forcedeth: tx max work
authorAyaz Abdulla <aabdulla@nvidia.com>
Tue, 23 Jan 2007 17:00:56 +0000 (12:00 -0500)
committerJeff Garzik <jeff@garzik.org>
Mon, 5 Feb 2007 21:58:48 +0000 (16:58 -0500)
This patch adds a limit to how much tx work can be done in each
iteration of tx processing. If the max limit is reached, remaining tx
completions will be handled by timer interrupt.

Signed-Off-By: Ayaz Abdulla <aabdulla@nvidia.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
drivers/net/forcedeth.c

index 7c0f0ccbbb297b849a753ee0f5855dcf6878842d..ae6cf7827e5793732aa3888e870730bc6d74f2af 100644 (file)
@@ -210,7 +210,7 @@ enum {
  * NVREG_POLL_DEFAULT=97 would result in an interval length of 1 ms
  */
        NvRegPollingInterval = 0x00c,
-#define NVREG_POLL_DEFAULT_THROUGHPUT  970
+#define NVREG_POLL_DEFAULT_THROUGHPUT  970 /* backup tx cleanup if loop max reached */
 #define NVREG_POLL_DEFAULT_CPU 13
        NvRegMSIMap0 = 0x020,
        NvRegMSIMap1 = 0x024,
@@ -1859,14 +1859,15 @@ static void nv_tx_done(struct net_device *dev)
        }
 }
 
-static void nv_tx_done_optimized(struct net_device *dev)
+static void nv_tx_done_optimized(struct net_device *dev, int limit)
 {
        struct fe_priv *np = netdev_priv(dev);
        u32 flags;
        struct ring_desc_ex* orig_get_tx = np->get_tx.ex;
 
        while ((np->get_tx.ex != np->put_tx.ex) &&
-              !((flags = le32_to_cpu(np->get_tx.ex->flaglen)) & NV_TX_VALID)) {
+              !((flags = le32_to_cpu(np->get_tx.ex->flaglen)) & NV_TX_VALID) &&
+              (limit-- > 0)) {
 
                dprintk(KERN_DEBUG "%s: nv_tx_done_optimized: flags 0x%x.\n",
                                        dev->name, flags);
@@ -1973,7 +1974,7 @@ static void nv_tx_timeout(struct net_device *dev)
        if (np->desc_ver == DESC_VER_1 || np->desc_ver == DESC_VER_2)
                nv_tx_done(dev);
        else
-               nv_tx_done_optimized(dev);
+               nv_tx_done_optimized(dev, np->tx_ring_size);
 
        /* 3) if there are dead entries: clear everything */
        if (np->get_tx_ctx != np->put_tx_ctx) {
@@ -2899,7 +2900,7 @@ static irqreturn_t nv_nic_irq_optimized(int foo, void *data)
                        break;
 
                spin_lock(&np->lock);
-               nv_tx_done_optimized(dev);
+               nv_tx_done_optimized(dev, TX_WORK_PER_LOOP);
                spin_unlock(&np->lock);
 
 #ifdef CONFIG_FORCEDETH_NAPI
@@ -3006,7 +3007,7 @@ static irqreturn_t nv_nic_irq_tx(int foo, void *data)
                        break;
 
                spin_lock_irqsave(&np->lock, flags);
-               nv_tx_done_optimized(dev);
+               nv_tx_done_optimized(dev, TX_WORK_PER_LOOP);
                spin_unlock_irqrestore(&np->lock, flags);
 
                if (unlikely(events & (NVREG_IRQ_TX_ERR))) {
@@ -3163,6 +3164,11 @@ static irqreturn_t nv_nic_irq_other(int foo, void *data)
                if (!(events & np->irqmask))
                        break;
 
+               /* check tx in case we reached max loop limit in tx isr */
+               spin_lock_irqsave(&np->lock, flags);
+               nv_tx_done_optimized(dev, TX_WORK_PER_LOOP);
+               spin_unlock_irqrestore(&np->lock, flags);
+
                if (events & NVREG_IRQ_LINK) {
                        spin_lock_irqsave(&np->lock, flags);
                        nv_link_irq(dev);