Re: 5.8-rc*: kernel BUG at kernel/signal.c:1917

[peterz@infradead.org]

On Mon, Jul 20, 2020 at 01:26:23PM +0200, peterz@infradead.org wrote:
>  kernel/sched/core.c | 34 ++++++++++++++++++++++++++++------
>  1 file changed, 28 insertions(+), 6 deletions(-)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index e15543cb84812..b5973d7fa521c 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -4100,9 +4100,9 @@ pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
>   */
>  static void __sched notrace __schedule(bool preempt)
>  {
> +	unsigned long prev_state, tmp_state;
>  	struct task_struct *prev, *next;
>  	unsigned long *switch_count;
> -	unsigned long prev_state;
>  	struct rq_flags rf;
>  	struct rq *rq;
>  	int cpu;
> @@ -4140,16 +4140,38 @@ static void __sched notrace __schedule(bool preempt)
>  	rq_lock(rq, &rf);
>  	smp_mb__after_spinlock();
>  
> +	/*
> +	 * We must re-load prev->state in case ttwu_remote() changed it
> +	 * before we acquired rq->lock.
> +	 */
> +	tmp_state = prev->state;
> +	if (unlikely(prev_state != tmp_state)) {
> +		/*
> +		 * ptrace_{,un}freeze_traced() think it is cool to change
> +		 * ->state around behind our backs between TASK_TRACED and
> +		 *  __TASK_TRACED.
> +		 *
> +		 * This is safe because this, as well as any __TASK_TRACED
> +		 * wakeups are under siglock.
> +		 *
> +		 * For any other case, a changed prev_state must be to
> +		 * TASK_RUNNING, such that when it blocks, the load has
> +		 * happened before the smp_mb().
> +		 *
> +		 * Also see the comment with deactivate_task().
> +		 */
> +		SCHED_WARN_ON(tmp_state && (prev_state & __TASK_TRACED &&
> +					   !(tmp_state & __TASK_TRACED)));
> +
> +		prev_state = tmp_state;

While trying to write a changelog for this thing, I can't convince
myself we don't need:

		smp_mb();

here. Consider:

CPU0				CPU1				CPU2

				schedule()
				  prev_state = prev->state;
				  spin_lock(rq->lock);
				  smp_mb__after_spin_lock();
ptrace_freeze_traced()
  spin_lock(siglock)
  task->state = __TASK_TRACED;
  spin_unlock(siglock);
				  tmp_state = prev->state;
				  if (prev_state != tmp_state)
				    prev_state = tmp_state;
				  /* NO SMP_MB */
				  if (prev_state)
				    deactivate_task()
				      prev->on_rq = 0;
								spin_lock(siglock);
								ttwu()
								  if (rq->on_rq && ...)
								    goto unlock;
								  smp_acquire__after_ctrl_dep();
								  p->state = TASK_WAKING;

Looses the ordering we previously relied upon. That is, CPU1's
prev->state load and prev->on_rq store can get reordered vs CPU2.

OTOH, we have a control dependency on CPU1 as well, that should provide
LOAD->STORE ordering, after all we only do the ->on_rq=0 store, IFF we
see prev_state.

So that is:

	if (p->state)			if (!p->on_rq)
		p->on_rq = 0;			p->state = TASK_WAKING

which matches a CTRL-DEP to a CTRL-DEP ...

But this then means we can simplify dbfb089d360 as well, but now my head
hurts.

> +	}
> +
>  	/* Promote REQ to ACT */
>  	rq->clock_update_flags <<= 1;
>  	update_rq_clock(rq);
>  
>  	switch_count = &prev->nivcsw;
> -	/*
> -	 * We must re-load prev->state in case ttwu_remote() changed it
> -	 * before we acquired rq->lock.
> -	 */
> -	if (!preempt && prev_state && prev_state == prev->state) {
> +	if (!preempt && prev_state) {
>  		if (signal_pending_state(prev_state, prev)) {
>  			prev->state = TASK_RUNNING;
>  		} else {