On Tue, Nov 17, 2020 at 07:32:16PM +0000, Valentin Schneider wrote: > > On 17/11/20 16:13, Peter Zijlstra wrote: > > On Tue, Nov 17, 2020 at 03:37:24PM +0000, Valentin Schneider wrote: > > > >> >> + /* > >> >> + * This field must not be in the scheduler word above due to > >> >> wakelist > >> >> + * queueing no longer being serialized by p->on_cpu. However: > >> >> + * > >> >> + * p->XXX = X; ttwu() > >> >> + * schedule() if (p->on_rq && ..) // false > >> >> + * smp_mb__after_spinlock(); if > >> >> (smp_load_acquire(&p->on_cpu) && //true > >> >> + * deactivate_task() ttwu_queue_wakelist()) > >> >> + * p->on_rq = 0; p->sched_remote_wakeup > >> >> = Y; > >> >> + * > >> >> + * guarantees all stores of 'current' are visible before > >> >> + * ->sched_remote_wakeup gets used, so it can be in this word. > >> >> + */ > >> > > >> > Isn't the control dep between that ttwu() p->on_rq read and > >> > p->sched_remote_wakeup write "sufficient"? > >> > >> smp_acquire__after_ctrl_dep() that is, since we need > >> ->on_rq load => 'current' bits load + store > > > > I don't think we need that extra barrier; after all, there will be a > > complete schedule() between waking the task and it actually becoming > > current. > > Apologies for the messy train of thought; what I was trying to say is that > we have already the following, which AIUI is sufficient: > > * p->XXX = X; ttwu() > * schedule() if (p->on_rq && ..) // false > * smp_mb__after_spinlock(); smp_acquire__after_ctrl_dep(); > * deactivate_task() ttwu_queue_wakelist() > * p->on_rq = 0; p->sched_remote_wakeup = Y; >
Ah, you meant the existing smp_acquire__after_ctrl_dep(). Yeah, that's not required here either ;-) The reason I had the ->on_cpu thing in there is because it shows we violate the regular ->on_cpu handoff rules, not for the acquire. The only ordering that matters on the RHS of that thing is the ->on_rq load to p->sched_remote_wakeup store ctrl dep. That, combined with the LHS, guarantees there is a strict order on the stores. Makes sense?
