Smarter and more scalable tlb flushing routines. No need to interrupt

other cpus when flushing tlbs for single threaded mm's. Rather, just make
sure the mm will have to do a new context allocation if it runs on the
other cpus.

1 files changed, 43 insertions, 12 deletions

diff --git a/arch/mips64/kernel/smp.c b/arch/mips64/kernel/smp.c
index c36337564..c72d46276 100644
--- a/arch/mips64/kernel/smp.c
+++ b/arch/mips64/kernel/smp.c
@@ -205,9 +205,26 @@ static void flush_tlb_mm_ipi(void *mm)
 	_flush_tlb_mm((struct mm_struct *)mm);
 }
 
+/*
+ * The following tlb flush calls are invoked when old translations are 
+ * being torn down, or pte attributes are changing. For single threaded
+ * address spaces, a new context is obtained on the current cpu, and tlb
+ * context on other cpus are invalidated to force a new context allocation
+ * at switch_mm time, should the mm ever be used on other cpus. For 
+ * multithreaded address spaces, intercpu interrupts have to be sent.
+ * Kanoj 07/00.
+ */
+
 void flush_tlb_mm(struct mm_struct *mm)
 {
-	smp_call_function(flush_tlb_mm_ipi, (void *)mm, 1, 1);
+	if (atomic_read(&mm->mm_users) != 1) {
+		smp_call_function(flush_tlb_mm_ipi, (void *)mm, 1, 1);
+	} else {
+		int i;
+		for (i = 0; i < smp_num_cpus; i++)
+			if (smp_processor_id() != i)
+				CPU_CONTEXT(i, mm) = 0;
+	}
 	_flush_tlb_mm(mm);
 }
 
@@ -227,12 +244,19 @@ static void flush_tlb_range_ipi(void *info)
 
 void flush_tlb_range(struct mm_struct *mm, unsigned long start, unsigned long end)
 {
-	struct flush_tlb_data fd;
-
-	fd.mm = mm;
-	fd.addr1 = start;
-	fd.addr2 = end;
-	smp_call_function(flush_tlb_range_ipi, (void *)&fd, 1, 1);
+	if (atomic_read(&mm->mm_users) != 1) {
+		struct flush_tlb_data fd;
+
+		fd.mm = mm;
+		fd.addr1 = start;
+		fd.addr2 = end;
+		smp_call_function(flush_tlb_range_ipi, (void *)&fd, 1, 1);
+	} else {
+		int i;
+		for (i = 0; i < smp_num_cpus; i++)
+			if (smp_processor_id() != i)
+				CPU_CONTEXT(i, mm) = 0;
+	}
 	_flush_tlb_range(mm, start, end);
 }
 
@@ -245,11 +269,18 @@ static void flush_tlb_page_ipi(void *info)
 
 void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 {
-	struct flush_tlb_data fd;
-
-	fd.vma = vma;
-	fd.addr1 = page;
-	smp_call_function(flush_tlb_page_ipi, (void *)&fd, 1, 1);
+	if (atomic_read(&vma->vm_mm->mm_users) != 1) {
+		struct flush_tlb_data fd;
+
+		fd.vma = vma;
+		fd.addr1 = page;
+		smp_call_function(flush_tlb_page_ipi, (void *)&fd, 1, 1);
+	} else {
+		int i;
+		for (i = 0; i < smp_num_cpus; i++)
+			if (smp_processor_id() != i)
+				CPU_CONTEXT(i, vma->vm_mm) = 0;
+	}
 	_flush_tlb_page(vma, page);
 }