net/sfc/base: optimize credits overflow check
[dpdk.git] / drivers / net / sfc / sfc_flow.c
index 6b20bae..f2050f6 100644 (file)
@@ -1,4 +1,6 @@
 /*-
+ *   BSD LICENSE
+ *
  * Copyright (c) 2017 Solarflare Communications Inc.
  * All rights reserved.
  *
@@ -55,6 +57,8 @@ enum sfc_flow_item_layers {
        SFC_FLOW_ITEM_ANY_LAYER,
        SFC_FLOW_ITEM_START_LAYER,
        SFC_FLOW_ITEM_L2,
+       SFC_FLOW_ITEM_L3,
+       SFC_FLOW_ITEM_L4,
 };
 
 typedef int (sfc_flow_item_parse)(const struct rte_flow_item *item,
@@ -70,6 +74,11 @@ struct sfc_flow_item {
 
 static sfc_flow_item_parse sfc_flow_parse_void;
 static sfc_flow_item_parse sfc_flow_parse_eth;
+static sfc_flow_item_parse sfc_flow_parse_vlan;
+static sfc_flow_item_parse sfc_flow_parse_ipv4;
+static sfc_flow_item_parse sfc_flow_parse_ipv6;
+static sfc_flow_item_parse sfc_flow_parse_tcp;
+static sfc_flow_item_parse sfc_flow_parse_udp;
 
 static boolean_t
 sfc_flow_is_zero(const uint8_t *buf, unsigned int size)
@@ -198,8 +207,10 @@ sfc_flow_parse_void(__rte_unused const struct rte_flow_item *item,
  *
  * @param item[in]
  *   Item specification. Only source and destination addresses and
- *   Ethernet type fields are supported. If the mask is NULL, default
- *   mask will be used. Ranging is not supported.
+ *   Ethernet type fields are supported. In addition to full and
+ *   empty masks of destination address, individual/group mask is
+ *   also supported. If the mask is NULL, default mask will be used.
+ *   Ranging is not supported.
  * @param efx_spec[in, out]
  *   EFX filter specification to update.
  * @param[out] error
@@ -218,6 +229,9 @@ sfc_flow_parse_eth(const struct rte_flow_item *item,
                .src.addr_bytes = { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff },
                .type = 0xffff,
        };
+       const uint8_t ig_mask[EFX_MAC_ADDR_LEN] = {
+               0x01, 0x00, 0x00, 0x00, 0x00, 0x00
+       };
 
        rc = sfc_flow_parse_init(item,
                                 (const void **)&spec,
@@ -237,6 +251,14 @@ sfc_flow_parse_eth(const struct rte_flow_item *item,
                efx_spec->efs_match_flags |= EFX_FILTER_MATCH_LOC_MAC;
                rte_memcpy(efx_spec->efs_loc_mac, spec->dst.addr_bytes,
                           EFX_MAC_ADDR_LEN);
+       } else if (memcmp(mask->dst.addr_bytes, ig_mask,
+                         EFX_MAC_ADDR_LEN) == 0) {
+               if (is_unicast_ether_addr(&spec->dst))
+                       efx_spec->efs_match_flags |=
+                               EFX_FILTER_MATCH_UNKNOWN_UCAST_DST;
+               else
+                       efx_spec->efs_match_flags |=
+                               EFX_FILTER_MATCH_UNKNOWN_MCAST_DST;
        } else if (!is_zero_ether_addr(&mask->dst)) {
                goto fail_bad_mask;
        }
@@ -269,6 +291,433 @@ fail_bad_mask:
        return -rte_errno;
 }
 
+/**
+ * Convert VLAN item to EFX filter specification.
+ *
+ * @param item[in]
+ *   Item specification. Only VID field is supported.
+ *   The mask can not be NULL. Ranging is not supported.
+ * @param efx_spec[in, out]
+ *   EFX filter specification to update.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ */
+static int
+sfc_flow_parse_vlan(const struct rte_flow_item *item,
+                   efx_filter_spec_t *efx_spec,
+                   struct rte_flow_error *error)
+{
+       int rc;
+       uint16_t vid;
+       const struct rte_flow_item_vlan *spec = NULL;
+       const struct rte_flow_item_vlan *mask = NULL;
+       const struct rte_flow_item_vlan supp_mask = {
+               .tci = rte_cpu_to_be_16(ETH_VLAN_ID_MAX),
+       };
+
+       rc = sfc_flow_parse_init(item,
+                                (const void **)&spec,
+                                (const void **)&mask,
+                                &supp_mask,
+                                NULL,
+                                sizeof(struct rte_flow_item_vlan),
+                                error);
+       if (rc != 0)
+               return rc;
+
+       /*
+        * VID is in big-endian byte order in item and
+        * in little-endian in efx_spec, so byte swap is used.
+        * If two VLAN items are included, the first matches
+        * the outer tag and the next matches the inner tag.
+        */
+       if (mask->tci == supp_mask.tci) {
+               vid = rte_bswap16(spec->tci);
+
+               if (!(efx_spec->efs_match_flags &
+                     EFX_FILTER_MATCH_OUTER_VID)) {
+                       efx_spec->efs_match_flags |= EFX_FILTER_MATCH_OUTER_VID;
+                       efx_spec->efs_outer_vid = vid;
+               } else if (!(efx_spec->efs_match_flags &
+                            EFX_FILTER_MATCH_INNER_VID)) {
+                       efx_spec->efs_match_flags |= EFX_FILTER_MATCH_INNER_VID;
+                       efx_spec->efs_inner_vid = vid;
+               } else {
+                       rte_flow_error_set(error, EINVAL,
+                                          RTE_FLOW_ERROR_TYPE_ITEM, item,
+                                          "More than two VLAN items");
+                       return -rte_errno;
+               }
+       } else {
+               rte_flow_error_set(error, EINVAL,
+                                  RTE_FLOW_ERROR_TYPE_ITEM, item,
+                                  "VLAN ID in TCI match is required");
+               return -rte_errno;
+       }
+
+       return 0;
+}
+
+/**
+ * Convert IPv4 item to EFX filter specification.
+ *
+ * @param item[in]
+ *   Item specification. Only source and destination addresses and
+ *   protocol fields are supported. If the mask is NULL, default
+ *   mask will be used. Ranging is not supported.
+ * @param efx_spec[in, out]
+ *   EFX filter specification to update.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ */
+static int
+sfc_flow_parse_ipv4(const struct rte_flow_item *item,
+                   efx_filter_spec_t *efx_spec,
+                   struct rte_flow_error *error)
+{
+       int rc;
+       const struct rte_flow_item_ipv4 *spec = NULL;
+       const struct rte_flow_item_ipv4 *mask = NULL;
+       const uint16_t ether_type_ipv4 = rte_cpu_to_le_16(EFX_ETHER_TYPE_IPV4);
+       const struct rte_flow_item_ipv4 supp_mask = {
+               .hdr = {
+                       .src_addr = 0xffffffff,
+                       .dst_addr = 0xffffffff,
+                       .next_proto_id = 0xff,
+               }
+       };
+
+       rc = sfc_flow_parse_init(item,
+                                (const void **)&spec,
+                                (const void **)&mask,
+                                &supp_mask,
+                                &rte_flow_item_ipv4_mask,
+                                sizeof(struct rte_flow_item_ipv4),
+                                error);
+       if (rc != 0)
+               return rc;
+
+       /*
+        * Filtering by IPv4 source and destination addresses requires
+        * the appropriate ETHER_TYPE in hardware filters
+        */
+       if (!(efx_spec->efs_match_flags & EFX_FILTER_MATCH_ETHER_TYPE)) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_ETHER_TYPE;
+               efx_spec->efs_ether_type = ether_type_ipv4;
+       } else if (efx_spec->efs_ether_type != ether_type_ipv4) {
+               rte_flow_error_set(error, EINVAL,
+                       RTE_FLOW_ERROR_TYPE_ITEM, item,
+                       "Ethertype in pattern with IPV4 item should be appropriate");
+               return -rte_errno;
+       }
+
+       if (spec == NULL)
+               return 0;
+
+       /*
+        * IPv4 addresses are in big-endian byte order in item and in
+        * efx_spec
+        */
+       if (mask->hdr.src_addr == supp_mask.hdr.src_addr) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_REM_HOST;
+               efx_spec->efs_rem_host.eo_u32[0] = spec->hdr.src_addr;
+       } else if (mask->hdr.src_addr != 0) {
+               goto fail_bad_mask;
+       }
+
+       if (mask->hdr.dst_addr == supp_mask.hdr.dst_addr) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_LOC_HOST;
+               efx_spec->efs_loc_host.eo_u32[0] = spec->hdr.dst_addr;
+       } else if (mask->hdr.dst_addr != 0) {
+               goto fail_bad_mask;
+       }
+
+       if (mask->hdr.next_proto_id == supp_mask.hdr.next_proto_id) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_IP_PROTO;
+               efx_spec->efs_ip_proto = spec->hdr.next_proto_id;
+       } else if (mask->hdr.next_proto_id != 0) {
+               goto fail_bad_mask;
+       }
+
+       return 0;
+
+fail_bad_mask:
+       rte_flow_error_set(error, EINVAL,
+                          RTE_FLOW_ERROR_TYPE_ITEM, item,
+                          "Bad mask in the IPV4 pattern item");
+       return -rte_errno;
+}
+
+/**
+ * Convert IPv6 item to EFX filter specification.
+ *
+ * @param item[in]
+ *   Item specification. Only source and destination addresses and
+ *   next header fields are supported. If the mask is NULL, default
+ *   mask will be used. Ranging is not supported.
+ * @param efx_spec[in, out]
+ *   EFX filter specification to update.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ */
+static int
+sfc_flow_parse_ipv6(const struct rte_flow_item *item,
+                   efx_filter_spec_t *efx_spec,
+                   struct rte_flow_error *error)
+{
+       int rc;
+       const struct rte_flow_item_ipv6 *spec = NULL;
+       const struct rte_flow_item_ipv6 *mask = NULL;
+       const uint16_t ether_type_ipv6 = rte_cpu_to_le_16(EFX_ETHER_TYPE_IPV6);
+       const struct rte_flow_item_ipv6 supp_mask = {
+               .hdr = {
+                       .src_addr = { 0xff, 0xff, 0xff, 0xff,
+                                     0xff, 0xff, 0xff, 0xff,
+                                     0xff, 0xff, 0xff, 0xff,
+                                     0xff, 0xff, 0xff, 0xff },
+                       .dst_addr = { 0xff, 0xff, 0xff, 0xff,
+                                     0xff, 0xff, 0xff, 0xff,
+                                     0xff, 0xff, 0xff, 0xff,
+                                     0xff, 0xff, 0xff, 0xff },
+                       .proto = 0xff,
+               }
+       };
+
+       rc = sfc_flow_parse_init(item,
+                                (const void **)&spec,
+                                (const void **)&mask,
+                                &supp_mask,
+                                &rte_flow_item_ipv6_mask,
+                                sizeof(struct rte_flow_item_ipv6),
+                                error);
+       if (rc != 0)
+               return rc;
+
+       /*
+        * Filtering by IPv6 source and destination addresses requires
+        * the appropriate ETHER_TYPE in hardware filters
+        */
+       if (!(efx_spec->efs_match_flags & EFX_FILTER_MATCH_ETHER_TYPE)) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_ETHER_TYPE;
+               efx_spec->efs_ether_type = ether_type_ipv6;
+       } else if (efx_spec->efs_ether_type != ether_type_ipv6) {
+               rte_flow_error_set(error, EINVAL,
+                       RTE_FLOW_ERROR_TYPE_ITEM, item,
+                       "Ethertype in pattern with IPV6 item should be appropriate");
+               return -rte_errno;
+       }
+
+       if (spec == NULL)
+               return 0;
+
+       /*
+        * IPv6 addresses are in big-endian byte order in item and in
+        * efx_spec
+        */
+       if (memcmp(mask->hdr.src_addr, supp_mask.hdr.src_addr,
+                  sizeof(mask->hdr.src_addr)) == 0) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_REM_HOST;
+
+               RTE_BUILD_BUG_ON(sizeof(efx_spec->efs_rem_host) !=
+                                sizeof(spec->hdr.src_addr));
+               rte_memcpy(&efx_spec->efs_rem_host, spec->hdr.src_addr,
+                          sizeof(efx_spec->efs_rem_host));
+       } else if (!sfc_flow_is_zero(mask->hdr.src_addr,
+                                    sizeof(mask->hdr.src_addr))) {
+               goto fail_bad_mask;
+       }
+
+       if (memcmp(mask->hdr.dst_addr, supp_mask.hdr.dst_addr,
+                  sizeof(mask->hdr.dst_addr)) == 0) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_LOC_HOST;
+
+               RTE_BUILD_BUG_ON(sizeof(efx_spec->efs_loc_host) !=
+                                sizeof(spec->hdr.dst_addr));
+               rte_memcpy(&efx_spec->efs_loc_host, spec->hdr.dst_addr,
+                          sizeof(efx_spec->efs_loc_host));
+       } else if (!sfc_flow_is_zero(mask->hdr.dst_addr,
+                                    sizeof(mask->hdr.dst_addr))) {
+               goto fail_bad_mask;
+       }
+
+       if (mask->hdr.proto == supp_mask.hdr.proto) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_IP_PROTO;
+               efx_spec->efs_ip_proto = spec->hdr.proto;
+       } else if (mask->hdr.proto != 0) {
+               goto fail_bad_mask;
+       }
+
+       return 0;
+
+fail_bad_mask:
+       rte_flow_error_set(error, EINVAL,
+                          RTE_FLOW_ERROR_TYPE_ITEM, item,
+                          "Bad mask in the IPV6 pattern item");
+       return -rte_errno;
+}
+
+/**
+ * Convert TCP item to EFX filter specification.
+ *
+ * @param item[in]
+ *   Item specification. Only source and destination ports fields
+ *   are supported. If the mask is NULL, default mask will be used.
+ *   Ranging is not supported.
+ * @param efx_spec[in, out]
+ *   EFX filter specification to update.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ */
+static int
+sfc_flow_parse_tcp(const struct rte_flow_item *item,
+                  efx_filter_spec_t *efx_spec,
+                  struct rte_flow_error *error)
+{
+       int rc;
+       const struct rte_flow_item_tcp *spec = NULL;
+       const struct rte_flow_item_tcp *mask = NULL;
+       const struct rte_flow_item_tcp supp_mask = {
+               .hdr = {
+                       .src_port = 0xffff,
+                       .dst_port = 0xffff,
+               }
+       };
+
+       rc = sfc_flow_parse_init(item,
+                                (const void **)&spec,
+                                (const void **)&mask,
+                                &supp_mask,
+                                &rte_flow_item_tcp_mask,
+                                sizeof(struct rte_flow_item_tcp),
+                                error);
+       if (rc != 0)
+               return rc;
+
+       /*
+        * Filtering by TCP source and destination ports requires
+        * the appropriate IP_PROTO in hardware filters
+        */
+       if (!(efx_spec->efs_match_flags & EFX_FILTER_MATCH_IP_PROTO)) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_IP_PROTO;
+               efx_spec->efs_ip_proto = EFX_IPPROTO_TCP;
+       } else if (efx_spec->efs_ip_proto != EFX_IPPROTO_TCP) {
+               rte_flow_error_set(error, EINVAL,
+                       RTE_FLOW_ERROR_TYPE_ITEM, item,
+                       "IP proto in pattern with TCP item should be appropriate");
+               return -rte_errno;
+       }
+
+       if (spec == NULL)
+               return 0;
+
+       /*
+        * Source and destination ports are in big-endian byte order in item and
+        * in little-endian in efx_spec, so byte swap is used
+        */
+       if (mask->hdr.src_port == supp_mask.hdr.src_port) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_REM_PORT;
+               efx_spec->efs_rem_port = rte_bswap16(spec->hdr.src_port);
+       } else if (mask->hdr.src_port != 0) {
+               goto fail_bad_mask;
+       }
+
+       if (mask->hdr.dst_port == supp_mask.hdr.dst_port) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_LOC_PORT;
+               efx_spec->efs_loc_port = rte_bswap16(spec->hdr.dst_port);
+       } else if (mask->hdr.dst_port != 0) {
+               goto fail_bad_mask;
+       }
+
+       return 0;
+
+fail_bad_mask:
+       rte_flow_error_set(error, EINVAL,
+                          RTE_FLOW_ERROR_TYPE_ITEM, item,
+                          "Bad mask in the TCP pattern item");
+       return -rte_errno;
+}
+
+/**
+ * Convert UDP item to EFX filter specification.
+ *
+ * @param item[in]
+ *   Item specification. Only source and destination ports fields
+ *   are supported. If the mask is NULL, default mask will be used.
+ *   Ranging is not supported.
+ * @param efx_spec[in, out]
+ *   EFX filter specification to update.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ */
+static int
+sfc_flow_parse_udp(const struct rte_flow_item *item,
+                  efx_filter_spec_t *efx_spec,
+                  struct rte_flow_error *error)
+{
+       int rc;
+       const struct rte_flow_item_udp *spec = NULL;
+       const struct rte_flow_item_udp *mask = NULL;
+       const struct rte_flow_item_udp supp_mask = {
+               .hdr = {
+                       .src_port = 0xffff,
+                       .dst_port = 0xffff,
+               }
+       };
+
+       rc = sfc_flow_parse_init(item,
+                                (const void **)&spec,
+                                (const void **)&mask,
+                                &supp_mask,
+                                &rte_flow_item_udp_mask,
+                                sizeof(struct rte_flow_item_udp),
+                                error);
+       if (rc != 0)
+               return rc;
+
+       /*
+        * Filtering by UDP source and destination ports requires
+        * the appropriate IP_PROTO in hardware filters
+        */
+       if (!(efx_spec->efs_match_flags & EFX_FILTER_MATCH_IP_PROTO)) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_IP_PROTO;
+               efx_spec->efs_ip_proto = EFX_IPPROTO_UDP;
+       } else if (efx_spec->efs_ip_proto != EFX_IPPROTO_UDP) {
+               rte_flow_error_set(error, EINVAL,
+                       RTE_FLOW_ERROR_TYPE_ITEM, item,
+                       "IP proto in pattern with UDP item should be appropriate");
+               return -rte_errno;
+       }
+
+       if (spec == NULL)
+               return 0;
+
+       /*
+        * Source and destination ports are in big-endian byte order in item and
+        * in little-endian in efx_spec, so byte swap is used
+        */
+       if (mask->hdr.src_port == supp_mask.hdr.src_port) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_REM_PORT;
+               efx_spec->efs_rem_port = rte_bswap16(spec->hdr.src_port);
+       } else if (mask->hdr.src_port != 0) {
+               goto fail_bad_mask;
+       }
+
+       if (mask->hdr.dst_port == supp_mask.hdr.dst_port) {
+               efx_spec->efs_match_flags |= EFX_FILTER_MATCH_LOC_PORT;
+               efx_spec->efs_loc_port = rte_bswap16(spec->hdr.dst_port);
+       } else if (mask->hdr.dst_port != 0) {
+               goto fail_bad_mask;
+       }
+
+       return 0;
+
+fail_bad_mask:
+       rte_flow_error_set(error, EINVAL,
+                          RTE_FLOW_ERROR_TYPE_ITEM, item,
+                          "Bad mask in the UDP pattern item");
+       return -rte_errno;
+}
+
 static const struct sfc_flow_item sfc_flow_items[] = {
        {
                .type = RTE_FLOW_ITEM_TYPE_VOID,
@@ -282,6 +731,36 @@ static const struct sfc_flow_item sfc_flow_items[] = {
                .layer = SFC_FLOW_ITEM_L2,
                .parse = sfc_flow_parse_eth,
        },
+       {
+               .type = RTE_FLOW_ITEM_TYPE_VLAN,
+               .prev_layer = SFC_FLOW_ITEM_L2,
+               .layer = SFC_FLOW_ITEM_L2,
+               .parse = sfc_flow_parse_vlan,
+       },
+       {
+               .type = RTE_FLOW_ITEM_TYPE_IPV4,
+               .prev_layer = SFC_FLOW_ITEM_L2,
+               .layer = SFC_FLOW_ITEM_L3,
+               .parse = sfc_flow_parse_ipv4,
+       },
+       {
+               .type = RTE_FLOW_ITEM_TYPE_IPV6,
+               .prev_layer = SFC_FLOW_ITEM_L2,
+               .layer = SFC_FLOW_ITEM_L3,
+               .parse = sfc_flow_parse_ipv6,
+       },
+       {
+               .type = RTE_FLOW_ITEM_TYPE_TCP,
+               .prev_layer = SFC_FLOW_ITEM_L3,
+               .layer = SFC_FLOW_ITEM_L4,
+               .parse = sfc_flow_parse_tcp,
+       },
+       {
+               .type = RTE_FLOW_ITEM_TYPE_UDP,
+               .prev_layer = SFC_FLOW_ITEM_L3,
+               .layer = SFC_FLOW_ITEM_L4,
+               .parse = sfc_flow_parse_udp,
+       },
 };
 
 /*
@@ -324,7 +803,7 @@ sfc_flow_parse_attr(const struct rte_flow_attr *attr,
        }
 
        flow->spec.efs_flags |= EFX_FILTER_FLAG_RX;
-       flow->spec.efs_rss_context = EFX_FILTER_SPEC_RSS_CONTEXT_DEFAULT;
+       flow->spec.efs_rss_context = EFX_RSS_CONTEXT_DEFAULT;
 
        return 0;
 }
@@ -358,8 +837,7 @@ sfc_flow_parse_pattern(const struct rte_flow_item pattern[],
                return -rte_errno;
        }
 
-       for (; pattern != NULL &&
-              pattern->type != RTE_FLOW_ITEM_TYPE_END; pattern++) {
+       for (; pattern->type != RTE_FLOW_ITEM_TYPE_END; pattern++) {
                item = sfc_flow_get_item(pattern->type);
                if (item == NULL) {
                        rte_flow_error_set(error, ENOTSUP,
@@ -389,13 +867,6 @@ sfc_flow_parse_pattern(const struct rte_flow_item pattern[],
                        prev_layer = item->layer;
        }
 
-       if (pattern == NULL) {
-               rte_flow_error_set(error, EINVAL,
-                                  RTE_FLOW_ERROR_TYPE_ITEM, NULL,
-                                  "NULL item");
-               return -rte_errno;
-       }
-
        return 0;
 }
 
@@ -415,6 +886,170 @@ sfc_flow_parse_queue(struct sfc_adapter *sa,
        return 0;
 }
 
+#if EFSYS_OPT_RX_SCALE
+static int
+sfc_flow_parse_rss(struct sfc_adapter *sa,
+                  const struct rte_flow_action_rss *rss,
+                  struct rte_flow *flow)
+{
+       unsigned int rxq_sw_index;
+       struct sfc_rxq *rxq;
+       unsigned int rxq_hw_index_min;
+       unsigned int rxq_hw_index_max;
+       const struct rte_eth_rss_conf *rss_conf = rss->rss_conf;
+       uint64_t rss_hf;
+       uint8_t *rss_key = NULL;
+       struct sfc_flow_rss *sfc_rss_conf = &flow->rss_conf;
+       unsigned int i;
+
+       if (rss->num == 0)
+               return -EINVAL;
+
+       rxq_sw_index = sa->rxq_count - 1;
+       rxq = sa->rxq_info[rxq_sw_index].rxq;
+       rxq_hw_index_min = rxq->hw_index;
+       rxq_hw_index_max = 0;
+
+       for (i = 0; i < rss->num; ++i) {
+               rxq_sw_index = rss->queue[i];
+
+               if (rxq_sw_index >= sa->rxq_count)
+                       return -EINVAL;
+
+               rxq = sa->rxq_info[rxq_sw_index].rxq;
+
+               if (rxq->hw_index < rxq_hw_index_min)
+                       rxq_hw_index_min = rxq->hw_index;
+
+               if (rxq->hw_index > rxq_hw_index_max)
+                       rxq_hw_index_max = rxq->hw_index;
+       }
+
+       rss_hf = (rss_conf != NULL) ? rss_conf->rss_hf : SFC_RSS_OFFLOADS;
+       if ((rss_hf & ~SFC_RSS_OFFLOADS) != 0)
+               return -EINVAL;
+
+       if (rss_conf != NULL) {
+               if (rss_conf->rss_key_len != sizeof(sa->rss_key))
+                       return -EINVAL;
+
+               rss_key = rss_conf->rss_key;
+       } else {
+               rss_key = sa->rss_key;
+       }
+
+       flow->rss = B_TRUE;
+
+       sfc_rss_conf->rxq_hw_index_min = rxq_hw_index_min;
+       sfc_rss_conf->rxq_hw_index_max = rxq_hw_index_max;
+       sfc_rss_conf->rss_hash_types = sfc_rte_to_efx_hash_type(rss_hf);
+       rte_memcpy(sfc_rss_conf->rss_key, rss_key, sizeof(sa->rss_key));
+
+       for (i = 0; i < RTE_DIM(sfc_rss_conf->rss_tbl); ++i) {
+               unsigned int rxq_sw_index = rss->queue[i % rss->num];
+               struct sfc_rxq *rxq = sa->rxq_info[rxq_sw_index].rxq;
+
+               sfc_rss_conf->rss_tbl[i] = rxq->hw_index - rxq_hw_index_min;
+       }
+
+       return 0;
+}
+#endif /* EFSYS_OPT_RX_SCALE */
+
+static int
+sfc_flow_filter_insert(struct sfc_adapter *sa,
+                      struct rte_flow *flow)
+{
+       efx_filter_spec_t *spec = &flow->spec;
+
+#if EFSYS_OPT_RX_SCALE
+       struct sfc_flow_rss *rss = &flow->rss_conf;
+       int rc = 0;
+
+       if (flow->rss) {
+               unsigned int rss_spread = MIN(rss->rxq_hw_index_max -
+                                             rss->rxq_hw_index_min + 1,
+                                             EFX_MAXRSS);
+
+               rc = efx_rx_scale_context_alloc(sa->nic,
+                                               EFX_RX_SCALE_EXCLUSIVE,
+                                               rss_spread,
+                                               &spec->efs_rss_context);
+               if (rc != 0)
+                       goto fail_scale_context_alloc;
+
+               rc = efx_rx_scale_mode_set(sa->nic, spec->efs_rss_context,
+                                          EFX_RX_HASHALG_TOEPLITZ,
+                                          rss->rss_hash_types, B_TRUE);
+               if (rc != 0)
+                       goto fail_scale_mode_set;
+
+               rc = efx_rx_scale_key_set(sa->nic, spec->efs_rss_context,
+                                         rss->rss_key,
+                                         sizeof(sa->rss_key));
+               if (rc != 0)
+                       goto fail_scale_key_set;
+
+               spec->efs_dmaq_id = rss->rxq_hw_index_min;
+               spec->efs_flags |= EFX_FILTER_FLAG_RX_RSS;
+       }
+
+       rc = efx_filter_insert(sa->nic, spec);
+       if (rc != 0)
+               goto fail_filter_insert;
+
+       if (flow->rss) {
+               /*
+                * Scale table is set after filter insertion because
+                * the table entries are relative to the base RxQ ID
+                * and the latter is submitted to the HW by means of
+                * inserting a filter, so by the time of the request
+                * the HW knows all the information needed to verify
+                * the table entries, and the operation will succeed
+                */
+               rc = efx_rx_scale_tbl_set(sa->nic, spec->efs_rss_context,
+                                         rss->rss_tbl, RTE_DIM(rss->rss_tbl));
+               if (rc != 0)
+                       goto fail_scale_tbl_set;
+       }
+
+       return 0;
+
+fail_scale_tbl_set:
+       efx_filter_remove(sa->nic, spec);
+
+fail_filter_insert:
+fail_scale_key_set:
+fail_scale_mode_set:
+       if (rss != NULL)
+               efx_rx_scale_context_free(sa->nic, spec->efs_rss_context);
+
+fail_scale_context_alloc:
+       return rc;
+#else /* !EFSYS_OPT_RX_SCALE */
+       return efx_filter_insert(sa->nic, spec);
+#endif /* EFSYS_OPT_RX_SCALE */
+}
+
+static int
+sfc_flow_filter_remove(struct sfc_adapter *sa,
+                      struct rte_flow *flow)
+{
+       efx_filter_spec_t *spec = &flow->spec;
+       int rc = 0;
+
+       rc = efx_filter_remove(sa->nic, spec);
+       if (rc != 0)
+               return rc;
+
+#if EFSYS_OPT_RX_SCALE
+       if (flow->rss)
+               rc = efx_rx_scale_context_free(sa->nic, spec->efs_rss_context);
+#endif /* EFSYS_OPT_RX_SCALE */
+
+       return rc;
+}
+
 static int
 sfc_flow_parse_actions(struct sfc_adapter *sa,
                       const struct rte_flow_action actions[],
@@ -448,6 +1083,20 @@ sfc_flow_parse_actions(struct sfc_adapter *sa,
                        is_specified = B_TRUE;
                        break;
 
+#if EFSYS_OPT_RX_SCALE
+               case RTE_FLOW_ACTION_TYPE_RSS:
+                       rc = sfc_flow_parse_rss(sa, actions->conf, flow);
+                       if (rc != 0) {
+                               rte_flow_error_set(error, rc,
+                                       RTE_FLOW_ERROR_TYPE_ACTION, actions,
+                                       "Bad RSS action");
+                               return -rte_errno;
+                       }
+
+                       is_specified = B_TRUE;
+                       break;
+#endif /* EFSYS_OPT_RX_SCALE */
+
                default:
                        rte_flow_error_set(error, ENOTSUP,
                                           RTE_FLOW_ERROR_TYPE_ACTION, actions,
@@ -542,7 +1191,7 @@ sfc_flow_create(struct rte_eth_dev *dev,
        sfc_adapter_lock(sa);
 
        if (sa->state == SFC_ADAPTER_STARTED) {
-               rc = efx_filter_insert(sa->nic, &flow->spec);
+               rc = sfc_flow_filter_insert(sa, flow);
                if (rc != 0) {
                        rte_flow_error_set(error, rc,
                                RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
@@ -576,7 +1225,7 @@ sfc_flow_remove(struct sfc_adapter *sa,
        SFC_ASSERT(sfc_adapter_is_locked(sa));
 
        if (sa->state == SFC_ADAPTER_STARTED) {
-               rc = efx_filter_remove(sa->nic, &flow->spec);
+               rc = sfc_flow_filter_remove(sa, flow);
                if (rc != 0)
                        rte_flow_error_set(error, rc,
                                RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
@@ -641,12 +1290,35 @@ sfc_flow_flush(struct rte_eth_dev *dev,
        return -ret;
 }
 
+static int
+sfc_flow_isolate(struct rte_eth_dev *dev, int enable,
+                struct rte_flow_error *error)
+{
+       struct sfc_adapter *sa = dev->data->dev_private;
+       struct sfc_port *port = &sa->port;
+       int ret = 0;
+
+       sfc_adapter_lock(sa);
+       if (sa->state != SFC_ADAPTER_INITIALIZED) {
+               rte_flow_error_set(error, EBUSY,
+                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                  NULL, "please close the port first");
+               ret = -rte_errno;
+       } else {
+               port->isolated = (enable) ? B_TRUE : B_FALSE;
+       }
+       sfc_adapter_unlock(sa);
+
+       return ret;
+}
+
 const struct rte_flow_ops sfc_flow_ops = {
        .validate = sfc_flow_validate,
        .create = sfc_flow_create,
        .destroy = sfc_flow_destroy,
        .flush = sfc_flow_flush,
        .query = NULL,
+       .isolate = sfc_flow_isolate,
 };
 
 void
@@ -678,7 +1350,7 @@ sfc_flow_stop(struct sfc_adapter *sa)
        SFC_ASSERT(sfc_adapter_is_locked(sa));
 
        TAILQ_FOREACH(flow, &sa->filter.flow_list, entries)
-               efx_filter_remove(sa->nic, &flow->spec);
+               sfc_flow_filter_remove(sa, flow);
 }
 
 int
@@ -692,7 +1364,7 @@ sfc_flow_start(struct sfc_adapter *sa)
        SFC_ASSERT(sfc_adapter_is_locked(sa));
 
        TAILQ_FOREACH(flow, &sa->filter.flow_list, entries) {
-               rc = efx_filter_insert(sa->nic, &flow->spec);
+               rc = sfc_flow_filter_insert(sa, flow);
                if (rc != 0)
                        goto fail_bad_flow;
        }