Add RX quotas to the net device poll() method. This avoids the problem

of alloc_pkb() exhaustion when e.g. an iSCSI-booted DOS session is left
idle for a long time at the C:\ prompt and builds up a huge packet
backlog.

6 files changed, 76 insertions, 37 deletions

diff --git a/src/arch/i386/drivers/net/undinet.c b/src/arch/i386/drivers/net/undinet.c
index c53ff53c..5ff9bd42 100644
--- a/src/arch/i386/drivers/net/undinet.c
+++ b/src/arch/i386/drivers/net/undinet.c
@@ -364,7 +364,8 @@ static int undinet_transmit ( struct net_device *netdev,
 /** 
  * Poll for received packets
  *
- * @v netdev	Network device
+ * @v netdev		Network device
+ * @v rx_quota		Maximum number of packets to receive
  *
  * Fun, fun, fun.  UNDI drivers don't use polling; they use
  * interrupts.  We therefore cheat and pretend that an interrupt has
@@ -382,7 +383,7 @@ static int undinet_transmit ( struct net_device *netdev,
  * of installing a genuine interrupt service routine and dealing with
  * the wonderful 8259 Programmable Interrupt Controller.  Joy.
  */
-static void undinet_poll ( struct net_device *netdev ) {
+static void undinet_poll ( struct net_device *netdev, unsigned int rx_quota ) {
 	struct undi_nic *undinic = netdev->priv;
 	struct s_PXENV_UNDI_ISR undi_isr;
 	struct pk_buff *pkb = NULL;
@@ -416,7 +417,7 @@ static void undinet_poll ( struct net_device *netdev ) {
 	}
 
 	/* Run through the ISR loop */
-	while ( 1 ) {
+	while ( rx_quota ) {
 		if ( ( rc = undinet_call ( undinic, PXENV_UNDI_ISR, &undi_isr,
 					   sizeof ( undi_isr ) ) ) != 0 )
 			break;
@@ -448,6 +449,7 @@ static void undinet_poll ( struct net_device *netdev ) {
 			if ( pkb_len ( pkb ) == len ) {
 				netdev_rx ( netdev, pkb );
 				pkb = NULL;
+				--rx_quota;
 			}
 			break;
 		case PXENV_UNDI_ISR_OUT_DONE:
@@ -480,8 +482,8 @@ static void undinet_poll ( struct net_device *netdev ) {
  */
 static int undinet_open ( struct net_device *netdev ) {
 	struct undi_nic *undinic = netdev->priv;
-	struct s_PXENV_UNDI_SET_STATION_ADDRESS set_address;
-	struct s_PXENV_UNDI_OPEN open;
+	struct s_PXENV_UNDI_SET_STATION_ADDRESS undi_set_address;
+	struct s_PXENV_UNDI_OPEN undi_open;
 	int rc;
 
 	/* Hook interrupt service routine and enable interrupt */
@@ -494,16 +496,16 @@ static int undinet_open ( struct net_device *netdev ) {
 	 * use it to set the MAC address to the card's permanent value
 	 * anyway.
 	 */
-	memcpy ( set_address.StationAddress, netdev->ll_addr,
-		 sizeof ( set_address.StationAddress ) );
+	memcpy ( undi_set_address.StationAddress, netdev->ll_addr,
+		 sizeof ( undi_set_address.StationAddress ) );
 	undinet_call ( undinic, PXENV_UNDI_SET_STATION_ADDRESS,
-		       &set_address, sizeof ( set_address ) );
+		       &undi_set_address, sizeof ( undi_set_address ) );
 
 	/* Open NIC */
-	memset ( &open, 0, sizeof ( open ) );
-	open.PktFilter = ( FLTR_DIRECTED | FLTR_BRDCST );
-	if ( ( rc = undinet_call ( undinic, PXENV_UNDI_OPEN, &open,
-				   sizeof ( open ) ) ) != 0 )
+	memset ( &undi_open, 0, sizeof ( undi_open ) );
+	undi_open.PktFilter = ( FLTR_DIRECTED | FLTR_BRDCST );
+	if ( ( rc = undinet_call ( undinic, PXENV_UNDI_OPEN, &undi_open,
+				   sizeof ( undi_open ) ) ) != 0 )
 		goto err;
 
 	return 0;
@@ -520,14 +522,31 @@ static int undinet_open ( struct net_device *netdev ) {
  */
 static void undinet_close ( struct net_device *netdev ) {
 	struct undi_nic *undinic = netdev->priv;
-	struct s_PXENV_UNDI_CLOSE close;
+	struct s_PXENV_UNDI_ISR undi_isr;
+	struct s_PXENV_UNDI_CLOSE undi_close;
+	int rc;
 
 	/* Ensure ISR has exited cleanly */
-	while ( undinic->isr_processing )
-		undinet_poll ( netdev );
+	while ( undinic->isr_processing ) {
+		undi_isr.FuncFlag = PXENV_UNDI_ISR_IN_GET_NEXT;
+		if ( ( rc = undinet_call ( undinic, PXENV_UNDI_ISR, &undi_isr,
+					   sizeof ( undi_isr ) ) ) != 0 )
+			break;
+		switch ( undi_isr.FuncFlag ) {
+		case PXENV_UNDI_ISR_OUT_TRANSMIT:
+		case PXENV_UNDI_ISR_OUT_RECEIVE:
+			/* Continue draining */
+			break;
+		default:
+			/* Stop processing */
+			undinic->isr_processing = 0;
+			break;
+		}
+	}
 
 	/* Close NIC */
-	undinet_call ( undinic, PXENV_UNDI_CLOSE, &close, sizeof ( close ) );
+	undinet_call ( undinic, PXENV_UNDI_CLOSE, &undi_close,
+		       sizeof ( undi_close ) );
 
 	/* Disable interrupt and unhook ISR */
 	disable_irq ( undinic->irq );
diff --git a/src/drivers/net/legacy.c b/src/drivers/net/legacy.c
index 2d952aac..091cdcc2 100644
--- a/src/drivers/net/legacy.c
+++ b/src/drivers/net/legacy.c
@@ -38,10 +38,13 @@ static int legacy_transmit ( struct net_device *netdev, struct pk_buff *pkb ) {
 	return 0;
 }
 
-static void legacy_poll ( struct net_device *netdev ) {
+static void legacy_poll ( struct net_device *netdev, unsigned int rx_quota ) {
 	struct nic *nic = netdev->priv;
 	struct pk_buff *pkb;
 
+	if ( ! rx_quota )
+		return;
+
 	pkb = alloc_pkb ( ETH_FRAME_LEN );
 	if ( ! pkb )
 		return;
diff --git a/src/drivers/net/pnic.c b/src/drivers/net/pnic.c
index 09ac92a4..c74e7b50 100644
--- a/src/drivers/net/pnic.c
+++ b/src/drivers/net/pnic.c
@@ -112,14 +112,14 @@ static int pnic_api_check ( uint16_t api_version ) {
 /**************************************************************************
 POLL - Wait for a frame
 ***************************************************************************/
-static void pnic_poll ( struct net_device *netdev ) {
+static void pnic_poll ( struct net_device *netdev, unsigned int rx_quota ) {
 	struct pnic *pnic = netdev->priv;
 	struct pk_buff *pkb;
 	uint16_t length;
 	uint16_t qlen;
 
 	/* Fetch all available packets */
-	while ( 1 ) {
+	while ( rx_quota ) {
 		if ( pnic_command ( pnic, PNIC_CMD_RECV_QLEN, NULL, 0,
 				    &qlen, sizeof ( qlen ), NULL )
 		     != PNIC_STATUS_OK )
@@ -139,6 +139,7 @@ static void pnic_poll ( struct net_device *netdev ) {
 		}
 		pkb_put ( pkb, length );
 		netdev_rx ( netdev, pkb );
+		--rx_quota;
 	}
 }
 
diff --git a/src/drivers/net/rtl8139.c b/src/drivers/net/rtl8139.c
index 4981329c..a25755c2 100644
--- a/src/drivers/net/rtl8139.c
+++ b/src/drivers/net/rtl8139.c
@@ -413,8 +413,9 @@ static int rtl_transmit ( struct net_device *netdev, struct pk_buff *pkb ) {
  * Poll for received packets
  *
  * @v netdev	Network device
+ * @v rx_quota	Maximum number of packets to receive
  */
-static void rtl_poll ( struct net_device *netdev ) {
+static void rtl_poll ( struct net_device *netdev, unsigned int rx_quota ) {
 	struct rtl8139_nic *rtl = netdev->priv;
 	unsigned int status;
 	unsigned int tsad;
@@ -441,7 +442,7 @@ static void rtl_poll ( struct net_device *netdev ) {
 	}
 
 	/* Handle received packets */
-	while ( ! ( inw ( rtl->ioaddr + ChipCmd ) & RxBufEmpty ) ) {
+	while ( rx_quota && ! ( inw ( rtl->ioaddr + ChipCmd ) & RxBufEmpty ) ){
 		rx_status = * ( ( uint16_t * )
 				( rtl->rx.ring + rtl->rx.offset ) );
 		rx_len = * ( ( uint16_t * )
diff --git a/src/include/gpxe/netdevice.h b/src/include/gpxe/netdevice.h
index ad1f9013..bc082728 100644
--- a/src/include/gpxe/netdevice.h
+++ b/src/include/gpxe/netdevice.h
@@ -181,15 +181,16 @@ struct net_device {
 	/** Poll for received packet
 	 *
 	 * @v netdev	Network device
+	 * @v rx_quota	Maximum number of packets to receive
 	 *
 	 * This method should cause the hardware to check for received
 	 * packets.  Any received packets should be delivered via
-	 * netdev_rx().
+	 * netdev_rx(), up to a maximum of @c rx_quota packets.
 	 *
 	 * This method is guaranteed to be called only when the device
 	 * is open.
 	 */
-	void ( * poll ) ( struct net_device *netdev );
+	void ( * poll ) ( struct net_device *netdev, unsigned int rx_quota );
 
 	/** Link-layer protocol */
 	struct ll_protocol *ll_protocol;
@@ -238,7 +239,7 @@ extern int netdev_tx ( struct net_device *netdev, struct pk_buff *pkb );
 void netdev_tx_complete ( struct net_device *netdev, struct pk_buff *pkb );
 void netdev_tx_complete_next ( struct net_device *netdev );
 extern void netdev_rx ( struct net_device *netdev, struct pk_buff *pkb );
-extern int netdev_poll ( struct net_device *netdev );
+extern int netdev_poll ( struct net_device *netdev, unsigned int rx_quota );
 extern struct pk_buff * netdev_rx_dequeue ( struct net_device *netdev );
 extern struct net_device * alloc_netdev ( size_t priv_size );
 extern int register_netdev ( struct net_device *netdev );
diff --git a/src/net/netdevice.c b/src/net/netdevice.c
index 6da2ddfd..40f836d4 100644
--- a/src/net/netdevice.c
+++ b/src/net/netdevice.c
@@ -126,16 +126,17 @@ void netdev_rx ( struct net_device *netdev, struct pk_buff *pkb ) {
  * Poll for packet on network device
  *
  * @v netdev		Network device
+ * @v rx_quota		Maximum number of packets to receive
  * @ret True		There are packets present in the receive queue
  * @ret False		There are no packets present in the receive queue
  *
  * Polls the network device for received packets.  Any received
  * packets will be added to the RX packet queue via netdev_rx().
  */
-int netdev_poll ( struct net_device *netdev ) {
+int netdev_poll ( struct net_device *netdev, unsigned int rx_quota ) {
 
 	if ( netdev->state & NETDEV_OPEN )
-		netdev->poll ( netdev );
+		netdev->poll ( netdev, rx_quota );
 
 	return ( ! list_empty ( &netdev->rx_queue ) );
 }
@@ -351,14 +352,9 @@ int net_rx ( struct pk_buff *pkb, struct net_device *netdev,
  *
  * @v process		Network stack process
  *
- * This polls all interfaces for any received packets, and processes
- * at most one packet from the RX queue.
+ * This polls all interfaces for received packets, and processes
+ * packets from the RX queue.
  *
- * We avoid processing all received packets, because processing the
- * received packet can trigger transmission of a new packet (e.g. an
- * ARP response).  Since TX completions will be processed as part of
- * the poll operation, it is easy to overflow small TX queues if
- * multiple packets are processed per poll.
  */
 static void net_step ( struct process *process ) {
 	struct net_device *netdev;
@@ -367,10 +363,28 @@ static void net_step ( struct process *process ) {
 	/* Poll and process each network device */
 	list_for_each_entry ( netdev, &net_devices, list ) {
 
-		/* Poll for new packets */
-		netdev_poll ( netdev );
-
-		/* Handle at most one received packet per poll */
+		/* Poll for new packets.  Limit RX queue size to a
+		 * single packet, because otherwise most drivers are
+		 * in serious danger of running out of memory and
+		 * having to drop packets.
+		 *
+		 * This limitation isn't relevant to devices that
+		 * preallocate packet buffers (i.e. devices with
+		 * descriptor-based RX datapaths).  We might at some
+		 * point want to relax the quota for such devices.
+		 */
+		netdev_poll ( netdev,
+			      ( list_empty ( &netdev->rx_queue ) ? 1 : 0 ) );
+
+		/* Handle at most one received packet per poll.  We
+		 * avoid processing more than one packet per call to
+		 * netdev_poll(), because processing the received
+		 * packet can trigger transmission of a new packet
+		 * (e.g. an ARP response).  Since TX completions will
+		 * be processed as part of the poll operation, it is
+		 * easy to overflow small TX queues if multiple
+		 * packets are processed per poll.
+		 */
 		if ( ( pkb = netdev_rx_dequeue ( netdev ) ) ) {
 			DBGC ( netdev, "NETDEV %p processing %p\n",
 			       netdev, pkb );