cpuidle: fix a suspicious RCU usage in menu governor
authorLi Zhong <zhong@linux.vnet.ibm.com>
Thu, 22 Nov 2012 23:05:03 +0000 (00:05 +0100)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Thu, 22 Nov 2012 23:05:03 +0000 (00:05 +0100)
I saw this suspicious RCU usage on the next tree of 11/15

[   67.123404] ===============================
[   67.123413] [ INFO: suspicious RCU usage. ]
[   67.123423] 3.7.0-rc5-next-20121115-dirty #1 Not tainted
[   67.123434] -------------------------------
[   67.123444] include/trace/events/timer.h:186 suspicious rcu_dereference_check() usage!
[   67.123458]
[   67.123458] other info that might help us debug this:
[   67.123458]
[   67.123474]
[   67.123474] RCU used illegally from idle CPU!
[   67.123474] rcu_scheduler_active = 1, debug_locks = 0
[   67.123493] RCU used illegally from extended quiescent state!
[   67.123507] 1 lock held by swapper/1/0:
[   67.123516]  #0:  (&cpu_base->lock){-.-...}, at: [<c0000000000979b0>] .__hrtimer_start_range_ns+0x28c/0x524
[   67.123555]
[   67.123555] stack backtrace:
[   67.123566] Call Trace:
[   67.123576] [c0000001e2ccb920] [c00000000001275c] .show_stack+0x78/0x184 (unreliable)
[   67.123599] [c0000001e2ccb9d0] [c0000000000c15a0] .lockdep_rcu_suspicious+0x120/0x148
[   67.123619] [c0000001e2ccba70] [c00000000009601c] .enqueue_hrtimer+0x1c0/0x1c8
[   67.123639] [c0000001e2ccbb00] [c000000000097aa0] .__hrtimer_start_range_ns+0x37c/0x524
[   67.123660] [c0000001e2ccbc20] [c0000000005c9698] .menu_select+0x508/0x5bc
[   67.123678] [c0000001e2ccbd20] [c0000000005c740c] .cpuidle_idle_call+0xa8/0x6e4
[   67.123699] [c0000001e2ccbdd0] [c0000000000459a0] .pSeries_idle+0x10/0x34
[   67.123717] [c0000001e2ccbe40] [c000000000014dc8] .cpu_idle+0x130/0x280
[   67.123738] [c0000001e2ccbee0] [c0000000006ffa8c] .start_secondary+0x378/0x384
[   67.123758] [c0000001e2ccbf90] [c00000000000936c] .start_secondary_prolog+0x10/0x14

hrtimer_start was added in 198fd638 and ae515197. The patch below tries
to use RCU_NONIDLE around it to avoid the above report.

Signed-off-by: Li Zhong <zhong@linux.vnet.ibm.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/cpuidle/governors/menu.c

index 2efee27..bd40b94 100644 (file)
@@ -407,8 +407,9 @@ static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
                perfect_us = perfect_cstate_ms * 1000;
 
                if (repeat && (4 * timer_us < data->expected_us)) {
-                       hrtimer_start(hrtmr, ns_to_ktime(1000 * timer_us),
-                               HRTIMER_MODE_REL_PINNED);
+                       RCU_NONIDLE(hrtimer_start(hrtmr,
+                               ns_to_ktime(1000 * timer_us),
+                               HRTIMER_MODE_REL_PINNED));
                        /* In repeat case, menu hrtimer is started */
                        per_cpu(hrtimer_status, cpu) = MENU_HRTIMER_REPEAT;
                } else if (perfect_us < data->expected_us) {
@@ -418,8 +419,9 @@ static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
                         * In that case, it makes sense to re-enter
                         * into a deeper C-state after some time.
                         */
-                       hrtimer_start(hrtmr, ns_to_ktime(1000 * timer_us),
-                               HRTIMER_MODE_REL_PINNED);
+                       RCU_NONIDLE(hrtimer_start(hrtmr,
+                               ns_to_ktime(1000 * timer_us),
+                               HRTIMER_MODE_REL_PINNED));
                        /* In general case, menu hrtimer is started */
                        per_cpu(hrtimer_status, cpu) = MENU_HRTIMER_GENERAL;
                }