1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) 2 /* Copyright (C) 2015-2019 Netronome Systems, Inc. */ 3 4 #include <linux/bpf_trace.h> 5 #include <linux/netdevice.h> 6 #include <linux/bitfield.h> 7 #include <net/xfrm.h> 8 9 #include "../nfp_app.h" 10 #include "../nfp_net.h" 11 #include "../nfp_net_dp.h" 12 #include "../nfp_net_xsk.h" 13 #include "../crypto/crypto.h" 14 #include "../crypto/fw.h" 15 #include "nfd3.h" 16 17 /* Transmit processing 18 * 19 * One queue controller peripheral queue is used for transmit. The 20 * driver en-queues packets for transmit by advancing the write 21 * pointer. The device indicates that packets have transmitted by 22 * advancing the read pointer. The driver maintains a local copy of 23 * the read and write pointer in @struct nfp_net_tx_ring. The driver 24 * keeps @wr_p in sync with the queue controller write pointer and can 25 * determine how many packets have been transmitted by comparing its 26 * copy of the read pointer @rd_p with the read pointer maintained by 27 * the queue controller peripheral. 28 */ 29 30 /* Wrappers for deciding when to stop and restart TX queues */ 31 static int nfp_nfd3_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring) 32 { 33 return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4); 34 } 35 36 static int nfp_nfd3_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring) 37 { 38 return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1); 39 } 40 41 /** 42 * nfp_nfd3_tx_ring_stop() - stop tx ring 43 * @nd_q: netdev queue 44 * @tx_ring: driver tx queue structure 45 * 46 * Safely stop TX ring. Remember that while we are running .start_xmit() 47 * someone else may be cleaning the TX ring completions so we need to be 48 * extra careful here. 49 */ 50 static void 51 nfp_nfd3_tx_ring_stop(struct netdev_queue *nd_q, 52 struct nfp_net_tx_ring *tx_ring) 53 { 54 netif_tx_stop_queue(nd_q); 55 56 /* We can race with the TX completion out of NAPI so recheck */ 57 smp_mb(); 58 if (unlikely(nfp_nfd3_tx_ring_should_wake(tx_ring))) 59 netif_tx_start_queue(nd_q); 60 } 61 62 /** 63 * nfp_nfd3_tx_tso() - Set up Tx descriptor for LSO 64 * @r_vec: per-ring structure 65 * @txbuf: Pointer to driver soft TX descriptor 66 * @txd: Pointer to HW TX descriptor 67 * @skb: Pointer to SKB 68 * @md_bytes: Prepend length 69 * 70 * Set up Tx descriptor for LSO, do nothing for non-LSO skbs. 71 * Return error on packet header greater than maximum supported LSO header size. 72 */ 73 static void 74 nfp_nfd3_tx_tso(struct nfp_net_r_vector *r_vec, struct nfp_nfd3_tx_buf *txbuf, 75 struct nfp_nfd3_tx_desc *txd, struct sk_buff *skb, u32 md_bytes) 76 { 77 u32 l3_offset, l4_offset, hdrlen; 78 u16 mss; 79 80 if (!skb_is_gso(skb)) 81 return; 82 83 if (!skb->encapsulation) { 84 l3_offset = skb_network_offset(skb); 85 l4_offset = skb_transport_offset(skb); 86 hdrlen = skb_tcp_all_headers(skb); 87 } else { 88 l3_offset = skb_inner_network_offset(skb); 89 l4_offset = skb_inner_transport_offset(skb); 90 hdrlen = skb_inner_tcp_all_headers(skb); 91 } 92 93 txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs; 94 txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1); 95 96 mss = skb_shinfo(skb)->gso_size & NFD3_DESC_TX_MSS_MASK; 97 txd->l3_offset = l3_offset - md_bytes; 98 txd->l4_offset = l4_offset - md_bytes; 99 txd->lso_hdrlen = hdrlen - md_bytes; 100 txd->mss = cpu_to_le16(mss); 101 txd->flags |= NFD3_DESC_TX_LSO; 102 103 u64_stats_update_begin(&r_vec->tx_sync); 104 r_vec->tx_lso++; 105 u64_stats_update_end(&r_vec->tx_sync); 106 } 107 108 /** 109 * nfp_nfd3_tx_csum() - Set TX CSUM offload flags in TX descriptor 110 * @dp: NFP Net data path struct 111 * @r_vec: per-ring structure 112 * @txbuf: Pointer to driver soft TX descriptor 113 * @txd: Pointer to TX descriptor 114 * @skb: Pointer to SKB 115 * 116 * This function sets the TX checksum flags in the TX descriptor based 117 * on the configuration and the protocol of the packet to be transmitted. 118 */ 119 static void 120 nfp_nfd3_tx_csum(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec, 121 struct nfp_nfd3_tx_buf *txbuf, struct nfp_nfd3_tx_desc *txd, 122 struct sk_buff *skb) 123 { 124 struct ipv6hdr *ipv6h; 125 struct iphdr *iph; 126 u8 l4_hdr; 127 128 if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM)) 129 return; 130 131 if (skb->ip_summed != CHECKSUM_PARTIAL) 132 return; 133 134 txd->flags |= NFD3_DESC_TX_CSUM; 135 if (skb->encapsulation) 136 txd->flags |= NFD3_DESC_TX_ENCAP; 137 138 iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb); 139 ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb); 140 141 if (iph->version == 4) { 142 txd->flags |= NFD3_DESC_TX_IP4_CSUM; 143 l4_hdr = iph->protocol; 144 } else if (ipv6h->version == 6) { 145 l4_hdr = ipv6h->nexthdr; 146 } else { 147 nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version); 148 return; 149 } 150 151 switch (l4_hdr) { 152 case IPPROTO_TCP: 153 txd->flags |= NFD3_DESC_TX_TCP_CSUM; 154 break; 155 case IPPROTO_UDP: 156 txd->flags |= NFD3_DESC_TX_UDP_CSUM; 157 break; 158 default: 159 nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr); 160 return; 161 } 162 163 u64_stats_update_begin(&r_vec->tx_sync); 164 if (skb->encapsulation) 165 r_vec->hw_csum_tx_inner += txbuf->pkt_cnt; 166 else 167 r_vec->hw_csum_tx += txbuf->pkt_cnt; 168 u64_stats_update_end(&r_vec->tx_sync); 169 } 170 171 static int nfp_nfd3_prep_tx_meta(struct nfp_net_dp *dp, struct sk_buff *skb, 172 u64 tls_handle, bool *ipsec) 173 { 174 struct metadata_dst *md_dst = skb_metadata_dst(skb); 175 struct nfp_ipsec_offload offload_info; 176 unsigned char *data; 177 bool vlan_insert; 178 u32 meta_id = 0; 179 int md_bytes; 180 181 #ifdef CONFIG_NFP_NET_IPSEC 182 if (xfrm_offload(skb)) 183 *ipsec = nfp_net_ipsec_tx_prep(dp, skb, &offload_info); 184 #endif 185 186 if (unlikely(md_dst && md_dst->type != METADATA_HW_PORT_MUX)) 187 md_dst = NULL; 188 189 vlan_insert = skb_vlan_tag_present(skb) && (dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN_V2); 190 191 if (!(md_dst || tls_handle || vlan_insert || *ipsec)) 192 return 0; 193 194 md_bytes = sizeof(meta_id) + 195 (!!md_dst ? NFP_NET_META_PORTID_SIZE : 0) + 196 (!!tls_handle ? NFP_NET_META_CONN_HANDLE_SIZE : 0) + 197 (vlan_insert ? NFP_NET_META_VLAN_SIZE : 0) + 198 (*ipsec ? NFP_NET_META_IPSEC_FIELD_SIZE : 0); 199 200 if (unlikely(skb_cow_head(skb, md_bytes))) 201 return -ENOMEM; 202 203 data = skb_push(skb, md_bytes) + md_bytes; 204 if (md_dst) { 205 data -= NFP_NET_META_PORTID_SIZE; 206 put_unaligned_be32(md_dst->u.port_info.port_id, data); 207 meta_id = NFP_NET_META_PORTID; 208 } 209 if (tls_handle) { 210 /* conn handle is opaque, we just use u64 to be able to quickly 211 * compare it to zero 212 */ 213 data -= NFP_NET_META_CONN_HANDLE_SIZE; 214 memcpy(data, &tls_handle, sizeof(tls_handle)); 215 meta_id <<= NFP_NET_META_FIELD_SIZE; 216 meta_id |= NFP_NET_META_CONN_HANDLE; 217 } 218 if (vlan_insert) { 219 data -= NFP_NET_META_VLAN_SIZE; 220 /* data type of skb->vlan_proto is __be16 221 * so it fills metadata without calling put_unaligned_be16 222 */ 223 memcpy(data, &skb->vlan_proto, sizeof(skb->vlan_proto)); 224 put_unaligned_be16(skb_vlan_tag_get(skb), data + sizeof(skb->vlan_proto)); 225 meta_id <<= NFP_NET_META_FIELD_SIZE; 226 meta_id |= NFP_NET_META_VLAN; 227 } 228 if (*ipsec) { 229 data -= NFP_NET_META_IPSEC_SIZE; 230 put_unaligned_be32(offload_info.seq_hi, data); 231 data -= NFP_NET_META_IPSEC_SIZE; 232 put_unaligned_be32(offload_info.seq_low, data); 233 data -= NFP_NET_META_IPSEC_SIZE; 234 put_unaligned_be32(offload_info.handle - 1, data); 235 meta_id <<= NFP_NET_META_IPSEC_FIELD_SIZE; 236 meta_id |= NFP_NET_META_IPSEC << 8 | NFP_NET_META_IPSEC << 4 | NFP_NET_META_IPSEC; 237 } 238 239 data -= sizeof(meta_id); 240 put_unaligned_be32(meta_id, data); 241 242 return md_bytes; 243 } 244 245 /** 246 * nfp_nfd3_tx() - Main transmit entry point 247 * @skb: SKB to transmit 248 * @netdev: netdev structure 249 * 250 * Return: NETDEV_TX_OK on success. 251 */ 252 netdev_tx_t nfp_nfd3_tx(struct sk_buff *skb, struct net_device *netdev) 253 { 254 struct nfp_net *nn = netdev_priv(netdev); 255 int f, nr_frags, wr_idx, md_bytes; 256 struct nfp_net_tx_ring *tx_ring; 257 struct nfp_net_r_vector *r_vec; 258 struct nfp_nfd3_tx_buf *txbuf; 259 struct nfp_nfd3_tx_desc *txd; 260 struct netdev_queue *nd_q; 261 const skb_frag_t *frag; 262 struct nfp_net_dp *dp; 263 dma_addr_t dma_addr; 264 unsigned int fsize; 265 u64 tls_handle = 0; 266 bool ipsec = false; 267 u16 qidx; 268 269 dp = &nn->dp; 270 qidx = skb_get_queue_mapping(skb); 271 tx_ring = &dp->tx_rings[qidx]; 272 r_vec = tx_ring->r_vec; 273 274 nr_frags = skb_shinfo(skb)->nr_frags; 275 276 if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) { 277 nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n", 278 qidx, tx_ring->wr_p, tx_ring->rd_p); 279 nd_q = netdev_get_tx_queue(dp->netdev, qidx); 280 netif_tx_stop_queue(nd_q); 281 nfp_net_tx_xmit_more_flush(tx_ring); 282 u64_stats_update_begin(&r_vec->tx_sync); 283 r_vec->tx_busy++; 284 u64_stats_update_end(&r_vec->tx_sync); 285 return NETDEV_TX_BUSY; 286 } 287 288 skb = nfp_net_tls_tx(dp, r_vec, skb, &tls_handle, &nr_frags); 289 if (unlikely(!skb)) { 290 nfp_net_tx_xmit_more_flush(tx_ring); 291 return NETDEV_TX_OK; 292 } 293 294 md_bytes = nfp_nfd3_prep_tx_meta(dp, skb, tls_handle, &ipsec); 295 if (unlikely(md_bytes < 0)) 296 goto err_flush; 297 298 /* Start with the head skbuf */ 299 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb), 300 DMA_TO_DEVICE); 301 if (dma_mapping_error(dp->dev, dma_addr)) 302 goto err_dma_err; 303 304 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 305 306 /* Stash the soft descriptor of the head then initialize it */ 307 txbuf = &tx_ring->txbufs[wr_idx]; 308 txbuf->skb = skb; 309 txbuf->dma_addr = dma_addr; 310 txbuf->fidx = -1; 311 txbuf->pkt_cnt = 1; 312 txbuf->real_len = skb->len; 313 314 /* Build TX descriptor */ 315 txd = &tx_ring->txds[wr_idx]; 316 txd->offset_eop = (nr_frags ? 0 : NFD3_DESC_TX_EOP) | md_bytes; 317 txd->dma_len = cpu_to_le16(skb_headlen(skb)); 318 nfp_desc_set_dma_addr_40b(txd, dma_addr); 319 txd->data_len = cpu_to_le16(skb->len); 320 321 txd->flags = 0; 322 txd->mss = 0; 323 txd->lso_hdrlen = 0; 324 325 /* Do not reorder - tso may adjust pkt cnt, vlan may override fields */ 326 nfp_nfd3_tx_tso(r_vec, txbuf, txd, skb, md_bytes); 327 if (ipsec) 328 nfp_nfd3_ipsec_tx(txd, skb); 329 else 330 nfp_nfd3_tx_csum(dp, r_vec, txbuf, txd, skb); 331 if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) { 332 txd->flags |= NFD3_DESC_TX_VLAN; 333 txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb)); 334 } 335 336 /* Gather DMA */ 337 if (nr_frags > 0) { 338 __le64 second_half; 339 340 /* all descs must match except for in addr, length and eop */ 341 second_half = txd->vals8[1]; 342 343 for (f = 0; f < nr_frags; f++) { 344 frag = &skb_shinfo(skb)->frags[f]; 345 fsize = skb_frag_size(frag); 346 347 dma_addr = skb_frag_dma_map(dp->dev, frag, 0, 348 fsize, DMA_TO_DEVICE); 349 if (dma_mapping_error(dp->dev, dma_addr)) 350 goto err_unmap; 351 352 wr_idx = D_IDX(tx_ring, wr_idx + 1); 353 tx_ring->txbufs[wr_idx].skb = skb; 354 tx_ring->txbufs[wr_idx].dma_addr = dma_addr; 355 tx_ring->txbufs[wr_idx].fidx = f; 356 357 txd = &tx_ring->txds[wr_idx]; 358 txd->dma_len = cpu_to_le16(fsize); 359 nfp_desc_set_dma_addr_40b(txd, dma_addr); 360 txd->offset_eop = md_bytes | 361 ((f == nr_frags - 1) ? NFD3_DESC_TX_EOP : 0); 362 txd->vals8[1] = second_half; 363 } 364 365 u64_stats_update_begin(&r_vec->tx_sync); 366 r_vec->tx_gather++; 367 u64_stats_update_end(&r_vec->tx_sync); 368 } 369 370 skb_tx_timestamp(skb); 371 372 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx); 373 374 tx_ring->wr_p += nr_frags + 1; 375 if (nfp_nfd3_tx_ring_should_stop(tx_ring)) 376 nfp_nfd3_tx_ring_stop(nd_q, tx_ring); 377 378 tx_ring->wr_ptr_add += nr_frags + 1; 379 if (__netdev_tx_sent_queue(nd_q, txbuf->real_len, netdev_xmit_more())) 380 nfp_net_tx_xmit_more_flush(tx_ring); 381 382 return NETDEV_TX_OK; 383 384 err_unmap: 385 while (--f >= 0) { 386 frag = &skb_shinfo(skb)->frags[f]; 387 dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr, 388 skb_frag_size(frag), DMA_TO_DEVICE); 389 tx_ring->txbufs[wr_idx].skb = NULL; 390 tx_ring->txbufs[wr_idx].dma_addr = 0; 391 tx_ring->txbufs[wr_idx].fidx = -2; 392 wr_idx = wr_idx - 1; 393 if (wr_idx < 0) 394 wr_idx += tx_ring->cnt; 395 } 396 dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr, 397 skb_headlen(skb), DMA_TO_DEVICE); 398 tx_ring->txbufs[wr_idx].skb = NULL; 399 tx_ring->txbufs[wr_idx].dma_addr = 0; 400 tx_ring->txbufs[wr_idx].fidx = -2; 401 err_dma_err: 402 nn_dp_warn(dp, "Failed to map DMA TX buffer\n"); 403 err_flush: 404 nfp_net_tx_xmit_more_flush(tx_ring); 405 u64_stats_update_begin(&r_vec->tx_sync); 406 r_vec->tx_errors++; 407 u64_stats_update_end(&r_vec->tx_sync); 408 nfp_net_tls_tx_undo(skb, tls_handle); 409 dev_kfree_skb_any(skb); 410 return NETDEV_TX_OK; 411 } 412 413 /** 414 * nfp_nfd3_tx_complete() - Handled completed TX packets 415 * @tx_ring: TX ring structure 416 * @budget: NAPI budget (only used as bool to determine if in NAPI context) 417 */ 418 void nfp_nfd3_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget) 419 { 420 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 421 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 422 u32 done_pkts = 0, done_bytes = 0; 423 struct netdev_queue *nd_q; 424 u32 qcp_rd_p; 425 int todo; 426 427 if (tx_ring->wr_p == tx_ring->rd_p) 428 return; 429 430 /* Work out how many descriptors have been transmitted */ 431 qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp); 432 433 if (qcp_rd_p == tx_ring->qcp_rd_p) 434 return; 435 436 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p); 437 438 while (todo--) { 439 const skb_frag_t *frag; 440 struct nfp_nfd3_tx_buf *tx_buf; 441 struct sk_buff *skb; 442 int fidx, nr_frags; 443 int idx; 444 445 idx = D_IDX(tx_ring, tx_ring->rd_p++); 446 tx_buf = &tx_ring->txbufs[idx]; 447 448 skb = tx_buf->skb; 449 if (!skb) 450 continue; 451 452 nr_frags = skb_shinfo(skb)->nr_frags; 453 fidx = tx_buf->fidx; 454 455 if (fidx == -1) { 456 /* unmap head */ 457 dma_unmap_single(dp->dev, tx_buf->dma_addr, 458 skb_headlen(skb), DMA_TO_DEVICE); 459 460 done_pkts += tx_buf->pkt_cnt; 461 done_bytes += tx_buf->real_len; 462 } else { 463 /* unmap fragment */ 464 frag = &skb_shinfo(skb)->frags[fidx]; 465 dma_unmap_page(dp->dev, tx_buf->dma_addr, 466 skb_frag_size(frag), DMA_TO_DEVICE); 467 } 468 469 /* check for last gather fragment */ 470 if (fidx == nr_frags - 1) 471 napi_consume_skb(skb, budget); 472 473 tx_buf->dma_addr = 0; 474 tx_buf->skb = NULL; 475 tx_buf->fidx = -2; 476 } 477 478 tx_ring->qcp_rd_p = qcp_rd_p; 479 480 u64_stats_update_begin(&r_vec->tx_sync); 481 r_vec->tx_bytes += done_bytes; 482 r_vec->tx_pkts += done_pkts; 483 u64_stats_update_end(&r_vec->tx_sync); 484 485 if (!dp->netdev) 486 return; 487 488 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx); 489 netdev_tx_completed_queue(nd_q, done_pkts, done_bytes); 490 if (nfp_nfd3_tx_ring_should_wake(tx_ring)) { 491 /* Make sure TX thread will see updated tx_ring->rd_p */ 492 smp_mb(); 493 494 if (unlikely(netif_tx_queue_stopped(nd_q))) 495 netif_tx_wake_queue(nd_q); 496 } 497 498 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt, 499 "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n", 500 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt); 501 } 502 503 static bool nfp_nfd3_xdp_complete(struct nfp_net_tx_ring *tx_ring) 504 { 505 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 506 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 507 u32 done_pkts = 0, done_bytes = 0; 508 bool done_all; 509 int idx, todo; 510 u32 qcp_rd_p; 511 512 /* Work out how many descriptors have been transmitted */ 513 qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp); 514 515 if (qcp_rd_p == tx_ring->qcp_rd_p) 516 return true; 517 518 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p); 519 520 done_all = todo <= NFP_NET_XDP_MAX_COMPLETE; 521 todo = min(todo, NFP_NET_XDP_MAX_COMPLETE); 522 523 tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo); 524 525 done_pkts = todo; 526 while (todo--) { 527 idx = D_IDX(tx_ring, tx_ring->rd_p); 528 tx_ring->rd_p++; 529 530 done_bytes += tx_ring->txbufs[idx].real_len; 531 } 532 533 u64_stats_update_begin(&r_vec->tx_sync); 534 r_vec->tx_bytes += done_bytes; 535 r_vec->tx_pkts += done_pkts; 536 u64_stats_update_end(&r_vec->tx_sync); 537 538 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt, 539 "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n", 540 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt); 541 542 return done_all; 543 } 544 545 /* Receive processing 546 */ 547 548 static void * 549 nfp_nfd3_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr) 550 { 551 void *frag; 552 553 if (!dp->xdp_prog) { 554 frag = napi_alloc_frag(dp->fl_bufsz); 555 if (unlikely(!frag)) 556 return NULL; 557 } else { 558 struct page *page; 559 560 page = dev_alloc_page(); 561 if (unlikely(!page)) 562 return NULL; 563 frag = page_address(page); 564 } 565 566 *dma_addr = nfp_net_dma_map_rx(dp, frag); 567 if (dma_mapping_error(dp->dev, *dma_addr)) { 568 nfp_net_free_frag(frag, dp->xdp_prog); 569 nn_dp_warn(dp, "Failed to map DMA RX buffer\n"); 570 return NULL; 571 } 572 573 return frag; 574 } 575 576 /** 577 * nfp_nfd3_rx_give_one() - Put mapped skb on the software and hardware rings 578 * @dp: NFP Net data path struct 579 * @rx_ring: RX ring structure 580 * @frag: page fragment buffer 581 * @dma_addr: DMA address of skb mapping 582 */ 583 static void 584 nfp_nfd3_rx_give_one(const struct nfp_net_dp *dp, 585 struct nfp_net_rx_ring *rx_ring, 586 void *frag, dma_addr_t dma_addr) 587 { 588 unsigned int wr_idx; 589 590 wr_idx = D_IDX(rx_ring, rx_ring->wr_p); 591 592 nfp_net_dma_sync_dev_rx(dp, dma_addr); 593 594 /* Stash SKB and DMA address away */ 595 rx_ring->rxbufs[wr_idx].frag = frag; 596 rx_ring->rxbufs[wr_idx].dma_addr = dma_addr; 597 598 /* Fill freelist descriptor */ 599 rx_ring->rxds[wr_idx].fld.reserved = 0; 600 rx_ring->rxds[wr_idx].fld.meta_len_dd = 0; 601 /* DMA address is expanded to 48-bit width in freelist for NFP3800, 602 * so the *_48b macro is used accordingly, it's also OK to fill 603 * a 40-bit address since the top 8 bits are get set to 0. 604 */ 605 nfp_desc_set_dma_addr_48b(&rx_ring->rxds[wr_idx].fld, 606 dma_addr + dp->rx_dma_off); 607 608 rx_ring->wr_p++; 609 if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) { 610 /* Update write pointer of the freelist queue. Make 611 * sure all writes are flushed before telling the hardware. 612 */ 613 wmb(); 614 nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH); 615 } 616 } 617 618 /** 619 * nfp_nfd3_rx_ring_fill_freelist() - Give buffers from the ring to FW 620 * @dp: NFP Net data path struct 621 * @rx_ring: RX ring to fill 622 */ 623 void nfp_nfd3_rx_ring_fill_freelist(struct nfp_net_dp *dp, 624 struct nfp_net_rx_ring *rx_ring) 625 { 626 unsigned int i; 627 628 if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx)) 629 return nfp_net_xsk_rx_ring_fill_freelist(rx_ring); 630 631 for (i = 0; i < rx_ring->cnt - 1; i++) 632 nfp_nfd3_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag, 633 rx_ring->rxbufs[i].dma_addr); 634 } 635 636 /** 637 * nfp_nfd3_rx_csum_has_errors() - group check if rxd has any csum errors 638 * @flags: RX descriptor flags field in CPU byte order 639 */ 640 static int nfp_nfd3_rx_csum_has_errors(u16 flags) 641 { 642 u16 csum_all_checked, csum_all_ok; 643 644 csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL; 645 csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK; 646 647 return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT); 648 } 649 650 /** 651 * nfp_nfd3_rx_csum() - set SKB checksum field based on RX descriptor flags 652 * @dp: NFP Net data path struct 653 * @r_vec: per-ring structure 654 * @rxd: Pointer to RX descriptor 655 * @meta: Parsed metadata prepend 656 * @skb: Pointer to SKB 657 */ 658 void 659 nfp_nfd3_rx_csum(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec, 660 const struct nfp_net_rx_desc *rxd, 661 const struct nfp_meta_parsed *meta, struct sk_buff *skb) 662 { 663 skb_checksum_none_assert(skb); 664 665 if (!(dp->netdev->features & NETIF_F_RXCSUM)) 666 return; 667 668 if (meta->csum_type) { 669 skb->ip_summed = meta->csum_type; 670 skb->csum = meta->csum; 671 u64_stats_update_begin(&r_vec->rx_sync); 672 r_vec->hw_csum_rx_complete++; 673 u64_stats_update_end(&r_vec->rx_sync); 674 return; 675 } 676 677 if (nfp_nfd3_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) { 678 u64_stats_update_begin(&r_vec->rx_sync); 679 r_vec->hw_csum_rx_error++; 680 u64_stats_update_end(&r_vec->rx_sync); 681 return; 682 } 683 684 /* Assume that the firmware will never report inner CSUM_OK unless outer 685 * L4 headers were successfully parsed. FW will always report zero UDP 686 * checksum as CSUM_OK. 687 */ 688 if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK || 689 rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) { 690 __skb_incr_checksum_unnecessary(skb); 691 u64_stats_update_begin(&r_vec->rx_sync); 692 r_vec->hw_csum_rx_ok++; 693 u64_stats_update_end(&r_vec->rx_sync); 694 } 695 696 if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK || 697 rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) { 698 __skb_incr_checksum_unnecessary(skb); 699 u64_stats_update_begin(&r_vec->rx_sync); 700 r_vec->hw_csum_rx_inner_ok++; 701 u64_stats_update_end(&r_vec->rx_sync); 702 } 703 } 704 705 static void 706 nfp_nfd3_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta, 707 unsigned int type, __be32 *hash) 708 { 709 if (!(netdev->features & NETIF_F_RXHASH)) 710 return; 711 712 switch (type) { 713 case NFP_NET_RSS_IPV4: 714 case NFP_NET_RSS_IPV6: 715 case NFP_NET_RSS_IPV6_EX: 716 meta->hash_type = PKT_HASH_TYPE_L3; 717 break; 718 default: 719 meta->hash_type = PKT_HASH_TYPE_L4; 720 break; 721 } 722 723 meta->hash = get_unaligned_be32(hash); 724 } 725 726 static void 727 nfp_nfd3_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta, 728 void *data, struct nfp_net_rx_desc *rxd) 729 { 730 struct nfp_net_rx_hash *rx_hash = data; 731 732 if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS)) 733 return; 734 735 nfp_nfd3_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type), 736 &rx_hash->hash); 737 } 738 739 bool 740 nfp_nfd3_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta, 741 void *data, void *pkt, unsigned int pkt_len, int meta_len) 742 { 743 u32 meta_info, vlan_info; 744 745 meta_info = get_unaligned_be32(data); 746 data += 4; 747 748 while (meta_info) { 749 switch (meta_info & NFP_NET_META_FIELD_MASK) { 750 case NFP_NET_META_HASH: 751 meta_info >>= NFP_NET_META_FIELD_SIZE; 752 nfp_nfd3_set_hash(netdev, meta, 753 meta_info & NFP_NET_META_FIELD_MASK, 754 (__be32 *)data); 755 data += 4; 756 break; 757 case NFP_NET_META_MARK: 758 meta->mark = get_unaligned_be32(data); 759 data += 4; 760 break; 761 case NFP_NET_META_VLAN: 762 vlan_info = get_unaligned_be32(data); 763 if (FIELD_GET(NFP_NET_META_VLAN_STRIP, vlan_info)) { 764 meta->vlan.stripped = true; 765 meta->vlan.tpid = FIELD_GET(NFP_NET_META_VLAN_TPID_MASK, 766 vlan_info); 767 meta->vlan.tci = FIELD_GET(NFP_NET_META_VLAN_TCI_MASK, 768 vlan_info); 769 } 770 data += 4; 771 break; 772 case NFP_NET_META_PORTID: 773 meta->portid = get_unaligned_be32(data); 774 data += 4; 775 break; 776 case NFP_NET_META_CSUM: 777 meta->csum_type = CHECKSUM_COMPLETE; 778 meta->csum = 779 (__force __wsum)__get_unaligned_cpu32(data); 780 data += 4; 781 break; 782 case NFP_NET_META_RESYNC_INFO: 783 if (nfp_net_tls_rx_resync_req(netdev, data, pkt, 784 pkt_len)) 785 return false; 786 data += sizeof(struct nfp_net_tls_resync_req); 787 break; 788 #ifdef CONFIG_NFP_NET_IPSEC 789 case NFP_NET_META_IPSEC: 790 /* Note: IPsec packet will have zero saidx, so need add 1 791 * to indicate packet is IPsec packet within driver. 792 */ 793 meta->ipsec_saidx = get_unaligned_be32(data) + 1; 794 data += 4; 795 break; 796 #endif 797 default: 798 return true; 799 } 800 801 meta_info >>= NFP_NET_META_FIELD_SIZE; 802 } 803 804 return data != pkt; 805 } 806 807 static void 808 nfp_nfd3_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec, 809 struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf, 810 struct sk_buff *skb) 811 { 812 u64_stats_update_begin(&r_vec->rx_sync); 813 r_vec->rx_drops++; 814 /* If we have both skb and rxbuf the replacement buffer allocation 815 * must have failed, count this as an alloc failure. 816 */ 817 if (skb && rxbuf) 818 r_vec->rx_replace_buf_alloc_fail++; 819 u64_stats_update_end(&r_vec->rx_sync); 820 821 /* skb is build based on the frag, free_skb() would free the frag 822 * so to be able to reuse it we need an extra ref. 823 */ 824 if (skb && rxbuf && skb->head == rxbuf->frag) 825 page_ref_inc(virt_to_head_page(rxbuf->frag)); 826 if (rxbuf) 827 nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr); 828 if (skb) 829 dev_kfree_skb_any(skb); 830 } 831 832 static bool 833 nfp_nfd3_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring, 834 struct nfp_net_tx_ring *tx_ring, 835 struct nfp_net_rx_buf *rxbuf, unsigned int dma_off, 836 unsigned int pkt_len, bool *completed) 837 { 838 unsigned int dma_map_sz = dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA; 839 struct nfp_nfd3_tx_buf *txbuf; 840 struct nfp_nfd3_tx_desc *txd; 841 int wr_idx; 842 843 /* Reject if xdp_adjust_tail grow packet beyond DMA area */ 844 if (pkt_len + dma_off > dma_map_sz) 845 return false; 846 847 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 848 if (!*completed) { 849 nfp_nfd3_xdp_complete(tx_ring); 850 *completed = true; 851 } 852 853 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 854 nfp_nfd3_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf, 855 NULL); 856 return false; 857 } 858 } 859 860 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 861 862 /* Stash the soft descriptor of the head then initialize it */ 863 txbuf = &tx_ring->txbufs[wr_idx]; 864 865 nfp_nfd3_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr); 866 867 txbuf->frag = rxbuf->frag; 868 txbuf->dma_addr = rxbuf->dma_addr; 869 txbuf->fidx = -1; 870 txbuf->pkt_cnt = 1; 871 txbuf->real_len = pkt_len; 872 873 dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off, 874 pkt_len, DMA_BIDIRECTIONAL); 875 876 /* Build TX descriptor */ 877 txd = &tx_ring->txds[wr_idx]; 878 txd->offset_eop = NFD3_DESC_TX_EOP; 879 txd->dma_len = cpu_to_le16(pkt_len); 880 nfp_desc_set_dma_addr_40b(txd, rxbuf->dma_addr + dma_off); 881 txd->data_len = cpu_to_le16(pkt_len); 882 883 txd->flags = 0; 884 txd->mss = 0; 885 txd->lso_hdrlen = 0; 886 887 tx_ring->wr_p++; 888 tx_ring->wr_ptr_add++; 889 return true; 890 } 891 892 /** 893 * nfp_nfd3_rx() - receive up to @budget packets on @rx_ring 894 * @rx_ring: RX ring to receive from 895 * @budget: NAPI budget 896 * 897 * Note, this function is separated out from the napi poll function to 898 * more cleanly separate packet receive code from other bookkeeping 899 * functions performed in the napi poll function. 900 * 901 * Return: Number of packets received. 902 */ 903 static int nfp_nfd3_rx(struct nfp_net_rx_ring *rx_ring, int budget) 904 { 905 struct nfp_net_r_vector *r_vec = rx_ring->r_vec; 906 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 907 struct nfp_net_tx_ring *tx_ring; 908 struct bpf_prog *xdp_prog; 909 int idx, pkts_polled = 0; 910 bool xdp_tx_cmpl = false; 911 unsigned int true_bufsz; 912 struct sk_buff *skb; 913 struct xdp_buff xdp; 914 915 xdp_prog = READ_ONCE(dp->xdp_prog); 916 true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz; 917 xdp_init_buff(&xdp, PAGE_SIZE - NFP_NET_RX_BUF_HEADROOM, 918 &rx_ring->xdp_rxq); 919 tx_ring = r_vec->xdp_ring; 920 921 while (pkts_polled < budget) { 922 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off; 923 struct nfp_net_rx_buf *rxbuf; 924 struct nfp_net_rx_desc *rxd; 925 struct nfp_meta_parsed meta; 926 bool redir_egress = false; 927 struct net_device *netdev; 928 dma_addr_t new_dma_addr; 929 u32 meta_len_xdp = 0; 930 void *new_frag; 931 932 idx = D_IDX(rx_ring, rx_ring->rd_p); 933 934 rxd = &rx_ring->rxds[idx]; 935 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD)) 936 break; 937 938 /* Memory barrier to ensure that we won't do other reads 939 * before the DD bit. 940 */ 941 dma_rmb(); 942 943 memset(&meta, 0, sizeof(meta)); 944 945 rx_ring->rd_p++; 946 pkts_polled++; 947 948 rxbuf = &rx_ring->rxbufs[idx]; 949 /* < meta_len > 950 * <-- [rx_offset] --> 951 * --------------------------------------------------------- 952 * | [XX] | metadata | packet | XXXX | 953 * --------------------------------------------------------- 954 * <---------------- data_len ---------------> 955 * 956 * The rx_offset is fixed for all packets, the meta_len can vary 957 * on a packet by packet basis. If rx_offset is set to zero 958 * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the 959 * buffer and is immediately followed by the packet (no [XX]). 960 */ 961 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK; 962 data_len = le16_to_cpu(rxd->rxd.data_len); 963 pkt_len = data_len - meta_len; 964 965 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off; 966 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 967 pkt_off += meta_len; 968 else 969 pkt_off += dp->rx_offset; 970 meta_off = pkt_off - meta_len; 971 972 /* Stats update */ 973 u64_stats_update_begin(&r_vec->rx_sync); 974 r_vec->rx_pkts++; 975 r_vec->rx_bytes += pkt_len; 976 u64_stats_update_end(&r_vec->rx_sync); 977 978 if (unlikely(meta_len > NFP_NET_MAX_PREPEND || 979 (dp->rx_offset && meta_len > dp->rx_offset))) { 980 nn_dp_warn(dp, "oversized RX packet metadata %u\n", 981 meta_len); 982 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 983 continue; 984 } 985 986 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, 987 data_len); 988 989 if (!dp->chained_metadata_format) { 990 nfp_nfd3_set_hash_desc(dp->netdev, &meta, 991 rxbuf->frag + meta_off, rxd); 992 } else if (meta_len) { 993 if (unlikely(nfp_nfd3_parse_meta(dp->netdev, &meta, 994 rxbuf->frag + meta_off, 995 rxbuf->frag + pkt_off, 996 pkt_len, meta_len))) { 997 nn_dp_warn(dp, "invalid RX packet metadata\n"); 998 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, 999 NULL); 1000 continue; 1001 } 1002 } 1003 1004 if (xdp_prog && !meta.portid) { 1005 void *orig_data = rxbuf->frag + pkt_off; 1006 unsigned int dma_off; 1007 int act; 1008 1009 xdp_prepare_buff(&xdp, 1010 rxbuf->frag + NFP_NET_RX_BUF_HEADROOM, 1011 pkt_off - NFP_NET_RX_BUF_HEADROOM, 1012 pkt_len, true); 1013 1014 act = bpf_prog_run_xdp(xdp_prog, &xdp); 1015 1016 pkt_len = xdp.data_end - xdp.data; 1017 pkt_off += xdp.data - orig_data; 1018 1019 switch (act) { 1020 case XDP_PASS: 1021 meta_len_xdp = xdp.data - xdp.data_meta; 1022 break; 1023 case XDP_TX: 1024 dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM; 1025 if (unlikely(!nfp_nfd3_tx_xdp_buf(dp, rx_ring, 1026 tx_ring, 1027 rxbuf, 1028 dma_off, 1029 pkt_len, 1030 &xdp_tx_cmpl))) 1031 trace_xdp_exception(dp->netdev, 1032 xdp_prog, act); 1033 continue; 1034 default: 1035 bpf_warn_invalid_xdp_action(dp->netdev, xdp_prog, act); 1036 fallthrough; 1037 case XDP_ABORTED: 1038 trace_xdp_exception(dp->netdev, xdp_prog, act); 1039 fallthrough; 1040 case XDP_DROP: 1041 nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag, 1042 rxbuf->dma_addr); 1043 continue; 1044 } 1045 } 1046 1047 if (likely(!meta.portid)) { 1048 netdev = dp->netdev; 1049 } else if (meta.portid == NFP_META_PORT_ID_CTRL) { 1050 struct nfp_net *nn = netdev_priv(dp->netdev); 1051 1052 nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off, 1053 pkt_len); 1054 nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag, 1055 rxbuf->dma_addr); 1056 continue; 1057 } else { 1058 struct nfp_net *nn; 1059 1060 nn = netdev_priv(dp->netdev); 1061 netdev = nfp_app_dev_get(nn->app, meta.portid, 1062 &redir_egress); 1063 if (unlikely(!netdev)) { 1064 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, 1065 NULL); 1066 continue; 1067 } 1068 1069 if (nfp_netdev_is_nfp_repr(netdev)) 1070 nfp_repr_inc_rx_stats(netdev, pkt_len); 1071 } 1072 1073 skb = build_skb(rxbuf->frag, true_bufsz); 1074 if (unlikely(!skb)) { 1075 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 1076 continue; 1077 } 1078 new_frag = nfp_nfd3_napi_alloc_one(dp, &new_dma_addr); 1079 if (unlikely(!new_frag)) { 1080 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, skb); 1081 continue; 1082 } 1083 1084 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr); 1085 1086 nfp_nfd3_rx_give_one(dp, rx_ring, new_frag, new_dma_addr); 1087 1088 skb_reserve(skb, pkt_off); 1089 skb_put(skb, pkt_len); 1090 1091 skb->mark = meta.mark; 1092 skb_set_hash(skb, meta.hash, meta.hash_type); 1093 1094 skb_record_rx_queue(skb, rx_ring->idx); 1095 skb->protocol = eth_type_trans(skb, netdev); 1096 1097 nfp_nfd3_rx_csum(dp, r_vec, rxd, &meta, skb); 1098 1099 #ifdef CONFIG_TLS_DEVICE 1100 if (rxd->rxd.flags & PCIE_DESC_RX_DECRYPTED) { 1101 skb->decrypted = true; 1102 u64_stats_update_begin(&r_vec->rx_sync); 1103 r_vec->hw_tls_rx++; 1104 u64_stats_update_end(&r_vec->rx_sync); 1105 } 1106 #endif 1107 1108 if (unlikely(!nfp_net_vlan_strip(skb, rxd, &meta))) { 1109 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, NULL, skb); 1110 continue; 1111 } 1112 1113 #ifdef CONFIG_NFP_NET_IPSEC 1114 if (meta.ipsec_saidx != 0 && unlikely(nfp_net_ipsec_rx(&meta, skb))) { 1115 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, NULL, skb); 1116 continue; 1117 } 1118 #endif 1119 1120 if (meta_len_xdp) 1121 skb_metadata_set(skb, meta_len_xdp); 1122 1123 if (likely(!redir_egress)) { 1124 napi_gro_receive(&rx_ring->r_vec->napi, skb); 1125 } else { 1126 skb->dev = netdev; 1127 skb_reset_network_header(skb); 1128 __skb_push(skb, ETH_HLEN); 1129 dev_queue_xmit(skb); 1130 } 1131 } 1132 1133 if (xdp_prog) { 1134 if (tx_ring->wr_ptr_add) 1135 nfp_net_tx_xmit_more_flush(tx_ring); 1136 else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) && 1137 !xdp_tx_cmpl) 1138 if (!nfp_nfd3_xdp_complete(tx_ring)) 1139 pkts_polled = budget; 1140 } 1141 1142 return pkts_polled; 1143 } 1144 1145 /** 1146 * nfp_nfd3_poll() - napi poll function 1147 * @napi: NAPI structure 1148 * @budget: NAPI budget 1149 * 1150 * Return: number of packets polled. 1151 */ 1152 int nfp_nfd3_poll(struct napi_struct *napi, int budget) 1153 { 1154 struct nfp_net_r_vector *r_vec = 1155 container_of(napi, struct nfp_net_r_vector, napi); 1156 unsigned int pkts_polled = 0; 1157 1158 if (r_vec->tx_ring) 1159 nfp_nfd3_tx_complete(r_vec->tx_ring, budget); 1160 if (r_vec->rx_ring) 1161 pkts_polled = nfp_nfd3_rx(r_vec->rx_ring, budget); 1162 1163 if (pkts_polled < budget) 1164 if (napi_complete_done(napi, pkts_polled)) 1165 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry); 1166 1167 if (r_vec->nfp_net->rx_coalesce_adapt_on && r_vec->rx_ring) { 1168 struct dim_sample dim_sample = {}; 1169 unsigned int start; 1170 u64 pkts, bytes; 1171 1172 do { 1173 start = u64_stats_fetch_begin(&r_vec->rx_sync); 1174 pkts = r_vec->rx_pkts; 1175 bytes = r_vec->rx_bytes; 1176 } while (u64_stats_fetch_retry(&r_vec->rx_sync, start)); 1177 1178 dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample); 1179 net_dim(&r_vec->rx_dim, dim_sample); 1180 } 1181 1182 if (r_vec->nfp_net->tx_coalesce_adapt_on && r_vec->tx_ring) { 1183 struct dim_sample dim_sample = {}; 1184 unsigned int start; 1185 u64 pkts, bytes; 1186 1187 do { 1188 start = u64_stats_fetch_begin(&r_vec->tx_sync); 1189 pkts = r_vec->tx_pkts; 1190 bytes = r_vec->tx_bytes; 1191 } while (u64_stats_fetch_retry(&r_vec->tx_sync, start)); 1192 1193 dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample); 1194 net_dim(&r_vec->tx_dim, dim_sample); 1195 } 1196 1197 return pkts_polled; 1198 } 1199 1200 /* Control device data path 1201 */ 1202 1203 bool 1204 nfp_nfd3_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec, 1205 struct sk_buff *skb, bool old) 1206 { 1207 unsigned int real_len = skb->len, meta_len = 0; 1208 struct nfp_net_tx_ring *tx_ring; 1209 struct nfp_nfd3_tx_buf *txbuf; 1210 struct nfp_nfd3_tx_desc *txd; 1211 struct nfp_net_dp *dp; 1212 dma_addr_t dma_addr; 1213 int wr_idx; 1214 1215 dp = &r_vec->nfp_net->dp; 1216 tx_ring = r_vec->tx_ring; 1217 1218 if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) { 1219 nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n"); 1220 goto err_free; 1221 } 1222 1223 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 1224 u64_stats_update_begin(&r_vec->tx_sync); 1225 r_vec->tx_busy++; 1226 u64_stats_update_end(&r_vec->tx_sync); 1227 if (!old) 1228 __skb_queue_tail(&r_vec->queue, skb); 1229 else 1230 __skb_queue_head(&r_vec->queue, skb); 1231 return true; 1232 } 1233 1234 if (nfp_app_ctrl_has_meta(nn->app)) { 1235 if (unlikely(skb_headroom(skb) < 8)) { 1236 nn_dp_warn(dp, "CTRL TX on skb without headroom\n"); 1237 goto err_free; 1238 } 1239 meta_len = 8; 1240 put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4)); 1241 put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4)); 1242 } 1243 1244 /* Start with the head skbuf */ 1245 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb), 1246 DMA_TO_DEVICE); 1247 if (dma_mapping_error(dp->dev, dma_addr)) 1248 goto err_dma_warn; 1249 1250 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 1251 1252 /* Stash the soft descriptor of the head then initialize it */ 1253 txbuf = &tx_ring->txbufs[wr_idx]; 1254 txbuf->skb = skb; 1255 txbuf->dma_addr = dma_addr; 1256 txbuf->fidx = -1; 1257 txbuf->pkt_cnt = 1; 1258 txbuf->real_len = real_len; 1259 1260 /* Build TX descriptor */ 1261 txd = &tx_ring->txds[wr_idx]; 1262 txd->offset_eop = meta_len | NFD3_DESC_TX_EOP; 1263 txd->dma_len = cpu_to_le16(skb_headlen(skb)); 1264 nfp_desc_set_dma_addr_40b(txd, dma_addr); 1265 txd->data_len = cpu_to_le16(skb->len); 1266 1267 txd->flags = 0; 1268 txd->mss = 0; 1269 txd->lso_hdrlen = 0; 1270 1271 tx_ring->wr_p++; 1272 tx_ring->wr_ptr_add++; 1273 nfp_net_tx_xmit_more_flush(tx_ring); 1274 1275 return false; 1276 1277 err_dma_warn: 1278 nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n"); 1279 err_free: 1280 u64_stats_update_begin(&r_vec->tx_sync); 1281 r_vec->tx_errors++; 1282 u64_stats_update_end(&r_vec->tx_sync); 1283 dev_kfree_skb_any(skb); 1284 return false; 1285 } 1286 1287 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec) 1288 { 1289 struct sk_buff *skb; 1290 1291 while ((skb = __skb_dequeue(&r_vec->queue))) 1292 if (nfp_nfd3_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true)) 1293 return; 1294 } 1295 1296 static bool 1297 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len) 1298 { 1299 u32 meta_type, meta_tag; 1300 1301 if (!nfp_app_ctrl_has_meta(nn->app)) 1302 return !meta_len; 1303 1304 if (meta_len != 8) 1305 return false; 1306 1307 meta_type = get_unaligned_be32(data); 1308 meta_tag = get_unaligned_be32(data + 4); 1309 1310 return (meta_type == NFP_NET_META_PORTID && 1311 meta_tag == NFP_META_PORT_ID_CTRL); 1312 } 1313 1314 static bool 1315 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp, 1316 struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring) 1317 { 1318 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off; 1319 struct nfp_net_rx_buf *rxbuf; 1320 struct nfp_net_rx_desc *rxd; 1321 dma_addr_t new_dma_addr; 1322 struct sk_buff *skb; 1323 void *new_frag; 1324 int idx; 1325 1326 idx = D_IDX(rx_ring, rx_ring->rd_p); 1327 1328 rxd = &rx_ring->rxds[idx]; 1329 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD)) 1330 return false; 1331 1332 /* Memory barrier to ensure that we won't do other reads 1333 * before the DD bit. 1334 */ 1335 dma_rmb(); 1336 1337 rx_ring->rd_p++; 1338 1339 rxbuf = &rx_ring->rxbufs[idx]; 1340 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK; 1341 data_len = le16_to_cpu(rxd->rxd.data_len); 1342 pkt_len = data_len - meta_len; 1343 1344 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off; 1345 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 1346 pkt_off += meta_len; 1347 else 1348 pkt_off += dp->rx_offset; 1349 meta_off = pkt_off - meta_len; 1350 1351 /* Stats update */ 1352 u64_stats_update_begin(&r_vec->rx_sync); 1353 r_vec->rx_pkts++; 1354 r_vec->rx_bytes += pkt_len; 1355 u64_stats_update_end(&r_vec->rx_sync); 1356 1357 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, data_len); 1358 1359 if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) { 1360 nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n", 1361 meta_len); 1362 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 1363 return true; 1364 } 1365 1366 skb = build_skb(rxbuf->frag, dp->fl_bufsz); 1367 if (unlikely(!skb)) { 1368 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 1369 return true; 1370 } 1371 new_frag = nfp_nfd3_napi_alloc_one(dp, &new_dma_addr); 1372 if (unlikely(!new_frag)) { 1373 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, skb); 1374 return true; 1375 } 1376 1377 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr); 1378 1379 nfp_nfd3_rx_give_one(dp, rx_ring, new_frag, new_dma_addr); 1380 1381 skb_reserve(skb, pkt_off); 1382 skb_put(skb, pkt_len); 1383 1384 nfp_app_ctrl_rx(nn->app, skb); 1385 1386 return true; 1387 } 1388 1389 static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec) 1390 { 1391 struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring; 1392 struct nfp_net *nn = r_vec->nfp_net; 1393 struct nfp_net_dp *dp = &nn->dp; 1394 unsigned int budget = 512; 1395 1396 while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--) 1397 continue; 1398 1399 return budget; 1400 } 1401 1402 void nfp_nfd3_ctrl_poll(struct tasklet_struct *t) 1403 { 1404 struct nfp_net_r_vector *r_vec = from_tasklet(r_vec, t, tasklet); 1405 1406 spin_lock(&r_vec->lock); 1407 nfp_nfd3_tx_complete(r_vec->tx_ring, 0); 1408 __nfp_ctrl_tx_queued(r_vec); 1409 spin_unlock(&r_vec->lock); 1410 1411 if (nfp_ctrl_rx(r_vec)) { 1412 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry); 1413 } else { 1414 tasklet_schedule(&r_vec->tasklet); 1415 nn_dp_warn(&r_vec->nfp_net->dp, 1416 "control message budget exceeded!\n"); 1417 } 1418 } 1419