net/tap: calculate checksums of multi segs packets
authorOphir Munk <ophirmu@mellanox.com>
Sat, 23 Jun 2018 23:17:40 +0000 (23:17 +0000)
committerFerruh Yigit <ferruh.yigit@intel.com>
Mon, 2 Jul 2018 23:35:58 +0000 (01:35 +0200)
Prior to this commit IP/UDP/TCP checksum offload calculations
were skipped in case of a multi segments packet.
This commit enables TAP checksum calculations for multi segments
packets.
The only restriction is that the first segment must contain
headers of layers 3 (IP) and 4 (UDP or TCP)

Reviewed-by: Raslan Darawsheh <rasland@mellanox.com>
Signed-off-by: Ophir Munk <ophirmu@mellanox.com>
Acked-by: Keith Wiles <keith.wiles@intel.com>
drivers/net/tap/rte_eth_tap.c

index df396bf..8903646 100644 (file)
@@ -415,9 +415,41 @@ tap_tx_offload_get_queue_capa(void)
               DEV_TX_OFFLOAD_TCP_CKSUM;
 }
 
+/* Finalize l4 checksum calculation */
 static void
-tap_tx_offload(char *packet, uint64_t ol_flags, unsigned int l2_len,
-              unsigned int l3_len)
+tap_tx_l4_cksum(uint16_t *l4_cksum, uint16_t l4_phdr_cksum,
+               uint32_t l4_raw_cksum)
+{
+       if (l4_cksum) {
+               uint32_t cksum;
+
+               cksum = __rte_raw_cksum_reduce(l4_raw_cksum);
+               cksum += l4_phdr_cksum;
+
+               cksum = ((cksum & 0xffff0000) >> 16) + (cksum & 0xffff);
+               cksum = (~cksum) & 0xffff;
+               if (cksum == 0)
+                       cksum = 0xffff;
+               *l4_cksum = cksum;
+       }
+}
+
+/* Accumaulate L4 raw checksums */
+static void
+tap_tx_l4_add_rcksum(char *l4_data, unsigned int l4_len, uint16_t *l4_cksum,
+                       uint32_t *l4_raw_cksum)
+{
+       if (l4_cksum == NULL)
+               return;
+
+       *l4_raw_cksum = __rte_raw_cksum(l4_data, l4_len, *l4_raw_cksum);
+}
+
+/* L3 and L4 pseudo headers checksum offloads */
+static void
+tap_tx_l3_cksum(char *packet, uint64_t ol_flags, unsigned int l2_len,
+               unsigned int l3_len, unsigned int l4_len, uint16_t **l4_cksum,
+               uint16_t *l4_phdr_cksum, uint32_t *l4_raw_cksum)
 {
        void *l3_hdr = packet + l2_len;
 
@@ -430,38 +462,21 @@ tap_tx_offload(char *packet, uint64_t ol_flags, unsigned int l2_len,
                iph->hdr_checksum = (cksum == 0xffff) ? cksum : ~cksum;
        }
        if (ol_flags & PKT_TX_L4_MASK) {
-               uint16_t l4_len;
-               uint32_t cksum;
-               uint16_t *l4_cksum;
                void *l4_hdr;
 
                l4_hdr = packet + l2_len + l3_len;
                if ((ol_flags & PKT_TX_L4_MASK) == PKT_TX_UDP_CKSUM)
-                       l4_cksum = &((struct udp_hdr *)l4_hdr)->dgram_cksum;
+                       *l4_cksum = &((struct udp_hdr *)l4_hdr)->dgram_cksum;
                else if ((ol_flags & PKT_TX_L4_MASK) == PKT_TX_TCP_CKSUM)
-                       l4_cksum = &((struct tcp_hdr *)l4_hdr)->cksum;
+                       *l4_cksum = &((struct tcp_hdr *)l4_hdr)->cksum;
                else
                        return;
-               *l4_cksum = 0;
-               if (ol_flags & PKT_TX_IPV4) {
-                       struct ipv4_hdr *iph = l3_hdr;
-
-                       l4_len = rte_be_to_cpu_16(iph->total_length) - l3_len;
-                       cksum = rte_ipv4_phdr_cksum(l3_hdr, 0);
-               } else {
-                       struct ipv6_hdr *ip6h = l3_hdr;
-
-                       /* payload_len does not include ext headers */
-                       l4_len = rte_be_to_cpu_16(ip6h->payload_len) -
-                               l3_len + sizeof(struct ipv6_hdr);
-                       cksum = rte_ipv6_phdr_cksum(l3_hdr, 0);
-               }
-               cksum += rte_raw_cksum(l4_hdr, l4_len);
-               cksum = ((cksum & 0xffff0000) >> 16) + (cksum & 0xffff);
-               cksum = (~cksum) & 0xffff;
-               if (cksum == 0)
-                       cksum = 0xffff;
-               *l4_cksum = cksum;
+               **l4_cksum = 0;
+               if (ol_flags & PKT_TX_IPV4)
+                       *l4_phdr_cksum = rte_ipv4_phdr_cksum(l3_hdr, 0);
+               else
+                       *l4_phdr_cksum = rte_ipv6_phdr_cksum(l3_hdr, 0);
+               *l4_raw_cksum = __rte_raw_cksum(l4_hdr, l4_len, 0);
        }
 }
 
@@ -482,17 +497,27 @@ pmd_tx_burst(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
        max_size = *txq->mtu + (ETHER_HDR_LEN + ETHER_CRC_LEN + 4);
        for (i = 0; i < nb_pkts; i++) {
                struct rte_mbuf *mbuf = bufs[num_tx];
-               struct iovec iovecs[mbuf->nb_segs + 1];
+               struct iovec iovecs[mbuf->nb_segs + 2];
                struct tun_pi pi = { .flags = 0, .proto = 0x00 };
                struct rte_mbuf *seg = mbuf;
                char m_copy[mbuf->data_len];
+               int proto;
                int n;
                int j;
+               int k; /* first index in iovecs for copying segments */
+               uint16_t l234_hlen; /* length of layers 2,3,4 headers */
+               uint16_t seg_len; /* length of first segment */
+               uint16_t nb_segs;
+               uint16_t *l4_cksum; /* l4 checksum (pseudo header + payload) */
+               uint32_t l4_raw_cksum = 0; /* TCP/UDP payload raw checksum */
+               uint16_t l4_phdr_cksum = 0; /* TCP/UDP pseudo header checksum */
+               uint16_t is_cksum = 0; /* in case cksum should be offloaded */
 
                /* stats.errs will be incremented */
                if (rte_pktmbuf_pkt_len(mbuf) > max_size)
                        break;
 
+               l4_cksum = NULL;
                if (txq->type == ETH_TUNTAP_TYPE_TUN) {
                        /*
                         * TUN and TAP are created with IFF_NO_PI disabled.
@@ -505,35 +530,77 @@ pmd_tx_burst(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
                         * is 4 or 6, then protocol field is updated.
                         */
                        char *buff_data = rte_pktmbuf_mtod(seg, void *);
-                       j = (*buff_data & 0xf0);
-                       pi.proto = (j == 0x40) ? rte_cpu_to_be_16(ETHER_TYPE_IPv4) :
-                               (j == 0x60) ? rte_cpu_to_be_16(ETHER_TYPE_IPv6) : 0x00;
+                       proto = (*buff_data & 0xf0);
+                       pi.proto = (proto == 0x40) ?
+                               rte_cpu_to_be_16(ETHER_TYPE_IPv4) :
+                               ((proto == 0x60) ?
+                                       rte_cpu_to_be_16(ETHER_TYPE_IPv6) :
+                                       0x00);
                }
 
-               iovecs[0].iov_base = &pi;
-               iovecs[0].iov_len = sizeof(pi);
-               for (j = 1; j <= mbuf->nb_segs; j++) {
-                       iovecs[j].iov_len = rte_pktmbuf_data_len(seg);
-                       iovecs[j].iov_base =
-                               rte_pktmbuf_mtod(seg, void *);
-                       seg = seg->next;
-               }
+               k = 0;
+               iovecs[k].iov_base = &pi;
+               iovecs[k].iov_len = sizeof(pi);
+               k++;
+
+               nb_segs = mbuf->nb_segs;
                if (txq->csum &&
                    ((mbuf->ol_flags & (PKT_TX_IP_CKSUM | PKT_TX_IPV4) ||
                     (mbuf->ol_flags & PKT_TX_L4_MASK) == PKT_TX_UDP_CKSUM ||
                     (mbuf->ol_flags & PKT_TX_L4_MASK) == PKT_TX_TCP_CKSUM))) {
-                       /* Support only packets with all data in the same seg */
-                       if (mbuf->nb_segs > 1)
+                       is_cksum = 1;
+
+                       /* Support only packets with at least layer 4
+                        * header included in the first segment
+                        */
+                       seg_len = rte_pktmbuf_data_len(mbuf);
+                       l234_hlen = mbuf->l2_len + mbuf->l3_len + mbuf->l4_len;
+                       if (seg_len < l234_hlen)
                                break;
-                       /* To change checksums, work on a copy of data. */
+
+                       /* To change checksums, work on a
+                        * copy of l2, l3 l4 headers.
+                        */
                        rte_memcpy(m_copy, rte_pktmbuf_mtod(mbuf, void *),
-                                  rte_pktmbuf_data_len(mbuf));
-                       tap_tx_offload(m_copy, mbuf->ol_flags,
-                                      mbuf->l2_len, mbuf->l3_len);
-                       iovecs[1].iov_base = m_copy;
+                                       l234_hlen);
+                       tap_tx_l3_cksum(m_copy, mbuf->ol_flags,
+                                      mbuf->l2_len, mbuf->l3_len, mbuf->l4_len,
+                                      &l4_cksum, &l4_phdr_cksum,
+                                      &l4_raw_cksum);
+                       iovecs[k].iov_base = m_copy;
+                       iovecs[k].iov_len = l234_hlen;
+                       k++;
+
+                       /* Update next iovecs[] beyond l2, l3, l4 headers */
+                       if (seg_len > l234_hlen) {
+                               iovecs[k].iov_len = seg_len - l234_hlen;
+                               iovecs[k].iov_base =
+                                       rte_pktmbuf_mtod(seg, char *) +
+                                               l234_hlen;
+                               tap_tx_l4_add_rcksum(iovecs[k].iov_base,
+                                       iovecs[k].iov_len, l4_cksum,
+                                       &l4_raw_cksum);
+                               k++;
+                               nb_segs++;
+                       }
+                       seg = seg->next;
+               }
+
+               for (j = k; j <= nb_segs; j++) {
+                       iovecs[j].iov_len = rte_pktmbuf_data_len(seg);
+                       iovecs[j].iov_base = rte_pktmbuf_mtod(seg, void *);
+                       if (is_cksum)
+                               tap_tx_l4_add_rcksum(iovecs[j].iov_base,
+                                       iovecs[j].iov_len, l4_cksum,
+                                       &l4_raw_cksum);
+                       seg = seg->next;
                }
+
+               if (is_cksum)
+                       tap_tx_l4_cksum(l4_cksum, l4_phdr_cksum, l4_raw_cksum);
+
                /* copy the tx frame data */
-               n = writev(txq->fd, iovecs, mbuf->nb_segs + 1);
+               n = writev(txq->fd, iovecs, j);
                if (n <= 0)
                        break;