sched/core: Add preempt checks in preempt_schedule() code
authorSteven Rostedt <rostedt@goodmis.org>
Mon, 21 Mar 2016 15:23:39 +0000 (11:23 -0400)
committerIngo Molnar <mingo@kernel.org>
Thu, 31 Mar 2016 08:49:45 +0000 (10:49 +0200)
While testing the tracer preemptoff, I hit this strange trace:

   <...>-259     0...1    0us : schedule <-worker_thread
   <...>-259     0d..1    0us : rcu_note_context_switch <-__schedule
   <...>-259     0d..1    0us : rcu_sched_qs <-rcu_note_context_switch
   <...>-259     0d..1    0us : rcu_preempt_qs <-rcu_note_context_switch
   <...>-259     0d..1    0us : _raw_spin_lock <-__schedule
   <...>-259     0d..1    0us : preempt_count_add <-_raw_spin_lock
   <...>-259     0d..2    0us : do_raw_spin_lock <-_raw_spin_lock
   <...>-259     0d..2    1us : deactivate_task <-__schedule
   <...>-259     0d..2    1us : update_rq_clock.part.84 <-deactivate_task
   <...>-259     0d..2    1us : dequeue_task_fair <-deactivate_task
   <...>-259     0d..2    1us : dequeue_entity <-dequeue_task_fair
   <...>-259     0d..2    1us : update_curr <-dequeue_entity
   <...>-259     0d..2    1us : update_min_vruntime <-update_curr
   <...>-259     0d..2    1us : cpuacct_charge <-update_curr
   <...>-259     0d..2    1us : __rcu_read_lock <-cpuacct_charge
   <...>-259     0d..2    1us : __rcu_read_unlock <-cpuacct_charge
   <...>-259     0d..2    1us : clear_buddies <-dequeue_entity
   <...>-259     0d..2    1us : account_entity_dequeue <-dequeue_entity
   <...>-259     0d..2    2us : update_min_vruntime <-dequeue_entity
   <...>-259     0d..2    2us : update_cfs_shares <-dequeue_entity
   <...>-259     0d..2    2us : hrtick_update <-dequeue_task_fair
   <...>-259     0d..2    2us : wq_worker_sleeping <-__schedule
   <...>-259     0d..2    2us : kthread_data <-wq_worker_sleeping
   <...>-259     0d..2    2us : pick_next_task_fair <-__schedule
   <...>-259     0d..2    2us : check_cfs_rq_runtime <-pick_next_task_fair
   <...>-259     0d..2    2us : pick_next_entity <-pick_next_task_fair
   <...>-259     0d..2    2us : clear_buddies <-pick_next_entity
   <...>-259     0d..2    2us : pick_next_entity <-pick_next_task_fair
   <...>-259     0d..2    2us : clear_buddies <-pick_next_entity
   <...>-259     0d..2    2us : set_next_entity <-pick_next_task_fair
   <...>-259     0d..2    3us : put_prev_entity <-pick_next_task_fair
   <...>-259     0d..2    3us : check_cfs_rq_runtime <-put_prev_entity
   <...>-259     0d..2    3us : set_next_entity <-pick_next_task_fair
gnome-sh-1031    0d..2    3us : finish_task_switch <-__schedule
gnome-sh-1031    0d..2    3us : _raw_spin_unlock_irq <-finish_task_switch
gnome-sh-1031    0d..2    3us : do_raw_spin_unlock <-_raw_spin_unlock_irq
gnome-sh-1031    0...2    3us!: preempt_count_sub <-_raw_spin_unlock_irq
gnome-sh-1031    0...1  582us : do_raw_spin_lock <-_raw_spin_lock
gnome-sh-1031    0...1  583us : _raw_spin_unlock <-drm_gem_object_lookup
gnome-sh-1031    0...1  583us : do_raw_spin_unlock <-_raw_spin_unlock
gnome-sh-1031    0...1  583us : preempt_count_sub <-_raw_spin_unlock
gnome-sh-1031    0...1  584us : _raw_spin_unlock <-drm_gem_object_lookup
gnome-sh-1031    0...1  584us+: trace_preempt_on <-drm_gem_object_lookup
gnome-sh-1031    0...1  603us : <stack trace>
 => preempt_count_sub
 => _raw_spin_unlock
 => drm_gem_object_lookup
 => i915_gem_madvise_ioctl
 => drm_ioctl
 => do_vfs_ioctl
 => SyS_ioctl
 => entry_SYSCALL_64_fastpath

As I'm tracing preemption disabled, it seemed incorrect that the trace
would go across a schedule and report not being in the scheduler.
Looking into this I discovered the problem.

schedule() calls preempt_disable() but the preempt_schedule() calls
preempt_enable_notrace(). What happened above was that the gnome-shell
task was preempted on another CPU, migrated over to the idle cpu. The
tracer stared with idle calling schedule(), which called
preempt_disable(), but then gnome-shell finished, and it enabled
preemption with preempt_enable_notrace() that does stop the trace, even
though preemption was enabled.

The purpose of the preempt_disable_notrace() in the preempt_schedule()
is to prevent function tracing from going into an infinite loop.
Because function tracing can trace the preempt_enable/disable() calls
that are traced. The problem with function tracing is:

  NEED_RESCHED set
  preempt_schedule()
    preempt_disable()
      preempt_count_inc()
        function trace (before incrementing preempt count)
          preempt_disable_notrace()
          preempt_enable_notrace()
            sees NEED_RESCHED set
               preempt_schedule() (repeat)

Now by breaking out the preempt off/on tracing into their own code:
preempt_disable_check() and preempt_enable_check(), we can add these to
the preempt_schedule() code. As preemption would then be disabled, even
if they were to be traced by the function tracer, the disabled
preemption would prevent the recursion.

Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20160321112339.6dc78ad6@gandalf.local.home
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/sched/core.c

index 8b489fc..b5cf01d 100644 (file)
@@ -2958,6 +2958,20 @@ u64 scheduler_tick_max_deferment(void)
 
 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
                                defined(CONFIG_PREEMPT_TRACER))
+/*
+ * If the value passed in is equal to the current preempt count
+ * then we just disabled preemption. Start timing the latency.
+ */
+static inline void preempt_latency_start(int val)
+{
+       if (preempt_count() == val) {
+               unsigned long ip = get_lock_parent_ip();
+#ifdef CONFIG_DEBUG_PREEMPT
+               current->preempt_disable_ip = ip;
+#endif
+               trace_preempt_off(CALLER_ADDR0, ip);
+       }
+}
 
 void preempt_count_add(int val)
 {
@@ -2976,17 +2990,21 @@ void preempt_count_add(int val)
        DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
                                PREEMPT_MASK - 10);
 #endif
-       if (preempt_count() == val) {
-               unsigned long ip = get_lock_parent_ip();
-#ifdef CONFIG_DEBUG_PREEMPT
-               current->preempt_disable_ip = ip;
-#endif
-               trace_preempt_off(CALLER_ADDR0, ip);
-       }
+       preempt_latency_start(val);
 }
 EXPORT_SYMBOL(preempt_count_add);
 NOKPROBE_SYMBOL(preempt_count_add);
 
+/*
+ * If the value passed in equals to the current preempt count
+ * then we just enabled preemption. Stop timing the latency.
+ */
+static inline void preempt_latency_stop(int val)
+{
+       if (preempt_count() == val)
+               trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
+}
+
 void preempt_count_sub(int val)
 {
 #ifdef CONFIG_DEBUG_PREEMPT
@@ -3003,13 +3021,15 @@ void preempt_count_sub(int val)
                return;
 #endif
 
-       if (preempt_count() == val)
-               trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
+       preempt_latency_stop(val);
        __preempt_count_sub(val);
 }
 EXPORT_SYMBOL(preempt_count_sub);
 NOKPROBE_SYMBOL(preempt_count_sub);
 
+#else
+static inline void preempt_latency_start(int val) { }
+static inline void preempt_latency_stop(int val) { }
 #endif
 
 /*
@@ -3284,8 +3304,23 @@ void __sched schedule_preempt_disabled(void)
 static void __sched notrace preempt_schedule_common(void)
 {
        do {
+               /*
+                * Because the function tracer can trace preempt_count_sub()
+                * and it also uses preempt_enable/disable_notrace(), if
+                * NEED_RESCHED is set, the preempt_enable_notrace() called
+                * by the function tracer will call this function again and
+                * cause infinite recursion.
+                *
+                * Preemption must be disabled here before the function
+                * tracer can trace. Break up preempt_disable() into two
+                * calls. One to disable preemption without fear of being
+                * traced. The other to still record the preemption latency,
+                * which can also be traced by the function tracer.
+                */
                preempt_disable_notrace();
+               preempt_latency_start(1);
                __schedule(true);
+               preempt_latency_stop(1);
                preempt_enable_no_resched_notrace();
 
                /*
@@ -3337,7 +3372,21 @@ asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
                return;
 
        do {
+               /*
+                * Because the function tracer can trace preempt_count_sub()
+                * and it also uses preempt_enable/disable_notrace(), if
+                * NEED_RESCHED is set, the preempt_enable_notrace() called
+                * by the function tracer will call this function again and
+                * cause infinite recursion.
+                *
+                * Preemption must be disabled here before the function
+                * tracer can trace. Break up preempt_disable() into two
+                * calls. One to disable preemption without fear of being
+                * traced. The other to still record the preemption latency,
+                * which can also be traced by the function tracer.
+                */
                preempt_disable_notrace();
+               preempt_latency_start(1);
                /*
                 * Needs preempt disabled in case user_exit() is traced
                 * and the tracer calls preempt_enable_notrace() causing
@@ -3347,6 +3396,7 @@ asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
                __schedule(true);
                exception_exit(prev_ctx);
 
+               preempt_latency_stop(1);
                preempt_enable_no_resched_notrace();
        } while (need_resched());
 }