net/tap: add basic flow API patterns and actions
authorPascal Mazon <pascal.mazon@6wind.com>
Thu, 23 Mar 2017 08:33:57 +0000 (09:33 +0100)
committerFerruh Yigit <ferruh.yigit@intel.com>
Tue, 4 Apr 2017 16:59:46 +0000 (18:59 +0200)
Supported flow rules are now mapped to TC rules on the tap netdevice.
The netlink message used for creating the TC rule is stored in struct
rte_flow. That way, by simply changing a metadata in it, we can require
for the rule deletion without further parsing.

Supported items:
- eth: src and dst (with variable masks), and eth_type (0xffff mask).
- vlan: vid, pcp, tpid, but not eid.
- ipv4/6: src and dst (with variable masks), and ip_proto (0xffff mask).
- udp/tcp: src and dst port (0xffff) mask.

Supported actions:
- DROP
- QUEUE
- PASSTHRU

It is generally not possible to provide a "last" item. However, if the
"last" item, once masked, is identical to the masked spec, then it is
supported.

Only IPv4/6 and MAC addresses can use a variable mask. All other
items need a full mask (exact match).

Support for VLAN requires kernel headers >= 4.9, checked using
auto-config.sh.

Signed-off-by: Pascal Mazon <pascal.mazon@6wind.com>
Acked-by: Olga Shern <olgas@mellanox.com>
Acked-by: Keith Wiles <keith.wiles@intel.com>
doc/guides/nics/tap.rst
drivers/net/tap/Makefile
drivers/net/tap/rte_eth_tap.c
drivers/net/tap/rte_eth_tap.h
drivers/net/tap/tap_flow.c
drivers/net/tap/tap_flow.h

index c4f207b..4986e47 100644 (file)
@@ -82,6 +82,51 @@ can utilize that stack to handle the network protocols. Plus you would be able
 to address the interface using an IP address assigned to the internal
 interface.
 
+Flow API support
+----------------
+
+The tap PMD supports major flow API pattern items and actions, when running on
+linux kernels above 4.2 ("Flower" classifier required). Supported items:
+
+- eth: src and dst (with variable masks), and eth_type (0xffff mask).
+- vlan: vid, pcp, tpid, but not eid. (requires kernel 4.9)
+- ipv4/6: src and dst (with variable masks), and ip_proto (0xffff mask).
+- udp/tcp: src and dst port (0xffff) mask.
+
+Supported actions:
+
+- DROP
+- QUEUE
+- PASSTHRU
+
+It is generally not possible to provide a "last" item. However, if the "last"
+item, once masked, is identical to the masked spec, then it is supported.
+
+Only IPv4/6 and MAC addresses can use a variable mask. All other items need a
+full mask (exact match).
+
+As rules are translated to TC, it is possible to show them with something like::
+
+   tc -s filter show dev tap1 parent 1:
+
+Examples of testpmd flow rules
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Drop packets for destination IP 192.168.0.1::
+
+   testpmd> flow create 0 priority 1 ingress pattern eth / ipv4 dst is 1.1.1.1 \
+            / end actions drop / end
+
+Ensure packets from a given MAC address are received on a queue 2::
+
+   testpmd> flow create 0 priority 2 ingress pattern eth src is 06:05:04:03:02:01 \
+            / end actions queue index 2 / end
+
+Drop UDP packets in vlan 3::
+
+   testpmd> flow create 0 priority 3 ingress pattern eth / vlan vid is 3 / \
+            ipv4 proto is 17 / end actions drop / end
+
 Example
 -------
 
index 3a33b56..c42a680 100644 (file)
@@ -41,6 +41,7 @@ LIBABIVER := 1
 
 CFLAGS += -O3
 CFLAGS += -I$(SRCDIR)
+CFLAGS += -I.
 CFLAGS += $(WERROR_FLAGS)
 
 #
@@ -58,5 +59,44 @@ DEPDIRS-$(CONFIG_RTE_LIBRTE_PMD_TAP) += lib/librte_mempool
 DEPDIRS-$(CONFIG_RTE_LIBRTE_PMD_TAP) += lib/librte_ether
 DEPDIRS-$(CONFIG_RTE_LIBRTE_PMD_TAP) += lib/librte_kvargs
 DEPDIRS-$(CONFIG_RTE_LIBRTE_PMD_TAP) += lib/librte_net
+DEPDIRS-$(CONFIG_RTE_LIBRTE_PMD_TAP) += lib/librte_hash
 
 include $(RTE_SDK)/mk/rte.lib.mk
+
+# Generate and clean-up tap_autoconf.h.
+
+export CC CFLAGS CPPFLAGS EXTRA_CFLAGS EXTRA_CPPFLAGS
+export AUTO_CONFIG_CFLAGS = -Wno-error
+
+ifndef V
+AUTOCONF_OUTPUT := >/dev/null
+endif
+
+tap_autoconf.h.new: FORCE
+
+tap_autoconf.h.new: $(RTE_SDK)/buildtools/auto-config-h.sh
+       $Q $(RM) -f -- '$@'
+       $Q sh -- '$<' '$@' \
+               HAVE_TC_FLOWER \
+               linux/pkt_cls.h \
+               enum TCA_FLOWER_UNSPEC \
+               $(AUTOCONF_OUTPUT)
+       $Q sh -- '$<' '$@' \
+               HAVE_TC_VLAN_ID \
+               linux/pkt_cls.h \
+               enum TCA_FLOWER_KEY_VLAN_PRIO \
+               $(AUTOCONF_OUTPUT)
+
+# Create tap_autoconf.h or update it in case it differs from the new one.
+
+tap_autoconf.h: tap_autoconf.h.new
+       $Q [ -f '$@' ] && \
+               cmp '$<' '$@' $(AUTOCONF_OUTPUT) || \
+               mv '$<' '$@'
+
+$(SRCS-$(CONFIG_RTE_LIBRTE_PMD_TAP):.c=.o): tap_autoconf.h
+
+clean_tap: FORCE
+       $Q rm -f -- tap_autoconf.h tap_autoconf.h.new
+
+clean: clean_tap
index 9127c73..c711b36 100644 (file)
 #include <sys/stat.h>
 #include <sys/socket.h>
 #include <sys/ioctl.h>
+#include <sys/utsname.h>
 #include <sys/mman.h>
 #include <errno.h>
 #include <signal.h>
 #include <stdint.h>
 #include <unistd.h>
 #include <arpa/inet.h>
-#include <linux/if.h>
+#include <net/if.h>
 #include <linux/if_tun.h>
 #include <linux/if_ether.h>
+#include <linux/version.h>
 #include <fcntl.h>
 
 #include <rte_eth_tap.h>
 #include <tap_flow.h>
+#include <tap_tcmsgs.h>
 
 /* Linux based path to the TUN device */
 #define TUN_TAP_DEV_PATH        "/dev/net/tun"
@@ -71,6 +74,9 @@
 #define RTE_PMD_TAP_MAX_QUEUES 1
 #endif
 
+#define FLOWER_KERNEL_VERSION KERNEL_VERSION(4, 2, 0)
+#define FLOWER_VLAN_KERNEL_VERSION KERNEL_VERSION(4, 9, 0)
+
 static struct rte_vdev_driver pmd_tap_drv;
 
 static const char *valid_arguments[] = {
@@ -209,6 +215,28 @@ tun_alloc(struct pmd_internals *pmd, uint16_t qid)
                        goto error;
                rte_memcpy(&pmd->eth_addr, ifr.ifr_hwaddr.sa_data,
                           ETHER_ADDR_LEN);
+
+               pmd->if_index = if_nametoindex(pmd->name);
+               if (!pmd->if_index) {
+                       RTE_LOG(ERR, PMD,
+                               "Could not find ifindex for %s: rte_flow won't be usable.\n",
+                               pmd->name);
+                       return fd;
+               }
+               if (!pmd->flower_support)
+                       return fd;
+               if (qdisc_create_multiq(pmd->nlsk_fd, pmd->if_index) < 0) {
+                       RTE_LOG(ERR, PMD,
+                               "Could not create multiq qdisc for %s: rte_flow won't be usable.\n",
+                               pmd->name);
+                       return fd;
+               }
+               if (qdisc_create_ingress(pmd->nlsk_fd, pmd->if_index) < 0) {
+                       RTE_LOG(ERR, PMD,
+                               "Could not create multiq qdisc for %s: rte_flow won't be usable.\n",
+                               pmd->name);
+                       return fd;
+               }
        }
 
        return fd;
@@ -811,6 +839,24 @@ static const struct eth_dev_ops ops = {
        .filter_ctrl            = tap_dev_filter_ctrl,
 };
 
+static int
+tap_kernel_support(struct pmd_internals *pmd)
+{
+       struct utsname utsname;
+       int ver[3];
+
+       if (uname(&utsname) == -1 ||
+           sscanf(utsname.release, "%d.%d.%d",
+                  &ver[0], &ver[1], &ver[2]) != 3)
+               return 0;
+       if (KERNEL_VERSION(ver[0], ver[1], ver[2]) >= FLOWER_KERNEL_VERSION)
+               pmd->flower_support = 1;
+       if (KERNEL_VERSION(ver[0], ver[1], ver[2]) >=
+           FLOWER_VLAN_KERNEL_VERSION)
+               pmd->flower_vlan_support = 1;
+       return 1;
+}
+
 static int
 eth_dev_tap_create(const char *name, char *tap_name)
 {
@@ -880,7 +926,15 @@ eth_dev_tap_create(const char *name, char *tap_name)
                pmd->txq[i].fd = -1;
        }
 
+       tap_kernel_support(pmd);
+       if (!pmd->flower_support)
+               return 0;
        LIST_INIT(&pmd->flows);
+       /*
+        * If no netlink socket can be created, then it will fail when
+        * creating/destroying flow rules.
+        */
+       pmd->nlsk_fd = nl_init();
 
        return 0;
 
@@ -995,7 +1049,10 @@ rte_pmd_tap_remove(const char *name)
                return 0;
 
        internals = eth_dev->data->dev_private;
-       tap_flow_flush(eth_dev, NULL);
+       if (internals->flower_support && internals->nlsk_fd) {
+               tap_flow_flush(eth_dev, NULL);
+               nl_final(internals->nlsk_fd);
+       }
        for (i = 0; i < internals->nb_queues; i++)
                if (internals->rxq[i].fd != -1)
                        close(internals->rxq[i].fd);
index bf82267..741ec53 100644 (file)
@@ -69,6 +69,9 @@ struct pmd_internals {
        struct ether_addr eth_addr;       /* Mac address of the device port */
        int if_index;                     /* IF_INDEX for the port */
        int ioctl_sock;                   /* socket for ioctl calls */
+       int nlsk_fd;                      /* Netlink socket fd */
+       int flower_support;               /* 1 if kernel supports, else 0 */
+       int flower_vlan_support;          /* 1 if kernel supports, else 0 */
        LIST_HEAD(tap_flows, rte_flow) flows;        /* rte_flow rules */
        struct rx_queue rxq[RTE_PMD_TAP_MAX_QUEUES]; /* List of RX queues */
        struct tx_queue txq[RTE_PMD_TAP_MAX_QUEUES]; /* List of TX queues */
index c32ed38..6adacdc 100644 (file)
 
 #include <sys/queue.h>
 
+#include <rte_byteorder.h>
+#include <rte_jhash.h>
 #include <rte_malloc.h>
 #include <rte_eth_tap.h>
 #include <tap_flow.h>
+#include <tap_autoconf.h>
+#include <tap_tcmsgs.h>
+
+#ifndef HAVE_TC_FLOWER
+/*
+ * For kernels < 4.2, this enum is not defined. Runtime checks will be made to
+ * avoid sending TC messages the kernel cannot understand.
+ */
+enum {
+       TCA_FLOWER_UNSPEC,
+       TCA_FLOWER_CLASSID,
+       TCA_FLOWER_INDEV,
+       TCA_FLOWER_ACT,
+       TCA_FLOWER_KEY_ETH_DST,         /* ETH_ALEN */
+       TCA_FLOWER_KEY_ETH_DST_MASK,    /* ETH_ALEN */
+       TCA_FLOWER_KEY_ETH_SRC,         /* ETH_ALEN */
+       TCA_FLOWER_KEY_ETH_SRC_MASK,    /* ETH_ALEN */
+       TCA_FLOWER_KEY_ETH_TYPE,        /* be16 */
+       TCA_FLOWER_KEY_IP_PROTO,        /* u8 */
+       TCA_FLOWER_KEY_IPV4_SRC,        /* be32 */
+       TCA_FLOWER_KEY_IPV4_SRC_MASK,   /* be32 */
+       TCA_FLOWER_KEY_IPV4_DST,        /* be32 */
+       TCA_FLOWER_KEY_IPV4_DST_MASK,   /* be32 */
+       TCA_FLOWER_KEY_IPV6_SRC,        /* struct in6_addr */
+       TCA_FLOWER_KEY_IPV6_SRC_MASK,   /* struct in6_addr */
+       TCA_FLOWER_KEY_IPV6_DST,        /* struct in6_addr */
+       TCA_FLOWER_KEY_IPV6_DST_MASK,   /* struct in6_addr */
+       TCA_FLOWER_KEY_TCP_SRC,         /* be16 */
+       TCA_FLOWER_KEY_TCP_DST,         /* be16 */
+       TCA_FLOWER_KEY_UDP_SRC,         /* be16 */
+       TCA_FLOWER_KEY_UDP_DST,         /* be16 */
+};
+#endif
+#ifndef HAVE_TC_VLAN_ID
+enum {
+       /* TCA_FLOWER_FLAGS, */
+       TCA_FLOWER_KEY_VLAN_ID = TCA_FLOWER_KEY_UDP_DST + 2, /* be16 */
+       TCA_FLOWER_KEY_VLAN_PRIO,       /* u8   */
+       TCA_FLOWER_KEY_VLAN_ETH_TYPE,   /* be16 */
+};
+#endif
 
 struct rte_flow {
        LIST_ENTRY(rte_flow) next; /* Pointer to the next rte_flow structure */
+       struct nlmsg msg;
+};
+
+struct convert_data {
+       uint16_t eth_type;
+       uint16_t ip_proto;
+       uint8_t vlan;
+       struct rte_flow *flow;
 };
 
+static int tap_flow_create_eth(const struct rte_flow_item *item, void *data);
+static int tap_flow_create_vlan(const struct rte_flow_item *item, void *data);
+static int tap_flow_create_ipv4(const struct rte_flow_item *item, void *data);
+static int tap_flow_create_ipv6(const struct rte_flow_item *item, void *data);
+static int tap_flow_create_udp(const struct rte_flow_item *item, void *data);
+static int tap_flow_create_tcp(const struct rte_flow_item *item, void *data);
 static int
 tap_flow_validate(struct rte_eth_dev *dev,
                  const struct rte_flow_attr *attr,
@@ -67,6 +124,752 @@ static const struct rte_flow_ops tap_flow_ops = {
        .flush = tap_flow_flush,
 };
 
+/* Static initializer for items. */
+#define ITEMS(...) \
+       (const enum rte_flow_item_type []){ \
+               __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
+       }
+
+/* Structure to generate a simple graph of layers supported by the NIC. */
+struct tap_flow_items {
+       /* Bit-mask corresponding to what is supported for this item. */
+       const void *mask;
+       const unsigned int mask_sz; /* Bit-mask size in bytes. */
+       /*
+        * Bit-mask corresponding to the default mask, if none is provided
+        * along with the item.
+        */
+       const void *default_mask;
+       /**
+        * Conversion function from rte_flow to netlink attributes.
+        *
+        * @param item
+        *   rte_flow item to convert.
+        * @param data
+        *   Internal structure to store the conversion.
+        *
+        * @return
+        *   0 on success, negative value otherwise.
+        */
+       int (*convert)(const struct rte_flow_item *item, void *data);
+       /** List of possible following items.  */
+       const enum rte_flow_item_type *const items;
+};
+
+/* Graph of supported items and associated actions. */
+static const struct tap_flow_items tap_flow_items[] = {
+       [RTE_FLOW_ITEM_TYPE_END] = {
+               .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
+       },
+       [RTE_FLOW_ITEM_TYPE_ETH] = {
+               .items = ITEMS(
+                       RTE_FLOW_ITEM_TYPE_VLAN,
+                       RTE_FLOW_ITEM_TYPE_IPV4,
+                       RTE_FLOW_ITEM_TYPE_IPV6),
+               .mask = &(const struct rte_flow_item_eth){
+                       .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
+                       .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
+                       .type = -1,
+               },
+               .mask_sz = sizeof(struct rte_flow_item_eth),
+               .default_mask = &rte_flow_item_eth_mask,
+               .convert = tap_flow_create_eth,
+       },
+       [RTE_FLOW_ITEM_TYPE_VLAN] = {
+               .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
+                              RTE_FLOW_ITEM_TYPE_IPV6),
+               .mask = &(const struct rte_flow_item_vlan){
+                       .tpid = -1,
+                       /* DEI matching is not supported */
+#if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
+                       .tci = 0xffef,
+#else
+                       .tci = 0xefff,
+#endif
+               },
+               .mask_sz = sizeof(struct rte_flow_item_vlan),
+               .default_mask = &rte_flow_item_vlan_mask,
+               .convert = tap_flow_create_vlan,
+       },
+       [RTE_FLOW_ITEM_TYPE_IPV4] = {
+               .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
+                              RTE_FLOW_ITEM_TYPE_TCP),
+               .mask = &(const struct rte_flow_item_ipv4){
+                       .hdr = {
+                               .src_addr = -1,
+                               .dst_addr = -1,
+                               .next_proto_id = -1,
+                       },
+               },
+               .mask_sz = sizeof(struct rte_flow_item_ipv4),
+               .default_mask = &rte_flow_item_ipv4_mask,
+               .convert = tap_flow_create_ipv4,
+       },
+       [RTE_FLOW_ITEM_TYPE_IPV6] = {
+               .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
+                              RTE_FLOW_ITEM_TYPE_TCP),
+               .mask = &(const struct rte_flow_item_ipv6){
+                       .hdr = {
+                               .src_addr = {
+                                       "\xff\xff\xff\xff\xff\xff\xff\xff"
+                                       "\xff\xff\xff\xff\xff\xff\xff\xff",
+                               },
+                               .dst_addr = {
+                                       "\xff\xff\xff\xff\xff\xff\xff\xff"
+                                       "\xff\xff\xff\xff\xff\xff\xff\xff",
+                               },
+                               .proto = -1,
+                       },
+               },
+               .mask_sz = sizeof(struct rte_flow_item_ipv6),
+               .default_mask = &rte_flow_item_ipv6_mask,
+               .convert = tap_flow_create_ipv6,
+       },
+       [RTE_FLOW_ITEM_TYPE_UDP] = {
+               .mask = &(const struct rte_flow_item_udp){
+                       .hdr = {
+                               .src_port = -1,
+                               .dst_port = -1,
+                       },
+               },
+               .mask_sz = sizeof(struct rte_flow_item_udp),
+               .default_mask = &rte_flow_item_udp_mask,
+               .convert = tap_flow_create_udp,
+       },
+       [RTE_FLOW_ITEM_TYPE_TCP] = {
+               .mask = &(const struct rte_flow_item_tcp){
+                       .hdr = {
+                               .src_port = -1,
+                               .dst_port = -1,
+                       },
+               },
+               .mask_sz = sizeof(struct rte_flow_item_tcp),
+               .default_mask = &rte_flow_item_tcp_mask,
+               .convert = tap_flow_create_tcp,
+       },
+};
+
+/**
+ * Make as much checks as possible on an Ethernet item, and if a flow is
+ * provided, fill it appropriately with Ethernet info.
+ *
+ * @param[in] item
+ *   Item specification.
+ * @param[in, out] data
+ *   Additional data structure to tell next layers we've been here.
+ *
+ * @return
+ *   0 if checks are alright, -1 otherwise.
+ */
+static int
+tap_flow_create_eth(const struct rte_flow_item *item, void *data)
+{
+       struct convert_data *info = (struct convert_data *)data;
+       const struct rte_flow_item_eth *spec = item->spec;
+       const struct rte_flow_item_eth *mask = item->mask;
+       struct rte_flow *flow = info->flow;
+       struct nlmsg *msg;
+
+       /* use default mask if none provided */
+       if (!mask)
+               mask = tap_flow_items[RTE_FLOW_ITEM_TYPE_ETH].default_mask;
+       /* TC does not support eth_type masking. Only accept if exact match. */
+       if (mask->type && mask->type != 0xffff)
+               return -1;
+       if (!spec)
+               return 0;
+       /* store eth_type for consistency if ipv4/6 pattern item comes next */
+       if (spec->type & mask->type)
+               info->eth_type = spec->type;
+       if (!flow)
+               return 0;
+       msg = &flow->msg;
+       if (spec->type & mask->type)
+               msg->t.tcm_info = TC_H_MAKE(msg->t.tcm_info,
+                                           (spec->type & mask->type));
+       if (!is_zero_ether_addr(&spec->dst)) {
+               nlattr_add(&msg->nh, TCA_FLOWER_KEY_ETH_DST, ETHER_ADDR_LEN,
+                          &spec->dst.addr_bytes);
+               nlattr_add(&msg->nh,
+                          TCA_FLOWER_KEY_ETH_DST_MASK, ETHER_ADDR_LEN,
+                          &mask->dst.addr_bytes);
+       }
+       if (!is_zero_ether_addr(&mask->src)) {
+               nlattr_add(&msg->nh, TCA_FLOWER_KEY_ETH_SRC, ETHER_ADDR_LEN,
+                          &spec->src.addr_bytes);
+               nlattr_add(&msg->nh,
+                          TCA_FLOWER_KEY_ETH_SRC_MASK, ETHER_ADDR_LEN,
+                          &mask->src.addr_bytes);
+       }
+       return 0;
+}
+
+/**
+ * Make as much checks as possible on a VLAN item, and if a flow is provided,
+ * fill it appropriately with VLAN info.
+ *
+ * @param[in] item
+ *   Item specification.
+ * @param[in, out] data
+ *   Additional data structure to tell next layers we've been here.
+ *
+ * @return
+ *   0 if checks are alright, -1 otherwise.
+ */
+static int
+tap_flow_create_vlan(const struct rte_flow_item *item, void *data)
+{
+       struct convert_data *info = (struct convert_data *)data;
+       const struct rte_flow_item_vlan *spec = item->spec;
+       const struct rte_flow_item_vlan *mask = item->mask;
+       struct rte_flow *flow = info->flow;
+       struct nlmsg *msg;
+
+       /* use default mask if none provided */
+       if (!mask)
+               mask = tap_flow_items[RTE_FLOW_ITEM_TYPE_VLAN].default_mask;
+       /* TC does not support tpid masking. Only accept if exact match. */
+       if (mask->tpid && mask->tpid != 0xffff)
+               return -1;
+       /* Double-tagging not supported. */
+       if (spec && mask->tpid && spec->tpid != htons(ETH_P_8021Q))
+               return -1;
+       info->vlan = 1;
+       if (!flow)
+               return 0;
+       msg = &flow->msg;
+       msg->t.tcm_info = TC_H_MAKE(msg->t.tcm_info, htons(ETH_P_8021Q));
+#define VLAN_PRIO(tci) ((tci) >> 13)
+#define VLAN_ID(tci) ((tci) & 0xfff)
+       if (!spec)
+               return 0;
+       if (spec->tci) {
+               uint16_t tci = ntohs(spec->tci) & mask->tci;
+               uint16_t prio = VLAN_PRIO(tci);
+               uint8_t vid = VLAN_ID(tci);
+
+               if (prio)
+                       nlattr_add8(&msg->nh, TCA_FLOWER_KEY_VLAN_PRIO, prio);
+               if (vid)
+                       nlattr_add16(&msg->nh, TCA_FLOWER_KEY_VLAN_ID, vid);
+       }
+       return 0;
+}
+
+/**
+ * Make as much checks as possible on an IPv4 item, and if a flow is provided,
+ * fill it appropriately with IPv4 info.
+ *
+ * @param[in] item
+ *   Item specification.
+ * @param[in, out] data
+ *   Additional data structure to tell next layers we've been here.
+ *
+ * @return
+ *   0 if checks are alright, -1 otherwise.
+ */
+static int
+tap_flow_create_ipv4(const struct rte_flow_item *item, void *data)
+{
+       struct convert_data *info = (struct convert_data *)data;
+       const struct rte_flow_item_ipv4 *spec = item->spec;
+       const struct rte_flow_item_ipv4 *mask = item->mask;
+       struct rte_flow *flow = info->flow;
+       struct nlmsg *msg;
+
+       /* use default mask if none provided */
+       if (!mask)
+               mask = tap_flow_items[RTE_FLOW_ITEM_TYPE_IPV4].default_mask;
+       /* check that previous eth type is compatible with ipv4 */
+       if (info->eth_type && info->eth_type != htons(ETH_P_IP))
+               return -1;
+       /* store ip_proto for consistency if udp/tcp pattern item comes next */
+       if (spec)
+               info->ip_proto = spec->hdr.next_proto_id;
+       if (!flow)
+               return 0;
+       msg = &flow->msg;
+       if (!info->eth_type)
+               info->eth_type = htons(ETH_P_IP);
+       if (!info->vlan)
+               msg->t.tcm_info = TC_H_MAKE(msg->t.tcm_info, htons(ETH_P_IP));
+       if (!spec)
+               return 0;
+       if (spec->hdr.dst_addr) {
+               nlattr_add32(&msg->nh, TCA_FLOWER_KEY_IPV4_DST,
+                            spec->hdr.dst_addr);
+               nlattr_add32(&msg->nh, TCA_FLOWER_KEY_IPV4_DST_MASK,
+                            mask->hdr.dst_addr);
+       }
+       if (spec->hdr.src_addr) {
+               nlattr_add32(&msg->nh, TCA_FLOWER_KEY_IPV4_SRC,
+                            spec->hdr.src_addr);
+               nlattr_add32(&msg->nh, TCA_FLOWER_KEY_IPV4_SRC_MASK,
+                            mask->hdr.src_addr);
+       }
+       if (spec->hdr.next_proto_id)
+               nlattr_add8(&msg->nh, TCA_FLOWER_KEY_IP_PROTO,
+                           spec->hdr.next_proto_id);
+       return 0;
+}
+
+/**
+ * Make as much checks as possible on an IPv6 item, and if a flow is provided,
+ * fill it appropriately with IPv6 info.
+ *
+ * @param[in] item
+ *   Item specification.
+ * @param[in, out] data
+ *   Additional data structure to tell next layers we've been here.
+ *
+ * @return
+ *   0 if checks are alright, -1 otherwise.
+ */
+static int
+tap_flow_create_ipv6(const struct rte_flow_item *item, void *data)
+{
+       struct convert_data *info = (struct convert_data *)data;
+       const struct rte_flow_item_ipv6 *spec = item->spec;
+       const struct rte_flow_item_ipv6 *mask = item->mask;
+       struct rte_flow *flow = info->flow;
+       uint8_t empty_addr[16] = { 0 };
+       struct nlmsg *msg;
+
+       /* use default mask if none provided */
+       if (!mask)
+               mask = tap_flow_items[RTE_FLOW_ITEM_TYPE_IPV6].default_mask;
+       /* check that previous eth type is compatible with ipv6 */
+       if (info->eth_type && info->eth_type != htons(ETH_P_IPV6))
+               return -1;
+       /* store ip_proto for consistency if udp/tcp pattern item comes next */
+       if (spec)
+               info->ip_proto = spec->hdr.proto;
+       if (!flow)
+               return 0;
+       msg = &flow->msg;
+       if (!info->eth_type)
+               info->eth_type = htons(ETH_P_IPV6);
+       if (!info->vlan)
+               msg->t.tcm_info = TC_H_MAKE(msg->t.tcm_info, htons(ETH_P_IPV6));
+       if (!spec)
+               return 0;
+       if (memcmp(spec->hdr.dst_addr, empty_addr, 16)) {
+               nlattr_add(&msg->nh, TCA_FLOWER_KEY_IPV6_DST,
+                          sizeof(spec->hdr.dst_addr), &spec->hdr.dst_addr);
+               nlattr_add(&msg->nh, TCA_FLOWER_KEY_IPV6_DST_MASK,
+                          sizeof(mask->hdr.dst_addr), &mask->hdr.dst_addr);
+       }
+       if (memcmp(spec->hdr.src_addr, empty_addr, 16)) {
+               nlattr_add(&msg->nh, TCA_FLOWER_KEY_IPV6_SRC,
+                          sizeof(spec->hdr.src_addr), &spec->hdr.src_addr);
+               nlattr_add(&msg->nh, TCA_FLOWER_KEY_IPV6_SRC_MASK,
+                          sizeof(mask->hdr.src_addr), &mask->hdr.src_addr);
+       }
+       if (spec->hdr.proto)
+               nlattr_add8(&msg->nh, TCA_FLOWER_KEY_IP_PROTO, spec->hdr.proto);
+       return 0;
+}
+
+/**
+ * Make as much checks as possible on a UDP item, and if a flow is provided,
+ * fill it appropriately with UDP info.
+ *
+ * @param[in] item
+ *   Item specification.
+ * @param[in, out] data
+ *   Additional data structure to tell next layers we've been here.
+ *
+ * @return
+ *   0 if checks are alright, -1 otherwise.
+ */
+static int
+tap_flow_create_udp(const struct rte_flow_item *item, void *data)
+{
+       struct convert_data *info = (struct convert_data *)data;
+       const struct rte_flow_item_udp *spec = item->spec;
+       const struct rte_flow_item_udp *mask = item->mask;
+       struct rte_flow *flow = info->flow;
+       struct nlmsg *msg;
+
+       /* use default mask if none provided */
+       if (!mask)
+               mask = tap_flow_items[RTE_FLOW_ITEM_TYPE_UDP].default_mask;
+       /* check that previous ip_proto is compatible with udp */
+       if (info->ip_proto && info->ip_proto != IPPROTO_UDP)
+               return -1;
+       if (!flow)
+               return 0;
+       msg = &flow->msg;
+       nlattr_add8(&msg->nh, TCA_FLOWER_KEY_IP_PROTO, IPPROTO_UDP);
+       if (!spec)
+               return 0;
+       if (spec->hdr.dst_port &&
+           (spec->hdr.dst_port & mask->hdr.dst_port) == spec->hdr.dst_port)
+               nlattr_add16(&msg->nh, TCA_FLOWER_KEY_UDP_DST,
+                            spec->hdr.dst_port);
+       if (spec->hdr.src_port &&
+           (spec->hdr.src_port & mask->hdr.src_port) == spec->hdr.src_port)
+               nlattr_add16(&msg->nh, TCA_FLOWER_KEY_UDP_SRC,
+                            spec->hdr.src_port);
+       return 0;
+}
+
+/**
+ * Make as much checks as possible on a TCP item, and if a flow is provided,
+ * fill it appropriately with TCP info.
+ *
+ * @param[in] item
+ *   Item specification.
+ * @param[in, out] data
+ *   Additional data structure to tell next layers we've been here.
+ *
+ * @return
+ *   0 if checks are alright, -1 otherwise.
+ */
+static int
+tap_flow_create_tcp(const struct rte_flow_item *item, void *data)
+{
+       struct convert_data *info = (struct convert_data *)data;
+       const struct rte_flow_item_tcp *spec = item->spec;
+       const struct rte_flow_item_tcp *mask = item->mask;
+       struct rte_flow *flow = info->flow;
+       struct nlmsg *msg;
+
+       /* use default mask if none provided */
+       if (!mask)
+               mask = tap_flow_items[RTE_FLOW_ITEM_TYPE_TCP].default_mask;
+       /* check that previous ip_proto is compatible with tcp */
+       if (info->ip_proto && info->ip_proto != IPPROTO_TCP)
+               return -1;
+       if (!flow)
+               return 0;
+       msg = &flow->msg;
+       nlattr_add8(&msg->nh, TCA_FLOWER_KEY_IP_PROTO, IPPROTO_TCP);
+       if (!spec)
+               return 0;
+       if (spec->hdr.dst_port &&
+           (spec->hdr.dst_port & mask->hdr.dst_port) == spec->hdr.dst_port)
+               nlattr_add16(&msg->nh, TCA_FLOWER_KEY_TCP_DST,
+                            spec->hdr.dst_port);
+       if (spec->hdr.src_port &&
+           (spec->hdr.src_port & mask->hdr.src_port) == spec->hdr.src_port)
+               nlattr_add16(&msg->nh, TCA_FLOWER_KEY_TCP_SRC,
+                            spec->hdr.src_port);
+       return 0;
+}
+
+/**
+ * Check support for a given item.
+ *
+ * @param[in] item
+ *   Item specification.
+ * @param size
+ *   Bit-Mask size in bytes.
+ * @param[in] supported_mask
+ *   Bit-mask covering supported fields to compare with spec, last and mask in
+ *   \item.
+ * @param[in] default_mask
+ *   Bit-mask default mask if none is provided in \item.
+ *
+ * @return
+ *   0 on success.
+ */
+static int
+tap_flow_item_validate(const struct rte_flow_item *item,
+                      unsigned int size,
+                      const uint8_t *supported_mask,
+                      const uint8_t *default_mask)
+{
+       int ret = 0;
+
+       /* An empty layer is allowed, as long as all fields are NULL */
+       if (!item->spec && (item->mask || item->last))
+               return -1;
+       /* Is the item spec compatible with what the NIC supports? */
+       if (item->spec && !item->mask) {
+               unsigned int i;
+               const uint8_t *spec = item->spec;
+
+               for (i = 0; i < size; ++i)
+                       if ((spec[i] | supported_mask[i]) != supported_mask[i])
+                               return -1;
+               /* Is the default mask compatible with what the NIC supports? */
+               for (i = 0; i < size; i++)
+                       if ((default_mask[i] | supported_mask[i]) !=
+                           supported_mask[i])
+                               return -1;
+       }
+       /* Is the item last compatible with what the NIC supports? */
+       if (item->last && !item->mask) {
+               unsigned int i;
+               const uint8_t *spec = item->last;
+
+               for (i = 0; i < size; ++i)
+                       if ((spec[i] | supported_mask[i]) != supported_mask[i])
+                               return -1;
+       }
+       /* Is the item mask compatible with what the NIC supports? */
+       if (item->mask) {
+               unsigned int i;
+               const uint8_t *spec = item->mask;
+
+               for (i = 0; i < size; ++i)
+                       if ((spec[i] | supported_mask[i]) != supported_mask[i])
+                               return -1;
+       }
+       /**
+        * Once masked, Are item spec and item last equal?
+        * TC does not support range so anything else is invalid.
+        */
+       if (item->spec && item->last) {
+               uint8_t spec[size];
+               uint8_t last[size];
+               const uint8_t *apply = default_mask;
+               unsigned int i;
+
+               if (item->mask)
+                       apply = item->mask;
+               for (i = 0; i < size; ++i) {
+                       spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
+                       last[i] = ((const uint8_t *)item->last)[i] & apply[i];
+               }
+               ret = memcmp(spec, last, size);
+       }
+       return ret;
+}
+
+/**
+ * Transform a DROP/PASSTHRU action item in the provided flow for TC.
+ *
+ * @param[in, out] flow
+ *   Flow to be filled.
+ * @param[in] action
+ *   Appropriate action to be set in the TCA_GACT_PARMS structure.
+ *
+ * @return
+ *   0 if checks are alright, -1 otherwise.
+ */
+static int
+add_action_gact(struct rte_flow *flow, int action)
+{
+       struct nlmsg *msg = &flow->msg;
+       size_t act_index = 1;
+       struct tc_gact p = {
+               .action = action
+       };
+
+       if (nlattr_nested_start(msg, TCA_FLOWER_ACT) < 0)
+               return -1;
+       if (nlattr_nested_start(msg, act_index++) < 0)
+               return -1;
+       nlattr_add(&msg->nh, TCA_ACT_KIND, sizeof("gact"), "gact");
+       if (nlattr_nested_start(msg, TCA_ACT_OPTIONS) < 0)
+               return -1;
+       nlattr_add(&msg->nh, TCA_GACT_PARMS, sizeof(p), &p);
+       nlattr_nested_finish(msg); /* nested TCA_ACT_OPTIONS */
+       nlattr_nested_finish(msg); /* nested act_index */
+       nlattr_nested_finish(msg); /* nested TCA_FLOWER_ACT */
+       return 0;
+}
+
+/**
+ * Transform a QUEUE action item in the provided flow for TC.
+ *
+ * @param[in, out] flow
+ *   Flow to be filled.
+ * @param[in] queue
+ *   Queue id to use.
+ *
+ * @return
+ *   0 if checks are alright, -1 otherwise.
+ */
+static int
+add_action_skbedit(struct rte_flow *flow, uint16_t queue)
+{
+       struct nlmsg *msg = &flow->msg;
+       size_t act_index = 1;
+       struct tc_skbedit p = {
+               .action = TC_ACT_PIPE
+       };
+
+       if (nlattr_nested_start(msg, TCA_FLOWER_ACT) < 0)
+               return -1;
+       if (nlattr_nested_start(msg, act_index++) < 0)
+               return -1;
+       nlattr_add(&msg->nh, TCA_ACT_KIND, sizeof("skbedit"), "skbedit");
+       if (nlattr_nested_start(msg, TCA_ACT_OPTIONS) < 0)
+               return -1;
+       nlattr_add(&msg->nh, TCA_SKBEDIT_PARMS, sizeof(p), &p);
+       nlattr_add16(&msg->nh, TCA_SKBEDIT_QUEUE_MAPPING, queue);
+       nlattr_nested_finish(msg); /* nested TCA_ACT_OPTIONS */
+       nlattr_nested_finish(msg); /* nested act_index */
+       nlattr_nested_finish(msg); /* nested TCA_FLOWER_ACT */
+       return 0;
+}
+
+/**
+ * Validate a flow supported by TC.
+ * If flow param is not NULL, then also fill the netlink message inside.
+ *
+ * @param pmd
+ *   Pointer to private structure.
+ * @param[in] attr
+ *   Flow rule attributes.
+ * @param[in] pattern
+ *   Pattern specification (list terminated by the END pattern item).
+ * @param[in] actions
+ *   Associated actions (list terminated by the END action).
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ * @param[in, out] flow
+ *   Flow structure to update.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static int
+priv_flow_process(struct pmd_internals *pmd,
+                 const struct rte_flow_attr *attr,
+                 const struct rte_flow_item items[],
+                 const struct rte_flow_action actions[],
+                 struct rte_flow_error *error,
+                 struct rte_flow *flow)
+{
+       const struct tap_flow_items *cur_item = tap_flow_items;
+       struct convert_data data = {
+               .eth_type = 0,
+               .ip_proto = 0,
+               .flow = flow,
+       };
+       int action = 0; /* Only one action authorized for now */
+
+       if (attr->group > MAX_GROUP) {
+               rte_flow_error_set(
+                       error, EINVAL, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
+                       NULL, "group value too big: cannot exceed 15");
+               return -rte_errno;
+       }
+       if (attr->priority > MAX_PRIORITY) {
+               rte_flow_error_set(
+                       error, EINVAL, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
+                       NULL, "priority value too big");
+               return -rte_errno;
+       } else if (flow) {
+               uint16_t group = attr->group << GROUP_SHIFT;
+               uint16_t prio = group | (attr->priority + PRIORITY_OFFSET);
+               flow->msg.t.tcm_info = TC_H_MAKE(prio << 16,
+                                                flow->msg.t.tcm_info);
+       }
+       if (!attr->ingress) {
+               rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR,
+                                  NULL, "direction should be ingress");
+               return -rte_errno;
+       }
+       /* rte_flow ingress is actually egress as seen in the kernel */
+       if (attr->ingress && flow)
+               flow->msg.t.tcm_parent = TC_H_MAKE(MULTIQ_MAJOR_HANDLE, 0);
+       if (flow) {
+               /* use flower filter type */
+               nlattr_add(&flow->msg.nh, TCA_KIND, sizeof("flower"), "flower");
+               if (nlattr_nested_start(&flow->msg, TCA_OPTIONS) < 0)
+                       goto exit_item_not_supported;
+       }
+       for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
+               const struct tap_flow_items *token = NULL;
+               unsigned int i;
+               int err = 0;
+
+               if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
+                       continue;
+               for (i = 0;
+                    cur_item->items &&
+                    cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
+                    ++i) {
+                       if (cur_item->items[i] == items->type) {
+                               token = &tap_flow_items[items->type];
+                               break;
+                       }
+               }
+               if (!token)
+                       goto exit_item_not_supported;
+               cur_item = token;
+               err = tap_flow_item_validate(
+                       items, cur_item->mask_sz,
+                       (const uint8_t *)cur_item->mask,
+                       (const uint8_t *)cur_item->default_mask);
+               if (err)
+                       goto exit_item_not_supported;
+               if (flow && cur_item->convert) {
+                       if (!pmd->flower_vlan_support &&
+                           cur_item->convert == tap_flow_create_vlan)
+                               goto exit_item_not_supported;
+                       err = cur_item->convert(items, &data);
+                       if (err)
+                               goto exit_item_not_supported;
+               }
+       }
+       if (flow) {
+               if (pmd->flower_vlan_support && data.vlan) {
+                       nlattr_add16(&flow->msg.nh, TCA_FLOWER_KEY_ETH_TYPE,
+                                    htons(ETH_P_8021Q));
+                       nlattr_add16(&flow->msg.nh,
+                                    TCA_FLOWER_KEY_VLAN_ETH_TYPE,
+                                    data.eth_type ?
+                                    data.eth_type : htons(ETH_P_ALL));
+               } else if (data.eth_type) {
+                       nlattr_add16(&flow->msg.nh, TCA_FLOWER_KEY_ETH_TYPE,
+                                    data.eth_type);
+               }
+       }
+       for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
+               int err = 0;
+
+               if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
+                       continue;
+               } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
+                       if (action)
+                               goto exit_action_not_supported;
+                       action = 1;
+                       if (flow)
+                               err = add_action_gact(flow, TC_ACT_SHOT);
+               } else if (actions->type == RTE_FLOW_ACTION_TYPE_PASSTHRU) {
+                       if (action)
+                               goto exit_action_not_supported;
+                       action = 1;
+                       if (flow)
+                               err = add_action_gact(flow, TC_ACT_UNSPEC);
+               } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
+                       const struct rte_flow_action_queue *queue =
+                               (const struct rte_flow_action_queue *)
+                               actions->conf;
+                       if (action)
+                               goto exit_action_not_supported;
+                       action = 1;
+                       if (!queue || (queue->index >= pmd->nb_queues))
+                               goto exit_action_not_supported;
+                       if (flow)
+                               err = add_action_skbedit(flow, queue->index);
+               } else {
+                       goto exit_action_not_supported;
+               }
+               if (err)
+                       goto exit_action_not_supported;
+       }
+       if (flow)
+               nlattr_nested_finish(&flow->msg); /* nested TCA_OPTIONS */
+       return 0;
+exit_item_not_supported:
+       rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
+                          items, "item not supported");
+       return -rte_errno;
+exit_action_not_supported:
+       rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
+                          actions, "action not supported");
+       return -rte_errno;
+}
+
+
+
 /**
  * Validate a flow.
  *
@@ -74,15 +877,46 @@ static const struct rte_flow_ops tap_flow_ops = {
  * @see rte_flow_ops
  */
 static int
-tap_flow_validate(struct rte_eth_dev *dev __rte_unused,
-                 const struct rte_flow_attr *attr __rte_unused,
-                 const struct rte_flow_item items[] __rte_unused,
-                 const struct rte_flow_action actions[] __rte_unused,
+tap_flow_validate(struct rte_eth_dev *dev,
+                 const struct rte_flow_attr *attr,
+                 const struct rte_flow_item items[],
+                 const struct rte_flow_action actions[],
                  struct rte_flow_error *error)
 {
-       return -rte_flow_error_set(error, ENOTSUP,
-                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-                                  NULL, "not implemented yet");
+       struct pmd_internals *pmd = dev->data->dev_private;
+
+       return priv_flow_process(pmd, attr, items, actions, error, NULL);
+}
+
+/**
+ * Set a unique handle in a flow.
+ *
+ * The kernel supports TC rules with equal priority, as long as they use the
+ * same matching fields (e.g.: dst mac and ipv4) with different values (and
+ * full mask to ensure no collision is possible).
+ * In those rules, the handle (uint32_t) is the part that would identify
+ * specifically each rule.
+ *
+ * On 32-bit architectures, the handle can simply be the flow's pointer address.
+ * On 64-bit architectures, we rely on jhash(flow) to find a (sufficiently)
+ * unique handle.
+ *
+ * @param[in, out] flow
+ *   The flow that needs its handle set.
+ */
+static void
+tap_flow_set_handle(struct rte_flow *flow)
+{
+       uint32_t handle = 0;
+
+       if (sizeof(flow) > 4)
+               handle = rte_jhash(&flow, sizeof(flow), 1);
+       else
+               handle = (uintptr_t)flow;
+       /* must be at least 1 to avoid letting the kernel choose one for us */
+       if (!handle)
+               handle = 1;
+       flow->msg.t.tcm_handle = handle;
 }
 
 /**
@@ -100,17 +934,46 @@ tap_flow_create(struct rte_eth_dev *dev,
 {
        struct pmd_internals *pmd = dev->data->dev_private;
        struct rte_flow *flow = NULL;
+       struct nlmsg *msg = NULL;
+       int err;
 
-       if (tap_flow_validate(dev, attr, items, actions, error))
-               return NULL;
+       if (!pmd->if_index) {
+               rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
+                                  NULL,
+                                  "can't create rule, ifindex not found");
+               goto fail;
+       }
        flow = rte_malloc(__func__, sizeof(struct rte_flow), 0);
        if (!flow) {
                rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
                                   NULL, "cannot allocate memory for rte_flow");
-               return NULL;
+               goto fail;
+       }
+       msg = &flow->msg;
+       tc_init_msg(msg, pmd->if_index, RTM_NEWTFILTER,
+                   NLM_F_REQUEST | NLM_F_ACK | NLM_F_EXCL | NLM_F_CREATE);
+       msg->t.tcm_info = TC_H_MAKE(0, htons(ETH_P_ALL));
+       tap_flow_set_handle(flow);
+       if (priv_flow_process(pmd, attr, items, actions, error, flow))
+               goto fail;
+       err = nl_send(pmd->nlsk_fd, &msg->nh);
+       if (err < 0) {
+               rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
+                                  NULL, "couldn't send request to kernel");
+               goto fail;
+       }
+       err = nl_recv_ack(pmd->nlsk_fd);
+       if (err < 0) {
+               rte_flow_error_set(error, EEXIST, RTE_FLOW_ERROR_TYPE_HANDLE,
+                                  NULL, "overlapping rules");
+               goto fail;
        }
        LIST_INSERT_HEAD(&pmd->flows, flow, next);
        return flow;
+fail:
+       if (flow)
+               rte_free(flow);
+       return NULL;
 }
 
 /**
@@ -120,13 +983,31 @@ tap_flow_create(struct rte_eth_dev *dev,
  * @see rte_flow_ops
  */
 static int
-tap_flow_destroy(struct rte_eth_dev *dev __rte_unused,
+tap_flow_destroy(struct rte_eth_dev *dev,
                 struct rte_flow *flow,
-                struct rte_flow_error *error __rte_unused)
+                struct rte_flow_error *error)
 {
+       struct pmd_internals *pmd = dev->data->dev_private;
+       int ret = 0;
+
        LIST_REMOVE(flow, next);
+       flow->msg.nh.nlmsg_flags = NLM_F_REQUEST | NLM_F_ACK;
+       flow->msg.nh.nlmsg_type = RTM_DELTFILTER;
+
+       ret = nl_send(pmd->nlsk_fd, &flow->msg.nh);
+       if (ret < 0) {
+               rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
+                                  NULL, "couldn't send request to kernel");
+               goto end;
+       }
+       ret = nl_recv_ack(pmd->nlsk_fd);
+       if (ret < 0)
+               rte_flow_error_set(
+                       error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
+                       "couldn't receive kernel ack to our request");
+end:
        rte_free(flow);
-       return 0;
+       return ret;
 }
 
 /**
@@ -170,6 +1051,10 @@ tap_dev_filter_ctrl(struct rte_eth_dev *dev,
                    enum rte_filter_op filter_op,
                    void *arg)
 {
+       struct pmd_internals *pmd = dev->data->dev_private;
+
+       if (!pmd->flower_support)
+               return -ENOTSUP;
        switch (filter_type) {
        case RTE_ETH_FILTER_GENERIC:
                if (filter_op != RTE_ETH_FILTER_GET)
index 377a9f7..a05e945 100644 (file)
 #include <rte_flow.h>
 #include <rte_flow_driver.h>
 
+/**
+ * In TC, priority 0 means we require the kernel to allocate one for us.
+ * In rte_flow, however, we want the priority 0 to be the most important one.
+ * Use an offset to have the most important priority being 1 in TC.
+ */
+#define PRIORITY_OFFSET 1
+#define PRIORITY_MASK (0xfff)
+#define MAX_PRIORITY (PRIORITY_MASK - PRIORITY_OFFSET)
+#define GROUP_MASK (0xf)
+#define GROUP_SHIFT 12
+#define MAX_GROUP GROUP_MASK
+
 int tap_dev_filter_ctrl(struct rte_eth_dev *dev,
                        enum rte_filter_type filter_type,
                        enum rte_filter_op filter_op,