2 files changed, 90 insertions, 5 deletions

diff --git a/arch/ia64/kernel/irq_ia64.c b/arch/ia64/kernel/irq_ia64.c
index 456f57b087c..9a5f41be760 100644
--- a/arch/ia64/kernel/irq_ia64.c
+++ b/arch/ia64/kernel/irq_ia64.c
@@ -39,6 +39,7 @@
 #include <asm/machvec.h>
 #include <asm/pgtable.h>
 #include <asm/system.h>
+#include <asm/tlbflush.h>
 
 #ifdef CONFIG_PERFMON
 # include <asm/perfmon.h>
@@ -127,8 +128,10 @@ void destroy_irq(unsigned int irq)
 
 #ifdef CONFIG_SMP
 #	define IS_RESCHEDULE(vec)	(vec == IA64_IPI_RESCHEDULE)
+#	define IS_LOCAL_TLB_FLUSH(vec)	(vec == IA64_IPI_LOCAL_TLB_FLUSH)
 #else
 #	define IS_RESCHEDULE(vec)	(0)
+#	define IS_LOCAL_TLB_FLUSH(vec)	(0)
 #endif
 /*
  * That's where the IVT branches when we get an external
@@ -180,8 +183,11 @@ ia64_handle_irq (ia64_vector vector, struct pt_regs *regs)
 	saved_tpr = ia64_getreg(_IA64_REG_CR_TPR);
 	ia64_srlz_d();
 	while (vector != IA64_SPURIOUS_INT_VECTOR) {
-		if (unlikely(IS_RESCHEDULE(vector)))
-			 kstat_this_cpu.irqs[vector]++;
+		if (unlikely(IS_LOCAL_TLB_FLUSH(vector))) {
+			smp_local_flush_tlb();
+			kstat_this_cpu.irqs[vector]++;
+		} else if (unlikely(IS_RESCHEDULE(vector)))
+			kstat_this_cpu.irqs[vector]++;
 		else {
 			ia64_setreg(_IA64_REG_CR_TPR, vector);
 			ia64_srlz_d();
@@ -227,8 +233,11 @@ void ia64_process_pending_intr(void)
 	  * Perform normal interrupt style processing
 	  */
 	while (vector != IA64_SPURIOUS_INT_VECTOR) {
-		if (unlikely(IS_RESCHEDULE(vector)))
-			 kstat_this_cpu.irqs[vector]++;
+		if (unlikely(IS_LOCAL_TLB_FLUSH(vector))) {
+			smp_local_flush_tlb();
+			kstat_this_cpu.irqs[vector]++;
+		} else if (unlikely(IS_RESCHEDULE(vector)))
+			kstat_this_cpu.irqs[vector]++;
 		else {
 			struct pt_regs *old_regs = set_irq_regs(NULL);
 
@@ -260,12 +269,12 @@ void ia64_process_pending_intr(void)
 
 
 #ifdef CONFIG_SMP
-extern irqreturn_t handle_IPI (int irq, void *dev_id);
 
 static irqreturn_t dummy_handler (int irq, void *dev_id)
 {
 	BUG();
 }
+extern irqreturn_t handle_IPI (int irq, void *dev_id);
 
 static struct irqaction ipi_irqaction = {
 	.handler =	handle_IPI,
@@ -278,6 +287,13 @@ static struct irqaction resched_irqaction = {
 	.flags =	IRQF_DISABLED,
 	.name =		"resched"
 };
+
+static struct irqaction tlb_irqaction = {
+	.handler =	dummy_handler,
+	.flags =	SA_INTERRUPT,
+	.name =		"tlb_flush"
+};
+
 #endif
 
 void
@@ -303,6 +319,7 @@ init_IRQ (void)
 #ifdef CONFIG_SMP
 	register_percpu_irq(IA64_IPI_VECTOR, &ipi_irqaction);
 	register_percpu_irq(IA64_IPI_RESCHEDULE, &resched_irqaction);
+	register_percpu_irq(IA64_IPI_LOCAL_TLB_FLUSH, &tlb_irqaction);
 #endif
 #ifdef CONFIG_PERFMON
 	pfm_init_percpu();
diff --git a/arch/ia64/kernel/smp.c b/arch/ia64/kernel/smp.c
index 55ddd809b02..221de380456 100644
--- a/arch/ia64/kernel/smp.c
+++ b/arch/ia64/kernel/smp.c
@@ -50,6 +50,18 @@
 #include <asm/mca.h>
 
 /*
+ * Note: alignment of 4 entries/cacheline was empirically determined
+ * to be a good tradeoff between hot cachelines & spreading the array
+ * across too many cacheline.
+ */
+static struct local_tlb_flush_counts {
+	unsigned int count;
+} __attribute__((__aligned__(32))) local_tlb_flush_counts[NR_CPUS];
+
+static DEFINE_PER_CPU(unsigned int, shadow_flush_counts[NR_CPUS]) ____cacheline_aligned;
+
+
+/*
  * Structure and data for smp_call_function(). This is designed to minimise static memory
  * requirements. It also looks cleaner.
  */
@@ -248,6 +260,62 @@ smp_send_reschedule (int cpu)
 	platform_send_ipi(cpu, IA64_IPI_RESCHEDULE, IA64_IPI_DM_INT, 0);
 }
 
+/*
+ * Called with preeemption disabled.
+ */
+static void
+smp_send_local_flush_tlb (int cpu)
+{
+	platform_send_ipi(cpu, IA64_IPI_LOCAL_TLB_FLUSH, IA64_IPI_DM_INT, 0);
+}
+
+void
+smp_local_flush_tlb(void)
+{
+	/*
+	 * Use atomic ops. Otherwise, the load/increment/store sequence from
+	 * a "++" operation can have the line stolen between the load & store.
+	 * The overhead of the atomic op in negligible in this case & offers
+	 * significant benefit for the brief periods where lots of cpus
+	 * are simultaneously flushing TLBs.
+	 */
+	ia64_fetchadd(1, &local_tlb_flush_counts[smp_processor_id()].count, acq);
+	local_flush_tlb_all();
+}
+
+#define FLUSH_DELAY	5 /* Usec backoff to eliminate excessive cacheline bouncing */
+
+void
+smp_flush_tlb_cpumask(cpumask_t xcpumask)
+{
+	unsigned int *counts = __ia64_per_cpu_var(shadow_flush_counts);
+	cpumask_t cpumask = xcpumask;
+	int mycpu, cpu, flush_mycpu = 0;
+
+	preempt_disable();
+	mycpu = smp_processor_id();
+
+	for_each_cpu_mask(cpu, cpumask)
+		counts[cpu] = local_tlb_flush_counts[cpu].count;
+
+	mb();
+	for_each_cpu_mask(cpu, cpumask) {
+		if (cpu == mycpu)
+			flush_mycpu = 1;
+		else
+			smp_send_local_flush_tlb(cpu);
+	}
+
+	if (flush_mycpu)
+		smp_local_flush_tlb();
+
+	for_each_cpu_mask(cpu, cpumask)
+		while(counts[cpu] == local_tlb_flush_counts[cpu].count)
+			udelay(FLUSH_DELAY);
+
+	preempt_enable();
+}
+
 void
 smp_flush_tlb_all (void)
 {