chelsio: NAPI speed improvement
authorStephen Hemminger <shemminger@osdl.org>
Mon, 8 Jan 2007 19:26:30 +0000 (11:26 -0800)
committerJeff Garzik <jeff@garzik.org>
Mon, 5 Feb 2007 21:58:46 +0000 (16:58 -0500)
Speedup and cleanup the receive processing by eliminating the
mmio read and a lock round trip.

Signed-off-by: Stephen Hemminger <shemminger@osdl.org>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
drivers/net/chelsio/sge.c

index f94d63971642c518cd4676b2e9eee2786f8b90d2..8e287e79e4e3988c89ece043969fcb4e6258cb07 100644 (file)
@@ -1559,6 +1559,14 @@ static int process_responses(struct adapter *adapter, int budget)
        return budget;
 }
 
+static inline int responses_pending(const struct adapter *adapter)
+{
+       const struct respQ *Q = &adapter->sge->respQ;
+       const struct respQ_e *e = &Q->entries[Q->cidx];
+
+       return (e->GenerationBit == Q->genbit);
+}
+
 #ifdef CONFIG_CHELSIO_T1_NAPI
 /*
  * A simpler version of process_responses() that handles only pure (i.e.,
@@ -1568,13 +1576,16 @@ static int process_responses(struct adapter *adapter, int budget)
  * which the caller must ensure is a valid pure response.  Returns 1 if it
  * encounters a valid data-carrying response, 0 otherwise.
  */
-static int process_pure_responses(struct adapter *adapter, struct respQ_e *e)
+static int process_pure_responses(struct adapter *adapter)
 {
        struct sge *sge = adapter->sge;
        struct respQ *q = &sge->respQ;
+       struct respQ_e *e = &q->entries[q->cidx];
        unsigned int flags = 0;
        unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
 
+       if (e->DataValid)
+               return 1;
        do {
                flags |= e->Qsleeping;
 
@@ -1610,23 +1621,20 @@ static int process_pure_responses(struct adapter *adapter, struct respQ_e *e)
 int t1_poll(struct net_device *dev, int *budget)
 {
        struct adapter *adapter = dev->priv;
-       int effective_budget = min(*budget, dev->quota);
-       int work_done = process_responses(adapter, effective_budget);
+       int work_done;
 
+       work_done = process_responses(adapter, min(*budget, dev->quota));
        *budget -= work_done;
        dev->quota -= work_done;
 
-       if (work_done >= effective_budget)
+       if (unlikely(responses_pending(adapter)))
                return 1;
 
-       spin_lock_irq(&adapter->async_lock);
-       __netif_rx_complete(dev);
+       netif_rx_complete(dev);
        writel(adapter->sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
-       writel(adapter->slow_intr_mask | F_PL_INTR_SGE_DATA,
-              adapter->regs + A_PL_ENABLE);
-       spin_unlock_irq(&adapter->async_lock);
 
        return 0;
+
 }
 
 /*
@@ -1635,44 +1643,33 @@ int t1_poll(struct net_device *dev, int *budget)
 irqreturn_t t1_interrupt(int irq, void *data)
 {
        struct adapter *adapter = data;
-       struct net_device *dev = adapter->sge->netdev;
        struct sge *sge = adapter->sge;
-       u32 cause;
-       int handled = 0;
-
-       cause = readl(adapter->regs + A_PL_CAUSE);
-       if (cause == 0 || cause == ~0)
-               return IRQ_NONE;
+       int handled;
 
-       spin_lock(&adapter->async_lock);
-       if (cause & F_PL_INTR_SGE_DATA) {
-               struct respQ *q = &adapter->sge->respQ;
-               struct respQ_e *e = &q->entries[q->cidx];
+       if (likely(responses_pending(adapter))) {
+               struct net_device *dev = sge->netdev;
 
-               handled = 1;
                writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
 
-               if (e->GenerationBit == q->genbit &&
-                   __netif_rx_schedule_prep(dev)) {
-                       if (e->DataValid || process_pure_responses(adapter, e)) {
-                               /* mask off data IRQ */
-                               writel(adapter->slow_intr_mask,
-                                      adapter->regs + A_PL_ENABLE);
-                               __netif_rx_schedule(sge->netdev);
-                               goto unlock;
+               if (__netif_rx_schedule_prep(dev)) {
+                       if (process_pure_responses(adapter))
+                               __netif_rx_schedule(dev);
+                       else {
+                               /* no data, no NAPI needed */
+                               writel(sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
+                               netif_poll_enable(dev); /* undo schedule_prep */
                        }
-                       /* no data, no NAPI needed */
-                       netif_poll_enable(dev);
-
                }
-               writel(q->cidx, adapter->regs + A_SG_SLEEPING);
-       } else
-               handled = t1_slow_intr_handler(adapter);
+               return IRQ_HANDLED;
+       }
+
+       spin_lock(&adapter->async_lock);
+       handled = t1_slow_intr_handler(adapter);
+       spin_unlock(&adapter->async_lock);
 
        if (!handled)
                sge->stats.unhandled_irqs++;
-unlock:
-       spin_unlock(&adapter->async_lock);
+
        return IRQ_RETVAL(handled != 0);
 }
 
@@ -1695,17 +1692,13 @@ unlock:
 irqreturn_t t1_interrupt(int irq, void *cookie)
 {
        int work_done;
-       struct respQ_e *e;
        struct adapter *adapter = cookie;
-       struct respQ *Q = &adapter->sge->respQ;
 
        spin_lock(&adapter->async_lock);
-       e = &Q->entries[Q->cidx];
-       prefetch(e);
 
        writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
 
-       if (likely(e->GenerationBit == Q->genbit))
+       if (likely(responses_pending(adapter))
                work_done = process_responses(adapter, -1);
        else
                work_done = t1_slow_intr_handler(adapter);