Re: [PATCH v5 1/3] tracing/hist: Add poll(POLLIN) support on hist file

[Steven Rostedt]

On Sun, 18 Aug 2024 22:18:34 +0900
"Masami Hiramatsu (Google)" <mhiramat@xxxxxxxxxx> wrote:

> From: Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
> 
> Add poll syscall support on the `hist` file. The Waiter will be waken
> up when the histogram is updated with POLLIN.
> 
> Currently, there is no way to wait for a specific event in userspace.
> So user needs to peek the `trace` periodicaly, or wait on `trace_pipe`.
> But that is not good idea to peek the `trace` for the event randomely
> happens. And `trace_pipe` is not coming back until a page is filled
> with events.
> 
> This allows user to wait for a specific events on `hist` file. User
> can set a histogram trigger on the event which they want to monitor.
> And poll() on its `hist` file. Since this poll() returns POLLIN,
> the next poll() will return soon unless you do read() on hist file.
> 
> NOTE: To read the hist file again, you must set the file offset to 0,
> but just for monitoring the event, you may not need to read the
> histogram.
> 
> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
> Reviewed-by: Tom Zanussi <zanussi@xxxxxxxxxx>
> ---
>  include/linux/trace_events.h     |    5 +++
>  kernel/trace/trace_events.c      |   18 +++++++++
>  kernel/trace/trace_events_hist.c |   76 +++++++++++++++++++++++++++++++++++++-
>  3 files changed, 96 insertions(+), 3 deletions(-)
> 
> diff --git a/include/linux/trace_events.h b/include/linux/trace_events.h
> index 42bedcddd511..f3ec67d34097 100644
> --- a/include/linux/trace_events.h
> +++ b/include/linux/trace_events.h
> @@ -663,6 +663,11 @@ struct trace_event_file {
>  	struct trace_subsystem_dir	*system;
>  	struct list_head		triggers;
>  
> +#ifdef CONFIG_HIST_TRIGGERS
> +	struct irq_work			hist_work;
> +	wait_queue_head_t		hist_wq;
> +#endif

I really hate to add this to the trace_event_file, as that is created for
every event. If there are 1853 events (that's what my machine shows), and
sizeof(struct irq_work) = 32 bytes, and sizeof(wait_queue_head_t) is 24
bytes, this ends up adding 103,768 bytes to every instance!

That's a bit of bloat.

Can we make this a global work queue, and perhaps add a flag to flags that
states there's a waiter? Then on poll wakeup, we can check if the change
happened to hist that the waiter is waiting on (which it actually already does!)

> +
>  	/*
>  	 * 32 bit flags:
>  	 *   bit 0:		enabled
> diff --git a/kernel/trace/trace_events.c b/kernel/trace/trace_events.c
> index 7266ec2a4eea..0f077b32eea4 100644
> --- a/kernel/trace/trace_events.c
> +++ b/kernel/trace/trace_events.c
> @@ -2972,6 +2972,20 @@ static bool event_in_systems(struct trace_event_call *call,
>  	return !*p || isspace(*p) || *p == ',';
>  }
>  
> +#ifdef CONFIG_HIST_TRIGGERS
> +/*
> + * Wake up waiter on the hist_wq from irq_work because the hist trigger
> + * may happen in any context.
> + */
> +static void hist_event_irq_work(struct irq_work *work)
> +{
> +	struct trace_event_file *event_file;
> +
> +	event_file = container_of(work, struct trace_event_file, hist_work);
> +	wake_up_all(&event_file->hist_wq);
> +}
> +#endif
> +
>  static struct trace_event_file *
>  trace_create_new_event(struct trace_event_call *call,
>  		       struct trace_array *tr)
> @@ -3004,6 +3018,10 @@ trace_create_new_event(struct trace_event_call *call,
>  	INIT_LIST_HEAD(&file->triggers);
>  	list_add(&file->list, &tr->events);
>  	refcount_set(&file->ref, 1);
> +#ifdef CONFIG_HIST_TRIGGERS
> +	init_irq_work(&file->hist_work, hist_event_irq_work);
> +	init_waitqueue_head(&file->hist_wq);
> +#endif
>  
>  	return file;
>  }
> diff --git a/kernel/trace/trace_events_hist.c b/kernel/trace/trace_events_hist.c
> index 5f9119eb7c67..d27b60f54f68 100644
> --- a/kernel/trace/trace_events_hist.c
> +++ b/kernel/trace/trace_events_hist.c
> @@ -5314,6 +5314,9 @@ static void event_hist_trigger(struct event_trigger_data *data,
>  
>  	if (resolve_var_refs(hist_data, key, var_ref_vals, true))
>  		hist_trigger_actions(hist_data, elt, buffer, rec, rbe, key, var_ref_vals);
> +
> +	if (hist_data->event_file && wq_has_sleeper(&hist_data->event_file->hist_wq))
> +		irq_work_queue(&hist_data->event_file->hist_work);

Instead of using wq_has_sleeper() we can use the event_file->flags instead.

>  }
>  
>  static void hist_trigger_stacktrace_print(struct seq_file *m,
> @@ -5593,15 +5596,36 @@ static void hist_trigger_show(struct seq_file *m,
>  		   n_entries, (u64)atomic64_read(&hist_data->map->drops));
>  }
>  
> +struct hist_file_data {
> +	struct file *file;
> +	u64 last_read;
> +};
> +
> +static u64 get_hist_hit_count(struct trace_event_file *event_file)
> +{
> +	struct hist_trigger_data *hist_data;
> +	struct event_trigger_data *data;
> +	u64 ret = 0;
> +
> +	list_for_each_entry(data, &event_file->triggers, list) {
> +		if (data->cmd_ops->trigger_type == ETT_EVENT_HIST) {
> +			hist_data = data->private_data;
> +			ret += atomic64_read(&hist_data->map->hits);
> +		}
> +	}
> +	return ret;
> +}
> +
>  static int hist_show(struct seq_file *m, void *v)
>  {
> +	struct hist_file_data *hist_file = m->private;
>  	struct event_trigger_data *data;
>  	struct trace_event_file *event_file;
>  	int n = 0, ret = 0;
>  
>  	mutex_lock(&event_mutex);
>  
> -	event_file = event_file_file(m->private);
> +	event_file = event_file_file(hist_file->file);
>  	if (unlikely(!event_file)) {
>  		ret = -ENODEV;
>  		goto out_unlock;
> @@ -5611,6 +5635,7 @@ static int hist_show(struct seq_file *m, void *v)
>  		if (data->cmd_ops->trigger_type == ETT_EVENT_HIST)
>  			hist_trigger_show(m, data, n++);
>  	}
> +	hist_file->last_read = get_hist_hit_count(event_file);
>  
>   out_unlock:
>  	mutex_unlock(&event_mutex);
> @@ -5618,24 +5643,69 @@ static int hist_show(struct seq_file *m, void *v)
>  	return ret;
>  }
>  
> +static __poll_t event_hist_poll(struct file *file, struct poll_table_struct *wait)
> +{
> +	struct trace_event_file *event_file;
> +	struct seq_file *m = file->private_data;
> +	struct hist_file_data *hist_file = m->private;
> +	__poll_t ret = 0;
> +
> +	mutex_lock(&event_mutex);
> +
> +	event_file = event_file_data(file);
> +	if (!event_file) {
> +		ret = EPOLLERR;
> +		goto out_unlock;
> +	}
> +
> +	poll_wait(file, &event_file->hist_wq, wait);
> +
> +	if (hist_file->last_read != get_hist_hit_count(event_file))
> +		ret = EPOLLIN | EPOLLRDNORM;

Even if this gets woken up early, it will still not go back to user space.

I don't expect a lot of hist waiters so this should not be an issue.

I don't see an issue if we just have the hist wait queue global.

-- Steve


> +
> +out_unlock:
> +	mutex_unlock(&event_mutex);
> +
> +	return ret;
> +}
> +
> +static int event_hist_release(struct inode *inode, struct file *file)
> +{
> +	struct seq_file *m = file->private_data;
> +	struct hist_file_data *hist_file = m->private;
> +
> +	kfree(hist_file);
> +	return tracing_single_release_file_tr(inode, file);
> +}
> +
>  static int event_hist_open(struct inode *inode, struct file *file)
>  {
> +	struct hist_file_data *hist_file;
>  	int ret;
>  
>  	ret = tracing_open_file_tr(inode, file);
>  	if (ret)
>  		return ret;
>  
> +	hist_file = kzalloc(sizeof(*hist_file), GFP_KERNEL);
> +	if (!hist_file)
> +		return -ENOMEM;
> +	hist_file->file = file;
> +
>  	/* Clear private_data to avoid warning in single_open() */
>  	file->private_data = NULL;
> -	return single_open(file, hist_show, file);
> +	ret = single_open(file, hist_show, hist_file);
> +	if (ret)
> +		kfree(hist_file);
> +	return ret;
>  }
>  
>  const struct file_operations event_hist_fops = {
>  	.open = event_hist_open,
>  	.read = seq_read,
>  	.llseek = seq_lseek,
> -	.release = tracing_single_release_file_tr,
> +	.release = event_hist_release,
> +	.poll = event_hist_poll,
>  };
>  
>  #ifdef CONFIG_HIST_TRIGGERS_DEBUG