Re: [PATCH v2] sched/fair: Fix non-empty throttled_limbo_list warning in tg_throttle_down()

[Aaron Lu]

On Tue, Oct 28, 2025 at 08:33:15AM +0530, K Prateek Nayak wrote:
> Hello Aaron,
> 
> On 10/27/2025 5:32 PM, Aaron Lu wrote:
> > On Mon, Oct 27, 2025 at 05:05:34PM +0800, Hao Jia wrote:
> >> @@ -6403,7 +6407,7 @@ static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
> >>   * expired/exceeded, otherwise it may be allowed to steal additional ticks of
> >>   * runtime as update_curr() throttling can not trigger until it's on-rq.
> >>   */
> >> -static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
> >> +static void check_enqueue_throttle(struct cfs_rq *cfs_rq, int flags)
> >>  {
> >>  	if (!cfs_bandwidth_used())
> >>  		return;
> >> @@ -6418,6 +6422,13 @@ static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
> >>  
> >>  	/* update runtime allocation */
> >>  	account_cfs_rq_runtime(cfs_rq, 0);
> >> +	/*
> >> +	 * Do not attempt to throttle on the cfs_rq unthrottle path.
> >> +	 * and it must be placed after account_cfs_rq_runtime() to
> >> +	 * prevent a possible missed start of the bandwidth timer.
> > 
> > Hi Prateek and Hao,
> > 
> > Does it matter to start the bw timer? If no cfs_rq gets throttled, the
> > timer doesn't look that useful.
> 
> Ack! But if we've reached here with ENQUEUE_THROTTLE set, we are in fact
> trying to enqueue a task during unthrottling and we only start the timer
> if we don't have any bandwidth since __account_cfs_rq_runtime() would bail
> out if "cfs_rq->runtime_remaining > 0"
> > Also, account_cfs_rq_runtime() calls
> > assign_cfs_rq_runtime() and if assign failed, it will do resched_curr()
> > but since we do not throttle cfs_rq here, that resched would be useless.
> 
> resched only happens if "cfs_rq->curr" is set andcheck_enqueue_throttle() already bails out if "cfs_rq->curr" is set so
> the resched is avoided.

Right, sorry I missed that.