Improve support for XSAVE with debuggers.

- Dump an NT_X86_XSTATE note if XSAVE is in use. This note is designed
  to match what Linux does in that 1) it dumps the entire XSAVE area
  including the fxsave state, and 2) it stashes a copy of the current
  xsave mask in the unused padding between the fxsave state and the
  xstate header at the same location used by Linux.
- Teach readelf() to recognize NT_X86_XSTATE notes.
- Change PT_GET/SETXSTATE to take the entire XSAVE state instead of
  only the extra portion. This avoids having to always make two
  ptrace() calls to get or set the full XSAVE state.
- Add a PT_GET_XSTATE_INFO which returns the length of the current
  XSTATE save area (so the size of the buffer needed for PT_GETXSTATE)
  and the current XSAVE mask (%xcr0).

Differential Revision:	https://reviews.freebsd.org/D1193
Reviewed by:	kib
MFC after:	2 weeks

15 files changed, 259 insertions, 38 deletions

diff --git a/contrib/binutils/binutils/readelf.c b/contrib/binutils/binutils/readelf.c
index aafd8d4502e9..09f4062a4d9f 100644
--- a/contrib/binutils/binutils/readelf.c
+++ b/contrib/binutils/binutils/readelf.c
@@ -9159,6 +9159,8 @@ get_freebsd_note_type (unsigned e_type)
 	return _("NT_PROCSTAT_PSSTRINGS (ps_strings data)");
       case NT_PROCSTAT_AUXV:
 	return _("NT_PROCSTAT_AUXV (auxv data)");
+      case NT_X86_XSTATE:
+	return _("NT_X86_XSTATE (x86 XSAVE extended state)");
       default:
 	return get_note_type(e_type);
       }
diff --git a/contrib/binutils/include/elf/common.h b/contrib/binutils/include/elf/common.h
index d4d9d7b0cc7a..157922af3f0b 100644
--- a/contrib/binutils/include/elf/common.h
+++ b/contrib/binutils/include/elf/common.h
@@ -414,6 +414,7 @@
 #define NT_PROCSTAT_OSREL	14
 #define NT_PROCSTAT_PSSTRINGS	15
 #define NT_PROCSTAT_AUXV	16
+#define	NT_X86_XSTATE		0x202
 
 
 /* Note segments for core files on NetBSD systems.  Note name
diff --git a/sys/amd64/amd64/elf_machdep.c b/sys/amd64/amd64/elf_machdep.c
index fdc4d56308dd..23fa39b4fd1e 100644
--- a/sys/amd64/amd64/elf_machdep.c
+++ b/sys/amd64/amd64/elf_machdep.c
@@ -44,6 +44,7 @@ __FBSDID("$FreeBSD$");
 #include <vm/vm_param.h>
 
 #include <machine/elf.h>
+#include <machine/fpu.h>
 #include <machine/md_var.h>
 
 struct sysentvec elf64_freebsd_sysvec = {
@@ -133,12 +134,27 @@ SYSINIT(kelf64, SI_SUB_EXEC, SI_ORDER_ANY,
 	&kfreebsd_brand_info);
 
 void
-elf64_dump_thread(struct thread *td __unused, void *dst __unused,
-    size_t *off __unused)
+elf64_dump_thread(struct thread *td, void *dst, size_t *off)
 {
+	void *buf;
+	size_t len;
+
+	len = 0;
+	if (use_xsave) {
+		if (dst != NULL) {
+			fpugetregs(td);
+			len += elf64_populate_note(NT_X86_XSTATE,
+			    get_pcb_user_save_td(td), dst,
+			    cpu_max_ext_state_size, &buf);
+			*(uint64_t *)((char *)buf + X86_XSTATE_XCR0_OFFSET) =
+			    xsave_mask;
+		} else
+			len += elf64_populate_note(NT_X86_XSTATE, NULL, NULL,
+			    cpu_max_ext_state_size, NULL);
+	}
+	*off = len;
 }
 
-
 /* Process one elf relocation with addend. */
 static int
 elf_reloc_internal(linker_file_t lf, Elf_Addr relocbase, const void *data,
diff --git a/sys/amd64/amd64/fpu.c b/sys/amd64/amd64/fpu.c
index b0f25ba244c8..f30c07370374 100644
--- a/sys/amd64/amd64/fpu.c
+++ b/sys/amd64/amd64/fpu.c
@@ -127,6 +127,13 @@ CTASSERT(sizeof(struct savefpu_ymm) == 832);
  */
 CTASSERT(sizeof(struct pcb) % XSAVE_AREA_ALIGN == 0);
 
+/*
+ * Ensure the copy of XCR0 saved in a core is contained in the padding
+ * area.
+ */
+CTASSERT(X86_XSTATE_XCR0_OFFSET >= offsetof(struct savefpu, sv_pad) &&
+    X86_XSTATE_XCR0_OFFSET + sizeof(uint64_t) <= sizeof(struct savefpu));
+
 static	void	fpu_clean_state(void);
 
 SYSCTL_INT(_hw, HW_FLOATINGPT, floatingpoint, CTLFLAG_RD,
diff --git a/sys/amd64/amd64/ptrace_machdep.c b/sys/amd64/amd64/ptrace_machdep.c
index 9fa1917d8bd4..c96fe263db92 100644
--- a/sys/amd64/amd64/ptrace_machdep.c
+++ b/sys/amd64/amd64/ptrace_machdep.c
@@ -42,6 +42,7 @@ __FBSDID("$FreeBSD$");
 static int
 cpu_ptrace_xstate(struct thread *td, int req, void *addr, int data)
 {
+	struct ptrace_xstate_info info;
 	char *savefpu;
 	int error;
 
@@ -49,14 +50,14 @@ cpu_ptrace_xstate(struct thread *td, int req, void *addr, int data)
 		return (EOPNOTSUPP);
 
 	switch (req) {
-	case PT_GETXSTATE:
+	case PT_GETXSTATE_OLD:
 		fpugetregs(td);
 		savefpu = (char *)(get_pcb_user_save_td(td) + 1);
 		error = copyout(savefpu, addr,
 		    cpu_max_ext_state_size - sizeof(struct savefpu));
 		break;
 
-	case PT_SETXSTATE:
+	case PT_SETXSTATE_OLD:
 		if (data > cpu_max_ext_state_size - sizeof(struct savefpu)) {
 			error = EINVAL;
 			break;
@@ -70,6 +71,36 @@ cpu_ptrace_xstate(struct thread *td, int req, void *addr, int data)
 		free(savefpu, M_TEMP);
 		break;
 
+	case PT_GETXSTATE_INFO:
+		if (data != sizeof(info)) {
+			error  = EINVAL;
+			break;
+		}
+		info.xsave_len = cpu_max_ext_state_size;
+		info.xsave_mask = xsave_mask;
+		error = copyout(&info, addr, data);
+		break;
+
+	case PT_GETXSTATE:
+		fpugetregs(td);
+		savefpu = (char *)(get_pcb_user_save_td(td));
+		error = copyout(savefpu, addr, cpu_max_ext_state_size);
+		break;
+
+	case PT_SETXSTATE:
+		if (data > cpu_max_ext_state_size) {
+			error = EINVAL;
+			break;
+		}
+		savefpu = malloc(data, M_TEMP, M_WAITOK);
+		error = copyin(addr, savefpu, data);
+		if (error == 0)
+			error = fpusetregs(td, (struct savefpu *)savefpu,
+			    savefpu + sizeof(struct savefpu), data -
+			    sizeof(struct savefpu));
+		free(savefpu, M_TEMP);
+		break;
+
 	default:
 		error = EINVAL;
 		break;
@@ -81,8 +112,6 @@ cpu_ptrace_xstate(struct thread *td, int req, void *addr, int data)
 #ifdef COMPAT_FREEBSD32
 #define PT_I386_GETXMMREGS	(PT_FIRSTMACH + 0)
 #define PT_I386_SETXMMREGS	(PT_FIRSTMACH + 1)
-#define PT_I386_GETXSTATE	(PT_FIRSTMACH + 2)
-#define PT_I386_SETXSTATE	(PT_FIRSTMACH + 3)
 
 static int
 cpu32_ptrace(struct thread *td, int req, void *addr, int data)
@@ -104,12 +133,12 @@ cpu32_ptrace(struct thread *td, int req, void *addr, int data)
 		fpstate->sv_env.en_mxcsr &= cpu_mxcsr_mask;
 		break;
 
-	case PT_I386_GETXSTATE:
-		error = cpu_ptrace_xstate(td, PT_GETXSTATE, addr, data);
-		break;
-
-	case PT_I386_SETXSTATE:
-		error = cpu_ptrace_xstate(td, PT_SETXSTATE, addr, data);
+	case PT_GETXSTATE_OLD:
+	case PT_SETXSTATE_OLD:
+	case PT_GETXSTATE_INFO:
+	case PT_GETXSTATE:
+	case PT_SETXSTATE:
+		error = cpu_ptrace_xstate(td, req, addr, data);
 		break;
 
 	default:
@@ -131,13 +160,16 @@ cpu_ptrace(struct thread *td, int req, void *addr, int data)
 		return (cpu32_ptrace(td, req, addr, data));
 #endif
 
-	/* Support old values of PT_GETXSTATE and PT_SETXSTATE. */
+	/* Support old values of PT_GETXSTATE_OLD and PT_SETXSTATE_OLD. */
 	if (req == PT_FIRSTMACH + 0)
-		req = PT_GETXSTATE;
+		req = PT_GETXSTATE_OLD;
 	if (req == PT_FIRSTMACH + 1)
-		req = PT_SETXSTATE;
+		req = PT_SETXSTATE_OLD;
 
 	switch (req) {
+	case PT_GETXSTATE_OLD:
+	case PT_SETXSTATE_OLD:
+	case PT_GETXSTATE_INFO:
 	case PT_GETXSTATE:
 	case PT_SETXSTATE:
 		error = cpu_ptrace_xstate(td, req, addr, data);
diff --git a/sys/compat/ia32/ia32_sysvec.c b/sys/compat/ia32/ia32_sysvec.c
index 770bbbf9e462..6dd02aff96a5 100644
--- a/sys/compat/ia32/ia32_sysvec.c
+++ b/sys/compat/ia32/ia32_sysvec.c
@@ -187,9 +187,25 @@ SYSINIT(kia32, SI_SUB_EXEC, SI_ORDER_ANY,
 	&kia32_brand_info);
 
 void
-elf32_dump_thread(struct thread *td __unused, void *dst __unused,
-    size_t *off __unused)
+elf32_dump_thread(struct thread *td, void *dst, size_t *off)
 {
+	void *buf;
+	size_t len;
+
+	len = 0;
+	if (use_xsave) {
+		if (dst != NULL) {
+			fpugetregs(td);
+			len += elf32_populate_note(NT_X86_XSTATE,
+			    get_pcb_user_save_td(td), dst,
+			    cpu_max_ext_state_size, &buf);
+			*(uint64_t *)((char *)buf + X86_XSTATE_XCR0_OFFSET) =
+			    xsave_mask;
+		} else
+			len += elf32_populate_note(NT_X86_XSTATE, NULL, NULL,
+			    cpu_max_ext_state_size, NULL);
+	}
+	*off = len;
 }
 
 void
diff --git a/sys/i386/i386/elf_machdep.c b/sys/i386/i386/elf_machdep.c
index 034b4c47d877..cd701aa3dd7c 100644
--- a/sys/i386/i386/elf_machdep.c
+++ b/sys/i386/i386/elf_machdep.c
@@ -45,6 +45,7 @@ __FBSDID("$FreeBSD$");
 
 #include <machine/elf.h>
 #include <machine/md_var.h>
+#include <machine/npx.h>
 
 struct sysentvec elf32_freebsd_sysvec = {
 	.sv_size	= SYS_MAXSYSCALL,
@@ -134,12 +135,27 @@ SYSINIT(kelf32, SI_SUB_EXEC, SI_ORDER_ANY,
 
 
 void
-elf32_dump_thread(struct thread *td __unused, void *dst __unused,
-    size_t *off __unused)
+elf32_dump_thread(struct thread *td, void *dst, size_t *off)
 {
+	void *buf;
+	size_t len;
+
+	len = 0;
+	if (use_xsave) {
+		if (dst != NULL) {
+			npxgetregs(td);
+			len += elf32_populate_note(NT_X86_XSTATE,
+			    get_pcb_user_save_td(td), dst,
+			    cpu_max_ext_state_size, &buf);
+			*(uint64_t *)((char *)buf + X86_XSTATE_XCR0_OFFSET) =
+			    xsave_mask;
+		} else
+			len += elf32_populate_note(NT_X86_XSTATE, NULL, NULL,
+			    cpu_max_ext_state_size, NULL);
+	}
+	*off = len;
 }
 
-
 /* Process one elf relocation with addend. */
 static int
 elf_reloc_internal(linker_file_t lf, Elf_Addr relocbase, const void *data,
diff --git a/sys/i386/i386/ptrace_machdep.c b/sys/i386/i386/ptrace_machdep.c
index 870d69b52a56..5cfeb4c299c7 100644
--- a/sys/i386/i386/ptrace_machdep.c
+++ b/sys/i386/i386/ptrace_machdep.c
@@ -46,6 +46,7 @@ __FBSDID("$FreeBSD$");
 static int
 cpu_ptrace_xstate(struct thread *td, int req, void *addr, int data)
 {
+	struct ptrace_xstate_info info;
 	char *savefpu;
 	int error;
 
@@ -53,14 +54,14 @@ cpu_ptrace_xstate(struct thread *td, int req, void *addr, int data)
 		return (EOPNOTSUPP);
 
 	switch (req) {
-	case PT_GETXSTATE:
+	case PT_GETXSTATE_OLD:
 		npxgetregs(td);
 		savefpu = (char *)(get_pcb_user_save_td(td) + 1);
 		error = copyout(savefpu, addr,
 		    cpu_max_ext_state_size - sizeof(union savefpu));
 		break;
 
-	case PT_SETXSTATE:
+	case PT_SETXSTATE_OLD:
 		if (data > cpu_max_ext_state_size - sizeof(union savefpu)) {
 			error = EINVAL;
 			break;
@@ -74,6 +75,36 @@ cpu_ptrace_xstate(struct thread *td, int req, void *addr, int data)
 		free(savefpu, M_TEMP);
 		break;
 
+	case PT_GETXSTATE_INFO:
+		if (data != sizeof(info)) {
+			error  = EINVAL;
+			break;
+		}
+		info.xsave_len = cpu_max_ext_state_size;
+		info.xsave_mask = xsave_mask;
+		error = copyout(&info, addr, data);
+		break;
+
+	case PT_GETXSTATE:
+		npxgetregs(td);
+		savefpu = (char *)(get_pcb_user_save_td(td));
+		error = copyout(savefpu, addr, cpu_max_ext_state_size);
+		break;
+
+	case PT_SETXSTATE:
+		if (data > cpu_max_ext_state_size) {
+			error = EINVAL;
+			break;
+		}
+		savefpu = malloc(data, M_TEMP, M_WAITOK);
+		error = copyin(addr, savefpu, data);
+		if (error == 0)
+			error = npxsetregs(td, (union savefpu *)savefpu,
+			    savefpu + sizeof(union savefpu), data -
+			    sizeof(union savefpu));
+		free(savefpu, M_TEMP);
+		break;
+
 	default:
 		error = EINVAL;
 		break;
@@ -106,6 +137,9 @@ cpu_ptrace(struct thread *td, int req, void *addr, int data)
 		fpstate->sv_env.en_mxcsr &= cpu_mxcsr_mask;
 		break;
 
+	case PT_GETXSTATE_OLD:
+	case PT_SETXSTATE_OLD:
+	case PT_GETXSTATE_INFO:
 	case PT_GETXSTATE:
 	case PT_SETXSTATE:
 		error = cpu_ptrace_xstate(td, req, addr, data);
diff --git a/sys/i386/isa/npx.c b/sys/i386/isa/npx.c
index 0034f2d674dd..8cc81185394b 100644
--- a/sys/i386/isa/npx.c
+++ b/sys/i386/isa/npx.c
@@ -201,6 +201,13 @@ CTASSERT(sizeof(struct savefpu_ymm) == 832);
  */
 CTASSERT(sizeof(struct pcb) % XSAVE_AREA_ALIGN == 0);
 
+/*
+ * Ensure the copy of XCR0 saved in a core is contained in the padding
+ * area.
+ */
+CTASSERT(X86_XSTATE_XCR0_OFFSET >= offsetof(struct savexmm, sv_pad) &&
+    X86_XSTATE_XCR0_OFFSET + sizeof(uint64_t) <= sizeof(struct savexmm));
+
 static	void	fpu_clean_state(void);
 #endif
 
diff --git a/sys/kern/imgact_elf.c b/sys/kern/imgact_elf.c
index e3494f69d07b..0bdd6edee8c5 100644
--- a/sys/kern/imgact_elf.c
+++ b/sys/kern/imgact_elf.c
@@ -1587,7 +1587,50 @@ register_note(struct note_info_list *list, int type, outfunc_t out, void *arg)
 		return (size);
 
 	notesize = sizeof(Elf_Note) +		/* note header */
-	    roundup2(8, ELF_NOTE_ROUNDSIZE) +	/* note name ("FreeBSD") */
+	    roundup2(sizeof(FREEBSD_ABI_VENDOR), ELF_NOTE_ROUNDSIZE) +
+						/* note name */
+	    roundup2(size, ELF_NOTE_ROUNDSIZE);	/* note description */
+
+	return (notesize);
+}
+
+static size_t
+append_note_data(const void *src, void *dst, size_t len)
+{
+	size_t padded_len;
+
+	padded_len = roundup2(len, ELF_NOTE_ROUNDSIZE);
+	if (dst != NULL) {
+		bcopy(src, dst, len);
+		bzero((char *)dst + len, padded_len - len);
+	}
+	return (padded_len);
+}
+
+size_t
+__elfN(populate_note)(int type, void *src, void *dst, size_t size, void **descp)
+{
+	Elf_Note *note;
+	char *buf;
+	size_t notesize;
+
+	buf = dst;
+	if (buf != NULL) {
+		note = (Elf_Note *)buf;
+		note->n_namesz = sizeof(FREEBSD_ABI_VENDOR);
+		note->n_descsz = size;
+		note->n_type = type;
+		buf += sizeof(*note);
+		buf += append_note_data(FREEBSD_ABI_VENDOR, buf,
+		    sizeof(FREEBSD_ABI_VENDOR));
+		append_note_data(src, buf, size);
+		if (descp != NULL)
+			*descp = buf;
+	}
+
+	notesize = sizeof(Elf_Note) +		/* note header */
+	    roundup2(sizeof(FREEBSD_ABI_VENDOR), ELF_NOTE_ROUNDSIZE) +
+						/* note name */
 	    roundup2(size, ELF_NOTE_ROUNDSIZE);	/* note description */
 
 	return (notesize);
@@ -1604,13 +1647,13 @@ __elfN(putnote)(struct note_info *ninfo, struct sbuf *sb)
 		return;
 	}
 
-	note.n_namesz = 8; /* strlen("FreeBSD") + 1 */
+	note.n_namesz = sizeof(FREEBSD_ABI_VENDOR);
 	note.n_descsz = ninfo->outsize;
 	note.n_type = ninfo->type;
 
 	sbuf_bcat(sb, &note, sizeof(note));
 	sbuf_start_section(sb, &old_len);
-	sbuf_bcat(sb, "FreeBSD", note.n_namesz);
+	sbuf_bcat(sb, FREEBSD_ABI_VENDOR, sizeof(FREEBSD_ABI_VENDOR));
 	sbuf_end_section(sb, old_len, ELF_NOTE_ROUNDSIZE, 0);
 	if (note.n_descsz == 0)
 		return;
diff --git a/sys/sys/elf_common.h b/sys/sys/elf_common.h
index 3f6eebbd6856..2a29dd42df03 100644
--- a/sys/sys/elf_common.h
+++ b/sys/sys/elf_common.h
@@ -508,6 +508,7 @@ typedef struct {
 #define	NT_PROCSTAT_OSREL	14	/* Procstat osreldate data. */
 #define	NT_PROCSTAT_PSSTRINGS	15	/* Procstat ps_strings data. */
 #define	NT_PROCSTAT_AUXV	16	/* Procstat auxv data. */
+#define	NT_X86_XSTATE	0x202	/* x86 XSAVE extended state. */
 
 /* Symbol Binding - ELFNN_ST_BIND - st_info */
 #define	STB_LOCAL	0	/* Local symbol */
diff --git a/sys/sys/imgact_elf.h b/sys/sys/imgact_elf.h
index 712a55e1bac4..e8d5567cc273 100644
--- a/sys/sys/imgact_elf.h
+++ b/sys/sys/imgact_elf.h
@@ -90,6 +90,7 @@ int	__elfN(insert_brand_entry)(Elf_Brandinfo *entry);
 int	__elfN(remove_brand_entry)(Elf_Brandinfo *entry);
 int	__elfN(freebsd_fixup)(register_t **, struct image_params *);
 int	__elfN(coredump)(struct thread *, struct vnode *, off_t, int);
+size_t	__elfN(populate_note)(int, void *, void *, size_t, void **);
 
 /* Machine specific function to dump per-thread information. */
 void	__elfN(dump_thread)(struct thread *, void *, size_t *);
diff --git a/sys/x86/include/fpu.h b/sys/x86/include/fpu.h
index 73af34c4fd20..0430108ec08e 100644
--- a/sys/x86/include/fpu.h
+++ b/sys/x86/include/fpu.h
@@ -63,15 +63,7 @@ struct save87 {
 	struct env87	sv_env;		/* floating point control/status */
 	struct fpacc87	sv_ac[8];	/* accumulator contents, 0-7 */
 	uint8_t		sv_pad0[4];	/* saved status word (now unused) */
-	/*
-	 * Bogus padding for emulators.  Emulators should use their own
-	 * struct and arrange to store into this struct (ending here)
-	 * before it is inspected for ptracing or for core dumps.  Some
-	 * emulators overwrite the whole struct.  We have no good way of
-	 * knowing how much padding to leave.  Leave just enough for the
-	 * GPL emulator's i387_union (176 bytes total).
-	 */
-	uint8_t		sv_pad[64];	/* padding; used by emulators */
+	uint8_t		sv_pad[64];
 };
 
 /* Contents of each SSE extended accumulator. */
@@ -215,4 +207,11 @@ struct savefpu_ymm {
 #define	__INITIAL_MXCSR__	0x1F80
 #define	__INITIAL_MXCSR_MASK__	0xFFBF
 
+/*
+ * The current value of %xcr0 is saved in the sv_pad[] field of the FPU
+ * state in the NT_X86_XSTATE note in core dumps.  This offset is chosen
+ * to match the offset used by NT_X86_XSTATE in other systems.
+ */
+#define	X86_XSTATE_XCR0_OFFSET	464
+
 #endif /* !_X86_FPU_H_ */
diff --git a/sys/x86/include/ptrace.h b/sys/x86/include/ptrace.h
index 16a04d1e9f6e..16f9c62bb213 100644
--- a/sys/x86/include/ptrace.h
+++ b/sys/x86/include/ptrace.h
@@ -37,14 +37,25 @@
 
 /*
  * On amd64 (PT_FIRSTMACH + 0) and (PT_FIRSTMACH + 1) are old values for
- * PT_GETXSTATE and PT_SETXSTATE.  They should not be (re)used.
+ * PT_GETXSTATE_OLD and PT_SETXSTATE_OLD.  They should not be (re)used.
  */
 
 #ifdef __i386__
 #define	PT_GETXMMREGS	(PT_FIRSTMACH + 0)
 #define	PT_SETXMMREGS	(PT_FIRSTMACH + 1)
 #endif
-#define	PT_GETXSTATE	(PT_FIRSTMACH + 2)
-#define	PT_SETXSTATE	(PT_FIRSTMACH + 3)
+#ifdef _KERNEL
+#define	PT_GETXSTATE_OLD (PT_FIRSTMACH + 2)
+#define	PT_SETXSTATE_OLD (PT_FIRSTMACH + 3)
+#endif
+#define	PT_GETXSTATE_INFO (PT_FIRSTMACH + 4)
+#define	PT_GETXSTATE	(PT_FIRSTMACH + 5)
+#define	PT_SETXSTATE	(PT_FIRSTMACH + 6)
+
+/* Argument structure for PT_GETXSTATE_INFO. */
+struct ptrace_xstate_info {
+	uint64_t	xsave_mask;
+	uint32_t	xsave_len;
+};
 
 #endif
diff --git a/usr.bin/gcore/elfcore.c b/usr.bin/gcore/elfcore.c
index 037854c34d03..fb3bf5067e47 100644
--- a/usr.bin/gcore/elfcore.c
+++ b/usr.bin/gcore/elfcore.c
@@ -47,6 +47,7 @@ __FBSDID("$FreeBSD$");
 #include <err.h>
 #include <errno.h>
 #include <fcntl.h>
+#include <stdbool.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -101,6 +102,9 @@ static void *elf_note_fpregset(void *, size_t *);
 static void *elf_note_prpsinfo(void *, size_t *);
 static void *elf_note_prstatus(void *, size_t *);
 static void *elf_note_thrmisc(void *, size_t *);
+#if defined(__i386__) || defined(__amd64__)
+static void *elf_note_x86_xstate(void *, size_t *);
+#endif
 static void *elf_note_procstat_auxv(void *, size_t *);
 static void *elf_note_procstat_files(void *, size_t *);
 static void *elf_note_procstat_groups(void *, size_t *);
@@ -341,6 +345,9 @@ elf_putnotes(pid_t pid, struct sbuf *sb, size_t *sizep)
 		elf_putnote(NT_PRSTATUS, elf_note_prstatus, tids + i, sb);
 		elf_putnote(NT_FPREGSET, elf_note_fpregset, tids + i, sb);
 		elf_putnote(NT_THRMISC, elf_note_thrmisc, tids + i, sb);
+#if defined(__i386__) || defined(__amd64__)
+		elf_putnote(NT_X86_XSTATE, elf_note_x86_xstate, tids + i, sb);
+#endif
 	}
 
 #ifndef ELFCORE_COMPAT_32
@@ -615,6 +622,34 @@ elf_note_thrmisc(void *arg, size_t *sizep)
 	return (thrmisc);
 }
 
+#if defined(__i386__) || defined(__amd64__)
+static void *
+elf_note_x86_xstate(void *arg, size_t *sizep)
+{
+	lwpid_t tid;
+	char *xstate;
+	static bool xsave_checked = false;
+	static struct ptrace_xstate_info info;
+
+	tid = *(lwpid_t *)arg;
+	if (!xsave_checked) {
+		if (ptrace(PT_GETXSTATE_INFO, tid, (void *)&info,
+		    sizeof(info)) != 0)
+			info.xsave_len = 0;
+		xsave_checked = true;
+	}
+	if (info.xsave_len == 0) {
+		*sizep = 0;
+		return (NULL);
+	}
+	xstate = calloc(1, info.xsave_len);
+	ptrace(PT_GETXSTATE, tid, xstate, 0);
+	*(uint64_t *)(xstate + X86_XSTATE_XCR0_OFFSET) = info.xsave_mask;
+	*sizep = info.xsave_len;
+	return (xstate);
+}
+#endif
+
 static void *
 procstat_sysctl(void *arg, int what, size_t structsz, size_t *sizep)
 {