[SPARC64]: Fix and re-enable dynamic TSB sizing.

This is good for up to %50 performance improvement of some test cases.
The problem has been the race conditions, and hopefully I've plugged
them all up here.

1) There was a serious race in switch_mm() wrt. lazy TLB
   switching to and from kernel threads.

   We could erroneously skip a tsb_context_switch() and thus
   use a stale TSB across a TSB grow event.

   There is a big comment now in that function describing
   exactly how it can happen.

2) All code paths that do something with the TSB need to be
   guarded with the mm->context.lock spinlock.  This makes
   page table flushing paths properly synchronize with both
   TSB growing and TLB context changes.

3) TSB growing events are moved to the end of successful fault
   processing.  Previously it was in update_mmu_cache() but
   that is deadlock prone.  At the end of do_sparc64_fault()
   we hold no spinlocks that could deadlock the TSB grow
   sequence.  We also have dropped the address space semaphore.

While we're here, add prefetching to the copy_tsb() routine
and put it in assembler into the tsb.S file.  This piece of
code is quite time critical.

There are some small negative side effects to this code which
can be improved upon.  In particular we grab the mm->context.lock
even for the tsb insert done by update_mmu_cache() now and that's
a bit excessive.  We can get rid of that locking, and the same
lock taking in flush_tsb_user(), by disabling PSTATE_IE around
the whole operation including the capturing of the tsb pointer
and tsb_nentries value.  That would work because anyone growing
the TSB won't free up the old TSB until all cpus respond to the
TSB change cross call.

I'm not quite so confident in that optimization to put it in
right now, but eventually we might be able to and the description
is here for reference.

This code seems very solid now.  It passes several parallel GCC
bootstrap builds, and our favorite "nut cruncher" stress test which is
a full "make -j8192" build of a "make allmodconfig" kernel.  That puts
about 256 processes on each cpu's run queue, makes lots of process cpu
migrations occur, causes lots of page table and TLB flushing activity,
incurs many context version number changes, and it swaps the machine
real far out to disk even though there is 16GB of ram on this test
system. :-)
Signed-off-by: David S. Miller <davem@davemloft.net>

arch/sparc64/kernel/tsb.S

@@ -34,8 +34,9 @@ tsb_miss_itlb:
 	 ldxa		[%g4] ASI_IMMU, %g4
 
 	/* At this point we have:
-	 * %g4 --	missing virtual address
 	 * %g1 --	TSB entry address
+	 * %g3 --	FAULT_CODE_{D,I}TLB
+	 * %g4 --	missing virtual address
 	 * %g6 --	TAG TARGET (vaddr >> 22)
 	 */
 tsb_miss_page_table_walk:
@@ -45,6 +46,12 @@ tsb_miss_page_table_walk:
 tsb_miss_page_table_walk_sun4v_fastpath:
 	USER_PGTABLE_WALK_TL1(%g4, %g7, %g5, %g2, tsb_do_fault)
 
+	/* At this point we have:
+	 * %g1 --	TSB entry address
+	 * %g3 --	FAULT_CODE_{D,I}TLB
+	 * %g5 --	physical address of PTE in Linux page tables
+	 * %g6 --	TAG TARGET (vaddr >> 22)
+	 */
 tsb_reload:
 	TSB_LOCK_TAG(%g1, %g2, %g7)
 
@@ -199,6 +206,7 @@ __tsb_insert:
 	wrpr	%o5, %pstate
 	retl
 	 nop
+	.size	__tsb_insert, .-__tsb_insert
 
 	/* Flush the given TSB entry if it has the matching
 	 * tag.
@@ -208,6 +216,7 @@ __tsb_insert:
 	 */
 	.align	32
 	.globl	tsb_flush
+	.type	tsb_flush,#function
 tsb_flush:
 	sethi	%hi(TSB_TAG_LOCK_HIGH), %g2
 1:	TSB_LOAD_TAG(%o0, %g1)
@@ -225,6 +234,7 @@ tsb_flush:
 	 nop
 2:	retl
 	 TSB_MEMBAR
+	.size	tsb_flush, .-tsb_flush
 
 	/* Reload MMU related context switch state at
 	 * schedule() time.
@@ -241,6 +251,7 @@ tsb_flush:
 	 */
 	.align	32
 	.globl	__tsb_context_switch
+	.type	__tsb_context_switch,#function
 __tsb_context_switch:
 	rdpr	%pstate, %o5
 	wrpr	%o5, PSTATE_IE, %pstate
@@ -302,3 +313,61 @@ __tsb_context_switch:
 
 	retl
 	 nop
+	.size	__tsb_context_switch, .-__tsb_context_switch
+
+#define TSB_PASS_BITS	((1 << TSB_TAG_LOCK_BIT) | \
+			 (1 << TSB_TAG_INVALID_BIT))
+
+	.align	32
+	.globl	copy_tsb
+	.type	copy_tsb,#function
+copy_tsb:		/* %o0=old_tsb_base, %o1=old_tsb_size
+			 * %o2=new_tsb_base, %o3=new_tsb_size
+			 */
+	sethi		%uhi(TSB_PASS_BITS), %g7
+	srlx		%o3, 4, %o3
+	add		%o0, %o1, %g1	/* end of old tsb */
+	sllx		%g7, 32, %g7
+	sub		%o3, 1, %o3	/* %o3 == new tsb hash mask */
+
+661:	prefetcha	[%o0] ASI_N, #one_read
+	.section	.tsb_phys_patch, "ax"
+	.word		661b
+	prefetcha	[%o0] ASI_PHYS_USE_EC, #one_read
+	.previous
+
+90:	andcc		%o0, (64 - 1), %g0
+	bne		1f
+	 add		%o0, 64, %o5
+
+661:	prefetcha	[%o5] ASI_N, #one_read
+	.section	.tsb_phys_patch, "ax"
+	.word		661b
+	prefetcha	[%o5] ASI_PHYS_USE_EC, #one_read
+	.previous
+
+1:	TSB_LOAD_QUAD(%o0, %g2)		/* %g2/%g3 == TSB entry */
+	andcc		%g2, %g7, %g0	/* LOCK or INVALID set? */
+	bne,pn		%xcc, 80f	/* Skip it */
+	 sllx		%g2, 22, %o4	/* TAG --> VADDR */
+
+	/* This can definitely be computed faster... */
+	srlx		%o0, 4, %o5	/* Build index */
+	and		%o5, 511, %o5	/* Mask index */
+	sllx		%o5, PAGE_SHIFT, %o5 /* Put into vaddr position */
+	or		%o4, %o5, %o4	/* Full VADDR. */
+	srlx		%o4, PAGE_SHIFT, %o4 /* Shift down to create index */
+	and		%o4, %o3, %o4	/* Mask with new_tsb_nents-1 */
+	sllx		%o4, 4, %o4	/* Shift back up into tsb ent offset */
+	TSB_STORE(%o2 + %o4, %g2)	/* Store TAG */
+	add		%o4, 0x8, %o4	/* Advance to TTE */
+	TSB_STORE(%o2 + %o4, %g3)	/* Store TTE */
+
+80:	add		%o0, 16, %o0
+	cmp		%o0, %g1
+	bne,pt		%xcc, 90b
+	 nop
+
+	retl
+	 TSB_MEMBAR
+	.size		copy_tsb, .-copy_tsb

arch/sparc64/mm/fault.c

@@ -29,6 +29,7 @@
 #include <asm/lsu.h>
 #include <asm/sections.h>
 #include <asm/kdebug.h>
+#include <asm/mmu_context.h>
 
 /*
  * To debug kernel to catch accesses to certain virtual/physical addresses.
@@ -258,7 +259,7 @@ asmlinkage void __kprobes do_sparc64_fault(struct pt_regs *regs)
 	struct vm_area_struct *vma;
 	unsigned int insn = 0;
 	int si_code, fault_code;
-	unsigned long address;
+	unsigned long address, mm_rss;
 
 	fault_code = get_thread_fault_code();
 
@@ -407,6 +408,11 @@ good_area:
 	}
 
 	up_read(&mm->mmap_sem);
+
+	mm_rss = get_mm_rss(mm);
+	if (unlikely(mm_rss >= mm->context.tsb_rss_limit))
+		tsb_grow(mm, mm_rss);
+
 	return;
 
 	/*

arch/sparc64/mm/init.c

@@ -279,7 +279,7 @@ void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t p
 {
 	struct mm_struct *mm;
 	struct tsb *tsb;
-	unsigned long tag;
+	unsigned long tag, flags;
 
 	if (tlb_type != hypervisor) {
 		unsigned long pfn = pte_pfn(pte);
@@ -308,10 +308,15 @@ void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t p
 	}
 
 	mm = vma->vm_mm;
+
+	spin_lock_irqsave(&mm->context.lock, flags);
+
 	tsb = &mm->context.tsb[(address >> PAGE_SHIFT) &
 			       (mm->context.tsb_nentries - 1UL)];
 	tag = (address >> 22UL);
 	tsb_insert(tsb, tag, pte_val(pte));
+
+	spin_unlock_irqrestore(&mm->context.lock, flags);
 }
 
 void flush_dcache_page(struct page *page)

arch/sparc64/mm/tsb.c

@@ -48,11 +48,15 @@ void flush_tsb_kernel_range(unsigned long start, unsigned long end)
 void flush_tsb_user(struct mmu_gather *mp)
 {
 	struct mm_struct *mm = mp->mm;
-	struct tsb *tsb = mm->context.tsb;
-	unsigned long nentries = mm->context.tsb_nentries;
-	unsigned long base;
+	unsigned long nentries, base, flags;
+	struct tsb *tsb;
 	int i;
 
+	spin_lock_irqsave(&mm->context.lock, flags);
+
+	tsb = mm->context.tsb;
+	nentries = mm->context.tsb_nentries;
+
 	if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 		base = __pa(tsb);
 	else
@@ -70,6 +74,8 @@ void flush_tsb_user(struct mmu_gather *mp)
 
 		tsb_flush(ent, tag);
 	}
+
+	spin_unlock_irqrestore(&mm->context.lock, flags);
 }
 
 static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_bytes)
@@ -201,86 +207,9 @@ static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_bytes)
 	}
 }
 
-/* The page tables are locked against modifications while this
- * runs.
- *
- * XXX do some prefetching...
- */
-static void copy_tsb(struct tsb *old_tsb, unsigned long old_size,
-		     struct tsb *new_tsb, unsigned long new_size)
-{
-	unsigned long old_nentries = old_size / sizeof(struct tsb);
-	unsigned long new_nentries = new_size / sizeof(struct tsb);
-	unsigned long i;
-
-	for (i = 0; i < old_nentries; i++) {
-		register unsigned long tag asm("o4");
-		register unsigned long pte asm("o5");
-		unsigned long v, hash;
-
-		if (tlb_type == hypervisor) {
-			__asm__ __volatile__(
-				"ldda [%2] %3, %0"
-				: "=r" (tag), "=r" (pte)
-				: "r" (__pa(&old_tsb[i])),
-				  "i" (ASI_QUAD_LDD_PHYS_4V));
-		} else 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 & ((1UL << TSB_TAG_LOCK_BIT) |
-			   (1UL << TSB_TAG_INVALID_BIT)))
-			continue;
-
-		/* We only put base page size PTEs into the TSB,
-		 * but that might change in the future.  This code
-		 * would need to be changed if we start putting larger
-		 * page size PTEs into there.
-		 */
-		WARN_ON((pte & _PAGE_ALL_SZ_BITS) != _PAGE_SZBITS);
-
-		/* The tag holds bits 22 to 63 of the virtual address
-		 * and the context.  Clear out the context, and shift
-		 * up to make a virtual address.
-		 */
-		v = (tag & ((1UL << 42UL) - 1UL)) << 22UL;
-
-		/* The implied bits of the tag (bits 13 to 21) are
-		 * determined by the TSB entry index, so fill that in.
-		 */
-		v |= (i & (512UL - 1UL)) << 13UL;
-
-		hash = tsb_hash(v, new_nentries);
-		if (tlb_type == cheetah_plus ||
-		    tlb_type == hypervisor) {
-			__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;
-		}
-	}
-}
-
 /* When the RSS of an address space exceeds mm->context.tsb_rss_limit,
- * update_mmu_cache() invokes this routine to try and grow the TSB.
+ * do_sparc64_fault() invokes this routine to try and grow the TSB.
+ *
  * When we reach the maximum TSB size supported, we stick ~0UL into
  * mm->context.tsb_rss_limit so the grow checks in update_mmu_cache()
  * will not trigger any longer.
@@ -293,12 +222,12 @@ static void copy_tsb(struct tsb *old_tsb, unsigned long old_size,
  * the number of entries that the current TSB can hold at once.  Currently,
  * we trigger when the RSS hits 3/4 of the TSB capacity.
  */
-void tsb_grow(struct mm_struct *mm, unsigned long rss, gfp_t gfp_flags)
+void tsb_grow(struct mm_struct *mm, unsigned long rss)
 {
 	unsigned long max_tsb_size = 1 * 1024 * 1024;
-	unsigned long size, old_size;
+	unsigned long size, old_size, flags;
 	struct page *page;
-	struct tsb *old_tsb;
+	struct tsb *old_tsb, *new_tsb;
 
 	if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
 		max_tsb_size = (PAGE_SIZE << MAX_ORDER);
@@ -311,12 +240,51 @@ void tsb_grow(struct mm_struct *mm, unsigned long rss, gfp_t gfp_flags)
 			break;
 	}
 
-	page = alloc_pages(gfp_flags, get_order(size));
+	page = alloc_pages(GFP_KERNEL, get_order(size));
 	if (unlikely(!page))
 		return;
 
 	/* Mark all tags as invalid.  */
-	memset(page_address(page), 0x40, size);
+	new_tsb = page_address(page);
+	memset(new_tsb, 0x40, size);
+
+	/* Ok, we are about to commit the changes.  If we are
+	 * growing an existing TSB the locking is very tricky,
+	 * so WATCH OUT!
+	 *
+	 * We have to hold mm->context.lock while committing to the
+	 * new TSB, this synchronizes us with processors in
+	 * flush_tsb_user() and switch_mm() for this address space.
+	 *
+	 * But even with that lock held, processors run asynchronously
+	 * accessing the old TSB via TLB miss handling.  This is OK
+	 * because those actions are just propagating state from the
+	 * Linux page tables into the TSB, page table mappings are not
+	 * being changed.  If a real fault occurs, the processor will
+	 * synchronize with us when it hits flush_tsb_user(), this is
+	 * also true for the case where vmscan is modifying the page
+	 * tables.  The only thing we need to be careful with is to
+	 * skip any locked TSB entries during copy_tsb().
+	 *
+	 * When we finish committing to the new TSB, we have to drop
+	 * the lock and ask all other cpus running this address space
+	 * to run tsb_context_switch() to see the new TSB table.
+	 */
+	spin_lock_irqsave(&mm->context.lock, flags);
+
+	old_tsb = mm->context.tsb;
+	old_size = mm->context.tsb_nentries * sizeof(struct tsb);
+
+	/* Handle multiple threads trying to grow the TSB at the same time.
+	 * One will get in here first, and bump the size and the RSS limit.
+	 * The others will get in here next and hit this check.
+	 */
+	if (unlikely(old_tsb && (rss < mm->context.tsb_rss_limit))) {
+		spin_unlock_irqrestore(&mm->context.lock, flags);
+
+		free_pages((unsigned long) new_tsb, get_order(size));
+		return;
+	}
 
 	if (size == max_tsb_size)
 		mm->context.tsb_rss_limit = ~0UL;
@@ -324,30 +292,37 @@ void tsb_grow(struct mm_struct *mm, unsigned long rss, gfp_t gfp_flags)
 		mm->context.tsb_rss_limit =
 			((size / sizeof(struct tsb)) * 3) / 4;
 
-	old_tsb = mm->context.tsb;
-	old_size = mm->context.tsb_nentries * sizeof(struct tsb);
-
-	if (old_tsb)
-		copy_tsb(old_tsb, old_size, page_address(page), size);
+	if (old_tsb) {
+		extern void copy_tsb(unsigned long old_tsb_base,
+				     unsigned long old_tsb_size,
+				     unsigned long new_tsb_base,
+				     unsigned long new_tsb_size);
+		unsigned long old_tsb_base = (unsigned long) old_tsb;
+		unsigned long new_tsb_base = (unsigned long) new_tsb;
+
+		if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
+			old_tsb_base = __pa(old_tsb_base);
+			new_tsb_base = __pa(new_tsb_base);
+		}
+		copy_tsb(old_tsb_base, old_size, new_tsb_base, size);
+	}
 
-	mm->context.tsb = page_address(page);
+	mm->context.tsb = new_tsb;
 	setup_tsb_params(mm, size);
 
+	spin_unlock_irqrestore(&mm->context.lock, flags);
+
 	/* If old_tsb is NULL, we're being invoked for the first time
 	 * from init_new_context().
 	 */
 	if (old_tsb) {
-		/* Now force all other processors to reload the new
-		 * TSB state.
-		 */
-		smp_tsb_sync(mm);
-
-		/* Finally reload it on the local cpu.  No further
-		 * references will remain to the old TSB and we can
-		 * thus free it up.
-		 */
+		/* Reload it on the local cpu.  */
 		tsb_context_switch(mm);
 
+		/* Now force other processors to do the same.  */
+		smp_tsb_sync(mm);
+
+		/* Now it is safe to free the old tsb.  */
 		free_pages((unsigned long) old_tsb, get_order(old_size));
 	}
 }
@@ -363,7 +338,11 @@ int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 	 * will be confused and think there is an older TSB to free up.
 	 */
 	mm->context.tsb = NULL;
-	tsb_grow(mm, 0, GFP_KERNEL);
+
+	/* If this is fork, inherit the parent's TSB size.  We would
+	 * grow it to that size on the first page fault anyways.
+	 */
+	tsb_grow(mm, get_mm_rss(mm));
 
 	if (unlikely(!mm->context.tsb))
 		return -ENOMEM;

include/asm-sparc64/mmu_context.h

@@ -42,7 +42,7 @@ static inline void tsb_context_switch(struct mm_struct *mm)
 			     __pa(&mm->context.tsb_descr));
 }
 
-extern void tsb_grow(struct mm_struct *mm, unsigned long mm_rss, gfp_t gfp_flags);
+extern void tsb_grow(struct mm_struct *mm, unsigned long mm_rss);
 #ifdef CONFIG_SMP
 extern void smp_tsb_sync(struct mm_struct *mm);
 #else
@@ -74,18 +74,43 @@ static inline void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm, str
 	ctx_valid = CTX_VALID(mm->context);
 	if (!ctx_valid)
 		get_new_mmu_context(mm);
-	spin_unlock_irqrestore(&mm->context.lock, flags);
 
-	if (!ctx_valid || (old_mm != mm)) {
-		load_secondary_context(mm);
-		tsb_context_switch(mm);
-	}
+	/* We have to be extremely careful here or else we will miss
+	 * a TSB grow if we switch back and forth between a kernel
+	 * thread and an address space which has it's TSB size increased
+	 * on another processor.
+	 *
+	 * It is possible to play some games in order to optimize the
+	 * switch, but the safest thing to do is to unconditionally
+	 * perform the secondary context load and the TSB context switch.
+	 *
+	 * For reference the bad case is, for address space "A":
+	 *
+	 *		CPU 0			CPU 1
+	 *	run address space A
+	 *	set cpu0's bits in cpu_vm_mask
+	 *	switch to kernel thread, borrow
+	 *	address space A via entry_lazy_tlb
+	 *					run address space A
+	 *					set cpu1's bit in cpu_vm_mask
+	 *					flush_tlb_pending()
+	 *					reset cpu_vm_mask to just cpu1
+	 *					TSB grow
+	 *	run address space A
+	 *	context was valid, so skip
+	 *	TSB context switch
+	 *
+	 * At that point cpu0 continues to use a stale TSB, the one from
+	 * before the TSB grow performed on cpu1.  cpu1 did not cross-call
+	 * cpu0 to update it's TSB because at that point the cpu_vm_mask
+	 * only had cpu1 set in it.
+	 */
+	load_secondary_context(mm);
+	tsb_context_switch(mm);
 
-	/* Even if (mm == old_mm) we _must_ check
-	 * the cpu_vm_mask.  If we do not we could
-	 * corrupt the TLB state because of how
-	 * smp_flush_tlb_{page,range,mm} on sparc64
-	 * and lazy tlb switches work. -DaveM
+	/* Any time a processor runs a context on an address space
+	 * for the first time, we must flush that context out of the
+	 * local TLB.
 	 */
 	cpu = smp_processor_id();
 	if (!ctx_valid || !cpu_isset(cpu, mm->cpu_vm_mask)) {
@@ -93,6 +118,7 @@ static inline void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm, str
 		__flush_tlb_mm(CTX_HWBITS(mm->context),
 			       SECONDARY_CONTEXT);
 	}
+	spin_unlock_irqrestore(&mm->context.lock, flags);
 }
 
 #define deactivate_mm(tsk,mm)	do { } while (0)
@@ -109,11 +135,11 @@ static inline void activate_mm(struct mm_struct *active_mm, struct mm_struct *mm
 	cpu = smp_processor_id();
 	if (!cpu_isset(cpu, mm->cpu_vm_mask))
 		cpu_set(cpu, mm->cpu_vm_mask);
-	spin_unlock_irqrestore(&mm->context.lock, flags);
 
 	load_secondary_context(mm);
 	__flush_tlb_mm(CTX_HWBITS(mm->context), SECONDARY_CONTEXT);
 	tsb_context_switch(mm);
+	spin_unlock_irqrestore(&mm->context.lock, flags);
 }
 
 #endif /* !(__ASSEMBLY__) */