I've cleaned up the winbond-840 interrupt handler and added a very
simple interrupt avoidance scheme:
* if more than 1.333*packet_limit packets arrive per second, then switch
off the per-packet rx interrupt, and poll with a 500 us timer.
* if the number of packets falls below packet_limit, switch back to rx
interrupts.
packet limit is right now 2000.
Result:
single stream tcp receive performance up from 9.2 MB/sec to 10.5 MB/sec
(mtu 1500).
In both cases, 0 packet drops (as expected - one tcp streams with a
window of 64 kB and a 64 entry rx queue makes packet drops impossible)
IMHO that shows that packet drops are not the right sign when to switch
on interrupt mitigation - packets aren't dropped, instead the tcp sender
just sends packets slower and performance goes down.
--
Manfred
--- 2.4/drivers/net/winbond-840.c Tue Oct 2 19:58:10 2001
+++ build-2.4/drivers/net/winbond-840.c Sat Oct 6 10:37:06 2001
@@ -63,7 +63,6 @@
These may be modified when a driver module is loaded.*/
static int debug = 1; /* 1 normal messages, 0 quiet .. 7 verbose. */
-static int max_interrupt_work = 20;
/* Maximum number of multicast addresses to filter (vs. Rx-all-multicast).
The '840 uses a 64 element hash table based on the Ethernet CRC. */
static int multicast_filter_limit = 32;
@@ -72,6 +71,7 @@
Setting to > 1518 effectively disables this feature. */
static int rx_copybreak;
+static int packet_limit = 2000;
/* Used to pass the media type, etc.
Both 'options[]' and 'full_duplex[]' should exist for driver
interoperability.
@@ -91,7 +91,7 @@
#define TX_RING_SIZE 16
#define TX_QUEUE_LEN 10 /* Limit ring entries actually used. */
#define TX_QUEUE_LEN_RESTART 5
-#define RX_RING_SIZE 32
+#define RX_RING_SIZE 64
#define TX_BUFLIMIT (1024-128)
@@ -151,15 +151,15 @@
MODULE_DESCRIPTION("Winbond W89c840 Ethernet driver");
MODULE_LICENSE("GPL");
-MODULE_PARM(max_interrupt_work, "i");
MODULE_PARM(debug, "i");
MODULE_PARM(rx_copybreak, "i");
MODULE_PARM(multicast_filter_limit, "i");
+MODULE_PARM(packet_limit, "i");
MODULE_PARM(options, "1-" __MODULE_STRING(MAX_UNITS) "i");
MODULE_PARM(full_duplex, "1-" __MODULE_STRING(MAX_UNITS) "i");
-MODULE_PARM_DESC(max_interrupt_work, "winbond-840 maximum events handled per interrupt");
MODULE_PARM_DESC(debug, "winbond-840 debug level (0-6)");
MODULE_PARM_DESC(rx_copybreak, "winbond-840 copy breakpoint for copy-only-tiny-frames");
+MODULE_PARM_DESC(packet_limit, "winbond-840 static interrupt mitigation");
MODULE_PARM_DESC(multicast_filter_limit, "winbond-840 maximum number of filtered multicast addresses");
MODULE_PARM_DESC(options, "winbond-840: Bits 0-3: media type, bit 17: full duplex");
MODULE_PARM_DESC(full_duplex, "winbond-840 full duplex setting(s) (1)");
@@ -304,6 +304,8 @@
TxFIFOUnderflow=0x20, RxErrIntr=0x10,
TxIdle=0x04, IntrTxStopped=0x02, IntrTxDone=0x01,
};
+#define INTR_MASK_NORMAL 0x1A0F1
+#define INTR_MASK_MITIGATED 0x1A8B1
/* Bits in the NetworkConfig register. */
enum rx_mode_bits {
@@ -338,13 +340,12 @@
DescIntr=0x80000000,
};
-#define PRIV_ALIGN 15 /* Required alignment mask */
#define MII_CNT 1 /* winbond only supports one MII */
struct netdev_private {
struct w840_rx_desc *rx_ring;
dma_addr_t rx_addr[RX_RING_SIZE];
struct w840_tx_desc *tx_ring;
- dma_addr_t tx_addr[RX_RING_SIZE];
+ dma_addr_t tx_addr[TX_RING_SIZE];
dma_addr_t ring_dma_addr;
/* The addresses of receive-in-place skbuffs. */
struct sk_buff* rx_skbuff[RX_RING_SIZE];
@@ -371,6 +372,9 @@
u16 advertising; /* NWay media advertisement */
unsigned char phys[MII_CNT]; /* MII device addresses, but only the first is used */
u32 mii;
+ int jiffy_packets;
+ int mitigate;
+ int jiffies;
};
static int eeprom_read(long ioaddr, int location);
@@ -970,8 +974,8 @@
update_csr6(dev, 0x00022002 | update_link(dev) | __set_rx_mode(dev));
/* Clear and Enable interrupts by setting the interrupt mask. */
- writel(0x1A0F5, ioaddr + IntrStatus);
- writel(0x1A0F5, ioaddr + IntrEnable);
+ writel(INTR_MASK_NORMAL, ioaddr + IntrStatus);
+ writel(INTR_MASK_NORMAL, ioaddr + IntrEnable);
writel(0, ioaddr + RxStartDemand);
}
@@ -1064,12 +1068,12 @@
np->tx_ring[entry].buffer1 = np->tx_addr[entry];
if (skb->len < TX_BUFLIMIT) {
- np->tx_ring[entry].length = DescWholePkt | skb->len;
+ np->tx_ring[entry].length = DescIntr | DescWholePkt | skb->len;
} else {
int len = skb->len - TX_BUFLIMIT;
np->tx_ring[entry].buffer2 = np->tx_addr[entry]+TX_BUFLIMIT;
- np->tx_ring[entry].length = DescWholePkt | (len << 11) | TX_BUFLIMIT;
+ np->tx_ring[entry].length = DescIntr | DescWholePkt | (len << 11) | TX_BUFLIMIT;
}
if(entry == TX_RING_SIZE-1)
np->tx_ring[entry].length |= DescEndRing;
@@ -1097,7 +1101,6 @@
if (np->cur_tx - np->dirty_tx > TX_QUEUE_LEN ||
((np->drv_flags & HasBrokenTx) && np->tx_q_bytes > TX_BUG_FIFO_LIMIT)) {
netif_stop_queue(dev);
- wmb();
np->tx_full = 1;
}
spin_unlock_irq(&np->lock);
@@ -1162,11 +1165,36 @@
np->tx_q_bytes < TX_BUG_FIFO_LIMIT) {
/* The ring is no longer full, clear tbusy. */
np->tx_full = 0;
- wmb();
netif_wake_queue(dev);
}
}
+void check_mitigation(struct net_device *dev)
+{
+ struct netdev_private *np = dev->priv;
+ int packets = np->jiffy_packets;
+ long ioaddr = dev->base_addr;
+
+ if (!netif_device_present(dev))
+ goto out; /* needed against suspend races */
+ np->jiffy_packets = 0;
+ if(packets > (packet_limit*4)/(3*HZ) && !np->mitigate) {
+ /* switch on mitigation */
+ writel(INTR_MASK_MITIGATED, ioaddr + IntrEnable);
+ np->mitigate = 1;
+ if (debug > 2)
+ printk(KERN_DEBUG "%s: enabling mitigation.\n", dev->name);
+ } else if (packets < packet_limit/HZ && np->mitigate) {
+ /* switch off mitigation */
+ writel(INTR_MASK_NORMAL, ioaddr + IntrEnable);
+ np->mitigate = 0;
+ if (debug > 2)
+ printk(KERN_DEBUG "%s: disabling mitigation.\n", dev->name);
+ }
+out:
+ np->jiffies = jiffies;
+}
+
/* The interrupt handler does all of the Rx thread work and cleans up
after the Tx thread. */
static void intr_handler(int irq, void *dev_instance, struct pt_regs *rgs)
@@ -1174,55 +1202,47 @@
struct net_device *dev = (struct net_device *)dev_instance;
struct netdev_private *np = dev->priv;
long ioaddr = dev->base_addr;
- int work_limit = max_interrupt_work;
+ u32 intr_status;
if (!netif_device_present(dev))
return;
- do {
- u32 intr_status = readl(ioaddr + IntrStatus);
-
- /* Acknowledge all of the current interrupt sources ASAP. */
- writel(intr_status & 0x001ffff, ioaddr + IntrStatus);
-
- if (debug > 4)
- printk(KERN_DEBUG "%s: Interrupt, status %4.4x.\n",
+ if (np->jiffies != jiffies)
+ check_mitigation(dev);
+ intr_status = readl(ioaddr + IntrStatus);
+ if (debug > 4)
+ printk(KERN_DEBUG "%s: Interrupt, status %4.4x.\n",
dev->name, intr_status);
- if ((intr_status & (NormalIntr|AbnormalIntr)) == 0)
- break;
+ if (!(intr_status & (NormalIntr|AbnormalIntr|IntrRxDone)))
+ goto out;
- if (intr_status & (IntrRxDone | RxNoBuf))
- netdev_rx(dev);
- if (intr_status & RxNoBuf)
- writel(0, ioaddr + RxStartDemand);
-
- if (intr_status & (TxIdle | IntrTxDone) &&
- np->cur_tx != np->dirty_tx) {
- spin_lock(&np->lock);
- netdev_tx_done(dev);
- spin_unlock(&np->lock);
- }
+ /* Acknowledge all of the current interrupt sources ASAP. */
+ writel(intr_status & 0x001ffff, ioaddr + IntrStatus);
- /* Abnormal error summary/uncommon events handlers. */
- if (intr_status & (AbnormalIntr | TxFIFOUnderflow | IntrPCIErr |
- TimerInt | IntrTxStopped))
- netdev_error(dev, intr_status);
-
- if (--work_limit < 0) {
- printk(KERN_WARNING "%s: Too much work at interrupt, "
- "status=0x%4.4x.\n", dev->name, intr_status);
- /* Set the timer to re-enable the other interrupts after
- 10*82usec ticks. */
- spin_lock(&np->lock);
- if (netif_device_present(dev)) {
- writel(AbnormalIntr | TimerInt, ioaddr + IntrEnable);
- writel(10, ioaddr + GPTimer);
- }
- spin_unlock(&np->lock);
- break;
- }
- } while (1);
+ /* and reload to make certain it has left the writebuffer */
+ readl(ioaddr + IntrStatus);
+ if (intr_status & (IntrRxDone | RxNoBuf))
+ netdev_rx(dev);
+ if (intr_status & RxNoBuf)
+ writel(0, ioaddr + RxStartDemand);
+
+ if (intr_status & (TxIdle | IntrTxDone) &&
+ np->cur_tx != np->dirty_tx) {
+ spin_lock(&np->lock);
+ netdev_tx_done(dev);
+ spin_unlock(&np->lock);
+ }
+ /* Abnormal error summary/uncommon events handlers. */
+ if (intr_status & (TxFIFOUnderflow | IntrPCIErr | IntrRxDied |
+ IntrTxStopped))
+ netdev_error(dev, intr_status);
+out:
+ if (np->mitigate) {
+ /* set 0x10000 (periodic) to guarantee that the timer arrives
+ * it seems that otherwise the timer is sometimes lost */
+ writel(0x10000|7, ioaddr + GPTimer);
+ }
if (debug > 3)
printk(KERN_DEBUG "%s: exiting interrupt, status=%#4.4x.\n",
dev->name, (int)readl(ioaddr + IntrStatus));
@@ -1255,7 +1275,8 @@
if ((status & 0x38000300) != 0x0300) {
/* Ingore earlier buffers. */
if ((status & 0xffff) != 0x7fff) {
- printk(KERN_WARNING "%s: Oversized Ethernet frame spanned "
+ if (debug > 1)
+ printk(KERN_INFO "%s: Oversized Ethernet frame spanned "
"multiple buffers, entry %#x status %4.4x!\n",
dev->name, np->cur_rx, status);
np->stats.rx_length_errors++;
@@ -1322,6 +1343,7 @@
dev->last_rx = jiffies;
np->stats.rx_packets++;
np->stats.rx_bytes += pkt_len;
+ np->jiffy_packets++;
}
entry = (++np->cur_rx) % RX_RING_SIZE;
np->rx_head_desc = &np->rx_ring[entry];
@@ -1382,11 +1404,6 @@
}
if (intr_status & IntrRxDied) { /* Missed a Rx frame. */
np->stats.rx_errors++;
- }
- if (intr_status & TimerInt) {
- /* Re-enable other interrupts. */
- if (netif_device_present(dev))
- writel(0x1A0F5, ioaddr + IntrEnable);
}
np->stats.rx_missed_errors += readl(ioaddr + RxMissed) & 0xffff;
writel(0, ioaddr + RxStartDemand);
