xnu-3247.1.106.tar.gz

[apple/xnu.git] / bsd / net / classq / classq_sfb.c

diff --git a/bsd/net/classq/classq_sfb.c b/bsd/net/classq/classq_sfb.c
index 4f705e0f13c5d13ce87924d639635fc8de75e131..7d12ba606f03add83bc16efb6e39c93ab995c01b 100644 (file)
--- a/bsd/net/classq/classq_sfb.c
+++ b/bsd/net/classq/classq_sfb.c
@@ -280,9 +280,9 @@ static u_int64_t sfb_hinterval;
 SYSCTL_QUAD(_net_classq_sfb, OID_AUTO, hinterval, CTLFLAG_RW|CTLFLAG_LOCKED,
     &sfb_hinterval, "SFB hash interval in nanoseconds");
 
-static u_int64_t sfb_target_qdelay;
+static u_int64_t sfb_target_qdelay = 0;
 SYSCTL_QUAD(_net_classq_sfb, OID_AUTO, target_qdelay, CTLFLAG_RW|CTLFLAG_LOCKED,
-    &sfb_target_qdelay, "SFB target queue delay in milliseconds");
+    &sfb_target_qdelay, "SFB target queue delay in nanoseconds");
 
 static u_int64_t sfb_update_interval;
 SYSCTL_QUAD(_net_classq_sfb, OID_AUTO, update_interval,
@@ -458,6 +458,15 @@ sfb_calc_target_qdelay(struct sfb *sp, u_int64_t out_bw)
        if (target_qdelay == 0)
                target_qdelay = IFQ_TARGET_DELAY;
 
+       /*
+        * If a delay has been added to ifnet start callback for
+        * coalescing, we have to add that to the pre-set target delay
+        * because the packets can be in the queue longer.
+        */
+       if ((ifp->if_eflags & IFEF_ENQUEUE_MULTI) &&
+               ifp->if_start_delay_timeout > 0)
+               target_qdelay += ifp->if_start_delay_timeout;
+
        sp->sfb_target_qdelay = target_qdelay;
 }
 
@@ -1147,8 +1156,7 @@ sfb_addq(struct sfb *sp, class_queue_t *q, struct mbuf *m, struct pf_mtag *t)
        u_int16_t pmin;
        int fc_adv = 0;
        int ret = CLASSQEQ_SUCCESS;
-
-       nanouptime(&now);
+       u_int32_t maxqsize = 0;
 
        s = sp->sfb_current;
        VERIFY((s + (s ^ 1)) == 1);
@@ -1157,6 +1165,13 @@ sfb_addq(struct sfb *sp, class_queue_t *q, struct mbuf *m, struct pf_mtag *t)
        VERIFY(!(pkt->pkt_flags & PKTF_PRIV_GUARDED));
        pkt->pkt_flags |= PKTF_PRIV_GUARDED;
 
+       if (pkt->pkt_enqueue_ts > 0) {
+               net_nsectimer(&pkt->pkt_enqueue_ts, &now); 
+       } else {
+               nanouptime(&now);
+               net_timernsec(&now, &pkt->pkt_enqueue_ts);
+       }
+
        /* time to swap the bins? */
        if (net_timercmp(&now, &sp->sfb_nextreset, >=)) {
                net_timeradd(&now, &sp->sfb_hinterval, &sp->sfb_nextreset);
@@ -1170,6 +1185,13 @@ sfb_addq(struct sfb *sp, class_queue_t *q, struct mbuf *m, struct pf_mtag *t)
                    &sp->sfb_update_time);
        }
 
+       /*
+        * If getq time is not set because this is the first packet
+        * or after idle time, set it now so that we can detect a stall.
+        */
+       if (qsize(q) == 0 && !net_timerisset(&sp->sfb_getqtime))
+               *(&sp->sfb_getqtime) = *(&now);
+
        pkt->pkt_sfb_flags = 0;
        pkt->pkt_sfb_hash16[s] =
            (SFB_HASH(&pkt->pkt_flowid, sizeof (pkt->pkt_flowid),
@@ -1218,25 +1240,33 @@ sfb_addq(struct sfb *sp, class_queue_t *q, struct mbuf *m, struct pf_mtag *t)
                sp->sfb_stats.drop_pbox++;
        }
 
-       /*
-        * if max queue size is static, make it a forced drop
-        * when the queue length hits the queue limit
-        */
-       if (!(SFB_QUEUE_DELAYBASED(sp)) &&
-           droptype == DTYPE_NODROP && qlen(q) >= qlimit(q)) {
-               droptype = DTYPE_FORCED;
-               sp->sfb_stats.drop_queue++;
-       }
+       if (SFB_QUEUE_DELAYBASED(sp))
+               maxqsize = SFB_QUEUE_DELAYBASED_MAXSIZE;
+       else
+               maxqsize = qlimit(q);
 
        /*
-        * delay based queues have a larger maximum size to
-        * allow for bursts
+        * When the queue length hits the queue limit, make it a forced
+        * drop
         */
-       if (SFB_QUEUE_DELAYBASED(sp) &&
-           droptype == DTYPE_NODROP &&
-           qlen(q) >= SFB_QUEUE_DELAYBASED_MAXSIZE) {
-               droptype = DTYPE_FORCED;
-               sp->sfb_stats.drop_queue++;
+       if (droptype == DTYPE_NODROP && qlen(q) >= maxqsize) {
+               if (pkt->pkt_proto == IPPROTO_TCP &&
+                   ((pkt->pkt_flags & PKTF_TCP_REXMT) ||
+                   (sp->sfb_flags & SFBF_LAST_PKT_DROPPED))) {
+                       /*
+                        * At some level, dropping packets will make the
+                        * flows backoff and will keep memory requirements
+                        * under control. But we should not cause a tail
+                        * drop because it can take a long time for a
+                        * TCP flow to recover. We should try to drop
+                        * alternate packets instead.
+                        */
+                       sp->sfb_flags &= ~SFBF_LAST_PKT_DROPPED;
+               } else {
+                       droptype = DTYPE_FORCED;
+                       sp->sfb_stats.drop_queue++;
+                       sp->sfb_flags |= SFBF_LAST_PKT_DROPPED;
+               }
        }
 
        if (fc_adv == 1 && droptype != DTYPE_FORCED &&
@@ -1255,7 +1285,6 @@ sfb_addq(struct sfb *sp, class_queue_t *q, struct mbuf *m, struct pf_mtag *t)
        }
        /* if successful enqueue this packet, else drop it */
        if (droptype == DTYPE_NODROP) {
-               net_timernsec(&now, &pkt->pkt_enqueue_ts);
                _addq(q, m);
        } else {
                IFCQ_CONVERT_LOCK(&sp->sfb_ifp->if_snd);
@@ -1346,6 +1375,7 @@ sfb_getq_flow(struct sfb *sp, class_queue_t *q, u_int32_t flow, boolean_t purge)
                        sp->sfb_min_qdelay = 0;
                }
        }
+       pkt->pkt_enqueue_ts = 0;
 
        /*
         * Clearpkts are the ones which were in the queue when the hash
@@ -1378,6 +1408,7 @@ sfb_getq_flow(struct sfb *sp, class_queue_t *q, u_int32_t flow, boolean_t purge)
                sp->sfb_min_qdelay = 0;
                sp->sfb_fc_threshold = 0;
                net_timerclear(&sp->sfb_update_time);
+               net_timerclear(&sp->sfb_getqtime);
        }
 
        return (m);