Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
tracing: Add better comments for the filtering temp buffer use case
When filtering is enabled, the event is copied into a temp buffer instead
of being written into the ring buffer directly, because the discarding of
events from the ring buffer is very expensive, and doing the extra copy is
much faster than having to discard most of the time.
As that logic is subtle, add comments to explain in more detail to what is
going on and how it works.
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
+35
-1
+35
-1
kernel/trace/trace.c
+35
-1
kernel/trace/trace.c
···
2734
2734
if (!tr->no_filter_buffering_ref &&
2735
2735
(trace_file->flags & (EVENT_FILE_FL_SOFT_DISABLED | EVENT_FILE_FL_FILTERED)) &&
2736
2736
(entry = this_cpu_read(trace_buffered_event))) {
2737
-
/* Try to use the per cpu buffer first */
2737
+
/*
2738
+
* Filtering is on, so try to use the per cpu buffer first.
2739
+
* This buffer will simulate a ring_buffer_event,
2740
+
* where the type_len is zero and the array[0] will
2741
+
* hold the full length.
2742
+
* (see include/linux/ring-buffer.h for details on
2743
+
* how the ring_buffer_event is structured).
2744
+
*
2745
+
* Using a temp buffer during filtering and copying it
2746
+
* on a matched filter is quicker than writing directly
2747
+
* into the ring buffer and then discarding it when
2748
+
* it doesn't match. That is because the discard
2749
+
* requires several atomic operations to get right.
2750
+
* Copying on match and doing nothing on a failed match
2751
+
* is still quicker than no copy on match, but having
2752
+
* to discard out of the ring buffer on a failed match.
2753
+
*/
2738
2754
int max_len = PAGE_SIZE - struct_size(entry, array, 1);
2739
2755
2740
2756
val = this_cpu_inc_return(trace_buffered_event_cnt);
2757
+
2758
+
/*
2759
+
* Preemption is disabled, but interrupts and NMIs
2760
+
* can still come in now. If that happens after
2761
+
* the above increment, then it will have to go
2762
+
* back to the old method of allocating the event
2763
+
* on the ring buffer, and if the filter fails, it
2764
+
* will have to call ring_buffer_discard_commit()
2765
+
* to remove it.
2766
+
*
2767
+
* Need to also check the unlikely case that the
2768
+
* length is bigger than the temp buffer size.
2769
+
* If that happens, then the reserve is pretty much
2770
+
* guaranteed to fail, as the ring buffer currently
2771
+
* only allows events less than a page. But that may
2772
+
* change in the future, so let the ring buffer reserve
2773
+
* handle the failure in that case.
2774
+
*/
2741
2775
if (val == 1 && likely(len <= max_len)) {
2742
2776
trace_event_setup(entry, type, trace_ctx);
2743
2777
entry->array[0] = len;