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 <linux/skbuff_ref.h> 30 #include <net/page_pool/helpers.h> 31 32 #define DRV_NAME "veth" 33 #define DRV_VERSION "1.0" 34 35 #define VETH_XDP_FLAG BIT(0) 36 #define VETH_RING_SIZE 256 37 #define VETH_XDP_HEADROOM (XDP_PACKET_HEADROOM + NET_IP_ALIGN) 38 39 #define VETH_XDP_TX_BULK_SIZE 16 40 #define VETH_XDP_BATCH 16 41 42 struct veth_stats { 43 u64 rx_drops; 44 /* xdp */ 45 u64 xdp_packets; 46 u64 xdp_bytes; 47 u64 xdp_redirect; 48 u64 xdp_drops; 49 u64 xdp_tx; 50 u64 xdp_tx_err; 51 u64 peer_tq_xdp_xmit; 52 u64 peer_tq_xdp_xmit_err; 53 }; 54 55 struct veth_rq_stats { 56 struct veth_stats vs; 57 struct u64_stats_sync syncp; 58 }; 59 60 struct veth_rq { 61 struct napi_struct xdp_napi; 62 struct napi_struct __rcu *napi; /* points to xdp_napi when the latter is initialized */ 63 struct net_device *dev; 64 struct bpf_prog __rcu *xdp_prog; 65 struct xdp_mem_info xdp_mem; 66 struct veth_rq_stats stats; 67 bool rx_notify_masked; 68 struct ptr_ring xdp_ring; 69 struct xdp_rxq_info xdp_rxq; 70 struct page_pool *page_pool; 71 }; 72 73 struct veth_priv { 74 struct net_device __rcu *peer; 75 atomic64_t dropped; 76 struct bpf_prog *_xdp_prog; 77 struct veth_rq *rq; 78 unsigned int requested_headroom; 79 }; 80 81 struct veth_xdp_tx_bq { 82 struct xdp_frame *q[VETH_XDP_TX_BULK_SIZE]; 83 unsigned int count; 84 }; 85 86 /* 87 * ethtool interface 88 */ 89 90 struct veth_q_stat_desc { 91 char desc[ETH_GSTRING_LEN]; 92 size_t offset; 93 }; 94 95 #define VETH_RQ_STAT(m) offsetof(struct veth_stats, m) 96 97 static const struct veth_q_stat_desc veth_rq_stats_desc[] = { 98 { "xdp_packets", VETH_RQ_STAT(xdp_packets) }, 99 { "xdp_bytes", VETH_RQ_STAT(xdp_bytes) }, 100 { "drops", VETH_RQ_STAT(rx_drops) }, 101 { "xdp_redirect", VETH_RQ_STAT(xdp_redirect) }, 102 { "xdp_drops", VETH_RQ_STAT(xdp_drops) }, 103 { "xdp_tx", VETH_RQ_STAT(xdp_tx) }, 104 { "xdp_tx_errors", VETH_RQ_STAT(xdp_tx_err) }, 105 }; 106 107 #define VETH_RQ_STATS_LEN ARRAY_SIZE(veth_rq_stats_desc) 108 109 static const struct veth_q_stat_desc veth_tq_stats_desc[] = { 110 { "xdp_xmit", VETH_RQ_STAT(peer_tq_xdp_xmit) }, 111 { "xdp_xmit_errors", VETH_RQ_STAT(peer_tq_xdp_xmit_err) }, 112 }; 113 114 #define VETH_TQ_STATS_LEN ARRAY_SIZE(veth_tq_stats_desc) 115 116 static struct { 117 const char string[ETH_GSTRING_LEN]; 118 } ethtool_stats_keys[] = { 119 { "peer_ifindex" }, 120 }; 121 122 struct veth_xdp_buff { 123 struct xdp_buff xdp; 124 struct sk_buff *skb; 125 }; 126 127 static int veth_get_link_ksettings(struct net_device *dev, 128 struct ethtool_link_ksettings *cmd) 129 { 130 cmd->base.speed = SPEED_10000; 131 cmd->base.duplex = DUPLEX_FULL; 132 cmd->base.port = PORT_TP; 133 cmd->base.autoneg = AUTONEG_DISABLE; 134 return 0; 135 } 136 137 static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info) 138 { 139 strscpy(info->driver, DRV_NAME, sizeof(info->driver)); 140 strscpy(info->version, DRV_VERSION, sizeof(info->version)); 141 } 142 143 static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf) 144 { 145 u8 *p = buf; 146 int i, j; 147 148 switch(stringset) { 149 case ETH_SS_STATS: 150 memcpy(p, ðtool_stats_keys, sizeof(ethtool_stats_keys)); 151 p += sizeof(ethtool_stats_keys); 152 for (i = 0; i < dev->real_num_rx_queues; i++) 153 for (j = 0; j < VETH_RQ_STATS_LEN; j++) 154 ethtool_sprintf(&p, "rx_queue_%u_%.18s", 155 i, veth_rq_stats_desc[j].desc); 156 157 for (i = 0; i < dev->real_num_tx_queues; i++) 158 for (j = 0; j < VETH_TQ_STATS_LEN; j++) 159 ethtool_sprintf(&p, "tx_queue_%u_%.18s", 160 i, veth_tq_stats_desc[j].desc); 161 162 page_pool_ethtool_stats_get_strings(p); 163 break; 164 } 165 } 166 167 static int veth_get_sset_count(struct net_device *dev, int sset) 168 { 169 switch (sset) { 170 case ETH_SS_STATS: 171 return ARRAY_SIZE(ethtool_stats_keys) + 172 VETH_RQ_STATS_LEN * dev->real_num_rx_queues + 173 VETH_TQ_STATS_LEN * dev->real_num_tx_queues + 174 page_pool_ethtool_stats_get_count(); 175 default: 176 return -EOPNOTSUPP; 177 } 178 } 179 180 static void veth_get_page_pool_stats(struct net_device *dev, u64 *data) 181 { 182 #ifdef CONFIG_PAGE_POOL_STATS 183 struct veth_priv *priv = netdev_priv(dev); 184 struct page_pool_stats pp_stats = {}; 185 int i; 186 187 for (i = 0; i < dev->real_num_rx_queues; i++) { 188 if (!priv->rq[i].page_pool) 189 continue; 190 page_pool_get_stats(priv->rq[i].page_pool, &pp_stats); 191 } 192 page_pool_ethtool_stats_get(data, &pp_stats); 193 #endif /* CONFIG_PAGE_POOL_STATS */ 194 } 195 196 static void veth_get_ethtool_stats(struct net_device *dev, 197 struct ethtool_stats *stats, u64 *data) 198 { 199 struct veth_priv *rcv_priv, *priv = netdev_priv(dev); 200 struct net_device *peer = rtnl_dereference(priv->peer); 201 int i, j, idx, pp_idx; 202 203 data[0] = peer ? peer->ifindex : 0; 204 idx = 1; 205 for (i = 0; i < dev->real_num_rx_queues; i++) { 206 const struct veth_rq_stats *rq_stats = &priv->rq[i].stats; 207 const void *stats_base = (void *)&rq_stats->vs; 208 unsigned int start; 209 size_t offset; 210 211 do { 212 start = u64_stats_fetch_begin(&rq_stats->syncp); 213 for (j = 0; j < VETH_RQ_STATS_LEN; j++) { 214 offset = veth_rq_stats_desc[j].offset; 215 data[idx + j] = *(u64 *)(stats_base + offset); 216 } 217 } while (u64_stats_fetch_retry(&rq_stats->syncp, start)); 218 idx += VETH_RQ_STATS_LEN; 219 } 220 pp_idx = idx; 221 222 if (!peer) 223 goto page_pool_stats; 224 225 rcv_priv = netdev_priv(peer); 226 for (i = 0; i < peer->real_num_rx_queues; i++) { 227 const struct veth_rq_stats *rq_stats = &rcv_priv->rq[i].stats; 228 const void *base = (void *)&rq_stats->vs; 229 unsigned int start, tx_idx = idx; 230 size_t offset; 231 232 tx_idx += (i % dev->real_num_tx_queues) * VETH_TQ_STATS_LEN; 233 do { 234 start = u64_stats_fetch_begin(&rq_stats->syncp); 235 for (j = 0; j < VETH_TQ_STATS_LEN; j++) { 236 offset = veth_tq_stats_desc[j].offset; 237 data[tx_idx + j] += *(u64 *)(base + offset); 238 } 239 } while (u64_stats_fetch_retry(&rq_stats->syncp, start)); 240 } 241 pp_idx = idx + dev->real_num_tx_queues * VETH_TQ_STATS_LEN; 242 243 page_pool_stats: 244 veth_get_page_pool_stats(dev, &data[pp_idx]); 245 } 246 247 static void veth_get_channels(struct net_device *dev, 248 struct ethtool_channels *channels) 249 { 250 channels->tx_count = dev->real_num_tx_queues; 251 channels->rx_count = dev->real_num_rx_queues; 252 channels->max_tx = dev->num_tx_queues; 253 channels->max_rx = dev->num_rx_queues; 254 } 255 256 static int veth_set_channels(struct net_device *dev, 257 struct ethtool_channels *ch); 258 259 static const struct ethtool_ops veth_ethtool_ops = { 260 .get_drvinfo = veth_get_drvinfo, 261 .get_link = ethtool_op_get_link, 262 .get_strings = veth_get_strings, 263 .get_sset_count = veth_get_sset_count, 264 .get_ethtool_stats = veth_get_ethtool_stats, 265 .get_link_ksettings = veth_get_link_ksettings, 266 .get_ts_info = ethtool_op_get_ts_info, 267 .get_channels = veth_get_channels, 268 .set_channels = veth_set_channels, 269 }; 270 271 /* general routines */ 272 273 static bool veth_is_xdp_frame(void *ptr) 274 { 275 return (unsigned long)ptr & VETH_XDP_FLAG; 276 } 277 278 static struct xdp_frame *veth_ptr_to_xdp(void *ptr) 279 { 280 return (void *)((unsigned long)ptr & ~VETH_XDP_FLAG); 281 } 282 283 static void *veth_xdp_to_ptr(struct xdp_frame *xdp) 284 { 285 return (void *)((unsigned long)xdp | VETH_XDP_FLAG); 286 } 287 288 static void veth_ptr_free(void *ptr) 289 { 290 if (veth_is_xdp_frame(ptr)) 291 xdp_return_frame(veth_ptr_to_xdp(ptr)); 292 else 293 kfree_skb(ptr); 294 } 295 296 static void __veth_xdp_flush(struct veth_rq *rq) 297 { 298 /* Write ptr_ring before reading rx_notify_masked */ 299 smp_mb(); 300 if (!READ_ONCE(rq->rx_notify_masked) && 301 napi_schedule_prep(&rq->xdp_napi)) { 302 WRITE_ONCE(rq->rx_notify_masked, true); 303 __napi_schedule(&rq->xdp_napi); 304 } 305 } 306 307 static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb) 308 { 309 if (unlikely(ptr_ring_produce(&rq->xdp_ring, skb))) { 310 dev_kfree_skb_any(skb); 311 return NET_RX_DROP; 312 } 313 314 return NET_RX_SUCCESS; 315 } 316 317 static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb, 318 struct veth_rq *rq, bool xdp) 319 { 320 return __dev_forward_skb(dev, skb) ?: xdp ? 321 veth_xdp_rx(rq, skb) : 322 __netif_rx(skb); 323 } 324 325 /* return true if the specified skb has chances of GRO aggregation 326 * Don't strive for accuracy, but try to avoid GRO overhead in the most 327 * common scenarios. 328 * When XDP is enabled, all traffic is considered eligible, as the xmit 329 * device has TSO off. 330 * When TSO is enabled on the xmit device, we are likely interested only 331 * in UDP aggregation, explicitly check for that if the skb is suspected 332 * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets - 333 * to belong to locally generated UDP traffic. 334 */ 335 static bool veth_skb_is_eligible_for_gro(const struct net_device *dev, 336 const struct net_device *rcv, 337 const struct sk_buff *skb) 338 { 339 return !(dev->features & NETIF_F_ALL_TSO) || 340 (skb->destructor == sock_wfree && 341 rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD)); 342 } 343 344 static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev) 345 { 346 struct veth_priv *rcv_priv, *priv = netdev_priv(dev); 347 struct veth_rq *rq = NULL; 348 int ret = NETDEV_TX_OK; 349 struct net_device *rcv; 350 int length = skb->len; 351 bool use_napi = false; 352 int rxq; 353 354 rcu_read_lock(); 355 rcv = rcu_dereference(priv->peer); 356 if (unlikely(!rcv) || !pskb_may_pull(skb, ETH_HLEN)) { 357 kfree_skb(skb); 358 goto drop; 359 } 360 361 rcv_priv = netdev_priv(rcv); 362 rxq = skb_get_queue_mapping(skb); 363 if (rxq < rcv->real_num_rx_queues) { 364 rq = &rcv_priv->rq[rxq]; 365 366 /* The napi pointer is available when an XDP program is 367 * attached or when GRO is enabled 368 * Don't bother with napi/GRO if the skb can't be aggregated 369 */ 370 use_napi = rcu_access_pointer(rq->napi) && 371 veth_skb_is_eligible_for_gro(dev, rcv, skb); 372 } 373 374 skb_tx_timestamp(skb); 375 if (likely(veth_forward_skb(rcv, skb, rq, use_napi) == NET_RX_SUCCESS)) { 376 if (!use_napi) 377 dev_sw_netstats_tx_add(dev, 1, length); 378 else 379 __veth_xdp_flush(rq); 380 } else { 381 drop: 382 atomic64_inc(&priv->dropped); 383 ret = NET_XMIT_DROP; 384 } 385 386 rcu_read_unlock(); 387 388 return ret; 389 } 390 391 static void veth_stats_rx(struct veth_stats *result, struct net_device *dev) 392 { 393 struct veth_priv *priv = netdev_priv(dev); 394 int i; 395 396 result->peer_tq_xdp_xmit_err = 0; 397 result->xdp_packets = 0; 398 result->xdp_tx_err = 0; 399 result->xdp_bytes = 0; 400 result->rx_drops = 0; 401 for (i = 0; i < dev->num_rx_queues; i++) { 402 u64 packets, bytes, drops, xdp_tx_err, peer_tq_xdp_xmit_err; 403 struct veth_rq_stats *stats = &priv->rq[i].stats; 404 unsigned int start; 405 406 do { 407 start = u64_stats_fetch_begin(&stats->syncp); 408 peer_tq_xdp_xmit_err = stats->vs.peer_tq_xdp_xmit_err; 409 xdp_tx_err = stats->vs.xdp_tx_err; 410 packets = stats->vs.xdp_packets; 411 bytes = stats->vs.xdp_bytes; 412 drops = stats->vs.rx_drops; 413 } while (u64_stats_fetch_retry(&stats->syncp, start)); 414 result->peer_tq_xdp_xmit_err += peer_tq_xdp_xmit_err; 415 result->xdp_tx_err += xdp_tx_err; 416 result->xdp_packets += packets; 417 result->xdp_bytes += bytes; 418 result->rx_drops += drops; 419 } 420 } 421 422 static void veth_get_stats64(struct net_device *dev, 423 struct rtnl_link_stats64 *tot) 424 { 425 struct veth_priv *priv = netdev_priv(dev); 426 struct net_device *peer; 427 struct veth_stats rx; 428 429 tot->tx_dropped = atomic64_read(&priv->dropped); 430 dev_fetch_sw_netstats(tot, dev->tstats); 431 432 veth_stats_rx(&rx, dev); 433 tot->tx_dropped += rx.xdp_tx_err; 434 tot->rx_dropped = rx.rx_drops + rx.peer_tq_xdp_xmit_err; 435 tot->rx_bytes += rx.xdp_bytes; 436 tot->rx_packets += rx.xdp_packets; 437 438 rcu_read_lock(); 439 peer = rcu_dereference(priv->peer); 440 if (peer) { 441 struct rtnl_link_stats64 tot_peer = {}; 442 443 dev_fetch_sw_netstats(&tot_peer, peer->tstats); 444 tot->rx_bytes += tot_peer.tx_bytes; 445 tot->rx_packets += tot_peer.tx_packets; 446 447 veth_stats_rx(&rx, peer); 448 tot->tx_dropped += rx.peer_tq_xdp_xmit_err; 449 tot->rx_dropped += rx.xdp_tx_err; 450 tot->tx_bytes += rx.xdp_bytes; 451 tot->tx_packets += rx.xdp_packets; 452 } 453 rcu_read_unlock(); 454 } 455 456 /* fake multicast ability */ 457 static void veth_set_multicast_list(struct net_device *dev) 458 { 459 } 460 461 static int veth_select_rxq(struct net_device *dev) 462 { 463 return smp_processor_id() % dev->real_num_rx_queues; 464 } 465 466 static struct net_device *veth_peer_dev(struct net_device *dev) 467 { 468 struct veth_priv *priv = netdev_priv(dev); 469 470 /* Callers must be under RCU read side. */ 471 return rcu_dereference(priv->peer); 472 } 473 474 static int veth_xdp_xmit(struct net_device *dev, int n, 475 struct xdp_frame **frames, 476 u32 flags, bool ndo_xmit) 477 { 478 struct veth_priv *rcv_priv, *priv = netdev_priv(dev); 479 int i, ret = -ENXIO, nxmit = 0; 480 struct net_device *rcv; 481 unsigned int max_len; 482 struct veth_rq *rq; 483 484 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) 485 return -EINVAL; 486 487 rcu_read_lock(); 488 rcv = rcu_dereference(priv->peer); 489 if (unlikely(!rcv)) 490 goto out; 491 492 rcv_priv = netdev_priv(rcv); 493 rq = &rcv_priv->rq[veth_select_rxq(rcv)]; 494 /* The napi pointer is set if NAPI is enabled, which ensures that 495 * xdp_ring is initialized on receive side and the peer device is up. 496 */ 497 if (!rcu_access_pointer(rq->napi)) 498 goto out; 499 500 max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN; 501 502 spin_lock(&rq->xdp_ring.producer_lock); 503 for (i = 0; i < n; i++) { 504 struct xdp_frame *frame = frames[i]; 505 void *ptr = veth_xdp_to_ptr(frame); 506 507 if (unlikely(xdp_get_frame_len(frame) > max_len || 508 __ptr_ring_produce(&rq->xdp_ring, ptr))) 509 break; 510 nxmit++; 511 } 512 spin_unlock(&rq->xdp_ring.producer_lock); 513 514 if (flags & XDP_XMIT_FLUSH) 515 __veth_xdp_flush(rq); 516 517 ret = nxmit; 518 if (ndo_xmit) { 519 u64_stats_update_begin(&rq->stats.syncp); 520 rq->stats.vs.peer_tq_xdp_xmit += nxmit; 521 rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit; 522 u64_stats_update_end(&rq->stats.syncp); 523 } 524 525 out: 526 rcu_read_unlock(); 527 528 return ret; 529 } 530 531 static int veth_ndo_xdp_xmit(struct net_device *dev, int n, 532 struct xdp_frame **frames, u32 flags) 533 { 534 int err; 535 536 err = veth_xdp_xmit(dev, n, frames, flags, true); 537 if (err < 0) { 538 struct veth_priv *priv = netdev_priv(dev); 539 540 atomic64_add(n, &priv->dropped); 541 } 542 543 return err; 544 } 545 546 static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq) 547 { 548 int sent, i, err = 0, drops; 549 550 sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false); 551 if (sent < 0) { 552 err = sent; 553 sent = 0; 554 } 555 556 for (i = sent; unlikely(i < bq->count); i++) 557 xdp_return_frame(bq->q[i]); 558 559 drops = bq->count - sent; 560 trace_xdp_bulk_tx(rq->dev, sent, drops, err); 561 562 u64_stats_update_begin(&rq->stats.syncp); 563 rq->stats.vs.xdp_tx += sent; 564 rq->stats.vs.xdp_tx_err += drops; 565 u64_stats_update_end(&rq->stats.syncp); 566 567 bq->count = 0; 568 } 569 570 static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq) 571 { 572 struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev); 573 struct net_device *rcv; 574 struct veth_rq *rcv_rq; 575 576 rcu_read_lock(); 577 veth_xdp_flush_bq(rq, bq); 578 rcv = rcu_dereference(priv->peer); 579 if (unlikely(!rcv)) 580 goto out; 581 582 rcv_priv = netdev_priv(rcv); 583 rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)]; 584 /* xdp_ring is initialized on receive side? */ 585 if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog))) 586 goto out; 587 588 __veth_xdp_flush(rcv_rq); 589 out: 590 rcu_read_unlock(); 591 } 592 593 static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp, 594 struct veth_xdp_tx_bq *bq) 595 { 596 struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp); 597 598 if (unlikely(!frame)) 599 return -EOVERFLOW; 600 601 if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE)) 602 veth_xdp_flush_bq(rq, bq); 603 604 bq->q[bq->count++] = frame; 605 606 return 0; 607 } 608 609 static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq, 610 struct xdp_frame *frame, 611 struct veth_xdp_tx_bq *bq, 612 struct veth_stats *stats) 613 { 614 struct xdp_frame orig_frame; 615 struct bpf_prog *xdp_prog; 616 617 rcu_read_lock(); 618 xdp_prog = rcu_dereference(rq->xdp_prog); 619 if (likely(xdp_prog)) { 620 struct veth_xdp_buff vxbuf; 621 struct xdp_buff *xdp = &vxbuf.xdp; 622 u32 act; 623 624 xdp_convert_frame_to_buff(frame, xdp); 625 xdp->rxq = &rq->xdp_rxq; 626 vxbuf.skb = NULL; 627 628 act = bpf_prog_run_xdp(xdp_prog, xdp); 629 630 switch (act) { 631 case XDP_PASS: 632 if (xdp_update_frame_from_buff(xdp, frame)) 633 goto err_xdp; 634 break; 635 case XDP_TX: 636 orig_frame = *frame; 637 xdp->rxq->mem = frame->mem; 638 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) { 639 trace_xdp_exception(rq->dev, xdp_prog, act); 640 frame = &orig_frame; 641 stats->rx_drops++; 642 goto err_xdp; 643 } 644 stats->xdp_tx++; 645 rcu_read_unlock(); 646 goto xdp_xmit; 647 case XDP_REDIRECT: 648 orig_frame = *frame; 649 xdp->rxq->mem = frame->mem; 650 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) { 651 frame = &orig_frame; 652 stats->rx_drops++; 653 goto err_xdp; 654 } 655 stats->xdp_redirect++; 656 rcu_read_unlock(); 657 goto xdp_xmit; 658 default: 659 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act); 660 fallthrough; 661 case XDP_ABORTED: 662 trace_xdp_exception(rq->dev, xdp_prog, act); 663 fallthrough; 664 case XDP_DROP: 665 stats->xdp_drops++; 666 goto err_xdp; 667 } 668 } 669 rcu_read_unlock(); 670 671 return frame; 672 err_xdp: 673 rcu_read_unlock(); 674 xdp_return_frame(frame); 675 xdp_xmit: 676 return NULL; 677 } 678 679 /* frames array contains VETH_XDP_BATCH at most */ 680 static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames, 681 int n_xdpf, struct veth_xdp_tx_bq *bq, 682 struct veth_stats *stats) 683 { 684 void *skbs[VETH_XDP_BATCH]; 685 int i; 686 687 if (xdp_alloc_skb_bulk(skbs, n_xdpf, 688 GFP_ATOMIC | __GFP_ZERO) < 0) { 689 for (i = 0; i < n_xdpf; i++) 690 xdp_return_frame(frames[i]); 691 stats->rx_drops += n_xdpf; 692 693 return; 694 } 695 696 for (i = 0; i < n_xdpf; i++) { 697 struct sk_buff *skb = skbs[i]; 698 699 skb = __xdp_build_skb_from_frame(frames[i], skb, 700 rq->dev); 701 if (!skb) { 702 xdp_return_frame(frames[i]); 703 stats->rx_drops++; 704 continue; 705 } 706 napi_gro_receive(&rq->xdp_napi, skb); 707 } 708 } 709 710 static void veth_xdp_get(struct xdp_buff *xdp) 711 { 712 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp); 713 int i; 714 715 get_page(virt_to_page(xdp->data)); 716 if (likely(!xdp_buff_has_frags(xdp))) 717 return; 718 719 for (i = 0; i < sinfo->nr_frags; i++) 720 __skb_frag_ref(&sinfo->frags[i]); 721 } 722 723 static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq, 724 struct xdp_buff *xdp, 725 struct sk_buff **pskb) 726 { 727 struct sk_buff *skb = *pskb; 728 u32 frame_sz; 729 730 if (skb_shared(skb) || skb_head_is_locked(skb) || 731 skb_shinfo(skb)->nr_frags || 732 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 733 if (skb_pp_cow_data(rq->page_pool, pskb, XDP_PACKET_HEADROOM)) 734 goto drop; 735 736 skb = *pskb; 737 } 738 739 /* SKB "head" area always have tailroom for skb_shared_info */ 740 frame_sz = skb_end_pointer(skb) - skb->head; 741 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 742 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq); 743 xdp_prepare_buff(xdp, skb->head, skb_headroom(skb), 744 skb_headlen(skb), true); 745 746 if (skb_is_nonlinear(skb)) { 747 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 748 xdp_buff_set_frags_flag(xdp); 749 } else { 750 xdp_buff_clear_frags_flag(xdp); 751 } 752 *pskb = skb; 753 754 return 0; 755 drop: 756 consume_skb(skb); 757 *pskb = NULL; 758 759 return -ENOMEM; 760 } 761 762 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq, 763 struct sk_buff *skb, 764 struct veth_xdp_tx_bq *bq, 765 struct veth_stats *stats) 766 { 767 void *orig_data, *orig_data_end; 768 struct bpf_prog *xdp_prog; 769 struct veth_xdp_buff vxbuf; 770 struct xdp_buff *xdp = &vxbuf.xdp; 771 u32 act, metalen; 772 int off; 773 774 skb_prepare_for_gro(skb); 775 776 rcu_read_lock(); 777 xdp_prog = rcu_dereference(rq->xdp_prog); 778 if (unlikely(!xdp_prog)) { 779 rcu_read_unlock(); 780 goto out; 781 } 782 783 __skb_push(skb, skb->data - skb_mac_header(skb)); 784 if (veth_convert_skb_to_xdp_buff(rq, xdp, &skb)) 785 goto drop; 786 vxbuf.skb = skb; 787 788 orig_data = xdp->data; 789 orig_data_end = xdp->data_end; 790 791 act = bpf_prog_run_xdp(xdp_prog, xdp); 792 793 switch (act) { 794 case XDP_PASS: 795 break; 796 case XDP_TX: 797 veth_xdp_get(xdp); 798 consume_skb(skb); 799 xdp->rxq->mem = rq->xdp_mem; 800 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) { 801 trace_xdp_exception(rq->dev, xdp_prog, act); 802 stats->rx_drops++; 803 goto err_xdp; 804 } 805 stats->xdp_tx++; 806 rcu_read_unlock(); 807 goto xdp_xmit; 808 case XDP_REDIRECT: 809 veth_xdp_get(xdp); 810 consume_skb(skb); 811 xdp->rxq->mem = rq->xdp_mem; 812 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) { 813 stats->rx_drops++; 814 goto err_xdp; 815 } 816 stats->xdp_redirect++; 817 rcu_read_unlock(); 818 goto xdp_xmit; 819 default: 820 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act); 821 fallthrough; 822 case XDP_ABORTED: 823 trace_xdp_exception(rq->dev, xdp_prog, act); 824 fallthrough; 825 case XDP_DROP: 826 stats->xdp_drops++; 827 goto xdp_drop; 828 } 829 rcu_read_unlock(); 830 831 /* check if bpf_xdp_adjust_head was used */ 832 off = orig_data - xdp->data; 833 if (off > 0) 834 __skb_push(skb, off); 835 else if (off < 0) 836 __skb_pull(skb, -off); 837 838 skb_reset_mac_header(skb); 839 840 /* check if bpf_xdp_adjust_tail was used */ 841 off = xdp->data_end - orig_data_end; 842 if (off != 0) 843 __skb_put(skb, off); /* positive on grow, negative on shrink */ 844 845 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 846 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 847 */ 848 if (xdp_buff_has_frags(xdp)) 849 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 850 else 851 skb->data_len = 0; 852 853 skb->protocol = eth_type_trans(skb, rq->dev); 854 855 metalen = xdp->data - xdp->data_meta; 856 if (metalen) 857 skb_metadata_set(skb, metalen); 858 out: 859 return skb; 860 drop: 861 stats->rx_drops++; 862 xdp_drop: 863 rcu_read_unlock(); 864 kfree_skb(skb); 865 return NULL; 866 err_xdp: 867 rcu_read_unlock(); 868 xdp_return_buff(xdp); 869 xdp_xmit: 870 return NULL; 871 } 872 873 static int veth_xdp_rcv(struct veth_rq *rq, int budget, 874 struct veth_xdp_tx_bq *bq, 875 struct veth_stats *stats) 876 { 877 int i, done = 0, n_xdpf = 0; 878 void *xdpf[VETH_XDP_BATCH]; 879 880 for (i = 0; i < budget; i++) { 881 void *ptr = __ptr_ring_consume(&rq->xdp_ring); 882 883 if (!ptr) 884 break; 885 886 if (veth_is_xdp_frame(ptr)) { 887 /* ndo_xdp_xmit */ 888 struct xdp_frame *frame = veth_ptr_to_xdp(ptr); 889 890 stats->xdp_bytes += xdp_get_frame_len(frame); 891 frame = veth_xdp_rcv_one(rq, frame, bq, stats); 892 if (frame) { 893 /* XDP_PASS */ 894 xdpf[n_xdpf++] = frame; 895 if (n_xdpf == VETH_XDP_BATCH) { 896 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, 897 bq, stats); 898 n_xdpf = 0; 899 } 900 } 901 } else { 902 /* ndo_start_xmit */ 903 struct sk_buff *skb = ptr; 904 905 stats->xdp_bytes += skb->len; 906 skb = veth_xdp_rcv_skb(rq, skb, bq, stats); 907 if (skb) { 908 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC)) 909 netif_receive_skb(skb); 910 else 911 napi_gro_receive(&rq->xdp_napi, skb); 912 } 913 } 914 done++; 915 } 916 917 if (n_xdpf) 918 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats); 919 920 u64_stats_update_begin(&rq->stats.syncp); 921 rq->stats.vs.xdp_redirect += stats->xdp_redirect; 922 rq->stats.vs.xdp_bytes += stats->xdp_bytes; 923 rq->stats.vs.xdp_drops += stats->xdp_drops; 924 rq->stats.vs.rx_drops += stats->rx_drops; 925 rq->stats.vs.xdp_packets += done; 926 u64_stats_update_end(&rq->stats.syncp); 927 928 return done; 929 } 930 931 static int veth_poll(struct napi_struct *napi, int budget) 932 { 933 struct veth_rq *rq = 934 container_of(napi, struct veth_rq, xdp_napi); 935 struct veth_stats stats = {}; 936 struct veth_xdp_tx_bq bq; 937 int done; 938 939 bq.count = 0; 940 941 xdp_set_return_frame_no_direct(); 942 done = veth_xdp_rcv(rq, budget, &bq, &stats); 943 944 if (stats.xdp_redirect > 0) 945 xdp_do_flush(); 946 947 if (done < budget && napi_complete_done(napi, done)) { 948 /* Write rx_notify_masked before reading ptr_ring */ 949 smp_store_mb(rq->rx_notify_masked, false); 950 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) { 951 if (napi_schedule_prep(&rq->xdp_napi)) { 952 WRITE_ONCE(rq->rx_notify_masked, true); 953 __napi_schedule(&rq->xdp_napi); 954 } 955 } 956 } 957 958 if (stats.xdp_tx > 0) 959 veth_xdp_flush(rq, &bq); 960 xdp_clear_return_frame_no_direct(); 961 962 return done; 963 } 964 965 static int veth_create_page_pool(struct veth_rq *rq) 966 { 967 struct page_pool_params pp_params = { 968 .order = 0, 969 .pool_size = VETH_RING_SIZE, 970 .nid = NUMA_NO_NODE, 971 .dev = &rq->dev->dev, 972 }; 973 974 rq->page_pool = page_pool_create(&pp_params); 975 if (IS_ERR(rq->page_pool)) { 976 int err = PTR_ERR(rq->page_pool); 977 978 rq->page_pool = NULL; 979 return err; 980 } 981 982 return 0; 983 } 984 985 static int __veth_napi_enable_range(struct net_device *dev, int start, int end) 986 { 987 struct veth_priv *priv = netdev_priv(dev); 988 int err, i; 989 990 for (i = start; i < end; i++) { 991 err = veth_create_page_pool(&priv->rq[i]); 992 if (err) 993 goto err_page_pool; 994 } 995 996 for (i = start; i < end; i++) { 997 struct veth_rq *rq = &priv->rq[i]; 998 999 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL); 1000 if (err) 1001 goto err_xdp_ring; 1002 } 1003 1004 for (i = start; i < end; i++) { 1005 struct veth_rq *rq = &priv->rq[i]; 1006 1007 napi_enable(&rq->xdp_napi); 1008 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi); 1009 } 1010 1011 return 0; 1012 1013 err_xdp_ring: 1014 for (i--; i >= start; i--) 1015 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free); 1016 i = end; 1017 err_page_pool: 1018 for (i--; i >= start; i--) { 1019 page_pool_destroy(priv->rq[i].page_pool); 1020 priv->rq[i].page_pool = NULL; 1021 } 1022 1023 return err; 1024 } 1025 1026 static int __veth_napi_enable(struct net_device *dev) 1027 { 1028 return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues); 1029 } 1030 1031 static void veth_napi_del_range(struct net_device *dev, int start, int end) 1032 { 1033 struct veth_priv *priv = netdev_priv(dev); 1034 int i; 1035 1036 for (i = start; i < end; i++) { 1037 struct veth_rq *rq = &priv->rq[i]; 1038 1039 rcu_assign_pointer(priv->rq[i].napi, NULL); 1040 napi_disable(&rq->xdp_napi); 1041 __netif_napi_del(&rq->xdp_napi); 1042 } 1043 synchronize_net(); 1044 1045 for (i = start; i < end; i++) { 1046 struct veth_rq *rq = &priv->rq[i]; 1047 1048 rq->rx_notify_masked = false; 1049 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free); 1050 } 1051 1052 for (i = start; i < end; i++) { 1053 page_pool_destroy(priv->rq[i].page_pool); 1054 priv->rq[i].page_pool = NULL; 1055 } 1056 } 1057 1058 static void veth_napi_del(struct net_device *dev) 1059 { 1060 veth_napi_del_range(dev, 0, dev->real_num_rx_queues); 1061 } 1062 1063 static bool veth_gro_requested(const struct net_device *dev) 1064 { 1065 return !!(dev->wanted_features & NETIF_F_GRO); 1066 } 1067 1068 static int veth_enable_xdp_range(struct net_device *dev, int start, int end, 1069 bool napi_already_on) 1070 { 1071 struct veth_priv *priv = netdev_priv(dev); 1072 int err, i; 1073 1074 for (i = start; i < end; i++) { 1075 struct veth_rq *rq = &priv->rq[i]; 1076 1077 if (!napi_already_on) 1078 netif_napi_add(dev, &rq->xdp_napi, veth_poll); 1079 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id); 1080 if (err < 0) 1081 goto err_rxq_reg; 1082 1083 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq, 1084 MEM_TYPE_PAGE_SHARED, 1085 NULL); 1086 if (err < 0) 1087 goto err_reg_mem; 1088 1089 /* Save original mem info as it can be overwritten */ 1090 rq->xdp_mem = rq->xdp_rxq.mem; 1091 } 1092 return 0; 1093 1094 err_reg_mem: 1095 xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq); 1096 err_rxq_reg: 1097 for (i--; i >= start; i--) { 1098 struct veth_rq *rq = &priv->rq[i]; 1099 1100 xdp_rxq_info_unreg(&rq->xdp_rxq); 1101 if (!napi_already_on) 1102 netif_napi_del(&rq->xdp_napi); 1103 } 1104 1105 return err; 1106 } 1107 1108 static void veth_disable_xdp_range(struct net_device *dev, int start, int end, 1109 bool delete_napi) 1110 { 1111 struct veth_priv *priv = netdev_priv(dev); 1112 int i; 1113 1114 for (i = start; i < end; i++) { 1115 struct veth_rq *rq = &priv->rq[i]; 1116 1117 rq->xdp_rxq.mem = rq->xdp_mem; 1118 xdp_rxq_info_unreg(&rq->xdp_rxq); 1119 1120 if (delete_napi) 1121 netif_napi_del(&rq->xdp_napi); 1122 } 1123 } 1124 1125 static int veth_enable_xdp(struct net_device *dev) 1126 { 1127 bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP); 1128 struct veth_priv *priv = netdev_priv(dev); 1129 int err, i; 1130 1131 if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) { 1132 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on); 1133 if (err) 1134 return err; 1135 1136 if (!napi_already_on) { 1137 err = __veth_napi_enable(dev); 1138 if (err) { 1139 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true); 1140 return err; 1141 } 1142 } 1143 } 1144 1145 for (i = 0; i < dev->real_num_rx_queues; i++) { 1146 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog); 1147 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi); 1148 } 1149 1150 return 0; 1151 } 1152 1153 static void veth_disable_xdp(struct net_device *dev) 1154 { 1155 struct veth_priv *priv = netdev_priv(dev); 1156 int i; 1157 1158 for (i = 0; i < dev->real_num_rx_queues; i++) 1159 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL); 1160 1161 if (!netif_running(dev) || !veth_gro_requested(dev)) 1162 veth_napi_del(dev); 1163 1164 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false); 1165 } 1166 1167 static int veth_napi_enable_range(struct net_device *dev, int start, int end) 1168 { 1169 struct veth_priv *priv = netdev_priv(dev); 1170 int err, i; 1171 1172 for (i = start; i < end; i++) { 1173 struct veth_rq *rq = &priv->rq[i]; 1174 1175 netif_napi_add(dev, &rq->xdp_napi, veth_poll); 1176 } 1177 1178 err = __veth_napi_enable_range(dev, start, end); 1179 if (err) { 1180 for (i = start; i < end; i++) { 1181 struct veth_rq *rq = &priv->rq[i]; 1182 1183 netif_napi_del(&rq->xdp_napi); 1184 } 1185 return err; 1186 } 1187 return err; 1188 } 1189 1190 static int veth_napi_enable(struct net_device *dev) 1191 { 1192 return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues); 1193 } 1194 1195 static void veth_disable_range_safe(struct net_device *dev, int start, int end) 1196 { 1197 struct veth_priv *priv = netdev_priv(dev); 1198 1199 if (start >= end) 1200 return; 1201 1202 if (priv->_xdp_prog) { 1203 veth_napi_del_range(dev, start, end); 1204 veth_disable_xdp_range(dev, start, end, false); 1205 } else if (veth_gro_requested(dev)) { 1206 veth_napi_del_range(dev, start, end); 1207 } 1208 } 1209 1210 static int veth_enable_range_safe(struct net_device *dev, int start, int end) 1211 { 1212 struct veth_priv *priv = netdev_priv(dev); 1213 int err; 1214 1215 if (start >= end) 1216 return 0; 1217 1218 if (priv->_xdp_prog) { 1219 /* these channels are freshly initialized, napi is not on there even 1220 * when GRO is requeste 1221 */ 1222 err = veth_enable_xdp_range(dev, start, end, false); 1223 if (err) 1224 return err; 1225 1226 err = __veth_napi_enable_range(dev, start, end); 1227 if (err) { 1228 /* on error always delete the newly added napis */ 1229 veth_disable_xdp_range(dev, start, end, true); 1230 return err; 1231 } 1232 } else if (veth_gro_requested(dev)) { 1233 return veth_napi_enable_range(dev, start, end); 1234 } 1235 return 0; 1236 } 1237 1238 static void veth_set_xdp_features(struct net_device *dev) 1239 { 1240 struct veth_priv *priv = netdev_priv(dev); 1241 struct net_device *peer; 1242 1243 peer = rtnl_dereference(priv->peer); 1244 if (peer && peer->real_num_tx_queues <= dev->real_num_rx_queues) { 1245 struct veth_priv *priv_peer = netdev_priv(peer); 1246 xdp_features_t val = NETDEV_XDP_ACT_BASIC | 1247 NETDEV_XDP_ACT_REDIRECT | 1248 NETDEV_XDP_ACT_RX_SG; 1249 1250 if (priv_peer->_xdp_prog || veth_gro_requested(peer)) 1251 val |= NETDEV_XDP_ACT_NDO_XMIT | 1252 NETDEV_XDP_ACT_NDO_XMIT_SG; 1253 xdp_set_features_flag(dev, val); 1254 } else { 1255 xdp_clear_features_flag(dev); 1256 } 1257 } 1258 1259 static int veth_set_channels(struct net_device *dev, 1260 struct ethtool_channels *ch) 1261 { 1262 struct veth_priv *priv = netdev_priv(dev); 1263 unsigned int old_rx_count, new_rx_count; 1264 struct veth_priv *peer_priv; 1265 struct net_device *peer; 1266 int err; 1267 1268 /* sanity check. Upper bounds are already enforced by the caller */ 1269 if (!ch->rx_count || !ch->tx_count) 1270 return -EINVAL; 1271 1272 /* avoid braking XDP, if that is enabled */ 1273 peer = rtnl_dereference(priv->peer); 1274 peer_priv = peer ? netdev_priv(peer) : NULL; 1275 if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues) 1276 return -EINVAL; 1277 1278 if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues) 1279 return -EINVAL; 1280 1281 old_rx_count = dev->real_num_rx_queues; 1282 new_rx_count = ch->rx_count; 1283 if (netif_running(dev)) { 1284 /* turn device off */ 1285 netif_carrier_off(dev); 1286 if (peer) 1287 netif_carrier_off(peer); 1288 1289 /* try to allocate new resurces, as needed*/ 1290 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count); 1291 if (err) 1292 goto out; 1293 } 1294 1295 err = netif_set_real_num_rx_queues(dev, ch->rx_count); 1296 if (err) 1297 goto revert; 1298 1299 err = netif_set_real_num_tx_queues(dev, ch->tx_count); 1300 if (err) { 1301 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count); 1302 1303 /* this error condition could happen only if rx and tx change 1304 * in opposite directions (e.g. tx nr raises, rx nr decreases) 1305 * and we can't do anything to fully restore the original 1306 * status 1307 */ 1308 if (err2) 1309 pr_warn("Can't restore rx queues config %d -> %d %d", 1310 new_rx_count, old_rx_count, err2); 1311 else 1312 goto revert; 1313 } 1314 1315 out: 1316 if (netif_running(dev)) { 1317 /* note that we need to swap the arguments WRT the enable part 1318 * to identify the range we have to disable 1319 */ 1320 veth_disable_range_safe(dev, new_rx_count, old_rx_count); 1321 netif_carrier_on(dev); 1322 if (peer) 1323 netif_carrier_on(peer); 1324 } 1325 1326 /* update XDP supported features */ 1327 veth_set_xdp_features(dev); 1328 if (peer) 1329 veth_set_xdp_features(peer); 1330 1331 return err; 1332 1333 revert: 1334 new_rx_count = old_rx_count; 1335 old_rx_count = ch->rx_count; 1336 goto out; 1337 } 1338 1339 static int veth_open(struct net_device *dev) 1340 { 1341 struct veth_priv *priv = netdev_priv(dev); 1342 struct net_device *peer = rtnl_dereference(priv->peer); 1343 int err; 1344 1345 if (!peer) 1346 return -ENOTCONN; 1347 1348 if (priv->_xdp_prog) { 1349 err = veth_enable_xdp(dev); 1350 if (err) 1351 return err; 1352 } else if (veth_gro_requested(dev)) { 1353 err = veth_napi_enable(dev); 1354 if (err) 1355 return err; 1356 } 1357 1358 if (peer->flags & IFF_UP) { 1359 netif_carrier_on(dev); 1360 netif_carrier_on(peer); 1361 } 1362 1363 veth_set_xdp_features(dev); 1364 1365 return 0; 1366 } 1367 1368 static int veth_close(struct net_device *dev) 1369 { 1370 struct veth_priv *priv = netdev_priv(dev); 1371 struct net_device *peer = rtnl_dereference(priv->peer); 1372 1373 netif_carrier_off(dev); 1374 if (peer) 1375 netif_carrier_off(peer); 1376 1377 if (priv->_xdp_prog) 1378 veth_disable_xdp(dev); 1379 else if (veth_gro_requested(dev)) 1380 veth_napi_del(dev); 1381 1382 return 0; 1383 } 1384 1385 static int is_valid_veth_mtu(int mtu) 1386 { 1387 return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU; 1388 } 1389 1390 static int veth_alloc_queues(struct net_device *dev) 1391 { 1392 struct veth_priv *priv = netdev_priv(dev); 1393 int i; 1394 1395 priv->rq = kvcalloc(dev->num_rx_queues, sizeof(*priv->rq), 1396 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 1397 if (!priv->rq) 1398 return -ENOMEM; 1399 1400 for (i = 0; i < dev->num_rx_queues; i++) { 1401 priv->rq[i].dev = dev; 1402 u64_stats_init(&priv->rq[i].stats.syncp); 1403 } 1404 1405 return 0; 1406 } 1407 1408 static void veth_free_queues(struct net_device *dev) 1409 { 1410 struct veth_priv *priv = netdev_priv(dev); 1411 1412 kvfree(priv->rq); 1413 } 1414 1415 static int veth_dev_init(struct net_device *dev) 1416 { 1417 netdev_lockdep_set_classes(dev); 1418 return veth_alloc_queues(dev); 1419 } 1420 1421 static void veth_dev_free(struct net_device *dev) 1422 { 1423 veth_free_queues(dev); 1424 } 1425 1426 #ifdef CONFIG_NET_POLL_CONTROLLER 1427 static void veth_poll_controller(struct net_device *dev) 1428 { 1429 /* veth only receives frames when its peer sends one 1430 * Since it has nothing to do with disabling irqs, we are guaranteed 1431 * never to have pending data when we poll for it so 1432 * there is nothing to do here. 1433 * 1434 * We need this though so netpoll recognizes us as an interface that 1435 * supports polling, which enables bridge devices in virt setups to 1436 * still use netconsole 1437 */ 1438 } 1439 #endif /* CONFIG_NET_POLL_CONTROLLER */ 1440 1441 static int veth_get_iflink(const struct net_device *dev) 1442 { 1443 struct veth_priv *priv = netdev_priv(dev); 1444 struct net_device *peer; 1445 int iflink; 1446 1447 rcu_read_lock(); 1448 peer = rcu_dereference(priv->peer); 1449 iflink = peer ? READ_ONCE(peer->ifindex) : 0; 1450 rcu_read_unlock(); 1451 1452 return iflink; 1453 } 1454 1455 static netdev_features_t veth_fix_features(struct net_device *dev, 1456 netdev_features_t features) 1457 { 1458 struct veth_priv *priv = netdev_priv(dev); 1459 struct net_device *peer; 1460 1461 peer = rtnl_dereference(priv->peer); 1462 if (peer) { 1463 struct veth_priv *peer_priv = netdev_priv(peer); 1464 1465 if (peer_priv->_xdp_prog) 1466 features &= ~NETIF_F_GSO_SOFTWARE; 1467 } 1468 1469 return features; 1470 } 1471 1472 static int veth_set_features(struct net_device *dev, 1473 netdev_features_t features) 1474 { 1475 netdev_features_t changed = features ^ dev->features; 1476 struct veth_priv *priv = netdev_priv(dev); 1477 struct net_device *peer; 1478 int err; 1479 1480 if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog) 1481 return 0; 1482 1483 peer = rtnl_dereference(priv->peer); 1484 if (features & NETIF_F_GRO) { 1485 err = veth_napi_enable(dev); 1486 if (err) 1487 return err; 1488 1489 if (peer) 1490 xdp_features_set_redirect_target(peer, true); 1491 } else { 1492 if (peer) 1493 xdp_features_clear_redirect_target(peer); 1494 veth_napi_del(dev); 1495 } 1496 return 0; 1497 } 1498 1499 static void veth_set_rx_headroom(struct net_device *dev, int new_hr) 1500 { 1501 struct veth_priv *peer_priv, *priv = netdev_priv(dev); 1502 struct net_device *peer; 1503 1504 if (new_hr < 0) 1505 new_hr = 0; 1506 1507 rcu_read_lock(); 1508 peer = rcu_dereference(priv->peer); 1509 if (unlikely(!peer)) 1510 goto out; 1511 1512 peer_priv = netdev_priv(peer); 1513 priv->requested_headroom = new_hr; 1514 new_hr = max(priv->requested_headroom, peer_priv->requested_headroom); 1515 dev->needed_headroom = new_hr; 1516 peer->needed_headroom = new_hr; 1517 1518 out: 1519 rcu_read_unlock(); 1520 } 1521 1522 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog, 1523 struct netlink_ext_ack *extack) 1524 { 1525 struct veth_priv *priv = netdev_priv(dev); 1526 struct bpf_prog *old_prog; 1527 struct net_device *peer; 1528 unsigned int max_mtu; 1529 int err; 1530 1531 old_prog = priv->_xdp_prog; 1532 priv->_xdp_prog = prog; 1533 peer = rtnl_dereference(priv->peer); 1534 1535 if (prog) { 1536 if (!peer) { 1537 NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached"); 1538 err = -ENOTCONN; 1539 goto err; 1540 } 1541 1542 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) - 1543 peer->hard_header_len; 1544 /* Allow increasing the max_mtu if the program supports 1545 * XDP fragments. 1546 */ 1547 if (prog->aux->xdp_has_frags) 1548 max_mtu += PAGE_SIZE * MAX_SKB_FRAGS; 1549 1550 if (peer->mtu > max_mtu) { 1551 NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP"); 1552 err = -ERANGE; 1553 goto err; 1554 } 1555 1556 if (dev->real_num_rx_queues < peer->real_num_tx_queues) { 1557 NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues"); 1558 err = -ENOSPC; 1559 goto err; 1560 } 1561 1562 if (dev->flags & IFF_UP) { 1563 err = veth_enable_xdp(dev); 1564 if (err) { 1565 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed"); 1566 goto err; 1567 } 1568 } 1569 1570 if (!old_prog) { 1571 peer->hw_features &= ~NETIF_F_GSO_SOFTWARE; 1572 peer->max_mtu = max_mtu; 1573 } 1574 1575 xdp_features_set_redirect_target(peer, true); 1576 } 1577 1578 if (old_prog) { 1579 if (!prog) { 1580 if (peer && !veth_gro_requested(dev)) 1581 xdp_features_clear_redirect_target(peer); 1582 1583 if (dev->flags & IFF_UP) 1584 veth_disable_xdp(dev); 1585 1586 if (peer) { 1587 peer->hw_features |= NETIF_F_GSO_SOFTWARE; 1588 peer->max_mtu = ETH_MAX_MTU; 1589 } 1590 } 1591 bpf_prog_put(old_prog); 1592 } 1593 1594 if ((!!old_prog ^ !!prog) && peer) 1595 netdev_update_features(peer); 1596 1597 return 0; 1598 err: 1599 priv->_xdp_prog = old_prog; 1600 1601 return err; 1602 } 1603 1604 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp) 1605 { 1606 switch (xdp->command) { 1607 case XDP_SETUP_PROG: 1608 return veth_xdp_set(dev, xdp->prog, xdp->extack); 1609 default: 1610 return -EINVAL; 1611 } 1612 } 1613 1614 static int veth_xdp_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp) 1615 { 1616 struct veth_xdp_buff *_ctx = (void *)ctx; 1617 1618 if (!_ctx->skb) 1619 return -ENODATA; 1620 1621 *timestamp = skb_hwtstamps(_ctx->skb)->hwtstamp; 1622 return 0; 1623 } 1624 1625 static int veth_xdp_rx_hash(const struct xdp_md *ctx, u32 *hash, 1626 enum xdp_rss_hash_type *rss_type) 1627 { 1628 struct veth_xdp_buff *_ctx = (void *)ctx; 1629 struct sk_buff *skb = _ctx->skb; 1630 1631 if (!skb) 1632 return -ENODATA; 1633 1634 *hash = skb_get_hash(skb); 1635 *rss_type = skb->l4_hash ? XDP_RSS_TYPE_L4_ANY : XDP_RSS_TYPE_NONE; 1636 1637 return 0; 1638 } 1639 1640 static int veth_xdp_rx_vlan_tag(const struct xdp_md *ctx, __be16 *vlan_proto, 1641 u16 *vlan_tci) 1642 { 1643 const struct veth_xdp_buff *_ctx = (void *)ctx; 1644 const struct sk_buff *skb = _ctx->skb; 1645 int err; 1646 1647 if (!skb) 1648 return -ENODATA; 1649 1650 err = __vlan_hwaccel_get_tag(skb, vlan_tci); 1651 if (err) 1652 return err; 1653 1654 *vlan_proto = skb->vlan_proto; 1655 return err; 1656 } 1657 1658 static const struct net_device_ops veth_netdev_ops = { 1659 .ndo_init = veth_dev_init, 1660 .ndo_open = veth_open, 1661 .ndo_stop = veth_close, 1662 .ndo_start_xmit = veth_xmit, 1663 .ndo_get_stats64 = veth_get_stats64, 1664 .ndo_set_rx_mode = veth_set_multicast_list, 1665 .ndo_set_mac_address = eth_mac_addr, 1666 #ifdef CONFIG_NET_POLL_CONTROLLER 1667 .ndo_poll_controller = veth_poll_controller, 1668 #endif 1669 .ndo_get_iflink = veth_get_iflink, 1670 .ndo_fix_features = veth_fix_features, 1671 .ndo_set_features = veth_set_features, 1672 .ndo_features_check = passthru_features_check, 1673 .ndo_set_rx_headroom = veth_set_rx_headroom, 1674 .ndo_bpf = veth_xdp, 1675 .ndo_xdp_xmit = veth_ndo_xdp_xmit, 1676 .ndo_get_peer_dev = veth_peer_dev, 1677 }; 1678 1679 static const struct xdp_metadata_ops veth_xdp_metadata_ops = { 1680 .xmo_rx_timestamp = veth_xdp_rx_timestamp, 1681 .xmo_rx_hash = veth_xdp_rx_hash, 1682 .xmo_rx_vlan_tag = veth_xdp_rx_vlan_tag, 1683 }; 1684 1685 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \ 1686 NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \ 1687 NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \ 1688 NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \ 1689 NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX ) 1690 1691 static void veth_setup(struct net_device *dev) 1692 { 1693 ether_setup(dev); 1694 1695 dev->priv_flags &= ~IFF_TX_SKB_SHARING; 1696 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE; 1697 dev->priv_flags |= IFF_NO_QUEUE; 1698 dev->priv_flags |= IFF_PHONY_HEADROOM; 1699 dev->priv_flags |= IFF_DISABLE_NETPOLL; 1700 dev->lltx = true; 1701 1702 dev->netdev_ops = &veth_netdev_ops; 1703 dev->xdp_metadata_ops = &veth_xdp_metadata_ops; 1704 dev->ethtool_ops = &veth_ethtool_ops; 1705 dev->features |= VETH_FEATURES; 1706 dev->vlan_features = dev->features & 1707 ~(NETIF_F_HW_VLAN_CTAG_TX | 1708 NETIF_F_HW_VLAN_STAG_TX | 1709 NETIF_F_HW_VLAN_CTAG_RX | 1710 NETIF_F_HW_VLAN_STAG_RX); 1711 dev->needs_free_netdev = true; 1712 dev->priv_destructor = veth_dev_free; 1713 dev->pcpu_stat_type = NETDEV_PCPU_STAT_TSTATS; 1714 dev->max_mtu = ETH_MAX_MTU; 1715 1716 dev->hw_features = VETH_FEATURES; 1717 dev->hw_enc_features = VETH_FEATURES; 1718 dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE; 1719 netif_set_tso_max_size(dev, GSO_MAX_SIZE); 1720 } 1721 1722 /* 1723 * netlink interface 1724 */ 1725 1726 static int veth_validate(struct nlattr *tb[], struct nlattr *data[], 1727 struct netlink_ext_ack *extack) 1728 { 1729 if (tb[IFLA_ADDRESS]) { 1730 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) 1731 return -EINVAL; 1732 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) 1733 return -EADDRNOTAVAIL; 1734 } 1735 if (tb[IFLA_MTU]) { 1736 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU]))) 1737 return -EINVAL; 1738 } 1739 return 0; 1740 } 1741 1742 static struct rtnl_link_ops veth_link_ops; 1743 1744 static void veth_disable_gro(struct net_device *dev) 1745 { 1746 dev->features &= ~NETIF_F_GRO; 1747 dev->wanted_features &= ~NETIF_F_GRO; 1748 netdev_update_features(dev); 1749 } 1750 1751 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[]) 1752 { 1753 int err; 1754 1755 if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) { 1756 err = netif_set_real_num_tx_queues(dev, 1); 1757 if (err) 1758 return err; 1759 } 1760 if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) { 1761 err = netif_set_real_num_rx_queues(dev, 1); 1762 if (err) 1763 return err; 1764 } 1765 return 0; 1766 } 1767 1768 static int veth_newlink(struct net *src_net, struct net_device *dev, 1769 struct nlattr *tb[], struct nlattr *data[], 1770 struct netlink_ext_ack *extack) 1771 { 1772 int err; 1773 struct net_device *peer; 1774 struct veth_priv *priv; 1775 char ifname[IFNAMSIZ]; 1776 struct nlattr *peer_tb[IFLA_MAX + 1], **tbp; 1777 unsigned char name_assign_type; 1778 struct ifinfomsg *ifmp; 1779 struct net *net; 1780 1781 /* 1782 * create and register peer first 1783 */ 1784 if (data != NULL && data[VETH_INFO_PEER] != NULL) { 1785 struct nlattr *nla_peer; 1786 1787 nla_peer = data[VETH_INFO_PEER]; 1788 ifmp = nla_data(nla_peer); 1789 err = rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack); 1790 if (err < 0) 1791 return err; 1792 1793 err = veth_validate(peer_tb, NULL, extack); 1794 if (err < 0) 1795 return err; 1796 1797 tbp = peer_tb; 1798 } else { 1799 ifmp = NULL; 1800 tbp = tb; 1801 } 1802 1803 if (ifmp && tbp[IFLA_IFNAME]) { 1804 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ); 1805 name_assign_type = NET_NAME_USER; 1806 } else { 1807 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d"); 1808 name_assign_type = NET_NAME_ENUM; 1809 } 1810 1811 net = rtnl_link_get_net(src_net, tbp); 1812 if (IS_ERR(net)) 1813 return PTR_ERR(net); 1814 1815 peer = rtnl_create_link(net, ifname, name_assign_type, 1816 &veth_link_ops, tbp, extack); 1817 if (IS_ERR(peer)) { 1818 put_net(net); 1819 return PTR_ERR(peer); 1820 } 1821 1822 if (!ifmp || !tbp[IFLA_ADDRESS]) 1823 eth_hw_addr_random(peer); 1824 1825 if (ifmp && (dev->ifindex != 0)) 1826 peer->ifindex = ifmp->ifi_index; 1827 1828 netif_inherit_tso_max(peer, dev); 1829 1830 err = register_netdevice(peer); 1831 put_net(net); 1832 net = NULL; 1833 if (err < 0) 1834 goto err_register_peer; 1835 1836 /* keep GRO disabled by default to be consistent with the established 1837 * veth behavior 1838 */ 1839 veth_disable_gro(peer); 1840 netif_carrier_off(peer); 1841 1842 err = rtnl_configure_link(peer, ifmp, 0, NULL); 1843 if (err < 0) 1844 goto err_configure_peer; 1845 1846 /* 1847 * register dev last 1848 * 1849 * note, that since we've registered new device the dev's name 1850 * should be re-allocated 1851 */ 1852 1853 if (tb[IFLA_ADDRESS] == NULL) 1854 eth_hw_addr_random(dev); 1855 1856 if (tb[IFLA_IFNAME]) 1857 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ); 1858 else 1859 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d"); 1860 1861 err = register_netdevice(dev); 1862 if (err < 0) 1863 goto err_register_dev; 1864 1865 netif_carrier_off(dev); 1866 1867 /* 1868 * tie the deviced together 1869 */ 1870 1871 priv = netdev_priv(dev); 1872 rcu_assign_pointer(priv->peer, peer); 1873 err = veth_init_queues(dev, tb); 1874 if (err) 1875 goto err_queues; 1876 1877 priv = netdev_priv(peer); 1878 rcu_assign_pointer(priv->peer, dev); 1879 err = veth_init_queues(peer, tb); 1880 if (err) 1881 goto err_queues; 1882 1883 veth_disable_gro(dev); 1884 /* update XDP supported features */ 1885 veth_set_xdp_features(dev); 1886 veth_set_xdp_features(peer); 1887 1888 return 0; 1889 1890 err_queues: 1891 unregister_netdevice(dev); 1892 err_register_dev: 1893 /* nothing to do */ 1894 err_configure_peer: 1895 unregister_netdevice(peer); 1896 return err; 1897 1898 err_register_peer: 1899 free_netdev(peer); 1900 return err; 1901 } 1902 1903 static void veth_dellink(struct net_device *dev, struct list_head *head) 1904 { 1905 struct veth_priv *priv; 1906 struct net_device *peer; 1907 1908 priv = netdev_priv(dev); 1909 peer = rtnl_dereference(priv->peer); 1910 1911 /* Note : dellink() is called from default_device_exit_batch(), 1912 * before a rcu_synchronize() point. The devices are guaranteed 1913 * not being freed before one RCU grace period. 1914 */ 1915 RCU_INIT_POINTER(priv->peer, NULL); 1916 unregister_netdevice_queue(dev, head); 1917 1918 if (peer) { 1919 priv = netdev_priv(peer); 1920 RCU_INIT_POINTER(priv->peer, NULL); 1921 unregister_netdevice_queue(peer, head); 1922 } 1923 } 1924 1925 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = { 1926 [VETH_INFO_PEER] = { .len = sizeof(struct ifinfomsg) }, 1927 }; 1928 1929 static struct net *veth_get_link_net(const struct net_device *dev) 1930 { 1931 struct veth_priv *priv = netdev_priv(dev); 1932 struct net_device *peer = rtnl_dereference(priv->peer); 1933 1934 return peer ? dev_net(peer) : dev_net(dev); 1935 } 1936 1937 static unsigned int veth_get_num_queues(void) 1938 { 1939 /* enforce the same queue limit as rtnl_create_link */ 1940 int queues = num_possible_cpus(); 1941 1942 if (queues > 4096) 1943 queues = 4096; 1944 return queues; 1945 } 1946 1947 static struct rtnl_link_ops veth_link_ops = { 1948 .kind = DRV_NAME, 1949 .priv_size = sizeof(struct veth_priv), 1950 .setup = veth_setup, 1951 .validate = veth_validate, 1952 .newlink = veth_newlink, 1953 .dellink = veth_dellink, 1954 .policy = veth_policy, 1955 .maxtype = VETH_INFO_MAX, 1956 .get_link_net = veth_get_link_net, 1957 .get_num_tx_queues = veth_get_num_queues, 1958 .get_num_rx_queues = veth_get_num_queues, 1959 }; 1960 1961 /* 1962 * init/fini 1963 */ 1964 1965 static __init int veth_init(void) 1966 { 1967 return rtnl_link_register(&veth_link_ops); 1968 } 1969 1970 static __exit void veth_exit(void) 1971 { 1972 rtnl_link_unregister(&veth_link_ops); 1973 } 1974 1975 module_init(veth_init); 1976 module_exit(veth_exit); 1977 1978 MODULE_DESCRIPTION("Virtual Ethernet Tunnel"); 1979 MODULE_LICENSE("GPL v2"); 1980 MODULE_ALIAS_RTNL_LINK(DRV_NAME); 1981