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_rxtx_vec_common.h"
  21
  22 #include "bnxt_txq.h"
  23 #include "bnxt_txr.h"
  24
  25 /*
  26  * RX Ring handling
  27  */
  28
  29 static inline void
  30 bnxt_rxq_rearm(struct bnxt_rx_queue *rxq, struct bnxt_rx_ring_info *rxr)
  31 {
  32         struct rx_prod_pkt_bd *rxbds = &rxr->rx_desc_ring[rxq->rxrearm_start];
  33         struct rte_mbuf **rx_bufs = &rxr->rx_buf_ring[rxq->rxrearm_start];
  34         struct rte_mbuf *mb0, *mb1;
  35         int nb, i;
  36
  37         const __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM, 0);
  38         const __m128i addrmask = _mm_set_epi64x(UINT64_MAX, 0);
  39
  40         /*
  41          * Number of mbufs to allocate must be a multiple of two. The
  42          * allocation must not go past the end of the ring.
  43          */
  44         nb = RTE_MIN(rxq->rxrearm_nb & ~0x1,
  45                      rxq->nb_rx_desc - rxq->rxrearm_start);
  46
  47         /* Allocate new mbufs into the software ring */
  48         if (rte_mempool_get_bulk(rxq->mb_pool, (void *)rx_bufs, nb) < 0) {
  49                 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed += nb;
  50
  51                 return;
  52         }
  53
  54         /* Initialize the mbufs in vector, process 2 mbufs in one loop */
  55         for (i = 0; i < nb; i += 2, rx_bufs += 2) {
  56                 __m128i buf_addr0, buf_addr1;
  57                 __m128i rxbd0, rxbd1;
  58
  59                 mb0 = rx_bufs[0];
  60                 mb1 = rx_bufs[1];
  61
  62                 /* Load address fields from both mbufs */
  63                 buf_addr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
  64                 buf_addr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
  65
  66                 /* Load both rx descriptors (preserving some existing fields) */
  67                 rxbd0 = _mm_loadu_si128((__m128i *)(rxbds + 0));
  68                 rxbd1 = _mm_loadu_si128((__m128i *)(rxbds + 1));
  69
  70                 /* Add default offset to buffer address. */
  71                 buf_addr0 = _mm_add_epi64(buf_addr0, hdr_room);
  72                 buf_addr1 = _mm_add_epi64(buf_addr1, hdr_room);
  73
  74                 /* Clear all fields except address. */
  75                 buf_addr0 =  _mm_and_si128(buf_addr0, addrmask);
  76                 buf_addr1 =  _mm_and_si128(buf_addr1, addrmask);
  77
  78                 /* Clear address field in descriptor. */
  79                 rxbd0 = _mm_andnot_si128(addrmask, rxbd0);
  80                 rxbd1 = _mm_andnot_si128(addrmask, rxbd1);
  81
  82                 /* Set address field in descriptor. */
  83                 rxbd0 = _mm_add_epi64(rxbd0, buf_addr0);
  84                 rxbd1 = _mm_add_epi64(rxbd1, buf_addr1);
  85
  86                 /* Store descriptors to memory. */
  87                 _mm_store_si128((__m128i *)(rxbds++), rxbd0);
  88                 _mm_store_si128((__m128i *)(rxbds++), rxbd1);
  89         }
  90
  91         rxq->rxrearm_start += nb;
  92         bnxt_db_write(&rxr->rx_db, rxq->rxrearm_start - 1);
  93         if (rxq->rxrearm_start >= rxq->nb_rx_desc)
  94                 rxq->rxrearm_start = 0;
  95
  96         rxq->rxrearm_nb -= nb;
  97 }
  98
  99 static __m128i
 100 bnxt_parse_pkt_type(__m128i mm_rxcmp, __m128i mm_rxcmp1)
 101 {
 102         uint32_t flags_type, flags2;
 103         uint8_t index;
 104
 105         flags_type = _mm_extract_epi16(mm_rxcmp, 0);
 106         flags2 = _mm_extract_epi32(mm_rxcmp1, 0);
 107
 108         /*
 109          * Index format:
 110          *     bit 0: RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC
 111          *     bit 1: RX_CMPL_FLAGS2_IP_TYPE
 112          *     bit 2: RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN
 113          *     bits 3-6: RX_PKT_CMPL_FLAGS_ITYPE
 114          */
 115         index = ((flags_type & RX_PKT_CMPL_FLAGS_ITYPE_MASK) >> 9) |
 116                 ((flags2 & (RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN |
 117                            RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC)) >> 2) |
 118                 ((flags2 & RX_PKT_CMPL_FLAGS2_IP_TYPE) >> 7);
 119
 120         return _mm_set_epi32(0, 0, 0, bnxt_ptype_table[index]);
 121 }
 122
 123 static void
 124 bnxt_parse_csum(struct rte_mbuf *mbuf, struct rx_pkt_cmpl_hi *rxcmp1)
 125 {
 126         uint32_t flags;
 127
 128         flags = flags2_0xf(rxcmp1);
 129         /* IP Checksum */
 130         if (likely(IS_IP_NONTUNNEL_PKT(flags))) {
 131                 if (unlikely(RX_CMP_IP_CS_ERROR(rxcmp1)))
 132                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
 133                 else
 134                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
 135         } else if (IS_IP_TUNNEL_PKT(flags)) {
 136                 if (unlikely(RX_CMP_IP_OUTER_CS_ERROR(rxcmp1) ||
 137                              RX_CMP_IP_CS_ERROR(rxcmp1)))
 138                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
 139                 else
 140                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
 141         } else if (unlikely(RX_CMP_IP_CS_UNKNOWN(rxcmp1))) {
 142                 mbuf->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 143         }
 144
 145         /* L4 Checksum */
 146         if (likely(IS_L4_NONTUNNEL_PKT(flags))) {
 147                 if (unlikely(RX_CMP_L4_INNER_CS_ERR2(rxcmp1)))
 148                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
 149                 else
 150                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 151         } else if (IS_L4_TUNNEL_PKT(flags)) {
 152                 if (unlikely(RX_CMP_L4_INNER_CS_ERR2(rxcmp1)))
 153                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
 154                 else
 155                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 156                 if (unlikely(RX_CMP_L4_OUTER_CS_ERR2(rxcmp1))) {
 157                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_BAD;
 158                 } else if (unlikely(IS_L4_TUNNEL_PKT_ONLY_INNER_L4_CS
 159                                     (flags))) {
 160                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_UNKNOWN;
 161                 } else {
 162                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_GOOD;
 163                 }
 164         } else if (unlikely(RX_CMP_L4_CS_UNKNOWN(rxcmp1))) {
 165                 mbuf->ol_flags |= PKT_RX_L4_CKSUM_UNKNOWN;
 166         }
 167 }
 168
 169 uint16_t
 170 bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 171                    uint16_t nb_pkts)
 172 {
 173         struct bnxt_rx_queue *rxq = rx_queue;
 174         struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
 175         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
 176         uint32_t raw_cons = cpr->cp_raw_cons;
 177         uint32_t cons;
 178         int nb_rx_pkts = 0;
 179         struct rx_pkt_cmpl *rxcmp;
 180         const __m128i mbuf_init = _mm_set_epi64x(0, rxq->mbuf_initializer);
 181         const __m128i shuf_msk =
 182                 _mm_set_epi8(15, 14, 13, 12,          /* rss */
 183                              0xFF, 0xFF,              /* vlan_tci (zeroes) */
 184                              3, 2,                    /* data_len */
 185                              0xFF, 0xFF, 3, 2,        /* pkt_len */
 186                              0xFF, 0xFF, 0xFF, 0xFF); /* pkt_type (zeroes) */
 187         int i;
 188
 189         /* If Rx Q was stopped return */
 190         if (unlikely(!rxq->rx_started))
 191                 return 0;
 192
 193         if (rxq->rxrearm_nb >= rxq->rx_free_thresh)
 194                 bnxt_rxq_rearm(rxq, rxr);
 195
 196         /* Return no more than RTE_BNXT_MAX_RX_BURST per call. */
 197         nb_pkts = RTE_MIN(nb_pkts, RTE_BNXT_MAX_RX_BURST);
 198
 199         /*
 200          * Make nb_pkts an integer multiple of RTE_BNXT_DESCS_PER_LOOP.
 201          * nb_pkts < RTE_BNXT_DESCS_PER_LOOP, just return no packet
 202          */
 203         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_BNXT_DESCS_PER_LOOP);
 204         if (!nb_pkts)
 205                 return 0;
 206
 207         /* Handle RX burst request */
 208         for (i = 0; i < nb_pkts; i++) {
 209                 struct rx_pkt_cmpl_hi *rxcmp1;
 210                 struct rte_mbuf *mbuf;
 211                 __m128i mm_rxcmp, mm_rxcmp1, pkt_mb, ptype;
 212
 213                 cons = RING_CMP(cpr->cp_ring_struct, raw_cons);
 214
 215                 rxcmp = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[cons];
 216                 rxcmp1 = (struct rx_pkt_cmpl_hi *)&cpr->cp_desc_ring[cons + 1];
 217
 218                 if (!CMP_VALID(rxcmp1, raw_cons + 1, cpr->cp_ring_struct))
 219                         break;
 220
 221                 mm_rxcmp = _mm_load_si128((__m128i *)rxcmp);
 222                 mm_rxcmp1 = _mm_load_si128((__m128i *)rxcmp1);
 223
 224                 raw_cons += 2;
 225                 cons = rxcmp->opaque;
 226
 227                 mbuf = rxr->rx_buf_ring[cons];
 228                 rte_prefetch0(mbuf);
 229                 rxr->rx_buf_ring[cons] = NULL;
 230
 231                 /* Set constant fields from mbuf initializer. */
 232                 _mm_store_si128((__m128i *)&mbuf->rearm_data, mbuf_init);
 233
 234                 /* Set mbuf pkt_len, data_len, and rss_hash fields. */
 235                 pkt_mb = _mm_shuffle_epi8(mm_rxcmp, shuf_msk);
 236                 ptype = bnxt_parse_pkt_type(mm_rxcmp, mm_rxcmp1);
 237                 pkt_mb = _mm_blend_epi16(pkt_mb, ptype, 0x3);
 238
 239                 _mm_storeu_si128((void *)&mbuf->rx_descriptor_fields1, pkt_mb);
 240
 241                 rte_compiler_barrier();
 242
 243                 if (rxcmp->flags_type & RX_PKT_CMPL_FLAGS_RSS_VALID)
 244                         mbuf->ol_flags |= PKT_RX_RSS_HASH;
 245
 246                 if (rxcmp1->flags2 &
 247                     RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN) {
 248                         mbuf->vlan_tci = rxcmp1->metadata &
 249                                 (RX_PKT_CMPL_METADATA_VID_MASK |
 250                                 RX_PKT_CMPL_METADATA_DE |
 251                                 RX_PKT_CMPL_METADATA_PRI_MASK);
 252                         mbuf->ol_flags |=
 253                                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
 254                 }
 255
 256                 bnxt_parse_csum(mbuf, rxcmp1);
 257                 rx_pkts[nb_rx_pkts++] = mbuf;
 258         }
 259
 260         if (nb_rx_pkts) {
 261                 rxr->rx_prod =
 262                         RING_ADV(rxr->rx_ring_struct, rxr->rx_prod, nb_rx_pkts);
 263
 264                 rxq->rxrearm_nb += nb_rx_pkts;
 265                 cpr->cp_raw_cons = raw_cons;
 266                 cpr->valid =
 267                         !!(cpr->cp_raw_cons & cpr->cp_ring_struct->ring_size);
 268                 bnxt_db_cq(cpr);
 269         }
 270
 271         return nb_rx_pkts;
 272 }
 273
 274 static void
 275 bnxt_tx_cmp_vec(struct bnxt_tx_queue *txq, int nr_pkts)
 276 {
 277         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 278         struct rte_mbuf **free = txq->free;
 279         uint16_t cons = txr->tx_cons;
 280         unsigned int blk = 0;
 281
 282         while (nr_pkts--) {
 283                 struct bnxt_sw_tx_bd *tx_buf;
 284                 struct rte_mbuf *mbuf;
 285
 286                 tx_buf = &txr->tx_buf_ring[cons];
 287                 cons = RING_NEXT(txr->tx_ring_struct, cons);
 288                 mbuf = rte_pktmbuf_prefree_seg(tx_buf->mbuf);
 289                 if (unlikely(mbuf == NULL))
 290                         continue;
 291                 tx_buf->mbuf = NULL;
 292
 293                 if (blk && mbuf->pool != free[0]->pool) {
 294                         rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
 295                         blk = 0;
 296                 }
 297                 free[blk++] = mbuf;
 298         }
 299         if (blk)
 300                 rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
 301
 302         txr->tx_cons = cons;
 303 }
 304
 305 static void
 306 bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
 307 {
 308         struct bnxt_cp_ring_info *cpr = txq->cp_ring;
 309         uint32_t raw_cons = cpr->cp_raw_cons;
 310         uint32_t cons;
 311         uint32_t nb_tx_pkts = 0;
 312         struct tx_cmpl *txcmp;
 313         struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
 314         struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
 315         uint32_t ring_mask = cp_ring_struct->ring_mask;
 316
 317         do {
 318                 cons = RING_CMPL(ring_mask, raw_cons);
 319                 txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
 320
 321                 if (!CMP_VALID(txcmp, raw_cons, cp_ring_struct))
 322                         break;
 323
 324                 if (likely(CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2))
 325                         nb_tx_pkts += txcmp->opaque;
 326                 else
 327                         RTE_LOG_DP(ERR, PMD,
 328                                    "Unhandled CMP type %02x\n",
 329                                    CMP_TYPE(txcmp));
 330                 raw_cons = NEXT_RAW_CMP(raw_cons);
 331         } while (nb_tx_pkts < ring_mask);
 332
 333         cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);
 334         if (nb_tx_pkts) {
 335                 bnxt_tx_cmp_vec(txq, nb_tx_pkts);
 336                 cpr->cp_raw_cons = raw_cons;
 337                 bnxt_db_cq(cpr);
 338         }
 339 }
 340
 341 static uint16_t
 342 bnxt_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 343                           uint16_t nb_pkts)
 344 {
 345         struct bnxt_tx_queue *txq = tx_queue;
 346         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 347         uint16_t prod = txr->tx_prod;
 348         struct rte_mbuf *tx_mbuf;
 349         struct tx_bd_long *txbd = NULL;
 350         struct bnxt_sw_tx_bd *tx_buf;
 351         uint16_t to_send;
 352
 353         nb_pkts = RTE_MIN(nb_pkts, bnxt_tx_avail(txq));
 354
 355         if (unlikely(nb_pkts == 0))
 356                 return 0;
 357
 358         /* Handle TX burst request */
 359         to_send = nb_pkts;
 360         while (to_send) {
 361                 tx_mbuf = *tx_pkts++;
 362                 rte_prefetch0(tx_mbuf);
 363
 364                 tx_buf = &txr->tx_buf_ring[prod];
 365                 tx_buf->mbuf = tx_mbuf;
 366                 tx_buf->nr_bds = 1;
 367
 368                 txbd = &txr->tx_desc_ring[prod];
 369                 txbd->address = tx_mbuf->buf_iova + tx_mbuf->data_off;
 370                 txbd->len = tx_mbuf->data_len;
 371                 txbd->flags_type = bnxt_xmit_flags_len(tx_mbuf->data_len,
 372                                                        TX_BD_FLAGS_NOCMPL);
 373                 prod = RING_NEXT(txr->tx_ring_struct, prod);
 374                 to_send--;
 375         }
 376
 377         /* Request a completion for last packet in burst */
 378         if (txbd) {
 379                 txbd->opaque = nb_pkts;
 380                 txbd->flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
 381         }
 382
 383         rte_compiler_barrier();
 384         bnxt_db_write(&txr->tx_db, prod);
 385
 386         txr->tx_prod = prod;
 387
 388         return nb_pkts;
 389 }
 390
 391 uint16_t
 392 bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 393                    uint16_t nb_pkts)
 394 {
 395         int nb_sent = 0;
 396         struct bnxt_tx_queue *txq = tx_queue;
 397
 398         /* Tx queue was stopped; wait for it to be restarted */
 399         if (unlikely(!txq->tx_started)) {
 400                 PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
 401                 return 0;
 402         }
 403
 404         /* Handle TX completions */
 405         if (bnxt_tx_bds_in_hw(txq) >= txq->tx_free_thresh)
 406                 bnxt_handle_tx_cp_vec(txq);
 407
 408         while (nb_pkts) {
 409                 uint16_t ret, num;
 410
 411                 num = RTE_MIN(nb_pkts, RTE_BNXT_MAX_TX_BURST);
 412                 ret = bnxt_xmit_fixed_burst_vec(tx_queue,
 413                                                 &tx_pkts[nb_sent],
 414                                                 num);
 415                 nb_sent += ret;
 416                 nb_pkts -= ret;
 417                 if (ret < num)
 418                         break;
 419         }
 420
 421         return nb_sent;
 422 }
 423
 424 int __rte_cold
 425 bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq)
 426 {
 427         return bnxt_rxq_vec_setup_common(rxq);
 428 }