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 /* 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 struct eth_tx_1st_bd *first_bd; 337 u16 idx = txq->sw_tx_prod; 338 u16 val; 339 340 if (!qed_chain_get_elem_left(&txq->tx_pbl)) { 341 txq->stopped_cnt++; 342 return -ENOMEM; 343 } 344 345 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl); 346 347 memset(first_bd, 0, sizeof(*first_bd)); 348 first_bd->data.bd_flags.bitfields = 349 BIT(ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT); 350 351 val = (length & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) << 352 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT; 353 354 first_bd->data.bitfields |= cpu_to_le16(val); 355 first_bd->data.nbds = 1; 356 357 /* We can safely ignore the offset, as it's 0 for XDP */ 358 BD_SET_UNMAP_ADDR_LEN(first_bd, metadata->mapping + padding, length); 359 360 /* Synchronize the buffer back to device, as program [probably] 361 * has changed it. 362 */ 363 dma_sync_single_for_device(&edev->pdev->dev, 364 metadata->mapping + padding, 365 length, PCI_DMA_TODEVICE); 366 367 txq->sw_tx_ring.xdp[idx].page = metadata->data; 368 txq->sw_tx_ring.xdp[idx].mapping = metadata->mapping; 369 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers; 370 371 /* Mark the fastpath for future XDP doorbell */ 372 fp->xdp_xmit = 1; 373 374 return 0; 375 } 376 377 int qede_txq_has_work(struct qede_tx_queue *txq) 378 { 379 u16 hw_bd_cons; 380 381 /* Tell compiler that consumer and producer can change */ 382 barrier(); 383 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 384 if (qed_chain_get_cons_idx(&txq->tx_pbl) == hw_bd_cons + 1) 385 return 0; 386 387 return hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl); 388 } 389 390 static void qede_xdp_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq) 391 { 392 u16 hw_bd_cons, idx; 393 394 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 395 barrier(); 396 397 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) { 398 qed_chain_consume(&txq->tx_pbl); 399 idx = txq->sw_tx_cons; 400 401 dma_unmap_page(&edev->pdev->dev, 402 txq->sw_tx_ring.xdp[idx].mapping, 403 PAGE_SIZE, DMA_BIDIRECTIONAL); 404 __free_page(txq->sw_tx_ring.xdp[idx].page); 405 406 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers; 407 txq->xmit_pkts++; 408 } 409 } 410 411 static int qede_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq) 412 { 413 struct netdev_queue *netdev_txq; 414 u16 hw_bd_cons; 415 unsigned int pkts_compl = 0, bytes_compl = 0; 416 int rc; 417 418 netdev_txq = netdev_get_tx_queue(edev->ndev, txq->index); 419 420 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr); 421 barrier(); 422 423 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) { 424 int len = 0; 425 426 rc = qede_free_tx_pkt(edev, txq, &len); 427 if (rc) { 428 DP_NOTICE(edev, "hw_bd_cons = %d, chain_cons=%d\n", 429 hw_bd_cons, 430 qed_chain_get_cons_idx(&txq->tx_pbl)); 431 break; 432 } 433 434 bytes_compl += len; 435 pkts_compl++; 436 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers; 437 txq->xmit_pkts++; 438 } 439 440 netdev_tx_completed_queue(netdev_txq, pkts_compl, bytes_compl); 441 442 /* Need to make the tx_bd_cons update visible to start_xmit() 443 * before checking for netif_tx_queue_stopped(). Without the 444 * memory barrier, there is a small possibility that 445 * start_xmit() will miss it and cause the queue to be stopped 446 * forever. 447 * On the other hand we need an rmb() here to ensure the proper 448 * ordering of bit testing in the following 449 * netif_tx_queue_stopped(txq) call. 450 */ 451 smp_mb(); 452 453 if (unlikely(netif_tx_queue_stopped(netdev_txq))) { 454 /* Taking tx_lock is needed to prevent reenabling the queue 455 * while it's empty. This could have happen if rx_action() gets 456 * suspended in qede_tx_int() after the condition before 457 * netif_tx_wake_queue(), while tx_action (qede_start_xmit()): 458 * 459 * stops the queue->sees fresh tx_bd_cons->releases the queue-> 460 * sends some packets consuming the whole queue again-> 461 * stops the queue 462 */ 463 464 __netif_tx_lock(netdev_txq, smp_processor_id()); 465 466 if ((netif_tx_queue_stopped(netdev_txq)) && 467 (edev->state == QEDE_STATE_OPEN) && 468 (qed_chain_get_elem_left(&txq->tx_pbl) 469 >= (MAX_SKB_FRAGS + 1))) { 470 netif_tx_wake_queue(netdev_txq); 471 DP_VERBOSE(edev, NETIF_MSG_TX_DONE, 472 "Wake queue was called\n"); 473 } 474 475 __netif_tx_unlock(netdev_txq); 476 } 477 478 return 0; 479 } 480 481 bool qede_has_rx_work(struct qede_rx_queue *rxq) 482 { 483 u16 hw_comp_cons, sw_comp_cons; 484 485 /* Tell compiler that status block fields can change */ 486 barrier(); 487 488 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr); 489 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 490 491 return hw_comp_cons != sw_comp_cons; 492 } 493 494 static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq) 495 { 496 qed_chain_consume(&rxq->rx_bd_ring); 497 rxq->sw_rx_cons++; 498 } 499 500 /* This function reuses the buffer(from an offset) from 501 * consumer index to producer index in the bd ring 502 */ 503 static inline void qede_reuse_page(struct qede_rx_queue *rxq, 504 struct sw_rx_data *curr_cons) 505 { 506 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring); 507 struct sw_rx_data *curr_prod; 508 dma_addr_t new_mapping; 509 510 curr_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX]; 511 *curr_prod = *curr_cons; 512 513 new_mapping = curr_prod->mapping + curr_prod->page_offset; 514 515 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping)); 516 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping) + 517 rxq->rx_headroom); 518 519 rxq->sw_rx_prod++; 520 curr_cons->data = NULL; 521 } 522 523 /* In case of allocation failures reuse buffers 524 * from consumer index to produce buffers for firmware 525 */ 526 void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count) 527 { 528 struct sw_rx_data *curr_cons; 529 530 for (; count > 0; count--) { 531 curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX]; 532 qede_reuse_page(rxq, curr_cons); 533 qede_rx_bd_ring_consume(rxq); 534 } 535 } 536 537 static inline int qede_realloc_rx_buffer(struct qede_rx_queue *rxq, 538 struct sw_rx_data *curr_cons) 539 { 540 /* Move to the next segment in the page */ 541 curr_cons->page_offset += rxq->rx_buf_seg_size; 542 543 if (curr_cons->page_offset == PAGE_SIZE) { 544 if (unlikely(qede_alloc_rx_buffer(rxq, true))) { 545 /* Since we failed to allocate new buffer 546 * current buffer can be used again. 547 */ 548 curr_cons->page_offset -= rxq->rx_buf_seg_size; 549 550 return -ENOMEM; 551 } 552 553 dma_unmap_page(rxq->dev, curr_cons->mapping, 554 PAGE_SIZE, rxq->data_direction); 555 } else { 556 /* Increment refcount of the page as we don't want 557 * network stack to take the ownership of the page 558 * which can be recycled multiple times by the driver. 559 */ 560 page_ref_inc(curr_cons->data); 561 qede_reuse_page(rxq, curr_cons); 562 } 563 564 return 0; 565 } 566 567 void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq) 568 { 569 u16 bd_prod = qed_chain_get_prod_idx(&rxq->rx_bd_ring); 570 u16 cqe_prod = qed_chain_get_prod_idx(&rxq->rx_comp_ring); 571 struct eth_rx_prod_data rx_prods = {0}; 572 573 /* Update producers */ 574 rx_prods.bd_prod = cpu_to_le16(bd_prod); 575 rx_prods.cqe_prod = cpu_to_le16(cqe_prod); 576 577 /* Make sure that the BD and SGE data is updated before updating the 578 * producers since FW might read the BD/SGE right after the producer 579 * is updated. 580 */ 581 wmb(); 582 583 internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods), 584 (u32 *)&rx_prods); 585 586 /* mmiowb is needed to synchronize doorbell writes from more than one 587 * processor. It guarantees that the write arrives to the device before 588 * the napi lock is released and another qede_poll is called (possibly 589 * on another CPU). Without this barrier, the next doorbell can bypass 590 * this doorbell. This is applicable to IA64/Altix systems. 591 */ 592 mmiowb(); 593 } 594 595 static void qede_get_rxhash(struct sk_buff *skb, u8 bitfields, __le32 rss_hash) 596 { 597 enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE; 598 enum rss_hash_type htype; 599 u32 hash = 0; 600 601 htype = GET_FIELD(bitfields, ETH_FAST_PATH_RX_REG_CQE_RSS_HASH_TYPE); 602 if (htype) { 603 hash_type = ((htype == RSS_HASH_TYPE_IPV4) || 604 (htype == RSS_HASH_TYPE_IPV6)) ? 605 PKT_HASH_TYPE_L3 : PKT_HASH_TYPE_L4; 606 hash = le32_to_cpu(rss_hash); 607 } 608 skb_set_hash(skb, hash, hash_type); 609 } 610 611 static void qede_set_skb_csum(struct sk_buff *skb, u8 csum_flag) 612 { 613 skb_checksum_none_assert(skb); 614 615 if (csum_flag & QEDE_CSUM_UNNECESSARY) 616 skb->ip_summed = CHECKSUM_UNNECESSARY; 617 618 if (csum_flag & QEDE_TUNN_CSUM_UNNECESSARY) { 619 skb->csum_level = 1; 620 skb->encapsulation = 1; 621 } 622 } 623 624 static inline void qede_skb_receive(struct qede_dev *edev, 625 struct qede_fastpath *fp, 626 struct qede_rx_queue *rxq, 627 struct sk_buff *skb, u16 vlan_tag) 628 { 629 if (vlan_tag) 630 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag); 631 632 napi_gro_receive(&fp->napi, skb); 633 } 634 635 static void qede_set_gro_params(struct qede_dev *edev, 636 struct sk_buff *skb, 637 struct eth_fast_path_rx_tpa_start_cqe *cqe) 638 { 639 u16 parsing_flags = le16_to_cpu(cqe->pars_flags.flags); 640 641 if (((parsing_flags >> PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) & 642 PARSING_AND_ERR_FLAGS_L3TYPE_MASK) == 2) 643 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6; 644 else 645 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4; 646 647 skb_shinfo(skb)->gso_size = __le16_to_cpu(cqe->len_on_first_bd) - 648 cqe->header_len; 649 } 650 651 static int qede_fill_frag_skb(struct qede_dev *edev, 652 struct qede_rx_queue *rxq, 653 u8 tpa_agg_index, u16 len_on_bd) 654 { 655 struct sw_rx_data *current_bd = &rxq->sw_rx_ring[rxq->sw_rx_cons & 656 NUM_RX_BDS_MAX]; 657 struct qede_agg_info *tpa_info = &rxq->tpa_info[tpa_agg_index]; 658 struct sk_buff *skb = tpa_info->skb; 659 660 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START)) 661 goto out; 662 663 /* Add one frag and update the appropriate fields in the skb */ 664 skb_fill_page_desc(skb, tpa_info->frag_id++, 665 current_bd->data, current_bd->page_offset, 666 len_on_bd); 667 668 if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) { 669 /* Incr page ref count to reuse on allocation failure 670 * so that it doesn't get freed while freeing SKB. 671 */ 672 page_ref_inc(current_bd->data); 673 goto out; 674 } 675 676 qed_chain_consume(&rxq->rx_bd_ring); 677 rxq->sw_rx_cons++; 678 679 skb->data_len += len_on_bd; 680 skb->truesize += rxq->rx_buf_seg_size; 681 skb->len += len_on_bd; 682 683 return 0; 684 685 out: 686 tpa_info->state = QEDE_AGG_STATE_ERROR; 687 qede_recycle_rx_bd_ring(rxq, 1); 688 689 return -ENOMEM; 690 } 691 692 static bool qede_tunn_exist(u16 flag) 693 { 694 return !!(flag & (PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK << 695 PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT)); 696 } 697 698 static u8 qede_check_tunn_csum(u16 flag) 699 { 700 u16 csum_flag = 0; 701 u8 tcsum = 0; 702 703 if (flag & (PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK << 704 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT)) 705 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK << 706 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT; 707 708 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK << 709 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) { 710 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK << 711 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT; 712 tcsum = QEDE_TUNN_CSUM_UNNECESSARY; 713 } 714 715 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK << 716 PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT | 717 PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK << 718 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT; 719 720 if (csum_flag & flag) 721 return QEDE_CSUM_ERROR; 722 723 return QEDE_CSUM_UNNECESSARY | tcsum; 724 } 725 726 static void qede_tpa_start(struct qede_dev *edev, 727 struct qede_rx_queue *rxq, 728 struct eth_fast_path_rx_tpa_start_cqe *cqe) 729 { 730 struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index]; 731 struct eth_rx_bd *rx_bd_cons = qed_chain_consume(&rxq->rx_bd_ring); 732 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring); 733 struct sw_rx_data *replace_buf = &tpa_info->buffer; 734 dma_addr_t mapping = tpa_info->buffer_mapping; 735 struct sw_rx_data *sw_rx_data_cons; 736 struct sw_rx_data *sw_rx_data_prod; 737 738 sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX]; 739 sw_rx_data_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX]; 740 741 /* Use pre-allocated replacement buffer - we can't release the agg. 742 * start until its over and we don't want to risk allocation failing 743 * here, so re-allocate when aggregation will be over. 744 */ 745 sw_rx_data_prod->mapping = replace_buf->mapping; 746 747 sw_rx_data_prod->data = replace_buf->data; 748 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(mapping)); 749 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(mapping)); 750 sw_rx_data_prod->page_offset = replace_buf->page_offset; 751 752 rxq->sw_rx_prod++; 753 754 /* move partial skb from cons to pool (don't unmap yet) 755 * save mapping, incase we drop the packet later on. 756 */ 757 tpa_info->buffer = *sw_rx_data_cons; 758 mapping = HILO_U64(le32_to_cpu(rx_bd_cons->addr.hi), 759 le32_to_cpu(rx_bd_cons->addr.lo)); 760 761 tpa_info->buffer_mapping = mapping; 762 rxq->sw_rx_cons++; 763 764 /* set tpa state to start only if we are able to allocate skb 765 * for this aggregation, otherwise mark as error and aggregation will 766 * be dropped 767 */ 768 tpa_info->skb = netdev_alloc_skb(edev->ndev, 769 le16_to_cpu(cqe->len_on_first_bd)); 770 if (unlikely(!tpa_info->skb)) { 771 DP_NOTICE(edev, "Failed to allocate SKB for gro\n"); 772 tpa_info->state = QEDE_AGG_STATE_ERROR; 773 goto cons_buf; 774 } 775 776 /* Start filling in the aggregation info */ 777 skb_put(tpa_info->skb, le16_to_cpu(cqe->len_on_first_bd)); 778 tpa_info->frag_id = 0; 779 tpa_info->state = QEDE_AGG_STATE_START; 780 781 /* Store some information from first CQE */ 782 tpa_info->start_cqe_placement_offset = cqe->placement_offset; 783 tpa_info->start_cqe_bd_len = le16_to_cpu(cqe->len_on_first_bd); 784 if ((le16_to_cpu(cqe->pars_flags.flags) >> 785 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) & 786 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK) 787 tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag); 788 else 789 tpa_info->vlan_tag = 0; 790 791 qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash); 792 793 /* This is needed in order to enable forwarding support */ 794 qede_set_gro_params(edev, tpa_info->skb, cqe); 795 796 cons_buf: /* We still need to handle bd_len_list to consume buffers */ 797 if (likely(cqe->ext_bd_len_list[0])) 798 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 799 le16_to_cpu(cqe->ext_bd_len_list[0])); 800 801 if (unlikely(cqe->ext_bd_len_list[1])) { 802 DP_ERR(edev, 803 "Unlikely - got a TPA aggregation with more than one ext_bd_len_list entry in the TPA start\n"); 804 tpa_info->state = QEDE_AGG_STATE_ERROR; 805 } 806 } 807 808 #ifdef CONFIG_INET 809 static void qede_gro_ip_csum(struct sk_buff *skb) 810 { 811 const struct iphdr *iph = ip_hdr(skb); 812 struct tcphdr *th; 813 814 skb_set_transport_header(skb, sizeof(struct iphdr)); 815 th = tcp_hdr(skb); 816 817 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb), 818 iph->saddr, iph->daddr, 0); 819 820 tcp_gro_complete(skb); 821 } 822 823 static void qede_gro_ipv6_csum(struct sk_buff *skb) 824 { 825 struct ipv6hdr *iph = ipv6_hdr(skb); 826 struct tcphdr *th; 827 828 skb_set_transport_header(skb, sizeof(struct ipv6hdr)); 829 th = tcp_hdr(skb); 830 831 th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb), 832 &iph->saddr, &iph->daddr, 0); 833 tcp_gro_complete(skb); 834 } 835 #endif 836 837 static void qede_gro_receive(struct qede_dev *edev, 838 struct qede_fastpath *fp, 839 struct sk_buff *skb, 840 u16 vlan_tag) 841 { 842 /* FW can send a single MTU sized packet from gro flow 843 * due to aggregation timeout/last segment etc. which 844 * is not expected to be a gro packet. If a skb has zero 845 * frags then simply push it in the stack as non gso skb. 846 */ 847 if (unlikely(!skb->data_len)) { 848 skb_shinfo(skb)->gso_type = 0; 849 skb_shinfo(skb)->gso_size = 0; 850 goto send_skb; 851 } 852 853 #ifdef CONFIG_INET 854 if (skb_shinfo(skb)->gso_size) { 855 skb_reset_network_header(skb); 856 857 switch (skb->protocol) { 858 case htons(ETH_P_IP): 859 qede_gro_ip_csum(skb); 860 break; 861 case htons(ETH_P_IPV6): 862 qede_gro_ipv6_csum(skb); 863 break; 864 default: 865 DP_ERR(edev, 866 "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n", 867 ntohs(skb->protocol)); 868 } 869 } 870 #endif 871 872 send_skb: 873 skb_record_rx_queue(skb, fp->rxq->rxq_id); 874 qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag); 875 } 876 877 static inline void qede_tpa_cont(struct qede_dev *edev, 878 struct qede_rx_queue *rxq, 879 struct eth_fast_path_rx_tpa_cont_cqe *cqe) 880 { 881 int i; 882 883 for (i = 0; cqe->len_list[i]; i++) 884 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 885 le16_to_cpu(cqe->len_list[i])); 886 887 if (unlikely(i > 1)) 888 DP_ERR(edev, 889 "Strange - TPA cont with more than a single len_list entry\n"); 890 } 891 892 static int qede_tpa_end(struct qede_dev *edev, 893 struct qede_fastpath *fp, 894 struct eth_fast_path_rx_tpa_end_cqe *cqe) 895 { 896 struct qede_rx_queue *rxq = fp->rxq; 897 struct qede_agg_info *tpa_info; 898 struct sk_buff *skb; 899 int i; 900 901 tpa_info = &rxq->tpa_info[cqe->tpa_agg_index]; 902 skb = tpa_info->skb; 903 904 for (i = 0; cqe->len_list[i]; i++) 905 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index, 906 le16_to_cpu(cqe->len_list[i])); 907 if (unlikely(i > 1)) 908 DP_ERR(edev, 909 "Strange - TPA emd with more than a single len_list entry\n"); 910 911 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START)) 912 goto err; 913 914 /* Sanity */ 915 if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1)) 916 DP_ERR(edev, 917 "Strange - TPA had %02x BDs, but SKB has only %d frags\n", 918 cqe->num_of_bds, tpa_info->frag_id); 919 if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len))) 920 DP_ERR(edev, 921 "Strange - total packet len [cqe] is %4x but SKB has len %04x\n", 922 le16_to_cpu(cqe->total_packet_len), skb->len); 923 924 memcpy(skb->data, 925 page_address(tpa_info->buffer.data) + 926 tpa_info->start_cqe_placement_offset + 927 tpa_info->buffer.page_offset, tpa_info->start_cqe_bd_len); 928 929 /* Finalize the SKB */ 930 skb->protocol = eth_type_trans(skb, edev->ndev); 931 skb->ip_summed = CHECKSUM_UNNECESSARY; 932 933 /* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count 934 * to skb_shinfo(skb)->gso_segs 935 */ 936 NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs); 937 938 qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag); 939 940 tpa_info->state = QEDE_AGG_STATE_NONE; 941 942 return 1; 943 err: 944 tpa_info->state = QEDE_AGG_STATE_NONE; 945 dev_kfree_skb_any(tpa_info->skb); 946 tpa_info->skb = NULL; 947 return 0; 948 } 949 950 static u8 qede_check_notunn_csum(u16 flag) 951 { 952 u16 csum_flag = 0; 953 u8 csum = 0; 954 955 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK << 956 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) { 957 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK << 958 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT; 959 csum = QEDE_CSUM_UNNECESSARY; 960 } 961 962 csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK << 963 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT; 964 965 if (csum_flag & flag) 966 return QEDE_CSUM_ERROR; 967 968 return csum; 969 } 970 971 static u8 qede_check_csum(u16 flag) 972 { 973 if (!qede_tunn_exist(flag)) 974 return qede_check_notunn_csum(flag); 975 else 976 return qede_check_tunn_csum(flag); 977 } 978 979 static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe, 980 u16 flag) 981 { 982 u8 tun_pars_flg = cqe->tunnel_pars_flags.flags; 983 984 if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK << 985 ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) || 986 (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK << 987 PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT))) 988 return true; 989 990 return false; 991 } 992 993 /* Return true iff packet is to be passed to stack */ 994 static bool qede_rx_xdp(struct qede_dev *edev, 995 struct qede_fastpath *fp, 996 struct qede_rx_queue *rxq, 997 struct bpf_prog *prog, 998 struct sw_rx_data *bd, 999 struct eth_fast_path_rx_reg_cqe *cqe, 1000 u16 *data_offset, u16 *len) 1001 { 1002 struct xdp_buff xdp; 1003 enum xdp_action act; 1004 1005 xdp.data_hard_start = page_address(bd->data); 1006 xdp.data = xdp.data_hard_start + *data_offset; 1007 xdp.data_end = xdp.data + *len; 1008 1009 /* Queues always have a full reset currently, so for the time 1010 * being until there's atomic program replace just mark read 1011 * side for map helpers. 1012 */ 1013 rcu_read_lock(); 1014 act = bpf_prog_run_xdp(prog, &xdp); 1015 rcu_read_unlock(); 1016 1017 /* Recalculate, as XDP might have changed the headers */ 1018 *data_offset = xdp.data - xdp.data_hard_start; 1019 *len = xdp.data_end - xdp.data; 1020 1021 if (act == XDP_PASS) 1022 return true; 1023 1024 /* Count number of packets not to be passed to stack */ 1025 rxq->xdp_no_pass++; 1026 1027 switch (act) { 1028 case XDP_TX: 1029 /* We need the replacement buffer before transmit. */ 1030 if (qede_alloc_rx_buffer(rxq, true)) { 1031 qede_recycle_rx_bd_ring(rxq, 1); 1032 trace_xdp_exception(edev->ndev, prog, act); 1033 return false; 1034 } 1035 1036 /* Now if there's a transmission problem, we'd still have to 1037 * throw current buffer, as replacement was already allocated. 1038 */ 1039 if (qede_xdp_xmit(edev, fp, bd, *data_offset, *len)) { 1040 dma_unmap_page(rxq->dev, bd->mapping, 1041 PAGE_SIZE, DMA_BIDIRECTIONAL); 1042 __free_page(bd->data); 1043 trace_xdp_exception(edev->ndev, prog, act); 1044 } 1045 1046 /* Regardless, we've consumed an Rx BD */ 1047 qede_rx_bd_ring_consume(rxq); 1048 return false; 1049 1050 default: 1051 bpf_warn_invalid_xdp_action(act); 1052 case XDP_ABORTED: 1053 trace_xdp_exception(edev->ndev, prog, act); 1054 case XDP_DROP: 1055 qede_recycle_rx_bd_ring(rxq, cqe->bd_num); 1056 } 1057 1058 return false; 1059 } 1060 1061 static struct sk_buff *qede_rx_allocate_skb(struct qede_dev *edev, 1062 struct qede_rx_queue *rxq, 1063 struct sw_rx_data *bd, u16 len, 1064 u16 pad) 1065 { 1066 unsigned int offset = bd->page_offset + pad; 1067 struct skb_frag_struct *frag; 1068 struct page *page = bd->data; 1069 unsigned int pull_len; 1070 struct sk_buff *skb; 1071 unsigned char *va; 1072 1073 /* Allocate a new SKB with a sufficient large header len */ 1074 skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE); 1075 if (unlikely(!skb)) 1076 return NULL; 1077 1078 /* Copy data into SKB - if it's small, we can simply copy it and 1079 * re-use the already allcoated & mapped memory. 1080 */ 1081 if (len + pad <= edev->rx_copybreak) { 1082 skb_put_data(skb, page_address(page) + offset, len); 1083 qede_reuse_page(rxq, bd); 1084 goto out; 1085 } 1086 1087 frag = &skb_shinfo(skb)->frags[0]; 1088 1089 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, 1090 page, offset, len, rxq->rx_buf_seg_size); 1091 1092 va = skb_frag_address(frag); 1093 pull_len = eth_get_headlen(va, QEDE_RX_HDR_SIZE); 1094 1095 /* Align the pull_len to optimize memcpy */ 1096 memcpy(skb->data, va, ALIGN(pull_len, sizeof(long))); 1097 1098 /* Correct the skb & frag sizes offset after the pull */ 1099 skb_frag_size_sub(frag, pull_len); 1100 frag->page_offset += pull_len; 1101 skb->data_len -= pull_len; 1102 skb->tail += pull_len; 1103 1104 if (unlikely(qede_realloc_rx_buffer(rxq, bd))) { 1105 /* Incr page ref count to reuse on allocation failure so 1106 * that it doesn't get freed while freeing SKB [as its 1107 * already mapped there]. 1108 */ 1109 page_ref_inc(page); 1110 dev_kfree_skb_any(skb); 1111 return NULL; 1112 } 1113 1114 out: 1115 /* We've consumed the first BD and prepared an SKB */ 1116 qede_rx_bd_ring_consume(rxq); 1117 return skb; 1118 } 1119 1120 static int qede_rx_build_jumbo(struct qede_dev *edev, 1121 struct qede_rx_queue *rxq, 1122 struct sk_buff *skb, 1123 struct eth_fast_path_rx_reg_cqe *cqe, 1124 u16 first_bd_len) 1125 { 1126 u16 pkt_len = le16_to_cpu(cqe->pkt_len); 1127 struct sw_rx_data *bd; 1128 u16 bd_cons_idx; 1129 u8 num_frags; 1130 1131 pkt_len -= first_bd_len; 1132 1133 /* We've already used one BD for the SKB. Now take care of the rest */ 1134 for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) { 1135 u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size : 1136 pkt_len; 1137 1138 if (unlikely(!cur_size)) { 1139 DP_ERR(edev, 1140 "Still got %d BDs for mapping jumbo, but length became 0\n", 1141 num_frags); 1142 goto out; 1143 } 1144 1145 /* We need a replacement buffer for each BD */ 1146 if (unlikely(qede_alloc_rx_buffer(rxq, true))) 1147 goto out; 1148 1149 /* Now that we've allocated the replacement buffer, 1150 * we can safely consume the next BD and map it to the SKB. 1151 */ 1152 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX; 1153 bd = &rxq->sw_rx_ring[bd_cons_idx]; 1154 qede_rx_bd_ring_consume(rxq); 1155 1156 dma_unmap_page(rxq->dev, bd->mapping, 1157 PAGE_SIZE, DMA_FROM_DEVICE); 1158 1159 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags++, 1160 bd->data, 0, cur_size); 1161 1162 skb->truesize += PAGE_SIZE; 1163 skb->data_len += cur_size; 1164 skb->len += cur_size; 1165 pkt_len -= cur_size; 1166 } 1167 1168 if (unlikely(pkt_len)) 1169 DP_ERR(edev, 1170 "Mapped all BDs of jumbo, but still have %d bytes\n", 1171 pkt_len); 1172 1173 out: 1174 return num_frags; 1175 } 1176 1177 static int qede_rx_process_tpa_cqe(struct qede_dev *edev, 1178 struct qede_fastpath *fp, 1179 struct qede_rx_queue *rxq, 1180 union eth_rx_cqe *cqe, 1181 enum eth_rx_cqe_type type) 1182 { 1183 switch (type) { 1184 case ETH_RX_CQE_TYPE_TPA_START: 1185 qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start); 1186 return 0; 1187 case ETH_RX_CQE_TYPE_TPA_CONT: 1188 qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont); 1189 return 0; 1190 case ETH_RX_CQE_TYPE_TPA_END: 1191 return qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end); 1192 default: 1193 return 0; 1194 } 1195 } 1196 1197 static int qede_rx_process_cqe(struct qede_dev *edev, 1198 struct qede_fastpath *fp, 1199 struct qede_rx_queue *rxq) 1200 { 1201 struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog); 1202 struct eth_fast_path_rx_reg_cqe *fp_cqe; 1203 u16 len, pad, bd_cons_idx, parse_flag; 1204 enum eth_rx_cqe_type cqe_type; 1205 union eth_rx_cqe *cqe; 1206 struct sw_rx_data *bd; 1207 struct sk_buff *skb; 1208 __le16 flags; 1209 u8 csum_flag; 1210 1211 /* Get the CQE from the completion ring */ 1212 cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring); 1213 cqe_type = cqe->fast_path_regular.type; 1214 1215 /* Process an unlikely slowpath event */ 1216 if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) { 1217 struct eth_slow_path_rx_cqe *sp_cqe; 1218 1219 sp_cqe = (struct eth_slow_path_rx_cqe *)cqe; 1220 edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe); 1221 return 0; 1222 } 1223 1224 /* Handle TPA cqes */ 1225 if (cqe_type != ETH_RX_CQE_TYPE_REGULAR) 1226 return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type); 1227 1228 /* Get the data from the SW ring; Consume it only after it's evident 1229 * we wouldn't recycle it. 1230 */ 1231 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX; 1232 bd = &rxq->sw_rx_ring[bd_cons_idx]; 1233 1234 fp_cqe = &cqe->fast_path_regular; 1235 len = le16_to_cpu(fp_cqe->len_on_first_bd); 1236 pad = fp_cqe->placement_offset + rxq->rx_headroom; 1237 1238 /* Run eBPF program if one is attached */ 1239 if (xdp_prog) 1240 if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe, 1241 &pad, &len)) 1242 return 0; 1243 1244 /* If this is an error packet then drop it */ 1245 flags = cqe->fast_path_regular.pars_flags.flags; 1246 parse_flag = le16_to_cpu(flags); 1247 1248 csum_flag = qede_check_csum(parse_flag); 1249 if (unlikely(csum_flag == QEDE_CSUM_ERROR)) { 1250 if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag)) { 1251 rxq->rx_ip_frags++; 1252 } else { 1253 DP_NOTICE(edev, 1254 "CQE has error, flags = %x, dropping incoming packet\n", 1255 parse_flag); 1256 rxq->rx_hw_errors++; 1257 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num); 1258 return 0; 1259 } 1260 } 1261 1262 /* Basic validation passed; Need to prepare an SKB. This would also 1263 * guarantee to finally consume the first BD upon success. 1264 */ 1265 skb = qede_rx_allocate_skb(edev, rxq, bd, len, pad); 1266 if (!skb) { 1267 rxq->rx_alloc_errors++; 1268 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num); 1269 return 0; 1270 } 1271 1272 /* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed 1273 * by a single cqe. 1274 */ 1275 if (fp_cqe->bd_num > 1) { 1276 u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb, 1277 fp_cqe, len); 1278 1279 if (unlikely(unmapped_frags > 0)) { 1280 qede_recycle_rx_bd_ring(rxq, unmapped_frags); 1281 dev_kfree_skb_any(skb); 1282 return 0; 1283 } 1284 } 1285 1286 /* The SKB contains all the data. Now prepare meta-magic */ 1287 skb->protocol = eth_type_trans(skb, edev->ndev); 1288 qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash); 1289 qede_set_skb_csum(skb, csum_flag); 1290 skb_record_rx_queue(skb, rxq->rxq_id); 1291 qede_ptp_record_rx_ts(edev, cqe, skb); 1292 1293 /* SKB is prepared - pass it to stack */ 1294 qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag)); 1295 1296 return 1; 1297 } 1298 1299 static int qede_rx_int(struct qede_fastpath *fp, int budget) 1300 { 1301 struct qede_rx_queue *rxq = fp->rxq; 1302 struct qede_dev *edev = fp->edev; 1303 int work_done = 0, rcv_pkts = 0; 1304 u16 hw_comp_cons, sw_comp_cons; 1305 1306 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr); 1307 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 1308 1309 /* Memory barrier to prevent the CPU from doing speculative reads of CQE 1310 * / BD in the while-loop before reading hw_comp_cons. If the CQE is 1311 * read before it is written by FW, then FW writes CQE and SB, and then 1312 * the CPU reads the hw_comp_cons, it will use an old CQE. 1313 */ 1314 rmb(); 1315 1316 /* Loop to complete all indicated BDs */ 1317 while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) { 1318 rcv_pkts += qede_rx_process_cqe(edev, fp, rxq); 1319 qed_chain_recycle_consumed(&rxq->rx_comp_ring); 1320 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring); 1321 work_done++; 1322 } 1323 1324 rxq->rcv_pkts += rcv_pkts; 1325 1326 /* Allocate replacement buffers */ 1327 while (rxq->num_rx_buffers - rxq->filled_buffers) 1328 if (qede_alloc_rx_buffer(rxq, false)) 1329 break; 1330 1331 /* Update producers */ 1332 qede_update_rx_prod(edev, rxq); 1333 1334 return work_done; 1335 } 1336 1337 static bool qede_poll_is_more_work(struct qede_fastpath *fp) 1338 { 1339 qed_sb_update_sb_idx(fp->sb_info); 1340 1341 /* *_has_*_work() reads the status block, thus we need to ensure that 1342 * status block indices have been actually read (qed_sb_update_sb_idx) 1343 * prior to this check (*_has_*_work) so that we won't write the 1344 * "newer" value of the status block to HW (if there was a DMA right 1345 * after qede_has_rx_work and if there is no rmb, the memory reading 1346 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb). 1347 * In this case there will never be another interrupt until there is 1348 * another update of the status block, while there is still unhandled 1349 * work. 1350 */ 1351 rmb(); 1352 1353 if (likely(fp->type & QEDE_FASTPATH_RX)) 1354 if (qede_has_rx_work(fp->rxq)) 1355 return true; 1356 1357 if (fp->type & QEDE_FASTPATH_XDP) 1358 if (qede_txq_has_work(fp->xdp_tx)) 1359 return true; 1360 1361 if (likely(fp->type & QEDE_FASTPATH_TX)) 1362 if (qede_txq_has_work(fp->txq)) 1363 return true; 1364 1365 return false; 1366 } 1367 1368 /********************* 1369 * NDO & API related * 1370 *********************/ 1371 int qede_poll(struct napi_struct *napi, int budget) 1372 { 1373 struct qede_fastpath *fp = container_of(napi, struct qede_fastpath, 1374 napi); 1375 struct qede_dev *edev = fp->edev; 1376 int rx_work_done = 0; 1377 1378 if (likely(fp->type & QEDE_FASTPATH_TX) && qede_txq_has_work(fp->txq)) 1379 qede_tx_int(edev, fp->txq); 1380 1381 if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx)) 1382 qede_xdp_tx_int(edev, fp->xdp_tx); 1383 1384 rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) && 1385 qede_has_rx_work(fp->rxq)) ? 1386 qede_rx_int(fp, budget) : 0; 1387 if (rx_work_done < budget) { 1388 if (!qede_poll_is_more_work(fp)) { 1389 napi_complete_done(napi, rx_work_done); 1390 1391 /* Update and reenable interrupts */ 1392 qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1); 1393 } else { 1394 rx_work_done = budget; 1395 } 1396 } 1397 1398 if (fp->xdp_xmit) { 1399 u16 xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl); 1400 1401 fp->xdp_xmit = 0; 1402 fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod); 1403 qede_update_tx_producer(fp->xdp_tx); 1404 } 1405 1406 return rx_work_done; 1407 } 1408 1409 irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie) 1410 { 1411 struct qede_fastpath *fp = fp_cookie; 1412 1413 qed_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0 /*do not update*/); 1414 1415 napi_schedule_irqoff(&fp->napi); 1416 return IRQ_HANDLED; 1417 } 1418 1419 /* Main transmit function */ 1420 netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev) 1421 { 1422 struct qede_dev *edev = netdev_priv(ndev); 1423 struct netdev_queue *netdev_txq; 1424 struct qede_tx_queue *txq; 1425 struct eth_tx_1st_bd *first_bd; 1426 struct eth_tx_2nd_bd *second_bd = NULL; 1427 struct eth_tx_3rd_bd *third_bd = NULL; 1428 struct eth_tx_bd *tx_data_bd = NULL; 1429 u16 txq_index, val = 0; 1430 u8 nbd = 0; 1431 dma_addr_t mapping; 1432 int rc, frag_idx = 0, ipv6_ext = 0; 1433 u8 xmit_type; 1434 u16 idx; 1435 u16 hlen; 1436 bool data_split = false; 1437 1438 /* Get tx-queue context and netdev index */ 1439 txq_index = skb_get_queue_mapping(skb); 1440 WARN_ON(txq_index >= QEDE_TSS_COUNT(edev)); 1441 txq = edev->fp_array[edev->fp_num_rx + txq_index].txq; 1442 netdev_txq = netdev_get_tx_queue(ndev, txq_index); 1443 1444 WARN_ON(qed_chain_get_elem_left(&txq->tx_pbl) < (MAX_SKB_FRAGS + 1)); 1445 1446 xmit_type = qede_xmit_type(skb, &ipv6_ext); 1447 1448 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET) 1449 if (qede_pkt_req_lin(skb, xmit_type)) { 1450 if (skb_linearize(skb)) { 1451 DP_NOTICE(edev, 1452 "SKB linearization failed - silently dropping this SKB\n"); 1453 dev_kfree_skb_any(skb); 1454 return NETDEV_TX_OK; 1455 } 1456 } 1457 #endif 1458 1459 /* Fill the entry in the SW ring and the BDs in the FW ring */ 1460 idx = txq->sw_tx_prod; 1461 txq->sw_tx_ring.skbs[idx].skb = skb; 1462 first_bd = (struct eth_tx_1st_bd *) 1463 qed_chain_produce(&txq->tx_pbl); 1464 memset(first_bd, 0, sizeof(*first_bd)); 1465 first_bd->data.bd_flags.bitfields = 1466 1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT; 1467 1468 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)) 1469 qede_ptp_tx_ts(edev, skb); 1470 1471 /* Map skb linear data for DMA and set in the first BD */ 1472 mapping = dma_map_single(txq->dev, skb->data, 1473 skb_headlen(skb), DMA_TO_DEVICE); 1474 if (unlikely(dma_mapping_error(txq->dev, mapping))) { 1475 DP_NOTICE(edev, "SKB mapping failed\n"); 1476 qede_free_failed_tx_pkt(txq, first_bd, 0, false); 1477 qede_update_tx_producer(txq); 1478 return NETDEV_TX_OK; 1479 } 1480 nbd++; 1481 BD_SET_UNMAP_ADDR_LEN(first_bd, mapping, skb_headlen(skb)); 1482 1483 /* In case there is IPv6 with extension headers or LSO we need 2nd and 1484 * 3rd BDs. 1485 */ 1486 if (unlikely((xmit_type & XMIT_LSO) | ipv6_ext)) { 1487 second_bd = (struct eth_tx_2nd_bd *) 1488 qed_chain_produce(&txq->tx_pbl); 1489 memset(second_bd, 0, sizeof(*second_bd)); 1490 1491 nbd++; 1492 third_bd = (struct eth_tx_3rd_bd *) 1493 qed_chain_produce(&txq->tx_pbl); 1494 memset(third_bd, 0, sizeof(*third_bd)); 1495 1496 nbd++; 1497 /* We need to fill in additional data in second_bd... */ 1498 tx_data_bd = (struct eth_tx_bd *)second_bd; 1499 } 1500 1501 if (skb_vlan_tag_present(skb)) { 1502 first_bd->data.vlan = cpu_to_le16(skb_vlan_tag_get(skb)); 1503 first_bd->data.bd_flags.bitfields |= 1504 1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT; 1505 } 1506 1507 /* Fill the parsing flags & params according to the requested offload */ 1508 if (xmit_type & XMIT_L4_CSUM) { 1509 /* We don't re-calculate IP checksum as it is already done by 1510 * the upper stack 1511 */ 1512 first_bd->data.bd_flags.bitfields |= 1513 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT; 1514 1515 if (xmit_type & XMIT_ENC) { 1516 first_bd->data.bd_flags.bitfields |= 1517 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT; 1518 1519 val |= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT); 1520 } 1521 1522 /* Legacy FW had flipped behavior in regard to this bit - 1523 * I.e., needed to set to prevent FW from touching encapsulated 1524 * packets when it didn't need to. 1525 */ 1526 if (unlikely(txq->is_legacy)) 1527 val ^= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT); 1528 1529 /* If the packet is IPv6 with extension header, indicate that 1530 * to FW and pass few params, since the device cracker doesn't 1531 * support parsing IPv6 with extension header/s. 1532 */ 1533 if (unlikely(ipv6_ext)) 1534 qede_set_params_for_ipv6_ext(skb, second_bd, third_bd); 1535 } 1536 1537 if (xmit_type & XMIT_LSO) { 1538 first_bd->data.bd_flags.bitfields |= 1539 (1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT); 1540 third_bd->data.lso_mss = 1541 cpu_to_le16(skb_shinfo(skb)->gso_size); 1542 1543 if (unlikely(xmit_type & XMIT_ENC)) { 1544 first_bd->data.bd_flags.bitfields |= 1545 1 << ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT; 1546 1547 if (xmit_type & XMIT_ENC_GSO_L4_CSUM) { 1548 u8 tmp = ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT; 1549 1550 first_bd->data.bd_flags.bitfields |= 1 << tmp; 1551 } 1552 hlen = qede_get_skb_hlen(skb, true); 1553 } else { 1554 first_bd->data.bd_flags.bitfields |= 1555 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT; 1556 hlen = qede_get_skb_hlen(skb, false); 1557 } 1558 1559 /* @@@TBD - if will not be removed need to check */ 1560 third_bd->data.bitfields |= 1561 cpu_to_le16(1 << ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT); 1562 1563 /* Make life easier for FW guys who can't deal with header and 1564 * data on same BD. If we need to split, use the second bd... 1565 */ 1566 if (unlikely(skb_headlen(skb) > hlen)) { 1567 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1568 "TSO split header size is %d (%x:%x)\n", 1569 first_bd->nbytes, first_bd->addr.hi, 1570 first_bd->addr.lo); 1571 1572 mapping = HILO_U64(le32_to_cpu(first_bd->addr.hi), 1573 le32_to_cpu(first_bd->addr.lo)) + 1574 hlen; 1575 1576 BD_SET_UNMAP_ADDR_LEN(tx_data_bd, mapping, 1577 le16_to_cpu(first_bd->nbytes) - 1578 hlen); 1579 1580 /* this marks the BD as one that has no 1581 * individual mapping 1582 */ 1583 txq->sw_tx_ring.skbs[idx].flags |= QEDE_TSO_SPLIT_BD; 1584 1585 first_bd->nbytes = cpu_to_le16(hlen); 1586 1587 tx_data_bd = (struct eth_tx_bd *)third_bd; 1588 data_split = true; 1589 } 1590 } else { 1591 val |= ((skb->len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) << 1592 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT); 1593 } 1594 1595 first_bd->data.bitfields = cpu_to_le16(val); 1596 1597 /* Handle fragmented skb */ 1598 /* special handle for frags inside 2nd and 3rd bds.. */ 1599 while (tx_data_bd && frag_idx < skb_shinfo(skb)->nr_frags) { 1600 rc = map_frag_to_bd(txq, 1601 &skb_shinfo(skb)->frags[frag_idx], 1602 tx_data_bd); 1603 if (rc) { 1604 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split); 1605 qede_update_tx_producer(txq); 1606 return NETDEV_TX_OK; 1607 } 1608 1609 if (tx_data_bd == (struct eth_tx_bd *)second_bd) 1610 tx_data_bd = (struct eth_tx_bd *)third_bd; 1611 else 1612 tx_data_bd = NULL; 1613 1614 frag_idx++; 1615 } 1616 1617 /* map last frags into 4th, 5th .... */ 1618 for (; frag_idx < skb_shinfo(skb)->nr_frags; frag_idx++, nbd++) { 1619 tx_data_bd = (struct eth_tx_bd *) 1620 qed_chain_produce(&txq->tx_pbl); 1621 1622 memset(tx_data_bd, 0, sizeof(*tx_data_bd)); 1623 1624 rc = map_frag_to_bd(txq, 1625 &skb_shinfo(skb)->frags[frag_idx], 1626 tx_data_bd); 1627 if (rc) { 1628 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split); 1629 qede_update_tx_producer(txq); 1630 return NETDEV_TX_OK; 1631 } 1632 } 1633 1634 /* update the first BD with the actual num BDs */ 1635 first_bd->data.nbds = nbd; 1636 1637 netdev_tx_sent_queue(netdev_txq, skb->len); 1638 1639 skb_tx_timestamp(skb); 1640 1641 /* Advance packet producer only before sending the packet since mapping 1642 * of pages may fail. 1643 */ 1644 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers; 1645 1646 /* 'next page' entries are counted in the producer value */ 1647 txq->tx_db.data.bd_prod = 1648 cpu_to_le16(qed_chain_get_prod_idx(&txq->tx_pbl)); 1649 1650 if (!skb->xmit_more || netif_xmit_stopped(netdev_txq)) 1651 qede_update_tx_producer(txq); 1652 1653 if (unlikely(qed_chain_get_elem_left(&txq->tx_pbl) 1654 < (MAX_SKB_FRAGS + 1))) { 1655 if (skb->xmit_more) 1656 qede_update_tx_producer(txq); 1657 1658 netif_tx_stop_queue(netdev_txq); 1659 txq->stopped_cnt++; 1660 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1661 "Stop queue was called\n"); 1662 /* paired memory barrier is in qede_tx_int(), we have to keep 1663 * ordering of set_bit() in netif_tx_stop_queue() and read of 1664 * fp->bd_tx_cons 1665 */ 1666 smp_mb(); 1667 1668 if ((qed_chain_get_elem_left(&txq->tx_pbl) >= 1669 (MAX_SKB_FRAGS + 1)) && 1670 (edev->state == QEDE_STATE_OPEN)) { 1671 netif_tx_wake_queue(netdev_txq); 1672 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED, 1673 "Wake queue was called\n"); 1674 } 1675 } 1676 1677 return NETDEV_TX_OK; 1678 } 1679 1680 /* 8B udp header + 8B base tunnel header + 32B option length */ 1681 #define QEDE_MAX_TUN_HDR_LEN 48 1682 1683 netdev_features_t qede_features_check(struct sk_buff *skb, 1684 struct net_device *dev, 1685 netdev_features_t features) 1686 { 1687 if (skb->encapsulation) { 1688 u8 l4_proto = 0; 1689 1690 switch (vlan_get_protocol(skb)) { 1691 case htons(ETH_P_IP): 1692 l4_proto = ip_hdr(skb)->protocol; 1693 break; 1694 case htons(ETH_P_IPV6): 1695 l4_proto = ipv6_hdr(skb)->nexthdr; 1696 break; 1697 default: 1698 return features; 1699 } 1700 1701 /* Disable offloads for geneve tunnels, as HW can't parse 1702 * the geneve header which has option length greater than 32b 1703 * and disable offloads for the ports which are not offloaded. 1704 */ 1705 if (l4_proto == IPPROTO_UDP) { 1706 struct qede_dev *edev = netdev_priv(dev); 1707 u16 hdrlen, vxln_port, gnv_port; 1708 1709 hdrlen = QEDE_MAX_TUN_HDR_LEN; 1710 vxln_port = edev->vxlan_dst_port; 1711 gnv_port = edev->geneve_dst_port; 1712 1713 if ((skb_inner_mac_header(skb) - 1714 skb_transport_header(skb)) > hdrlen || 1715 (ntohs(udp_hdr(skb)->dest) != vxln_port && 1716 ntohs(udp_hdr(skb)->dest) != gnv_port)) 1717 return features & ~(NETIF_F_CSUM_MASK | 1718 NETIF_F_GSO_MASK); 1719 } 1720 } 1721 1722 return features; 1723 } 1724