drivers/net/bnxt/bnxt_rxtx_vec_sse.c

   1 /* SPDX-License-Identifier: BSD-3-Clause */
   2 /* Copyright(c) 2019-2021 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 #include <rte_vect.h>
  12
  13 #include "bnxt.h"
  14 #include "bnxt_cpr.h"
  15 #include "bnxt_ring.h"
  16
  17 #include "bnxt_txq.h"
  18 #include "bnxt_txr.h"
  19 #include "bnxt_rxtx_vec_common.h"
  20
  21 /*
  22  * RX Ring handling
  23  */
  24
  25 #define GET_OL_FLAGS(rss_flags, ol_index, errors, pi, ol_flags)                \
  26 {                                                                              \
  27         uint32_t tmp, of;                                                      \
  28                                                                                \
  29         of = _mm_extract_epi32((rss_flags), (pi)) |                            \
  30                 rxr->ol_flags_table[_mm_extract_epi32((ol_index), (pi))];      \
  31                                                                                \
  32         tmp = _mm_extract_epi32((errors), (pi));                               \
  33         if (tmp)                                                               \
  34                 of |= rxr->ol_flags_err_table[tmp];                            \
  35         (ol_flags) = of;                                                       \
  36 }
  37
  38 #define GET_DESC_FIELDS(rxcmp, rxcmp1, shuf_msk, ptype_idx, pi, ret)           \
  39 {                                                                              \
  40         uint32_t ptype;                                                        \
  41         __m128i r;                                                             \
  42                                                                                \
  43         /* Set mbuf pkt_len, data_len, and rss_hash fields. */                 \
  44         r = _mm_shuffle_epi8((rxcmp), (shuf_msk));                             \
  45                                                                                \
  46         /* Set packet type. */                                                 \
  47         ptype = bnxt_ptype_table[_mm_extract_epi32((ptype_idx), (pi))];        \
  48         r = _mm_blend_epi16(r, _mm_set_epi32(0, 0, 0, ptype), 0x3);            \
  49                                                                                \
  50         /* Set vlan_tci. */                                                    \
  51         r = _mm_blend_epi16(r, _mm_slli_si128((rxcmp1), 6), 0x20);             \
  52         (ret) = r;                                                             \
  53 }
  54
  55 static inline void
  56 descs_to_mbufs(__m128i mm_rxcmp[4], __m128i mm_rxcmp1[4],
  57                __m128i mbuf_init, struct rte_mbuf **mbuf,
  58                struct bnxt_rx_ring_info *rxr)
  59 {
  60         const __m128i shuf_msk =
  61                 _mm_set_epi8(15, 14, 13, 12,          /* rss */
  62                              0xFF, 0xFF,              /* vlan_tci (zeroes) */
  63                              3, 2,                    /* data_len */
  64                              0xFF, 0xFF, 3, 2,        /* pkt_len */
  65                              0xFF, 0xFF, 0xFF, 0xFF); /* pkt_type (zeroes) */
  66         const __m128i flags_type_mask =
  67                 _mm_set1_epi32(RX_PKT_CMPL_FLAGS_ITYPE_MASK);
  68         const __m128i flags2_mask1 =
  69                 _mm_set1_epi32(CMPL_FLAGS2_VLAN_TUN_MSK);
  70         const __m128i flags2_mask2 =
  71                 _mm_set1_epi32(RX_PKT_CMPL_FLAGS2_IP_TYPE);
  72         const __m128i rss_mask =
  73                 _mm_set1_epi32(RX_PKT_CMPL_FLAGS_RSS_VALID);
  74         __m128i t0, t1, flags_type, flags2, index, errors, rss_flags;
  75         __m128i ptype_idx, is_tunnel;
  76         uint32_t ol_flags;
  77
  78         /* Validate ptype table indexing at build time. */
  79         bnxt_check_ptype_constants();
  80
  81         /* Compute packet type table indexes for four packets */
  82         t0 = _mm_unpacklo_epi32(mm_rxcmp[0], mm_rxcmp[1]);
  83         t1 = _mm_unpacklo_epi32(mm_rxcmp[2], mm_rxcmp[3]);
  84         flags_type = _mm_unpacklo_epi64(t0, t1);
  85         ptype_idx = _mm_srli_epi32(_mm_and_si128(flags_type, flags_type_mask),
  86                         RX_PKT_CMPL_FLAGS_ITYPE_SFT - BNXT_PTYPE_TBL_TYPE_SFT);
  87
  88         t0 = _mm_unpacklo_epi32(mm_rxcmp1[0], mm_rxcmp1[1]);
  89         t1 = _mm_unpacklo_epi32(mm_rxcmp1[2], mm_rxcmp1[3]);
  90         flags2 = _mm_unpacklo_epi64(t0, t1);
  91
  92         ptype_idx = _mm_or_si128(ptype_idx,
  93                         _mm_srli_epi32(_mm_and_si128(flags2, flags2_mask1),
  94                                        RX_PKT_CMPL_FLAGS2_META_FORMAT_SFT -
  95                                        BNXT_PTYPE_TBL_VLAN_SFT));
  96         ptype_idx = _mm_or_si128(ptype_idx,
  97                         _mm_srli_epi32(_mm_and_si128(flags2, flags2_mask2),
  98                                        RX_PKT_CMPL_FLAGS2_IP_TYPE_SFT -
  99                                        BNXT_PTYPE_TBL_IP_VER_SFT));
 100
 101         /* Extract RSS valid flags for four packets. */
 102         rss_flags = _mm_srli_epi32(_mm_and_si128(flags_type, rss_mask), 9);
 103
 104         /* Extract errors_v2 fields for four packets. */
 105         t0 = _mm_unpackhi_epi32(mm_rxcmp1[0], mm_rxcmp1[1]);
 106         t1 = _mm_unpackhi_epi32(mm_rxcmp1[2], mm_rxcmp1[3]);
 107
 108         /* Compute ol_flags and checksum error indexes for four packets. */
 109         is_tunnel = _mm_and_si128(flags2, _mm_set1_epi32(4));
 110         is_tunnel = _mm_slli_epi32(is_tunnel, 3);
 111         flags2 = _mm_and_si128(flags2, _mm_set1_epi32(0x1F));
 112
 113         errors = _mm_srli_epi32(_mm_unpacklo_epi64(t0, t1), 4);
 114         errors = _mm_and_si128(errors, _mm_set1_epi32(0xF));
 115         errors = _mm_and_si128(errors, flags2);
 116
 117         index = _mm_andnot_si128(errors, flags2);
 118         errors = _mm_or_si128(errors, _mm_srli_epi32(is_tunnel, 1));
 119         index = _mm_or_si128(index, is_tunnel);
 120
 121         /* Update mbuf rearm_data for four packets. */
 122         GET_OL_FLAGS(rss_flags, index, errors, 0, ol_flags);
 123         _mm_store_si128((void *)&mbuf[0]->rearm_data,
 124                         _mm_or_si128(mbuf_init, _mm_set_epi64x(ol_flags, 0)));
 125
 126         GET_OL_FLAGS(rss_flags, index, errors, 1, ol_flags);
 127         _mm_store_si128((void *)&mbuf[1]->rearm_data,
 128                         _mm_or_si128(mbuf_init, _mm_set_epi64x(ol_flags, 0)));
 129
 130         GET_OL_FLAGS(rss_flags, index, errors, 2, ol_flags);
 131         _mm_store_si128((void *)&mbuf[2]->rearm_data,
 132                         _mm_or_si128(mbuf_init, _mm_set_epi64x(ol_flags, 0)));
 133
 134         GET_OL_FLAGS(rss_flags, index, errors, 3, ol_flags);
 135         _mm_store_si128((void *)&mbuf[3]->rearm_data,
 136                         _mm_or_si128(mbuf_init, _mm_set_epi64x(ol_flags, 0)));
 137
 138         /* Update mbuf rx_descriptor_fields1 for four packes. */
 139         GET_DESC_FIELDS(mm_rxcmp[0], mm_rxcmp1[0], shuf_msk, ptype_idx, 0, t0);
 140         _mm_store_si128((void *)&mbuf[0]->rx_descriptor_fields1, t0);
 141
 142         GET_DESC_FIELDS(mm_rxcmp[1], mm_rxcmp1[1], shuf_msk, ptype_idx, 1, t0);
 143         _mm_store_si128((void *)&mbuf[1]->rx_descriptor_fields1, t0);
 144
 145         GET_DESC_FIELDS(mm_rxcmp[2], mm_rxcmp1[2], shuf_msk, ptype_idx, 2, t0);
 146         _mm_store_si128((void *)&mbuf[2]->rx_descriptor_fields1, t0);
 147
 148         GET_DESC_FIELDS(mm_rxcmp[3], mm_rxcmp1[3], shuf_msk, ptype_idx, 3, t0);
 149         _mm_store_si128((void *)&mbuf[3]->rx_descriptor_fields1, t0);
 150 }
 151
 152 uint16_t
 153 bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 154                    uint16_t nb_pkts)
 155 {
 156         struct bnxt_rx_queue *rxq = rx_queue;
 157         const __m128i mbuf_init = _mm_set_epi64x(0, rxq->mbuf_initializer);
 158         struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
 159         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
 160         uint16_t cp_ring_size = cpr->cp_ring_struct->ring_size;
 161         uint16_t rx_ring_size = rxr->rx_ring_struct->ring_size;
 162         struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
 163         uint64_t valid, desc_valid_mask = ~0ULL;
 164         const __m128i info3_v_mask = _mm_set1_epi32(CMPL_BASE_V);
 165         uint32_t raw_cons = cpr->cp_raw_cons;
 166         uint32_t cons, mbcons;
 167         int nb_rx_pkts = 0;
 168         const __m128i valid_target =
 169                 _mm_set1_epi32(!!(raw_cons & cp_ring_size));
 170         int i;
 171
 172         /* If Rx Q was stopped return */
 173         if (unlikely(!rxq->rx_started))
 174                 return 0;
 175
 176         if (rxq->rxrearm_nb >= rxq->rx_free_thresh)
 177                 bnxt_rxq_rearm(rxq, rxr);
 178
 179         /* Return no more than RTE_BNXT_MAX_RX_BURST per call. */
 180         nb_pkts = RTE_MIN(nb_pkts, RTE_BNXT_MAX_RX_BURST);
 181
 182         cons = raw_cons & (cp_ring_size - 1);
 183         mbcons = (raw_cons / 2) & (rx_ring_size - 1);
 184
 185         /* Prefetch first four descriptor pairs. */
 186         rte_prefetch0(&cp_desc_ring[cons]);
 187         rte_prefetch0(&cp_desc_ring[cons + 4]);
 188
 189         /* Ensure that we do not go past the ends of the rings. */
 190         nb_pkts = RTE_MIN(nb_pkts, RTE_MIN(rx_ring_size - mbcons,
 191                                            (cp_ring_size - cons) / 2));
 192         /*
 193          * If we are at the end of the ring, ensure that descriptors after the
 194          * last valid entry are not treated as valid. Otherwise, force the
 195          * maximum number of packets to receive to be a multiple of the per-
 196          * loop count.
 197          */
 198         if (nb_pkts < RTE_BNXT_DESCS_PER_LOOP)
 199                 desc_valid_mask >>= 16 * (RTE_BNXT_DESCS_PER_LOOP - nb_pkts);
 200         else
 201                 nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_BNXT_DESCS_PER_LOOP);
 202
 203         /* Handle RX burst request */
 204         for (i = 0; i < nb_pkts; i += RTE_BNXT_DESCS_PER_LOOP,
 205                                   cons += RTE_BNXT_DESCS_PER_LOOP * 2,
 206                                   mbcons += RTE_BNXT_DESCS_PER_LOOP) {
 207                 __m128i rxcmp1[RTE_BNXT_DESCS_PER_LOOP];
 208                 __m128i rxcmp[RTE_BNXT_DESCS_PER_LOOP];
 209                 __m128i tmp0, tmp1, info3_v;
 210                 uint32_t num_valid;
 211
 212                 /* Copy four mbuf pointers to output array. */
 213                 tmp0 = _mm_loadu_si128((void *)&rxr->rx_buf_ring[mbcons]);
 214 #ifdef RTE_ARCH_X86_64
 215                 tmp1 = _mm_loadu_si128((void *)&rxr->rx_buf_ring[mbcons + 2]);
 216 #endif
 217                 _mm_storeu_si128((void *)&rx_pkts[i], tmp0);
 218 #ifdef RTE_ARCH_X86_64
 219                 _mm_storeu_si128((void *)&rx_pkts[i + 2], tmp1);
 220 #endif
 221
 222                 /* Prefetch four descriptor pairs for next iteration. */
 223                 if (i + RTE_BNXT_DESCS_PER_LOOP < nb_pkts) {
 224                         rte_prefetch0(&cp_desc_ring[cons + 8]);
 225                         rte_prefetch0(&cp_desc_ring[cons + 12]);
 226                 }
 227
 228                 /*
 229                  * Load the four current descriptors into SSE registers in
 230                  * reverse order to ensure consistent state.
 231                  */
 232                 rxcmp1[3] = _mm_load_si128((void *)&cp_desc_ring[cons + 7]);
 233                 rte_compiler_barrier();
 234                 rxcmp[3] = _mm_load_si128((void *)&cp_desc_ring[cons + 6]);
 235
 236                 rxcmp1[2] = _mm_load_si128((void *)&cp_desc_ring[cons + 5]);
 237                 rte_compiler_barrier();
 238                 rxcmp[2] = _mm_load_si128((void *)&cp_desc_ring[cons + 4]);
 239
 240                 tmp1 = _mm_unpackhi_epi32(rxcmp1[2], rxcmp1[3]);
 241
 242                 rxcmp1[1] = _mm_load_si128((void *)&cp_desc_ring[cons + 3]);
 243                 rte_compiler_barrier();
 244                 rxcmp[1] = _mm_load_si128((void *)&cp_desc_ring[cons + 2]);
 245
 246                 rxcmp1[0] = _mm_load_si128((void *)&cp_desc_ring[cons + 1]);
 247                 rte_compiler_barrier();
 248                 rxcmp[0] = _mm_load_si128((void *)&cp_desc_ring[cons + 0]);
 249
 250                 tmp0 = _mm_unpackhi_epi32(rxcmp1[0], rxcmp1[1]);
 251
 252                 /* Isolate descriptor valid flags. */
 253                 info3_v = _mm_and_si128(_mm_unpacklo_epi64(tmp0, tmp1),
 254                                         info3_v_mask);
 255                 info3_v = _mm_xor_si128(info3_v, valid_target);
 256
 257                 /*
 258                  * Pack the 128-bit array of valid descriptor flags into 64
 259                  * bits and count the number of set bits in order to determine
 260                  * the number of valid descriptors.
 261                  */
 262                 valid = _mm_cvtsi128_si64(_mm_packs_epi32(info3_v, info3_v));
 263                 num_valid = __builtin_popcountll(valid & desc_valid_mask);
 264
 265                 if (num_valid == 0)
 266                         break;
 267
 268                 descs_to_mbufs(rxcmp, rxcmp1, mbuf_init, &rx_pkts[nb_rx_pkts],
 269                                rxr);
 270                 nb_rx_pkts += num_valid;
 271
 272                 if (num_valid < RTE_BNXT_DESCS_PER_LOOP)
 273                         break;
 274         }
 275
 276         if (nb_rx_pkts) {
 277                 rxr->rx_raw_prod = RING_ADV(rxr->rx_raw_prod, nb_rx_pkts);
 278
 279                 rxq->rxrearm_nb += nb_rx_pkts;
 280                 cpr->cp_raw_cons += 2 * nb_rx_pkts;
 281                 cpr->valid =
 282                         !!(cpr->cp_raw_cons & cpr->cp_ring_struct->ring_size);
 283                 bnxt_db_cq(cpr);
 284         }
 285
 286         return nb_rx_pkts;
 287 }
 288
 289 static void
 290 bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
 291 {
 292         struct bnxt_cp_ring_info *cpr = txq->cp_ring;
 293         uint32_t raw_cons = cpr->cp_raw_cons;
 294         uint32_t cons;
 295         uint32_t nb_tx_pkts = 0;
 296         struct tx_cmpl *txcmp;
 297         struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
 298         struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
 299         uint32_t ring_mask = cp_ring_struct->ring_mask;
 300
 301         do {
 302                 cons = RING_CMPL(ring_mask, raw_cons);
 303                 txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
 304
 305                 if (!CMP_VALID(txcmp, raw_cons, cp_ring_struct))
 306                         break;
 307
 308                 if (likely(CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2))
 309                         nb_tx_pkts += txcmp->opaque;
 310                 else
 311                         RTE_LOG_DP(ERR, PMD,
 312                                    "Unhandled CMP type %02x\n",
 313                                    CMP_TYPE(txcmp));
 314                 raw_cons = NEXT_RAW_CMP(raw_cons);
 315         } while (nb_tx_pkts < ring_mask);
 316
 317         cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);
 318         if (nb_tx_pkts) {
 319                 if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
 320                         bnxt_tx_cmp_vec_fast(txq, nb_tx_pkts);
 321                 else
 322                         bnxt_tx_cmp_vec(txq, nb_tx_pkts);
 323                 cpr->cp_raw_cons = raw_cons;
 324                 bnxt_db_cq(cpr);
 325         }
 326 }
 327
 328 static inline void
 329 bnxt_xmit_one(struct rte_mbuf *mbuf, struct tx_bd_long *txbd,
 330               struct rte_mbuf **tx_buf)
 331 {
 332         __m128i desc;
 333
 334         *tx_buf = mbuf;
 335
 336         desc = _mm_set_epi64x(mbuf->buf_iova + mbuf->data_off,
 337                               bnxt_xmit_flags_len(mbuf->data_len,
 338                                                   TX_BD_FLAGS_NOCMPL));
 339         desc = _mm_blend_epi16(desc, _mm_set_epi16(0, 0, 0, 0, 0, 0,
 340                                                    mbuf->data_len, 0), 0x02);
 341         _mm_store_si128((void *)txbd, desc);
 342 }
 343
 344 static uint16_t
 345 bnxt_xmit_fixed_burst_vec(struct bnxt_tx_queue *txq, struct rte_mbuf **tx_pkts,
 346                           uint16_t nb_pkts)
 347 {
 348         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 349         uint16_t tx_prod, tx_raw_prod = txr->tx_raw_prod;
 350         struct tx_bd_long *txbd;
 351         struct rte_mbuf **tx_buf;
 352         uint16_t to_send;
 353
 354         tx_prod = RING_IDX(txr->tx_ring_struct, tx_raw_prod);
 355         txbd = &txr->tx_desc_ring[tx_prod];
 356         tx_buf = &txr->tx_buf_ring[tx_prod];
 357
 358         /* Prefetch next transmit buffer descriptors. */
 359         rte_prefetch0(txbd);
 360         rte_prefetch0(txbd + 3);
 361
 362         nb_pkts = RTE_MIN(nb_pkts, bnxt_tx_avail(txq));
 363
 364         if (unlikely(nb_pkts == 0))
 365                 return 0;
 366
 367         /* Handle TX burst request */
 368         to_send = nb_pkts;
 369         while (to_send >= RTE_BNXT_DESCS_PER_LOOP) {
 370                 /* Prefetch next transmit buffer descriptors. */
 371                 rte_prefetch0(txbd + 4);
 372                 rte_prefetch0(txbd + 7);
 373
 374                 bnxt_xmit_one(tx_pkts[0], txbd++, tx_buf++);
 375                 bnxt_xmit_one(tx_pkts[1], txbd++, tx_buf++);
 376                 bnxt_xmit_one(tx_pkts[2], txbd++, tx_buf++);
 377                 bnxt_xmit_one(tx_pkts[3], txbd++, tx_buf++);
 378
 379                 to_send -= RTE_BNXT_DESCS_PER_LOOP;
 380                 tx_pkts += RTE_BNXT_DESCS_PER_LOOP;
 381         }
 382
 383         while (to_send) {
 384                 bnxt_xmit_one(tx_pkts[0], txbd++, tx_buf++);
 385                 to_send--;
 386                 tx_pkts++;
 387         }
 388
 389         /* Request a completion for the final packet of burst. */
 390         rte_compiler_barrier();
 391         txbd[-1].opaque = nb_pkts;
 392         txbd[-1].flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
 393
 394         tx_raw_prod += nb_pkts;
 395         bnxt_db_write(&txr->tx_db, tx_raw_prod);
 396
 397         txr->tx_raw_prod = tx_raw_prod;
 398
 399         return nb_pkts;
 400 }
 401
 402 uint16_t
 403 bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 404                    uint16_t nb_pkts)
 405 {
 406         int nb_sent = 0;
 407         struct bnxt_tx_queue *txq = tx_queue;
 408         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 409         uint16_t ring_size = txr->tx_ring_struct->ring_size;
 410
 411         /* Tx queue was stopped; wait for it to be restarted */
 412         if (unlikely(!txq->tx_started)) {
 413                 PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
 414                 return 0;
 415         }
 416
 417         /* Handle TX completions */
 418         if (bnxt_tx_bds_in_hw(txq) >= txq->tx_free_thresh)
 419                 bnxt_handle_tx_cp_vec(txq);
 420
 421         while (nb_pkts) {
 422                 uint16_t ret, num;
 423
 424                 /*
 425                  * Ensure that no more than RTE_BNXT_MAX_TX_BURST packets
 426                  * are transmitted before the next completion.
 427                  */
 428                 num = RTE_MIN(nb_pkts, RTE_BNXT_MAX_TX_BURST);
 429
 430                 /*
 431                  * Ensure that a ring wrap does not occur within a call to
 432                  * bnxt_xmit_fixed_burst_vec().
 433                  */
 434                 num = RTE_MIN(num, ring_size -
 435                                    (txr->tx_raw_prod & (ring_size - 1)));
 436                 ret = bnxt_xmit_fixed_burst_vec(txq, &tx_pkts[nb_sent], num);
 437                 nb_sent += ret;
 438                 nb_pkts -= ret;
 439                 if (ret < num)
 440                         break;
 441         }
 442
 443         return nb_sent;
 444 }
 445
 446 int __rte_cold
 447 bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq)
 448 {
 449         return bnxt_rxq_vec_setup_common(rxq);
 450 }