[netdevice] Add netdev_tx_defer() to allow drivers to defer transmissions

Devices with small transmit descriptor rings may temporarily run out
of space.  Provide netdev_tx_defer() to allow drivers to defer packets
for retransmission as soon as a descriptor becomes available.

Signed-off-by: Michael Brown <mcb30@ipxe.org>

src/include/ipxe/netdevice.h

@@ -346,6 +346,8 @@ struct net_device {
 	size_t max_pkt_len;
 	/** TX packet queue */
 	struct list_head tx_queue;
+	/** Deferred TX packet queue */
+	struct list_head tx_deferred;
 	/** RX packet queue */
 	struct list_head rx_queue;
 	/** TX statistics */
@@ -605,6 +607,8 @@ netdev_rx_frozen ( struct net_device *netdev ) {
 extern void netdev_link_err ( struct net_device *netdev, int rc );
 extern void netdev_link_down ( struct net_device *netdev );
 extern int netdev_tx ( struct net_device *netdev, struct io_buffer *iobuf );
+extern void netdev_tx_defer ( struct net_device *netdev,
+			      struct io_buffer *iobuf );
 extern void netdev_tx_err ( struct net_device *netdev,
 			    struct io_buffer *iobuf, int rc );
 extern void netdev_tx_complete_err ( struct net_device *netdev,

src/net/netdevice.c

@@ -31,6 +31,7 @@ FILE_LICENCE ( GPL2_OR_LATER );
 #include <ipxe/tables.h>
 #include <ipxe/process.h>
 #include <ipxe/init.h>
+#include <ipxe/malloc.h>
 #include <ipxe/device.h>
 #include <ipxe/errortab.h>
 #include <ipxe/vlan.h>
@@ -212,6 +213,43 @@ int netdev_tx ( struct net_device *netdev, struct io_buffer *iobuf ) {
 	return rc;
 }
 
+/**
+ * Defer transmitted packet
+ *
+ * @v netdev		Network device
+ * @v iobuf		I/O buffer
+ *
+ * Drivers may call netdev_tx_defer() if there is insufficient space
+ * in the transmit descriptor ring.  Any packets deferred in this way
+ * will be automatically retransmitted as soon as space becomes
+ * available (i.e. as soon as the driver calls netdev_tx_complete()).
+ *
+ * The packet must currently be in the network device's TX queue.
+ *
+ * Drivers utilising netdev_tx_defer() must ensure that space in the
+ * transmit descriptor ring is freed up @b before calling
+ * netdev_tx_complete().  For example, if the ring is modelled using a
+ * producer counter and a consumer counter, then the consumer counter
+ * must be incremented before the call to netdev_tx_complete().
+ * Failure to do this will cause the retransmitted packet to be
+ * immediately redeferred (which will result in out-of-order
+ * transmissions and other nastiness).
+ */
+void netdev_tx_defer ( struct net_device *netdev, struct io_buffer *iobuf ) {
+
+	/* Catch data corruption as early as possible */
+	list_check_contains_entry ( iobuf, &netdev->tx_queue, list );
+
+	/* Remove from transmit queue */
+	list_del ( &iobuf->list );
+
+	/* Add to deferred transmit queue */
+	list_add_tail ( &iobuf->list, &netdev->tx_deferred );
+
+	/* Record "out of space" statistic */
+	netdev_tx_err ( netdev, NULL, -ENOBUFS );
+}
+
 /**
  * Discard transmitted packet
  *
@@ -257,6 +295,13 @@ void netdev_tx_complete_err ( struct net_device *netdev,
 	/* Dequeue and free I/O buffer */
 	list_del ( &iobuf->list );
 	netdev_tx_err ( netdev, iobuf, rc );
+
+	/* Transmit first pending packet, if any */
+	if ( ( iobuf = list_first_entry ( &netdev->tx_deferred,
+					  struct io_buffer, list ) ) != NULL ) {
+		list_del ( &iobuf->list );
+		netdev_tx ( netdev, iobuf );
+	}
 }
 
 /**
@@ -270,9 +315,9 @@ void netdev_tx_complete_err ( struct net_device *netdev,
 void netdev_tx_complete_next_err ( struct net_device *netdev, int rc ) {
 	struct io_buffer *iobuf;
 
-	list_for_each_entry ( iobuf, &netdev->tx_queue, list ) {
+	if ( ( iobuf = list_first_entry ( &netdev->tx_queue, struct io_buffer,
+					  list ) ) != NULL ) {
 		netdev_tx_complete_err ( netdev, iobuf, rc );
-		return;
 	}
 }
 
@@ -283,10 +328,15 @@ void netdev_tx_complete_next_err ( struct net_device *netdev, int rc ) {
  */
 static void netdev_tx_flush ( struct net_device *netdev ) {
 
-	/* Discard any packets in the TX queue */
+	/* Discard any packets in the TX queue.  This will also cause
+	 * any packets in the deferred TX queue to be discarded
+	 * automatically.
+	 */
 	while ( ! list_empty ( &netdev->tx_queue ) ) {
 		netdev_tx_complete_next_err ( netdev, -ECANCELED );
 	}
+	assert ( list_empty ( &netdev->tx_queue ) );
+	assert ( list_empty ( &netdev->tx_deferred ) );
 }
 
 /**
@@ -424,6 +474,7 @@ struct net_device * alloc_netdev ( size_t priv_size ) {
 		ref_init ( &netdev->refcnt, free_netdev );
 		netdev->link_rc = -EUNKNOWN_LINK_STATUS;
 		INIT_LIST_HEAD ( &netdev->tx_queue );
+		INIT_LIST_HEAD ( &netdev->tx_deferred );
 		INIT_LIST_HEAD ( &netdev->rx_queue );
 		netdev_settings_init ( netdev );
 		netdev->priv = ( ( ( void * ) netdev ) + sizeof ( *netdev ) );
@@ -817,3 +868,36 @@ __weak struct net_device * vlan_find ( struct net_device *trunk __unused,
 
 /** Networking stack process */
 PERMANENT_PROCESS ( net_process, net_step );
+
+/**
+ * Discard some cached network device data
+ *
+ * @ret discarded	Number of cached items discarded
+ */
+static unsigned int net_discard ( void ) {
+	struct net_device *netdev;
+	struct io_buffer *iobuf;
+	unsigned int discarded = 0;
+
+	/* Try to drop one deferred TX packet from each network device */
+	for_each_netdev ( netdev ) {
+		if ( ( iobuf = list_first_entry ( &netdev->tx_deferred,
+						  struct io_buffer,
+						  list ) ) != NULL ) {
+
+			/* Discard first deferred packet */
+			list_del ( &iobuf->list );
+			free ( iobuf );
+
+			/* Report discard */
+			discarded++;
+		}
+	}
+
+	return discarded;
+}
+
+/** Network device cache discarder */
+struct cache_discarder net_discarder __cache_discarder ( CACHE_NORMAL ) = {
+	.discard = net_discard,
+};