On Tue, Dec 07, 2021, Lai Jiangshan wrote:
> diff --git a/arch/x86/kvm/mmu.h b/arch/x86/kvm/mmu.h
> index b70b36734bc0..0cb2c52377c8 100644
> --- a/arch/x86/kvm/mmu.h
> +++ b/arch/x86/kvm/mmu.h
> @@ -252,23 +252,26 @@ static inline u8 permission_fault(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
> unsigned pte_access, unsigned pte_pkey,
> unsigned pfec)
> {
> - int cpl = static_call(kvm_x86_get_cpl)(vcpu);
> unsigned long rflags = static_call(kvm_x86_get_rflags)(vcpu);
>
> /*
> - * If CPL < 3, SMAP prevention are disabled if EFLAGS.AC = 1.
> + * If explicit supervisor accesses, SMAP is disabled
Slight reword, and each clause can fit on one line.
* For explicit supervisor accesses, SMAP is disabled if EFLAGS.AC = 1.
*
* For implicit supervisor accesses, SMAP cannot be overridden.
> + * if EFLAGS.AC = 1.
> *
> - * If CPL = 3, SMAP applies to all supervisor-mode data accesses
> - * (these are implicit supervisor accesses) regardless of the value
> - * of EFLAGS.AC.
> + * If implicit supervisor accesses, SMAP can not be disabled
> + * regardless of the value EFLAGS.AC.
> *
> - * This computes (cpl < 3) && (rflags & X86_EFLAGS_AC), leaving
> + * SMAP works on supervisor accesses only, and not_smap can
> + * be set or not set when user access with neither has any bearing
> + * on the result.
This is quite jumbled, I'd just drop it entirely, the interesting bits are
the rules for implicit vs. explicit and the blurb below that describes the magic.
> + *
> + * This computes explicit_access && (rflags & X86_EFLAGS_AC), leaving
Too many &&, the logic below is a bitwise &, not a logical &&.
> * the result in X86_EFLAGS_AC. We then insert it in place of
> * the PFERR_RSVD_MASK bit; this bit will always be zero in pfec,
> * but it will be one in index if SMAP checks are being overridden.
> * It is important to keep this branchless.
Heh, so important that it incurs multiple branches and possible VMREADs in
vmx_get_cpl() and vmx_get_rflags(). And before static_call, multiple retpolines
to boot. Probably a net win now as only the first permission_fault() check for
a given VM-Exit be penalized, but the comment is amusing nonetheless.
> */
> - unsigned long not_smap = (cpl - 3) & (rflags & X86_EFLAGS_AC);
> + u32 not_smap = (rflags & X86_EFLAGS_AC) & vcpu->arch.explicit_access;
I really, really dislike shoving this into vcpu->arch. I'd much prefer to make
this a property of the access, even if that means adding another param or doing
something gross with @access (@pfec here).
> int index = (pfec >> 1) +
> (not_smap >> (X86_EFLAGS_AC_BIT - PFERR_RSVD_BIT + 1));
> bool fault = (mmu->permissions[index] >> pte_access) & 1;
