Merge branch 'cbq-fixes'

Vasily Averin says:

====================
cbq: incorrectly low bandwidth blocks limited traffic

v2: patch description changes
Fixes: f0f6ee1f70c4 ("cbq: incorrect processing of high limits")

Mainstream commit f0f6ee1f70c4 ("cbq: incorrect processing of high limits")
have side effect: if cbq bandwidth setting is less than real interface
throughput non-limited traffic can delay limited traffic for a very long time.

This happen because of q->now changes incorrectly in cbq_dequeue():
in described scenario L2T is much greater than real time delay,
and q->now gets an extra boost for each transmitted packet.

Accumulated boost prevents update q->now, and blocked class can wait
very long time until (q->now >= cl->undertime) will be true again.

More detailed problem description can be found here:
http://www.spinics.net/lists/netdev/msg292493.html

Following patches should fix the problem.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 14 insertions, 34 deletions

diff --git a/net/sched/sch_cbq.c b/net/sched/sch_cbq.c
index ead526467cca..762a04bb8f6d 100644
--- a/net/sched/sch_cbq.c
+++ b/net/sched/sch_cbq.c
@@ -159,7 +159,6 @@ struct cbq_sched_data {
         struct cbq_class        *tx_borrowed;
         int                     tx_len;
         psched_time_t           now;            /* Cached timestamp */
-        psched_time_t           now_rt;         /* Cached real time */
         unsigned int            pmask;
 
         struct hrtimer          delay_timer;
@@ -353,12 +352,7 @@ cbq_mark_toplevel(struct cbq_sched_data *q, struct cbq_class *cl)
         int toplevel = q->toplevel;
 
         if (toplevel > cl->level && !(qdisc_is_throttled(cl->q))) {
-                psched_time_t now;
-                psched_tdiff_t incr;
-
-                now = psched_get_time();
-                incr = now - q->now_rt;
-                now = q->now + incr;
+                psched_time_t now = psched_get_time();
 
                 do {
                         if (cl->undertime < now) {
@@ -700,8 +694,13 @@ cbq_update(struct cbq_sched_data *q)
         struct cbq_class *this = q->tx_class;
         struct cbq_class *cl = this;
         int len = q->tx_len;
+        psched_time_t now;
 
         q->tx_class = NULL;
+        /* Time integrator. We calculate EOS time
+         * by adding expected packet transmission time.
+         */
+        now = q->now + L2T(&q->link, len);
 
         for ( ; cl; cl = cl->share) {
                 long avgidle = cl->avgidle;
@@ -717,7 +716,7 @@ cbq_update(struct cbq_sched_data *q)
                  *      idle = (now - last) - last_pktlen/rate
                  */
 
-                idle = q->now - cl->last;
+                idle = now - cl->last;
                 if ((unsigned long)idle > 128*1024*1024) {
                         avgidle = cl->maxidle;
                 } else {
@@ -761,7 +760,7 @@ cbq_update(struct cbq_sched_data *q)
                         idle -= L2T(&q->link, len);
                         idle += L2T(cl, len);
 
-                        cl->undertime = q->now + idle;
+                        cl->undertime = now + idle;
                 } else {
                         /* Underlimit */
 
@@ -771,7 +770,8 @@ cbq_update(struct cbq_sched_data *q)
                         else
                                 cl->avgidle = avgidle;
                 }
-                cl->last = q->now;
+                if ((s64)(now - cl->last) > 0)
+                        cl->last = now;
         }
 
         cbq_update_toplevel(q, this, q->tx_borrowed);
@@ -943,31 +943,13 @@ cbq_dequeue(struct Qdisc *sch)
         struct sk_buff *skb;
         struct cbq_sched_data *q = qdisc_priv(sch);
         psched_time_t now;
-        psched_tdiff_t incr;
 
         now = psched_get_time();
-        incr = now - q->now_rt;
-
-        if (q->tx_class) {
-                psched_tdiff_t incr2;
-                /* Time integrator. We calculate EOS time
-                 * by adding expected packet transmission time.
-                 * If real time is greater, we warp artificial clock,
-                 * so that:
-                 *
-                 * cbq_time = max(real_time, work);
-                 */
-                incr2 = L2T(&q->link, q->tx_len);
-                q->now += incr2;
+
+        if (q->tx_class)
                 cbq_update(q);
-                if ((incr -= incr2) < 0)
-                        incr = 0;
-                q->now += incr;
-        } else {
-                if (now > q->now)
-                        q->now = now;
-        }
-        q->now_rt = now;
+
+        q->now = now;
 
         for (;;) {
                 q->wd_expires = 0;
@@ -1223,7 +1205,6 @@ cbq_reset(struct Qdisc *sch)
         hrtimer_cancel(&q->delay_timer);
         q->toplevel = TC_CBQ_MAXLEVEL;
         q->now = psched_get_time();
-        q->now_rt = q->now;
 
         for (prio = 0; prio <= TC_CBQ_MAXPRIO; prio++)
                 q->active[prio] = NULL;
@@ -1407,7 +1388,6 @@ static int cbq_init(struct Qdisc *sch, struct nlattr *opt)
         q->delay_timer.function = cbq_undelay;
         q->toplevel = TC_CBQ_MAXLEVEL;
         q->now = psched_get_time();
-        q->now_rt = q->now;
 
         cbq_link_class(&q->link);