Re: [PATCH] tracing: Fix to avoid wakeup loop in splice read of per-cpu buffer

[Masami Hiramatsu (Google) <mhiramat@kernel.org>]

On Mon, 21 Aug 2023 23:19:18 +0900
"Masami Hiramatsu (Google)" <mhiramat@kernel.org> wrote:

> From: Masami Hiramatsu (Google) <mhiramat@kernel.org>
> 
> ftrace user can set 0 or small number to the 'buffer_percent' for quick
> response for the ring buffer. In that case wait_on_pipe() will return
> before filling a page of the ring buffer. That is too soon for splice()
> because ring_buffer_read_page() will fail again.
> This leads unnecessary loop in tracing_buffers_splice_read().
> 
> Set a minimum percentage of the buffer which is enough to fill a page to
> wait_on_pipe() to avoid this situation.
> 
> Fixes: 03329f993978 ("tracing: Add tracefs file buffer_percentage")
> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
> ---
>  kernel/trace/trace.c |   12 +++++++++++-
>  1 file changed, 11 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> index b8870078ef58..88448e8d8214 100644
> --- a/kernel/trace/trace.c
> +++ b/kernel/trace/trace.c
> @@ -8462,6 +8462,8 @@ tracing_buffers_splice_read(struct file *file, loff_t *ppos,
>  	/* did we read anything? */
>  	if (!spd.nr_pages) {
>  		long wait_index;
> +		size_t nr_pages;
> +		size_t full;
>  
>  		if (ret)
>  			goto out;
> @@ -8472,7 +8474,15 @@ tracing_buffers_splice_read(struct file *file, loff_t *ppos,
>  
>  		wait_index = READ_ONCE(iter->wait_index);
>  
> -		ret = wait_on_pipe(iter, iter->tr->buffer_percent);
> +		/* For splice, we have to ensure at least 1 page is filled */
> +		nr_pages = ring_buffer_nr_pages(iter->array_buffer->buffer, iter->cpu_file);
> +		if (nr_pages * iter->tr->buffer_percent < 100) {
> +			full = nr_pages + 99;
> +			do_div(full, nr_pages);
> +		} else
> +			full = iter->tr->buffer_percent;

Ah I must have to take a sleep well. What I need is to ensure full >= 1.

static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
{
...
        return (dirty * 100) > (full * nr_pages);
}

If dirty = 0, this always false.
But I think if full == 0, this should return true.

If dirty = 1,

- nr_pages < 100, this is always true and that is good.

- nr_pages > 100, even if full is 1 (smallest), it doesn't true. But that is OK
  because dirty page number will be increased later.

- nr_pages == 100 is the corner case. I think this should be

  return (dirty * 100) >= (full * nr_pages);

Let me update the patch.

Thank you,

> +
> +		ret = wait_on_pipe(iter, full);
>  		if (ret)
>  			goto out;
>  
> 


-- 
Masami Hiramatsu (Google) <mhiramat@kernel.org>