KVM: x86/xcr0: Don't make XFEATURE_MASK_SSE a mandatory bit setting
PATCH 1/1 KVM: x86/xcr0: Don't make XFEATURE_MASK_SSE a mandatory bit setting
From: Like Xu <[email protected]>
If "bits 1 and 2 of XCR0" are cleared and set only "bit 0 of XCR0", the x86 hardware does work properly and the CPUID leaf of the states size will be updated as follows:
xsetbv(XCR_XFEATURE_ENABLED_MASK, 0x2e7);
CPUID[0d, 00]: eax=000002e7 ebx=00000a88 ecx=00000a88 edx=00000000xsetbv(XCR_XFEATURE_ENABLED_MASK, 0x1);
CPUID[0d, 00]: eax=000002e7 ebx=00000240 ecx=00000a88 edx=00000000
According to Intel SDM "XSETBV - Set Extended Control Register", the truth emerges: only the "bit 0 of XCR0 (corresponding to x87 state) must be set to 1", and in addition, it's necessary to set both bits (XCR0[1] and XCR0[2]) to use AVX instructions".
Therefore, it's pretty safe for KVM to run a user-defined guest w/o using AVX instructions and in that case, the kvm_update_cpuid_runtime() cannot work as expected due to the wrong macro definition.
Remove the XFEATURE_MASK_SSE bit as part of the XFEATURE_MASK_EXTEND and opportunistically, move it into the context of its unique user KVM.
--- arch/x86/include/asm/fpu/xstate.h | 3 --- arch/x86/kvm/cpuid.h | 8 ++++++++ 2 files changed, 8 insertions(+), 3 deletions(-)
arch/x86/include/asm/fpu/xstate.h +0 -3
@@ -9,9 +9,6 @@#include <asm/fpu/api.h> #include <asm/user.h> -/* Bit 63 of XCR0 is reserved for future expansion */ -#define XFEATURE_MASK_EXTEND (~(XFEATURE_MASK_FPSSE | (1ULL << 63))) -#define XSTATE_CPUID 0x0000000d#define TILE_CPUID 0x0000001d
arch/x86/kvm/cpuid.h +8 -0
@@ -8,6 +8,14 @@#include <asm/processor.h> #include <uapi/asm/kvm_para.h> +/* +* Bit 63 of XCR0 is reserved for future expansion, +* and will not represent a processor state component. +* +* Only bit 0 of XCR0 (x87 state) must be set to 1. +*/ +#define XFEATURE_MASK_EXTEND (~(XFEATURE_MASK_FP | (1ULL << 63))) +extern u32 kvm_cpu_caps[NR_KVM_CPU_CAPS ]__read_mostly ;void kvm_set_cpu_caps (void );
Signature · 1 line
2.33.1
Quoted 2 lines · Like Xu
Remove the
XFEATURE_MASK_SSEbit as part of theXFEATURE_MASK_EXTENDand opportunistically, move it into the context of its unique user KVM.
Is this a problem for xstate_required_size()? The rules for the CPUID sub-functions <=1 are different than those for >1. Most importantly, 'eax' doesn't enumerate the size of the feature for the XFEATURE_SSE sub-leaf.
I think XFEATURE_MASK_EXTEND was being used to avoid that oddity:
Quoted 24 lines
u32
xstate_required_size(u64 xstate_bv, bool compacted){ int feature_bit = 0; u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET; xstate_bv &= XFEATURE_MASK_EXTEND; while (xstate_bv) { if (xstate_bv & 0x1) { u32 eax, ebx, ecx, edx, offset; cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx); /* ECX[1]: 64B alignment in compacted form */ if (compacted) offset = (ecx & 0x2) ? ALIGN(ret, 64) : ret; else offset = ebx; ret = max(ret, offset + eax); } xstate_bv >>= 1; feature_bit++; } return ret; }
Quoted 8 lines · Dave Hansen
Quoted 2 lines · Like Xu
Remove the
XFEATURE_MASK_SSEbit as part of theXFEATURE_MASK_EXTENDand opportunistically, move it into the context of its unique user KVM.Is this a problem for
xstate_required_size()? The rules for the CPUID sub-functions <=1 are different than those for >1. Most importantly, 'eax' doesn't enumerate the size of the feature for theXFEATURE_SSEsub-leaf.
Indeed.
Quoted 2 lines
I think
XFEATURE_MASK_EXTENDwas being used to avoid that oddity:
It seems that the cpuid.0xd.0.ebx size update for the SSE+AVX state needs to be triggered by setting bit 2 which is quite odd:
XCR0 = 001B, ebx=00000240
XCR0 = 011B, ebx=00000240
XCR0 = 111B, ebx=00000340
Thank you and sorry for the noise.
Quoted 25 lines
Quoted 25 lines
u32
xstate_required_size(u64 xstate_bv, bool compacted){ int feature_bit = 0; u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET; xstate_bv &= XFEATURE_MASK_EXTEND; while (xstate_bv) { if (xstate_bv & 0x1) { u32 eax, ebx, ecx, edx, offset; cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx); /* ECX[1]: 64B alignment in compacted form */ if (compacted) offset = (ecx & 0x2) ? ALIGN(ret, 64) : ret; else offset = ebx; ret = max(ret, offset + eax); } xstate_bv >>= 1; feature_bit++; } return ret; }