bitkeeper revision 1.1159.258.154 (429ae8764I9WsS60DkRARFDRTCFT1Q)

Fix Xen idle loop to enter/exit tickless mode in same way as s390,
which interfaces properly with the RCU subsystem.

This includes a patch that fixes a race in the generic RCU code: it
was created by IBM for s390, and is being pushed upstream by them.

Signed-off-by: Keir Fraser <keir@xensource.com>

4 files changed, 41 insertions, 36 deletions

diff --git a/.rootkeys b/.rootkeys
index b4fa0cad90..54ae38da8c 100644
--- a/.rootkeys
+++ b/.rootkeys
@@ -369,6 +369,7 @@
 42372652KCUP-IOH9RN19YQmGhs4aA patches/linux-2.6.11/iomap.patch
 428359d4b3fDYtazwXi4UUmSWaOUew patches/linux-2.6.11/linux-2.6.11.10.patch
 418abc69J3F638vPO9MYoDGeYilxoQ patches/linux-2.6.11/nettel.patch
+429ae875I9ZrqrRDjGD34IC2kzDREw patches/linux-2.6.11/rcu-nohz.patch
 3f776bd1Hy9rn69ntXBhPReUFw9IEA tools/Makefile
 40e1b09db5mN69Ijj0X_Eol-S7dXiw tools/Rules.mk
 4124b307nRyK3dhn1hAsvrY76NuV3g tools/check/Makefile
diff --git a/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/process.c b/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/process.c
index 9a0e5d1322..0ef40b7035 100644
--- a/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/process.c
+++ b/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/process.c
@@ -91,25 +91,18 @@ void enable_hlt(void)
 EXPORT_SYMBOL(enable_hlt);
 
 /* XXX XEN doesn't use default_idle(), poll_idle(). Use xen_idle() instead. */
-extern int set_timeout_timer(void);
+extern void stop_hz_timer(void);
+extern void start_hz_timer(void);
 void xen_idle(void)
 {
-	int cpu;
-
 	local_irq_disable();
 
-	cpu = smp_processor_id();
-	if (rcu_pending(cpu))
-		rcu_check_callbacks(cpu, 0);
-
 	if (need_resched()) {
 		local_irq_enable();
-	} else if (set_timeout_timer() == 0) {
-		/* NB. Blocking reenable events in a race-free manner. */
-		HYPERVISOR_block();
 	} else {
-		local_irq_enable();
-		HYPERVISOR_yield();
+		stop_hz_timer();
+		HYPERVISOR_block(); /* implicit local_irq_enable() */
+		start_hz_timer();
 	}
 }
 
diff --git a/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/time.c b/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/time.c
index e013191f7b..acf5dd3130 100644
--- a/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/time.c
+++ b/linux-2.6.11-xen-sparse/arch/xen/i386/kernel/time.c
@@ -665,36 +665,31 @@ static inline u64 __jiffies_to_st(unsigned long j)
 }
 
 /*
- * This function works out when the the next timer function has to be
- * executed (by looking at the timer list) and sets the Xen one-shot
- * domain timer to the appropriate value. This is typically called in
- * cpu_idle() before the domain blocks.
- * 
- * The function returns a non-0 value on error conditions.
- * 
- * It must be called with interrupts disabled.
+ * stop_hz_timer / start_hz_timer - enter/exit 'tickless mode' on an idle cpu
+ * These functions are based on implementations from arch/s390/kernel/time.c
  */
-int set_timeout_timer(void)
+void stop_hz_timer(void)
 {
-	u64 alarm = 0;
-	int ret = 0;
+	unsigned int cpu = smp_processor_id();
 	unsigned long j;
 
-	/*
-	 * This is safe against long blocking (since calculations are
-	 * not based on TSC deltas). It is also safe against warped
-	 * system time since suspend-resume is cooperative and we
-	 * would first get locked out. It is safe against normal
-	 * updates of jiffies since interrupts are off.
-	 */
-	j = next_timer_interrupt();
-	alarm = __jiffies_to_st(j);
+	/* s390 does this /before/ checking rcu_pending(). We copy them. */
+	cpu_set(cpu, nohz_cpu_mask);
 
-	/* Failure is pretty bad, but we'd best soldier on. */
-	if ( HYPERVISOR_set_timer_op(alarm) != 0 )
-		ret = -1;
+	/* Leave ourselves in 'tick mode' if rcu or softirq pending. */
+	if (rcu_pending(cpu) || local_softirq_pending()) {
+		cpu_clear(cpu, nohz_cpu_mask);
+		j = jiffies + 1;
+	} else {
+		j = next_timer_interrupt();
+	}
 
-	return ret;
+	BUG_ON(HYPERVISOR_set_timer_op(__jiffies_to_st(j)) != 0);
+}
+
+void start_hz_timer(void)
+{
+	cpu_clear(smp_processor_id(), nohz_cpu_mask);
 }
 
 void time_suspend(void)
diff --git a/patches/linux-2.6.11/rcu-nohz.patch b/patches/linux-2.6.11/rcu-nohz.patch
new file mode 100644
index 0000000000..d7bafb3a62
--- /dev/null
+++ b/patches/linux-2.6.11/rcu-nohz.patch
@@ -0,0 +1,16 @@
+diff -ur linux-2.6.11/kernel/rcupdate.c linux-2.6.11-rcu/kernel/rcupdate.c
+--- linux-2.6.11/kernel/rcupdate.c	2005-05-30 10:51:41 +01:00
++++ linux-2.6.11-rcu/kernel/rcupdate.c	2005-05-30 10:53:53 +01:00
+@@ -202,8 +202,11 @@
+  */
+ static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp, struct rcu_state *rsp)
+ {
++	cpumask_t mask;
++
+ 	cpu_clear(cpu, rsp->cpumask);
+-	if (cpus_empty(rsp->cpumask)) {
++	cpus_andnot(mask, rsp->cpumask, nohz_cpu_mask);
++	if (cpus_empty(mask)) {
+ 		/* batch completed ! */
+ 		rcp->completed = rcp->cur;
+ 		rcu_start_batch(rcp, rsp, 0);