1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-3-Clause) 2 /* QLogic qede NIC Driver 3 * Copyright (c) 2015-2017 QLogic Corporation 4 * Copyright (c) 2019-2020 Marvell International Ltd. 5 */ 6 7 #include <linux/netdevice.h> 8 #include <linux/etherdevice.h> 9 #include <linux/skbuff.h> 10 #include <linux/bpf_trace.h> 11 #include <net/udp_tunnel.h> 12 #include <linux/ip.h> 13 #include <net/ipv6.h> 14 #include <net/tcp.h> 15 #include <linux/if_ether.h> 16 #include <linux/if_vlan.h> 17 #include <net/ip6_checksum.h> 18 #include "qede_ptp.h" 19 20 #include <linux/qed/qed_if.h> 21 #include "qede.h" 22 /********************************* 23 * Content also used by slowpath * 24 *********************************/ 25 26 int qede_alloc_rx_buffer(struct qede_rx_queue *rxq, bool allow_lazy) 27 { 28 struct sw_rx_data *sw_rx_data; 29 struct eth_rx_bd *rx_bd; 30 dma_addr_t mapping; 31 struct page *data; 32 33 /* In case lazy-allocation is allowed, postpone allocation until the 34 * end of the NAPI run. We'd still need to make sure the Rx ring has 35 * sufficient buffers to guarantee an additional Rx interrupt. 36 */ 37 if (allow_lazy && likely(rxq->filled_buffers > 12)) { 38 rxq->filled_buffers--; 39 return 0; 40 } 41 42 data = alloc_pages(GFP_ATOMIC, 0); 43 if (unlikely(!data)) 44 return -ENOMEM; 45 46 /* Map the entire page as it would be used 47 * for multiple RX buffer segment size mapping. 48 */ 49 mapping = dma_map_page(rxq->dev, data, 0, 50 PAGE_SIZE, rxq->data_direction); 51 if (unlikely(dma_mapping_error(rxq->dev, mapping))) { 52 __free_page(data); 53 return -ENOMEM; 54 } 55 56 sw_rx_data = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX]; 57 sw_rx_data->page_offset = 0; 58 sw_rx_data->data = data; 59 sw_rx_data->mapping = mapping; 60 61 /* Advance PROD and get BD pointer */ 62 rx_bd = (struct eth_rx_bd *)qed_chain_produce(&rxq->rx_bd_ring); 63 WARN_ON(!rx_bd); 64 rx_bd->addr.hi = cpu_to_le32(upper_32_bits(mapping)); 65 rx_bd->addr.lo = cpu_to_le32(lower_32_bits(mapping) + 66 rxq->rx_headroom); 67 68 rxq->sw_rx_prod++; 69 rxq->filled_buffers++; 70 71 return 0; 72 } 73 74 /* Unmap the data and free skb */ 75 int qede_free_tx_pkt(struct qede_dev *edev, struct qede_tx_queue *txq, int *len) 76 { 77 u16 idx = txq->sw_tx_cons; 78 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb; 79 struct eth_tx_1st_bd *first_bd; 80 struct eth_tx_bd *tx_data_bd; 81 int bds_consumed = 0; 82 int nbds; 83 bool data_split = txq->sw_tx_ring.skbs[idx].flags & QEDE_TSO_SPLIT_BD; 84 int i, split_bd_len = 0; 85 86 if (unlikely(!skb)) { 87 DP_ERR(edev, 88 "skb is null for txq idx=%d txq->sw_tx_cons=%d txq->sw_tx_prod=%d\n", 89 idx, txq->sw_tx_cons, txq->sw_tx_prod); 90 return -1; 91 } 92 93 *len = skb->len; 94 95 first_bd = (struct eth_tx_1st_bd *)qed_chain_consume(&txq->tx_pbl); 96 97 bds_consumed++; 98 99 nbds = first_bd->data.nbds; 100 101 if (data_split) { 102 struct eth_tx_bd *split = (struct eth_tx_bd *) 103 qed_chain_consume(&txq->tx_pbl); 104 split_bd_len = BD_UNMAP_LEN(split); 105 bds_consumed++; 106 } 107 dma_unmap_single(&edev->pdev->dev, BD_UNMAP_ADDR(first_bd), 108 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE); 109 110 /* Unmap the data of the skb frags */ 111 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, bds_consumed++) { 112 tx_data_bd = (struct eth_tx_bd *) 113 qed_chain_consume(&txq->tx_pbl); 114 dma_unmap_page(&edev->pdev->dev, BD_UNMAP_ADDR(tx_data_bd), 115 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE); 116 } 117 118 while (bds_consumed++ < nbds) 119 qed_chain_consume(&txq->tx_pbl); 120 121 /* Free skb */ 122 dev_kfree_skb_any(skb); 123 txq->sw_tx_ring.skbs[idx].skb = NULL; 124 txq->sw_tx_ring.skbs[idx].flags = 0; 125 126 return 0; 127 } 128 129 /* Unmap the data and free skb when mapping failed during start_xmit */ 130 static void qede_free_failed_tx_pkt(struct qede_tx_queue *txq, 131 struct eth_tx_1st_bd *first_bd, 132 int nbd, bool data_split) 133 { 134 u16 idx = txq->sw_tx_prod; 135 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb; 136 struct eth_tx_bd *tx_data_bd; 137 int i, split_bd_len = 0; 138 139 /* Return prod to its position before this skb was handled */ 140 qed_chain_set_prod(&txq->tx_pbl, 141 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd); 142 143 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl); 144 145 if (data_split) { 146 struct eth_tx_bd *split = (struct eth_tx_bd *) 147 qed_chain_produce(&txq->tx_pbl); 148 split_bd_len = BD_UNMAP_LEN(split); 149 nbd--; 150 } 151 152 dma_unmap_single(txq->dev, BD_UNMAP_ADDR(first_bd), 153 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE); 154 155 /* Unmap the data of the skb frags */ 156 for (i = 0; i < nbd; i++) { 157 tx_data_bd = (struct eth_tx_bd *) 158 qed_chain_produce(&txq->tx_pbl); 159 if (tx_data_bd->nbytes) 160 dma_unmap_page(txq->dev, 161 BD_UNMAP_ADDR(tx_data_bd), 162 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE); 163 } 164 165 /* Return again prod to its position before this skb was handled */ 166 qed_chain_set_prod(&txq->tx_pbl, 167 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd); 168 169 /* Free skb */ 170 dev_kfree_skb_any(skb); 171 txq->sw_tx_ring.skbs[idx].skb = NULL; 172 txq->sw_tx_ring.skbs[idx].flags = 0; 173 } 174 175 static u32 qede_xmit_type(struct sk_buff *skb, int *ipv6_ext) 176 { 177 u32 rc = XMIT_L4_CSUM; 178 __be16 l3_proto; 179 180 if (skb->ip_summed != CHECKSUM_PARTIAL) 181 return XMIT_PLAIN; 182 183 l3_proto = vlan_get_protocol(skb); 184 if (l3_proto == htons(ETH_P_IPV6) && 185 (ipv6_hdr(skb)->nexthdr == NEXTHDR_IPV6)) 186 *ipv6_ext = 1; 187 188 if (skb->encapsulation) { 189 rc |= XMIT_ENC; 190 if (skb_is_gso(skb)) { 191 unsigned short gso_type = skb_shinfo(skb)->gso_type; 192 193 if ((gso_type & SKB_GSO_UDP_TUNNEL_CSUM) || 194 (gso_type & SKB_GSO_GRE_CSUM)) 195 rc |= XMIT_ENC_GSO_L4_CSUM; 196 197 rc |= XMIT_LSO; 198 return rc; 199 } 200 } 201 202 if (skb_is_gso(skb)) 203 rc |= XMIT_LSO; 204 205 return rc; 206 } 207 208 static void qede_set_params_for_ipv6_ext(struct sk_buff *skb, 209 struct eth_tx_2nd_bd *second_bd, 210 struct eth_tx_3rd_bd *third_bd) 211 { 212 u8 l4_proto; 213 u16 bd2_bits1 = 0, bd2_bits2 = 0; 214 215 bd2_bits1 |= (1 << ETH_TX_DATA_2ND_BD_IPV6_EXT_SHIFT); 216 217 bd2_bits2 |= ((((u8 *)skb_transport_header(skb) - skb->data) >> 1) & 218 ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK) 219 << ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_SHIFT; 220 221 bd2_bits1 |= (ETH_L4_PSEUDO_CSUM_CORRECT_LENGTH << 222 ETH_TX_DATA_2ND_BD_L4_PSEUDO_CSUM_MODE_SHIFT); 223 224 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6)) 225 l4_proto = ipv6_hdr(skb)->nexthdr; 226 else 227 l4_proto = ip_hdr(skb)->protocol; 228 229 if (l4_proto == IPPROTO_UDP) 230 bd2_bits1 |= 1 << ETH_TX_DATA_2ND_BD_L4_UDP_SHIFT; 231 232 if (third_bd) 233 third_bd->data.bitfields |= 234 cpu_to_le16(((tcp_hdrlen(skb) / 4) & 235 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_MASK) << 236 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_SHIFT); 237 238 second_bd->data.bitfields1 = cpu_to_le16(bd2_bits1); 239 second_bd->data.bitfields2 = cpu_to_le16(bd2_bits2); 240 } 241 242 static int map_frag_to_bd(struct qede_tx_queue *txq, 243 skb_frag_t *frag, struct eth_tx_bd *bd) 244 { 245 dma_addr_t mapping; 246 247 /* Map skb non-linear frag data for DMA */ 248 mapping = skb_frag_dma_map(txq->dev, frag, 0, 249 skb_frag_size(frag), DMA_TO_DEVICE); 250 if (unlikely(dma_mapping_error(txq->dev, mapping))) 251 return -ENOMEM; 252 253 /* Setup the data pointer of the frag data */ 254 BD_SET_UNMAP_ADDR_LEN(bd, mapping, skb_frag_size(frag)); 255 256 return 0; 257 } 258 259 static u16 qede_get_skb_hlen(struct sk_buff *skb, bool is_encap_pkt) 260 { 261 if (is_encap_pkt) 262 return (skb_inner_transport_header(skb) + 263 inner_tcp_hdrlen(skb) - skb->data); 264 else 265 return (skb_transport_header(skb) + 266 tcp_hdrlen(skb) - skb->data); 267 } 268 269 /* +2 for 1st BD for headers and 2nd BD for headlen (if required) */ 270 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET) 271 static bool qede_pkt_req_lin(struct sk_buff *skb, u8 xmit_type) 272 { 273 int allowed_frags = ETH_TX_MAX_BDS_PER_NON_LSO_PACKET - 1; 274 275 if (xmit_type & XMIT_LSO) { 276 int hlen; 277 278 hlen = qede_get_skb_hlen(skb, xmit_type & XMIT_ENC); 279 280 /* linear payload would require its own BD */ 281 if (skb_headlen(skb) > hlen) 282 allowed_frags--; 283 } 284 285 return (skb_shinfo(skb)->nr_frags > allowed_frags); 286 } 287 #endif 288 289 static inline void qede_update_tx_producer(struct qede_tx_queue *txq) 290 { 291 /* wmb makes sure that the BDs data is updated before updating the 292 * producer, otherwise FW may read old data from the BDs. 293 */ 294 wmb(); 295 barrier(); 296 writel(txq->tx_db.raw, txq->doorbell_addr); 297 298 /* Fence required to flush the write combined buffer, since another 299 * CPU may write to the same doorbell address and data may be lost 300 * due to relaxed order nature of write combined bar. 301 */ 302 wmb(); 303 } 304 305 static int qede_xdp_xmit(struct qede_dev *edev, struct qede_fastpath *fp, 306 struct sw_rx_data *metadata, u16 padding, u16 length) 307 { 308 struct qede_tx_queue *txq = fp->xdp_tx; 309 struct eth_tx_1st_bd *first_bd; 310 u16 idx = txq->sw_tx_prod; 311 u16 val; 312 313 if (!qed_chain_get_elem_left(&txq->tx_pbl)) { 314 txq->stopped_cnt++; 315 return -ENOMEM; 316 } 317 318 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl); 319 320 memset(first_bd, 0, sizeof(*first_bd)); 321 first_bd->data.bd_flags.bitfields = 322 BIT(ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT); 323 324 val = (length & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) << 325 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT; 326 327 first_bd->data.bitfields |= cpu_to_le16(val); 328 first_bd->data.nbds = 1; 329 330 /* We can safely ignore the offset, as it's 0 for XDP */ 331 BD_SET_UNMAP_ADDR_LEN(first_bd, metadata->mapping + padding, length); 332 333 /* Synchronize the buffer back to device, as program [probably] 334 * has changed it. 335 */ 336 dma_sync_single_for_device(&edev->pdev->dev, 337 metadata->mapping + padding, 338 length, PCI_DMA_TODEVICE); 339 340 txq->sw_tx_ring.xdp[idx].page = metadata->data; 341 txq->sw_tx_ring.xdp[idx].mapping = metadata->mapping; 342 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers; 343 344 /* Mark the fastpath for future XDP doorbell */ 345 fp->xdp_xmit = 1; 346 347 return 0; 348 } 349 350 int qede_txq_has_work(struct qede_tx_queue *txq) 351 { 352 u16 hw_bd_cons; 353 354 /* Tell compiler that consumer and producer can change */ 355 barrier(); 356 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 357 if (qed_chain_get_cons_idx(&txq->tx_pbl) == hw_bd_cons + 1) 358 return 0; 359 360 return hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl); 361 } 362 363 static void qede_xdp_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq) 364 { 365 u16 hw_bd_cons, idx; 366 367 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 368 barrier(); 369 370 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) { 371 qed_chain_consume(&txq->tx_pbl); 372 idx = txq->sw_tx_cons; 373 374 dma_unmap_page(&edev->pdev->dev, 375 txq->sw_tx_ring.xdp[idx].mapping, 376 PAGE_SIZE, DMA_BIDIRECTIONAL); 377 __free_page(txq->sw_tx_ring.xdp[idx].page); 378 379 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers; 380 txq->xmit_pkts++; 381 } 382 } 383 384 static int qede_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq) 385 { 386 unsigned int pkts_compl = 0, bytes_compl = 0; 387 struct netdev_queue *netdev_txq; 388 u16 hw_bd_cons; 389 int rc; 390 391 netdev_txq = netdev_get_tx_queue(edev->ndev, txq->ndev_txq_id); 392 393 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 394 barrier(); 395 396 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) { 397 int len = 0; 398 399 rc = qede_free_tx_pkt(edev, txq, &len); 400 if (rc) { 401 DP_NOTICE(edev, "hw_bd_cons = %d, chain_cons=%d\n", 402 hw_bd_cons, 403 qed_chain_get_cons_idx(&txq->tx_pbl)); 404 break; 405 } 406 407 bytes_compl += len; 408 pkts_compl++; 409 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers; 410 txq->xmit_pkts++; 411 } 412 413 netdev_tx_completed_queue(netdev_txq, pkts_compl, bytes_compl); 414 415 /* Need to make the tx_bd_cons update visible to start_xmit() 416 * before checking for netif_tx_queue_stopped(). Without the 417 * memory barrier, there is a small possibility that 418 * start_xmit() will miss it and cause the queue to be stopped 419 * forever. 420 * On the other hand we need an rmb() here to ensure the proper 421 * ordering of bit testing in the following 422 * netif_tx_queue_stopped(txq) call. 423 */ 424 smp_mb(); 425 426 if (unlikely(netif_tx_queue_stopped(netdev_txq))) { 427 /* Taking tx_lock is needed to prevent reenabling the queue 428 * while it's empty. This could have happen if rx_action() gets 429 * suspended in qede_tx_int() after the condition before 430 * netif_tx_wake_queue(), while tx_action (qede_start_xmit()): 431 * 432 * stops the queue->sees fresh tx_bd_cons->releases the queue-> 433 * sends some packets consuming the whole queue again-> 434 * stops the queue 435 */ 436 437 __netif_tx_lock(netdev_txq, smp_processor_id()); 438 439 if ((netif_tx_queue_stopped(netdev_txq)) && 440 (edev->state == QEDE_STATE_OPEN) && 441 (qed_chain_get_elem_left(&txq->tx_pbl) 442 >= (MAX_SKB_FRAGS + 1))) { 443 netif_tx_wake_queue(netdev_txq); 444 DP_VERBOSE(edev, NETIF_MSG_TX_DONE, 445 "Wake queue was called\n"); 446 } 447 448 __netif_tx_unlock(netdev_txq); 449 } 450 451 return 0; 452 } 453 454 bool qede_has_rx_work(struct qede_rx_queue *rxq) 455 { 456 u16 hw_comp_cons, sw_comp_cons; 457 458 /* Tell compiler that status block fields can change */ 459 barrier(); 460 461 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr); 462 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 463 464 return hw_comp_cons != sw_comp_cons; 465 } 466 467 static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq) 468 { 469 qed_chain_consume(&rxq->rx_bd_ring); 470 rxq->sw_rx_cons++; 471 } 472 473 /* This function reuses the buffer(from an offset) from 474 * consumer index to producer index in the bd ring 475 */ 476 static inline void qede_reuse_page(struct qede_rx_queue *rxq, 477 struct sw_rx_data *curr_cons) 478 { 479 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring); 480 struct sw_rx_data *curr_prod; 481 dma_addr_t new_mapping; 482 483 curr_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX]; 484 *curr_prod = *curr_cons; 485 486 new_mapping = curr_prod->mapping + curr_prod->page_offset; 487 488 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping)); 489 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping) + 490 rxq->rx_headroom); 491 492 rxq->sw_rx_prod++; 493 curr_cons->data = NULL; 494 } 495 496 /* In case of allocation failures reuse buffers 497 * from consumer index to produce buffers for firmware 498 */ 499 void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count) 500 { 501 struct sw_rx_data *curr_cons; 502 503 for (; count > 0; count--) { 504 curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX]; 505 qede_reuse_page(rxq, curr_cons); 506 qede_rx_bd_ring_consume(rxq); 507 } 508 } 509 510 static inline int qede_realloc_rx_buffer(struct qede_rx_queue *rxq, 511 struct sw_rx_data *curr_cons) 512 { 513 /* Move to the next segment in the page */ 514 curr_cons->page_offset += rxq->rx_buf_seg_size; 515 516 if (curr_cons->page_offset == PAGE_SIZE) { 517 if (unlikely(qede_alloc_rx_buffer(rxq, true))) { 518 /* Since we failed to allocate new buffer 519 * current buffer can be used again. 520 */ 521 curr_cons->page_offset -= rxq->rx_buf_seg_size; 522 523 return -ENOMEM; 524 } 525 526 dma_unmap_page(rxq->dev, curr_cons->mapping, 527 PAGE_SIZE, rxq->data_direction); 528 } else { 529 /* Increment refcount of the page as we don't want 530 * network stack to take the ownership of the page 531 * which can be recycled multiple times by the driver. 532 */ 533 page_ref_inc(curr_cons->data); 534 qede_reuse_page(rxq, curr_cons); 535 } 536 537 return 0; 538 } 539 540 void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq) 541 { 542 u16 bd_prod = qed_chain_get_prod_idx(&rxq->rx_bd_ring); 543 u16 cqe_prod = qed_chain_get_prod_idx(&rxq->rx_comp_ring); 544 struct eth_rx_prod_data rx_prods = {0}; 545 546 /* Update producers */ 547 rx_prods.bd_prod = cpu_to_le16(bd_prod); 548 rx_prods.cqe_prod = cpu_to_le16(cqe_prod); 549 550 /* Make sure that the BD and SGE data is updated before updating the 551 * producers since FW might read the BD/SGE right after the producer 552 * is updated. 553 */ 554 wmb(); 555 556 internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods), 557 (u32 *)&rx_prods); 558 } 559 560 static void qede_get_rxhash(struct sk_buff *skb, u8 bitfields, __le32 rss_hash) 561 { 562 enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE; 563 enum rss_hash_type htype; 564 u32 hash = 0; 565 566 htype = GET_FIELD(bitfields, ETH_FAST_PATH_RX_REG_CQE_RSS_HASH_TYPE); 567 if (htype) { 568 hash_type = ((htype == RSS_HASH_TYPE_IPV4) || 569 (htype == RSS_HASH_TYPE_IPV6)) ? 570 PKT_HASH_TYPE_L3 : PKT_HASH_TYPE_L4; 571 hash = le32_to_cpu(rss_hash); 572 } 573 skb_set_hash(skb, hash, hash_type); 574 } 575 576 static void qede_set_skb_csum(struct sk_buff *skb, u8 csum_flag) 577 { 578 skb_checksum_none_assert(skb); 579 580 if (csum_flag & QEDE_CSUM_UNNECESSARY) 581 skb->ip_summed = CHECKSUM_UNNECESSARY; 582 583 if (csum_flag & QEDE_TUNN_CSUM_UNNECESSARY) { 584 skb->csum_level = 1; 585 skb->encapsulation = 1; 586 } 587 } 588 589 static inline void qede_skb_receive(struct qede_dev *edev, 590 struct qede_fastpath *fp, 591 struct qede_rx_queue *rxq, 592 struct sk_buff *skb, u16 vlan_tag) 593 { 594 if (vlan_tag) 595 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag); 596 597 napi_gro_receive(&fp->napi, skb); 598 } 599 600 static void qede_set_gro_params(struct qede_dev *edev, 601 struct sk_buff *skb, 602 struct eth_fast_path_rx_tpa_start_cqe *cqe) 603 { 604 u16 parsing_flags = le16_to_cpu(cqe->pars_flags.flags); 605 606 if (((parsing_flags >> PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) & 607 PARSING_AND_ERR_FLAGS_L3TYPE_MASK) == 2) 608 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6; 609 else 610 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4; 611 612 skb_shinfo(skb)->gso_size = __le16_to_cpu(cqe->len_on_first_bd) - 613 cqe->header_len; 614 } 615 616 static int qede_fill_frag_skb(struct qede_dev *edev, 617 struct qede_rx_queue *rxq, 618 u8 tpa_agg_index, u16 len_on_bd) 619 { 620 struct sw_rx_data *current_bd = &rxq->sw_rx_ring[rxq->sw_rx_cons & 621 NUM_RX_BDS_MAX]; 622 struct qede_agg_info *tpa_info = &rxq->tpa_info[tpa_agg_index]; 623 struct sk_buff *skb = tpa_info->skb; 624 625 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START)) 626 goto out; 627 628 /* Add one frag and update the appropriate fields in the skb */ 629 skb_fill_page_desc(skb, tpa_info->frag_id++, 630 current_bd->data, 631 current_bd->page_offset + rxq->rx_headroom, 632 len_on_bd); 633 634 if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) { 635 /* Incr page ref count to reuse on allocation failure 636 * so that it doesn't get freed while freeing SKB. 637 */ 638 page_ref_inc(current_bd->data); 639 goto out; 640 } 641 642 qede_rx_bd_ring_consume(rxq); 643 644 skb->data_len += len_on_bd; 645 skb->truesize += rxq->rx_buf_seg_size; 646 skb->len += len_on_bd; 647 648 return 0; 649 650 out: 651 tpa_info->state = QEDE_AGG_STATE_ERROR; 652 qede_recycle_rx_bd_ring(rxq, 1); 653 654 return -ENOMEM; 655 } 656 657 static bool qede_tunn_exist(u16 flag) 658 { 659 return !!(flag & (PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK << 660 PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT)); 661 } 662 663 static u8 qede_check_tunn_csum(u16 flag) 664 { 665 u16 csum_flag = 0; 666 u8 tcsum = 0; 667 668 if (flag & (PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK << 669 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT)) 670 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK << 671 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT; 672 673 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK << 674 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) { 675 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK << 676 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT; 677 tcsum = QEDE_TUNN_CSUM_UNNECESSARY; 678 } 679 680 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK << 681 PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT | 682 PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK << 683 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT; 684 685 if (csum_flag & flag) 686 return QEDE_CSUM_ERROR; 687 688 return QEDE_CSUM_UNNECESSARY | tcsum; 689 } 690 691 static inline struct sk_buff * 692 qede_build_skb(struct qede_rx_queue *rxq, 693 struct sw_rx_data *bd, u16 len, u16 pad) 694 { 695 struct sk_buff *skb; 696 void *buf; 697 698 buf = page_address(bd->data) + bd->page_offset; 699 skb = build_skb(buf, rxq->rx_buf_seg_size); 700 701 skb_reserve(skb, pad); 702 skb_put(skb, len); 703 704 return skb; 705 } 706 707 static struct sk_buff * 708 qede_tpa_rx_build_skb(struct qede_dev *edev, 709 struct qede_rx_queue *rxq, 710 struct sw_rx_data *bd, u16 len, u16 pad, 711 bool alloc_skb) 712 { 713 struct sk_buff *skb; 714 715 skb = qede_build_skb(rxq, bd, len, pad); 716 bd->page_offset += rxq->rx_buf_seg_size; 717 718 if (bd->page_offset == PAGE_SIZE) { 719 if (unlikely(qede_alloc_rx_buffer(rxq, true))) { 720 DP_NOTICE(edev, 721 "Failed to allocate RX buffer for tpa start\n"); 722 bd->page_offset -= rxq->rx_buf_seg_size; 723 page_ref_inc(bd->data); 724 dev_kfree_skb_any(skb); 725 return NULL; 726 } 727 } else { 728 page_ref_inc(bd->data); 729 qede_reuse_page(rxq, bd); 730 } 731 732 /* We've consumed the first BD and prepared an SKB */ 733 qede_rx_bd_ring_consume(rxq); 734 735 return skb; 736 } 737 738 static struct sk_buff * 739 qede_rx_build_skb(struct qede_dev *edev, 740 struct qede_rx_queue *rxq, 741 struct sw_rx_data *bd, u16 len, u16 pad) 742 { 743 struct sk_buff *skb = NULL; 744 745 /* For smaller frames still need to allocate skb, memcpy 746 * data and benefit in reusing the page segment instead of 747 * un-mapping it. 748 */ 749 if ((len + pad <= edev->rx_copybreak)) { 750 unsigned int offset = bd->page_offset + pad; 751 752 skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE); 753 if (unlikely(!skb)) 754 return NULL; 755 756 skb_reserve(skb, pad); 757 skb_put_data(skb, page_address(bd->data) + offset, len); 758 qede_reuse_page(rxq, bd); 759 goto out; 760 } 761 762 skb = qede_build_skb(rxq, bd, len, pad); 763 764 if (unlikely(qede_realloc_rx_buffer(rxq, bd))) { 765 /* Incr page ref count to reuse on allocation failure so 766 * that it doesn't get freed while freeing SKB [as its 767 * already mapped there]. 768 */ 769 page_ref_inc(bd->data); 770 dev_kfree_skb_any(skb); 771 return NULL; 772 } 773 out: 774 /* We've consumed the first BD and prepared an SKB */ 775 qede_rx_bd_ring_consume(rxq); 776 777 return skb; 778 } 779 780 static void qede_tpa_start(struct qede_dev *edev, 781 struct qede_rx_queue *rxq, 782 struct eth_fast_path_rx_tpa_start_cqe *cqe) 783 { 784 struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index]; 785 struct sw_rx_data *sw_rx_data_cons; 786 u16 pad; 787 788 sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX]; 789 pad = cqe->placement_offset + rxq->rx_headroom; 790 791 tpa_info->skb = qede_tpa_rx_build_skb(edev, rxq, sw_rx_data_cons, 792 le16_to_cpu(cqe->len_on_first_bd), 793 pad, false); 794 tpa_info->buffer.page_offset = sw_rx_data_cons->page_offset; 795 tpa_info->buffer.mapping = sw_rx_data_cons->mapping; 796 797 if (unlikely(!tpa_info->skb)) { 798 DP_NOTICE(edev, "Failed to allocate SKB for gro\n"); 799 800 /* Consume from ring but do not produce since 801 * this might be used by FW still, it will be re-used 802 * at TPA end. 803 */ 804 tpa_info->tpa_start_fail = true; 805 qede_rx_bd_ring_consume(rxq); 806 tpa_info->state = QEDE_AGG_STATE_ERROR; 807 goto cons_buf; 808 } 809 810 tpa_info->frag_id = 0; 811 tpa_info->state = QEDE_AGG_STATE_START; 812 813 if ((le16_to_cpu(cqe->pars_flags.flags) >> 814 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) & 815 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK) 816 tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag); 817 else 818 tpa_info->vlan_tag = 0; 819 820 qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash); 821 822 /* This is needed in order to enable forwarding support */ 823 qede_set_gro_params(edev, tpa_info->skb, cqe); 824 825 cons_buf: /* We still need to handle bd_len_list to consume buffers */ 826 if (likely(cqe->bw_ext_bd_len_list[0])) 827 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 828 le16_to_cpu(cqe->bw_ext_bd_len_list[0])); 829 830 if (unlikely(cqe->bw_ext_bd_len_list[1])) { 831 DP_ERR(edev, 832 "Unlikely - got a TPA aggregation with more than one bw_ext_bd_len_list entry in the TPA start\n"); 833 tpa_info->state = QEDE_AGG_STATE_ERROR; 834 } 835 } 836 837 #ifdef CONFIG_INET 838 static void qede_gro_ip_csum(struct sk_buff *skb) 839 { 840 const struct iphdr *iph = ip_hdr(skb); 841 struct tcphdr *th; 842 843 skb_set_transport_header(skb, sizeof(struct iphdr)); 844 th = tcp_hdr(skb); 845 846 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb), 847 iph->saddr, iph->daddr, 0); 848 849 tcp_gro_complete(skb); 850 } 851 852 static void qede_gro_ipv6_csum(struct sk_buff *skb) 853 { 854 struct ipv6hdr *iph = ipv6_hdr(skb); 855 struct tcphdr *th; 856 857 skb_set_transport_header(skb, sizeof(struct ipv6hdr)); 858 th = tcp_hdr(skb); 859 860 th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb), 861 &iph->saddr, &iph->daddr, 0); 862 tcp_gro_complete(skb); 863 } 864 #endif 865 866 static void qede_gro_receive(struct qede_dev *edev, 867 struct qede_fastpath *fp, 868 struct sk_buff *skb, 869 u16 vlan_tag) 870 { 871 /* FW can send a single MTU sized packet from gro flow 872 * due to aggregation timeout/last segment etc. which 873 * is not expected to be a gro packet. If a skb has zero 874 * frags then simply push it in the stack as non gso skb. 875 */ 876 if (unlikely(!skb->data_len)) { 877 skb_shinfo(skb)->gso_type = 0; 878 skb_shinfo(skb)->gso_size = 0; 879 goto send_skb; 880 } 881 882 #ifdef CONFIG_INET 883 if (skb_shinfo(skb)->gso_size) { 884 skb_reset_network_header(skb); 885 886 switch (skb->protocol) { 887 case htons(ETH_P_IP): 888 qede_gro_ip_csum(skb); 889 break; 890 case htons(ETH_P_IPV6): 891 qede_gro_ipv6_csum(skb); 892 break; 893 default: 894 DP_ERR(edev, 895 "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n", 896 ntohs(skb->protocol)); 897 } 898 } 899 #endif 900 901 send_skb: 902 skb_record_rx_queue(skb, fp->rxq->rxq_id); 903 qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag); 904 } 905 906 static inline void qede_tpa_cont(struct qede_dev *edev, 907 struct qede_rx_queue *rxq, 908 struct eth_fast_path_rx_tpa_cont_cqe *cqe) 909 { 910 int i; 911 912 for (i = 0; cqe->len_list[i]; i++) 913 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 914 le16_to_cpu(cqe->len_list[i])); 915 916 if (unlikely(i > 1)) 917 DP_ERR(edev, 918 "Strange - TPA cont with more than a single len_list entry\n"); 919 } 920 921 static int qede_tpa_end(struct qede_dev *edev, 922 struct qede_fastpath *fp, 923 struct eth_fast_path_rx_tpa_end_cqe *cqe) 924 { 925 struct qede_rx_queue *rxq = fp->rxq; 926 struct qede_agg_info *tpa_info; 927 struct sk_buff *skb; 928 int i; 929 930 tpa_info = &rxq->tpa_info[cqe->tpa_agg_index]; 931 skb = tpa_info->skb; 932 933 if (tpa_info->buffer.page_offset == PAGE_SIZE) 934 dma_unmap_page(rxq->dev, tpa_info->buffer.mapping, 935 PAGE_SIZE, rxq->data_direction); 936 937 for (i = 0; cqe->len_list[i]; i++) 938 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 939 le16_to_cpu(cqe->len_list[i])); 940 if (unlikely(i > 1)) 941 DP_ERR(edev, 942 "Strange - TPA emd with more than a single len_list entry\n"); 943 944 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START)) 945 goto err; 946 947 /* Sanity */ 948 if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1)) 949 DP_ERR(edev, 950 "Strange - TPA had %02x BDs, but SKB has only %d frags\n", 951 cqe->num_of_bds, tpa_info->frag_id); 952 if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len))) 953 DP_ERR(edev, 954 "Strange - total packet len [cqe] is %4x but SKB has len %04x\n", 955 le16_to_cpu(cqe->total_packet_len), skb->len); 956 957 /* Finalize the SKB */ 958 skb->protocol = eth_type_trans(skb, edev->ndev); 959 skb->ip_summed = CHECKSUM_UNNECESSARY; 960 961 /* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count 962 * to skb_shinfo(skb)->gso_segs 963 */ 964 NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs); 965 966 qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag); 967 968 tpa_info->state = QEDE_AGG_STATE_NONE; 969 970 return 1; 971 err: 972 tpa_info->state = QEDE_AGG_STATE_NONE; 973 974 if (tpa_info->tpa_start_fail) { 975 qede_reuse_page(rxq, &tpa_info->buffer); 976 tpa_info->tpa_start_fail = false; 977 } 978 979 dev_kfree_skb_any(tpa_info->skb); 980 tpa_info->skb = NULL; 981 return 0; 982 } 983 984 static u8 qede_check_notunn_csum(u16 flag) 985 { 986 u16 csum_flag = 0; 987 u8 csum = 0; 988 989 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK << 990 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) { 991 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK << 992 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT; 993 csum = QEDE_CSUM_UNNECESSARY; 994 } 995 996 csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK << 997 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT; 998 999 if (csum_flag & flag) 1000 return QEDE_CSUM_ERROR; 1001 1002 return csum; 1003 } 1004 1005 static u8 qede_check_csum(u16 flag) 1006 { 1007 if (!qede_tunn_exist(flag)) 1008 return qede_check_notunn_csum(flag); 1009 else 1010 return qede_check_tunn_csum(flag); 1011 } 1012 1013 static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe, 1014 u16 flag) 1015 { 1016 u8 tun_pars_flg = cqe->tunnel_pars_flags.flags; 1017 1018 if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK << 1019 ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) || 1020 (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK << 1021 PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT))) 1022 return true; 1023 1024 return false; 1025 } 1026 1027 /* Return true iff packet is to be passed to stack */ 1028 static bool qede_rx_xdp(struct qede_dev *edev, 1029 struct qede_fastpath *fp, 1030 struct qede_rx_queue *rxq, 1031 struct bpf_prog *prog, 1032 struct sw_rx_data *bd, 1033 struct eth_fast_path_rx_reg_cqe *cqe, 1034 u16 *data_offset, u16 *len) 1035 { 1036 struct xdp_buff xdp; 1037 enum xdp_action act; 1038 1039 xdp.data_hard_start = page_address(bd->data); 1040 xdp.data = xdp.data_hard_start + *data_offset; 1041 xdp_set_data_meta_invalid(&xdp); 1042 xdp.data_end = xdp.data + *len; 1043 xdp.rxq = &rxq->xdp_rxq; 1044 xdp.frame_sz = rxq->rx_buf_seg_size; /* PAGE_SIZE when XDP enabled */ 1045 1046 /* Queues always have a full reset currently, so for the time 1047 * being until there's atomic program replace just mark read 1048 * side for map helpers. 1049 */ 1050 rcu_read_lock(); 1051 act = bpf_prog_run_xdp(prog, &xdp); 1052 rcu_read_unlock(); 1053 1054 /* Recalculate, as XDP might have changed the headers */ 1055 *data_offset = xdp.data - xdp.data_hard_start; 1056 *len = xdp.data_end - xdp.data; 1057 1058 if (act == XDP_PASS) 1059 return true; 1060 1061 /* Count number of packets not to be passed to stack */ 1062 rxq->xdp_no_pass++; 1063 1064 switch (act) { 1065 case XDP_TX: 1066 /* We need the replacement buffer before transmit. */ 1067 if (qede_alloc_rx_buffer(rxq, true)) { 1068 qede_recycle_rx_bd_ring(rxq, 1); 1069 trace_xdp_exception(edev->ndev, prog, act); 1070 return false; 1071 } 1072 1073 /* Now if there's a transmission problem, we'd still have to 1074 * throw current buffer, as replacement was already allocated. 1075 */ 1076 if (qede_xdp_xmit(edev, fp, bd, *data_offset, *len)) { 1077 dma_unmap_page(rxq->dev, bd->mapping, 1078 PAGE_SIZE, DMA_BIDIRECTIONAL); 1079 __free_page(bd->data); 1080 trace_xdp_exception(edev->ndev, prog, act); 1081 } 1082 1083 /* Regardless, we've consumed an Rx BD */ 1084 qede_rx_bd_ring_consume(rxq); 1085 return false; 1086 1087 default: 1088 bpf_warn_invalid_xdp_action(act); 1089 /* Fall through */ 1090 case XDP_ABORTED: 1091 trace_xdp_exception(edev->ndev, prog, act); 1092 /* Fall through */ 1093 case XDP_DROP: 1094 qede_recycle_rx_bd_ring(rxq, cqe->bd_num); 1095 } 1096 1097 return false; 1098 } 1099 1100 static int qede_rx_build_jumbo(struct qede_dev *edev, 1101 struct qede_rx_queue *rxq, 1102 struct sk_buff *skb, 1103 struct eth_fast_path_rx_reg_cqe *cqe, 1104 u16 first_bd_len) 1105 { 1106 u16 pkt_len = le16_to_cpu(cqe->pkt_len); 1107 struct sw_rx_data *bd; 1108 u16 bd_cons_idx; 1109 u8 num_frags; 1110 1111 pkt_len -= first_bd_len; 1112 1113 /* We've already used one BD for the SKB. Now take care of the rest */ 1114 for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) { 1115 u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size : 1116 pkt_len; 1117 1118 if (unlikely(!cur_size)) { 1119 DP_ERR(edev, 1120 "Still got %d BDs for mapping jumbo, but length became 0\n", 1121 num_frags); 1122 goto out; 1123 } 1124 1125 /* We need a replacement buffer for each BD */ 1126 if (unlikely(qede_alloc_rx_buffer(rxq, true))) 1127 goto out; 1128 1129 /* Now that we've allocated the replacement buffer, 1130 * we can safely consume the next BD and map it to the SKB. 1131 */ 1132 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX; 1133 bd = &rxq->sw_rx_ring[bd_cons_idx]; 1134 qede_rx_bd_ring_consume(rxq); 1135 1136 dma_unmap_page(rxq->dev, bd->mapping, 1137 PAGE_SIZE, DMA_FROM_DEVICE); 1138 1139 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags++, 1140 bd->data, rxq->rx_headroom, cur_size); 1141 1142 skb->truesize += PAGE_SIZE; 1143 skb->data_len += cur_size; 1144 skb->len += cur_size; 1145 pkt_len -= cur_size; 1146 } 1147 1148 if (unlikely(pkt_len)) 1149 DP_ERR(edev, 1150 "Mapped all BDs of jumbo, but still have %d bytes\n", 1151 pkt_len); 1152 1153 out: 1154 return num_frags; 1155 } 1156 1157 static int qede_rx_process_tpa_cqe(struct qede_dev *edev, 1158 struct qede_fastpath *fp, 1159 struct qede_rx_queue *rxq, 1160 union eth_rx_cqe *cqe, 1161 enum eth_rx_cqe_type type) 1162 { 1163 switch (type) { 1164 case ETH_RX_CQE_TYPE_TPA_START: 1165 qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start); 1166 return 0; 1167 case ETH_RX_CQE_TYPE_TPA_CONT: 1168 qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont); 1169 return 0; 1170 case ETH_RX_CQE_TYPE_TPA_END: 1171 return qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end); 1172 default: 1173 return 0; 1174 } 1175 } 1176 1177 static int qede_rx_process_cqe(struct qede_dev *edev, 1178 struct qede_fastpath *fp, 1179 struct qede_rx_queue *rxq) 1180 { 1181 struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog); 1182 struct eth_fast_path_rx_reg_cqe *fp_cqe; 1183 u16 len, pad, bd_cons_idx, parse_flag; 1184 enum eth_rx_cqe_type cqe_type; 1185 union eth_rx_cqe *cqe; 1186 struct sw_rx_data *bd; 1187 struct sk_buff *skb; 1188 __le16 flags; 1189 u8 csum_flag; 1190 1191 /* Get the CQE from the completion ring */ 1192 cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring); 1193 cqe_type = cqe->fast_path_regular.type; 1194 1195 /* Process an unlikely slowpath event */ 1196 if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) { 1197 struct eth_slow_path_rx_cqe *sp_cqe; 1198 1199 sp_cqe = (struct eth_slow_path_rx_cqe *)cqe; 1200 edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe); 1201 return 0; 1202 } 1203 1204 /* Handle TPA cqes */ 1205 if (cqe_type != ETH_RX_CQE_TYPE_REGULAR) 1206 return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type); 1207 1208 /* Get the data from the SW ring; Consume it only after it's evident 1209 * we wouldn't recycle it. 1210 */ 1211 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX; 1212 bd = &rxq->sw_rx_ring[bd_cons_idx]; 1213 1214 fp_cqe = &cqe->fast_path_regular; 1215 len = le16_to_cpu(fp_cqe->len_on_first_bd); 1216 pad = fp_cqe->placement_offset + rxq->rx_headroom; 1217 1218 /* Run eBPF program if one is attached */ 1219 if (xdp_prog) 1220 if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe, 1221 &pad, &len)) 1222 return 0; 1223 1224 /* If this is an error packet then drop it */ 1225 flags = cqe->fast_path_regular.pars_flags.flags; 1226 parse_flag = le16_to_cpu(flags); 1227 1228 csum_flag = qede_check_csum(parse_flag); 1229 if (unlikely(csum_flag == QEDE_CSUM_ERROR)) { 1230 if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag)) 1231 rxq->rx_ip_frags++; 1232 else 1233 rxq->rx_hw_errors++; 1234 } 1235 1236 /* Basic validation passed; Need to prepare an SKB. This would also 1237 * guarantee to finally consume the first BD upon success. 1238 */ 1239 skb = qede_rx_build_skb(edev, rxq, bd, len, pad); 1240 if (!skb) { 1241 rxq->rx_alloc_errors++; 1242 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num); 1243 return 0; 1244 } 1245 1246 /* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed 1247 * by a single cqe. 1248 */ 1249 if (fp_cqe->bd_num > 1) { 1250 u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb, 1251 fp_cqe, len); 1252 1253 if (unlikely(unmapped_frags > 0)) { 1254 qede_recycle_rx_bd_ring(rxq, unmapped_frags); 1255 dev_kfree_skb_any(skb); 1256 return 0; 1257 } 1258 } 1259 1260 /* The SKB contains all the data. Now prepare meta-magic */ 1261 skb->protocol = eth_type_trans(skb, edev->ndev); 1262 qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash); 1263 qede_set_skb_csum(skb, csum_flag); 1264 skb_record_rx_queue(skb, rxq->rxq_id); 1265 qede_ptp_record_rx_ts(edev, cqe, skb); 1266 1267 /* SKB is prepared - pass it to stack */ 1268 qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag)); 1269 1270 return 1; 1271 } 1272 1273 static int qede_rx_int(struct qede_fastpath *fp, int budget) 1274 { 1275 struct qede_rx_queue *rxq = fp->rxq; 1276 struct qede_dev *edev = fp->edev; 1277 int work_done = 0, rcv_pkts = 0; 1278 u16 hw_comp_cons, sw_comp_cons; 1279 1280 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr); 1281 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 1282 1283 /* Memory barrier to prevent the CPU from doing speculative reads of CQE 1284 * / BD in the while-loop before reading hw_comp_cons. If the CQE is 1285 * read before it is written by FW, then FW writes CQE and SB, and then 1286 * the CPU reads the hw_comp_cons, it will use an old CQE. 1287 */ 1288 rmb(); 1289 1290 /* Loop to complete all indicated BDs */ 1291 while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) { 1292 rcv_pkts += qede_rx_process_cqe(edev, fp, rxq); 1293 qed_chain_recycle_consumed(&rxq->rx_comp_ring); 1294 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 1295 work_done++; 1296 } 1297 1298 rxq->rcv_pkts += rcv_pkts; 1299 1300 /* Allocate replacement buffers */ 1301 while (rxq->num_rx_buffers - rxq->filled_buffers) 1302 if (qede_alloc_rx_buffer(rxq, false)) 1303 break; 1304 1305 /* Update producers */ 1306 qede_update_rx_prod(edev, rxq); 1307 1308 return work_done; 1309 } 1310 1311 static bool qede_poll_is_more_work(struct qede_fastpath *fp) 1312 { 1313 qed_sb_update_sb_idx(fp->sb_info); 1314 1315 /* *_has_*_work() reads the status block, thus we need to ensure that 1316 * status block indices have been actually read (qed_sb_update_sb_idx) 1317 * prior to this check (*_has_*_work) so that we won't write the 1318 * "newer" value of the status block to HW (if there was a DMA right 1319 * after qede_has_rx_work and if there is no rmb, the memory reading 1320 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb). 1321 * In this case there will never be another interrupt until there is 1322 * another update of the status block, while there is still unhandled 1323 * work. 1324 */ 1325 rmb(); 1326 1327 if (likely(fp->type & QEDE_FASTPATH_RX)) 1328 if (qede_has_rx_work(fp->rxq)) 1329 return true; 1330 1331 if (fp->type & QEDE_FASTPATH_XDP) 1332 if (qede_txq_has_work(fp->xdp_tx)) 1333 return true; 1334 1335 if (likely(fp->type & QEDE_FASTPATH_TX)) { 1336 int cos; 1337 1338 for_each_cos_in_txq(fp->edev, cos) { 1339 if (qede_txq_has_work(&fp->txq[cos])) 1340 return true; 1341 } 1342 } 1343 1344 return false; 1345 } 1346 1347 /********************* 1348 * NDO & API related * 1349 *********************/ 1350 int qede_poll(struct napi_struct *napi, int budget) 1351 { 1352 struct qede_fastpath *fp = container_of(napi, struct qede_fastpath, 1353 napi); 1354 struct qede_dev *edev = fp->edev; 1355 int rx_work_done = 0; 1356 1357 if (likely(fp->type & QEDE_FASTPATH_TX)) { 1358 int cos; 1359 1360 for_each_cos_in_txq(fp->edev, cos) { 1361 if (qede_txq_has_work(&fp->txq[cos])) 1362 qede_tx_int(edev, &fp->txq[cos]); 1363 } 1364 } 1365 1366 if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx)) 1367 qede_xdp_tx_int(edev, fp->xdp_tx); 1368 1369 rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) && 1370 qede_has_rx_work(fp->rxq)) ? 1371 qede_rx_int(fp, budget) : 0; 1372 if (rx_work_done < budget) { 1373 if (!qede_poll_is_more_work(fp)) { 1374 napi_complete_done(napi, rx_work_done); 1375 1376 /* Update and reenable interrupts */ 1377 qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1); 1378 } else { 1379 rx_work_done = budget; 1380 } 1381 } 1382 1383 if (fp->xdp_xmit) { 1384 u16 xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl); 1385 1386 fp->xdp_xmit = 0; 1387 fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod); 1388 qede_update_tx_producer(fp->xdp_tx); 1389 } 1390 1391 return rx_work_done; 1392 } 1393 1394 irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie) 1395 { 1396 struct qede_fastpath *fp = fp_cookie; 1397 1398 qed_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0 /*do not update*/); 1399 1400 napi_schedule_irqoff(&fp->napi); 1401 return IRQ_HANDLED; 1402 } 1403 1404 /* Main transmit function */ 1405 netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev) 1406 { 1407 struct qede_dev *edev = netdev_priv(ndev); 1408 struct netdev_queue *netdev_txq; 1409 struct qede_tx_queue *txq; 1410 struct eth_tx_1st_bd *first_bd; 1411 struct eth_tx_2nd_bd *second_bd = NULL; 1412 struct eth_tx_3rd_bd *third_bd = NULL; 1413 struct eth_tx_bd *tx_data_bd = NULL; 1414 u16 txq_index, val = 0; 1415 u8 nbd = 0; 1416 dma_addr_t mapping; 1417 int rc, frag_idx = 0, ipv6_ext = 0; 1418 u8 xmit_type; 1419 u16 idx; 1420 u16 hlen; 1421 bool data_split = false; 1422 1423 /* Get tx-queue context and netdev index */ 1424 txq_index = skb_get_queue_mapping(skb); 1425 WARN_ON(txq_index >= QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc); 1426 txq = QEDE_NDEV_TXQ_ID_TO_TXQ(edev, txq_index); 1427 netdev_txq = netdev_get_tx_queue(ndev, txq_index); 1428 1429 WARN_ON(qed_chain_get_elem_left(&txq->tx_pbl) < (MAX_SKB_FRAGS + 1)); 1430 1431 xmit_type = qede_xmit_type(skb, &ipv6_ext); 1432 1433 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET) 1434 if (qede_pkt_req_lin(skb, xmit_type)) { 1435 if (skb_linearize(skb)) { 1436 txq->tx_mem_alloc_err++; 1437 1438 dev_kfree_skb_any(skb); 1439 return NETDEV_TX_OK; 1440 } 1441 } 1442 #endif 1443 1444 /* Fill the entry in the SW ring and the BDs in the FW ring */ 1445 idx = txq->sw_tx_prod; 1446 txq->sw_tx_ring.skbs[idx].skb = skb; 1447 first_bd = (struct eth_tx_1st_bd *) 1448 qed_chain_produce(&txq->tx_pbl); 1449 memset(first_bd, 0, sizeof(*first_bd)); 1450 first_bd->data.bd_flags.bitfields = 1451 1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT; 1452 1453 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)) 1454 qede_ptp_tx_ts(edev, skb); 1455 1456 /* Map skb linear data for DMA and set in the first BD */ 1457 mapping = dma_map_single(txq->dev, skb->data, 1458 skb_headlen(skb), DMA_TO_DEVICE); 1459 if (unlikely(dma_mapping_error(txq->dev, mapping))) { 1460 DP_NOTICE(edev, "SKB mapping failed\n"); 1461 qede_free_failed_tx_pkt(txq, first_bd, 0, false); 1462 qede_update_tx_producer(txq); 1463 return NETDEV_TX_OK; 1464 } 1465 nbd++; 1466 BD_SET_UNMAP_ADDR_LEN(first_bd, mapping, skb_headlen(skb)); 1467 1468 /* In case there is IPv6 with extension headers or LSO we need 2nd and 1469 * 3rd BDs. 1470 */ 1471 if (unlikely((xmit_type & XMIT_LSO) | ipv6_ext)) { 1472 second_bd = (struct eth_tx_2nd_bd *) 1473 qed_chain_produce(&txq->tx_pbl); 1474 memset(second_bd, 0, sizeof(*second_bd)); 1475 1476 nbd++; 1477 third_bd = (struct eth_tx_3rd_bd *) 1478 qed_chain_produce(&txq->tx_pbl); 1479 memset(third_bd, 0, sizeof(*third_bd)); 1480 1481 nbd++; 1482 /* We need to fill in additional data in second_bd... */ 1483 tx_data_bd = (struct eth_tx_bd *)second_bd; 1484 } 1485 1486 if (skb_vlan_tag_present(skb)) { 1487 first_bd->data.vlan = cpu_to_le16(skb_vlan_tag_get(skb)); 1488 first_bd->data.bd_flags.bitfields |= 1489 1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT; 1490 } 1491 1492 /* Fill the parsing flags & params according to the requested offload */ 1493 if (xmit_type & XMIT_L4_CSUM) { 1494 /* We don't re-calculate IP checksum as it is already done by 1495 * the upper stack 1496 */ 1497 first_bd->data.bd_flags.bitfields |= 1498 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT; 1499 1500 if (xmit_type & XMIT_ENC) { 1501 first_bd->data.bd_flags.bitfields |= 1502 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT; 1503 1504 val |= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT); 1505 } 1506 1507 /* Legacy FW had flipped behavior in regard to this bit - 1508 * I.e., needed to set to prevent FW from touching encapsulated 1509 * packets when it didn't need to. 1510 */ 1511 if (unlikely(txq->is_legacy)) 1512 val ^= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT); 1513 1514 /* If the packet is IPv6 with extension header, indicate that 1515 * to FW and pass few params, since the device cracker doesn't 1516 * support parsing IPv6 with extension header/s. 1517 */ 1518 if (unlikely(ipv6_ext)) 1519 qede_set_params_for_ipv6_ext(skb, second_bd, third_bd); 1520 } 1521 1522 if (xmit_type & XMIT_LSO) { 1523 first_bd->data.bd_flags.bitfields |= 1524 (1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT); 1525 third_bd->data.lso_mss = 1526 cpu_to_le16(skb_shinfo(skb)->gso_size); 1527 1528 if (unlikely(xmit_type & XMIT_ENC)) { 1529 first_bd->data.bd_flags.bitfields |= 1530 1 << ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT; 1531 1532 if (xmit_type & XMIT_ENC_GSO_L4_CSUM) { 1533 u8 tmp = ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT; 1534 1535 first_bd->data.bd_flags.bitfields |= 1 << tmp; 1536 } 1537 hlen = qede_get_skb_hlen(skb, true); 1538 } else { 1539 first_bd->data.bd_flags.bitfields |= 1540 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT; 1541 hlen = qede_get_skb_hlen(skb, false); 1542 } 1543 1544 /* @@@TBD - if will not be removed need to check */ 1545 third_bd->data.bitfields |= 1546 cpu_to_le16(1 << ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT); 1547 1548 /* Make life easier for FW guys who can't deal with header and 1549 * data on same BD. If we need to split, use the second bd... 1550 */ 1551 if (unlikely(skb_headlen(skb) > hlen)) { 1552 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1553 "TSO split header size is %d (%x:%x)\n", 1554 first_bd->nbytes, first_bd->addr.hi, 1555 first_bd->addr.lo); 1556 1557 mapping = HILO_U64(le32_to_cpu(first_bd->addr.hi), 1558 le32_to_cpu(first_bd->addr.lo)) + 1559 hlen; 1560 1561 BD_SET_UNMAP_ADDR_LEN(tx_data_bd, mapping, 1562 le16_to_cpu(first_bd->nbytes) - 1563 hlen); 1564 1565 /* this marks the BD as one that has no 1566 * individual mapping 1567 */ 1568 txq->sw_tx_ring.skbs[idx].flags |= QEDE_TSO_SPLIT_BD; 1569 1570 first_bd->nbytes = cpu_to_le16(hlen); 1571 1572 tx_data_bd = (struct eth_tx_bd *)third_bd; 1573 data_split = true; 1574 } 1575 } else { 1576 val |= ((skb->len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) << 1577 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT); 1578 } 1579 1580 first_bd->data.bitfields = cpu_to_le16(val); 1581 1582 /* Handle fragmented skb */ 1583 /* special handle for frags inside 2nd and 3rd bds.. */ 1584 while (tx_data_bd && frag_idx < skb_shinfo(skb)->nr_frags) { 1585 rc = map_frag_to_bd(txq, 1586 &skb_shinfo(skb)->frags[frag_idx], 1587 tx_data_bd); 1588 if (rc) { 1589 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split); 1590 qede_update_tx_producer(txq); 1591 return NETDEV_TX_OK; 1592 } 1593 1594 if (tx_data_bd == (struct eth_tx_bd *)second_bd) 1595 tx_data_bd = (struct eth_tx_bd *)third_bd; 1596 else 1597 tx_data_bd = NULL; 1598 1599 frag_idx++; 1600 } 1601 1602 /* map last frags into 4th, 5th .... */ 1603 for (; frag_idx < skb_shinfo(skb)->nr_frags; frag_idx++, nbd++) { 1604 tx_data_bd = (struct eth_tx_bd *) 1605 qed_chain_produce(&txq->tx_pbl); 1606 1607 memset(tx_data_bd, 0, sizeof(*tx_data_bd)); 1608 1609 rc = map_frag_to_bd(txq, 1610 &skb_shinfo(skb)->frags[frag_idx], 1611 tx_data_bd); 1612 if (rc) { 1613 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split); 1614 qede_update_tx_producer(txq); 1615 return NETDEV_TX_OK; 1616 } 1617 } 1618 1619 /* update the first BD with the actual num BDs */ 1620 first_bd->data.nbds = nbd; 1621 1622 netdev_tx_sent_queue(netdev_txq, skb->len); 1623 1624 skb_tx_timestamp(skb); 1625 1626 /* Advance packet producer only before sending the packet since mapping 1627 * of pages may fail. 1628 */ 1629 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers; 1630 1631 /* 'next page' entries are counted in the producer value */ 1632 txq->tx_db.data.bd_prod = 1633 cpu_to_le16(qed_chain_get_prod_idx(&txq->tx_pbl)); 1634 1635 if (!netdev_xmit_more() || netif_xmit_stopped(netdev_txq)) 1636 qede_update_tx_producer(txq); 1637 1638 if (unlikely(qed_chain_get_elem_left(&txq->tx_pbl) 1639 < (MAX_SKB_FRAGS + 1))) { 1640 if (netdev_xmit_more()) 1641 qede_update_tx_producer(txq); 1642 1643 netif_tx_stop_queue(netdev_txq); 1644 txq->stopped_cnt++; 1645 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1646 "Stop queue was called\n"); 1647 /* paired memory barrier is in qede_tx_int(), we have to keep 1648 * ordering of set_bit() in netif_tx_stop_queue() and read of 1649 * fp->bd_tx_cons 1650 */ 1651 smp_mb(); 1652 1653 if ((qed_chain_get_elem_left(&txq->tx_pbl) >= 1654 (MAX_SKB_FRAGS + 1)) && 1655 (edev->state == QEDE_STATE_OPEN)) { 1656 netif_tx_wake_queue(netdev_txq); 1657 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1658 "Wake queue was called\n"); 1659 } 1660 } 1661 1662 return NETDEV_TX_OK; 1663 } 1664 1665 u16 qede_select_queue(struct net_device *dev, struct sk_buff *skb, 1666 struct net_device *sb_dev) 1667 { 1668 struct qede_dev *edev = netdev_priv(dev); 1669 int total_txq; 1670 1671 total_txq = QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc; 1672 1673 return QEDE_TSS_COUNT(edev) ? 1674 netdev_pick_tx(dev, skb, NULL) % total_txq : 0; 1675 } 1676 1677 /* 8B udp header + 8B base tunnel header + 32B option length */ 1678 #define QEDE_MAX_TUN_HDR_LEN 48 1679 1680 netdev_features_t qede_features_check(struct sk_buff *skb, 1681 struct net_device *dev, 1682 netdev_features_t features) 1683 { 1684 if (skb->encapsulation) { 1685 u8 l4_proto = 0; 1686 1687 switch (vlan_get_protocol(skb)) { 1688 case htons(ETH_P_IP): 1689 l4_proto = ip_hdr(skb)->protocol; 1690 break; 1691 case htons(ETH_P_IPV6): 1692 l4_proto = ipv6_hdr(skb)->nexthdr; 1693 break; 1694 default: 1695 return features; 1696 } 1697 1698 /* Disable offloads for geneve tunnels, as HW can't parse 1699 * the geneve header which has option length greater than 32b 1700 * and disable offloads for the ports which are not offloaded. 1701 */ 1702 if (l4_proto == IPPROTO_UDP) { 1703 struct qede_dev *edev = netdev_priv(dev); 1704 u16 hdrlen, vxln_port, gnv_port; 1705 1706 hdrlen = QEDE_MAX_TUN_HDR_LEN; 1707 vxln_port = edev->vxlan_dst_port; 1708 gnv_port = edev->geneve_dst_port; 1709 1710 if ((skb_inner_mac_header(skb) - 1711 skb_transport_header(skb)) > hdrlen || 1712 (ntohs(udp_hdr(skb)->dest) != vxln_port && 1713 ntohs(udp_hdr(skb)->dest) != gnv_port)) 1714 return features & ~(NETIF_F_CSUM_MASK | 1715 NETIF_F_GSO_MASK); 1716 } 1717 } 1718 1719 return features; 1720 } 1721