forcedeth: add interrupt moderation logic
This patch adds the logic to moderate the interrupts by changing the
mode between throughput and poll. If there has been a large amount of
time without any burst of network load, the code will transition to pure
throughput mode (where each tx/rx/other will cause an interrupt). If
bursts of network load occurs, it will transition to poll based mode to
help reduce cpu utilization (it will not interrupt on each packet) while
maintaining the optimum network bandwidth utilization.
Signed-off-by: Ayaz Abdulla <aabdulla@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/drivers/net/forcedeth.c b/drivers/net/forcedeth.c
index 2c3a7f8..341a351 100644
--- a/drivers/net/forcedeth.c
+++ b/drivers/net/forcedeth.c
@@ -593,6 +593,9 @@
#define NV_TX_LIMIT_COUNT 16
+#define NV_DYNAMIC_THRESHOLD 4
+#define NV_DYNAMIC_MAX_QUIET_COUNT 2048
+
/* statistics */
struct nv_ethtool_str {
char name[ETH_GSTRING_LEN];
@@ -750,6 +753,7 @@
u16 gigabit;
int intr_test;
int recover_error;
+ int quiet_count;
/* General data: RO fields */
dma_addr_t ring_addr;
@@ -832,7 +836,7 @@
* Maximum number of loops until we assume that a bit in the irq mask
* is stuck. Overridable with module param.
*/
-static int max_interrupt_work = 15;
+static int max_interrupt_work = 4;
/*
* Optimization can be either throuput mode or cpu mode
@@ -3418,11 +3422,43 @@
}
}
+static inline int nv_change_interrupt_mode(struct net_device *dev, int total_work)
+{
+ struct fe_priv *np = netdev_priv(dev);
+
+ if (optimization_mode == NV_OPTIMIZATION_MODE_DYNAMIC) {
+ if (total_work > NV_DYNAMIC_THRESHOLD) {
+ /* transition to poll based interrupts */
+ np->quiet_count = 0;
+ if (np->irqmask != NVREG_IRQMASK_CPU) {
+ np->irqmask = NVREG_IRQMASK_CPU;
+ return 1;
+ }
+ } else {
+ if (np->quiet_count < NV_DYNAMIC_MAX_QUIET_COUNT) {
+ np->quiet_count++;
+ } else {
+ /* reached a period of low activity, switch
+ to per tx/rx packet interrupts */
+ if (np->irqmask != NVREG_IRQMASK_THROUGHPUT) {
+ np->irqmask = NVREG_IRQMASK_THROUGHPUT;
+ return 1;
+ }
+ }
+ }
+ }
+ return 0;
+}
+
static irqreturn_t nv_nic_irq(int foo, void *data)
{
struct net_device *dev = (struct net_device *) data;
struct fe_priv *np = netdev_priv(dev);
u8 __iomem *base = get_hwbase(dev);
+#ifndef CONFIG_FORCEDETH_NAPI
+ int total_work = 0;
+ int loop_count = 0;
+#endif
dprintk(KERN_DEBUG "%s: nv_nic_irq\n", dev->name);
@@ -3449,19 +3485,35 @@
spin_unlock(&np->lock);
- return IRQ_HANDLED;
#else
- spin_lock(&np->lock);
- nv_tx_done(dev, np->tx_ring_size);
- spin_unlock(&np->lock);
-
- if (nv_rx_process(dev, RX_WORK_PER_LOOP)) {
- if (unlikely(nv_alloc_rx(dev))) {
- spin_lock(&np->lock);
- if (!np->in_shutdown)
- mod_timer(&np->oom_kick, jiffies + OOM_REFILL);
- spin_unlock(&np->lock);
+ do
+ {
+ int work = 0;
+ if ((work = nv_rx_process(dev, RX_WORK_PER_LOOP))) {
+ if (unlikely(nv_alloc_rx(dev))) {
+ spin_lock(&np->lock);
+ if (!np->in_shutdown)
+ mod_timer(&np->oom_kick, jiffies + OOM_REFILL);
+ spin_unlock(&np->lock);
+ }
}
+
+ spin_lock(&np->lock);
+ work += nv_tx_done(dev, TX_WORK_PER_LOOP);
+ spin_unlock(&np->lock);
+
+ if (!work)
+ break;
+
+ total_work += work;
+
+ loop_count++;
+ }
+ while (loop_count < max_interrupt_work);
+
+ if (nv_change_interrupt_mode(dev, total_work)) {
+ /* setup new irq mask */
+ writel(np->irqmask, base + NvRegIrqMask);
}
if (unlikely(np->events & NVREG_IRQ_LINK)) {
@@ -3507,6 +3559,10 @@
struct net_device *dev = (struct net_device *) data;
struct fe_priv *np = netdev_priv(dev);
u8 __iomem *base = get_hwbase(dev);
+#ifndef CONFIG_FORCEDETH_NAPI
+ int total_work = 0;
+ int loop_count = 0;
+#endif
dprintk(KERN_DEBUG "%s: nv_nic_irq_optimized\n", dev->name);
@@ -3533,19 +3589,35 @@
spin_unlock(&np->lock);
- return IRQ_HANDLED;
#else
- spin_lock(&np->lock);
- nv_tx_done_optimized(dev, TX_WORK_PER_LOOP);
- spin_unlock(&np->lock);
-
- if (nv_rx_process_optimized(dev, RX_WORK_PER_LOOP)) {
- if (unlikely(nv_alloc_rx_optimized(dev))) {
- spin_lock(&np->lock);
- if (!np->in_shutdown)
- mod_timer(&np->oom_kick, jiffies + OOM_REFILL);
- spin_unlock(&np->lock);
+ do
+ {
+ int work = 0;
+ if ((work = nv_rx_process_optimized(dev, RX_WORK_PER_LOOP))) {
+ if (unlikely(nv_alloc_rx_optimized(dev))) {
+ spin_lock(&np->lock);
+ if (!np->in_shutdown)
+ mod_timer(&np->oom_kick, jiffies + OOM_REFILL);
+ spin_unlock(&np->lock);
+ }
}
+
+ spin_lock(&np->lock);
+ work += nv_tx_done_optimized(dev, TX_WORK_PER_LOOP);
+ spin_unlock(&np->lock);
+
+ if (!work)
+ break;
+
+ total_work += work;
+
+ loop_count++;
+ }
+ while (loop_count < max_interrupt_work);
+
+ if (nv_change_interrupt_mode(dev, total_work)) {
+ /* setup new irq mask */
+ writel(np->irqmask, base + NvRegIrqMask);
}
if (unlikely(np->events & NVREG_IRQ_LINK)) {
@@ -3632,21 +3704,22 @@
struct net_device *dev = np->dev;
u8 __iomem *base = get_hwbase(dev);
unsigned long flags;
- int pkts, retcode;
+ int retcode;
+ int tx_work, rx_work;
if (!nv_optimized(np)) {
spin_lock_irqsave(&np->lock, flags);
- nv_tx_done(dev, np->tx_ring_size);
+ tx_work = nv_tx_done(dev, np->tx_ring_size);
spin_unlock_irqrestore(&np->lock, flags);
- pkts = nv_rx_process(dev, budget);
+ rx_work = nv_rx_process(dev, budget);
retcode = nv_alloc_rx(dev);
} else {
spin_lock_irqsave(&np->lock, flags);
- nv_tx_done_optimized(dev, np->tx_ring_size);
+ tx_work = nv_tx_done_optimized(dev, np->tx_ring_size);
spin_unlock_irqrestore(&np->lock, flags);
- pkts = nv_rx_process_optimized(dev, budget);
+ rx_work = nv_rx_process_optimized(dev, budget);
retcode = nv_alloc_rx_optimized(dev);
}
@@ -3657,6 +3730,8 @@
spin_unlock_irqrestore(&np->lock, flags);
}
+ nv_change_interrupt_mode(dev, tx_work + rx_work);
+
if (unlikely(np->events & NVREG_IRQ_LINK)) {
spin_lock_irqsave(&np->lock, flags);
nv_link_irq(dev);
@@ -3677,10 +3752,10 @@
}
spin_unlock_irqrestore(&np->lock, flags);
__napi_complete(napi);
- return pkts;
+ return rx_work;
}
- if (pkts < budget) {
+ if (rx_work < budget) {
/* re-enable interrupts
(msix not enabled in napi) */
spin_lock_irqsave(&np->lock, flags);
@@ -3691,7 +3766,7 @@
spin_unlock_irqrestore(&np->lock, flags);
}
- return pkts;
+ return rx_work;
}
#endif