gro: support VXLAN UDP/IPv4
authorYi Yang <yangyi01@inspur.com>
Thu, 24 Sep 2020 08:57:39 +0000 (16:57 +0800)
committerThomas Monjalon <thomas@monjalon.net>
Tue, 6 Oct 2020 19:51:03 +0000 (21:51 +0200)
VXLAN UDP/IPv4 GRO can help improve VM-to-VM UDP
performance when UFO or GSO is enabled in VM, GRO
must be supported if UFO or GSO is enabled,
otherwise, performance can't get big improvement
if only GSO is there.

With this enabled in DPDK, OVS DPDK can leverage it
to improve VM-to-VM UDP performance, it will reassemble
VXLAN UDP/IPv4 fragments immediate after they are
received from a physical NIC. It is very helpful in
OVS DPDK VXLAN use case.

Signed-off-by: Yi Yang <yangyi01@inspur.com>
Acked-by: Jiayu Hu <jiayu.hu@intel.com>
doc/guides/prog_guide/generic_receive_offload_lib.rst
doc/guides/rel_notes/release_20_11.rst
lib/librte_gro/gro_udp4.h
lib/librte_gro/gro_vxlan_udp4.c [new file with mode: 0644]
lib/librte_gro/gro_vxlan_udp4.h [new file with mode: 0644]
lib/librte_gro/meson.build
lib/librte_gro/rte_gro.c
lib/librte_gro/rte_gro.h

index abe1668..98a5d29 100644 (file)
@@ -32,7 +32,7 @@ fragmentation is possible (i.e., DF==0). Additionally, it complies RFC
 
 Currently, the GRO library provides GRO supports for TCP/IPv4 and UDP/IPv4
 packets as well as VxLAN packets which contain an outer IPv4 header and an
-inner TCP/IPv4 packet.
+inner TCP/IPv4 or UDP/IPv4 packet.
 
 Two Sets of API
 ---------------
index b5d7648..cdf2040 100644 (file)
@@ -84,8 +84,9 @@ New Features
 
   Added NTB device support (4th generation) for Intel Ice Lake platform.
 
-* **Added UDP/IPv4 GRO support for non-VxLAN packets.**
+* **Added UDP/IPv4 GRO support for VxLAN and non-VxLAN packets.**
 
+  For VxLAN packets, added inner UDP/IPv4 support.
   For non-VxLAN packets, added UDP/IPv4 support.
 
 * **Extended flow-perf application.**
index 0a078e4..d38b393 100644 (file)
@@ -7,6 +7,7 @@
 
 #include <rte_ip.h>
 #include <rte_udp.h>
+#include <rte_vxlan.h>
 
 #define INVALID_ARRAY_INDEX 0xffffffffUL
 #define GRO_UDP4_TBL_MAX_ITEM_NUM (1024UL * 1024UL)
diff --git a/lib/librte_gro/gro_vxlan_udp4.c b/lib/librte_gro/gro_vxlan_udp4.c
new file mode 100644 (file)
index 0000000..3747636
--- /dev/null
@@ -0,0 +1,545 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Inspur Corporation
+ */
+
+#include <rte_malloc.h>
+#include <rte_mbuf.h>
+#include <rte_cycles.h>
+#include <rte_ethdev.h>
+#include <rte_udp.h>
+
+#include "gro_vxlan_udp4.h"
+
+void *
+gro_vxlan_udp4_tbl_create(uint16_t socket_id,
+               uint16_t max_flow_num,
+               uint16_t max_item_per_flow)
+{
+       struct gro_vxlan_udp4_tbl *tbl;
+       size_t size;
+       uint32_t entries_num, i;
+
+       entries_num = max_flow_num * max_item_per_flow;
+       entries_num = RTE_MIN(entries_num, GRO_VXLAN_UDP4_TBL_MAX_ITEM_NUM);
+
+       if (entries_num == 0)
+               return NULL;
+
+       tbl = rte_zmalloc_socket(__func__,
+                       sizeof(struct gro_vxlan_udp4_tbl),
+                       RTE_CACHE_LINE_SIZE,
+                       socket_id);
+       if (tbl == NULL)
+               return NULL;
+
+       size = sizeof(struct gro_vxlan_udp4_item) * entries_num;
+       tbl->items = rte_zmalloc_socket(__func__,
+                       size,
+                       RTE_CACHE_LINE_SIZE,
+                       socket_id);
+       if (tbl->items == NULL) {
+               rte_free(tbl);
+               return NULL;
+       }
+       tbl->max_item_num = entries_num;
+
+       size = sizeof(struct gro_vxlan_udp4_flow) * entries_num;
+       tbl->flows = rte_zmalloc_socket(__func__,
+                       size,
+                       RTE_CACHE_LINE_SIZE,
+                       socket_id);
+       if (tbl->flows == NULL) {
+               rte_free(tbl->items);
+               rte_free(tbl);
+               return NULL;
+       }
+
+       for (i = 0; i < entries_num; i++)
+               tbl->flows[i].start_index = INVALID_ARRAY_INDEX;
+       tbl->max_flow_num = entries_num;
+
+       return tbl;
+}
+
+void
+gro_vxlan_udp4_tbl_destroy(void *tbl)
+{
+       struct gro_vxlan_udp4_tbl *vxlan_tbl = tbl;
+
+       if (vxlan_tbl) {
+               rte_free(vxlan_tbl->items);
+               rte_free(vxlan_tbl->flows);
+       }
+       rte_free(vxlan_tbl);
+}
+
+static inline uint32_t
+find_an_empty_item(struct gro_vxlan_udp4_tbl *tbl)
+{
+       uint32_t max_item_num = tbl->max_item_num, i;
+
+       for (i = 0; i < max_item_num; i++)
+               if (tbl->items[i].inner_item.firstseg == NULL)
+                       return i;
+       return INVALID_ARRAY_INDEX;
+}
+
+static inline uint32_t
+find_an_empty_flow(struct gro_vxlan_udp4_tbl *tbl)
+{
+       uint32_t max_flow_num = tbl->max_flow_num, i;
+
+       for (i = 0; i < max_flow_num; i++)
+               if (tbl->flows[i].start_index == INVALID_ARRAY_INDEX)
+                       return i;
+       return INVALID_ARRAY_INDEX;
+}
+
+static inline uint32_t
+insert_new_item(struct gro_vxlan_udp4_tbl *tbl,
+               struct rte_mbuf *pkt,
+               uint64_t start_time,
+               uint32_t prev_idx,
+               uint16_t frag_offset,
+               uint8_t is_last_frag)
+{
+       uint32_t item_idx;
+
+       item_idx = find_an_empty_item(tbl);
+       if (unlikely(item_idx == INVALID_ARRAY_INDEX))
+               return INVALID_ARRAY_INDEX;
+
+       tbl->items[item_idx].inner_item.firstseg = pkt;
+       tbl->items[item_idx].inner_item.lastseg = rte_pktmbuf_lastseg(pkt);
+       tbl->items[item_idx].inner_item.start_time = start_time;
+       tbl->items[item_idx].inner_item.next_pkt_idx = INVALID_ARRAY_INDEX;
+       tbl->items[item_idx].inner_item.frag_offset = frag_offset;
+       tbl->items[item_idx].inner_item.is_last_frag = is_last_frag;
+       tbl->items[item_idx].inner_item.nb_merged = 1;
+       tbl->item_num++;
+
+       /* If the previous packet exists, chain the new one with it. */
+       if (prev_idx != INVALID_ARRAY_INDEX) {
+               tbl->items[item_idx].inner_item.next_pkt_idx =
+                       tbl->items[prev_idx].inner_item.next_pkt_idx;
+               tbl->items[prev_idx].inner_item.next_pkt_idx = item_idx;
+       }
+
+       return item_idx;
+}
+
+static inline uint32_t
+delete_item(struct gro_vxlan_udp4_tbl *tbl,
+               uint32_t item_idx,
+               uint32_t prev_item_idx)
+{
+       uint32_t next_idx = tbl->items[item_idx].inner_item.next_pkt_idx;
+
+       /* NULL indicates an empty item. */
+       tbl->items[item_idx].inner_item.firstseg = NULL;
+       tbl->item_num--;
+       if (prev_item_idx != INVALID_ARRAY_INDEX)
+               tbl->items[prev_item_idx].inner_item.next_pkt_idx = next_idx;
+
+       return next_idx;
+}
+
+static inline uint32_t
+insert_new_flow(struct gro_vxlan_udp4_tbl *tbl,
+               struct vxlan_udp4_flow_key *src,
+               uint32_t item_idx)
+{
+       struct vxlan_udp4_flow_key *dst;
+       uint32_t flow_idx;
+
+       flow_idx = find_an_empty_flow(tbl);
+       if (unlikely(flow_idx == INVALID_ARRAY_INDEX))
+               return INVALID_ARRAY_INDEX;
+
+       dst = &(tbl->flows[flow_idx].key);
+
+       rte_ether_addr_copy(&(src->inner_key.eth_saddr),
+                       &(dst->inner_key.eth_saddr));
+       rte_ether_addr_copy(&(src->inner_key.eth_daddr),
+                       &(dst->inner_key.eth_daddr));
+       dst->inner_key.ip_src_addr = src->inner_key.ip_src_addr;
+       dst->inner_key.ip_dst_addr = src->inner_key.ip_dst_addr;
+       dst->inner_key.ip_id = src->inner_key.ip_id;
+
+       dst->vxlan_hdr.vx_flags = src->vxlan_hdr.vx_flags;
+       dst->vxlan_hdr.vx_vni = src->vxlan_hdr.vx_vni;
+       rte_ether_addr_copy(&(src->outer_eth_saddr), &(dst->outer_eth_saddr));
+       rte_ether_addr_copy(&(src->outer_eth_daddr), &(dst->outer_eth_daddr));
+       dst->outer_ip_src_addr = src->outer_ip_src_addr;
+       dst->outer_ip_dst_addr = src->outer_ip_dst_addr;
+       dst->outer_dst_port = src->outer_dst_port;
+
+       tbl->flows[flow_idx].start_index = item_idx;
+       tbl->flow_num++;
+
+       return flow_idx;
+}
+
+static inline int
+is_same_vxlan_udp4_flow(struct vxlan_udp4_flow_key k1,
+               struct vxlan_udp4_flow_key k2)
+{
+       /* For VxLAN packet, outer udp src port is calculated from
+        * inner packet RSS hash, udp src port of the first UDP
+        * fragment is different from one of other UDP fragments
+        * even if they are same flow, so we have to skip outer udp
+        * src port comparison here.
+        */
+       return (rte_is_same_ether_addr(&k1.outer_eth_saddr,
+                                       &k2.outer_eth_saddr) &&
+                       rte_is_same_ether_addr(&k1.outer_eth_daddr,
+                               &k2.outer_eth_daddr) &&
+                       (k1.outer_ip_src_addr == k2.outer_ip_src_addr) &&
+                       (k1.outer_ip_dst_addr == k2.outer_ip_dst_addr) &&
+                       (k1.outer_dst_port == k2.outer_dst_port) &&
+                       (k1.vxlan_hdr.vx_flags == k2.vxlan_hdr.vx_flags) &&
+                       (k1.vxlan_hdr.vx_vni == k2.vxlan_hdr.vx_vni) &&
+                       is_same_udp4_flow(k1.inner_key, k2.inner_key));
+}
+
+static inline int
+udp4_check_vxlan_neighbor(struct gro_vxlan_udp4_item *item,
+               uint16_t frag_offset,
+               uint16_t ip_dl)
+{
+       struct rte_mbuf *pkt = item->inner_item.firstseg;
+       int cmp;
+       uint16_t l2_offset;
+       int ret = 0;
+
+       /* Note: if outer DF bit is set, i.e outer_is_atomic is 0,
+        * we needn't compare outer_ip_id because they are same,
+        * for the case outer_is_atomic is 1, we also have no way
+        * to compare outer_ip_id because the difference between
+        * outer_ip_ids of two received packets isn't always +/-1.
+        * So skip outer_ip_id comparison here.
+        */
+
+       l2_offset = pkt->outer_l2_len + pkt->outer_l3_len;
+       cmp = udp4_check_neighbor(&item->inner_item, frag_offset, ip_dl,
+                                       l2_offset);
+       if (cmp > 0)
+               /* Append the new packet. */
+               ret = 1;
+       else if (cmp < 0)
+               /* Prepend the new packet. */
+               ret = -1;
+
+       return ret;
+}
+
+static inline int
+merge_two_vxlan_udp4_packets(struct gro_vxlan_udp4_item *item,
+               struct rte_mbuf *pkt,
+               int cmp,
+               uint16_t frag_offset,
+               uint8_t is_last_frag)
+{
+       if (merge_two_udp4_packets(&item->inner_item, pkt, cmp, frag_offset,
+                               is_last_frag,
+                               pkt->outer_l2_len + pkt->outer_l3_len)) {
+               return 1;
+       }
+
+       return 0;
+}
+
+static inline void
+update_vxlan_header(struct gro_vxlan_udp4_item *item)
+{
+       struct rte_ipv4_hdr *ipv4_hdr;
+       struct rte_udp_hdr *udp_hdr;
+       struct rte_mbuf *pkt = item->inner_item.firstseg;
+       uint16_t len;
+       uint16_t frag_offset;
+
+       /* Update the outer IPv4 header. */
+       len = pkt->pkt_len - pkt->outer_l2_len;
+       ipv4_hdr = (struct rte_ipv4_hdr *)(rte_pktmbuf_mtod(pkt, char *) +
+                       pkt->outer_l2_len);
+       ipv4_hdr->total_length = rte_cpu_to_be_16(len);
+
+       /* Update the outer UDP header. */
+       len -= pkt->outer_l3_len;
+       udp_hdr = (struct rte_udp_hdr *)((char *)ipv4_hdr + pkt->outer_l3_len);
+       udp_hdr->dgram_len = rte_cpu_to_be_16(len);
+
+       /* Update the inner IPv4 header. */
+       len -= pkt->l2_len;
+       ipv4_hdr = (struct rte_ipv4_hdr *)((char *)udp_hdr + pkt->l2_len);
+       ipv4_hdr->total_length = rte_cpu_to_be_16(len);
+
+       /* Clear MF bit if it is last fragment */
+       if (item->inner_item.is_last_frag) {
+               frag_offset = rte_be_to_cpu_16(ipv4_hdr->fragment_offset);
+               ipv4_hdr->fragment_offset =
+                       rte_cpu_to_be_16(frag_offset & ~RTE_IPV4_HDR_MF_FLAG);
+       }
+}
+
+int32_t
+gro_vxlan_udp4_reassemble(struct rte_mbuf *pkt,
+               struct gro_vxlan_udp4_tbl *tbl,
+               uint64_t start_time)
+{
+       struct rte_ether_hdr *outer_eth_hdr, *eth_hdr;
+       struct rte_ipv4_hdr *outer_ipv4_hdr, *ipv4_hdr;
+       struct rte_udp_hdr *udp_hdr;
+       struct rte_vxlan_hdr *vxlan_hdr;
+       uint16_t frag_offset;
+       uint8_t is_last_frag;
+       int16_t ip_dl;
+       uint16_t ip_id;
+
+       struct vxlan_udp4_flow_key key;
+       uint32_t cur_idx, prev_idx, item_idx;
+       uint32_t i, max_flow_num, remaining_flow_num;
+       int cmp;
+       uint16_t hdr_len;
+       uint8_t find;
+
+       outer_eth_hdr = rte_pktmbuf_mtod(pkt, struct rte_ether_hdr *);
+       outer_ipv4_hdr = (struct rte_ipv4_hdr *)((char *)outer_eth_hdr +
+                       pkt->outer_l2_len);
+
+       udp_hdr = (struct rte_udp_hdr *)((char *)outer_ipv4_hdr +
+                       pkt->outer_l3_len);
+       vxlan_hdr = (struct rte_vxlan_hdr *)((char *)udp_hdr +
+                       sizeof(struct rte_udp_hdr));
+       eth_hdr = (struct rte_ether_hdr *)((char *)vxlan_hdr +
+                       sizeof(struct rte_vxlan_hdr));
+       /* l2_len = outer udp hdr len + vxlan hdr len + inner l2 len */
+       ipv4_hdr = (struct rte_ipv4_hdr *)((char *)udp_hdr + pkt->l2_len);
+
+       /*
+        * Don't process the packet which has non-fragment inner IP.
+        */
+       if (!is_ipv4_fragment(ipv4_hdr))
+               return -1;
+
+       hdr_len = pkt->outer_l2_len + pkt->outer_l3_len + pkt->l2_len +
+               pkt->l3_len;
+       /*
+        * Don't process the packet whose payload length is less than or
+        * equal to 0.
+        */
+       if (pkt->pkt_len <= hdr_len)
+               return -1;
+
+       ip_dl = pkt->pkt_len - hdr_len;
+
+       ip_id = rte_be_to_cpu_16(ipv4_hdr->packet_id);
+       frag_offset = rte_be_to_cpu_16(ipv4_hdr->fragment_offset);
+       is_last_frag = ((frag_offset & RTE_IPV4_HDR_MF_FLAG) == 0) ? 1 : 0;
+       frag_offset = (uint16_t)(frag_offset & RTE_IPV4_HDR_OFFSET_MASK) << 3;
+
+       rte_ether_addr_copy(&(eth_hdr->s_addr), &(key.inner_key.eth_saddr));
+       rte_ether_addr_copy(&(eth_hdr->d_addr), &(key.inner_key.eth_daddr));
+       key.inner_key.ip_src_addr = ipv4_hdr->src_addr;
+       key.inner_key.ip_dst_addr = ipv4_hdr->dst_addr;
+       key.inner_key.ip_id = ip_id;
+
+       key.vxlan_hdr.vx_flags = vxlan_hdr->vx_flags;
+       key.vxlan_hdr.vx_vni = vxlan_hdr->vx_vni;
+       rte_ether_addr_copy(&(outer_eth_hdr->s_addr), &(key.outer_eth_saddr));
+       rte_ether_addr_copy(&(outer_eth_hdr->d_addr), &(key.outer_eth_daddr));
+       key.outer_ip_src_addr = outer_ipv4_hdr->src_addr;
+       key.outer_ip_dst_addr = outer_ipv4_hdr->dst_addr;
+       /* Note: It is unnecessary to save outer_src_port here because it can
+        * be different for VxLAN UDP fragments from the same flow.
+        */
+       key.outer_dst_port = udp_hdr->dst_port;
+
+       /* Search for a matched flow. */
+       max_flow_num = tbl->max_flow_num;
+       remaining_flow_num = tbl->flow_num;
+       find = 0;
+       for (i = 0; i < max_flow_num && remaining_flow_num; i++) {
+               if (tbl->flows[i].start_index != INVALID_ARRAY_INDEX) {
+                       if (is_same_vxlan_udp4_flow(tbl->flows[i].key, key)) {
+                               find = 1;
+                               break;
+                       }
+                       remaining_flow_num--;
+               }
+       }
+
+       /*
+        * Can't find a matched flow. Insert a new flow and store the
+        * packet into the flow.
+        */
+       if (find == 0) {
+               item_idx = insert_new_item(tbl, pkt, start_time,
+                               INVALID_ARRAY_INDEX, frag_offset,
+                               is_last_frag);
+               if (unlikely(item_idx == INVALID_ARRAY_INDEX))
+                       return -1;
+               if (insert_new_flow(tbl, &key, item_idx) ==
+                               INVALID_ARRAY_INDEX) {
+                       /*
+                        * Fail to insert a new flow, so
+                        * delete the inserted packet.
+                        */
+                       delete_item(tbl, item_idx, INVALID_ARRAY_INDEX);
+                       return -1;
+               }
+               return 0;
+       }
+
+       /* Check all packets in the flow and try to find a neighbor. */
+       cur_idx = tbl->flows[i].start_index;
+       prev_idx = cur_idx;
+       do {
+               cmp = udp4_check_vxlan_neighbor(&(tbl->items[cur_idx]),
+                               frag_offset, ip_dl);
+               if (cmp) {
+                       if (merge_two_vxlan_udp4_packets(
+                                               &(tbl->items[cur_idx]),
+                                               pkt, cmp, frag_offset,
+                                               is_last_frag)) {
+                               return 1;
+                       }
+                       /*
+                        * Can't merge two packets, as the packet
+                        * length will be greater than the max value.
+                        * Insert the packet into the flow.
+                        */
+                       if (insert_new_item(tbl, pkt, start_time, prev_idx,
+                                               frag_offset, is_last_frag) ==
+                                       INVALID_ARRAY_INDEX)
+                               return -1;
+                       return 0;
+               }
+
+               /* Ensure inserted items are ordered by frag_offset */
+               if (frag_offset
+                       < tbl->items[cur_idx].inner_item.frag_offset) {
+                       break;
+               }
+
+               prev_idx = cur_idx;
+               cur_idx = tbl->items[cur_idx].inner_item.next_pkt_idx;
+       } while (cur_idx != INVALID_ARRAY_INDEX);
+
+       /* Can't find neighbor. Insert the packet into the flow. */
+       if (cur_idx == tbl->flows[i].start_index) {
+               /* Insert it before the first packet of the flow */
+               item_idx = insert_new_item(tbl, pkt, start_time,
+                               INVALID_ARRAY_INDEX, frag_offset,
+                               is_last_frag);
+               if (unlikely(item_idx == INVALID_ARRAY_INDEX))
+                       return -1;
+               tbl->items[item_idx].inner_item.next_pkt_idx = cur_idx;
+               tbl->flows[i].start_index = item_idx;
+       } else {
+               if (insert_new_item(tbl, pkt, start_time, prev_idx,
+                                       frag_offset, is_last_frag
+                                       ) == INVALID_ARRAY_INDEX)
+                       return -1;
+       }
+
+       return 0;
+}
+
+static int
+gro_vxlan_udp4_merge_items(struct gro_vxlan_udp4_tbl *tbl,
+                          uint32_t start_idx)
+{
+       uint16_t frag_offset;
+       uint8_t is_last_frag;
+       int16_t ip_dl;
+       struct rte_mbuf *pkt;
+       int cmp;
+       uint32_t item_idx;
+       uint16_t hdr_len;
+
+       item_idx = tbl->items[start_idx].inner_item.next_pkt_idx;
+       while (item_idx != INVALID_ARRAY_INDEX) {
+               pkt = tbl->items[item_idx].inner_item.firstseg;
+               hdr_len = pkt->outer_l2_len + pkt->outer_l3_len + pkt->l2_len +
+                       pkt->l3_len;
+               ip_dl = pkt->pkt_len - hdr_len;
+               frag_offset = tbl->items[item_idx].inner_item.frag_offset;
+               is_last_frag = tbl->items[item_idx].inner_item.is_last_frag;
+               cmp = udp4_check_vxlan_neighbor(&(tbl->items[start_idx]),
+                                       frag_offset, ip_dl);
+               if (cmp) {
+                       if (merge_two_vxlan_udp4_packets(
+                                       &(tbl->items[start_idx]),
+                                       pkt, cmp, frag_offset,
+                                       is_last_frag)) {
+                               item_idx = delete_item(tbl, item_idx,
+                                                       INVALID_ARRAY_INDEX);
+                               tbl->items[start_idx].inner_item.next_pkt_idx
+                                       = item_idx;
+                       } else
+                               return 0;
+               } else
+                       return 0;
+       }
+
+       return 0;
+}
+
+uint16_t
+gro_vxlan_udp4_tbl_timeout_flush(struct gro_vxlan_udp4_tbl *tbl,
+               uint64_t flush_timestamp,
+               struct rte_mbuf **out,
+               uint16_t nb_out)
+{
+       uint16_t k = 0;
+       uint32_t i, j;
+       uint32_t max_flow_num = tbl->max_flow_num;
+
+       for (i = 0; i < max_flow_num; i++) {
+               if (unlikely(tbl->flow_num == 0))
+                       return k;
+
+               j = tbl->flows[i].start_index;
+               while (j != INVALID_ARRAY_INDEX) {
+                       if (tbl->items[j].inner_item.start_time <=
+                                       flush_timestamp) {
+                               gro_vxlan_udp4_merge_items(tbl, j);
+                               out[k++] = tbl->items[j].inner_item.firstseg;
+                               if (tbl->items[j].inner_item.nb_merged > 1)
+                                       update_vxlan_header(&(tbl->items[j]));
+                               /*
+                                * Delete the item and get the next packet
+                                * index.
+                                */
+                               j = delete_item(tbl, j, INVALID_ARRAY_INDEX);
+                               tbl->flows[i].start_index = j;
+                               if (j == INVALID_ARRAY_INDEX)
+                                       tbl->flow_num--;
+
+                               if (unlikely(k == nb_out))
+                                       return k;
+                       } else
+                               /*
+                                * Flushing packets does not strictly follow
+                                * timestamp. It does not flush left packets of
+                                * the flow this time once it finds one item
+                                * whose start_time is greater than
+                                * flush_timestamp. So go to check other flows.
+                                */
+                               break;
+               }
+       }
+       return k;
+}
+
+uint32_t
+gro_vxlan_udp4_tbl_pkt_count(void *tbl)
+{
+       struct gro_vxlan_udp4_tbl *gro_tbl = tbl;
+
+       if (gro_tbl)
+               return gro_tbl->item_num;
+
+       return 0;
+}
diff --git a/lib/librte_gro/gro_vxlan_udp4.h b/lib/librte_gro/gro_vxlan_udp4.h
new file mode 100644 (file)
index 0000000..d045221
--- /dev/null
@@ -0,0 +1,153 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Inspur Corporation
+ */
+
+#ifndef _GRO_VXLAN_UDP4_H_
+#define _GRO_VXLAN_UDP4_H_
+
+#include "gro_udp4.h"
+
+#define GRO_VXLAN_UDP4_TBL_MAX_ITEM_NUM (1024UL * 1024UL)
+
+/* Header fields representing a VxLAN flow */
+struct vxlan_udp4_flow_key {
+       struct udp4_flow_key inner_key;
+       struct rte_vxlan_hdr vxlan_hdr;
+
+       struct rte_ether_addr outer_eth_saddr;
+       struct rte_ether_addr outer_eth_daddr;
+
+       uint32_t outer_ip_src_addr;
+       uint32_t outer_ip_dst_addr;
+
+       /* Note: It is unnecessary to save outer_src_port here because it can
+        * be different for VxLAN UDP fragments from the same flow.
+        */
+       uint16_t outer_dst_port;
+};
+
+struct gro_vxlan_udp4_flow {
+       struct vxlan_udp4_flow_key key;
+       /*
+        * The index of the first packet in the flow. INVALID_ARRAY_INDEX
+        * indicates an empty flow.
+        */
+       uint32_t start_index;
+};
+
+struct gro_vxlan_udp4_item {
+       struct gro_udp4_item inner_item;
+       /* Note: VXLAN UDP/IPv4 GRO needn't check outer_ip_id because
+        * the difference between outer_ip_ids of two received packets
+        * isn't always +/-1 in case of OVS DPDK. So no outer_ip_id
+        * and outer_is_atomic fields here.
+        */
+};
+
+/*
+ * VxLAN (with an outer IPv4 header and an inner UDP/IPv4 packet)
+ * reassembly table structure
+ */
+struct gro_vxlan_udp4_tbl {
+       /* item array */
+       struct gro_vxlan_udp4_item *items;
+       /* flow array */
+       struct gro_vxlan_udp4_flow *flows;
+       /* current item number */
+       uint32_t item_num;
+       /* current flow number */
+       uint32_t flow_num;
+       /* the maximum item number */
+       uint32_t max_item_num;
+       /* the maximum flow number */
+       uint32_t max_flow_num;
+};
+
+/**
+ * This function creates a VxLAN reassembly table for VxLAN packets
+ * which have an outer IPv4 header and an inner UDP/IPv4 packet.
+ *
+ * @param socket_id
+ *  Socket index for allocating the table
+ * @param max_flow_num
+ *  The maximum number of flows in the table
+ * @param max_item_per_flow
+ *  The maximum number of packets per flow
+ *
+ * @return
+ *  - Return the table pointer on success.
+ *  - Return NULL on failure.
+ */
+void *gro_vxlan_udp4_tbl_create(uint16_t socket_id,
+               uint16_t max_flow_num,
+               uint16_t max_item_per_flow);
+
+/**
+ * This function destroys a VxLAN reassembly table.
+ *
+ * @param tbl
+ *  Pointer pointing to the VxLAN reassembly table
+ */
+void gro_vxlan_udp4_tbl_destroy(void *tbl);
+
+/**
+ * This function merges a VxLAN packet which has an outer IPv4 header and
+ * an inner UDP/IPv4 packet. It does not process the packet which does not
+ * have payload.
+ *
+ * This function does not check if the packet has correct checksums and
+ * does not re-calculate checksums for the merged packet. It returns the
+ * packet if there is no available space in the table.
+ *
+ * @param pkt
+ *  Packet to reassemble
+ * @param tbl
+ *  Pointer pointing to the VxLAN reassembly table
+ * @start_time
+ *  The time when the packet is inserted into the table
+ *
+ * @return
+ *  - Return a positive value if the packet is merged.
+ *  - Return zero if the packet isn't merged but stored in the table.
+ *  - Return a negative value for invalid parameters or no available
+ *    space in the table.
+ */
+int32_t gro_vxlan_udp4_reassemble(struct rte_mbuf *pkt,
+               struct gro_vxlan_udp4_tbl *tbl,
+               uint64_t start_time);
+
+/**
+ * This function flushes timeout packets in the VxLAN reassembly table,
+ * and without updating checksums.
+ *
+ * @param tbl
+ *  Pointer pointing to a VxLAN GRO table
+ * @param flush_timestamp
+ *  This function flushes packets which are inserted into the table
+ *  before or at the flush_timestamp.
+ * @param out
+ *  Pointer array used to keep flushed packets
+ * @param nb_out
+ *  The element number in 'out'. It also determines the maximum number of
+ *  packets that can be flushed finally.
+ *
+ * @return
+ *  The number of flushed packets
+ */
+uint16_t gro_vxlan_udp4_tbl_timeout_flush(struct gro_vxlan_udp4_tbl *tbl,
+               uint64_t flush_timestamp,
+               struct rte_mbuf **out,
+               uint16_t nb_out);
+
+/**
+ * This function returns the number of the packets in a VxLAN
+ * reassembly table.
+ *
+ * @param tbl
+ *  Pointer pointing to the VxLAN reassembly table
+ *
+ * @return
+ *  The number of packets in the table
+ */
+uint32_t gro_vxlan_udp4_tbl_pkt_count(void *tbl);
+#endif
index 0d18dc2..ea8b45c 100644 (file)
@@ -1,6 +1,6 @@
 # SPDX-License-Identifier: BSD-3-Clause
 # Copyright(c) 2017 Intel Corporation
 
-sources = files('rte_gro.c', 'gro_tcp4.c', 'gro_udp4.c', 'gro_vxlan_tcp4.c')
+sources = files('rte_gro.c', 'gro_tcp4.c', 'gro_udp4.c', 'gro_vxlan_tcp4.c', 'gro_vxlan_udp4.c')
 headers = files('rte_gro.h')
 deps += ['ethdev']
index ac23df1..e56bd20 100644 (file)
@@ -11,6 +11,7 @@
 #include "gro_tcp4.h"
 #include "gro_udp4.h"
 #include "gro_vxlan_tcp4.h"
+#include "gro_vxlan_udp4.h"
 
 typedef void *(*gro_tbl_create_fn)(uint16_t socket_id,
                uint16_t max_flow_num,
@@ -20,14 +21,14 @@ typedef uint32_t (*gro_tbl_pkt_count_fn)(void *tbl);
 
 static gro_tbl_create_fn tbl_create_fn[RTE_GRO_TYPE_MAX_NUM] = {
                gro_tcp4_tbl_create, gro_vxlan_tcp4_tbl_create,
-               gro_udp4_tbl_create, NULL};
+               gro_udp4_tbl_create, gro_vxlan_udp4_tbl_create, NULL};
 static gro_tbl_destroy_fn tbl_destroy_fn[RTE_GRO_TYPE_MAX_NUM] = {
                        gro_tcp4_tbl_destroy, gro_vxlan_tcp4_tbl_destroy,
-                       gro_udp4_tbl_destroy,
+                       gro_udp4_tbl_destroy, gro_vxlan_udp4_tbl_destroy,
                        NULL};
 static gro_tbl_pkt_count_fn tbl_pkt_count_fn[RTE_GRO_TYPE_MAX_NUM] = {
                        gro_tcp4_tbl_pkt_count, gro_vxlan_tcp4_tbl_pkt_count,
-                       gro_udp4_tbl_pkt_count,
+                       gro_udp4_tbl_pkt_count, gro_vxlan_udp4_tbl_pkt_count,
                        NULL};
 
 #define IS_IPV4_TCP_PKT(ptype) (RTE_ETH_IS_IPV4_HDR(ptype) && \
@@ -47,6 +48,16 @@ static gro_tbl_pkt_count_fn tbl_pkt_count_fn[RTE_GRO_TYPE_MAX_NUM] = {
                     RTE_PTYPE_INNER_L3_IPV4_EXT | \
                     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN)) != 0))
 
+#define IS_IPV4_VXLAN_UDP4_PKT(ptype) (RTE_ETH_IS_IPV4_HDR(ptype) && \
+               ((ptype & RTE_PTYPE_L4_UDP) == RTE_PTYPE_L4_UDP) && \
+               ((ptype & RTE_PTYPE_TUNNEL_VXLAN) == \
+                RTE_PTYPE_TUNNEL_VXLAN) && \
+                ((ptype & RTE_PTYPE_INNER_L4_UDP) == \
+                 RTE_PTYPE_INNER_L4_UDP) && \
+                 (((ptype & RTE_PTYPE_INNER_L3_MASK) & \
+                   (RTE_PTYPE_INNER_L3_IPV4 | \
+                    RTE_PTYPE_INNER_L3_IPV4_EXT | \
+                    RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN)) != 0))
 
 /*
  * GRO context structure. It keeps the table structures, which are
@@ -137,19 +148,27 @@ rte_gro_reassemble_burst(struct rte_mbuf **pkts,
        struct gro_udp4_item udp_items[RTE_GRO_MAX_BURST_ITEM_NUM] = {{0} };
 
        /* Allocate a reassembly table for VXLAN TCP GRO */
-       struct gro_vxlan_tcp4_tbl vxlan_tbl;
-       struct gro_vxlan_tcp4_flow vxlan_flows[RTE_GRO_MAX_BURST_ITEM_NUM];
-       struct gro_vxlan_tcp4_item vxlan_items[RTE_GRO_MAX_BURST_ITEM_NUM]
+       struct gro_vxlan_tcp4_tbl vxlan_tcp_tbl;
+       struct gro_vxlan_tcp4_flow vxlan_tcp_flows[RTE_GRO_MAX_BURST_ITEM_NUM];
+       struct gro_vxlan_tcp4_item vxlan_tcp_items[RTE_GRO_MAX_BURST_ITEM_NUM]
                        = {{{0}, 0, 0} };
 
+       /* Allocate a reassembly table for VXLAN UDP GRO */
+       struct gro_vxlan_udp4_tbl vxlan_udp_tbl;
+       struct gro_vxlan_udp4_flow vxlan_udp_flows[RTE_GRO_MAX_BURST_ITEM_NUM];
+       struct gro_vxlan_udp4_item vxlan_udp_items[RTE_GRO_MAX_BURST_ITEM_NUM]
+                       = {{{0}} };
+
        struct rte_mbuf *unprocess_pkts[nb_pkts];
        uint32_t item_num;
        int32_t ret;
        uint16_t i, unprocess_num = 0, nb_after_gro = nb_pkts;
-       uint8_t do_tcp4_gro = 0, do_vxlan_gro = 0, do_udp4_gro = 0;
+       uint8_t do_tcp4_gro = 0, do_vxlan_tcp_gro = 0, do_udp4_gro = 0,
+               do_vxlan_udp_gro = 0;
 
        if (unlikely((param->gro_types & (RTE_GRO_IPV4_VXLAN_TCP_IPV4 |
                                        RTE_GRO_TCP_IPV4 |
+                                       RTE_GRO_IPV4_VXLAN_UDP_IPV4 |
                                        RTE_GRO_UDP_IPV4)) == 0))
                return nb_pkts;
 
@@ -160,15 +179,28 @@ rte_gro_reassemble_burst(struct rte_mbuf **pkts,
 
        if (param->gro_types & RTE_GRO_IPV4_VXLAN_TCP_IPV4) {
                for (i = 0; i < item_num; i++)
-                       vxlan_flows[i].start_index = INVALID_ARRAY_INDEX;
-
-               vxlan_tbl.flows = vxlan_flows;
-               vxlan_tbl.items = vxlan_items;
-               vxlan_tbl.flow_num = 0;
-               vxlan_tbl.item_num = 0;
-               vxlan_tbl.max_flow_num = item_num;
-               vxlan_tbl.max_item_num = item_num;
-               do_vxlan_gro = 1;
+                       vxlan_tcp_flows[i].start_index = INVALID_ARRAY_INDEX;
+
+               vxlan_tcp_tbl.flows = vxlan_tcp_flows;
+               vxlan_tcp_tbl.items = vxlan_tcp_items;
+               vxlan_tcp_tbl.flow_num = 0;
+               vxlan_tcp_tbl.item_num = 0;
+               vxlan_tcp_tbl.max_flow_num = item_num;
+               vxlan_tcp_tbl.max_item_num = item_num;
+               do_vxlan_tcp_gro = 1;
+       }
+
+       if (param->gro_types & RTE_GRO_IPV4_VXLAN_UDP_IPV4) {
+               for (i = 0; i < item_num; i++)
+                       vxlan_udp_flows[i].start_index = INVALID_ARRAY_INDEX;
+
+               vxlan_udp_tbl.flows = vxlan_udp_flows;
+               vxlan_udp_tbl.items = vxlan_udp_items;
+               vxlan_udp_tbl.flow_num = 0;
+               vxlan_udp_tbl.item_num = 0;
+               vxlan_udp_tbl.max_flow_num = item_num;
+               vxlan_udp_tbl.max_item_num = item_num;
+               do_vxlan_udp_gro = 1;
        }
 
        if (param->gro_types & RTE_GRO_TCP_IPV4) {
@@ -204,9 +236,18 @@ rte_gro_reassemble_burst(struct rte_mbuf **pkts,
                 * will be flushed from the tables.
                 */
                if (IS_IPV4_VXLAN_TCP4_PKT(pkts[i]->packet_type) &&
-                               do_vxlan_gro) {
+                               do_vxlan_tcp_gro) {
                        ret = gro_vxlan_tcp4_reassemble(pkts[i],
-                                                       &vxlan_tbl, 0);
+                                                       &vxlan_tcp_tbl, 0);
+                       if (ret > 0)
+                               /* Merge successfully */
+                               nb_after_gro--;
+                       else if (ret < 0)
+                               unprocess_pkts[unprocess_num++] = pkts[i];
+               } else if (IS_IPV4_VXLAN_UDP4_PKT(pkts[i]->packet_type) &&
+                               do_vxlan_udp_gro) {
+                       ret = gro_vxlan_udp4_reassemble(pkts[i],
+                                                       &vxlan_udp_tbl, 0);
                        if (ret > 0)
                                /* Merge successfully */
                                nb_after_gro--;
@@ -236,11 +277,17 @@ rte_gro_reassemble_burst(struct rte_mbuf **pkts,
                 || (unprocess_num < nb_pkts)) {
                i = 0;
                /* Flush all packets from the tables */
-               if (do_vxlan_gro) {
-                       i = gro_vxlan_tcp4_tbl_timeout_flush(&vxlan_tbl,
+               if (do_vxlan_tcp_gro) {
+                       i = gro_vxlan_tcp4_tbl_timeout_flush(&vxlan_tcp_tbl,
                                        0, pkts, nb_pkts);
                }
 
+               if (do_vxlan_udp_gro) {
+                       i += gro_vxlan_udp4_tbl_timeout_flush(&vxlan_udp_tbl,
+                                       0, &pkts[i], nb_pkts - i);
+
+               }
+
                if (do_tcp4_gro) {
                        i += gro_tcp4_tbl_timeout_flush(&tcp_tbl, 0,
                                        &pkts[i], nb_pkts - i);
@@ -269,33 +316,42 @@ rte_gro_reassemble(struct rte_mbuf **pkts,
 {
        struct rte_mbuf *unprocess_pkts[nb_pkts];
        struct gro_ctx *gro_ctx = ctx;
-       void *tcp_tbl, *udp_tbl, *vxlan_tbl;
+       void *tcp_tbl, *udp_tbl, *vxlan_tcp_tbl, *vxlan_udp_tbl;
        uint64_t current_time;
        uint16_t i, unprocess_num = 0;
-       uint8_t do_tcp4_gro, do_vxlan_gro, do_udp4_gro;
+       uint8_t do_tcp4_gro, do_vxlan_tcp_gro, do_udp4_gro, do_vxlan_udp_gro;
 
        if (unlikely((gro_ctx->gro_types & (RTE_GRO_IPV4_VXLAN_TCP_IPV4 |
                                        RTE_GRO_TCP_IPV4 |
+                                       RTE_GRO_IPV4_VXLAN_UDP_IPV4 |
                                        RTE_GRO_UDP_IPV4)) == 0))
                return nb_pkts;
 
        tcp_tbl = gro_ctx->tbls[RTE_GRO_TCP_IPV4_INDEX];
-       vxlan_tbl = gro_ctx->tbls[RTE_GRO_IPV4_VXLAN_TCP_IPV4_INDEX];
+       vxlan_tcp_tbl = gro_ctx->tbls[RTE_GRO_IPV4_VXLAN_TCP_IPV4_INDEX];
        udp_tbl = gro_ctx->tbls[RTE_GRO_UDP_IPV4_INDEX];
+       vxlan_udp_tbl = gro_ctx->tbls[RTE_GRO_IPV4_VXLAN_UDP_IPV4_INDEX];
 
        do_tcp4_gro = (gro_ctx->gro_types & RTE_GRO_TCP_IPV4) ==
                RTE_GRO_TCP_IPV4;
-       do_vxlan_gro = (gro_ctx->gro_types & RTE_GRO_IPV4_VXLAN_TCP_IPV4) ==
+       do_vxlan_tcp_gro = (gro_ctx->gro_types & RTE_GRO_IPV4_VXLAN_TCP_IPV4) ==
                RTE_GRO_IPV4_VXLAN_TCP_IPV4;
        do_udp4_gro = (gro_ctx->gro_types & RTE_GRO_UDP_IPV4) ==
                RTE_GRO_UDP_IPV4;
+       do_vxlan_udp_gro = (gro_ctx->gro_types & RTE_GRO_IPV4_VXLAN_UDP_IPV4) ==
+               RTE_GRO_IPV4_VXLAN_UDP_IPV4;
 
        current_time = rte_rdtsc();
 
        for (i = 0; i < nb_pkts; i++) {
                if (IS_IPV4_VXLAN_TCP4_PKT(pkts[i]->packet_type) &&
-                               do_vxlan_gro) {
-                       if (gro_vxlan_tcp4_reassemble(pkts[i], vxlan_tbl,
+                               do_vxlan_tcp_gro) {
+                       if (gro_vxlan_tcp4_reassemble(pkts[i], vxlan_tcp_tbl,
+                                               current_time) < 0)
+                               unprocess_pkts[unprocess_num++] = pkts[i];
+               } else if (IS_IPV4_VXLAN_UDP4_PKT(pkts[i]->packet_type) &&
+                               do_vxlan_udp_gro) {
+                       if (gro_vxlan_udp4_reassemble(pkts[i], vxlan_udp_tbl,
                                                current_time) < 0)
                                unprocess_pkts[unprocess_num++] = pkts[i];
                } else if (IS_IPV4_TCP_PKT(pkts[i]->packet_type) &&
@@ -341,6 +397,13 @@ rte_gro_timeout_flush(void *ctx,
                left_nb_out = max_nb_out - num;
        }
 
+       if ((gro_types & RTE_GRO_IPV4_VXLAN_UDP_IPV4) && left_nb_out > 0) {
+               num += gro_vxlan_udp4_tbl_timeout_flush(gro_ctx->tbls[
+                               RTE_GRO_IPV4_VXLAN_UDP_IPV4_INDEX],
+                               flush_timestamp, &out[num], left_nb_out);
+               left_nb_out = max_nb_out - num;
+       }
+
        /* If no available space in 'out', stop flushing. */
        if ((gro_types & RTE_GRO_TCP_IPV4) && left_nb_out > 0) {
                num += gro_tcp4_tbl_timeout_flush(
index 470f3ed..9f9ed49 100644 (file)
@@ -35,6 +35,9 @@ extern "C" {
 #define RTE_GRO_UDP_IPV4_INDEX 2
 #define RTE_GRO_UDP_IPV4 (1ULL << RTE_GRO_UDP_IPV4_INDEX)
 /**< UDP/IPv4 GRO flag */
+#define RTE_GRO_IPV4_VXLAN_UDP_IPV4_INDEX 3
+#define RTE_GRO_IPV4_VXLAN_UDP_IPV4 (1ULL << RTE_GRO_IPV4_VXLAN_UDP_IPV4_INDEX)
+/**< VxLAN UDP/IPv4 GRO flag. */
 
 /**
  * Structure used to create GRO context objects or used to pass