Index: sys/amd64/linux/linux.h
===================================================================
--- sys/amd64/linux/linux.h
+++ sys/amd64/linux/linux.h
@@ -233,6 +233,7 @@
 	l_stack_t	uc_stack;
 	struct l_sigcontext	uc_mcontext;
 	l_sigset_t	uc_sigmask;
+	struct l_fpstate	uc_fpregs;
 };
 
 #define LINUX_SI_PREAMBLE_SIZE	(4 * sizeof(int))
Index: sys/amd64/linux/linux_sysvec.c
===================================================================
--- sys/amd64/linux/linux_sysvec.c
+++ sys/amd64/linux/linux_sysvec.c
@@ -513,9 +513,7 @@
 }
 
 /*
- * Copied from amd64/amd64/machdep.c
- *
- * XXX fpu state need? don't think so
+ * Based on amd64/amd64/exec_machdep.c
  */
 int
 linux_rt_sigreturn(struct thread *td, struct linux_rt_sigreturn_args *args)
@@ -524,8 +522,10 @@
 	struct l_ucontext uc;
 	struct l_sigcontext *context;
 	struct trapframe *regs;
+	mcontext_t mcontext;
+	struct savefpu *savefpu;
 	unsigned long rflags;
-	int error;
+	int error, ret;
 	ksiginfo_t ksi;
 
 	regs = td->td_frame;
@@ -573,6 +573,34 @@
 		return (EINVAL);
 	}
 
+	if (uc.uc_mcontext.sc_fpstate != NULL) {
+		memset(&mcontext, 0, sizeof(mcontext));
+		mcontext.mc_flags |= _MC_HASFPXSTATE;
+		mcontext.mc_fpformat = _MC_FPFMT_XMM;
+		mcontext.mc_ownedfp = _MC_FPOWNED_FPU;
+
+		savefpu = (struct savefpu *)mcontext.mc_fpstate;
+		savefpu->sv_env.en_cw = uc.uc_fpregs.cwd;
+		savefpu->sv_env.en_sw = uc.uc_fpregs.swd;
+		savefpu->sv_env.en_tw = uc.uc_fpregs.twd;
+		savefpu->sv_env.en_opcode = uc.uc_fpregs.fop;
+		savefpu->sv_env.en_rip = uc.uc_fpregs.rip;
+		savefpu->sv_env.en_rdp = uc.uc_fpregs.rdp;
+		savefpu->sv_env.en_mxcsr = uc.uc_fpregs.mxcsr;
+		savefpu->sv_env.en_mxcsr_mask = uc.uc_fpregs.mxcsr_mask;
+		CTASSERT(sizeof(uc.uc_fpregs.st_space) == sizeof(savefpu->sv_fp));
+		memcpy(savefpu->sv_fp, uc.uc_fpregs.st_space, sizeof(savefpu->sv_fp));
+		CTASSERT(sizeof(uc.uc_fpregs.xmm_space) == sizeof(savefpu->sv_xmm));
+		memcpy(savefpu->sv_xmm, uc.uc_fpregs.xmm_space, sizeof(savefpu->sv_xmm));
+
+		ret = set_fpcontext(td, &mcontext, NULL, 0);
+		if (ret != 0) {
+			uprintf("pid %d (%s): sigreturn set_fpcontext err %d\n",
+					p->p_pid, td->td_name, ret);
+			return (ret);
+		}
+	}
+
 	PROC_LOCK(p);
 	linux_to_bsd_sigset(&uc.uc_sigmask, &td->td_sigmask);
 	SIG_CANTMASK(td->td_sigmask);
@@ -605,7 +633,7 @@
 }
 
 /*
- * copied from amd64/amd64/machdep.c
+ * Based on amd64/amd64/exec_machdep.c
  *
  * Send an interrupt to process.
  */
@@ -618,6 +646,8 @@
 	struct sigacts *psp;
 	caddr_t sp;
 	struct trapframe *regs;
+	mcontext_t mcontext;
+	const struct savefpu *savefpu;
 	int sig, code;
 	int oonstack;
 
@@ -667,6 +697,28 @@
 	sf.sf_sc.uc_mcontext.sc_trapno = bsd_to_linux_trapcode(code);
 	sf.sf_sc.uc_mcontext.sc_cr2    = (register_t)ksi->ksi_addr;
 
+	get_fpcontext(td, &mcontext, NULL, NULL);
+	if (mcontext.mc_fpformat == _MC_FPFMT_XMM) {
+		savefpu = (const struct savefpu *)mcontext.mc_fpstate;
+
+		sf.sf_sc.uc_fpregs.cwd         = savefpu->sv_env.en_cw;
+		sf.sf_sc.uc_fpregs.swd         = savefpu->sv_env.en_sw;
+		sf.sf_sc.uc_fpregs.twd         = savefpu->sv_env.en_tw;
+		sf.sf_sc.uc_fpregs.fop         = savefpu->sv_env.en_opcode;
+		sf.sf_sc.uc_fpregs.rip         = savefpu->sv_env.en_rip;
+		sf.sf_sc.uc_fpregs.rdp         = savefpu->sv_env.en_rdp;
+		sf.sf_sc.uc_fpregs.mxcsr       = savefpu->sv_env.en_mxcsr;
+		sf.sf_sc.uc_fpregs.mxcsr_mask  = savefpu->sv_env.en_mxcsr_mask;
+		CTASSERT(sizeof(sf.sf_sc.uc_fpregs.st_space) ==
+		    sizeof(savefpu->sv_fp));
+		memcpy(sf.sf_sc.uc_fpregs.st_space,
+		    savefpu->sv_fp, sizeof(savefpu->sv_fp));
+		CTASSERT(sizeof(sf.sf_sc.uc_fpregs.xmm_space) ==
+		    sizeof(savefpu->sv_xmm));
+		memcpy(sf.sf_sc.uc_fpregs.xmm_space,
+		    savefpu->sv_xmm, sizeof(savefpu->sv_xmm));
+	}
+
 	/* Allocate space for the signal handler context. */
 	if ((td->td_pflags & TDP_ALTSTACK) != 0 && !oonstack &&
 	    SIGISMEMBER(psp->ps_sigonstack, sig)) {
@@ -693,6 +745,11 @@
 	mtx_unlock(&psp->ps_mtx);
 	PROC_UNLOCK(p);
 
+	if (mcontext.mc_fpformat == _MC_FPFMT_XMM) {
+		sf.sf_sc.uc_mcontext.sc_fpstate = (void *)((char *)sfp +
+		    ((char *)&sf.sf_sc.uc_fpregs - (char *)&sf));
+	}
+
 	/* Copy the sigframe out to the user's stack. */
 	if (copyout(&sf, sfp, sizeof(*sfp)) != 0) {
 		uprintf("pid %d comm %s has trashed its stack, killing\n",