Re: [PATCH 2/5] perf: generic intel uncore support

["Yan, Zheng" <zheng.z.yan@intel.com>]

On 04/16/2012 08:07 PM, Peter Zijlstra wrote:
> On Sun, 2012-04-01 at 11:11 +0800, Yan, Zheng wrote:
>> Any hints how to do this. I'm afraid it requires big changes to perf core.
> 
> 
> Sorry for taking so long..
> 
> 
> I think something like the (completely untested) below should suffice..
> 
> In your driver, have hotplug notifiers keep track of what cpu is the
> active cpu for your node, if it needs to change due to it going offline,
> pick a new one and use the below function to migrate the events.
> 
> The only missing piece is not doing the normal
> perf_event_exit_cpu_context() thing for these PMUs, except of course
> once there's no cpus left in your node.
> 
> Doing that might want an extra struct pmu method, which if not set
> defaults to perf_event_exit_cpu_context, and otherwise does your custom
> migrate/exit.
> 
> ---
>  kernel/events/core.c |   33 +++++++++++++++++++++++++++++++++
>  1 files changed, 33 insertions(+), 0 deletions(-)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index a6a9ec4..824becf 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -1641,6 +1641,8 @@ perf_install_in_context(struct perf_event_context *ctx,
>  	lockdep_assert_held(&ctx->mutex);
>  
>  	event->ctx = ctx;
> +	if (event->cpu != -1)
> +		event->cpu = cpu;
>  
>  	if (!task) {
>  		/*
> @@ -6375,6 +6377,8 @@ SYSCALL_DEFINE5(perf_event_open,
>  	mutex_lock(&ctx->mutex);
>  
>  	if (move_group) {
> +		synchronize_rcu();
> +
>  		perf_install_in_context(ctx, group_leader, cpu);
>  		get_ctx(ctx);
>  		list_for_each_entry(sibling, &group_leader->sibling_list,
> @@ -6477,6 +6481,35 @@ perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
>  }
>  EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
>  
> +void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
> +{
> +	struct perf_cpu_context *src_cpuctx = per_cpu(pmu->pmu_cpu_context, src_cpu);
> +	struct perf_cpu_context *dst_cpuctx = per_cpu(pmu->pmu_cpu_context, dst_cpu);
> +	struct perf_event_context *src_ctx = &src_cpuctx->ctx;
> +	struct perf_event_context *dst_ctx = &dst_cpuctx->ctx;
> +	struct perf_event *event, *tmp;
> +	LIST_HEAD(events);
> +
> +	mutex_lock(&src_ctx->mutex);
> +	list_for_each_entry_safe(event, tmp, &src_ctx->event_list, event_entry) {
> +		perf_remove_from_context(event);
> +		put_ctx(src_ctx);
> +		list_add(&event->event_entry, &events);
> +	}
> +	mutex_unlock(&src_ctx->mutex);
> +
> +	synchronize_rcu();
> +
> +	mutex_lock(&dst_ctx->mutex);
> +	list_for_each_entry_safe(event, tmp, &events, event_entry) {
> +		list_del(&event->event_entry);
> +		perf_install_in_context(dst_ctx, event, dst_cpu);
> +		get_ctx(dst_ctx);
> +	}
> +	mutex_unlock(&dst_ctx->mutex);
> +}
> +EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
> +
>  static void sync_child_event(struct perf_event *child_event,
>  			       struct task_struct *child)
>  {
> 

Thank you every much.

How about interpreting the parameter 'cpu' for perf_event_open() as target socket
instead of target cpu? So that we can get rid of raw_spin_lock in the uncore_box.
The event_init() pmu callback can do this job easily.

Regards
Yan, Zheng