target/linux/generic/patches-3.3/046-fq_codel-qdisc-backlog.patch

   1 From 7bd90773f89001ea4960ed47676b550137f3facb Mon Sep 17 00:00:00 2001
   2 From: Eric Dumazet <edumazet@google.com>
   3 Date: Wed, 16 May 2012 04:39:09 +0000
   4 Subject: [PATCH] fq_codel: should use qdisc backlog as threshold
   5
   6 commit 865ec5523dadbedefbc5710a68969f686a28d928 upstream.
   7
   8 codel_should_drop() logic allows a packet being not dropped if queue
   9 size is under max packet size.
  10
  11 In fq_codel, we have two possible backlogs : The qdisc global one, and
  12 the flow local one.
  13
  14 The meaningful one for codel_should_drop() should be the global backlog,
  15 not the per flow one, so that thin flows can have a non zero drop/mark
  16 probability.
  17
  18 Signed-off-by: Eric Dumazet <edumazet@google.com>
  19 Cc: Dave Taht <dave.taht@bufferbloat.net>
  20 Cc: Kathleen Nichols <nichols@pollere.com>
  21 Cc: Van Jacobson <van@pollere.net>
  22 Signed-off-by: David S. Miller <davem@davemloft.net>
  23 ---
  24  include/net/codel.h      |   15 +++++++--------
  25  net/sched/sch_codel.c    |    4 ++--
  26  net/sched/sch_fq_codel.c |    5 +++--
  27  3 files changed, 12 insertions(+), 12 deletions(-)
  28
  29 --- a/include/net/codel.h
  30 +++ b/include/net/codel.h
  31 @@ -205,7 +205,7 @@ static codel_time_t codel_control_law(co
  32
  33
  34  static bool codel_should_drop(const struct sk_buff *skb,
  35 -                             unsigned int *backlog,
  36 +                             struct Qdisc *sch,
  37                               struct codel_vars *vars,
  38                               struct codel_params *params,
  39                               struct codel_stats *stats,
  40 @@ -219,13 +219,13 @@ static bool codel_should_drop(const stru
  41         }
  42
  43         vars->ldelay = now - codel_get_enqueue_time(skb);
  44 -       *backlog -= qdisc_pkt_len(skb);
  45 +       sch->qstats.backlog -= qdisc_pkt_len(skb);
  46
  47         if (unlikely(qdisc_pkt_len(skb) > stats->maxpacket))
  48                 stats->maxpacket = qdisc_pkt_len(skb);
  49
  50         if (codel_time_before(vars->ldelay, params->target) ||
  51 -           *backlog <= stats->maxpacket) {
  52 +           sch->qstats.backlog <= stats->maxpacket) {
  53                 /* went below - stay below for at least interval */
  54                 vars->first_above_time = 0;
  55                 return false;
  56 @@ -249,8 +249,7 @@ static struct sk_buff *codel_dequeue(str
  57                                      struct codel_params *params,
  58                                      struct codel_vars *vars,
  59                                      struct codel_stats *stats,
  60 -                                    codel_skb_dequeue_t dequeue_func,
  61 -                                    u32 *backlog)
  62 +                                    codel_skb_dequeue_t dequeue_func)
  63  {
  64         struct sk_buff *skb = dequeue_func(vars, sch);
  65         codel_time_t now;
  66 @@ -261,7 +260,7 @@ static struct sk_buff *codel_dequeue(str
  67                 return skb;
  68         }
  69         now = codel_get_time();
  70 -       drop = codel_should_drop(skb, backlog, vars, params, stats, now);
  71 +       drop = codel_should_drop(skb, sch, vars, params, stats, now);
  72         if (vars->dropping) {
  73                 if (!drop) {
  74                         /* sojourn time below target - leave dropping state */
  75 @@ -292,7 +291,7 @@ static struct sk_buff *codel_dequeue(str
  76                                 qdisc_drop(skb, sch);
  77                                 stats->drop_count++;
  78                                 skb = dequeue_func(vars, sch);
  79 -                               if (!codel_should_drop(skb, backlog,
  80 +                               if (!codel_should_drop(skb, sch,
  81                                                        vars, params, stats, now)) {
  82                                         /* leave dropping state */
  83                                         vars->dropping = false;
  84 @@ -313,7 +312,7 @@ static struct sk_buff *codel_dequeue(str
  85                         stats->drop_count++;
  86
  87                         skb = dequeue_func(vars, sch);
  88 -                       drop = codel_should_drop(skb, backlog, vars, params,
  89 +                       drop = codel_should_drop(skb, sch, vars, params,
  90                                                  stats, now);
  91                 }
  92                 vars->dropping = true;
  93 --- a/net/sched/sch_codel.c
  94 +++ b/net/sched/sch_codel.c
  95 @@ -77,8 +77,8 @@ static struct sk_buff *codel_qdisc_deque
  96         struct codel_sched_data *q = qdisc_priv(sch);
  97         struct sk_buff *skb;
  98
  99 -       skb = codel_dequeue(sch, &q->params, &q->vars, &q->stats,
 100 -                           dequeue, &sch->qstats.backlog);
 101 +       skb = codel_dequeue(sch, &q->params, &q->vars, &q->stats, dequeue);
 102 +
 103         /* We cant call qdisc_tree_decrease_qlen() if our qlen is 0,
 104          * or HTB crashes. Defer it for next round.
 105          */
 106 --- a/net/sched/sch_fq_codel.c
 107 +++ b/net/sched/sch_fq_codel.c
 108 @@ -217,13 +217,14 @@ static int fq_codel_enqueue(struct sk_bu
 109   */
 110  static struct sk_buff *dequeue(struct codel_vars *vars, struct Qdisc *sch)
 111  {
 112 +       struct fq_codel_sched_data *q = qdisc_priv(sch);
 113         struct fq_codel_flow *flow;
 114         struct sk_buff *skb = NULL;
 115
 116         flow = container_of(vars, struct fq_codel_flow, cvars);
 117         if (flow->head) {
 118                 skb = dequeue_head(flow);
 119 -               sch->qstats.backlog -= qdisc_pkt_len(skb);
 120 +               q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
 121                 sch->q.qlen--;
 122         }
 123         return skb;
 124 @@ -256,7 +257,7 @@ begin:
 125         prev_ecn_mark = q->cstats.ecn_mark;
 126
 127         skb = codel_dequeue(sch, &q->cparams, &flow->cvars, &q->cstats,
 128 -                           dequeue, &q->backlogs[flow - q->flows]);
 129 +                           dequeue);
 130
 131         flow->dropped += q->cstats.drop_count - prev_drop_count;
 132         flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;