Re: [PATCH 2/2] tracing - putting cond_resched into tace_pipe loop

[Oleg Nesterov <oleg@redhat.com>]

On 03/12, Jiri Olsa wrote:
>
> --- a/kernel/trace/trace.c
> +++ b/kernel/trace/trace.c
> @@ -3237,10 +3237,23 @@ waitagain:
>  		 * One of the trace_seq_* functions is not used properly.
>  		 */
>  		WARN_ON(iter->seq.full);
> +
> +		/*
> +		 * There's a chance this loop might get quite tight,
> +		 * causing latency in non preemptive kernel.
> +		 */
> +		cond_resched();
> +		if (signal_pending(current)) {
> +			sret = -EINTR;
> +			break;

First of all: I do not pretend I understand this code ;) Still, a
couple of nits.

-EINTR doesn't look exactly right, I'd suggest -ERESTARTSYS. The same
for tracing_wait_pipe() btw, I think it should be fixed.



I wonder if it makes sense to simply "break" if signal_pending(), it
is possible we already have something to report via trace_seq_to_user().
Then we could do

	-	if (sret == -EBUSY)
	-		goto waitagain;
	+	if (sret == -EBUSY) {
	+		if (!signal_pending())
	+			goto waitagain;
	+		sret = -ERESTARTSYS;
	+	}

Or we can change tracing_wait_pipe() to check signal_pending()
uncondditionally, I dunno.

Up to you, but note that otherwise the logic looks a bit strange.
Suppose that signal_pending() is already true when we call
tracing_wait_pipe(). In this case we are going to do the "unnecessary"
work and then return EINTR/ERESTART. This is correct, the next
invocation does trace_seq_to_user() before anything else, just
looks a bit strange.

Oleg.