drivers/net/bnxt/bnxt_rxtx_vec_sse.c

   1 /* SPDX-License-Identifier: BSD-3-Clause */
   2 /* Copyright(c) 2019-2020 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                 bnxt_ol_flags_table[_mm_extract_epi32((ol_index), (pi))];      \
  31                                                                                \
  32         tmp = _mm_extract_epi32((errors), (pi));                               \
  33         if (tmp)                                                               \
  34                 of |= bnxt_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 {
  59         const __m128i shuf_msk =
  60                 _mm_set_epi8(15, 14, 13, 12,          /* rss */
  61                              0xFF, 0xFF,              /* vlan_tci (zeroes) */
  62                              3, 2,                    /* data_len */
  63                              0xFF, 0xFF, 3, 2,        /* pkt_len */
  64                              0xFF, 0xFF, 0xFF, 0xFF); /* pkt_type (zeroes) */
  65         const __m128i flags_type_mask =
  66                 _mm_set1_epi32(RX_PKT_CMPL_FLAGS_ITYPE_MASK);
  67         const __m128i flags2_mask1 =
  68                 _mm_set1_epi32(RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN |
  69                                RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC);
  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;
  76         uint32_t ol_flags;
  77
  78         /* Compute packet type table indexes for four packets */
  79         t0 = _mm_unpacklo_epi32(mm_rxcmp[0], mm_rxcmp[1]);
  80         t1 = _mm_unpacklo_epi32(mm_rxcmp[2], mm_rxcmp[3]);
  81         flags_type = _mm_unpacklo_epi64(t0, t1);
  82         ptype_idx =
  83                 _mm_srli_epi32(_mm_and_si128(flags_type, flags_type_mask), 9);
  84
  85         t0 = _mm_unpacklo_epi32(mm_rxcmp1[0], mm_rxcmp1[1]);
  86         t1 = _mm_unpacklo_epi32(mm_rxcmp1[2], mm_rxcmp1[3]);
  87         flags2 = _mm_unpacklo_epi64(t0, t1);
  88
  89         ptype_idx = _mm_or_si128(ptype_idx,
  90                         _mm_srli_epi32(_mm_and_si128(flags2, flags2_mask1), 2));
  91         ptype_idx = _mm_or_si128(ptype_idx,
  92                         _mm_srli_epi32(_mm_and_si128(flags2, flags2_mask2), 7));
  93
  94         /* Extract RSS valid flags for four packets. */
  95         rss_flags = _mm_srli_epi32(_mm_and_si128(flags_type, rss_mask), 9);
  96
  97         /* Extract errors_v2 fields for four packets. */
  98         t0 = _mm_unpackhi_epi32(mm_rxcmp1[0], mm_rxcmp1[1]);
  99         t1 = _mm_unpackhi_epi32(mm_rxcmp1[2], mm_rxcmp1[3]);
 100
 101         /* Compute ol_flags and checksum error indexes for four packets. */
 102         flags2 = _mm_and_si128(flags2, _mm_set1_epi32(0x1F));
 103
 104         errors = _mm_srli_epi32(_mm_unpacklo_epi64(t0, t1), 4);
 105         errors = _mm_and_si128(errors, _mm_set1_epi32(0xF));
 106         errors = _mm_and_si128(errors, flags2);
 107
 108         index = _mm_andnot_si128(errors, flags2);
 109
 110         /* Update mbuf rearm_data for four packets. */
 111         GET_OL_FLAGS(rss_flags, index, errors, 0, ol_flags);
 112         _mm_store_si128((void *)&mbuf[0]->rearm_data,
 113                         _mm_or_si128(mbuf_init, _mm_set_epi64x(ol_flags, 0)));
 114
 115         GET_OL_FLAGS(rss_flags, index, errors, 1, ol_flags);
 116         _mm_store_si128((void *)&mbuf[1]->rearm_data,
 117                         _mm_or_si128(mbuf_init, _mm_set_epi64x(ol_flags, 0)));
 118
 119         GET_OL_FLAGS(rss_flags, index, errors, 2, ol_flags);
 120         _mm_store_si128((void *)&mbuf[2]->rearm_data,
 121                         _mm_or_si128(mbuf_init, _mm_set_epi64x(ol_flags, 0)));
 122
 123         GET_OL_FLAGS(rss_flags, index, errors, 3, ol_flags);
 124         _mm_store_si128((void *)&mbuf[3]->rearm_data,
 125                         _mm_or_si128(mbuf_init, _mm_set_epi64x(ol_flags, 0)));
 126
 127         /* Update mbuf rx_descriptor_fields1 for four packes. */
 128         GET_DESC_FIELDS(mm_rxcmp[0], mm_rxcmp1[0], shuf_msk, ptype_idx, 0, t0);
 129         _mm_store_si128((void *)&mbuf[0]->rx_descriptor_fields1, t0);
 130
 131         GET_DESC_FIELDS(mm_rxcmp[1], mm_rxcmp1[1], shuf_msk, ptype_idx, 1, t0);
 132         _mm_store_si128((void *)&mbuf[1]->rx_descriptor_fields1, t0);
 133
 134         GET_DESC_FIELDS(mm_rxcmp[2], mm_rxcmp1[2], shuf_msk, ptype_idx, 2, t0);
 135         _mm_store_si128((void *)&mbuf[2]->rx_descriptor_fields1, t0);
 136
 137         GET_DESC_FIELDS(mm_rxcmp[3], mm_rxcmp1[3], shuf_msk, ptype_idx, 3, t0);
 138         _mm_store_si128((void *)&mbuf[3]->rx_descriptor_fields1, t0);
 139 }
 140
 141 uint16_t
 142 bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 143                    uint16_t nb_pkts)
 144 {
 145         struct bnxt_rx_queue *rxq = rx_queue;
 146         const __m128i mbuf_init = _mm_set_epi64x(0, rxq->mbuf_initializer);
 147         struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
 148         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
 149         uint16_t cp_ring_size = cpr->cp_ring_struct->ring_size;
 150         uint16_t rx_ring_size = rxr->rx_ring_struct->ring_size;
 151         struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
 152         uint64_t valid, desc_valid_mask = ~0ULL;
 153         const __m128i info3_v_mask = _mm_set1_epi32(CMPL_BASE_V);
 154         uint32_t raw_cons = cpr->cp_raw_cons;
 155         uint32_t cons, mbcons;
 156         int nb_rx_pkts = 0;
 157         const __m128i valid_target =
 158                 _mm_set1_epi32(!!(raw_cons & cp_ring_size));
 159         int i;
 160
 161         /* If Rx Q was stopped return */
 162         if (unlikely(!rxq->rx_started))
 163                 return 0;
 164
 165         if (rxq->rxrearm_nb >= rxq->rx_free_thresh)
 166                 bnxt_rxq_rearm(rxq, rxr);
 167
 168         /* Return no more than RTE_BNXT_MAX_RX_BURST per call. */
 169         nb_pkts = RTE_MIN(nb_pkts, RTE_BNXT_MAX_RX_BURST);
 170
 171         cons = raw_cons & (cp_ring_size - 1);
 172         mbcons = (raw_cons / 2) & (rx_ring_size - 1);
 173
 174         /* Prefetch first four descriptor pairs. */
 175         rte_prefetch0(&cp_desc_ring[cons]);
 176         rte_prefetch0(&cp_desc_ring[cons + 4]);
 177
 178         /* Ensure that we do not go past the ends of the rings. */
 179         nb_pkts = RTE_MIN(nb_pkts, RTE_MIN(rx_ring_size - mbcons,
 180                                            (cp_ring_size - cons) / 2));
 181         /*
 182          * If we are at the end of the ring, ensure that descriptors after the
 183          * last valid entry are not treated as valid. Otherwise, force the
 184          * maximum number of packets to receive to be a multiple of the per-
 185          * loop count.
 186          */
 187         if (nb_pkts < RTE_BNXT_DESCS_PER_LOOP)
 188                 desc_valid_mask >>= 16 * (RTE_BNXT_DESCS_PER_LOOP - nb_pkts);
 189         else
 190                 nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_BNXT_DESCS_PER_LOOP);
 191
 192         /* Handle RX burst request */
 193         for (i = 0; i < nb_pkts; i += RTE_BNXT_DESCS_PER_LOOP,
 194                                   cons += RTE_BNXT_DESCS_PER_LOOP * 2,
 195                                   mbcons += RTE_BNXT_DESCS_PER_LOOP) {
 196                 __m128i rxcmp1[RTE_BNXT_DESCS_PER_LOOP];
 197                 __m128i rxcmp[RTE_BNXT_DESCS_PER_LOOP];
 198                 __m128i tmp0, tmp1, info3_v;
 199                 uint32_t num_valid;
 200
 201                 /* Copy four mbuf pointers to output array. */
 202                 tmp0 = _mm_loadu_si128((void *)&rxr->rx_buf_ring[mbcons]);
 203 #ifdef RTE_ARCH_X86_64
 204                 tmp1 = _mm_loadu_si128((void *)&rxr->rx_buf_ring[mbcons + 2]);
 205 #endif
 206                 _mm_storeu_si128((void *)&rx_pkts[i], tmp0);
 207 #ifdef RTE_ARCH_X86_64
 208                 _mm_storeu_si128((void *)&rx_pkts[i + 2], tmp1);
 209 #endif
 210
 211                 /* Prefetch four descriptor pairs for next iteration. */
 212                 if (i + RTE_BNXT_DESCS_PER_LOOP < nb_pkts) {
 213                         rte_prefetch0(&cp_desc_ring[cons + 8]);
 214                         rte_prefetch0(&cp_desc_ring[cons + 12]);
 215                 }
 216
 217                 /*
 218                  * Load the four current descriptors into SSE registers in
 219                  * reverse order to ensure consistent state.
 220                  */
 221                 rxcmp1[3] = _mm_load_si128((void *)&cp_desc_ring[cons + 7]);
 222                 rte_compiler_barrier();
 223                 rxcmp[3] = _mm_load_si128((void *)&cp_desc_ring[cons + 6]);
 224
 225                 rxcmp1[2] = _mm_load_si128((void *)&cp_desc_ring[cons + 5]);
 226                 rte_compiler_barrier();
 227                 rxcmp[2] = _mm_load_si128((void *)&cp_desc_ring[cons + 4]);
 228
 229                 tmp1 = _mm_unpackhi_epi32(rxcmp1[2], rxcmp1[3]);
 230
 231                 rxcmp1[1] = _mm_load_si128((void *)&cp_desc_ring[cons + 3]);
 232                 rte_compiler_barrier();
 233                 rxcmp[1] = _mm_load_si128((void *)&cp_desc_ring[cons + 2]);
 234
 235                 rxcmp1[0] = _mm_load_si128((void *)&cp_desc_ring[cons + 1]);
 236                 rte_compiler_barrier();
 237                 rxcmp[0] = _mm_load_si128((void *)&cp_desc_ring[cons + 0]);
 238
 239                 tmp0 = _mm_unpackhi_epi32(rxcmp1[0], rxcmp1[1]);
 240
 241                 /* Isolate descriptor valid flags. */
 242                 info3_v = _mm_and_si128(_mm_unpacklo_epi64(tmp0, tmp1),
 243                                         info3_v_mask);
 244                 info3_v = _mm_xor_si128(info3_v, valid_target);
 245
 246                 /*
 247                  * Pack the 128-bit array of valid descriptor flags into 64
 248                  * bits and count the number of set bits in order to determine
 249                  * the number of valid descriptors.
 250                  */
 251                 valid = _mm_cvtsi128_si64(_mm_packs_epi32(info3_v, info3_v));
 252                 num_valid = __builtin_popcountll(valid & desc_valid_mask);
 253
 254                 switch (num_valid) {
 255                 case 4:
 256                         rxr->rx_buf_ring[mbcons + 3] = NULL;
 257                         /* FALLTHROUGH */
 258                 case 3:
 259                         rxr->rx_buf_ring[mbcons + 2] = NULL;
 260                         /* FALLTHROUGH */
 261                 case 2:
 262                         rxr->rx_buf_ring[mbcons + 1] = NULL;
 263                         /* FALLTHROUGH */
 264                 case 1:
 265                         rxr->rx_buf_ring[mbcons + 0] = NULL;
 266                         break;
 267                 case 0:
 268                         goto out;
 269                 }
 270
 271                 descs_to_mbufs(rxcmp, rxcmp1, mbuf_init, &rx_pkts[nb_rx_pkts]);
 272                 nb_rx_pkts += num_valid;
 273
 274                 if (num_valid < RTE_BNXT_DESCS_PER_LOOP)
 275                         break;
 276         }
 277
 278 out:
 279         if (nb_rx_pkts) {
 280                 rxr->rx_prod =
 281                         RING_ADV(rxr->rx_ring_struct, rxr->rx_prod, nb_rx_pkts);
 282
 283                 rxq->rxrearm_nb += nb_rx_pkts;
 284                 cpr->cp_raw_cons += 2 * nb_rx_pkts;
 285                 cpr->valid =
 286                         !!(cpr->cp_raw_cons & cpr->cp_ring_struct->ring_size);
 287                 bnxt_db_cq(cpr);
 288         }
 289
 290         return nb_rx_pkts;
 291 }
 292
 293 static void
 294 bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
 295 {
 296         struct bnxt_cp_ring_info *cpr = txq->cp_ring;
 297         uint32_t raw_cons = cpr->cp_raw_cons;
 298         uint32_t cons;
 299         uint32_t nb_tx_pkts = 0;
 300         struct tx_cmpl *txcmp;
 301         struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
 302         struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
 303         uint32_t ring_mask = cp_ring_struct->ring_mask;
 304
 305         do {
 306                 cons = RING_CMPL(ring_mask, raw_cons);
 307                 txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
 308
 309                 if (!CMP_VALID(txcmp, raw_cons, cp_ring_struct))
 310                         break;
 311
 312                 if (likely(CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2))
 313                         nb_tx_pkts += txcmp->opaque;
 314                 else
 315                         RTE_LOG_DP(ERR, PMD,
 316                                    "Unhandled CMP type %02x\n",
 317                                    CMP_TYPE(txcmp));
 318                 raw_cons = NEXT_RAW_CMP(raw_cons);
 319         } while (nb_tx_pkts < ring_mask);
 320
 321         cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);
 322         if (nb_tx_pkts) {
 323                 if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
 324                         bnxt_tx_cmp_vec_fast(txq, nb_tx_pkts);
 325                 else
 326                         bnxt_tx_cmp_vec(txq, nb_tx_pkts);
 327                 cpr->cp_raw_cons = raw_cons;
 328                 bnxt_db_cq(cpr);
 329         }
 330 }
 331
 332 static inline void
 333 bnxt_xmit_one(struct rte_mbuf *mbuf, struct tx_bd_long *txbd,
 334               struct bnxt_sw_tx_bd *tx_buf)
 335 {
 336         __m128i desc;
 337
 338         tx_buf->mbuf = mbuf;
 339         tx_buf->nr_bds = 1;
 340
 341         desc = _mm_set_epi64x(mbuf->buf_iova + mbuf->data_off,
 342                               bnxt_xmit_flags_len(mbuf->data_len,
 343                                                   TX_BD_FLAGS_NOCMPL));
 344         desc = _mm_blend_epi16(desc, _mm_set_epi16(0, 0, 0, 0, 0, 0,
 345                                                    mbuf->data_len, 0), 0x02);
 346         _mm_store_si128((void *)txbd, desc);
 347 }
 348
 349 static uint16_t
 350 bnxt_xmit_fixed_burst_vec(struct bnxt_tx_queue *txq, struct rte_mbuf **tx_pkts,
 351                           uint16_t nb_pkts)
 352 {
 353         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 354         uint16_t tx_prod = txr->tx_prod;
 355         struct tx_bd_long *txbd;
 356         struct bnxt_sw_tx_bd *tx_buf;
 357         uint16_t to_send;
 358
 359         txbd = &txr->tx_desc_ring[tx_prod];
 360         tx_buf = &txr->tx_buf_ring[tx_prod];
 361
 362         /* Prefetch next transmit buffer descriptors. */
 363         rte_prefetch0(txbd);
 364         rte_prefetch0(txbd + 3);
 365
 366         nb_pkts = RTE_MIN(nb_pkts, bnxt_tx_avail(txq));
 367
 368         if (unlikely(nb_pkts == 0))
 369                 return 0;
 370
 371         /* Handle TX burst request */
 372         to_send = nb_pkts;
 373         while (to_send >= RTE_BNXT_DESCS_PER_LOOP) {
 374                 /* Prefetch next transmit buffer descriptors. */
 375                 rte_prefetch0(txbd + 4);
 376                 rte_prefetch0(txbd + 7);
 377
 378                 bnxt_xmit_one(tx_pkts[0], txbd++, tx_buf++);
 379                 bnxt_xmit_one(tx_pkts[1], txbd++, tx_buf++);
 380                 bnxt_xmit_one(tx_pkts[2], txbd++, tx_buf++);
 381                 bnxt_xmit_one(tx_pkts[3], txbd++, tx_buf++);
 382
 383                 to_send -= RTE_BNXT_DESCS_PER_LOOP;
 384                 tx_pkts += RTE_BNXT_DESCS_PER_LOOP;
 385         }
 386
 387         while (to_send) {
 388                 bnxt_xmit_one(tx_pkts[0], txbd++, tx_buf++);
 389                 to_send--;
 390                 tx_pkts++;
 391         }
 392
 393         /* Request a completion for the final packet of burst. */
 394         rte_compiler_barrier();
 395         txbd[-1].opaque = nb_pkts;
 396         txbd[-1].flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
 397
 398         tx_prod = RING_ADV(txr->tx_ring_struct, tx_prod, nb_pkts);
 399         bnxt_db_write(&txr->tx_db, tx_prod);
 400
 401         txr->tx_prod = tx_prod;
 402
 403         return nb_pkts;
 404 }
 405
 406 uint16_t
 407 bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 408                    uint16_t nb_pkts)
 409 {
 410         int nb_sent = 0;
 411         struct bnxt_tx_queue *txq = tx_queue;
 412         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 413         uint16_t ring_size = txr->tx_ring_struct->ring_size;
 414
 415         /* Tx queue was stopped; wait for it to be restarted */
 416         if (unlikely(!txq->tx_started)) {
 417                 PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
 418                 return 0;
 419         }
 420
 421         /* Handle TX completions */
 422         if (bnxt_tx_bds_in_hw(txq) >= txq->tx_free_thresh)
 423                 bnxt_handle_tx_cp_vec(txq);
 424
 425         while (nb_pkts) {
 426                 uint16_t ret, num;
 427
 428                 /*
 429                  * Ensure that no more than RTE_BNXT_MAX_TX_BURST packets
 430                  * are transmitted before the next completion.
 431                  */
 432                 num = RTE_MIN(nb_pkts, RTE_BNXT_MAX_TX_BURST);
 433
 434                 /*
 435                  * Ensure that a ring wrap does not occur within a call to
 436                  * bnxt_xmit_fixed_burst_vec().
 437                  */
 438                 num = RTE_MIN(num,
 439                               ring_size - (txr->tx_prod & (ring_size - 1)));
 440                 ret = bnxt_xmit_fixed_burst_vec(txq, &tx_pkts[nb_sent], num);
 441                 nb_sent += ret;
 442                 nb_pkts -= ret;
 443                 if (ret < num)
 444                         break;
 445         }
 446
 447         return nb_sent;
 448 }
 449
 450 int __rte_cold
 451 bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq)
 452 {
 453         return bnxt_rxq_vec_setup_common(rxq);
 454 }