[SPARC64]: Access TSB with physical addresses when possible.

This way we don't need to lock the TSB into the TLB.
The trick is that every TSB load/store is registered into
a special instruction patch section.  The default uses
virtual addresses, and the patch instructions use physical
address load/stores.

We can't do this on all chips because only cheetah+ and later
have the physical variant of the atomic quad load.

Signed-off-by: David S. Miller <davem@davemloft.net>

9 files changed, 234 insertions, 53 deletions

diff --git a/arch/sparc64/kernel/dtlb_miss.S b/arch/sparc64/kernel/dtlb_miss.S
index d0f1565cb564..2ef6f6e6e72b 100644
--- a/arch/sparc64/kernel/dtlb_miss.S
+++ b/arch/sparc64/kernel/dtlb_miss.S
@@ -4,7 +4,7 @@
 	srlx	%g6, 48, %g5			! Get context
 	brz,pn	%g5, kvmap_dtlb			! Context 0 processing
 	 nop					! Delay slot (fill me)
-	ldda	[%g1] ASI_NUCLEUS_QUAD_LDD, %g4	! Load TSB entry
+	TSB_LOAD_QUAD(%g1, %g4)			! Load TSB entry
 	nop					! Push branch to next I$ line
 	cmp	%g4, %g6			! Compare TAG
 
diff --git a/arch/sparc64/kernel/itlb_miss.S b/arch/sparc64/kernel/itlb_miss.S
index 6b6c8fee04bd..97facce27aad 100644
--- a/arch/sparc64/kernel/itlb_miss.S
+++ b/arch/sparc64/kernel/itlb_miss.S
@@ -4,7 +4,7 @@
 	srlx	%g6, 48, %g5			! Get context
 	brz,pn	%g5, kvmap_itlb			! Context 0 processing
 	 nop					! Delay slot (fill me)
-	ldda	[%g1] ASI_NUCLEUS_QUAD_LDD, %g4	! Load TSB entry
+	TSB_LOAD_QUAD(%g1, %g4)			! Load TSB entry
 	cmp	%g4, %g6			! Compare TAG
 	sethi	%hi(_PAGE_EXEC), %g4		! Setup exec check
 
diff --git a/arch/sparc64/kernel/ktlb.S b/arch/sparc64/kernel/ktlb.S
index 2b5e71b68882..9b415ab6db6b 100644
--- a/arch/sparc64/kernel/ktlb.S
+++ b/arch/sparc64/kernel/ktlb.S
@@ -44,14 +44,14 @@ kvmap_itlb_tsb_miss:
 kvmap_itlb_vmalloc_addr:
 	KERN_PGTABLE_WALK(%g4, %g5, %g2, kvmap_itlb_longpath)
 
-	TSB_LOCK_TAG(%g1, %g2, %g4)
+	KTSB_LOCK_TAG(%g1, %g2, %g4)
 
 	/* Load and check PTE.  */
 	ldxa		[%g5] ASI_PHYS_USE_EC, %g5
 	brgez,a,pn	%g5, kvmap_itlb_longpath
-	 stx		%g0, [%g1]
+	 KTSB_STORE(%g1, %g0)
 
-	TSB_WRITE(%g1, %g5, %g6)
+	KTSB_WRITE(%g1, %g5, %g6)
 
 	/* fallthrough to TLB load */
 
@@ -69,9 +69,9 @@ kvmap_itlb_longpath:
 kvmap_itlb_obp:
 	OBP_TRANS_LOOKUP(%g4, %g5, %g2, %g3, kvmap_itlb_longpath)
 
-	TSB_LOCK_TAG(%g1, %g2, %g4)
+	KTSB_LOCK_TAG(%g1, %g2, %g4)
 
-	TSB_WRITE(%g1, %g5, %g6)
+	KTSB_WRITE(%g1, %g5, %g6)
 
 	ba,pt		%xcc, kvmap_itlb_load
 	 nop
@@ -79,9 +79,9 @@ kvmap_itlb_obp:
 kvmap_dtlb_obp:
 	OBP_TRANS_LOOKUP(%g4, %g5, %g2, %g3, kvmap_dtlb_longpath)
 
-	TSB_LOCK_TAG(%g1, %g2, %g4)
+	KTSB_LOCK_TAG(%g1, %g2, %g4)
 
-	TSB_WRITE(%g1, %g5, %g6)
+	KTSB_WRITE(%g1, %g5, %g6)
 
 	ba,pt		%xcc, kvmap_dtlb_load
 	 nop
@@ -114,14 +114,14 @@ kvmap_linear_patch:
 kvmap_dtlb_vmalloc_addr:
 	KERN_PGTABLE_WALK(%g4, %g5, %g2, kvmap_dtlb_longpath)
 
-	TSB_LOCK_TAG(%g1, %g2, %g4)
+	KTSB_LOCK_TAG(%g1, %g2, %g4)
 
 	/* Load and check PTE.  */
 	ldxa		[%g5] ASI_PHYS_USE_EC, %g5
 	brgez,a,pn	%g5, kvmap_dtlb_longpath
-	 stx		%g0, [%g1]
+	 KTSB_STORE(%g1, %g0)
 
-	TSB_WRITE(%g1, %g5, %g6)
+	KTSB_WRITE(%g1, %g5, %g6)
 
 	/* fallthrough to TLB load */
 
diff --git a/arch/sparc64/kernel/tsb.S b/arch/sparc64/kernel/tsb.S
index e1dd37f5e535..ff6a79beb98d 100644
--- a/arch/sparc64/kernel/tsb.S
+++ b/arch/sparc64/kernel/tsb.S
@@ -53,7 +53,7 @@ tsb_reload:
 	/* Load and check PTE.  */
 	ldxa		[%g5] ASI_PHYS_USE_EC, %g5
 	brgez,a,pn	%g5, tsb_do_fault
-	 stx		%g0, [%g1]
+	 TSB_STORE(%g1, %g0)
 
 	/* If it is larger than the base page size, don't
 	 * bother putting it into the TSB.
@@ -64,7 +64,7 @@ tsb_reload:
 	and		%g2, %g4, %g2
 	cmp		%g2, %g7
 	bne,a,pn	%xcc, tsb_tlb_reload
-	 stx		%g0, [%g1]
+	 TSB_STORE(%g1, %g0)
 
 	TSB_WRITE(%g1, %g5, %g6)
 
@@ -131,13 +131,13 @@ winfix_trampoline:
 
 	/* Insert an entry into the TSB.
 	 *
-	 * %o0: TSB entry pointer
+	 * %o0: TSB entry pointer (virt or phys address)
 	 * %o1: tag
 	 * %o2:	pte
 	 */
 	.align	32
-	.globl	tsb_insert
-tsb_insert:
+	.globl	__tsb_insert
+__tsb_insert:
 	rdpr	%pstate, %o5
 	wrpr	%o5, PSTATE_IE, %pstate
 	TSB_LOCK_TAG(%o0, %g2, %g3)
@@ -146,6 +146,31 @@ tsb_insert:
 	retl
 	 nop
 
+	/* Flush the given TSB entry if it has the matching
+	 * tag.
+	 *
+	 * %o0: TSB entry pointer (virt or phys address)
+	 * %o1:	tag
+	 */
+	.align	32
+	.globl	tsb_flush
+tsb_flush:
+	sethi	%hi(TSB_TAG_LOCK_HIGH), %g2
+1:	TSB_LOAD_TAG(%o0, %g1)
+	srlx	%g1, 32, %o3
+	andcc	%o3, %g2, %g0
+	bne,pn	%icc, 1b
+	 membar	#LoadLoad
+	cmp	%g1, %o1
+	bne,pt	%xcc, 2f
+	 clr	%o3
+	TSB_CAS_TAG(%o0, %g1, %o3)
+	cmp	%g1, %o3
+	bne,pn	%xcc, 1b
+	 nop
+2:	retl
+	 TSB_MEMBAR
+
 	/* Reload MMU related context switch state at
 	 * schedule() time.
 	 *
diff --git a/arch/sparc64/kernel/vmlinux.lds.S b/arch/sparc64/kernel/vmlinux.lds.S
index 467d13a0d5c1..71b943f1c9b1 100644
--- a/arch/sparc64/kernel/vmlinux.lds.S
+++ b/arch/sparc64/kernel/vmlinux.lds.S
@@ -70,6 +70,10 @@ SECTIONS
   .con_initcall.init : { *(.con_initcall.init) }
   __con_initcall_end = .;
   SECURITY_INIT
+  . = ALIGN(4);
+  __tsb_phys_patch = .;
+  .tsb_phys_patch : { *(.tsb_phys_patch) }
+  __tsb_phys_patch_end = .;
   . = ALIGN(8192); 
   __initramfs_start = .;
   .init.ramfs : { *(.init.ramfs) }
diff --git a/arch/sparc64/mm/init.c b/arch/sparc64/mm/init.c
index 2c21d85de78f..4893f3e2c336 100644
--- a/arch/sparc64/mm/init.c
+++ b/arch/sparc64/mm/init.c
@@ -39,6 +39,7 @@
 #include <asm/tlb.h>
 #include <asm/spitfire.h>
 #include <asm/sections.h>
+#include <asm/tsb.h>
 
 extern void device_scan(void);
 
@@ -244,6 +245,16 @@ static __inline__ void clear_dcache_dirty_cpu(struct page *page, unsigned long c
 			     : "g1", "g7");
 }
 
+static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
+{
+	unsigned long tsb_addr = (unsigned long) ent;
+
+	if (tlb_type == cheetah_plus)
+		tsb_addr = __pa(tsb_addr);
+
+	__tsb_insert(tsb_addr, tag, pte);
+}
+
 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t pte)
 {
 	struct mm_struct *mm;
@@ -1040,6 +1051,24 @@ unsigned long __init find_ecache_flush_span(unsigned long size)
 	return ~0UL;
 }
 
+static void __init tsb_phys_patch(void)
+{
+	struct tsb_phys_patch_entry *p;
+
+	p = &__tsb_phys_patch;
+	while (p < &__tsb_phys_patch_end) {
+		unsigned long addr = p->addr;
+
+		*(unsigned int *) addr = p->insn;
+		wmb();
+		__asm__ __volatile__("flush	%0"
+				     : /* no outputs */
+				     : "r" (addr));
+
+		p++;
+	}
+}
+
 /* paging_init() sets up the page tables */
 
 extern void cheetah_ecache_flush_init(void);
@@ -1052,6 +1081,9 @@ void __init paging_init(void)
 	unsigned long end_pfn, pages_avail, shift;
 	unsigned long real_end, i;
 
+	if (tlb_type == cheetah_plus)
+		tsb_phys_patch();
+
 	/* Find available physical memory... */
 	read_obp_memory("available", &pavail[0], &pavail_ents);
 
diff --git a/arch/sparc64/mm/tsb.c b/arch/sparc64/mm/tsb.c
index 1c4e5c2dfc53..787533f01049 100644
--- a/arch/sparc64/mm/tsb.c
+++ b/arch/sparc64/mm/tsb.c
@@ -20,12 +20,9 @@ static inline unsigned long tsb_hash(unsigned long vaddr, unsigned long nentries
 	return vaddr & (nentries - 1);
 }
 
-static inline int tag_compare(struct tsb *entry, unsigned long vaddr, unsigned long context)
+static inline int tag_compare(unsigned long tag, unsigned long vaddr, unsigned long context)
 {
-	if (context == ~0UL)
-		return 1;
-
-	return (entry->tag == ((vaddr >> 22) | (context << 48)));
+	return (tag == ((vaddr >> 22) | (context << 48)));
 }
 
 /* TSB flushes need only occur on the processor initiating the address
@@ -41,7 +38,7 @@ void flush_tsb_kernel_range(unsigned long start, unsigned long end)
 		unsigned long hash = tsb_hash(v, KERNEL_TSB_NENTRIES);
 		struct tsb *ent = &swapper_tsb[hash];
 
-		if (tag_compare(ent, v, 0)) {
+		if (tag_compare(ent->tag, v, 0)) {
 			ent->tag = 0UL;
 			membar_storeload_storestore();
 		}
@@ -52,24 +49,31 @@ void flush_tsb_user(struct mmu_gather *mp)
 {
 	struct mm_struct *mm = mp->mm;
 	struct tsb *tsb = mm->context.tsb;
-	unsigned long ctx = ~0UL;
 	unsigned long nentries = mm->context.tsb_nentries;
+	unsigned long ctx, base;
 	int i;
 
-	if (CTX_VALID(mm->context))
-		ctx = CTX_HWBITS(mm->context);
+	if (unlikely(!CTX_VALID(mm->context)))
+		return;
+
+	ctx = CTX_HWBITS(mm->context);
 
+	if (tlb_type == cheetah_plus)
+		base = __pa(tsb);
+	else
+		base = (unsigned long) tsb;
+	
 	for (i = 0; i < mp->tlb_nr; i++) {
 		unsigned long v = mp->vaddrs[i];
-		struct tsb *ent;
+		unsigned long tag, ent, hash;
 
 		v &= ~0x1UL;
 
-		ent = &tsb[tsb_hash(v, nentries)];
-		if (tag_compare(ent, v, ctx)) {
-			ent->tag = 0UL;
-			membar_storeload_storestore();
-		}
+		hash = tsb_hash(v, nentries);
+		ent = base + (hash * sizeof(struct tsb));
+		tag = (v >> 22UL) | (ctx << 48UL);
+
+		tsb_flush(ent, tag);
 	}
 }
 
@@ -84,6 +88,7 @@ static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_bytes)
 	tte = (_PAGE_VALID | _PAGE_L | _PAGE_CP |
 	       _PAGE_CV    | _PAGE_P | _PAGE_W);
 	tsb_paddr = __pa(mm->context.tsb);
+	BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
 
 	/* Use the smallest page size that can map the whole TSB
 	 * in one TLB entry.
@@ -144,13 +149,23 @@ static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_bytes)
 		BUG();
 	};
 
-	tsb_reg |= base;
-	tsb_reg |= (tsb_paddr & (page_sz - 1UL));
-	tte |= (tsb_paddr & ~(page_sz - 1UL));
+	if (tlb_type == cheetah_plus) {
+		/* Physical mapping, no locked TLB entry for TSB.  */
+		tsb_reg |= tsb_paddr;
+
+		mm->context.tsb_reg_val = tsb_reg;
+		mm->context.tsb_map_vaddr = 0;
+		mm->context.tsb_map_pte = 0;
+	} else {
+		tsb_reg |= base;
+		tsb_reg |= (tsb_paddr & (page_sz - 1UL));
+		tte |= (tsb_paddr & ~(page_sz - 1UL));
+
+		mm->context.tsb_reg_val = tsb_reg;
+		mm->context.tsb_map_vaddr = base;
+		mm->context.tsb_map_pte = tte;
+	}
 
-	mm->context.tsb_reg_val = tsb_reg;
-	mm->context.tsb_map_vaddr = base;
-	mm->context.tsb_map_pte = tte;
 }
 
 /* The page tables are locked against modifications while this
@@ -168,13 +183,21 @@ static void copy_tsb(struct tsb *old_tsb, unsigned long old_size,
 	for (i = 0; i < old_nentries; i++) {
 		register unsigned long tag asm("o4");
 		register unsigned long pte asm("o5");
-		unsigned long v;
-		unsigned int hash;
-
-		__asm__ __volatile__(
-			"ldda [%2] %3, %0"
-			: "=r" (tag), "=r" (pte)
-			: "r" (&old_tsb[i]), "i" (ASI_NUCLEUS_QUAD_LDD));
+		unsigned long v, hash;
+
+		if (tlb_type == cheetah_plus) {
+			__asm__ __volatile__(
+				"ldda [%2] %3, %0"
+				: "=r" (tag), "=r" (pte)
+				: "r" (__pa(&old_tsb[i])),
+				  "i" (ASI_QUAD_LDD_PHYS));
+		} else {
+			__asm__ __volatile__(
+				"ldda [%2] %3, %0"
+				: "=r" (tag), "=r" (pte)
+				: "r" (&old_tsb[i]),
+				  "i" (ASI_NUCLEUS_QUAD_LDD));
+		}
 
 		if (!tag || (tag & (1UL << TSB_TAG_LOCK_BIT)))
 			continue;
@@ -198,8 +221,20 @@ static void copy_tsb(struct tsb *old_tsb, unsigned long old_size,
 		v |= (i & (512UL - 1UL)) << 13UL;
 
 		hash = tsb_hash(v, new_nentries);
-		new_tsb[hash].tag = tag;
-		new_tsb[hash].pte = pte;
+		if (tlb_type == cheetah_plus) {
+			__asm__ __volatile__(
+				"stxa	%0, [%1] %2\n\t"
+				"stxa	%3, [%4] %2"
+				: /* no outputs */
+				: "r" (tag),
+				  "r" (__pa(&new_tsb[hash].tag)),
+				  "i" (ASI_PHYS_USE_EC),
+				  "r" (pte),
+				  "r" (__pa(&new_tsb[hash].pte)));
+		} else {
+			new_tsb[hash].tag = tag;
+			new_tsb[hash].pte = pte;
+		}
 	}
 }
 
diff --git a/include/asm-sparc64/mmu.h b/include/asm-sparc64/mmu.h
index 18f98edfbcda..55e622711b96 100644
--- a/include/asm-sparc64/mmu.h
+++ b/include/asm-sparc64/mmu.h
@@ -97,7 +97,8 @@ struct tsb {
 	unsigned long pte;
 } __attribute__((aligned(TSB_ENTRY_ALIGNMENT)));
 
-extern void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte);
+extern void __tsb_insert(unsigned long ent, unsigned long tag, unsigned long pte);
+extern void tsb_flush(unsigned long ent, unsigned long tag);
 
 typedef struct {
 	unsigned long	sparc64_ctx_val;
diff --git a/include/asm-sparc64/tsb.h b/include/asm-sparc64/tsb.h
index f384565212fe..44709cde5617 100644
--- a/include/asm-sparc64/tsb.h
+++ b/include/asm-sparc64/tsb.h
@@ -44,7 +44,89 @@
 
 #define TSB_MEMBAR	membar	#StoreStore
 
+/* Some cpus support physical address quad loads.  We want to use
+ * those if possible so we don't need to hard-lock the TSB mapping
+ * into the TLB.  We encode some instruction patching in order to
+ * support this.
+ *
+ * The kernel TSB is locked into the TLB by virtue of being in the
+ * kernel image, so we don't play these games for swapper_tsb access.
+ */
+#ifndef __ASSEMBLY__
+struct tsb_phys_patch_entry {
+	unsigned int	addr;
+	unsigned int	insn;
+};
+extern struct tsb_phys_patch_entry __tsb_phys_patch, __tsb_phys_patch_end;
+#endif
+#define TSB_LOAD_QUAD(TSB, REG)	\
+661:	ldda		[TSB] ASI_NUCLEUS_QUAD_LDD, REG; \
+	.section	.tsb_phys_patch, "ax"; \
+	.word		661b; \
+	ldda		[TSB] ASI_QUAD_LDD_PHYS, REG; \
+	.previous
+
+#define TSB_LOAD_TAG_HIGH(TSB, REG) \
+661:	lduwa		[TSB] ASI_N, REG; \
+	.section	.tsb_phys_patch, "ax"; \
+	.word		661b; \
+	lduwa		[TSB] ASI_PHYS_USE_EC, REG; \
+	.previous
+
+#define TSB_LOAD_TAG(TSB, REG) \
+661:	ldxa		[TSB] ASI_N, REG; \
+	.section	.tsb_phys_patch, "ax"; \
+	.word		661b; \
+	ldxa		[TSB] ASI_PHYS_USE_EC, REG; \
+	.previous
+
+#define TSB_CAS_TAG_HIGH(TSB, REG1, REG2) \
+661:	casa		[TSB] ASI_N, REG1, REG2; \
+	.section	.tsb_phys_patch, "ax"; \
+	.word		661b; \
+	casa		[TSB] ASI_PHYS_USE_EC, REG1, REG2; \
+	.previous
+
+#define TSB_CAS_TAG(TSB, REG1, REG2) \
+661:	casxa		[TSB] ASI_N, REG1, REG2; \
+	.section	.tsb_phys_patch, "ax"; \
+	.word		661b; \
+	casxa		[TSB] ASI_PHYS_USE_EC, REG1, REG2; \
+	.previous
+
+#define TSB_STORE(ADDR, VAL) \
+661:	stxa		VAL, [ADDR] ASI_N; \
+	.section	.tsb_phys_patch, "ax"; \
+	.word		661b; \
+	stxa		VAL, [ADDR] ASI_PHYS_USE_EC; \
+	.previous
+
 #define TSB_LOCK_TAG(TSB, REG1, REG2)	\
+99:	TSB_LOAD_TAG_HIGH(TSB, REG1);	\
+	sethi	%hi(TSB_TAG_LOCK_HIGH), REG2;\
+	andcc	REG1, REG2, %g0;	\
+	bne,pn	%icc, 99b;		\
+	 nop;				\
+	TSB_CAS_TAG_HIGH(TSB, REG1, REG2);	\
+	cmp	REG1, REG2;		\
+	bne,pn	%icc, 99b;		\
+	 nop;				\
+	TSB_MEMBAR
+
+#define TSB_WRITE(TSB, TTE, TAG) \
+	add	TSB, 0x8, TSB;   \
+	TSB_STORE(TSB, TTE);     \
+	sub	TSB, 0x8, TSB;   \
+	TSB_MEMBAR;              \
+	TSB_STORE(TSB, TAG);
+
+#define KTSB_LOAD_QUAD(TSB, REG) \
+	ldda		[TSB] ASI_NUCLEUS_QUAD_LDD, REG;
+
+#define KTSB_STORE(ADDR, VAL) \
+	stxa		VAL, [ADDR] ASI_N;
+
+#define KTSB_LOCK_TAG(TSB, REG1, REG2)	\
 99:	lduwa	[TSB] ASI_N, REG1;	\
 	sethi	%hi(TSB_TAG_LOCK_HIGH), REG2;\
 	andcc	REG1, REG2, %g0;	\
@@ -56,10 +138,12 @@
 	 nop;				\
 	TSB_MEMBAR
 
-#define TSB_WRITE(TSB, TTE, TAG)	   \
-	stx		TTE, [TSB + 0x08]; \
-	TSB_MEMBAR;			   \
-	stx		TAG, [TSB + 0x00];
+#define KTSB_WRITE(TSB, TTE, TAG) \
+	add	TSB, 0x8, TSB;   \
+	stxa	TTE, [TSB] ASI_N;     \
+	sub	TSB, 0x8, TSB;   \
+	TSB_MEMBAR;              \
+	stxa	TAG, [TSB] ASI_N;
 
 	/* Do a kernel page table walk.  Leaves physical PTE pointer in
 	 * REG1.  Jumps to FAIL_LABEL on early page table walk termination.
@@ -157,7 +241,7 @@
 	and		REG2, (KERNEL_TSB_NENTRIES - 1), REG2; \
 	sllx		REG2, 4, REG2; \
 	add		REG1, REG2, REG2; \
-	ldda		[REG2] ASI_NUCLEUS_QUAD_LDD, REG3; \
+	KTSB_LOAD_QUAD(REG2, REG3); \
 	cmp		REG3, TAG; \
 	be,a,pt		%xcc, OK_LABEL; \
 	 mov		REG4, REG1;