Re: [PATCH] tracing - fix function graph trace to properly display failed entries

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, Nov 24, 2009 at 10:14:36PM +0100, Frederic Weisbecker wrote:
> But I guess it doesn't happen because the function graph tracer
> has too large entries, or it would do an endless loop while
> reading the trace file.


Ah actually it won't as we don't check the return value
of trace_print_line() from the seq show handler.
There is no retry.

And also there is no need to retry. The buffer gets never
entirely filled in the trace file reading path, as it
is flushed after each entry. And no single trace fills the
whole buffer. I just made some tests and we never reach that
point (no TRACE_TYPE_PARTIAL_LINE) returned. So no overflow
as far as I tested (there might be rare corner cases).

But it happens with trace_pipe reading.

There is another bug. Here is a sample from
trace reading:

 0)               |  sys_newfstat() {
 0)               |    vfs_fstat() {
 0)   1.187 us    |      fget();
 0)               |      vfs_getattr() {
 0)               |        security_inode_getattr() {
 0)   0.608 us    |              }
 0)   2.951 us    |            }
 0) + 65.349 us   |          }
 0)               |          path_put() {
 0)   0.608 us    |            dput();
 0)   0.548 us    |            mntput_no_expire();
 0)   2.748 us    |          }
 0) + 74.472 us   |        }


I think we are loosing some traces here, between security_inode_getattr()
and path_put().

I'm not sure why. The problem related by Olsa would explain such
symptoms, and what olsa reported is a real bug, but only
in trace_pipe. There is also something else, probably more
related to the fast path.