Commit d306ebc2 authored Feb 10, 2010 by Pallipadi, Venkatesh Committed by Len Brown Feb 22, 2010

ACPI: Be in TS_POLLING state during mwait based C-state entry

ACPI deep C-state entry had a long standing bug/missing feature, wherein we were sending
resched IPIs when an idle CPU is in mwait based deep C-state. Only mwait based C1 was using
the write to the monitored address to wake up mwait'ing CPU.

This patch changes the code to retain TS_POLLING bit if we are entering an mwait based
deep C-state.

The patch has been verified to reduce the number of resched IPIs in general and also
improves the performance/power on workloads with low system utilization (i.e., when mwait based
deep C-states are being used).

Fixes "netperf ~50% regression with 2.6.33-rc1, bisect to 1b9508f6"
http://marc.info/?l=linux-kernel&m=126441481427331&w=4

Reported-by: Lin Ming <ming.m.lin@intel.com>
Tested-by: Alex Shi <alex.shi@intel.com>
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Len Brown <len.brown@intel.com>

parent 627fa177

drivers/acpi/processor_idle.c

+16 −12

Original line number	Diff line number	Diff line
		@@ -880,12 +880,14 @@ static int acpi_idle_enter_simple(struct cpuidle_device *dev,
		return(acpi_idle_enter_c1(dev, state));

		local_irq_disable();
		if (cx->entry_method != ACPI_CSTATE_FFH) {
		current_thread_info()->status &= ~TS_POLLING;
		/*
		* TS_POLLING-cleared state must be visible before we test
		* NEED_RESCHED:
		*/
		smp_mb();
		}

		if (unlikely(need_resched())) {
		current_thread_info()->status \|= TS_POLLING;
		@@ -965,12 +967,14 @@ static int acpi_idle_enter_bm(struct cpuidle_device *dev,
		}

		local_irq_disable();
		if (cx->entry_method != ACPI_CSTATE_FFH) {
		current_thread_info()->status &= ~TS_POLLING;
		/*
		* TS_POLLING-cleared state must be visible before we test
		* NEED_RESCHED:
		*/
		smp_mb();
		}

		if (unlikely(need_resched())) {
		current_thread_info()->status \|= TS_POLLING;

Admin message