Re: [tracing] 06e0a548ba: WARNING:at_kernel/trace/ring_buffer.c:#ring_buffer_iter_peek

[Sven Schnelle]

Hi Steve,

On Wed, May 13, 2020 at 03:30:33PM -0400, Steven Rostedt wrote:
> On Wed, 13 May 2020 18:15:57 +0200
> Sven Schnelle <svens@xxxxxxxxxxxxx> wrote:
> 
> > Thanks for looking into this. I've attached my /proc/config.gz to this Mail.
> > The x86 system is my Laptop which is a Thinkpad X280 with 4 HT CPUs (so 8 cpus
> > in total). I've tried disabling preemption, but this didn't help.
> > 
> > It's always this check that causes the loop:
> > 
> > if (iter->head >= rb_page_size(iter->head_page)) {
> > 	rb_inc_iter(iter);
> > 	goto again;
> > }
> > 
> > On the first loop iter->head is some value > 0 and rb_page_size returns
> > 0, afterwards it gets twice to this check with both values 0. The third
> > time the warning is triggered. Maybe that information helps.
> 
> I figured out what was causing this, and that's just that the writer and
> the iterator could end up almost "in sync" where the writer writes to each
> of the pages the iterator is trying to read, and this can trigger the three
> failures of "zero commits" per page.
> 
> I had a way to detect this, but then realized that it may be possible for
> an active writer to possibly trigger the other failures to get an event,
> that I just decided to force it to try three times, and simply return NULL
> if an active writer is messing with the iterator. The iterator is a "best
> effort" to read the buffer if there's an active writer. The consumer read
> (trace_pipe) is made for that.
> 
> This patch should solve you issues.
> 
> (care to give a Tested-by: if it works for you?)

Well, as there's no longer a RB_WARN_ON that indeed fixes the issue :-)

Tested-by: Sven Schnelle <svens@xxxxxxxxxxxxx>

Thanks!

Sven