These changes are the raw update to qemu-2.6.

Collission happened in the following patches:

migration: do cleanup operation after completion(738df5b9)
Bug fix.(1750c932f86)
kvmclock: add a new function to update env->tsc.(b52baab2)

The code provided by the patches was already in the upstreamed
version.

Change-Id: I3cc11841a6a76ae20887b2e245710199e1ea7f9a
Signed-off-by: José Pekkarinen <jose.pekkarinen@nokia.com>

6 files changed, 368 insertions, 407 deletions

diff --git a/qemu/roms/openbios/arch/sparc64/call-client.S b/qemu/roms/openbios/arch/sparc64/call-client.S
index f365e3cb1..a8c0348e4 100644
--- a/qemu/roms/openbios/arch/sparc64/call-client.S
+++ b/qemu/roms/openbios/arch/sparc64/call-client.S
@@ -1,3 +1,5 @@
+#include "cpustate.h"
+
 	.globl	sparc64_of_client_interface, client_tba
 
 
@@ -9,151 +11,9 @@
  * behaviour of OBP.
  */
 
-#define SAVE_WINDOW_STATE(type) \
-	setx	client_window, %g6, %g1; \
-	rdpr	%cwp, %g7; \
-	stx	%g7, [%g1]; \
-	rdpr	%cansave, %g7; \
-	stx	%g7, [%g1 + 0x8]; \
-	rdpr	%canrestore, %g7; \
-	stx	%g7, [%g1 + 0x10]; \
-	rdpr	%otherwin, %g7; \
-	stx	%g7, [%g1 + 0x18]; \
-	rdpr	%wstate, %g7; \
-	stx	%g7, [%g1 + 0x20]; \
-	rdpr	%cleanwin, %g7; \
-	stx	%g7, [%g1 + 0x28]; \
-	\
-	stx	%o0, [%g1 + 0x30]; \
-	stx	%o1, [%g1 + 0x38]; \
-	stx	%o2, [%g1 + 0x40]; \
-	stx	%o3, [%g1 + 0x48]; \
-	stx	%o4, [%g1 + 0x50]; \
-	stx	%o5, [%g1 + 0x58]; \
-	stx	%o6, [%g1 + 0x60]; \
-	stx	%o7, [%g1 + 0x68]; \
-	\
-	rdpr	%pstate, %g7; \
-	stx	%g7, [%g1 + 0x70]; \
-	rd	%y, %g7; \
-	stx	%g7, [%g1 + 0x78]; \
-	rd	%fprs, %g7; \
-	stx	%g7, [%g1 + 0x80]; \
-	\
-	/* Now iterate through all of the windows saving all l and i registers */ \
-	add	%g1, 0x90, %g5; \
-	\
-	/* Get the number of windows in %g6 */ \
-	rdpr	%ver, %g6; \
-	and	%g6, 0xf, %g6; \
-	inc	%g6; \
-	\
-save_cpu_window_##type: \
-	deccc	%g6; \
-	wrpr	%g6, %cwp; \
-	stx	%l0, [%g5]; \
-	stx	%l1, [%g5 + 0x8]; \
-	stx	%l2, [%g5 + 0x10]; \
-	stx	%l3, [%g5 + 0x18]; \
-	stx	%l4, [%g5 + 0x20]; \
-	stx	%l5, [%g5 + 0x28]; \
-	stx	%l6, [%g5 + 0x30]; \
-	stx	%l7, [%g5 + 0x38]; \
-	stx	%i0, [%g5 + 0x40]; \
-	stx	%i1, [%g5 + 0x48]; \
-	stx	%i2, [%g5 + 0x50]; \
-	stx	%i3, [%g5 + 0x58]; \
-	stx	%i4, [%g5 + 0x60]; \
-	stx	%i5, [%g5 + 0x68]; \
-	stx	%i6, [%g5 + 0x70]; \
-	stx	%i7, [%g5 + 0x78]; \
-	bne	save_cpu_window_##type; \
-	 add	%g5, 0x80, %g5; \
-	\
-	/* For 8 windows with 16 registers to save in the window, memory required \
-	is 16*8*8 = 0x400 bytes */ \
-	\
-	/* Now we should be in window 0 so update the other window registers */ \
-	rdpr	%ver, %g6; \
-	and	%g6, 0xf, %g6; \
-	dec	%g6; \
-	wrpr	%g6, %cansave; \
-	\
-	wrpr	%g0, %cleanwin; \
-	wrpr	%g0, %canrestore; \
-	wrpr	%g0, %otherwin;
-
-
-#define RESTORE_WINDOW_STATE(type) \
-	setx	client_window, %g6, %g1; \
-	\
-	/* Get the number of windows in %g6 */ \
-	rdpr	%ver, %g6; \
-	and	%g6, 0xf, %g6; \
-	inc	%g6; \
-	\
-	/* Now iterate through all of the windows restoring all l and i registers */ \
-	add	%g1, 0x90, %g5; \
-	\
-restore_cpu_window_##type: \
-	deccc	%g6; \
-	wrpr	%g6, %cwp; \
-	ldx	[%g5], %l0; \
-	ldx	[%g5 + 0x8], %l1; \
-	ldx	[%g5 + 0x10], %l2; \
-	ldx	[%g5 + 0x18], %l3; \
-	ldx	[%g5 + 0x20], %l4; \
-	ldx	[%g5 + 0x28], %l5; \
-	ldx	[%g5 + 0x30], %l6; \
-	ldx	[%g5 + 0x38], %l7; \
-	ldx	[%g5 + 0x40], %i0; \
-	ldx	[%g5 + 0x48], %i1; \
-	ldx	[%g5 + 0x50], %i2; \
-	ldx	[%g5 + 0x58], %i3; \
-	ldx	[%g5 + 0x60], %i4; \
-	ldx	[%g5 + 0x68], %i5; \
-	ldx	[%g5 + 0x70], %i6; \
-	ldx	[%g5 + 0x78], %i7; \
-	bne	restore_cpu_window_##type; \
-	 add	%g5, 0x80, %g5; \
-	\
-	/* Restore the window registers to their original value */ \
-	ldx	[%g1], %g7; \
-	wrpr	%g7, %cwp; \
-	ldx	[%g1 + 0x8], %g7; \
-	wrpr	%g7, %cansave; \
-	ldx	[%g1 + 0x10], %g7; \
-	wrpr	%g7, %canrestore; \
-	ldx	[%g1 + 0x18], %g7; \
-	wrpr	%g7, %otherwin; \
-	ldx	[%g1 + 0x20], %g7; \
-	wrpr	%g7, %wstate; \
-	ldx	[%g1 + 0x28], %g7; \
-	wrpr	%g7, %cleanwin; \
-	\
-	ldx	[%g1 + 0x30], %o0; \
-	ldx	[%g1 + 0x38], %o1; \
-	ldx	[%g1 + 0x40], %o2; \
-	ldx	[%g1 + 0x48], %o3; \
-	ldx	[%g1 + 0x50], %o4; \
-	ldx	[%g1 + 0x58], %o5; \
-	ldx	[%g1 + 0x60], %o6; \
-	ldx	[%g1 + 0x68], %o7; \
-	\
-	ldx	[%g1 + 0x70], %g7; \
-	wrpr	%g7, %pstate; \
-	ldx	[%g1 + 0x78], %g7; \
-	wr	%g7, 0, %y; \
-	ldx	[%g1 + 0x80], %g7; \
-	wr	%g7, 0, %fprs
-
-
 	.data
 	.align	8
 
-	.skip	16384
-openbios_stack:
-
 client_stack:
 	.xword	0
 client_tba:
@@ -176,15 +36,15 @@ client_window:
 sparc64_of_client_interface:
 
 	/* Save globals on callers stack */
-	add	%sp, -56, %sp
+	add	%sp, -248, %sp
 
-	stx	%g1, [%sp + 2047 + 0]
-	stx	%g2, [%sp + 2047 + 8]
-	stx	%g3, [%sp + 2047 + 16]
-	stx	%g4, [%sp + 2047 + 24]
-	stx	%g5, [%sp + 2047 + 32]
-	stx	%g6, [%sp + 2047 + 40]
-	stx	%g7, [%sp + 2047 + 48]
+	stx	%g1, [%sp + 2047 + 192]
+	stx	%g2, [%sp + 2047 + 200]
+	stx	%g3, [%sp + 2047 + 208]
+	stx	%g4, [%sp + 2047 + 216]
+	stx	%g5, [%sp + 2047 + 224]
+	stx	%g6, [%sp + 2047 + 232]
+	stx	%g7, [%sp + 2047 + 240]
 
 	/* Save client trap table */
 	setx	client_tba, %g6, %g7
@@ -196,22 +56,44 @@ sparc64_of_client_interface:
 	stx	%sp, [%g7]
 
 	/* Save windows */
-	SAVE_WINDOW_STATE(cif)
-
-	/* Move to OpenBIOS stack */
-	setx	openbios_stack - 2047 - 192, %g6, %g7
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g1
+	add	%g1, -CONTEXT_STATE_SIZE, %g1
+	stx	%g1, [%g7]
+	
+	SAVE_CPU_WINDOW_STATE(cif)
+
+	/* Move to OpenBIOS context stack */
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g6
+	setx	CONTEXT_STACK_SIZE, %g4, %g5
+	sub	%g6, %g5, %g6
+	stx	%g6, [%g7]
+	
+	setx	- 2047 - 192, %g6, %g7
+	add	%g1, %g7, %g7
 	mov	%g7, %sp
 
 	/* Call client inteface */
 	call of_client_interface
 	 ldx	[%g1 + 0x30], %o0
 
-	setx	client_window, %g6, %g1
-	stx	%o0, [%g1 + 0x30]
-
 	/* Restore windows */
-	RESTORE_WINDOW_STATE(cif)
-
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g1
+	setx	CONTEXT_STACK_SIZE, %g4, %g5
+	add	%g1, %g5, %g1
+	stx	%g1, [%g7]
+	
+	/* Return value */
+	stx	%o0, [%g1 + 0x30]
+	
+	RESTORE_CPU_WINDOW_STATE(cif)
+	
+	add	%g1, CONTEXT_STATE_SIZE, %g1
+	setx	_fcstack_ptr, %g6, %g7
+	stx	%g1, [%g7]
+	
 	/* Restore stack */
 	setx	client_stack, %g6, %g7
 	ldx	[%g7], %sp
@@ -222,15 +104,15 @@ sparc64_of_client_interface:
 	wrpr	%g6, %tba
 
 	/* Restore globals */
-	ldx	[%sp + 2047 + 0], %g1
-	ldx	[%sp + 2047 + 8], %g2
-	ldx	[%sp + 2047 + 16], %g3
-	ldx	[%sp + 2047 + 24], %g4
-	ldx	[%sp + 2047 + 32], %g5
-	ldx	[%sp + 2047 + 40], %g6
-	ldx	[%sp + 2047 + 48], %g7
-
-	add	%sp, 56, %sp
+	ldx	[%sp + 2047 + 192], %g1
+	ldx	[%sp + 2047 + 200], %g2
+	ldx	[%sp + 2047 + 208], %g3
+	ldx	[%sp + 2047 + 216], %g4
+	ldx	[%sp + 2047 + 224], %g5
+	ldx	[%sp + 2047 + 232], %g6
+	ldx	[%sp + 2047 + 240], %g7
+
+	add	%sp, 248, %sp
 
 	jmp	%o7+8
 	 nop
diff --git a/qemu/roms/openbios/arch/sparc64/context.c b/qemu/roms/openbios/arch/sparc64/context.c
index 2e7668958..98932ee9c 100644
--- a/qemu/roms/openbios/arch/sparc64/context.c
+++ b/qemu/roms/openbios/arch/sparc64/context.c
@@ -40,6 +40,10 @@ static uint8_t image_stack[IMAGE_STACK_SIZE];
 /* Pointer to startup context (physical address) */
 unsigned long __boot_ctx;
 
+/* Pointer to Forth context stack */
+void *_fcstack_ptr = &_efcstack;
+
+
 /*
  * Main starter
  * This is the C function that runs first.
diff --git a/qemu/roms/openbios/arch/sparc64/cpustate.h b/qemu/roms/openbios/arch/sparc64/cpustate.h
new file mode 100644
index 000000000..0c276bfd1
--- /dev/null
+++ b/qemu/roms/openbios/arch/sparc64/cpustate.h
@@ -0,0 +1,244 @@
+/*
+ *   Save/restore CPU state macros
+ *
+ *   Copyright (C) 2015 Mark Cave-Ayland (mark.cave-ayland@ilande.co.uk>)
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   version 2
+ *
+ */
+
+/* State size for context (see below) */
+#define CONTEXT_STATE_SIZE 0x510
+
+/* Stack size for context (allocated inline of the context stack) */
+#define CONTEXT_STACK_SIZE 0x2000
+
+/*
+ * SAVE_CPU_STATE and RESTORE_CPU_STATE are macros used to enable a context switch
+ * to C to occur within the MMU I/D TLB miss handlers.
+ *
+ * Because these handlers are called on a TLB miss, we cannot use flushw to store
+ * processor window state on the stack, as the memory areas used by each window's
+ * stack pointer may not be in the TLB, causing recursive TLB miss traps.
+ *
+ * For this reason, we save window state by manually rotating the window registers
+ * and saving their contents (along with other vital registers) into a special
+ * tlb_handler_stack defined above which is guaranteed to be locked in the TLB, and
+ * so won't cause issues with trap recursion.
+ *
+ * Once this process is complete, we remain in a TL=0, CWP=0 state (with IE=1 to allow
+ * window fill/spill traps if required), switch to our safe tlb_handler_stack and 
+ * invoke the miss handler.
+ */
+
+#define SAVE_CPU_WINDOW_STATE(type) \
+	/* Save window state into context at %g1 */ \
+	rdpr	%cwp, %g7; \
+	stx	%g7, [%g1]; \
+	rdpr	%cansave, %g7; \
+	stx	%g7, [%g1 + 0x8]; \
+	rdpr	%canrestore, %g7; \
+	stx	%g7, [%g1 + 0x10]; \
+	rdpr	%otherwin, %g7; \
+	stx	%g7, [%g1 + 0x18]; \
+	rdpr	%wstate, %g7; \
+	stx	%g7, [%g1 + 0x20]; \
+	rdpr	%cleanwin, %g7; \
+	stx	%g7, [%g1 + 0x28]; \
+	\
+	stx	%o0, [%g1 + 0x30]; \
+	stx	%o1, [%g1 + 0x38]; \
+	stx	%o2, [%g1 + 0x40]; \
+	stx	%o3, [%g1 + 0x48]; \
+	stx	%o4, [%g1 + 0x50]; \
+	stx	%o5, [%g1 + 0x58]; \
+	stx	%o6, [%g1 + 0x60]; \
+	stx	%o7, [%g1 + 0x68]; \
+	\
+	rdpr	%pstate, %g7; \
+	stx	%g7, [%g1 + 0x70]; \
+	rd	%y, %g7; \
+	stx	%g7, [%g1 + 0x78]; \
+	rd	%fprs, %g7; \
+	stx	%g7, [%g1 + 0x80]; \
+	rdpr    %tl, %g7; \
+	stx     %g7, [%g1 + 0x88]; \
+	\
+	/* Now iterate through all of the windows saving all l and i registers */ \
+	add	%g1, 0x90, %g5; \
+	\
+	/* Get the number of windows in %g6 */ \
+	rdpr	%ver, %g6; \
+	and	%g6, 0xf, %g6; \
+	\
+	mov     %g6, %g4; \
+	inc     %g4; \
+	\
+	/* Starting cwp in g7 */ \
+	rdpr    %cwp, %g7; \
+	\
+save_cpu_window_##type: \
+	wrpr    %g7, %cwp; \
+	stx	%l0, [%g5]; \
+	stx	%l1, [%g5 + 0x8]; \
+	stx	%l2, [%g5 + 0x10]; \
+	stx	%l3, [%g5 + 0x18]; \
+	stx	%l4, [%g5 + 0x20]; \
+	stx	%l5, [%g5 + 0x28]; \
+	stx	%l6, [%g5 + 0x30]; \
+	stx	%l7, [%g5 + 0x38]; \
+	stx	%i0, [%g5 + 0x40]; \
+	stx	%i1, [%g5 + 0x48]; \
+	stx	%i2, [%g5 + 0x50]; \
+	stx	%i3, [%g5 + 0x58]; \
+	stx	%i4, [%g5 + 0x60]; \
+	stx	%i5, [%g5 + 0x68]; \
+	stx	%i6, [%g5 + 0x70]; \
+	stx	%i7, [%g5 + 0x78]; \
+	dec	%g7; \
+	and	%g7, %g6, %g7; \
+	subcc	%g4, 1, %g4; \
+	bne	save_cpu_window_##type; \
+	 add	%g5, 0x80, %g5; \
+	\
+	/* For 8 windows with 16 registers to save in the window, memory required \
+	is 16*8*8 = 0x400 bytes */ \
+	\
+	/* Now we should be in window 0 so update the other window registers */ \
+	rdpr	%ver, %g6; \
+	and	%g6, 0xf, %g6; \
+	dec	%g6; \
+	wrpr	%g6, %cansave; \
+	\
+	wrpr	%g0, %cleanwin; \
+	wrpr	%g0, %canrestore; \
+	wrpr	%g0, %otherwin; \
+
+	
+#define SAVE_CPU_TRAP_STATE(type) \
+	/* Save trap state into context at %g1 */ \
+	add	%g1, 0x490, %g5; \
+	mov	4, %g6; \
+	\
+save_trap_state_##type: \
+	deccc	%g6; \
+	wrpr	%g6, %tl; \
+	rdpr	%tpc, %g7; \
+	stx	%g7, [%g5]; \
+	rdpr	%tnpc, %g7; \
+	stx	%g7, [%g5 + 0x8]; \
+	rdpr	%tstate, %g7; \
+	stx	%g7, [%g5 + 0x10]; \
+	rdpr	%tt, %g7; \
+	stx	%g7, [%g5 + 0x18]; \
+	bne	save_trap_state_##type; \
+	 add	%g5, 0x20, %g5; \
+	\
+	/* For 4 trap levels with 4 registers, memory required is \
+	4*8*4 = 0x80 bytes */
+
+/* Save all state into context at %g1 */
+#define SAVE_CPU_STATE(type) \
+	SAVE_CPU_WINDOW_STATE(type); \
+	SAVE_CPU_TRAP_STATE(type);
+
+
+#define RESTORE_CPU_WINDOW_STATE(type) \
+	/* Restore window state from context at %g1 */ \
+	\
+	/* Get the number of windows in %g6 */ \
+	rdpr	%ver, %g6; \
+	and	%g6, 0xf, %g6; \
+	\
+	mov	%g6, %g4; \
+	inc	%g4; \
+	\
+	/* Set starting window */ \
+	ldx	[%g1], %g7; \
+	\
+	/* Now iterate through all of the windows restoring all l and i registers */ \
+	add	%g1, 0x90, %g5; \
+	\
+restore_cpu_window_##type: \
+	wrpr	%g7, %cwp; \
+	ldx	[%g5], %l0; \
+	ldx	[%g5 + 0x8], %l1; \
+	ldx	[%g5 + 0x10], %l2; \
+	ldx	[%g5 + 0x18], %l3; \
+	ldx	[%g5 + 0x20], %l4; \
+	ldx	[%g5 + 0x28], %l5; \
+	ldx	[%g5 + 0x30], %l6; \
+	ldx	[%g5 + 0x38], %l7; \
+	ldx	[%g5 + 0x40], %i0; \
+	ldx	[%g5 + 0x48], %i1; \
+	ldx	[%g5 + 0x50], %i2; \
+	ldx	[%g5 + 0x58], %i3; \
+	ldx	[%g5 + 0x60], %i4; \
+	ldx	[%g5 + 0x68], %i5; \
+	ldx	[%g5 + 0x70], %i6; \
+	ldx	[%g5 + 0x78], %i7; \
+	dec	%g7; \
+	and	%g7, %g6, %g7; \
+	subcc	%g4, 1, %g4; \
+	bne	restore_cpu_window_##type; \
+	 add	%g5, 0x80, %g5; \
+	\
+	/* Restore the window registers to their original value */ \
+	ldx	[%g1], %g7; \
+	wrpr	%g7, %cwp; \
+	ldx	[%g1 + 0x8], %g7; \
+	wrpr	%g7, %cansave; \
+	ldx	[%g1 + 0x10], %g7; \
+	wrpr	%g7, %canrestore; \
+	ldx	[%g1 + 0x18], %g7; \
+	wrpr	%g7, %otherwin; \
+	ldx	[%g1 + 0x20], %g7; \
+	wrpr	%g7, %wstate; \
+	ldx	[%g1 + 0x28], %g7; \
+	wrpr	%g7, %cleanwin; \
+	\
+	ldx	[%g1 + 0x30], %o0; \
+	ldx	[%g1 + 0x38], %o1; \
+	ldx	[%g1 + 0x40], %o2; \
+	ldx	[%g1 + 0x48], %o3; \
+	ldx	[%g1 + 0x50], %o4; \
+	ldx	[%g1 + 0x58], %o5; \
+	ldx	[%g1 + 0x60], %o6; \
+	ldx	[%g1 + 0x68], %o7; \
+	\
+	ldx	[%g1 + 0x70], %g7; \
+	wrpr	%g7, %pstate; \
+	ldx	[%g1 + 0x78], %g7; \
+	wr	%g7, 0, %y; \
+	ldx	[%g1 + 0x80], %g7; \
+	wr	%g7, 0, %fprs; \
+
+
+#define RESTORE_CPU_TRAP_STATE(type) \
+	/* Restore trap state from context at %g1 */ \
+	add	%g1, 0x490, %g5; \
+	mov	4, %g6; \
+	\
+restore_trap_state_##type: \
+	deccc	%g6; \
+	wrpr	%g6, %tl; \
+	ldx	[%g5], %g7; \
+	wrpr	%g7, %tpc; \
+	ldx	[%g5 + 0x8], %g7; \
+	wrpr	%g7, %tnpc; \
+	ldx	[%g5 + 0x10], %g7; \
+	wrpr	%g7, %tstate; \
+	ldx	[%g5 + 0x18], %g7; \
+	wrpr	%g7, %tt; \
+	bne	restore_trap_state_##type; \
+	 add	%g5, 0x20, %g5; \
+	\
+	ldx	[%g1 + 0x88], %g7; \
+	wrpr	%g7, %tl
+
+/* Restore all state from context at %g1 */
+#define RESTORE_CPU_STATE(type) \
+	RESTORE_CPU_WINDOW_STATE(type); \
+	RESTORE_CPU_TRAP_STATE(type);
diff --git a/qemu/roms/openbios/arch/sparc64/ldscript b/qemu/roms/openbios/arch/sparc64/ldscript
index 54288e825..c5cc6a5ce 100644
--- a/qemu/roms/openbios/arch/sparc64/ldscript
+++ b/qemu/roms/openbios/arch/sparc64/ldscript
@@ -50,6 +50,11 @@ SECTIONS
 	*(.bss.*)
 	*(COMMON)
 
+	_fcstack = .;
+	. += 32768;
+	. = ALIGN(16);
+	_efcstack = .;
+	
 	_stack = .;
 	. += STACK_SIZE;
 	. = ALIGN(16);
diff --git a/qemu/roms/openbios/arch/sparc64/lib.c b/qemu/roms/openbios/arch/sparc64/lib.c
index e9101af52..4709ca8fe 100644
--- a/qemu/roms/openbios/arch/sparc64/lib.c
+++ b/qemu/roms/openbios/arch/sparc64/lib.c
@@ -458,10 +458,10 @@ NODE_METHODS(mmu) = {
 void ob_mmu_init(const char *cpuname, uint64_t ram_size)
 {
     /* memory node */
-    REGISTER_NODE_METHODS(memory, "/memory");
+    REGISTER_NODE(memory);
 
     /* MMU node */
-    REGISTER_NODE_METHODS(mmu, "/virtual-memory");
+    REGISTER_NODE(mmu);
 
     ofmem_register(find_dev("/memory"), find_dev("/virtual-memory"));
 
diff --git a/qemu/roms/openbios/arch/sparc64/vectors.S b/qemu/roms/openbios/arch/sparc64/vectors.S
index 927c1cdc7..9d86b6bd8 100644
--- a/qemu/roms/openbios/arch/sparc64/vectors.S
+++ b/qemu/roms/openbios/arch/sparc64/vectors.S
@@ -24,6 +24,7 @@
  */
 
 #define __ASSEMBLY__
+#include "cpustate.h"
 #include "pstate.h"
 #include <asm/asi.h>
 #define ASI_BP ASI_PHYS_BYPASS_EC_E
@@ -276,17 +277,8 @@ tl1_resv1f0:	BTRAPS(0x1f0) BTRAPS(0x1f8)
 
 	.section ".data"
 	.align 8
-	.globl tlb_handler_stack_top, tlb_handler_stack_pointer, obp_ticks_pointer
+	.globl obp_ticks_pointer
 
-	! Stack for the tlb MMU trap handlers
-tlb_handler_stack_bottom:
-	.skip 8192
-tlb_handler_stack_top:
-	.skip 8
-
-	! MMU trap handler stack pointer
-tlb_handler_stack_pointer:
-	.xword tlb_handler_stack_top
 
 	! Pointer to current tick value
 obp_ticks_pointer:
@@ -336,234 +328,30 @@ fill_32bit:
         restored
         retry
 
-/*
- * SAVE_CPU_STATE and RESTORE_CPU_STATE are macros used to enable a context switch
- * to C to occur within the MMU I/D TLB miss handlers.
- *
- * Because these handlers are called on a TLB miss, we cannot use flushw to store
- * processor window state on the stack, as the memory areas used by each window's
- * stack pointer may not be in the TLB, causing recursive TLB miss traps.
- *
- * For this reason, we save window state by manually rotating the window registers
- * and saving their contents (along with other vital registers) into a special
- * tlb_handler_stack defined above which is guaranteed to be locked in the TLB, and
- * so won't cause issues with trap recursion.
- *
- * Once this process is complete, we remain in a TL=0, CWP=0 state (with IE=1 to allow
- * window fill/spill traps if required), switch to our safe tlb_handler_stack and 
- * invoke the miss handler.
- */
-
-#define SAVE_CPU_STATE(type) \
-	/* Set up our exception stack pointer in %g1 */ \
-	setx	tlb_handler_stack_pointer, %g7, %g6; \
-	ldx	[%g6], %g1; \
-	add	%g1, -0x510, %g1; \
-	\
-	/* First save the various state registers */ \
-	rdpr	%cwp, %g7; \
-	stx	%g7, [%g1]; \
-	rdpr	%cansave, %g7; \
-	stx	%g7, [%g1 + 0x8]; \
-	rdpr	%canrestore, %g7; \
-	stx	%g7, [%g1 + 0x10]; \
-	rdpr	%otherwin, %g7; \
-	stx	%g7, [%g1 + 0x18]; \
-	rdpr	%wstate, %g7; \
-	stx	%g7, [%g1 + 0x20]; \
-	rdpr	%cleanwin, %g7; \
-	stx	%g7, [%g1 + 0x28]; \
-	rdpr	%pstate, %g7; \
-	stx	%g7, [%g1 + 0x30]; \
-	\
-	rd	%y, %g7; \
-	stx	%g7, [%g1 + 0x38]; \
-	rd	%fprs, %g7; \
-	stx	%g7, [%g1 + 0x40]; \
-	\
-	rdpr	%tl, %g7; \
-	stx	%g7, [%g1 + 0x48]; \
-	\
-	/* Trap state */ \
-	add	%g1, 0x50, %g5; \
-	mov	4, %g6; \
-	\
-save_trap_state_##type: \
-	deccc	%g6; \
-	wrpr	%g6, %tl; \
-	rdpr	%tpc, %g7; \
-	stx	%g7, [%g5]; \
-	rdpr	%tnpc, %g7; \
-	stx	%g7, [%g5 + 0x8]; \
-	rdpr	%tstate, %g7; \
-	stx	%g7, [%g5 + 0x10]; \
-	rdpr	%tt, %g7; \
-	stx	%g7, [%g5 + 0x18]; \
-	bne	save_trap_state_##type; \
-	 add	%g5, 0x20, %g5; \
-	\
-	/* For 4 trap levels with 4 registers, memory required is 
-	4*8*4 = 0x80 bytes */ \
-	\
-	/* Save the o registers */ \
-	stx	%o0, [%g1 + 0xd0]; \
-	stx	%o1, [%g1 + 0xd8]; \
-	stx	%o2, [%g1 + 0xe0]; \
-	stx	%o3, [%g1 + 0xe8]; \
-	stx	%o4, [%g1 + 0xf0]; \
-	stx	%o5, [%g1 + 0xf8]; \
-	stx	%o6, [%g1 + 0x100]; \
-	stx	%o7, [%g1 + 0x108]; \
-	\
-	/* Now iterate through all of the windows saving all l and i registers */ \
-	add	%g1, 0x110, %g5; \
-	\
-	/* Get the number of windows in %g6 */ \
-	rdpr	%ver, %g6; \
-	and	%g6, 0xf, %g6; \
-	inc	%g6; \
-	\
-save_cpu_window_##type: \
-	deccc	%g6; \
-	wrpr	%g6, %cwp; \
-	stx	%l0, [%g5]; \
-	stx	%l1, [%g5 + 0x8]; \
-	stx	%l2, [%g5 + 0x10]; \
-	stx	%l3, [%g5 + 0x18]; \
-	stx	%l4, [%g5 + 0x20]; \
-	stx	%l5, [%g5 + 0x28]; \
-	stx	%l6, [%g5 + 0x30]; \
-	stx	%l7, [%g5 + 0x38]; \
-	stx	%i0, [%g5 + 0x40]; \
-	stx	%i1, [%g5 + 0x48]; \
-	stx	%i2, [%g5 + 0x50]; \
-	stx	%i3, [%g5 + 0x58]; \
-	stx	%i4, [%g5 + 0x60]; \
-	stx	%i5, [%g5 + 0x68]; \
-	stx	%i6, [%g5 + 0x70]; \
-	stx	%i7, [%g5 + 0x78]; \
-	bne	save_cpu_window_##type; \
-	 add	%g5, 0x80, %g5; \
-	\
-	/* For 8 windows with 16 registers to save in the window, memory required
-	is 16*8*8 = 0x400 bytes */ \
-	\
-	/* Now we should be in window 0 so update the other window registers */ \
-	rdpr	%ver, %g6; \
-	and	%g6, 0xf, %g6; \
-	dec	%g6; \
-	wrpr	%g6, %cansave; \
-	\
-	wrpr	%g0, %cleanwin; \
-	wrpr	%g0, %canrestore; \
-	wrpr	%g0, %otherwin; \
-	\
-	/* Update our exception stack pointer */ \
-	setx	tlb_handler_stack_pointer, %g7, %g6; \
-	stx	%g1, [%g6];
-
-
-#define RESTORE_CPU_STATE(type) \
-	/* Set up our exception stack pointer in %g1 */ \
-	setx	tlb_handler_stack_pointer, %g7, %g6; \
-	ldx	[%g6], %g1; \
-	\
-	/* Get the number of windows in %g6 */ \
-	rdpr	%ver, %g6; \
-	and	%g6, 0xf, %g6; \
-	inc	%g6; \
-	\
-	/* Now iterate through all of the windows restoring all l and i registers */ \
-	add	%g1, 0x110, %g5; \
-	\
-restore_cpu_window_##type: \
-	deccc	%g6; \
-	wrpr	%g6, %cwp; \
-	ldx	[%g5], %l0; \
-	ldx	[%g5 + 0x8], %l1; \
-	ldx	[%g5 + 0x10], %l2; \
-	ldx	[%g5 + 0x18], %l3; \
-	ldx	[%g5 + 0x20], %l4; \
-	ldx	[%g5 + 0x28], %l5; \
-	ldx	[%g5 + 0x30], %l6; \
-	ldx	[%g5 + 0x38], %l7; \
-	ldx	[%g5 + 0x40], %i0; \
-	ldx	[%g5 + 0x48], %i1; \
-	ldx	[%g5 + 0x50], %i2; \
-	ldx	[%g5 + 0x58], %i3; \
-	ldx	[%g5 + 0x60], %i4; \
-	ldx	[%g5 + 0x68], %i5; \
-	ldx	[%g5 + 0x70], %i6; \
-	ldx	[%g5 + 0x78], %i7; \
-	bne	restore_cpu_window_##type; \
-	 add	%g5, 0x80, %g5; \
-	\
-	/* Restore the window registers to their original value */ \
-	ldx	[%g1], %g7; \
-	wrpr	%g7, %cwp; \
-	ldx	[%g1 + 0x8], %g7; \
-	wrpr	%g7, %cansave; \
-	ldx	[%g1 + 0x10], %g7; \
-	wrpr	%g7, %canrestore; \
-	ldx	[%g1 + 0x18], %g7; \
-	wrpr	%g7, %otherwin; \
-	ldx	[%g1 + 0x20], %g7; \
-	wrpr	%g7, %wstate; \
-	ldx	[%g1 + 0x28], %g7; \
-	wrpr	%g7, %cleanwin; \
-	ldx	[%g1 + 0x30], %g7; \
-	wrpr	%g7, %pstate; \
-	\
-	/* Restore the o registers */ \
-	ldx	[%g1 + 0xd0], %o0; \
-	ldx	[%g1 + 0xd8], %o1; \
-	ldx	[%g1 + 0xe0], %o2; \
-	ldx	[%g1 + 0xe8], %o3; \
-	ldx	[%g1 + 0xf0], %o4; \
-	ldx	[%g1 + 0xf8], %o5; \
-	ldx	[%g1 + 0x100], %o6; \
-	ldx	[%g1 + 0x108], %o7; \
-	\
-	/* Restore the trap state */ \
-	add	%g1, 0x50, %g5; \
-	mov	4, %g6; \
-	\
-restore_trap_state_##type: \
-	deccc	%g6; \
-	wrpr	%g6, %tl; \
-	ldx	[%g5], %g7; \
-	wrpr	%g7, %tpc; \
-	ldx	[%g5 + 0x8], %g7; \
-	wrpr	%g7, %tnpc; \
-	ldx	[%g5 + 0x10], %g7; \
-	wrpr	%g7, %tstate; \
-	ldx	[%g5 + 0x18], %g7; \
-	wrpr	%g7, %tt; \
-	bne	restore_trap_state_##type; \
-	 add	%g5, 0x20, %g5; \
-	\
-	ldx	[%g1 + 0x38], %g7; \
-	wr	%g7, 0, %y; \
-	ldx	[%g1 + 0x40], %g7; \
-	wr	%g7, 0, %fprs; \
-	ldx	[%g1 + 0x48], %g7; \
-	wrpr	%g7, %tl; \
-	\
-	/* Restore exception stack pointer to previous value */ \
-	setx	tlb_handler_stack_pointer, %g7, %g6; \
-	add	%g1, 0x510, %g1; \
-	stx	%g1, [%g6];
-
 
         .globl reload_DMMU_tlb, reload_IMMU_tlb, bug
 
 reload_DMMU_tlb:
-
+	
+	/* Save CPU state to stack */
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g1
+	add	%g1, -CONTEXT_STATE_SIZE, %g1
+	stx	%g1, [%g7]
+	
 	SAVE_CPU_STATE(dtlb)
 
-	/* Switch to TLB locked stack space (note we add an additional 192 bytes required for
+	/* Switch to 8K TLB locked OpenBIOS stack (note we add an additional 192 bytes required for
 	   gcc to save its arguments when building with -O0) */
-	add 	%g1, -STACK_BIAS - 192, %sp
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g6
+	setx	CONTEXT_STACK_SIZE, %g4, %g5
+	sub	%g6, %g5, %g6
+	stx	%g6, [%g7]
+	
+	setx	- 2047 - 192, %g6, %g7
+	add	%g1, %g7, %g7
+	mov	%g7, %sp
 
 	/* Enable interrupts for window spill/fill traps */
 	rdpr	%pstate, %g7
@@ -577,18 +365,44 @@ reload_DMMU_tlb:
 	rdpr	%pstate, %g7
 	andn	%g7, PSTATE_IE, %g7
 	wrpr	%g7, %pstate
-
+	
+	/* Restore CPU state from stack */
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g1
+	setx	CONTEXT_STACK_SIZE, %g4, %g5
+	add	%g1, %g5, %g1
+	stx	%g1, [%g7]
+	
 	RESTORE_CPU_STATE(dtlb)
-
+	
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g1
+	add	%g1, CONTEXT_STATE_SIZE, %g1
+	stx	%g1, [%g7]
+	
         retry
 
 reload_IMMU_tlb:
-
+	
+	/* Save CPU state to stack */
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g1
+	add	%g1, -CONTEXT_STATE_SIZE, %g1
+	stx	%g1, [%g7]
+	
 	SAVE_CPU_STATE(itlb)
 
-	/* Switch to TLB locked stack space (note we add an additional 192 bytes required for
+	/* Switch to 8K TLB locked OpenBIOS stack (note we add an additional 192 bytes required for
 	   gcc to save its arguments when building with -O0) */
-	add 	%g1, -STACK_BIAS - 192, %sp
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g6
+	setx	CONTEXT_STACK_SIZE, %g4, %g5
+	sub	%g6, %g5, %g6
+	stx	%g6, [%g7]
+	
+	setx	- 2047 - 192, %g6, %g7
+	add	%g1, %g7, %g7
+	mov	%g7, %sp
 
 	/* Enable interrupts for window spill/fill traps */
 	rdpr	%pstate, %g7
@@ -602,8 +416,20 @@ reload_IMMU_tlb:
 	rdpr	%pstate, %g7
 	andn	%g7, PSTATE_IE, %g7
 	wrpr	%g7, %pstate
-
+	
+	/* Restore CPU state from stack */
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g1
+	setx	CONTEXT_STACK_SIZE, %g4, %g5
+	add	%g1, %g5, %g1
+	stx	%g1, [%g7]
+	
 	RESTORE_CPU_STATE(itlb)
+	
+	setx	_fcstack_ptr, %g6, %g7
+	ldx	[%g7], %g1
+	add	%g1, CONTEXT_STATE_SIZE, %g1
+	stx	%g1, [%g7]
 
         retry