queue-4.14/perf-ring_buffer-add-ordering-to-rb-nest-increment.patch

   1 From 27cc54201727bcc4aedc236c16c91460138f3849 Mon Sep 17 00:00:00 2001
   2 From: Peter Zijlstra <peterz@infradead.org>
   3 Date: Fri, 17 May 2019 13:52:32 +0200
   4 Subject: perf/ring_buffer: Add ordering to rb->nest increment
   5
   6 [ Upstream commit 3f9fbe9bd86c534eba2faf5d840fd44c6049f50e ]
   7
   8 Similar to how decrementing rb->next too early can cause data_head to
   9 (temporarily) be observed to go backward, so too can this happen when
  10 we increment too late.
  11
  12 This barrier() ensures the rb->head load happens after the increment,
  13 both the one in the 'goto again' path, as the one from
  14 perf_output_get_handle() -- albeit very unlikely to matter for the
  15 latter.
  16
  17 Suggested-by: Yabin Cui <yabinc@google.com>
  18 Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
  19 Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
  20 Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
  21 Cc: Jiri Olsa <jolsa@redhat.com>
  22 Cc: Linus Torvalds <torvalds@linux-foundation.org>
  23 Cc: Peter Zijlstra <peterz@infradead.org>
  24 Cc: Stephane Eranian <eranian@google.com>
  25 Cc: Thomas Gleixner <tglx@linutronix.de>
  26 Cc: Vince Weaver <vincent.weaver@maine.edu>
  27 Cc: acme@kernel.org
  28 Cc: mark.rutland@arm.com
  29 Cc: namhyung@kernel.org
  30 Fixes: ef60777c9abd ("perf: Optimize the perf_output() path by removing IRQ-disables")
  31 Link: http://lkml.kernel.org/r/20190517115418.309516009@infradead.org
  32 Signed-off-by: Ingo Molnar <mingo@kernel.org>
  33 Signed-off-by: Sasha Levin <sashal@kernel.org>
  34 ---
  35  kernel/events/ring_buffer.c | 9 +++++++++
  36  1 file changed, 9 insertions(+)
  37
  38 diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
  39 index fde853270c09..aef2af80a927 100644
  40 --- a/kernel/events/ring_buffer.c
  41 +++ b/kernel/events/ring_buffer.c
  42 @@ -49,6 +49,15 @@ static void perf_output_put_handle(struct perf_output_handle *handle)
  43         unsigned long head;
  44
  45  again:
  46 +       /*
  47 +        * In order to avoid publishing a head value that goes backwards,
  48 +        * we must ensure the load of @rb->head happens after we've
  49 +        * incremented @rb->nest.
  50 +        *
  51 +        * Otherwise we can observe a @rb->head value before one published
  52 +        * by an IRQ/NMI happening between the load and the increment.
  53 +        */
  54 +       barrier();
  55         head = local_read(&rb->head);
  56
  57         /*
  58 --
  59 2.20.1
  60