Re: [PATCH v7 2/6] sched_ext: Implement scx_bpf_now()

[Peter Zijlstra <peterz@infradead.org>]

> +__bpf_kfunc u64 scx_bpf_now(void)
> +{
> +	struct rq *rq;
> +	u64 clock;
> +
> +	preempt_disable();
> +
> +	rq = this_rq();
> +	if (READ_ONCE(rq->scx.flags) & SCX_RQ_CLK_VALID) {
> +		/*
> +		 * If the rq clock is valid, use the cached rq clock.
> +		 *
> +		 * Note that scx_bpf_now() is re-entrant between a process
> +		 * context and an interrupt context (e.g., timer interrupt).
> +		 * However, we don't need to consider the race between them
> +		 * because such race is not observable from a caller.
> +		 */
> +		clock = READ_ONCE(rq->scx.clock);
> +	} else {
> +		/*
> +		 * Otherwise, return a fresh rq clock.
> +		 *
> +		 * The rq clock is updated outside of the rq lock.
> +		 * In this case, keep the updated rq clock invalid so the next
> +		 * kfunc call outside the rq lock gets a fresh rq clock.
> +		 */
> +		clock = sched_clock_cpu(cpu_of(rq));
> +	}
> +
> +	preempt_enable();
> +
> +	return clock;
> +}

> +static inline void scx_rq_clock_update(struct rq *rq, u64 clock)
> +{
> +	if (!scx_enabled())
> +		return;
> +	WRITE_ONCE(rq->scx.clock, clock);
> +	WRITE_ONCE(rq->scx.flags, rq->scx.flags | SCX_RQ_CLK_VALID);
> +}


AFAICT it is possible to be used like:


  CPU0					CPU1

  lock(rq1->lock);
  ...
  scx_rq_clock_update(...);		scx_bpf_now();
  ...
  unlock(rq->lock);


Which then enables the following ordering problem:

  CPU0					CPU1

  WRITE_ONCE(rq->scx.clock, clock);	if (rq->scx.flags & VALID)
  WRITE_ONCE(rq->scx.flags, VALID);	  return rq->scx.clock;

Where it then becomes possible to observe VALID before clock is written.

That is, I rather think you need:

> +static inline void scx_rq_clock_update(struct rq *rq, u64 clock)
> +{
> +	if (!scx_enabled())
> +		return;
> +	WRITE_ONCE(rq->scx.clock, clock);
> +	smp_store_release(&rq->scx.flags, rq->scx.flags | SCX_RQ_CLK_VALID);
> +}

and:

	if (smp_load_acquire(&rq->scx.flags) & SCX_RQ_CLK_VALID) {
> +		/*
> +		 * If the rq clock is valid, use the cached rq clock.
> +		 *
> +		 * Note that scx_bpf_now() is re-entrant between a process
> +		 * context and an interrupt context (e.g., timer interrupt).
> +		 * However, we don't need to consider the race between them
> +		 * because such race is not observable from a caller.
> +		 */
> +		clock = READ_ONCE(rq->scx.clock);

Such that if you ovbserve VALID, you must then also observe the clock.