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 & NUM_TX_BDS_MAX; 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 & NUM_TX_BDS_MAX; 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 /* mmiowb is needed to synchronize doorbell writes from more than one 324 * processor. It guarantees that the write arrives to the device before 325 * the queue lock is released and another start_xmit is called (possibly 326 * on another CPU). Without this barrier, the next doorbell can bypass 327 * this doorbell. This is applicable to IA64/Altix systems. 328 */ 329 mmiowb(); 330 } 331 332 static int qede_xdp_xmit(struct qede_dev *edev, struct qede_fastpath *fp, 333 struct sw_rx_data *metadata, u16 padding, u16 length) 334 { 335 struct qede_tx_queue *txq = fp->xdp_tx; 336 u16 idx = txq->sw_tx_prod & NUM_TX_BDS_MAX; 337 struct eth_tx_1st_bd *first_bd; 338 339 if (!qed_chain_get_elem_left(&txq->tx_pbl)) { 340 txq->stopped_cnt++; 341 return -ENOMEM; 342 } 343 344 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl); 345 346 memset(first_bd, 0, sizeof(*first_bd)); 347 first_bd->data.bd_flags.bitfields = 348 BIT(ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT); 349 first_bd->data.bitfields |= 350 (length & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) << 351 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT; 352 first_bd->data.nbds = 1; 353 354 /* We can safely ignore the offset, as it's 0 for XDP */ 355 BD_SET_UNMAP_ADDR_LEN(first_bd, metadata->mapping + padding, length); 356 357 /* Synchronize the buffer back to device, as program [probably] 358 * has changed it. 359 */ 360 dma_sync_single_for_device(&edev->pdev->dev, 361 metadata->mapping + padding, 362 length, PCI_DMA_TODEVICE); 363 364 txq->sw_tx_ring.xdp[idx].page = metadata->data; 365 txq->sw_tx_ring.xdp[idx].mapping = metadata->mapping; 366 txq->sw_tx_prod++; 367 368 /* Mark the fastpath for future XDP doorbell */ 369 fp->xdp_xmit = 1; 370 371 return 0; 372 } 373 374 int qede_txq_has_work(struct qede_tx_queue *txq) 375 { 376 u16 hw_bd_cons; 377 378 /* Tell compiler that consumer and producer can change */ 379 barrier(); 380 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 381 if (qed_chain_get_cons_idx(&txq->tx_pbl) == hw_bd_cons + 1) 382 return 0; 383 384 return hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl); 385 } 386 387 static void qede_xdp_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq) 388 { 389 u16 hw_bd_cons, idx; 390 391 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 392 barrier(); 393 394 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) { 395 qed_chain_consume(&txq->tx_pbl); 396 idx = txq->sw_tx_cons & NUM_TX_BDS_MAX; 397 398 dma_unmap_page(&edev->pdev->dev, 399 txq->sw_tx_ring.xdp[idx].mapping, 400 PAGE_SIZE, DMA_BIDIRECTIONAL); 401 __free_page(txq->sw_tx_ring.xdp[idx].page); 402 403 txq->sw_tx_cons++; 404 txq->xmit_pkts++; 405 } 406 } 407 408 static int qede_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq) 409 { 410 struct netdev_queue *netdev_txq; 411 u16 hw_bd_cons; 412 unsigned int pkts_compl = 0, bytes_compl = 0; 413 int rc; 414 415 netdev_txq = netdev_get_tx_queue(edev->ndev, txq->index); 416 417 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 418 barrier(); 419 420 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) { 421 int len = 0; 422 423 rc = qede_free_tx_pkt(edev, txq, &len); 424 if (rc) { 425 DP_NOTICE(edev, "hw_bd_cons = %d, chain_cons=%d\n", 426 hw_bd_cons, 427 qed_chain_get_cons_idx(&txq->tx_pbl)); 428 break; 429 } 430 431 bytes_compl += len; 432 pkts_compl++; 433 txq->sw_tx_cons++; 434 txq->xmit_pkts++; 435 } 436 437 netdev_tx_completed_queue(netdev_txq, pkts_compl, bytes_compl); 438 439 /* Need to make the tx_bd_cons update visible to start_xmit() 440 * before checking for netif_tx_queue_stopped(). Without the 441 * memory barrier, there is a small possibility that 442 * start_xmit() will miss it and cause the queue to be stopped 443 * forever. 444 * On the other hand we need an rmb() here to ensure the proper 445 * ordering of bit testing in the following 446 * netif_tx_queue_stopped(txq) call. 447 */ 448 smp_mb(); 449 450 if (unlikely(netif_tx_queue_stopped(netdev_txq))) { 451 /* Taking tx_lock is needed to prevent reenabling the queue 452 * while it's empty. This could have happen if rx_action() gets 453 * suspended in qede_tx_int() after the condition before 454 * netif_tx_wake_queue(), while tx_action (qede_start_xmit()): 455 * 456 * stops the queue->sees fresh tx_bd_cons->releases the queue-> 457 * sends some packets consuming the whole queue again-> 458 * stops the queue 459 */ 460 461 __netif_tx_lock(netdev_txq, smp_processor_id()); 462 463 if ((netif_tx_queue_stopped(netdev_txq)) && 464 (edev->state == QEDE_STATE_OPEN) && 465 (qed_chain_get_elem_left(&txq->tx_pbl) 466 >= (MAX_SKB_FRAGS + 1))) { 467 netif_tx_wake_queue(netdev_txq); 468 DP_VERBOSE(edev, NETIF_MSG_TX_DONE, 469 "Wake queue was called\n"); 470 } 471 472 __netif_tx_unlock(netdev_txq); 473 } 474 475 return 0; 476 } 477 478 bool qede_has_rx_work(struct qede_rx_queue *rxq) 479 { 480 u16 hw_comp_cons, sw_comp_cons; 481 482 /* Tell compiler that status block fields can change */ 483 barrier(); 484 485 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr); 486 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 487 488 return hw_comp_cons != sw_comp_cons; 489 } 490 491 static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq) 492 { 493 qed_chain_consume(&rxq->rx_bd_ring); 494 rxq->sw_rx_cons++; 495 } 496 497 /* This function reuses the buffer(from an offset) from 498 * consumer index to producer index in the bd ring 499 */ 500 static inline void qede_reuse_page(struct qede_rx_queue *rxq, 501 struct sw_rx_data *curr_cons) 502 { 503 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring); 504 struct sw_rx_data *curr_prod; 505 dma_addr_t new_mapping; 506 507 curr_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX]; 508 *curr_prod = *curr_cons; 509 510 new_mapping = curr_prod->mapping + curr_prod->page_offset; 511 512 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping)); 513 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping) + 514 rxq->rx_headroom); 515 516 rxq->sw_rx_prod++; 517 curr_cons->data = NULL; 518 } 519 520 /* In case of allocation failures reuse buffers 521 * from consumer index to produce buffers for firmware 522 */ 523 void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count) 524 { 525 struct sw_rx_data *curr_cons; 526 527 for (; count > 0; count--) { 528 curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX]; 529 qede_reuse_page(rxq, curr_cons); 530 qede_rx_bd_ring_consume(rxq); 531 } 532 } 533 534 static inline int qede_realloc_rx_buffer(struct qede_rx_queue *rxq, 535 struct sw_rx_data *curr_cons) 536 { 537 /* Move to the next segment in the page */ 538 curr_cons->page_offset += rxq->rx_buf_seg_size; 539 540 if (curr_cons->page_offset == PAGE_SIZE) { 541 if (unlikely(qede_alloc_rx_buffer(rxq, true))) { 542 /* Since we failed to allocate new buffer 543 * current buffer can be used again. 544 */ 545 curr_cons->page_offset -= rxq->rx_buf_seg_size; 546 547 return -ENOMEM; 548 } 549 550 dma_unmap_page(rxq->dev, curr_cons->mapping, 551 PAGE_SIZE, rxq->data_direction); 552 } else { 553 /* Increment refcount of the page as we don't want 554 * network stack to take the ownership of the page 555 * which can be recycled multiple times by the driver. 556 */ 557 page_ref_inc(curr_cons->data); 558 qede_reuse_page(rxq, curr_cons); 559 } 560 561 return 0; 562 } 563 564 void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq) 565 { 566 u16 bd_prod = qed_chain_get_prod_idx(&rxq->rx_bd_ring); 567 u16 cqe_prod = qed_chain_get_prod_idx(&rxq->rx_comp_ring); 568 struct eth_rx_prod_data rx_prods = {0}; 569 570 /* Update producers */ 571 rx_prods.bd_prod = cpu_to_le16(bd_prod); 572 rx_prods.cqe_prod = cpu_to_le16(cqe_prod); 573 574 /* Make sure that the BD and SGE data is updated before updating the 575 * producers since FW might read the BD/SGE right after the producer 576 * is updated. 577 */ 578 wmb(); 579 580 internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods), 581 (u32 *)&rx_prods); 582 583 /* mmiowb is needed to synchronize doorbell writes from more than one 584 * processor. It guarantees that the write arrives to the device before 585 * the napi lock is released and another qede_poll is called (possibly 586 * on another CPU). Without this barrier, the next doorbell can bypass 587 * this doorbell. This is applicable to IA64/Altix systems. 588 */ 589 mmiowb(); 590 } 591 592 static void qede_get_rxhash(struct sk_buff *skb, u8 bitfields, __le32 rss_hash) 593 { 594 enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE; 595 enum rss_hash_type htype; 596 u32 hash = 0; 597 598 htype = GET_FIELD(bitfields, ETH_FAST_PATH_RX_REG_CQE_RSS_HASH_TYPE); 599 if (htype) { 600 hash_type = ((htype == RSS_HASH_TYPE_IPV4) || 601 (htype == RSS_HASH_TYPE_IPV6)) ? 602 PKT_HASH_TYPE_L3 : PKT_HASH_TYPE_L4; 603 hash = le32_to_cpu(rss_hash); 604 } 605 skb_set_hash(skb, hash, hash_type); 606 } 607 608 static void qede_set_skb_csum(struct sk_buff *skb, u8 csum_flag) 609 { 610 skb_checksum_none_assert(skb); 611 612 if (csum_flag & QEDE_CSUM_UNNECESSARY) 613 skb->ip_summed = CHECKSUM_UNNECESSARY; 614 615 if (csum_flag & QEDE_TUNN_CSUM_UNNECESSARY) { 616 skb->csum_level = 1; 617 skb->encapsulation = 1; 618 } 619 } 620 621 static inline void qede_skb_receive(struct qede_dev *edev, 622 struct qede_fastpath *fp, 623 struct qede_rx_queue *rxq, 624 struct sk_buff *skb, u16 vlan_tag) 625 { 626 if (vlan_tag) 627 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag); 628 629 napi_gro_receive(&fp->napi, skb); 630 } 631 632 static void qede_set_gro_params(struct qede_dev *edev, 633 struct sk_buff *skb, 634 struct eth_fast_path_rx_tpa_start_cqe *cqe) 635 { 636 u16 parsing_flags = le16_to_cpu(cqe->pars_flags.flags); 637 638 if (((parsing_flags >> PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) & 639 PARSING_AND_ERR_FLAGS_L3TYPE_MASK) == 2) 640 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6; 641 else 642 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4; 643 644 skb_shinfo(skb)->gso_size = __le16_to_cpu(cqe->len_on_first_bd) - 645 cqe->header_len; 646 } 647 648 static int qede_fill_frag_skb(struct qede_dev *edev, 649 struct qede_rx_queue *rxq, 650 u8 tpa_agg_index, u16 len_on_bd) 651 { 652 struct sw_rx_data *current_bd = &rxq->sw_rx_ring[rxq->sw_rx_cons & 653 NUM_RX_BDS_MAX]; 654 struct qede_agg_info *tpa_info = &rxq->tpa_info[tpa_agg_index]; 655 struct sk_buff *skb = tpa_info->skb; 656 657 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START)) 658 goto out; 659 660 /* Add one frag and update the appropriate fields in the skb */ 661 skb_fill_page_desc(skb, tpa_info->frag_id++, 662 current_bd->data, current_bd->page_offset, 663 len_on_bd); 664 665 if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) { 666 /* Incr page ref count to reuse on allocation failure 667 * so that it doesn't get freed while freeing SKB. 668 */ 669 page_ref_inc(current_bd->data); 670 goto out; 671 } 672 673 qed_chain_consume(&rxq->rx_bd_ring); 674 rxq->sw_rx_cons++; 675 676 skb->data_len += len_on_bd; 677 skb->truesize += rxq->rx_buf_seg_size; 678 skb->len += len_on_bd; 679 680 return 0; 681 682 out: 683 tpa_info->state = QEDE_AGG_STATE_ERROR; 684 qede_recycle_rx_bd_ring(rxq, 1); 685 686 return -ENOMEM; 687 } 688 689 static bool qede_tunn_exist(u16 flag) 690 { 691 return !!(flag & (PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK << 692 PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT)); 693 } 694 695 static u8 qede_check_tunn_csum(u16 flag) 696 { 697 u16 csum_flag = 0; 698 u8 tcsum = 0; 699 700 if (flag & (PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK << 701 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT)) 702 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK << 703 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT; 704 705 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK << 706 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) { 707 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK << 708 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT; 709 tcsum = QEDE_TUNN_CSUM_UNNECESSARY; 710 } 711 712 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK << 713 PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT | 714 PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK << 715 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT; 716 717 if (csum_flag & flag) 718 return QEDE_CSUM_ERROR; 719 720 return QEDE_CSUM_UNNECESSARY | tcsum; 721 } 722 723 static void qede_tpa_start(struct qede_dev *edev, 724 struct qede_rx_queue *rxq, 725 struct eth_fast_path_rx_tpa_start_cqe *cqe) 726 { 727 struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index]; 728 struct eth_rx_bd *rx_bd_cons = qed_chain_consume(&rxq->rx_bd_ring); 729 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring); 730 struct sw_rx_data *replace_buf = &tpa_info->buffer; 731 dma_addr_t mapping = tpa_info->buffer_mapping; 732 struct sw_rx_data *sw_rx_data_cons; 733 struct sw_rx_data *sw_rx_data_prod; 734 735 sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX]; 736 sw_rx_data_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX]; 737 738 /* Use pre-allocated replacement buffer - we can't release the agg. 739 * start until its over and we don't want to risk allocation failing 740 * here, so re-allocate when aggregation will be over. 741 */ 742 sw_rx_data_prod->mapping = replace_buf->mapping; 743 744 sw_rx_data_prod->data = replace_buf->data; 745 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(mapping)); 746 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(mapping)); 747 sw_rx_data_prod->page_offset = replace_buf->page_offset; 748 749 rxq->sw_rx_prod++; 750 751 /* move partial skb from cons to pool (don't unmap yet) 752 * save mapping, incase we drop the packet later on. 753 */ 754 tpa_info->buffer = *sw_rx_data_cons; 755 mapping = HILO_U64(le32_to_cpu(rx_bd_cons->addr.hi), 756 le32_to_cpu(rx_bd_cons->addr.lo)); 757 758 tpa_info->buffer_mapping = mapping; 759 rxq->sw_rx_cons++; 760 761 /* set tpa state to start only if we are able to allocate skb 762 * for this aggregation, otherwise mark as error and aggregation will 763 * be dropped 764 */ 765 tpa_info->skb = netdev_alloc_skb(edev->ndev, 766 le16_to_cpu(cqe->len_on_first_bd)); 767 if (unlikely(!tpa_info->skb)) { 768 DP_NOTICE(edev, "Failed to allocate SKB for gro\n"); 769 tpa_info->state = QEDE_AGG_STATE_ERROR; 770 goto cons_buf; 771 } 772 773 /* Start filling in the aggregation info */ 774 skb_put(tpa_info->skb, le16_to_cpu(cqe->len_on_first_bd)); 775 tpa_info->frag_id = 0; 776 tpa_info->state = QEDE_AGG_STATE_START; 777 778 /* Store some information from first CQE */ 779 tpa_info->start_cqe_placement_offset = cqe->placement_offset; 780 tpa_info->start_cqe_bd_len = le16_to_cpu(cqe->len_on_first_bd); 781 if ((le16_to_cpu(cqe->pars_flags.flags) >> 782 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) & 783 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK) 784 tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag); 785 else 786 tpa_info->vlan_tag = 0; 787 788 qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash); 789 790 /* This is needed in order to enable forwarding support */ 791 qede_set_gro_params(edev, tpa_info->skb, cqe); 792 793 cons_buf: /* We still need to handle bd_len_list to consume buffers */ 794 if (likely(cqe->ext_bd_len_list[0])) 795 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 796 le16_to_cpu(cqe->ext_bd_len_list[0])); 797 798 if (unlikely(cqe->ext_bd_len_list[1])) { 799 DP_ERR(edev, 800 "Unlikely - got a TPA aggregation with more than one ext_bd_len_list entry in the TPA start\n"); 801 tpa_info->state = QEDE_AGG_STATE_ERROR; 802 } 803 } 804 805 #ifdef CONFIG_INET 806 static void qede_gro_ip_csum(struct sk_buff *skb) 807 { 808 const struct iphdr *iph = ip_hdr(skb); 809 struct tcphdr *th; 810 811 skb_set_transport_header(skb, sizeof(struct iphdr)); 812 th = tcp_hdr(skb); 813 814 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb), 815 iph->saddr, iph->daddr, 0); 816 817 tcp_gro_complete(skb); 818 } 819 820 static void qede_gro_ipv6_csum(struct sk_buff *skb) 821 { 822 struct ipv6hdr *iph = ipv6_hdr(skb); 823 struct tcphdr *th; 824 825 skb_set_transport_header(skb, sizeof(struct ipv6hdr)); 826 th = tcp_hdr(skb); 827 828 th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb), 829 &iph->saddr, &iph->daddr, 0); 830 tcp_gro_complete(skb); 831 } 832 #endif 833 834 static void qede_gro_receive(struct qede_dev *edev, 835 struct qede_fastpath *fp, 836 struct sk_buff *skb, 837 u16 vlan_tag) 838 { 839 /* FW can send a single MTU sized packet from gro flow 840 * due to aggregation timeout/last segment etc. which 841 * is not expected to be a gro packet. If a skb has zero 842 * frags then simply push it in the stack as non gso skb. 843 */ 844 if (unlikely(!skb->data_len)) { 845 skb_shinfo(skb)->gso_type = 0; 846 skb_shinfo(skb)->gso_size = 0; 847 goto send_skb; 848 } 849 850 #ifdef CONFIG_INET 851 if (skb_shinfo(skb)->gso_size) { 852 skb_reset_network_header(skb); 853 854 switch (skb->protocol) { 855 case htons(ETH_P_IP): 856 qede_gro_ip_csum(skb); 857 break; 858 case htons(ETH_P_IPV6): 859 qede_gro_ipv6_csum(skb); 860 break; 861 default: 862 DP_ERR(edev, 863 "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n", 864 ntohs(skb->protocol)); 865 } 866 } 867 #endif 868 869 send_skb: 870 skb_record_rx_queue(skb, fp->rxq->rxq_id); 871 qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag); 872 } 873 874 static inline void qede_tpa_cont(struct qede_dev *edev, 875 struct qede_rx_queue *rxq, 876 struct eth_fast_path_rx_tpa_cont_cqe *cqe) 877 { 878 int i; 879 880 for (i = 0; cqe->len_list[i]; i++) 881 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 882 le16_to_cpu(cqe->len_list[i])); 883 884 if (unlikely(i > 1)) 885 DP_ERR(edev, 886 "Strange - TPA cont with more than a single len_list entry\n"); 887 } 888 889 static int qede_tpa_end(struct qede_dev *edev, 890 struct qede_fastpath *fp, 891 struct eth_fast_path_rx_tpa_end_cqe *cqe) 892 { 893 struct qede_rx_queue *rxq = fp->rxq; 894 struct qede_agg_info *tpa_info; 895 struct sk_buff *skb; 896 int i; 897 898 tpa_info = &rxq->tpa_info[cqe->tpa_agg_index]; 899 skb = tpa_info->skb; 900 901 for (i = 0; cqe->len_list[i]; i++) 902 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 903 le16_to_cpu(cqe->len_list[i])); 904 if (unlikely(i > 1)) 905 DP_ERR(edev, 906 "Strange - TPA emd with more than a single len_list entry\n"); 907 908 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START)) 909 goto err; 910 911 /* Sanity */ 912 if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1)) 913 DP_ERR(edev, 914 "Strange - TPA had %02x BDs, but SKB has only %d frags\n", 915 cqe->num_of_bds, tpa_info->frag_id); 916 if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len))) 917 DP_ERR(edev, 918 "Strange - total packet len [cqe] is %4x but SKB has len %04x\n", 919 le16_to_cpu(cqe->total_packet_len), skb->len); 920 921 memcpy(skb->data, 922 page_address(tpa_info->buffer.data) + 923 tpa_info->start_cqe_placement_offset + 924 tpa_info->buffer.page_offset, tpa_info->start_cqe_bd_len); 925 926 /* Finalize the SKB */ 927 skb->protocol = eth_type_trans(skb, edev->ndev); 928 skb->ip_summed = CHECKSUM_UNNECESSARY; 929 930 /* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count 931 * to skb_shinfo(skb)->gso_segs 932 */ 933 NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs); 934 935 qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag); 936 937 tpa_info->state = QEDE_AGG_STATE_NONE; 938 939 return 1; 940 err: 941 tpa_info->state = QEDE_AGG_STATE_NONE; 942 dev_kfree_skb_any(tpa_info->skb); 943 tpa_info->skb = NULL; 944 return 0; 945 } 946 947 static u8 qede_check_notunn_csum(u16 flag) 948 { 949 u16 csum_flag = 0; 950 u8 csum = 0; 951 952 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK << 953 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) { 954 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK << 955 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT; 956 csum = QEDE_CSUM_UNNECESSARY; 957 } 958 959 csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK << 960 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT; 961 962 if (csum_flag & flag) 963 return QEDE_CSUM_ERROR; 964 965 return csum; 966 } 967 968 static u8 qede_check_csum(u16 flag) 969 { 970 if (!qede_tunn_exist(flag)) 971 return qede_check_notunn_csum(flag); 972 else 973 return qede_check_tunn_csum(flag); 974 } 975 976 static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe, 977 u16 flag) 978 { 979 u8 tun_pars_flg = cqe->tunnel_pars_flags.flags; 980 981 if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK << 982 ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) || 983 (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK << 984 PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT))) 985 return true; 986 987 return false; 988 } 989 990 /* Return true iff packet is to be passed to stack */ 991 static bool qede_rx_xdp(struct qede_dev *edev, 992 struct qede_fastpath *fp, 993 struct qede_rx_queue *rxq, 994 struct bpf_prog *prog, 995 struct sw_rx_data *bd, 996 struct eth_fast_path_rx_reg_cqe *cqe, 997 u16 *data_offset, u16 *len) 998 { 999 struct xdp_buff xdp; 1000 enum xdp_action act; 1001 1002 xdp.data_hard_start = page_address(bd->data); 1003 xdp.data = xdp.data_hard_start + *data_offset; 1004 xdp.data_end = xdp.data + *len; 1005 1006 /* Queues always have a full reset currently, so for the time 1007 * being until there's atomic program replace just mark read 1008 * side for map helpers. 1009 */ 1010 rcu_read_lock(); 1011 act = bpf_prog_run_xdp(prog, &xdp); 1012 rcu_read_unlock(); 1013 1014 /* Recalculate, as XDP might have changed the headers */ 1015 *data_offset = xdp.data - xdp.data_hard_start; 1016 *len = xdp.data_end - xdp.data; 1017 1018 if (act == XDP_PASS) 1019 return true; 1020 1021 /* Count number of packets not to be passed to stack */ 1022 rxq->xdp_no_pass++; 1023 1024 switch (act) { 1025 case XDP_TX: 1026 /* We need the replacement buffer before transmit. */ 1027 if (qede_alloc_rx_buffer(rxq, true)) { 1028 qede_recycle_rx_bd_ring(rxq, 1); 1029 trace_xdp_exception(edev->ndev, prog, act); 1030 return false; 1031 } 1032 1033 /* Now if there's a transmission problem, we'd still have to 1034 * throw current buffer, as replacement was already allocated. 1035 */ 1036 if (qede_xdp_xmit(edev, fp, bd, *data_offset, *len)) { 1037 dma_unmap_page(rxq->dev, bd->mapping, 1038 PAGE_SIZE, DMA_BIDIRECTIONAL); 1039 __free_page(bd->data); 1040 trace_xdp_exception(edev->ndev, prog, act); 1041 } 1042 1043 /* Regardless, we've consumed an Rx BD */ 1044 qede_rx_bd_ring_consume(rxq); 1045 return false; 1046 1047 default: 1048 bpf_warn_invalid_xdp_action(act); 1049 case XDP_ABORTED: 1050 trace_xdp_exception(edev->ndev, prog, act); 1051 case XDP_DROP: 1052 qede_recycle_rx_bd_ring(rxq, cqe->bd_num); 1053 } 1054 1055 return false; 1056 } 1057 1058 static struct sk_buff *qede_rx_allocate_skb(struct qede_dev *edev, 1059 struct qede_rx_queue *rxq, 1060 struct sw_rx_data *bd, u16 len, 1061 u16 pad) 1062 { 1063 unsigned int offset = bd->page_offset + pad; 1064 struct skb_frag_struct *frag; 1065 struct page *page = bd->data; 1066 unsigned int pull_len; 1067 struct sk_buff *skb; 1068 unsigned char *va; 1069 1070 /* Allocate a new SKB with a sufficient large header len */ 1071 skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE); 1072 if (unlikely(!skb)) 1073 return NULL; 1074 1075 /* Copy data into SKB - if it's small, we can simply copy it and 1076 * re-use the already allcoated & mapped memory. 1077 */ 1078 if (len + pad <= edev->rx_copybreak) { 1079 memcpy(skb_put(skb, len), 1080 page_address(page) + offset, len); 1081 qede_reuse_page(rxq, bd); 1082 goto out; 1083 } 1084 1085 frag = &skb_shinfo(skb)->frags[0]; 1086 1087 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, 1088 page, offset, len, rxq->rx_buf_seg_size); 1089 1090 va = skb_frag_address(frag); 1091 pull_len = eth_get_headlen(va, QEDE_RX_HDR_SIZE); 1092 1093 /* Align the pull_len to optimize memcpy */ 1094 memcpy(skb->data, va, ALIGN(pull_len, sizeof(long))); 1095 1096 /* Correct the skb & frag sizes offset after the pull */ 1097 skb_frag_size_sub(frag, pull_len); 1098 frag->page_offset += pull_len; 1099 skb->data_len -= pull_len; 1100 skb->tail += pull_len; 1101 1102 if (unlikely(qede_realloc_rx_buffer(rxq, bd))) { 1103 /* Incr page ref count to reuse on allocation failure so 1104 * that it doesn't get freed while freeing SKB [as its 1105 * already mapped there]. 1106 */ 1107 page_ref_inc(page); 1108 dev_kfree_skb_any(skb); 1109 return NULL; 1110 } 1111 1112 out: 1113 /* We've consumed the first BD and prepared an SKB */ 1114 qede_rx_bd_ring_consume(rxq); 1115 return skb; 1116 } 1117 1118 static int qede_rx_build_jumbo(struct qede_dev *edev, 1119 struct qede_rx_queue *rxq, 1120 struct sk_buff *skb, 1121 struct eth_fast_path_rx_reg_cqe *cqe, 1122 u16 first_bd_len) 1123 { 1124 u16 pkt_len = le16_to_cpu(cqe->pkt_len); 1125 struct sw_rx_data *bd; 1126 u16 bd_cons_idx; 1127 u8 num_frags; 1128 1129 pkt_len -= first_bd_len; 1130 1131 /* We've already used one BD for the SKB. Now take care of the rest */ 1132 for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) { 1133 u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size : 1134 pkt_len; 1135 1136 if (unlikely(!cur_size)) { 1137 DP_ERR(edev, 1138 "Still got %d BDs for mapping jumbo, but length became 0\n", 1139 num_frags); 1140 goto out; 1141 } 1142 1143 /* We need a replacement buffer for each BD */ 1144 if (unlikely(qede_alloc_rx_buffer(rxq, true))) 1145 goto out; 1146 1147 /* Now that we've allocated the replacement buffer, 1148 * we can safely consume the next BD and map it to the SKB. 1149 */ 1150 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX; 1151 bd = &rxq->sw_rx_ring[bd_cons_idx]; 1152 qede_rx_bd_ring_consume(rxq); 1153 1154 dma_unmap_page(rxq->dev, bd->mapping, 1155 PAGE_SIZE, DMA_FROM_DEVICE); 1156 1157 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags++, 1158 bd->data, 0, cur_size); 1159 1160 skb->truesize += PAGE_SIZE; 1161 skb->data_len += cur_size; 1162 skb->len += cur_size; 1163 pkt_len -= cur_size; 1164 } 1165 1166 if (unlikely(pkt_len)) 1167 DP_ERR(edev, 1168 "Mapped all BDs of jumbo, but still have %d bytes\n", 1169 pkt_len); 1170 1171 out: 1172 return num_frags; 1173 } 1174 1175 static int qede_rx_process_tpa_cqe(struct qede_dev *edev, 1176 struct qede_fastpath *fp, 1177 struct qede_rx_queue *rxq, 1178 union eth_rx_cqe *cqe, 1179 enum eth_rx_cqe_type type) 1180 { 1181 switch (type) { 1182 case ETH_RX_CQE_TYPE_TPA_START: 1183 qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start); 1184 return 0; 1185 case ETH_RX_CQE_TYPE_TPA_CONT: 1186 qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont); 1187 return 0; 1188 case ETH_RX_CQE_TYPE_TPA_END: 1189 return qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end); 1190 default: 1191 return 0; 1192 } 1193 } 1194 1195 static int qede_rx_process_cqe(struct qede_dev *edev, 1196 struct qede_fastpath *fp, 1197 struct qede_rx_queue *rxq) 1198 { 1199 struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog); 1200 struct eth_fast_path_rx_reg_cqe *fp_cqe; 1201 u16 len, pad, bd_cons_idx, parse_flag; 1202 enum eth_rx_cqe_type cqe_type; 1203 union eth_rx_cqe *cqe; 1204 struct sw_rx_data *bd; 1205 struct sk_buff *skb; 1206 __le16 flags; 1207 u8 csum_flag; 1208 1209 /* Get the CQE from the completion ring */ 1210 cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring); 1211 cqe_type = cqe->fast_path_regular.type; 1212 1213 /* Process an unlikely slowpath event */ 1214 if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) { 1215 struct eth_slow_path_rx_cqe *sp_cqe; 1216 1217 sp_cqe = (struct eth_slow_path_rx_cqe *)cqe; 1218 edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe); 1219 return 0; 1220 } 1221 1222 /* Handle TPA cqes */ 1223 if (cqe_type != ETH_RX_CQE_TYPE_REGULAR) 1224 return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type); 1225 1226 /* Get the data from the SW ring; Consume it only after it's evident 1227 * we wouldn't recycle it. 1228 */ 1229 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX; 1230 bd = &rxq->sw_rx_ring[bd_cons_idx]; 1231 1232 fp_cqe = &cqe->fast_path_regular; 1233 len = le16_to_cpu(fp_cqe->len_on_first_bd); 1234 pad = fp_cqe->placement_offset + rxq->rx_headroom; 1235 1236 /* Run eBPF program if one is attached */ 1237 if (xdp_prog) 1238 if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe, 1239 &pad, &len)) 1240 return 0; 1241 1242 /* If this is an error packet then drop it */ 1243 flags = cqe->fast_path_regular.pars_flags.flags; 1244 parse_flag = le16_to_cpu(flags); 1245 1246 csum_flag = qede_check_csum(parse_flag); 1247 if (unlikely(csum_flag == QEDE_CSUM_ERROR)) { 1248 if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag)) { 1249 rxq->rx_ip_frags++; 1250 } else { 1251 DP_NOTICE(edev, 1252 "CQE has error, flags = %x, dropping incoming packet\n", 1253 parse_flag); 1254 rxq->rx_hw_errors++; 1255 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num); 1256 return 0; 1257 } 1258 } 1259 1260 /* Basic validation passed; Need to prepare an SKB. This would also 1261 * guarantee to finally consume the first BD upon success. 1262 */ 1263 skb = qede_rx_allocate_skb(edev, rxq, bd, len, pad); 1264 if (!skb) { 1265 rxq->rx_alloc_errors++; 1266 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num); 1267 return 0; 1268 } 1269 1270 /* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed 1271 * by a single cqe. 1272 */ 1273 if (fp_cqe->bd_num > 1) { 1274 u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb, 1275 fp_cqe, len); 1276 1277 if (unlikely(unmapped_frags > 0)) { 1278 qede_recycle_rx_bd_ring(rxq, unmapped_frags); 1279 dev_kfree_skb_any(skb); 1280 return 0; 1281 } 1282 } 1283 1284 /* The SKB contains all the data. Now prepare meta-magic */ 1285 skb->protocol = eth_type_trans(skb, edev->ndev); 1286 qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash); 1287 qede_set_skb_csum(skb, csum_flag); 1288 skb_record_rx_queue(skb, rxq->rxq_id); 1289 qede_ptp_record_rx_ts(edev, cqe, skb); 1290 1291 /* SKB is prepared - pass it to stack */ 1292 qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag)); 1293 1294 return 1; 1295 } 1296 1297 static int qede_rx_int(struct qede_fastpath *fp, int budget) 1298 { 1299 struct qede_rx_queue *rxq = fp->rxq; 1300 struct qede_dev *edev = fp->edev; 1301 int work_done = 0, rcv_pkts = 0; 1302 u16 hw_comp_cons, sw_comp_cons; 1303 1304 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr); 1305 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 1306 1307 /* Memory barrier to prevent the CPU from doing speculative reads of CQE 1308 * / BD in the while-loop before reading hw_comp_cons. If the CQE is 1309 * read before it is written by FW, then FW writes CQE and SB, and then 1310 * the CPU reads the hw_comp_cons, it will use an old CQE. 1311 */ 1312 rmb(); 1313 1314 /* Loop to complete all indicated BDs */ 1315 while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) { 1316 rcv_pkts += qede_rx_process_cqe(edev, fp, rxq); 1317 qed_chain_recycle_consumed(&rxq->rx_comp_ring); 1318 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 1319 work_done++; 1320 } 1321 1322 rxq->rcv_pkts += rcv_pkts; 1323 1324 /* Allocate replacement buffers */ 1325 while (rxq->num_rx_buffers - rxq->filled_buffers) 1326 if (qede_alloc_rx_buffer(rxq, false)) 1327 break; 1328 1329 /* Update producers */ 1330 qede_update_rx_prod(edev, rxq); 1331 1332 return work_done; 1333 } 1334 1335 static bool qede_poll_is_more_work(struct qede_fastpath *fp) 1336 { 1337 qed_sb_update_sb_idx(fp->sb_info); 1338 1339 /* *_has_*_work() reads the status block, thus we need to ensure that 1340 * status block indices have been actually read (qed_sb_update_sb_idx) 1341 * prior to this check (*_has_*_work) so that we won't write the 1342 * "newer" value of the status block to HW (if there was a DMA right 1343 * after qede_has_rx_work and if there is no rmb, the memory reading 1344 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb). 1345 * In this case there will never be another interrupt until there is 1346 * another update of the status block, while there is still unhandled 1347 * work. 1348 */ 1349 rmb(); 1350 1351 if (likely(fp->type & QEDE_FASTPATH_RX)) 1352 if (qede_has_rx_work(fp->rxq)) 1353 return true; 1354 1355 if (fp->type & QEDE_FASTPATH_XDP) 1356 if (qede_txq_has_work(fp->xdp_tx)) 1357 return true; 1358 1359 if (likely(fp->type & QEDE_FASTPATH_TX)) 1360 if (qede_txq_has_work(fp->txq)) 1361 return true; 1362 1363 return false; 1364 } 1365 1366 /********************* 1367 * NDO & API related * 1368 *********************/ 1369 int qede_poll(struct napi_struct *napi, int budget) 1370 { 1371 struct qede_fastpath *fp = container_of(napi, struct qede_fastpath, 1372 napi); 1373 struct qede_dev *edev = fp->edev; 1374 int rx_work_done = 0; 1375 1376 if (likely(fp->type & QEDE_FASTPATH_TX) && qede_txq_has_work(fp->txq)) 1377 qede_tx_int(edev, fp->txq); 1378 1379 if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx)) 1380 qede_xdp_tx_int(edev, fp->xdp_tx); 1381 1382 rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) && 1383 qede_has_rx_work(fp->rxq)) ? 1384 qede_rx_int(fp, budget) : 0; 1385 if (rx_work_done < budget) { 1386 if (!qede_poll_is_more_work(fp)) { 1387 napi_complete_done(napi, rx_work_done); 1388 1389 /* Update and reenable interrupts */ 1390 qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1); 1391 } else { 1392 rx_work_done = budget; 1393 } 1394 } 1395 1396 if (fp->xdp_xmit) { 1397 u16 xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl); 1398 1399 fp->xdp_xmit = 0; 1400 fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod); 1401 qede_update_tx_producer(fp->xdp_tx); 1402 } 1403 1404 return rx_work_done; 1405 } 1406 1407 irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie) 1408 { 1409 struct qede_fastpath *fp = fp_cookie; 1410 1411 qed_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0 /*do not update*/); 1412 1413 napi_schedule_irqoff(&fp->napi); 1414 return IRQ_HANDLED; 1415 } 1416 1417 /* Main transmit function */ 1418 netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev) 1419 { 1420 struct qede_dev *edev = netdev_priv(ndev); 1421 struct netdev_queue *netdev_txq; 1422 struct qede_tx_queue *txq; 1423 struct eth_tx_1st_bd *first_bd; 1424 struct eth_tx_2nd_bd *second_bd = NULL; 1425 struct eth_tx_3rd_bd *third_bd = NULL; 1426 struct eth_tx_bd *tx_data_bd = NULL; 1427 u16 txq_index; 1428 u8 nbd = 0; 1429 dma_addr_t mapping; 1430 int rc, frag_idx = 0, ipv6_ext = 0; 1431 u8 xmit_type; 1432 u16 idx; 1433 u16 hlen; 1434 bool data_split = false; 1435 1436 /* Get tx-queue context and netdev index */ 1437 txq_index = skb_get_queue_mapping(skb); 1438 WARN_ON(txq_index >= QEDE_TSS_COUNT(edev)); 1439 txq = edev->fp_array[edev->fp_num_rx + txq_index].txq; 1440 netdev_txq = netdev_get_tx_queue(ndev, txq_index); 1441 1442 WARN_ON(qed_chain_get_elem_left(&txq->tx_pbl) < (MAX_SKB_FRAGS + 1)); 1443 1444 xmit_type = qede_xmit_type(skb, &ipv6_ext); 1445 1446 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET) 1447 if (qede_pkt_req_lin(skb, xmit_type)) { 1448 if (skb_linearize(skb)) { 1449 DP_NOTICE(edev, 1450 "SKB linearization failed - silently dropping this SKB\n"); 1451 dev_kfree_skb_any(skb); 1452 return NETDEV_TX_OK; 1453 } 1454 } 1455 #endif 1456 1457 /* Fill the entry in the SW ring and the BDs in the FW ring */ 1458 idx = txq->sw_tx_prod & NUM_TX_BDS_MAX; 1459 txq->sw_tx_ring.skbs[idx].skb = skb; 1460 first_bd = (struct eth_tx_1st_bd *) 1461 qed_chain_produce(&txq->tx_pbl); 1462 memset(first_bd, 0, sizeof(*first_bd)); 1463 first_bd->data.bd_flags.bitfields = 1464 1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT; 1465 1466 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)) 1467 qede_ptp_tx_ts(edev, skb); 1468 1469 /* Map skb linear data for DMA and set in the first BD */ 1470 mapping = dma_map_single(txq->dev, skb->data, 1471 skb_headlen(skb), DMA_TO_DEVICE); 1472 if (unlikely(dma_mapping_error(txq->dev, mapping))) { 1473 DP_NOTICE(edev, "SKB mapping failed\n"); 1474 qede_free_failed_tx_pkt(txq, first_bd, 0, false); 1475 qede_update_tx_producer(txq); 1476 return NETDEV_TX_OK; 1477 } 1478 nbd++; 1479 BD_SET_UNMAP_ADDR_LEN(first_bd, mapping, skb_headlen(skb)); 1480 1481 /* In case there is IPv6 with extension headers or LSO we need 2nd and 1482 * 3rd BDs. 1483 */ 1484 if (unlikely((xmit_type & XMIT_LSO) | ipv6_ext)) { 1485 second_bd = (struct eth_tx_2nd_bd *) 1486 qed_chain_produce(&txq->tx_pbl); 1487 memset(second_bd, 0, sizeof(*second_bd)); 1488 1489 nbd++; 1490 third_bd = (struct eth_tx_3rd_bd *) 1491 qed_chain_produce(&txq->tx_pbl); 1492 memset(third_bd, 0, sizeof(*third_bd)); 1493 1494 nbd++; 1495 /* We need to fill in additional data in second_bd... */ 1496 tx_data_bd = (struct eth_tx_bd *)second_bd; 1497 } 1498 1499 if (skb_vlan_tag_present(skb)) { 1500 first_bd->data.vlan = cpu_to_le16(skb_vlan_tag_get(skb)); 1501 first_bd->data.bd_flags.bitfields |= 1502 1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT; 1503 } 1504 1505 /* Fill the parsing flags & params according to the requested offload */ 1506 if (xmit_type & XMIT_L4_CSUM) { 1507 /* We don't re-calculate IP checksum as it is already done by 1508 * the upper stack 1509 */ 1510 first_bd->data.bd_flags.bitfields |= 1511 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT; 1512 1513 if (xmit_type & XMIT_ENC) { 1514 first_bd->data.bd_flags.bitfields |= 1515 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT; 1516 first_bd->data.bitfields |= 1517 1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT; 1518 } 1519 1520 /* Legacy FW had flipped behavior in regard to this bit - 1521 * I.e., needed to set to prevent FW from touching encapsulated 1522 * packets when it didn't need to. 1523 */ 1524 if (unlikely(txq->is_legacy)) 1525 first_bd->data.bitfields ^= 1526 1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT; 1527 1528 /* If the packet is IPv6 with extension header, indicate that 1529 * to FW and pass few params, since the device cracker doesn't 1530 * support parsing IPv6 with extension header/s. 1531 */ 1532 if (unlikely(ipv6_ext)) 1533 qede_set_params_for_ipv6_ext(skb, second_bd, third_bd); 1534 } 1535 1536 if (xmit_type & XMIT_LSO) { 1537 first_bd->data.bd_flags.bitfields |= 1538 (1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT); 1539 third_bd->data.lso_mss = 1540 cpu_to_le16(skb_shinfo(skb)->gso_size); 1541 1542 if (unlikely(xmit_type & XMIT_ENC)) { 1543 first_bd->data.bd_flags.bitfields |= 1544 1 << ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT; 1545 1546 if (xmit_type & XMIT_ENC_GSO_L4_CSUM) { 1547 u8 tmp = ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT; 1548 1549 first_bd->data.bd_flags.bitfields |= 1 << tmp; 1550 } 1551 hlen = qede_get_skb_hlen(skb, true); 1552 } else { 1553 first_bd->data.bd_flags.bitfields |= 1554 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT; 1555 hlen = qede_get_skb_hlen(skb, false); 1556 } 1557 1558 /* @@@TBD - if will not be removed need to check */ 1559 third_bd->data.bitfields |= 1560 cpu_to_le16(1 << ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT); 1561 1562 /* Make life easier for FW guys who can't deal with header and 1563 * data on same BD. If we need to split, use the second bd... 1564 */ 1565 if (unlikely(skb_headlen(skb) > hlen)) { 1566 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1567 "TSO split header size is %d (%x:%x)\n", 1568 first_bd->nbytes, first_bd->addr.hi, 1569 first_bd->addr.lo); 1570 1571 mapping = HILO_U64(le32_to_cpu(first_bd->addr.hi), 1572 le32_to_cpu(first_bd->addr.lo)) + 1573 hlen; 1574 1575 BD_SET_UNMAP_ADDR_LEN(tx_data_bd, mapping, 1576 le16_to_cpu(first_bd->nbytes) - 1577 hlen); 1578 1579 /* this marks the BD as one that has no 1580 * individual mapping 1581 */ 1582 txq->sw_tx_ring.skbs[idx].flags |= QEDE_TSO_SPLIT_BD; 1583 1584 first_bd->nbytes = cpu_to_le16(hlen); 1585 1586 tx_data_bd = (struct eth_tx_bd *)third_bd; 1587 data_split = true; 1588 } 1589 } else { 1590 first_bd->data.bitfields |= 1591 (skb->len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) << 1592 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT; 1593 } 1594 1595 /* Handle fragmented skb */ 1596 /* special handle for frags inside 2nd and 3rd bds.. */ 1597 while (tx_data_bd && frag_idx < skb_shinfo(skb)->nr_frags) { 1598 rc = map_frag_to_bd(txq, 1599 &skb_shinfo(skb)->frags[frag_idx], 1600 tx_data_bd); 1601 if (rc) { 1602 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split); 1603 qede_update_tx_producer(txq); 1604 return NETDEV_TX_OK; 1605 } 1606 1607 if (tx_data_bd == (struct eth_tx_bd *)second_bd) 1608 tx_data_bd = (struct eth_tx_bd *)third_bd; 1609 else 1610 tx_data_bd = NULL; 1611 1612 frag_idx++; 1613 } 1614 1615 /* map last frags into 4th, 5th .... */ 1616 for (; frag_idx < skb_shinfo(skb)->nr_frags; frag_idx++, nbd++) { 1617 tx_data_bd = (struct eth_tx_bd *) 1618 qed_chain_produce(&txq->tx_pbl); 1619 1620 memset(tx_data_bd, 0, sizeof(*tx_data_bd)); 1621 1622 rc = map_frag_to_bd(txq, 1623 &skb_shinfo(skb)->frags[frag_idx], 1624 tx_data_bd); 1625 if (rc) { 1626 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split); 1627 qede_update_tx_producer(txq); 1628 return NETDEV_TX_OK; 1629 } 1630 } 1631 1632 /* update the first BD with the actual num BDs */ 1633 first_bd->data.nbds = nbd; 1634 1635 netdev_tx_sent_queue(netdev_txq, skb->len); 1636 1637 skb_tx_timestamp(skb); 1638 1639 /* Advance packet producer only before sending the packet since mapping 1640 * of pages may fail. 1641 */ 1642 txq->sw_tx_prod++; 1643 1644 /* 'next page' entries are counted in the producer value */ 1645 txq->tx_db.data.bd_prod = 1646 cpu_to_le16(qed_chain_get_prod_idx(&txq->tx_pbl)); 1647 1648 if (!skb->xmit_more || netif_xmit_stopped(netdev_txq)) 1649 qede_update_tx_producer(txq); 1650 1651 if (unlikely(qed_chain_get_elem_left(&txq->tx_pbl) 1652 < (MAX_SKB_FRAGS + 1))) { 1653 if (skb->xmit_more) 1654 qede_update_tx_producer(txq); 1655 1656 netif_tx_stop_queue(netdev_txq); 1657 txq->stopped_cnt++; 1658 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1659 "Stop queue was called\n"); 1660 /* paired memory barrier is in qede_tx_int(), we have to keep 1661 * ordering of set_bit() in netif_tx_stop_queue() and read of 1662 * fp->bd_tx_cons 1663 */ 1664 smp_mb(); 1665 1666 if ((qed_chain_get_elem_left(&txq->tx_pbl) >= 1667 (MAX_SKB_FRAGS + 1)) && 1668 (edev->state == QEDE_STATE_OPEN)) { 1669 netif_tx_wake_queue(netdev_txq); 1670 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1671 "Wake queue was called\n"); 1672 } 1673 } 1674 1675 return NETDEV_TX_OK; 1676 } 1677 1678 /* 8B udp header + 8B base tunnel header + 32B option length */ 1679 #define QEDE_MAX_TUN_HDR_LEN 48 1680 1681 netdev_features_t qede_features_check(struct sk_buff *skb, 1682 struct net_device *dev, 1683 netdev_features_t features) 1684 { 1685 if (skb->encapsulation) { 1686 u8 l4_proto = 0; 1687 1688 switch (vlan_get_protocol(skb)) { 1689 case htons(ETH_P_IP): 1690 l4_proto = ip_hdr(skb)->protocol; 1691 break; 1692 case htons(ETH_P_IPV6): 1693 l4_proto = ipv6_hdr(skb)->nexthdr; 1694 break; 1695 default: 1696 return features; 1697 } 1698 1699 /* Disable offloads for geneve tunnels, as HW can't parse 1700 * the geneve header which has option length greater than 32b 1701 * and disable offloads for the ports which are not offloaded. 1702 */ 1703 if (l4_proto == IPPROTO_UDP) { 1704 struct qede_dev *edev = netdev_priv(dev); 1705 u16 hdrlen, vxln_port, gnv_port; 1706 1707 hdrlen = QEDE_MAX_TUN_HDR_LEN; 1708 vxln_port = edev->vxlan_dst_port; 1709 gnv_port = edev->geneve_dst_port; 1710 1711 if ((skb_inner_mac_header(skb) - 1712 skb_transport_header(skb)) > hdrlen || 1713 (ntohs(udp_hdr(skb)->dest) != vxln_port && 1714 ntohs(udp_hdr(skb)->dest) != gnv_port)) 1715 return features & ~(NETIF_F_CSUM_MASK | 1716 NETIF_F_GSO_MASK); 1717 } 1718 } 1719 1720 return features; 1721 } 1722