ibmvnic: Fix overflowing firmware/hardware TX queue
authorThomas Falcon <tlfalcon@linux.vnet.ibm.com>
Sun, 5 Mar 2017 18:18:41 +0000 (12:18 -0600)
committerDavid S. Miller <davem@davemloft.net>
Tue, 7 Mar 2017 22:14:30 +0000 (14:14 -0800)
Use a counter to track the number of outstanding transmissions sent
that have not received completions. If the counter reaches the maximum
number of queue entries, stop transmissions on that queue. As we receive
more completions from firmware, wake the queue once the counter reaches
an acceptable level.

This patch prevents hardware/firmware TX queue from filling up and
and generating errors.  Since incorporating this fix, internal testing
has reported that these firmware errors have stopped.

Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/ibm/ibmvnic.c
drivers/net/ethernet/ibm/ibmvnic.h

index 9198e6b..0e87b83 100644 (file)
@@ -705,6 +705,7 @@ static int ibmvnic_xmit(struct sk_buff *skb, struct net_device *netdev)
        u8 *hdrs = (u8 *)&adapter->tx_rx_desc_req;
        struct device *dev = &adapter->vdev->dev;
        struct ibmvnic_tx_buff *tx_buff = NULL;
+       struct ibmvnic_sub_crq_queue *tx_scrq;
        struct ibmvnic_tx_pool *tx_pool;
        unsigned int tx_send_failed = 0;
        unsigned int tx_map_failed = 0;
@@ -724,6 +725,7 @@ static int ibmvnic_xmit(struct sk_buff *skb, struct net_device *netdev)
        int ret = 0;
 
        tx_pool = &adapter->tx_pool[queue_num];
+       tx_scrq = adapter->tx_scrq[queue_num];
        txq = netdev_get_tx_queue(netdev, skb_get_queue_mapping(skb));
        handle_array = (u64 *)((u8 *)(adapter->login_rsp_buf) +
                                   be32_to_cpu(adapter->login_rsp_buf->
@@ -826,6 +828,14 @@ static int ibmvnic_xmit(struct sk_buff *skb, struct net_device *netdev)
                ret = NETDEV_TX_BUSY;
                goto out;
        }
+
+       atomic_inc(&tx_scrq->used);
+
+       if (atomic_read(&tx_scrq->used) >= adapter->req_tx_entries_per_subcrq) {
+               netdev_info(netdev, "Stopping queue %d\n", queue_num);
+               netif_stop_subqueue(netdev, queue_num);
+       }
+
        tx_packets++;
        tx_bytes += skb->len;
        txq->trans_start = jiffies;
@@ -1213,6 +1223,7 @@ static struct ibmvnic_sub_crq_queue *init_sub_crq_queue(struct ibmvnic_adapter
        scrq->adapter = adapter;
        scrq->size = 4 * PAGE_SIZE / sizeof(*scrq->msgs);
        scrq->cur = 0;
+       atomic_set(&scrq->used, 0);
        scrq->rx_skb_top = NULL;
        spin_lock_init(&scrq->lock);
 
@@ -1355,8 +1366,22 @@ restart_loop:
                                                 DMA_TO_DEVICE);
                        }
 
-                       if (txbuff->last_frag)
+                       if (txbuff->last_frag) {
+                               atomic_dec(&scrq->used);
+
+                               if (atomic_read(&scrq->used) <=
+                                   (adapter->req_tx_entries_per_subcrq / 2) &&
+                                   netif_subqueue_stopped(adapter->netdev,
+                                                          txbuff->skb)) {
+                                       netif_wake_subqueue(adapter->netdev,
+                                                           scrq->pool_index);
+                                       netdev_dbg(adapter->netdev,
+                                                  "Started queue %d\n",
+                                                  scrq->pool_index);
+                               }
+
                                dev_kfree_skb_any(txbuff->skb);
+                       }
 
                        adapter->tx_pool[pool].free_map[adapter->tx_pool[pool].
                                                     producer_index] = index;
index 422824f..1993b42 100644 (file)
@@ -863,6 +863,7 @@ struct ibmvnic_sub_crq_queue {
        spinlock_t lock;
        struct sk_buff *rx_skb_top;
        struct ibmvnic_adapter *adapter;
+       atomic_t used;
 };
 
 struct ibmvnic_long_term_buff {