--- sys/amd64/amd64/fpu.c.orig
+++ sys/amd64/amd64/fpu.c
@@ -744,6 +744,7 @@
 	int max_ext_n, i, owned;
 
 	pcb = td->td_pcb;
+	critical_enter();
 	if ((pcb->pcb_flags & PCB_USERFPUINITDONE) == 0) {
 		bcopy(fpu_initialstate, get_pcb_user_save_pcb(pcb),
 		    cpu_max_ext_state_size);
@@ -750,9 +751,9 @@
 		get_pcb_user_save_pcb(pcb)->sv_env.en_cw =
 		    pcb->pcb_initial_fpucw;
 		fpuuserinited(td);
+		critical_exit();
 		return (_MC_FPOWNED_PCB);
 	}
-	critical_enter();
 	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
 		fpusave(get_pcb_user_save_pcb(pcb));
 		owned = _MC_FPOWNED_FPU;
@@ -759,7 +760,6 @@
 	} else {
 		owned = _MC_FPOWNED_PCB;
 	}
-	critical_exit();
 	if (use_xsave) {
 		/*
 		 * Handle partially saved state.
@@ -779,6 +779,7 @@
 			*xstate_bv |= bit;
 		}
 	}
+	critical_exit();
 	return (owned);
 }
 
@@ -787,6 +788,7 @@
 {
 	struct pcb *pcb;
 
+	CRITICAL_ASSERT(td);
 	pcb = td->td_pcb;
 	if (PCB_USER_FPU(pcb))
 		set_pcb_flags(pcb,
@@ -845,26 +847,25 @@
 
 	addr->sv_env.en_mxcsr &= cpu_mxcsr_mask;
 	pcb = td->td_pcb;
+	error = 0;
 	critical_enter();
 	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
 		error = fpusetxstate(td, xfpustate, xfpustate_size);
-		if (error != 0) {
-			critical_exit();
-			return (error);
+		if (error == 0) {
+			bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
+			fpurestore(get_pcb_user_save_td(td));
+			set_pcb_flags(pcb, PCB_FPUINITDONE |
+			    PCB_USERFPUINITDONE);
 		}
-		bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
-		fpurestore(get_pcb_user_save_td(td));
-		critical_exit();
-		set_pcb_flags(pcb, PCB_FPUINITDONE | PCB_USERFPUINITDONE);
 	} else {
-		critical_exit();
 		error = fpusetxstate(td, xfpustate, xfpustate_size);
-		if (error != 0)
-			return (error);
-		bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
-		fpuuserinited(td);
+		if (error == 0) {
+			bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
+			fpuuserinited(td);
+		}
 	}
-	return (0);
+	critical_exit();
+	return (error);
 }
 
 /*
@@ -1037,6 +1038,7 @@
 		ctx->flags = FPU_KERN_CTX_DUMMY | FPU_KERN_CTX_INUSE;
 		return (0);
 	}
+	critical_enter();
 	KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save ==
 	    get_pcb_user_save_pcb(pcb), ("mangled pcb_save"));
 	ctx->flags = FPU_KERN_CTX_INUSE;
@@ -1047,6 +1049,7 @@
 	pcb->pcb_save = fpu_kern_ctx_savefpu(ctx);
 	set_pcb_flags(pcb, PCB_KERNFPU);
 	clear_pcb_flags(pcb, PCB_FPUINITDONE);
+	critical_exit();
 	return (0);
 }
 
@@ -1065,7 +1068,6 @@
 
 		clear_pcb_flags(pcb,  PCB_FPUNOSAVE | PCB_FPUINITDONE);
 		start_emulating();
-		critical_exit();
 	} else {
 		KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0,
 		    ("leaving not inuse ctx"));
@@ -1079,7 +1081,6 @@
 		critical_enter();
 		if (curthread == PCPU_GET(fpcurthread))
 			fpudrop();
-		critical_exit();
 		pcb->pcb_save = ctx->prev;
 	}
 
@@ -1096,6 +1097,7 @@
 			clear_pcb_flags(pcb, PCB_FPUINITDONE);
 		KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave"));
 	}
+	critical_exit();
 	return (0);
 }
 
--- sys/amd64/amd64/machdep.c.orig
+++ sys/amd64/amd64/machdep.c
@@ -2158,8 +2158,10 @@
 set_fpregs(struct thread *td, struct fpreg *fpregs)
 {
 
+	critical_enter();
 	set_fpregs_xmm(fpregs, get_pcb_user_save_td(td));
 	fpuuserinited(td);
+	critical_exit();
 	return (0);
 }
 
--- sys/i386/i386/machdep.c.orig
+++ sys/i386/i386/machdep.c
@@ -3004,6 +3004,7 @@
 set_fpregs(struct thread *td, struct fpreg *fpregs)
 {
 
+	critical_enter();
 	if (cpu_fxsr)
 		npx_set_fpregs_xmm((struct save87 *)fpregs,
 		    &get_pcb_user_save_td(td)->sv_xmm);
@@ -3011,6 +3012,7 @@
 		bcopy(fpregs, &get_pcb_user_save_td(td)->sv_87,
 		    sizeof(*fpregs));
 	npxuserinited(td);
+	critical_exit();
 	return (0);
 }
 
--- sys/i386/isa/npx.c.orig
+++ sys/i386/isa/npx.c
@@ -974,14 +974,15 @@
 		return (_MC_FPOWNED_NONE);
 
 	pcb = td->td_pcb;
+	critical_enter();
 	if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
 		bcopy(npx_initialstate, get_pcb_user_save_pcb(pcb),
 		    cpu_max_ext_state_size);
 		SET_FPU_CW(get_pcb_user_save_pcb(pcb), pcb->pcb_initial_npxcw);
 		npxuserinited(td);
+		critical_exit();
 		return (_MC_FPOWNED_PCB);
 	}
-	critical_enter();
 	if (td == PCPU_GET(fpcurthread)) {
 		fpusave(get_pcb_user_save_pcb(pcb));
 		if (!cpu_fxsr)
@@ -995,7 +996,6 @@
 	} else {
 		owned = _MC_FPOWNED_PCB;
 	}
-	critical_exit();
 	if (use_xsave) {
 		/*
 		 * Handle partially saved state.
@@ -1018,6 +1018,7 @@
 			*xstate_bv |= bit;
 		}
 	}
+	critical_exit();
 	return (owned);
 }
 
@@ -1026,6 +1027,7 @@
 {
 	struct pcb *pcb;
 
+	CRITICAL_ASSERT(td);
 	pcb = td->td_pcb;
 	if (PCB_USER_FPU(pcb))
 		pcb->pcb_flags |= PCB_NPXINITDONE;
@@ -1083,28 +1085,26 @@
 	if (cpu_fxsr)
 		addr->sv_xmm.sv_env.en_mxcsr &= cpu_mxcsr_mask;
 	pcb = td->td_pcb;
+	error = 0;
 	critical_enter();
 	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
 		error = npxsetxstate(td, xfpustate, xfpustate_size);
-		if (error != 0) {
-			critical_exit();
-			return (error);
+		if (error == 0) {
+			if (!cpu_fxsr)
+				fnclex();	/* As in npxdrop(). */
+			bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
+			fpurstor(get_pcb_user_save_td(td));
+			pcb->pcb_flags |= PCB_NPXUSERINITDONE | PCB_NPXINITDONE;
 		}
-		if (!cpu_fxsr)
-			fnclex();	/* As in npxdrop(). */
-		bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
-		fpurstor(get_pcb_user_save_td(td));
-		critical_exit();
-		pcb->pcb_flags |= PCB_NPXUSERINITDONE | PCB_NPXINITDONE;
 	} else {
-		critical_exit();
 		error = npxsetxstate(td, xfpustate, xfpustate_size);
-		if (error != 0)
-			return (error);
-		bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
-		npxuserinited(td);
+		if (error == 0) {
+			bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
+			npxuserinited(td);
+		}
 	}
-	return (0);
+	critical_exit();
+	return (error);
 }
 
 static void
@@ -1373,6 +1373,7 @@
 		return (0);
 	}
 	pcb = td->td_pcb;
+	critical_enter();
 	KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save ==
 	    get_pcb_user_save_pcb(pcb), ("mangled pcb_save"));
 	ctx->flags = FPU_KERN_CTX_INUSE;
@@ -1383,6 +1384,7 @@
 	pcb->pcb_save = fpu_kern_ctx_savefpu(ctx);
 	pcb->pcb_flags |= PCB_KERNNPX;
 	pcb->pcb_flags &= ~PCB_NPXINITDONE;
+	critical_exit();
 	return (0);
 }
 
@@ -1401,7 +1403,6 @@
 	critical_enter();
 	if (curthread == PCPU_GET(fpcurthread))
 		npxdrop();
-	critical_exit();
 	pcb->pcb_save = ctx->prev;
 	if (pcb->pcb_save == get_pcb_user_save_pcb(pcb)) {
 		if ((pcb->pcb_flags & PCB_NPXUSERINITDONE) != 0)
@@ -1416,6 +1417,7 @@
 			pcb->pcb_flags &= ~PCB_NPXINITDONE;
 		KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave"));
 	}
+	critical_exit();
 	return (0);
 }