hrtimer: timeout too long when using HRTIMER_CB_SOFTIRQ

When using hrtimer with timer->cb_mode == HRTIMER_CB_SOFTIRQ
in some cases the clockevent is not programmed.
This happens, if:
 - a timer is rearmed while it's state is HRTIMER_STATE_CALLBACK
 - hrtimer_reprogram() returns -ETIME, when it is called after
   CALLBACK is finished. This occurs if the new timer->expires
   is in the past when CALLBACK is done.
In this case, the timer needs to be removed from the tree and put
onto the pending list again.

The patch is against 2.6.22.5, but AFAICS, it is relevant
for 2.6.25 also (in run_hrtimer_pending()).

Signed-off-by: Bodo Stroesser <bstroesser@fujitsu-siemens.com>
Cc: stable@kernel.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

1 files changed, 13 insertions, 2 deletions

diff --git a/kernel/hrtimer.c b/kernel/hrtimer.c
index f78777abe769..e379ef0e9c20 100644
--- a/kernel/hrtimer.c
+++ b/kernel/hrtimer.c
@@ -1080,8 +1080,19 @@ static void run_hrtimer_pending(struct hrtimer_cpu_base *cpu_base)
                          * If the timer was rearmed on another CPU, reprogram
                          * the event device.
                          */
-                        if (timer->base->first == &timer->node)
-                                hrtimer_reprogram(timer, timer->base);
+                        struct hrtimer_clock_base *base = timer->base;
+
+                        if (base->first == &timer->node &&
+                            hrtimer_reprogram(timer, base)) {
+                                /*
+                                 * Timer is expired. Thus move it from tree to
+                                 * pending list again.
+                                 */
+                                __remove_hrtimer(timer, base,
+                                                 HRTIMER_STATE_PENDING, 0);
+                                list_add_tail(&timer->cb_entry,
+                                              &base->cpu_base->cb_pending);
+                        }
                 }
         }
         spin_unlock_irq(&cpu_base->lock);