Messages in this thread | | | From | Namhyung Kim <> | Date | Tue, 22 Mar 2022 15:18:45 -0700 | Subject | Re: [PATCH v2 2/6] perf/core: Introduce percpu perf_cgroup |
| |
On Tue, Mar 22, 2022 at 5:10 AM Chengming Zhou <zhouchengming@bytedance.com> wrote: > > Although we don't have incosistency problem any more, we can > have other problem like: > > CPU1 CPU2 > (in context_switch) (attach running task) > prev->cgroups = cgrp2 > perf_cgroup_sched_switch(prev, next) > cgrp2 == cgrp2 is True > > If perf_cgroup of prev task changes from cgrp1 to cgrp2, > perf_cgroup_sched_switch() will skip perf_cgroup_switch(), > so the CPU would still schedule the cgrp1 events, but we should > schedule the cgrp2 events.
Ah ok, now I see the problem in changing prev->cgroup too.
> > The reason of this problem is that we shouldn't use the changeable > prev->cgroups to decide whether skip perf_cgroup_switch(). > > This patch introduces a percpu perf_cgroup to cache the perf_cgroup > that scheduled in cpuctxes, which later used to compare with the > perf_cgroup of next task to decide whether skip perf_cgroup_switch(). > > Since the perf_cgroup_switch() can be called after the context switch, > the cgroup events might be scheduled already. So we put the comparison > of perf_cgroups in perf_cgroup_switch(), and delete the unused function > perf_cgroup_sched_switch(). > > We must clear the percpu perf_cgroup cache when the last cgroup event > disabled. > > Fixes: a8d757ef076f ("perf events: Fix slow and broken cgroup context switch code") > Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com> > --- > kernel/events/core.c | 63 ++++++++++++++++---------------------------- > 1 file changed, 22 insertions(+), 41 deletions(-) > > diff --git a/kernel/events/core.c b/kernel/events/core.c > index 8b5cf2aedfe6..848a3bfa9513 100644 > --- a/kernel/events/core.c > +++ b/kernel/events/core.c > @@ -826,6 +826,7 @@ perf_cgroup_set_timestamp(struct task_struct *task, > } > } > > +static DEFINE_PER_CPU(struct perf_cgroup *, cpu_perf_cgroup); > static DEFINE_PER_CPU(struct list_head, cgrp_cpuctx_list); > > /* > @@ -833,6 +834,7 @@ static DEFINE_PER_CPU(struct list_head, cgrp_cpuctx_list); > */ > static void perf_cgroup_switch(struct task_struct *task) > { > + struct perf_cgroup *cgrp; > struct perf_cpu_context *cpuctx, *tmp; > struct list_head *list; > unsigned long flags; > @@ -843,11 +845,21 @@ static void perf_cgroup_switch(struct task_struct *task) > */ > local_irq_save(flags); > > + cgrp = perf_cgroup_from_task(task, NULL); > + if (cgrp == __this_cpu_read(cpu_perf_cgroup)) > + goto out; > + > + __this_cpu_write(cpu_perf_cgroup, cgrp); > + > list = this_cpu_ptr(&cgrp_cpuctx_list); > list_for_each_entry_safe(cpuctx, tmp, list, cgrp_cpuctx_entry) { > WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0); > > perf_ctx_lock(cpuctx, cpuctx->task_ctx); > + > + if (cpuctx->cgrp == cgrp)
Missing perf_ctx_unlock().
Thanks, Namhyung
> + continue; > + > perf_pmu_disable(cpuctx->ctx.pmu); > > cpu_ctx_sched_out(cpuctx, EVENT_ALL); > @@ -855,14 +867,11 @@ static void perf_cgroup_switch(struct task_struct *task) > * must not be done before ctxswout due > * to event_filter_match() in event_sched_out() > */ > - cpuctx->cgrp = perf_cgroup_from_task(task, > - &cpuctx->ctx); > + cpuctx->cgrp = cgrp; > /* > * set cgrp before ctxsw in to allow > * event_filter_match() to not have to pass > * task around > - * we pass the cpuctx->ctx to perf_cgroup_from_task() > - * because cgroup events are only per-cpu > */ > cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); > > @@ -870,35 +879,10 @@ static void perf_cgroup_switch(struct task_struct *task) > perf_ctx_unlock(cpuctx, cpuctx->task_ctx); > } > > +out: > local_irq_restore(flags); > }
| |