Re: [PATCH] tracing/uprobes: Support ftrace_event_file base multibuffer

[Oleg Nesterov <oleg@redhat.com>]

On 06/14, zhangwei(Jovi) wrote:
>
> Support multi-buffer on uprobe-based dynamic events by
> using ftrace_event_file.
>
> The code change is based on kprobe-based dynamic events
> multibuffer support work commited by Masami(commit 41a7dd420c)

And the change in probe_event_enable() doesn't look right, but
let me repeat I didn't read the patch carefully yet.

> +static DEFINE_MUTEX(uprobe_enable_lock);
> +
>  static inline bool is_trace_uprobe_enabled(struct trace_uprobe *tu)
>  {
>  	return tu->flags & (TP_FLAG_TRACE | TP_FLAG_PROFILE);
> @@ -607,33 +655,123 @@ typedef bool (*filter_func_t)(struct uprobe_consumer *self,
>  				struct mm_struct *mm);
> 
>  static int
> -probe_event_enable(struct trace_uprobe *tu, int flag, filter_func_t filter)
> +probe_event_enable(struct trace_uprobe *tu, struct ftrace_event_file *file,
> +		   filter_func_t filter)
>  {
> +	int enabled = 0;
>  	int ret = 0;
> 
> +	mutex_lock(&uprobe_enable_lock);

Do we really need this? Can't we really on mutex_event hold by the caller?

>  	if (is_trace_uprobe_enabled(tu))
> -		return -EINTR;
> +		enabled = 1;
> +
> +	if (file) {
> +		struct ftrace_event_file **new, **old;
> +		int n = trace_uprobe_nr_files(tu);
> +
> +		old = rcu_dereference_raw(tu->files);
> +		/* 1 is for new one and 1 is for stopper */
> +		new = kzalloc((n + 2) * sizeof(struct ftrace_event_file *),
> +			      GFP_KERNEL);
> +		if (!new) {
> +			ret = -ENOMEM;
> +			goto out_unlock;
> +		}
> +		memcpy(new, old, n * sizeof(struct ftrace_event_file *));
> +		new[n] = file;
> +		/* The last one keeps a NULL */
> +
> +		rcu_assign_pointer(tu->files, new);
> +		tu->flags |= TP_FLAG_TRACE;
> +
> +		if (old) {
> +			/* Make sure the probe is done with old files */
> +			synchronize_sched();
> +			kfree(old);
> +		}
> +	} else
> +		tu->flags |= TP_FLAG_PROFILE;

So it can set both TP_FLAG_TRACE and TP_FLAG_PROFILE, yes?

If yes, this is not right. Until we change the pre-filtering at least.
Currently TP_FLAG_TRACE/TP_FLAG_PROFILE are mutually exclusive.

I think it makes sense to remove this limitation anyway, and in fact
I do not remember why I didn't do this... But this needs a separate
change.

Oleg.