Re: [RFC PATCH 3/8] perf: Init/fini PMU specific data

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Nov 28, 2019 at 07:14:26AM -0800, kan.liang@linux.intel.com wrote:
> +static int
> +init_system_wide_ctx_data(size_t ctx_size)
> +{
> +	struct task_struct *g, *p;
> +	int failed_alloc = 0;
> +	unsigned long flags;
> +
> +	/*
> +	 * Allocate perf_ctx_data for all existing threads by the first event.
> +	 *
> +	 * The perf_ctx_data for new thread will be allocated in
> +	 * perf_event_fork(). The perf_event_fork() is called after the thread
> +	 * is added into the tasklist. It guarantees that any new threads will
> +	 * not be missed.
> +	 */
> +	raw_spin_lock_irqsave(&task_data_events_lock, flags);
> +	if (atomic_inc_return(&nr_task_data_events) > 1)
> +		goto unlock;
> +
> +	read_lock(&tasklist_lock);
> +
> +	for_each_process_thread(g, p) {
> +		/*
> +		 * The PMU specific data may already be allocated by
> +		 * per-process event. Need to update refcounter.
> +		 * init_task_ctx_data_rcu() is called here.
> +		 * Do a quick allocation in first round with GFP_ATOMIC.
> +		 */
> +		if (init_task_ctx_data_rcu(p, ctx_size, GFP_ATOMIC))
> +			failed_alloc++;
> +	}

This is atricous crap. Also it is completely broken for -RT.