[SPARC64]: Elminate all usage of hard-coded trap globals.

UltraSPARC has special sets of global registers which are switched to
for certain trap types.  There is one set for MMU related traps, one
set of Interrupt Vector processing, and another set (called the
Alternate globals) for all other trap types.

For what seems like forever we've hard coded the values in some of
these trap registers.  Some examples include:

1) Interrupt Vector global %g6 holds current processors interrupt
   work struct where received interrupts are managed for IRQ handler
   dispatch.

2) MMU global %g7 holds the base of the page tables of the currently
   active address space.

3) Alternate global %g6 held the current_thread_info() value.

Such hardcoding has resulted in some serious issues in many areas.
There are some code sequences where having another register available
would help clean up the implementation.  Taking traps such as
cross-calls from the OBP firmware requires some trick code sequences
wherein we have to save away and restore all of the special sets of
global registers when we enter/exit OBP.

We were also using the IMMU TSB register on SMP to hold the per-cpu
area base address, which doesn't work any longer now that we actually
use the TSB facility of the cpu.

The implementation is pretty straight forward.  One tricky bit is
getting the current processor ID as that is different on different cpu
variants.  We use a stub with a fancy calling convention which we
patch at boot time.  The calling convention is that the stub is
branched to and the (PC - 4) to return to is in register %g1.  The cpu
number is left in %g6.  This stub can be invoked by using the
__GET_CPUID macro.

We use an array of per-cpu trap state to store the current thread and
physical address of the current address space's page tables.  The
TRAP_LOAD_THREAD_REG loads %g6 with the current thread from this
table, it uses __GET_CPUID and also clobbers %g1.

TRAP_LOAD_IRQ_WORK is used by the interrupt vector processing to load
the current processor's IRQ software state into %g6.  It also uses
__GET_CPUID and clobbers %g1.

Finally, TRAP_LOAD_PGD_PHYS loads the physical address base of the
current address space's page tables into %g7, it clobbers %g1 and uses
__GET_CPUID.

Many refinements are possible, as well as some tuning, with this stuff
in place.

Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/arch/sparc64/kernel/entry.S b/arch/sparc64/kernel/entry.S
index a73553a..906b64f 100644
--- a/arch/sparc64/kernel/entry.S
+++ b/arch/sparc64/kernel/entry.S
@@ -50,7 +50,8 @@
 	add		%g0, %g0, %g0
 	ba,a,pt		%xcc, rtrap_clr_l6
 
-1:	ldub		[%g6 + TI_FPSAVED], %g5
+1:	TRAP_LOAD_THREAD_REG
+	ldub		[%g6 + TI_FPSAVED], %g5
 	wr		%g0, FPRS_FEF, %fprs
 	andcc		%g5, FPRS_FEF, %g0
 	be,a,pt		%icc, 1f
@@ -189,6 +190,7 @@
 	.globl		do_fpother_check_fitos
 	.align		32
 do_fpother_check_fitos:
+	TRAP_LOAD_THREAD_REG
 	sethi		%hi(fp_other_bounce - 4), %g7
 	or		%g7, %lo(fp_other_bounce - 4), %g7
 
@@ -353,8 +355,6 @@
 	 *
 	 * With this method we can do most of the cross-call tlb/cache
 	 * flushing very quickly.
-	 *
-	 * Current CPU's IRQ worklist table is locked into %g6, don't touch.
 	 */
 	.text
 	.align		32
@@ -378,6 +378,8 @@
 	sllx		%g2, %g4, %g2
 	sllx		%g4, 2, %g4
 
+	TRAP_LOAD_IRQ_WORK
+
 	lduw		[%g6 + %g4], %g5	/* g5 = irq_work(cpu, pil) */
 	stw		%g5, [%g3 + 0x00]	/* bucket->irq_chain = g5 */
 	stw		%g3, [%g6 + %g4]	/* irq_work(cpu, pil) = bucket */
@@ -488,9 +490,24 @@
 	retl
 	 stx		%o1, [%o0 + PT_V9_TSTATE]
 
-	.globl		utrap, utrap_ill
-utrap:	brz,pn		%g1, etrap
+	.globl		utrap_trap
+utrap_trap:		/* %g3=handler,%g4=level */
+	TRAP_LOAD_THREAD_REG
+	ldx		[%g6 + TI_UTRAPS], %g1
+	brnz,pt		%g1, invoke_utrap
 	 nop
+
+	ba,pt		%xcc, etrap
+	 rd		%pc, %g7
+	mov		%l4, %o1
+        call		bad_trap
+	 add		%sp, PTREGS_OFF, %o0
+	ba,pt		%xcc, rtrap
+	 clr		%l6
+
+invoke_utrap:
+	sllx		%g3, 3, %g3
+	ldx		[%g1 + %g3], %g1
 	save		%sp, -128, %sp
 	rdpr		%tstate, %l6
 	rdpr		%cwp, %l7
@@ -500,17 +517,6 @@
 	rdpr		%tnpc, %l7
 	wrpr		%g1, 0, %tnpc
 	done
-utrap_ill:
-        call		bad_trap
-	 add		%sp, PTREGS_OFF, %o0
-	ba,pt		%xcc, rtrap
-	 clr		%l6
-
-	/* XXX Here is stuff we still need to write... -DaveM XXX */
-	.globl		netbsd_syscall
-netbsd_syscall:
-	retl
-	 nop
 
 	/* We need to carefully read the error status, ACK
 	 * the errors, prevent recursive traps, and pass the
@@ -1001,7 +1007,7 @@
 	 * %g3:		scratch
 	 * %g4:		AFSR
 	 * %g5:		AFAR
-	 * %g6:		current thread ptr
+	 * %g6:		unused, will have current thread ptr after etrap
 	 * %g7:		scratch
 	 */
 __cheetah_log_error:
@@ -1690,3 +1696,85 @@
 	 restore	%g0, %g0, %g0
 2:	retl
 	 nop
+
+	/* Read cpu ID from hardware, return in %g6.
+	 * (callers_pc - 4) is in %g1.  Patched at boot time.
+	 *
+	 * Default is spitfire implementation.
+	 *
+	 * The instruction sequence needs to be 5 instructions
+	 * in order to fit the longest implementation, which is
+	 * currently starfire.
+	 */
+	.align		32
+	.globl		__get_cpu_id
+__get_cpu_id:
+	ldxa		[%g0] ASI_UPA_CONFIG, %g6
+	srlx		%g6, 17, %g6
+	jmpl		%g1 + 0x4, %g0
+	 and		%g6, 0x1f, %g6
+	nop
+
+__get_cpu_id_cheetah_safari:
+	ldxa		[%g0] ASI_SAFARI_CONFIG, %g6
+	srlx		%g6, 17, %g6
+	jmpl		%g1 + 0x4, %g0
+	 and		%g6, 0x3ff, %g6
+	nop
+
+__get_cpu_id_cheetah_jbus:
+	ldxa		[%g0] ASI_JBUS_CONFIG, %g6
+	srlx		%g6, 17, %g6
+	jmpl		%g1 + 0x4, %g0
+	 and		%g6, 0x1f, %g6
+	nop
+
+__get_cpu_id_starfire:
+	sethi		%hi(0x1fff40000d0 >> 9), %g6
+	sllx		%g6, 9, %g6
+	or		%g6, 0xd0, %g6
+	jmpl		%g1 + 0x4, %g0
+	 lduwa		[%g6] ASI_PHYS_BYPASS_EC_E, %g6
+
+	.globl		per_cpu_patch
+per_cpu_patch:
+	sethi		%hi(this_is_starfire), %o0
+	lduw		[%o0 + %lo(this_is_starfire)], %o1
+	sethi		%hi(__get_cpu_id_starfire), %o0
+	brnz,pn		%o1, 10f
+	 or		%o0, %lo(__get_cpu_id_starfire), %o0
+	sethi		%hi(tlb_type), %o0
+	lduw		[%o0 + %lo(tlb_type)], %o1
+	brz,pt		%o1, 11f
+	 nop
+	rdpr		%ver, %o0
+	srlx		%o0, 32, %o0
+	sethi		%hi(0x003e0016), %o1
+	or		%o1, %lo(0x003e0016), %o1
+	cmp		%o0, %o1
+	sethi		%hi(__get_cpu_id_cheetah_jbus), %o0
+	be,pn		%icc, 10f
+	 or		%o0, %lo(__get_cpu_id_cheetah_jbus), %o0
+	sethi		%hi(__get_cpu_id_cheetah_safari), %o0
+	or		%o0, %lo(__get_cpu_id_cheetah_safari), %o0
+10:
+	sethi		%hi(__get_cpu_id), %o1
+	or		%o1, %lo(__get_cpu_id), %o1
+	lduw		[%o0 + 0x00], %o2
+	stw		%o2, [%o1 + 0x00]
+	flush		%o1 + 0x00
+	lduw		[%o0 + 0x04], %o2
+	stw		%o2, [%o1 + 0x04]
+	flush		%o1 + 0x04
+	lduw		[%o0 + 0x08], %o2
+	stw		%o2, [%o1 + 0x08]
+	flush		%o1 + 0x08
+	lduw		[%o0 + 0x0c], %o2
+	stw		%o2, [%o1 + 0x0c]
+	flush		%o1 + 0x0c
+	lduw		[%o0 + 0x10], %o2
+	stw		%o2, [%o1 + 0x10]
+	flush		%o1 + 0x10
+11:
+	retl
+	 nop
diff --git a/arch/sparc64/kernel/etrap.S b/arch/sparc64/kernel/etrap.S
index 567dbb7..8b3b6d7 100644
--- a/arch/sparc64/kernel/etrap.S
+++ b/arch/sparc64/kernel/etrap.S
@@ -31,6 +31,7 @@
 		.globl	etrap, etrap_irq, etraptl1
 etrap:		rdpr	%pil, %g2
 etrap_irq:
+		TRAP_LOAD_THREAD_REG
 		rdpr	%tstate, %g1
 		sllx	%g2, 20, %g3
 		andcc	%g1, TSTATE_PRIV, %g0
@@ -98,11 +99,7 @@
 		stx	%i7, [%sp + PTREGS_OFF + PT_V9_I7]
 		wrpr	%g0, ETRAP_PSTATE2, %pstate
 		mov	%l6, %g6
-#ifdef CONFIG_SMP
-#error IMMU TSB usage must be fixed
-		mov	TSB_REG, %g3
-		ldxa	[%g3] ASI_IMMU, %g5
-#endif
+		LOAD_PER_CPU_BASE(%g4, %g3)
 		jmpl	%l2 + 0x4, %g0
 		 ldx	[%g6 + TI_TASK], %g4
 
@@ -126,6 +123,7 @@
 		 *	0x58	TL4's TT
 		 *	0x60	TL
 		 */
+		TRAP_LOAD_THREAD_REG
 		sub	%sp, ((4 * 8) * 4) + 8, %g2
 		rdpr	%tl, %g1
 
@@ -179,7 +177,9 @@
 
 		.align	64
 		.globl	scetrap
-scetrap:	rdpr	%pil, %g2
+scetrap:
+		TRAP_LOAD_THREAD_REG
+		rdpr	%pil, %g2
 		rdpr	%tstate, %g1
 		sllx	%g2, 20, %g3
 		andcc	%g1, TSTATE_PRIV, %g0
@@ -248,11 +248,7 @@
 		stx	%i6, [%sp + PTREGS_OFF + PT_V9_I6]
 		mov	%l6, %g6
 		stx	%i7, [%sp + PTREGS_OFF + PT_V9_I7]
-#ifdef CONFIG_SMP
-#error IMMU TSB usage must be fixed
-		mov	TSB_REG, %g3
-		ldxa	[%g3] ASI_IMMU, %g5
-#endif
+		LOAD_PER_CPU_BASE(%g4, %g3)
 		ldx	[%g6 + TI_TASK], %g4
 		done
 
diff --git a/arch/sparc64/kernel/head.S b/arch/sparc64/kernel/head.S
index d00e206..82ce5bc 100644
--- a/arch/sparc64/kernel/head.S
+++ b/arch/sparc64/kernel/head.S
@@ -26,6 +26,7 @@
 #include <asm/head.h>
 #include <asm/ttable.h>
 #include <asm/mmu.h>
+#include <asm/cpudata.h>
 	
 /* This section from from _start to sparc64_boot_end should fit into
  * 0x0000000000404000 to 0x0000000000408000.
@@ -421,24 +422,6 @@
 	stxa	%g2, [%g1] ASI_DMMU
 	membar	#Sync
 
-	/* The Linux trap handlers expect various trap global registers
-	 * to be setup with some fixed values.  So here we set these
-	 * up very carefully.  These globals are:
-	 *
-	 * Alternate Globals (PSTATE_AG):
-	 *
-	 * %g6			--> current_thread_info()
-	 *
-	 * Interrupt Globals (PSTATE_IG, setup by init_irqwork_curcpu()):
-	 *
-	 * %g6			--> __irq_work[smp_processor_id()]
-	 */
-
-	rdpr	%pstate, %o1
-	mov	%g6, %o2
-	wrpr	%o1, PSTATE_AG, %pstate
-	mov	%o2, %g6
-
 	/* Kill PROM timer */
 	sethi	%hi(0x80000000), %o2
 	sllx	%o2, 32, %o2
@@ -457,7 +440,6 @@
 
 2:
 	wrpr	%g0, %g0, %wstate
-	wrpr	%o1, 0x0, %pstate
 
 	call	init_irqwork_curcpu
 	 nop
diff --git a/arch/sparc64/kernel/irq.c b/arch/sparc64/kernel/irq.c
index f7490ef..3e48af2 100644
--- a/arch/sparc64/kernel/irq.c
+++ b/arch/sparc64/kernel/irq.c
@@ -848,33 +848,9 @@
 
 void init_irqwork_curcpu(void)
 {
-	register struct irq_work_struct *workp asm("o2");
-	register unsigned long tmp asm("o3");
 	int cpu = hard_smp_processor_id();
 
-	memset(__irq_work + cpu, 0, sizeof(*workp));
-
-	/* Make sure we are called with PSTATE_IE disabled.  */
-	__asm__ __volatile__("rdpr	%%pstate, %0\n\t"
-			     : "=r" (tmp));
-	if (tmp & PSTATE_IE) {
-		prom_printf("BUG: init_irqwork_curcpu() called with "
-			    "PSTATE_IE enabled, bailing.\n");
-		__asm__ __volatile__("mov	%%i7, %0\n\t"
-				     : "=r" (tmp));
-		prom_printf("BUG: Called from %lx\n", tmp);
-		prom_halt();
-	}
-
-	/* Set interrupt globals.  */
-	workp = &__irq_work[cpu];
-	__asm__ __volatile__(
-	"rdpr	%%pstate, %0\n\t"
-	"wrpr	%0, %1, %%pstate\n\t"
-	"mov	%2, %%g6\n\t"
-	"wrpr	%0, 0x0, %%pstate\n\t"
-	: "=&r" (tmp)
-	: "i" (PSTATE_IG), "r" (workp));
+	memset(__irq_work + cpu, 0, sizeof(struct irq_work_struct));
 }
 
 /* Only invoked on boot processor. */
diff --git a/arch/sparc64/kernel/rtrap.S b/arch/sparc64/kernel/rtrap.S
index 213eb4a..5a62ec5d 100644
--- a/arch/sparc64/kernel/rtrap.S
+++ b/arch/sparc64/kernel/rtrap.S
@@ -223,12 +223,10 @@
 		ldx			[%sp + PTREGS_OFF + PT_V9_G3], %g3
 		ldx			[%sp + PTREGS_OFF + PT_V9_G4], %g4
 		ldx			[%sp + PTREGS_OFF + PT_V9_G5], %g5
-#ifdef CONFIG_SMP
-#error IMMU TSB usage must be fixed
-		mov			TSB_REG, %g6
-		brnz,a,pn		%l3, 1f
-		 ldxa			[%g6] ASI_IMMU, %g5
-#endif
+		brz,pt			%l3, 1f
+		 nop
+		/* Must do this before thread reg is clobbered below.  */
+		LOAD_PER_CPU_BASE(%g6, %g7)
 1:
 		ldx			[%sp + PTREGS_OFF + PT_V9_G6], %g6
 		ldx			[%sp + PTREGS_OFF + PT_V9_G7], %g7
diff --git a/arch/sparc64/kernel/setup.c b/arch/sparc64/kernel/setup.c
index 158bd31..59a7030 100644
--- a/arch/sparc64/kernel/setup.c
+++ b/arch/sparc64/kernel/setup.c
@@ -507,6 +507,11 @@
 	/* Work out if we are starfire early on */
 	check_if_starfire();
 
+	/* Now we know enough to patch the __get_cpu_id()
+	 * trampoline used by trap code.
+	 */
+	per_cpu_patch();
+
 	boot_flags_init(*cmdline_p);
 
 	idprom_init();
@@ -545,6 +550,9 @@
 	smp_setup_cpu_possible_map();
 
 	paging_init();
+
+	/* Get boot processor trap_block[] setup.  */
+	init_cur_cpu_trap();
 }
 
 static int __init set_preferred_console(void)
diff --git a/arch/sparc64/kernel/smp.c b/arch/sparc64/kernel/smp.c
index d2d3369..8c24585 100644
--- a/arch/sparc64/kernel/smp.c
+++ b/arch/sparc64/kernel/smp.c
@@ -38,6 +38,7 @@
 #include <asm/timer.h>
 #include <asm/starfire.h>
 #include <asm/tlb.h>
+#include <asm/sections.h>
 
 extern void calibrate_delay(void);
 
@@ -87,10 +88,6 @@
 	cpu_data(id).clock_tick = prom_getintdefault(cpu_node,
 						     "clock-frequency", 0);
 
-	cpu_data(id).pgcache_size		= 0;
-	cpu_data(id).pte_cache[0]		= NULL;
-	cpu_data(id).pte_cache[1]		= NULL;
-	cpu_data(id).pgd_cache			= NULL;
 	cpu_data(id).idle_volume		= 1;
 
 	cpu_data(id).dcache_size = prom_getintdefault(cpu_node, "dcache-size",
@@ -121,26 +118,15 @@
 
 extern void inherit_locked_prom_mappings(int save_p);
 
-static inline void cpu_setup_percpu_base(unsigned long cpu_id)
-{
-#error IMMU TSB usage must be fixed
-	__asm__ __volatile__("mov	%0, %%g5\n\t"
-			     "stxa	%0, [%1] %2\n\t"
-			     "membar	#Sync"
-			     : /* no outputs */
-			     : "r" (__per_cpu_offset(cpu_id)),
-			       "r" (TSB_REG), "i" (ASI_IMMU));
-}
-
 void __init smp_callin(void)
 {
 	int cpuid = hard_smp_processor_id();
 
 	inherit_locked_prom_mappings(0);
 
-	__flush_tlb_all();
+	__local_per_cpu_offset = __per_cpu_offset(cpuid);
 
-	cpu_setup_percpu_base(cpuid);
+	__flush_tlb_all();
 
 	smp_setup_percpu_timer();
 
@@ -1107,12 +1093,15 @@
 
 void __devinit smp_prepare_boot_cpu(void)
 {
-	if (hard_smp_processor_id() >= NR_CPUS) {
+	int cpu = hard_smp_processor_id();
+
+	if (cpu >= NR_CPUS) {
 		prom_printf("Serious problem, boot cpu id >= NR_CPUS\n");
 		prom_halt();
 	}
 
-	current_thread_info()->cpu = hard_smp_processor_id();
+	current_thread_info()->cpu = cpu;
+	__local_per_cpu_offset = __per_cpu_offset(cpu);
 
 	cpu_set(smp_processor_id(), cpu_online_map);
 	cpu_set(smp_processor_id(), phys_cpu_present_map);
@@ -1173,12 +1162,9 @@
 {
 	unsigned long goal, size, i;
 	char *ptr;
-	/* Created by linker magic */
-	extern char __per_cpu_start[], __per_cpu_end[];
 
 	/* Copy section for each CPU (we discard the original) */
-	goal = ALIGN(__per_cpu_end - __per_cpu_start, PAGE_SIZE);
-
+	goal = ALIGN(__per_cpu_end - __per_cpu_start, SMP_CACHE_BYTES);
 #ifdef CONFIG_MODULES
 	if (goal < PERCPU_ENOUGH_ROOM)
 		goal = PERCPU_ENOUGH_ROOM;
@@ -1187,31 +1173,10 @@
 	for (size = 1UL; size < goal; size <<= 1UL)
 		__per_cpu_shift++;
 
-	/* Make sure the resulting __per_cpu_base value
-	 * will fit in the 43-bit sign extended IMMU
-	 * TSB register.
-	 */
-	ptr = __alloc_bootmem(size * NR_CPUS, PAGE_SIZE,
-			      (unsigned long) __per_cpu_start);
+	ptr = alloc_bootmem(size * NR_CPUS);
 
 	__per_cpu_base = ptr - __per_cpu_start;
 
-	if ((__per_cpu_shift < PAGE_SHIFT) ||
-	    (__per_cpu_base & ~PAGE_MASK) ||
-	    (__per_cpu_base != (((long) __per_cpu_base << 20) >> 20))) {
-		prom_printf("PER_CPU: Invalid layout, "
-			    "ptr[%p] shift[%lx] base[%lx]\n",
-			    ptr, __per_cpu_shift, __per_cpu_base);
-		prom_halt();
-	}
-
 	for (i = 0; i < NR_CPUS; i++, ptr += size)
 		memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
-
-	/* Finally, load in the boot cpu's base value.
-	 * We abuse the IMMU TSB register for trap handler
-	 * entry and exit loading of %g5.  That is why it
-	 * has to be page aligned.
-	 */
-	cpu_setup_percpu_base(hard_smp_processor_id());
 }
diff --git a/arch/sparc64/kernel/trampoline.S b/arch/sparc64/kernel/trampoline.S
index 782d8c4..18c333f 100644
--- a/arch/sparc64/kernel/trampoline.S
+++ b/arch/sparc64/kernel/trampoline.S
@@ -287,21 +287,18 @@
 	wrpr		%g0, 0, %wstate
 	wrpr		%g0, 0, %tl
 
-	/* Setup the trap globals, then we can resurface. */
-	rdpr		%pstate, %o1
-	mov		%g6, %o2
-	wrpr		%o1, PSTATE_AG, %pstate
+	/* Load TBA, then we can resurface. */
 	sethi		%hi(sparc64_ttable_tl0), %g5
 	wrpr		%g5, %tba
-	mov		%o2, %g6
 
-	wrpr		%o1, 0x0, %pstate
 	ldx		[%g6 + TI_TASK], %g4
 
 	wrpr		%g0, 0, %wstate
 
 	call		init_irqwork_curcpu
 	 nop
+	call		init_cur_cpu_trap
+	 nop
 
 	/* Start using proper page size encodings in ctx register.  */
 	sethi	%hi(sparc64_kern_pri_context), %g3
diff --git a/arch/sparc64/kernel/traps.c b/arch/sparc64/kernel/traps.c
index 8d44ae5..f47f4874 100644
--- a/arch/sparc64/kernel/traps.c
+++ b/arch/sparc64/kernel/traps.c
@@ -2130,7 +2130,22 @@
 	}
 }
 
+struct trap_per_cpu trap_block[NR_CPUS];
+
+/* This can get invoked before sched_init() so play it super safe
+ * and use hard_smp_processor_id().
+ */
+void init_cur_cpu_trap(void)
+{
+	int cpu = hard_smp_processor_id();
+	struct trap_per_cpu *p = &trap_block[cpu];
+
+	p->thread = current_thread_info();
+	p->pgd_paddr = 0;
+}
+
 extern void thread_info_offsets_are_bolixed_dave(void);
+extern void trap_per_cpu_offsets_are_bolixed_dave(void);
 
 /* Only invoked on boot processor. */
 void __init trap_init(void)
@@ -2165,6 +2180,10 @@
 	    (TI_FPREGS & (64 - 1)))
 		thread_info_offsets_are_bolixed_dave();
 
+	if (TRAP_PER_CPU_THREAD != offsetof(struct trap_per_cpu, thread) ||
+	    TRAP_PER_CPU_PGD_PADDR != offsetof(struct trap_per_cpu, pgd_paddr))
+		trap_per_cpu_offsets_are_bolixed_dave();
+
 	/* Attach to the address space of init_task.  On SMP we
 	 * do this in smp.c:smp_callin for other cpus.
 	 */
diff --git a/arch/sparc64/kernel/tsb.S b/arch/sparc64/kernel/tsb.S
index 44b9e6f..50752c5 100644
--- a/arch/sparc64/kernel/tsb.S
+++ b/arch/sparc64/kernel/tsb.S
@@ -36,6 +36,15 @@
 	 nop
 
 tsb_miss_page_table_walk:
+	/* This clobbers %g1 and %g6, preserve them... */
+	mov		%g1, %g5
+	mov		%g6, %g2
+
+	TRAP_LOAD_PGD_PHYS
+
+	mov		%g2, %g6
+	mov		%g5, %g1
+
 	USER_PGTABLE_WALK_TL1(%g4, %g7, %g5, %g2, tsb_do_fault)
 
 tsb_reload:
@@ -112,15 +121,20 @@
 	 * %o0: page table physical address
 	 * %o1:	TSB address
 	 */
+	.align	32
 	.globl	tsb_context_switch
 tsb_context_switch:
-	wrpr	%g0, PSTATE_MG | PSTATE_RMO | PSTATE_PEF | PSTATE_PRIV, %pstate
+	rdpr	%pstate, %o5
+	wrpr	%o5, PSTATE_IE, %pstate
 
-	/* Set page table base alternate global.  */
-	mov	%o0, %g7
+	ldub	[%g6 + TI_CPU], %o3
+	sethi	%hi(trap_block), %o4
+	sllx	%o3, TRAP_BLOCK_SZ_SHIFT, %o3
+	or	%o4, %lo(trap_block), %o4
+	add	%o4, %o3, %o4
+	stx	%o0, [%o4 + TRAP_PER_CPU_PGD_PADDR]
 
-	/* XXX can this happen?  */
-	brz,pn	%o1, 9f
+	brgez	%o1, 9f
 	 nop
 
 	/* Lock TSB into D-TLB.  */
@@ -163,7 +177,7 @@
 	membar		#Sync
 
 9:
-	wrpr	%g0, PSTATE_RMO | PSTATE_PEF | PSTATE_PRIV | PSTATE_IE, %pstate
+	wrpr	%o5, %pstate
 
 	retl
 	 mov	%o2, %o0
diff --git a/arch/sparc64/kernel/ttable.S b/arch/sparc64/kernel/ttable.S
index 56f060c..2fb7a33 100644
--- a/arch/sparc64/kernel/ttable.S
+++ b/arch/sparc64/kernel/ttable.S
@@ -128,7 +128,7 @@
 tl0_resv104:	BTRAP(0x104) BTRAP(0x105) BTRAP(0x106) BTRAP(0x107)
 		.globl tl0_solaris
 tl0_solaris:	SOLARIS_SYSCALL_TRAP
-tl0_netbsd:	NETBSD_SYSCALL_TRAP
+tl0_resv109:	BTRAP(0x109)
 tl0_resv10a:	BTRAP(0x10a) BTRAP(0x10b) BTRAP(0x10c) BTRAP(0x10d) BTRAP(0x10e)
 tl0_resv10f:	BTRAP(0x10f)
 tl0_linux32:	LINUX_32BIT_SYSCALL_TRAP
diff --git a/arch/sparc64/kernel/winfixup.S b/arch/sparc64/kernel/winfixup.S
index f5d93aa..de58803 100644
--- a/arch/sparc64/kernel/winfixup.S
+++ b/arch/sparc64/kernel/winfixup.S
@@ -39,6 +39,7 @@
 	 */
 	.globl	fill_fixup, spill_fixup
 fill_fixup:
+	TRAP_LOAD_THREAD_REG
 	rdpr		%tstate, %g1
 	andcc		%g1, TSTATE_PRIV, %g0
 	or		%g4, FAULT_CODE_WINFIXUP, %g4
@@ -84,11 +85,7 @@
 	wrpr		%l1, (PSTATE_IE | PSTATE_AG | PSTATE_RMO), %pstate
 	mov		%o7, %g6
 	ldx		[%g6 + TI_TASK], %g4
-#ifdef CONFIG_SMP
-#error IMMU TSB usage must be fixed
-	mov		TSB_REG, %g1
-	ldxa		[%g1] ASI_IMMU, %g5
-#endif
+	LOAD_PER_CPU_BASE(%g1, %g2)
 
 	/* This is the same as below, except we handle this a bit special
 	 * since we must preserve %l5 and %l6, see comment above.
@@ -107,6 +104,7 @@
 	 * do not touch %g7 or %g2 so we handle the two cases fine.
 	 */
 spill_fixup:
+	TRAP_LOAD_THREAD_REG
 	ldx		[%g6 + TI_FLAGS], %g1
 	andcc		%g1, _TIF_32BIT, %g0
 	ldub		[%g6 + TI_WSAVED], %g1
@@ -182,6 +180,7 @@
 	wrpr		%g3, %tnpc
 	done
 fill_fixup_mna:
+	TRAP_LOAD_THREAD_REG
 	rdpr		%tstate, %g1
 	andcc		%g1, TSTATE_PRIV, %g0
 	be,pt		%xcc, window_mna_from_user_common
@@ -209,17 +208,14 @@
 	wrpr		%l1, (PSTATE_IE | PSTATE_AG | PSTATE_RMO), %pstate
 	mov		%o7, %g6			! Get current back.
 	ldx		[%g6 + TI_TASK], %g4		! Finish it.
-#ifdef CONFIG_SMP
-#error IMMU TSB usage must be fixed
-	mov		TSB_REG, %g1
-	ldxa		[%g1] ASI_IMMU, %g5
-#endif
+	LOAD_PER_CPU_BASE(%g1, %g2)
 	call		mem_address_unaligned
 	 add		%sp, PTREGS_OFF, %o0
 
 	b,pt		%xcc, rtrap
 	 nop						! yes, the nop is correct
 spill_fixup_mna:
+	TRAP_LOAD_THREAD_REG
 	ldx		[%g6 + TI_FLAGS], %g1
 	andcc		%g1, _TIF_32BIT, %g0
 	ldub		[%g6 + TI_WSAVED], %g1
@@ -287,6 +283,7 @@
 	wrpr		%g3, %tnpc
 	done
 fill_fixup_dax:
+	TRAP_LOAD_THREAD_REG
 	rdpr		%tstate, %g1
 	andcc		%g1, TSTATE_PRIV, %g0
 	be,pt		%xcc, window_dax_from_user_common
@@ -314,17 +311,14 @@
 	wrpr		%l1, (PSTATE_IE | PSTATE_AG | PSTATE_RMO), %pstate
 	mov		%o7, %g6			! Get current back.
 	ldx		[%g6 + TI_TASK], %g4		! Finish it.
-#ifdef CONFIG_SMP
-#error IMMU TSB usage must be fixed
-	mov		TSB_REG, %g1
-	ldxa		[%g1] ASI_IMMU, %g5
-#endif
+	LOAD_PER_CPU_BASE(%g1, %g2)
 	call		spitfire_data_access_exception
 	 add		%sp, PTREGS_OFF, %o0
 
 	b,pt		%xcc, rtrap
 	 nop						! yes, the nop is correct
 spill_fixup_dax:
+	TRAP_LOAD_THREAD_REG
 	ldx		[%g6 + TI_FLAGS], %g1
 	andcc		%g1, _TIF_32BIT, %g0
 	ldub		[%g6 + TI_WSAVED], %g1
diff --git a/arch/sparc64/mm/ultra.S b/arch/sparc64/mm/ultra.S
index 22791f2..a873948 100644
--- a/arch/sparc64/mm/ultra.S
+++ b/arch/sparc64/mm/ultra.S
@@ -295,12 +295,10 @@
 	 *   %g1	address arg 1	(tlb page and range flushes)
 	 *   %g7	address arg 2	(tlb range flush only)
 	 *
-	 *   %g6	ivector table, don't touch
-	 *   %g2	scratch 1
-	 *   %g3	scratch 2
-	 *   %g4	scratch 3
-	 *
-	 * TODO: Make xcall TLB range flushes use the tricks above... -DaveM
+	 *   %g6	scratch 1
+	 *   %g2	scratch 2
+	 *   %g3	scratch 3
+	 *   %g4	scratch 4
 	 */
 	.align		32
 	.globl		xcall_flush_tlb_mm
diff --git a/include/asm-sparc64/cpudata.h b/include/asm-sparc64/cpudata.h
index f7c0faed..6c57cbb 100644
--- a/include/asm-sparc64/cpudata.h
+++ b/include/asm-sparc64/cpudata.h
@@ -1,12 +1,15 @@
 /* cpudata.h: Per-cpu parameters.
  *
- * Copyright (C) 2003, 2005 David S. Miller (davem@redhat.com)
+ * Copyright (C) 2003, 2005, 2006 David S. Miller (davem@davemloft.net)
  */
 
 #ifndef _SPARC64_CPUDATA_H
 #define _SPARC64_CPUDATA_H
 
+#ifndef __ASSEMBLY__
+
 #include <linux/percpu.h>
+#include <linux/threads.h>
 
 typedef struct {
 	/* Dcache line 1 */
@@ -32,4 +35,85 @@
 #define cpu_data(__cpu)		per_cpu(__cpu_data, (__cpu))
 #define local_cpu_data()	__get_cpu_var(__cpu_data)
 
+/* Trap handling code needs to get at a few critical values upon
+ * trap entry and to process TSB misses.  These cannot be in the
+ * per_cpu() area as we really need to lock them into the TLB and
+ * thus make them part of the main kernel image.  As a result we
+ * try to make this as small as possible.
+ *
+ * This is padded out and aligned to 64-bytes to avoid false sharing
+ * on SMP.
+ */
+
+/* If you modify the size of this structure, please update
+ * TRAP_BLOCK_SZ_SHIFT below.
+ */
+struct thread_info;
+struct trap_per_cpu {
+/* D-cache line 1 */
+	struct thread_info	*thread;
+	unsigned long		pgd_paddr;
+	unsigned long		__pad1[2];
+
+/* D-cache line 2 */
+	unsigned long		__pad2[4];
+} __attribute__((aligned(64)));
+extern struct trap_per_cpu trap_block[NR_CPUS];
+extern void init_cur_cpu_trap(void);
+extern void per_cpu_patch(void);
+
+#endif /* !(__ASSEMBLY__) */
+
+#define TRAP_PER_CPU_THREAD	0x00
+#define TRAP_PER_CPU_PGD_PADDR	0x08
+
+#define TRAP_BLOCK_SZ_SHIFT	6
+
+/* Clobbers %g1, loads %g6 with local processor's cpuid */
+#define __GET_CPUID			\
+	ba,pt	%xcc, __get_cpu_id;	\
+	 rd	%pc, %g1;
+
+/* Clobbers %g1, current address space PGD phys address into %g7.  */
+#define TRAP_LOAD_PGD_PHYS			\
+	__GET_CPUID				\
+	sllx	%g6, TRAP_BLOCK_SZ_SHIFT, %g6;	\
+	sethi	%hi(trap_block), %g7;		\
+	or	%g7, %lo(trap_block), %g7;	\
+	add	%g7, %g6, %g7;			\
+	ldx	[%g7 + TRAP_PER_CPU_PGD_PADDR], %g7;
+
+/* Clobbers %g1, loads local processor's IRQ work area into %g6.  */
+#define TRAP_LOAD_IRQ_WORK			\
+	__GET_CPUID				\
+	sethi	%hi(__irq_work), %g1;		\
+	sllx	%g6, 6, %g6;			\
+	or	%g1, %lo(__irq_work), %g1;	\
+	add	%g1, %g6, %g6;
+
+/* Clobbers %g1, loads %g6 with current thread info pointer.  */
+#define TRAP_LOAD_THREAD_REG			\
+	__GET_CPUID				\
+	sllx	%g6, TRAP_BLOCK_SZ_SHIFT, %g6;	\
+	sethi	%hi(trap_block), %g1;		\
+	or	%g1, %lo(trap_block), %g1;	\
+	ldx	[%g1 + %g6], %g6;
+
+/* Given the current thread info pointer in %g6, load the per-cpu
+ * area base of the current processor into %g5.  REG1 and REG2 are
+ * clobbered.
+ */
+#ifdef CONFIG_SMP
+#define LOAD_PER_CPU_BASE(REG1, REG2)			\
+	ldub	[%g6 + TI_CPU], REG1;			\
+	sethi	%hi(__per_cpu_shift), %g5;		\
+	sethi	%hi(__per_cpu_base), REG2;		\
+	ldx	[%g5 + %lo(__per_cpu_shift)], %g5;	\
+	ldx	[REG2 + %lo(__per_cpu_base)], REG2;	\
+	sllx	REG1, %g5, %g5;				\
+	add	%g5, REG2, %g5;
+#else
+#define LOAD_PER_CPU_BASE(REG1, REG2)
+#endif
+
 #endif /* _SPARC64_CPUDATA_H */
diff --git a/include/asm-sparc64/system.h b/include/asm-sparc64/system.h
index af254e5..26c0807 100644
--- a/include/asm-sparc64/system.h
+++ b/include/asm-sparc64/system.h
@@ -209,6 +209,8 @@
 	/* so that ASI is only written if it changes, think again. */	\
 	__asm__ __volatile__("wr %%g0, %0, %%asi"			\
 	: : "r" (__thread_flag_byte_ptr(task_thread_info(next))[TI_FLAG_BYTE_CURRENT_DS]));\
+	trap_block[current_thread_info()->cpu].thread =			\
+		task_thread_info(next);					\
 	__asm__ __volatile__(						\
 	"mov	%%g4, %%g7\n\t"						\
 	"wrpr	%%g0, 0x95, %%pstate\n\t"				\
diff --git a/include/asm-sparc64/ttable.h b/include/asm-sparc64/ttable.h
index 2784f80..f557db4 100644
--- a/include/asm-sparc64/ttable.h
+++ b/include/asm-sparc64/ttable.h
@@ -109,14 +109,14 @@
 	nop;nop;nop;
 	
 #define TRAP_UTRAP(handler,lvl)				\
-	ldx	[%g6 + TI_UTRAPS], %g1;			\
-	sethi	%hi(109f), %g7;				\
-	brz,pn	%g1, utrap;				\
-	 or	%g7, %lo(109f), %g7;			\
-	ba,pt	%xcc, utrap;				\
-109:	 ldx	[%g1 + handler*8], %g1;			\
-	ba,pt	%xcc, utrap_ill;			\
-	 mov	lvl, %o1;
+	mov	handler, %g3;				\
+	ba,pt	%xcc, utrap_trap;			\
+	 mov	lvl, %g4;				\
+	nop;						\
+	nop;						\
+	nop;						\
+	nop;						\
+	nop;
 
 #ifdef CONFIG_SUNOS_EMUL
 #define SUNOS_SYSCALL_TRAP SYSCALL_TRAP(linux_sparc_syscall32, sunos_sys_table)
@@ -136,8 +136,6 @@
 #else
 #define SOLARIS_SYSCALL_TRAP TRAP(solaris_syscall)
 #endif
-/* FIXME: Write these actually */	
-#define NETBSD_SYSCALL_TRAP TRAP(netbsd_syscall)
 #define BREAKPOINT_TRAP TRAP(breakpoint_trap)
 
 #define TRAP_IRQ(routine, level)			\