1 From jejb@kernel.org Tue Apr 29 09:51:24 2008
2 From: Thomas Gleixner <tglx@linutronix.de>
3 Date: Tue, 29 Apr 2008 01:15:10 GMT
4 Subject: hrtimer: raise softirq unlocked to avoid circular lock dependency
5 To: jejb@kernel.org, stable@kernel.org
6 Message-ID: <200804290115.m3T1FAaX028758@hera.kernel.org>
8 From: Thomas Gleixner <tglx@linutronix.de>
10 commit 0c96c5979a522c3323c30a078a70120e29b5bdbc upstream
12 The scheduler hrtimer bits in 2.6.25 introduced a circular lock
13 dependency in a rare code path:
15 =======================================================
16 [ INFO: possible circular locking dependency detected ]
17 2.6.25-sched-devel.git-x86-latest.git #19
18 -------------------------------------------------------
19 X/2980 is trying to acquire lock:
20 (&rq->rq_lock_key#2){++..}, at: [<ffffffff80230146>] task_rq_lock+0x56/0xa0
22 but task is already holding lock:
23 (&cpu_base->lock){++..}, at: [<ffffffff80257ae1>] lock_hrtimer_base+0x31/0x60
25 which lock already depends on the new lock.
27 The scenario which leads to this is:
29 posix-timer signal is delivered
30 -> posix-timer is rearmed
31 timer is already expired in hrtimer_enqueue()
34 To prevent this we need to move the raise of the softirq out of the
35 base->lock protected code path.
37 Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
38 Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
39 Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
42 kernel/hrtimer.c | 19 +++++++++++++++++--
43 1 file changed, 17 insertions(+), 2 deletions(-)
45 --- a/kernel/hrtimer.c
46 +++ b/kernel/hrtimer.c
47 @@ -590,7 +590,6 @@ static inline int hrtimer_enqueue_reprog
48 list_add_tail(&timer->cb_entry,
49 &base->cpu_base->cb_pending);
50 timer->state = HRTIMER_STATE_PENDING;
51 - raise_softirq(HRTIMER_SOFTIRQ);
55 @@ -633,6 +632,11 @@ static int hrtimer_switch_to_hres(void)
59 +static inline void hrtimer_raise_softirq(void)
61 + raise_softirq(HRTIMER_SOFTIRQ);
66 static inline int hrtimer_hres_active(void) { return 0; }
67 @@ -651,6 +655,7 @@ static inline int hrtimer_reprogram(stru
71 +static inline void hrtimer_raise_softirq(void) { }
73 #endif /* CONFIG_HIGH_RES_TIMERS */
75 @@ -850,7 +855,7 @@ hrtimer_start(struct hrtimer *timer, kti
77 struct hrtimer_clock_base *base, *new_base;
82 base = lock_hrtimer_base(timer, &flags);
84 @@ -884,8 +889,18 @@ hrtimer_start(struct hrtimer *timer, kti
85 enqueue_hrtimer(timer, new_base,
86 new_base->cpu_base == &__get_cpu_var(hrtimer_bases));
89 + * The timer may be expired and moved to the cb_pending
90 + * list. We can not raise the softirq with base lock held due
91 + * to a possible deadlock with runqueue lock.
93 + raise = timer->state == HRTIMER_STATE_PENDING;
95 unlock_hrtimer_base(timer, &flags);
98 + hrtimer_raise_softirq();
102 EXPORT_SYMBOL_GPL(hrtimer_start);