Re: [PATCH v5 23/29] sched/deadline: Move a task's bandwidth when it must leave its CPU

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

On 2026-11-08 14:41, sashiko-bot@kernel.org wrote:
 > stale dl_bw_cpu on same-rd return in dl_task_move_bw()

Thanks. The record can name a CPU the task no longer runs on. That
matters only for a task that departs before the rebuild. The rebuild
recounts a task in the class on the CPU it is on. A departed task whose
recorded CPU is inactive loses its bandwidth until its 0-lag time. The
comment above dl_departed_list says so. The release from def_root_domain
happens only between the detach and dl_clear_root_domains(), in the same
rebuild, which clears it. Still, the next version records the CPU a task
departs from, which covers this case.

 > missed bw move in migrate_task_rq_dl() on task wakeup

Thanks, both shapes are real. The test should not ask whether the task
leaves an allowed CPU. It should ask whether the new CPU's root domain
counts the bandwidth. The next version compares the root domain of the
recorded CPU with the new CPU's, in migrate_task_rq_dl(). It does the
same check in task_woken_dl(), for a wakeup with no migration. The same
state also comes up with admission control on. A task is moved asleep
into a cpuset of an isolated CPU. Then cpuset.cpus gets back the CPU the
task sleeps on. Two new cases, hotplug_fallback_in_mask and
mask_back_asleep, fail on v5 and pass with the fix.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com