drivers/net/enic/enic_rxtx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2008-2017 Cisco Systems, Inc.  All rights reserved.
   3  * Copyright 2007 Nuova Systems, Inc.  All rights reserved.
   4  */
   5
   6 #include <rte_mbuf.h>
   7 #include <rte_ethdev_driver.h>
   8 #include <rte_net.h>
   9 #include <rte_prefetch.h>
  10
  11 #include "enic_compat.h"
  12 #include "rq_enet_desc.h"
  13 #include "enic.h"
  14 #include <rte_ether.h>
  15 #include <rte_ip.h>
  16 #include <rte_tcp.h>
  17
  18 #define RTE_PMD_USE_PREFETCH
  19
  20 #ifdef RTE_PMD_USE_PREFETCH
  21 /*Prefetch a cache line into all cache levels. */
  22 #define rte_enic_prefetch(p) rte_prefetch0(p)
  23 #else
  24 #define rte_enic_prefetch(p) do {} while (0)
  25 #endif
  26
  27 #ifdef RTE_PMD_PACKET_PREFETCH
  28 #define rte_packet_prefetch(p) rte_prefetch1(p)
  29 #else
  30 #define rte_packet_prefetch(p) do {} while (0)
  31 #endif
  32
  33 static inline uint16_t
  34 enic_cq_rx_desc_ciflags(struct cq_enet_rq_desc *crd)
  35 {
  36         return le16_to_cpu(crd->completed_index_flags) & ~CQ_DESC_COMP_NDX_MASK;
  37 }
  38
  39 static inline uint16_t
  40 enic_cq_rx_desc_bwflags(struct cq_enet_rq_desc *crd)
  41 {
  42         return le16_to_cpu(crd->bytes_written_flags) &
  43                            ~CQ_ENET_RQ_DESC_BYTES_WRITTEN_MASK;
  44 }
  45
  46 static inline uint8_t
  47 enic_cq_rx_desc_packet_error(uint16_t bwflags)
  48 {
  49         return (bwflags & CQ_ENET_RQ_DESC_FLAGS_TRUNCATED) ==
  50                 CQ_ENET_RQ_DESC_FLAGS_TRUNCATED;
  51 }
  52
  53 static inline uint8_t
  54 enic_cq_rx_desc_eop(uint16_t ciflags)
  55 {
  56         return (ciflags & CQ_ENET_RQ_DESC_FLAGS_EOP)
  57                 == CQ_ENET_RQ_DESC_FLAGS_EOP;
  58 }
  59
  60 static inline uint8_t
  61 enic_cq_rx_desc_csum_not_calc(struct cq_enet_rq_desc *cqrd)
  62 {
  63         return (le16_to_cpu(cqrd->q_number_rss_type_flags) &
  64                 CQ_ENET_RQ_DESC_FLAGS_CSUM_NOT_CALC) ==
  65                 CQ_ENET_RQ_DESC_FLAGS_CSUM_NOT_CALC;
  66 }
  67
  68 static inline uint8_t
  69 enic_cq_rx_desc_ipv4_csum_ok(struct cq_enet_rq_desc *cqrd)
  70 {
  71         return (cqrd->flags & CQ_ENET_RQ_DESC_FLAGS_IPV4_CSUM_OK) ==
  72                 CQ_ENET_RQ_DESC_FLAGS_IPV4_CSUM_OK;
  73 }
  74
  75 static inline uint8_t
  76 enic_cq_rx_desc_tcp_udp_csum_ok(struct cq_enet_rq_desc *cqrd)
  77 {
  78         return (cqrd->flags & CQ_ENET_RQ_DESC_FLAGS_TCP_UDP_CSUM_OK) ==
  79                 CQ_ENET_RQ_DESC_FLAGS_TCP_UDP_CSUM_OK;
  80 }
  81
  82 static inline uint8_t
  83 enic_cq_rx_desc_rss_type(struct cq_enet_rq_desc *cqrd)
  84 {
  85         return (uint8_t)((le16_to_cpu(cqrd->q_number_rss_type_flags) >>
  86                 CQ_DESC_Q_NUM_BITS) & CQ_ENET_RQ_DESC_RSS_TYPE_MASK);
  87 }
  88
  89 static inline uint32_t
  90 enic_cq_rx_desc_rss_hash(struct cq_enet_rq_desc *cqrd)
  91 {
  92         return le32_to_cpu(cqrd->rss_hash);
  93 }
  94
  95 static inline uint16_t
  96 enic_cq_rx_desc_vlan(struct cq_enet_rq_desc *cqrd)
  97 {
  98         return le16_to_cpu(cqrd->vlan);
  99 }
 100
 101 static inline uint16_t
 102 enic_cq_rx_desc_n_bytes(struct cq_desc *cqd)
 103 {
 104         struct cq_enet_rq_desc *cqrd = (struct cq_enet_rq_desc *)cqd;
 105         return le16_to_cpu(cqrd->bytes_written_flags) &
 106                 CQ_ENET_RQ_DESC_BYTES_WRITTEN_MASK;
 107 }
 108
 109
 110 static inline uint8_t
 111 enic_cq_rx_check_err(struct cq_desc *cqd)
 112 {
 113         struct cq_enet_rq_desc *cqrd = (struct cq_enet_rq_desc *)cqd;
 114         uint16_t bwflags;
 115
 116         bwflags = enic_cq_rx_desc_bwflags(cqrd);
 117         if (unlikely(enic_cq_rx_desc_packet_error(bwflags)))
 118                 return 1;
 119         return 0;
 120 }
 121
 122 /* Lookup table to translate RX CQ flags to mbuf flags. */
 123 static inline uint32_t
 124 enic_cq_rx_flags_to_pkt_type(struct cq_desc *cqd, uint8_t tnl)
 125 {
 126         struct cq_enet_rq_desc *cqrd = (struct cq_enet_rq_desc *)cqd;
 127         uint8_t cqrd_flags = cqrd->flags;
 128         /*
 129          * Odd-numbered entries are for tunnel packets. All packet type info
 130          * applies to the inner packet, and there is no info on the outer
 131          * packet. The outer flags in these entries exist only to avoid
 132          * changing enic_cq_rx_to_pkt_flags(). They are cleared from mbuf
 133          * afterwards.
 134          *
 135          * Also, as there is no tunnel type info (VXLAN, NVGRE, or GENEVE), set
 136          * RTE_PTYPE_TUNNEL_GRENAT..
 137          */
 138         static const uint32_t cq_type_table[128] __rte_cache_aligned = {
 139                 [0x00] = RTE_PTYPE_UNKNOWN,
 140                 [0x01] = RTE_PTYPE_UNKNOWN |
 141                          RTE_PTYPE_TUNNEL_GRENAT |
 142                          RTE_PTYPE_INNER_L2_ETHER,
 143                 [0x20] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_NONFRAG,
 144                 [0x21] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_NONFRAG |
 145                          RTE_PTYPE_TUNNEL_GRENAT |
 146                          RTE_PTYPE_INNER_L2_ETHER |
 147                          RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 148                          RTE_PTYPE_INNER_L4_NONFRAG,
 149                 [0x22] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_UDP,
 150                 [0x23] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_UDP |
 151                          RTE_PTYPE_TUNNEL_GRENAT |
 152                          RTE_PTYPE_INNER_L2_ETHER |
 153                          RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 154                          RTE_PTYPE_INNER_L4_UDP,
 155                 [0x24] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_TCP,
 156                 [0x25] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_TCP |
 157                          RTE_PTYPE_TUNNEL_GRENAT |
 158                          RTE_PTYPE_INNER_L2_ETHER |
 159                          RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 160                          RTE_PTYPE_INNER_L4_TCP,
 161                 [0x60] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG,
 162                 [0x61] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG |
 163                          RTE_PTYPE_TUNNEL_GRENAT |
 164                          RTE_PTYPE_INNER_L2_ETHER |
 165                          RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 166                          RTE_PTYPE_INNER_L4_FRAG,
 167                 [0x62] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG,
 168                 [0x63] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG |
 169                          RTE_PTYPE_TUNNEL_GRENAT |
 170                          RTE_PTYPE_INNER_L2_ETHER |
 171                          RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 172                          RTE_PTYPE_INNER_L4_FRAG,
 173                 [0x64] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG,
 174                 [0x65] = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG |
 175                          RTE_PTYPE_TUNNEL_GRENAT |
 176                          RTE_PTYPE_INNER_L2_ETHER |
 177                          RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 178                          RTE_PTYPE_INNER_L4_FRAG,
 179                 [0x10] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_NONFRAG,
 180                 [0x11] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_NONFRAG |
 181                          RTE_PTYPE_TUNNEL_GRENAT |
 182                          RTE_PTYPE_INNER_L2_ETHER |
 183                          RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 184                          RTE_PTYPE_INNER_L4_NONFRAG,
 185                 [0x12] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_UDP,
 186                 [0x13] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_UDP |
 187                          RTE_PTYPE_TUNNEL_GRENAT |
 188                          RTE_PTYPE_INNER_L2_ETHER |
 189                          RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 190                          RTE_PTYPE_INNER_L4_UDP,
 191                 [0x14] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_TCP,
 192                 [0x15] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_TCP |
 193                          RTE_PTYPE_TUNNEL_GRENAT |
 194                          RTE_PTYPE_INNER_L2_ETHER |
 195                          RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 196                          RTE_PTYPE_INNER_L4_TCP,
 197                 [0x50] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG,
 198                 [0x51] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG |
 199                          RTE_PTYPE_TUNNEL_GRENAT |
 200                          RTE_PTYPE_INNER_L2_ETHER |
 201                          RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 202                          RTE_PTYPE_INNER_L4_FRAG,
 203                 [0x52] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG,
 204                 [0x53] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG |
 205                          RTE_PTYPE_TUNNEL_GRENAT |
 206                          RTE_PTYPE_INNER_L2_ETHER |
 207                          RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 208                          RTE_PTYPE_INNER_L4_FRAG,
 209                 [0x54] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG,
 210                 [0x55] = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN | RTE_PTYPE_L4_FRAG |
 211                          RTE_PTYPE_TUNNEL_GRENAT |
 212                          RTE_PTYPE_INNER_L2_ETHER |
 213                          RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 214                          RTE_PTYPE_INNER_L4_FRAG,
 215                 /* All others reserved */
 216         };
 217         cqrd_flags &= CQ_ENET_RQ_DESC_FLAGS_IPV4_FRAGMENT
 218                 | CQ_ENET_RQ_DESC_FLAGS_IPV4 | CQ_ENET_RQ_DESC_FLAGS_IPV6
 219                 | CQ_ENET_RQ_DESC_FLAGS_TCP | CQ_ENET_RQ_DESC_FLAGS_UDP;
 220         return cq_type_table[cqrd_flags + tnl];
 221 }
 222
 223 static inline void
 224 enic_cq_rx_to_pkt_flags(struct cq_desc *cqd, struct rte_mbuf *mbuf)
 225 {
 226         struct cq_enet_rq_desc *cqrd = (struct cq_enet_rq_desc *)cqd;
 227         uint16_t bwflags, pkt_flags = 0, vlan_tci;
 228         bwflags = enic_cq_rx_desc_bwflags(cqrd);
 229         vlan_tci = enic_cq_rx_desc_vlan(cqrd);
 230
 231         /* VLAN STRIPPED flag. The L2 packet type updated here also */
 232         if (bwflags & CQ_ENET_RQ_DESC_FLAGS_VLAN_STRIPPED) {
 233                 pkt_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
 234                 mbuf->packet_type |= RTE_PTYPE_L2_ETHER;
 235         } else {
 236                 if (vlan_tci != 0)
 237                         mbuf->packet_type |= RTE_PTYPE_L2_ETHER_VLAN;
 238                 else
 239                         mbuf->packet_type |= RTE_PTYPE_L2_ETHER;
 240         }
 241         mbuf->vlan_tci = vlan_tci;
 242
 243         if ((cqd->type_color & CQ_DESC_TYPE_MASK) == CQ_DESC_TYPE_CLASSIFIER) {
 244                 struct cq_enet_rq_clsf_desc *clsf_cqd;
 245                 uint16_t filter_id;
 246                 clsf_cqd = (struct cq_enet_rq_clsf_desc *)cqd;
 247                 filter_id = clsf_cqd->filter_id;
 248                 if (filter_id) {
 249                         pkt_flags |= PKT_RX_FDIR;
 250                         if (filter_id != ENIC_MAGIC_FILTER_ID) {
 251                                 mbuf->hash.fdir.hi = clsf_cqd->filter_id;
 252                                 pkt_flags |= PKT_RX_FDIR_ID;
 253                         }
 254                 }
 255         } else if (enic_cq_rx_desc_rss_type(cqrd)) {
 256                 /* RSS flag */
 257                 pkt_flags |= PKT_RX_RSS_HASH;
 258                 mbuf->hash.rss = enic_cq_rx_desc_rss_hash(cqrd);
 259         }
 260
 261         /* checksum flags */
 262         if (mbuf->packet_type & (RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L3_IPV6)) {
 263                 if (!enic_cq_rx_desc_csum_not_calc(cqrd)) {
 264                         uint32_t l4_flags;
 265                         l4_flags = mbuf->packet_type & RTE_PTYPE_L4_MASK;
 266
 267                         /*
 268                          * When overlay offload is enabled, the NIC may
 269                          * set ipv4_csum_ok=1 if the inner packet is IPv6..
 270                          * So, explicitly check for IPv4 before checking
 271                          * ipv4_csum_ok.
 272                          */
 273                         if (mbuf->packet_type & RTE_PTYPE_L3_IPV4) {
 274                                 if (enic_cq_rx_desc_ipv4_csum_ok(cqrd))
 275                                         pkt_flags |= PKT_RX_IP_CKSUM_GOOD;
 276                                 else
 277                                         pkt_flags |= PKT_RX_IP_CKSUM_BAD;
 278                         }
 279
 280                         if (l4_flags == RTE_PTYPE_L4_UDP ||
 281                             l4_flags == RTE_PTYPE_L4_TCP) {
 282                                 if (enic_cq_rx_desc_tcp_udp_csum_ok(cqrd))
 283                                         pkt_flags |= PKT_RX_L4_CKSUM_GOOD;
 284                                 else
 285                                         pkt_flags |= PKT_RX_L4_CKSUM_BAD;
 286                         }
 287                 }
 288         }
 289
 290         mbuf->ol_flags = pkt_flags;
 291 }
 292
 293 /* dummy receive function to replace actual function in
 294  * order to do safe reconfiguration operations.
 295  */
 296 uint16_t
 297 enic_dummy_recv_pkts(__rte_unused void *rx_queue,
 298                      __rte_unused struct rte_mbuf **rx_pkts,
 299                      __rte_unused uint16_t nb_pkts)
 300 {
 301         return 0;
 302 }
 303
 304 uint16_t
 305 enic_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 306                uint16_t nb_pkts)
 307 {
 308         struct vnic_rq *sop_rq = rx_queue;
 309         struct vnic_rq *data_rq;
 310         struct vnic_rq *rq;
 311         struct enic *enic = vnic_dev_priv(sop_rq->vdev);
 312         uint16_t cq_idx;
 313         uint16_t rq_idx, max_rx;
 314         uint16_t rq_num;
 315         struct rte_mbuf *nmb, *rxmb;
 316         uint16_t nb_rx = 0;
 317         struct vnic_cq *cq;
 318         volatile struct cq_desc *cqd_ptr;
 319         uint8_t color;
 320         uint8_t tnl;
 321         uint16_t seg_length;
 322         struct rte_mbuf *first_seg = sop_rq->pkt_first_seg;
 323         struct rte_mbuf *last_seg = sop_rq->pkt_last_seg;
 324
 325         cq = &enic->cq[enic_cq_rq(enic, sop_rq->index)];
 326         cq_idx = cq->to_clean;          /* index of cqd, rqd, mbuf_table */
 327         cqd_ptr = (struct cq_desc *)(cq->ring.descs) + cq_idx;
 328         color = cq->last_color;
 329
 330         data_rq = &enic->rq[sop_rq->data_queue_idx];
 331
 332         /* Receive until the end of the ring, at most. */
 333         max_rx = RTE_MIN(nb_pkts, cq->ring.desc_count - cq_idx);
 334
 335         while (max_rx) {
 336                 volatile struct rq_enet_desc *rqd_ptr;
 337                 struct cq_desc cqd;
 338                 uint8_t packet_error;
 339                 uint16_t ciflags;
 340
 341                 max_rx--;
 342
 343                 /* Check for pkts available */
 344                 if ((cqd_ptr->type_color & CQ_DESC_COLOR_MASK_NOSHIFT) == color)
 345                         break;
 346
 347                 /* Get the cq descriptor and extract rq info from it */
 348                 cqd = *cqd_ptr;
 349                 rq_num = cqd.q_number & CQ_DESC_Q_NUM_MASK;
 350                 rq_idx = cqd.completed_index & CQ_DESC_COMP_NDX_MASK;
 351
 352                 rq = &enic->rq[rq_num];
 353                 rqd_ptr = ((struct rq_enet_desc *)rq->ring.descs) + rq_idx;
 354
 355                 /* allocate a new mbuf */
 356                 nmb = rte_mbuf_raw_alloc(rq->mp);
 357                 if (nmb == NULL) {
 358                         rte_atomic64_inc(&enic->soft_stats.rx_nombuf);
 359                         break;
 360                 }
 361
 362                 /* A packet error means descriptor and data are untrusted */
 363                 packet_error = enic_cq_rx_check_err(&cqd);
 364
 365                 /* Get the mbuf to return and replace with one just allocated */
 366                 rxmb = rq->mbuf_ring[rq_idx];
 367                 rq->mbuf_ring[rq_idx] = nmb;
 368                 cq_idx++;
 369
 370                 /* Prefetch next mbuf & desc while processing current one */
 371                 cqd_ptr = (struct cq_desc *)(cq->ring.descs) + cq_idx;
 372                 rte_enic_prefetch(cqd_ptr);
 373
 374                 ciflags = enic_cq_rx_desc_ciflags(
 375                         (struct cq_enet_rq_desc *)&cqd);
 376
 377                 /* Push descriptor for newly allocated mbuf */
 378                 nmb->data_off = RTE_PKTMBUF_HEADROOM;
 379                 /*
 380                  * Only the address needs to be refilled. length_type of the
 381                  * descriptor it set during initialization
 382                  * (enic_alloc_rx_queue_mbufs) and does not change.
 383                  */
 384                 rqd_ptr->address = rte_cpu_to_le_64(nmb->buf_iova +
 385                                                     RTE_PKTMBUF_HEADROOM);
 386
 387                 /* Fill in the rest of the mbuf */
 388                 seg_length = enic_cq_rx_desc_n_bytes(&cqd);
 389
 390                 if (rq->is_sop) {
 391                         first_seg = rxmb;
 392                         first_seg->pkt_len = seg_length;
 393                 } else {
 394                         first_seg->pkt_len = (uint16_t)(first_seg->pkt_len
 395                                                         + seg_length);
 396                         first_seg->nb_segs++;
 397                         last_seg->next = rxmb;
 398                 }
 399
 400                 rxmb->port = enic->port_id;
 401                 rxmb->data_len = seg_length;
 402
 403                 rq->rx_nb_hold++;
 404
 405                 if (!(enic_cq_rx_desc_eop(ciflags))) {
 406                         last_seg = rxmb;
 407                         continue;
 408                 }
 409
 410                 /*
 411                  * When overlay offload is enabled, CQ.fcoe indicates the
 412                  * packet is tunnelled.
 413                  */
 414                 tnl = enic->overlay_offload &&
 415                         (ciflags & CQ_ENET_RQ_DESC_FLAGS_FCOE) != 0;
 416                 /* cq rx flags are only valid if eop bit is set */
 417                 first_seg->packet_type =
 418                         enic_cq_rx_flags_to_pkt_type(&cqd, tnl);
 419                 enic_cq_rx_to_pkt_flags(&cqd, first_seg);
 420
 421                 /* Wipe the outer types set by enic_cq_rx_flags_to_pkt_type() */
 422                 if (tnl) {
 423                         first_seg->packet_type &= ~(RTE_PTYPE_L3_MASK |
 424                                                     RTE_PTYPE_L4_MASK);
 425                 }
 426                 if (unlikely(packet_error)) {
 427                         rte_pktmbuf_free(first_seg);
 428                         rte_atomic64_inc(&enic->soft_stats.rx_packet_errors);
 429                         continue;
 430                 }
 431
 432
 433                 /* prefetch mbuf data for caller */
 434                 rte_packet_prefetch(RTE_PTR_ADD(first_seg->buf_addr,
 435                                     RTE_PKTMBUF_HEADROOM));
 436
 437                 /* store the mbuf address into the next entry of the array */
 438                 rx_pkts[nb_rx++] = first_seg;
 439         }
 440         if (unlikely(cq_idx == cq->ring.desc_count)) {
 441                 cq_idx = 0;
 442                 cq->last_color ^= CQ_DESC_COLOR_MASK_NOSHIFT;
 443         }
 444
 445         sop_rq->pkt_first_seg = first_seg;
 446         sop_rq->pkt_last_seg = last_seg;
 447
 448         cq->to_clean = cq_idx;
 449
 450         if ((sop_rq->rx_nb_hold + data_rq->rx_nb_hold) >
 451             sop_rq->rx_free_thresh) {
 452                 if (data_rq->in_use) {
 453                         data_rq->posted_index =
 454                                 enic_ring_add(data_rq->ring.desc_count,
 455                                               data_rq->posted_index,
 456                                               data_rq->rx_nb_hold);
 457                         data_rq->rx_nb_hold = 0;
 458                 }
 459                 sop_rq->posted_index = enic_ring_add(sop_rq->ring.desc_count,
 460                                                      sop_rq->posted_index,
 461                                                      sop_rq->rx_nb_hold);
 462                 sop_rq->rx_nb_hold = 0;
 463
 464                 rte_mb();
 465                 if (data_rq->in_use)
 466                         iowrite32_relaxed(data_rq->posted_index,
 467                                           &data_rq->ctrl->posted_index);
 468                 rte_compiler_barrier();
 469                 iowrite32_relaxed(sop_rq->posted_index,
 470                                   &sop_rq->ctrl->posted_index);
 471         }
 472
 473
 474         return nb_rx;
 475 }
 476
 477 uint16_t
 478 enic_noscatter_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 479                          uint16_t nb_pkts)
 480 {
 481         struct rte_mbuf *mb, **rx, **rxmb;
 482         uint16_t cq_idx, nb_rx, max_rx;
 483         struct cq_enet_rq_desc *cqd;
 484         struct rq_enet_desc *rqd;
 485         unsigned int port_id;
 486         struct vnic_cq *cq;
 487         struct vnic_rq *rq;
 488         struct enic *enic;
 489         uint8_t color;
 490         bool overlay;
 491         bool tnl;
 492
 493         rq = rx_queue;
 494         enic = vnic_dev_priv(rq->vdev);
 495         cq = &enic->cq[enic_cq_rq(enic, rq->index)];
 496         cq_idx = cq->to_clean;
 497
 498         /*
 499          * Fill up the reserve of free mbufs. Below, we restock the receive
 500          * ring with these mbufs to avoid allocation failures.
 501          */
 502         if (rq->num_free_mbufs == 0) {
 503                 if (rte_mempool_get_bulk(rq->mp, (void **)rq->free_mbufs,
 504                                          ENIC_RX_BURST_MAX))
 505                         return 0;
 506                 rq->num_free_mbufs = ENIC_RX_BURST_MAX;
 507         }
 508
 509         /* Receive until the end of the ring, at most. */
 510         max_rx = RTE_MIN(nb_pkts, rq->num_free_mbufs);
 511         max_rx = RTE_MIN(max_rx, cq->ring.desc_count - cq_idx);
 512
 513         cqd = (struct cq_enet_rq_desc *)(cq->ring.descs) + cq_idx;
 514         color = cq->last_color;
 515         rxmb = rq->mbuf_ring + cq_idx;
 516         port_id = enic->port_id;
 517         overlay = enic->overlay_offload;
 518
 519         rx = rx_pkts;
 520         while (max_rx) {
 521                 max_rx--;
 522                 if ((cqd->type_color & CQ_DESC_COLOR_MASK_NOSHIFT) == color)
 523                         break;
 524                 if (unlikely(cqd->bytes_written_flags &
 525                              CQ_ENET_RQ_DESC_FLAGS_TRUNCATED)) {
 526                         rte_pktmbuf_free(*rxmb++);
 527                         rte_atomic64_inc(&enic->soft_stats.rx_packet_errors);
 528                         cqd++;
 529                         continue;
 530                 }
 531
 532                 mb = *rxmb++;
 533                 /* prefetch mbuf data for caller */
 534                 rte_packet_prefetch(RTE_PTR_ADD(mb->buf_addr,
 535                                     RTE_PKTMBUF_HEADROOM));
 536                 mb->data_len = cqd->bytes_written_flags &
 537                         CQ_ENET_RQ_DESC_BYTES_WRITTEN_MASK;
 538                 mb->pkt_len = mb->data_len;
 539                 mb->port = port_id;
 540                 tnl = overlay && (cqd->completed_index_flags &
 541                                   CQ_ENET_RQ_DESC_FLAGS_FCOE) != 0;
 542                 mb->packet_type =
 543                         enic_cq_rx_flags_to_pkt_type((struct cq_desc *)cqd,
 544                                                      tnl);
 545                 enic_cq_rx_to_pkt_flags((struct cq_desc *)cqd, mb);
 546                 /* Wipe the outer types set by enic_cq_rx_flags_to_pkt_type() */
 547                 if (tnl) {
 548                         mb->packet_type &= ~(RTE_PTYPE_L3_MASK |
 549                                              RTE_PTYPE_L4_MASK);
 550                 }
 551                 cqd++;
 552                 *rx++ = mb;
 553         }
 554         /* Number of descriptors visited */
 555         nb_rx = cqd - (struct cq_enet_rq_desc *)(cq->ring.descs) - cq_idx;
 556         if (nb_rx == 0)
 557                 return 0;
 558         rqd = ((struct rq_enet_desc *)rq->ring.descs) + cq_idx;
 559         rxmb = rq->mbuf_ring + cq_idx;
 560         cq_idx += nb_rx;
 561         rq->rx_nb_hold += nb_rx;
 562         if (unlikely(cq_idx == cq->ring.desc_count)) {
 563                 cq_idx = 0;
 564                 cq->last_color ^= CQ_DESC_COLOR_MASK_NOSHIFT;
 565         }
 566         cq->to_clean = cq_idx;
 567
 568         memcpy(rxmb, rq->free_mbufs + ENIC_RX_BURST_MAX - rq->num_free_mbufs,
 569                sizeof(struct rte_mbuf *) * nb_rx);
 570         rq->num_free_mbufs -= nb_rx;
 571         while (nb_rx) {
 572                 nb_rx--;
 573                 mb = *rxmb++;
 574                 mb->data_off = RTE_PKTMBUF_HEADROOM;
 575                 rqd->address = mb->buf_iova + RTE_PKTMBUF_HEADROOM;
 576                 rqd++;
 577         }
 578         if (rq->rx_nb_hold > rq->rx_free_thresh) {
 579                 rq->posted_index = enic_ring_add(rq->ring.desc_count,
 580                                                  rq->posted_index,
 581                                                  rq->rx_nb_hold);
 582                 rq->rx_nb_hold = 0;
 583                 rte_wmb();
 584                 iowrite32_relaxed(rq->posted_index,
 585                                   &rq->ctrl->posted_index);
 586         }
 587
 588         return rx - rx_pkts;
 589 }
 590
 591 static void enic_fast_free_wq_bufs(struct vnic_wq *wq, u16 completed_index)
 592 {
 593         unsigned int desc_count, n, nb_to_free, tail_idx;
 594         struct rte_mempool *pool;
 595         struct rte_mbuf **m;
 596
 597         desc_count = wq->ring.desc_count;
 598         nb_to_free = enic_ring_sub(desc_count, wq->tail_idx, completed_index)
 599                                    + 1;
 600         tail_idx = wq->tail_idx;
 601         wq->tail_idx += nb_to_free;
 602         wq->ring.desc_avail += nb_to_free;
 603         if (wq->tail_idx >= desc_count)
 604                 wq->tail_idx -= desc_count;
 605         /* First, free at most until the end of ring */
 606         m = &wq->bufs[tail_idx];
 607         pool = (*m)->pool;
 608         n = RTE_MIN(nb_to_free, desc_count - tail_idx);
 609         rte_mempool_put_bulk(pool, (void **)m, n);
 610         n = nb_to_free - n;
 611         /* Then wrap and free the rest */
 612         if (unlikely(n))
 613                 rte_mempool_put_bulk(pool, (void **)wq->bufs, n);
 614 }
 615
 616 static inline void enic_free_wq_bufs(struct vnic_wq *wq, u16 completed_index)
 617 {
 618         struct rte_mbuf *buf;
 619         struct rte_mbuf *m, *free[ENIC_MAX_WQ_DESCS];
 620         unsigned int nb_to_free, nb_free = 0, i;
 621         struct rte_mempool *pool;
 622         unsigned int tail_idx;
 623         unsigned int desc_count = wq->ring.desc_count;
 624
 625         nb_to_free = enic_ring_sub(desc_count, wq->tail_idx, completed_index)
 626                                    + 1;
 627         tail_idx = wq->tail_idx;
 628         pool = wq->bufs[tail_idx]->pool;
 629         for (i = 0; i < nb_to_free; i++) {
 630                 buf = wq->bufs[tail_idx];
 631                 m = rte_pktmbuf_prefree_seg(buf);
 632                 if (unlikely(m == NULL)) {
 633                         tail_idx = enic_ring_incr(desc_count, tail_idx);
 634                         continue;
 635                 }
 636
 637                 if (likely(m->pool == pool)) {
 638                         RTE_ASSERT(nb_free < ENIC_MAX_WQ_DESCS);
 639                         free[nb_free++] = m;
 640                 } else {
 641                         rte_mempool_put_bulk(pool, (void *)free, nb_free);
 642                         free[0] = m;
 643                         nb_free = 1;
 644                         pool = m->pool;
 645                 }
 646                 tail_idx = enic_ring_incr(desc_count, tail_idx);
 647         }
 648
 649         if (nb_free > 0)
 650                 rte_mempool_put_bulk(pool, (void **)free, nb_free);
 651
 652         wq->tail_idx = tail_idx;
 653         wq->ring.desc_avail += nb_to_free;
 654 }
 655
 656 unsigned int enic_cleanup_wq(__rte_unused struct enic *enic, struct vnic_wq *wq)
 657 {
 658         u16 completed_index;
 659
 660         completed_index = *((uint32_t *)wq->cqmsg_rz->addr) & 0xffff;
 661
 662         if (wq->last_completed_index != completed_index) {
 663                 if (wq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
 664                         enic_fast_free_wq_bufs(wq, completed_index);
 665                 else
 666                         enic_free_wq_bufs(wq, completed_index);
 667                 wq->last_completed_index = completed_index;
 668         }
 669         return 0;
 670 }
 671
 672 uint16_t enic_prep_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 673                         uint16_t nb_pkts)
 674 {
 675         struct vnic_wq *wq = (struct vnic_wq *)tx_queue;
 676         int32_t ret;
 677         uint16_t i;
 678         uint64_t ol_flags;
 679         struct rte_mbuf *m;
 680
 681         for (i = 0; i != nb_pkts; i++) {
 682                 m = tx_pkts[i];
 683                 if (unlikely(m->pkt_len > ENIC_TX_MAX_PKT_SIZE)) {
 684                         rte_errno = EINVAL;
 685                         return i;
 686                 }
 687                 ol_flags = m->ol_flags;
 688                 if (ol_flags & wq->tx_offload_notsup_mask) {
 689                         rte_errno = ENOTSUP;
 690                         return i;
 691                 }
 692 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
 693                 ret = rte_validate_tx_offload(m);
 694                 if (ret != 0) {
 695                         rte_errno = ret;
 696                         return i;
 697                 }
 698 #endif
 699                 ret = rte_net_intel_cksum_prepare(m);
 700                 if (ret != 0) {
 701                         rte_errno = ret;
 702                         return i;
 703                 }
 704         }
 705
 706         return i;
 707 }
 708
 709 uint16_t enic_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 710         uint16_t nb_pkts)
 711 {
 712         uint16_t index;
 713         unsigned int pkt_len, data_len;
 714         unsigned int nb_segs;
 715         struct rte_mbuf *tx_pkt;
 716         struct vnic_wq *wq = (struct vnic_wq *)tx_queue;
 717         struct enic *enic = vnic_dev_priv(wq->vdev);
 718         unsigned short vlan_id;
 719         uint64_t ol_flags;
 720         uint64_t ol_flags_mask;
 721         unsigned int wq_desc_avail;
 722         int head_idx;
 723         unsigned int desc_count;
 724         struct wq_enet_desc *descs, *desc_p, desc_tmp;
 725         uint16_t mss;
 726         uint8_t vlan_tag_insert;
 727         uint8_t eop, cq;
 728         uint64_t bus_addr;
 729         uint8_t offload_mode;
 730         uint16_t header_len;
 731         uint64_t tso;
 732         rte_atomic64_t *tx_oversized;
 733
 734         enic_cleanup_wq(enic, wq);
 735         wq_desc_avail = vnic_wq_desc_avail(wq);
 736         head_idx = wq->head_idx;
 737         desc_count = wq->ring.desc_count;
 738         ol_flags_mask = PKT_TX_VLAN_PKT | PKT_TX_IP_CKSUM | PKT_TX_L4_MASK;
 739         tx_oversized = &enic->soft_stats.tx_oversized;
 740
 741         nb_pkts = RTE_MIN(nb_pkts, ENIC_TX_XMIT_MAX);
 742
 743         for (index = 0; index < nb_pkts; index++) {
 744                 tx_pkt = *tx_pkts++;
 745                 pkt_len = tx_pkt->pkt_len;
 746                 data_len = tx_pkt->data_len;
 747                 ol_flags = tx_pkt->ol_flags;
 748                 nb_segs = tx_pkt->nb_segs;
 749                 tso = ol_flags & PKT_TX_TCP_SEG;
 750
 751                 /* drop packet if it's too big to send */
 752                 if (unlikely(!tso && pkt_len > ENIC_TX_MAX_PKT_SIZE)) {
 753                         rte_pktmbuf_free(tx_pkt);
 754                         rte_atomic64_inc(tx_oversized);
 755                         continue;
 756                 }
 757
 758                 if (nb_segs > wq_desc_avail) {
 759                         if (index > 0)
 760                                 goto post;
 761                         goto done;
 762                 }
 763
 764                 mss = 0;
 765                 vlan_id = tx_pkt->vlan_tci;
 766                 vlan_tag_insert = !!(ol_flags & PKT_TX_VLAN_PKT);
 767                 bus_addr = (dma_addr_t)
 768                            (tx_pkt->buf_iova + tx_pkt->data_off);
 769
 770                 descs = (struct wq_enet_desc *)wq->ring.descs;
 771                 desc_p = descs + head_idx;
 772
 773                 eop = (data_len == pkt_len);
 774                 offload_mode = WQ_ENET_OFFLOAD_MODE_CSUM;
 775                 header_len = 0;
 776
 777                 if (tso) {
 778                         header_len = tx_pkt->l2_len + tx_pkt->l3_len +
 779                                      tx_pkt->l4_len;
 780
 781                         /* Drop if non-TCP packet or TSO seg size is too big */
 782                         if (unlikely(header_len == 0 || ((tx_pkt->tso_segsz +
 783                             header_len) > ENIC_TX_MAX_PKT_SIZE))) {
 784                                 rte_pktmbuf_free(tx_pkt);
 785                                 rte_atomic64_inc(tx_oversized);
 786                                 continue;
 787                         }
 788
 789                         offload_mode = WQ_ENET_OFFLOAD_MODE_TSO;
 790                         mss = tx_pkt->tso_segsz;
 791                         /* For tunnel, need the size of outer+inner headers */
 792                         if (ol_flags & PKT_TX_TUNNEL_MASK) {
 793                                 header_len += tx_pkt->outer_l2_len +
 794                                         tx_pkt->outer_l3_len;
 795                         }
 796                 }
 797
 798                 if ((ol_flags & ol_flags_mask) && (header_len == 0)) {
 799                         if (ol_flags & PKT_TX_IP_CKSUM)
 800                                 mss |= ENIC_CALC_IP_CKSUM;
 801
 802                         /* Nic uses just 1 bit for UDP and TCP */
 803                         switch (ol_flags & PKT_TX_L4_MASK) {
 804                         case PKT_TX_TCP_CKSUM:
 805                         case PKT_TX_UDP_CKSUM:
 806                                 mss |= ENIC_CALC_TCP_UDP_CKSUM;
 807                                 break;
 808                         }
 809                 }
 810                 wq->cq_pend++;
 811                 cq = 0;
 812                 if (eop && wq->cq_pend >= ENIC_WQ_CQ_THRESH) {
 813                         cq = 1;
 814                         wq->cq_pend = 0;
 815                 }
 816                 wq_enet_desc_enc(&desc_tmp, bus_addr, data_len, mss, header_len,
 817                                  offload_mode, eop, cq, 0, vlan_tag_insert,
 818                                  vlan_id, 0);
 819
 820                 *desc_p = desc_tmp;
 821                 wq->bufs[head_idx] = tx_pkt;
 822                 head_idx = enic_ring_incr(desc_count, head_idx);
 823                 wq_desc_avail--;
 824
 825                 if (!eop) {
 826                         for (tx_pkt = tx_pkt->next; tx_pkt; tx_pkt =
 827                             tx_pkt->next) {
 828                                 data_len = tx_pkt->data_len;
 829
 830                                 wq->cq_pend++;
 831                                 cq = 0;
 832                                 if (tx_pkt->next == NULL) {
 833                                         eop = 1;
 834                                         if (wq->cq_pend >= ENIC_WQ_CQ_THRESH) {
 835                                                 cq = 1;
 836                                                 wq->cq_pend = 0;
 837                                         }
 838                                 }
 839                                 desc_p = descs + head_idx;
 840                                 bus_addr = (dma_addr_t)(tx_pkt->buf_iova
 841                                            + tx_pkt->data_off);
 842                                 wq_enet_desc_enc((struct wq_enet_desc *)
 843                                                  &desc_tmp, bus_addr, data_len,
 844                                                  mss, 0, offload_mode, eop, cq,
 845                                                  0, vlan_tag_insert, vlan_id,
 846                                                  0);
 847
 848                                 *desc_p = desc_tmp;
 849                                 wq->bufs[head_idx] = tx_pkt;
 850                                 head_idx = enic_ring_incr(desc_count, head_idx);
 851                                 wq_desc_avail--;
 852                         }
 853                 }
 854         }
 855  post:
 856         rte_wmb();
 857         iowrite32_relaxed(head_idx, &wq->ctrl->posted_index);
 858  done:
 859         wq->ring.desc_avail = wq_desc_avail;
 860         wq->head_idx = head_idx;
 861
 862         return index;
 863 }
 864
 865 static void enqueue_simple_pkts(struct rte_mbuf **pkts,
 866                                 struct wq_enet_desc *desc,
 867                                 uint16_t n,
 868                                 struct enic *enic)
 869 {
 870         struct rte_mbuf *p;
 871
 872         while (n) {
 873                 n--;
 874                 p = *pkts++;
 875                 desc->address = p->buf_iova + p->data_off;
 876                 desc->length = p->pkt_len;
 877                 /*
 878                  * The app should not send oversized
 879                  * packets. tx_pkt_prepare includes a check as
 880                  * well. But some apps ignore the device max size and
 881                  * tx_pkt_prepare. Oversized packets cause WQ errrors
 882                  * and the NIC ends up disabling the whole WQ. So
 883                  * truncate packets..
 884                  */
 885                 if (unlikely(p->pkt_len > ENIC_TX_MAX_PKT_SIZE)) {
 886                         desc->length = ENIC_TX_MAX_PKT_SIZE;
 887                         rte_atomic64_inc(&enic->soft_stats.tx_oversized);
 888                 }
 889                 desc++;
 890         }
 891 }
 892
 893 uint16_t enic_simple_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 894                                uint16_t nb_pkts)
 895 {
 896         unsigned int head_idx, desc_count;
 897         struct wq_enet_desc *desc;
 898         struct vnic_wq *wq;
 899         struct enic *enic;
 900         uint16_t rem, n;
 901
 902         wq = (struct vnic_wq *)tx_queue;
 903         enic = vnic_dev_priv(wq->vdev);
 904         enic_cleanup_wq(enic, wq);
 905         /* Will enqueue this many packets in this call */
 906         nb_pkts = RTE_MIN(nb_pkts, wq->ring.desc_avail);
 907         if (nb_pkts == 0)
 908                 return 0;
 909
 910         head_idx = wq->head_idx;
 911         desc_count = wq->ring.desc_count;
 912
 913         /* Descriptors until the end of the ring */
 914         n = desc_count - head_idx;
 915         n = RTE_MIN(nb_pkts, n);
 916
 917         /* Save mbuf pointers to free later */
 918         memcpy(wq->bufs + head_idx, tx_pkts, sizeof(struct rte_mbuf *) * n);
 919
 920         /* Enqueue until the ring end */
 921         rem = nb_pkts - n;
 922         desc = ((struct wq_enet_desc *)wq->ring.descs) + head_idx;
 923         enqueue_simple_pkts(tx_pkts, desc, n, enic);
 924
 925         /* Wrap to the start of the ring */
 926         if (rem) {
 927                 tx_pkts += n;
 928                 memcpy(wq->bufs, tx_pkts, sizeof(struct rte_mbuf *) * rem);
 929                 desc = (struct wq_enet_desc *)wq->ring.descs;
 930                 enqueue_simple_pkts(tx_pkts, desc, rem, enic);
 931         }
 932         rte_wmb();
 933
 934         /* Update head_idx and desc_avail */
 935         wq->ring.desc_avail -= nb_pkts;
 936         head_idx += nb_pkts;
 937         if (head_idx >= desc_count)
 938                 head_idx -= desc_count;
 939         wq->head_idx = head_idx;
 940         iowrite32_relaxed(head_idx, &wq->ctrl->posted_index);
 941         return nb_pkts;
 942 }