Messages in this thread | | | Date | Thu, 4 Nov 2021 11:00:15 +0100 | Subject | Re: [PATCH] crypto: x86/aes-ni: fix AVX detection | From | Paolo Bonzini <> |
| |
On 11/3/21 15:52, Herbert Xu wrote: > On Wed, Nov 03, 2021 at 07:43:43AM -0700, Dave Hansen wrote: >> The kernel shouldn't crash in this case. We've got a software >> dependency which should disable AVX2 if AVX is off: > It's qemu, I thought it was a qemu bug but Paulo disagrees.
I don't disagree that QEMU could do better in preventing nonsensical combinations; however, independent of that you shouldn't use AVX2 without factoring XCR0 in the decision. This is true no matter if userspace/kernel or even bare metal/virt.
That does not have to be done directly with cpu_has_xfeatures as in Maxim's patch: if boot_cpu_has(AVX) returns false after xsave_cpuid_features has filtered it out, or if boot_cpu_has(AVX2) returns false after cpuid_deps has filtered it out, that's fine. I guess Maxim can look at the pointers that Dave provided and check if there's another bug somewhere in arch/x86/kernel.
Paolo
| |