igb: Make Tx budget for NAPI user adjustable
authorAlexander Duyck <alexander.h.duyck@intel.com>
Wed, 5 Oct 2011 13:35:24 +0000 (13:35 +0000)
committerJeff Kirsher <jeffrey.t.kirsher@intel.com>
Fri, 7 Oct 2011 05:46:55 +0000 (22:46 -0700)
This change is to make the NAPI budget limits for transmit
adjustable.  Currently they are only set to 128, and when
the changes/improvements to NAPI occur to allow for adjustability,
it would be possible to tune the value for optimal
performance with applications such as routing.

v2: remove tie between NAPI and interrupt moderation
    fix work limit define name (s/IXGBE/IGB/)
    Update patch description to better reflect patch

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

index b725937856006d66ce830e81626efa8579b620d3..beab918e09abfac45226b0195a230cce3852a834 100644 (file)
@@ -47,6 +47,7 @@ struct igb_adapter;
 
 /* TX/RX descriptor defines */
 #define IGB_DEFAULT_TXD                  256
+#define IGB_DEFAULT_TX_WORK             128
 #define IGB_MIN_TXD                       80
 #define IGB_MAX_TXD                     4096
 
@@ -177,6 +178,7 @@ struct igb_q_vector {
 
        u32 eims_value;
        u16 cpu;
+       u16 tx_work_limit;
 
        u16 itr_val;
        u8 set_itr;
@@ -266,6 +268,7 @@ struct igb_adapter {
        u16 rx_itr;
 
        /* TX */
+       u16 tx_work_limit;
        u32 tx_timeout_count;
        int num_tx_queues;
        struct igb_ring *tx_ring[16];
index f231d82cc6cfe077759059dcef7c7adb614c0bd4..a445c4fbf19e610527fe69a48ff4830a675503a5 100644 (file)
@@ -2011,6 +2011,7 @@ static int igb_set_coalesce(struct net_device *netdev,
 
        for (i = 0; i < adapter->num_q_vectors; i++) {
                struct igb_q_vector *q_vector = adapter->q_vector[i];
+               q_vector->tx_work_limit = adapter->tx_work_limit;
                if (q_vector->rx_ring)
                        q_vector->itr_val = adapter->rx_itr_setting;
                else
index 7ad25e867add8689864d777365647fe6e675c717..12faa99cac53bfd8c14ba409a9f47c2ebf54118d 100644 (file)
@@ -136,8 +136,8 @@ static irqreturn_t igb_msix_ring(int irq, void *);
 static void igb_update_dca(struct igb_q_vector *);
 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_tx_irq(struct igb_q_vector *);
 static bool igb_clean_rx_irq(struct igb_q_vector *, int);
 static int igb_ioctl(struct net_device *, struct ifreq *, int cmd);
 static void igb_tx_timeout(struct net_device *);
@@ -1120,6 +1120,7 @@ static void igb_map_tx_ring_to_vector(struct igb_adapter *adapter,
        q_vector->tx_ring = adapter->tx_ring[ring_idx];
        q_vector->tx_ring->q_vector = q_vector;
        q_vector->itr_val = adapter->tx_itr_setting;
+       q_vector->tx_work_limit = adapter->tx_work_limit;
        if (q_vector->itr_val && q_vector->itr_val <= 3)
                q_vector->itr_val = IGB_START_ITR;
 }
@@ -2388,11 +2389,17 @@ static int __devinit igb_sw_init(struct igb_adapter *adapter)
 
        pci_read_config_word(pdev, PCI_COMMAND, &hw->bus.pci_cmd_word);
 
+       /* set default ring sizes */
        adapter->tx_ring_count = IGB_DEFAULT_TXD;
        adapter->rx_ring_count = IGB_DEFAULT_RXD;
+
+       /* set default ITR values */
        adapter->rx_itr_setting = IGB_DEFAULT_ITR;
        adapter->tx_itr_setting = IGB_DEFAULT_ITR;
 
+       /* set default work limits */
+       adapter->tx_work_limit = IGB_DEFAULT_TX_WORK;
+
        adapter->max_frame_size = netdev->mtu + ETH_HLEN + ETH_FCS_LEN +
                                  VLAN_HLEN;
        adapter->min_frame_size = ETH_ZLEN + ETH_FCS_LEN;
@@ -5496,7 +5503,7 @@ static int igb_poll(struct napi_struct *napi, int budget)
                igb_update_dca(q_vector);
 #endif
        if (q_vector->tx_ring)
-               clean_complete = !!igb_clean_tx_irq(q_vector);
+               clean_complete = igb_clean_tx_irq(q_vector);
 
        if (q_vector->rx_ring)
                clean_complete &= igb_clean_rx_irq(q_vector, budget);
@@ -5578,64 +5585,69 @@ static bool igb_clean_tx_irq(struct igb_q_vector *q_vector)
 {
        struct igb_adapter *adapter = q_vector->adapter;
        struct igb_ring *tx_ring = q_vector->tx_ring;
-       struct net_device *netdev = tx_ring->netdev;
-       struct e1000_hw *hw = &adapter->hw;
-       struct igb_buffer *buffer_info;
-       union e1000_adv_tx_desc *tx_desc, *eop_desc;
+       struct igb_buffer *tx_buffer;
+       union e1000_adv_tx_desc *tx_desc;
        unsigned int total_bytes = 0, total_packets = 0;
-       unsigned int i, eop, count = 0;
-       bool cleaned = false;
+       unsigned int budget = q_vector->tx_work_limit;
+       u16 i = tx_ring->next_to_clean;
 
-       i = tx_ring->next_to_clean;
-       eop = tx_ring->buffer_info[i].next_to_watch;
-       eop_desc = IGB_TX_DESC(tx_ring, eop);
+       if (test_bit(__IGB_DOWN, &adapter->state))
+               return true;
 
-       while ((eop_desc->wb.status & cpu_to_le32(E1000_TXD_STAT_DD)) &&
-              (count < tx_ring->count)) {
-               rmb();  /* read buffer_info after eop_desc status */
-               for (cleaned = false; !cleaned; count++) {
-                       tx_desc = IGB_TX_DESC(tx_ring, i);
-                       buffer_info = &tx_ring->buffer_info[i];
-                       cleaned = (i == eop);
+       tx_buffer = &tx_ring->buffer_info[i];
+       tx_desc = IGB_TX_DESC(tx_ring, i);
 
-                       if (buffer_info->skb) {
-                               total_bytes += buffer_info->bytecount;
-                               /* gso_segs is currently only valid for tcp */
-                               total_packets += buffer_info->gso_segs;
-                               igb_tx_hwtstamp(q_vector, buffer_info);
-                       }
+       for (; budget; budget--) {
+               u16 eop = tx_buffer->next_to_watch;
+               union e1000_adv_tx_desc *eop_desc;
+
+               eop_desc = IGB_TX_DESC(tx_ring, eop);
+
+               /* if DD is not set pending work has not been completed */
+               if (!(eop_desc->wb.status & cpu_to_le32(E1000_TXD_STAT_DD)))
+                       break;
+
+               /* prevent any other reads prior to eop_desc being verified */
+               rmb();
 
-                       igb_unmap_and_free_tx_resource(tx_ring, buffer_info);
+               do {
                        tx_desc->wb.status = 0;
+                       if (likely(tx_desc == eop_desc)) {
+                               eop_desc = NULL;
+
+                               total_bytes += tx_buffer->bytecount;
+                               total_packets += tx_buffer->gso_segs;
+                               igb_tx_hwtstamp(q_vector, tx_buffer);
+                       }
+
+                       igb_unmap_and_free_tx_resource(tx_ring, tx_buffer);
 
+                       tx_buffer++;
+                       tx_desc++;
                        i++;
-                       if (i == tx_ring->count)
+                       if (unlikely(i == tx_ring->count)) {
                                i = 0;
-               }
-               eop = tx_ring->buffer_info[i].next_to_watch;
-               eop_desc = IGB_TX_DESC(tx_ring, eop);
+                               tx_buffer = tx_ring->buffer_info;
+                               tx_desc = IGB_TX_DESC(tx_ring, 0);
+                       }
+               } while (eop_desc);
        }
 
        tx_ring->next_to_clean = i;
+       u64_stats_update_begin(&tx_ring->tx_syncp);
+       tx_ring->tx_stats.bytes += total_bytes;
+       tx_ring->tx_stats.packets += total_packets;
+       u64_stats_update_end(&tx_ring->tx_syncp);
+       tx_ring->total_bytes += total_bytes;
+       tx_ring->total_packets += total_packets;
 
-       if (unlikely(count &&
-                    netif_carrier_ok(netdev) &&
-                    igb_desc_unused(tx_ring) >= IGB_TX_QUEUE_WAKE)) {
-               /* Make sure that anybody stopping the queue after this
-                * sees the new next_to_clean.
-                */
-               smp_mb();
-               if (__netif_subqueue_stopped(netdev, tx_ring->queue_index) &&
-                   !(test_bit(__IGB_DOWN, &adapter->state))) {
-                       netif_wake_subqueue(netdev, tx_ring->queue_index);
+       if (tx_ring->detect_tx_hung) {
+               struct e1000_hw *hw = &adapter->hw;
+               u16 eop = tx_ring->buffer_info[i].next_to_watch;
+               union e1000_adv_tx_desc *eop_desc;
 
-                       u64_stats_update_begin(&tx_ring->tx_syncp);
-                       tx_ring->tx_stats.restart_queue++;
-                       u64_stats_update_end(&tx_ring->tx_syncp);
-               }
-       }
+               eop_desc = IGB_TX_DESC(tx_ring, eop);
 
-       if (tx_ring->detect_tx_hung) {
                /* Detect a transmit hang in hardware, this serializes the
                 * check with the clearing of time_stamp and movement of i */
                tx_ring->detect_tx_hung = false;
@@ -5666,16 +5678,34 @@ static bool igb_clean_tx_irq(struct igb_q_vector *q_vector)
                                eop,
                                jiffies,
                                eop_desc->wb.status);
-                       netif_stop_subqueue(netdev, tx_ring->queue_index);
+                       netif_stop_subqueue(tx_ring->netdev,
+                                           tx_ring->queue_index);
+
+                       /* we are about to reset, no point in enabling stuff */
+                       return true;
                }
        }
-       tx_ring->total_bytes += total_bytes;
-       tx_ring->total_packets += total_packets;
-       u64_stats_update_begin(&tx_ring->tx_syncp);
-       tx_ring->tx_stats.bytes += total_bytes;
-       tx_ring->tx_stats.packets += total_packets;
-       u64_stats_update_end(&tx_ring->tx_syncp);
-       return count < tx_ring->count;
+
+       if (unlikely(total_packets &&
+                    netif_carrier_ok(tx_ring->netdev) &&
+                    igb_desc_unused(tx_ring) >= IGB_TX_QUEUE_WAKE)) {
+               /* Make sure that anybody stopping the queue after this
+                * sees the new next_to_clean.
+                */
+               smp_mb();
+               if (__netif_subqueue_stopped(tx_ring->netdev,
+                                            tx_ring->queue_index) &&
+                   !(test_bit(__IGB_DOWN, &adapter->state))) {
+                       netif_wake_subqueue(tx_ring->netdev,
+                                           tx_ring->queue_index);
+
+                       u64_stats_update_begin(&tx_ring->tx_syncp);
+                       tx_ring->tx_stats.restart_queue++;
+                       u64_stats_update_end(&tx_ring->tx_syncp);
+               }
+       }
+
+       return !!budget;
 }
 
 static inline void igb_rx_checksum(struct igb_ring *ring,