proc: use CPUID to determine ZMM_Hi256 region offset (#3831)

The offset of state component i can be found via
CPUID.(EAX=0DH,ECX=i):EBX. The ZMM_Hi256 is state component 6, so we use
CPUID to enumerate the offset instead of hardcoding.

For core dumps, we guess the ZMM_Hi256 offset based on xcr0 and the
length of xsave region. The logic comes from binutils-gdb.

Fixes #3827.

pkg/proc/amd64util/xsave.go

@@ -79,7 +79,6 @@ const (
 	_XSAVE_HEADER_LEN              = 64
 	_XSAVE_EXTENDED_REGION_START   = 576
 	_XSAVE_SSE_REGION_LEN          = 416
-	_XSAVE_AVX512_ZMM_REGION_START = 1152
 )
 
 // AMD64XstateRead reads a byte array containing an XSAVE area into regset.
@@ -87,7 +86,8 @@ const (
 // contents of the legacy region of the XSAVE area.
 // See Section 13.1 (and following) of Intel® 64 and IA-32 Architectures
 // Software Developer’s Manual, Volume 1: Basic Architecture.
-func AMD64XstateRead(xstateargs []byte, readLegacy bool, regset *AMD64Xstate) error {
+// If xstateZMMHi256Offset is zero, it will be guessed.
+func AMD64XstateRead(xstateargs []byte, readLegacy bool, regset *AMD64Xstate, xstateZMMHi256Offset int) error {
 	if _XSAVE_HEADER_START+_XSAVE_HEADER_LEN >= len(xstateargs) {
 		return nil
 	}
@@ -120,7 +120,19 @@ func AMD64XstateRead(xstateargs []byte, readLegacy bool, regset *AMD64Xstate) er
 		return nil
 	}
 
-	avx512state := xstateargs[_XSAVE_AVX512_ZMM_REGION_START:]
+	if xstateZMMHi256Offset == 0 {
+		// Guess ZMM_Hi256 component offset
+		// ref: https://github.com/bminor/binutils-gdb/blob/df89bdf0baf106c3b0a9fae53e4e48607a7f3f87/gdb/i387-tdep.c#L916
+		if xstate_bv&(1<<9) != 0 && len(xstateargs) == 2440 {
+			// AMD CPUs supporting PKRU
+			xstateZMMHi256Offset = 896
+		} else {
+			// Intel CPUs supporting AVX512
+			xstateZMMHi256Offset = 1152
+		}
+	}
+
+	avx512state := xstateargs[xstateZMMHi256Offset:]
 	regset.Avx512State = true
 	copy(regset.ZmmSpace[:], avx512state[:len(regset.ZmmSpace)])
 
@@ -180,7 +192,7 @@ func (xstate *AMD64Xstate) SetXmmRegister(n int, value []byte) error {
 	// Copy bytes [32, 64) to Xsave area
 
 	zmmval := rest
-	zmmpos := _XSAVE_AVX512_ZMM_REGION_START + (n * 32)
+	zmmpos := AMD64XstateZMMHi256Offset() + (n * 32)
 	if zmmpos >= len(xstate.Xsave) {
 		return fmt.Errorf("could not set XMM%d: bytes 32..%d not in XSAVE area", n, 32+len(zmmval))
 	}

pkg/proc/amd64util/xsave_amd64.go

@@ -28,3 +28,28 @@ func AMD64XstateMaxSize() int {
 	})
 	return xstateMaxSize
 }
+
+var xstateZMMHi256Offset int
+var loadXstateZMMHi256OffsetOnce sync.Once
+
+// AMD64XstateZMMHi256Offset probes ZMM_Hi256 offset of the current CPU.  Beware
+// that core dumps may be generated from a different CPU.
+func AMD64XstateZMMHi256Offset() int {
+	loadXstateZMMHi256OffsetOnce.Do(func() {
+		// See Intel 64 and IA-32 Architecture Software Developer's Manual, Vol. 1
+		// chapter 13.2 and Vol. 2A CPUID instruction for a description of all the
+		// magic constants.
+
+		_, _, cx, _ := cpuid(0x01, 0x00)
+
+		if cx&(1<<26) == 0 { // Vol. 2A, Table 3-10, XSAVE enabled bit check
+			// XSAVE not supported by this processor
+			xstateZMMHi256Offset = 0
+			return
+		}
+
+		_, bx, _, _ := cpuid(0x0d, 0x06) // ZMM_Hi256 is component #6
+		xstateZMMHi256Offset = int(bx)
+	})
+	return xstateZMMHi256Offset
+}

pkg/proc/amd64util/xsave_other.go

@@ -5,3 +5,8 @@ package amd64util
 func AMD64XstateMaxSize() int {
 	return _XSTATE_MAX_KNOWN_SIZE
 }
+
+func AMD64XstateZMMHi256Offset() int {
+	// AVX-512 not supported
+	return 0
+}

pkg/proc/core/linux_core.go

@@ -338,7 +338,7 @@ func readNote(r io.ReadSeeker, machineType elf.Machine) (*note, error) {
 	case _NT_X86_XSTATE:
 		if machineType == _EM_X86_64 {
 			var fpregs amd64util.AMD64Xstate
-			if err := amd64util.AMD64XstateRead(desc, true, &fpregs); err != nil {
+			if err := amd64util.AMD64XstateRead(desc, true, &fpregs, 0); err != nil {
 				return nil, err
 			}
 			note.Desc = &fpregs

pkg/proc/native/ptrace_freebsd_amd64.go

@@ -73,7 +73,7 @@ func ptraceGetRegset(id int) (*amd64util.AMD64Xstate, error) {
 	if err != nil {
 		return nil, err
 	}
-	err = amd64util.AMD64XstateRead(regset.Xsave, false, &regset)
+	err = amd64util.AMD64XstateRead(regset.Xsave, false, &regset, amd64util.AMD64XstateZMMHi256Offset())
 	return &regset, err
 }
 

pkg/proc/native/ptrace_linux_386.go

@@ -38,7 +38,7 @@ func ptraceGetRegset(tid int) (regset amd64util.AMD64Xstate, err error) {
 	}
 
 	regset.Xsave = xstateargs[:iov.Len]
-	err = amd64util.AMD64XstateRead(regset.Xsave, false, &regset)
+	err = amd64util.AMD64XstateRead(regset.Xsave, false, &regset, amd64util.AMD64XstateZMMHi256Offset())
 	return
 }
 

pkg/proc/native/ptrace_linux_amd64.go

@@ -37,6 +37,6 @@ func ptraceGetRegset(tid int) (regset amd64util.AMD64Xstate, err error) {
 	}
 
 	regset.Xsave = xstateargs[:iov.Len]
-	err = amd64util.AMD64XstateRead(regset.Xsave, false, &regset)
+	err = amd64util.AMD64XstateRead(regset.Xsave, false, &regset, amd64util.AMD64XstateZMMHi256Offset())
 	return
 }