trace_pipe tentative fix (Re: [PATCH -tip 2/3] Tracing/ftrace: Adapt mmiotrace to the new type of print_line)

[Pekka Paalanen <pq@iki.fi>]

On Sat, 27 Sep 2008 14:17:47 +0200
"Frédéric Weisbecker" <fweisbec@gmail.com> wrote:

> 2008/9/26 Pekka Paalanen <pq@iki.fi>:
> > Yes, I do disagree about printing stuff that doesn't belong there.
> > First, I doubt printing bogus text is a right way to fix the
> > early pipe EOF problem. Second, if you really have to do that,
> > do it so that it obeys the mmiotrace log format specification.
> > (I'd recommend a MARK entry.) The spec is in
> > Documentation/tracers/mmiotrace.txt.
> 
> No problem, I have one other option: ignoring and sleep again before
> receiving another entries.
> I can do another patch this week end to implement that.
> The third patch only concerns the boot tracer. It relays to other
> output functions.
> 
> Just tell me if you agree with the ignoring...

If I understand you suggestion, it looks like the right thing to do.
Here is a tentative fix, which has not even been compile-tested.

Is it so that the problem is triggered by consuming a trace entry
which does not produce any output? If that entry is all there is
in the ring at a time of a read call, then the last call to
trace_seq_to_user() returns -EBUSY, because there is nothing to
copy to user. What I failed to understand when I wrote that
piece of code, is that returning 0 means EOF. The only cases
when we do want to return an EOF are near the
	while (trace_empty(iter)) {
loop.

Frederic, could you test the fix, and if it works, send it to Ingo?


Thanks.

---

diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
index 6ada059..d85659a 100644
--- a/kernel/trace/trace.c
+++ b/kernel/trace/trace.c
@@ -2616,6 +2616,7 @@ tracing_read_pipe(struct file *filp, char __user *ubuf,
                        goto out;
        }
 
+waitagain:
        while (trace_empty(iter)) {
 
                if ((filp->f_flags & O_NONBLOCK)) {
@@ -2749,8 +2750,13 @@ tracing_read_pipe(struct file *filp, char __user *ubuf,
        sret = trace_seq_to_user(&iter->seq, ubuf, cnt);
        if (iter->seq.readpos >= iter->seq.len)
                trace_seq_reset(&iter->seq);
+
+       /*
+        * If there was nothing to send to user, inspite of consuming trace
+        * entries, go back to wait for more entries.
+        */
        if (sret == -EBUSY)
-               sret = 0;
+               goto waitagain;
 
 out:
        mutex_unlock(&trace_types_lock);