1 /* QLogic qede NIC Driver 2 * Copyright (c) 2015-2017 QLogic Corporation 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and /or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 */ 32 #include <linux/netdevice.h> 33 #include <linux/etherdevice.h> 34 #include <linux/skbuff.h> 35 #include <linux/bpf_trace.h> 36 #include <net/udp_tunnel.h> 37 #include <linux/ip.h> 38 #include <net/ipv6.h> 39 #include <net/tcp.h> 40 #include <linux/if_ether.h> 41 #include <linux/if_vlan.h> 42 #include <net/ip6_checksum.h> 43 #include "qede_ptp.h" 44 45 #include <linux/qed/qed_if.h> 46 #include "qede.h" 47 /********************************* 48 * Content also used by slowpath * 49 *********************************/ 50 51 int qede_alloc_rx_buffer(struct qede_rx_queue *rxq, bool allow_lazy) 52 { 53 struct sw_rx_data *sw_rx_data; 54 struct eth_rx_bd *rx_bd; 55 dma_addr_t mapping; 56 struct page *data; 57 58 /* In case lazy-allocation is allowed, postpone allocation until the 59 * end of the NAPI run. We'd still need to make sure the Rx ring has 60 * sufficient buffers to guarantee an additional Rx interrupt. 61 */ 62 if (allow_lazy && likely(rxq->filled_buffers > 12)) { 63 rxq->filled_buffers--; 64 return 0; 65 } 66 67 data = alloc_pages(GFP_ATOMIC, 0); 68 if (unlikely(!data)) 69 return -ENOMEM; 70 71 /* Map the entire page as it would be used 72 * for multiple RX buffer segment size mapping. 73 */ 74 mapping = dma_map_page(rxq->dev, data, 0, 75 PAGE_SIZE, rxq->data_direction); 76 if (unlikely(dma_mapping_error(rxq->dev, mapping))) { 77 __free_page(data); 78 return -ENOMEM; 79 } 80 81 sw_rx_data = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX]; 82 sw_rx_data->page_offset = 0; 83 sw_rx_data->data = data; 84 sw_rx_data->mapping = mapping; 85 86 /* Advance PROD and get BD pointer */ 87 rx_bd = (struct eth_rx_bd *)qed_chain_produce(&rxq->rx_bd_ring); 88 WARN_ON(!rx_bd); 89 rx_bd->addr.hi = cpu_to_le32(upper_32_bits(mapping)); 90 rx_bd->addr.lo = cpu_to_le32(lower_32_bits(mapping) + 91 rxq->rx_headroom); 92 93 rxq->sw_rx_prod++; 94 rxq->filled_buffers++; 95 96 return 0; 97 } 98 99 /* Unmap the data and free skb */ 100 int qede_free_tx_pkt(struct qede_dev *edev, struct qede_tx_queue *txq, int *len) 101 { 102 u16 idx = txq->sw_tx_cons; 103 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb; 104 struct eth_tx_1st_bd *first_bd; 105 struct eth_tx_bd *tx_data_bd; 106 int bds_consumed = 0; 107 int nbds; 108 bool data_split = txq->sw_tx_ring.skbs[idx].flags & QEDE_TSO_SPLIT_BD; 109 int i, split_bd_len = 0; 110 111 if (unlikely(!skb)) { 112 DP_ERR(edev, 113 "skb is null for txq idx=%d txq->sw_tx_cons=%d txq->sw_tx_prod=%d\n", 114 idx, txq->sw_tx_cons, txq->sw_tx_prod); 115 return -1; 116 } 117 118 *len = skb->len; 119 120 first_bd = (struct eth_tx_1st_bd *)qed_chain_consume(&txq->tx_pbl); 121 122 bds_consumed++; 123 124 nbds = first_bd->data.nbds; 125 126 if (data_split) { 127 struct eth_tx_bd *split = (struct eth_tx_bd *) 128 qed_chain_consume(&txq->tx_pbl); 129 split_bd_len = BD_UNMAP_LEN(split); 130 bds_consumed++; 131 } 132 dma_unmap_single(&edev->pdev->dev, BD_UNMAP_ADDR(first_bd), 133 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE); 134 135 /* Unmap the data of the skb frags */ 136 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, bds_consumed++) { 137 tx_data_bd = (struct eth_tx_bd *) 138 qed_chain_consume(&txq->tx_pbl); 139 dma_unmap_page(&edev->pdev->dev, BD_UNMAP_ADDR(tx_data_bd), 140 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE); 141 } 142 143 while (bds_consumed++ < nbds) 144 qed_chain_consume(&txq->tx_pbl); 145 146 /* Free skb */ 147 dev_kfree_skb_any(skb); 148 txq->sw_tx_ring.skbs[idx].skb = NULL; 149 txq->sw_tx_ring.skbs[idx].flags = 0; 150 151 return 0; 152 } 153 154 /* Unmap the data and free skb when mapping failed during start_xmit */ 155 static void qede_free_failed_tx_pkt(struct qede_tx_queue *txq, 156 struct eth_tx_1st_bd *first_bd, 157 int nbd, bool data_split) 158 { 159 u16 idx = txq->sw_tx_prod; 160 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb; 161 struct eth_tx_bd *tx_data_bd; 162 int i, split_bd_len = 0; 163 164 /* Return prod to its position before this skb was handled */ 165 qed_chain_set_prod(&txq->tx_pbl, 166 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd); 167 168 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl); 169 170 if (data_split) { 171 struct eth_tx_bd *split = (struct eth_tx_bd *) 172 qed_chain_produce(&txq->tx_pbl); 173 split_bd_len = BD_UNMAP_LEN(split); 174 nbd--; 175 } 176 177 dma_unmap_single(txq->dev, BD_UNMAP_ADDR(first_bd), 178 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE); 179 180 /* Unmap the data of the skb frags */ 181 for (i = 0; i < nbd; i++) { 182 tx_data_bd = (struct eth_tx_bd *) 183 qed_chain_produce(&txq->tx_pbl); 184 if (tx_data_bd->nbytes) 185 dma_unmap_page(txq->dev, 186 BD_UNMAP_ADDR(tx_data_bd), 187 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE); 188 } 189 190 /* Return again prod to its position before this skb was handled */ 191 qed_chain_set_prod(&txq->tx_pbl, 192 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd); 193 194 /* Free skb */ 195 dev_kfree_skb_any(skb); 196 txq->sw_tx_ring.skbs[idx].skb = NULL; 197 txq->sw_tx_ring.skbs[idx].flags = 0; 198 } 199 200 static u32 qede_xmit_type(struct sk_buff *skb, int *ipv6_ext) 201 { 202 u32 rc = XMIT_L4_CSUM; 203 __be16 l3_proto; 204 205 if (skb->ip_summed != CHECKSUM_PARTIAL) 206 return XMIT_PLAIN; 207 208 l3_proto = vlan_get_protocol(skb); 209 if (l3_proto == htons(ETH_P_IPV6) && 210 (ipv6_hdr(skb)->nexthdr == NEXTHDR_IPV6)) 211 *ipv6_ext = 1; 212 213 if (skb->encapsulation) { 214 rc |= XMIT_ENC; 215 if (skb_is_gso(skb)) { 216 unsigned short gso_type = skb_shinfo(skb)->gso_type; 217 218 if ((gso_type & SKB_GSO_UDP_TUNNEL_CSUM) || 219 (gso_type & SKB_GSO_GRE_CSUM)) 220 rc |= XMIT_ENC_GSO_L4_CSUM; 221 222 rc |= XMIT_LSO; 223 return rc; 224 } 225 } 226 227 if (skb_is_gso(skb)) 228 rc |= XMIT_LSO; 229 230 return rc; 231 } 232 233 static void qede_set_params_for_ipv6_ext(struct sk_buff *skb, 234 struct eth_tx_2nd_bd *second_bd, 235 struct eth_tx_3rd_bd *third_bd) 236 { 237 u8 l4_proto; 238 u16 bd2_bits1 = 0, bd2_bits2 = 0; 239 240 bd2_bits1 |= (1 << ETH_TX_DATA_2ND_BD_IPV6_EXT_SHIFT); 241 242 bd2_bits2 |= ((((u8 *)skb_transport_header(skb) - skb->data) >> 1) & 243 ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK) 244 << ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_SHIFT; 245 246 bd2_bits1 |= (ETH_L4_PSEUDO_CSUM_CORRECT_LENGTH << 247 ETH_TX_DATA_2ND_BD_L4_PSEUDO_CSUM_MODE_SHIFT); 248 249 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6)) 250 l4_proto = ipv6_hdr(skb)->nexthdr; 251 else 252 l4_proto = ip_hdr(skb)->protocol; 253 254 if (l4_proto == IPPROTO_UDP) 255 bd2_bits1 |= 1 << ETH_TX_DATA_2ND_BD_L4_UDP_SHIFT; 256 257 if (third_bd) 258 third_bd->data.bitfields |= 259 cpu_to_le16(((tcp_hdrlen(skb) / 4) & 260 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_MASK) << 261 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_SHIFT); 262 263 second_bd->data.bitfields1 = cpu_to_le16(bd2_bits1); 264 second_bd->data.bitfields2 = cpu_to_le16(bd2_bits2); 265 } 266 267 static int map_frag_to_bd(struct qede_tx_queue *txq, 268 skb_frag_t *frag, struct eth_tx_bd *bd) 269 { 270 dma_addr_t mapping; 271 272 /* Map skb non-linear frag data for DMA */ 273 mapping = skb_frag_dma_map(txq->dev, frag, 0, 274 skb_frag_size(frag), DMA_TO_DEVICE); 275 if (unlikely(dma_mapping_error(txq->dev, mapping))) 276 return -ENOMEM; 277 278 /* Setup the data pointer of the frag data */ 279 BD_SET_UNMAP_ADDR_LEN(bd, mapping, skb_frag_size(frag)); 280 281 return 0; 282 } 283 284 static u16 qede_get_skb_hlen(struct sk_buff *skb, bool is_encap_pkt) 285 { 286 if (is_encap_pkt) 287 return (skb_inner_transport_header(skb) + 288 inner_tcp_hdrlen(skb) - skb->data); 289 else 290 return (skb_transport_header(skb) + 291 tcp_hdrlen(skb) - skb->data); 292 } 293 294 /* +2 for 1st BD for headers and 2nd BD for headlen (if required) */ 295 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET) 296 static bool qede_pkt_req_lin(struct sk_buff *skb, u8 xmit_type) 297 { 298 int allowed_frags = ETH_TX_MAX_BDS_PER_NON_LSO_PACKET - 1; 299 300 if (xmit_type & XMIT_LSO) { 301 int hlen; 302 303 hlen = qede_get_skb_hlen(skb, xmit_type & XMIT_ENC); 304 305 /* linear payload would require its own BD */ 306 if (skb_headlen(skb) > hlen) 307 allowed_frags--; 308 } 309 310 return (skb_shinfo(skb)->nr_frags > allowed_frags); 311 } 312 #endif 313 314 static inline void qede_update_tx_producer(struct qede_tx_queue *txq) 315 { 316 /* wmb makes sure that the BDs data is updated before updating the 317 * producer, otherwise FW may read old data from the BDs. 318 */ 319 wmb(); 320 barrier(); 321 writel(txq->tx_db.raw, txq->doorbell_addr); 322 323 /* Fence required to flush the write combined buffer, since another 324 * CPU may write to the same doorbell address and data may be lost 325 * due to relaxed order nature of write combined bar. 326 */ 327 wmb(); 328 } 329 330 static int qede_xdp_xmit(struct qede_dev *edev, struct qede_fastpath *fp, 331 struct sw_rx_data *metadata, u16 padding, u16 length) 332 { 333 struct qede_tx_queue *txq = fp->xdp_tx; 334 struct eth_tx_1st_bd *first_bd; 335 u16 idx = txq->sw_tx_prod; 336 u16 val; 337 338 if (!qed_chain_get_elem_left(&txq->tx_pbl)) { 339 txq->stopped_cnt++; 340 return -ENOMEM; 341 } 342 343 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl); 344 345 memset(first_bd, 0, sizeof(*first_bd)); 346 first_bd->data.bd_flags.bitfields = 347 BIT(ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT); 348 349 val = (length & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) << 350 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT; 351 352 first_bd->data.bitfields |= cpu_to_le16(val); 353 first_bd->data.nbds = 1; 354 355 /* We can safely ignore the offset, as it's 0 for XDP */ 356 BD_SET_UNMAP_ADDR_LEN(first_bd, metadata->mapping + padding, length); 357 358 /* Synchronize the buffer back to device, as program [probably] 359 * has changed it. 360 */ 361 dma_sync_single_for_device(&edev->pdev->dev, 362 metadata->mapping + padding, 363 length, PCI_DMA_TODEVICE); 364 365 txq->sw_tx_ring.xdp[idx].page = metadata->data; 366 txq->sw_tx_ring.xdp[idx].mapping = metadata->mapping; 367 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers; 368 369 /* Mark the fastpath for future XDP doorbell */ 370 fp->xdp_xmit = 1; 371 372 return 0; 373 } 374 375 int qede_txq_has_work(struct qede_tx_queue *txq) 376 { 377 u16 hw_bd_cons; 378 379 /* Tell compiler that consumer and producer can change */ 380 barrier(); 381 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 382 if (qed_chain_get_cons_idx(&txq->tx_pbl) == hw_bd_cons + 1) 383 return 0; 384 385 return hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl); 386 } 387 388 static void qede_xdp_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq) 389 { 390 u16 hw_bd_cons, idx; 391 392 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 393 barrier(); 394 395 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) { 396 qed_chain_consume(&txq->tx_pbl); 397 idx = txq->sw_tx_cons; 398 399 dma_unmap_page(&edev->pdev->dev, 400 txq->sw_tx_ring.xdp[idx].mapping, 401 PAGE_SIZE, DMA_BIDIRECTIONAL); 402 __free_page(txq->sw_tx_ring.xdp[idx].page); 403 404 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers; 405 txq->xmit_pkts++; 406 } 407 } 408 409 static int qede_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq) 410 { 411 unsigned int pkts_compl = 0, bytes_compl = 0; 412 struct netdev_queue *netdev_txq; 413 u16 hw_bd_cons; 414 int rc; 415 416 netdev_txq = netdev_get_tx_queue(edev->ndev, txq->ndev_txq_id); 417 418 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 419 barrier(); 420 421 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) { 422 int len = 0; 423 424 rc = qede_free_tx_pkt(edev, txq, &len); 425 if (rc) { 426 DP_NOTICE(edev, "hw_bd_cons = %d, chain_cons=%d\n", 427 hw_bd_cons, 428 qed_chain_get_cons_idx(&txq->tx_pbl)); 429 break; 430 } 431 432 bytes_compl += len; 433 pkts_compl++; 434 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers; 435 txq->xmit_pkts++; 436 } 437 438 netdev_tx_completed_queue(netdev_txq, pkts_compl, bytes_compl); 439 440 /* Need to make the tx_bd_cons update visible to start_xmit() 441 * before checking for netif_tx_queue_stopped(). Without the 442 * memory barrier, there is a small possibility that 443 * start_xmit() will miss it and cause the queue to be stopped 444 * forever. 445 * On the other hand we need an rmb() here to ensure the proper 446 * ordering of bit testing in the following 447 * netif_tx_queue_stopped(txq) call. 448 */ 449 smp_mb(); 450 451 if (unlikely(netif_tx_queue_stopped(netdev_txq))) { 452 /* Taking tx_lock is needed to prevent reenabling the queue 453 * while it's empty. This could have happen if rx_action() gets 454 * suspended in qede_tx_int() after the condition before 455 * netif_tx_wake_queue(), while tx_action (qede_start_xmit()): 456 * 457 * stops the queue->sees fresh tx_bd_cons->releases the queue-> 458 * sends some packets consuming the whole queue again-> 459 * stops the queue 460 */ 461 462 __netif_tx_lock(netdev_txq, smp_processor_id()); 463 464 if ((netif_tx_queue_stopped(netdev_txq)) && 465 (edev->state == QEDE_STATE_OPEN) && 466 (qed_chain_get_elem_left(&txq->tx_pbl) 467 >= (MAX_SKB_FRAGS + 1))) { 468 netif_tx_wake_queue(netdev_txq); 469 DP_VERBOSE(edev, NETIF_MSG_TX_DONE, 470 "Wake queue was called\n"); 471 } 472 473 __netif_tx_unlock(netdev_txq); 474 } 475 476 return 0; 477 } 478 479 bool qede_has_rx_work(struct qede_rx_queue *rxq) 480 { 481 u16 hw_comp_cons, sw_comp_cons; 482 483 /* Tell compiler that status block fields can change */ 484 barrier(); 485 486 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr); 487 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 488 489 return hw_comp_cons != sw_comp_cons; 490 } 491 492 static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq) 493 { 494 qed_chain_consume(&rxq->rx_bd_ring); 495 rxq->sw_rx_cons++; 496 } 497 498 /* This function reuses the buffer(from an offset) from 499 * consumer index to producer index in the bd ring 500 */ 501 static inline void qede_reuse_page(struct qede_rx_queue *rxq, 502 struct sw_rx_data *curr_cons) 503 { 504 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring); 505 struct sw_rx_data *curr_prod; 506 dma_addr_t new_mapping; 507 508 curr_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX]; 509 *curr_prod = *curr_cons; 510 511 new_mapping = curr_prod->mapping + curr_prod->page_offset; 512 513 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping)); 514 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping) + 515 rxq->rx_headroom); 516 517 rxq->sw_rx_prod++; 518 curr_cons->data = NULL; 519 } 520 521 /* In case of allocation failures reuse buffers 522 * from consumer index to produce buffers for firmware 523 */ 524 void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count) 525 { 526 struct sw_rx_data *curr_cons; 527 528 for (; count > 0; count--) { 529 curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX]; 530 qede_reuse_page(rxq, curr_cons); 531 qede_rx_bd_ring_consume(rxq); 532 } 533 } 534 535 static inline int qede_realloc_rx_buffer(struct qede_rx_queue *rxq, 536 struct sw_rx_data *curr_cons) 537 { 538 /* Move to the next segment in the page */ 539 curr_cons->page_offset += rxq->rx_buf_seg_size; 540 541 if (curr_cons->page_offset == PAGE_SIZE) { 542 if (unlikely(qede_alloc_rx_buffer(rxq, true))) { 543 /* Since we failed to allocate new buffer 544 * current buffer can be used again. 545 */ 546 curr_cons->page_offset -= rxq->rx_buf_seg_size; 547 548 return -ENOMEM; 549 } 550 551 dma_unmap_page(rxq->dev, curr_cons->mapping, 552 PAGE_SIZE, rxq->data_direction); 553 } else { 554 /* Increment refcount of the page as we don't want 555 * network stack to take the ownership of the page 556 * which can be recycled multiple times by the driver. 557 */ 558 page_ref_inc(curr_cons->data); 559 qede_reuse_page(rxq, curr_cons); 560 } 561 562 return 0; 563 } 564 565 void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq) 566 { 567 u16 bd_prod = qed_chain_get_prod_idx(&rxq->rx_bd_ring); 568 u16 cqe_prod = qed_chain_get_prod_idx(&rxq->rx_comp_ring); 569 struct eth_rx_prod_data rx_prods = {0}; 570 571 /* Update producers */ 572 rx_prods.bd_prod = cpu_to_le16(bd_prod); 573 rx_prods.cqe_prod = cpu_to_le16(cqe_prod); 574 575 /* Make sure that the BD and SGE data is updated before updating the 576 * producers since FW might read the BD/SGE right after the producer 577 * is updated. 578 */ 579 wmb(); 580 581 internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods), 582 (u32 *)&rx_prods); 583 } 584 585 static void qede_get_rxhash(struct sk_buff *skb, u8 bitfields, __le32 rss_hash) 586 { 587 enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE; 588 enum rss_hash_type htype; 589 u32 hash = 0; 590 591 htype = GET_FIELD(bitfields, ETH_FAST_PATH_RX_REG_CQE_RSS_HASH_TYPE); 592 if (htype) { 593 hash_type = ((htype == RSS_HASH_TYPE_IPV4) || 594 (htype == RSS_HASH_TYPE_IPV6)) ? 595 PKT_HASH_TYPE_L3 : PKT_HASH_TYPE_L4; 596 hash = le32_to_cpu(rss_hash); 597 } 598 skb_set_hash(skb, hash, hash_type); 599 } 600 601 static void qede_set_skb_csum(struct sk_buff *skb, u8 csum_flag) 602 { 603 skb_checksum_none_assert(skb); 604 605 if (csum_flag & QEDE_CSUM_UNNECESSARY) 606 skb->ip_summed = CHECKSUM_UNNECESSARY; 607 608 if (csum_flag & QEDE_TUNN_CSUM_UNNECESSARY) { 609 skb->csum_level = 1; 610 skb->encapsulation = 1; 611 } 612 } 613 614 static inline void qede_skb_receive(struct qede_dev *edev, 615 struct qede_fastpath *fp, 616 struct qede_rx_queue *rxq, 617 struct sk_buff *skb, u16 vlan_tag) 618 { 619 if (vlan_tag) 620 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag); 621 622 napi_gro_receive(&fp->napi, skb); 623 } 624 625 static void qede_set_gro_params(struct qede_dev *edev, 626 struct sk_buff *skb, 627 struct eth_fast_path_rx_tpa_start_cqe *cqe) 628 { 629 u16 parsing_flags = le16_to_cpu(cqe->pars_flags.flags); 630 631 if (((parsing_flags >> PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) & 632 PARSING_AND_ERR_FLAGS_L3TYPE_MASK) == 2) 633 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6; 634 else 635 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4; 636 637 skb_shinfo(skb)->gso_size = __le16_to_cpu(cqe->len_on_first_bd) - 638 cqe->header_len; 639 } 640 641 static int qede_fill_frag_skb(struct qede_dev *edev, 642 struct qede_rx_queue *rxq, 643 u8 tpa_agg_index, u16 len_on_bd) 644 { 645 struct sw_rx_data *current_bd = &rxq->sw_rx_ring[rxq->sw_rx_cons & 646 NUM_RX_BDS_MAX]; 647 struct qede_agg_info *tpa_info = &rxq->tpa_info[tpa_agg_index]; 648 struct sk_buff *skb = tpa_info->skb; 649 650 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START)) 651 goto out; 652 653 /* Add one frag and update the appropriate fields in the skb */ 654 skb_fill_page_desc(skb, tpa_info->frag_id++, 655 current_bd->data, 656 current_bd->page_offset + rxq->rx_headroom, 657 len_on_bd); 658 659 if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) { 660 /* Incr page ref count to reuse on allocation failure 661 * so that it doesn't get freed while freeing SKB. 662 */ 663 page_ref_inc(current_bd->data); 664 goto out; 665 } 666 667 qede_rx_bd_ring_consume(rxq); 668 669 skb->data_len += len_on_bd; 670 skb->truesize += rxq->rx_buf_seg_size; 671 skb->len += len_on_bd; 672 673 return 0; 674 675 out: 676 tpa_info->state = QEDE_AGG_STATE_ERROR; 677 qede_recycle_rx_bd_ring(rxq, 1); 678 679 return -ENOMEM; 680 } 681 682 static bool qede_tunn_exist(u16 flag) 683 { 684 return !!(flag & (PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK << 685 PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT)); 686 } 687 688 static u8 qede_check_tunn_csum(u16 flag) 689 { 690 u16 csum_flag = 0; 691 u8 tcsum = 0; 692 693 if (flag & (PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK << 694 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT)) 695 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK << 696 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT; 697 698 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK << 699 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) { 700 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK << 701 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT; 702 tcsum = QEDE_TUNN_CSUM_UNNECESSARY; 703 } 704 705 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK << 706 PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT | 707 PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK << 708 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT; 709 710 if (csum_flag & flag) 711 return QEDE_CSUM_ERROR; 712 713 return QEDE_CSUM_UNNECESSARY | tcsum; 714 } 715 716 static inline struct sk_buff * 717 qede_build_skb(struct qede_rx_queue *rxq, 718 struct sw_rx_data *bd, u16 len, u16 pad) 719 { 720 struct sk_buff *skb; 721 void *buf; 722 723 buf = page_address(bd->data) + bd->page_offset; 724 skb = build_skb(buf, rxq->rx_buf_seg_size); 725 726 skb_reserve(skb, pad); 727 skb_put(skb, len); 728 729 return skb; 730 } 731 732 static struct sk_buff * 733 qede_tpa_rx_build_skb(struct qede_dev *edev, 734 struct qede_rx_queue *rxq, 735 struct sw_rx_data *bd, u16 len, u16 pad, 736 bool alloc_skb) 737 { 738 struct sk_buff *skb; 739 740 skb = qede_build_skb(rxq, bd, len, pad); 741 bd->page_offset += rxq->rx_buf_seg_size; 742 743 if (bd->page_offset == PAGE_SIZE) { 744 if (unlikely(qede_alloc_rx_buffer(rxq, true))) { 745 DP_NOTICE(edev, 746 "Failed to allocate RX buffer for tpa start\n"); 747 bd->page_offset -= rxq->rx_buf_seg_size; 748 page_ref_inc(bd->data); 749 dev_kfree_skb_any(skb); 750 return NULL; 751 } 752 } else { 753 page_ref_inc(bd->data); 754 qede_reuse_page(rxq, bd); 755 } 756 757 /* We've consumed the first BD and prepared an SKB */ 758 qede_rx_bd_ring_consume(rxq); 759 760 return skb; 761 } 762 763 static struct sk_buff * 764 qede_rx_build_skb(struct qede_dev *edev, 765 struct qede_rx_queue *rxq, 766 struct sw_rx_data *bd, u16 len, u16 pad) 767 { 768 struct sk_buff *skb = NULL; 769 770 /* For smaller frames still need to allocate skb, memcpy 771 * data and benefit in reusing the page segment instead of 772 * un-mapping it. 773 */ 774 if ((len + pad <= edev->rx_copybreak)) { 775 unsigned int offset = bd->page_offset + pad; 776 777 skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE); 778 if (unlikely(!skb)) 779 return NULL; 780 781 skb_reserve(skb, pad); 782 skb_put_data(skb, page_address(bd->data) + offset, len); 783 qede_reuse_page(rxq, bd); 784 goto out; 785 } 786 787 skb = qede_build_skb(rxq, bd, len, pad); 788 789 if (unlikely(qede_realloc_rx_buffer(rxq, bd))) { 790 /* Incr page ref count to reuse on allocation failure so 791 * that it doesn't get freed while freeing SKB [as its 792 * already mapped there]. 793 */ 794 page_ref_inc(bd->data); 795 dev_kfree_skb_any(skb); 796 return NULL; 797 } 798 out: 799 /* We've consumed the first BD and prepared an SKB */ 800 qede_rx_bd_ring_consume(rxq); 801 802 return skb; 803 } 804 805 static void qede_tpa_start(struct qede_dev *edev, 806 struct qede_rx_queue *rxq, 807 struct eth_fast_path_rx_tpa_start_cqe *cqe) 808 { 809 struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index]; 810 struct sw_rx_data *sw_rx_data_cons; 811 u16 pad; 812 813 sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX]; 814 pad = cqe->placement_offset + rxq->rx_headroom; 815 816 tpa_info->skb = qede_tpa_rx_build_skb(edev, rxq, sw_rx_data_cons, 817 le16_to_cpu(cqe->len_on_first_bd), 818 pad, false); 819 tpa_info->buffer.page_offset = sw_rx_data_cons->page_offset; 820 tpa_info->buffer.mapping = sw_rx_data_cons->mapping; 821 822 if (unlikely(!tpa_info->skb)) { 823 DP_NOTICE(edev, "Failed to allocate SKB for gro\n"); 824 825 /* Consume from ring but do not produce since 826 * this might be used by FW still, it will be re-used 827 * at TPA end. 828 */ 829 tpa_info->tpa_start_fail = true; 830 qede_rx_bd_ring_consume(rxq); 831 tpa_info->state = QEDE_AGG_STATE_ERROR; 832 goto cons_buf; 833 } 834 835 tpa_info->frag_id = 0; 836 tpa_info->state = QEDE_AGG_STATE_START; 837 838 if ((le16_to_cpu(cqe->pars_flags.flags) >> 839 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) & 840 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK) 841 tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag); 842 else 843 tpa_info->vlan_tag = 0; 844 845 qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash); 846 847 /* This is needed in order to enable forwarding support */ 848 qede_set_gro_params(edev, tpa_info->skb, cqe); 849 850 cons_buf: /* We still need to handle bd_len_list to consume buffers */ 851 if (likely(cqe->bw_ext_bd_len_list[0])) 852 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 853 le16_to_cpu(cqe->bw_ext_bd_len_list[0])); 854 855 if (unlikely(cqe->bw_ext_bd_len_list[1])) { 856 DP_ERR(edev, 857 "Unlikely - got a TPA aggregation with more than one bw_ext_bd_len_list entry in the TPA start\n"); 858 tpa_info->state = QEDE_AGG_STATE_ERROR; 859 } 860 } 861 862 #ifdef CONFIG_INET 863 static void qede_gro_ip_csum(struct sk_buff *skb) 864 { 865 const struct iphdr *iph = ip_hdr(skb); 866 struct tcphdr *th; 867 868 skb_set_transport_header(skb, sizeof(struct iphdr)); 869 th = tcp_hdr(skb); 870 871 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb), 872 iph->saddr, iph->daddr, 0); 873 874 tcp_gro_complete(skb); 875 } 876 877 static void qede_gro_ipv6_csum(struct sk_buff *skb) 878 { 879 struct ipv6hdr *iph = ipv6_hdr(skb); 880 struct tcphdr *th; 881 882 skb_set_transport_header(skb, sizeof(struct ipv6hdr)); 883 th = tcp_hdr(skb); 884 885 th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb), 886 &iph->saddr, &iph->daddr, 0); 887 tcp_gro_complete(skb); 888 } 889 #endif 890 891 static void qede_gro_receive(struct qede_dev *edev, 892 struct qede_fastpath *fp, 893 struct sk_buff *skb, 894 u16 vlan_tag) 895 { 896 /* FW can send a single MTU sized packet from gro flow 897 * due to aggregation timeout/last segment etc. which 898 * is not expected to be a gro packet. If a skb has zero 899 * frags then simply push it in the stack as non gso skb. 900 */ 901 if (unlikely(!skb->data_len)) { 902 skb_shinfo(skb)->gso_type = 0; 903 skb_shinfo(skb)->gso_size = 0; 904 goto send_skb; 905 } 906 907 #ifdef CONFIG_INET 908 if (skb_shinfo(skb)->gso_size) { 909 skb_reset_network_header(skb); 910 911 switch (skb->protocol) { 912 case htons(ETH_P_IP): 913 qede_gro_ip_csum(skb); 914 break; 915 case htons(ETH_P_IPV6): 916 qede_gro_ipv6_csum(skb); 917 break; 918 default: 919 DP_ERR(edev, 920 "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n", 921 ntohs(skb->protocol)); 922 } 923 } 924 #endif 925 926 send_skb: 927 skb_record_rx_queue(skb, fp->rxq->rxq_id); 928 qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag); 929 } 930 931 static inline void qede_tpa_cont(struct qede_dev *edev, 932 struct qede_rx_queue *rxq, 933 struct eth_fast_path_rx_tpa_cont_cqe *cqe) 934 { 935 int i; 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])); 940 941 if (unlikely(i > 1)) 942 DP_ERR(edev, 943 "Strange - TPA cont with more than a single len_list entry\n"); 944 } 945 946 static int qede_tpa_end(struct qede_dev *edev, 947 struct qede_fastpath *fp, 948 struct eth_fast_path_rx_tpa_end_cqe *cqe) 949 { 950 struct qede_rx_queue *rxq = fp->rxq; 951 struct qede_agg_info *tpa_info; 952 struct sk_buff *skb; 953 int i; 954 955 tpa_info = &rxq->tpa_info[cqe->tpa_agg_index]; 956 skb = tpa_info->skb; 957 958 if (tpa_info->buffer.page_offset == PAGE_SIZE) 959 dma_unmap_page(rxq->dev, tpa_info->buffer.mapping, 960 PAGE_SIZE, rxq->data_direction); 961 962 for (i = 0; cqe->len_list[i]; i++) 963 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 964 le16_to_cpu(cqe->len_list[i])); 965 if (unlikely(i > 1)) 966 DP_ERR(edev, 967 "Strange - TPA emd with more than a single len_list entry\n"); 968 969 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START)) 970 goto err; 971 972 /* Sanity */ 973 if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1)) 974 DP_ERR(edev, 975 "Strange - TPA had %02x BDs, but SKB has only %d frags\n", 976 cqe->num_of_bds, tpa_info->frag_id); 977 if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len))) 978 DP_ERR(edev, 979 "Strange - total packet len [cqe] is %4x but SKB has len %04x\n", 980 le16_to_cpu(cqe->total_packet_len), skb->len); 981 982 /* Finalize the SKB */ 983 skb->protocol = eth_type_trans(skb, edev->ndev); 984 skb->ip_summed = CHECKSUM_UNNECESSARY; 985 986 /* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count 987 * to skb_shinfo(skb)->gso_segs 988 */ 989 NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs); 990 991 qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag); 992 993 tpa_info->state = QEDE_AGG_STATE_NONE; 994 995 return 1; 996 err: 997 tpa_info->state = QEDE_AGG_STATE_NONE; 998 999 if (tpa_info->tpa_start_fail) { 1000 qede_reuse_page(rxq, &tpa_info->buffer); 1001 tpa_info->tpa_start_fail = false; 1002 } 1003 1004 dev_kfree_skb_any(tpa_info->skb); 1005 tpa_info->skb = NULL; 1006 return 0; 1007 } 1008 1009 static u8 qede_check_notunn_csum(u16 flag) 1010 { 1011 u16 csum_flag = 0; 1012 u8 csum = 0; 1013 1014 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK << 1015 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) { 1016 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK << 1017 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT; 1018 csum = QEDE_CSUM_UNNECESSARY; 1019 } 1020 1021 csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK << 1022 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT; 1023 1024 if (csum_flag & flag) 1025 return QEDE_CSUM_ERROR; 1026 1027 return csum; 1028 } 1029 1030 static u8 qede_check_csum(u16 flag) 1031 { 1032 if (!qede_tunn_exist(flag)) 1033 return qede_check_notunn_csum(flag); 1034 else 1035 return qede_check_tunn_csum(flag); 1036 } 1037 1038 static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe, 1039 u16 flag) 1040 { 1041 u8 tun_pars_flg = cqe->tunnel_pars_flags.flags; 1042 1043 if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK << 1044 ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) || 1045 (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK << 1046 PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT))) 1047 return true; 1048 1049 return false; 1050 } 1051 1052 /* Return true iff packet is to be passed to stack */ 1053 static bool qede_rx_xdp(struct qede_dev *edev, 1054 struct qede_fastpath *fp, 1055 struct qede_rx_queue *rxq, 1056 struct bpf_prog *prog, 1057 struct sw_rx_data *bd, 1058 struct eth_fast_path_rx_reg_cqe *cqe, 1059 u16 *data_offset, u16 *len) 1060 { 1061 struct xdp_buff xdp; 1062 enum xdp_action act; 1063 1064 xdp.data_hard_start = page_address(bd->data); 1065 xdp.data = xdp.data_hard_start + *data_offset; 1066 xdp_set_data_meta_invalid(&xdp); 1067 xdp.data_end = xdp.data + *len; 1068 xdp.rxq = &rxq->xdp_rxq; 1069 xdp.frame_sz = rxq->rx_buf_seg_size; /* PAGE_SIZE when XDP enabled */ 1070 1071 /* Queues always have a full reset currently, so for the time 1072 * being until there's atomic program replace just mark read 1073 * side for map helpers. 1074 */ 1075 rcu_read_lock(); 1076 act = bpf_prog_run_xdp(prog, &xdp); 1077 rcu_read_unlock(); 1078 1079 /* Recalculate, as XDP might have changed the headers */ 1080 *data_offset = xdp.data - xdp.data_hard_start; 1081 *len = xdp.data_end - xdp.data; 1082 1083 if (act == XDP_PASS) 1084 return true; 1085 1086 /* Count number of packets not to be passed to stack */ 1087 rxq->xdp_no_pass++; 1088 1089 switch (act) { 1090 case XDP_TX: 1091 /* We need the replacement buffer before transmit. */ 1092 if (qede_alloc_rx_buffer(rxq, true)) { 1093 qede_recycle_rx_bd_ring(rxq, 1); 1094 trace_xdp_exception(edev->ndev, prog, act); 1095 return false; 1096 } 1097 1098 /* Now if there's a transmission problem, we'd still have to 1099 * throw current buffer, as replacement was already allocated. 1100 */ 1101 if (qede_xdp_xmit(edev, fp, bd, *data_offset, *len)) { 1102 dma_unmap_page(rxq->dev, bd->mapping, 1103 PAGE_SIZE, DMA_BIDIRECTIONAL); 1104 __free_page(bd->data); 1105 trace_xdp_exception(edev->ndev, prog, act); 1106 } 1107 1108 /* Regardless, we've consumed an Rx BD */ 1109 qede_rx_bd_ring_consume(rxq); 1110 return false; 1111 1112 default: 1113 bpf_warn_invalid_xdp_action(act); 1114 /* Fall through */ 1115 case XDP_ABORTED: 1116 trace_xdp_exception(edev->ndev, prog, act); 1117 /* Fall through */ 1118 case XDP_DROP: 1119 qede_recycle_rx_bd_ring(rxq, cqe->bd_num); 1120 } 1121 1122 return false; 1123 } 1124 1125 static int qede_rx_build_jumbo(struct qede_dev *edev, 1126 struct qede_rx_queue *rxq, 1127 struct sk_buff *skb, 1128 struct eth_fast_path_rx_reg_cqe *cqe, 1129 u16 first_bd_len) 1130 { 1131 u16 pkt_len = le16_to_cpu(cqe->pkt_len); 1132 struct sw_rx_data *bd; 1133 u16 bd_cons_idx; 1134 u8 num_frags; 1135 1136 pkt_len -= first_bd_len; 1137 1138 /* We've already used one BD for the SKB. Now take care of the rest */ 1139 for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) { 1140 u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size : 1141 pkt_len; 1142 1143 if (unlikely(!cur_size)) { 1144 DP_ERR(edev, 1145 "Still got %d BDs for mapping jumbo, but length became 0\n", 1146 num_frags); 1147 goto out; 1148 } 1149 1150 /* We need a replacement buffer for each BD */ 1151 if (unlikely(qede_alloc_rx_buffer(rxq, true))) 1152 goto out; 1153 1154 /* Now that we've allocated the replacement buffer, 1155 * we can safely consume the next BD and map it to the SKB. 1156 */ 1157 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX; 1158 bd = &rxq->sw_rx_ring[bd_cons_idx]; 1159 qede_rx_bd_ring_consume(rxq); 1160 1161 dma_unmap_page(rxq->dev, bd->mapping, 1162 PAGE_SIZE, DMA_FROM_DEVICE); 1163 1164 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags++, 1165 bd->data, rxq->rx_headroom, cur_size); 1166 1167 skb->truesize += PAGE_SIZE; 1168 skb->data_len += cur_size; 1169 skb->len += cur_size; 1170 pkt_len -= cur_size; 1171 } 1172 1173 if (unlikely(pkt_len)) 1174 DP_ERR(edev, 1175 "Mapped all BDs of jumbo, but still have %d bytes\n", 1176 pkt_len); 1177 1178 out: 1179 return num_frags; 1180 } 1181 1182 static int qede_rx_process_tpa_cqe(struct qede_dev *edev, 1183 struct qede_fastpath *fp, 1184 struct qede_rx_queue *rxq, 1185 union eth_rx_cqe *cqe, 1186 enum eth_rx_cqe_type type) 1187 { 1188 switch (type) { 1189 case ETH_RX_CQE_TYPE_TPA_START: 1190 qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start); 1191 return 0; 1192 case ETH_RX_CQE_TYPE_TPA_CONT: 1193 qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont); 1194 return 0; 1195 case ETH_RX_CQE_TYPE_TPA_END: 1196 return qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end); 1197 default: 1198 return 0; 1199 } 1200 } 1201 1202 static int qede_rx_process_cqe(struct qede_dev *edev, 1203 struct qede_fastpath *fp, 1204 struct qede_rx_queue *rxq) 1205 { 1206 struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog); 1207 struct eth_fast_path_rx_reg_cqe *fp_cqe; 1208 u16 len, pad, bd_cons_idx, parse_flag; 1209 enum eth_rx_cqe_type cqe_type; 1210 union eth_rx_cqe *cqe; 1211 struct sw_rx_data *bd; 1212 struct sk_buff *skb; 1213 __le16 flags; 1214 u8 csum_flag; 1215 1216 /* Get the CQE from the completion ring */ 1217 cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring); 1218 cqe_type = cqe->fast_path_regular.type; 1219 1220 /* Process an unlikely slowpath event */ 1221 if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) { 1222 struct eth_slow_path_rx_cqe *sp_cqe; 1223 1224 sp_cqe = (struct eth_slow_path_rx_cqe *)cqe; 1225 edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe); 1226 return 0; 1227 } 1228 1229 /* Handle TPA cqes */ 1230 if (cqe_type != ETH_RX_CQE_TYPE_REGULAR) 1231 return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type); 1232 1233 /* Get the data from the SW ring; Consume it only after it's evident 1234 * we wouldn't recycle it. 1235 */ 1236 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX; 1237 bd = &rxq->sw_rx_ring[bd_cons_idx]; 1238 1239 fp_cqe = &cqe->fast_path_regular; 1240 len = le16_to_cpu(fp_cqe->len_on_first_bd); 1241 pad = fp_cqe->placement_offset + rxq->rx_headroom; 1242 1243 /* Run eBPF program if one is attached */ 1244 if (xdp_prog) 1245 if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe, 1246 &pad, &len)) 1247 return 0; 1248 1249 /* If this is an error packet then drop it */ 1250 flags = cqe->fast_path_regular.pars_flags.flags; 1251 parse_flag = le16_to_cpu(flags); 1252 1253 csum_flag = qede_check_csum(parse_flag); 1254 if (unlikely(csum_flag == QEDE_CSUM_ERROR)) { 1255 if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag)) 1256 rxq->rx_ip_frags++; 1257 else 1258 rxq->rx_hw_errors++; 1259 } 1260 1261 /* Basic validation passed; Need to prepare an SKB. This would also 1262 * guarantee to finally consume the first BD upon success. 1263 */ 1264 skb = qede_rx_build_skb(edev, rxq, bd, len, pad); 1265 if (!skb) { 1266 rxq->rx_alloc_errors++; 1267 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num); 1268 return 0; 1269 } 1270 1271 /* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed 1272 * by a single cqe. 1273 */ 1274 if (fp_cqe->bd_num > 1) { 1275 u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb, 1276 fp_cqe, len); 1277 1278 if (unlikely(unmapped_frags > 0)) { 1279 qede_recycle_rx_bd_ring(rxq, unmapped_frags); 1280 dev_kfree_skb_any(skb); 1281 return 0; 1282 } 1283 } 1284 1285 /* The SKB contains all the data. Now prepare meta-magic */ 1286 skb->protocol = eth_type_trans(skb, edev->ndev); 1287 qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash); 1288 qede_set_skb_csum(skb, csum_flag); 1289 skb_record_rx_queue(skb, rxq->rxq_id); 1290 qede_ptp_record_rx_ts(edev, cqe, skb); 1291 1292 /* SKB is prepared - pass it to stack */ 1293 qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag)); 1294 1295 return 1; 1296 } 1297 1298 static int qede_rx_int(struct qede_fastpath *fp, int budget) 1299 { 1300 struct qede_rx_queue *rxq = fp->rxq; 1301 struct qede_dev *edev = fp->edev; 1302 int work_done = 0, rcv_pkts = 0; 1303 u16 hw_comp_cons, sw_comp_cons; 1304 1305 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr); 1306 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 1307 1308 /* Memory barrier to prevent the CPU from doing speculative reads of CQE 1309 * / BD in the while-loop before reading hw_comp_cons. If the CQE is 1310 * read before it is written by FW, then FW writes CQE and SB, and then 1311 * the CPU reads the hw_comp_cons, it will use an old CQE. 1312 */ 1313 rmb(); 1314 1315 /* Loop to complete all indicated BDs */ 1316 while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) { 1317 rcv_pkts += qede_rx_process_cqe(edev, fp, rxq); 1318 qed_chain_recycle_consumed(&rxq->rx_comp_ring); 1319 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 1320 work_done++; 1321 } 1322 1323 rxq->rcv_pkts += rcv_pkts; 1324 1325 /* Allocate replacement buffers */ 1326 while (rxq->num_rx_buffers - rxq->filled_buffers) 1327 if (qede_alloc_rx_buffer(rxq, false)) 1328 break; 1329 1330 /* Update producers */ 1331 qede_update_rx_prod(edev, rxq); 1332 1333 return work_done; 1334 } 1335 1336 static bool qede_poll_is_more_work(struct qede_fastpath *fp) 1337 { 1338 qed_sb_update_sb_idx(fp->sb_info); 1339 1340 /* *_has_*_work() reads the status block, thus we need to ensure that 1341 * status block indices have been actually read (qed_sb_update_sb_idx) 1342 * prior to this check (*_has_*_work) so that we won't write the 1343 * "newer" value of the status block to HW (if there was a DMA right 1344 * after qede_has_rx_work and if there is no rmb, the memory reading 1345 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb). 1346 * In this case there will never be another interrupt until there is 1347 * another update of the status block, while there is still unhandled 1348 * work. 1349 */ 1350 rmb(); 1351 1352 if (likely(fp->type & QEDE_FASTPATH_RX)) 1353 if (qede_has_rx_work(fp->rxq)) 1354 return true; 1355 1356 if (fp->type & QEDE_FASTPATH_XDP) 1357 if (qede_txq_has_work(fp->xdp_tx)) 1358 return true; 1359 1360 if (likely(fp->type & QEDE_FASTPATH_TX)) { 1361 int cos; 1362 1363 for_each_cos_in_txq(fp->edev, cos) { 1364 if (qede_txq_has_work(&fp->txq[cos])) 1365 return true; 1366 } 1367 } 1368 1369 return false; 1370 } 1371 1372 /********************* 1373 * NDO & API related * 1374 *********************/ 1375 int qede_poll(struct napi_struct *napi, int budget) 1376 { 1377 struct qede_fastpath *fp = container_of(napi, struct qede_fastpath, 1378 napi); 1379 struct qede_dev *edev = fp->edev; 1380 int rx_work_done = 0; 1381 1382 if (likely(fp->type & QEDE_FASTPATH_TX)) { 1383 int cos; 1384 1385 for_each_cos_in_txq(fp->edev, cos) { 1386 if (qede_txq_has_work(&fp->txq[cos])) 1387 qede_tx_int(edev, &fp->txq[cos]); 1388 } 1389 } 1390 1391 if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx)) 1392 qede_xdp_tx_int(edev, fp->xdp_tx); 1393 1394 rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) && 1395 qede_has_rx_work(fp->rxq)) ? 1396 qede_rx_int(fp, budget) : 0; 1397 if (rx_work_done < budget) { 1398 if (!qede_poll_is_more_work(fp)) { 1399 napi_complete_done(napi, rx_work_done); 1400 1401 /* Update and reenable interrupts */ 1402 qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1); 1403 } else { 1404 rx_work_done = budget; 1405 } 1406 } 1407 1408 if (fp->xdp_xmit) { 1409 u16 xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl); 1410 1411 fp->xdp_xmit = 0; 1412 fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod); 1413 qede_update_tx_producer(fp->xdp_tx); 1414 } 1415 1416 return rx_work_done; 1417 } 1418 1419 irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie) 1420 { 1421 struct qede_fastpath *fp = fp_cookie; 1422 1423 qed_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0 /*do not update*/); 1424 1425 napi_schedule_irqoff(&fp->napi); 1426 return IRQ_HANDLED; 1427 } 1428 1429 /* Main transmit function */ 1430 netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev) 1431 { 1432 struct qede_dev *edev = netdev_priv(ndev); 1433 struct netdev_queue *netdev_txq; 1434 struct qede_tx_queue *txq; 1435 struct eth_tx_1st_bd *first_bd; 1436 struct eth_tx_2nd_bd *second_bd = NULL; 1437 struct eth_tx_3rd_bd *third_bd = NULL; 1438 struct eth_tx_bd *tx_data_bd = NULL; 1439 u16 txq_index, val = 0; 1440 u8 nbd = 0; 1441 dma_addr_t mapping; 1442 int rc, frag_idx = 0, ipv6_ext = 0; 1443 u8 xmit_type; 1444 u16 idx; 1445 u16 hlen; 1446 bool data_split = false; 1447 1448 /* Get tx-queue context and netdev index */ 1449 txq_index = skb_get_queue_mapping(skb); 1450 WARN_ON(txq_index >= QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc); 1451 txq = QEDE_NDEV_TXQ_ID_TO_TXQ(edev, txq_index); 1452 netdev_txq = netdev_get_tx_queue(ndev, txq_index); 1453 1454 WARN_ON(qed_chain_get_elem_left(&txq->tx_pbl) < (MAX_SKB_FRAGS + 1)); 1455 1456 xmit_type = qede_xmit_type(skb, &ipv6_ext); 1457 1458 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET) 1459 if (qede_pkt_req_lin(skb, xmit_type)) { 1460 if (skb_linearize(skb)) { 1461 txq->tx_mem_alloc_err++; 1462 1463 dev_kfree_skb_any(skb); 1464 return NETDEV_TX_OK; 1465 } 1466 } 1467 #endif 1468 1469 /* Fill the entry in the SW ring and the BDs in the FW ring */ 1470 idx = txq->sw_tx_prod; 1471 txq->sw_tx_ring.skbs[idx].skb = skb; 1472 first_bd = (struct eth_tx_1st_bd *) 1473 qed_chain_produce(&txq->tx_pbl); 1474 memset(first_bd, 0, sizeof(*first_bd)); 1475 first_bd->data.bd_flags.bitfields = 1476 1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT; 1477 1478 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)) 1479 qede_ptp_tx_ts(edev, skb); 1480 1481 /* Map skb linear data for DMA and set in the first BD */ 1482 mapping = dma_map_single(txq->dev, skb->data, 1483 skb_headlen(skb), DMA_TO_DEVICE); 1484 if (unlikely(dma_mapping_error(txq->dev, mapping))) { 1485 DP_NOTICE(edev, "SKB mapping failed\n"); 1486 qede_free_failed_tx_pkt(txq, first_bd, 0, false); 1487 qede_update_tx_producer(txq); 1488 return NETDEV_TX_OK; 1489 } 1490 nbd++; 1491 BD_SET_UNMAP_ADDR_LEN(first_bd, mapping, skb_headlen(skb)); 1492 1493 /* In case there is IPv6 with extension headers or LSO we need 2nd and 1494 * 3rd BDs. 1495 */ 1496 if (unlikely((xmit_type & XMIT_LSO) | ipv6_ext)) { 1497 second_bd = (struct eth_tx_2nd_bd *) 1498 qed_chain_produce(&txq->tx_pbl); 1499 memset(second_bd, 0, sizeof(*second_bd)); 1500 1501 nbd++; 1502 third_bd = (struct eth_tx_3rd_bd *) 1503 qed_chain_produce(&txq->tx_pbl); 1504 memset(third_bd, 0, sizeof(*third_bd)); 1505 1506 nbd++; 1507 /* We need to fill in additional data in second_bd... */ 1508 tx_data_bd = (struct eth_tx_bd *)second_bd; 1509 } 1510 1511 if (skb_vlan_tag_present(skb)) { 1512 first_bd->data.vlan = cpu_to_le16(skb_vlan_tag_get(skb)); 1513 first_bd->data.bd_flags.bitfields |= 1514 1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT; 1515 } 1516 1517 /* Fill the parsing flags & params according to the requested offload */ 1518 if (xmit_type & XMIT_L4_CSUM) { 1519 /* We don't re-calculate IP checksum as it is already done by 1520 * the upper stack 1521 */ 1522 first_bd->data.bd_flags.bitfields |= 1523 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT; 1524 1525 if (xmit_type & XMIT_ENC) { 1526 first_bd->data.bd_flags.bitfields |= 1527 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT; 1528 1529 val |= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT); 1530 } 1531 1532 /* Legacy FW had flipped behavior in regard to this bit - 1533 * I.e., needed to set to prevent FW from touching encapsulated 1534 * packets when it didn't need to. 1535 */ 1536 if (unlikely(txq->is_legacy)) 1537 val ^= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT); 1538 1539 /* If the packet is IPv6 with extension header, indicate that 1540 * to FW and pass few params, since the device cracker doesn't 1541 * support parsing IPv6 with extension header/s. 1542 */ 1543 if (unlikely(ipv6_ext)) 1544 qede_set_params_for_ipv6_ext(skb, second_bd, third_bd); 1545 } 1546 1547 if (xmit_type & XMIT_LSO) { 1548 first_bd->data.bd_flags.bitfields |= 1549 (1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT); 1550 third_bd->data.lso_mss = 1551 cpu_to_le16(skb_shinfo(skb)->gso_size); 1552 1553 if (unlikely(xmit_type & XMIT_ENC)) { 1554 first_bd->data.bd_flags.bitfields |= 1555 1 << ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT; 1556 1557 if (xmit_type & XMIT_ENC_GSO_L4_CSUM) { 1558 u8 tmp = ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT; 1559 1560 first_bd->data.bd_flags.bitfields |= 1 << tmp; 1561 } 1562 hlen = qede_get_skb_hlen(skb, true); 1563 } else { 1564 first_bd->data.bd_flags.bitfields |= 1565 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT; 1566 hlen = qede_get_skb_hlen(skb, false); 1567 } 1568 1569 /* @@@TBD - if will not be removed need to check */ 1570 third_bd->data.bitfields |= 1571 cpu_to_le16(1 << ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT); 1572 1573 /* Make life easier for FW guys who can't deal with header and 1574 * data on same BD. If we need to split, use the second bd... 1575 */ 1576 if (unlikely(skb_headlen(skb) > hlen)) { 1577 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1578 "TSO split header size is %d (%x:%x)\n", 1579 first_bd->nbytes, first_bd->addr.hi, 1580 first_bd->addr.lo); 1581 1582 mapping = HILO_U64(le32_to_cpu(first_bd->addr.hi), 1583 le32_to_cpu(first_bd->addr.lo)) + 1584 hlen; 1585 1586 BD_SET_UNMAP_ADDR_LEN(tx_data_bd, mapping, 1587 le16_to_cpu(first_bd->nbytes) - 1588 hlen); 1589 1590 /* this marks the BD as one that has no 1591 * individual mapping 1592 */ 1593 txq->sw_tx_ring.skbs[idx].flags |= QEDE_TSO_SPLIT_BD; 1594 1595 first_bd->nbytes = cpu_to_le16(hlen); 1596 1597 tx_data_bd = (struct eth_tx_bd *)third_bd; 1598 data_split = true; 1599 } 1600 } else { 1601 val |= ((skb->len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) << 1602 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT); 1603 } 1604 1605 first_bd->data.bitfields = cpu_to_le16(val); 1606 1607 /* Handle fragmented skb */ 1608 /* special handle for frags inside 2nd and 3rd bds.. */ 1609 while (tx_data_bd && frag_idx < skb_shinfo(skb)->nr_frags) { 1610 rc = map_frag_to_bd(txq, 1611 &skb_shinfo(skb)->frags[frag_idx], 1612 tx_data_bd); 1613 if (rc) { 1614 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split); 1615 qede_update_tx_producer(txq); 1616 return NETDEV_TX_OK; 1617 } 1618 1619 if (tx_data_bd == (struct eth_tx_bd *)second_bd) 1620 tx_data_bd = (struct eth_tx_bd *)third_bd; 1621 else 1622 tx_data_bd = NULL; 1623 1624 frag_idx++; 1625 } 1626 1627 /* map last frags into 4th, 5th .... */ 1628 for (; frag_idx < skb_shinfo(skb)->nr_frags; frag_idx++, nbd++) { 1629 tx_data_bd = (struct eth_tx_bd *) 1630 qed_chain_produce(&txq->tx_pbl); 1631 1632 memset(tx_data_bd, 0, sizeof(*tx_data_bd)); 1633 1634 rc = map_frag_to_bd(txq, 1635 &skb_shinfo(skb)->frags[frag_idx], 1636 tx_data_bd); 1637 if (rc) { 1638 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split); 1639 qede_update_tx_producer(txq); 1640 return NETDEV_TX_OK; 1641 } 1642 } 1643 1644 /* update the first BD with the actual num BDs */ 1645 first_bd->data.nbds = nbd; 1646 1647 netdev_tx_sent_queue(netdev_txq, skb->len); 1648 1649 skb_tx_timestamp(skb); 1650 1651 /* Advance packet producer only before sending the packet since mapping 1652 * of pages may fail. 1653 */ 1654 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers; 1655 1656 /* 'next page' entries are counted in the producer value */ 1657 txq->tx_db.data.bd_prod = 1658 cpu_to_le16(qed_chain_get_prod_idx(&txq->tx_pbl)); 1659 1660 if (!netdev_xmit_more() || netif_xmit_stopped(netdev_txq)) 1661 qede_update_tx_producer(txq); 1662 1663 if (unlikely(qed_chain_get_elem_left(&txq->tx_pbl) 1664 < (MAX_SKB_FRAGS + 1))) { 1665 if (netdev_xmit_more()) 1666 qede_update_tx_producer(txq); 1667 1668 netif_tx_stop_queue(netdev_txq); 1669 txq->stopped_cnt++; 1670 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1671 "Stop queue was called\n"); 1672 /* paired memory barrier is in qede_tx_int(), we have to keep 1673 * ordering of set_bit() in netif_tx_stop_queue() and read of 1674 * fp->bd_tx_cons 1675 */ 1676 smp_mb(); 1677 1678 if ((qed_chain_get_elem_left(&txq->tx_pbl) >= 1679 (MAX_SKB_FRAGS + 1)) && 1680 (edev->state == QEDE_STATE_OPEN)) { 1681 netif_tx_wake_queue(netdev_txq); 1682 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1683 "Wake queue was called\n"); 1684 } 1685 } 1686 1687 return NETDEV_TX_OK; 1688 } 1689 1690 u16 qede_select_queue(struct net_device *dev, struct sk_buff *skb, 1691 struct net_device *sb_dev) 1692 { 1693 struct qede_dev *edev = netdev_priv(dev); 1694 int total_txq; 1695 1696 total_txq = QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc; 1697 1698 return QEDE_TSS_COUNT(edev) ? 1699 netdev_pick_tx(dev, skb, NULL) % total_txq : 0; 1700 } 1701 1702 /* 8B udp header + 8B base tunnel header + 32B option length */ 1703 #define QEDE_MAX_TUN_HDR_LEN 48 1704 1705 netdev_features_t qede_features_check(struct sk_buff *skb, 1706 struct net_device *dev, 1707 netdev_features_t features) 1708 { 1709 if (skb->encapsulation) { 1710 u8 l4_proto = 0; 1711 1712 switch (vlan_get_protocol(skb)) { 1713 case htons(ETH_P_IP): 1714 l4_proto = ip_hdr(skb)->protocol; 1715 break; 1716 case htons(ETH_P_IPV6): 1717 l4_proto = ipv6_hdr(skb)->nexthdr; 1718 break; 1719 default: 1720 return features; 1721 } 1722 1723 /* Disable offloads for geneve tunnels, as HW can't parse 1724 * the geneve header which has option length greater than 32b 1725 * and disable offloads for the ports which are not offloaded. 1726 */ 1727 if (l4_proto == IPPROTO_UDP) { 1728 struct qede_dev *edev = netdev_priv(dev); 1729 u16 hdrlen, vxln_port, gnv_port; 1730 1731 hdrlen = QEDE_MAX_TUN_HDR_LEN; 1732 vxln_port = edev->vxlan_dst_port; 1733 gnv_port = edev->geneve_dst_port; 1734 1735 if ((skb_inner_mac_header(skb) - 1736 skb_transport_header(skb)) > hdrlen || 1737 (ntohs(udp_hdr(skb)->dest) != vxln_port && 1738 ntohs(udp_hdr(skb)->dest) != gnv_port)) 1739 return features & ~(NETIF_F_CSUM_MASK | 1740 NETIF_F_GSO_MASK); 1741 } 1742 } 1743 1744 return features; 1745 } 1746