drivers/net/bnxt/bnxt_rxtx_vec_sse.c

   1 // SPDX-License-Identifier: BSD-3-Clause
   2 /* Copyright(c) 2019 Broadcom All rights reserved. */
   3
   4 #include <inttypes.h>
   5 #include <stdbool.h>
   6
   7 #include <rte_bitmap.h>
   8 #include <rte_byteorder.h>
   9 #include <rte_malloc.h>
  10 #include <rte_memory.h>
  11 #if defined(RTE_ARCH_X86)
  12 #include <tmmintrin.h>
  13 #else
  14 #error "bnxt vector pmd: unsupported target."
  15 #endif
  16
  17 #include "bnxt.h"
  18 #include "bnxt_cpr.h"
  19 #include "bnxt_ring.h"
  20 #include "bnxt_rxr.h"
  21 #include "bnxt_rxq.h"
  22 #include "hsi_struct_def_dpdk.h"
  23 #include "bnxt_rxtx_vec_common.h"
  24
  25 #include "bnxt_txq.h"
  26 #include "bnxt_txr.h"
  27
  28 /*
  29  * RX Ring handling
  30  */
  31
  32 static inline void
  33 bnxt_rxq_rearm(struct bnxt_rx_queue *rxq, struct bnxt_rx_ring_info *rxr)
  34 {
  35         struct rx_prod_pkt_bd *rxbds = &rxr->rx_desc_ring[rxq->rxrearm_start];
  36         struct rte_mbuf **rx_bufs = &rxr->rx_buf_ring[rxq->rxrearm_start];
  37         struct rte_mbuf *mb0, *mb1;
  38         int i;
  39
  40         const __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM, 0);
  41         const __m128i addrmask = _mm_set_epi64x(UINT64_MAX, 0);
  42
  43         /* Pull RTE_BNXT_RXQ_REARM_THRESH more mbufs into the software ring */
  44         if (rte_mempool_get_bulk(rxq->mb_pool,
  45                                  (void *)rx_bufs,
  46                                  RTE_BNXT_RXQ_REARM_THRESH) < 0) {
  47                 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
  48                         RTE_BNXT_RXQ_REARM_THRESH;
  49
  50                 return;
  51         }
  52
  53         /* Initialize the mbufs in vector, process 2 mbufs in one loop */
  54         for (i = 0; i < RTE_BNXT_RXQ_REARM_THRESH; i += 2, rx_bufs += 2) {
  55                 __m128i buf_addr0, buf_addr1;
  56                 __m128i rxbd0, rxbd1;
  57
  58                 mb0 = rx_bufs[0];
  59                 mb1 = rx_bufs[1];
  60
  61                 /* Load address fields from both mbufs */
  62                 buf_addr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
  63                 buf_addr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
  64
  65                 /* Load both rx descriptors (preserving some existing fields) */
  66                 rxbd0 = _mm_loadu_si128((__m128i *)(rxbds + 0));
  67                 rxbd1 = _mm_loadu_si128((__m128i *)(rxbds + 1));
  68
  69                 /* Add default offset to buffer address. */
  70                 buf_addr0 = _mm_add_epi64(buf_addr0, hdr_room);
  71                 buf_addr1 = _mm_add_epi64(buf_addr1, hdr_room);
  72
  73                 /* Clear all fields except address. */
  74                 buf_addr0 =  _mm_and_si128(buf_addr0, addrmask);
  75                 buf_addr1 =  _mm_and_si128(buf_addr1, addrmask);
  76
  77                 /* Clear address field in descriptor. */
  78                 rxbd0 = _mm_andnot_si128(addrmask, rxbd0);
  79                 rxbd1 = _mm_andnot_si128(addrmask, rxbd1);
  80
  81                 /* Set address field in descriptor. */
  82                 rxbd0 = _mm_add_epi64(rxbd0, buf_addr0);
  83                 rxbd1 = _mm_add_epi64(rxbd1, buf_addr1);
  84
  85                 /* Store descriptors to memory. */
  86                 _mm_store_si128((__m128i *)(rxbds++), rxbd0);
  87                 _mm_store_si128((__m128i *)(rxbds++), rxbd1);
  88         }
  89
  90         rxq->rxrearm_start += RTE_BNXT_RXQ_REARM_THRESH;
  91         bnxt_db_write(&rxr->rx_db, rxq->rxrearm_start - 1);
  92         if (rxq->rxrearm_start >= rxq->nb_rx_desc)
  93                 rxq->rxrearm_start = 0;
  94
  95         rxq->rxrearm_nb -= RTE_BNXT_RXQ_REARM_THRESH;
  96 }
  97
  98 static uint32_t
  99 bnxt_parse_pkt_type(struct rx_pkt_cmpl *rxcmp, struct rx_pkt_cmpl_hi *rxcmp1)
 100 {
 101         uint32_t l3, pkt_type = 0;
 102         uint32_t t_ipcs = 0, ip6 = 0, vlan = 0;
 103         uint32_t flags_type;
 104
 105         vlan = !!(rxcmp1->flags2 &
 106                 rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN));
 107         pkt_type |= vlan ? RTE_PTYPE_L2_ETHER_VLAN : RTE_PTYPE_L2_ETHER;
 108
 109         t_ipcs = !!(rxcmp1->flags2 &
 110                 rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC));
 111         ip6 = !!(rxcmp1->flags2 &
 112                  rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_IP_TYPE));
 113
 114         flags_type = rxcmp->flags_type &
 115                 rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS_ITYPE_MASK);
 116
 117         if (!t_ipcs && !ip6)
 118                 l3 = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
 119         else if (!t_ipcs && ip6)
 120                 l3 = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
 121         else if (t_ipcs && !ip6)
 122                 l3 = RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN;
 123         else
 124                 l3 = RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN;
 125
 126         switch (flags_type) {
 127         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_ICMP):
 128                 if (!t_ipcs)
 129                         pkt_type |= l3 | RTE_PTYPE_L4_ICMP;
 130                 else
 131                         pkt_type |= l3 | RTE_PTYPE_INNER_L4_ICMP;
 132                 break;
 133
 134         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_TCP):
 135                 if (!t_ipcs)
 136                         pkt_type |= l3 | RTE_PTYPE_L4_TCP;
 137                 else
 138                         pkt_type |= l3 | RTE_PTYPE_INNER_L4_TCP;
 139                 break;
 140
 141         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_UDP):
 142                 if (!t_ipcs)
 143                         pkt_type |= l3 | RTE_PTYPE_L4_UDP;
 144                 else
 145                         pkt_type |= l3 | RTE_PTYPE_INNER_L4_UDP;
 146                 break;
 147
 148         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_IP):
 149                 pkt_type |= l3;
 150                 break;
 151         }
 152
 153         return pkt_type;
 154 }
 155
 156 static void
 157 bnxt_parse_csum(struct rte_mbuf *mbuf, struct rx_pkt_cmpl_hi *rxcmp1)
 158 {
 159         uint32_t flags;
 160
 161         flags = flags2_0xf(rxcmp1);
 162         /* IP Checksum */
 163         if (likely(IS_IP_NONTUNNEL_PKT(flags))) {
 164                 if (unlikely(RX_CMP_IP_CS_ERROR(rxcmp1)))
 165                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
 166                 else
 167                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
 168         } else if (IS_IP_TUNNEL_PKT(flags)) {
 169                 if (unlikely(RX_CMP_IP_OUTER_CS_ERROR(rxcmp1) ||
 170                              RX_CMP_IP_CS_ERROR(rxcmp1)))
 171                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
 172                 else
 173                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
 174         } else if (unlikely(RX_CMP_IP_CS_UNKNOWN(rxcmp1))) {
 175                 mbuf->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 176         }
 177
 178         /* L4 Checksum */
 179         if (likely(IS_L4_NONTUNNEL_PKT(flags))) {
 180                 if (unlikely(RX_CMP_L4_INNER_CS_ERR2(rxcmp1)))
 181                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
 182                 else
 183                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 184         } else if (IS_L4_TUNNEL_PKT(flags)) {
 185                 if (unlikely(RX_CMP_L4_INNER_CS_ERR2(rxcmp1)))
 186                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
 187                 else
 188                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 189                 if (unlikely(RX_CMP_L4_OUTER_CS_ERR2(rxcmp1))) {
 190                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_BAD;
 191                 } else if (unlikely(IS_L4_TUNNEL_PKT_ONLY_INNER_L4_CS
 192                                     (flags))) {
 193                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_UNKNOWN;
 194                 } else {
 195                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_GOOD;
 196                 }
 197         } else if (unlikely(RX_CMP_L4_CS_UNKNOWN(rxcmp1))) {
 198                 mbuf->ol_flags |= PKT_RX_L4_CKSUM_UNKNOWN;
 199         }
 200 }
 201
 202 uint16_t
 203 bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 204                    uint16_t nb_pkts)
 205 {
 206         struct bnxt_rx_queue *rxq = rx_queue;
 207         struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
 208         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
 209         uint32_t raw_cons = cpr->cp_raw_cons;
 210         uint32_t cons;
 211         int nb_rx_pkts = 0;
 212         struct rx_pkt_cmpl *rxcmp;
 213         bool evt = false;
 214         const __m128i mbuf_init = _mm_set_epi64x(0, rxq->mbuf_initializer);
 215         const __m128i shuf_msk =
 216                 _mm_set_epi8(15, 14, 13, 12,          /* rss */
 217                              0xFF, 0xFF,              /* vlan_tci (zeroes) */
 218                              3, 2,                    /* data_len */
 219                              0xFF, 0xFF, 3, 2,        /* pkt_len */
 220                              0xFF, 0xFF, 0xFF, 0xFF); /* pkt_type (zeroes) */
 221
 222         /* If Rx Q was stopped return */
 223         if (unlikely(!rxq->rx_started))
 224                 return 0;
 225
 226         if (rxq->rxrearm_nb >= RTE_BNXT_RXQ_REARM_THRESH)
 227                 bnxt_rxq_rearm(rxq, rxr);
 228
 229         /* Return no more than RTE_BNXT_MAX_RX_BURST per call. */
 230         nb_pkts = RTE_MIN(nb_pkts, RTE_BNXT_MAX_RX_BURST);
 231
 232         /*
 233          * Make nb_pkts an integer multiple of RTE_BNXT_DESCS_PER_LOOP.
 234          * nb_pkts < RTE_BNXT_DESCS_PER_LOOP, just return no packet
 235          */
 236         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_BNXT_DESCS_PER_LOOP);
 237         if (!nb_pkts)
 238                 return 0;
 239
 240         /* Handle RX burst request */
 241         while (1) {
 242                 cons = RING_CMP(cpr->cp_ring_struct, raw_cons);
 243
 244                 rxcmp = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[cons];
 245
 246                 if (!CMP_VALID(rxcmp, raw_cons, cpr->cp_ring_struct))
 247                         break;
 248
 249                 if (likely(CMP_TYPE(rxcmp) == RX_PKT_CMPL_TYPE_RX_L2)) {
 250                         struct rx_pkt_cmpl_hi *rxcmp1;
 251                         uint32_t tmp_raw_cons;
 252                         uint16_t cp_cons;
 253                         struct rte_mbuf *mbuf;
 254                         __m128i mm_rxcmp, pkt_mb;
 255
 256                         tmp_raw_cons = NEXT_RAW_CMP(raw_cons);
 257                         cp_cons = RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
 258                         rxcmp1 = (struct rx_pkt_cmpl_hi *)
 259                                                 &cpr->cp_desc_ring[cp_cons];
 260
 261                         if (!CMP_VALID(rxcmp1, tmp_raw_cons,
 262                                        cpr->cp_ring_struct))
 263                                 break;
 264
 265                         raw_cons = tmp_raw_cons;
 266                         cons = rxcmp->opaque;
 267
 268                         mbuf = rxr->rx_buf_ring[cons];
 269                         rte_prefetch0(mbuf);
 270                         rxr->rx_buf_ring[cons] = NULL;
 271
 272                         /* Set constant fields from mbuf initializer. */
 273                         _mm_store_si128((__m128i *)&mbuf->rearm_data,
 274                                         mbuf_init);
 275
 276                         /* Set mbuf pkt_len, data_len, and rss_hash fields. */
 277                         mm_rxcmp = _mm_load_si128((__m128i *)rxcmp);
 278                         pkt_mb = _mm_shuffle_epi8(mm_rxcmp, shuf_msk);
 279                         _mm_storeu_si128((void *)&mbuf->rx_descriptor_fields1,
 280                                          pkt_mb);
 281
 282                         rte_compiler_barrier();
 283
 284                         if (rxcmp->flags_type & RX_PKT_CMPL_FLAGS_RSS_VALID)
 285                                 mbuf->ol_flags |= PKT_RX_RSS_HASH;
 286
 287                         if (rxcmp1->flags2 &
 288                             RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN) {
 289                                 mbuf->vlan_tci = rxcmp1->metadata &
 290                                         (RX_PKT_CMPL_METADATA_VID_MASK |
 291                                         RX_PKT_CMPL_METADATA_DE |
 292                                         RX_PKT_CMPL_METADATA_PRI_MASK);
 293                                 mbuf->ol_flags |=
 294                                         PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
 295                         }
 296
 297                         bnxt_parse_csum(mbuf, rxcmp1);
 298                         mbuf->packet_type = bnxt_parse_pkt_type(rxcmp, rxcmp1);
 299
 300                         rx_pkts[nb_rx_pkts++] = mbuf;
 301                 } else if (!BNXT_NUM_ASYNC_CPR(rxq->bp)) {
 302                         evt =
 303                         bnxt_event_hwrm_resp_handler(rxq->bp,
 304                                                      (struct cmpl_base *)rxcmp);
 305                 }
 306
 307                 raw_cons = NEXT_RAW_CMP(raw_cons);
 308                 if (nb_rx_pkts == nb_pkts || evt)
 309                         break;
 310         }
 311         rxr->rx_prod = RING_ADV(rxr->rx_ring_struct, rxr->rx_prod, nb_rx_pkts);
 312
 313         rxq->rxrearm_nb += nb_rx_pkts;
 314         cpr->cp_raw_cons = raw_cons;
 315         cpr->valid = !!(cpr->cp_raw_cons & cpr->cp_ring_struct->ring_size);
 316         if (nb_rx_pkts || evt)
 317                 bnxt_db_cq(cpr);
 318
 319         return nb_rx_pkts;
 320 }
 321
 322 static void
 323 bnxt_tx_cmp_vec(struct bnxt_tx_queue *txq, int nr_pkts)
 324 {
 325         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 326         struct rte_mbuf **free = txq->free;
 327         uint16_t cons = txr->tx_cons;
 328         unsigned int blk = 0;
 329
 330         while (nr_pkts--) {
 331                 struct bnxt_sw_tx_bd *tx_buf;
 332                 struct rte_mbuf *mbuf;
 333
 334                 tx_buf = &txr->tx_buf_ring[cons];
 335                 cons = RING_NEXT(txr->tx_ring_struct, cons);
 336                 mbuf = rte_pktmbuf_prefree_seg(tx_buf->mbuf);
 337                 if (unlikely(mbuf == NULL))
 338                         continue;
 339                 tx_buf->mbuf = NULL;
 340
 341                 if (blk && mbuf->pool != free[0]->pool) {
 342                         rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
 343                         blk = 0;
 344                 }
 345                 free[blk++] = mbuf;
 346         }
 347         if (blk)
 348                 rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
 349
 350         txr->tx_cons = cons;
 351 }
 352
 353 static void
 354 bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
 355 {
 356         struct bnxt_cp_ring_info *cpr = txq->cp_ring;
 357         uint32_t raw_cons = cpr->cp_raw_cons;
 358         uint32_t cons;
 359         uint32_t nb_tx_pkts = 0;
 360         struct tx_cmpl *txcmp;
 361         struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
 362         struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
 363         uint32_t ring_mask = cp_ring_struct->ring_mask;
 364
 365         do {
 366                 cons = RING_CMPL(ring_mask, raw_cons);
 367                 txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
 368
 369                 if (!CMP_VALID(txcmp, raw_cons, cp_ring_struct))
 370                         break;
 371
 372                 if (likely(CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2))
 373                         nb_tx_pkts += txcmp->opaque;
 374                 else
 375                         RTE_LOG_DP(ERR, PMD,
 376                                    "Unhandled CMP type %02x\n",
 377                                    CMP_TYPE(txcmp));
 378                 raw_cons = NEXT_RAW_CMP(raw_cons);
 379         } while (nb_tx_pkts < ring_mask);
 380
 381         cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);
 382         if (nb_tx_pkts) {
 383                 bnxt_tx_cmp_vec(txq, nb_tx_pkts);
 384                 cpr->cp_raw_cons = raw_cons;
 385                 bnxt_db_cq(cpr);
 386         }
 387 }
 388
 389 static uint16_t
 390 bnxt_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 391                           uint16_t nb_pkts)
 392 {
 393         struct bnxt_tx_queue *txq = tx_queue;
 394         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 395         uint16_t prod = txr->tx_prod;
 396         struct rte_mbuf *tx_mbuf;
 397         struct tx_bd_long *txbd = NULL;
 398         struct bnxt_sw_tx_bd *tx_buf;
 399         uint16_t to_send;
 400
 401         nb_pkts = RTE_MIN(nb_pkts, bnxt_tx_avail(txq));
 402
 403         if (unlikely(nb_pkts == 0))
 404                 return 0;
 405
 406         /* Handle TX burst request */
 407         to_send = nb_pkts;
 408         while (to_send) {
 409                 tx_mbuf = *tx_pkts++;
 410                 rte_prefetch0(tx_mbuf);
 411
 412                 tx_buf = &txr->tx_buf_ring[prod];
 413                 tx_buf->mbuf = tx_mbuf;
 414                 tx_buf->nr_bds = 1;
 415
 416                 txbd = &txr->tx_desc_ring[prod];
 417                 txbd->address = tx_mbuf->buf_iova + tx_mbuf->data_off;
 418                 txbd->len = tx_mbuf->data_len;
 419                 txbd->flags_type = bnxt_xmit_flags_len(tx_mbuf->data_len,
 420                                                        TX_BD_FLAGS_NOCMPL);
 421                 prod = RING_NEXT(txr->tx_ring_struct, prod);
 422                 to_send--;
 423         }
 424
 425         /* Request a completion for last packet in burst */
 426         if (txbd) {
 427                 txbd->opaque = nb_pkts;
 428                 txbd->flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
 429         }
 430
 431         rte_compiler_barrier();
 432         bnxt_db_write(&txr->tx_db, prod);
 433
 434         txr->tx_prod = prod;
 435
 436         return nb_pkts;
 437 }
 438
 439 uint16_t
 440 bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 441                    uint16_t nb_pkts)
 442 {
 443         int nb_sent = 0;
 444         struct bnxt_tx_queue *txq = tx_queue;
 445
 446         /* Tx queue was stopped; wait for it to be restarted */
 447         if (unlikely(!txq->tx_started)) {
 448                 PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
 449                 return 0;
 450         }
 451
 452         /* Handle TX completions */
 453         if (bnxt_tx_bds_in_hw(txq) >= txq->tx_free_thresh)
 454                 bnxt_handle_tx_cp_vec(txq);
 455
 456         while (nb_pkts) {
 457                 uint16_t ret, num;
 458
 459                 num = RTE_MIN(nb_pkts, RTE_BNXT_MAX_TX_BURST);
 460                 ret = bnxt_xmit_fixed_burst_vec(tx_queue,
 461                                                 &tx_pkts[nb_sent],
 462                                                 num);
 463                 nb_sent += ret;
 464                 nb_pkts -= ret;
 465                 if (ret < num)
 466                         break;
 467         }
 468
 469         return nb_sent;
 470 }
 471
 472 int __rte_cold
 473 bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq)
 474 {
 475         return bnxt_rxq_vec_setup_common(rxq);
 476 }