Skip to contentSkip to conversation

KVM: x86: Fix stalled descriptor table accesses to MMIO

PATCH Superseded weichenchen 2 participants 2 messages Nov 19

Subsystems (3)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 KVM: x86: Fix stalled descriptor table accesses to MMIO

Use the emulated memory accessors for the explicit operands of LGDT, LIDT, SGDT and SIDT. When the operand has no backing memslot, the standard accessors return X86EMUL_IO_NEEDED without queuing an MMIO fragment. LIDT can repeat the generic operand read at the same RIP, and SIDT can re-enter emulation without sending a write to userspace.

Check the segment bounds of both fields before reading the limit or base. If the operand crosses a page boundary, pass the whole operand to the emulated read so a fault on the second page is reported before an MMIO exit to userspace. Keep the separate limit and base reads on a single page. Use the emulator's read cache to resume after MMIO and set NoAccess to skip the redundant generic source read.

Use the destination operand for common writeback of SGDT and SIDT. This keeps the entire descriptor available until userspace completes a ten-byte MMIO write.

Fixes c3cd7ffaf57a ("KVM: x86 emulator: x86_emulate_insn() return -1 only in case of emulation failure")
Fixes 129a72a0d3c8 ("KVM: x86: Introduce segmented_write_std")

---
On Intel VMX/EPT, LIDT and SIDT both reached the following HLT with the expected IDTR data for RAM and MMIO operands. LIDT completed a 2-byte and an 8-byte MMIO read. SIDT completed an 8-byte and a 2-byte write.

With only the limit field in range in 32-bit compatibility mode, LIDT delivered #GP(0) at the expected RIP before any userspace MMIO exit. The same selftest passed with forced emulation. A local extension to set_memory_region_test also covered all four descriptor-table instructions and a split RAM/MMIO LIDT.

A standalone test of the emulator and its MMIO fragment handling used mock page translation to check a userspace MMIO page followed by a non-present page. All 36 cross-page load cases returned a page fault without preparing a userspace MMIO exit or updating GDTR or IDTR.

Six KVM selftests passed. Eight kvm-unit-tests suites at a2891d9528ef reported 445 passes, five skips and no failures. Both vendor modules built with KVM_WERROR. Runtime testing used Intel.

Codex assisted with the emulator changes and regression testing.

 arch/x86/kvm/emulate.c | 57 ++++++++++++++++++++++--------------------
 1 file changed, 30 insertions(+), 27 deletions(-)
arch/x86/kvm/emulate.c +30 -27
index c1b2128218..3bd720d983 100644
@@ -918,23 +918,6 @@ static void *decode_register(struct x86_emulate_ctxt *ctxt, u8 modrm_reg,
 	return p;
 }
 
-static int read_descriptor(struct x86_emulate_ctxt *ctxt,
-			   struct segmented_address addr,
-			   u16 *size, unsigned long *address, int op_bytes)
-{
-	int rc;
-
-	if (op_bytes == 2)
-		op_bytes = 3;
-	*address = 0;
-	rc = segmented_read_std(ctxt, addr, size, 2);
-	if (rc != X86EMUL_CONTINUE)
-		return rc;
-	addr.ea += 2;
-	rc = segmented_read_std(ctxt, addr, address, op_bytes);
-	return rc;
-}
-
 EM_ASM_2(add);
 EM_ASM_2(or);
 EM_ASM_2(adc);
@@ -3447,10 +3430,9 @@ static int emulate_store_desc_ptr(struct x86_emulate_ctxt *ctxt,
 		ctxt->op_bytes = 4;
 		desc_ptr.address &= 0x00ffffff;
 	}
-	/* Disable writeback. */
-	ctxt->dst.type = OP_NONE;
-	return segmented_write_std(ctxt, ctxt->dst.addr.mem,
-				   &desc_ptr, 2 + ctxt->op_bytes);
+	ctxt->dst.bytes = 2 + ctxt->op_bytes;
+	memcpy(ctxt->dst.valptr, &desc_ptr, ctxt->dst.bytes);
+	return X86EMUL_CONTINUE;
 }
 
 static int em_sgdt(struct x86_emulate_ctxt *ctxt)
@@ -3465,14 +3447,35 @@ static int em_sidt(struct x86_emulate_ctxt *ctxt)
 
 static int em_lgdt_lidt(struct x86_emulate_ctxt *ctxt, bool lgdt)
 {
-	struct desc_ptr desc_ptr;
+	struct segmented_address addr = ctxt->src.addr.mem;
+	struct desc_ptr desc_ptr = {};
+	unsigned long limit_addr, base_addr;
+	unsigned int base_bytes, desc_bytes;
 	int rc;
 
 	if (ctxt->mode == X86EMUL_MODE_PROT64)
 		ctxt->op_bytes = 8;
-	rc = read_descriptor(ctxt, ctxt->src.addr.mem,
-			     &desc_ptr.size, &desc_ptr.address,
-			     ctxt->op_bytes);
+	base_bytes = ctxt->op_bytes == 2 ? 3 : ctxt->op_bytes;
+	desc_bytes = sizeof(desc_ptr.size) + base_bytes;
+
+	/* Check both fields before an MMIO read can have side effects. */
+	rc = linearize(ctxt, addr, sizeof(desc_ptr.size), false, &limit_addr);
+	if (rc != X86EMUL_CONTINUE)
+		return rc;
+	addr.ea += sizeof(desc_ptr.size);
+	rc = linearize(ctxt, addr, base_bytes, false, &base_addr);
+	if (rc != X86EMUL_CONTINUE)
+		return rc;
+
+	/* Defer userspace MMIO until both pages have been translated. */
+	if (offset_in_page(limit_addr) + desc_bytes > PAGE_SIZE) {
+		rc = read_emulated(ctxt, limit_addr, &desc_ptr, desc_bytes);
+	} else {
+		rc = read_emulated(ctxt, limit_addr, &desc_ptr.size, sizeof(desc_ptr.size));
+		if (rc != X86EMUL_CONTINUE)
+			return rc;
+		rc = read_emulated(ctxt, base_addr, &desc_ptr.address, base_bytes);
+	}
 	if (rc != X86EMUL_CONTINUE)
 		return rc;
 	if (ctxt->mode == X86EMUL_MODE_PROT64 &&
@@ -4074,8 +4077,8 @@ static const struct opcode group6[] = {
 static const struct group_dual group7 = { {
 	II(Mov | DstMem,			em_sgdt, sgdt),
 	II(Mov | DstMem,			em_sidt, sidt),
-	II(SrcMem | Priv,			em_lgdt, lgdt),
-	II(SrcMem | Priv,			em_lidt, lidt),
+	II(SrcMem | Priv | NoAccess,		em_lgdt, lgdt),
+	II(SrcMem | Priv | NoAccess,		em_lidt, lidt),
 	II(SrcNone | DstMem | Mov,		em_smsw, smsw), N,
 	II(SrcMem16 | Mov | Priv,		em_lmsw, lmsw),
 	II(SrcMem | ByteOp | Priv | NoAccess,	em_invlpg, invlpg),

base-commit: cbe54fc3462b3949ab27b9ae81fb9ddb509040c7

Signature · 1 line
2.53.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue to consider: - [Medium] KVM: x86: split single-page reads in em_lgdt_lidt() miss saved GPA --

Thank you for your contribution! Sashiko AI review found 1 potential issue to consider:
- [Medium] KVM: x86: split single-page reads in em_lgdt_lidt() miss saved GPA
--

commit db71a6d851801384a40299e7d8acce741f81c256
Author: weichenchen <[email protected]>

KVM: x86: Fix stalled descriptor table accesses to MMIO

This commit switches LGDT, LIDT, SGDT and SIDT from standard guest memory accessors to emulated memory accessors so that descriptor table accesses targeting MMIO queue MMIO fragments and complete properly across userspace exits.

Quoted 4 lines
arch/x86/kvm/emulate.c +0 -0
index c1b21282187f3..3bd720d983f8e 100644

[ ... ]

Quoted 27 lines
arch/x86/kvm/emulate.c +15 -4
@@ -3465,14 +3447,35 @@ static int em_sidt(struct x86_emulate_ctxt *ctxt)
 
 static int em_lgdt_lidt(struct x86_emulate_ctxt *ctxt, bool lgdt)
 {
-	struct desc_ptr desc_ptr;
+	struct segmented_address addr = ctxt->src.addr.mem;
+	struct desc_ptr desc_ptr = {};
+	unsigned long limit_addr, base_addr;
+	unsigned int base_bytes, desc_bytes;
 	int rc;
 
 	if (ctxt->mode == X86EMUL_MODE_PROT64)
 		ctxt->op_bytes = 8;
-	rc = read_descriptor(ctxt, ctxt->src.addr.mem,
-			     &desc_ptr.size, &desc_ptr.address,
-			     ctxt->op_bytes);
+	base_bytes = ctxt->op_bytes == 2 ? 3 : ctxt->op_bytes;
+	desc_bytes = sizeof(desc_ptr.size) + base_bytes;
+
+	/* Check both fields before an MMIO read can have side effects. */
+	rc = linearize(ctxt, addr, sizeof(desc_ptr.size), false, &limit_addr);
+	if (rc != X86EMUL_CONTINUE)
+		return rc;
+	addr.ea += sizeof(desc_ptr.size);
+	rc = linearize(ctxt, addr, base_bytes, false, &base_addr);
+	if (rc != X86EMUL_CONTINUE)
+		return rc;

arch/x86/kvm/emulate.c static int em_sidt(struct x86_emulate_ctxt *ctxt) · line 3478

Quoted 10 lines
arch/x86/kvm/emulate.c +10 -0
+
+	/* Defer userspace MMIO until both pages have been translated. */
+	if (offset_in_page(limit_addr) + desc_bytes > PAGE_SIZE) {
+		rc = read_emulated(ctxt, limit_addr, &desc_ptr, desc_bytes);
+	} else {
+		rc = read_emulated(ctxt, limit_addr, &desc_ptr.size, sizeof(desc_ptr.size));
+		if (rc != X86EMUL_CONTINUE)
+			return rc;
+		rc = read_emulated(ctxt, base_addr, &desc_ptr.address, base_bytes);
+	}

[Severity: Medium]
Why split the single-page case into two read_emulated() calls instead of always issuing a single read_emulated(ctxt, limit_addr, &desc_ptr, desc_bytes)?

On a single-page MMIO operand, the first read at limit_addr matches the hardware-saved GPA in emulator_read_write_onepage(), but the second read on resume passes base_addr (limit_addr + 2), which misses the page-offset check:

arch/x86/kvm/x86.c:emulator_read_write_onepage() {
    ...
	if ((ctxt->gpa_access & (write ? ACC_WRITE_MASK : ACC_READ_MASK)) &&
	    emulator_can_use_gpa(ctxt) &&
	    (addr & ~PAGE_MASK) == (ctxt->gpa_val & ~PAGE_MASK)) {
		gpa = ctxt->gpa_val;
		ret = vcpu_is_mmio_gpa(vcpu, addr, gpa, write);
	} else {
		ret = vcpu_mmio_gva_to_gpa(vcpu, addr, &gpa, exception, write);
    ...
}

Because vcpu->arch.mmio_gva is 0 when TDP is enabled, vcpu_mmio_gva_to_gpa() falls back to walking the guest page tables in software after the first 2-byte MMIO read has already exited to userspace. On SEV guests with encrypted guest page tables, that software walk reads ciphertext as PTEs and injects a spurious #PF or translates a bogus GPA; on SMP guests, a concurrent PTE update between the two MMIO exits can fault mid- instruction after the first MMIO read has already taken effect.

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1