KVM: x86: Emulator performs privilege checks on __linearize

When segment is accessed, real hardware does not perform any privilege level
checks.  In contrast, KVM emulator does. This causes some discrepencies from
real hardware. For instance, reading from readable code segment may fail due to
incorrect segment checks. In addition, it introduces unnecassary overhead.

To reference Intel SDM 5.5 ("Privilege Levels"): "Privilege levels are checked
when the segment selector of a segment descriptor is loaded into a segment
register." The SDM never mentions privilege level checks during memory access,
except for loading far pointers in section 5.10 ("Pointer Validation"). Those
are actually segment selector loads and are emulated in the similarily (i.e.,
regardless to __linearize checks).

This behavior was also checked using sysexit. A data-segment whose DPL=0 was
loaded, and after sysexit (CPL=3) it is still accessible.

Therefore, all the privilege level checks in __linearize are removed.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

arch/x86/kvm/emulate.c

@@ -663,7 +663,6 @@ static __always_inline int __linearize(struct x86_emulate_ctxt *ctxt,
 	ulong la;
 	u32 lim;
 	u16 sel;
-	unsigned cpl;
 
 	la = seg_base(ctxt, addr.seg) + addr.ea;
 	*max_size = 0;
@@ -705,20 +704,6 @@ static __always_inline int __linearize(struct x86_emulate_ctxt *ctxt,
 		}
 		if (size > *max_size)
 			goto bad;
-		cpl = ctxt->ops->cpl(ctxt);
-		if (!fetch) {
-			/* data segment or readable code segment */
-			if (cpl > desc.dpl)
-				goto bad;
-		} else if ((desc.type & 8) && !(desc.type & 4)) {
-			/* nonconforming code segment */
-			if (cpl != desc.dpl)
-				goto bad;
-		} else if ((desc.type & 8) && (desc.type & 4)) {
-			/* conforming code segment */
-			if (cpl < desc.dpl)
-				goto bad;
-		}
 		break;
 	}
 	if (ctxt->mode != X86EMUL_MODE_PROT64)