1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2018-2021 Beijing WangXun Technology Co., Ltd. 3 * Copyright(c) 2010-2017 Intel Corporation 4 */ 5 6 #include <sys/queue.h> 7 8 #include <stdint.h> 9 #include <rte_ethdev.h> 10 #include <ethdev_driver.h> 11 #include <rte_malloc.h> 12 #include <rte_net.h> 13 14 #include "ngbe_logs.h" 15 #include "base/ngbe.h" 16 #include "ngbe_ethdev.h" 17 #include "ngbe_rxtx.h" 18 19 /* Bit Mask to indicate what bits required for building Tx context */ 20 static const u64 NGBE_TX_OFFLOAD_MASK = (RTE_MBUF_F_TX_IP_CKSUM | 21 RTE_MBUF_F_TX_OUTER_IPV6 | 22 RTE_MBUF_F_TX_OUTER_IPV4 | 23 RTE_MBUF_F_TX_IPV6 | 24 RTE_MBUF_F_TX_IPV4 | 25 RTE_MBUF_F_TX_VLAN | 26 RTE_MBUF_F_TX_L4_MASK | 27 RTE_MBUF_F_TX_TCP_SEG | 28 RTE_MBUF_F_TX_TUNNEL_MASK | 29 RTE_MBUF_F_TX_OUTER_IP_CKSUM); 30 #define NGBE_TX_OFFLOAD_NOTSUP_MASK \ 31 (RTE_MBUF_F_TX_OFFLOAD_MASK ^ NGBE_TX_OFFLOAD_MASK) 32 33 /* 34 * Prefetch a cache line into all cache levels. 35 */ 36 #define rte_ngbe_prefetch(p) rte_prefetch0(p) 37 38 /********************************************************************* 39 * 40 * Tx functions 41 * 42 **********************************************************************/ 43 44 /* 45 * Check for descriptors with their DD bit set and free mbufs. 46 * Return the total number of buffers freed. 47 */ 48 static __rte_always_inline int 49 ngbe_tx_free_bufs(struct ngbe_tx_queue *txq) 50 { 51 struct ngbe_tx_entry *txep; 52 uint32_t status; 53 int i, nb_free = 0; 54 struct rte_mbuf *m, *free[RTE_NGBE_TX_MAX_FREE_BUF_SZ]; 55 56 /* check DD bit on threshold descriptor */ 57 status = txq->tx_ring[txq->tx_next_dd].dw3; 58 if (!(status & rte_cpu_to_le_32(NGBE_TXD_DD))) { 59 if (txq->nb_tx_free >> 1 < txq->tx_free_thresh) 60 ngbe_set32_masked(txq->tdc_reg_addr, 61 NGBE_TXCFG_FLUSH, NGBE_TXCFG_FLUSH); 62 return 0; 63 } 64 65 /* 66 * first buffer to free from S/W ring is at index 67 * tx_next_dd - (tx_free_thresh-1) 68 */ 69 txep = &txq->sw_ring[txq->tx_next_dd - (txq->tx_free_thresh - 1)]; 70 for (i = 0; i < txq->tx_free_thresh; ++i, ++txep) { 71 /* free buffers one at a time */ 72 m = rte_pktmbuf_prefree_seg(txep->mbuf); 73 txep->mbuf = NULL; 74 75 if (unlikely(m == NULL)) 76 continue; 77 78 if (nb_free >= RTE_NGBE_TX_MAX_FREE_BUF_SZ || 79 (nb_free > 0 && m->pool != free[0]->pool)) { 80 rte_mempool_put_bulk(free[0]->pool, 81 (void **)free, nb_free); 82 nb_free = 0; 83 } 84 85 free[nb_free++] = m; 86 } 87 88 if (nb_free > 0) 89 rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free); 90 91 /* buffers were freed, update counters */ 92 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_free_thresh); 93 txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_free_thresh); 94 if (txq->tx_next_dd >= txq->nb_tx_desc) 95 txq->tx_next_dd = (uint16_t)(txq->tx_free_thresh - 1); 96 97 return txq->tx_free_thresh; 98 } 99 100 /* Populate 4 descriptors with data from 4 mbufs */ 101 static inline void 102 tx4(volatile struct ngbe_tx_desc *txdp, struct rte_mbuf **pkts) 103 { 104 uint64_t buf_dma_addr; 105 uint32_t pkt_len; 106 int i; 107 108 for (i = 0; i < 4; ++i, ++txdp, ++pkts) { 109 buf_dma_addr = rte_mbuf_data_iova(*pkts); 110 pkt_len = (*pkts)->data_len; 111 112 /* write data to descriptor */ 113 txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr); 114 txdp->dw2 = cpu_to_le32(NGBE_TXD_FLAGS | 115 NGBE_TXD_DATLEN(pkt_len)); 116 txdp->dw3 = cpu_to_le32(NGBE_TXD_PAYLEN(pkt_len)); 117 118 rte_prefetch0(&(*pkts)->pool); 119 } 120 } 121 122 /* Populate 1 descriptor with data from 1 mbuf */ 123 static inline void 124 tx1(volatile struct ngbe_tx_desc *txdp, struct rte_mbuf **pkts) 125 { 126 uint64_t buf_dma_addr; 127 uint32_t pkt_len; 128 129 buf_dma_addr = rte_mbuf_data_iova(*pkts); 130 pkt_len = (*pkts)->data_len; 131 132 /* write data to descriptor */ 133 txdp->qw0 = cpu_to_le64(buf_dma_addr); 134 txdp->dw2 = cpu_to_le32(NGBE_TXD_FLAGS | 135 NGBE_TXD_DATLEN(pkt_len)); 136 txdp->dw3 = cpu_to_le32(NGBE_TXD_PAYLEN(pkt_len)); 137 138 rte_prefetch0(&(*pkts)->pool); 139 } 140 141 /* 142 * Fill H/W descriptor ring with mbuf data. 143 * Copy mbuf pointers to the S/W ring. 144 */ 145 static inline void 146 ngbe_tx_fill_hw_ring(struct ngbe_tx_queue *txq, struct rte_mbuf **pkts, 147 uint16_t nb_pkts) 148 { 149 volatile struct ngbe_tx_desc *txdp = &txq->tx_ring[txq->tx_tail]; 150 struct ngbe_tx_entry *txep = &txq->sw_ring[txq->tx_tail]; 151 const int N_PER_LOOP = 4; 152 const int N_PER_LOOP_MASK = N_PER_LOOP - 1; 153 int mainpart, leftover; 154 int i, j; 155 156 /* 157 * Process most of the packets in chunks of N pkts. Any 158 * leftover packets will get processed one at a time. 159 */ 160 mainpart = (nb_pkts & ((uint32_t)~N_PER_LOOP_MASK)); 161 leftover = (nb_pkts & ((uint32_t)N_PER_LOOP_MASK)); 162 for (i = 0; i < mainpart; i += N_PER_LOOP) { 163 /* Copy N mbuf pointers to the S/W ring */ 164 for (j = 0; j < N_PER_LOOP; ++j) 165 (txep + i + j)->mbuf = *(pkts + i + j); 166 tx4(txdp + i, pkts + i); 167 } 168 169 if (unlikely(leftover > 0)) { 170 for (i = 0; i < leftover; ++i) { 171 (txep + mainpart + i)->mbuf = *(pkts + mainpart + i); 172 tx1(txdp + mainpart + i, pkts + mainpart + i); 173 } 174 } 175 } 176 177 static inline uint16_t 178 tx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, 179 uint16_t nb_pkts) 180 { 181 struct ngbe_tx_queue *txq = (struct ngbe_tx_queue *)tx_queue; 182 uint16_t n = 0; 183 184 /* 185 * Begin scanning the H/W ring for done descriptors when the 186 * number of available descriptors drops below tx_free_thresh. 187 * For each done descriptor, free the associated buffer. 188 */ 189 if (txq->nb_tx_free < txq->tx_free_thresh) 190 ngbe_tx_free_bufs(txq); 191 192 /* Only use descriptors that are available */ 193 nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts); 194 if (unlikely(nb_pkts == 0)) 195 return 0; 196 197 /* Use exactly nb_pkts descriptors */ 198 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts); 199 200 /* 201 * At this point, we know there are enough descriptors in the 202 * ring to transmit all the packets. This assumes that each 203 * mbuf contains a single segment, and that no new offloads 204 * are expected, which would require a new context descriptor. 205 */ 206 207 /* 208 * See if we're going to wrap-around. If so, handle the top 209 * of the descriptor ring first, then do the bottom. If not, 210 * the processing looks just like the "bottom" part anyway... 211 */ 212 if ((txq->tx_tail + nb_pkts) > txq->nb_tx_desc) { 213 n = (uint16_t)(txq->nb_tx_desc - txq->tx_tail); 214 ngbe_tx_fill_hw_ring(txq, tx_pkts, n); 215 txq->tx_tail = 0; 216 } 217 218 /* Fill H/W descriptor ring with mbuf data */ 219 ngbe_tx_fill_hw_ring(txq, tx_pkts + n, (uint16_t)(nb_pkts - n)); 220 txq->tx_tail = (uint16_t)(txq->tx_tail + (nb_pkts - n)); 221 222 /* 223 * Check for wrap-around. This would only happen if we used 224 * up to the last descriptor in the ring, no more, no less. 225 */ 226 if (txq->tx_tail >= txq->nb_tx_desc) 227 txq->tx_tail = 0; 228 229 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u", 230 (uint16_t)txq->port_id, (uint16_t)txq->queue_id, 231 (uint16_t)txq->tx_tail, (uint16_t)nb_pkts); 232 233 /* update tail pointer */ 234 rte_wmb(); 235 ngbe_set32_relaxed(txq->tdt_reg_addr, txq->tx_tail); 236 237 return nb_pkts; 238 } 239 240 uint16_t 241 ngbe_xmit_pkts_simple(void *tx_queue, struct rte_mbuf **tx_pkts, 242 uint16_t nb_pkts) 243 { 244 uint16_t nb_tx; 245 246 /* Try to transmit at least chunks of TX_MAX_BURST pkts */ 247 if (likely(nb_pkts <= RTE_PMD_NGBE_TX_MAX_BURST)) 248 return tx_xmit_pkts(tx_queue, tx_pkts, nb_pkts); 249 250 /* transmit more than the max burst, in chunks of TX_MAX_BURST */ 251 nb_tx = 0; 252 while (nb_pkts != 0) { 253 uint16_t ret, n; 254 255 n = (uint16_t)RTE_MIN(nb_pkts, RTE_PMD_NGBE_TX_MAX_BURST); 256 ret = tx_xmit_pkts(tx_queue, &tx_pkts[nb_tx], n); 257 nb_tx = (uint16_t)(nb_tx + ret); 258 nb_pkts = (uint16_t)(nb_pkts - ret); 259 if (ret < n) 260 break; 261 } 262 263 return nb_tx; 264 } 265 266 static inline void 267 ngbe_set_xmit_ctx(struct ngbe_tx_queue *txq, 268 volatile struct ngbe_tx_ctx_desc *ctx_txd, 269 uint64_t ol_flags, union ngbe_tx_offload tx_offload) 270 { 271 union ngbe_tx_offload tx_offload_mask; 272 uint32_t type_tucmd_mlhl; 273 uint32_t mss_l4len_idx; 274 uint32_t ctx_idx; 275 uint32_t vlan_macip_lens; 276 uint32_t tunnel_seed; 277 278 ctx_idx = txq->ctx_curr; 279 tx_offload_mask.data[0] = 0; 280 tx_offload_mask.data[1] = 0; 281 282 /* Specify which HW CTX to upload. */ 283 mss_l4len_idx = NGBE_TXD_IDX(ctx_idx); 284 type_tucmd_mlhl = NGBE_TXD_CTXT; 285 286 tx_offload_mask.ptid |= ~0; 287 type_tucmd_mlhl |= NGBE_TXD_PTID(tx_offload.ptid); 288 289 /* check if TCP segmentation required for this packet */ 290 if (ol_flags & RTE_MBUF_F_TX_TCP_SEG) { 291 tx_offload_mask.l2_len |= ~0; 292 tx_offload_mask.l3_len |= ~0; 293 tx_offload_mask.l4_len |= ~0; 294 tx_offload_mask.tso_segsz |= ~0; 295 mss_l4len_idx |= NGBE_TXD_MSS(tx_offload.tso_segsz); 296 mss_l4len_idx |= NGBE_TXD_L4LEN(tx_offload.l4_len); 297 } else { /* no TSO, check if hardware checksum is needed */ 298 if (ol_flags & RTE_MBUF_F_TX_IP_CKSUM) { 299 tx_offload_mask.l2_len |= ~0; 300 tx_offload_mask.l3_len |= ~0; 301 } 302 303 switch (ol_flags & RTE_MBUF_F_TX_L4_MASK) { 304 case RTE_MBUF_F_TX_UDP_CKSUM: 305 mss_l4len_idx |= 306 NGBE_TXD_L4LEN(sizeof(struct rte_udp_hdr)); 307 tx_offload_mask.l2_len |= ~0; 308 tx_offload_mask.l3_len |= ~0; 309 break; 310 case RTE_MBUF_F_TX_TCP_CKSUM: 311 mss_l4len_idx |= 312 NGBE_TXD_L4LEN(sizeof(struct rte_tcp_hdr)); 313 tx_offload_mask.l2_len |= ~0; 314 tx_offload_mask.l3_len |= ~0; 315 break; 316 case RTE_MBUF_F_TX_SCTP_CKSUM: 317 mss_l4len_idx |= 318 NGBE_TXD_L4LEN(sizeof(struct rte_sctp_hdr)); 319 tx_offload_mask.l2_len |= ~0; 320 tx_offload_mask.l3_len |= ~0; 321 break; 322 default: 323 break; 324 } 325 } 326 327 vlan_macip_lens = NGBE_TXD_IPLEN(tx_offload.l3_len >> 1); 328 329 if (ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK) { 330 tx_offload_mask.outer_tun_len |= ~0; 331 tx_offload_mask.outer_l2_len |= ~0; 332 tx_offload_mask.outer_l3_len |= ~0; 333 tx_offload_mask.l2_len |= ~0; 334 tunnel_seed = NGBE_TXD_ETUNLEN(tx_offload.outer_tun_len >> 1); 335 tunnel_seed |= NGBE_TXD_EIPLEN(tx_offload.outer_l3_len >> 2); 336 337 switch (ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK) { 338 case RTE_MBUF_F_TX_TUNNEL_IPIP: 339 /* for non UDP / GRE tunneling, set to 0b */ 340 break; 341 default: 342 PMD_TX_LOG(ERR, "Tunnel type not supported"); 343 return; 344 } 345 vlan_macip_lens |= NGBE_TXD_MACLEN(tx_offload.outer_l2_len); 346 } else { 347 tunnel_seed = 0; 348 vlan_macip_lens |= NGBE_TXD_MACLEN(tx_offload.l2_len); 349 } 350 351 if (ol_flags & RTE_MBUF_F_TX_VLAN) { 352 tx_offload_mask.vlan_tci |= ~0; 353 vlan_macip_lens |= NGBE_TXD_VLAN(tx_offload.vlan_tci); 354 } 355 356 txq->ctx_cache[ctx_idx].flags = ol_flags; 357 txq->ctx_cache[ctx_idx].tx_offload.data[0] = 358 tx_offload_mask.data[0] & tx_offload.data[0]; 359 txq->ctx_cache[ctx_idx].tx_offload.data[1] = 360 tx_offload_mask.data[1] & tx_offload.data[1]; 361 txq->ctx_cache[ctx_idx].tx_offload_mask = tx_offload_mask; 362 363 ctx_txd->dw0 = rte_cpu_to_le_32(vlan_macip_lens); 364 ctx_txd->dw1 = rte_cpu_to_le_32(tunnel_seed); 365 ctx_txd->dw2 = rte_cpu_to_le_32(type_tucmd_mlhl); 366 ctx_txd->dw3 = rte_cpu_to_le_32(mss_l4len_idx); 367 } 368 369 /* 370 * Check which hardware context can be used. Use the existing match 371 * or create a new context descriptor. 372 */ 373 static inline uint32_t 374 what_ctx_update(struct ngbe_tx_queue *txq, uint64_t flags, 375 union ngbe_tx_offload tx_offload) 376 { 377 /* If match with the current used context */ 378 if (likely(txq->ctx_cache[txq->ctx_curr].flags == flags && 379 (txq->ctx_cache[txq->ctx_curr].tx_offload.data[0] == 380 (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data[0] 381 & tx_offload.data[0])) && 382 (txq->ctx_cache[txq->ctx_curr].tx_offload.data[1] == 383 (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data[1] 384 & tx_offload.data[1])))) 385 return txq->ctx_curr; 386 387 /* What if match with the next context */ 388 txq->ctx_curr ^= 1; 389 if (likely(txq->ctx_cache[txq->ctx_curr].flags == flags && 390 (txq->ctx_cache[txq->ctx_curr].tx_offload.data[0] == 391 (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data[0] 392 & tx_offload.data[0])) && 393 (txq->ctx_cache[txq->ctx_curr].tx_offload.data[1] == 394 (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data[1] 395 & tx_offload.data[1])))) 396 return txq->ctx_curr; 397 398 /* Mismatch, use the previous context */ 399 return NGBE_CTX_NUM; 400 } 401 402 static inline uint32_t 403 tx_desc_cksum_flags_to_olinfo(uint64_t ol_flags) 404 { 405 uint32_t tmp = 0; 406 407 if ((ol_flags & RTE_MBUF_F_TX_L4_MASK) != RTE_MBUF_F_TX_L4_NO_CKSUM) { 408 tmp |= NGBE_TXD_CC; 409 tmp |= NGBE_TXD_L4CS; 410 } 411 if (ol_flags & RTE_MBUF_F_TX_IP_CKSUM) { 412 tmp |= NGBE_TXD_CC; 413 tmp |= NGBE_TXD_IPCS; 414 } 415 if (ol_flags & RTE_MBUF_F_TX_OUTER_IP_CKSUM) { 416 tmp |= NGBE_TXD_CC; 417 tmp |= NGBE_TXD_EIPCS; 418 } 419 if (ol_flags & RTE_MBUF_F_TX_TCP_SEG) { 420 tmp |= NGBE_TXD_CC; 421 /* implies IPv4 cksum */ 422 if (ol_flags & RTE_MBUF_F_TX_IPV4) 423 tmp |= NGBE_TXD_IPCS; 424 tmp |= NGBE_TXD_L4CS; 425 } 426 if (ol_flags & RTE_MBUF_F_TX_VLAN) 427 tmp |= NGBE_TXD_CC; 428 429 return tmp; 430 } 431 432 static inline uint32_t 433 tx_desc_ol_flags_to_cmdtype(uint64_t ol_flags) 434 { 435 uint32_t cmdtype = 0; 436 437 if (ol_flags & RTE_MBUF_F_TX_VLAN) 438 cmdtype |= NGBE_TXD_VLE; 439 if (ol_flags & RTE_MBUF_F_TX_TCP_SEG) 440 cmdtype |= NGBE_TXD_TSE; 441 return cmdtype; 442 } 443 444 static inline uint8_t 445 tx_desc_ol_flags_to_ptid(uint64_t oflags, uint32_t ptype) 446 { 447 bool tun; 448 449 if (ptype) 450 return ngbe_encode_ptype(ptype); 451 452 /* Only support flags in NGBE_TX_OFFLOAD_MASK */ 453 tun = !!(oflags & RTE_MBUF_F_TX_TUNNEL_MASK); 454 455 /* L2 level */ 456 ptype = RTE_PTYPE_L2_ETHER; 457 if (oflags & RTE_MBUF_F_TX_VLAN) 458 ptype |= RTE_PTYPE_L2_ETHER_VLAN; 459 460 /* L3 level */ 461 if (oflags & (RTE_MBUF_F_TX_OUTER_IPV4 | RTE_MBUF_F_TX_OUTER_IP_CKSUM)) 462 ptype |= RTE_PTYPE_L3_IPV4; 463 else if (oflags & (RTE_MBUF_F_TX_OUTER_IPV6)) 464 ptype |= RTE_PTYPE_L3_IPV6; 465 466 if (oflags & (RTE_MBUF_F_TX_IPV4 | RTE_MBUF_F_TX_IP_CKSUM)) 467 ptype |= (tun ? RTE_PTYPE_INNER_L3_IPV4 : RTE_PTYPE_L3_IPV4); 468 else if (oflags & (RTE_MBUF_F_TX_IPV6)) 469 ptype |= (tun ? RTE_PTYPE_INNER_L3_IPV6 : RTE_PTYPE_L3_IPV6); 470 471 /* L4 level */ 472 switch (oflags & (RTE_MBUF_F_TX_L4_MASK)) { 473 case RTE_MBUF_F_TX_TCP_CKSUM: 474 ptype |= (tun ? RTE_PTYPE_INNER_L4_TCP : RTE_PTYPE_L4_TCP); 475 break; 476 case RTE_MBUF_F_TX_UDP_CKSUM: 477 ptype |= (tun ? RTE_PTYPE_INNER_L4_UDP : RTE_PTYPE_L4_UDP); 478 break; 479 case RTE_MBUF_F_TX_SCTP_CKSUM: 480 ptype |= (tun ? RTE_PTYPE_INNER_L4_SCTP : RTE_PTYPE_L4_SCTP); 481 break; 482 } 483 484 if (oflags & RTE_MBUF_F_TX_TCP_SEG) 485 ptype |= (tun ? RTE_PTYPE_INNER_L4_TCP : RTE_PTYPE_L4_TCP); 486 487 /* Tunnel */ 488 switch (oflags & RTE_MBUF_F_TX_TUNNEL_MASK) { 489 case RTE_MBUF_F_TX_TUNNEL_IPIP: 490 case RTE_MBUF_F_TX_TUNNEL_IP: 491 ptype |= RTE_PTYPE_L2_ETHER | 492 RTE_PTYPE_L3_IPV4 | 493 RTE_PTYPE_TUNNEL_IP; 494 break; 495 } 496 497 return ngbe_encode_ptype(ptype); 498 } 499 500 /* Reset transmit descriptors after they have been used */ 501 static inline int 502 ngbe_xmit_cleanup(struct ngbe_tx_queue *txq) 503 { 504 struct ngbe_tx_entry *sw_ring = txq->sw_ring; 505 volatile struct ngbe_tx_desc *txr = txq->tx_ring; 506 uint16_t last_desc_cleaned = txq->last_desc_cleaned; 507 uint16_t nb_tx_desc = txq->nb_tx_desc; 508 uint16_t desc_to_clean_to; 509 uint16_t nb_tx_to_clean; 510 uint32_t status; 511 512 /* Determine the last descriptor needing to be cleaned */ 513 desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_free_thresh); 514 if (desc_to_clean_to >= nb_tx_desc) 515 desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc); 516 517 /* Check to make sure the last descriptor to clean is done */ 518 desc_to_clean_to = sw_ring[desc_to_clean_to].last_id; 519 status = txr[desc_to_clean_to].dw3; 520 if (!(status & rte_cpu_to_le_32(NGBE_TXD_DD))) { 521 PMD_TX_LOG(DEBUG, 522 "Tx descriptor %4u is not done" 523 "(port=%d queue=%d)", 524 desc_to_clean_to, 525 txq->port_id, txq->queue_id); 526 if (txq->nb_tx_free >> 1 < txq->tx_free_thresh) 527 ngbe_set32_masked(txq->tdc_reg_addr, 528 NGBE_TXCFG_FLUSH, NGBE_TXCFG_FLUSH); 529 /* Failed to clean any descriptors, better luck next time */ 530 return -(1); 531 } 532 533 /* Figure out how many descriptors will be cleaned */ 534 if (last_desc_cleaned > desc_to_clean_to) 535 nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) + 536 desc_to_clean_to); 537 else 538 nb_tx_to_clean = (uint16_t)(desc_to_clean_to - 539 last_desc_cleaned); 540 541 PMD_TX_LOG(DEBUG, 542 "Cleaning %4u Tx descriptors: %4u to %4u (port=%d queue=%d)", 543 nb_tx_to_clean, last_desc_cleaned, desc_to_clean_to, 544 txq->port_id, txq->queue_id); 545 546 /* 547 * The last descriptor to clean is done, so that means all the 548 * descriptors from the last descriptor that was cleaned 549 * up to the last descriptor with the RS bit set 550 * are done. Only reset the threshold descriptor. 551 */ 552 txr[desc_to_clean_to].dw3 = 0; 553 554 /* Update the txq to reflect the last descriptor that was cleaned */ 555 txq->last_desc_cleaned = desc_to_clean_to; 556 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean); 557 558 /* No Error */ 559 return 0; 560 } 561 562 uint16_t 563 ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, 564 uint16_t nb_pkts) 565 { 566 struct ngbe_tx_queue *txq; 567 struct ngbe_tx_entry *sw_ring; 568 struct ngbe_tx_entry *txe, *txn; 569 volatile struct ngbe_tx_desc *txr; 570 volatile struct ngbe_tx_desc *txd; 571 struct rte_mbuf *tx_pkt; 572 struct rte_mbuf *m_seg; 573 uint64_t buf_dma_addr; 574 uint32_t olinfo_status; 575 uint32_t cmd_type_len; 576 uint32_t pkt_len; 577 uint16_t slen; 578 uint64_t ol_flags; 579 uint16_t tx_id; 580 uint16_t tx_last; 581 uint16_t nb_tx; 582 uint16_t nb_used; 583 uint64_t tx_ol_req; 584 uint32_t ctx = 0; 585 uint32_t new_ctx; 586 union ngbe_tx_offload tx_offload; 587 588 tx_offload.data[0] = 0; 589 tx_offload.data[1] = 0; 590 txq = tx_queue; 591 sw_ring = txq->sw_ring; 592 txr = txq->tx_ring; 593 tx_id = txq->tx_tail; 594 txe = &sw_ring[tx_id]; 595 596 /* Determine if the descriptor ring needs to be cleaned. */ 597 if (txq->nb_tx_free < txq->tx_free_thresh) 598 ngbe_xmit_cleanup(txq); 599 600 rte_prefetch0(&txe->mbuf->pool); 601 602 /* Tx loop */ 603 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) { 604 new_ctx = 0; 605 tx_pkt = *tx_pkts++; 606 pkt_len = tx_pkt->pkt_len; 607 608 /* 609 * Determine how many (if any) context descriptors 610 * are needed for offload functionality. 611 */ 612 ol_flags = tx_pkt->ol_flags; 613 614 /* If hardware offload required */ 615 tx_ol_req = ol_flags & NGBE_TX_OFFLOAD_MASK; 616 if (tx_ol_req) { 617 tx_offload.ptid = tx_desc_ol_flags_to_ptid(tx_ol_req, 618 tx_pkt->packet_type); 619 tx_offload.l2_len = tx_pkt->l2_len; 620 tx_offload.l3_len = tx_pkt->l3_len; 621 tx_offload.l4_len = tx_pkt->l4_len; 622 tx_offload.vlan_tci = tx_pkt->vlan_tci; 623 tx_offload.tso_segsz = tx_pkt->tso_segsz; 624 tx_offload.outer_l2_len = tx_pkt->outer_l2_len; 625 tx_offload.outer_l3_len = tx_pkt->outer_l3_len; 626 tx_offload.outer_tun_len = 0; 627 628 /* If new context need be built or reuse the exist ctx*/ 629 ctx = what_ctx_update(txq, tx_ol_req, tx_offload); 630 /* Only allocate context descriptor if required */ 631 new_ctx = (ctx == NGBE_CTX_NUM); 632 ctx = txq->ctx_curr; 633 } 634 635 /* 636 * Keep track of how many descriptors are used this loop 637 * This will always be the number of segments + the number of 638 * Context descriptors required to transmit the packet 639 */ 640 nb_used = (uint16_t)(tx_pkt->nb_segs + new_ctx); 641 642 /* 643 * The number of descriptors that must be allocated for a 644 * packet is the number of segments of that packet, plus 1 645 * Context Descriptor for the hardware offload, if any. 646 * Determine the last Tx descriptor to allocate in the Tx ring 647 * for the packet, starting from the current position (tx_id) 648 * in the ring. 649 */ 650 tx_last = (uint16_t)(tx_id + nb_used - 1); 651 652 /* Circular ring */ 653 if (tx_last >= txq->nb_tx_desc) 654 tx_last = (uint16_t)(tx_last - txq->nb_tx_desc); 655 656 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u pktlen=%u" 657 " tx_first=%u tx_last=%u", 658 (uint16_t)txq->port_id, 659 (uint16_t)txq->queue_id, 660 (uint32_t)pkt_len, 661 (uint16_t)tx_id, 662 (uint16_t)tx_last); 663 664 /* 665 * Make sure there are enough Tx descriptors available to 666 * transmit the entire packet. 667 * nb_used better be less than or equal to txq->tx_free_thresh 668 */ 669 if (nb_used > txq->nb_tx_free) { 670 PMD_TX_LOG(DEBUG, 671 "Not enough free Tx descriptors " 672 "nb_used=%4u nb_free=%4u " 673 "(port=%d queue=%d)", 674 nb_used, txq->nb_tx_free, 675 txq->port_id, txq->queue_id); 676 677 if (ngbe_xmit_cleanup(txq) != 0) { 678 /* Could not clean any descriptors */ 679 if (nb_tx == 0) 680 return 0; 681 goto end_of_tx; 682 } 683 684 /* nb_used better be <= txq->tx_free_thresh */ 685 if (unlikely(nb_used > txq->tx_free_thresh)) { 686 PMD_TX_LOG(DEBUG, 687 "The number of descriptors needed to " 688 "transmit the packet exceeds the " 689 "RS bit threshold. This will impact " 690 "performance." 691 "nb_used=%4u nb_free=%4u " 692 "tx_free_thresh=%4u. " 693 "(port=%d queue=%d)", 694 nb_used, txq->nb_tx_free, 695 txq->tx_free_thresh, 696 txq->port_id, txq->queue_id); 697 /* 698 * Loop here until there are enough Tx 699 * descriptors or until the ring cannot be 700 * cleaned. 701 */ 702 while (nb_used > txq->nb_tx_free) { 703 if (ngbe_xmit_cleanup(txq) != 0) { 704 /* 705 * Could not clean any 706 * descriptors 707 */ 708 if (nb_tx == 0) 709 return 0; 710 goto end_of_tx; 711 } 712 } 713 } 714 } 715 716 /* 717 * By now there are enough free Tx descriptors to transmit 718 * the packet. 719 */ 720 721 /* 722 * Set common flags of all Tx Data Descriptors. 723 * 724 * The following bits must be set in the first Data Descriptor 725 * and are ignored in the other ones: 726 * - NGBE_TXD_FCS 727 * 728 * The following bits must only be set in the last Data 729 * Descriptor: 730 * - NGBE_TXD_EOP 731 */ 732 cmd_type_len = NGBE_TXD_FCS; 733 734 olinfo_status = 0; 735 if (tx_ol_req) { 736 if (ol_flags & RTE_MBUF_F_TX_TCP_SEG) { 737 /* when TSO is on, paylen in descriptor is the 738 * not the packet len but the tcp payload len 739 */ 740 pkt_len -= (tx_offload.l2_len + 741 tx_offload.l3_len + tx_offload.l4_len); 742 pkt_len -= 743 (tx_pkt->ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK) 744 ? tx_offload.outer_l2_len + 745 tx_offload.outer_l3_len : 0; 746 } 747 748 /* 749 * Setup the Tx Context Descriptor if required 750 */ 751 if (new_ctx) { 752 volatile struct ngbe_tx_ctx_desc *ctx_txd; 753 754 ctx_txd = (volatile struct ngbe_tx_ctx_desc *) 755 &txr[tx_id]; 756 757 txn = &sw_ring[txe->next_id]; 758 rte_prefetch0(&txn->mbuf->pool); 759 760 if (txe->mbuf != NULL) { 761 rte_pktmbuf_free_seg(txe->mbuf); 762 txe->mbuf = NULL; 763 } 764 765 ngbe_set_xmit_ctx(txq, ctx_txd, tx_ol_req, 766 tx_offload); 767 768 txe->last_id = tx_last; 769 tx_id = txe->next_id; 770 txe = txn; 771 } 772 773 /* 774 * Setup the Tx Data Descriptor, 775 * This path will go through 776 * whatever new/reuse the context descriptor 777 */ 778 cmd_type_len |= tx_desc_ol_flags_to_cmdtype(ol_flags); 779 olinfo_status |= 780 tx_desc_cksum_flags_to_olinfo(ol_flags); 781 olinfo_status |= NGBE_TXD_IDX(ctx); 782 } 783 784 olinfo_status |= NGBE_TXD_PAYLEN(pkt_len); 785 786 m_seg = tx_pkt; 787 do { 788 txd = &txr[tx_id]; 789 txn = &sw_ring[txe->next_id]; 790 rte_prefetch0(&txn->mbuf->pool); 791 792 if (txe->mbuf != NULL) 793 rte_pktmbuf_free_seg(txe->mbuf); 794 txe->mbuf = m_seg; 795 796 /* 797 * Set up Transmit Data Descriptor. 798 */ 799 slen = m_seg->data_len; 800 buf_dma_addr = rte_mbuf_data_iova(m_seg); 801 txd->qw0 = rte_cpu_to_le_64(buf_dma_addr); 802 txd->dw2 = rte_cpu_to_le_32(cmd_type_len | slen); 803 txd->dw3 = rte_cpu_to_le_32(olinfo_status); 804 txe->last_id = tx_last; 805 tx_id = txe->next_id; 806 txe = txn; 807 m_seg = m_seg->next; 808 } while (m_seg != NULL); 809 810 /* 811 * The last packet data descriptor needs End Of Packet (EOP) 812 */ 813 cmd_type_len |= NGBE_TXD_EOP; 814 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used); 815 816 txd->dw2 |= rte_cpu_to_le_32(cmd_type_len); 817 } 818 819 end_of_tx: 820 821 rte_wmb(); 822 823 /* 824 * Set the Transmit Descriptor Tail (TDT) 825 */ 826 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u", 827 (uint16_t)txq->port_id, (uint16_t)txq->queue_id, 828 (uint16_t)tx_id, (uint16_t)nb_tx); 829 ngbe_set32_relaxed(txq->tdt_reg_addr, tx_id); 830 txq->tx_tail = tx_id; 831 832 return nb_tx; 833 } 834 835 /********************************************************************* 836 * 837 * Tx prep functions 838 * 839 **********************************************************************/ 840 uint16_t 841 ngbe_prep_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts) 842 { 843 int i, ret; 844 uint64_t ol_flags; 845 struct rte_mbuf *m; 846 struct ngbe_tx_queue *txq = (struct ngbe_tx_queue *)tx_queue; 847 848 for (i = 0; i < nb_pkts; i++) { 849 m = tx_pkts[i]; 850 ol_flags = m->ol_flags; 851 852 /** 853 * Check if packet meets requirements for number of segments 854 * 855 * NOTE: for ngbe it's always (40 - WTHRESH) for both TSO and 856 * non-TSO 857 */ 858 859 if (m->nb_segs > NGBE_TX_MAX_SEG - txq->wthresh) { 860 rte_errno = -EINVAL; 861 return i; 862 } 863 864 if (ol_flags & NGBE_TX_OFFLOAD_NOTSUP_MASK) { 865 rte_errno = -ENOTSUP; 866 return i; 867 } 868 869 #ifdef RTE_ETHDEV_DEBUG_TX 870 ret = rte_validate_tx_offload(m); 871 if (ret != 0) { 872 rte_errno = ret; 873 return i; 874 } 875 #endif 876 ret = rte_net_intel_cksum_prepare(m); 877 if (ret != 0) { 878 rte_errno = ret; 879 return i; 880 } 881 } 882 883 return i; 884 } 885 886 /********************************************************************* 887 * 888 * Rx functions 889 * 890 **********************************************************************/ 891 static inline uint32_t 892 ngbe_rxd_pkt_info_to_pkt_type(uint32_t pkt_info, uint16_t ptid_mask) 893 { 894 uint16_t ptid = NGBE_RXD_PTID(pkt_info); 895 896 ptid &= ptid_mask; 897 898 return ngbe_decode_ptype(ptid); 899 } 900 901 static inline uint64_t 902 rx_desc_status_to_pkt_flags(uint32_t rx_status, uint64_t vlan_flags) 903 { 904 uint64_t pkt_flags; 905 906 /* 907 * Check if VLAN present only. 908 * Do not check whether L3/L4 rx checksum done by NIC or not, 909 * That can be found from rte_eth_rxmode.offloads flag 910 */ 911 pkt_flags = (rx_status & NGBE_RXD_STAT_VLAN && 912 vlan_flags & RTE_MBUF_F_RX_VLAN_STRIPPED) 913 ? vlan_flags : 0; 914 915 return pkt_flags; 916 } 917 918 static inline uint64_t 919 rx_desc_error_to_pkt_flags(uint32_t rx_status) 920 { 921 uint64_t pkt_flags = 0; 922 923 /* checksum offload can't be disabled */ 924 if (rx_status & NGBE_RXD_STAT_IPCS) 925 pkt_flags |= (rx_status & NGBE_RXD_ERR_IPCS 926 ? RTE_MBUF_F_RX_IP_CKSUM_BAD : RTE_MBUF_F_RX_IP_CKSUM_GOOD); 927 928 if (rx_status & NGBE_RXD_STAT_L4CS) 929 pkt_flags |= (rx_status & NGBE_RXD_ERR_L4CS 930 ? RTE_MBUF_F_RX_L4_CKSUM_BAD : RTE_MBUF_F_RX_L4_CKSUM_GOOD); 931 932 if (rx_status & NGBE_RXD_STAT_EIPCS && 933 rx_status & NGBE_RXD_ERR_EIPCS) 934 pkt_flags |= RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD; 935 936 return pkt_flags; 937 } 938 939 /* 940 * LOOK_AHEAD defines how many desc statuses to check beyond the 941 * current descriptor. 942 * It must be a pound define for optimal performance. 943 * Do not change the value of LOOK_AHEAD, as the ngbe_rx_scan_hw_ring 944 * function only works with LOOK_AHEAD=8. 945 */ 946 #define LOOK_AHEAD 8 947 #if (LOOK_AHEAD != 8) 948 #error "PMD NGBE: LOOK_AHEAD must be 8\n" 949 #endif 950 static inline int 951 ngbe_rx_scan_hw_ring(struct ngbe_rx_queue *rxq) 952 { 953 volatile struct ngbe_rx_desc *rxdp; 954 struct ngbe_rx_entry *rxep; 955 struct rte_mbuf *mb; 956 uint16_t pkt_len; 957 uint64_t pkt_flags; 958 int nb_dd; 959 uint32_t s[LOOK_AHEAD]; 960 uint32_t pkt_info[LOOK_AHEAD]; 961 int i, j, nb_rx = 0; 962 uint32_t status; 963 964 /* get references to current descriptor and S/W ring entry */ 965 rxdp = &rxq->rx_ring[rxq->rx_tail]; 966 rxep = &rxq->sw_ring[rxq->rx_tail]; 967 968 status = rxdp->qw1.lo.status; 969 /* check to make sure there is at least 1 packet to receive */ 970 if (!(status & rte_cpu_to_le_32(NGBE_RXD_STAT_DD))) 971 return 0; 972 973 /* 974 * Scan LOOK_AHEAD descriptors at a time to determine which descriptors 975 * reference packets that are ready to be received. 976 */ 977 for (i = 0; i < RTE_PMD_NGBE_RX_MAX_BURST; 978 i += LOOK_AHEAD, rxdp += LOOK_AHEAD, rxep += LOOK_AHEAD) { 979 /* Read desc statuses backwards to avoid race condition */ 980 for (j = 0; j < LOOK_AHEAD; j++) 981 s[j] = rte_le_to_cpu_32(rxdp[j].qw1.lo.status); 982 983 rte_atomic_thread_fence(__ATOMIC_ACQUIRE); 984 985 /* Compute how many status bits were set */ 986 for (nb_dd = 0; nb_dd < LOOK_AHEAD && 987 (s[nb_dd] & NGBE_RXD_STAT_DD); nb_dd++) 988 ; 989 990 for (j = 0; j < nb_dd; j++) 991 pkt_info[j] = rte_le_to_cpu_32(rxdp[j].qw0.dw0); 992 993 nb_rx += nb_dd; 994 995 /* Translate descriptor info to mbuf format */ 996 for (j = 0; j < nb_dd; ++j) { 997 mb = rxep[j].mbuf; 998 pkt_len = rte_le_to_cpu_16(rxdp[j].qw1.hi.len) - 999 rxq->crc_len; 1000 mb->data_len = pkt_len; 1001 mb->pkt_len = pkt_len; 1002 mb->vlan_tci = rte_le_to_cpu_16(rxdp[j].qw1.hi.tag); 1003 1004 /* convert descriptor fields to rte mbuf flags */ 1005 pkt_flags = rx_desc_status_to_pkt_flags(s[j], 1006 rxq->vlan_flags); 1007 pkt_flags |= rx_desc_error_to_pkt_flags(s[j]); 1008 mb->ol_flags = pkt_flags; 1009 mb->packet_type = 1010 ngbe_rxd_pkt_info_to_pkt_type(pkt_info[j], 1011 NGBE_PTID_MASK); 1012 } 1013 1014 /* Move mbuf pointers from the S/W ring to the stage */ 1015 for (j = 0; j < LOOK_AHEAD; ++j) 1016 rxq->rx_stage[i + j] = rxep[j].mbuf; 1017 1018 /* stop if all requested packets could not be received */ 1019 if (nb_dd != LOOK_AHEAD) 1020 break; 1021 } 1022 1023 /* clear software ring entries so we can cleanup correctly */ 1024 for (i = 0; i < nb_rx; ++i) 1025 rxq->sw_ring[rxq->rx_tail + i].mbuf = NULL; 1026 1027 return nb_rx; 1028 } 1029 1030 static inline int 1031 ngbe_rx_alloc_bufs(struct ngbe_rx_queue *rxq, bool reset_mbuf) 1032 { 1033 volatile struct ngbe_rx_desc *rxdp; 1034 struct ngbe_rx_entry *rxep; 1035 struct rte_mbuf *mb; 1036 uint16_t alloc_idx; 1037 __le64 dma_addr; 1038 int diag, i; 1039 1040 /* allocate buffers in bulk directly into the S/W ring */ 1041 alloc_idx = rxq->rx_free_trigger - (rxq->rx_free_thresh - 1); 1042 rxep = &rxq->sw_ring[alloc_idx]; 1043 diag = rte_mempool_get_bulk(rxq->mb_pool, (void *)rxep, 1044 rxq->rx_free_thresh); 1045 if (unlikely(diag != 0)) 1046 return -ENOMEM; 1047 1048 rxdp = &rxq->rx_ring[alloc_idx]; 1049 for (i = 0; i < rxq->rx_free_thresh; ++i) { 1050 /* populate the static rte mbuf fields */ 1051 mb = rxep[i].mbuf; 1052 if (reset_mbuf) 1053 mb->port = rxq->port_id; 1054 1055 rte_mbuf_refcnt_set(mb, 1); 1056 mb->data_off = RTE_PKTMBUF_HEADROOM; 1057 1058 /* populate the descriptors */ 1059 dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(mb)); 1060 NGBE_RXD_HDRADDR(&rxdp[i], 0); 1061 NGBE_RXD_PKTADDR(&rxdp[i], dma_addr); 1062 } 1063 1064 /* update state of internal queue structure */ 1065 rxq->rx_free_trigger = rxq->rx_free_trigger + rxq->rx_free_thresh; 1066 if (rxq->rx_free_trigger >= rxq->nb_rx_desc) 1067 rxq->rx_free_trigger = rxq->rx_free_thresh - 1; 1068 1069 /* no errors */ 1070 return 0; 1071 } 1072 1073 static inline uint16_t 1074 ngbe_rx_fill_from_stage(struct ngbe_rx_queue *rxq, struct rte_mbuf **rx_pkts, 1075 uint16_t nb_pkts) 1076 { 1077 struct rte_mbuf **stage = &rxq->rx_stage[rxq->rx_next_avail]; 1078 int i; 1079 1080 /* how many packets are ready to return? */ 1081 nb_pkts = (uint16_t)RTE_MIN(nb_pkts, rxq->rx_nb_avail); 1082 1083 /* copy mbuf pointers to the application's packet list */ 1084 for (i = 0; i < nb_pkts; ++i) 1085 rx_pkts[i] = stage[i]; 1086 1087 /* update internal queue state */ 1088 rxq->rx_nb_avail = (uint16_t)(rxq->rx_nb_avail - nb_pkts); 1089 rxq->rx_next_avail = (uint16_t)(rxq->rx_next_avail + nb_pkts); 1090 1091 return nb_pkts; 1092 } 1093 1094 static inline uint16_t 1095 ngbe_rx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, 1096 uint16_t nb_pkts) 1097 { 1098 struct ngbe_rx_queue *rxq = (struct ngbe_rx_queue *)rx_queue; 1099 struct rte_eth_dev *dev = &rte_eth_devices[rxq->port_id]; 1100 uint16_t nb_rx = 0; 1101 1102 /* Any previously recv'd pkts will be returned from the Rx stage */ 1103 if (rxq->rx_nb_avail) 1104 return ngbe_rx_fill_from_stage(rxq, rx_pkts, nb_pkts); 1105 1106 /* Scan the H/W ring for packets to receive */ 1107 nb_rx = (uint16_t)ngbe_rx_scan_hw_ring(rxq); 1108 1109 /* update internal queue state */ 1110 rxq->rx_next_avail = 0; 1111 rxq->rx_nb_avail = nb_rx; 1112 rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_rx); 1113 1114 /* if required, allocate new buffers to replenish descriptors */ 1115 if (rxq->rx_tail > rxq->rx_free_trigger) { 1116 uint16_t cur_free_trigger = rxq->rx_free_trigger; 1117 1118 if (ngbe_rx_alloc_bufs(rxq, true) != 0) { 1119 int i, j; 1120 1121 PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u " 1122 "queue_id=%u", (uint16_t)rxq->port_id, 1123 (uint16_t)rxq->queue_id); 1124 1125 dev->data->rx_mbuf_alloc_failed += 1126 rxq->rx_free_thresh; 1127 1128 /* 1129 * Need to rewind any previous receives if we cannot 1130 * allocate new buffers to replenish the old ones. 1131 */ 1132 rxq->rx_nb_avail = 0; 1133 rxq->rx_tail = (uint16_t)(rxq->rx_tail - nb_rx); 1134 for (i = 0, j = rxq->rx_tail; i < nb_rx; ++i, ++j) 1135 rxq->sw_ring[j].mbuf = rxq->rx_stage[i]; 1136 1137 return 0; 1138 } 1139 1140 /* update tail pointer */ 1141 rte_wmb(); 1142 ngbe_set32_relaxed(rxq->rdt_reg_addr, cur_free_trigger); 1143 } 1144 1145 if (rxq->rx_tail >= rxq->nb_rx_desc) 1146 rxq->rx_tail = 0; 1147 1148 /* received any packets this loop? */ 1149 if (rxq->rx_nb_avail) 1150 return ngbe_rx_fill_from_stage(rxq, rx_pkts, nb_pkts); 1151 1152 return 0; 1153 } 1154 1155 /* split requests into chunks of size RTE_PMD_NGBE_RX_MAX_BURST */ 1156 uint16_t 1157 ngbe_recv_pkts_bulk_alloc(void *rx_queue, struct rte_mbuf **rx_pkts, 1158 uint16_t nb_pkts) 1159 { 1160 uint16_t nb_rx; 1161 1162 if (unlikely(nb_pkts == 0)) 1163 return 0; 1164 1165 if (likely(nb_pkts <= RTE_PMD_NGBE_RX_MAX_BURST)) 1166 return ngbe_rx_recv_pkts(rx_queue, rx_pkts, nb_pkts); 1167 1168 /* request is relatively large, chunk it up */ 1169 nb_rx = 0; 1170 while (nb_pkts) { 1171 uint16_t ret, n; 1172 1173 n = (uint16_t)RTE_MIN(nb_pkts, RTE_PMD_NGBE_RX_MAX_BURST); 1174 ret = ngbe_rx_recv_pkts(rx_queue, &rx_pkts[nb_rx], n); 1175 nb_rx = (uint16_t)(nb_rx + ret); 1176 nb_pkts = (uint16_t)(nb_pkts - ret); 1177 if (ret < n) 1178 break; 1179 } 1180 1181 return nb_rx; 1182 } 1183 1184 uint16_t 1185 ngbe_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, 1186 uint16_t nb_pkts) 1187 { 1188 struct ngbe_rx_queue *rxq; 1189 volatile struct ngbe_rx_desc *rx_ring; 1190 volatile struct ngbe_rx_desc *rxdp; 1191 struct ngbe_rx_entry *sw_ring; 1192 struct ngbe_rx_entry *rxe; 1193 struct rte_mbuf *rxm; 1194 struct rte_mbuf *nmb; 1195 struct ngbe_rx_desc rxd; 1196 uint64_t dma_addr; 1197 uint32_t staterr; 1198 uint32_t pkt_info; 1199 uint16_t pkt_len; 1200 uint16_t rx_id; 1201 uint16_t nb_rx; 1202 uint16_t nb_hold; 1203 uint64_t pkt_flags; 1204 1205 nb_rx = 0; 1206 nb_hold = 0; 1207 rxq = rx_queue; 1208 rx_id = rxq->rx_tail; 1209 rx_ring = rxq->rx_ring; 1210 sw_ring = rxq->sw_ring; 1211 struct rte_eth_dev *dev = &rte_eth_devices[rxq->port_id]; 1212 while (nb_rx < nb_pkts) { 1213 /* 1214 * The order of operations here is important as the DD status 1215 * bit must not be read after any other descriptor fields. 1216 * rx_ring and rxdp are pointing to volatile data so the order 1217 * of accesses cannot be reordered by the compiler. If they were 1218 * not volatile, they could be reordered which could lead to 1219 * using invalid descriptor fields when read from rxd. 1220 */ 1221 rxdp = &rx_ring[rx_id]; 1222 staterr = rxdp->qw1.lo.status; 1223 if (!(staterr & rte_cpu_to_le_32(NGBE_RXD_STAT_DD))) 1224 break; 1225 rxd = *rxdp; 1226 1227 /* 1228 * End of packet. 1229 * 1230 * If the NGBE_RXD_STAT_EOP flag is not set, the Rx packet 1231 * is likely to be invalid and to be dropped by the various 1232 * validation checks performed by the network stack. 1233 * 1234 * Allocate a new mbuf to replenish the RX ring descriptor. 1235 * If the allocation fails: 1236 * - arrange for that Rx descriptor to be the first one 1237 * being parsed the next time the receive function is 1238 * invoked [on the same queue]. 1239 * 1240 * - Stop parsing the Rx ring and return immediately. 1241 * 1242 * This policy do not drop the packet received in the Rx 1243 * descriptor for which the allocation of a new mbuf failed. 1244 * Thus, it allows that packet to be later retrieved if 1245 * mbuf have been freed in the mean time. 1246 * As a side effect, holding Rx descriptors instead of 1247 * systematically giving them back to the NIC may lead to 1248 * Rx ring exhaustion situations. 1249 * However, the NIC can gracefully prevent such situations 1250 * to happen by sending specific "back-pressure" flow control 1251 * frames to its peer(s). 1252 */ 1253 PMD_RX_LOG(DEBUG, 1254 "port_id=%u queue_id=%u rx_id=%u ext_err_stat=0x%08x pkt_len=%u", 1255 (uint16_t)rxq->port_id, (uint16_t)rxq->queue_id, 1256 (uint16_t)rx_id, (uint32_t)staterr, 1257 (uint16_t)rte_le_to_cpu_16(rxd.qw1.hi.len)); 1258 1259 nmb = rte_mbuf_raw_alloc(rxq->mb_pool); 1260 if (nmb == NULL) { 1261 PMD_RX_LOG(DEBUG, 1262 "Rx mbuf alloc failed port_id=%u queue_id=%u", 1263 (uint16_t)rxq->port_id, 1264 (uint16_t)rxq->queue_id); 1265 dev->data->rx_mbuf_alloc_failed++; 1266 break; 1267 } 1268 1269 nb_hold++; 1270 rxe = &sw_ring[rx_id]; 1271 rx_id++; 1272 if (rx_id == rxq->nb_rx_desc) 1273 rx_id = 0; 1274 1275 /* Prefetch next mbuf while processing current one. */ 1276 rte_ngbe_prefetch(sw_ring[rx_id].mbuf); 1277 1278 /* 1279 * When next Rx descriptor is on a cache-line boundary, 1280 * prefetch the next 4 Rx descriptors and the next 8 pointers 1281 * to mbufs. 1282 */ 1283 if ((rx_id & 0x3) == 0) { 1284 rte_ngbe_prefetch(&rx_ring[rx_id]); 1285 rte_ngbe_prefetch(&sw_ring[rx_id]); 1286 } 1287 1288 rxm = rxe->mbuf; 1289 rxe->mbuf = nmb; 1290 dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb)); 1291 NGBE_RXD_HDRADDR(rxdp, 0); 1292 NGBE_RXD_PKTADDR(rxdp, dma_addr); 1293 1294 /* 1295 * Initialize the returned mbuf. 1296 * 1) setup generic mbuf fields: 1297 * - number of segments, 1298 * - next segment, 1299 * - packet length, 1300 * - Rx port identifier. 1301 * 2) integrate hardware offload data, if any: 1302 * - IP checksum flag, 1303 * - VLAN TCI, if any, 1304 * - error flags. 1305 */ 1306 pkt_len = (uint16_t)(rte_le_to_cpu_16(rxd.qw1.hi.len) - 1307 rxq->crc_len); 1308 rxm->data_off = RTE_PKTMBUF_HEADROOM; 1309 rte_packet_prefetch((char *)rxm->buf_addr + rxm->data_off); 1310 rxm->nb_segs = 1; 1311 rxm->next = NULL; 1312 rxm->pkt_len = pkt_len; 1313 rxm->data_len = pkt_len; 1314 rxm->port = rxq->port_id; 1315 1316 pkt_info = rte_le_to_cpu_32(rxd.qw0.dw0); 1317 /* Only valid if RTE_MBUF_F_RX_VLAN set in pkt_flags */ 1318 rxm->vlan_tci = rte_le_to_cpu_16(rxd.qw1.hi.tag); 1319 1320 pkt_flags = rx_desc_status_to_pkt_flags(staterr, 1321 rxq->vlan_flags); 1322 pkt_flags |= rx_desc_error_to_pkt_flags(staterr); 1323 rxm->ol_flags = pkt_flags; 1324 rxm->packet_type = ngbe_rxd_pkt_info_to_pkt_type(pkt_info, 1325 NGBE_PTID_MASK); 1326 1327 /* 1328 * Store the mbuf address into the next entry of the array 1329 * of returned packets. 1330 */ 1331 rx_pkts[nb_rx++] = rxm; 1332 } 1333 rxq->rx_tail = rx_id; 1334 1335 /* 1336 * If the number of free Rx descriptors is greater than the Rx free 1337 * threshold of the queue, advance the Receive Descriptor Tail (RDT) 1338 * register. 1339 * Update the RDT with the value of the last processed Rx descriptor 1340 * minus 1, to guarantee that the RDT register is never equal to the 1341 * RDH register, which creates a "full" ring situation from the 1342 * hardware point of view... 1343 */ 1344 nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold); 1345 if (nb_hold > rxq->rx_free_thresh) { 1346 PMD_RX_LOG(DEBUG, 1347 "port_id=%u queue_id=%u rx_tail=%u nb_hold=%u nb_rx=%u", 1348 (uint16_t)rxq->port_id, (uint16_t)rxq->queue_id, 1349 (uint16_t)rx_id, (uint16_t)nb_hold, 1350 (uint16_t)nb_rx); 1351 rx_id = (uint16_t)((rx_id == 0) ? 1352 (rxq->nb_rx_desc - 1) : (rx_id - 1)); 1353 ngbe_set32(rxq->rdt_reg_addr, rx_id); 1354 nb_hold = 0; 1355 } 1356 rxq->nb_rx_hold = nb_hold; 1357 return nb_rx; 1358 } 1359 1360 /** 1361 * ngbe_fill_cluster_head_buf - fill the first mbuf of the returned packet 1362 * 1363 * Fill the following info in the HEAD buffer of the Rx cluster: 1364 * - RX port identifier 1365 * - hardware offload data, if any: 1366 * - IP checksum flag 1367 * - VLAN TCI, if any 1368 * - error flags 1369 * @head HEAD of the packet cluster 1370 * @desc HW descriptor to get data from 1371 * @rxq Pointer to the Rx queue 1372 */ 1373 static inline void 1374 ngbe_fill_cluster_head_buf(struct rte_mbuf *head, struct ngbe_rx_desc *desc, 1375 struct ngbe_rx_queue *rxq, uint32_t staterr) 1376 { 1377 uint32_t pkt_info; 1378 uint64_t pkt_flags; 1379 1380 head->port = rxq->port_id; 1381 1382 /* The vlan_tci field is only valid when RTE_MBUF_F_RX_VLAN is 1383 * set in the pkt_flags field. 1384 */ 1385 head->vlan_tci = rte_le_to_cpu_16(desc->qw1.hi.tag); 1386 pkt_info = rte_le_to_cpu_32(desc->qw0.dw0); 1387 pkt_flags = rx_desc_status_to_pkt_flags(staterr, rxq->vlan_flags); 1388 pkt_flags |= rx_desc_error_to_pkt_flags(staterr); 1389 head->ol_flags = pkt_flags; 1390 head->packet_type = ngbe_rxd_pkt_info_to_pkt_type(pkt_info, 1391 NGBE_PTID_MASK); 1392 } 1393 1394 /** 1395 * ngbe_recv_pkts_sc - receive handler for scatter case. 1396 * 1397 * @rx_queue Rx queue handle 1398 * @rx_pkts table of received packets 1399 * @nb_pkts size of rx_pkts table 1400 * @bulk_alloc if TRUE bulk allocation is used for a HW ring refilling 1401 * 1402 * Returns the number of received packets/clusters (according to the "bulk 1403 * receive" interface). 1404 */ 1405 static inline uint16_t 1406 ngbe_recv_pkts_sc(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts, 1407 bool bulk_alloc) 1408 { 1409 struct ngbe_rx_queue *rxq = rx_queue; 1410 struct rte_eth_dev *dev = &rte_eth_devices[rxq->port_id]; 1411 volatile struct ngbe_rx_desc *rx_ring = rxq->rx_ring; 1412 struct ngbe_rx_entry *sw_ring = rxq->sw_ring; 1413 struct ngbe_scattered_rx_entry *sw_sc_ring = rxq->sw_sc_ring; 1414 uint16_t rx_id = rxq->rx_tail; 1415 uint16_t nb_rx = 0; 1416 uint16_t nb_hold = rxq->nb_rx_hold; 1417 uint16_t prev_id = rxq->rx_tail; 1418 1419 while (nb_rx < nb_pkts) { 1420 bool eop; 1421 struct ngbe_rx_entry *rxe; 1422 struct ngbe_scattered_rx_entry *sc_entry; 1423 struct ngbe_scattered_rx_entry *next_sc_entry = NULL; 1424 struct ngbe_rx_entry *next_rxe = NULL; 1425 struct rte_mbuf *first_seg; 1426 struct rte_mbuf *rxm; 1427 struct rte_mbuf *nmb = NULL; 1428 struct ngbe_rx_desc rxd; 1429 uint16_t data_len; 1430 uint16_t next_id; 1431 volatile struct ngbe_rx_desc *rxdp; 1432 uint32_t staterr; 1433 1434 next_desc: 1435 rxdp = &rx_ring[rx_id]; 1436 staterr = rte_le_to_cpu_32(rxdp->qw1.lo.status); 1437 1438 if (!(staterr & NGBE_RXD_STAT_DD)) 1439 break; 1440 1441 rxd = *rxdp; 1442 1443 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u " 1444 "staterr=0x%x data_len=%u", 1445 rxq->port_id, rxq->queue_id, rx_id, staterr, 1446 rte_le_to_cpu_16(rxd.qw1.hi.len)); 1447 1448 if (!bulk_alloc) { 1449 nmb = rte_mbuf_raw_alloc(rxq->mb_pool); 1450 if (nmb == NULL) { 1451 PMD_RX_LOG(DEBUG, "Rx mbuf alloc failed " 1452 "port_id=%u queue_id=%u", 1453 rxq->port_id, rxq->queue_id); 1454 1455 dev->data->rx_mbuf_alloc_failed++; 1456 break; 1457 } 1458 } else if (nb_hold > rxq->rx_free_thresh) { 1459 uint16_t next_rdt = rxq->rx_free_trigger; 1460 1461 if (!ngbe_rx_alloc_bufs(rxq, false)) { 1462 rte_wmb(); 1463 ngbe_set32_relaxed(rxq->rdt_reg_addr, 1464 next_rdt); 1465 nb_hold -= rxq->rx_free_thresh; 1466 } else { 1467 PMD_RX_LOG(DEBUG, "Rx bulk alloc failed " 1468 "port_id=%u queue_id=%u", 1469 rxq->port_id, rxq->queue_id); 1470 1471 dev->data->rx_mbuf_alloc_failed++; 1472 break; 1473 } 1474 } 1475 1476 nb_hold++; 1477 rxe = &sw_ring[rx_id]; 1478 eop = staterr & NGBE_RXD_STAT_EOP; 1479 1480 next_id = rx_id + 1; 1481 if (next_id == rxq->nb_rx_desc) 1482 next_id = 0; 1483 1484 /* Prefetch next mbuf while processing current one. */ 1485 rte_ngbe_prefetch(sw_ring[next_id].mbuf); 1486 1487 /* 1488 * When next Rx descriptor is on a cache-line boundary, 1489 * prefetch the next 4 RX descriptors and the next 4 pointers 1490 * to mbufs. 1491 */ 1492 if ((next_id & 0x3) == 0) { 1493 rte_ngbe_prefetch(&rx_ring[next_id]); 1494 rte_ngbe_prefetch(&sw_ring[next_id]); 1495 } 1496 1497 rxm = rxe->mbuf; 1498 1499 if (!bulk_alloc) { 1500 __le64 dma = 1501 rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb)); 1502 /* 1503 * Update Rx descriptor with the physical address of the 1504 * new data buffer of the new allocated mbuf. 1505 */ 1506 rxe->mbuf = nmb; 1507 1508 rxm->data_off = RTE_PKTMBUF_HEADROOM; 1509 NGBE_RXD_HDRADDR(rxdp, 0); 1510 NGBE_RXD_PKTADDR(rxdp, dma); 1511 } else { 1512 rxe->mbuf = NULL; 1513 } 1514 1515 /* 1516 * Set data length & data buffer address of mbuf. 1517 */ 1518 data_len = rte_le_to_cpu_16(rxd.qw1.hi.len); 1519 rxm->data_len = data_len; 1520 1521 if (!eop) { 1522 uint16_t nextp_id; 1523 1524 nextp_id = next_id; 1525 next_sc_entry = &sw_sc_ring[nextp_id]; 1526 next_rxe = &sw_ring[nextp_id]; 1527 rte_ngbe_prefetch(next_rxe); 1528 } 1529 1530 sc_entry = &sw_sc_ring[rx_id]; 1531 first_seg = sc_entry->fbuf; 1532 sc_entry->fbuf = NULL; 1533 1534 /* 1535 * If this is the first buffer of the received packet, 1536 * set the pointer to the first mbuf of the packet and 1537 * initialize its context. 1538 * Otherwise, update the total length and the number of segments 1539 * of the current scattered packet, and update the pointer to 1540 * the last mbuf of the current packet. 1541 */ 1542 if (first_seg == NULL) { 1543 first_seg = rxm; 1544 first_seg->pkt_len = data_len; 1545 first_seg->nb_segs = 1; 1546 } else { 1547 first_seg->pkt_len += data_len; 1548 first_seg->nb_segs++; 1549 } 1550 1551 prev_id = rx_id; 1552 rx_id = next_id; 1553 1554 /* 1555 * If this is not the last buffer of the received packet, update 1556 * the pointer to the first mbuf at the NEXTP entry in the 1557 * sw_sc_ring and continue to parse the Rx ring. 1558 */ 1559 if (!eop && next_rxe) { 1560 rxm->next = next_rxe->mbuf; 1561 next_sc_entry->fbuf = first_seg; 1562 goto next_desc; 1563 } 1564 1565 /* Initialize the first mbuf of the returned packet */ 1566 ngbe_fill_cluster_head_buf(first_seg, &rxd, rxq, staterr); 1567 1568 /* Deal with the case, when HW CRC srip is disabled. */ 1569 first_seg->pkt_len -= rxq->crc_len; 1570 if (unlikely(rxm->data_len <= rxq->crc_len)) { 1571 struct rte_mbuf *lp; 1572 1573 for (lp = first_seg; lp->next != rxm; lp = lp->next) 1574 ; 1575 1576 first_seg->nb_segs--; 1577 lp->data_len -= rxq->crc_len - rxm->data_len; 1578 lp->next = NULL; 1579 rte_pktmbuf_free_seg(rxm); 1580 } else { 1581 rxm->data_len -= rxq->crc_len; 1582 } 1583 1584 /* Prefetch data of first segment, if configured to do so. */ 1585 rte_packet_prefetch((char *)first_seg->buf_addr + 1586 first_seg->data_off); 1587 1588 /* 1589 * Store the mbuf address into the next entry of the array 1590 * of returned packets. 1591 */ 1592 rx_pkts[nb_rx++] = first_seg; 1593 } 1594 1595 /* 1596 * Record index of the next Rx descriptor to probe. 1597 */ 1598 rxq->rx_tail = rx_id; 1599 1600 /* 1601 * If the number of free Rx descriptors is greater than the Rx free 1602 * threshold of the queue, advance the Receive Descriptor Tail (RDT) 1603 * register. 1604 * Update the RDT with the value of the last processed Rx descriptor 1605 * minus 1, to guarantee that the RDT register is never equal to the 1606 * RDH register, which creates a "full" ring situation from the 1607 * hardware point of view... 1608 */ 1609 if (!bulk_alloc && nb_hold > rxq->rx_free_thresh) { 1610 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u " 1611 "nb_hold=%u nb_rx=%u", 1612 rxq->port_id, rxq->queue_id, rx_id, nb_hold, nb_rx); 1613 1614 rte_wmb(); 1615 ngbe_set32_relaxed(rxq->rdt_reg_addr, prev_id); 1616 nb_hold = 0; 1617 } 1618 1619 rxq->nb_rx_hold = nb_hold; 1620 return nb_rx; 1621 } 1622 1623 uint16_t 1624 ngbe_recv_pkts_sc_single_alloc(void *rx_queue, struct rte_mbuf **rx_pkts, 1625 uint16_t nb_pkts) 1626 { 1627 return ngbe_recv_pkts_sc(rx_queue, rx_pkts, nb_pkts, false); 1628 } 1629 1630 uint16_t 1631 ngbe_recv_pkts_sc_bulk_alloc(void *rx_queue, struct rte_mbuf **rx_pkts, 1632 uint16_t nb_pkts) 1633 { 1634 return ngbe_recv_pkts_sc(rx_queue, rx_pkts, nb_pkts, true); 1635 } 1636 1637 /********************************************************************* 1638 * 1639 * Queue management functions 1640 * 1641 **********************************************************************/ 1642 1643 static void 1644 ngbe_tx_queue_release_mbufs(struct ngbe_tx_queue *txq) 1645 { 1646 unsigned int i; 1647 1648 if (txq->sw_ring != NULL) { 1649 for (i = 0; i < txq->nb_tx_desc; i++) { 1650 if (txq->sw_ring[i].mbuf != NULL) { 1651 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf); 1652 txq->sw_ring[i].mbuf = NULL; 1653 } 1654 } 1655 } 1656 } 1657 1658 static void 1659 ngbe_tx_free_swring(struct ngbe_tx_queue *txq) 1660 { 1661 if (txq != NULL) 1662 rte_free(txq->sw_ring); 1663 } 1664 1665 static void 1666 ngbe_tx_queue_release(struct ngbe_tx_queue *txq) 1667 { 1668 if (txq != NULL) { 1669 if (txq->ops != NULL) { 1670 txq->ops->release_mbufs(txq); 1671 txq->ops->free_swring(txq); 1672 } 1673 rte_free(txq); 1674 } 1675 } 1676 1677 void 1678 ngbe_dev_tx_queue_release(struct rte_eth_dev *dev, uint16_t qid) 1679 { 1680 ngbe_tx_queue_release(dev->data->tx_queues[qid]); 1681 } 1682 1683 /* (Re)set dynamic ngbe_tx_queue fields to defaults */ 1684 static void 1685 ngbe_reset_tx_queue(struct ngbe_tx_queue *txq) 1686 { 1687 static const struct ngbe_tx_desc zeroed_desc = {0}; 1688 struct ngbe_tx_entry *txe = txq->sw_ring; 1689 uint16_t prev, i; 1690 1691 /* Zero out HW ring memory */ 1692 for (i = 0; i < txq->nb_tx_desc; i++) 1693 txq->tx_ring[i] = zeroed_desc; 1694 1695 /* Initialize SW ring entries */ 1696 prev = (uint16_t)(txq->nb_tx_desc - 1); 1697 for (i = 0; i < txq->nb_tx_desc; i++) { 1698 /* the ring can also be modified by hardware */ 1699 volatile struct ngbe_tx_desc *txd = &txq->tx_ring[i]; 1700 1701 txd->dw3 = rte_cpu_to_le_32(NGBE_TXD_DD); 1702 txe[i].mbuf = NULL; 1703 txe[i].last_id = i; 1704 txe[prev].next_id = i; 1705 prev = i; 1706 } 1707 1708 txq->tx_next_dd = (uint16_t)(txq->tx_free_thresh - 1); 1709 txq->tx_tail = 0; 1710 1711 /* 1712 * Always allow 1 descriptor to be un-allocated to avoid 1713 * a H/W race condition 1714 */ 1715 txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1); 1716 txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1); 1717 txq->ctx_curr = 0; 1718 memset((void *)&txq->ctx_cache, 0, 1719 NGBE_CTX_NUM * sizeof(struct ngbe_ctx_info)); 1720 } 1721 1722 static const struct ngbe_txq_ops def_txq_ops = { 1723 .release_mbufs = ngbe_tx_queue_release_mbufs, 1724 .free_swring = ngbe_tx_free_swring, 1725 .reset = ngbe_reset_tx_queue, 1726 }; 1727 1728 /* Takes an ethdev and a queue and sets up the tx function to be used based on 1729 * the queue parameters. Used in tx_queue_setup by primary process and then 1730 * in dev_init by secondary process when attaching to an existing ethdev. 1731 */ 1732 void 1733 ngbe_set_tx_function(struct rte_eth_dev *dev, struct ngbe_tx_queue *txq) 1734 { 1735 /* Use a simple Tx queue (no offloads, no multi segs) if possible */ 1736 if (txq->offloads == 0 && 1737 txq->tx_free_thresh >= RTE_PMD_NGBE_TX_MAX_BURST) { 1738 PMD_INIT_LOG(DEBUG, "Using simple tx code path"); 1739 dev->tx_pkt_burst = ngbe_xmit_pkts_simple; 1740 dev->tx_pkt_prepare = NULL; 1741 } else { 1742 PMD_INIT_LOG(DEBUG, "Using full-featured tx code path"); 1743 PMD_INIT_LOG(DEBUG, 1744 " - offloads = 0x%" PRIx64, 1745 txq->offloads); 1746 PMD_INIT_LOG(DEBUG, 1747 " - tx_free_thresh = %lu [RTE_PMD_NGBE_TX_MAX_BURST=%lu]", 1748 (unsigned long)txq->tx_free_thresh, 1749 (unsigned long)RTE_PMD_NGBE_TX_MAX_BURST); 1750 dev->tx_pkt_burst = ngbe_xmit_pkts; 1751 dev->tx_pkt_prepare = ngbe_prep_pkts; 1752 } 1753 } 1754 1755 static const struct { 1756 eth_tx_burst_t pkt_burst; 1757 const char *info; 1758 } ngbe_tx_burst_infos[] = { 1759 { ngbe_xmit_pkts_simple, "Scalar Simple"}, 1760 { ngbe_xmit_pkts, "Scalar"}, 1761 }; 1762 1763 int 1764 ngbe_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id, 1765 struct rte_eth_burst_mode *mode) 1766 { 1767 eth_tx_burst_t pkt_burst = dev->tx_pkt_burst; 1768 int ret = -EINVAL; 1769 unsigned int i; 1770 1771 for (i = 0; i < RTE_DIM(ngbe_tx_burst_infos); ++i) { 1772 if (pkt_burst == ngbe_tx_burst_infos[i].pkt_burst) { 1773 snprintf(mode->info, sizeof(mode->info), "%s", 1774 ngbe_tx_burst_infos[i].info); 1775 ret = 0; 1776 break; 1777 } 1778 } 1779 1780 return ret; 1781 } 1782 1783 uint64_t 1784 ngbe_get_tx_port_offloads(struct rte_eth_dev *dev) 1785 { 1786 uint64_t tx_offload_capa; 1787 struct ngbe_hw *hw = ngbe_dev_hw(dev); 1788 1789 tx_offload_capa = 1790 RTE_ETH_TX_OFFLOAD_VLAN_INSERT | 1791 RTE_ETH_TX_OFFLOAD_IPV4_CKSUM | 1792 RTE_ETH_TX_OFFLOAD_UDP_CKSUM | 1793 RTE_ETH_TX_OFFLOAD_TCP_CKSUM | 1794 RTE_ETH_TX_OFFLOAD_SCTP_CKSUM | 1795 RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM | 1796 RTE_ETH_TX_OFFLOAD_TCP_TSO | 1797 RTE_ETH_TX_OFFLOAD_UDP_TSO | 1798 RTE_ETH_TX_OFFLOAD_UDP_TNL_TSO | 1799 RTE_ETH_TX_OFFLOAD_IP_TNL_TSO | 1800 RTE_ETH_TX_OFFLOAD_IPIP_TNL_TSO | 1801 RTE_ETH_TX_OFFLOAD_MULTI_SEGS; 1802 1803 if (hw->is_pf) 1804 tx_offload_capa |= RTE_ETH_TX_OFFLOAD_QINQ_INSERT; 1805 1806 return tx_offload_capa; 1807 } 1808 1809 int 1810 ngbe_dev_tx_queue_setup(struct rte_eth_dev *dev, 1811 uint16_t queue_idx, 1812 uint16_t nb_desc, 1813 unsigned int socket_id, 1814 const struct rte_eth_txconf *tx_conf) 1815 { 1816 const struct rte_memzone *tz; 1817 struct ngbe_tx_queue *txq; 1818 struct ngbe_hw *hw; 1819 uint16_t tx_free_thresh; 1820 uint64_t offloads; 1821 1822 PMD_INIT_FUNC_TRACE(); 1823 hw = ngbe_dev_hw(dev); 1824 1825 offloads = tx_conf->offloads | dev->data->dev_conf.txmode.offloads; 1826 1827 /* 1828 * The Tx descriptor ring will be cleaned after txq->tx_free_thresh 1829 * descriptors are used or if the number of descriptors required 1830 * to transmit a packet is greater than the number of free Tx 1831 * descriptors. 1832 * One descriptor in the Tx ring is used as a sentinel to avoid a 1833 * H/W race condition, hence the maximum threshold constraints. 1834 * When set to zero use default values. 1835 */ 1836 tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ? 1837 tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH); 1838 if (tx_free_thresh >= (nb_desc - 3)) { 1839 PMD_INIT_LOG(ERR, 1840 "tx_free_thresh must be less than the number of TX descriptors minus 3. (tx_free_thresh=%u port=%d queue=%d)", 1841 (unsigned int)tx_free_thresh, 1842 (int)dev->data->port_id, (int)queue_idx); 1843 return -(EINVAL); 1844 } 1845 1846 if (nb_desc % tx_free_thresh != 0) { 1847 PMD_INIT_LOG(ERR, 1848 "tx_free_thresh must be a divisor of the number of Tx descriptors. (tx_free_thresh=%u port=%d queue=%d)", 1849 (unsigned int)tx_free_thresh, 1850 (int)dev->data->port_id, (int)queue_idx); 1851 return -(EINVAL); 1852 } 1853 1854 /* Free memory prior to re-allocation if needed... */ 1855 if (dev->data->tx_queues[queue_idx] != NULL) { 1856 ngbe_tx_queue_release(dev->data->tx_queues[queue_idx]); 1857 dev->data->tx_queues[queue_idx] = NULL; 1858 } 1859 1860 /* First allocate the Tx queue data structure */ 1861 txq = rte_zmalloc_socket("ethdev Tx queue", 1862 sizeof(struct ngbe_tx_queue), 1863 RTE_CACHE_LINE_SIZE, socket_id); 1864 if (txq == NULL) 1865 return -ENOMEM; 1866 1867 /* 1868 * Allocate Tx ring hardware descriptors. A memzone large enough to 1869 * handle the maximum ring size is allocated in order to allow for 1870 * resizing in later calls to the queue setup function. 1871 */ 1872 tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx, 1873 sizeof(struct ngbe_tx_desc) * NGBE_RING_DESC_MAX, 1874 NGBE_ALIGN, socket_id); 1875 if (tz == NULL) { 1876 ngbe_tx_queue_release(txq); 1877 return -ENOMEM; 1878 } 1879 1880 txq->nb_tx_desc = nb_desc; 1881 txq->tx_free_thresh = tx_free_thresh; 1882 txq->pthresh = tx_conf->tx_thresh.pthresh; 1883 txq->hthresh = tx_conf->tx_thresh.hthresh; 1884 txq->wthresh = tx_conf->tx_thresh.wthresh; 1885 txq->queue_id = queue_idx; 1886 txq->reg_idx = queue_idx; 1887 txq->port_id = dev->data->port_id; 1888 txq->offloads = offloads; 1889 txq->ops = &def_txq_ops; 1890 txq->tx_deferred_start = tx_conf->tx_deferred_start; 1891 1892 txq->tdt_reg_addr = NGBE_REG_ADDR(hw, NGBE_TXWP(txq->reg_idx)); 1893 txq->tdc_reg_addr = NGBE_REG_ADDR(hw, NGBE_TXCFG(txq->reg_idx)); 1894 1895 txq->tx_ring_phys_addr = TMZ_PADDR(tz); 1896 txq->tx_ring = (struct ngbe_tx_desc *)TMZ_VADDR(tz); 1897 1898 /* Allocate software ring */ 1899 txq->sw_ring = rte_zmalloc_socket("txq->sw_ring", 1900 sizeof(struct ngbe_tx_entry) * nb_desc, 1901 RTE_CACHE_LINE_SIZE, socket_id); 1902 if (txq->sw_ring == NULL) { 1903 ngbe_tx_queue_release(txq); 1904 return -ENOMEM; 1905 } 1906 PMD_INIT_LOG(DEBUG, 1907 "sw_ring=%p hw_ring=%p dma_addr=0x%" PRIx64, 1908 txq->sw_ring, txq->tx_ring, txq->tx_ring_phys_addr); 1909 1910 /* set up scalar Tx function as appropriate */ 1911 ngbe_set_tx_function(dev, txq); 1912 1913 txq->ops->reset(txq); 1914 1915 dev->data->tx_queues[queue_idx] = txq; 1916 1917 return 0; 1918 } 1919 1920 /** 1921 * ngbe_free_sc_cluster - free the not-yet-completed scattered cluster 1922 * 1923 * The "next" pointer of the last segment of (not-yet-completed) RSC clusters 1924 * in the sw_sc_ring is not set to NULL but rather points to the next 1925 * mbuf of this RSC aggregation (that has not been completed yet and still 1926 * resides on the HW ring). So, instead of calling for rte_pktmbuf_free() we 1927 * will just free first "nb_segs" segments of the cluster explicitly by calling 1928 * an rte_pktmbuf_free_seg(). 1929 * 1930 * @m scattered cluster head 1931 */ 1932 static void 1933 ngbe_free_sc_cluster(struct rte_mbuf *m) 1934 { 1935 uint16_t i, nb_segs = m->nb_segs; 1936 struct rte_mbuf *next_seg; 1937 1938 for (i = 0; i < nb_segs; i++) { 1939 next_seg = m->next; 1940 rte_pktmbuf_free_seg(m); 1941 m = next_seg; 1942 } 1943 } 1944 1945 static void 1946 ngbe_rx_queue_release_mbufs(struct ngbe_rx_queue *rxq) 1947 { 1948 unsigned int i; 1949 1950 if (rxq->sw_ring != NULL) { 1951 for (i = 0; i < rxq->nb_rx_desc; i++) { 1952 if (rxq->sw_ring[i].mbuf != NULL) { 1953 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf); 1954 rxq->sw_ring[i].mbuf = NULL; 1955 } 1956 } 1957 for (i = 0; i < rxq->rx_nb_avail; ++i) { 1958 struct rte_mbuf *mb; 1959 1960 mb = rxq->rx_stage[rxq->rx_next_avail + i]; 1961 rte_pktmbuf_free_seg(mb); 1962 } 1963 rxq->rx_nb_avail = 0; 1964 } 1965 1966 if (rxq->sw_sc_ring != NULL) 1967 for (i = 0; i < rxq->nb_rx_desc; i++) 1968 if (rxq->sw_sc_ring[i].fbuf != NULL) { 1969 ngbe_free_sc_cluster(rxq->sw_sc_ring[i].fbuf); 1970 rxq->sw_sc_ring[i].fbuf = NULL; 1971 } 1972 } 1973 1974 static void 1975 ngbe_rx_queue_release(struct ngbe_rx_queue *rxq) 1976 { 1977 if (rxq != NULL) { 1978 ngbe_rx_queue_release_mbufs(rxq); 1979 rte_free(rxq->sw_ring); 1980 rte_free(rxq->sw_sc_ring); 1981 rte_free(rxq); 1982 } 1983 } 1984 1985 void 1986 ngbe_dev_rx_queue_release(struct rte_eth_dev *dev, uint16_t qid) 1987 { 1988 ngbe_rx_queue_release(dev->data->rx_queues[qid]); 1989 } 1990 1991 /* 1992 * Check if Rx Burst Bulk Alloc function can be used. 1993 * Return 1994 * 0: the preconditions are satisfied and the bulk allocation function 1995 * can be used. 1996 * -EINVAL: the preconditions are NOT satisfied and the default Rx burst 1997 * function must be used. 1998 */ 1999 static inline int 2000 check_rx_burst_bulk_alloc_preconditions(struct ngbe_rx_queue *rxq) 2001 { 2002 int ret = 0; 2003 2004 /* 2005 * Make sure the following pre-conditions are satisfied: 2006 * rxq->rx_free_thresh >= RTE_PMD_NGBE_RX_MAX_BURST 2007 * rxq->rx_free_thresh < rxq->nb_rx_desc 2008 * (rxq->nb_rx_desc % rxq->rx_free_thresh) == 0 2009 * Scattered packets are not supported. This should be checked 2010 * outside of this function. 2011 */ 2012 if (rxq->rx_free_thresh < RTE_PMD_NGBE_RX_MAX_BURST) { 2013 PMD_INIT_LOG(DEBUG, 2014 "Rx Burst Bulk Alloc Preconditions: rxq->rx_free_thresh=%d, RTE_PMD_NGBE_RX_MAX_BURST=%d", 2015 rxq->rx_free_thresh, RTE_PMD_NGBE_RX_MAX_BURST); 2016 ret = -EINVAL; 2017 } else if (rxq->rx_free_thresh >= rxq->nb_rx_desc) { 2018 PMD_INIT_LOG(DEBUG, 2019 "Rx Burst Bulk Alloc Preconditions: rxq->rx_free_thresh=%d, rxq->nb_rx_desc=%d", 2020 rxq->rx_free_thresh, rxq->nb_rx_desc); 2021 ret = -EINVAL; 2022 } else if ((rxq->nb_rx_desc % rxq->rx_free_thresh) != 0) { 2023 PMD_INIT_LOG(DEBUG, 2024 "Rx Burst Bulk Alloc Preconditions: rxq->nb_rx_desc=%d, rxq->rx_free_thresh=%d", 2025 rxq->nb_rx_desc, rxq->rx_free_thresh); 2026 ret = -EINVAL; 2027 } 2028 2029 return ret; 2030 } 2031 2032 /* Reset dynamic ngbe_rx_queue fields back to defaults */ 2033 static void 2034 ngbe_reset_rx_queue(struct ngbe_adapter *adapter, struct ngbe_rx_queue *rxq) 2035 { 2036 static const struct ngbe_rx_desc zeroed_desc = { 2037 {{0}, {0} }, {{0}, {0} } }; 2038 unsigned int i; 2039 uint16_t len = rxq->nb_rx_desc; 2040 2041 /* 2042 * By default, the Rx queue setup function allocates enough memory for 2043 * NGBE_RING_DESC_MAX. The Rx Burst bulk allocation function requires 2044 * extra memory at the end of the descriptor ring to be zero'd out. 2045 */ 2046 if (adapter->rx_bulk_alloc_allowed) 2047 /* zero out extra memory */ 2048 len += RTE_PMD_NGBE_RX_MAX_BURST; 2049 2050 /* 2051 * Zero out HW ring memory. Zero out extra memory at the end of 2052 * the H/W ring so look-ahead logic in Rx Burst bulk alloc function 2053 * reads extra memory as zeros. 2054 */ 2055 for (i = 0; i < len; i++) 2056 rxq->rx_ring[i] = zeroed_desc; 2057 2058 /* 2059 * initialize extra software ring entries. Space for these extra 2060 * entries is always allocated 2061 */ 2062 memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf)); 2063 for (i = rxq->nb_rx_desc; i < len; ++i) 2064 rxq->sw_ring[i].mbuf = &rxq->fake_mbuf; 2065 2066 rxq->rx_nb_avail = 0; 2067 rxq->rx_next_avail = 0; 2068 rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1); 2069 rxq->rx_tail = 0; 2070 rxq->nb_rx_hold = 0; 2071 rxq->pkt_first_seg = NULL; 2072 rxq->pkt_last_seg = NULL; 2073 } 2074 2075 uint64_t 2076 ngbe_get_rx_queue_offloads(struct rte_eth_dev *dev __rte_unused) 2077 { 2078 return RTE_ETH_RX_OFFLOAD_VLAN_STRIP; 2079 } 2080 2081 uint64_t 2082 ngbe_get_rx_port_offloads(struct rte_eth_dev *dev) 2083 { 2084 uint64_t offloads; 2085 struct ngbe_hw *hw = ngbe_dev_hw(dev); 2086 2087 offloads = RTE_ETH_RX_OFFLOAD_IPV4_CKSUM | 2088 RTE_ETH_RX_OFFLOAD_UDP_CKSUM | 2089 RTE_ETH_RX_OFFLOAD_TCP_CKSUM | 2090 RTE_ETH_RX_OFFLOAD_KEEP_CRC | 2091 RTE_ETH_RX_OFFLOAD_VLAN_FILTER | 2092 RTE_ETH_RX_OFFLOAD_SCATTER; 2093 2094 if (hw->is_pf) 2095 offloads |= (RTE_ETH_RX_OFFLOAD_QINQ_STRIP | 2096 RTE_ETH_RX_OFFLOAD_VLAN_EXTEND); 2097 2098 return offloads; 2099 } 2100 2101 int 2102 ngbe_dev_rx_queue_setup(struct rte_eth_dev *dev, 2103 uint16_t queue_idx, 2104 uint16_t nb_desc, 2105 unsigned int socket_id, 2106 const struct rte_eth_rxconf *rx_conf, 2107 struct rte_mempool *mp) 2108 { 2109 const struct rte_memzone *rz; 2110 struct ngbe_rx_queue *rxq; 2111 struct ngbe_hw *hw; 2112 uint16_t len; 2113 struct ngbe_adapter *adapter = ngbe_dev_adapter(dev); 2114 uint64_t offloads; 2115 2116 PMD_INIT_FUNC_TRACE(); 2117 hw = ngbe_dev_hw(dev); 2118 2119 offloads = rx_conf->offloads | dev->data->dev_conf.rxmode.offloads; 2120 2121 /* Free memory prior to re-allocation if needed... */ 2122 if (dev->data->rx_queues[queue_idx] != NULL) { 2123 ngbe_rx_queue_release(dev->data->rx_queues[queue_idx]); 2124 dev->data->rx_queues[queue_idx] = NULL; 2125 } 2126 2127 /* First allocate the Rx queue data structure */ 2128 rxq = rte_zmalloc_socket("ethdev RX queue", 2129 sizeof(struct ngbe_rx_queue), 2130 RTE_CACHE_LINE_SIZE, socket_id); 2131 if (rxq == NULL) 2132 return -ENOMEM; 2133 rxq->mb_pool = mp; 2134 rxq->nb_rx_desc = nb_desc; 2135 rxq->rx_free_thresh = rx_conf->rx_free_thresh; 2136 rxq->queue_id = queue_idx; 2137 rxq->reg_idx = queue_idx; 2138 rxq->port_id = dev->data->port_id; 2139 if (dev->data->dev_conf.rxmode.offloads & RTE_ETH_RX_OFFLOAD_KEEP_CRC) 2140 rxq->crc_len = RTE_ETHER_CRC_LEN; 2141 else 2142 rxq->crc_len = 0; 2143 rxq->drop_en = rx_conf->rx_drop_en; 2144 rxq->rx_deferred_start = rx_conf->rx_deferred_start; 2145 rxq->offloads = offloads; 2146 2147 /* 2148 * Allocate Rx ring hardware descriptors. A memzone large enough to 2149 * handle the maximum ring size is allocated in order to allow for 2150 * resizing in later calls to the queue setup function. 2151 */ 2152 rz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx, 2153 RX_RING_SZ, NGBE_ALIGN, socket_id); 2154 if (rz == NULL) { 2155 ngbe_rx_queue_release(rxq); 2156 return -ENOMEM; 2157 } 2158 2159 /* 2160 * Zero init all the descriptors in the ring. 2161 */ 2162 memset(rz->addr, 0, RX_RING_SZ); 2163 2164 rxq->rdt_reg_addr = NGBE_REG_ADDR(hw, NGBE_RXWP(rxq->reg_idx)); 2165 rxq->rdh_reg_addr = NGBE_REG_ADDR(hw, NGBE_RXRP(rxq->reg_idx)); 2166 2167 rxq->rx_ring_phys_addr = TMZ_PADDR(rz); 2168 rxq->rx_ring = (struct ngbe_rx_desc *)TMZ_VADDR(rz); 2169 2170 /* 2171 * Certain constraints must be met in order to use the bulk buffer 2172 * allocation Rx burst function. If any of Rx queues doesn't meet them 2173 * the feature should be disabled for the whole port. 2174 */ 2175 if (check_rx_burst_bulk_alloc_preconditions(rxq)) { 2176 PMD_INIT_LOG(DEBUG, 2177 "queue[%d] doesn't meet Rx Bulk Alloc preconditions - canceling the feature for the whole port[%d]", 2178 rxq->queue_id, rxq->port_id); 2179 adapter->rx_bulk_alloc_allowed = false; 2180 } 2181 2182 /* 2183 * Allocate software ring. Allow for space at the end of the 2184 * S/W ring to make sure look-ahead logic in bulk alloc Rx burst 2185 * function does not access an invalid memory region. 2186 */ 2187 len = nb_desc; 2188 if (adapter->rx_bulk_alloc_allowed) 2189 len += RTE_PMD_NGBE_RX_MAX_BURST; 2190 2191 rxq->sw_ring = rte_zmalloc_socket("rxq->sw_ring", 2192 sizeof(struct ngbe_rx_entry) * len, 2193 RTE_CACHE_LINE_SIZE, socket_id); 2194 if (rxq->sw_ring == NULL) { 2195 ngbe_rx_queue_release(rxq); 2196 return -ENOMEM; 2197 } 2198 2199 /* 2200 * Always allocate even if it's not going to be needed in order to 2201 * simplify the code. 2202 * 2203 * This ring is used in Scattered Rx cases and Scattered Rx may 2204 * be requested in ngbe_dev_rx_init(), which is called later from 2205 * dev_start() flow. 2206 */ 2207 rxq->sw_sc_ring = 2208 rte_zmalloc_socket("rxq->sw_sc_ring", 2209 sizeof(struct ngbe_scattered_rx_entry) * len, 2210 RTE_CACHE_LINE_SIZE, socket_id); 2211 if (rxq->sw_sc_ring == NULL) { 2212 ngbe_rx_queue_release(rxq); 2213 return -ENOMEM; 2214 } 2215 2216 PMD_INIT_LOG(DEBUG, 2217 "sw_ring=%p sw_sc_ring=%p hw_ring=%p dma_addr=0x%" PRIx64, 2218 rxq->sw_ring, rxq->sw_sc_ring, rxq->rx_ring, 2219 rxq->rx_ring_phys_addr); 2220 2221 dev->data->rx_queues[queue_idx] = rxq; 2222 2223 ngbe_reset_rx_queue(adapter, rxq); 2224 2225 return 0; 2226 } 2227 2228 void 2229 ngbe_dev_clear_queues(struct rte_eth_dev *dev) 2230 { 2231 unsigned int i; 2232 struct ngbe_adapter *adapter = ngbe_dev_adapter(dev); 2233 2234 PMD_INIT_FUNC_TRACE(); 2235 2236 for (i = 0; i < dev->data->nb_tx_queues; i++) { 2237 struct ngbe_tx_queue *txq = dev->data->tx_queues[i]; 2238 2239 if (txq != NULL) { 2240 txq->ops->release_mbufs(txq); 2241 txq->ops->reset(txq); 2242 } 2243 } 2244 2245 for (i = 0; i < dev->data->nb_rx_queues; i++) { 2246 struct ngbe_rx_queue *rxq = dev->data->rx_queues[i]; 2247 2248 if (rxq != NULL) { 2249 ngbe_rx_queue_release_mbufs(rxq); 2250 ngbe_reset_rx_queue(adapter, rxq); 2251 } 2252 } 2253 } 2254 2255 void 2256 ngbe_dev_free_queues(struct rte_eth_dev *dev) 2257 { 2258 unsigned int i; 2259 2260 PMD_INIT_FUNC_TRACE(); 2261 2262 for (i = 0; i < dev->data->nb_rx_queues; i++) { 2263 ngbe_dev_rx_queue_release(dev, i); 2264 dev->data->rx_queues[i] = NULL; 2265 } 2266 dev->data->nb_rx_queues = 0; 2267 2268 for (i = 0; i < dev->data->nb_tx_queues; i++) { 2269 ngbe_dev_tx_queue_release(dev, i); 2270 dev->data->tx_queues[i] = NULL; 2271 } 2272 dev->data->nb_tx_queues = 0; 2273 } 2274 2275 void ngbe_configure_port(struct rte_eth_dev *dev) 2276 { 2277 struct ngbe_hw *hw = ngbe_dev_hw(dev); 2278 int i = 0; 2279 uint16_t tpids[8] = {RTE_ETHER_TYPE_VLAN, RTE_ETHER_TYPE_QINQ, 2280 0x9100, 0x9200, 2281 0x0000, 0x0000, 2282 0x0000, 0x0000}; 2283 2284 PMD_INIT_FUNC_TRACE(); 2285 2286 /* default outer vlan tpid */ 2287 wr32(hw, NGBE_EXTAG, 2288 NGBE_EXTAG_ETAG(RTE_ETHER_TYPE_ETAG) | 2289 NGBE_EXTAG_VLAN(RTE_ETHER_TYPE_QINQ)); 2290 2291 /* default inner vlan tpid */ 2292 wr32m(hw, NGBE_VLANCTL, 2293 NGBE_VLANCTL_TPID_MASK, 2294 NGBE_VLANCTL_TPID(RTE_ETHER_TYPE_VLAN)); 2295 wr32m(hw, NGBE_DMATXCTRL, 2296 NGBE_DMATXCTRL_TPID_MASK, 2297 NGBE_DMATXCTRL_TPID(RTE_ETHER_TYPE_VLAN)); 2298 2299 /* default vlan tpid filters */ 2300 for (i = 0; i < 8; i++) { 2301 wr32m(hw, NGBE_TAGTPID(i / 2), 2302 (i % 2 ? NGBE_TAGTPID_MSB_MASK 2303 : NGBE_TAGTPID_LSB_MASK), 2304 (i % 2 ? NGBE_TAGTPID_MSB(tpids[i]) 2305 : NGBE_TAGTPID_LSB(tpids[i]))); 2306 } 2307 } 2308 2309 static int 2310 ngbe_alloc_rx_queue_mbufs(struct ngbe_rx_queue *rxq) 2311 { 2312 struct ngbe_rx_entry *rxe = rxq->sw_ring; 2313 uint64_t dma_addr; 2314 unsigned int i; 2315 2316 /* Initialize software ring entries */ 2317 for (i = 0; i < rxq->nb_rx_desc; i++) { 2318 /* the ring can also be modified by hardware */ 2319 volatile struct ngbe_rx_desc *rxd; 2320 struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mb_pool); 2321 2322 if (mbuf == NULL) { 2323 PMD_INIT_LOG(ERR, "Rx mbuf alloc failed queue_id=%u port_id=%u", 2324 (unsigned int)rxq->queue_id, 2325 (unsigned int)rxq->port_id); 2326 return -ENOMEM; 2327 } 2328 2329 mbuf->data_off = RTE_PKTMBUF_HEADROOM; 2330 mbuf->port = rxq->port_id; 2331 2332 dma_addr = 2333 rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf)); 2334 rxd = &rxq->rx_ring[i]; 2335 NGBE_RXD_HDRADDR(rxd, 0); 2336 NGBE_RXD_PKTADDR(rxd, dma_addr); 2337 rxe[i].mbuf = mbuf; 2338 } 2339 2340 return 0; 2341 } 2342 2343 void 2344 ngbe_set_rx_function(struct rte_eth_dev *dev) 2345 { 2346 struct ngbe_adapter *adapter = ngbe_dev_adapter(dev); 2347 2348 if (dev->data->scattered_rx) { 2349 /* 2350 * Set the scattered callback: there are bulk and 2351 * single allocation versions. 2352 */ 2353 if (adapter->rx_bulk_alloc_allowed) { 2354 PMD_INIT_LOG(DEBUG, "Using a Scattered with bulk " 2355 "allocation callback (port=%d).", 2356 dev->data->port_id); 2357 dev->rx_pkt_burst = ngbe_recv_pkts_sc_bulk_alloc; 2358 } else { 2359 PMD_INIT_LOG(DEBUG, "Using Regular (non-vector, " 2360 "single allocation) " 2361 "Scattered Rx callback " 2362 "(port=%d).", 2363 dev->data->port_id); 2364 2365 dev->rx_pkt_burst = ngbe_recv_pkts_sc_single_alloc; 2366 } 2367 /* 2368 * Below we set "simple" callbacks according to port/queues parameters. 2369 * If parameters allow we are going to choose between the following 2370 * callbacks: 2371 * - Bulk Allocation 2372 * - Single buffer allocation (the simplest one) 2373 */ 2374 } else if (adapter->rx_bulk_alloc_allowed) { 2375 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are " 2376 "satisfied. Rx Burst Bulk Alloc function " 2377 "will be used on port=%d.", 2378 dev->data->port_id); 2379 2380 dev->rx_pkt_burst = ngbe_recv_pkts_bulk_alloc; 2381 } else { 2382 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are not " 2383 "satisfied, or Scattered Rx is requested " 2384 "(port=%d).", 2385 dev->data->port_id); 2386 2387 dev->rx_pkt_burst = ngbe_recv_pkts; 2388 } 2389 } 2390 2391 static const struct { 2392 eth_rx_burst_t pkt_burst; 2393 const char *info; 2394 } ngbe_rx_burst_infos[] = { 2395 { ngbe_recv_pkts_sc_single_alloc, "Scalar Scattered"}, 2396 { ngbe_recv_pkts_sc_bulk_alloc, "Scalar Scattered Bulk Alloc"}, 2397 { ngbe_recv_pkts_bulk_alloc, "Scalar Bulk Alloc"}, 2398 { ngbe_recv_pkts, "Scalar"}, 2399 }; 2400 2401 int 2402 ngbe_rx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id, 2403 struct rte_eth_burst_mode *mode) 2404 { 2405 eth_rx_burst_t pkt_burst = dev->rx_pkt_burst; 2406 int ret = -EINVAL; 2407 unsigned int i; 2408 2409 for (i = 0; i < RTE_DIM(ngbe_rx_burst_infos); ++i) { 2410 if (pkt_burst == ngbe_rx_burst_infos[i].pkt_burst) { 2411 snprintf(mode->info, sizeof(mode->info), "%s", 2412 ngbe_rx_burst_infos[i].info); 2413 ret = 0; 2414 break; 2415 } 2416 } 2417 2418 return ret; 2419 } 2420 2421 /* 2422 * Initializes Receive Unit. 2423 */ 2424 int 2425 ngbe_dev_rx_init(struct rte_eth_dev *dev) 2426 { 2427 struct ngbe_hw *hw; 2428 struct ngbe_rx_queue *rxq; 2429 uint64_t bus_addr; 2430 uint32_t fctrl; 2431 uint32_t hlreg0; 2432 uint32_t srrctl; 2433 uint32_t rdrxctl; 2434 uint32_t rxcsum; 2435 uint16_t buf_size; 2436 uint16_t i; 2437 struct rte_eth_rxmode *rx_conf = &dev->data->dev_conf.rxmode; 2438 2439 PMD_INIT_FUNC_TRACE(); 2440 hw = ngbe_dev_hw(dev); 2441 2442 /* 2443 * Make sure receives are disabled while setting 2444 * up the Rx context (registers, descriptor rings, etc.). 2445 */ 2446 wr32m(hw, NGBE_MACRXCFG, NGBE_MACRXCFG_ENA, 0); 2447 wr32m(hw, NGBE_PBRXCTL, NGBE_PBRXCTL_ENA, 0); 2448 2449 /* Enable receipt of broadcasted frames */ 2450 fctrl = rd32(hw, NGBE_PSRCTL); 2451 fctrl |= NGBE_PSRCTL_BCA; 2452 wr32(hw, NGBE_PSRCTL, fctrl); 2453 2454 /* 2455 * Configure CRC stripping, if any. 2456 */ 2457 hlreg0 = rd32(hw, NGBE_SECRXCTL); 2458 if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_KEEP_CRC) 2459 hlreg0 &= ~NGBE_SECRXCTL_CRCSTRIP; 2460 else 2461 hlreg0 |= NGBE_SECRXCTL_CRCSTRIP; 2462 hlreg0 &= ~NGBE_SECRXCTL_XDSA; 2463 wr32(hw, NGBE_SECRXCTL, hlreg0); 2464 2465 /* 2466 * Configure jumbo frame support, if any. 2467 */ 2468 wr32m(hw, NGBE_FRMSZ, NGBE_FRMSZ_MAX_MASK, 2469 NGBE_FRMSZ_MAX(dev->data->mtu + NGBE_ETH_OVERHEAD)); 2470 2471 /* 2472 * Assume no header split and no VLAN strip support 2473 * on any Rx queue first . 2474 */ 2475 rx_conf->offloads &= ~RTE_ETH_RX_OFFLOAD_VLAN_STRIP; 2476 2477 /* Setup Rx queues */ 2478 for (i = 0; i < dev->data->nb_rx_queues; i++) { 2479 rxq = dev->data->rx_queues[i]; 2480 2481 /* 2482 * Reset crc_len in case it was changed after queue setup by a 2483 * call to configure. 2484 */ 2485 if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_KEEP_CRC) 2486 rxq->crc_len = RTE_ETHER_CRC_LEN; 2487 else 2488 rxq->crc_len = 0; 2489 2490 /* Setup the Base and Length of the Rx Descriptor Rings */ 2491 bus_addr = rxq->rx_ring_phys_addr; 2492 wr32(hw, NGBE_RXBAL(rxq->reg_idx), 2493 (uint32_t)(bus_addr & BIT_MASK32)); 2494 wr32(hw, NGBE_RXBAH(rxq->reg_idx), 2495 (uint32_t)(bus_addr >> 32)); 2496 wr32(hw, NGBE_RXRP(rxq->reg_idx), 0); 2497 wr32(hw, NGBE_RXWP(rxq->reg_idx), 0); 2498 2499 srrctl = NGBE_RXCFG_RNGLEN(rxq->nb_rx_desc); 2500 2501 /* Set if packets are dropped when no descriptors available */ 2502 if (rxq->drop_en) 2503 srrctl |= NGBE_RXCFG_DROP; 2504 2505 /* 2506 * Configure the Rx buffer size in the PKTLEN field of 2507 * the RXCFG register of the queue. 2508 * The value is in 1 KB resolution. Valid values can be from 2509 * 1 KB to 16 KB. 2510 */ 2511 buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mb_pool) - 2512 RTE_PKTMBUF_HEADROOM); 2513 buf_size = ROUND_DOWN(buf_size, 0x1 << 10); 2514 srrctl |= NGBE_RXCFG_PKTLEN(buf_size); 2515 2516 wr32(hw, NGBE_RXCFG(rxq->reg_idx), srrctl); 2517 2518 /* It adds dual VLAN length for supporting dual VLAN */ 2519 if (dev->data->mtu + NGBE_ETH_OVERHEAD + 2520 2 * NGBE_VLAN_TAG_SIZE > buf_size) 2521 dev->data->scattered_rx = 1; 2522 if (rxq->offloads & RTE_ETH_RX_OFFLOAD_VLAN_STRIP) 2523 rx_conf->offloads |= RTE_ETH_RX_OFFLOAD_VLAN_STRIP; 2524 } 2525 2526 if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_SCATTER) 2527 dev->data->scattered_rx = 1; 2528 /* 2529 * Setup the Checksum Register. 2530 * Enable IP/L4 checksum computation by hardware if requested to do so. 2531 */ 2532 rxcsum = rd32(hw, NGBE_PSRCTL); 2533 rxcsum |= NGBE_PSRCTL_PCSD; 2534 if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_CHECKSUM) 2535 rxcsum |= NGBE_PSRCTL_L4CSUM; 2536 else 2537 rxcsum &= ~NGBE_PSRCTL_L4CSUM; 2538 2539 wr32(hw, NGBE_PSRCTL, rxcsum); 2540 2541 if (hw->is_pf) { 2542 rdrxctl = rd32(hw, NGBE_SECRXCTL); 2543 if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_KEEP_CRC) 2544 rdrxctl &= ~NGBE_SECRXCTL_CRCSTRIP; 2545 else 2546 rdrxctl |= NGBE_SECRXCTL_CRCSTRIP; 2547 wr32(hw, NGBE_SECRXCTL, rdrxctl); 2548 } 2549 2550 ngbe_set_rx_function(dev); 2551 2552 return 0; 2553 } 2554 2555 /* 2556 * Initializes Transmit Unit. 2557 */ 2558 void 2559 ngbe_dev_tx_init(struct rte_eth_dev *dev) 2560 { 2561 struct ngbe_hw *hw; 2562 struct ngbe_tx_queue *txq; 2563 uint64_t bus_addr; 2564 uint16_t i; 2565 2566 PMD_INIT_FUNC_TRACE(); 2567 hw = ngbe_dev_hw(dev); 2568 2569 wr32m(hw, NGBE_SECTXCTL, NGBE_SECTXCTL_ODSA, NGBE_SECTXCTL_ODSA); 2570 wr32m(hw, NGBE_SECTXCTL, NGBE_SECTXCTL_XDSA, 0); 2571 2572 /* Setup the Base and Length of the Tx Descriptor Rings */ 2573 for (i = 0; i < dev->data->nb_tx_queues; i++) { 2574 txq = dev->data->tx_queues[i]; 2575 2576 bus_addr = txq->tx_ring_phys_addr; 2577 wr32(hw, NGBE_TXBAL(txq->reg_idx), 2578 (uint32_t)(bus_addr & BIT_MASK32)); 2579 wr32(hw, NGBE_TXBAH(txq->reg_idx), 2580 (uint32_t)(bus_addr >> 32)); 2581 wr32m(hw, NGBE_TXCFG(txq->reg_idx), NGBE_TXCFG_BUFLEN_MASK, 2582 NGBE_TXCFG_BUFLEN(txq->nb_tx_desc)); 2583 /* Setup the HW Tx Head and TX Tail descriptor pointers */ 2584 wr32(hw, NGBE_TXRP(txq->reg_idx), 0); 2585 wr32(hw, NGBE_TXWP(txq->reg_idx), 0); 2586 } 2587 } 2588 2589 /* 2590 * Start Transmit and Receive Units. 2591 */ 2592 int 2593 ngbe_dev_rxtx_start(struct rte_eth_dev *dev) 2594 { 2595 struct ngbe_hw *hw; 2596 struct ngbe_tx_queue *txq; 2597 struct ngbe_rx_queue *rxq; 2598 uint32_t dmatxctl; 2599 uint32_t rxctrl; 2600 uint16_t i; 2601 int ret = 0; 2602 2603 PMD_INIT_FUNC_TRACE(); 2604 hw = ngbe_dev_hw(dev); 2605 2606 for (i = 0; i < dev->data->nb_tx_queues; i++) { 2607 txq = dev->data->tx_queues[i]; 2608 /* Setup Transmit Threshold Registers */ 2609 wr32m(hw, NGBE_TXCFG(txq->reg_idx), 2610 NGBE_TXCFG_HTHRESH_MASK | 2611 NGBE_TXCFG_WTHRESH_MASK, 2612 NGBE_TXCFG_HTHRESH(txq->hthresh) | 2613 NGBE_TXCFG_WTHRESH(txq->wthresh)); 2614 } 2615 2616 dmatxctl = rd32(hw, NGBE_DMATXCTRL); 2617 dmatxctl |= NGBE_DMATXCTRL_ENA; 2618 wr32(hw, NGBE_DMATXCTRL, dmatxctl); 2619 2620 for (i = 0; i < dev->data->nb_tx_queues; i++) { 2621 txq = dev->data->tx_queues[i]; 2622 if (txq->tx_deferred_start == 0) { 2623 ret = ngbe_dev_tx_queue_start(dev, i); 2624 if (ret < 0) 2625 return ret; 2626 } 2627 } 2628 2629 for (i = 0; i < dev->data->nb_rx_queues; i++) { 2630 rxq = dev->data->rx_queues[i]; 2631 if (rxq->rx_deferred_start == 0) { 2632 ret = ngbe_dev_rx_queue_start(dev, i); 2633 if (ret < 0) 2634 return ret; 2635 } 2636 } 2637 2638 /* Enable Receive engine */ 2639 rxctrl = rd32(hw, NGBE_PBRXCTL); 2640 rxctrl |= NGBE_PBRXCTL_ENA; 2641 hw->mac.enable_rx_dma(hw, rxctrl); 2642 2643 return 0; 2644 } 2645 2646 void 2647 ngbe_dev_save_rx_queue(struct ngbe_hw *hw, uint16_t rx_queue_id) 2648 { 2649 u32 *reg = &hw->q_rx_regs[rx_queue_id * 8]; 2650 *(reg++) = rd32(hw, NGBE_RXBAL(rx_queue_id)); 2651 *(reg++) = rd32(hw, NGBE_RXBAH(rx_queue_id)); 2652 *(reg++) = rd32(hw, NGBE_RXCFG(rx_queue_id)); 2653 } 2654 2655 void 2656 ngbe_dev_store_rx_queue(struct ngbe_hw *hw, uint16_t rx_queue_id) 2657 { 2658 u32 *reg = &hw->q_rx_regs[rx_queue_id * 8]; 2659 wr32(hw, NGBE_RXBAL(rx_queue_id), *(reg++)); 2660 wr32(hw, NGBE_RXBAH(rx_queue_id), *(reg++)); 2661 wr32(hw, NGBE_RXCFG(rx_queue_id), *(reg++) & ~NGBE_RXCFG_ENA); 2662 } 2663 2664 void 2665 ngbe_dev_save_tx_queue(struct ngbe_hw *hw, uint16_t tx_queue_id) 2666 { 2667 u32 *reg = &hw->q_tx_regs[tx_queue_id * 8]; 2668 *(reg++) = rd32(hw, NGBE_TXBAL(tx_queue_id)); 2669 *(reg++) = rd32(hw, NGBE_TXBAH(tx_queue_id)); 2670 *(reg++) = rd32(hw, NGBE_TXCFG(tx_queue_id)); 2671 } 2672 2673 void 2674 ngbe_dev_store_tx_queue(struct ngbe_hw *hw, uint16_t tx_queue_id) 2675 { 2676 u32 *reg = &hw->q_tx_regs[tx_queue_id * 8]; 2677 wr32(hw, NGBE_TXBAL(tx_queue_id), *(reg++)); 2678 wr32(hw, NGBE_TXBAH(tx_queue_id), *(reg++)); 2679 wr32(hw, NGBE_TXCFG(tx_queue_id), *(reg++) & ~NGBE_TXCFG_ENA); 2680 } 2681 2682 /* 2683 * Start Receive Units for specified queue. 2684 */ 2685 int 2686 ngbe_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id) 2687 { 2688 struct ngbe_hw *hw = ngbe_dev_hw(dev); 2689 struct ngbe_rx_queue *rxq; 2690 uint32_t rxdctl; 2691 int poll_ms; 2692 2693 PMD_INIT_FUNC_TRACE(); 2694 2695 rxq = dev->data->rx_queues[rx_queue_id]; 2696 2697 /* Allocate buffers for descriptor rings */ 2698 if (ngbe_alloc_rx_queue_mbufs(rxq) != 0) { 2699 PMD_INIT_LOG(ERR, "Could not alloc mbuf for queue:%d", 2700 rx_queue_id); 2701 return -1; 2702 } 2703 rxdctl = rd32(hw, NGBE_RXCFG(rxq->reg_idx)); 2704 rxdctl |= NGBE_RXCFG_ENA; 2705 wr32(hw, NGBE_RXCFG(rxq->reg_idx), rxdctl); 2706 2707 /* Wait until Rx Enable ready */ 2708 poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS; 2709 do { 2710 rte_delay_ms(1); 2711 rxdctl = rd32(hw, NGBE_RXCFG(rxq->reg_idx)); 2712 } while (--poll_ms && !(rxdctl & NGBE_RXCFG_ENA)); 2713 if (poll_ms == 0) 2714 PMD_INIT_LOG(ERR, "Could not enable Rx Queue %d", rx_queue_id); 2715 rte_wmb(); 2716 wr32(hw, NGBE_RXRP(rxq->reg_idx), 0); 2717 wr32(hw, NGBE_RXWP(rxq->reg_idx), rxq->nb_rx_desc - 1); 2718 dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED; 2719 2720 return 0; 2721 } 2722 2723 /* 2724 * Stop Receive Units for specified queue. 2725 */ 2726 int 2727 ngbe_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id) 2728 { 2729 struct ngbe_hw *hw = ngbe_dev_hw(dev); 2730 struct ngbe_adapter *adapter = ngbe_dev_adapter(dev); 2731 struct ngbe_rx_queue *rxq; 2732 uint32_t rxdctl; 2733 int poll_ms; 2734 2735 PMD_INIT_FUNC_TRACE(); 2736 2737 rxq = dev->data->rx_queues[rx_queue_id]; 2738 2739 ngbe_dev_save_rx_queue(hw, rxq->reg_idx); 2740 wr32m(hw, NGBE_RXCFG(rxq->reg_idx), NGBE_RXCFG_ENA, 0); 2741 2742 /* Wait until Rx Enable bit clear */ 2743 poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS; 2744 do { 2745 rte_delay_ms(1); 2746 rxdctl = rd32(hw, NGBE_RXCFG(rxq->reg_idx)); 2747 } while (--poll_ms && (rxdctl & NGBE_RXCFG_ENA)); 2748 if (poll_ms == 0) 2749 PMD_INIT_LOG(ERR, "Could not disable Rx Queue %d", rx_queue_id); 2750 2751 rte_delay_us(RTE_NGBE_WAIT_100_US); 2752 ngbe_dev_store_rx_queue(hw, rxq->reg_idx); 2753 2754 ngbe_rx_queue_release_mbufs(rxq); 2755 ngbe_reset_rx_queue(adapter, rxq); 2756 dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED; 2757 2758 return 0; 2759 } 2760 2761 /* 2762 * Start Transmit Units for specified queue. 2763 */ 2764 int 2765 ngbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id) 2766 { 2767 struct ngbe_hw *hw = ngbe_dev_hw(dev); 2768 struct ngbe_tx_queue *txq; 2769 uint32_t txdctl; 2770 int poll_ms; 2771 2772 PMD_INIT_FUNC_TRACE(); 2773 2774 txq = dev->data->tx_queues[tx_queue_id]; 2775 wr32m(hw, NGBE_TXCFG(txq->reg_idx), NGBE_TXCFG_ENA, NGBE_TXCFG_ENA); 2776 2777 /* Wait until Tx Enable ready */ 2778 poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS; 2779 do { 2780 rte_delay_ms(1); 2781 txdctl = rd32(hw, NGBE_TXCFG(txq->reg_idx)); 2782 } while (--poll_ms && !(txdctl & NGBE_TXCFG_ENA)); 2783 if (poll_ms == 0) 2784 PMD_INIT_LOG(ERR, "Could not enable Tx Queue %d", 2785 tx_queue_id); 2786 2787 rte_wmb(); 2788 wr32(hw, NGBE_TXWP(txq->reg_idx), txq->tx_tail); 2789 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED; 2790 2791 return 0; 2792 } 2793 2794 /* 2795 * Stop Transmit Units for specified queue. 2796 */ 2797 int 2798 ngbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id) 2799 { 2800 struct ngbe_hw *hw = ngbe_dev_hw(dev); 2801 struct ngbe_tx_queue *txq; 2802 uint32_t txdctl; 2803 uint32_t txtdh, txtdt; 2804 int poll_ms; 2805 2806 PMD_INIT_FUNC_TRACE(); 2807 2808 txq = dev->data->tx_queues[tx_queue_id]; 2809 2810 /* Wait until Tx queue is empty */ 2811 poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS; 2812 do { 2813 rte_delay_us(RTE_NGBE_WAIT_100_US); 2814 txtdh = rd32(hw, NGBE_TXRP(txq->reg_idx)); 2815 txtdt = rd32(hw, NGBE_TXWP(txq->reg_idx)); 2816 } while (--poll_ms && (txtdh != txtdt)); 2817 if (poll_ms == 0) 2818 PMD_INIT_LOG(ERR, "Tx Queue %d is not empty when stopping.", 2819 tx_queue_id); 2820 2821 ngbe_dev_save_tx_queue(hw, txq->reg_idx); 2822 wr32m(hw, NGBE_TXCFG(txq->reg_idx), NGBE_TXCFG_ENA, 0); 2823 2824 /* Wait until Tx Enable bit clear */ 2825 poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS; 2826 do { 2827 rte_delay_ms(1); 2828 txdctl = rd32(hw, NGBE_TXCFG(txq->reg_idx)); 2829 } while (--poll_ms && (txdctl & NGBE_TXCFG_ENA)); 2830 if (poll_ms == 0) 2831 PMD_INIT_LOG(ERR, "Could not disable Tx Queue %d", 2832 tx_queue_id); 2833 2834 rte_delay_us(RTE_NGBE_WAIT_100_US); 2835 ngbe_dev_store_tx_queue(hw, txq->reg_idx); 2836 2837 if (txq->ops != NULL) { 2838 txq->ops->release_mbufs(txq); 2839 txq->ops->reset(txq); 2840 } 2841 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED; 2842 2843 return 0; 2844 } 2845