drivers/net/bnxt/bnxt_rxtx_vec_neon.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 #include "bnxt_rxr.h"
  17 #include "bnxt_rxq.h"
  18 #include "hsi_struct_def_dpdk.h"
  19 #include "bnxt_rxtx_vec_common.h"
  20
  21 #include "bnxt_txq.h"
  22 #include "bnxt_txr.h"
  23
  24 /*
  25  * RX Ring handling
  26  */
  27
  28 static inline void
  29 bnxt_rxq_rearm(struct bnxt_rx_queue *rxq, struct bnxt_rx_ring_info *rxr)
  30 {
  31         struct rx_prod_pkt_bd *rxbds = &rxr->rx_desc_ring[rxq->rxrearm_start];
  32         struct rte_mbuf **rx_bufs = &rxr->rx_buf_ring[rxq->rxrearm_start];
  33         struct rte_mbuf *mb0, *mb1;
  34         int i;
  35
  36         const uint64x2_t hdr_room = {0, RTE_PKTMBUF_HEADROOM};
  37         const uint64x2_t addrmask = {0, UINT64_MAX};
  38
  39         /* Pull RTE_BNXT_RXQ_REARM_THRESH more mbufs into the software ring */
  40         if (rte_mempool_get_bulk(rxq->mb_pool,
  41                                  (void *)rx_bufs,
  42                                  RTE_BNXT_RXQ_REARM_THRESH) < 0) {
  43                 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
  44                         RTE_BNXT_RXQ_REARM_THRESH;
  45
  46                 return;
  47         }
  48
  49         /* Initialize the mbufs in vector, process 2 mbufs in one loop */
  50         for (i = 0; i < RTE_BNXT_RXQ_REARM_THRESH; i += 2, rx_bufs += 2) {
  51                 uint64x2_t buf_addr0, buf_addr1;
  52                 uint64x2_t rxbd0, rxbd1;
  53
  54                 mb0 = rx_bufs[0];
  55                 mb1 = rx_bufs[1];
  56
  57                 /* Load address fields from both mbufs */
  58                 buf_addr0 = vld1q_u64((uint64_t *)&mb0->buf_addr);
  59                 buf_addr1 = vld1q_u64((uint64_t *)&mb1->buf_addr);
  60
  61                 /* Load both rx descriptors (preserving some existing fields) */
  62                 rxbd0 = vld1q_u64((uint64_t *)(rxbds + 0));
  63                 rxbd1 = vld1q_u64((uint64_t *)(rxbds + 1));
  64
  65                 /* Add default offset to buffer address. */
  66                 buf_addr0 = vaddq_u64(buf_addr0, hdr_room);
  67                 buf_addr1 = vaddq_u64(buf_addr1, hdr_room);
  68
  69                 /* Clear all fields except address. */
  70                 buf_addr0 =  vandq_u64(buf_addr0, addrmask);
  71                 buf_addr1 =  vandq_u64(buf_addr1, addrmask);
  72
  73                 /* Clear address field in descriptor. */
  74                 rxbd0 = vbicq_u64(rxbd0, addrmask);
  75                 rxbd1 = vbicq_u64(rxbd1, addrmask);
  76
  77                 /* Set address field in descriptor. */
  78                 rxbd0 = vaddq_u64(rxbd0, buf_addr0);
  79                 rxbd1 = vaddq_u64(rxbd1, buf_addr1);
  80
  81                 /* Store descriptors to memory. */
  82                 vst1q_u64((uint64_t *)(rxbds++), rxbd0);
  83                 vst1q_u64((uint64_t *)(rxbds++), rxbd1);
  84         }
  85
  86         rxq->rxrearm_start += RTE_BNXT_RXQ_REARM_THRESH;
  87         bnxt_db_write(&rxr->rx_db, rxq->rxrearm_start - 1);
  88         if (rxq->rxrearm_start >= rxq->nb_rx_desc)
  89                 rxq->rxrearm_start = 0;
  90
  91         rxq->rxrearm_nb -= RTE_BNXT_RXQ_REARM_THRESH;
  92 }
  93
  94 static uint32_t
  95 bnxt_parse_pkt_type(struct rx_pkt_cmpl *rxcmp, struct rx_pkt_cmpl_hi *rxcmp1)
  96 {
  97         uint32_t l3, pkt_type = 0;
  98         uint32_t t_ipcs = 0, ip6 = 0, vlan = 0;
  99         uint32_t flags_type;
 100
 101         vlan = !!(rxcmp1->flags2 &
 102                 rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN));
 103         pkt_type |= vlan ? RTE_PTYPE_L2_ETHER_VLAN : RTE_PTYPE_L2_ETHER;
 104
 105         t_ipcs = !!(rxcmp1->flags2 &
 106                 rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC));
 107         ip6 = !!(rxcmp1->flags2 &
 108                  rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_IP_TYPE));
 109
 110         flags_type = rxcmp->flags_type &
 111                 rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS_ITYPE_MASK);
 112
 113         if (!t_ipcs && !ip6)
 114                 l3 = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
 115         else if (!t_ipcs && ip6)
 116                 l3 = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
 117         else if (t_ipcs && !ip6)
 118                 l3 = RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN;
 119         else
 120                 l3 = RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN;
 121
 122         switch (flags_type) {
 123         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_ICMP):
 124                 if (!t_ipcs)
 125                         pkt_type |= l3 | RTE_PTYPE_L4_ICMP;
 126                 else
 127                         pkt_type |= l3 | RTE_PTYPE_INNER_L4_ICMP;
 128                 break;
 129
 130         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_TCP):
 131                 if (!t_ipcs)
 132                         pkt_type |= l3 | RTE_PTYPE_L4_TCP;
 133                 else
 134                         pkt_type |= l3 | RTE_PTYPE_INNER_L4_TCP;
 135                 break;
 136
 137         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_UDP):
 138                 if (!t_ipcs)
 139                         pkt_type |= l3 | RTE_PTYPE_L4_UDP;
 140                 else
 141                         pkt_type |= l3 | RTE_PTYPE_INNER_L4_UDP;
 142                 break;
 143
 144         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_IP):
 145                 pkt_type |= l3;
 146                 break;
 147         }
 148
 149         return pkt_type;
 150 }
 151
 152 static void
 153 bnxt_parse_csum(struct rte_mbuf *mbuf, struct rx_pkt_cmpl_hi *rxcmp1)
 154 {
 155         uint32_t flags;
 156
 157         flags = flags2_0xf(rxcmp1);
 158         /* IP Checksum */
 159         if (likely(IS_IP_NONTUNNEL_PKT(flags))) {
 160                 if (unlikely(RX_CMP_IP_CS_ERROR(rxcmp1)))
 161                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
 162                 else
 163                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
 164         } else if (IS_IP_TUNNEL_PKT(flags)) {
 165                 if (unlikely(RX_CMP_IP_OUTER_CS_ERROR(rxcmp1) ||
 166                              RX_CMP_IP_CS_ERROR(rxcmp1)))
 167                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
 168                 else
 169                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
 170         } else if (unlikely(RX_CMP_IP_CS_UNKNOWN(rxcmp1))) {
 171                 mbuf->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 172         }
 173
 174         /* L4 Checksum */
 175         if (likely(IS_L4_NONTUNNEL_PKT(flags))) {
 176                 if (unlikely(RX_CMP_L4_INNER_CS_ERR2(rxcmp1)))
 177                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
 178                 else
 179                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 180         } else if (IS_L4_TUNNEL_PKT(flags)) {
 181                 if (unlikely(RX_CMP_L4_INNER_CS_ERR2(rxcmp1)))
 182                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
 183                 else
 184                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 185                 if (unlikely(RX_CMP_L4_OUTER_CS_ERR2(rxcmp1))) {
 186                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_BAD;
 187                 } else if (unlikely(IS_L4_TUNNEL_PKT_ONLY_INNER_L4_CS
 188                                     (flags))) {
 189                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_UNKNOWN;
 190                 } else {
 191                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_GOOD;
 192                 }
 193         } else if (unlikely(RX_CMP_L4_CS_UNKNOWN(rxcmp1))) {
 194                 mbuf->ol_flags |= PKT_RX_L4_CKSUM_UNKNOWN;
 195         }
 196 }
 197
 198 uint16_t
 199 bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 200                    uint16_t nb_pkts)
 201 {
 202         struct bnxt_rx_queue *rxq = rx_queue;
 203         struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
 204         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
 205         uint32_t raw_cons = cpr->cp_raw_cons;
 206         uint32_t cons;
 207         int nb_rx_pkts = 0;
 208         struct rx_pkt_cmpl *rxcmp;
 209         const uint64x2_t mbuf_init = {rxq->mbuf_initializer, 0};
 210         const uint8x16_t shuf_msk = {
 211                 0xFF, 0xFF, 0xFF, 0xFF,    /* pkt_type (zeroes) */
 212                 2, 3, 0xFF, 0xFF,          /* pkt_len */
 213                 2, 3,                      /* data_len */
 214                 0xFF, 0xFF,                /* vlan_tci (zeroes) */
 215                 12, 13, 14, 15             /* rss hash */
 216         };
 217         int i;
 218
 219         /* If Rx Q was stopped return */
 220         if (unlikely(!rxq->rx_started))
 221                 return 0;
 222
 223         if (rxq->rxrearm_nb >= RTE_BNXT_RXQ_REARM_THRESH)
 224                 bnxt_rxq_rearm(rxq, rxr);
 225
 226         /* Return no more than RTE_BNXT_MAX_RX_BURST per call. */
 227         nb_pkts = RTE_MIN(nb_pkts, RTE_BNXT_MAX_RX_BURST);
 228
 229         /* Make nb_pkts an integer multiple of RTE_BNXT_DESCS_PER_LOOP. */
 230         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_BNXT_DESCS_PER_LOOP);
 231         if (!nb_pkts)
 232                 return 0;
 233
 234         /* Handle RX burst request */
 235         for (i = 0; i < nb_pkts; i++) {
 236                 struct rx_pkt_cmpl_hi *rxcmp1;
 237                 struct rte_mbuf *mbuf;
 238                 uint64x2_t mm_rxcmp;
 239                 uint8x16_t pkt_mb;
 240
 241                 cons = RING_CMP(cpr->cp_ring_struct, raw_cons);
 242
 243                 rxcmp = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[cons];
 244                 rxcmp1 = (struct rx_pkt_cmpl_hi *)&cpr->cp_desc_ring[cons + 1];
 245
 246                 if (!CMP_VALID(rxcmp1, raw_cons + 1, cpr->cp_ring_struct))
 247                         break;
 248
 249                 raw_cons += 2;
 250                 cons = rxcmp->opaque;
 251
 252                 mbuf = rxr->rx_buf_ring[cons];
 253                 rte_prefetch0(mbuf);
 254                 rxr->rx_buf_ring[cons] = NULL;
 255
 256                 /* Set constant fields from mbuf initializer. */
 257                 vst1q_u64((uint64_t *)&mbuf->rearm_data, mbuf_init);
 258
 259                 /* Set mbuf pkt_len, data_len, and rss_hash fields. */
 260                 mm_rxcmp = vld1q_u64((uint64_t *)rxcmp);
 261                 pkt_mb = vqtbl1q_u8(vreinterpretq_u8_u64(mm_rxcmp), shuf_msk);
 262                 vst1q_u64((uint64_t *)&mbuf->rx_descriptor_fields1,
 263                           vreinterpretq_u64_u8(pkt_mb));
 264
 265                 rte_compiler_barrier();
 266
 267                 if (rxcmp->flags_type & RX_PKT_CMPL_FLAGS_RSS_VALID)
 268                         mbuf->ol_flags |= PKT_RX_RSS_HASH;
 269
 270                 if (rxcmp1->flags2 &
 271                     RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN) {
 272                         mbuf->vlan_tci = rxcmp1->metadata &
 273                                 (RX_PKT_CMPL_METADATA_VID_MASK |
 274                                 RX_PKT_CMPL_METADATA_DE |
 275                                 RX_PKT_CMPL_METADATA_PRI_MASK);
 276                         mbuf->ol_flags |=
 277                                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
 278                 }
 279
 280                 bnxt_parse_csum(mbuf, rxcmp1);
 281                 mbuf->packet_type = bnxt_parse_pkt_type(rxcmp, rxcmp1);
 282
 283                 rx_pkts[nb_rx_pkts++] = mbuf;
 284         }
 285
 286         if (nb_rx_pkts) {
 287                 rxr->rx_prod =
 288                         RING_ADV(rxr->rx_ring_struct, rxr->rx_prod, nb_rx_pkts);
 289
 290                 rxq->rxrearm_nb += nb_rx_pkts;
 291                 cpr->cp_raw_cons = raw_cons;
 292                 cpr->valid =
 293                         !!(cpr->cp_raw_cons & cpr->cp_ring_struct->ring_size);
 294                 bnxt_db_cq(cpr);
 295         }
 296
 297         return nb_rx_pkts;
 298 }
 299
 300 static void
 301 bnxt_tx_cmp_vec(struct bnxt_tx_queue *txq, int nr_pkts)
 302 {
 303         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 304         struct rte_mbuf **free = txq->free;
 305         uint16_t cons = txr->tx_cons;
 306         unsigned int blk = 0;
 307
 308         while (nr_pkts--) {
 309                 struct bnxt_sw_tx_bd *tx_buf;
 310                 struct rte_mbuf *mbuf;
 311
 312                 tx_buf = &txr->tx_buf_ring[cons];
 313                 cons = RING_NEXT(txr->tx_ring_struct, cons);
 314                 mbuf = rte_pktmbuf_prefree_seg(tx_buf->mbuf);
 315                 if (unlikely(mbuf == NULL))
 316                         continue;
 317                 tx_buf->mbuf = NULL;
 318
 319                 if (blk && mbuf->pool != free[0]->pool) {
 320                         rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
 321                         blk = 0;
 322                 }
 323                 free[blk++] = mbuf;
 324         }
 325         if (blk)
 326                 rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
 327
 328         txr->tx_cons = cons;
 329 }
 330
 331 static void
 332 bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
 333 {
 334         struct bnxt_cp_ring_info *cpr = txq->cp_ring;
 335         uint32_t raw_cons = cpr->cp_raw_cons;
 336         uint32_t cons;
 337         uint32_t nb_tx_pkts = 0;
 338         struct tx_cmpl *txcmp;
 339         struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
 340         struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
 341         uint32_t ring_mask = cp_ring_struct->ring_mask;
 342
 343         do {
 344                 cons = RING_CMPL(ring_mask, raw_cons);
 345                 txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
 346
 347                 if (!CMP_VALID(txcmp, raw_cons, cp_ring_struct))
 348                         break;
 349
 350                 if (likely(CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2))
 351                         nb_tx_pkts += txcmp->opaque;
 352                 else
 353                         RTE_LOG_DP(ERR, PMD,
 354                                    "Unhandled CMP type %02x\n",
 355                                    CMP_TYPE(txcmp));
 356                 raw_cons = NEXT_RAW_CMP(raw_cons);
 357         } while (nb_tx_pkts < ring_mask);
 358
 359         cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);
 360         if (nb_tx_pkts) {
 361                 bnxt_tx_cmp_vec(txq, nb_tx_pkts);
 362                 cpr->cp_raw_cons = raw_cons;
 363                 bnxt_db_cq(cpr);
 364         }
 365 }
 366
 367 static uint16_t
 368 bnxt_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 369                           uint16_t nb_pkts)
 370 {
 371         struct bnxt_tx_queue *txq = tx_queue;
 372         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 373         uint16_t prod = txr->tx_prod;
 374         struct rte_mbuf *tx_mbuf;
 375         struct tx_bd_long *txbd = NULL;
 376         struct bnxt_sw_tx_bd *tx_buf;
 377         uint16_t to_send;
 378
 379         nb_pkts = RTE_MIN(nb_pkts, bnxt_tx_avail(txq));
 380
 381         if (unlikely(nb_pkts == 0))
 382                 return 0;
 383
 384         /* Handle TX burst request */
 385         to_send = nb_pkts;
 386         while (to_send) {
 387                 tx_mbuf = *tx_pkts++;
 388                 rte_prefetch0(tx_mbuf);
 389
 390                 tx_buf = &txr->tx_buf_ring[prod];
 391                 tx_buf->mbuf = tx_mbuf;
 392                 tx_buf->nr_bds = 1;
 393
 394                 txbd = &txr->tx_desc_ring[prod];
 395                 txbd->address = tx_mbuf->buf_iova + tx_mbuf->data_off;
 396                 txbd->len = tx_mbuf->data_len;
 397                 txbd->flags_type = bnxt_xmit_flags_len(tx_mbuf->data_len,
 398                                                        TX_BD_FLAGS_NOCMPL);
 399                 prod = RING_NEXT(txr->tx_ring_struct, prod);
 400                 to_send--;
 401         }
 402
 403         /* Request a completion for last packet in burst */
 404         if (txbd) {
 405                 txbd->opaque = nb_pkts;
 406                 txbd->flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
 407         }
 408
 409         rte_compiler_barrier();
 410         bnxt_db_write(&txr->tx_db, prod);
 411
 412         txr->tx_prod = prod;
 413
 414         return nb_pkts;
 415 }
 416
 417 uint16_t
 418 bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 419                    uint16_t nb_pkts)
 420 {
 421         int nb_sent = 0;
 422         struct bnxt_tx_queue *txq = tx_queue;
 423
 424         /* Tx queue was stopped; wait for it to be restarted */
 425         if (unlikely(!txq->tx_started)) {
 426                 PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
 427                 return 0;
 428         }
 429
 430         /* Handle TX completions */
 431         if (bnxt_tx_bds_in_hw(txq) >= txq->tx_free_thresh)
 432                 bnxt_handle_tx_cp_vec(txq);
 433
 434         while (nb_pkts) {
 435                 uint16_t ret, num;
 436
 437                 num = RTE_MIN(nb_pkts, RTE_BNXT_MAX_TX_BURST);
 438                 ret = bnxt_xmit_fixed_burst_vec(tx_queue,
 439                                                 &tx_pkts[nb_sent],
 440                                                 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 }