target/linux/generic/patches-3.3/049-codel-refine-one-condition-to-avoid-a-nul-rec_inv_sqrt.patch

   1 From patchwork Mon Jul 30 06:52:21 2012
   2 Content-Type: text/plain; charset="utf-8"
   3 MIME-Version: 1.0
   4 Content-Transfer-Encoding: 7bit
   5 Subject: codel: refine one condition to avoid a nul rec_inv_sqrt
   6 Date: Sun, 29 Jul 2012 20:52:21 -0000
   7 From: Eric Dumazet <eric.dumazet@gmail.com>
   8 X-Patchwork-Id: 173968
   9 Message-Id: <1343631141.2626.13293.camel@edumazet-glaptop>
  10 To: David Miller <davem@davemloft.net>
  11 Cc: netdev <netdev@vger.kernel.org>, Anton Mich <lp2s1h@gmail.com>
  12
  13 From: Eric Dumazet <edumazet@google.com>
  14
  15 One condition before codel_Newton_step() was not good if
  16 we never left the dropping state for a flow. As a result
  17 rec_inv_sqrt was 0, instead of the ~0 initial value.
  18
  19 codel control law was then set to a very aggressive mode, dropping
  20 many packets before reaching 'target' and recovering from this problem.
  21
  22 To keep codel_vars_init() as efficient as possible, refine
  23 the condition to make sure rec_inv_sqrt initial value is correct
  24
  25 Many thanks to Anton Mich for discovering the issue and suggesting
  26 a fix.
  27
  28 Reported-by: Anton Mich <lp2s1h@gmail.com>
  29 Signed-off-by: Eric Dumazet <edumazet@google.com>
  30
  31 ---
  32 include/net/codel.h |    8 ++++++--
  33  1 file changed, 6 insertions(+), 2 deletions(-)
  34
  35
  36
  37 --
  38 To unsubscribe from this list: send the line "unsubscribe netdev" in
  39 the body of a message to majordomo@vger.kernel.org
  40 More majordomo info at  http://vger.kernel.org/majordomo-info.html
  41
  42 --- a/include/net/codel.h
  43 +++ b/include/net/codel.h
  44 @@ -305,6 +305,8 @@ static struct sk_buff *codel_dequeue(str
  45                         }
  46                 }
  47         } else if (drop) {
  48 +               u32 delta;
  49 +
  50                 if (params->ecn && INET_ECN_set_ce(skb)) {
  51                         stats->ecn_mark++;
  52                 } else {
  53 @@ -320,9 +322,11 @@ static struct sk_buff *codel_dequeue(str
  54                  * assume that the drop rate that controlled the queue on the
  55                  * last cycle is a good starting point to control it now.
  56                  */
  57 -               if (codel_time_before(now - vars->drop_next,
  58 +               delta = vars->count - vars->lastcount;
  59 +               if (delta > 1 &&
  60 +                   codel_time_before(now - vars->drop_next,
  61                                       16 * params->interval)) {
  62 -                       vars->count = (vars->count - vars->lastcount) | 1;
  63 +                       vars->count = delta;
  64                         /* we dont care if rec_inv_sqrt approximation
  65                          * is not very precise :
  66                          * Next Newton steps will correct it quadratically.