1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Huawei HiNIC PCI Express Linux driver 4 * Copyright(c) 2017 Huawei Technologies Co., Ltd 5 */ 6 7 #include <linux/kernel.h> 8 #include <linux/types.h> 9 #include <linux/errno.h> 10 #include <linux/pci.h> 11 #include <linux/device.h> 12 #include <linux/netdevice.h> 13 #include <linux/etherdevice.h> 14 #include <linux/u64_stats_sync.h> 15 #include <linux/slab.h> 16 #include <linux/interrupt.h> 17 #include <linux/skbuff.h> 18 #include <linux/dma-mapping.h> 19 #include <linux/prefetch.h> 20 #include <linux/cpumask.h> 21 #include <linux/if_vlan.h> 22 #include <asm/barrier.h> 23 24 #include "hinic_common.h" 25 #include "hinic_hw_if.h" 26 #include "hinic_hw_wqe.h" 27 #include "hinic_hw_wq.h" 28 #include "hinic_hw_qp.h" 29 #include "hinic_hw_dev.h" 30 #include "hinic_rx.h" 31 #include "hinic_dev.h" 32 33 #define RX_IRQ_NO_PENDING 0 34 #define RX_IRQ_NO_COALESC 0 35 #define RX_IRQ_NO_LLI_TIMER 0 36 #define RX_IRQ_NO_CREDIT 0 37 #define RX_IRQ_NO_RESEND_TIMER 0 38 #define HINIC_RX_BUFFER_WRITE 16 39 40 #define HINIC_RX_IPV6_PKT 7 41 #define LRO_PKT_HDR_LEN_IPV4 66 42 #define LRO_PKT_HDR_LEN_IPV6 86 43 #define LRO_REPLENISH_THLD 256 44 45 #define LRO_PKT_HDR_LEN(cqe) \ 46 (HINIC_GET_RX_PKT_TYPE(be32_to_cpu((cqe)->offload_type)) == \ 47 HINIC_RX_IPV6_PKT ? LRO_PKT_HDR_LEN_IPV6 : LRO_PKT_HDR_LEN_IPV4) 48 49 /** 50 * hinic_rxq_clean_stats - Clean the statistics of specific queue 51 * @rxq: Logical Rx Queue 52 **/ 53 void hinic_rxq_clean_stats(struct hinic_rxq *rxq) 54 { 55 struct hinic_rxq_stats *rxq_stats = &rxq->rxq_stats; 56 57 u64_stats_update_begin(&rxq_stats->syncp); 58 rxq_stats->pkts = 0; 59 rxq_stats->bytes = 0; 60 rxq_stats->errors = 0; 61 rxq_stats->csum_errors = 0; 62 rxq_stats->other_errors = 0; 63 u64_stats_update_end(&rxq_stats->syncp); 64 } 65 66 /** 67 * hinic_rxq_get_stats - get statistics of Rx Queue 68 * @rxq: Logical Rx Queue 69 * @stats: return updated stats here 70 **/ 71 void hinic_rxq_get_stats(struct hinic_rxq *rxq, struct hinic_rxq_stats *stats) 72 { 73 struct hinic_rxq_stats *rxq_stats = &rxq->rxq_stats; 74 unsigned int start; 75 76 u64_stats_update_begin(&stats->syncp); 77 do { 78 start = u64_stats_fetch_begin(&rxq_stats->syncp); 79 stats->pkts = rxq_stats->pkts; 80 stats->bytes = rxq_stats->bytes; 81 stats->errors = rxq_stats->csum_errors + 82 rxq_stats->other_errors; 83 stats->csum_errors = rxq_stats->csum_errors; 84 stats->other_errors = rxq_stats->other_errors; 85 } while (u64_stats_fetch_retry(&rxq_stats->syncp, start)); 86 u64_stats_update_end(&stats->syncp); 87 } 88 89 /** 90 * rxq_stats_init - Initialize the statistics of specific queue 91 * @rxq: Logical Rx Queue 92 **/ 93 static void rxq_stats_init(struct hinic_rxq *rxq) 94 { 95 struct hinic_rxq_stats *rxq_stats = &rxq->rxq_stats; 96 97 u64_stats_init(&rxq_stats->syncp); 98 hinic_rxq_clean_stats(rxq); 99 } 100 101 static void rx_csum(struct hinic_rxq *rxq, u32 status, 102 struct sk_buff *skb) 103 { 104 struct net_device *netdev = rxq->netdev; 105 u32 csum_err; 106 107 csum_err = HINIC_RQ_CQE_STATUS_GET(status, CSUM_ERR); 108 109 if (!(netdev->features & NETIF_F_RXCSUM)) 110 return; 111 112 if (!csum_err) { 113 skb->ip_summed = CHECKSUM_UNNECESSARY; 114 } else { 115 if (!(csum_err & (HINIC_RX_CSUM_HW_CHECK_NONE | 116 HINIC_RX_CSUM_IPSU_OTHER_ERR))) 117 rxq->rxq_stats.csum_errors++; 118 skb->ip_summed = CHECKSUM_NONE; 119 } 120 } 121 122 /** 123 * rx_alloc_skb - allocate skb and map it to dma address 124 * @rxq: rx queue 125 * @dma_addr: returned dma address for the skb 126 * 127 * Return skb 128 **/ 129 static struct sk_buff *rx_alloc_skb(struct hinic_rxq *rxq, 130 dma_addr_t *dma_addr) 131 { 132 struct hinic_dev *nic_dev = netdev_priv(rxq->netdev); 133 struct hinic_hwdev *hwdev = nic_dev->hwdev; 134 struct hinic_hwif *hwif = hwdev->hwif; 135 struct pci_dev *pdev = hwif->pdev; 136 struct sk_buff *skb; 137 dma_addr_t addr; 138 int err; 139 140 skb = netdev_alloc_skb_ip_align(rxq->netdev, rxq->rq->buf_sz); 141 if (!skb) 142 return NULL; 143 144 addr = dma_map_single(&pdev->dev, skb->data, rxq->rq->buf_sz, 145 DMA_FROM_DEVICE); 146 err = dma_mapping_error(&pdev->dev, addr); 147 if (err) { 148 dev_err(&pdev->dev, "Failed to map Rx DMA, err = %d\n", err); 149 goto err_rx_map; 150 } 151 152 *dma_addr = addr; 153 return skb; 154 155 err_rx_map: 156 dev_kfree_skb_any(skb); 157 return NULL; 158 } 159 160 /** 161 * rx_unmap_skb - unmap the dma address of the skb 162 * @rxq: rx queue 163 * @dma_addr: dma address of the skb 164 **/ 165 static void rx_unmap_skb(struct hinic_rxq *rxq, dma_addr_t dma_addr) 166 { 167 struct hinic_dev *nic_dev = netdev_priv(rxq->netdev); 168 struct hinic_hwdev *hwdev = nic_dev->hwdev; 169 struct hinic_hwif *hwif = hwdev->hwif; 170 struct pci_dev *pdev = hwif->pdev; 171 172 dma_unmap_single(&pdev->dev, dma_addr, rxq->rq->buf_sz, 173 DMA_FROM_DEVICE); 174 } 175 176 /** 177 * rx_free_skb - unmap and free skb 178 * @rxq: rx queue 179 * @skb: skb to free 180 * @dma_addr: dma address of the skb 181 **/ 182 static void rx_free_skb(struct hinic_rxq *rxq, struct sk_buff *skb, 183 dma_addr_t dma_addr) 184 { 185 rx_unmap_skb(rxq, dma_addr); 186 dev_kfree_skb_any(skb); 187 } 188 189 /** 190 * rx_alloc_pkts - allocate pkts in rx queue 191 * @rxq: rx queue 192 * 193 * Return number of skbs allocated 194 **/ 195 static int rx_alloc_pkts(struct hinic_rxq *rxq) 196 { 197 struct hinic_dev *nic_dev = netdev_priv(rxq->netdev); 198 struct hinic_rq_wqe *rq_wqe; 199 unsigned int free_wqebbs; 200 struct hinic_sge sge; 201 dma_addr_t dma_addr; 202 struct sk_buff *skb; 203 u16 prod_idx; 204 int i; 205 206 free_wqebbs = hinic_get_rq_free_wqebbs(rxq->rq); 207 208 /* Limit the allocation chunks */ 209 if (free_wqebbs > nic_dev->rx_weight) 210 free_wqebbs = nic_dev->rx_weight; 211 212 for (i = 0; i < free_wqebbs; i++) { 213 skb = rx_alloc_skb(rxq, &dma_addr); 214 if (!skb) 215 goto skb_out; 216 217 hinic_set_sge(&sge, dma_addr, skb->len); 218 219 rq_wqe = hinic_rq_get_wqe(rxq->rq, HINIC_RQ_WQE_SIZE, 220 &prod_idx); 221 if (!rq_wqe) { 222 rx_free_skb(rxq, skb, dma_addr); 223 goto skb_out; 224 } 225 226 hinic_rq_prepare_wqe(rxq->rq, prod_idx, rq_wqe, &sge); 227 228 hinic_rq_write_wqe(rxq->rq, prod_idx, rq_wqe, skb); 229 } 230 231 skb_out: 232 if (i) { 233 wmb(); /* write all the wqes before update PI */ 234 235 hinic_rq_update(rxq->rq, prod_idx); 236 } 237 238 return i; 239 } 240 241 /** 242 * free_all_rx_skbs - free all skbs in rx queue 243 * @rxq: rx queue 244 **/ 245 static void free_all_rx_skbs(struct hinic_rxq *rxq) 246 { 247 struct hinic_rq *rq = rxq->rq; 248 struct hinic_hw_wqe *hw_wqe; 249 struct hinic_sge sge; 250 u16 ci; 251 252 while ((hw_wqe = hinic_read_wqe(rq->wq, HINIC_RQ_WQE_SIZE, &ci))) { 253 if (IS_ERR(hw_wqe)) 254 break; 255 256 hinic_rq_get_sge(rq, &hw_wqe->rq_wqe, ci, &sge); 257 258 hinic_put_wqe(rq->wq, HINIC_RQ_WQE_SIZE); 259 260 rx_free_skb(rxq, rq->saved_skb[ci], hinic_sge_to_dma(&sge)); 261 } 262 } 263 264 /** 265 * rx_recv_jumbo_pkt - Rx handler for jumbo pkt 266 * @rxq: rx queue 267 * @head_skb: the first skb in the list 268 * @left_pkt_len: left size of the pkt exclude head skb 269 * @ci: consumer index 270 * 271 * Return number of wqes that used for the left of the pkt 272 **/ 273 static int rx_recv_jumbo_pkt(struct hinic_rxq *rxq, struct sk_buff *head_skb, 274 unsigned int left_pkt_len, u16 ci) 275 { 276 struct sk_buff *skb, *curr_skb = head_skb; 277 struct hinic_rq_wqe *rq_wqe; 278 unsigned int curr_len; 279 struct hinic_sge sge; 280 int num_wqes = 0; 281 282 while (left_pkt_len > 0) { 283 rq_wqe = hinic_rq_read_next_wqe(rxq->rq, HINIC_RQ_WQE_SIZE, 284 &skb, &ci); 285 286 num_wqes++; 287 288 hinic_rq_get_sge(rxq->rq, rq_wqe, ci, &sge); 289 290 rx_unmap_skb(rxq, hinic_sge_to_dma(&sge)); 291 292 prefetch(skb->data); 293 294 curr_len = (left_pkt_len > HINIC_RX_BUF_SZ) ? HINIC_RX_BUF_SZ : 295 left_pkt_len; 296 297 left_pkt_len -= curr_len; 298 299 __skb_put(skb, curr_len); 300 301 if (curr_skb == head_skb) 302 skb_shinfo(head_skb)->frag_list = skb; 303 else 304 curr_skb->next = skb; 305 306 head_skb->len += skb->len; 307 head_skb->data_len += skb->len; 308 head_skb->truesize += skb->truesize; 309 310 curr_skb = skb; 311 } 312 313 return num_wqes; 314 } 315 316 static void hinic_copy_lp_data(struct hinic_dev *nic_dev, 317 struct sk_buff *skb) 318 { 319 struct net_device *netdev = nic_dev->netdev; 320 u8 *lb_buf = nic_dev->lb_test_rx_buf; 321 int lb_len = nic_dev->lb_pkt_len; 322 int pkt_offset, frag_len, i; 323 void *frag_data = NULL; 324 325 if (nic_dev->lb_test_rx_idx == LP_PKT_CNT) { 326 nic_dev->lb_test_rx_idx = 0; 327 netif_warn(nic_dev, drv, netdev, "Loopback test warning, receive too more test pkts\n"); 328 } 329 330 if (skb->len != nic_dev->lb_pkt_len) { 331 netif_warn(nic_dev, drv, netdev, "Wrong packet length\n"); 332 nic_dev->lb_test_rx_idx++; 333 return; 334 } 335 336 pkt_offset = nic_dev->lb_test_rx_idx * lb_len; 337 frag_len = (int)skb_headlen(skb); 338 memcpy(lb_buf + pkt_offset, skb->data, frag_len); 339 pkt_offset += frag_len; 340 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 341 frag_data = skb_frag_address(&skb_shinfo(skb)->frags[i]); 342 frag_len = (int)skb_frag_size(&skb_shinfo(skb)->frags[i]); 343 memcpy((lb_buf + pkt_offset), frag_data, frag_len); 344 pkt_offset += frag_len; 345 } 346 nic_dev->lb_test_rx_idx++; 347 } 348 349 /** 350 * rxq_recv - Rx handler 351 * @rxq: rx queue 352 * @budget: maximum pkts to process 353 * 354 * Return number of pkts received 355 **/ 356 static int rxq_recv(struct hinic_rxq *rxq, int budget) 357 { 358 struct hinic_qp *qp = container_of(rxq->rq, struct hinic_qp, rq); 359 struct net_device *netdev = rxq->netdev; 360 u64 pkt_len = 0, rx_bytes = 0; 361 struct hinic_rq *rq = rxq->rq; 362 struct hinic_rq_wqe *rq_wqe; 363 struct hinic_dev *nic_dev; 364 unsigned int free_wqebbs; 365 struct hinic_rq_cqe *cqe; 366 int num_wqes, pkts = 0; 367 struct hinic_sge sge; 368 unsigned int status; 369 struct sk_buff *skb; 370 u32 offload_type; 371 u16 ci, num_lro; 372 u16 num_wqe = 0; 373 u32 vlan_len; 374 u16 vid; 375 376 nic_dev = netdev_priv(netdev); 377 378 while (pkts < budget) { 379 num_wqes = 0; 380 381 rq_wqe = hinic_rq_read_wqe(rxq->rq, HINIC_RQ_WQE_SIZE, &skb, 382 &ci); 383 if (!rq_wqe) 384 break; 385 386 /* make sure we read rx_done before packet length */ 387 dma_rmb(); 388 389 cqe = rq->cqe[ci]; 390 status = be32_to_cpu(cqe->status); 391 hinic_rq_get_sge(rxq->rq, rq_wqe, ci, &sge); 392 393 rx_unmap_skb(rxq, hinic_sge_to_dma(&sge)); 394 395 rx_csum(rxq, status, skb); 396 397 prefetch(skb->data); 398 399 pkt_len = sge.len; 400 401 if (pkt_len <= HINIC_RX_BUF_SZ) { 402 __skb_put(skb, pkt_len); 403 } else { 404 __skb_put(skb, HINIC_RX_BUF_SZ); 405 num_wqes = rx_recv_jumbo_pkt(rxq, skb, pkt_len - 406 HINIC_RX_BUF_SZ, ci); 407 } 408 409 hinic_rq_put_wqe(rq, ci, 410 (num_wqes + 1) * HINIC_RQ_WQE_SIZE); 411 412 offload_type = be32_to_cpu(cqe->offload_type); 413 vlan_len = be32_to_cpu(cqe->len); 414 if ((netdev->features & NETIF_F_HW_VLAN_CTAG_RX) && 415 HINIC_GET_RX_VLAN_OFFLOAD_EN(offload_type)) { 416 vid = HINIC_GET_RX_VLAN_TAG(vlan_len); 417 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vid); 418 } 419 420 if (unlikely(nic_dev->flags & HINIC_LP_TEST)) 421 hinic_copy_lp_data(nic_dev, skb); 422 423 skb_record_rx_queue(skb, qp->q_id); 424 skb->protocol = eth_type_trans(skb, rxq->netdev); 425 426 napi_gro_receive(&rxq->napi, skb); 427 428 pkts++; 429 rx_bytes += pkt_len; 430 431 num_lro = HINIC_GET_RX_NUM_LRO(status); 432 if (num_lro) { 433 rx_bytes += ((num_lro - 1) * 434 LRO_PKT_HDR_LEN(cqe)); 435 436 num_wqe += 437 (u16)(pkt_len >> rxq->rx_buff_shift) + 438 ((pkt_len & (rxq->buf_len - 1)) ? 1 : 0); 439 } 440 441 cqe->status = 0; 442 443 if (num_wqe >= LRO_REPLENISH_THLD) 444 break; 445 } 446 447 free_wqebbs = hinic_get_rq_free_wqebbs(rxq->rq); 448 if (free_wqebbs > HINIC_RX_BUFFER_WRITE) 449 rx_alloc_pkts(rxq); 450 451 u64_stats_update_begin(&rxq->rxq_stats.syncp); 452 rxq->rxq_stats.pkts += pkts; 453 rxq->rxq_stats.bytes += rx_bytes; 454 u64_stats_update_end(&rxq->rxq_stats.syncp); 455 456 return pkts; 457 } 458 459 static int rx_poll(struct napi_struct *napi, int budget) 460 { 461 struct hinic_rxq *rxq = container_of(napi, struct hinic_rxq, napi); 462 struct hinic_dev *nic_dev = netdev_priv(rxq->netdev); 463 struct hinic_rq *rq = rxq->rq; 464 int pkts; 465 466 pkts = rxq_recv(rxq, budget); 467 if (pkts >= budget) 468 return budget; 469 470 napi_complete(napi); 471 472 if (!HINIC_IS_VF(nic_dev->hwdev->hwif)) 473 hinic_hwdev_set_msix_state(nic_dev->hwdev, 474 rq->msix_entry, 475 HINIC_MSIX_ENABLE); 476 477 return pkts; 478 } 479 480 static void rx_add_napi(struct hinic_rxq *rxq) 481 { 482 struct hinic_dev *nic_dev = netdev_priv(rxq->netdev); 483 484 netif_napi_add_weight(rxq->netdev, &rxq->napi, rx_poll, 485 nic_dev->rx_weight); 486 napi_enable(&rxq->napi); 487 } 488 489 static void rx_del_napi(struct hinic_rxq *rxq) 490 { 491 napi_disable(&rxq->napi); 492 netif_napi_del(&rxq->napi); 493 } 494 495 static irqreturn_t rx_irq(int irq, void *data) 496 { 497 struct hinic_rxq *rxq = (struct hinic_rxq *)data; 498 struct hinic_rq *rq = rxq->rq; 499 struct hinic_dev *nic_dev; 500 501 /* Disable the interrupt until napi will be completed */ 502 nic_dev = netdev_priv(rxq->netdev); 503 if (!HINIC_IS_VF(nic_dev->hwdev->hwif)) 504 hinic_hwdev_set_msix_state(nic_dev->hwdev, 505 rq->msix_entry, 506 HINIC_MSIX_DISABLE); 507 508 nic_dev = netdev_priv(rxq->netdev); 509 hinic_hwdev_msix_cnt_set(nic_dev->hwdev, rq->msix_entry); 510 511 napi_schedule(&rxq->napi); 512 return IRQ_HANDLED; 513 } 514 515 static int rx_request_irq(struct hinic_rxq *rxq) 516 { 517 struct hinic_dev *nic_dev = netdev_priv(rxq->netdev); 518 struct hinic_msix_config interrupt_info = {0}; 519 struct hinic_intr_coal_info *intr_coal = NULL; 520 struct hinic_hwdev *hwdev = nic_dev->hwdev; 521 struct hinic_rq *rq = rxq->rq; 522 struct hinic_qp *qp; 523 int err; 524 525 qp = container_of(rq, struct hinic_qp, rq); 526 527 rx_add_napi(rxq); 528 529 hinic_hwdev_msix_set(hwdev, rq->msix_entry, 530 RX_IRQ_NO_PENDING, RX_IRQ_NO_COALESC, 531 RX_IRQ_NO_LLI_TIMER, RX_IRQ_NO_CREDIT, 532 RX_IRQ_NO_RESEND_TIMER); 533 534 intr_coal = &nic_dev->rx_intr_coalesce[qp->q_id]; 535 interrupt_info.msix_index = rq->msix_entry; 536 interrupt_info.coalesce_timer_cnt = intr_coal->coalesce_timer_cfg; 537 interrupt_info.pending_cnt = intr_coal->pending_limt; 538 interrupt_info.resend_timer_cnt = intr_coal->resend_timer_cfg; 539 540 err = hinic_set_interrupt_cfg(hwdev, &interrupt_info); 541 if (err) { 542 netif_err(nic_dev, drv, rxq->netdev, 543 "Failed to set RX interrupt coalescing attribute\n"); 544 goto err_req_irq; 545 } 546 547 err = request_irq(rq->irq, rx_irq, 0, rxq->irq_name, rxq); 548 if (err) 549 goto err_req_irq; 550 551 cpumask_set_cpu(qp->q_id % num_online_cpus(), &rq->affinity_mask); 552 err = irq_set_affinity_and_hint(rq->irq, &rq->affinity_mask); 553 if (err) 554 goto err_irq_affinity; 555 556 return 0; 557 558 err_irq_affinity: 559 free_irq(rq->irq, rxq); 560 err_req_irq: 561 rx_del_napi(rxq); 562 return err; 563 } 564 565 static void rx_free_irq(struct hinic_rxq *rxq) 566 { 567 struct hinic_rq *rq = rxq->rq; 568 569 irq_update_affinity_hint(rq->irq, NULL); 570 free_irq(rq->irq, rxq); 571 rx_del_napi(rxq); 572 } 573 574 /** 575 * hinic_init_rxq - Initialize the Rx Queue 576 * @rxq: Logical Rx Queue 577 * @rq: Hardware Rx Queue to connect the Logical queue with 578 * @netdev: network device to connect the Logical queue with 579 * 580 * Return 0 - Success, negative - Failure 581 **/ 582 int hinic_init_rxq(struct hinic_rxq *rxq, struct hinic_rq *rq, 583 struct net_device *netdev) 584 { 585 struct hinic_qp *qp = container_of(rq, struct hinic_qp, rq); 586 int err, pkts; 587 588 rxq->netdev = netdev; 589 rxq->rq = rq; 590 rxq->buf_len = HINIC_RX_BUF_SZ; 591 rxq->rx_buff_shift = ilog2(HINIC_RX_BUF_SZ); 592 593 rxq_stats_init(rxq); 594 595 rxq->irq_name = devm_kasprintf(&netdev->dev, GFP_KERNEL, 596 "%s_rxq%d", netdev->name, qp->q_id); 597 if (!rxq->irq_name) 598 return -ENOMEM; 599 600 pkts = rx_alloc_pkts(rxq); 601 if (!pkts) { 602 err = -ENOMEM; 603 goto err_rx_pkts; 604 } 605 606 err = rx_request_irq(rxq); 607 if (err) { 608 netdev_err(netdev, "Failed to request Rx irq\n"); 609 goto err_req_rx_irq; 610 } 611 612 return 0; 613 614 err_req_rx_irq: 615 err_rx_pkts: 616 free_all_rx_skbs(rxq); 617 devm_kfree(&netdev->dev, rxq->irq_name); 618 return err; 619 } 620 621 /** 622 * hinic_clean_rxq - Clean the Rx Queue 623 * @rxq: Logical Rx Queue 624 **/ 625 void hinic_clean_rxq(struct hinic_rxq *rxq) 626 { 627 struct net_device *netdev = rxq->netdev; 628 629 rx_free_irq(rxq); 630 631 free_all_rx_skbs(rxq); 632 devm_kfree(&netdev->dev, rxq->irq_name); 633 } 634