KVM: x86: obey KVM_X86_QUIRK_CD_NW_CLEARED in kvm_set_cr0()

Commit b18d5431acc7 ("KVM: x86: fix CR0.CD virtualization") was
technically correct, but it broke OVMF guests by slowing down various
parts of the firmware.

Commit fb279950ba02 ("KVM: vmx: obey KVM_QUIRK_CD_NW_CLEARED") quirked the
first function modified by b18d5431acc7, vmx_get_mt_mask(), for OVMF's
sake. This restored the speed of the OVMF code that runs before
PlatformPei (including the memory intensive LZMA decompression in SEC).

This patch extends the quirk to the second function modified by
b18d5431acc7, kvm_set_cr0(). It eliminates the intrusive slowdown that
hits the EFI_MP_SERVICES_PROTOCOL implementation of edk2's
UefiCpuPkg/CpuDxe -- which is built into OVMF --, when CpuDxe starts up
all APs at once for initialization, in order to count them.

We also carry over the kvm_arch_has_noncoherent_dma() sub-condition from
the other half of the original commit b18d5431acc7.

Fixes: b18d5431acc7a2fd22767925f3a6f597aa4bd29e
Cc: [e-mail]
Cc: Jordan Justen <[e-mail]>
Cc: Alex Williamson <[e-mail]>
Reviewed-by: Xiao Guangrong <[e-mail]>
Tested-by: Janusz Mocek <[e-mail]>
Signed-off-by: Laszlo Ersek <[e-mail]>#
Signed-off-by: Paolo Bonzini <[e-mail]>

arch/x86/kvm/x86.c

···
 	if ((cr0 ^ old_cr0) & update_bits)
 		kvm_mmu_reset_context(vcpu);
 
-	if ((cr0 ^ old_cr0) & X86_CR0_CD)
+	if (((cr0 ^ old_cr0) & X86_CR0_CD) &&
+	    kvm_arch_has_noncoherent_dma(vcpu->kvm) &&
+	    !kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))
 		kvm_zap_gfn_range(vcpu->kvm, 0, ~0ULL);
 
 	return 0;