Andi reported that when creating a lot of events, a lot of time is spend in IPIs and asked if it would be possible to elide some of that.
Now when, as for example the perf-tool always does, events are created disabled, then these events will not need to be scheduled when added to the context (they're still disable) and therefore the IPI is not required -- except for the very first event, that will need to set ctx->is_active. ( it might be possible to set ctx->is_active remotely for cpu_ctx, but we really need the IPI for task_ctx, so lets not make that distinction ) Also use __perf_effective_state() since group events depend on the state of the leader, if the leader is OFF, the whole group is OFF. So when sibling events are created enabled (XXX check tool) then we only need a single IPI to create and enable the whole group (+ that initial IPI to initialize the context). Reported-by: Andi Kleen <a...@firstfloor.org> Suggested-by: Andi Kleen <a...@firstfloor.org> Signed-off-by: Peter Zijlstra (Intel) <pet...@infradead.org> Cc: kan.li...@linux.intel.com Cc: jo...@redhat.com Cc: a...@kernel.org --- kernel/events/core.c | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) --- a/kernel/events/core.c +++ b/kernel/events/core.c @@ -2666,6 +2666,25 @@ perf_install_in_context(struct perf_even */ smp_store_release(&event->ctx, ctx); + /* + * perf_event_attr::disabled events will not run and can be initialized + * without IPI. Except when this is the first event for the context, in + * that case we need the magic of the IPI to set ctx->is_active. + * + * The IOC_ENABLE that is sure to follow the creation of a disabled + * event will issue the IPI and reprogram the hardware. + */ + if (__perf_effective_state(event) == PERF_EVENT_STATE_OFF && ctx->nr_events) { + raw_spin_lock_irq(&ctx->lock); + if (task && ctx->task == TASK_TOMBSTONE) { + raw_spin_unlock_irq(&ctx->lock); + return; + } + add_event_to_ctx(event, ctx); + raw_spin_unlock_irq(&ctx->lock); + return; + } + if (!task) { cpu_function_call(cpu, __perf_install_in_context, event); return;