Skip to contentSkip to conversation

x86/fpu: Fix MXCSR handling and SSE component definition

PATCH Superseded Chang S. Bae 2 participants 8 messages Sep 22, 2022 – Sep 28, 2022

Subsystems (4)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

v2 cover x86/fpu: Fix MXCSR handling and SSE component definition

Hi all,

Changes from v1:
* patch3: revert the function name change and add a warning. Then update the changelog and title.

Thanks Sean Christopherson for the feedback.

== Regression ==

Recently the XSTATE copy functions were unitized together [2]. At a glance, this change appears to relapse the ptrace write on the MXCSR state when the non-compacted format is used in the kernel.

But, this regression appears to root in the XSAVES-enabling code [3] that introduced the XSTATE conversion along with the MXCSR mistreatment.

== MXCSR Hindsight ==

MXCSR is architecturally part of the SSE component. The MXCSR association of XSTATE_BV depends on the XSAVE format.

The change [3], however, presumed MXCSR as part of the X87 component and made the kernel referencing XSTATE_BV regardless of the format.

== Patches ==

* Fix the MXCSR conversion code along with adding the test case.
* Then, fixing MXCSR, one of the other call sites is also updated to exclude legacy states.
* The hard-coded legacy state offset and size are adjusted in the end.

These patches can be also found in this repository:
git://github.com/intel/amx-linux.git mxcsr

Thanks,
Chang

[1] v1: https://lore.kernel.org/lkml/[email protected]/
[2] Commit 43be46e89698 ("x86/fpu: Sanitize xstateregs_set()")
[3] Commit 91c3dba7dbc1 ("x86/fpu/xstate: Fix PTRACE frames for XSAVES")

Chang S. Bae (4):
x86/fpu: Fix the MXCSR state reshuffling between userspace and kernel
buffers
selftests/x86/mxcsr: Test the MXCSR state write via ptrace x86/fpu: Disallow legacy states from fpstate_clear_xstate_component() x86/fpu: Correct the legacy state offset and size information

 arch/x86/kernel/fpu/xstate.c         | 100 ++++++++++----
 tools/testing/selftests/x86/Makefile |   2 +-
 tools/testing/selftests/x86/mxcsr.c  | 200 +++++++++++++++++++++++++++
 3 files changed, 274 insertions(+), 28 deletions(-)
 create mode 100644 tools/testing/selftests/x86/mxcsr.c

base-commit: 08ed00508bc1fa0a0bc6dd2420f982b55051de23

Signature · 1 line
2.17.1
#1 lore raw

PATCH v2 1/4 x86/fpu: Fix the MXCSR state reshuffling between userspace and kernel buffers

== Hardware Background ==

The MXCSR state, as part of the SSE component, has different treatments with XSAVE*/XRSTOR* between the two XSAVE formats:

- When the MXCSR state is XSAVEd in the non-compacted format, the feature bit in XSTATE_BV pertains to the XMM registers. XRSTOR restores the MXCSR state without referencing XSTATE_BV.

- But, on XSAVE* with the compacted format, the SSE bit is set in XSTATE_BV if MXCSR is not in the init state on XSAVE*. Then, on XRSTOR*, the state is restored only when the SSE bit is set in XSTATE_BV.

== Regression ==

The XSTATE copy routine between userspace and kernel buffers used to be separate for different XSAVE formats. Commit 43be46e89698 ("x86/fpu:
Sanitize xstateregs_set()") combined them together.

This change appears to be a regression on XSAVES-less systems. But, the merged code is based on the original conversion code with commit 91c3dba7dbc1 ("x86/fpu/xstate: Fix PTRACE frames for XSAVES").

That has such oversight as:
- Mistreating MXCSR as part of the FP state instead of the SSE component.
- Taking the SSE bit in XSTATE_BV always for all the SSE states.

== Correction ==

Update the XSTATE conversion code:
- Refactor the copy routine for legacy states. Treat MXCSR as part of SSE.
- Copying MXCSR, reference XSTATE_BV only with the compacted format.
- Also, flip the SSE bit in XSTATE_BV accordingly to the format.

---
 arch/x86/kernel/fpu/xstate.c | 70 +++++++++++++++++++++++++-----------
 1 file changed, 49 insertions(+), 21 deletions(-)
arch/x86/kernel/fpu/xstate.c +49 -21
index c8340156bfd2..d7676cfc32eb 100644
@@ -1064,6 +1064,7 @@ void __copy_xstate_to_uabi_buf(struct membuf to, struct fpstate *fpstate,
 			       u32 pkru_val, enum xstate_copy_mode copy_mode)
 {
 	const unsigned int off_mxcsr = offsetof(struct fxregs_state, mxcsr);
+	bool compacted = cpu_feature_enabled(X86_FEATURE_XCOMPACTED);
 	struct xregs_state *xinit = &init_fpstate.regs.xsave;
 	struct xregs_state *xsave = &fpstate->regs.xsave;
 	struct xstate_header header;
@@ -1093,8 +1094,13 @@ void __copy_xstate_to_uabi_buf(struct membuf to, struct fpstate *fpstate,
 	copy_feature(header.xfeatures & XFEATURE_MASK_FP, &to, &xsave->i387,
 		     &xinit->i387, off_mxcsr);
 
-	/* Copy MXCSR when SSE or YMM are set in the feature mask */
-	copy_feature(header.xfeatures & (XFEATURE_MASK_SSE | XFEATURE_MASK_YMM),
+	/*
+	 * Copy MXCSR depending on the XSAVE format. If compacted,
+	 * reference the feature mask. Otherwise, check if any of related
+	 * features is valid.
+	 */
+	copy_feature(compacted ? header.xfeatures & XFEATURE_MASK_SSE :
+		     fpstate->user_xfeatures & (XFEATURE_MASK_SSE | XFEATURE_MASK_YMM),
 		     &to, &xsave->i387.mxcsr, &xinit->i387.mxcsr,
 		     MXCSR_AND_FLAGS_SIZE);
 
@@ -1199,6 +1205,11 @@ static int copy_from_buffer(void *dst, unsigned int offset, unsigned int size,
 static int copy_uabi_to_xstate(struct fpstate *fpstate, const void *kbuf,
 			       const void __user *ubuf)
 {
+	const unsigned int off_stspace = offsetof(struct fxregs_state, st_space);
+	const unsigned int off_xmm = offsetof(struct fxregs_state, xmm_space);
+	const unsigned int off_mxcsr = offsetof(struct fxregs_state, mxcsr);
+	bool compacted = cpu_feature_enabled(X86_FEATURE_XCOMPACTED);
+	struct fxregs_state *fxsave = &fpstate->regs.fxsave;
 	struct xregs_state *xsave = &fpstate->regs.xsave;
 	unsigned int offset, size;
 	struct xstate_header hdr;
@@ -1212,38 +1223,48 @@ static int copy_uabi_to_xstate(struct fpstate *fpstate, const void *kbuf,
 	if (validate_user_xstate_header(&hdr, fpstate))
 		return -EINVAL;
 
-	/* Validate MXCSR when any of the related features is in use */
-	mask = XFEATURE_MASK_FP | XFEATURE_MASK_SSE | XFEATURE_MASK_YMM;
-	if (hdr.xfeatures & mask) {
+	if (hdr.xfeatures & XFEATURE_MASK_FP) {
+		if (copy_from_buffer(fxsave, 0, off_mxcsr, kbuf, ubuf))
+			return -EINVAL;
+		if (copy_from_buffer(fxsave->st_space, off_stspace, sizeof(fxsave->st_space),
+				     kbuf, ubuf))
+			return -EINVAL;
+	}
+
+	/* Validate MXCSR when any of the related features is valid. */
+	mask = XFEATURE_MASK_SSE | XFEATURE_MASK_YMM;
+	if (fpstate->user_xfeatures & mask) {
 		u32 mxcsr[2];
 
-		offset = offsetof(struct fxregs_state, mxcsr);
-		if (copy_from_buffer(mxcsr, offset, sizeof(mxcsr), kbuf, ubuf))
+		if (copy_from_buffer(mxcsr, off_mxcsr, sizeof(mxcsr), kbuf, ubuf))
 			return -EFAULT;
 
 		/* Reserved bits in MXCSR must be zero. */
 		if (mxcsr[0] & ~mxcsr_feature_mask)
 			return -EINVAL;
 
-		/* SSE and YMM require MXCSR even when FP is not in use. */
-		if (!(hdr.xfeatures & XFEATURE_MASK_FP)) {
-			xsave->i387.mxcsr = mxcsr[0];
-			xsave->i387.mxcsr_mask = mxcsr[1];
-		}
+		/*
+		 * Copy MXCSR regardless of the feature mask as userspace
+		 * uses the uncompacted format.
+		 */
+		fxsave->mxcsr = mxcsr[0];
+		fxsave->mxcsr_mask = mxcsr[1];
 	}
 
-	for (i = 0; i < XFEATURE_MAX; i++) {
-		mask = BIT_ULL(i);
+	if (hdr.xfeatures & XFEATURE_MASK_SSE) {
+		if (copy_from_buffer(fxsave->xmm_space, off_xmm, sizeof(fxsave->xmm_space),
+				     kbuf, ubuf))
+			return -EINVAL;
+	}
 
-		if (hdr.xfeatures & mask) {
-			void *dst = __raw_xsave_addr(xsave, i);
+	for_each_extended_xfeature(i, hdr.xfeatures) {
+		void *dst = __raw_xsave_addr(xsave, i);
 
-			offset = xstate_offsets[i];
-			size = xstate_sizes[i];
+		offset = xstate_offsets[i];
+		size = xstate_sizes[i];
 
-			if (copy_from_buffer(dst, offset, size, kbuf, ubuf))
-				return -EFAULT;
-		}
+		if (copy_from_buffer(dst, offset, size, kbuf, ubuf))
+			return -EFAULT;
 	}
 
 	/*
@@ -1256,6 +1277,13 @@ static int copy_uabi_to_xstate(struct fpstate *fpstate, const void *kbuf,
 	 * Add back in the features that came in from userspace:
 	 */
 	xsave->header.xfeatures |= hdr.xfeatures;
+	/*
+	 * Convert the SSE bit in the feature mask as it implies
+	 * differently between the formats. It indicates the MXCSR state
+	 * if compacted; otherwise, it pertains to XMM registers.
+	 */
+	if (compacted && fxsave->mxcsr != MXCSR_DEFAULT)
+		xsave->header.xfeatures |= XFEATURE_MASK_SSE;
 
 	return 0;
 }
Signature · 1 line
2.17.1
#2 lore raw

PATCH v2 2/4 selftests/x86/mxcsr: Test the MXCSR state write via ptrace

In reply to Chang S. Bae

The ptrace buffer is in the non-compacted format. The MXCSR state should be written to the target thread when either SSE or AVX component is enabled.

Write an MXCSR value to the target and read back. Then it is validated with the XRSTOR/XSAVE result on the current.

---

If this is acceptable, I will also follow up to move some of the helper functions to a .h file from this and other test cases because duplicating what is shareable should be avoided.

---
 tools/testing/selftests/x86/Makefile |   2 +-
 tools/testing/selftests/x86/mxcsr.c  | 200 +++++++++++++++++++++++++++
 2 files changed, 201 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/selftests/x86/mxcsr.c
tools/testing/selftests/x86/Makefile +1 -1
index 0388c4d60af0..621c47960be3 100644
@@ -13,7 +13,7 @@ CAN_BUILD_WITH_NOPIE := $(shell ./check_cc.sh "$(CC)" trivial_program.c -no-pie)
 TARGETS_C_BOTHBITS := single_step_syscall sysret_ss_attrs syscall_nt test_mremap_vdso \
 			check_initial_reg_state sigreturn iopl ioperm \
 			test_vsyscall mov_ss_trap \
-			syscall_arg_fault fsgsbase_restore sigaltstack
+			syscall_arg_fault fsgsbase_restore sigaltstack mxcsr
 TARGETS_C_32BIT_ONLY := entry_from_vm86 test_syscall_vdso unwind_vdso \
 			test_FCMOV test_FCOMI test_FISTTP \
 			vdso_restorer
tools/testing/selftests/x86/mxcsr.c new file +200 -0

200 lines not on this page · raw message

Signature · 1 line
2.17.1
#3 lore raw

PATCH v2 3/4 x86/fpu: Disallow legacy states from fpstate_clear_xstate_component()

In reply to Chang S. Bae

Commit 087df48c298c ("x86/fpu: Replace KVMs xstate component clearing") refactored the MPX state clearing code.

But, legacy states are not warranted in this routine:
- It presumes every state is contiguous but that's not true for the legacy states. While MXCSR belongs to SSE, the state is located in the XSAVE buffer as surrounded by FP states.
- Also, zeroing out legacy states is not meaningful as their init state is non-zero.

It is possible to adjust the code to support them. Then, there is no use for clearing legacy states yet. To make it simple, explicitly disallow legacy states.

---
Changes from v1 (Sean Christopherson):
* Revert the name change.
* Add a warning.
* Update title/changelog.

---
 arch/x86/kernel/fpu/xstate.c | 9 +++++++++
 1 file changed, 9 insertions(+)
arch/x86/kernel/fpu/xstate.c +9 -0
index d7676cfc32eb..a3f7045d1f8e 100644
@@ -1375,6 +1375,15 @@ void fpstate_clear_xstate_component(struct fpstate *fps, unsigned int xfeature)
 {
 	void *addr = get_xsave_addr(&fps->regs.xsave, xfeature);
 
+	/*
+	 * Allow extended states only, because:
+	 * (1) Each legacy state is not contiguously located in the buffer.
+	 * (2) Zeroing those states is not meaningful as their init states
+	 *     are not zero.
+	 */
+	if (WARN_ON_ONCE(xfeature <= XFEATURE_SSE))
+		return;
+
 	if (addr)
 		memset(addr, 0, xstate_sizes[xfeature]);
 }
Signature · 1 line
2.17.1
#4 lore raw

PATCH v2 4/4 x86/fpu: Correct the legacy state offset and size information

In reply to Chang S. Bae

MXCSR is architecturally part of the SSE state. But, the kernel code presumes it as part of the FP component. Adjust the offset and size for these legacy states.

Notably, each legacy component area is not contiguous, unlike extended components. Add a warning message when these offset and size are referenced.

Fixes ac73b27aea4e ("x86/fpu/xstate: Fix xstate_offsets, xstate_sizes for non-extended xstates")
---
 arch/x86/kernel/fpu/xstate.c | 21 +++++++++++++++------
 1 file changed, 15 insertions(+), 6 deletions(-)
arch/x86/kernel/fpu/xstate.c +15 -6
index a3f7045d1f8e..ac2ec5d6e7e4 100644
@@ -143,8 +143,13 @@ static unsigned int xfeature_get_offset(u64 xcomp_bv, int xfeature)
 	 * offsets.
 	 */
 	if (!cpu_feature_enabled(X86_FEATURE_XCOMPACTED) ||
-	    xfeature <= XFEATURE_SSE)
+	    xfeature <= XFEATURE_SSE) {
+		if (xfeature <= XFEATURE_SSE)
+			pr_warn("The legacy state (%d) is discontiguously located.\n",
+				xfeature);
+
 		return xstate_offsets[xfeature];
+	}
 
 	/*
 	 * Compacted format offsets depend on the actual content of the
@@ -217,14 +222,18 @@ static void __init setup_xstate_cache(void)
 	 * The FP xstates and SSE xstates are legacy states. They are always
 	 * in the fixed offsets in the xsave area in either compacted form
 	 * or standard form.
+	 *
+	 * But, while MXCSR is part of the SSE state, it is located in
+	 * between the FP states. Note that it is erroneous assuming that
+	 * each legacy area is contiguous.
 	 */
 	xstate_offsets[XFEATURE_FP]	= 0;
-	xstate_sizes[XFEATURE_FP]	= offsetof(struct fxregs_state,
-						   xmm_space);
+	xstate_sizes[XFEATURE_FP]	= offsetof(struct fxregs_state, mxcsr) +
+					  sizeof_field(struct fxregs_state, st_space);
 
-	xstate_offsets[XFEATURE_SSE]	= xstate_sizes[XFEATURE_FP];
-	xstate_sizes[XFEATURE_SSE]	= sizeof_field(struct fxregs_state,
-						       xmm_space);
+	xstate_offsets[XFEATURE_SSE]	= offsetof(struct fxregs_state, mxcsr);
+	xstate_sizes[XFEATURE_SSE]	= MXCSR_AND_FLAGS_SIZE +
+					  sizeof_field(struct fxregs_state, xmm_space);
 
 	for_each_extended_xfeature(i, fpu_kernel_cfg.max_features) {
 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
Signature · 1 line
2.17.1
#5 lore raw

Re: [PATCH v2 4/4] x86/fpu: Correct the legacy state offset and size information

Quoted 21 lines · Chang S. Bae

MXCSR is architecturally part of the SSE state. But, the kernel code presumes it as part of the FP component. Adjust the offset and size for these legacy states.

Notably, each legacy component area is not contiguous, unlike extended components. Add a warning message when these offset and size are referenced.

Fixes ac73b27aea4e ("x86/fpu/xstate: Fix xstate_offsets, xstate_sizes for non-extended xstates")
---
 arch/x86/kernel/fpu/xstate.c | 21 +++++++++++++++------
 1 file changed, 15 insertions(+), 6 deletions(-)
arch/x86/kernel/fpu/xstate.c +0 -0
index a3f7045d1f8e..ac2ec5d6e7e4 100644

arch/x86/kernel/fpu/xstate.c static unsigned int xfeature_get_offset(u64 xcomp_bv, int xfeature) · line 149

Quoted 8 lines
arch/x86/kernel/fpu/xstate.c +4 -1
@@ -143,8 +143,13 @@ static unsigned int xfeature_get_offset(u64 xcomp_bv, int xfeature)
 	 * offsets.
 	 */
 	if (!cpu_feature_enabled(X86_FEATURE_XCOMPACTED) ||
-	    xfeature <= XFEATURE_SSE)
+	    xfeature <= XFEATURE_SSE) {
+		if (xfeature <= XFEATURE_SSE)
+			pr_warn("The legacy state (%d) is discontiguously located.\n",
+				xfeature);

pr_warn() here isn't warranted. copy_uabi_to_xstate() calls this with non-extended features, which is perfectly fine since it manually handles MXCSR. And that helper is directly reachable by userspace, i.e. userspace can spam the pr_warn().

#6 lore raw

Re: [PATCH v2 4/4] x86/fpu: Correct the legacy state offset and size information

Quoted 18 lines · Sean Christopherson
Quoted 6 lines · Chang S. Bae
arch/x86/kernel/fpu/xstate.c +0 -0
index a3f7045d1f8e..ac2ec5d6e7e4 100644

arch/x86/kernel/fpu/xstate.c static unsigned int xfeature_get_offset(u64 xcomp_bv, int xfeature) · line 149

Quoted 8 lines
arch/x86/kernel/fpu/xstate.c +4 -1
@@ -143,8 +143,13 @@ static unsigned int xfeature_get_offset(u64 xcomp_bv, int xfeature)
  	 * offsets.
  	 */
  	if (!cpu_feature_enabled(X86_FEATURE_XCOMPACTED) ||
-	    xfeature <= XFEATURE_SSE)
+	    xfeature <= XFEATURE_SSE) {
+		if (xfeature <= XFEATURE_SSE)
+			pr_warn("The legacy state (%d) is discontiguously located.\n",
+				xfeature);

pr_warn() here isn't warranted. copy_uabi_to_xstate() calls this with non-extended features,

I think patch1 makes changes not to call this for legacy features anymore.

Quoted 2 lines

which is perfectly fine since it manually handles MXCSR. And that helper is directly reachable by userspace, i.e. userspace can spam the pr_warn().

I don't think I get your point. I assume that helper is __raw_xsave_addr(). But then I'm missing how it can be directly reached by userspace.

Thanks,
Chang

#7 lore raw

Re: [PATCH v2 4/4] x86/fpu: Correct the legacy state offset and size information

Quoted 20 lines · Chang S. Bae
Quoted 18 lines · Sean Christopherson
Quoted 6 lines · Chang S. Bae
arch/x86/kernel/fpu/xstate.c +0 -0
index a3f7045d1f8e..ac2ec5d6e7e4 100644

arch/x86/kernel/fpu/xstate.c static unsigned int xfeature_get_offset(u64 xcomp_bv, int xfeature) · line 149

Quoted 8 lines
arch/x86/kernel/fpu/xstate.c +4 -1
@@ -143,8 +143,13 @@ static unsigned int xfeature_get_offset(u64 xcomp_bv, int xfeature)
  	 * offsets.
  	 */
  	if (!cpu_feature_enabled(X86_FEATURE_XCOMPACTED) ||
-	    xfeature <= XFEATURE_SSE)
+	    xfeature <= XFEATURE_SSE) {
+		if (xfeature <= XFEATURE_SSE)
+			pr_warn("The legacy state (%d) is discontiguously located.\n",
+				xfeature);

pr_warn() here isn't warranted. copy_uabi_to_xstate() calls this with non-extended features,

Quoted 1 line

I think patch1 makes changes not to call this for legacy features anymore.

Oh, even better! In that case, drop patch 3 and WARN here, because with that call site gone, all paths that lead to xfeature_get_offset() avoid calling it with legacy features.

The comment above this probably needs to be updated too.

I was going to make that suggestion in my original response, but I didn't apply the series and so didn't see that that last remaining wart was fixed. Thanks, and sorry for the runaround!

Quoted 5 lines
Quoted 2 lines

which is perfectly fine since it manually handles MXCSR. And that helper is directly reachable by userspace, i.e. userspace can spam the pr_warn().

I don't think I get your point. I assume that helper is __raw_xsave_addr(). But then I'm missing how it can be directly reached by userspace.

ioctl(KVM_GET_XSAVE) can reach this code, i.e. if there were a bug then userspace could invoke that ioctl() in a tight loop to spam the kernel log.