Merge branch 'upstream'

14 files changed, 422 insertions, 509 deletions

diff --git a/arch/sparc64/kernel/dtlb_base.S b/arch/sparc64/kernel/dtlb_base.S
index 702d349c1e88..6528786840c0 100644
--- a/arch/sparc64/kernel/dtlb_base.S
+++ b/arch/sparc64/kernel/dtlb_base.S
@@ -53,19 +53,18 @@
  * be guaranteed to be 0 ... mmu_context.h does guarantee this
  * by only using 10 bits in the hwcontext value.
  */
-#define CREATE_VPTE_OFFSET1(r1, r2)
+#define CREATE_VPTE_OFFSET1(r1, r2) nop
 #define CREATE_VPTE_OFFSET2(r1, r2) \
 				srax	r1, 10, r2
-#define CREATE_VPTE_NOP		nop
 #else
 #define CREATE_VPTE_OFFSET1(r1, r2) \
 				srax	r1, PAGE_SHIFT, r2
 #define CREATE_VPTE_OFFSET2(r1, r2) \
 				sllx	r2, 3, r2
-#define CREATE_VPTE_NOP
 #endif
 
 /* DTLB ** ICACHE line 1: Quick user TLB misses		*/
+	mov		TLB_SFSR, %g1
 	ldxa		[%g1 + %g1] ASI_DMMU, %g4	! Get TAG_ACCESS
 	andcc		%g4, TAG_CONTEXT_BITS, %g0	! From Nucleus?
 from_tl1_trap:
@@ -74,18 +73,16 @@ from_tl1_trap:
 	be,pn		%xcc, kvmap			! Yep, special processing
 	 CREATE_VPTE_OFFSET2(%g4, %g6)			! Create VPTE offset
 	cmp		%g5, 4				! Last trap level?
-	be,pn		%xcc, longpath			! Yep, cannot risk VPTE miss
-	 nop						! delay slot
 
 /* DTLB ** ICACHE line 2: User finish + quick kernel TLB misses	*/
+	be,pn		%xcc, longpath			! Yep, cannot risk VPTE miss
+	 nop						! delay slot
 	ldxa		[%g3 + %g6] ASI_S, %g5		! Load VPTE
 1:	brgez,pn	%g5, longpath			! Invalid, branch out
 	 nop						! Delay-slot
 9:	stxa		%g5, [%g0] ASI_DTLB_DATA_IN	! Reload TLB
 	retry						! Trap return
 	nop
-	nop
-	nop
 
 /* DTLB ** ICACHE line 3: winfixups+real_faults		*/
 longpath:
@@ -106,8 +103,7 @@ longpath:
 	nop
 	nop
 	nop
-	CREATE_VPTE_NOP
+	nop
 
 #undef CREATE_VPTE_OFFSET1
 #undef CREATE_VPTE_OFFSET2
-#undef CREATE_VPTE_NOP
diff --git a/arch/sparc64/kernel/dtlb_prot.S b/arch/sparc64/kernel/dtlb_prot.S
index d848bb7374bb..e0a920162604 100644
--- a/arch/sparc64/kernel/dtlb_prot.S
+++ b/arch/sparc64/kernel/dtlb_prot.S
@@ -14,14 +14,14 @@
  */
 
 /* PROT ** ICACHE line 1: User DTLB protection trap	*/
-	stxa		%g0, [%g1] ASI_DMMU		! Clear SFSR FaultValid bit
-	membar		#Sync				! Synchronize ASI stores
-	rdpr		%pstate, %g5			! Move into alternate globals
+	mov		TLB_SFSR, %g1
+	stxa		%g0, [%g1] ASI_DMMU		! Clear FaultValid bit
+	membar		#Sync				! Synchronize stores
+	rdpr		%pstate, %g5			! Move into alt-globals
 	wrpr		%g5, PSTATE_AG|PSTATE_MG, %pstate
-	rdpr		%tl, %g1			! Need to do a winfixup?
+	rdpr		%tl, %g1			! Need a winfixup?
 	cmp		%g1, 1				! Trap level >1?
-	mov		TLB_TAG_ACCESS, %g4		! Prepare reload of vaddr
-	nop
+	mov		TLB_TAG_ACCESS, %g4		! For reload of vaddr
 
 /* PROT ** ICACHE line 2: More real fault processing */
 	bgu,pn		%xcc, winfix_trampoline		! Yes, perform winfixup
diff --git a/arch/sparc64/kernel/entry.S b/arch/sparc64/kernel/entry.S
index f685035dbdb8..11a848402fb1 100644
--- a/arch/sparc64/kernel/entry.S
+++ b/arch/sparc64/kernel/entry.S
@@ -33,7 +33,7 @@
 	/* This is trivial with the new code... */
 	.globl		do_fpdis
 do_fpdis:
-	sethi		%hi(TSTATE_PEF), %g4					! IEU0
+	sethi		%hi(TSTATE_PEF), %g4
 	rdpr		%tstate, %g5
 	andcc		%g5, %g4, %g0
 	be,pt		%xcc, 1f
@@ -50,18 +50,18 @@ do_fpdis:
 	add		%g0, %g0, %g0
 	ba,a,pt		%xcc, rtrap_clr_l6
 
-1:	ldub		[%g6 + TI_FPSAVED], %g5					! Load	Group
-	wr		%g0, FPRS_FEF, %fprs					! LSU	Group+4bubbles
-	andcc		%g5, FPRS_FEF, %g0					! IEU1	Group
-	be,a,pt		%icc, 1f						! CTI
-	 clr		%g7							! IEU0
-	ldx		[%g6 + TI_GSR], %g7					! Load	Group
-1:	andcc		%g5, FPRS_DL, %g0					! IEU1
-	bne,pn		%icc, 2f						! CTI
-	 fzero		%f0							! FPA
-	andcc		%g5, FPRS_DU, %g0					! IEU1  Group
-	bne,pn		%icc, 1f						! CTI
-	 fzero		%f2							! FPA
+1:	ldub		[%g6 + TI_FPSAVED], %g5
+	wr		%g0, FPRS_FEF, %fprs
+	andcc		%g5, FPRS_FEF, %g0
+	be,a,pt		%icc, 1f
+	 clr		%g7
+	ldx		[%g6 + TI_GSR], %g7
+1:	andcc		%g5, FPRS_DL, %g0
+	bne,pn		%icc, 2f
+	 fzero		%f0
+	andcc		%g5, FPRS_DU, %g0
+	bne,pn		%icc, 1f
+	 fzero		%f2
 	faddd		%f0, %f2, %f4
 	fmuld		%f0, %f2, %f6
 	faddd		%f0, %f2, %f8
@@ -104,8 +104,10 @@ do_fpdis:
 	add		%g6, TI_FPREGS + 0xc0, %g2
 	faddd		%f0, %f2, %f8
 	fmuld		%f0, %f2, %f10
-	ldda		[%g1] ASI_BLK_S, %f32	! grrr, where is ASI_BLK_NUCLEUS 8-(
+	membar		#Sync
+	ldda		[%g1] ASI_BLK_S, %f32
 	ldda		[%g2] ASI_BLK_S, %f48
+	membar		#Sync
 	faddd		%f0, %f2, %f12
 	fmuld		%f0, %f2, %f14
 	faddd		%f0, %f2, %f16
@@ -116,7 +118,6 @@ do_fpdis:
 	fmuld		%f0, %f2, %f26
 	faddd		%f0, %f2, %f28
 	fmuld		%f0, %f2, %f30
-	membar		#Sync
 	b,pt		%xcc, fpdis_exit
 	 nop
 2:	andcc		%g5, FPRS_DU, %g0
@@ -133,8 +134,10 @@ do_fpdis:
 	add		%g6, TI_FPREGS + 0x40, %g2
 	faddd		%f32, %f34, %f36
 	fmuld		%f32, %f34, %f38
-	ldda		[%g1] ASI_BLK_S, %f0	! grrr, where is ASI_BLK_NUCLEUS 8-(
+	membar		#Sync
+	ldda		[%g1] ASI_BLK_S, %f0
 	ldda		[%g2] ASI_BLK_S, %f16
+	membar		#Sync
 	faddd		%f32, %f34, %f40
 	fmuld		%f32, %f34, %f42
 	faddd		%f32, %f34, %f44
@@ -147,7 +150,6 @@ do_fpdis:
 	fmuld		%f32, %f34, %f58
 	faddd		%f32, %f34, %f60
 	fmuld		%f32, %f34, %f62
-	membar		#Sync
 	ba,pt		%xcc, fpdis_exit
 	 nop
 3:	mov		SECONDARY_CONTEXT, %g3
@@ -158,7 +160,8 @@ do_fpdis:
 	stxa		%g2, [%g3] ASI_DMMU
 	membar		#Sync
 	mov		0x40, %g2
-	ldda		[%g1] ASI_BLK_S, %f0		! grrr, where is ASI_BLK_NUCLEUS 8-(
+	membar		#Sync
+	ldda		[%g1] ASI_BLK_S, %f0
 	ldda		[%g1 + %g2] ASI_BLK_S, %f16
 	add		%g1, 0x80, %g1
 	ldda		[%g1] ASI_BLK_S, %f32
diff --git a/arch/sparc64/kernel/head.S b/arch/sparc64/kernel/head.S
index 24340496cdd3..b49dcd4504b0 100644
--- a/arch/sparc64/kernel/head.S
+++ b/arch/sparc64/kernel/head.S
@@ -28,19 +28,14 @@
 #include <asm/mmu.h>
 	
 /* This section from from _start to sparc64_boot_end should fit into
- * 0x0000.0000.0040.4000 to 0x0000.0000.0040.8000 and will be sharing space
- * with bootup_user_stack, which is from 0x0000.0000.0040.4000 to
- * 0x0000.0000.0040.6000 and empty_bad_page, which is from
- * 0x0000.0000.0040.6000 to 0x0000.0000.0040.8000. 
+ * 0x0000000000404000 to 0x0000000000408000.
  */
-
 	.text
 	.globl	start, _start, stext, _stext
 _start:
 start:
 _stext:
 stext:
-bootup_user_stack:
 ! 0x0000000000404000
 	b	sparc64_boot
 	 flushw					/* Flush register file.      */
@@ -191,8 +186,9 @@ prom_boot_mapping_phys_low:
 	stx	%l3, [%sp + 2047 + 128 + 0x10]	! num_rets, 5
 	stx	%l2, [%sp + 2047 + 128 + 0x18]	! arg1: "translate"
 	stx	%l5, [%sp + 2047 + 128 + 0x20]	! arg2: prom_mmu_ihandle_cache
-	srlx	%l0, 22, %l3
-	sllx	%l3, 22, %l3
+	/* PAGE align */
+	srlx	%l0, 13, %l3
+	sllx	%l3, 13, %l3
 	stx	%l3, [%sp + 2047 + 128 + 0x28]	! arg3: vaddr, our PC
 	stx	%g0, [%sp + 2047 + 128 + 0x30]	! res1
 	stx	%g0, [%sp + 2047 + 128 + 0x38]	! res2
@@ -211,6 +207,9 @@ prom_boot_mapping_phys_low:
 	ldx	[%sp + 2047 + 128 + 0x48], %l2	! physaddr high
 	stx	%l2, [%l4 + 0x0]
 	ldx	[%sp + 2047 + 128 + 0x50], %l3	! physaddr low
+	/* 4MB align */
+	srlx	%l3, 22, %l3
+	sllx	%l3, 22, %l3
 	stx	%l3, [%l4 + 0x8]
 
 	/* Leave service as-is, "call-method" */
@@ -382,32 +381,78 @@ tlb_fixup_done:
 	 nop
 	/* Not reached... */
 
-/* IMPORTANT NOTE: Whenever making changes here, check
- * trampoline.S as well. -jj */
-	.globl	setup_tba
-setup_tba:	/* i0 = is_starfire */
-	save	%sp, -160, %sp
+	/* This is meant to allow the sharing of this code between
+	 * boot processor invocation (via setup_tba() below) and
+	 * secondary processor startup (via trampoline.S).  The
+	 * former does use this code, the latter does not yet due
+	 * to some complexities.  That should be fixed up at some
+	 * point.
+	 *
+	 * There used to be enormous complexity wrt. transferring
+	 * over from the firwmare's trap table to the Linux kernel's.
+	 * For example, there was a chicken & egg problem wrt. building
+	 * the OBP page tables, yet needing to be on the Linux kernel
+	 * trap table (to translate PAGE_OFFSET addresses) in order to
+	 * do that.
+	 *
+	 * We now handle OBP tlb misses differently, via linear lookups
+	 * into the prom_trans[] array.  So that specific problem no
+	 * longer exists.  Yet, unfortunately there are still some issues
+	 * preventing trampoline.S from using this code... ho hum.
+	 */
+	.globl	setup_trap_table
+setup_trap_table:
+	save	%sp, -192, %sp
 
-	rdpr	%tba, %g7
-	sethi	%hi(prom_tba), %o1
-	or	%o1, %lo(prom_tba), %o1
-	stx	%g7, [%o1]
+	/* Force interrupts to be disabled. */
+	rdpr	%pstate, %o1
+	andn	%o1, PSTATE_IE, %o1
+	wrpr	%o1, 0x0, %pstate
+	wrpr	%g0, 15, %pil
+
+	/* Make the firmware call to jump over to the Linux trap table.  */
+	call	prom_set_trap_table
+	 sethi	%hi(sparc64_ttable_tl0), %o0
+
+	/* Start using proper page size encodings in ctx register.  */
+	sethi	%hi(sparc64_kern_pri_context), %g3
+	ldx	[%g3 + %lo(sparc64_kern_pri_context)], %g2
+	mov	PRIMARY_CONTEXT, %g1
+	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()
+	 *
+	 * MMU Globals (PSTATE_MG):
+	 *
+	 * %g1			--> TLB_SFSR
+	 * %g2			--> ((_PAGE_VALID | _PAGE_SZ4MB |
+	 *			      _PAGE_CP | _PAGE_CV | _PAGE_P | _PAGE_W)
+	 *			     ^ 0xfffff80000000000)
+	 * (this %g2 value is used for computing the PAGE_OFFSET kernel
+	 *  TLB entries quickly, the virtual address of the fault XOR'd
+	 *  with this %g2 value is the PTE to load into the TLB)
+	 * %g3			--> VPTE_BASE_CHEETAH or VPTE_BASE_SPITFIRE
+	 *
+	 * Interrupt Globals (PSTATE_IG, setup by init_irqwork_curcpu()):
+	 *
+	 * %g6			--> __irq_work[smp_processor_id()]
+	 */
 
-	/* Setup "Linux" globals 8-) */
 	rdpr	%pstate, %o1
 	mov	%g6, %o2
-	wrpr	%o1, (PSTATE_AG|PSTATE_IE), %pstate
-	sethi	%hi(sparc64_ttable_tl0), %g1
-	wrpr	%g1, %tba
+	wrpr	%o1, PSTATE_AG, %pstate
 	mov	%o2, %g6
 
-	/* Set up MMU globals */
-	wrpr	%o1, (PSTATE_MG|PSTATE_IE), %pstate
-
-	/* Set fixed globals used by dTLB miss handler. */
 #define KERN_HIGHBITS		((_PAGE_VALID|_PAGE_SZ4MB)^0xfffff80000000000)
 #define KERN_LOWBITS		(_PAGE_CP | _PAGE_CV | _PAGE_P | _PAGE_W)
-
+	wrpr	%o1, PSTATE_MG, %pstate
 	mov	TSB_REG, %g1
 	stxa	%g0, [%g1] ASI_DMMU
 	membar	#Sync
@@ -419,17 +464,17 @@ setup_tba:	/* i0 = is_starfire */
 	sllx	%g2, 32, %g2
 	or	%g2, KERN_LOWBITS, %g2
 
-	BRANCH_IF_ANY_CHEETAH(g3,g7,cheetah_vpte_base)
-	ba,pt	%xcc, spitfire_vpte_base
+	BRANCH_IF_ANY_CHEETAH(g3,g7,8f)
+	ba,pt	%xcc, 9f
 	 nop
 
-cheetah_vpte_base:
+8:
 	sethi		%uhi(VPTE_BASE_CHEETAH), %g3
 	or		%g3, %ulo(VPTE_BASE_CHEETAH), %g3
 	ba,pt		%xcc, 2f
 	 sllx		%g3, 32, %g3
 
-spitfire_vpte_base:
+9:
 	sethi		%uhi(VPTE_BASE_SPITFIRE), %g3
 	or		%g3, %ulo(VPTE_BASE_SPITFIRE), %g3
 	sllx		%g3, 32, %g3
@@ -455,41 +500,55 @@ spitfire_vpte_base:
 	sllx	%o2, 32, %o2
 	wr	%o2, %asr25
 
-	/* Ok, we're done setting up all the state our trap mechanims needs,
-	 * now get back into normal globals and let the PROM know what is up.
-	 */
 2:
 	wrpr	%g0, %g0, %wstate
-	wrpr	%o1, PSTATE_IE, %pstate
+	wrpr	%o1, 0x0, %pstate
 
 	call	init_irqwork_curcpu
 	 nop
 
-	call	prom_set_trap_table
-	 sethi	%hi(sparc64_ttable_tl0), %o0
-
-	/* Start using proper page size encodings in ctx register.  */
-	sethi	%hi(sparc64_kern_pri_context), %g3
-	ldx	[%g3 + %lo(sparc64_kern_pri_context)], %g2
-	mov	PRIMARY_CONTEXT, %g1
-	stxa	%g2, [%g1] ASI_DMMU
-	membar	#Sync
-
+	/* Now we can turn interrupts back on. */
 	rdpr	%pstate, %o1
 	or	%o1, PSTATE_IE, %o1
 	wrpr	%o1, 0, %pstate
+	wrpr	%g0, 0x0, %pil
 
 	ret
 	 restore
 
+	.globl	setup_tba
+setup_tba:	/* i0 = is_starfire */
+	save	%sp, -192, %sp
+
+	/* The boot processor is the only cpu which invokes this
+	 * routine, the other cpus set things up via trampoline.S.
+	 * So save the OBP trap table address here.
+	 */
+	rdpr	%tba, %g7
+	sethi	%hi(prom_tba), %o1
+	or	%o1, %lo(prom_tba), %o1
+	stx	%g7, [%o1]
+
+	call	setup_trap_table
+	 nop
+
+	ret
+	 restore
+sparc64_boot_end:
+
+#include "systbls.S"
+#include "ktlb.S"
+#include "etrap.S"
+#include "rtrap.S"
+#include "winfixup.S"
+#include "entry.S"
+
 /*
- * The following skips make sure the trap table in ttable.S is aligned
+ * The following skip makes sure the trap table in ttable.S is aligned
  * on a 32K boundary as required by the v9 specs for TBA register.
  */
-sparc64_boot_end:
-	.skip	0x2000 + _start - sparc64_boot_end
-bootup_user_stack_end:
-	.skip	0x2000
+1:
+	.skip	0x4000 + _start - 1b
 
 #ifdef CONFIG_SBUS
 /* This is just a hack to fool make depend config.h discovering
@@ -501,15 +560,6 @@ bootup_user_stack_end:
 ! 0x0000000000408000
 
 #include "ttable.S"
-#include "systbls.S"
-#include "ktlb.S"
-#include "etrap.S"
-#include "rtrap.S"
-#include "winfixup.S"
-#include "entry.S"
-
-	/* This is just anal retentiveness on my part... */
-	.align	16384
 
 	.data
 	.align	8
diff --git a/arch/sparc64/kernel/irq.c b/arch/sparc64/kernel/irq.c
index c9b69167632a..233526ba3abe 100644
--- a/arch/sparc64/kernel/irq.c
+++ b/arch/sparc64/kernel/irq.c
@@ -27,6 +27,7 @@
 #include <asm/atomic.h>
 #include <asm/system.h>
 #include <asm/irq.h>
+#include <asm/io.h>
 #include <asm/sbus.h>
 #include <asm/iommu.h>
 #include <asm/upa.h>
diff --git a/arch/sparc64/kernel/itlb_base.S b/arch/sparc64/kernel/itlb_base.S
index b5e32dfa4fbc..4951ff8f6877 100644
--- a/arch/sparc64/kernel/itlb_base.S
+++ b/arch/sparc64/kernel/itlb_base.S
@@ -15,14 +15,12 @@
  */
 #define CREATE_VPTE_OFFSET1(r1, r2) \
 				srax	r1, 10, r2
-#define CREATE_VPTE_OFFSET2(r1, r2)
-#define CREATE_VPTE_NOP		nop
+#define CREATE_VPTE_OFFSET2(r1, r2) nop
 #else /* PAGE_SHIFT */
 #define CREATE_VPTE_OFFSET1(r1, r2) \
 				srax	r1, PAGE_SHIFT, r2
 #define CREATE_VPTE_OFFSET2(r1, r2) \
 				sllx	r2, 3, r2
-#define CREATE_VPTE_NOP
 #endif /* PAGE_SHIFT */
 
 
@@ -36,6 +34,7 @@
  */
 
 /* ITLB ** ICACHE line 1: Quick user TLB misses		*/
+	mov		TLB_SFSR, %g1
 	ldxa		[%g1 + %g1] ASI_IMMU, %g4	! Get TAG_ACCESS
 	CREATE_VPTE_OFFSET1(%g4, %g6)			! Create VPTE offset
 	CREATE_VPTE_OFFSET2(%g4, %g6)			! Create VPTE offset
@@ -43,41 +42,38 @@
 1:	brgez,pn	%g5, 3f				! Not valid, branch out
 	 sethi		%hi(_PAGE_EXEC), %g4		! Delay-slot
 	andcc		%g5, %g4, %g0			! Executable?
+
+/* ITLB ** ICACHE line 2: Real faults			*/
 	be,pn		%xcc, 3f			! Nope, branch.
 	 nop						! Delay-slot
 2:	stxa		%g5, [%g0] ASI_ITLB_DATA_IN	! Load PTE into TLB
 	retry						! Trap return
-3:	rdpr		%pstate, %g4			! Move into alternate globals
-
-/* ITLB ** ICACHE line 2: Real faults			*/
+3:	rdpr		%pstate, %g4			! Move into alt-globals
 	wrpr		%g4, PSTATE_AG|PSTATE_MG, %pstate
 	rdpr		%tpc, %g5			! And load faulting VA
 	mov		FAULT_CODE_ITLB, %g4		! It was read from ITLB
-sparc64_realfault_common:				! Called by TL0 dtlb_miss too
+
+/* ITLB ** ICACHE line 3: Finish faults	*/
+sparc64_realfault_common:				! Called by dtlb_miss
 	stb		%g4, [%g6 + TI_FAULT_CODE]
 	stx		%g5, [%g6 + TI_FAULT_ADDR]
 	ba,pt		%xcc, etrap			! Save state
 1:	 rd		%pc, %g7			! ...
-	nop
-
-/* ITLB ** ICACHE line 3: Finish faults + window fixups	*/
 	call		do_sparc64_fault		! Call fault handler
 	 add		%sp, PTREGS_OFF, %o0! Compute pt_regs arg
 	ba,pt		%xcc, rtrap_clr_l6		! Restore cpu state
 	 nop
+
+/* ITLB ** ICACHE line 4: Window fixups */
 winfix_trampoline:
 	rdpr		%tpc, %g3			! Prepare winfixup TNPC
-	or		%g3, 0x7c, %g3			! Compute offset to branch
+	or		%g3, 0x7c, %g3			! Compute branch offset
 	wrpr		%g3, %tnpc			! Write it into TNPC
 	done						! Do it to it
-
-/* ITLB ** ICACHE line 4: Unused...	*/
 	nop
 	nop
 	nop
 	nop
-	CREATE_VPTE_NOP
 
 #undef CREATE_VPTE_OFFSET1
 #undef CREATE_VPTE_OFFSET2
-#undef CREATE_VPTE_NOP
diff --git a/arch/sparc64/kernel/ktlb.S b/arch/sparc64/kernel/ktlb.S
index 7796b37f478c..d9244d3c9f73 100644
--- a/arch/sparc64/kernel/ktlb.S
+++ b/arch/sparc64/kernel/ktlb.S
@@ -58,9 +58,6 @@ vpte_noent:
 	done
 
 vpte_insn_obp:
-	sethi		%hi(prom_pmd_phys), %g5
-	ldx		[%g5 + %lo(prom_pmd_phys)], %g5
-
 	/* Behave as if we are at TL0.  */
 	wrpr		%g0, 1, %tl
 	rdpr		%tpc, %g4	/* Find original faulting iaddr */
@@ -71,58 +68,57 @@ vpte_insn_obp:
 	mov		TLB_SFSR, %g1
 	stxa		%g4, [%g1 + %g1] ASI_IMMU
 
-	/* Get PMD offset.  */
-	srlx		%g4, 23, %g6
-	and		%g6, 0x7ff, %g6
-	sllx		%g6, 2, %g6
-
-	/* Load PMD, is it valid?  */
-	lduwa		[%g5 + %g6] ASI_PHYS_USE_EC, %g5
-	brz,pn		%g5, longpath
-	 sllx		%g5, 11, %g5
-
-	/* Get PTE offset.  */
-	srlx		%g4, 13, %g6
-	and		%g6, 0x3ff, %g6
-	sllx		%g6, 3, %g6
-
-	/* Load PTE.  */
-	ldxa		[%g5 + %g6] ASI_PHYS_USE_EC, %g5
-	brgez,pn	%g5, longpath
-	 nop
-
-	/* TLB load and return from trap.  */
+	sethi		%hi(prom_trans), %g5
+	or		%g5, %lo(prom_trans), %g5
+
+1:	ldx		[%g5 + 0x00], %g6	! base
+	brz,a,pn	%g6, longpath		! no more entries, fail
+	 mov		TLB_SFSR, %g1		! and restore %g1
+	ldx		[%g5 + 0x08], %g1	! len
+	add		%g6, %g1, %g1		! end
+	cmp		%g6, %g4
+	bgu,pt		%xcc, 2f
+	 cmp		%g4, %g1
+	bgeu,pt		%xcc, 2f
+	 ldx		[%g5 + 0x10], %g1	! PTE
+
+	/* TLB load, restore %g1, and return from trap.  */
+	sub		%g4, %g6, %g6
+	add		%g1, %g6, %g5
+	mov		TLB_SFSR, %g1
 	stxa		%g5, [%g0] ASI_ITLB_DATA_IN
 	retry
 
-kvmap_do_obp:
-	sethi		%hi(prom_pmd_phys), %g5
-	ldx		[%g5 + %lo(prom_pmd_phys)], %g5
-
-	/* Get PMD offset.  */
-	srlx		%g4, 23, %g6
-	and		%g6, 0x7ff, %g6
-	sllx		%g6, 2, %g6
-
-	/* Load PMD, is it valid?  */
-	lduwa		[%g5 + %g6] ASI_PHYS_USE_EC, %g5
-	brz,pn		%g5, longpath
-	 sllx		%g5, 11, %g5
-
-	/* Get PTE offset.  */
-	srlx		%g4, 13, %g6
-	and		%g6, 0x3ff, %g6
-	sllx		%g6, 3, %g6
-
-	/* Load PTE.  */
-	ldxa		[%g5 + %g6] ASI_PHYS_USE_EC, %g5
-	brgez,pn	%g5, longpath
-	 nop
+2:	ba,pt		%xcc, 1b
+	 add		%g5, (3 * 8), %g5	! next entry
 
-	/* TLB load and return from trap.  */
+kvmap_do_obp:
+	sethi		%hi(prom_trans), %g5
+	or		%g5, %lo(prom_trans), %g5
+	srlx		%g4, 13, %g4
+	sllx		%g4, 13, %g4
+
+1:	ldx		[%g5 + 0x00], %g6	! base
+	brz,a,pn	%g6, longpath		! no more entries, fail
+	 mov		TLB_SFSR, %g1		! and restore %g1
+	ldx		[%g5 + 0x08], %g1	! len
+	add		%g6, %g1, %g1		! end
+	cmp		%g6, %g4
+	bgu,pt		%xcc, 2f
+	 cmp		%g4, %g1
+	bgeu,pt		%xcc, 2f
+	 ldx		[%g5 + 0x10], %g1	! PTE
+
+	/* TLB load, restore %g1, and return from trap.  */
+	sub		%g4, %g6, %g6
+	add		%g1, %g6, %g5
+	mov		TLB_SFSR, %g1
 	stxa		%g5, [%g0] ASI_DTLB_DATA_IN
 	retry
 
+2:	ba,pt		%xcc, 1b
+	 add		%g5, (3 * 8), %g5	! next entry
+
 /*
  * On a first level data miss, check whether this is to the OBP range (note
  * that such accesses can be made by prom, as well as by kernel using
diff --git a/arch/sparc64/kernel/pci_iommu.c b/arch/sparc64/kernel/pci_iommu.c
index 425c60cfea19..a11910be1013 100644
--- a/arch/sparc64/kernel/pci_iommu.c
+++ b/arch/sparc64/kernel/pci_iommu.c
@@ -49,12 +49,6 @@ static void __iommu_flushall(struct pci_iommu *iommu)
 
 	/* Ensure completion of previous PIO writes. */
 	(void) pci_iommu_read(iommu->write_complete_reg);
-
-	/* Now update everyone's flush point. */
-	for (entry = 0; entry < PBM_NCLUSTERS; entry++) {
-		iommu->alloc_info[entry].flush =
-			iommu->alloc_info[entry].next;
-	}
 }
 
 #define IOPTE_CONSISTENT(CTX) \
@@ -80,120 +74,117 @@ static void inline iopte_make_dummy(struct pci_iommu *iommu, iopte_t *iopte)
 	iopte_val(*iopte) = val;
 }
 
-void pci_iommu_table_init(struct pci_iommu *iommu, int tsbsize)
+/* Based largely upon the ppc64 iommu allocator.  */
+static long pci_arena_alloc(struct pci_iommu *iommu, unsigned long npages)
 {
-	int i;
-
-	tsbsize /= sizeof(iopte_t);
-
-	for (i = 0; i < tsbsize; i++)
-		iopte_make_dummy(iommu, &iommu->page_table[i]);
-}
-
-static iopte_t *alloc_streaming_cluster(struct pci_iommu *iommu, unsigned long npages)
-{
-	iopte_t *iopte, *limit, *first;
-	unsigned long cnum, ent, flush_point;
-
-	cnum = 0;
-	while ((1UL << cnum) < npages)
-		cnum++;
-	iopte  = (iommu->page_table +
-		  (cnum << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
-
-	if (cnum == 0)
-		limit = (iommu->page_table +
-			 iommu->lowest_consistent_map);
-	else
-		limit = (iopte +
-			 (1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
-
-	iopte += ((ent = iommu->alloc_info[cnum].next) << cnum);
-	flush_point = iommu->alloc_info[cnum].flush;
-	
-	first = iopte;
-	for (;;) {
-		if (IOPTE_IS_DUMMY(iommu, iopte)) {
-			if ((iopte + (1 << cnum)) >= limit)
-				ent = 0;
-			else
-				ent = ent + 1;
-			iommu->alloc_info[cnum].next = ent;
-			if (ent == flush_point)
-				__iommu_flushall(iommu);
-			break;
+	struct pci_iommu_arena *arena = &iommu->arena;
+	unsigned long n, i, start, end, limit;
+	int pass;
+
+	limit = arena->limit;
+	start = arena->hint;
+	pass = 0;
+
+again:
+	n = find_next_zero_bit(arena->map, limit, start);
+	end = n + npages;
+	if (unlikely(end >= limit)) {
+		if (likely(pass < 1)) {
+			limit = start;
+			start = 0;
+			__iommu_flushall(iommu);
+			pass++;
+			goto again;
+		} else {
+			/* Scanned the whole thing, give up. */
+			return -1;
 		}
-		iopte += (1 << cnum);
-		ent++;
-		if (iopte >= limit) {
-			iopte = (iommu->page_table +
-				 (cnum <<
-				  (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
-			ent = 0;
+	}
+
+	for (i = n; i < end; i++) {
+		if (test_bit(i, arena->map)) {
+			start = i + 1;
+			goto again;
 		}
-		if (ent == flush_point)
-			__iommu_flushall(iommu);
-		if (iopte == first)
-			goto bad;
 	}
 
-	/* I've got your streaming cluster right here buddy boy... */
-	return iopte;
+	for (i = n; i < end; i++)
+		__set_bit(i, arena->map);
 
-bad:
-	printk(KERN_EMERG "pci_iommu: alloc_streaming_cluster of npages(%ld) failed!\n",
-	       npages);
-	return NULL;
+	arena->hint = end;
+
+	return n;
 }
 
-static void free_streaming_cluster(struct pci_iommu *iommu, dma_addr_t base,
-				   unsigned long npages, unsigned long ctx)
+static void pci_arena_free(struct pci_iommu_arena *arena, unsigned long base, unsigned long npages)
 {
-	unsigned long cnum, ent;
+	unsigned long i;
 
-	cnum = 0;
-	while ((1UL << cnum) < npages)
-		cnum++;
+	for (i = base; i < (base + npages); i++)
+		__clear_bit(i, arena->map);
+}
 
-	ent = (base << (32 - IO_PAGE_SHIFT + PBM_LOGCLUSTERS - iommu->page_table_sz_bits))
-		>> (32 + PBM_LOGCLUSTERS + cnum - iommu->page_table_sz_bits);
+void pci_iommu_table_init(struct pci_iommu *iommu, int tsbsize, u32 dma_offset, u32 dma_addr_mask)
+{
+	unsigned long i, tsbbase, order, sz, num_tsb_entries;
+
+	num_tsb_entries = tsbsize / sizeof(iopte_t);
+
+	/* Setup initial software IOMMU state. */
+	spin_lock_init(&iommu->lock);
+	iommu->ctx_lowest_free = 1;
+	iommu->page_table_map_base = dma_offset;
+	iommu->dma_addr_mask = dma_addr_mask;
+
+	/* Allocate and initialize the free area map.  */
+	sz = num_tsb_entries / 8;
+	sz = (sz + 7UL) & ~7UL;
+	iommu->arena.map = kmalloc(sz, GFP_KERNEL);
+	if (!iommu->arena.map) {
+		prom_printf("PCI_IOMMU: Error, kmalloc(arena.map) failed.\n");
+		prom_halt();
+	}
+	memset(iommu->arena.map, 0, sz);
+	iommu->arena.limit = num_tsb_entries;
 
-	/* If the global flush might not have caught this entry,
-	 * adjust the flush point such that we will flush before
-	 * ever trying to reuse it.
+	/* Allocate and initialize the dummy page which we
+	 * set inactive IO PTEs to point to.
 	 */
-#define between(X,Y,Z)	(((Z) - (Y)) >= ((X) - (Y)))
-	if (between(ent, iommu->alloc_info[cnum].next, iommu->alloc_info[cnum].flush))
-		iommu->alloc_info[cnum].flush = ent;
-#undef between
+	iommu->dummy_page = __get_free_pages(GFP_KERNEL, 0);
+	if (!iommu->dummy_page) {
+		prom_printf("PCI_IOMMU: Error, gfp(dummy_page) failed.\n");
+		prom_halt();
+	}
+	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
+	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
+
+	/* Now allocate and setup the IOMMU page table itself.  */
+	order = get_order(tsbsize);
+	tsbbase = __get_free_pages(GFP_KERNEL, order);
+	if (!tsbbase) {
+		prom_printf("PCI_IOMMU: Error, gfp(tsb) failed.\n");
+		prom_halt();
+	}
+	iommu->page_table = (iopte_t *)tsbbase;
+
+	for (i = 0; i < num_tsb_entries; i++)
+		iopte_make_dummy(iommu, &iommu->page_table[i]);
 }
 
-/* We allocate consistent mappings from the end of cluster zero. */
-static iopte_t *alloc_consistent_cluster(struct pci_iommu *iommu, unsigned long npages)
+static inline iopte_t *alloc_npages(struct pci_iommu *iommu, unsigned long npages)
 {
-	iopte_t *iopte;
+	long entry;
 
-	iopte = iommu->page_table + (1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS));
-	while (iopte > iommu->page_table) {
-		iopte--;
-		if (IOPTE_IS_DUMMY(iommu, iopte)) {
-			unsigned long tmp = npages;
+	entry = pci_arena_alloc(iommu, npages);
+	if (unlikely(entry < 0))
+		return NULL;
 
-			while (--tmp) {
-				iopte--;
-				if (!IOPTE_IS_DUMMY(iommu, iopte))
-					break;
-			}
-			if (tmp == 0) {
-				u32 entry = (iopte - iommu->page_table);
+	return iommu->page_table + entry;
+}
 
-				if (entry < iommu->lowest_consistent_map)
-					iommu->lowest_consistent_map = entry;
-				return iopte;
-			}
-		}
-	}
-	return NULL;
+static inline void free_npages(struct pci_iommu *iommu, dma_addr_t base, unsigned long npages)
+{
+	pci_arena_free(&iommu->arena, base >> IO_PAGE_SHIFT, npages);
 }
 
 static int iommu_alloc_ctx(struct pci_iommu *iommu)
@@ -233,7 +224,7 @@ void *pci_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_ad
 	struct pcidev_cookie *pcp;
 	struct pci_iommu *iommu;
 	iopte_t *iopte;
-	unsigned long flags, order, first_page, ctx;
+	unsigned long flags, order, first_page;
 	void *ret;
 	int npages;
 
@@ -251,9 +242,10 @@ void *pci_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_ad
 	iommu = pcp->pbm->iommu;
 
 	spin_lock_irqsave(&iommu->lock, flags);
-	iopte = alloc_consistent_cluster(iommu, size >> IO_PAGE_SHIFT);
-	if (iopte == NULL) {
-		spin_unlock_irqrestore(&iommu->lock, flags);
+	iopte = alloc_npages(iommu, size >> IO_PAGE_SHIFT);
+	spin_unlock_irqrestore(&iommu->lock, flags);
+
+	if (unlikely(iopte == NULL)) {
 		free_pages(first_page, order);
 		return NULL;
 	}
@@ -262,31 +254,15 @@ void *pci_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_ad
 		      ((iopte - iommu->page_table) << IO_PAGE_SHIFT));
 	ret = (void *) first_page;
 	npages = size >> IO_PAGE_SHIFT;
-	ctx = 0;
-	if (iommu->iommu_ctxflush)
-		ctx = iommu_alloc_ctx(iommu);
 	first_page = __pa(first_page);
 	while (npages--) {
-		iopte_val(*iopte) = (IOPTE_CONSISTENT(ctx) |
+		iopte_val(*iopte) = (IOPTE_CONSISTENT(0UL) |
 				     IOPTE_WRITE |
 				     (first_page & IOPTE_PAGE));
 		iopte++;
 		first_page += IO_PAGE_SIZE;
 	}
 
-	{
-		int i;
-		u32 daddr = *dma_addrp;
-
-		npages = size >> IO_PAGE_SHIFT;
-		for (i = 0; i < npages; i++) {
-			pci_iommu_write(iommu->iommu_flush, daddr);
-			daddr += IO_PAGE_SIZE;
-		}
-	}
-
-	spin_unlock_irqrestore(&iommu->lock, flags);
-
 	return ret;
 }
 
@@ -296,7 +272,7 @@ void pci_free_consistent(struct pci_dev *pdev, size_t size, void *cpu, dma_addr_
 	struct pcidev_cookie *pcp;
 	struct pci_iommu *iommu;
 	iopte_t *iopte;
-	unsigned long flags, order, npages, i, ctx;
+	unsigned long flags, order, npages;
 
 	npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
 	pcp = pdev->sysdata;
@@ -306,46 +282,7 @@ void pci_free_consistent(struct pci_dev *pdev, size_t size, void *cpu, dma_addr_
 
 	spin_lock_irqsave(&iommu->lock, flags);
 
-	if ((iopte - iommu->page_table) ==
-	    iommu->lowest_consistent_map) {
-		iopte_t *walk = iopte + npages;
-		iopte_t *limit;
-
-		limit = (iommu->page_table +
-			 (1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
-		while (walk < limit) {
-			if (!IOPTE_IS_DUMMY(iommu, walk))
-				break;
-			walk++;
-		}
-		iommu->lowest_consistent_map =
-			(walk - iommu->page_table);
-	}
-
-	/* Data for consistent mappings cannot enter the streaming
-	 * buffers, so we only need to update the TSB.  We flush
-	 * the IOMMU here as well to prevent conflicts with the
-	 * streaming mapping deferred tlb flush scheme.
-	 */
-
-	ctx = 0;
-	if (iommu->iommu_ctxflush)
-		ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
-
-	for (i = 0; i < npages; i++, iopte++)
-		iopte_make_dummy(iommu, iopte);
-
-	if (iommu->iommu_ctxflush) {
-		pci_iommu_write(iommu->iommu_ctxflush, ctx);
-	} else {
-		for (i = 0; i < npages; i++) {
-			u32 daddr = dvma + (i << IO_PAGE_SHIFT);
-
-			pci_iommu_write(iommu->iommu_flush, daddr);
-		}
-	}
-
-	iommu_free_ctx(iommu, ctx);
+	free_npages(iommu, dvma, npages);
 
 	spin_unlock_irqrestore(&iommu->lock, flags);
 
@@ -372,25 +309,27 @@ dma_addr_t pci_map_single(struct pci_dev *pdev, void *ptr, size_t sz, int direct
 	iommu = pcp->pbm->iommu;
 	strbuf = &pcp->pbm->stc;
 
-	if (direction == PCI_DMA_NONE)
-		BUG();
+	if (unlikely(direction == PCI_DMA_NONE))
+		goto bad_no_ctx;
 
 	oaddr = (unsigned long)ptr;
 	npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK);
 	npages >>= IO_PAGE_SHIFT;
 
 	spin_lock_irqsave(&iommu->lock, flags);
+	base = alloc_npages(iommu, npages);
+	ctx = 0;
+	if (iommu->iommu_ctxflush)
+		ctx = iommu_alloc_ctx(iommu);
+	spin_unlock_irqrestore(&iommu->lock, flags);
 
-	base = alloc_streaming_cluster(iommu, npages);
-	if (base == NULL)
+	if (unlikely(!base))
 		goto bad;
+
 	bus_addr = (iommu->page_table_map_base +
 		    ((base - iommu->page_table) << IO_PAGE_SHIFT));
 	ret = bus_addr | (oaddr & ~IO_PAGE_MASK);
 	base_paddr = __pa(oaddr & IO_PAGE_MASK);
-	ctx = 0;
-	if (iommu->iommu_ctxflush)
-		ctx = iommu_alloc_ctx(iommu);
 	if (strbuf->strbuf_enabled)
 		iopte_protection = IOPTE_STREAMING(ctx);
 	else
@@ -401,12 +340,13 @@ dma_addr_t pci_map_single(struct pci_dev *pdev, void *ptr, size_t sz, int direct
 	for (i = 0; i < npages; i++, base++, base_paddr += IO_PAGE_SIZE)
 		iopte_val(*base) = iopte_protection | base_paddr;
 
-	spin_unlock_irqrestore(&iommu->lock, flags);
-
 	return ret;
 
 bad:
-	spin_unlock_irqrestore(&iommu->lock, flags);
+	iommu_free_ctx(iommu, ctx);
+bad_no_ctx:
+	if (printk_ratelimit())
+		WARN_ON(1);
 	return PCI_DMA_ERROR_CODE;
 }
 
@@ -481,10 +421,13 @@ void pci_unmap_single(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int
 	struct pci_iommu *iommu;
 	struct pci_strbuf *strbuf;
 	iopte_t *base;
-	unsigned long flags, npages, ctx;
+	unsigned long flags, npages, ctx, i;
 
-	if (direction == PCI_DMA_NONE)
-		BUG();
+	if (unlikely(direction == PCI_DMA_NONE)) {
+		if (printk_ratelimit())
+			WARN_ON(1);
+		return;
+	}
 
 	pcp = pdev->sysdata;
 	iommu = pcp->pbm->iommu;
@@ -510,13 +453,14 @@ void pci_unmap_single(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int
 
 	/* Step 1: Kick data out of streaming buffers if necessary. */
 	if (strbuf->strbuf_enabled)
-		pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
+		pci_strbuf_flush(strbuf, iommu, bus_addr, ctx,
+				 npages, direction);
 
-	/* Step 2: Clear out first TSB entry. */
-	iopte_make_dummy(iommu, base);
+	/* Step 2: Clear out TSB entries. */
+	for (i = 0; i < npages; i++)
+		iopte_make_dummy(iommu, base + i);
 
-	free_streaming_cluster(iommu, bus_addr - iommu->page_table_map_base,
-			       npages, ctx);
+	free_npages(iommu, bus_addr - iommu->page_table_map_base, npages);
 
 	iommu_free_ctx(iommu, ctx);
 
@@ -621,6 +565,8 @@ int pci_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int
 			pci_map_single(pdev,
 				       (page_address(sglist->page) + sglist->offset),
 				       sglist->length, direction);
+		if (unlikely(sglist->dma_address == PCI_DMA_ERROR_CODE))
+			return 0;
 		sglist->dma_length = sglist->length;
 		return 1;
 	}
@@ -629,21 +575,29 @@ int pci_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int
 	iommu = pcp->pbm->iommu;
 	strbuf = &pcp->pbm->stc;
 	
-	if (direction == PCI_DMA_NONE)
-		BUG();
+	if (unlikely(direction == PCI_DMA_NONE))
+		goto bad_no_ctx;
 
 	/* Step 1: Prepare scatter list. */
 
 	npages = prepare_sg(sglist, nelems);
 
-	/* Step 2: Allocate a cluster. */
+	/* Step 2: Allocate a cluster and context, if necessary. */
 
 	spin_lock_irqsave(&iommu->lock, flags);
 
-	base = alloc_streaming_cluster(iommu, npages);
+	base = alloc_npages(iommu, npages);
+	ctx = 0;
+	if (iommu->iommu_ctxflush)
+		ctx = iommu_alloc_ctx(iommu);
+
+	spin_unlock_irqrestore(&iommu->lock, flags);
+
 	if (base == NULL)
 		goto bad;
-	dma_base = iommu->page_table_map_base + ((base - iommu->page_table) << IO_PAGE_SHIFT);
+
+	dma_base = iommu->page_table_map_base +
+		((base - iommu->page_table) << IO_PAGE_SHIFT);
 
 	/* Step 3: Normalize DMA addresses. */
 	used = nelems;
@@ -656,30 +610,28 @@ int pci_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int
 	}
 	used = nelems - used;
 
-	/* Step 4: Choose a context if necessary. */
-	ctx = 0;
-	if (iommu->iommu_ctxflush)
-		ctx = iommu_alloc_ctx(iommu);
-
-	/* Step 5: Create the mappings. */
+	/* Step 4: Create the mappings. */
 	if (strbuf->strbuf_enabled)
 		iopte_protection = IOPTE_STREAMING(ctx);
 	else
 		iopte_protection = IOPTE_CONSISTENT(ctx);
 	if (direction != PCI_DMA_TODEVICE)
 		iopte_protection |= IOPTE_WRITE;
-	fill_sg (base, sglist, used, nelems, iopte_protection);
+
+	fill_sg(base, sglist, used, nelems, iopte_protection);
+
 #ifdef VERIFY_SG
 	verify_sglist(sglist, nelems, base, npages);
 #endif
 
-	spin_unlock_irqrestore(&iommu->lock, flags);
-
 	return used;
 
 bad:
-	spin_unlock_irqrestore(&iommu->lock, flags);
-	return PCI_DMA_ERROR_CODE;
+	iommu_free_ctx(iommu, ctx);
+bad_no_ctx:
+	if (printk_ratelimit())
+		WARN_ON(1);
+	return 0;
 }
 
 /* Unmap a set of streaming mode DMA translations. */
@@ -692,8 +644,10 @@ void pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems,
 	unsigned long flags, ctx, i, npages;
 	u32 bus_addr;
 
-	if (direction == PCI_DMA_NONE)
-		BUG();
+	if (unlikely(direction == PCI_DMA_NONE)) {
+		if (printk_ratelimit())
+			WARN_ON(1);
+	}
 
 	pcp = pdev->sysdata;
 	iommu = pcp->pbm->iommu;
@@ -705,7 +659,8 @@ void pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems,
 		if (sglist[i].dma_length == 0)
 			break;
 	i--;
-	npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) - bus_addr) >> IO_PAGE_SHIFT;
+	npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) -
+		  bus_addr) >> IO_PAGE_SHIFT;
 
 	base = iommu->page_table +
 		((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
@@ -726,11 +681,11 @@ void pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems,
 	if (strbuf->strbuf_enabled)
 		pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
 
-	/* Step 2: Clear out first TSB entry. */
-	iopte_make_dummy(iommu, base);
+	/* Step 2: Clear out the TSB entries. */
+	for (i = 0; i < npages; i++)
+		iopte_make_dummy(iommu, base + i);
 
-	free_streaming_cluster(iommu, bus_addr - iommu->page_table_map_base,
-			       npages, ctx);
+	free_npages(iommu, bus_addr - iommu->page_table_map_base, npages);
 
 	iommu_free_ctx(iommu, ctx);
 
diff --git a/arch/sparc64/kernel/pci_psycho.c b/arch/sparc64/kernel/pci_psycho.c
index 6ed1ef25e0ac..c03ed5f49d31 100644
--- a/arch/sparc64/kernel/pci_psycho.c
+++ b/arch/sparc64/kernel/pci_psycho.c
@@ -1207,13 +1207,9 @@ static void psycho_scan_bus(struct pci_controller_info *p)
 static void psycho_iommu_init(struct pci_controller_info *p)
 {
 	struct pci_iommu *iommu = p->pbm_A.iommu;
-	unsigned long tsbbase, i;
+	unsigned long i;
 	u64 control;
 
-	/* Setup initial software IOMMU state. */
-	spin_lock_init(&iommu->lock);
-	iommu->ctx_lowest_free = 1;
-
 	/* Register addresses. */
 	iommu->iommu_control  = p->pbm_A.controller_regs + PSYCHO_IOMMU_CONTROL;
 	iommu->iommu_tsbbase  = p->pbm_A.controller_regs + PSYCHO_IOMMU_TSBBASE;
@@ -1240,40 +1236,10 @@ static void psycho_iommu_init(struct pci_controller_info *p)
 	/* Leave diag mode enabled for full-flushing done
 	 * in pci_iommu.c
 	 */
+	pci_iommu_table_init(iommu, IO_TSB_SIZE, 0xc0000000, 0xffffffff);
 
-	iommu->dummy_page = __get_free_pages(GFP_KERNEL, 0);
-	if (!iommu->dummy_page) {
-		prom_printf("PSYCHO_IOMMU: Error, gfp(dummy_page) failed.\n");
-		prom_halt();
-	}
-	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
-	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
-
-	/* Using assumed page size 8K with 128K entries we need 1MB iommu page
-	 * table (128K ioptes * 8 bytes per iopte).  This is
-	 * page order 7 on UltraSparc.
-	 */
-	tsbbase = __get_free_pages(GFP_KERNEL, get_order(IO_TSB_SIZE));
-	if (!tsbbase) {
-		prom_printf("PSYCHO_IOMMU: Error, gfp(tsb) failed.\n");
-		prom_halt();
-	}
-	iommu->page_table = (iopte_t *)tsbbase;
-	iommu->page_table_sz_bits = 17;
-	iommu->page_table_map_base = 0xc0000000;
-	iommu->dma_addr_mask = 0xffffffff;
-	pci_iommu_table_init(iommu, IO_TSB_SIZE);
-
-	/* We start with no consistent mappings. */
-	iommu->lowest_consistent_map =
-		1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS);
-
-	for (i = 0; i < PBM_NCLUSTERS; i++) {
-		iommu->alloc_info[i].flush = 0;
-		iommu->alloc_info[i].next = 0;
-	}
-
-	psycho_write(p->pbm_A.controller_regs + PSYCHO_IOMMU_TSBBASE, __pa(tsbbase));
+	psycho_write(p->pbm_A.controller_regs + PSYCHO_IOMMU_TSBBASE,
+		     __pa(iommu->page_table));
 
 	control = psycho_read(p->pbm_A.controller_regs + PSYCHO_IOMMU_CONTROL);
 	control &= ~(PSYCHO_IOMMU_CTRL_TSBSZ | PSYCHO_IOMMU_CTRL_TBWSZ);
@@ -1281,7 +1247,7 @@ static void psycho_iommu_init(struct pci_controller_info *p)
 	psycho_write(p->pbm_A.controller_regs + PSYCHO_IOMMU_CONTROL, control);
 
 	/* If necessary, hook us up for starfire IRQ translations. */
-	if(this_is_starfire)
+	if (this_is_starfire)
 		p->starfire_cookie = starfire_hookup(p->pbm_A.portid);
 	else
 		p->starfire_cookie = NULL;
diff --git a/arch/sparc64/kernel/pci_sabre.c b/arch/sparc64/kernel/pci_sabre.c
index 0ee6bd5b9ac6..da8e1364194f 100644
--- a/arch/sparc64/kernel/pci_sabre.c
+++ b/arch/sparc64/kernel/pci_sabre.c
@@ -1267,13 +1267,9 @@ static void sabre_iommu_init(struct pci_controller_info *p,
 			     u32 dma_mask)
 {
 	struct pci_iommu *iommu = p->pbm_A.iommu;
-	unsigned long tsbbase, i, order;
+	unsigned long i;
 	u64 control;
 
-	/* Setup initial software IOMMU state. */
-	spin_lock_init(&iommu->lock);
-	iommu->ctx_lowest_free = 1;
-
 	/* Register addresses. */
 	iommu->iommu_control  = p->pbm_A.controller_regs + SABRE_IOMMU_CONTROL;
 	iommu->iommu_tsbbase  = p->pbm_A.controller_regs + SABRE_IOMMU_TSBBASE;
@@ -1295,26 +1291,10 @@ static void sabre_iommu_init(struct pci_controller_info *p,
 	/* Leave diag mode enabled for full-flushing done
 	 * in pci_iommu.c
 	 */
+	pci_iommu_table_init(iommu, tsbsize * 1024 * 8, dvma_offset, dma_mask);
 
-	iommu->dummy_page = __get_free_pages(GFP_KERNEL, 0);
-	if (!iommu->dummy_page) {
-		prom_printf("PSYCHO_IOMMU: Error, gfp(dummy_page) failed.\n");
-		prom_halt();
-	}
-	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
-	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
-
-	tsbbase = __get_free_pages(GFP_KERNEL, order = get_order(tsbsize * 1024 * 8));
-	if (!tsbbase) {
-		prom_printf("SABRE_IOMMU: Error, gfp(tsb) failed.\n");
-		prom_halt();
-	}
-	iommu->page_table = (iopte_t *)tsbbase;
-	iommu->page_table_map_base = dvma_offset;
-	iommu->dma_addr_mask = dma_mask;
-	pci_iommu_table_init(iommu, PAGE_SIZE << order);
-
-	sabre_write(p->pbm_A.controller_regs + SABRE_IOMMU_TSBBASE, __pa(tsbbase));
+	sabre_write(p->pbm_A.controller_regs + SABRE_IOMMU_TSBBASE,
+		    __pa(iommu->page_table));
 
 	control = sabre_read(p->pbm_A.controller_regs + SABRE_IOMMU_CONTROL);
 	control &= ~(SABRE_IOMMUCTRL_TSBSZ | SABRE_IOMMUCTRL_TBWSZ);
@@ -1322,11 +1302,9 @@ static void sabre_iommu_init(struct pci_controller_info *p,
 	switch(tsbsize) {
 	case 64:
 		control |= SABRE_IOMMU_TSBSZ_64K;
-		iommu->page_table_sz_bits = 16;
 		break;
 	case 128:
 		control |= SABRE_IOMMU_TSBSZ_128K;
-		iommu->page_table_sz_bits = 17;
 		break;
 	default:
 		prom_printf("iommu_init: Illegal TSB size %d\n", tsbsize);
@@ -1334,15 +1312,6 @@ static void sabre_iommu_init(struct pci_controller_info *p,
 		break;
 	}
 	sabre_write(p->pbm_A.controller_regs + SABRE_IOMMU_CONTROL, control);
-
-	/* We start with no consistent mappings. */
-	iommu->lowest_consistent_map =
-		1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS);
-
-	for (i = 0; i < PBM_NCLUSTERS; i++) {
-		iommu->alloc_info[i].flush = 0;
-		iommu->alloc_info[i].next = 0;
-	}
 }
 
 static void pbm_register_toplevel_resources(struct pci_controller_info *p,
diff --git a/arch/sparc64/kernel/pci_schizo.c b/arch/sparc64/kernel/pci_schizo.c
index cae5b61fe2f0..d8c4e0919b4e 100644
--- a/arch/sparc64/kernel/pci_schizo.c
+++ b/arch/sparc64/kernel/pci_schizo.c
@@ -1765,7 +1765,7 @@ static void schizo_pbm_strbuf_init(struct pci_pbm_info *pbm)
 static void schizo_pbm_iommu_init(struct pci_pbm_info *pbm)
 {
 	struct pci_iommu *iommu = pbm->iommu;
-	unsigned long tsbbase, i, tagbase, database, order;
+	unsigned long i, tagbase, database;
 	u32 vdma[2], dma_mask;
 	u64 control;
 	int err, tsbsize;
@@ -1800,10 +1800,6 @@ static void schizo_pbm_iommu_init(struct pci_pbm_info *pbm)
 			prom_halt();
 	};
 
-	/* Setup initial software IOMMU state. */
-	spin_lock_init(&iommu->lock);
-	iommu->ctx_lowest_free = 1;
-
 	/* Register addresses, SCHIZO has iommu ctx flushing. */
 	iommu->iommu_control  = pbm->pbm_regs + SCHIZO_IOMMU_CONTROL;
 	iommu->iommu_tsbbase  = pbm->pbm_regs + SCHIZO_IOMMU_TSBBASE;
@@ -1832,56 +1828,9 @@ static void schizo_pbm_iommu_init(struct pci_pbm_info *pbm)
 	/* Leave diag mode enabled for full-flushing done
 	 * in pci_iommu.c
 	 */
+	pci_iommu_table_init(iommu, tsbsize * 8 * 1024, vdma[0], dma_mask);
 
-	iommu->dummy_page = __get_free_pages(GFP_KERNEL, 0);
-	if (!iommu->dummy_page) {
-		prom_printf("PSYCHO_IOMMU: Error, gfp(dummy_page) failed.\n");
-		prom_halt();
-	}
-	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
-	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
-
-	/* Using assumed page size 8K with 128K entries we need 1MB iommu page
-	 * table (128K ioptes * 8 bytes per iopte).  This is
-	 * page order 7 on UltraSparc.
-	 */
-	order = get_order(tsbsize * 8 * 1024);
-	tsbbase = __get_free_pages(GFP_KERNEL, order);
-	if (!tsbbase) {
-		prom_printf("%s: Error, gfp(tsb) failed.\n", pbm->name);
-		prom_halt();
-	}
-
-	iommu->page_table = (iopte_t *)tsbbase;
-	iommu->page_table_map_base = vdma[0];
-	iommu->dma_addr_mask = dma_mask;
-	pci_iommu_table_init(iommu, PAGE_SIZE << order);
-
-	switch (tsbsize) {
-	case 64:
-		iommu->page_table_sz_bits = 16;
-		break;
-
-	case 128:
-		iommu->page_table_sz_bits = 17;
-		break;
-
-	default:
-		prom_printf("iommu_init: Illegal TSB size %d\n", tsbsize);
-		prom_halt();
-		break;
-	};
-
-	/* We start with no consistent mappings. */
-	iommu->lowest_consistent_map =
-		1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS);
-
-	for (i = 0; i < PBM_NCLUSTERS; i++) {
-		iommu->alloc_info[i].flush = 0;
-		iommu->alloc_info[i].next = 0;
-	}
-
-	schizo_write(iommu->iommu_tsbbase, __pa(tsbbase));
+	schizo_write(iommu->iommu_tsbbase, __pa(iommu->page_table));
 
 	control = schizo_read(iommu->iommu_control);
 	control &= ~(SCHIZO_IOMMU_CTRL_TSBSZ | SCHIZO_IOMMU_CTRL_TBWSZ);
diff --git a/arch/sparc64/kernel/power.c b/arch/sparc64/kernel/power.c
index 946cee0257ea..9e8362ea3104 100644
--- a/arch/sparc64/kernel/power.c
+++ b/arch/sparc64/kernel/power.c
@@ -17,6 +17,7 @@
 
 #include <asm/system.h>
 #include <asm/ebus.h>
+#include <asm/isa.h>
 #include <asm/auxio.h>
 
 #include <linux/unistd.h>
@@ -100,46 +101,83 @@ again:
 	return 0;
 }
 
-static int __init has_button_interrupt(struct linux_ebus_device *edev)
+static int __init has_button_interrupt(unsigned int irq, int prom_node)
 {
-	if (edev->irqs[0] == PCI_IRQ_NONE)
+	if (irq == PCI_IRQ_NONE)
 		return 0;
-	if (!prom_node_has_property(edev->prom_node, "button"))
+	if (!prom_node_has_property(prom_node, "button"))
 		return 0;
 
 	return 1;
 }
 
-void __init power_init(void)
+static int __init power_probe_ebus(struct resource **resp, unsigned int *irq_p, int *prom_node_p)
 {
 	struct linux_ebus *ebus;
 	struct linux_ebus_device *edev;
+
+	for_each_ebus(ebus) {
+		for_each_ebusdev(edev, ebus) {
+			if (!strcmp(edev->prom_name, "power")) {
+				*resp = &edev->resource[0];
+				*irq_p = edev->irqs[0];
+				*prom_node_p = edev->prom_node;
+				return 0;
+			}
+		}
+	}
+	return -ENODEV;
+}
+
+static int __init power_probe_isa(struct resource **resp, unsigned int *irq_p, int *prom_node_p)
+{
+	struct sparc_isa_bridge *isa_bus;
+	struct sparc_isa_device *isa_dev;
+
+	for_each_isa(isa_bus) {
+		for_each_isadev(isa_dev, isa_bus) {
+			if (!strcmp(isa_dev->prom_name, "power")) {
+				*resp = &isa_dev->resource;
+				*irq_p = isa_dev->irq;
+				*prom_node_p = isa_dev->prom_node;
+				return 0;
+			}
+		}
+	}
+	return -ENODEV;
+}
+
+void __init power_init(void)
+{
+	struct resource *res = NULL;
+	unsigned int irq;
+	int prom_node;
 	static int invoked;
 
 	if (invoked)
 		return;
 	invoked = 1;
 
-	for_each_ebus(ebus) {
-		for_each_ebusdev(edev, ebus) {
-			if (!strcmp(edev->prom_name, "power"))
-				goto found;
-		}
-	}
+	if (!power_probe_ebus(&res, &irq, &prom_node))
+		goto found;
+
+	if (!power_probe_isa(&res, &irq, &prom_node))
+		goto found;
+
 	return;
 
 found:
-	power_reg = ioremap(edev->resource[0].start, 0x4);
+	power_reg = ioremap(res->start, 0x4);
 	printk("power: Control reg at %p ... ", power_reg);
 	poweroff_method = machine_halt;  /* able to use the standard halt */
-	if (has_button_interrupt(edev)) {
+	if (has_button_interrupt(irq, prom_node)) {
 		if (kernel_thread(powerd, NULL, CLONE_FS) < 0) {
 			printk("Failed to start power daemon.\n");
 			return;
 		}
 		printk("powerd running.\n");
 
-		if (request_irq(edev->irqs[0],
+		if (request_irq(irq,
 				power_handler, SA_SHIRQ, "power", NULL) < 0)
 			printk("power: Error, cannot register IRQ handler.\n");
 	} else {
diff --git a/arch/sparc64/kernel/rtrap.S b/arch/sparc64/kernel/rtrap.S
index ecfb42a69a44..090dcca00d2a 100644
--- a/arch/sparc64/kernel/rtrap.S
+++ b/arch/sparc64/kernel/rtrap.S
@@ -312,32 +312,33 @@ kern_fpucheck:	ldub			[%g6 + TI_FPDEPTH], %l5
 		wr			%g1, FPRS_FEF, %fprs
 		ldx			[%o1 + %o5], %g1
 		add			%g6, TI_XFSR, %o1
-		membar			#StoreLoad | #LoadLoad
 		sll			%o0, 8, %o2
 		add			%g6, TI_FPREGS, %o3
 		brz,pn			%l6, 1f
 		 add			%g6, TI_FPREGS+0x40, %o4
 
+		membar			#Sync
 		ldda			[%o3 + %o2] ASI_BLK_P, %f0
 		ldda			[%o4 + %o2] ASI_BLK_P, %f16
+		membar			#Sync
 1:		andcc			%l2, FPRS_DU, %g0
 		be,pn			%icc, 1f
 		 wr			%g1, 0, %gsr
 		add			%o2, 0x80, %o2
+		membar			#Sync
 		ldda			[%o3 + %o2] ASI_BLK_P, %f32
 		ldda			[%o4 + %o2] ASI_BLK_P, %f48
-
 1:		membar			#Sync
 		ldx			[%o1 + %o5], %fsr
 2:		stb			%l5, [%g6 + TI_FPDEPTH]
 		ba,pt			%xcc, rt_continue
 		 nop
 5:		wr			%g0, FPRS_FEF, %fprs
-		membar			#StoreLoad | #LoadLoad
 		sll			%o0, 8, %o2
 
 		add			%g6, TI_FPREGS+0x80, %o3
 		add			%g6, TI_FPREGS+0xc0, %o4
+		membar			#Sync
 		ldda			[%o3 + %o2] ASI_BLK_P, %f32
 		ldda			[%o4 + %o2] ASI_BLK_P, %f48
 		membar			#Sync
diff --git a/arch/sparc64/kernel/smp.c b/arch/sparc64/kernel/smp.c
index 590df5a16f5a..b137fd63f5e1 100644
--- a/arch/sparc64/kernel/smp.c
+++ b/arch/sparc64/kernel/smp.c
@@ -1001,13 +1001,6 @@ void smp_penguin_jailcell(int irq, struct pt_regs *regs)
 	preempt_enable();
 }
 
-extern unsigned long xcall_promstop;
-
-void smp_promstop_others(void)
-{
-	smp_cross_call(&xcall_promstop, 0, 0, 0);
-}
-
 #define prof_multiplier(__cpu)		cpu_data(__cpu).multiplier
 #define prof_counter(__cpu)		cpu_data(__cpu).counter