igb: Refactor clean_rx_irq to reduce overhead and improve performance
authorAlexander Duyck <alexander.h.duyck@intel.com>
Fri, 26 Aug 2011 07:43:54 +0000 (07:43 +0000)
committerJeff Kirsher <jeffrey.t.kirsher@intel.com>
Tue, 20 Sep 2011 06:58:20 +0000 (23:58 -0700)
This change is meant to be a general cleanup and performance improvement
for clean_rx_irq.  The previous patch should have updated the allocation so
that the rings can be treated as read-only within the clean_rx_irq
function.  In addition I am re-ordering the operations such that several
goals are accomplished including reducing the overhead for packet
accounting, reducing the number of items on the stack, and improving
overall performance.

Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
drivers/net/ethernet/intel/igb/igb_main.c

index 9fa2ad01c6b7e9e4ec9059fee938eea86bf4fd4a..dd85df0ed7f24139271fc9f1d4e822b665c0d90d 100644 (file)
@@ -138,7 +138,7 @@ static void igb_setup_dca(struct igb_adapter *);
 #endif /* CONFIG_IGB_DCA */
 static bool igb_clean_tx_irq(struct igb_q_vector *);
 static int igb_poll(struct napi_struct *, int);
-static bool igb_clean_rx_irq_adv(struct igb_q_vector *, int *, int);
+static bool igb_clean_rx_irq_adv(struct igb_q_vector *, int);
 static int igb_ioctl(struct net_device *, struct ifreq *, int cmd);
 static void igb_tx_timeout(struct net_device *);
 static void igb_reset_task(struct work_struct *);
@@ -5481,28 +5481,27 @@ static int igb_poll(struct napi_struct *napi, int budget)
        struct igb_q_vector *q_vector = container_of(napi,
                                                     struct igb_q_vector,
                                                     napi);
-       int tx_clean_complete = 1, work_done = 0;
+       bool clean_complete = true;
 
 #ifdef CONFIG_IGB_DCA
        if (q_vector->adapter->flags & IGB_FLAG_DCA_ENABLED)
                igb_update_dca(q_vector);
 #endif
        if (q_vector->tx_ring)
-               tx_clean_complete = igb_clean_tx_irq(q_vector);
+               clean_complete = !!igb_clean_tx_irq(q_vector);
 
        if (q_vector->rx_ring)
-               igb_clean_rx_irq_adv(q_vector, &work_done, budget);
+               clean_complete &= igb_clean_rx_irq_adv(q_vector, budget);
 
-       if (!tx_clean_complete)
-               work_done = budget;
+       /* If all work not completed, return budget and keep polling */
+       if (!clean_complete)
+               return budget;
 
        /* If not enough Rx work done, exit the polling mode */
-       if (work_done < budget) {
-               napi_complete(napi);
-               igb_ring_irq_enable(q_vector);
-       }
+       napi_complete(napi);
+       igb_ring_irq_enable(q_vector);
 
-       return work_done;
+       return 0;
 }
 
 /**
@@ -5751,37 +5750,26 @@ static inline u16 igb_get_hlen(union e1000_adv_rx_desc *rx_desc)
        return hlen;
 }
 
-static bool igb_clean_rx_irq_adv(struct igb_q_vector *q_vector,
-                                 int *work_done, int budget)
+static bool igb_clean_rx_irq_adv(struct igb_q_vector *q_vector, int budget)
 {
        struct igb_ring *rx_ring = q_vector->rx_ring;
-       struct net_device *netdev = rx_ring->netdev;
-       struct device *dev = rx_ring->dev;
-       union e1000_adv_rx_desc *rx_desc , *next_rxd;
-       struct igb_buffer *buffer_info , *next_buffer;
-       struct sk_buff *skb;
-       bool cleaned = false;
-       u16 cleaned_count = igb_desc_unused(rx_ring);
-       int current_node = numa_node_id();
+       union e1000_adv_rx_desc *rx_desc;
+       const int current_node = numa_node_id();
        unsigned int total_bytes = 0, total_packets = 0;
-       unsigned int i;
        u32 staterr;
-       u16 length;
+       u16 cleaned_count = igb_desc_unused(rx_ring);
+       u16 i = rx_ring->next_to_clean;
 
-       i = rx_ring->next_to_clean;
-       buffer_info = &rx_ring->buffer_info[i];
        rx_desc = E1000_RX_DESC_ADV(*rx_ring, i);
        staterr = le32_to_cpu(rx_desc->wb.upper.status_error);
 
        while (staterr & E1000_RXD_STAT_DD) {
-               if (*work_done >= budget)
-                       break;
-               (*work_done)++;
-               rmb(); /* read descriptor and rx_buffer_info after status DD */
+               struct igb_buffer *buffer_info = &rx_ring->buffer_info[i];
+               struct sk_buff *skb = buffer_info->skb;
+               union e1000_adv_rx_desc *next_rxd;
 
-               skb = buffer_info->skb;
-               prefetch(skb->data - NET_IP_ALIGN);
                buffer_info->skb = NULL;
+               prefetch(skb->data);
 
                i++;
                if (i == rx_ring->count)
@@ -5789,42 +5777,48 @@ static bool igb_clean_rx_irq_adv(struct igb_q_vector *q_vector,
 
                next_rxd = E1000_RX_DESC_ADV(*rx_ring, i);
                prefetch(next_rxd);
-               next_buffer = &rx_ring->buffer_info[i];
 
-               length = le16_to_cpu(rx_desc->wb.upper.length);
-               cleaned = true;
-               cleaned_count++;
+               /*
+                * This memory barrier is needed to keep us from reading
+                * any other fields out of the rx_desc until we know the
+                * RXD_STAT_DD bit is set
+                */
+               rmb();
 
-               if (buffer_info->dma) {
-                       dma_unmap_single(dev, buffer_info->dma,
+               if (!skb_is_nonlinear(skb)) {
+                       __skb_put(skb, igb_get_hlen(rx_desc));
+                       dma_unmap_single(rx_ring->dev, buffer_info->dma,
                                         IGB_RX_HDR_LEN,
                                         DMA_FROM_DEVICE);
                        buffer_info->dma = 0;
-                       skb_put(skb, igb_get_hlen(rx_desc));
                }
 
-               if (length) {
-                       dma_unmap_page(dev, buffer_info->page_dma,
-                                      PAGE_SIZE / 2, DMA_FROM_DEVICE);
-                       buffer_info->page_dma = 0;
+               if (rx_desc->wb.upper.length) {
+                       u16 length = le16_to_cpu(rx_desc->wb.upper.length);
 
                        skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
                                                buffer_info->page,
                                                buffer_info->page_offset,
                                                length);
 
+                       skb->len += length;
+                       skb->data_len += length;
+                       skb->truesize += length;
+
                        if ((page_count(buffer_info->page) != 1) ||
                            (page_to_nid(buffer_info->page) != current_node))
                                buffer_info->page = NULL;
                        else
                                get_page(buffer_info->page);
 
-                       skb->len += length;
-                       skb->data_len += length;
-                       skb->truesize += length;
+                       dma_unmap_page(rx_ring->dev, buffer_info->page_dma,
+                                      PAGE_SIZE / 2, DMA_FROM_DEVICE);
+                       buffer_info->page_dma = 0;
                }
 
                if (!(staterr & E1000_RXD_STAT_EOP)) {
+                       struct igb_buffer *next_buffer;
+                       next_buffer = &rx_ring->buffer_info[i];
                        buffer_info->skb = next_buffer->skb;
                        buffer_info->dma = next_buffer->dma;
                        next_buffer->skb = skb;
@@ -5833,7 +5827,7 @@ static bool igb_clean_rx_irq_adv(struct igb_q_vector *q_vector,
                }
 
                if (staterr & E1000_RXDEXT_ERR_FRAME_ERR_MASK) {
-                       dev_kfree_skb_irq(skb);
+                       dev_kfree_skb_any(skb);
                        goto next_desc;
                }
 
@@ -5844,7 +5838,7 @@ static bool igb_clean_rx_irq_adv(struct igb_q_vector *q_vector,
 
                igb_rx_checksum_adv(rx_ring, staterr, skb);
 
-               skb->protocol = eth_type_trans(skb, netdev);
+               skb->protocol = eth_type_trans(skb, rx_ring->netdev);
 
                if (staterr & E1000_RXD_STAT_VP) {
                        u16 vid = le16_to_cpu(rx_desc->wb.upper.vlan);
@@ -5853,7 +5847,12 @@ static bool igb_clean_rx_irq_adv(struct igb_q_vector *q_vector,
                }
                napi_gro_receive(&q_vector->napi, skb);
 
+               budget--;
 next_desc:
+               if (!budget)
+                       break;
+
+               cleaned_count++;
                /* return some buffers to hardware, one at a time is too slow */
                if (cleaned_count >= IGB_RX_BUFFER_WRITE) {
                        igb_alloc_rx_buffers_adv(rx_ring, cleaned_count);
@@ -5862,7 +5861,6 @@ next_desc:
 
                /* use prefetched values */
                rx_desc = next_rxd;
-               buffer_info = next_buffer;
                staterr = le32_to_cpu(rx_desc->wb.upper.status_error);
        }
 
@@ -5877,7 +5875,7 @@ next_desc:
        if (cleaned_count)
                igb_alloc_rx_buffers_adv(rx_ring, cleaned_count);
 
-       return cleaned;
+       return !!budget;
 }
 
 static bool igb_alloc_mapped_skb(struct igb_ring *rx_ring,