Subtle changes to the powerpc fpu handling, basically fpuproc is handled

in a non-raceable manner inside save_fpu and enable_fpu so that the
eventual SMP case will not grab a half loaded fpu context.

6 files changed, 37 insertions, 26 deletions

diff --git a/sys/arch/macppc/macppc/machdep.c b/sys/arch/macppc/macppc/machdep.c
index 713f346e033..f1b741f673f 100644
--- a/sys/arch/macppc/macppc/machdep.c
+++ b/sys/arch/macppc/macppc/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.76 2005/10/09 14:17:32 drahn Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.77 2005/10/09 14:52:12 drahn Exp $	*/
 /*	$NetBSD: machdep.c,v 1.4 1996/10/16 19:33:11 ws Exp $	*/
 
 /*
@@ -93,7 +93,6 @@
  */
 struct pcb *curpcb;
 struct pmap *curpm;
-struct proc *fpuproc;
 
 extern struct user *proc0paddr;
 struct pool ppc_vecpl;
diff --git a/sys/arch/powerpc/include/fpu.h b/sys/arch/powerpc/include/fpu.h
index b845691b37a..82cd06facb3 100644
--- a/sys/arch/powerpc/include/fpu.h
+++ b/sys/arch/powerpc/include/fpu.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: fpu.h,v 1.5 2001/09/01 15:49:05 drahn Exp $	*/
+/*	$OpenBSD: fpu.h,v 1.6 2005/10/09 14:52:12 drahn Exp $	*/
 
 /*-
  * Copyright (C) 1996 Wolfgang Solfrank.
@@ -67,5 +67,5 @@
 #define	FPCSR_RN	0x00000003
 
 void enable_fpu(struct proc *p);
-void save_fpu(struct proc *p);
+void save_fpu(void);
 #endif	/* _POWERPC_FPU_H_ */
diff --git a/sys/arch/powerpc/powerpc/fpu.c b/sys/arch/powerpc/powerpc/fpu.c
index 8b09ff065da..c5672584c62 100644
--- a/sys/arch/powerpc/powerpc/fpu.c
+++ b/sys/arch/powerpc/powerpc/fpu.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: fpu.c,v 1.8 2003/10/15 02:43:09 drahn Exp $	*/
+/*	$OpenBSD: fpu.c,v 1.9 2005/10/09 14:52:12 drahn Exp $	*/
 /*	$NetBSD: fpu.c,v 1.1 1996/09/30 16:34:44 ws Exp $	*/
 
 /*
@@ -45,13 +45,12 @@ enable_fpu(struct proc *p)
 	struct pcb *pcb = &p->p_addr->u_pcb;
 	struct trapframe *tf = trapframe(p);
 	
-	tf->srr1 |= PSL_FP;
 	if (!(pcb->pcb_flags & PCB_FPU)) {
 		bzero(&pcb->pcb_fpu, sizeof pcb->pcb_fpu);
 		pcb->pcb_flags |= PCB_FPU;
 	}
 	msr = ppc_mfmsr();
-	ppc_mtmsr(msr | PSL_FP);
+	ppc_mtmsr((msr  & ~PSL_EE) | PSL_FP);
 	__asm volatile("isync");
 
 	asm volatile ("lfd 0,0(%0); mtfsf 0xff,0" :: "b"(&pcb->pcb_fpu.fpcsr));
@@ -87,23 +86,32 @@ enable_fpu(struct proc *p)
 	     "lfd 29,232(%0);"
 	     "lfd 30,240(%0);"
 	     "lfd 31,248(%0)" :: "b"(&pcb->pcb_fpu.fpr[0]));
+	fpuproc = p;
+	tf->srr1 |= PSL_FP;
 	ppc_mtmsr(msr);
 	__asm volatile("isync");
 }
 
 void
-save_fpu(struct proc *p)
+save_fpu()
 {
 	int msr;
 	struct pcb *pcb;
+	struct proc *p;
+	struct trapframe *tf;
+		
+	msr = ppc_mfmsr();
+	ppc_mtmsr((msr  & ~PSL_EE) | PSL_FP);
+
+	p = fpuproc;
 
-	if (p == NULL)
+	if (p == NULL) {
+		ppc_mtmsr(msr);
 		return;
+	}
 
 	pcb = &p->p_addr->u_pcb;
 	
-	msr = ppc_mfmsr();
-	ppc_mtmsr(msr | PSL_FP);
 	__asm volatile("isync");
 
 	asm ("stfd 0,0(%0);"
@@ -139,6 +147,12 @@ save_fpu(struct proc *p)
 	     "stfd 30,240(%0);"
 	     "stfd 31,248(%0)" :: "b"(&pcb->pcb_fpu.fpr[0]));
 	asm volatile ("mffs 0; stfd 0,0(%0)" :: "b"(&pcb->pcb_fpu.fpcsr));
+	asm ("lfd 0,0(%0);" :: "b"(&pcb->pcb_fpu.fpr[0]));
+
+	tf = trapframe(fpuproc);
+	tf->srr1 &= ~PSL_FP;
+	fpuproc = NULL;
+
 	ppc_mtmsr(msr);
 	__asm volatile("isync");
 }
diff --git a/sys/arch/powerpc/powerpc/process_machdep.c b/sys/arch/powerpc/powerpc/process_machdep.c
index 8c5ea10bb46..b7815091c3d 100644
--- a/sys/arch/powerpc/powerpc/process_machdep.c
+++ b/sys/arch/powerpc/powerpc/process_machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: process_machdep.c,v 1.9 2005/06/20 20:02:04 kettenis Exp $	*/
+/*	$OpenBSD: process_machdep.c,v 1.10 2005/10/09 14:52:12 drahn Exp $	*/
 /*	$NetBSD: process_machdep.c,v 1.1 1996/09/30 16:34:53 ws Exp $	*/
 
 /*
@@ -54,7 +54,7 @@ process_read_regs(struct proc *p, struct reg *regs)
 		bzero(regs->fpr, sizeof(regs->fpr));
 	} else {
 		if (p == fpuproc)
-			save_fpu(fpuproc);
+			save_fpu();
 		bcopy(pcb->pcb_fpu.fpr, regs->fpr, sizeof(regs->fpr));
 	}
 
@@ -104,7 +104,7 @@ process_write_regs(struct proc *p, struct reg *regs)
 	bcopy(regs->gpr, tf->fixreg, sizeof(regs->gpr));
 
 	if (p == fpuproc) {	/* release the fpu */
-		save_fpu(fpuproc);
+		save_fpu();
 		fpuproc = NULL;
 	}
 
diff --git a/sys/arch/powerpc/powerpc/trap.c b/sys/arch/powerpc/powerpc/trap.c
index 575da48d5a7..c1592769f57 100644
--- a/sys/arch/powerpc/powerpc/trap.c
+++ b/sys/arch/powerpc/powerpc/trap.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: trap.c,v 1.69 2005/09/15 21:09:29 miod Exp $	*/
+/*	$OpenBSD: trap.c,v 1.70 2005/10/09 14:52:12 drahn Exp $	*/
 /*	$NetBSD: trap.c,v 1.3 1996/10/13 03:31:37 christos Exp $	*/
 
 /*
@@ -73,6 +73,7 @@ void trap(struct trapframe *frame);
 
 volatile int want_resched;
 struct proc *ppc_vecproc;
+struct proc *fpuproc;
 
 #ifdef DDB
 void ppc_dumpbt(struct trapframe *frame);
@@ -523,8 +524,7 @@ syscall_bad:
 
 	case EXC_FPU|EXC_USER:
 		if (fpuproc)
-			save_fpu(fpuproc);
-		fpuproc = p;
+			save_fpu();
 		uvmexp.fpswtch++;
 		enable_fpu(p);
 		break;
@@ -675,8 +675,6 @@ for (i = 0; i < errnum; i++) {
 	 */
 	if (p != ppc_vecproc)
 		frame->srr1 &= ~PSL_VEC;
-	else
-		frame->srr1 |= PSL_VEC;
 #endif /* ALTIVEC */
 }
 
@@ -755,21 +753,21 @@ fix_unaligned(struct proc *p, struct trapframe *frame)
 			 */
 			if (fpuproc != p) {
 				if (fpuproc)
-					save_fpu(fpuproc);
+					save_fpu();
 				enable_fpu(p);
 			}
-			save_fpu(p);
+			save_fpu();
 
 			if (indicator == EXC_ALI_LFD) {
 				if (copyin((void *)frame->dar, fpr,
 				    sizeof(double)) != 0)
 					return -1;
-				enable_fpu(p);
 			} else {
 				if (copyout(fpr, (void *)frame->dar,
 				    sizeof(double)) != 0)
 					return -1;
 			}
+			enable_fpu(p);
 			return 0;
 		}
 		break;
diff --git a/sys/arch/powerpc/powerpc/vm_machdep.c b/sys/arch/powerpc/powerpc/vm_machdep.c
index dfc77a82d2f..c3ff197a4ad 100644
--- a/sys/arch/powerpc/powerpc/vm_machdep.c
+++ b/sys/arch/powerpc/powerpc/vm_machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: vm_machdep.c,v 1.38 2004/06/24 22:35:56 drahn Exp $	*/
+/*	$OpenBSD: vm_machdep.c,v 1.39 2005/10/09 14:52:12 drahn Exp $	*/
 /*	$NetBSD: vm_machdep.c,v 1.1 1996/09/30 16:34:57 ws Exp $	*/
 
 /*
@@ -61,7 +61,7 @@ cpu_fork(struct proc *p1, struct proc *p2, void *stack, size_t stacksize,
 	struct pcb *pcb = &p2->p_addr->u_pcb;
 
 	if (p1 == fpuproc)
-		save_fpu(p1);
+		save_fpu();
 	*pcb = p1->p_addr->u_pcb;
 	
 #ifdef ALTIVEC
@@ -170,8 +170,8 @@ cpu_exit(struct proc *p)
 	struct pcb *pcb = &p->p_addr->u_pcb;
 #endif /* ALTIVEC */
 	
-	if (p == fpuproc)	/* release the fpu */
-		fpuproc = 0;
+	if (p == fpuproc) 	/* release the fpu */
+		fpuproc = NULL;
 
 #ifdef ALTIVEC
 	if (p == ppc_vecproc)