netdev: netdev_send accepts multiple packets

The netdev_send function has been modified to accept multiple packets, to
allow netdev providers to amortize locking and queuing costs.
This is especially true for netdev-dpdk.

Later commits exploit the new API.

Signed-off-by: Daniele Di Proietto <ddiproietto@vmware.com>
Acked-by: Pravin B Shelar <pshelar@nicira.com>

lib/dpif-netdev.c

@@ -2123,7 +2123,7 @@ dp_execute_cb(void *aux_, struct dpif_packet *packet,
     case OVS_ACTION_ATTR_OUTPUT:
         p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
         if (p) {
-            netdev_send(p->netdev, packet, may_steal);
+            netdev_send(p->netdev, &packet, 1, may_steal);
         }
         break;
 

lib/netdev-bsd.c

@@ -686,14 +686,13 @@ netdev_bsd_rxq_drain(struct netdev_rxq *rxq_)
  * system or a tap device.
  */
 static int
-netdev_bsd_send(struct netdev *netdev_, struct dpif_packet *pkt,
+netdev_bsd_send(struct netdev *netdev_, struct dpif_packet **pkts, int cnt,
                 bool may_steal)
 {
     struct netdev_bsd *dev = netdev_bsd_cast(netdev_);
     const char *name = netdev_get_name(netdev_);
-    const void *data = ofpbuf_data(&pkt->ofpbuf);
-    size_t size = ofpbuf_size(&pkt->ofpbuf);
     int error;
+    int i;
 
     ovs_mutex_lock(&dev->mutex);
     if (dev->tap_fd < 0 && !dev->pcap) {
@@ -702,6 +701,10 @@ netdev_bsd_send(struct netdev *netdev_, struct dpif_packet *pkt,
         error = 0;
     }
 
+    for (i = 0; i < cnt; i++) {
+        const void *data = ofpbuf_data(&pkts[i]->ofpbuf);
+        size_t size = ofpbuf_size(&pkts[i]->ofpbuf);
+
         while (!error) {
             ssize_t retval;
             if (dev->tap_fd >= 0) {
@@ -715,22 +718,26 @@ netdev_bsd_send(struct netdev *netdev_, struct dpif_packet *pkt,
                 } else {
                     error = errno;
                     if (error != EAGAIN) {
-                    VLOG_WARN_RL(&rl, "error sending Ethernet packet on %s: "
+                        VLOG_WARN_RL(&rl, "error sending Ethernet packet on"
-                                 "%s", name, ovs_strerror(error));
+                                     " %s: %s", name, ovs_strerror(error));
                     }
                 }
             } else if (retval != size) {
-            VLOG_WARN_RL(&rl, "sent partial Ethernet packet (%"PRIuSIZE" bytes of "
+                VLOG_WARN_RL(&rl, "sent partial Ethernet packet "
+                                  "(%"PRIuSIZE" bytes of "
                                   "%"PRIuSIZE") on %s", retval, size, name);
                 error = EMSGSIZE;
             } else {
                 break;
             }
         }
+    }
 
     ovs_mutex_unlock(&dev->mutex);
     if (may_steal) {
-        dpif_packet_delete(pkt);
+        for (i = 0; i < cnt; i++) {
+            dpif_packet_delete(pkts[i]);
+        }
     }
 
     return error;

lib/netdev-dpdk.c

@@ -206,8 +206,8 @@ dpdk_rte_mzalloc(size_t sz)
 void
 free_dpdk_buf(struct dpif_packet *p)
 {
-    struct ofpbuf *b = &p->ofpbuf;
+    struct ofpbuf *ofp = &p->ofpbuf;
-    struct rte_mbuf *pkt = (struct rte_mbuf *) b->dpdk_buf;
+    struct rte_mbuf *pkt = (struct rte_mbuf *) ofp->dpdk_buf;
 
     rte_mempool_put(pkt->pool, pkt);
 }
@@ -612,16 +612,27 @@ netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dpif_packet **packets,
 }
 
 inline static void
-dpdk_queue_pkt(struct netdev_dpdk *dev, int qid,
+dpdk_queue_pkts(struct netdev_dpdk *dev, int qid,
-               struct rte_mbuf *pkt)
+               struct rte_mbuf **pkts, int cnt)
 {
     struct dpdk_tx_queue *txq = &dev->tx_q[qid];
     uint64_t diff_tsc;
     uint64_t cur_tsc;
     uint32_t nb_tx;
 
+    int i = 0;
+
     rte_spinlock_lock(&txq->tx_lock);
-    txq->burst_pkts[txq->count++] = pkt;
+    while (i < cnt) {
+        int freeslots = MAX_TX_QUEUE_LEN - txq->count;
+        int tocopy = MIN(freeslots, cnt-i);
+
+        memcpy(&txq->burst_pkts[txq->count], &pkts[i],
+               tocopy * sizeof (struct rte_mbuf *));
+
+        txq->count += tocopy;
+        i += tocopy;
+
         if (txq->count == MAX_TX_QUEUE_LEN) {
             goto flush;
         }
@@ -633,11 +644,11 @@ dpdk_queue_pkt(struct netdev_dpdk *dev, int qid,
         if (diff_tsc >= DRAIN_TSC) {
             goto flush;
         }
-    rte_spinlock_unlock(&txq->tx_lock);
+        continue;
-    return;
 
     flush:
-    nb_tx = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts, txq->count);
+        nb_tx = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts,
+                                 txq->count);
         if (nb_tx != txq->count) {
             /* free buffers if we couldn't transmit packets */
             rte_mempool_put_bulk(dev->dpdk_mp->mp,
@@ -645,18 +656,34 @@ flush:
                                  (txq->count - nb_tx));
         }
         txq->count = 0;
+    }
     rte_spinlock_unlock(&txq->tx_lock);
 }
 
 /* Tx function. Transmit packets indefinitely */
 static void
-dpdk_do_tx_copy(struct netdev *netdev, char *buf, int size)
+dpdk_do_tx_copy(struct netdev *netdev, struct dpif_packet ** pkts, int cnt)
 {
     struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
-    struct rte_mbuf *pkt;
+    struct rte_mbuf *mbufs[cnt];
+    int i, newcnt = 0;
 
-    pkt = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
+    for (i = 0; i < cnt; i++) {
-    if (!pkt) {
+        int size = ofpbuf_size(&pkts[i]->ofpbuf);
+        if (size > dev->max_packet_len) {
+            VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
+                         (int)size , dev->max_packet_len);
+
+            ovs_mutex_lock(&dev->mutex);
+            dev->stats.tx_dropped++;
+            ovs_mutex_unlock(&dev->mutex);
+
+            continue;
+        }
+
+        mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
+
+        if (!mbufs[newcnt]) {
             ovs_mutex_lock(&dev->mutex);
             dev->stats.tx_dropped++;
             ovs_mutex_unlock(&dev->mutex);
@@ -664,52 +691,72 @@ dpdk_do_tx_copy(struct netdev *netdev, char *buf, int size)
         }
 
         /* We have to do a copy for now */
-    memcpy(pkt->pkt.data, buf, size);
+        memcpy(mbufs[newcnt]->pkt.data, ofpbuf_data(&pkts[i]->ofpbuf), size);
 
-    rte_pktmbuf_data_len(pkt) = size;
+        rte_pktmbuf_data_len(mbufs[newcnt]) = size;
-    rte_pktmbuf_pkt_len(pkt) = size;
+        rte_pktmbuf_pkt_len(mbufs[newcnt]) = size;
 
-    dpdk_queue_pkt(dev, NON_PMD_THREAD_TX_QUEUE, pkt);
+        newcnt++;
+    }
+
+    dpdk_queue_pkts(dev, NON_PMD_THREAD_TX_QUEUE, mbufs, newcnt);
     dpdk_queue_flush(dev, NON_PMD_THREAD_TX_QUEUE);
 }
 
 static int
-netdev_dpdk_send(struct netdev *netdev,
+netdev_dpdk_send(struct netdev *netdev, struct dpif_packet **pkts, int cnt,
-                 struct dpif_packet *packet, bool may_steal)
+                 bool may_steal)
 {
     struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
-    struct ofpbuf *ofpbuf = &packet->ofpbuf;
     int ret;
+    int i;
 
-    if (ofpbuf_size(ofpbuf) > dev->max_packet_len) {
+    if (!may_steal || pkts[0]->ofpbuf.source != OFPBUF_DPDK) {
-        VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
+        dpdk_do_tx_copy(netdev, pkts, cnt);
-                     (int)ofpbuf_size(ofpbuf) , dev->max_packet_len);
-
-        ovs_mutex_lock(&dev->mutex);
-        dev->stats.tx_dropped++;
-        ovs_mutex_unlock(&dev->mutex);
-
-        ret = E2BIG;
-        goto out;
-    }
-
-    if (!may_steal || ofpbuf->source != OFPBUF_DPDK) {
-        dpdk_do_tx_copy(netdev, (char *) ofpbuf_data(ofpbuf), ofpbuf_size(ofpbuf));
 
         if (may_steal) {
-            dpif_packet_delete(packet);
+            for (i = 0; i < cnt; i++) {
+                dpif_packet_delete(pkts[i]);
+            }
         }
     } else {
         int qid;
+        int next_tx_idx = 0;
+        int dropped = 0;
 
         qid = rte_lcore_id() % NR_QUEUE;
 
-        dpdk_queue_pkt(dev, qid, (struct rte_mbuf *)ofpbuf);
+        for (i = 0; i < cnt; i++) {
+            int size = ofpbuf_size(&pkts[i]->ofpbuf);
+            if (OVS_UNLIKELY(size > dev->max_packet_len)) {
+                if (next_tx_idx != i) {
+                    dpdk_queue_pkts(dev, qid,
+                                    (struct rte_mbuf **)&pkts[next_tx_idx],
+                                    i-next_tx_idx);
 
+                    VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d",
+                                 (int)size , dev->max_packet_len);
+
+                    dpif_packet_delete(pkts[i]);
+                    dropped++;
+                }
+                next_tx_idx = i + 1;
+            }
+        }
+        if (next_tx_idx != cnt) {
+           dpdk_queue_pkts(dev, qid,
+                            (struct rte_mbuf **)&pkts[next_tx_idx],
+                            cnt-next_tx_idx);
+        }
+
+        if (OVS_UNLIKELY(dropped)) {
+            ovs_mutex_lock(&dev->mutex);
+            dev->stats.tx_dropped += dropped;
+            ovs_mutex_unlock(&dev->mutex);
+        }
     }
     ret = 0;
 
-out:
     return ret;
 }
 

lib/netdev-dummy.c

@@ -846,15 +846,20 @@ netdev_dummy_rxq_drain(struct netdev_rxq *rxq_)
 }
 
 static int
-netdev_dummy_send(struct netdev *netdev, struct dpif_packet *pkt,
+netdev_dummy_send(struct netdev *netdev, struct dpif_packet **pkts, int cnt,
                   bool may_steal)
 {
     struct netdev_dummy *dev = netdev_dummy_cast(netdev);
-    const void *buffer = ofpbuf_data(&pkt->ofpbuf);
+    int error = 0;
-    size_t size = ofpbuf_size(&pkt->ofpbuf);
+    int i;
+
+    for (i = 0; i < cnt; i++) {
+        const void *buffer = ofpbuf_data(&pkts[i]->ofpbuf);
+        size_t size = ofpbuf_size(&pkts[i]->ofpbuf);
 
         if (size < ETH_HEADER_LEN) {
-        return EMSGSIZE;
+            error = EMSGSIZE;
+            break;
         } else {
             const struct eth_header *eth = buffer;
             int max_size;
@@ -867,7 +872,8 @@ netdev_dummy_send(struct netdev *netdev, struct dpif_packet *pkt,
                 max_size += VLAN_HEADER_LEN;
             }
             if (size > max_size) {
-            return EMSGSIZE;
+                error = EMSGSIZE;
+                break;
             }
         }
 
@@ -886,11 +892,15 @@ netdev_dummy_send(struct netdev *netdev, struct dpif_packet *pkt,
         }
 
         ovs_mutex_unlock(&dev->mutex);
-    if (may_steal) {
-        dpif_packet_delete(pkt);
     }
 
-    return 0;
+    if (may_steal) {
+        for (i = 0; i < cnt; i++) {
+            dpif_packet_delete(pkts[i]);
+        }
+    }
+
+    return error;
 }
 
 static int

lib/netdev-linux.c

@@ -1012,7 +1012,7 @@ netdev_linux_rxq_recv(struct netdev_rxq *rxq_, struct dpif_packet **packets,
             VLOG_WARN_RL(&rl, "error receiving Ethernet packet on %s: %s",
                          ovs_strerror(errno), netdev_rxq_get_name(rxq_));
         }
-        ofpbuf_delete(buffer);
+        dpif_packet_delete(packet);
     } else {
         dp_packet_pad(buffer);
         packets[0] = packet;
@@ -1057,13 +1057,16 @@ netdev_linux_rxq_drain(struct netdev_rxq *rxq_)
  * The kernel maintains a packet transmission queue, so the caller is not
  * expected to do additional queuing of packets. */
 static int
-netdev_linux_send(struct netdev *netdev_, struct dpif_packet *pkt,
+netdev_linux_send(struct netdev *netdev_, struct dpif_packet **pkts, int cnt,
                   bool may_steal)
 {
-    const void *data = ofpbuf_data(&pkt->ofpbuf);
+    int i;
-    size_t size = ofpbuf_size(&pkt->ofpbuf);
+    int error = 0;
 
-    for (;;) {
+    /* 'i' is incremented only if there's no error */
+    for (i = 0; i < cnt;) {
+        const void *data = ofpbuf_data(&pkts[i]->ofpbuf);
+        size_t size = ofpbuf_size(&pkts[i]->ofpbuf);
         ssize_t retval;
 
         if (!is_tap_netdev(netdev_)) {
@@ -1113,31 +1116,41 @@ netdev_linux_send(struct netdev *netdev_, struct dpif_packet *pkt,
             retval = write(netdev->tap_fd, data, size);
         }
 
-        if (may_steal) {
-            dpif_packet_delete(pkt);
-        }
-
         if (retval < 0) {
             /* The Linux AF_PACKET implementation never blocks waiting for room
              * for packets, instead returning ENOBUFS.  Translate this into
              * EAGAIN for the caller. */
-            if (errno == ENOBUFS) {
+            error = errno == ENOBUFS ? EAGAIN : errno;
-                return EAGAIN;
+            if (error == EINTR) {
-            } else if (errno == EINTR) {
+                /* continue without incrementing 'i', i.e. retry this packet */
                 continue;
-            } else if (errno != EAGAIN) {
-                VLOG_WARN_RL(&rl, "error sending Ethernet packet on %s: %s",
-                             netdev_get_name(netdev_), ovs_strerror(errno));
             }
-            return errno;
+            break;
         } else if (retval != size) {
-            VLOG_WARN_RL(&rl, "sent partial Ethernet packet (%"PRIuSIZE" bytes of "
+            VLOG_WARN_RL(&rl, "sent partial Ethernet packet (%"PRIuSIZE" bytes"
-                         "%"PRIuSIZE") on %s", retval, size, netdev_get_name(netdev_));
+                              " of %"PRIuSIZE") on %s", retval, size,
-            return EMSGSIZE;
+                         netdev_get_name(netdev_));
-        } else {
+            error = EMSGSIZE;
-            return 0;
+            break;
+        }
+
+        /* Process the next packet in the batch */
+        i++;
+    }
+
+    if (may_steal) {
+        for (i = 0; i < cnt; i++) {
+            dpif_packet_delete(pkts[i]);
         }
     }
+
+    if (error && error != EAGAIN) {
+            VLOG_WARN_RL(&rl, "error sending Ethernet packet on %s: %s",
+                         netdev_get_name(netdev_), ovs_strerror(error));
+    }
+
+    return error;
+
 }
 
 /* Registers with the poll loop to wake up from the next call to poll_block()

lib/netdev-provider.h

@@ -250,13 +250,16 @@ struct netdev_class {
     const struct netdev_tunnel_config *
         (*get_tunnel_config)(const struct netdev *netdev);
 
-    /* Sends the buffer on 'netdev'.
+    /* Sends buffers on 'netdev'.
-     * Returns 0 if successful, otherwise a positive errno value.  Returns
+     * Returns 0 if successful (for every buffer), otherwise a positive errno value.
-     * EAGAIN without blocking if the packet cannot be queued immediately.
+     * Returns EAGAIN without blocking if one or more packets cannot be
-     * Returns EMSGSIZE if a partial packet was transmitted or if the packet
+     * queued immediately. Returns EMSGSIZE if a partial packet was transmitted
-     * is too big or too small to transmit on the device.
+     * or if a packet is too big or too small to transmit on the device.
      *
-     * To retain ownership of 'buffer' caller can set may_steal to false.
+     * If the function returns a non-zero value, some of the packets might have
+     * been sent anyway.
+     *
+     * To retain ownership of 'buffers' caller can set may_steal to false.
      *
      * The network device is expected to maintain a packet transmission queue,
      * so that the caller does not ordinarily have to do additional queuing of
@@ -268,7 +271,7 @@ struct netdev_class {
      * network device from being usefully used by the netdev-based "userspace
      * datapath".  It will also prevent the OVS implementation of bonding from
      * working properly over 'netdev'.) */
-    int (*send)(struct netdev *netdev, struct dpif_packet *buffer,
+    int (*send)(struct netdev *netdev, struct dpif_packet **buffers, int cnt,
                 bool may_steal);
 
     /* Registers with the poll loop to wake up from the next call to

lib/netdev.c

@@ -650,10 +650,14 @@ netdev_rxq_drain(struct netdev_rxq *rx)
             : 0);
 }
 
-/* Sends 'buffer' on 'netdev'.  Returns 0 if successful, otherwise a positive
+/* Sends 'buffers' on 'netdev'.  Returns 0 if successful (for every packet),
- * errno value.  Returns EAGAIN without blocking if the packet cannot be queued
+ * otherwise a positive errno value.  Returns EAGAIN without blocking if
- * immediately.  Returns EMSGSIZE if a partial packet was transmitted or if
+ * at least one the packets cannot be queued immediately.  Returns EMSGSIZE
- * the packet is too big or too small to transmit on the device.
+ * if a partial packet was transmitted or if a packet is too big or too small
+ * to transmit on the device.
+ *
+ * If the function returns a non-zero value, some of the packets might have
+ * been sent anyway.
  *
  * To retain ownership of 'buffer' caller can set may_steal to false.
  *
@@ -663,12 +667,13 @@ netdev_rxq_drain(struct netdev_rxq *rx)
  * Some network devices may not implement support for this function.  In such
  * cases this function will always return EOPNOTSUPP. */
 int
-netdev_send(struct netdev *netdev, struct dpif_packet *buffer, bool may_steal)
+netdev_send(struct netdev *netdev, struct dpif_packet **buffers, int cnt,
+            bool may_steal)
 {
     int error;
 
     error = (netdev->netdev_class->send
-             ? netdev->netdev_class->send(netdev, buffer, may_steal)
+             ? netdev->netdev_class->send(netdev, buffers, cnt, may_steal)
              : EOPNOTSUPP);
     if (!error) {
         COVERAGE_INC(netdev_sent);

lib/netdev.h

@@ -173,7 +173,8 @@ void netdev_rxq_wait(struct netdev_rxq *);
 int netdev_rxq_drain(struct netdev_rxq *);
 
 /* Packet transmission. */
-int netdev_send(struct netdev *, struct dpif_packet *, bool may_steal);
+int netdev_send(struct netdev *, struct dpif_packet **, int cnt,
+                bool may_steal);
 void netdev_send_wait(struct netdev *);
 
 /* Hardware address. */