitions for HWCAP_F8MM8 and HWCAP_F8MM4 but did > > > not include the crucial registration in arm64_elf_hwcaps. Add it. > > > Looking again, it looks like KERNEL_HWCAP_FPRCVT also isn't wired up? > > Indeed. I'll add that separately. Given that KERNEL_HWCAP_FPRCVT was also added by commit 819935464cb2, I assume you mean you'll send a v2 of this patch doing that. Mark.
On Tue, Feb 11, 2025 at 11:37:44AM +0000, Mark Rutland wrote: > On Mon, Feb 10, 2025 at 07:14:58PM +0000, Mark Brown wrote: > > On Mon, Feb 10, 2025 at 06:41:09PM +0000, Mark Rutland wrote: > > > Looking again, it looks like KERNEL_HWCAP_FPRCVT also isn't wired up? > > Indeed. I'll add that separately. > Given that KERNEL_HWCAP_FPRCVT was also added by commit 819935464cb2, > I assume you mean you'll send a v2 of this patch doing that. Well, I was actually intending to send a separate patch given that otherwise it ends up looking like two different changelogs in the same patch and the diffs shouldn't overlap.
On Tue, Feb 11, 2025 at 12:58:10PM +0000, Mark Brown wrote: > On Tue, Feb 11, 2025 at 11:37:44AM +0000, Mark Rutland wrote: > > On Mon, Feb 10, 2025 at 07:14:58PM +0000, Mark Brown wrote: > > > On Mon, Feb 10, 2025 at 06:41:09PM +0000, Mark Rutland wrote: > > > > > Looking again, it looks like KERNEL_HWCAP_FPRCVT also isn't wired up? > > > > Indeed. I'll add that separately. > > > Given that KERNEL_HWCAP_FPRCVT was also added by commit 819935464cb2, > > I assume you mean you'll send a v2 of this patch doing that. > > Well, I was actually intending to send a separate patch given that > otherwise it ends up looking like two different changelogs in the same > patch and the diffs shouldn't overlap. It would be simpler for others if you sent a single patch that handled the three hwcaps that we failed to wire up in commit 819935464cb2. That way they only have one thing to track. I don't understand what you mean by "two different changelogs"; surely the commit message is a trivial variation of the one for this v1 patch: | arm64: Add missing hwcap registrations | | Commit 819935464cb2 ("arm64/hwcap: Describe 2024 dpISA extensions to | userspace") added definitions for HWCAP_F8MM8, HWCAP_F8MM4, and | HWCAP_FPRCVT, but did not include the crucial registration in | arm64_elf_hwcaps. Add them. ... and the diff from v1 is "Handle HWCAP_FPRCVT". Mark.
On Tue, Feb 11, 2025 at 01:15:52PM +0000, Mark Rutland wrote: > On Tue, Feb 11, 2025 at 12:58:10PM +0000, Mark Brown wrote: > > Well, I was actually intending to send a separate patch given that > > otherwise it ends up looking like two different changelogs in the same > > patch and the diffs shouldn't overlap. > It would be simpler for others if you sent a single patch that handled > the three hwcaps that we failed to wire up in commit 819935464cb2. That > way they only have one thing to track. > I don't understand what you mean by "two different changelogs"; surely the > commit message is a trivial variation of the one for this v1 patch: My first scan through had made it look like there was more stuff missing for FPRCVT but now that I go to write the change that's not actually the case, it's different handling of the different ID registers.
diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c index f0910f20fbf8c18fbeb63bcee18abf13371b1d5e..2cde1e6e309a1b840e0e8dc3a5a55f07f0aadd82 100644 --- a/arch/arm64/kernel/cpufeature.c +++ b/arch/arm64/kernel/cpufeature.c @@ -3190,6 +3190,8 @@ static const struct arm64_cpu_capabilities arm64_elf_hwcaps[] = { HWCAP_CAP(ID_AA64FPFR0_EL1, F8FMA, IMP, CAP_HWCAP, KERNEL_HWCAP_F8FMA), HWCAP_CAP(ID_AA64FPFR0_EL1, F8DP4, IMP, CAP_HWCAP, KERNEL_HWCAP_F8DP4), HWCAP_CAP(ID_AA64FPFR0_EL1, F8DP2, IMP, CAP_HWCAP, KERNEL_HWCAP_F8DP2), + HWCAP_CAP(ID_AA64FPFR0_EL1, F8MM8, IMP, CAP_HWCAP, KERNEL_HWCAP_F8MM8), + HWCAP_CAP(ID_AA64FPFR0_EL1, F8MM4, IMP, CAP_HWCAP, KERNEL_HWCAP_F8MM4), HWCAP_CAP(ID_AA64FPFR0_EL1, F8E4M3, IMP, CAP_HWCAP, KERNEL_HWCAP_F8E4M3), HWCAP_CAP(ID_AA64FPFR0_EL1, F8E5M2, IMP, CAP_HWCAP, KERNEL_HWCAP_F8E5M2), #ifdef CONFIG_ARM64_POE
Commit 819935464cb2 ("arm64/hwcap: Describe 2024 dpISA extensions to userspace") added definitions for HWCAP_F8MM8 and HWCAP_F8MM4 but did not include the crucial registration in arm64_elf_hwcaps. Add it. Fixes: 819935464cb2 ("arm64/hwcap: Describe 2024 dpISA extensions to userspace") Reported-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Mark Brown <broonie@kernel.org> --- arch/arm64/kernel/cpufeature.c | 2 ++ 1 file changed, 2 insertions(+) --- base-commit: 21fed7c223e20e694b91dbf25936d922a50c8b19 change-id: 20250207-arm64-fix-2024-dpisa-635a5961609c Best regards,