diff --git a/arch/sparc64/kernel/ktlb.S b/arch/sparc64/kernel/ktlb.S
index 9b415ab6db6b0e09fb85bf59a221a7f163b6f4f1..c1335432124e3b4e1002a26b0714a2312cddff11 100644
--- a/arch/sparc64/kernel/ktlb.S
+++ b/arch/sparc64/kernel/ktlb.S
@@ -60,8 +60,15 @@ kvmap_itlb_load:
 	retry
 
 kvmap_itlb_longpath:
-	rdpr	%pstate, %g5
+
+661:	rdpr	%pstate, %g5
 	wrpr	%g5, PSTATE_AG | PSTATE_MG, %pstate
+	.section .gl_2insn_patch, "ax"
+	.word	661b
+	nop
+	nop
+	.previous
+
 	rdpr	%tpc, %g5
 	ba,pt	%xcc, sparc64_realfault_common
 	 mov	FAULT_CODE_ITLB, %g4
@@ -161,8 +168,15 @@ kvmap_check_obp:
 	 nop
 
 kvmap_dtlb_longpath:
-	rdpr	%pstate, %g5
+
+661:	rdpr	%pstate, %g5
 	wrpr	%g5, PSTATE_AG | PSTATE_MG, %pstate
+	.section .gl_2insn_patch, "ax"
+	.word	661b
+	nop
+	nop
+	.previous
+
 	rdpr	%tl, %g4
 	cmp	%g4, 1
 	mov	TLB_TAG_ACCESS, %g4
diff --git a/arch/sparc64/kernel/setup.c b/arch/sparc64/kernel/setup.c
index aaab319ad885e3bd0c9cb19dd4d48f32d6294920..e22bf5fc92ce8213d023d8ce0675f26a83a586f9 100644
--- a/arch/sparc64/kernel/setup.c
+++ b/arch/sparc64/kernel/setup.c
@@ -547,19 +547,33 @@ static void __init per_cpu_patch(void)
 
 static void __init gl_patch(void)
 {
-	struct gl_1insn_patch_entry *p;
+	struct gl_1insn_patch_entry *p1;
+	struct gl_2insn_patch_entry *p2;
 
 	if (tlb_type != hypervisor)
 		return;
 
-	p = &__gl_1insn_patch;
-	while (p < &__gl_1insn_patch_end) {
-		unsigned long addr = p->addr;
+	p1 = &__gl_1insn_patch;
+	while (p1 < &__gl_1insn_patch_end) {
+		unsigned long addr = p1->addr;
 
-		*(unsigned int *) (addr +  0) = p->insn;
+		*(unsigned int *) (addr +  0) = p1->insn;
 		__asm__ __volatile__("flush	%0" : : "r" (addr +  0));
 
-		p++;
+		p1++;
+	}
+
+	p2 = &__gl_2insn_patch;
+	while (p2 < &__gl_2insn_patch_end) {
+		unsigned long addr = p2->addr;
+
+		*(unsigned int *) (addr +  0) = p2->insns[0];
+		__asm__ __volatile__("flush	%0" : : "r" (addr +  0));
+
+		*(unsigned int *) (addr +  3) = p2->insns[1];
+		__asm__ __volatile__("flush	%0" : : "r" (addr +  4));
+
+		p2++;
 	}
 }
 
diff --git a/arch/sparc64/kernel/tsb.S b/arch/sparc64/kernel/tsb.S
index 3b45db98005a5a7a6e481ee9c53c2cd178b2a92b..96e63168d8b2644ad1fd958409b54f1a5f2276dc 100644
--- a/arch/sparc64/kernel/tsb.S
+++ b/arch/sparc64/kernel/tsb.S
@@ -82,9 +82,17 @@ tsb_itlb_load:
 	.globl		tsb_do_fault
 tsb_do_fault:
 	cmp		%g3, FAULT_CODE_DTLB
-	rdpr		%pstate, %g5
+
+661:	rdpr		%pstate, %g5
+	wrpr		%g5, PSTATE_AG | PSTATE_MG, %pstate
+	.section	.gl_2insn_patch, "ax"
+	.word		661b
+	nop
+	nop
+	.previous
+
 	bne,pn		%xcc, tsb_do_itlb_fault
-	 wrpr		%g5, PSTATE_AG | PSTATE_MG, %pstate
+	 nop
 
 tsb_do_dtlb_fault:
 	rdpr	%tl, %g4
diff --git a/arch/sparc64/kernel/vmlinux.lds.S b/arch/sparc64/kernel/vmlinux.lds.S
index 482d1ed87f4da6b59040c9c0818d5a87cb8a920c..686bf6b3b03f4cd82b8ca84208d3d47b502cd08e 100644
--- a/arch/sparc64/kernel/vmlinux.lds.S
+++ b/arch/sparc64/kernel/vmlinux.lds.S
@@ -80,6 +80,9 @@ SECTIONS
   __gl_1insn_patch = .;
   .gl_1insn_patch : { *(.gl_1insn_patch) }
   __gl_1insn_patch_end = .;
+  __gl_2insn_patch = .;
+  .gl_2insn_patch : { *(.gl_2insn_patch) }
+  __gl_2insn_patch_end = .;
   . = ALIGN(8192); 
   __initramfs_start = .;
   .init.ramfs : { *(.init.ramfs) }
diff --git a/arch/sparc64/mm/ultra.S b/arch/sparc64/mm/ultra.S
index cac58d66fca904f178c4a5b7b46acf01665a347e..5dd86ad0d29f83a1a75c0d25c2761af3cb6434f0 100644
--- a/arch/sparc64/mm/ultra.S
+++ b/arch/sparc64/mm/ultra.S
@@ -444,8 +444,15 @@ xcall_flush_tlb_kernel_range:	/* 22 insns */
 	 */
 	.globl		xcall_sync_tick
 xcall_sync_tick:
-	rdpr		%pstate, %g2
+
+661:	rdpr		%pstate, %g2
 	wrpr		%g2, PSTATE_IG | PSTATE_AG, %pstate
+	.section	.gl_2insn_patch, "ax"
+	.word		661b
+	nop
+	nop
+	.previous
+
 	rdpr		%pil, %g2
 	wrpr		%g0, 15, %pil
 	sethi		%hi(109f), %g7
@@ -468,8 +475,15 @@ xcall_sync_tick:
 	 */
 	.globl		xcall_report_regs
 xcall_report_regs:
-	rdpr		%pstate, %g2
+
+661:	rdpr		%pstate, %g2
 	wrpr		%g2, PSTATE_IG | PSTATE_AG, %pstate
+	.section	.gl_2insn_patch, "ax"
+	.word		661b
+	nop
+	nop
+	.previous
+
 	rdpr		%pil, %g2
 	wrpr		%g0, 15, %pil
 	sethi		%hi(109f), %g7
diff --git a/arch/sparc64/prom/cif.S b/arch/sparc64/prom/cif.S
index 29d0ae74aed8d8dc32b28f9cc64f4c8aadda7e34..5f27ad779c0c578097218e44b005eaed770acf7d 100644
--- a/arch/sparc64/prom/cif.S
+++ b/arch/sparc64/prom/cif.S
@@ -1,10 +1,12 @@
 /* cif.S: PROM entry/exit assembler trampolines.
  *
- * Copyright (C) 1996,1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
- * Copyright (C) 2005 David S. Miller <davem@davemloft.net>
+ * Copyright (C) 1996, 1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
+ * Copyright (C) 2005, 2006 David S. Miller <davem@davemloft.net>
  */
 
 #include <asm/pstate.h>
+#include <asm/cpudata.h>
+#include <asm/thread_info.h>
 
 	.text
 	.globl	prom_cif_interface
@@ -12,78 +14,16 @@ prom_cif_interface:
 	sethi	%hi(p1275buf), %o0
 	or	%o0, %lo(p1275buf), %o0
 	ldx	[%o0 + 0x010], %o1	! prom_cif_stack
-	save	%o1, -0x190, %sp
+	save	%o1, -192, %sp
 	ldx	[%i0 + 0x008], %l2	! prom_cif_handler
-	rdpr	%pstate, %l4
-	wrpr	%g0, 0x15, %pstate	! save alternate globals
-	stx	%g1, [%sp + 2047 + 0x0b0]
-	stx	%g2, [%sp + 2047 + 0x0b8]
-	stx	%g3, [%sp + 2047 + 0x0c0]
-	stx	%g4, [%sp + 2047 + 0x0c8]
-	stx	%g5, [%sp + 2047 + 0x0d0]
-	stx	%g6, [%sp + 2047 + 0x0d8]
-	stx	%g7, [%sp + 2047 + 0x0e0]
-	wrpr	%g0, 0x814, %pstate	! save interrupt globals
-	stx	%g1, [%sp + 2047 + 0x0e8]
-	stx	%g2, [%sp + 2047 + 0x0f0]
-	stx	%g3, [%sp + 2047 + 0x0f8]
-	stx	%g4, [%sp + 2047 + 0x100]
-	stx	%g5, [%sp + 2047 + 0x108]
-	stx	%g6, [%sp + 2047 + 0x110]
-	stx	%g7, [%sp + 2047 + 0x118]
-	wrpr	%g0, 0x14, %pstate	! save normal globals
-	stx	%g1, [%sp + 2047 + 0x120]
-	stx	%g2, [%sp + 2047 + 0x128]
-	stx	%g3, [%sp + 2047 + 0x130]
-	stx	%g4, [%sp + 2047 + 0x138]
-	stx	%g5, [%sp + 2047 + 0x140]
-	stx	%g6, [%sp + 2047 + 0x148]
-	stx	%g7, [%sp + 2047 + 0x150]
-	wrpr	%g0, 0x414, %pstate	! save mmu globals
-	stx	%g1, [%sp + 2047 + 0x158]
-	stx	%g2, [%sp + 2047 + 0x160]
-	stx	%g3, [%sp + 2047 + 0x168]
-	stx	%g4, [%sp + 2047 + 0x170]
-	stx	%g5, [%sp + 2047 + 0x178]
-	stx	%g6, [%sp + 2047 + 0x180]
-	stx	%g7, [%sp + 2047 + 0x188]
-	mov	%g1, %l0		! also save to locals, so we can handle
-	mov	%g2, %l1		! tlb faults later on, when accessing
-	mov	%g3, %l3		! the stack.
-	mov	%g7, %l5
-	wrpr	%l4, PSTATE_IE, %pstate	! turn off interrupts
+	mov	%g4, %l0
+	mov	%g5, %l1
+	mov	%g6, %l3
 	call	%l2
 	 add	%i0, 0x018, %o0		! prom_args
-	wrpr	%g0, 0x414, %pstate	! restore mmu globals
-	mov	%l0, %g1
-	mov	%l1, %g2
-	mov	%l3, %g3
-	mov	%l5, %g7
-	wrpr	%g0, 0x14, %pstate	! restore normal globals
-	ldx	[%sp + 2047 + 0x120], %g1
-	ldx	[%sp + 2047 + 0x128], %g2
-	ldx	[%sp + 2047 + 0x130], %g3
-	ldx	[%sp + 2047 + 0x138], %g4
-	ldx	[%sp + 2047 + 0x140], %g5
-	ldx	[%sp + 2047 + 0x148], %g6
-	ldx	[%sp + 2047 + 0x150], %g7
-	wrpr	%g0, 0x814, %pstate	! restore interrupt globals
-	ldx	[%sp + 2047 + 0x0e8], %g1
-	ldx	[%sp + 2047 + 0x0f0], %g2
-	ldx	[%sp + 2047 + 0x0f8], %g3
-	ldx	[%sp + 2047 + 0x100], %g4
-	ldx	[%sp + 2047 + 0x108], %g5
-	ldx	[%sp + 2047 + 0x110], %g6
-	ldx	[%sp + 2047 + 0x118], %g7
-	wrpr	%g0, 0x15, %pstate	! restore alternate globals
-	ldx	[%sp + 2047 + 0x0b0], %g1
-	ldx	[%sp + 2047 + 0x0b8], %g2
-	ldx	[%sp + 2047 + 0x0c0], %g3
-	ldx	[%sp + 2047 + 0x0c8], %g4
-	ldx	[%sp + 2047 + 0x0d0], %g5
-	ldx	[%sp + 2047 + 0x0d8], %g6
-	ldx	[%sp + 2047 + 0x0e0], %g7
-	wrpr	%l4, 0, %pstate	! restore original pstate
+	mov	%l0, %g4
+	mov	%l1, %g5
+	mov	%l3, %g6
 	ret
 	 restore
 
@@ -91,135 +31,18 @@ prom_cif_interface:
 prom_cif_callback:
 	sethi	%hi(p1275buf), %o1
 	or	%o1, %lo(p1275buf), %o1
-	save	%sp, -0x270, %sp
-	rdpr	%pstate, %l4
-	wrpr	%g0, 0x15, %pstate	! save PROM alternate globals
-	stx	%g1, [%sp + 2047 + 0x0b0]
-	stx	%g2, [%sp + 2047 + 0x0b8]
-	stx	%g3, [%sp + 2047 + 0x0c0]
-	stx	%g4, [%sp + 2047 + 0x0c8]
-	stx	%g5, [%sp + 2047 + 0x0d0]
-	stx	%g6, [%sp + 2047 + 0x0d8]
-	stx	%g7, [%sp + 2047 + 0x0e0]
-					! restore Linux alternate globals
-	ldx	[%sp + 2047 + 0x190], %g1
-	ldx	[%sp + 2047 + 0x198], %g2
-	ldx	[%sp + 2047 + 0x1a0], %g3
-	ldx	[%sp + 2047 + 0x1a8], %g4
-	ldx	[%sp + 2047 + 0x1b0], %g5
-	ldx	[%sp + 2047 + 0x1b8], %g6
-	ldx	[%sp + 2047 + 0x1c0], %g7
-	wrpr	%g0, 0x814, %pstate	! save PROM interrupt globals
-	stx	%g1, [%sp + 2047 + 0x0e8]
-	stx	%g2, [%sp + 2047 + 0x0f0]
-	stx	%g3, [%sp + 2047 + 0x0f8]
-	stx	%g4, [%sp + 2047 + 0x100]
-	stx	%g5, [%sp + 2047 + 0x108]
-	stx	%g6, [%sp + 2047 + 0x110]
-	stx	%g7, [%sp + 2047 + 0x118]
-					! restore Linux interrupt globals
-	ldx	[%sp + 2047 + 0x1c8], %g1
-	ldx	[%sp + 2047 + 0x1d0], %g2
-	ldx	[%sp + 2047 + 0x1d8], %g3
-	ldx	[%sp + 2047 + 0x1e0], %g4
-	ldx	[%sp + 2047 + 0x1e8], %g5
-	ldx	[%sp + 2047 + 0x1f0], %g6
-	ldx	[%sp + 2047 + 0x1f8], %g7
-	wrpr	%g0, 0x14, %pstate	! save PROM normal globals
-	stx	%g1, [%sp + 2047 + 0x120]
-	stx	%g2, [%sp + 2047 + 0x128]
-	stx	%g3, [%sp + 2047 + 0x130]
-	stx	%g4, [%sp + 2047 + 0x138]
-	stx	%g5, [%sp + 2047 + 0x140]
-	stx	%g6, [%sp + 2047 + 0x148]
-	stx	%g7, [%sp + 2047 + 0x150]
-					! restore Linux normal globals
-	ldx	[%sp + 2047 + 0x200], %g1
-	ldx	[%sp + 2047 + 0x208], %g2
-	ldx	[%sp + 2047 + 0x210], %g3
-	ldx	[%sp + 2047 + 0x218], %g4
-	ldx	[%sp + 2047 + 0x220], %g5
-	ldx	[%sp + 2047 + 0x228], %g6
-	ldx	[%sp + 2047 + 0x230], %g7
-	wrpr	%g0, 0x414, %pstate	! save PROM mmu globals
-	stx	%g1, [%sp + 2047 + 0x158]
-	stx	%g2, [%sp + 2047 + 0x160]
-	stx	%g3, [%sp + 2047 + 0x168]
-	stx	%g4, [%sp + 2047 + 0x170]
-	stx	%g5, [%sp + 2047 + 0x178]
-	stx	%g6, [%sp + 2047 + 0x180]
-	stx	%g7, [%sp + 2047 + 0x188]
-					! restore Linux mmu globals
-	ldx	[%sp + 2047 + 0x238], %o0
-	ldx	[%sp + 2047 + 0x240], %o1
-	ldx	[%sp + 2047 + 0x248], %l2
-	ldx	[%sp + 2047 + 0x250], %l3
-	ldx	[%sp + 2047 + 0x258], %l5
-	ldx	[%sp + 2047 + 0x260], %l6
-	ldx	[%sp + 2047 + 0x268], %l7
-					! switch to Linux tba
-	sethi	%hi(sparc64_ttable_tl0), %l1
-	rdpr	%tba, %l0		! save PROM tba
-	mov	%o0, %g1
-	mov	%o1, %g2
-	mov	%l2, %g3
-	mov	%l3, %g4
-	mov	%l5, %g5
-	mov	%l6, %g6
-	mov	%l7, %g7
-	wrpr	%l1, %tba		! install Linux tba
-	wrpr	%l4, 0, %pstate		! restore PSTATE
+	save	%sp, -192, %sp
+	TRAP_LOAD_THREAD_REG(%g6, %g1)
+	LOAD_PER_CPU_BASE(%g5, %g6, %g4, %g3, %o0)
+	ldx	[%g6 + TI_TASK], %g4
 	call	prom_world
-	 mov	%g0, %o0
+	 mov	0, %o0
 	ldx	[%i1 + 0x000], %l2
 	call	%l2
 	 mov	%i0, %o0
 	mov	%o0, %l1
 	call	prom_world
-	 or	%g0, 1, %o0
-	wrpr	%g0, 0x14, %pstate	! interrupts off
-					! restore PROM mmu globals
-	ldx	[%sp + 2047 + 0x158], %o0
-	ldx	[%sp + 2047 + 0x160], %o1
-	ldx	[%sp + 2047 + 0x168], %l2
-	ldx	[%sp + 2047 + 0x170], %l3
-	ldx	[%sp + 2047 + 0x178], %l5
-	ldx	[%sp + 2047 + 0x180], %l6
-	ldx	[%sp + 2047 + 0x188], %l7
-	wrpr	%g0, 0x414, %pstate	! restore PROM mmu globals
-	mov	%o0, %g1
-	mov	%o1, %g2
-	mov	%l2, %g3
-	mov	%l3, %g4
-	mov	%l5, %g5
-	mov	%l6, %g6
-	mov	%l7, %g7
-	wrpr	%l0, %tba		! restore PROM tba
-	wrpr	%g0, 0x14, %pstate	! restore PROM normal globals
-	ldx	[%sp + 2047 + 0x120], %g1
-	ldx	[%sp + 2047 + 0x128], %g2
-	ldx	[%sp + 2047 + 0x130], %g3
-	ldx	[%sp + 2047 + 0x138], %g4
-	ldx	[%sp + 2047 + 0x140], %g5
-	ldx	[%sp + 2047 + 0x148], %g6
-	ldx	[%sp + 2047 + 0x150], %g7
-	wrpr	%g0, 0x814, %pstate	! restore PROM interrupt globals
-	ldx	[%sp + 2047 + 0x0e8], %g1
-	ldx	[%sp + 2047 + 0x0f0], %g2
-	ldx	[%sp + 2047 + 0x0f8], %g3
-	ldx	[%sp + 2047 + 0x100], %g4
-	ldx	[%sp + 2047 + 0x108], %g5
-	ldx	[%sp + 2047 + 0x110], %g6
-	ldx	[%sp + 2047 + 0x118], %g7
-	wrpr	%g0, 0x15, %pstate	! restore PROM alternate globals
-	ldx	[%sp + 2047 + 0x0b0], %g1
-	ldx	[%sp + 2047 + 0x0b8], %g2
-	ldx	[%sp + 2047 + 0x0c0], %g3
-	ldx	[%sp + 2047 + 0x0c8], %g4
-	ldx	[%sp + 2047 + 0x0d0], %g5
-	ldx	[%sp + 2047 + 0x0d8], %g6
-	ldx	[%sp + 2047 + 0x0e0], %g7
-	wrpr	%l4, 0, %pstate
+	 mov	1, %o0
 	ret
 	 restore %l1, 0, %o0
 
diff --git a/include/asm-sparc64/cpudata.h b/include/asm-sparc64/cpudata.h
index 8666440c89af1c7dfc331b047b9f08eb839aacb3..998145b92653d4d5d5d1a7ef5f5678471d78f9db 100644
--- a/include/asm-sparc64/cpudata.h
+++ b/include/asm-sparc64/cpudata.h
@@ -78,6 +78,12 @@ struct gl_1insn_patch_entry {
 	unsigned int	insn;
 };
 extern struct gl_1insn_patch_entry __gl_1insn_patch, __gl_1insn_patch_end;
+
+struct gl_2insn_patch_entry {
+	unsigned int	addr;
+	unsigned int	insns[2];
+};
+extern struct gl_2insn_patch_entry __gl_2insn_patch, __gl_2insn_patch_end;
 #endif /* !(__ASSEMBLY__) */
 
 #define TRAP_PER_CPU_THREAD	0x00