perf/ring_buffer: Use local_try_cmpxchg in __perf_output_begin
authorUros Bizjak <ubizjak@gmail.com>
Sat, 8 Jul 2023 09:00:37 +0000 (11:00 +0200)
committerPeter Zijlstra <peterz@infradead.org>
Mon, 10 Jul 2023 07:52:36 +0000 (09:52 +0200)
Use local_try_cmpxchg instead of local_cmpxchg (*ptr, old, new) == old
in __perf_output_begin.  x86 CMPXCHG instruction returns success in ZF
flag, so this change saves a compare after cmpxchg (and related move
instruction in front of cmpxchg).

Also, try_cmpxchg implicitly assigns old *ptr value to "old" when cmpxchg
fails. There is no need to re-read the value in the loop.

No functional change intended.

Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20230708090048.63046-2-ubizjak@gmail.com
kernel/events/ring_buffer.c

index a0433f37b0243a334a6bf45a95bd08ab527a0008..fb1e180b5f0af767a71b3afe53ae87cc17f99f67 100644 (file)
@@ -191,9 +191,10 @@ __perf_output_begin(struct perf_output_handle *handle,
 
        perf_output_get_handle(handle);
 
+       offset = local_read(&rb->head);
        do {
+               head = offset;
                tail = READ_ONCE(rb->user_page->data_tail);
-               offset = head = local_read(&rb->head);
                if (!rb->overwrite) {
                        if (unlikely(!ring_buffer_has_space(head, tail,
                                                            perf_data_size(rb),
@@ -217,7 +218,7 @@ __perf_output_begin(struct perf_output_handle *handle,
                        head += size;
                else
                        head -= size;
-       } while (local_cmpxchg(&rb->head, offset, head) != offset);
+       } while (!local_try_cmpxchg(&rb->head, &offset, head));
 
        if (backward) {
                offset = head;