Merge branch 'for-davem' of git://git.kernel.org/pub/scm/linux/kernel/git/bwh/sfc-next-2.6

drivers/net/sfc/efx.h

@@ -74,8 +74,7 @@ extern int efx_filter_insert_filter(struct efx_nic *efx,
 				    bool replace);
 extern int efx_filter_remove_filter(struct efx_nic *efx,
 				    struct efx_filter_spec *spec);
-extern void efx_filter_table_clear(struct efx_nic *efx,
-				   enum efx_filter_table_id table_id,
+extern void efx_filter_clear_rx(struct efx_nic *efx,
 				enum efx_filter_priority priority);
 
 /* Channels */

drivers/net/sfc/ethtool.c

@@ -11,6 +11,7 @@
 #include <linux/netdevice.h>
 #include <linux/ethtool.h>
 #include <linux/rtnetlink.h>
+#include <linux/in.h>
 #include "net_driver.h"
 #include "workarounds.h"
 #include "selftest.h"
@@ -558,12 +559,8 @@ static int efx_ethtool_set_flags(struct net_device *net_dev, u32 data)
 	if (rc)
 		return rc;
 
-	if (!(data & ETH_FLAG_NTUPLE)) {
-		efx_filter_table_clear(efx, EFX_FILTER_TABLE_RX_IP,
-				       EFX_FILTER_PRI_MANUAL);
-		efx_filter_table_clear(efx, EFX_FILTER_TABLE_RX_MAC,
-				       EFX_FILTER_PRI_MANUAL);
-	}
+	if (!(data & ETH_FLAG_NTUPLE))
+		efx_filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
 
 	return 0;
 }
@@ -582,6 +579,9 @@ static void efx_ethtool_self_test(struct net_device *net_dev,
 		goto fail1;
 	}
 
+	netif_info(efx, drv, efx->net_dev, "starting %sline testing\n",
+		   (test->flags & ETH_TEST_FL_OFFLINE) ? "off" : "on");
+
 	/* We need rx buffers and interrupts. */
 	already_up = (efx->net_dev->flags & IFF_UP);
 	if (!already_up) {
@@ -600,7 +600,7 @@ static void efx_ethtool_self_test(struct net_device *net_dev,
 	if (!already_up)
 		dev_close(efx->net_dev);
 
-	netif_dbg(efx, drv, efx->net_dev, "%s %sline self-tests\n",
+	netif_info(efx, drv, efx->net_dev, "%s %sline self-tests\n",
 		   rc == 0 ? "passed" : "failed",
 		   (test->flags & ETH_TEST_FL_OFFLINE) ? "off" : "on");
 
@@ -921,6 +921,7 @@ static int efx_ethtool_set_rx_ntuple(struct net_device *net_dev,
 	struct ethhdr *mac_entry = &ntuple->fs.h_u.ether_spec;
 	struct ethhdr *mac_mask = &ntuple->fs.m_u.ether_spec;
 	struct efx_filter_spec filter;
+	int rc;
 
 	/* Range-check action */
 	if (ntuple->fs.action < ETHTOOL_RXNTUPLE_ACTION_CLEAR ||
@@ -930,9 +931,16 @@ static int efx_ethtool_set_rx_ntuple(struct net_device *net_dev,
 	if (~ntuple->fs.data_mask)
 		return -EINVAL;
 
+	efx_filter_init_rx(&filter, EFX_FILTER_PRI_MANUAL, 0,
+			   (ntuple->fs.action == ETHTOOL_RXNTUPLE_ACTION_DROP) ?
+			   0xfff : ntuple->fs.action);
+
 	switch (ntuple->fs.flow_type) {
 	case TCP_V4_FLOW:
-	case UDP_V4_FLOW:
+	case UDP_V4_FLOW: {
+		u8 proto = (ntuple->fs.flow_type == TCP_V4_FLOW ?
+			    IPPROTO_TCP : IPPROTO_UDP);
+
 		/* Must match all of destination, */
 		if (ip_mask->ip4dst | ip_mask->pdst)
 			return -EINVAL;
@@ -944,7 +952,22 @@ static int efx_ethtool_set_rx_ntuple(struct net_device *net_dev,
 		/* and nothing else */
 		if ((u8)~ip_mask->tos | (u16)~ntuple->fs.vlan_tag_mask)
 			return -EINVAL;
+
+		if (!ip_mask->ip4src)
+			rc = efx_filter_set_ipv4_full(&filter, proto,
+						      ip_entry->ip4dst,
+						      ip_entry->pdst,
+						      ip_entry->ip4src,
+						      ip_entry->psrc);
+		else
+			rc = efx_filter_set_ipv4_local(&filter, proto,
+						       ip_entry->ip4dst,
+						       ip_entry->pdst);
+		if (rc)
+			return rc;
 		break;
+	}
+
 	case ETHER_FLOW:
 		/* Must match all of destination, */
 		if (!is_zero_ether_addr(mac_mask->h_dest))
@@ -957,58 +980,24 @@ static int efx_ethtool_set_rx_ntuple(struct net_device *net_dev,
 		if (!is_broadcast_ether_addr(mac_mask->h_source) ||
 		    mac_mask->h_proto != htons(0xffff))
 			return -EINVAL;
-		break;
-	default:
-		return -EINVAL;
-	}
-
-	filter.priority = EFX_FILTER_PRI_MANUAL;
-	filter.flags = 0;
 
-	switch (ntuple->fs.flow_type) {
-	case TCP_V4_FLOW:
-		if (!ip_mask->ip4src)
-			efx_filter_set_rx_tcp_full(&filter,
-						   htonl(ip_entry->ip4src),
-						   htons(ip_entry->psrc),
-						   htonl(ip_entry->ip4dst),
-						   htons(ip_entry->pdst));
-		else
-			efx_filter_set_rx_tcp_wild(&filter,
-						   htonl(ip_entry->ip4dst),
-						   htons(ip_entry->pdst));
-		break;
-	case UDP_V4_FLOW:
-		if (!ip_mask->ip4src)
-			efx_filter_set_rx_udp_full(&filter,
-						   htonl(ip_entry->ip4src),
-						   htons(ip_entry->psrc),
-						   htonl(ip_entry->ip4dst),
-						   htons(ip_entry->pdst));
-		else
-			efx_filter_set_rx_udp_wild(&filter,
-						   htonl(ip_entry->ip4dst),
-						   htons(ip_entry->pdst));
-		break;
-	case ETHER_FLOW:
-		if (ntuple->fs.vlan_tag_mask == 0xf000)
-			efx_filter_set_rx_mac_full(&filter,
-						   ntuple->fs.vlan_tag & 0xfff,
+		rc = efx_filter_set_eth_local(
+			&filter,
+			(ntuple->fs.vlan_tag_mask == 0xf000) ?
+			ntuple->fs.vlan_tag : EFX_FILTER_VID_UNSPEC,
 			mac_entry->h_dest);
-		else
-			efx_filter_set_rx_mac_wild(&filter, mac_entry->h_dest);
+		if (rc)
+			return rc;
 		break;
+
+	default:
+		return -EINVAL;
 	}
 
-	if (ntuple->fs.action == ETHTOOL_RXNTUPLE_ACTION_CLEAR) {
+	if (ntuple->fs.action == ETHTOOL_RXNTUPLE_ACTION_CLEAR)
 		return efx_filter_remove_filter(efx, &filter);
-	} else {
-		if (ntuple->fs.action == ETHTOOL_RXNTUPLE_ACTION_DROP)
-			filter.dmaq_id = 0xfff;
 	else
-			filter.dmaq_id = ntuple->fs.action;
 		return efx_filter_insert_filter(efx, &filter, true);
-	}
 }
 
 static int efx_ethtool_get_rxfh_indir(struct net_device *net_dev,

drivers/net/sfc/filter.c

@@ -7,6 +7,7 @@
  * by the Free Software Foundation, incorporated herein by reference.
  */
 
+#include <linux/in.h>
 #include "efx.h"
 #include "filter.h"
 #include "io.h"
@@ -26,19 +27,26 @@
  */
 #define FILTER_CTL_SRCH_MAX 200
 
+enum efx_filter_table_id {
+	EFX_FILTER_TABLE_RX_IP = 0,
+	EFX_FILTER_TABLE_RX_MAC,
+	EFX_FILTER_TABLE_COUNT,
+};
+
 struct efx_filter_table {
+	enum efx_filter_table_id id;
 	u32		offset;		/* address of table relative to BAR */
 	unsigned	size;		/* number of entries */
 	unsigned	step;		/* step between entries */
 	unsigned	used;		/* number currently used */
 	unsigned long	*used_bitmap;
 	struct efx_filter_spec *spec;
+	unsigned	search_depth[EFX_FILTER_TYPE_COUNT];
 };
 
 struct efx_filter_state {
 	spinlock_t	lock;
 	struct efx_filter_table table[EFX_FILTER_TABLE_COUNT];
-	unsigned	search_depth[EFX_FILTER_TYPE_COUNT];
 };
 
 /* The filter hash function is LFSR polynomial x^16 + x^3 + 1 of a 32-bit
@@ -65,68 +73,203 @@ static u16 efx_filter_increment(u32 key)
 }
 
 static enum efx_filter_table_id
-efx_filter_type_table_id(enum efx_filter_type type)
+efx_filter_spec_table_id(const struct efx_filter_spec *spec)
+{
+	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_IP != (EFX_FILTER_TCP_FULL >> 2));
+	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_IP != (EFX_FILTER_TCP_WILD >> 2));
+	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_IP != (EFX_FILTER_UDP_FULL >> 2));
+	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_IP != (EFX_FILTER_UDP_WILD >> 2));
+	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_MAC != (EFX_FILTER_MAC_FULL >> 2));
+	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_MAC != (EFX_FILTER_MAC_WILD >> 2));
+	EFX_BUG_ON_PARANOID(spec->type == EFX_FILTER_UNSPEC);
+	return spec->type >> 2;
+}
+
+static struct efx_filter_table *
+efx_filter_spec_table(struct efx_filter_state *state,
+		      const struct efx_filter_spec *spec)
 {
-	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_IP != (EFX_FILTER_RX_TCP_FULL >> 2));
-	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_IP != (EFX_FILTER_RX_TCP_WILD >> 2));
-	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_IP != (EFX_FILTER_RX_UDP_FULL >> 2));
-	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_IP != (EFX_FILTER_RX_UDP_WILD >> 2));
-	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_MAC != (EFX_FILTER_RX_MAC_FULL >> 2));
-	BUILD_BUG_ON(EFX_FILTER_TABLE_RX_MAC != (EFX_FILTER_RX_MAC_WILD >> 2));
-	return type >> 2;
+	if (spec->type == EFX_FILTER_UNSPEC)
+		return NULL;
+	else
+		return &state->table[efx_filter_spec_table_id(spec)];
 }
 
-static void
-efx_filter_table_reset_search_depth(struct efx_filter_state *state,
-				    enum efx_filter_table_id table_id)
+static void efx_filter_table_reset_search_depth(struct efx_filter_table *table)
 {
-	memset(state->search_depth + (table_id << 2), 0,
-	       sizeof(state->search_depth[0]) << 2);
+	memset(table->search_depth, 0, sizeof(table->search_depth));
 }
 
 static void efx_filter_push_rx_limits(struct efx_nic *efx)
 {
 	struct efx_filter_state *state = efx->filter_state;
+	struct efx_filter_table *table;
 	efx_oword_t filter_ctl;
 
 	efx_reado(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
 
+	table = &state->table[EFX_FILTER_TABLE_RX_IP];
 	EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_FULL_SRCH_LIMIT,
-			    state->search_depth[EFX_FILTER_RX_TCP_FULL] +
+			    table->search_depth[EFX_FILTER_TCP_FULL] +
 			    FILTER_CTL_SRCH_FUDGE_FULL);
 	EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_WILD_SRCH_LIMIT,
-			    state->search_depth[EFX_FILTER_RX_TCP_WILD] +
+			    table->search_depth[EFX_FILTER_TCP_WILD] +
 			    FILTER_CTL_SRCH_FUDGE_WILD);
 	EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_FULL_SRCH_LIMIT,
-			    state->search_depth[EFX_FILTER_RX_UDP_FULL] +
+			    table->search_depth[EFX_FILTER_UDP_FULL] +
 			    FILTER_CTL_SRCH_FUDGE_FULL);
 	EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_WILD_SRCH_LIMIT,
-			    state->search_depth[EFX_FILTER_RX_UDP_WILD] +
+			    table->search_depth[EFX_FILTER_UDP_WILD] +
 			    FILTER_CTL_SRCH_FUDGE_WILD);
 
-	if (state->table[EFX_FILTER_TABLE_RX_MAC].size) {
+	table = &state->table[EFX_FILTER_TABLE_RX_MAC];
+	if (table->size) {
 		EFX_SET_OWORD_FIELD(
 			filter_ctl, FRF_CZ_ETHERNET_FULL_SEARCH_LIMIT,
-			state->search_depth[EFX_FILTER_RX_MAC_FULL] +
+			table->search_depth[EFX_FILTER_MAC_FULL] +
 			FILTER_CTL_SRCH_FUDGE_FULL);
 		EFX_SET_OWORD_FIELD(
 			filter_ctl, FRF_CZ_ETHERNET_WILDCARD_SEARCH_LIMIT,
-			state->search_depth[EFX_FILTER_RX_MAC_WILD] +
+			table->search_depth[EFX_FILTER_MAC_WILD] +
 			FILTER_CTL_SRCH_FUDGE_WILD);
 	}
 
 	efx_writeo(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
 }
 
+static inline void __efx_filter_set_ipv4(struct efx_filter_spec *spec,
+					 __be32 host1, __be16 port1,
+					 __be32 host2, __be16 port2)
+{
+	spec->data[0] = ntohl(host1) << 16 | ntohs(port1);
+	spec->data[1] = ntohs(port2) << 16 | ntohl(host1) >> 16;
+	spec->data[2] = ntohl(host2);
+}
+
+/**
+ * efx_filter_set_ipv4_local - specify IPv4 host, transport protocol and port
+ * @spec: Specification to initialise
+ * @proto: Transport layer protocol number
+ * @host: Local host address (network byte order)
+ * @port: Local port (network byte order)
+ */
+int efx_filter_set_ipv4_local(struct efx_filter_spec *spec, u8 proto,
+			      __be32 host, __be16 port)
+{
+	__be32 host1;
+	__be16 port1;
+
+	EFX_BUG_ON_PARANOID(!(spec->flags & EFX_FILTER_FLAG_RX));
+
+	/* This cannot currently be combined with other filtering */
+	if (spec->type != EFX_FILTER_UNSPEC)
+		return -EPROTONOSUPPORT;
+
+	if (port == 0)
+		return -EINVAL;
+
+	switch (proto) {
+	case IPPROTO_TCP:
+		spec->type = EFX_FILTER_TCP_WILD;
+		break;
+	case IPPROTO_UDP:
+		spec->type = EFX_FILTER_UDP_WILD;
+		break;
+	default:
+		return -EPROTONOSUPPORT;
+	}
+
+	/* Filter is constructed in terms of source and destination,
+	 * with the odd wrinkle that the ports are swapped in a UDP
+	 * wildcard filter.  We need to convert from local and remote
+	 * (= zero for wildcard) addresses.
+	 */
+	host1 = 0;
+	if (proto != IPPROTO_UDP) {
+		port1 = 0;
+	} else {
+		port1 = port;
+		port = 0;
+	}
+
+	__efx_filter_set_ipv4(spec, host1, port1, host, port);
+	return 0;
+}
+
+/**
+ * efx_filter_set_ipv4_full - specify IPv4 hosts, transport protocol and ports
+ * @spec: Specification to initialise
+ * @proto: Transport layer protocol number
+ * @host: Local host address (network byte order)
+ * @port: Local port (network byte order)
+ * @rhost: Remote host address (network byte order)
+ * @rport: Remote port (network byte order)
+ */
+int efx_filter_set_ipv4_full(struct efx_filter_spec *spec, u8 proto,
+			     __be32 host, __be16 port,
+			     __be32 rhost, __be16 rport)
+{
+	EFX_BUG_ON_PARANOID(!(spec->flags & EFX_FILTER_FLAG_RX));
+
+	/* This cannot currently be combined with other filtering */
+	if (spec->type != EFX_FILTER_UNSPEC)
+		return -EPROTONOSUPPORT;
+
+	if (port == 0 || rport == 0)
+		return -EINVAL;
+
+	switch (proto) {
+	case IPPROTO_TCP:
+		spec->type = EFX_FILTER_TCP_FULL;
+		break;
+	case IPPROTO_UDP:
+		spec->type = EFX_FILTER_UDP_FULL;
+		break;
+	default:
+		return -EPROTONOSUPPORT;
+	}
+
+	__efx_filter_set_ipv4(spec, rhost, rport, host, port);
+	return 0;
+}
+
+/**
+ * efx_filter_set_eth_local - specify local Ethernet address and optional VID
+ * @spec: Specification to initialise
+ * @vid: VLAN ID to match, or %EFX_FILTER_VID_UNSPEC
+ * @addr: Local Ethernet MAC address
+ */
+int efx_filter_set_eth_local(struct efx_filter_spec *spec,
+			     u16 vid, const u8 *addr)
+{
+	EFX_BUG_ON_PARANOID(!(spec->flags & EFX_FILTER_FLAG_RX));
+
+	/* This cannot currently be combined with other filtering */
+	if (spec->type != EFX_FILTER_UNSPEC)
+		return -EPROTONOSUPPORT;
+
+	if (vid == EFX_FILTER_VID_UNSPEC) {
+		spec->type = EFX_FILTER_MAC_WILD;
+		spec->data[0] = 0;
+	} else {
+		spec->type = EFX_FILTER_MAC_FULL;
+		spec->data[0] = vid;
+	}
+
+	spec->data[1] = addr[2] << 24 | addr[3] << 16 | addr[4] << 8 | addr[5];
+	spec->data[2] = addr[0] << 8 | addr[1];
+	return 0;
+}
+
 /* Build a filter entry and return its n-tuple key. */
 static u32 efx_filter_build(efx_oword_t *filter, struct efx_filter_spec *spec)
 {
 	u32 data3;
 
-	switch (efx_filter_type_table_id(spec->type)) {
+	switch (efx_filter_spec_table_id(spec)) {
 	case EFX_FILTER_TABLE_RX_IP: {
-		bool is_udp = (spec->type == EFX_FILTER_RX_UDP_FULL ||
-			       spec->type == EFX_FILTER_RX_UDP_WILD);
+		bool is_udp = (spec->type == EFX_FILTER_UDP_FULL ||
+			       spec->type == EFX_FILTER_UDP_WILD);
 		EFX_POPULATE_OWORD_7(
 			*filter,
 			FRF_BZ_RSS_EN,
@@ -143,7 +286,7 @@ static u32 efx_filter_build(efx_oword_t *filter, struct efx_filter_spec *spec)
 	}
 
 	case EFX_FILTER_TABLE_RX_MAC: {
-		bool is_wild = spec->type == EFX_FILTER_RX_MAC_WILD;
+		bool is_wild = spec->type == EFX_FILTER_MAC_WILD;
 		EFX_POPULATE_OWORD_8(
 			*filter,
 			FRF_CZ_RMFT_RSS_EN,
@@ -206,6 +349,14 @@ static int efx_filter_search(struct efx_filter_table *table,
 	return filter_idx;
 }
 
+/* Construct/deconstruct external filter IDs */
+
+static inline int
+efx_filter_make_id(enum efx_filter_table_id table_id, unsigned index)
+{
+	return table_id << 16 | index;
+}
+
 /**
  * efx_filter_insert_filter - add or replace a filter
  * @efx: NIC in which to insert the filter
@@ -213,30 +364,28 @@ static int efx_filter_search(struct efx_filter_table *table,
  * @replace: Flag for whether the specified filter may replace a filter
  *	with an identical match expression and equal or lower priority
  *
- * On success, return the filter index within its table.
+ * On success, return the filter ID.
  * On failure, return a negative error code.
  */
 int efx_filter_insert_filter(struct efx_nic *efx, struct efx_filter_spec *spec,
 			     bool replace)
 {
 	struct efx_filter_state *state = efx->filter_state;
-	enum efx_filter_table_id table_id =
-		efx_filter_type_table_id(spec->type);
-	struct efx_filter_table *table = &state->table[table_id];
+	struct efx_filter_table *table = efx_filter_spec_table(state, spec);
 	struct efx_filter_spec *saved_spec;
 	efx_oword_t filter;
 	int filter_idx, depth;
 	u32 key;
 	int rc;
 
-	if (table->size == 0)
+	if (!table || table->size == 0)
 		return -EINVAL;
 
 	key = efx_filter_build(&filter, spec);
 
 	netif_vdbg(efx, hw, efx->net_dev,
 		   "%s: type %d search_depth=%d", __func__, spec->type,
-		   state->search_depth[spec->type]);
+		   table->search_depth[spec->type]);
 
 	spin_lock_bh(&state->lock);
 
@@ -263,8 +412,8 @@ int efx_filter_insert_filter(struct efx_nic *efx, struct efx_filter_spec *spec,
 	}
 	*saved_spec = *spec;
 
-	if (state->search_depth[spec->type] < depth) {
-		state->search_depth[spec->type] = depth;
+	if (table->search_depth[spec->type] < depth) {
+		table->search_depth[spec->type] = depth;
 		efx_filter_push_rx_limits(efx);
 	}
 
@@ -273,6 +422,7 @@ int efx_filter_insert_filter(struct efx_nic *efx, struct efx_filter_spec *spec,
 	netif_vdbg(efx, hw, efx->net_dev,
 		   "%s: filter type %d index %d rxq %u set",
 		   __func__, spec->type, filter_idx, spec->dmaq_id);
+	rc = efx_filter_make_id(table->id, filter_idx);
 
 out:
 	spin_unlock_bh(&state->lock);
@@ -306,15 +456,16 @@ static void efx_filter_table_clear_entry(struct efx_nic *efx,
 int efx_filter_remove_filter(struct efx_nic *efx, struct efx_filter_spec *spec)
 {
 	struct efx_filter_state *state = efx->filter_state;
-	enum efx_filter_table_id table_id =
-		efx_filter_type_table_id(spec->type);
-	struct efx_filter_table *table = &state->table[table_id];
+	struct efx_filter_table *table = efx_filter_spec_table(state, spec);
 	struct efx_filter_spec *saved_spec;
 	efx_oword_t filter;
 	int filter_idx, depth;
 	u32 key;
 	int rc;
 
+	if (!table)
+		return -EINVAL;
+
 	key = efx_filter_build(&filter, spec);
 
 	spin_lock_bh(&state->lock);
@@ -332,7 +483,7 @@ int efx_filter_remove_filter(struct efx_nic *efx, struct efx_filter_spec *spec)
 
 	efx_filter_table_clear_entry(efx, table, filter_idx);
 	if (table->used == 0)
-		efx_filter_table_reset_search_depth(state, table_id);
+		efx_filter_table_reset_search_depth(table);
 	rc = 0;
 
 out:
@@ -340,13 +491,7 @@ int efx_filter_remove_filter(struct efx_nic *efx, struct efx_filter_spec *spec)
 	return rc;
 }
 
-/**
- * efx_filter_table_clear - remove filters from a table by priority
- * @efx: NIC from which to remove the filters
- * @table_id: Table from which to remove the filters
- * @priority: Maximum priority to remove
- */
-void efx_filter_table_clear(struct efx_nic *efx,
+static void efx_filter_table_clear(struct efx_nic *efx,
 				   enum efx_filter_table_id table_id,
 				   enum efx_filter_priority priority)
 {
@@ -360,11 +505,22 @@ void efx_filter_table_clear(struct efx_nic *efx,
 		if (table->spec[filter_idx].priority <= priority)
 			efx_filter_table_clear_entry(efx, table, filter_idx);
 	if (table->used == 0)
-		efx_filter_table_reset_search_depth(state, table_id);
+		efx_filter_table_reset_search_depth(table);
 
 	spin_unlock_bh(&state->lock);
 }
 
+/**
+ * efx_filter_clear_rx - remove RX filters by priority
+ * @efx: NIC from which to remove the filters
+ * @priority: Maximum priority to remove
+ */
+void efx_filter_clear_rx(struct efx_nic *efx, enum efx_filter_priority priority)
+{
+	efx_filter_table_clear(efx, EFX_FILTER_TABLE_RX_IP, priority);
+	efx_filter_table_clear(efx, EFX_FILTER_TABLE_RX_MAC, priority);
+}
+
 /* Restore filter stater after reset */
 void efx_restore_filters(struct efx_nic *efx)
 {
@@ -407,6 +563,7 @@ int efx_probe_filters(struct efx_nic *efx)
 
 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
 		table = &state->table[EFX_FILTER_TABLE_RX_IP];
+		table->id = EFX_FILTER_TABLE_RX_IP;
 		table->offset = FR_BZ_RX_FILTER_TBL0;
 		table->size = FR_BZ_RX_FILTER_TBL0_ROWS;
 		table->step = FR_BZ_RX_FILTER_TBL0_STEP;
@@ -414,6 +571,7 @@ int efx_probe_filters(struct efx_nic *efx)
 
 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
 		table = &state->table[EFX_FILTER_TABLE_RX_MAC];
+		table->id = EFX_FILTER_TABLE_RX_MAC;
 		table->offset = FR_CZ_RX_MAC_FILTER_TBL0;
 		table->size = FR_CZ_RX_MAC_FILTER_TBL0_ROWS;
 		table->step = FR_CZ_RX_MAC_FILTER_TBL0_STEP;

drivers/net/sfc/filter.h

@@ -12,31 +12,27 @@
 
 #include <linux/types.h>
 
-enum efx_filter_table_id {
-	EFX_FILTER_TABLE_RX_IP = 0,
-	EFX_FILTER_TABLE_RX_MAC,
-	EFX_FILTER_TABLE_COUNT,
-};
-
 /**
  * enum efx_filter_type - type of hardware filter
- * @EFX_FILTER_RX_TCP_FULL: RX, matching TCP/IPv4 4-tuple
- * @EFX_FILTER_RX_TCP_WILD: RX, matching TCP/IPv4 destination (host, port)
- * @EFX_FILTER_RX_UDP_FULL: RX, matching UDP/IPv4 4-tuple
- * @EFX_FILTER_RX_UDP_WILD: RX, matching UDP/IPv4 destination (host, port)
- * @EFX_FILTER_RX_MAC_FULL: RX, matching Ethernet destination MAC address, VID
- * @EFX_FILTER_RX_MAC_WILD: RX, matching Ethernet destination MAC address
+ * @EFX_FILTER_TCP_FULL: Matching TCP/IPv4 4-tuple
+ * @EFX_FILTER_TCP_WILD: Matching TCP/IPv4 destination (host, port)
+ * @EFX_FILTER_UDP_FULL: Matching UDP/IPv4 4-tuple
+ * @EFX_FILTER_UDP_WILD: Matching UDP/IPv4 destination (host, port)
+ * @EFX_FILTER_MAC_FULL: Matching Ethernet destination MAC address, VID
+ * @EFX_FILTER_MAC_WILD: Matching Ethernet destination MAC address
+ * @EFX_FILTER_UNSPEC: Match type is unspecified
  *
- * Falcon NICs only support the RX TCP/IPv4 and UDP/IPv4 filter types.
+ * Falcon NICs only support the TCP/IPv4 and UDP/IPv4 filter types.
  */
 enum efx_filter_type {
-	EFX_FILTER_RX_TCP_FULL = 0,
-	EFX_FILTER_RX_TCP_WILD,
-	EFX_FILTER_RX_UDP_FULL,
-	EFX_FILTER_RX_UDP_WILD,
-	EFX_FILTER_RX_MAC_FULL = 4,
-	EFX_FILTER_RX_MAC_WILD,
-	EFX_FILTER_TYPE_COUNT,
+	EFX_FILTER_TCP_FULL = 0,
+	EFX_FILTER_TCP_WILD,
+	EFX_FILTER_UDP_FULL,
+	EFX_FILTER_UDP_WILD,
+	EFX_FILTER_MAC_FULL = 4,
+	EFX_FILTER_MAC_WILD,
+	EFX_FILTER_TYPE_COUNT,		/* number of specific types */
+	EFX_FILTER_UNSPEC = 0xf,
 };
 
 /**
@@ -63,13 +59,13 @@ enum efx_filter_priority {
  * @EFX_FILTER_FLAG_RX_OVERRIDE_IP: Enables a MAC filter to override
  *	any IP filter that matches the same packet.  By default, IP
  *	filters take precedence.
- *
- * Currently, no flags are defined for TX filters.
+ * @EFX_FILTER_FLAG_RX: Filter is for RX
  */
 enum efx_filter_flags {
 	EFX_FILTER_FLAG_RX_RSS = 0x01,
 	EFX_FILTER_FLAG_RX_SCATTER = 0x02,
 	EFX_FILTER_FLAG_RX_OVERRIDE_IP = 0x04,
+	EFX_FILTER_FLAG_RX = 0x08,
 };
 
 /**
@@ -91,99 +87,26 @@ struct efx_filter_spec {
 	u32	data[3];
 };
 
-/**
- * efx_filter_set_rx_tcp_full - specify RX filter with TCP/IPv4 full match
- * @spec: Specification to initialise
- * @shost: Source host address (host byte order)
- * @sport: Source port (host byte order)
- * @dhost: Destination host address (host byte order)
- * @dport: Destination port (host byte order)
- */
-static inline void
-efx_filter_set_rx_tcp_full(struct efx_filter_spec *spec,
-			   u32 shost, u16 sport, u32 dhost, u16 dport)
-{
-	spec->type = EFX_FILTER_RX_TCP_FULL;
-	spec->data[0] = sport | shost << 16;
-	spec->data[1] = dport << 16 | shost >> 16;
-	spec->data[2] = dhost;
-}
-
-/**
- * efx_filter_set_rx_tcp_wild - specify RX filter with TCP/IPv4 wildcard match
- * @spec: Specification to initialise
- * @dhost: Destination host address (host byte order)
- * @dport: Destination port (host byte order)
- */
-static inline void
-efx_filter_set_rx_tcp_wild(struct efx_filter_spec *spec, u32 dhost, u16 dport)
-{
-	spec->type = EFX_FILTER_RX_TCP_WILD;
-	spec->data[0] = 0;
-	spec->data[1] = dport << 16;
-	spec->data[2] = dhost;
-}
-
-/**
- * efx_filter_set_rx_udp_full - specify RX filter with UDP/IPv4 full match
- * @spec: Specification to initialise
- * @shost: Source host address (host byte order)
- * @sport: Source port (host byte order)
- * @dhost: Destination host address (host byte order)
- * @dport: Destination port (host byte order)
- */
-static inline void
-efx_filter_set_rx_udp_full(struct efx_filter_spec *spec,
-			   u32 shost, u16 sport, u32 dhost, u16 dport)
-{
-	spec->type = EFX_FILTER_RX_UDP_FULL;
-	spec->data[0] = sport | shost << 16;
-	spec->data[1] = dport << 16 | shost >> 16;
-	spec->data[2] = dhost;
-}
-
-/**
- * efx_filter_set_rx_udp_wild - specify RX filter with UDP/IPv4 wildcard match
- * @spec: Specification to initialise
- * @dhost: Destination host address (host byte order)
- * @dport: Destination port (host byte order)
- */
-static inline void
-efx_filter_set_rx_udp_wild(struct efx_filter_spec *spec, u32 dhost, u16 dport)
+static inline void efx_filter_init_rx(struct efx_filter_spec *spec,
+				      enum efx_filter_priority priority,
+				      enum efx_filter_flags flags,
+				      unsigned rxq_id)
 {
-	spec->type = EFX_FILTER_RX_UDP_WILD;
-	spec->data[0] = dport;
-	spec->data[1] = 0;
-	spec->data[2] = dhost;
+	spec->type = EFX_FILTER_UNSPEC;
+	spec->priority = priority;
+	spec->flags = EFX_FILTER_FLAG_RX | flags;
+	spec->dmaq_id = rxq_id;
 }
 
-/**
- * efx_filter_set_rx_mac_full - specify RX filter with MAC full match
- * @spec: Specification to initialise
- * @vid: VLAN ID
- * @addr: Destination MAC address
- */
-static inline void efx_filter_set_rx_mac_full(struct efx_filter_spec *spec,
-					      u16 vid, const u8 *addr)
-{
-	spec->type = EFX_FILTER_RX_MAC_FULL;
-	spec->data[0] = vid;
-	spec->data[1] = addr[2] << 24 | addr[3] << 16 | addr[4] << 8 | addr[5];
-	spec->data[2] = addr[0] << 8 | addr[1];
-}
-
-/**
- * efx_filter_set_rx_mac_full - specify RX filter with MAC wildcard match
- * @spec: Specification to initialise
- * @addr: Destination MAC address
- */
-static inline void efx_filter_set_rx_mac_wild(struct efx_filter_spec *spec,
-					      const u8 *addr)
-{
-	spec->type = EFX_FILTER_RX_MAC_WILD;
-	spec->data[0] = 0;
-	spec->data[1] = addr[2] << 24 | addr[3] << 16 | addr[4] << 8 | addr[5];
-	spec->data[2] = addr[0] << 8 | addr[1];
-}
+extern int efx_filter_set_ipv4_local(struct efx_filter_spec *spec, u8 proto,
+				     __be32 host, __be16 port);
+extern int efx_filter_set_ipv4_full(struct efx_filter_spec *spec, u8 proto,
+				    __be32 host, __be16 port,
+				    __be32 rhost, __be16 rport);
+extern int efx_filter_set_eth_local(struct efx_filter_spec *spec,
+				    u16 vid, const u8 *addr);
+enum {
+	EFX_FILTER_VID_UNSPEC = 0xffff,
+};
 
 #endif /* EFX_FILTER_H */

drivers/net/sfc/io.h

@@ -22,28 +22,39 @@
  *
  * Notes on locking strategy:
  *
- * Most NIC registers require 16-byte (or 8-byte, for SRAM) atomic writes
- * which necessitates locking.
- * Under normal operation few writes to NIC registers are made and these
- * registers (EVQ_RPTR_REG, RX_DESC_UPD_REG and TX_DESC_UPD_REG) are special
- * cased to allow 4-byte (hence lockless) accesses.
+ * Most CSRs are 128-bit (oword) and therefore cannot be read or
+ * written atomically.  Access from the host is buffered by the Bus
+ * Interface Unit (BIU).  Whenever the host reads from the lowest
+ * address of such a register, or from the address of a different such
+ * register, the BIU latches the register's value.  Subsequent reads
+ * from higher addresses of the same register will read the latched
+ * value.  Whenever the host writes part of such a register, the BIU
+ * collects the written value and does not write to the underlying
+ * register until all 4 dwords have been written.  A similar buffering
+ * scheme applies to host access to the NIC's 64-bit SRAM.
  *
- * It *is* safe to write to these 4-byte registers in the middle of an
- * access to an 8-byte or 16-byte register.  We therefore use a
- * spinlock to protect accesses to the larger registers, but no locks
- * for the 4-byte registers.
+ * Access to different CSRs and 64-bit SRAM words must be serialised,
+ * since interleaved access can result in lost writes or lost
+ * information from read-to-clear fields.  We use efx_nic::biu_lock
+ * for this.  (We could use separate locks for read and write, but
+ * this is not normally a performance bottleneck.)
  *
- * A write barrier is needed to ensure that DW3 is written after DW0/1/2
- * due to the way the 16byte registers are "collected" in the BIU.
+ * The DMA descriptor pointers (RX_DESC_UPD and TX_DESC_UPD) are
+ * 128-bit but are special-cased in the BIU to avoid the need for
+ * locking in the host:
  *
- * We also lock when carrying out reads, to ensure consistency of the
- * data (made possible since the BIU reads all 128 bits into a cache).
- * Reads are very rare, so this isn't a significant performance
- * impact.  (Most data transferred from NIC to host is DMAed directly
- * into host memory).
- *
- * I/O BAR access uses locks for both reads and writes (but is only provided
- * for testing purposes).
+ * - They are write-only.
+ * - The semantics of writing to these registers are such that
+ *   replacing the low 96 bits with zero does not affect functionality.
+ * - If the host writes to the last dword address of such a register
+ *   (i.e. the high 32 bits) the underlying register will always be
+ *   written.  If the collector does not hold values for the low 96
+ *   bits of the register, they will be written as zero.  Writing to
+ *   the last qword does not have this effect and must not be done.
+ * - If the host writes to the address of any other part of such a
+ *   register while the collector already holds values for some other
+ *   register, the write is discarded and the collector maintains its
+ *   current state.
  */
 
 #if BITS_PER_LONG == 64
@@ -72,7 +83,7 @@ static inline __le32 _efx_readd(struct efx_nic *efx, unsigned int reg)
 	return (__force __le32)__raw_readl(efx->membase + reg);
 }
 
-/* Writes to a normal 16-byte Efx register, locking as appropriate. */
+/* Write a normal 128-bit CSR, locking as appropriate. */
 static inline void efx_writeo(struct efx_nic *efx, efx_oword_t *value,
 			      unsigned int reg)
 {
@@ -85,21 +96,18 @@ static inline void efx_writeo(struct efx_nic *efx, efx_oword_t *value,
 	spin_lock_irqsave(&efx->biu_lock, flags);
 #ifdef EFX_USE_QWORD_IO
 	_efx_writeq(efx, value->u64[0], reg + 0);
-	wmb();
 	_efx_writeq(efx, value->u64[1], reg + 8);
 #else
 	_efx_writed(efx, value->u32[0], reg + 0);
 	_efx_writed(efx, value->u32[1], reg + 4);
 	_efx_writed(efx, value->u32[2], reg + 8);
-	wmb();
 	_efx_writed(efx, value->u32[3], reg + 12);
 #endif
 	mmiowb();
 	spin_unlock_irqrestore(&efx->biu_lock, flags);
 }
 
-/* Write an 8-byte NIC SRAM entry through the supplied mapping,
- * locking as appropriate. */
+/* Write 64-bit SRAM through the supplied mapping, locking as appropriate. */
 static inline void efx_sram_writeq(struct efx_nic *efx, void __iomem *membase,
 				   efx_qword_t *value, unsigned int index)
 {
@@ -115,36 +123,25 @@ static inline void efx_sram_writeq(struct efx_nic *efx, void __iomem *membase,
 	__raw_writeq((__force u64)value->u64[0], membase + addr);
 #else
 	__raw_writel((__force u32)value->u32[0], membase + addr);
-	wmb();
 	__raw_writel((__force u32)value->u32[1], membase + addr + 4);
 #endif
 	mmiowb();
 	spin_unlock_irqrestore(&efx->biu_lock, flags);
 }
 
-/* Write dword to NIC register that allows partial writes
- *
- * Some registers (EVQ_RPTR_REG, RX_DESC_UPD_REG and
- * TX_DESC_UPD_REG) can be written to as a single dword.  This allows
- * for lockless writes.
- */
+/* Write a 32-bit CSR or the last dword of a special 128-bit CSR */
 static inline void efx_writed(struct efx_nic *efx, efx_dword_t *value,
 			      unsigned int reg)
 {
 	netif_vdbg(efx, hw, efx->net_dev,
-		   "writing partial register %x with "EFX_DWORD_FMT"\n",
+		   "writing register %x with "EFX_DWORD_FMT"\n",
 		   reg, EFX_DWORD_VAL(*value));
 
 	/* No lock required */
 	_efx_writed(efx, value->u32[0], reg);
 }
 
-/* Read from a NIC register
- *
- * This reads an entire 16-byte register in one go, locking as
- * appropriate.  It is essential to read the first dword first, as this
- * prompts the NIC to load the current value into the shadow register.
- */
+/* Read a 128-bit CSR, locking as appropriate. */
 static inline void efx_reado(struct efx_nic *efx, efx_oword_t *value,
 			     unsigned int reg)
 {
@@ -152,7 +149,6 @@ static inline void efx_reado(struct efx_nic *efx, efx_oword_t *value,
 
 	spin_lock_irqsave(&efx->biu_lock, flags);
 	value->u32[0] = _efx_readd(efx, reg + 0);
-	rmb();
 	value->u32[1] = _efx_readd(efx, reg + 4);
 	value->u32[2] = _efx_readd(efx, reg + 8);
 	value->u32[3] = _efx_readd(efx, reg + 12);
@@ -163,8 +159,7 @@ static inline void efx_reado(struct efx_nic *efx, efx_oword_t *value,
 		   EFX_OWORD_VAL(*value));
 }
 
-/* Read an 8-byte SRAM entry through supplied mapping,
- * locking as appropriate. */
+/* Read 64-bit SRAM through the supplied mapping, locking as appropriate. */
 static inline void efx_sram_readq(struct efx_nic *efx, void __iomem *membase,
 				  efx_qword_t *value, unsigned int index)
 {
@@ -176,7 +171,6 @@ static inline void efx_sram_readq(struct efx_nic *efx, void __iomem *membase,
 	value->u64[0] = (__force __le64)__raw_readq(membase + addr);
 #else
 	value->u32[0] = (__force __le32)__raw_readl(membase + addr);
-	rmb();
 	value->u32[1] = (__force __le32)__raw_readl(membase + addr + 4);
 #endif
 	spin_unlock_irqrestore(&efx->biu_lock, flags);
@@ -186,7 +180,7 @@ static inline void efx_sram_readq(struct efx_nic *efx, void __iomem *membase,
 		   addr, EFX_QWORD_VAL(*value));
 }
 
-/* Read dword from register that allows partial writes (sic) */
+/* Read a 32-bit CSR or SRAM */
 static inline void efx_readd(struct efx_nic *efx, efx_dword_t *value,
 				unsigned int reg)
 {
@@ -196,28 +190,28 @@ static inline void efx_readd(struct efx_nic *efx, efx_dword_t *value,
 		   reg, EFX_DWORD_VAL(*value));
 }
 
-/* Write to a register forming part of a table */
+/* Write a 128-bit CSR forming part of a table */
 static inline void efx_writeo_table(struct efx_nic *efx, efx_oword_t *value,
 				      unsigned int reg, unsigned int index)
 {
 	efx_writeo(efx, value, reg + index * sizeof(efx_oword_t));
 }
 
-/* Read to a register forming part of a table */
+/* Read a 128-bit CSR forming part of a table */
 static inline void efx_reado_table(struct efx_nic *efx, efx_oword_t *value,
 				     unsigned int reg, unsigned int index)
 {
 	efx_reado(efx, value, reg + index * sizeof(efx_oword_t));
 }
 
-/* Write to a dword register forming part of a table */
+/* Write a 32-bit CSR forming part of a table, or 32-bit SRAM */
 static inline void efx_writed_table(struct efx_nic *efx, efx_dword_t *value,
 				       unsigned int reg, unsigned int index)
 {
 	efx_writed(efx, value, reg + index * sizeof(efx_oword_t));
 }
 
-/* Read from a dword register forming part of a table */
+/* Read a 32-bit CSR forming part of a table, or 32-bit SRAM */
 static inline void efx_readd_table(struct efx_nic *efx, efx_dword_t *value,
 				   unsigned int reg, unsigned int index)
 {
@@ -231,26 +225,51 @@ static inline void efx_readd_table(struct efx_nic *efx, efx_dword_t *value,
 #define EFX_PAGED_REG(page, reg) \
 	((page) * EFX_PAGE_BLOCK_SIZE + (reg))
 
-/* As for efx_writeo(), but for a page-mapped register. */
-static inline void efx_writeo_page(struct efx_nic *efx, efx_oword_t *value,
+/* Write the whole of RX_DESC_UPD or TX_DESC_UPD */
+static inline void _efx_writeo_page(struct efx_nic *efx, efx_oword_t *value,
 				    unsigned int reg, unsigned int page)
 {
-	efx_writeo(efx, value, EFX_PAGED_REG(page, reg));
-}
+	reg = EFX_PAGED_REG(page, reg);
 
-/* As for efx_writed(), but for a page-mapped register. */
-static inline void efx_writed_page(struct efx_nic *efx, efx_dword_t *value,
+	netif_vdbg(efx, hw, efx->net_dev,
+		   "writing register %x with " EFX_OWORD_FMT "\n", reg,
+		   EFX_OWORD_VAL(*value));
+
+#ifdef EFX_USE_QWORD_IO
+	_efx_writeq(efx, value->u64[0], reg + 0);
+#else
+	_efx_writed(efx, value->u32[0], reg + 0);
+	_efx_writed(efx, value->u32[1], reg + 4);
+#endif
+	_efx_writed(efx, value->u32[2], reg + 8);
+	_efx_writed(efx, value->u32[3], reg + 12);
+}
+#define efx_writeo_page(efx, value, reg, page)				\
+	_efx_writeo_page(efx, value,					\
+			 reg +						\
+			 BUILD_BUG_ON_ZERO((reg) != 0x830 && (reg) != 0xa10), \
+			 page)
+
+/* Write a page-mapped 32-bit CSR (EVQ_RPTR or the high bits of
+ * RX_DESC_UPD or TX_DESC_UPD)
+ */
+static inline void _efx_writed_page(struct efx_nic *efx, efx_dword_t *value,
 				    unsigned int reg, unsigned int page)
 {
 	efx_writed(efx, value, EFX_PAGED_REG(page, reg));
 }
-
-/* Write dword to page-mapped register with an extra lock.
- *
- * As for efx_writed_page(), but for a register that suffers from
- * SFC bug 3181. Take out a lock so the BIU collector cannot be
- * confused. */
-static inline void efx_writed_page_locked(struct efx_nic *efx,
+#define efx_writed_page(efx, value, reg, page)				\
+	_efx_writed_page(efx, value,					\
+			 reg +						\
+			 BUILD_BUG_ON_ZERO((reg) != 0x400 && (reg) != 0x83c \
+					   && (reg) != 0xa1c),		\
+			 page)
+
+/* Write TIMER_COMMAND.  This is a page-mapped 32-bit CSR, but a bug
+ * in the BIU means that writes to TIMER_COMMAND[0] invalidate the
+ * collector register.
+ */
+static inline void _efx_writed_page_locked(struct efx_nic *efx,
 					   efx_dword_t *value,
 					   unsigned int reg,
 					   unsigned int page)
@@ -265,5 +284,9 @@ static inline void efx_writed_page_locked(struct efx_nic *efx,
 		efx_writed(efx, value, EFX_PAGED_REG(page, reg));
 	}
 }
+#define efx_writed_page_locked(efx, value, reg, page)			\
+	_efx_writed_page_locked(efx, value,				\
+				reg + BUILD_BUG_ON_ZERO((reg) != 0x420), \
+				page)
 
 #endif /* EFX_IO_H */

drivers/net/sfc/net_driver.h

@@ -142,6 +142,12 @@ struct efx_tx_buffer {
  * @flushed: Used when handling queue flushing
  * @read_count: Current read pointer.
  *	This is the number of buffers that have been removed from both rings.
+ * @old_write_count: The value of @write_count when last checked.
+ *	This is here for performance reasons.  The xmit path will
+ *	only get the up-to-date value of @write_count if this
+ *	variable indicates that the queue is empty.  This is to
+ *	avoid cache-line ping-pong between the xmit path and the
+ *	completion path.
  * @stopped: Stopped count.
  *	Set if this TX queue is currently stopping its port.
  * @insert_count: Current insert pointer
@@ -163,6 +169,10 @@ struct efx_tx_buffer {
  * @tso_long_headers: Number of packets with headers too long for standard
  *	blocks
  * @tso_packets: Number of packets via the TSO xmit path
+ * @pushes: Number of times the TX push feature has been used
+ * @empty_read_count: If the completion path has seen the queue as empty
+ *	and the transmission path has not yet checked this, the value of
+ *	@read_count bitwise-added to %EFX_EMPTY_COUNT_VALID; otherwise 0.
  */
 struct efx_tx_queue {
 	/* Members which don't change on the fast path */
@@ -177,6 +187,7 @@ struct efx_tx_queue {
 
 	/* Members used mainly on the completion path */
 	unsigned int read_count ____cacheline_aligned_in_smp;
+	unsigned int old_write_count;
 	int stopped;
 
 	/* Members used only on the xmit path */
@@ -187,6 +198,11 @@ struct efx_tx_queue {
 	unsigned int tso_bursts;
 	unsigned int tso_long_headers;
 	unsigned int tso_packets;
+	unsigned int pushes;
+
+	/* Members shared between paths and sometimes updated */
+	unsigned int empty_read_count ____cacheline_aligned_in_smp;
+#define EFX_EMPTY_COUNT_VALID 0x80000000
 };
 
 /**
@@ -626,10 +642,8 @@ struct efx_filter_state;
  *	Work items do not hold and must not acquire RTNL.
  * @workqueue_name: Name of workqueue
  * @reset_work: Scheduled reset workitem
- * @monitor_work: Hardware monitor workitem
  * @membase_phys: Memory BAR value as physical address
  * @membase: Memory BAR value
- * @biu_lock: BIU (bus interface unit) lock
  * @interrupt_mode: Interrupt mode
  * @irq_rx_adaptive: Adaptive IRQ moderation enabled for RX event queues
  * @irq_rx_moderation: IRQ moderation time for RX event queues
@@ -653,14 +667,9 @@ struct efx_filter_state;
  * @int_error_count: Number of internal errors seen recently
  * @int_error_expire: Time at which error count will be expired
  * @irq_status: Interrupt status buffer
- * @last_irq_cpu: Last CPU to handle interrupt.
- *	This register is written with the SMP processor ID whenever an
- *	interrupt is handled.  It is used by efx_nic_test_interrupt()
- *	to verify that an interrupt has occurred.
  * @irq_zero_count: Number of legacy IRQs seen with queue flags == 0
  * @fatal_irq_level: IRQ level (bit number) used for serious errors
  * @mtd_list: List of MTDs attached to the NIC
- * @n_rx_nodesc_drop_cnt: RX no descriptor drop count
  * @nic_data: Hardware dependant state
  * @mac_lock: MAC access lock. Protects @port_enabled, @phy_mode,
  *	@port_inhibited, efx_monitor() and efx_reconfigure_port()
@@ -673,11 +682,7 @@ struct efx_filter_state;
  * @port_initialized: Port initialized?
  * @net_dev: Operating system network device. Consider holding the rtnl lock
  * @rx_checksum_enabled: RX checksumming enabled
- * @mac_stats: MAC statistics. These include all statistics the MACs
- *	can provide.  Generic code converts these into a standard
- *	&struct net_device_stats.
  * @stats_buffer: DMA buffer for statistics
- * @stats_lock: Statistics update lock. Serialises statistics fetches
  * @mac_op: MAC interface
  * @phy_type: PHY type
  * @phy_op: PHY interface
@@ -695,10 +700,23 @@ struct efx_filter_state;
  * @loopback_mode: Loopback status
  * @loopback_modes: Supported loopback mode bitmask
  * @loopback_selftest: Offline self-test private state
+ * @monitor_work: Hardware monitor workitem
+ * @biu_lock: BIU (bus interface unit) lock
+ * @last_irq_cpu: Last CPU to handle interrupt.
+ *	This register is written with the SMP processor ID whenever an
+ *	interrupt is handled.  It is used by efx_nic_test_interrupt()
+ *	to verify that an interrupt has occurred.
+ * @n_rx_nodesc_drop_cnt: RX no descriptor drop count
+ * @mac_stats: MAC statistics. These include all statistics the MACs
+ *	can provide.  Generic code converts these into a standard
+ *	&struct net_device_stats.
+ * @stats_lock: Statistics update lock. Serialises statistics fetches
  *
  * This is stored in the private area of the &struct net_device.
  */
 struct efx_nic {
+	/* The following fields should be written very rarely */
+
 	char name[IFNAMSIZ];
 	struct pci_dev *pci_dev;
 	const struct efx_nic_type *type;
@@ -707,10 +725,9 @@ struct efx_nic {
 	struct workqueue_struct *workqueue;
 	char workqueue_name[16];
 	struct work_struct reset_work;
-	struct delayed_work monitor_work;
 	resource_size_t membase_phys;
 	void __iomem *membase;
-	spinlock_t biu_lock;
+
 	enum efx_int_mode interrupt_mode;
 	bool irq_rx_adaptive;
 	unsigned int irq_rx_moderation;
@@ -737,7 +754,6 @@ struct efx_nic {
 	unsigned long int_error_expire;
 
 	struct efx_buffer irq_status;
-	volatile signed int last_irq_cpu;
 	unsigned irq_zero_count;
 	unsigned fatal_irq_level;
 
@@ -745,8 +761,6 @@ struct efx_nic {
 	struct list_head mtd_list;
 #endif
 
-	unsigned n_rx_nodesc_drop_cnt;
-
 	void *nic_data;
 
 	struct mutex mac_lock;
@@ -758,9 +772,7 @@ struct efx_nic {
 	struct net_device *net_dev;
 	bool rx_checksum_enabled;
 
-	struct efx_mac_stats mac_stats;
 	struct efx_buffer stats_buffer;
-	spinlock_t stats_lock;
 
 	struct efx_mac_operations *mac_op;
 
@@ -786,6 +798,15 @@ struct efx_nic {
 	void *loopback_selftest;
 
 	struct efx_filter_state *filter_state;
+
+	/* The following fields may be written more often */
+
+	struct delayed_work monitor_work ____cacheline_aligned_in_smp;
+	spinlock_t biu_lock;
+	volatile signed int last_irq_cpu;
+	unsigned n_rx_nodesc_drop_cnt;
+	struct efx_mac_stats mac_stats;
+	spinlock_t stats_lock;
 };
 
 static inline int efx_dev_registered(struct efx_nic *efx)

drivers/net/sfc/nic.c

@@ -362,6 +362,35 @@ static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
 			FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
 }
 
+/* Write pointer and first descriptor for TX descriptor ring */
+static inline void efx_push_tx_desc(struct efx_tx_queue *tx_queue,
+				    const efx_qword_t *txd)
+{
+	unsigned write_ptr;
+	efx_oword_t reg;
+
+	BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
+	BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
+
+	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
+	EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
+			     FRF_AZ_TX_DESC_WPTR, write_ptr);
+	reg.qword[0] = *txd;
+	efx_writeo_page(tx_queue->efx, &reg,
+			FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
+}
+
+static inline bool
+efx_may_push_tx_desc(struct efx_tx_queue *tx_queue, unsigned int write_count)
+{
+	unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
+
+	if (empty_read_count == 0)
+		return false;
+
+	tx_queue->empty_read_count = 0;
+	return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0;
+}
 
 /* For each entry inserted into the software descriptor ring, create a
  * descriptor in the hardware TX descriptor ring (in host memory), and
@@ -373,6 +402,7 @@ void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
 	struct efx_tx_buffer *buffer;
 	efx_qword_t *txd;
 	unsigned write_ptr;
+	unsigned old_write_count = tx_queue->write_count;
 
 	BUG_ON(tx_queue->write_count == tx_queue->insert_count);
 
@@ -391,7 +421,15 @@ void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
 	} while (tx_queue->write_count != tx_queue->insert_count);
 
 	wmb(); /* Ensure descriptors are written before they are fetched */
+
+	if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
+		txd = efx_tx_desc(tx_queue,
+				  old_write_count & tx_queue->ptr_mask);
+		efx_push_tx_desc(tx_queue, txd);
+		++tx_queue->pushes;
+	} else {
 		efx_notify_tx_desc(tx_queue);
+	}
 }
 
 /* Allocate hardware resources for a TX queue */
@@ -1632,7 +1670,7 @@ void efx_nic_init_common(struct efx_nic *efx)
 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
-	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 0);
+	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
 	/* Enable SW_EV to inherit in char driver - assume harmless here */
 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);

drivers/net/sfc/tx.c

@@ -240,8 +240,7 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 				 * of read_count. */
 				smp_mb();
 				tx_queue->old_read_count =
-					*(volatile unsigned *)
-					&tx_queue->read_count;
+					ACCESS_ONCE(tx_queue->read_count);
 				fill_level = (tx_queue->insert_count
 					      - tx_queue->old_read_count);
 				q_space = efx->txq_entries - 1 - fill_level;
@@ -429,6 +428,16 @@ void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 			__netif_tx_unlock(queue);
 		}
 	}
+
+	/* Check whether the hardware queue is now empty */
+	if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
+		tx_queue->old_write_count = ACCESS_ONCE(tx_queue->write_count);
+		if (tx_queue->read_count == tx_queue->old_write_count) {
+			smp_mb();
+			tx_queue->empty_read_count =
+				tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
+		}
+	}
 }
 
 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
@@ -474,8 +483,10 @@ void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
 
 	tx_queue->insert_count = 0;
 	tx_queue->write_count = 0;
+	tx_queue->old_write_count = 0;
 	tx_queue->read_count = 0;
 	tx_queue->old_read_count = 0;
+	tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
 	BUG_ON(tx_queue->stopped);
 
 	/* Set up TX descriptor ring */
@@ -764,7 +775,7 @@ static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
 			 * stopped from the access of read_count. */
 			smp_mb();
 			tx_queue->old_read_count =
-				*(volatile unsigned *)&tx_queue->read_count;
+				ACCESS_ONCE(tx_queue->read_count);
 			fill_level = (tx_queue->insert_count
 				      - tx_queue->old_read_count);
 			q_space = efx->txq_entries - 1 - fill_level;