1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * drivers/net/veth.c 4 * 5 * Copyright (C) 2007 OpenVZ http://openvz.org, SWsoft Inc 6 * 7 * Author: Pavel Emelianov <xemul@openvz.org> 8 * Ethtool interface from: Eric W. Biederman <ebiederm@xmission.com> 9 * 10 */ 11 12 #include <linux/netdevice.h> 13 #include <linux/slab.h> 14 #include <linux/ethtool.h> 15 #include <linux/etherdevice.h> 16 #include <linux/u64_stats_sync.h> 17 18 #include <net/rtnetlink.h> 19 #include <net/dst.h> 20 #include <net/xfrm.h> 21 #include <net/xdp.h> 22 #include <linux/veth.h> 23 #include <linux/module.h> 24 #include <linux/bpf.h> 25 #include <linux/filter.h> 26 #include <linux/ptr_ring.h> 27 #include <linux/bpf_trace.h> 28 #include <linux/net_tstamp.h> 29 #include <net/page_pool/helpers.h> 30 31 #define DRV_NAME "veth" 32 #define DRV_VERSION "1.0" 33 34 #define VETH_XDP_FLAG BIT(0) 35 #define VETH_RING_SIZE 256 36 #define VETH_XDP_HEADROOM (XDP_PACKET_HEADROOM + NET_IP_ALIGN) 37 38 #define VETH_XDP_TX_BULK_SIZE 16 39 #define VETH_XDP_BATCH 16 40 41 struct veth_stats { 42 u64 rx_drops; 43 /* xdp */ 44 u64 xdp_packets; 45 u64 xdp_bytes; 46 u64 xdp_redirect; 47 u64 xdp_drops; 48 u64 xdp_tx; 49 u64 xdp_tx_err; 50 u64 peer_tq_xdp_xmit; 51 u64 peer_tq_xdp_xmit_err; 52 }; 53 54 struct veth_rq_stats { 55 struct veth_stats vs; 56 struct u64_stats_sync syncp; 57 }; 58 59 struct veth_rq { 60 struct napi_struct xdp_napi; 61 struct napi_struct __rcu *napi; /* points to xdp_napi when the latter is initialized */ 62 struct net_device *dev; 63 struct bpf_prog __rcu *xdp_prog; 64 struct xdp_mem_info xdp_mem; 65 struct veth_rq_stats stats; 66 bool rx_notify_masked; 67 struct ptr_ring xdp_ring; 68 struct xdp_rxq_info xdp_rxq; 69 struct page_pool *page_pool; 70 }; 71 72 struct veth_priv { 73 struct net_device __rcu *peer; 74 atomic64_t dropped; 75 struct bpf_prog *_xdp_prog; 76 struct veth_rq *rq; 77 unsigned int requested_headroom; 78 }; 79 80 struct veth_xdp_tx_bq { 81 struct xdp_frame *q[VETH_XDP_TX_BULK_SIZE]; 82 unsigned int count; 83 }; 84 85 /* 86 * ethtool interface 87 */ 88 89 struct veth_q_stat_desc { 90 char desc[ETH_GSTRING_LEN]; 91 size_t offset; 92 }; 93 94 #define VETH_RQ_STAT(m) offsetof(struct veth_stats, m) 95 96 static const struct veth_q_stat_desc veth_rq_stats_desc[] = { 97 { "xdp_packets", VETH_RQ_STAT(xdp_packets) }, 98 { "xdp_bytes", VETH_RQ_STAT(xdp_bytes) }, 99 { "drops", VETH_RQ_STAT(rx_drops) }, 100 { "xdp_redirect", VETH_RQ_STAT(xdp_redirect) }, 101 { "xdp_drops", VETH_RQ_STAT(xdp_drops) }, 102 { "xdp_tx", VETH_RQ_STAT(xdp_tx) }, 103 { "xdp_tx_errors", VETH_RQ_STAT(xdp_tx_err) }, 104 }; 105 106 #define VETH_RQ_STATS_LEN ARRAY_SIZE(veth_rq_stats_desc) 107 108 static const struct veth_q_stat_desc veth_tq_stats_desc[] = { 109 { "xdp_xmit", VETH_RQ_STAT(peer_tq_xdp_xmit) }, 110 { "xdp_xmit_errors", VETH_RQ_STAT(peer_tq_xdp_xmit_err) }, 111 }; 112 113 #define VETH_TQ_STATS_LEN ARRAY_SIZE(veth_tq_stats_desc) 114 115 static struct { 116 const char string[ETH_GSTRING_LEN]; 117 } ethtool_stats_keys[] = { 118 { "peer_ifindex" }, 119 }; 120 121 struct veth_xdp_buff { 122 struct xdp_buff xdp; 123 struct sk_buff *skb; 124 }; 125 126 static int veth_get_link_ksettings(struct net_device *dev, 127 struct ethtool_link_ksettings *cmd) 128 { 129 cmd->base.speed = SPEED_10000; 130 cmd->base.duplex = DUPLEX_FULL; 131 cmd->base.port = PORT_TP; 132 cmd->base.autoneg = AUTONEG_DISABLE; 133 return 0; 134 } 135 136 static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info) 137 { 138 strscpy(info->driver, DRV_NAME, sizeof(info->driver)); 139 strscpy(info->version, DRV_VERSION, sizeof(info->version)); 140 } 141 142 static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf) 143 { 144 u8 *p = buf; 145 int i, j; 146 147 switch(stringset) { 148 case ETH_SS_STATS: 149 memcpy(p, ðtool_stats_keys, sizeof(ethtool_stats_keys)); 150 p += sizeof(ethtool_stats_keys); 151 for (i = 0; i < dev->real_num_rx_queues; i++) 152 for (j = 0; j < VETH_RQ_STATS_LEN; j++) 153 ethtool_sprintf(&p, "rx_queue_%u_%.18s", 154 i, veth_rq_stats_desc[j].desc); 155 156 for (i = 0; i < dev->real_num_tx_queues; i++) 157 for (j = 0; j < VETH_TQ_STATS_LEN; j++) 158 ethtool_sprintf(&p, "tx_queue_%u_%.18s", 159 i, veth_tq_stats_desc[j].desc); 160 161 page_pool_ethtool_stats_get_strings(p); 162 break; 163 } 164 } 165 166 static int veth_get_sset_count(struct net_device *dev, int sset) 167 { 168 switch (sset) { 169 case ETH_SS_STATS: 170 return ARRAY_SIZE(ethtool_stats_keys) + 171 VETH_RQ_STATS_LEN * dev->real_num_rx_queues + 172 VETH_TQ_STATS_LEN * dev->real_num_tx_queues + 173 page_pool_ethtool_stats_get_count(); 174 default: 175 return -EOPNOTSUPP; 176 } 177 } 178 179 static void veth_get_page_pool_stats(struct net_device *dev, u64 *data) 180 { 181 #ifdef CONFIG_PAGE_POOL_STATS 182 struct veth_priv *priv = netdev_priv(dev); 183 struct page_pool_stats pp_stats = {}; 184 int i; 185 186 for (i = 0; i < dev->real_num_rx_queues; i++) { 187 if (!priv->rq[i].page_pool) 188 continue; 189 page_pool_get_stats(priv->rq[i].page_pool, &pp_stats); 190 } 191 page_pool_ethtool_stats_get(data, &pp_stats); 192 #endif /* CONFIG_PAGE_POOL_STATS */ 193 } 194 195 static void veth_get_ethtool_stats(struct net_device *dev, 196 struct ethtool_stats *stats, u64 *data) 197 { 198 struct veth_priv *rcv_priv, *priv = netdev_priv(dev); 199 struct net_device *peer = rtnl_dereference(priv->peer); 200 int i, j, idx, pp_idx; 201 202 data[0] = peer ? peer->ifindex : 0; 203 idx = 1; 204 for (i = 0; i < dev->real_num_rx_queues; i++) { 205 const struct veth_rq_stats *rq_stats = &priv->rq[i].stats; 206 const void *stats_base = (void *)&rq_stats->vs; 207 unsigned int start; 208 size_t offset; 209 210 do { 211 start = u64_stats_fetch_begin(&rq_stats->syncp); 212 for (j = 0; j < VETH_RQ_STATS_LEN; j++) { 213 offset = veth_rq_stats_desc[j].offset; 214 data[idx + j] = *(u64 *)(stats_base + offset); 215 } 216 } while (u64_stats_fetch_retry(&rq_stats->syncp, start)); 217 idx += VETH_RQ_STATS_LEN; 218 } 219 pp_idx = idx; 220 221 if (!peer) 222 goto page_pool_stats; 223 224 rcv_priv = netdev_priv(peer); 225 for (i = 0; i < peer->real_num_rx_queues; i++) { 226 const struct veth_rq_stats *rq_stats = &rcv_priv->rq[i].stats; 227 const void *base = (void *)&rq_stats->vs; 228 unsigned int start, tx_idx = idx; 229 size_t offset; 230 231 tx_idx += (i % dev->real_num_tx_queues) * VETH_TQ_STATS_LEN; 232 do { 233 start = u64_stats_fetch_begin(&rq_stats->syncp); 234 for (j = 0; j < VETH_TQ_STATS_LEN; j++) { 235 offset = veth_tq_stats_desc[j].offset; 236 data[tx_idx + j] += *(u64 *)(base + offset); 237 } 238 } while (u64_stats_fetch_retry(&rq_stats->syncp, start)); 239 } 240 pp_idx = idx + dev->real_num_tx_queues * VETH_TQ_STATS_LEN; 241 242 page_pool_stats: 243 veth_get_page_pool_stats(dev, &data[pp_idx]); 244 } 245 246 static void veth_get_channels(struct net_device *dev, 247 struct ethtool_channels *channels) 248 { 249 channels->tx_count = dev->real_num_tx_queues; 250 channels->rx_count = dev->real_num_rx_queues; 251 channels->max_tx = dev->num_tx_queues; 252 channels->max_rx = dev->num_rx_queues; 253 } 254 255 static int veth_set_channels(struct net_device *dev, 256 struct ethtool_channels *ch); 257 258 static const struct ethtool_ops veth_ethtool_ops = { 259 .get_drvinfo = veth_get_drvinfo, 260 .get_link = ethtool_op_get_link, 261 .get_strings = veth_get_strings, 262 .get_sset_count = veth_get_sset_count, 263 .get_ethtool_stats = veth_get_ethtool_stats, 264 .get_link_ksettings = veth_get_link_ksettings, 265 .get_ts_info = ethtool_op_get_ts_info, 266 .get_channels = veth_get_channels, 267 .set_channels = veth_set_channels, 268 }; 269 270 /* general routines */ 271 272 static bool veth_is_xdp_frame(void *ptr) 273 { 274 return (unsigned long)ptr & VETH_XDP_FLAG; 275 } 276 277 static struct xdp_frame *veth_ptr_to_xdp(void *ptr) 278 { 279 return (void *)((unsigned long)ptr & ~VETH_XDP_FLAG); 280 } 281 282 static void *veth_xdp_to_ptr(struct xdp_frame *xdp) 283 { 284 return (void *)((unsigned long)xdp | VETH_XDP_FLAG); 285 } 286 287 static void veth_ptr_free(void *ptr) 288 { 289 if (veth_is_xdp_frame(ptr)) 290 xdp_return_frame(veth_ptr_to_xdp(ptr)); 291 else 292 kfree_skb(ptr); 293 } 294 295 static void __veth_xdp_flush(struct veth_rq *rq) 296 { 297 /* Write ptr_ring before reading rx_notify_masked */ 298 smp_mb(); 299 if (!READ_ONCE(rq->rx_notify_masked) && 300 napi_schedule_prep(&rq->xdp_napi)) { 301 WRITE_ONCE(rq->rx_notify_masked, true); 302 __napi_schedule(&rq->xdp_napi); 303 } 304 } 305 306 static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb) 307 { 308 if (unlikely(ptr_ring_produce(&rq->xdp_ring, skb))) { 309 dev_kfree_skb_any(skb); 310 return NET_RX_DROP; 311 } 312 313 return NET_RX_SUCCESS; 314 } 315 316 static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb, 317 struct veth_rq *rq, bool xdp) 318 { 319 return __dev_forward_skb(dev, skb) ?: xdp ? 320 veth_xdp_rx(rq, skb) : 321 __netif_rx(skb); 322 } 323 324 /* return true if the specified skb has chances of GRO aggregation 325 * Don't strive for accuracy, but try to avoid GRO overhead in the most 326 * common scenarios. 327 * When XDP is enabled, all traffic is considered eligible, as the xmit 328 * device has TSO off. 329 * When TSO is enabled on the xmit device, we are likely interested only 330 * in UDP aggregation, explicitly check for that if the skb is suspected 331 * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets - 332 * to belong to locally generated UDP traffic. 333 */ 334 static bool veth_skb_is_eligible_for_gro(const struct net_device *dev, 335 const struct net_device *rcv, 336 const struct sk_buff *skb) 337 { 338 return !(dev->features & NETIF_F_ALL_TSO) || 339 (skb->destructor == sock_wfree && 340 rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD)); 341 } 342 343 static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev) 344 { 345 struct veth_priv *rcv_priv, *priv = netdev_priv(dev); 346 struct veth_rq *rq = NULL; 347 int ret = NETDEV_TX_OK; 348 struct net_device *rcv; 349 int length = skb->len; 350 bool use_napi = false; 351 int rxq; 352 353 rcu_read_lock(); 354 rcv = rcu_dereference(priv->peer); 355 if (unlikely(!rcv) || !pskb_may_pull(skb, ETH_HLEN)) { 356 kfree_skb(skb); 357 goto drop; 358 } 359 360 rcv_priv = netdev_priv(rcv); 361 rxq = skb_get_queue_mapping(skb); 362 if (rxq < rcv->real_num_rx_queues) { 363 rq = &rcv_priv->rq[rxq]; 364 365 /* The napi pointer is available when an XDP program is 366 * attached or when GRO is enabled 367 * Don't bother with napi/GRO if the skb can't be aggregated 368 */ 369 use_napi = rcu_access_pointer(rq->napi) && 370 veth_skb_is_eligible_for_gro(dev, rcv, skb); 371 } 372 373 skb_tx_timestamp(skb); 374 if (likely(veth_forward_skb(rcv, skb, rq, use_napi) == NET_RX_SUCCESS)) { 375 if (!use_napi) 376 dev_sw_netstats_tx_add(dev, 1, length); 377 else 378 __veth_xdp_flush(rq); 379 } else { 380 drop: 381 atomic64_inc(&priv->dropped); 382 ret = NET_XMIT_DROP; 383 } 384 385 rcu_read_unlock(); 386 387 return ret; 388 } 389 390 static void veth_stats_rx(struct veth_stats *result, struct net_device *dev) 391 { 392 struct veth_priv *priv = netdev_priv(dev); 393 int i; 394 395 result->peer_tq_xdp_xmit_err = 0; 396 result->xdp_packets = 0; 397 result->xdp_tx_err = 0; 398 result->xdp_bytes = 0; 399 result->rx_drops = 0; 400 for (i = 0; i < dev->num_rx_queues; i++) { 401 u64 packets, bytes, drops, xdp_tx_err, peer_tq_xdp_xmit_err; 402 struct veth_rq_stats *stats = &priv->rq[i].stats; 403 unsigned int start; 404 405 do { 406 start = u64_stats_fetch_begin(&stats->syncp); 407 peer_tq_xdp_xmit_err = stats->vs.peer_tq_xdp_xmit_err; 408 xdp_tx_err = stats->vs.xdp_tx_err; 409 packets = stats->vs.xdp_packets; 410 bytes = stats->vs.xdp_bytes; 411 drops = stats->vs.rx_drops; 412 } while (u64_stats_fetch_retry(&stats->syncp, start)); 413 result->peer_tq_xdp_xmit_err += peer_tq_xdp_xmit_err; 414 result->xdp_tx_err += xdp_tx_err; 415 result->xdp_packets += packets; 416 result->xdp_bytes += bytes; 417 result->rx_drops += drops; 418 } 419 } 420 421 static void veth_get_stats64(struct net_device *dev, 422 struct rtnl_link_stats64 *tot) 423 { 424 struct veth_priv *priv = netdev_priv(dev); 425 struct net_device *peer; 426 struct veth_stats rx; 427 428 tot->tx_dropped = atomic64_read(&priv->dropped); 429 dev_fetch_sw_netstats(tot, dev->tstats); 430 431 veth_stats_rx(&rx, dev); 432 tot->tx_dropped += rx.xdp_tx_err; 433 tot->rx_dropped = rx.rx_drops + rx.peer_tq_xdp_xmit_err; 434 tot->rx_bytes += rx.xdp_bytes; 435 tot->rx_packets += rx.xdp_packets; 436 437 rcu_read_lock(); 438 peer = rcu_dereference(priv->peer); 439 if (peer) { 440 struct rtnl_link_stats64 tot_peer = {}; 441 442 dev_fetch_sw_netstats(&tot_peer, peer->tstats); 443 tot->rx_bytes += tot_peer.tx_bytes; 444 tot->rx_packets += tot_peer.tx_packets; 445 446 veth_stats_rx(&rx, peer); 447 tot->tx_dropped += rx.peer_tq_xdp_xmit_err; 448 tot->rx_dropped += rx.xdp_tx_err; 449 tot->tx_bytes += rx.xdp_bytes; 450 tot->tx_packets += rx.xdp_packets; 451 } 452 rcu_read_unlock(); 453 } 454 455 /* fake multicast ability */ 456 static void veth_set_multicast_list(struct net_device *dev) 457 { 458 } 459 460 static int veth_select_rxq(struct net_device *dev) 461 { 462 return smp_processor_id() % dev->real_num_rx_queues; 463 } 464 465 static struct net_device *veth_peer_dev(struct net_device *dev) 466 { 467 struct veth_priv *priv = netdev_priv(dev); 468 469 /* Callers must be under RCU read side. */ 470 return rcu_dereference(priv->peer); 471 } 472 473 static int veth_xdp_xmit(struct net_device *dev, int n, 474 struct xdp_frame **frames, 475 u32 flags, bool ndo_xmit) 476 { 477 struct veth_priv *rcv_priv, *priv = netdev_priv(dev); 478 int i, ret = -ENXIO, nxmit = 0; 479 struct net_device *rcv; 480 unsigned int max_len; 481 struct veth_rq *rq; 482 483 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) 484 return -EINVAL; 485 486 rcu_read_lock(); 487 rcv = rcu_dereference(priv->peer); 488 if (unlikely(!rcv)) 489 goto out; 490 491 rcv_priv = netdev_priv(rcv); 492 rq = &rcv_priv->rq[veth_select_rxq(rcv)]; 493 /* The napi pointer is set if NAPI is enabled, which ensures that 494 * xdp_ring is initialized on receive side and the peer device is up. 495 */ 496 if (!rcu_access_pointer(rq->napi)) 497 goto out; 498 499 max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN; 500 501 spin_lock(&rq->xdp_ring.producer_lock); 502 for (i = 0; i < n; i++) { 503 struct xdp_frame *frame = frames[i]; 504 void *ptr = veth_xdp_to_ptr(frame); 505 506 if (unlikely(xdp_get_frame_len(frame) > max_len || 507 __ptr_ring_produce(&rq->xdp_ring, ptr))) 508 break; 509 nxmit++; 510 } 511 spin_unlock(&rq->xdp_ring.producer_lock); 512 513 if (flags & XDP_XMIT_FLUSH) 514 __veth_xdp_flush(rq); 515 516 ret = nxmit; 517 if (ndo_xmit) { 518 u64_stats_update_begin(&rq->stats.syncp); 519 rq->stats.vs.peer_tq_xdp_xmit += nxmit; 520 rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit; 521 u64_stats_update_end(&rq->stats.syncp); 522 } 523 524 out: 525 rcu_read_unlock(); 526 527 return ret; 528 } 529 530 static int veth_ndo_xdp_xmit(struct net_device *dev, int n, 531 struct xdp_frame **frames, u32 flags) 532 { 533 int err; 534 535 err = veth_xdp_xmit(dev, n, frames, flags, true); 536 if (err < 0) { 537 struct veth_priv *priv = netdev_priv(dev); 538 539 atomic64_add(n, &priv->dropped); 540 } 541 542 return err; 543 } 544 545 static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq) 546 { 547 int sent, i, err = 0, drops; 548 549 sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false); 550 if (sent < 0) { 551 err = sent; 552 sent = 0; 553 } 554 555 for (i = sent; unlikely(i < bq->count); i++) 556 xdp_return_frame(bq->q[i]); 557 558 drops = bq->count - sent; 559 trace_xdp_bulk_tx(rq->dev, sent, drops, err); 560 561 u64_stats_update_begin(&rq->stats.syncp); 562 rq->stats.vs.xdp_tx += sent; 563 rq->stats.vs.xdp_tx_err += drops; 564 u64_stats_update_end(&rq->stats.syncp); 565 566 bq->count = 0; 567 } 568 569 static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq) 570 { 571 struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev); 572 struct net_device *rcv; 573 struct veth_rq *rcv_rq; 574 575 rcu_read_lock(); 576 veth_xdp_flush_bq(rq, bq); 577 rcv = rcu_dereference(priv->peer); 578 if (unlikely(!rcv)) 579 goto out; 580 581 rcv_priv = netdev_priv(rcv); 582 rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)]; 583 /* xdp_ring is initialized on receive side? */ 584 if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog))) 585 goto out; 586 587 __veth_xdp_flush(rcv_rq); 588 out: 589 rcu_read_unlock(); 590 } 591 592 static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp, 593 struct veth_xdp_tx_bq *bq) 594 { 595 struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp); 596 597 if (unlikely(!frame)) 598 return -EOVERFLOW; 599 600 if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE)) 601 veth_xdp_flush_bq(rq, bq); 602 603 bq->q[bq->count++] = frame; 604 605 return 0; 606 } 607 608 static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq, 609 struct xdp_frame *frame, 610 struct veth_xdp_tx_bq *bq, 611 struct veth_stats *stats) 612 { 613 struct xdp_frame orig_frame; 614 struct bpf_prog *xdp_prog; 615 616 rcu_read_lock(); 617 xdp_prog = rcu_dereference(rq->xdp_prog); 618 if (likely(xdp_prog)) { 619 struct veth_xdp_buff vxbuf; 620 struct xdp_buff *xdp = &vxbuf.xdp; 621 u32 act; 622 623 xdp_convert_frame_to_buff(frame, xdp); 624 xdp->rxq = &rq->xdp_rxq; 625 vxbuf.skb = NULL; 626 627 act = bpf_prog_run_xdp(xdp_prog, xdp); 628 629 switch (act) { 630 case XDP_PASS: 631 if (xdp_update_frame_from_buff(xdp, frame)) 632 goto err_xdp; 633 break; 634 case XDP_TX: 635 orig_frame = *frame; 636 xdp->rxq->mem = frame->mem; 637 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) { 638 trace_xdp_exception(rq->dev, xdp_prog, act); 639 frame = &orig_frame; 640 stats->rx_drops++; 641 goto err_xdp; 642 } 643 stats->xdp_tx++; 644 rcu_read_unlock(); 645 goto xdp_xmit; 646 case XDP_REDIRECT: 647 orig_frame = *frame; 648 xdp->rxq->mem = frame->mem; 649 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) { 650 frame = &orig_frame; 651 stats->rx_drops++; 652 goto err_xdp; 653 } 654 stats->xdp_redirect++; 655 rcu_read_unlock(); 656 goto xdp_xmit; 657 default: 658 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act); 659 fallthrough; 660 case XDP_ABORTED: 661 trace_xdp_exception(rq->dev, xdp_prog, act); 662 fallthrough; 663 case XDP_DROP: 664 stats->xdp_drops++; 665 goto err_xdp; 666 } 667 } 668 rcu_read_unlock(); 669 670 return frame; 671 err_xdp: 672 rcu_read_unlock(); 673 xdp_return_frame(frame); 674 xdp_xmit: 675 return NULL; 676 } 677 678 /* frames array contains VETH_XDP_BATCH at most */ 679 static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames, 680 int n_xdpf, struct veth_xdp_tx_bq *bq, 681 struct veth_stats *stats) 682 { 683 void *skbs[VETH_XDP_BATCH]; 684 int i; 685 686 if (xdp_alloc_skb_bulk(skbs, n_xdpf, 687 GFP_ATOMIC | __GFP_ZERO) < 0) { 688 for (i = 0; i < n_xdpf; i++) 689 xdp_return_frame(frames[i]); 690 stats->rx_drops += n_xdpf; 691 692 return; 693 } 694 695 for (i = 0; i < n_xdpf; i++) { 696 struct sk_buff *skb = skbs[i]; 697 698 skb = __xdp_build_skb_from_frame(frames[i], skb, 699 rq->dev); 700 if (!skb) { 701 xdp_return_frame(frames[i]); 702 stats->rx_drops++; 703 continue; 704 } 705 napi_gro_receive(&rq->xdp_napi, skb); 706 } 707 } 708 709 static void veth_xdp_get(struct xdp_buff *xdp) 710 { 711 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp); 712 int i; 713 714 get_page(virt_to_page(xdp->data)); 715 if (likely(!xdp_buff_has_frags(xdp))) 716 return; 717 718 for (i = 0; i < sinfo->nr_frags; i++) 719 __skb_frag_ref(&sinfo->frags[i]); 720 } 721 722 static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq, 723 struct xdp_buff *xdp, 724 struct sk_buff **pskb) 725 { 726 struct sk_buff *skb = *pskb; 727 u32 frame_sz; 728 729 if (skb_shared(skb) || skb_head_is_locked(skb) || 730 skb_shinfo(skb)->nr_frags || 731 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 732 if (skb_pp_cow_data(rq->page_pool, pskb, XDP_PACKET_HEADROOM)) 733 goto drop; 734 735 skb = *pskb; 736 } 737 738 /* SKB "head" area always have tailroom for skb_shared_info */ 739 frame_sz = skb_end_pointer(skb) - skb->head; 740 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 741 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq); 742 xdp_prepare_buff(xdp, skb->head, skb_headroom(skb), 743 skb_headlen(skb), true); 744 745 if (skb_is_nonlinear(skb)) { 746 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 747 xdp_buff_set_frags_flag(xdp); 748 } else { 749 xdp_buff_clear_frags_flag(xdp); 750 } 751 *pskb = skb; 752 753 return 0; 754 drop: 755 consume_skb(skb); 756 *pskb = NULL; 757 758 return -ENOMEM; 759 } 760 761 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq, 762 struct sk_buff *skb, 763 struct veth_xdp_tx_bq *bq, 764 struct veth_stats *stats) 765 { 766 void *orig_data, *orig_data_end; 767 struct bpf_prog *xdp_prog; 768 struct veth_xdp_buff vxbuf; 769 struct xdp_buff *xdp = &vxbuf.xdp; 770 u32 act, metalen; 771 int off; 772 773 skb_prepare_for_gro(skb); 774 775 rcu_read_lock(); 776 xdp_prog = rcu_dereference(rq->xdp_prog); 777 if (unlikely(!xdp_prog)) { 778 rcu_read_unlock(); 779 goto out; 780 } 781 782 __skb_push(skb, skb->data - skb_mac_header(skb)); 783 if (veth_convert_skb_to_xdp_buff(rq, xdp, &skb)) 784 goto drop; 785 vxbuf.skb = skb; 786 787 orig_data = xdp->data; 788 orig_data_end = xdp->data_end; 789 790 act = bpf_prog_run_xdp(xdp_prog, xdp); 791 792 switch (act) { 793 case XDP_PASS: 794 break; 795 case XDP_TX: 796 veth_xdp_get(xdp); 797 consume_skb(skb); 798 xdp->rxq->mem = rq->xdp_mem; 799 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) { 800 trace_xdp_exception(rq->dev, xdp_prog, act); 801 stats->rx_drops++; 802 goto err_xdp; 803 } 804 stats->xdp_tx++; 805 rcu_read_unlock(); 806 goto xdp_xmit; 807 case XDP_REDIRECT: 808 veth_xdp_get(xdp); 809 consume_skb(skb); 810 xdp->rxq->mem = rq->xdp_mem; 811 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) { 812 stats->rx_drops++; 813 goto err_xdp; 814 } 815 stats->xdp_redirect++; 816 rcu_read_unlock(); 817 goto xdp_xmit; 818 default: 819 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act); 820 fallthrough; 821 case XDP_ABORTED: 822 trace_xdp_exception(rq->dev, xdp_prog, act); 823 fallthrough; 824 case XDP_DROP: 825 stats->xdp_drops++; 826 goto xdp_drop; 827 } 828 rcu_read_unlock(); 829 830 /* check if bpf_xdp_adjust_head was used */ 831 off = orig_data - xdp->data; 832 if (off > 0) 833 __skb_push(skb, off); 834 else if (off < 0) 835 __skb_pull(skb, -off); 836 837 skb_reset_mac_header(skb); 838 839 /* check if bpf_xdp_adjust_tail was used */ 840 off = xdp->data_end - orig_data_end; 841 if (off != 0) 842 __skb_put(skb, off); /* positive on grow, negative on shrink */ 843 844 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 845 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 846 */ 847 if (xdp_buff_has_frags(xdp)) 848 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 849 else 850 skb->data_len = 0; 851 852 skb->protocol = eth_type_trans(skb, rq->dev); 853 854 metalen = xdp->data - xdp->data_meta; 855 if (metalen) 856 skb_metadata_set(skb, metalen); 857 out: 858 return skb; 859 drop: 860 stats->rx_drops++; 861 xdp_drop: 862 rcu_read_unlock(); 863 kfree_skb(skb); 864 return NULL; 865 err_xdp: 866 rcu_read_unlock(); 867 xdp_return_buff(xdp); 868 xdp_xmit: 869 return NULL; 870 } 871 872 static int veth_xdp_rcv(struct veth_rq *rq, int budget, 873 struct veth_xdp_tx_bq *bq, 874 struct veth_stats *stats) 875 { 876 int i, done = 0, n_xdpf = 0; 877 void *xdpf[VETH_XDP_BATCH]; 878 879 for (i = 0; i < budget; i++) { 880 void *ptr = __ptr_ring_consume(&rq->xdp_ring); 881 882 if (!ptr) 883 break; 884 885 if (veth_is_xdp_frame(ptr)) { 886 /* ndo_xdp_xmit */ 887 struct xdp_frame *frame = veth_ptr_to_xdp(ptr); 888 889 stats->xdp_bytes += xdp_get_frame_len(frame); 890 frame = veth_xdp_rcv_one(rq, frame, bq, stats); 891 if (frame) { 892 /* XDP_PASS */ 893 xdpf[n_xdpf++] = frame; 894 if (n_xdpf == VETH_XDP_BATCH) { 895 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, 896 bq, stats); 897 n_xdpf = 0; 898 } 899 } 900 } else { 901 /* ndo_start_xmit */ 902 struct sk_buff *skb = ptr; 903 904 stats->xdp_bytes += skb->len; 905 skb = veth_xdp_rcv_skb(rq, skb, bq, stats); 906 if (skb) { 907 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC)) 908 netif_receive_skb(skb); 909 else 910 napi_gro_receive(&rq->xdp_napi, skb); 911 } 912 } 913 done++; 914 } 915 916 if (n_xdpf) 917 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats); 918 919 u64_stats_update_begin(&rq->stats.syncp); 920 rq->stats.vs.xdp_redirect += stats->xdp_redirect; 921 rq->stats.vs.xdp_bytes += stats->xdp_bytes; 922 rq->stats.vs.xdp_drops += stats->xdp_drops; 923 rq->stats.vs.rx_drops += stats->rx_drops; 924 rq->stats.vs.xdp_packets += done; 925 u64_stats_update_end(&rq->stats.syncp); 926 927 return done; 928 } 929 930 static int veth_poll(struct napi_struct *napi, int budget) 931 { 932 struct veth_rq *rq = 933 container_of(napi, struct veth_rq, xdp_napi); 934 struct veth_stats stats = {}; 935 struct veth_xdp_tx_bq bq; 936 int done; 937 938 bq.count = 0; 939 940 xdp_set_return_frame_no_direct(); 941 done = veth_xdp_rcv(rq, budget, &bq, &stats); 942 943 if (stats.xdp_redirect > 0) 944 xdp_do_flush(); 945 946 if (done < budget && napi_complete_done(napi, done)) { 947 /* Write rx_notify_masked before reading ptr_ring */ 948 smp_store_mb(rq->rx_notify_masked, false); 949 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) { 950 if (napi_schedule_prep(&rq->xdp_napi)) { 951 WRITE_ONCE(rq->rx_notify_masked, true); 952 __napi_schedule(&rq->xdp_napi); 953 } 954 } 955 } 956 957 if (stats.xdp_tx > 0) 958 veth_xdp_flush(rq, &bq); 959 xdp_clear_return_frame_no_direct(); 960 961 return done; 962 } 963 964 static int veth_create_page_pool(struct veth_rq *rq) 965 { 966 struct page_pool_params pp_params = { 967 .order = 0, 968 .pool_size = VETH_RING_SIZE, 969 .nid = NUMA_NO_NODE, 970 .dev = &rq->dev->dev, 971 }; 972 973 rq->page_pool = page_pool_create(&pp_params); 974 if (IS_ERR(rq->page_pool)) { 975 int err = PTR_ERR(rq->page_pool); 976 977 rq->page_pool = NULL; 978 return err; 979 } 980 981 return 0; 982 } 983 984 static int __veth_napi_enable_range(struct net_device *dev, int start, int end) 985 { 986 struct veth_priv *priv = netdev_priv(dev); 987 int err, i; 988 989 for (i = start; i < end; i++) { 990 err = veth_create_page_pool(&priv->rq[i]); 991 if (err) 992 goto err_page_pool; 993 } 994 995 for (i = start; i < end; i++) { 996 struct veth_rq *rq = &priv->rq[i]; 997 998 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL); 999 if (err) 1000 goto err_xdp_ring; 1001 } 1002 1003 for (i = start; i < end; i++) { 1004 struct veth_rq *rq = &priv->rq[i]; 1005 1006 napi_enable(&rq->xdp_napi); 1007 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi); 1008 } 1009 1010 return 0; 1011 1012 err_xdp_ring: 1013 for (i--; i >= start; i--) 1014 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free); 1015 i = end; 1016 err_page_pool: 1017 for (i--; i >= start; i--) { 1018 page_pool_destroy(priv->rq[i].page_pool); 1019 priv->rq[i].page_pool = NULL; 1020 } 1021 1022 return err; 1023 } 1024 1025 static int __veth_napi_enable(struct net_device *dev) 1026 { 1027 return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues); 1028 } 1029 1030 static void veth_napi_del_range(struct net_device *dev, int start, int end) 1031 { 1032 struct veth_priv *priv = netdev_priv(dev); 1033 int i; 1034 1035 for (i = start; i < end; i++) { 1036 struct veth_rq *rq = &priv->rq[i]; 1037 1038 rcu_assign_pointer(priv->rq[i].napi, NULL); 1039 napi_disable(&rq->xdp_napi); 1040 __netif_napi_del(&rq->xdp_napi); 1041 } 1042 synchronize_net(); 1043 1044 for (i = start; i < end; i++) { 1045 struct veth_rq *rq = &priv->rq[i]; 1046 1047 rq->rx_notify_masked = false; 1048 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free); 1049 } 1050 1051 for (i = start; i < end; i++) { 1052 page_pool_destroy(priv->rq[i].page_pool); 1053 priv->rq[i].page_pool = NULL; 1054 } 1055 } 1056 1057 static void veth_napi_del(struct net_device *dev) 1058 { 1059 veth_napi_del_range(dev, 0, dev->real_num_rx_queues); 1060 } 1061 1062 static bool veth_gro_requested(const struct net_device *dev) 1063 { 1064 return !!(dev->wanted_features & NETIF_F_GRO); 1065 } 1066 1067 static int veth_enable_xdp_range(struct net_device *dev, int start, int end, 1068 bool napi_already_on) 1069 { 1070 struct veth_priv *priv = netdev_priv(dev); 1071 int err, i; 1072 1073 for (i = start; i < end; i++) { 1074 struct veth_rq *rq = &priv->rq[i]; 1075 1076 if (!napi_already_on) 1077 netif_napi_add(dev, &rq->xdp_napi, veth_poll); 1078 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id); 1079 if (err < 0) 1080 goto err_rxq_reg; 1081 1082 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq, 1083 MEM_TYPE_PAGE_SHARED, 1084 NULL); 1085 if (err < 0) 1086 goto err_reg_mem; 1087 1088 /* Save original mem info as it can be overwritten */ 1089 rq->xdp_mem = rq->xdp_rxq.mem; 1090 } 1091 return 0; 1092 1093 err_reg_mem: 1094 xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq); 1095 err_rxq_reg: 1096 for (i--; i >= start; i--) { 1097 struct veth_rq *rq = &priv->rq[i]; 1098 1099 xdp_rxq_info_unreg(&rq->xdp_rxq); 1100 if (!napi_already_on) 1101 netif_napi_del(&rq->xdp_napi); 1102 } 1103 1104 return err; 1105 } 1106 1107 static void veth_disable_xdp_range(struct net_device *dev, int start, int end, 1108 bool delete_napi) 1109 { 1110 struct veth_priv *priv = netdev_priv(dev); 1111 int i; 1112 1113 for (i = start; i < end; i++) { 1114 struct veth_rq *rq = &priv->rq[i]; 1115 1116 rq->xdp_rxq.mem = rq->xdp_mem; 1117 xdp_rxq_info_unreg(&rq->xdp_rxq); 1118 1119 if (delete_napi) 1120 netif_napi_del(&rq->xdp_napi); 1121 } 1122 } 1123 1124 static int veth_enable_xdp(struct net_device *dev) 1125 { 1126 bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP); 1127 struct veth_priv *priv = netdev_priv(dev); 1128 int err, i; 1129 1130 if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) { 1131 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on); 1132 if (err) 1133 return err; 1134 1135 if (!napi_already_on) { 1136 err = __veth_napi_enable(dev); 1137 if (err) { 1138 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true); 1139 return err; 1140 } 1141 } 1142 } 1143 1144 for (i = 0; i < dev->real_num_rx_queues; i++) { 1145 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog); 1146 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi); 1147 } 1148 1149 return 0; 1150 } 1151 1152 static void veth_disable_xdp(struct net_device *dev) 1153 { 1154 struct veth_priv *priv = netdev_priv(dev); 1155 int i; 1156 1157 for (i = 0; i < dev->real_num_rx_queues; i++) 1158 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL); 1159 1160 if (!netif_running(dev) || !veth_gro_requested(dev)) 1161 veth_napi_del(dev); 1162 1163 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false); 1164 } 1165 1166 static int veth_napi_enable_range(struct net_device *dev, int start, int end) 1167 { 1168 struct veth_priv *priv = netdev_priv(dev); 1169 int err, i; 1170 1171 for (i = start; i < end; i++) { 1172 struct veth_rq *rq = &priv->rq[i]; 1173 1174 netif_napi_add(dev, &rq->xdp_napi, veth_poll); 1175 } 1176 1177 err = __veth_napi_enable_range(dev, start, end); 1178 if (err) { 1179 for (i = start; i < end; i++) { 1180 struct veth_rq *rq = &priv->rq[i]; 1181 1182 netif_napi_del(&rq->xdp_napi); 1183 } 1184 return err; 1185 } 1186 return err; 1187 } 1188 1189 static int veth_napi_enable(struct net_device *dev) 1190 { 1191 return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues); 1192 } 1193 1194 static void veth_disable_range_safe(struct net_device *dev, int start, int end) 1195 { 1196 struct veth_priv *priv = netdev_priv(dev); 1197 1198 if (start >= end) 1199 return; 1200 1201 if (priv->_xdp_prog) { 1202 veth_napi_del_range(dev, start, end); 1203 veth_disable_xdp_range(dev, start, end, false); 1204 } else if (veth_gro_requested(dev)) { 1205 veth_napi_del_range(dev, start, end); 1206 } 1207 } 1208 1209 static int veth_enable_range_safe(struct net_device *dev, int start, int end) 1210 { 1211 struct veth_priv *priv = netdev_priv(dev); 1212 int err; 1213 1214 if (start >= end) 1215 return 0; 1216 1217 if (priv->_xdp_prog) { 1218 /* these channels are freshly initialized, napi is not on there even 1219 * when GRO is requeste 1220 */ 1221 err = veth_enable_xdp_range(dev, start, end, false); 1222 if (err) 1223 return err; 1224 1225 err = __veth_napi_enable_range(dev, start, end); 1226 if (err) { 1227 /* on error always delete the newly added napis */ 1228 veth_disable_xdp_range(dev, start, end, true); 1229 return err; 1230 } 1231 } else if (veth_gro_requested(dev)) { 1232 return veth_napi_enable_range(dev, start, end); 1233 } 1234 return 0; 1235 } 1236 1237 static void veth_set_xdp_features(struct net_device *dev) 1238 { 1239 struct veth_priv *priv = netdev_priv(dev); 1240 struct net_device *peer; 1241 1242 peer = rtnl_dereference(priv->peer); 1243 if (peer && peer->real_num_tx_queues <= dev->real_num_rx_queues) { 1244 struct veth_priv *priv_peer = netdev_priv(peer); 1245 xdp_features_t val = NETDEV_XDP_ACT_BASIC | 1246 NETDEV_XDP_ACT_REDIRECT | 1247 NETDEV_XDP_ACT_RX_SG; 1248 1249 if (priv_peer->_xdp_prog || veth_gro_requested(peer)) 1250 val |= NETDEV_XDP_ACT_NDO_XMIT | 1251 NETDEV_XDP_ACT_NDO_XMIT_SG; 1252 xdp_set_features_flag(dev, val); 1253 } else { 1254 xdp_clear_features_flag(dev); 1255 } 1256 } 1257 1258 static int veth_set_channels(struct net_device *dev, 1259 struct ethtool_channels *ch) 1260 { 1261 struct veth_priv *priv = netdev_priv(dev); 1262 unsigned int old_rx_count, new_rx_count; 1263 struct veth_priv *peer_priv; 1264 struct net_device *peer; 1265 int err; 1266 1267 /* sanity check. Upper bounds are already enforced by the caller */ 1268 if (!ch->rx_count || !ch->tx_count) 1269 return -EINVAL; 1270 1271 /* avoid braking XDP, if that is enabled */ 1272 peer = rtnl_dereference(priv->peer); 1273 peer_priv = peer ? netdev_priv(peer) : NULL; 1274 if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues) 1275 return -EINVAL; 1276 1277 if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues) 1278 return -EINVAL; 1279 1280 old_rx_count = dev->real_num_rx_queues; 1281 new_rx_count = ch->rx_count; 1282 if (netif_running(dev)) { 1283 /* turn device off */ 1284 netif_carrier_off(dev); 1285 if (peer) 1286 netif_carrier_off(peer); 1287 1288 /* try to allocate new resurces, as needed*/ 1289 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count); 1290 if (err) 1291 goto out; 1292 } 1293 1294 err = netif_set_real_num_rx_queues(dev, ch->rx_count); 1295 if (err) 1296 goto revert; 1297 1298 err = netif_set_real_num_tx_queues(dev, ch->tx_count); 1299 if (err) { 1300 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count); 1301 1302 /* this error condition could happen only if rx and tx change 1303 * in opposite directions (e.g. tx nr raises, rx nr decreases) 1304 * and we can't do anything to fully restore the original 1305 * status 1306 */ 1307 if (err2) 1308 pr_warn("Can't restore rx queues config %d -> %d %d", 1309 new_rx_count, old_rx_count, err2); 1310 else 1311 goto revert; 1312 } 1313 1314 out: 1315 if (netif_running(dev)) { 1316 /* note that we need to swap the arguments WRT the enable part 1317 * to identify the range we have to disable 1318 */ 1319 veth_disable_range_safe(dev, new_rx_count, old_rx_count); 1320 netif_carrier_on(dev); 1321 if (peer) 1322 netif_carrier_on(peer); 1323 } 1324 1325 /* update XDP supported features */ 1326 veth_set_xdp_features(dev); 1327 if (peer) 1328 veth_set_xdp_features(peer); 1329 1330 return err; 1331 1332 revert: 1333 new_rx_count = old_rx_count; 1334 old_rx_count = ch->rx_count; 1335 goto out; 1336 } 1337 1338 static int veth_open(struct net_device *dev) 1339 { 1340 struct veth_priv *priv = netdev_priv(dev); 1341 struct net_device *peer = rtnl_dereference(priv->peer); 1342 int err; 1343 1344 if (!peer) 1345 return -ENOTCONN; 1346 1347 if (priv->_xdp_prog) { 1348 err = veth_enable_xdp(dev); 1349 if (err) 1350 return err; 1351 } else if (veth_gro_requested(dev)) { 1352 err = veth_napi_enable(dev); 1353 if (err) 1354 return err; 1355 } 1356 1357 if (peer->flags & IFF_UP) { 1358 netif_carrier_on(dev); 1359 netif_carrier_on(peer); 1360 } 1361 1362 veth_set_xdp_features(dev); 1363 1364 return 0; 1365 } 1366 1367 static int veth_close(struct net_device *dev) 1368 { 1369 struct veth_priv *priv = netdev_priv(dev); 1370 struct net_device *peer = rtnl_dereference(priv->peer); 1371 1372 netif_carrier_off(dev); 1373 if (peer) 1374 netif_carrier_off(peer); 1375 1376 if (priv->_xdp_prog) 1377 veth_disable_xdp(dev); 1378 else if (veth_gro_requested(dev)) 1379 veth_napi_del(dev); 1380 1381 return 0; 1382 } 1383 1384 static int is_valid_veth_mtu(int mtu) 1385 { 1386 return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU; 1387 } 1388 1389 static int veth_alloc_queues(struct net_device *dev) 1390 { 1391 struct veth_priv *priv = netdev_priv(dev); 1392 int i; 1393 1394 priv->rq = kvcalloc(dev->num_rx_queues, sizeof(*priv->rq), 1395 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 1396 if (!priv->rq) 1397 return -ENOMEM; 1398 1399 for (i = 0; i < dev->num_rx_queues; i++) { 1400 priv->rq[i].dev = dev; 1401 u64_stats_init(&priv->rq[i].stats.syncp); 1402 } 1403 1404 return 0; 1405 } 1406 1407 static void veth_free_queues(struct net_device *dev) 1408 { 1409 struct veth_priv *priv = netdev_priv(dev); 1410 1411 kvfree(priv->rq); 1412 } 1413 1414 static int veth_dev_init(struct net_device *dev) 1415 { 1416 netdev_lockdep_set_classes(dev); 1417 return veth_alloc_queues(dev); 1418 } 1419 1420 static void veth_dev_free(struct net_device *dev) 1421 { 1422 veth_free_queues(dev); 1423 } 1424 1425 #ifdef CONFIG_NET_POLL_CONTROLLER 1426 static void veth_poll_controller(struct net_device *dev) 1427 { 1428 /* veth only receives frames when its peer sends one 1429 * Since it has nothing to do with disabling irqs, we are guaranteed 1430 * never to have pending data when we poll for it so 1431 * there is nothing to do here. 1432 * 1433 * We need this though so netpoll recognizes us as an interface that 1434 * supports polling, which enables bridge devices in virt setups to 1435 * still use netconsole 1436 */ 1437 } 1438 #endif /* CONFIG_NET_POLL_CONTROLLER */ 1439 1440 static int veth_get_iflink(const struct net_device *dev) 1441 { 1442 struct veth_priv *priv = netdev_priv(dev); 1443 struct net_device *peer; 1444 int iflink; 1445 1446 rcu_read_lock(); 1447 peer = rcu_dereference(priv->peer); 1448 iflink = peer ? READ_ONCE(peer->ifindex) : 0; 1449 rcu_read_unlock(); 1450 1451 return iflink; 1452 } 1453 1454 static netdev_features_t veth_fix_features(struct net_device *dev, 1455 netdev_features_t features) 1456 { 1457 struct veth_priv *priv = netdev_priv(dev); 1458 struct net_device *peer; 1459 1460 peer = rtnl_dereference(priv->peer); 1461 if (peer) { 1462 struct veth_priv *peer_priv = netdev_priv(peer); 1463 1464 if (peer_priv->_xdp_prog) 1465 features &= ~NETIF_F_GSO_SOFTWARE; 1466 } 1467 1468 return features; 1469 } 1470 1471 static int veth_set_features(struct net_device *dev, 1472 netdev_features_t features) 1473 { 1474 netdev_features_t changed = features ^ dev->features; 1475 struct veth_priv *priv = netdev_priv(dev); 1476 struct net_device *peer; 1477 int err; 1478 1479 if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog) 1480 return 0; 1481 1482 peer = rtnl_dereference(priv->peer); 1483 if (features & NETIF_F_GRO) { 1484 err = veth_napi_enable(dev); 1485 if (err) 1486 return err; 1487 1488 if (peer) 1489 xdp_features_set_redirect_target(peer, true); 1490 } else { 1491 if (peer) 1492 xdp_features_clear_redirect_target(peer); 1493 veth_napi_del(dev); 1494 } 1495 return 0; 1496 } 1497 1498 static void veth_set_rx_headroom(struct net_device *dev, int new_hr) 1499 { 1500 struct veth_priv *peer_priv, *priv = netdev_priv(dev); 1501 struct net_device *peer; 1502 1503 if (new_hr < 0) 1504 new_hr = 0; 1505 1506 rcu_read_lock(); 1507 peer = rcu_dereference(priv->peer); 1508 if (unlikely(!peer)) 1509 goto out; 1510 1511 peer_priv = netdev_priv(peer); 1512 priv->requested_headroom = new_hr; 1513 new_hr = max(priv->requested_headroom, peer_priv->requested_headroom); 1514 dev->needed_headroom = new_hr; 1515 peer->needed_headroom = new_hr; 1516 1517 out: 1518 rcu_read_unlock(); 1519 } 1520 1521 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog, 1522 struct netlink_ext_ack *extack) 1523 { 1524 struct veth_priv *priv = netdev_priv(dev); 1525 struct bpf_prog *old_prog; 1526 struct net_device *peer; 1527 unsigned int max_mtu; 1528 int err; 1529 1530 old_prog = priv->_xdp_prog; 1531 priv->_xdp_prog = prog; 1532 peer = rtnl_dereference(priv->peer); 1533 1534 if (prog) { 1535 if (!peer) { 1536 NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached"); 1537 err = -ENOTCONN; 1538 goto err; 1539 } 1540 1541 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) - 1542 peer->hard_header_len; 1543 /* Allow increasing the max_mtu if the program supports 1544 * XDP fragments. 1545 */ 1546 if (prog->aux->xdp_has_frags) 1547 max_mtu += PAGE_SIZE * MAX_SKB_FRAGS; 1548 1549 if (peer->mtu > max_mtu) { 1550 NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP"); 1551 err = -ERANGE; 1552 goto err; 1553 } 1554 1555 if (dev->real_num_rx_queues < peer->real_num_tx_queues) { 1556 NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues"); 1557 err = -ENOSPC; 1558 goto err; 1559 } 1560 1561 if (dev->flags & IFF_UP) { 1562 err = veth_enable_xdp(dev); 1563 if (err) { 1564 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed"); 1565 goto err; 1566 } 1567 } 1568 1569 if (!old_prog) { 1570 peer->hw_features &= ~NETIF_F_GSO_SOFTWARE; 1571 peer->max_mtu = max_mtu; 1572 } 1573 1574 xdp_features_set_redirect_target(peer, true); 1575 } 1576 1577 if (old_prog) { 1578 if (!prog) { 1579 if (peer && !veth_gro_requested(dev)) 1580 xdp_features_clear_redirect_target(peer); 1581 1582 if (dev->flags & IFF_UP) 1583 veth_disable_xdp(dev); 1584 1585 if (peer) { 1586 peer->hw_features |= NETIF_F_GSO_SOFTWARE; 1587 peer->max_mtu = ETH_MAX_MTU; 1588 } 1589 } 1590 bpf_prog_put(old_prog); 1591 } 1592 1593 if ((!!old_prog ^ !!prog) && peer) 1594 netdev_update_features(peer); 1595 1596 return 0; 1597 err: 1598 priv->_xdp_prog = old_prog; 1599 1600 return err; 1601 } 1602 1603 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp) 1604 { 1605 switch (xdp->command) { 1606 case XDP_SETUP_PROG: 1607 return veth_xdp_set(dev, xdp->prog, xdp->extack); 1608 default: 1609 return -EINVAL; 1610 } 1611 } 1612 1613 static int veth_xdp_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp) 1614 { 1615 struct veth_xdp_buff *_ctx = (void *)ctx; 1616 1617 if (!_ctx->skb) 1618 return -ENODATA; 1619 1620 *timestamp = skb_hwtstamps(_ctx->skb)->hwtstamp; 1621 return 0; 1622 } 1623 1624 static int veth_xdp_rx_hash(const struct xdp_md *ctx, u32 *hash, 1625 enum xdp_rss_hash_type *rss_type) 1626 { 1627 struct veth_xdp_buff *_ctx = (void *)ctx; 1628 struct sk_buff *skb = _ctx->skb; 1629 1630 if (!skb) 1631 return -ENODATA; 1632 1633 *hash = skb_get_hash(skb); 1634 *rss_type = skb->l4_hash ? XDP_RSS_TYPE_L4_ANY : XDP_RSS_TYPE_NONE; 1635 1636 return 0; 1637 } 1638 1639 static int veth_xdp_rx_vlan_tag(const struct xdp_md *ctx, __be16 *vlan_proto, 1640 u16 *vlan_tci) 1641 { 1642 const struct veth_xdp_buff *_ctx = (void *)ctx; 1643 const struct sk_buff *skb = _ctx->skb; 1644 int err; 1645 1646 if (!skb) 1647 return -ENODATA; 1648 1649 err = __vlan_hwaccel_get_tag(skb, vlan_tci); 1650 if (err) 1651 return err; 1652 1653 *vlan_proto = skb->vlan_proto; 1654 return err; 1655 } 1656 1657 static const struct net_device_ops veth_netdev_ops = { 1658 .ndo_init = veth_dev_init, 1659 .ndo_open = veth_open, 1660 .ndo_stop = veth_close, 1661 .ndo_start_xmit = veth_xmit, 1662 .ndo_get_stats64 = veth_get_stats64, 1663 .ndo_set_rx_mode = veth_set_multicast_list, 1664 .ndo_set_mac_address = eth_mac_addr, 1665 #ifdef CONFIG_NET_POLL_CONTROLLER 1666 .ndo_poll_controller = veth_poll_controller, 1667 #endif 1668 .ndo_get_iflink = veth_get_iflink, 1669 .ndo_fix_features = veth_fix_features, 1670 .ndo_set_features = veth_set_features, 1671 .ndo_features_check = passthru_features_check, 1672 .ndo_set_rx_headroom = veth_set_rx_headroom, 1673 .ndo_bpf = veth_xdp, 1674 .ndo_xdp_xmit = veth_ndo_xdp_xmit, 1675 .ndo_get_peer_dev = veth_peer_dev, 1676 }; 1677 1678 static const struct xdp_metadata_ops veth_xdp_metadata_ops = { 1679 .xmo_rx_timestamp = veth_xdp_rx_timestamp, 1680 .xmo_rx_hash = veth_xdp_rx_hash, 1681 .xmo_rx_vlan_tag = veth_xdp_rx_vlan_tag, 1682 }; 1683 1684 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \ 1685 NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \ 1686 NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \ 1687 NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \ 1688 NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX ) 1689 1690 static void veth_setup(struct net_device *dev) 1691 { 1692 ether_setup(dev); 1693 1694 dev->priv_flags &= ~IFF_TX_SKB_SHARING; 1695 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE; 1696 dev->priv_flags |= IFF_NO_QUEUE; 1697 dev->priv_flags |= IFF_PHONY_HEADROOM; 1698 1699 dev->netdev_ops = &veth_netdev_ops; 1700 dev->xdp_metadata_ops = &veth_xdp_metadata_ops; 1701 dev->ethtool_ops = &veth_ethtool_ops; 1702 dev->features |= NETIF_F_LLTX; 1703 dev->features |= VETH_FEATURES; 1704 dev->vlan_features = dev->features & 1705 ~(NETIF_F_HW_VLAN_CTAG_TX | 1706 NETIF_F_HW_VLAN_STAG_TX | 1707 NETIF_F_HW_VLAN_CTAG_RX | 1708 NETIF_F_HW_VLAN_STAG_RX); 1709 dev->needs_free_netdev = true; 1710 dev->priv_destructor = veth_dev_free; 1711 dev->pcpu_stat_type = NETDEV_PCPU_STAT_TSTATS; 1712 dev->max_mtu = ETH_MAX_MTU; 1713 1714 dev->hw_features = VETH_FEATURES; 1715 dev->hw_enc_features = VETH_FEATURES; 1716 dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE; 1717 netif_set_tso_max_size(dev, GSO_MAX_SIZE); 1718 } 1719 1720 /* 1721 * netlink interface 1722 */ 1723 1724 static int veth_validate(struct nlattr *tb[], struct nlattr *data[], 1725 struct netlink_ext_ack *extack) 1726 { 1727 if (tb[IFLA_ADDRESS]) { 1728 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) 1729 return -EINVAL; 1730 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) 1731 return -EADDRNOTAVAIL; 1732 } 1733 if (tb[IFLA_MTU]) { 1734 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU]))) 1735 return -EINVAL; 1736 } 1737 return 0; 1738 } 1739 1740 static struct rtnl_link_ops veth_link_ops; 1741 1742 static void veth_disable_gro(struct net_device *dev) 1743 { 1744 dev->features &= ~NETIF_F_GRO; 1745 dev->wanted_features &= ~NETIF_F_GRO; 1746 netdev_update_features(dev); 1747 } 1748 1749 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[]) 1750 { 1751 int err; 1752 1753 if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) { 1754 err = netif_set_real_num_tx_queues(dev, 1); 1755 if (err) 1756 return err; 1757 } 1758 if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) { 1759 err = netif_set_real_num_rx_queues(dev, 1); 1760 if (err) 1761 return err; 1762 } 1763 return 0; 1764 } 1765 1766 static int veth_newlink(struct net *src_net, struct net_device *dev, 1767 struct nlattr *tb[], struct nlattr *data[], 1768 struct netlink_ext_ack *extack) 1769 { 1770 int err; 1771 struct net_device *peer; 1772 struct veth_priv *priv; 1773 char ifname[IFNAMSIZ]; 1774 struct nlattr *peer_tb[IFLA_MAX + 1], **tbp; 1775 unsigned char name_assign_type; 1776 struct ifinfomsg *ifmp; 1777 struct net *net; 1778 1779 /* 1780 * create and register peer first 1781 */ 1782 if (data != NULL && data[VETH_INFO_PEER] != NULL) { 1783 struct nlattr *nla_peer; 1784 1785 nla_peer = data[VETH_INFO_PEER]; 1786 ifmp = nla_data(nla_peer); 1787 err = rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack); 1788 if (err < 0) 1789 return err; 1790 1791 err = veth_validate(peer_tb, NULL, extack); 1792 if (err < 0) 1793 return err; 1794 1795 tbp = peer_tb; 1796 } else { 1797 ifmp = NULL; 1798 tbp = tb; 1799 } 1800 1801 if (ifmp && tbp[IFLA_IFNAME]) { 1802 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ); 1803 name_assign_type = NET_NAME_USER; 1804 } else { 1805 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d"); 1806 name_assign_type = NET_NAME_ENUM; 1807 } 1808 1809 net = rtnl_link_get_net(src_net, tbp); 1810 if (IS_ERR(net)) 1811 return PTR_ERR(net); 1812 1813 peer = rtnl_create_link(net, ifname, name_assign_type, 1814 &veth_link_ops, tbp, extack); 1815 if (IS_ERR(peer)) { 1816 put_net(net); 1817 return PTR_ERR(peer); 1818 } 1819 1820 if (!ifmp || !tbp[IFLA_ADDRESS]) 1821 eth_hw_addr_random(peer); 1822 1823 if (ifmp && (dev->ifindex != 0)) 1824 peer->ifindex = ifmp->ifi_index; 1825 1826 netif_inherit_tso_max(peer, dev); 1827 1828 err = register_netdevice(peer); 1829 put_net(net); 1830 net = NULL; 1831 if (err < 0) 1832 goto err_register_peer; 1833 1834 /* keep GRO disabled by default to be consistent with the established 1835 * veth behavior 1836 */ 1837 veth_disable_gro(peer); 1838 netif_carrier_off(peer); 1839 1840 err = rtnl_configure_link(peer, ifmp, 0, NULL); 1841 if (err < 0) 1842 goto err_configure_peer; 1843 1844 /* 1845 * register dev last 1846 * 1847 * note, that since we've registered new device the dev's name 1848 * should be re-allocated 1849 */ 1850 1851 if (tb[IFLA_ADDRESS] == NULL) 1852 eth_hw_addr_random(dev); 1853 1854 if (tb[IFLA_IFNAME]) 1855 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ); 1856 else 1857 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d"); 1858 1859 err = register_netdevice(dev); 1860 if (err < 0) 1861 goto err_register_dev; 1862 1863 netif_carrier_off(dev); 1864 1865 /* 1866 * tie the deviced together 1867 */ 1868 1869 priv = netdev_priv(dev); 1870 rcu_assign_pointer(priv->peer, peer); 1871 err = veth_init_queues(dev, tb); 1872 if (err) 1873 goto err_queues; 1874 1875 priv = netdev_priv(peer); 1876 rcu_assign_pointer(priv->peer, dev); 1877 err = veth_init_queues(peer, tb); 1878 if (err) 1879 goto err_queues; 1880 1881 veth_disable_gro(dev); 1882 /* update XDP supported features */ 1883 veth_set_xdp_features(dev); 1884 veth_set_xdp_features(peer); 1885 1886 return 0; 1887 1888 err_queues: 1889 unregister_netdevice(dev); 1890 err_register_dev: 1891 /* nothing to do */ 1892 err_configure_peer: 1893 unregister_netdevice(peer); 1894 return err; 1895 1896 err_register_peer: 1897 free_netdev(peer); 1898 return err; 1899 } 1900 1901 static void veth_dellink(struct net_device *dev, struct list_head *head) 1902 { 1903 struct veth_priv *priv; 1904 struct net_device *peer; 1905 1906 priv = netdev_priv(dev); 1907 peer = rtnl_dereference(priv->peer); 1908 1909 /* Note : dellink() is called from default_device_exit_batch(), 1910 * before a rcu_synchronize() point. The devices are guaranteed 1911 * not being freed before one RCU grace period. 1912 */ 1913 RCU_INIT_POINTER(priv->peer, NULL); 1914 unregister_netdevice_queue(dev, head); 1915 1916 if (peer) { 1917 priv = netdev_priv(peer); 1918 RCU_INIT_POINTER(priv->peer, NULL); 1919 unregister_netdevice_queue(peer, head); 1920 } 1921 } 1922 1923 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = { 1924 [VETH_INFO_PEER] = { .len = sizeof(struct ifinfomsg) }, 1925 }; 1926 1927 static struct net *veth_get_link_net(const struct net_device *dev) 1928 { 1929 struct veth_priv *priv = netdev_priv(dev); 1930 struct net_device *peer = rtnl_dereference(priv->peer); 1931 1932 return peer ? dev_net(peer) : dev_net(dev); 1933 } 1934 1935 static unsigned int veth_get_num_queues(void) 1936 { 1937 /* enforce the same queue limit as rtnl_create_link */ 1938 int queues = num_possible_cpus(); 1939 1940 if (queues > 4096) 1941 queues = 4096; 1942 return queues; 1943 } 1944 1945 static struct rtnl_link_ops veth_link_ops = { 1946 .kind = DRV_NAME, 1947 .priv_size = sizeof(struct veth_priv), 1948 .setup = veth_setup, 1949 .validate = veth_validate, 1950 .newlink = veth_newlink, 1951 .dellink = veth_dellink, 1952 .policy = veth_policy, 1953 .maxtype = VETH_INFO_MAX, 1954 .get_link_net = veth_get_link_net, 1955 .get_num_tx_queues = veth_get_num_queues, 1956 .get_num_rx_queues = veth_get_num_queues, 1957 }; 1958 1959 /* 1960 * init/fini 1961 */ 1962 1963 static __init int veth_init(void) 1964 { 1965 return rtnl_link_register(&veth_link_ops); 1966 } 1967 1968 static __exit void veth_exit(void) 1969 { 1970 rtnl_link_unregister(&veth_link_ops); 1971 } 1972 1973 module_init(veth_init); 1974 module_exit(veth_exit); 1975 1976 MODULE_DESCRIPTION("Virtual Ethernet Tunnel"); 1977 MODULE_LICENSE("GPL v2"); 1978 MODULE_ALIAS_RTNL_LINK(DRV_NAME); 1979