tracing: Use temp buffer when filtering events

Filtering of events requires the data to be written to the ring buffer
before it can be decided to filter or not. This is because the parameters of
the filter are based on the result that is written to the ring buffer and
not on the parameters that are passed into the trace functions.

The ftrace ring buffer is optimized for writing into the ring buffer and
committing. The discard procedure used when filtering decides the event
should be discarded is much more heavy weight. Thus, using a temporary
filter when filtering events can speed things up drastically.

Without a temp buffer we have:

 # trace-cmd start -p nop
 # perf stat -r 10 hackbench 50
       0.790706626 seconds time elapsed ( +-  0.71% )

 # trace-cmd start -e all
 # perf stat -r 10 hackbench 50
       1.566904059 seconds time elapsed ( +-  0.27% )

 # trace-cmd start -e all -f 'common_preempt_count==20'
 # perf stat -r 10 hackbench 50
       1.690598511 seconds time elapsed ( +-  0.19% )

 # trace-cmd start -e all -f 'common_preempt_count!=20'
 # perf stat -r 10 hackbench 50
       1.707486364 seconds time elapsed ( +-  0.30% )

The first run above is without any tracing, just to get a based figure.
hackbench takes ~0.79 seconds to run on the system.

The second run enables tracing all events where nothing is filtered. This
increases the time by 100% and hackbench takes 1.57 seconds to run.

The third run filters all events where the preempt count will equal "20"
(this should never happen) thus all events are discarded. This takes 1.69
seconds to run. This is 10% slower than just committing the events!

The last run enables all events and filters where the filter will commit all
events, and this takes 1.70 seconds to run. The filtering overhead is
approximately 10%. Thus, the discard and commit of an event from the ring
buffer may be about the same time.

With this patch, the numbers change:

 # trace-cmd start -p nop
 # perf stat -r 10 hackbench 50
       0.778233033 seconds time elapsed ( +-  0.38% )

 # trace-cmd start -e all
 # perf stat -r 10 hackbench 50
       1.582102692 seconds time elapsed ( +-  0.28% )

 # trace-cmd start -e all -f 'common_preempt_count==20'
 # perf stat -r 10 hackbench 50
       1.309230710 seconds time elapsed ( +-  0.22% )

 # trace-cmd start -e all -f 'common_preempt_count!=20'
 # perf stat -r 10 hackbench 50
       1.786001924 seconds time elapsed ( +-  0.20% )

The first run is again the base with no tracing.

The second run is all tracing with no filtering. It is a little slower, but
that may be well within the noise.

The third run shows that discarding all events only took 1.3 seconds. This
is a speed up of 23%! The discard is much faster than even the commit.

The one downside is shown in the last run. Events that are not discarded by
the filter will take longer to add, this is due to the extra copy of the
event.

Cc: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>

1 files changed, 10 insertions, 0 deletions

diff --git a/kernel/trace/trace_events_filter.c b/kernel/trace/trace_events_filter.c
index d1d27bf37a19..9daa9b3bc6d9 100644
--- a/kernel/trace/trace_events_filter.c
+++ b/kernel/trace/trace_events_filter.c
@@ -823,7 +823,12 @@ static void __free_preds(struct event_filter *filter)
 
 static void filter_disable(struct trace_event_file *file)
 {
+	unsigned long old_flags = file->flags;
+
 	file->flags &= ~EVENT_FILE_FL_FILTERED;
+
+	if (old_flags != file->flags)
+		trace_buffered_event_disable();
 }
 
 static void __free_filter(struct event_filter *filter)
@@ -1698,7 +1703,12 @@ fail:
 
 static inline void event_set_filtered_flag(struct trace_event_file *file)
 {
+	unsigned long old_flags = file->flags;
+
 	file->flags |= EVENT_FILE_FL_FILTERED;
+
+	if (old_flags != file->flags)
+		trace_buffered_event_enable();
 }
 
 static inline void event_set_filter(struct trace_event_file *file,