1 // SPDX-License-Identifier: GPL-2.0 2 /* Marvell RVU Ethernet driver 3 * 4 * Copyright (C) 2020 Marvell. 5 * 6 */ 7 8 #include <linux/etherdevice.h> 9 #include <net/ip.h> 10 #include <net/tso.h> 11 #include <linux/bpf.h> 12 #include <linux/bpf_trace.h> 13 #include <net/ip6_checksum.h> 14 15 #include "otx2_reg.h" 16 #include "otx2_common.h" 17 #include "otx2_struct.h" 18 #include "otx2_txrx.h" 19 #include "otx2_ptp.h" 20 #include "cn10k.h" 21 22 #define CQE_ADDR(CQ, idx) ((CQ)->cqe_base + ((CQ)->cqe_size * (idx))) 23 #define PTP_PORT 0x13F 24 /* PTPv2 header Original Timestamp starts at byte offset 34 and 25 * contains 6 byte seconds field and 4 byte nano seconds field. 26 */ 27 #define PTP_SYNC_SEC_OFFSET 34 28 29 static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf, 30 struct bpf_prog *prog, 31 struct nix_cqe_rx_s *cqe, 32 struct otx2_cq_queue *cq); 33 34 static int otx2_nix_cq_op_status(struct otx2_nic *pfvf, 35 struct otx2_cq_queue *cq) 36 { 37 u64 incr = (u64)(cq->cq_idx) << 32; 38 u64 status; 39 40 status = otx2_atomic64_fetch_add(incr, pfvf->cq_op_addr); 41 42 if (unlikely(status & BIT_ULL(CQ_OP_STAT_OP_ERR) || 43 status & BIT_ULL(CQ_OP_STAT_CQ_ERR))) { 44 dev_err(pfvf->dev, "CQ stopped due to error"); 45 return -EINVAL; 46 } 47 48 cq->cq_tail = status & 0xFFFFF; 49 cq->cq_head = (status >> 20) & 0xFFFFF; 50 if (cq->cq_tail < cq->cq_head) 51 cq->pend_cqe = (cq->cqe_cnt - cq->cq_head) + 52 cq->cq_tail; 53 else 54 cq->pend_cqe = cq->cq_tail - cq->cq_head; 55 56 return 0; 57 } 58 59 static struct nix_cqe_hdr_s *otx2_get_next_cqe(struct otx2_cq_queue *cq) 60 { 61 struct nix_cqe_hdr_s *cqe_hdr; 62 63 cqe_hdr = (struct nix_cqe_hdr_s *)CQE_ADDR(cq, cq->cq_head); 64 if (cqe_hdr->cqe_type == NIX_XQE_TYPE_INVALID) 65 return NULL; 66 67 cq->cq_head++; 68 cq->cq_head &= (cq->cqe_cnt - 1); 69 70 return cqe_hdr; 71 } 72 73 static unsigned int frag_num(unsigned int i) 74 { 75 #ifdef __BIG_ENDIAN 76 return (i & ~3) + 3 - (i & 3); 77 #else 78 return i; 79 #endif 80 } 81 82 static dma_addr_t otx2_dma_map_skb_frag(struct otx2_nic *pfvf, 83 struct sk_buff *skb, int seg, int *len) 84 { 85 const skb_frag_t *frag; 86 struct page *page; 87 int offset; 88 89 /* First segment is always skb->data */ 90 if (!seg) { 91 page = virt_to_page(skb->data); 92 offset = offset_in_page(skb->data); 93 *len = skb_headlen(skb); 94 } else { 95 frag = &skb_shinfo(skb)->frags[seg - 1]; 96 page = skb_frag_page(frag); 97 offset = skb_frag_off(frag); 98 *len = skb_frag_size(frag); 99 } 100 return otx2_dma_map_page(pfvf, page, offset, *len, DMA_TO_DEVICE); 101 } 102 103 static void otx2_dma_unmap_skb_frags(struct otx2_nic *pfvf, struct sg_list *sg) 104 { 105 int seg; 106 107 for (seg = 0; seg < sg->num_segs; seg++) { 108 otx2_dma_unmap_page(pfvf, sg->dma_addr[seg], 109 sg->size[seg], DMA_TO_DEVICE); 110 } 111 sg->num_segs = 0; 112 } 113 114 static void otx2_xdp_snd_pkt_handler(struct otx2_nic *pfvf, 115 struct otx2_snd_queue *sq, 116 struct nix_cqe_tx_s *cqe) 117 { 118 struct nix_send_comp_s *snd_comp = &cqe->comp; 119 struct sg_list *sg; 120 struct page *page; 121 u64 pa; 122 123 sg = &sq->sg[snd_comp->sqe_id]; 124 125 pa = otx2_iova_to_phys(pfvf->iommu_domain, sg->dma_addr[0]); 126 otx2_dma_unmap_page(pfvf, sg->dma_addr[0], 127 sg->size[0], DMA_TO_DEVICE); 128 page = virt_to_page(phys_to_virt(pa)); 129 put_page(page); 130 } 131 132 static void otx2_snd_pkt_handler(struct otx2_nic *pfvf, 133 struct otx2_cq_queue *cq, 134 struct otx2_snd_queue *sq, 135 struct nix_cqe_tx_s *cqe, 136 int budget, int *tx_pkts, int *tx_bytes) 137 { 138 struct nix_send_comp_s *snd_comp = &cqe->comp; 139 struct skb_shared_hwtstamps ts; 140 struct sk_buff *skb = NULL; 141 u64 timestamp, tsns; 142 struct sg_list *sg; 143 int err; 144 145 if (unlikely(snd_comp->status) && netif_msg_tx_err(pfvf)) 146 net_err_ratelimited("%s: TX%d: Error in send CQ status:%x\n", 147 pfvf->netdev->name, cq->cint_idx, 148 snd_comp->status); 149 150 sg = &sq->sg[snd_comp->sqe_id]; 151 skb = (struct sk_buff *)sg->skb; 152 if (unlikely(!skb)) 153 return; 154 155 if (skb_shinfo(skb)->tx_flags & SKBTX_IN_PROGRESS) { 156 timestamp = ((u64 *)sq->timestamps->base)[snd_comp->sqe_id]; 157 if (timestamp != 1) { 158 timestamp = pfvf->ptp->convert_tx_ptp_tstmp(timestamp); 159 err = otx2_ptp_tstamp2time(pfvf, timestamp, &tsns); 160 if (!err) { 161 memset(&ts, 0, sizeof(ts)); 162 ts.hwtstamp = ns_to_ktime(tsns); 163 skb_tstamp_tx(skb, &ts); 164 } 165 } 166 } 167 168 *tx_bytes += skb->len; 169 (*tx_pkts)++; 170 otx2_dma_unmap_skb_frags(pfvf, sg); 171 napi_consume_skb(skb, budget); 172 sg->skb = (u64)NULL; 173 } 174 175 static void otx2_set_rxtstamp(struct otx2_nic *pfvf, 176 struct sk_buff *skb, void *data) 177 { 178 u64 timestamp, tsns; 179 int err; 180 181 if (!(pfvf->flags & OTX2_FLAG_RX_TSTAMP_ENABLED)) 182 return; 183 184 timestamp = pfvf->ptp->convert_rx_ptp_tstmp(*(u64 *)data); 185 /* The first 8 bytes is the timestamp */ 186 err = otx2_ptp_tstamp2time(pfvf, timestamp, &tsns); 187 if (err) 188 return; 189 190 skb_hwtstamps(skb)->hwtstamp = ns_to_ktime(tsns); 191 } 192 193 static bool otx2_skb_add_frag(struct otx2_nic *pfvf, struct sk_buff *skb, 194 u64 iova, int len, struct nix_rx_parse_s *parse, 195 int qidx) 196 { 197 struct page *page; 198 int off = 0; 199 void *va; 200 201 va = phys_to_virt(otx2_iova_to_phys(pfvf->iommu_domain, iova)); 202 203 if (likely(!skb_shinfo(skb)->nr_frags)) { 204 /* Check if data starts at some nonzero offset 205 * from the start of the buffer. For now the 206 * only possible offset is 8 bytes in the case 207 * where packet is prepended by a timestamp. 208 */ 209 if (parse->laptr) { 210 otx2_set_rxtstamp(pfvf, skb, va); 211 off = OTX2_HW_TIMESTAMP_LEN; 212 } 213 } 214 215 page = virt_to_page(va); 216 if (likely(skb_shinfo(skb)->nr_frags < MAX_SKB_FRAGS)) { 217 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, 218 va - page_address(page) + off, 219 len - off, pfvf->rbsize); 220 221 otx2_dma_unmap_page(pfvf, iova - OTX2_HEAD_ROOM, 222 pfvf->rbsize, DMA_FROM_DEVICE); 223 return true; 224 } 225 226 /* If more than MAX_SKB_FRAGS fragments are received then 227 * give back those buffer pointers to hardware for reuse. 228 */ 229 pfvf->hw_ops->aura_freeptr(pfvf, qidx, iova & ~0x07ULL); 230 231 return false; 232 } 233 234 static void otx2_set_rxhash(struct otx2_nic *pfvf, 235 struct nix_cqe_rx_s *cqe, struct sk_buff *skb) 236 { 237 enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE; 238 struct otx2_rss_info *rss; 239 u32 hash = 0; 240 241 if (!(pfvf->netdev->features & NETIF_F_RXHASH)) 242 return; 243 244 rss = &pfvf->hw.rss_info; 245 if (rss->flowkey_cfg) { 246 if (rss->flowkey_cfg & 247 ~(NIX_FLOW_KEY_TYPE_IPV4 | NIX_FLOW_KEY_TYPE_IPV6)) 248 hash_type = PKT_HASH_TYPE_L4; 249 else 250 hash_type = PKT_HASH_TYPE_L3; 251 hash = cqe->hdr.flow_tag; 252 } 253 skb_set_hash(skb, hash, hash_type); 254 } 255 256 static void otx2_free_rcv_seg(struct otx2_nic *pfvf, struct nix_cqe_rx_s *cqe, 257 int qidx) 258 { 259 struct nix_rx_sg_s *sg = &cqe->sg; 260 void *end, *start; 261 u64 *seg_addr; 262 int seg; 263 264 start = (void *)sg; 265 end = start + ((cqe->parse.desc_sizem1 + 1) * 16); 266 while (start < end) { 267 sg = (struct nix_rx_sg_s *)start; 268 seg_addr = &sg->seg_addr; 269 for (seg = 0; seg < sg->segs; seg++, seg_addr++) 270 pfvf->hw_ops->aura_freeptr(pfvf, qidx, 271 *seg_addr & ~0x07ULL); 272 start += sizeof(*sg); 273 } 274 } 275 276 static bool otx2_check_rcv_errors(struct otx2_nic *pfvf, 277 struct nix_cqe_rx_s *cqe, int qidx) 278 { 279 struct otx2_drv_stats *stats = &pfvf->hw.drv_stats; 280 struct nix_rx_parse_s *parse = &cqe->parse; 281 282 if (netif_msg_rx_err(pfvf)) 283 netdev_err(pfvf->netdev, 284 "RQ%d: Error pkt with errlev:0x%x errcode:0x%x\n", 285 qidx, parse->errlev, parse->errcode); 286 287 if (parse->errlev == NPC_ERRLVL_RE) { 288 switch (parse->errcode) { 289 case ERRCODE_FCS: 290 case ERRCODE_FCS_RCV: 291 atomic_inc(&stats->rx_fcs_errs); 292 break; 293 case ERRCODE_UNDERSIZE: 294 atomic_inc(&stats->rx_undersize_errs); 295 break; 296 case ERRCODE_OVERSIZE: 297 atomic_inc(&stats->rx_oversize_errs); 298 break; 299 case ERRCODE_OL2_LEN_MISMATCH: 300 atomic_inc(&stats->rx_len_errs); 301 break; 302 default: 303 atomic_inc(&stats->rx_other_errs); 304 break; 305 } 306 } else if (parse->errlev == NPC_ERRLVL_NIX) { 307 switch (parse->errcode) { 308 case ERRCODE_OL3_LEN: 309 case ERRCODE_OL4_LEN: 310 case ERRCODE_IL3_LEN: 311 case ERRCODE_IL4_LEN: 312 atomic_inc(&stats->rx_len_errs); 313 break; 314 case ERRCODE_OL4_CSUM: 315 case ERRCODE_IL4_CSUM: 316 atomic_inc(&stats->rx_csum_errs); 317 break; 318 default: 319 atomic_inc(&stats->rx_other_errs); 320 break; 321 } 322 } else { 323 atomic_inc(&stats->rx_other_errs); 324 /* For now ignore all the NPC parser errors and 325 * pass the packets to stack. 326 */ 327 return false; 328 } 329 330 /* If RXALL is enabled pass on packets to stack. */ 331 if (pfvf->netdev->features & NETIF_F_RXALL) 332 return false; 333 334 /* Free buffer back to pool */ 335 if (cqe->sg.segs) 336 otx2_free_rcv_seg(pfvf, cqe, qidx); 337 return true; 338 } 339 340 static void otx2_rcv_pkt_handler(struct otx2_nic *pfvf, 341 struct napi_struct *napi, 342 struct otx2_cq_queue *cq, 343 struct nix_cqe_rx_s *cqe) 344 { 345 struct nix_rx_parse_s *parse = &cqe->parse; 346 struct nix_rx_sg_s *sg = &cqe->sg; 347 struct sk_buff *skb = NULL; 348 void *end, *start; 349 u64 *seg_addr; 350 u16 *seg_size; 351 int seg; 352 353 if (unlikely(parse->errlev || parse->errcode)) { 354 if (otx2_check_rcv_errors(pfvf, cqe, cq->cq_idx)) 355 return; 356 } 357 358 if (pfvf->xdp_prog) 359 if (otx2_xdp_rcv_pkt_handler(pfvf, pfvf->xdp_prog, cqe, cq)) 360 return; 361 362 skb = napi_get_frags(napi); 363 if (unlikely(!skb)) 364 return; 365 366 start = (void *)sg; 367 end = start + ((cqe->parse.desc_sizem1 + 1) * 16); 368 while (start < end) { 369 sg = (struct nix_rx_sg_s *)start; 370 seg_addr = &sg->seg_addr; 371 seg_size = (void *)sg; 372 for (seg = 0; seg < sg->segs; seg++, seg_addr++) { 373 if (otx2_skb_add_frag(pfvf, skb, *seg_addr, 374 seg_size[seg], parse, cq->cq_idx)) 375 cq->pool_ptrs++; 376 } 377 start += sizeof(*sg); 378 } 379 otx2_set_rxhash(pfvf, cqe, skb); 380 381 skb_record_rx_queue(skb, cq->cq_idx); 382 if (pfvf->netdev->features & NETIF_F_RXCSUM) 383 skb->ip_summed = CHECKSUM_UNNECESSARY; 384 385 napi_gro_frags(napi); 386 } 387 388 static int otx2_rx_napi_handler(struct otx2_nic *pfvf, 389 struct napi_struct *napi, 390 struct otx2_cq_queue *cq, int budget) 391 { 392 struct nix_cqe_rx_s *cqe; 393 int processed_cqe = 0; 394 395 if (cq->pend_cqe >= budget) 396 goto process_cqe; 397 398 if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) 399 return 0; 400 401 process_cqe: 402 while (likely(processed_cqe < budget) && cq->pend_cqe) { 403 cqe = (struct nix_cqe_rx_s *)CQE_ADDR(cq, cq->cq_head); 404 if (cqe->hdr.cqe_type == NIX_XQE_TYPE_INVALID || 405 !cqe->sg.seg_addr) { 406 if (!processed_cqe) 407 return 0; 408 break; 409 } 410 cq->cq_head++; 411 cq->cq_head &= (cq->cqe_cnt - 1); 412 413 otx2_rcv_pkt_handler(pfvf, napi, cq, cqe); 414 415 cqe->hdr.cqe_type = NIX_XQE_TYPE_INVALID; 416 cqe->sg.seg_addr = 0x00; 417 processed_cqe++; 418 cq->pend_cqe--; 419 } 420 421 /* Free CQEs to HW */ 422 otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR, 423 ((u64)cq->cq_idx << 32) | processed_cqe); 424 425 return processed_cqe; 426 } 427 428 void otx2_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq) 429 { 430 struct otx2_nic *pfvf = dev; 431 dma_addr_t bufptr; 432 433 while (cq->pool_ptrs) { 434 if (otx2_alloc_buffer(pfvf, cq, &bufptr)) 435 break; 436 otx2_aura_freeptr(pfvf, cq->cq_idx, bufptr + OTX2_HEAD_ROOM); 437 cq->pool_ptrs--; 438 } 439 } 440 441 static int otx2_tx_napi_handler(struct otx2_nic *pfvf, 442 struct otx2_cq_queue *cq, int budget) 443 { 444 int tx_pkts = 0, tx_bytes = 0, qidx; 445 struct otx2_snd_queue *sq; 446 struct nix_cqe_tx_s *cqe; 447 int processed_cqe = 0; 448 449 if (cq->pend_cqe >= budget) 450 goto process_cqe; 451 452 if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) 453 return 0; 454 455 process_cqe: 456 qidx = cq->cq_idx - pfvf->hw.rx_queues; 457 sq = &pfvf->qset.sq[qidx]; 458 459 while (likely(processed_cqe < budget) && cq->pend_cqe) { 460 cqe = (struct nix_cqe_tx_s *)otx2_get_next_cqe(cq); 461 if (unlikely(!cqe)) { 462 if (!processed_cqe) 463 return 0; 464 break; 465 } 466 467 if (cq->cq_type == CQ_XDP) { 468 otx2_xdp_snd_pkt_handler(pfvf, sq, cqe); 469 } else { 470 otx2_snd_pkt_handler(pfvf, cq, sq, cqe, budget, 471 &tx_pkts, &tx_bytes); 472 } 473 474 cqe->hdr.cqe_type = NIX_XQE_TYPE_INVALID; 475 processed_cqe++; 476 cq->pend_cqe--; 477 478 sq->cons_head++; 479 sq->cons_head &= (sq->sqe_cnt - 1); 480 } 481 482 /* Free CQEs to HW */ 483 otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR, 484 ((u64)cq->cq_idx << 32) | processed_cqe); 485 486 if (likely(tx_pkts)) { 487 struct netdev_queue *txq; 488 489 txq = netdev_get_tx_queue(pfvf->netdev, cq->cint_idx); 490 netdev_tx_completed_queue(txq, tx_pkts, tx_bytes); 491 /* Check if queue was stopped earlier due to ring full */ 492 smp_mb(); 493 if (netif_tx_queue_stopped(txq) && 494 netif_carrier_ok(pfvf->netdev)) 495 netif_tx_wake_queue(txq); 496 } 497 return 0; 498 } 499 500 static void otx2_adjust_adaptive_coalese(struct otx2_nic *pfvf, struct otx2_cq_poll *cq_poll) 501 { 502 struct dim_sample dim_sample; 503 u64 rx_frames, rx_bytes; 504 505 rx_frames = OTX2_GET_RX_STATS(RX_BCAST) + OTX2_GET_RX_STATS(RX_MCAST) + 506 OTX2_GET_RX_STATS(RX_UCAST); 507 rx_bytes = OTX2_GET_RX_STATS(RX_OCTS); 508 dim_update_sample(pfvf->napi_events, rx_frames, rx_bytes, &dim_sample); 509 net_dim(&cq_poll->dim, dim_sample); 510 } 511 512 int otx2_napi_handler(struct napi_struct *napi, int budget) 513 { 514 struct otx2_cq_queue *rx_cq = NULL; 515 struct otx2_cq_poll *cq_poll; 516 int workdone = 0, cq_idx, i; 517 struct otx2_cq_queue *cq; 518 struct otx2_qset *qset; 519 struct otx2_nic *pfvf; 520 521 cq_poll = container_of(napi, struct otx2_cq_poll, napi); 522 pfvf = (struct otx2_nic *)cq_poll->dev; 523 qset = &pfvf->qset; 524 525 for (i = 0; i < CQS_PER_CINT; i++) { 526 cq_idx = cq_poll->cq_ids[i]; 527 if (unlikely(cq_idx == CINT_INVALID_CQ)) 528 continue; 529 cq = &qset->cq[cq_idx]; 530 if (cq->cq_type == CQ_RX) { 531 rx_cq = cq; 532 workdone += otx2_rx_napi_handler(pfvf, napi, 533 cq, budget); 534 } else { 535 workdone += otx2_tx_napi_handler(pfvf, cq, budget); 536 } 537 } 538 539 if (rx_cq && rx_cq->pool_ptrs) 540 pfvf->hw_ops->refill_pool_ptrs(pfvf, rx_cq); 541 /* Clear the IRQ */ 542 otx2_write64(pfvf, NIX_LF_CINTX_INT(cq_poll->cint_idx), BIT_ULL(0)); 543 544 if (workdone < budget && napi_complete_done(napi, workdone)) { 545 /* If interface is going down, don't re-enable IRQ */ 546 if (pfvf->flags & OTX2_FLAG_INTF_DOWN) 547 return workdone; 548 549 /* Check for adaptive interrupt coalesce */ 550 if (workdone != 0 && 551 ((pfvf->flags & OTX2_FLAG_ADPTV_INT_COAL_ENABLED) == 552 OTX2_FLAG_ADPTV_INT_COAL_ENABLED)) { 553 /* Adjust irq coalese using net_dim */ 554 otx2_adjust_adaptive_coalese(pfvf, cq_poll); 555 /* Update irq coalescing */ 556 for (i = 0; i < pfvf->hw.cint_cnt; i++) 557 otx2_config_irq_coalescing(pfvf, i); 558 } 559 560 /* Re-enable interrupts */ 561 otx2_write64(pfvf, NIX_LF_CINTX_ENA_W1S(cq_poll->cint_idx), 562 BIT_ULL(0)); 563 } 564 return workdone; 565 } 566 567 void otx2_sqe_flush(void *dev, struct otx2_snd_queue *sq, 568 int size, int qidx) 569 { 570 u64 status; 571 572 /* Packet data stores should finish before SQE is flushed to HW */ 573 dma_wmb(); 574 575 do { 576 memcpy(sq->lmt_addr, sq->sqe_base, size); 577 status = otx2_lmt_flush(sq->io_addr); 578 } while (status == 0); 579 580 sq->head++; 581 sq->head &= (sq->sqe_cnt - 1); 582 } 583 584 #define MAX_SEGS_PER_SG 3 585 /* Add SQE scatter/gather subdescriptor structure */ 586 static bool otx2_sqe_add_sg(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, 587 struct sk_buff *skb, int num_segs, int *offset) 588 { 589 struct nix_sqe_sg_s *sg = NULL; 590 u64 dma_addr, *iova = NULL; 591 u16 *sg_lens = NULL; 592 int seg, len; 593 594 sq->sg[sq->head].num_segs = 0; 595 596 for (seg = 0; seg < num_segs; seg++) { 597 if ((seg % MAX_SEGS_PER_SG) == 0) { 598 sg = (struct nix_sqe_sg_s *)(sq->sqe_base + *offset); 599 sg->ld_type = NIX_SEND_LDTYPE_LDD; 600 sg->subdc = NIX_SUBDC_SG; 601 sg->segs = 0; 602 sg_lens = (void *)sg; 603 iova = (void *)sg + sizeof(*sg); 604 /* Next subdc always starts at a 16byte boundary. 605 * So if sg->segs is whether 2 or 3, offset += 16bytes. 606 */ 607 if ((num_segs - seg) >= (MAX_SEGS_PER_SG - 1)) 608 *offset += sizeof(*sg) + (3 * sizeof(u64)); 609 else 610 *offset += sizeof(*sg) + sizeof(u64); 611 } 612 dma_addr = otx2_dma_map_skb_frag(pfvf, skb, seg, &len); 613 if (dma_mapping_error(pfvf->dev, dma_addr)) 614 return false; 615 616 sg_lens[frag_num(seg % MAX_SEGS_PER_SG)] = len; 617 sg->segs++; 618 *iova++ = dma_addr; 619 620 /* Save DMA mapping info for later unmapping */ 621 sq->sg[sq->head].dma_addr[seg] = dma_addr; 622 sq->sg[sq->head].size[seg] = len; 623 sq->sg[sq->head].num_segs++; 624 } 625 626 sq->sg[sq->head].skb = (u64)skb; 627 return true; 628 } 629 630 /* Add SQE extended header subdescriptor */ 631 static void otx2_sqe_add_ext(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, 632 struct sk_buff *skb, int *offset) 633 { 634 struct nix_sqe_ext_s *ext; 635 636 ext = (struct nix_sqe_ext_s *)(sq->sqe_base + *offset); 637 ext->subdc = NIX_SUBDC_EXT; 638 if (skb_shinfo(skb)->gso_size) { 639 ext->lso = 1; 640 ext->lso_sb = skb_tcp_all_headers(skb); 641 ext->lso_mps = skb_shinfo(skb)->gso_size; 642 643 /* Only TSOv4 and TSOv6 GSO offloads are supported */ 644 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 645 ext->lso_format = pfvf->hw.lso_tsov4_idx; 646 647 /* HW adds payload size to 'ip_hdr->tot_len' while 648 * sending TSO segment, hence set payload length 649 * in IP header of the packet to just header length. 650 */ 651 ip_hdr(skb)->tot_len = 652 htons(ext->lso_sb - skb_network_offset(skb)); 653 } else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) { 654 ext->lso_format = pfvf->hw.lso_tsov6_idx; 655 ipv6_hdr(skb)->payload_len = htons(tcp_hdrlen(skb)); 656 } else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) { 657 __be16 l3_proto = vlan_get_protocol(skb); 658 struct udphdr *udph = udp_hdr(skb); 659 u16 iplen; 660 661 ext->lso_sb = skb_transport_offset(skb) + 662 sizeof(struct udphdr); 663 664 /* HW adds payload size to length fields in IP and 665 * UDP headers while segmentation, hence adjust the 666 * lengths to just header sizes. 667 */ 668 iplen = htons(ext->lso_sb - skb_network_offset(skb)); 669 if (l3_proto == htons(ETH_P_IP)) { 670 ip_hdr(skb)->tot_len = iplen; 671 ext->lso_format = pfvf->hw.lso_udpv4_idx; 672 } else { 673 ipv6_hdr(skb)->payload_len = iplen; 674 ext->lso_format = pfvf->hw.lso_udpv6_idx; 675 } 676 677 udph->len = htons(sizeof(struct udphdr)); 678 } 679 } else if (skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) { 680 ext->tstmp = 1; 681 } 682 683 #define OTX2_VLAN_PTR_OFFSET (ETH_HLEN - ETH_TLEN) 684 if (skb_vlan_tag_present(skb)) { 685 if (skb->vlan_proto == htons(ETH_P_8021Q)) { 686 ext->vlan1_ins_ena = 1; 687 ext->vlan1_ins_ptr = OTX2_VLAN_PTR_OFFSET; 688 ext->vlan1_ins_tci = skb_vlan_tag_get(skb); 689 } else if (skb->vlan_proto == htons(ETH_P_8021AD)) { 690 ext->vlan0_ins_ena = 1; 691 ext->vlan0_ins_ptr = OTX2_VLAN_PTR_OFFSET; 692 ext->vlan0_ins_tci = skb_vlan_tag_get(skb); 693 } 694 } 695 696 *offset += sizeof(*ext); 697 } 698 699 static void otx2_sqe_add_mem(struct otx2_snd_queue *sq, int *offset, 700 int alg, u64 iova, int ptp_offset, 701 u64 base_ns, bool udp_csum_crt) 702 { 703 struct nix_sqe_mem_s *mem; 704 705 mem = (struct nix_sqe_mem_s *)(sq->sqe_base + *offset); 706 mem->subdc = NIX_SUBDC_MEM; 707 mem->alg = alg; 708 mem->wmem = 1; /* wait for the memory operation */ 709 mem->addr = iova; 710 711 if (ptp_offset) { 712 mem->start_offset = ptp_offset; 713 mem->udp_csum_crt = !!udp_csum_crt; 714 mem->base_ns = base_ns; 715 mem->step_type = 1; 716 } 717 718 *offset += sizeof(*mem); 719 } 720 721 /* Add SQE header subdescriptor structure */ 722 static void otx2_sqe_add_hdr(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, 723 struct nix_sqe_hdr_s *sqe_hdr, 724 struct sk_buff *skb, u16 qidx) 725 { 726 int proto = 0; 727 728 /* Check if SQE was framed before, if yes then no need to 729 * set these constants again and again. 730 */ 731 if (!sqe_hdr->total) { 732 /* Don't free Tx buffers to Aura */ 733 sqe_hdr->df = 1; 734 sqe_hdr->aura = sq->aura_id; 735 /* Post a CQE Tx after pkt transmission */ 736 sqe_hdr->pnc = 1; 737 sqe_hdr->sq = qidx; 738 } 739 sqe_hdr->total = skb->len; 740 /* Set SQE identifier which will be used later for freeing SKB */ 741 sqe_hdr->sqe_id = sq->head; 742 743 /* Offload TCP/UDP checksum to HW */ 744 if (skb->ip_summed == CHECKSUM_PARTIAL) { 745 sqe_hdr->ol3ptr = skb_network_offset(skb); 746 sqe_hdr->ol4ptr = skb_transport_offset(skb); 747 /* get vlan protocol Ethertype */ 748 if (eth_type_vlan(skb->protocol)) 749 skb->protocol = vlan_get_protocol(skb); 750 751 if (skb->protocol == htons(ETH_P_IP)) { 752 proto = ip_hdr(skb)->protocol; 753 /* In case of TSO, HW needs this to be explicitly set. 754 * So set this always, instead of adding a check. 755 */ 756 sqe_hdr->ol3type = NIX_SENDL3TYPE_IP4_CKSUM; 757 } else if (skb->protocol == htons(ETH_P_IPV6)) { 758 proto = ipv6_hdr(skb)->nexthdr; 759 sqe_hdr->ol3type = NIX_SENDL3TYPE_IP6; 760 } 761 762 if (proto == IPPROTO_TCP) 763 sqe_hdr->ol4type = NIX_SENDL4TYPE_TCP_CKSUM; 764 else if (proto == IPPROTO_UDP) 765 sqe_hdr->ol4type = NIX_SENDL4TYPE_UDP_CKSUM; 766 } 767 } 768 769 static int otx2_dma_map_tso_skb(struct otx2_nic *pfvf, 770 struct otx2_snd_queue *sq, 771 struct sk_buff *skb, int sqe, int hdr_len) 772 { 773 int num_segs = skb_shinfo(skb)->nr_frags + 1; 774 struct sg_list *sg = &sq->sg[sqe]; 775 u64 dma_addr; 776 int seg, len; 777 778 sg->num_segs = 0; 779 780 /* Get payload length at skb->data */ 781 len = skb_headlen(skb) - hdr_len; 782 783 for (seg = 0; seg < num_segs; seg++) { 784 /* Skip skb->data, if there is no payload */ 785 if (!seg && !len) 786 continue; 787 dma_addr = otx2_dma_map_skb_frag(pfvf, skb, seg, &len); 788 if (dma_mapping_error(pfvf->dev, dma_addr)) 789 goto unmap; 790 791 /* Save DMA mapping info for later unmapping */ 792 sg->dma_addr[sg->num_segs] = dma_addr; 793 sg->size[sg->num_segs] = len; 794 sg->num_segs++; 795 } 796 return 0; 797 unmap: 798 otx2_dma_unmap_skb_frags(pfvf, sg); 799 return -EINVAL; 800 } 801 802 static u64 otx2_tso_frag_dma_addr(struct otx2_snd_queue *sq, 803 struct sk_buff *skb, int seg, 804 u64 seg_addr, int hdr_len, int sqe) 805 { 806 struct sg_list *sg = &sq->sg[sqe]; 807 const skb_frag_t *frag; 808 int offset; 809 810 if (seg < 0) 811 return sg->dma_addr[0] + (seg_addr - (u64)skb->data); 812 813 frag = &skb_shinfo(skb)->frags[seg]; 814 offset = seg_addr - (u64)skb_frag_address(frag); 815 if (skb_headlen(skb) - hdr_len) 816 seg++; 817 return sg->dma_addr[seg] + offset; 818 } 819 820 static void otx2_sqe_tso_add_sg(struct otx2_snd_queue *sq, 821 struct sg_list *list, int *offset) 822 { 823 struct nix_sqe_sg_s *sg = NULL; 824 u16 *sg_lens = NULL; 825 u64 *iova = NULL; 826 int seg; 827 828 /* Add SG descriptors with buffer addresses */ 829 for (seg = 0; seg < list->num_segs; seg++) { 830 if ((seg % MAX_SEGS_PER_SG) == 0) { 831 sg = (struct nix_sqe_sg_s *)(sq->sqe_base + *offset); 832 sg->ld_type = NIX_SEND_LDTYPE_LDD; 833 sg->subdc = NIX_SUBDC_SG; 834 sg->segs = 0; 835 sg_lens = (void *)sg; 836 iova = (void *)sg + sizeof(*sg); 837 /* Next subdc always starts at a 16byte boundary. 838 * So if sg->segs is whether 2 or 3, offset += 16bytes. 839 */ 840 if ((list->num_segs - seg) >= (MAX_SEGS_PER_SG - 1)) 841 *offset += sizeof(*sg) + (3 * sizeof(u64)); 842 else 843 *offset += sizeof(*sg) + sizeof(u64); 844 } 845 sg_lens[frag_num(seg % MAX_SEGS_PER_SG)] = list->size[seg]; 846 *iova++ = list->dma_addr[seg]; 847 sg->segs++; 848 } 849 } 850 851 static void otx2_sq_append_tso(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, 852 struct sk_buff *skb, u16 qidx) 853 { 854 struct netdev_queue *txq = netdev_get_tx_queue(pfvf->netdev, qidx); 855 int hdr_len, tcp_data, seg_len, pkt_len, offset; 856 struct nix_sqe_hdr_s *sqe_hdr; 857 int first_sqe = sq->head; 858 struct sg_list list; 859 struct tso_t tso; 860 861 hdr_len = tso_start(skb, &tso); 862 863 /* Map SKB's fragments to DMA. 864 * It's done here to avoid mapping for every TSO segment's packet. 865 */ 866 if (otx2_dma_map_tso_skb(pfvf, sq, skb, first_sqe, hdr_len)) { 867 dev_kfree_skb_any(skb); 868 return; 869 } 870 871 netdev_tx_sent_queue(txq, skb->len); 872 873 tcp_data = skb->len - hdr_len; 874 while (tcp_data > 0) { 875 char *hdr; 876 877 seg_len = min_t(int, skb_shinfo(skb)->gso_size, tcp_data); 878 tcp_data -= seg_len; 879 880 /* Set SQE's SEND_HDR */ 881 memset(sq->sqe_base, 0, sq->sqe_size); 882 sqe_hdr = (struct nix_sqe_hdr_s *)(sq->sqe_base); 883 otx2_sqe_add_hdr(pfvf, sq, sqe_hdr, skb, qidx); 884 offset = sizeof(*sqe_hdr); 885 886 /* Add TSO segment's pkt header */ 887 hdr = sq->tso_hdrs->base + (sq->head * TSO_HEADER_SIZE); 888 tso_build_hdr(skb, hdr, &tso, seg_len, tcp_data == 0); 889 list.dma_addr[0] = 890 sq->tso_hdrs->iova + (sq->head * TSO_HEADER_SIZE); 891 list.size[0] = hdr_len; 892 list.num_segs = 1; 893 894 /* Add TSO segment's payload data fragments */ 895 pkt_len = hdr_len; 896 while (seg_len > 0) { 897 int size; 898 899 size = min_t(int, tso.size, seg_len); 900 901 list.size[list.num_segs] = size; 902 list.dma_addr[list.num_segs] = 903 otx2_tso_frag_dma_addr(sq, skb, 904 tso.next_frag_idx - 1, 905 (u64)tso.data, hdr_len, 906 first_sqe); 907 list.num_segs++; 908 pkt_len += size; 909 seg_len -= size; 910 tso_build_data(skb, &tso, size); 911 } 912 sqe_hdr->total = pkt_len; 913 otx2_sqe_tso_add_sg(sq, &list, &offset); 914 915 /* DMA mappings and skb needs to be freed only after last 916 * TSO segment is transmitted out. So set 'PNC' only for 917 * last segment. Also point last segment's sqe_id to first 918 * segment's SQE index where skb address and DMA mappings 919 * are saved. 920 */ 921 if (!tcp_data) { 922 sqe_hdr->pnc = 1; 923 sqe_hdr->sqe_id = first_sqe; 924 sq->sg[first_sqe].skb = (u64)skb; 925 } else { 926 sqe_hdr->pnc = 0; 927 } 928 929 sqe_hdr->sizem1 = (offset / 16) - 1; 930 931 /* Flush SQE to HW */ 932 pfvf->hw_ops->sqe_flush(pfvf, sq, offset, qidx); 933 } 934 } 935 936 static bool is_hw_tso_supported(struct otx2_nic *pfvf, 937 struct sk_buff *skb) 938 { 939 int payload_len, last_seg_size; 940 941 if (test_bit(HW_TSO, &pfvf->hw.cap_flag)) 942 return true; 943 944 /* On 96xx A0, HW TSO not supported */ 945 if (!is_96xx_B0(pfvf->pdev)) 946 return false; 947 948 /* HW has an issue due to which when the payload of the last LSO 949 * segment is shorter than 16 bytes, some header fields may not 950 * be correctly modified, hence don't offload such TSO segments. 951 */ 952 953 payload_len = skb->len - skb_tcp_all_headers(skb); 954 last_seg_size = payload_len % skb_shinfo(skb)->gso_size; 955 if (last_seg_size && last_seg_size < 16) 956 return false; 957 958 return true; 959 } 960 961 static int otx2_get_sqe_count(struct otx2_nic *pfvf, struct sk_buff *skb) 962 { 963 if (!skb_shinfo(skb)->gso_size) 964 return 1; 965 966 /* HW TSO */ 967 if (is_hw_tso_supported(pfvf, skb)) 968 return 1; 969 970 /* SW TSO */ 971 return skb_shinfo(skb)->gso_segs; 972 } 973 974 static bool otx2_validate_network_transport(struct sk_buff *skb) 975 { 976 if ((ip_hdr(skb)->protocol == IPPROTO_UDP) || 977 (ipv6_hdr(skb)->nexthdr == IPPROTO_UDP)) { 978 struct udphdr *udph = udp_hdr(skb); 979 980 if (udph->source == htons(PTP_PORT) && 981 udph->dest == htons(PTP_PORT)) 982 return true; 983 } 984 985 return false; 986 } 987 988 static bool otx2_ptp_is_sync(struct sk_buff *skb, int *offset, bool *udp_csum_crt) 989 { 990 struct ethhdr *eth = (struct ethhdr *)(skb->data); 991 u16 nix_offload_hlen = 0, inner_vhlen = 0; 992 bool udp_hdr_present = false, is_sync; 993 u8 *data = skb->data, *msgtype; 994 __be16 proto = eth->h_proto; 995 int network_depth = 0; 996 997 /* NIX is programmed to offload outer VLAN header 998 * in case of single vlan protocol field holds Network header ETH_IP/V6 999 * in case of stacked vlan protocol field holds Inner vlan (8100) 1000 */ 1001 if (skb->dev->features & NETIF_F_HW_VLAN_CTAG_TX && 1002 skb->dev->features & NETIF_F_HW_VLAN_STAG_TX) { 1003 if (skb->vlan_proto == htons(ETH_P_8021AD)) { 1004 /* Get vlan protocol */ 1005 proto = __vlan_get_protocol(skb, eth->h_proto, NULL); 1006 /* SKB APIs like skb_transport_offset does not include 1007 * offloaded vlan header length. Need to explicitly add 1008 * the length 1009 */ 1010 nix_offload_hlen = VLAN_HLEN; 1011 inner_vhlen = VLAN_HLEN; 1012 } else if (skb->vlan_proto == htons(ETH_P_8021Q)) { 1013 nix_offload_hlen = VLAN_HLEN; 1014 } 1015 } else if (eth_type_vlan(eth->h_proto)) { 1016 proto = __vlan_get_protocol(skb, eth->h_proto, &network_depth); 1017 } 1018 1019 switch (ntohs(proto)) { 1020 case ETH_P_1588: 1021 if (network_depth) 1022 *offset = network_depth; 1023 else 1024 *offset = ETH_HLEN + nix_offload_hlen + 1025 inner_vhlen; 1026 break; 1027 case ETH_P_IP: 1028 case ETH_P_IPV6: 1029 if (!otx2_validate_network_transport(skb)) 1030 return false; 1031 1032 *offset = nix_offload_hlen + skb_transport_offset(skb) + 1033 sizeof(struct udphdr); 1034 udp_hdr_present = true; 1035 1036 } 1037 1038 msgtype = data + *offset; 1039 /* Check PTP messageId is SYNC or not */ 1040 is_sync = !(*msgtype & 0xf); 1041 if (is_sync) 1042 *udp_csum_crt = udp_hdr_present; 1043 else 1044 *offset = 0; 1045 1046 return is_sync; 1047 } 1048 1049 static void otx2_set_txtstamp(struct otx2_nic *pfvf, struct sk_buff *skb, 1050 struct otx2_snd_queue *sq, int *offset) 1051 { 1052 struct ethhdr *eth = (struct ethhdr *)(skb->data); 1053 struct ptpv2_tstamp *origin_tstamp; 1054 bool udp_csum_crt = false; 1055 unsigned int udphoff; 1056 struct timespec64 ts; 1057 int ptp_offset = 0; 1058 __wsum skb_csum; 1059 u64 iova; 1060 1061 if (unlikely(!skb_shinfo(skb)->gso_size && 1062 (skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))) { 1063 if (unlikely(pfvf->flags & OTX2_FLAG_PTP_ONESTEP_SYNC && 1064 otx2_ptp_is_sync(skb, &ptp_offset, &udp_csum_crt))) { 1065 origin_tstamp = (struct ptpv2_tstamp *) 1066 ((u8 *)skb->data + ptp_offset + 1067 PTP_SYNC_SEC_OFFSET); 1068 ts = ns_to_timespec64(pfvf->ptp->tstamp); 1069 origin_tstamp->seconds_msb = htons((ts.tv_sec >> 32) & 0xffff); 1070 origin_tstamp->seconds_lsb = htonl(ts.tv_sec & 0xffffffff); 1071 origin_tstamp->nanoseconds = htonl(ts.tv_nsec); 1072 /* Point to correction field in PTP packet */ 1073 ptp_offset += 8; 1074 1075 /* When user disables hw checksum, stack calculates the csum, 1076 * but it does not cover ptp timestamp which is added later. 1077 * Recalculate the checksum manually considering the timestamp. 1078 */ 1079 if (udp_csum_crt) { 1080 struct udphdr *uh = udp_hdr(skb); 1081 1082 if (skb->ip_summed != CHECKSUM_PARTIAL && uh->check != 0) { 1083 udphoff = skb_transport_offset(skb); 1084 uh->check = 0; 1085 skb_csum = skb_checksum(skb, udphoff, skb->len - udphoff, 1086 0); 1087 if (ntohs(eth->h_proto) == ETH_P_IPV6) 1088 uh->check = csum_ipv6_magic(&ipv6_hdr(skb)->saddr, 1089 &ipv6_hdr(skb)->daddr, 1090 skb->len - udphoff, 1091 ipv6_hdr(skb)->nexthdr, 1092 skb_csum); 1093 else 1094 uh->check = csum_tcpudp_magic(ip_hdr(skb)->saddr, 1095 ip_hdr(skb)->daddr, 1096 skb->len - udphoff, 1097 IPPROTO_UDP, 1098 skb_csum); 1099 } 1100 } 1101 } else { 1102 skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS; 1103 } 1104 iova = sq->timestamps->iova + (sq->head * sizeof(u64)); 1105 otx2_sqe_add_mem(sq, offset, NIX_SENDMEMALG_E_SETTSTMP, iova, 1106 ptp_offset, pfvf->ptp->base_ns, udp_csum_crt); 1107 } else { 1108 skb_tx_timestamp(skb); 1109 } 1110 } 1111 1112 bool otx2_sq_append_skb(struct net_device *netdev, struct otx2_snd_queue *sq, 1113 struct sk_buff *skb, u16 qidx) 1114 { 1115 struct netdev_queue *txq = netdev_get_tx_queue(netdev, qidx); 1116 struct otx2_nic *pfvf = netdev_priv(netdev); 1117 int offset, num_segs, free_desc; 1118 struct nix_sqe_hdr_s *sqe_hdr; 1119 1120 /* Check if there is enough room between producer 1121 * and consumer index. 1122 */ 1123 free_desc = (sq->cons_head - sq->head - 1 + sq->sqe_cnt) & (sq->sqe_cnt - 1); 1124 if (free_desc < sq->sqe_thresh) 1125 return false; 1126 1127 if (free_desc < otx2_get_sqe_count(pfvf, skb)) 1128 return false; 1129 1130 num_segs = skb_shinfo(skb)->nr_frags + 1; 1131 1132 /* If SKB doesn't fit in a single SQE, linearize it. 1133 * TODO: Consider adding JUMP descriptor instead. 1134 */ 1135 if (unlikely(num_segs > OTX2_MAX_FRAGS_IN_SQE)) { 1136 if (__skb_linearize(skb)) { 1137 dev_kfree_skb_any(skb); 1138 return true; 1139 } 1140 num_segs = skb_shinfo(skb)->nr_frags + 1; 1141 } 1142 1143 if (skb_shinfo(skb)->gso_size && !is_hw_tso_supported(pfvf, skb)) { 1144 /* Insert vlan tag before giving pkt to tso */ 1145 if (skb_vlan_tag_present(skb)) 1146 skb = __vlan_hwaccel_push_inside(skb); 1147 otx2_sq_append_tso(pfvf, sq, skb, qidx); 1148 return true; 1149 } 1150 1151 /* Set SQE's SEND_HDR. 1152 * Do not clear the first 64bit as it contains constant info. 1153 */ 1154 memset(sq->sqe_base + 8, 0, sq->sqe_size - 8); 1155 sqe_hdr = (struct nix_sqe_hdr_s *)(sq->sqe_base); 1156 otx2_sqe_add_hdr(pfvf, sq, sqe_hdr, skb, qidx); 1157 offset = sizeof(*sqe_hdr); 1158 1159 /* Add extended header if needed */ 1160 otx2_sqe_add_ext(pfvf, sq, skb, &offset); 1161 1162 /* Add SG subdesc with data frags */ 1163 if (!otx2_sqe_add_sg(pfvf, sq, skb, num_segs, &offset)) { 1164 otx2_dma_unmap_skb_frags(pfvf, &sq->sg[sq->head]); 1165 return false; 1166 } 1167 1168 otx2_set_txtstamp(pfvf, skb, sq, &offset); 1169 1170 sqe_hdr->sizem1 = (offset / 16) - 1; 1171 1172 netdev_tx_sent_queue(txq, skb->len); 1173 1174 /* Flush SQE to HW */ 1175 pfvf->hw_ops->sqe_flush(pfvf, sq, offset, qidx); 1176 1177 return true; 1178 } 1179 EXPORT_SYMBOL(otx2_sq_append_skb); 1180 1181 void otx2_cleanup_rx_cqes(struct otx2_nic *pfvf, struct otx2_cq_queue *cq) 1182 { 1183 struct nix_cqe_rx_s *cqe; 1184 int processed_cqe = 0; 1185 u64 iova, pa; 1186 1187 if (pfvf->xdp_prog) 1188 xdp_rxq_info_unreg(&cq->xdp_rxq); 1189 1190 if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) 1191 return; 1192 1193 while (cq->pend_cqe) { 1194 cqe = (struct nix_cqe_rx_s *)otx2_get_next_cqe(cq); 1195 processed_cqe++; 1196 cq->pend_cqe--; 1197 1198 if (!cqe) 1199 continue; 1200 if (cqe->sg.segs > 1) { 1201 otx2_free_rcv_seg(pfvf, cqe, cq->cq_idx); 1202 continue; 1203 } 1204 iova = cqe->sg.seg_addr - OTX2_HEAD_ROOM; 1205 pa = otx2_iova_to_phys(pfvf->iommu_domain, iova); 1206 otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize, DMA_FROM_DEVICE); 1207 put_page(virt_to_page(phys_to_virt(pa))); 1208 } 1209 1210 /* Free CQEs to HW */ 1211 otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR, 1212 ((u64)cq->cq_idx << 32) | processed_cqe); 1213 } 1214 1215 void otx2_cleanup_tx_cqes(struct otx2_nic *pfvf, struct otx2_cq_queue *cq) 1216 { 1217 struct sk_buff *skb = NULL; 1218 struct otx2_snd_queue *sq; 1219 struct nix_cqe_tx_s *cqe; 1220 int processed_cqe = 0; 1221 struct sg_list *sg; 1222 1223 sq = &pfvf->qset.sq[cq->cint_idx]; 1224 1225 if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) 1226 return; 1227 1228 while (cq->pend_cqe) { 1229 cqe = (struct nix_cqe_tx_s *)otx2_get_next_cqe(cq); 1230 processed_cqe++; 1231 cq->pend_cqe--; 1232 1233 if (!cqe) 1234 continue; 1235 sg = &sq->sg[cqe->comp.sqe_id]; 1236 skb = (struct sk_buff *)sg->skb; 1237 if (skb) { 1238 otx2_dma_unmap_skb_frags(pfvf, sg); 1239 dev_kfree_skb_any(skb); 1240 sg->skb = (u64)NULL; 1241 } 1242 } 1243 1244 /* Free CQEs to HW */ 1245 otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR, 1246 ((u64)cq->cq_idx << 32) | processed_cqe); 1247 } 1248 1249 int otx2_rxtx_enable(struct otx2_nic *pfvf, bool enable) 1250 { 1251 struct msg_req *msg; 1252 int err; 1253 1254 mutex_lock(&pfvf->mbox.lock); 1255 if (enable) 1256 msg = otx2_mbox_alloc_msg_nix_lf_start_rx(&pfvf->mbox); 1257 else 1258 msg = otx2_mbox_alloc_msg_nix_lf_stop_rx(&pfvf->mbox); 1259 1260 if (!msg) { 1261 mutex_unlock(&pfvf->mbox.lock); 1262 return -ENOMEM; 1263 } 1264 1265 err = otx2_sync_mbox_msg(&pfvf->mbox); 1266 mutex_unlock(&pfvf->mbox.lock); 1267 return err; 1268 } 1269 1270 static void otx2_xdp_sqe_add_sg(struct otx2_snd_queue *sq, u64 dma_addr, 1271 int len, int *offset) 1272 { 1273 struct nix_sqe_sg_s *sg = NULL; 1274 u64 *iova = NULL; 1275 1276 sg = (struct nix_sqe_sg_s *)(sq->sqe_base + *offset); 1277 sg->ld_type = NIX_SEND_LDTYPE_LDD; 1278 sg->subdc = NIX_SUBDC_SG; 1279 sg->segs = 1; 1280 sg->seg1_size = len; 1281 iova = (void *)sg + sizeof(*sg); 1282 *iova = dma_addr; 1283 *offset += sizeof(*sg) + sizeof(u64); 1284 1285 sq->sg[sq->head].dma_addr[0] = dma_addr; 1286 sq->sg[sq->head].size[0] = len; 1287 sq->sg[sq->head].num_segs = 1; 1288 } 1289 1290 bool otx2_xdp_sq_append_pkt(struct otx2_nic *pfvf, u64 iova, int len, u16 qidx) 1291 { 1292 struct nix_sqe_hdr_s *sqe_hdr; 1293 struct otx2_snd_queue *sq; 1294 int offset, free_sqe; 1295 1296 sq = &pfvf->qset.sq[qidx]; 1297 free_sqe = (sq->num_sqbs - *sq->aura_fc_addr) * sq->sqe_per_sqb; 1298 if (free_sqe < sq->sqe_thresh) 1299 return false; 1300 1301 memset(sq->sqe_base + 8, 0, sq->sqe_size - 8); 1302 1303 sqe_hdr = (struct nix_sqe_hdr_s *)(sq->sqe_base); 1304 1305 if (!sqe_hdr->total) { 1306 sqe_hdr->aura = sq->aura_id; 1307 sqe_hdr->df = 1; 1308 sqe_hdr->sq = qidx; 1309 sqe_hdr->pnc = 1; 1310 } 1311 sqe_hdr->total = len; 1312 sqe_hdr->sqe_id = sq->head; 1313 1314 offset = sizeof(*sqe_hdr); 1315 1316 otx2_xdp_sqe_add_sg(sq, iova, len, &offset); 1317 sqe_hdr->sizem1 = (offset / 16) - 1; 1318 pfvf->hw_ops->sqe_flush(pfvf, sq, offset, qidx); 1319 1320 return true; 1321 } 1322 1323 static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf, 1324 struct bpf_prog *prog, 1325 struct nix_cqe_rx_s *cqe, 1326 struct otx2_cq_queue *cq) 1327 { 1328 unsigned char *hard_start, *data; 1329 int qidx = cq->cq_idx; 1330 struct xdp_buff xdp; 1331 struct page *page; 1332 u64 iova, pa; 1333 u32 act; 1334 int err; 1335 1336 iova = cqe->sg.seg_addr - OTX2_HEAD_ROOM; 1337 pa = otx2_iova_to_phys(pfvf->iommu_domain, iova); 1338 page = virt_to_page(phys_to_virt(pa)); 1339 1340 xdp_init_buff(&xdp, pfvf->rbsize, &cq->xdp_rxq); 1341 1342 data = (unsigned char *)phys_to_virt(pa); 1343 hard_start = page_address(page); 1344 xdp_prepare_buff(&xdp, hard_start, data - hard_start, 1345 cqe->sg.seg_size, false); 1346 1347 act = bpf_prog_run_xdp(prog, &xdp); 1348 1349 switch (act) { 1350 case XDP_PASS: 1351 break; 1352 case XDP_TX: 1353 qidx += pfvf->hw.tx_queues; 1354 cq->pool_ptrs++; 1355 return otx2_xdp_sq_append_pkt(pfvf, iova, 1356 cqe->sg.seg_size, qidx); 1357 case XDP_REDIRECT: 1358 cq->pool_ptrs++; 1359 err = xdp_do_redirect(pfvf->netdev, &xdp, prog); 1360 1361 otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize, 1362 DMA_FROM_DEVICE); 1363 if (!err) 1364 return true; 1365 put_page(page); 1366 break; 1367 default: 1368 bpf_warn_invalid_xdp_action(pfvf->netdev, prog, act); 1369 break; 1370 case XDP_ABORTED: 1371 trace_xdp_exception(pfvf->netdev, prog, act); 1372 break; 1373 case XDP_DROP: 1374 otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize, 1375 DMA_FROM_DEVICE); 1376 put_page(page); 1377 cq->pool_ptrs++; 1378 return true; 1379 } 1380 return false; 1381 } 1382