On Mon, 2015-11-02 at 13:09 -0800, Cong Wang wrote:

(Sorry for the delay)

On Fri, Oct 30, 2015 at 12:30 PM, Eric Dumazet [off-list ref] wrote:

quoted

On Fri, 2015-10-30 at 11:22 -0700, Cong Wang wrote:

quoted

diff --git a/net/sched/sch_sfq.c b/net/sched/sch_sfq.c
index 3abab53..498f0a2 100644
--- a/net/sched/sch_sfq.c
+++ b/net/sched/sch_sfq.c

@@ -346,7 +346,7 @@ static int
 sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 {
      struct sfq_sched_data *q = qdisc_priv(sch);
-     unsigned int hash;
+     unsigned int hash, dropped;
      sfq_index x, qlen;
      struct sfq_slot *slot;
      int uninitialized_var(ret);

@@ -461,7 +461,7 @@ sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
              return NET_XMIT_SUCCESS;

      qlen = slot->qlen;
-     sfq_drop(sch);
+     dropped = sfq_drop(sch);
      /* Return Congestion Notification only if we dropped a packet
       * from this flow.
       */

@@ -469,7 +469,7 @@ sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
              return NET_XMIT_CN;

I believe you missed the NET_XMIT_CN cases.

SFQ can drop a prior packet, and queue current packet.

qdisc_tree_reduce_backlog() wont be called to update parents.

Not sure about other qdisc(s)

Are you saying some qdisc_tree_reduce_backlog() is missing? It could be,
since I don't audit that, if so, I'd do that in a separated patch, because this
patch is supposed to fix existing qdisc_tree_reduce_backlog().

Does this make sense for you?

Well, before your changes, the qdisc_tree_decrease_qlen(sch, 0) would
have been useless, since qdisc_tree_decrease_qlen() does nothing in this
case [1]

But after your changes, we need to change the backlog, even if the qlen
does not change.

Maybe I missed something, but your patch is not really obvious ;)

[1]
void qdisc_tree_decrease_qlen(struct Qdisc *sch, unsigned int n)
{
        ...
        if (n == 0)
                return;