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 u32 size, len, max_head_size, off, truesize, page_offset; 733 struct sk_buff *nskb; 734 struct page *page; 735 int i, head_off; 736 void *va; 737 738 /* We need a private copy of the skb and data buffers since 739 * the ebpf program can modify it. We segment the original skb 740 * into order-0 pages without linearize it. 741 * 742 * Make sure we have enough space for linear and paged area 743 */ 744 max_head_size = SKB_WITH_OVERHEAD(PAGE_SIZE - 745 VETH_XDP_HEADROOM); 746 if (skb->len > PAGE_SIZE * MAX_SKB_FRAGS + max_head_size) 747 goto drop; 748 749 size = min_t(u32, skb->len, max_head_size); 750 truesize = SKB_HEAD_ALIGN(size) + VETH_XDP_HEADROOM; 751 752 /* Allocate skb head */ 753 va = page_pool_dev_alloc_va(rq->page_pool, &truesize); 754 if (!va) 755 goto drop; 756 757 nskb = napi_build_skb(va, truesize); 758 if (!nskb) { 759 page_pool_free_va(rq->page_pool, va, true); 760 goto drop; 761 } 762 763 skb_reserve(nskb, VETH_XDP_HEADROOM); 764 skb_copy_header(nskb, skb); 765 skb_mark_for_recycle(nskb); 766 767 if (skb_copy_bits(skb, 0, nskb->data, size)) { 768 consume_skb(nskb); 769 goto drop; 770 } 771 skb_put(nskb, size); 772 773 head_off = skb_headroom(nskb) - skb_headroom(skb); 774 skb_headers_offset_update(nskb, head_off); 775 776 /* Allocate paged area of new skb */ 777 off = size; 778 len = skb->len - off; 779 780 for (i = 0; i < MAX_SKB_FRAGS && off < skb->len; i++) { 781 size = min_t(u32, len, PAGE_SIZE); 782 truesize = size; 783 784 page = page_pool_dev_alloc(rq->page_pool, &page_offset, 785 &truesize); 786 if (!page) { 787 consume_skb(nskb); 788 goto drop; 789 } 790 791 skb_add_rx_frag(nskb, i, page, page_offset, size, 792 truesize); 793 if (skb_copy_bits(skb, off, page_address(page), 794 size)) { 795 consume_skb(nskb); 796 goto drop; 797 } 798 799 len -= size; 800 off += size; 801 } 802 803 consume_skb(skb); 804 skb = nskb; 805 } 806 807 /* SKB "head" area always have tailroom for skb_shared_info */ 808 frame_sz = skb_end_pointer(skb) - skb->head; 809 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 810 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq); 811 xdp_prepare_buff(xdp, skb->head, skb_headroom(skb), 812 skb_headlen(skb), true); 813 814 if (skb_is_nonlinear(skb)) { 815 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 816 xdp_buff_set_frags_flag(xdp); 817 } else { 818 xdp_buff_clear_frags_flag(xdp); 819 } 820 *pskb = skb; 821 822 return 0; 823 drop: 824 consume_skb(skb); 825 *pskb = NULL; 826 827 return -ENOMEM; 828 } 829 830 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq, 831 struct sk_buff *skb, 832 struct veth_xdp_tx_bq *bq, 833 struct veth_stats *stats) 834 { 835 void *orig_data, *orig_data_end; 836 struct bpf_prog *xdp_prog; 837 struct veth_xdp_buff vxbuf; 838 struct xdp_buff *xdp = &vxbuf.xdp; 839 u32 act, metalen; 840 int off; 841 842 skb_prepare_for_gro(skb); 843 844 rcu_read_lock(); 845 xdp_prog = rcu_dereference(rq->xdp_prog); 846 if (unlikely(!xdp_prog)) { 847 rcu_read_unlock(); 848 goto out; 849 } 850 851 __skb_push(skb, skb->data - skb_mac_header(skb)); 852 if (veth_convert_skb_to_xdp_buff(rq, xdp, &skb)) 853 goto drop; 854 vxbuf.skb = skb; 855 856 orig_data = xdp->data; 857 orig_data_end = xdp->data_end; 858 859 act = bpf_prog_run_xdp(xdp_prog, xdp); 860 861 switch (act) { 862 case XDP_PASS: 863 break; 864 case XDP_TX: 865 veth_xdp_get(xdp); 866 consume_skb(skb); 867 xdp->rxq->mem = rq->xdp_mem; 868 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) { 869 trace_xdp_exception(rq->dev, xdp_prog, act); 870 stats->rx_drops++; 871 goto err_xdp; 872 } 873 stats->xdp_tx++; 874 rcu_read_unlock(); 875 goto xdp_xmit; 876 case XDP_REDIRECT: 877 veth_xdp_get(xdp); 878 consume_skb(skb); 879 xdp->rxq->mem = rq->xdp_mem; 880 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) { 881 stats->rx_drops++; 882 goto err_xdp; 883 } 884 stats->xdp_redirect++; 885 rcu_read_unlock(); 886 goto xdp_xmit; 887 default: 888 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act); 889 fallthrough; 890 case XDP_ABORTED: 891 trace_xdp_exception(rq->dev, xdp_prog, act); 892 fallthrough; 893 case XDP_DROP: 894 stats->xdp_drops++; 895 goto xdp_drop; 896 } 897 rcu_read_unlock(); 898 899 /* check if bpf_xdp_adjust_head was used */ 900 off = orig_data - xdp->data; 901 if (off > 0) 902 __skb_push(skb, off); 903 else if (off < 0) 904 __skb_pull(skb, -off); 905 906 skb_reset_mac_header(skb); 907 908 /* check if bpf_xdp_adjust_tail was used */ 909 off = xdp->data_end - orig_data_end; 910 if (off != 0) 911 __skb_put(skb, off); /* positive on grow, negative on shrink */ 912 913 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 914 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 915 */ 916 if (xdp_buff_has_frags(xdp)) 917 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 918 else 919 skb->data_len = 0; 920 921 skb->protocol = eth_type_trans(skb, rq->dev); 922 923 metalen = xdp->data - xdp->data_meta; 924 if (metalen) 925 skb_metadata_set(skb, metalen); 926 out: 927 return skb; 928 drop: 929 stats->rx_drops++; 930 xdp_drop: 931 rcu_read_unlock(); 932 kfree_skb(skb); 933 return NULL; 934 err_xdp: 935 rcu_read_unlock(); 936 xdp_return_buff(xdp); 937 xdp_xmit: 938 return NULL; 939 } 940 941 static int veth_xdp_rcv(struct veth_rq *rq, int budget, 942 struct veth_xdp_tx_bq *bq, 943 struct veth_stats *stats) 944 { 945 int i, done = 0, n_xdpf = 0; 946 void *xdpf[VETH_XDP_BATCH]; 947 948 for (i = 0; i < budget; i++) { 949 void *ptr = __ptr_ring_consume(&rq->xdp_ring); 950 951 if (!ptr) 952 break; 953 954 if (veth_is_xdp_frame(ptr)) { 955 /* ndo_xdp_xmit */ 956 struct xdp_frame *frame = veth_ptr_to_xdp(ptr); 957 958 stats->xdp_bytes += xdp_get_frame_len(frame); 959 frame = veth_xdp_rcv_one(rq, frame, bq, stats); 960 if (frame) { 961 /* XDP_PASS */ 962 xdpf[n_xdpf++] = frame; 963 if (n_xdpf == VETH_XDP_BATCH) { 964 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, 965 bq, stats); 966 n_xdpf = 0; 967 } 968 } 969 } else { 970 /* ndo_start_xmit */ 971 struct sk_buff *skb = ptr; 972 973 stats->xdp_bytes += skb->len; 974 skb = veth_xdp_rcv_skb(rq, skb, bq, stats); 975 if (skb) { 976 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC)) 977 netif_receive_skb(skb); 978 else 979 napi_gro_receive(&rq->xdp_napi, skb); 980 } 981 } 982 done++; 983 } 984 985 if (n_xdpf) 986 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats); 987 988 u64_stats_update_begin(&rq->stats.syncp); 989 rq->stats.vs.xdp_redirect += stats->xdp_redirect; 990 rq->stats.vs.xdp_bytes += stats->xdp_bytes; 991 rq->stats.vs.xdp_drops += stats->xdp_drops; 992 rq->stats.vs.rx_drops += stats->rx_drops; 993 rq->stats.vs.xdp_packets += done; 994 u64_stats_update_end(&rq->stats.syncp); 995 996 return done; 997 } 998 999 static int veth_poll(struct napi_struct *napi, int budget) 1000 { 1001 struct veth_rq *rq = 1002 container_of(napi, struct veth_rq, xdp_napi); 1003 struct veth_stats stats = {}; 1004 struct veth_xdp_tx_bq bq; 1005 int done; 1006 1007 bq.count = 0; 1008 1009 xdp_set_return_frame_no_direct(); 1010 done = veth_xdp_rcv(rq, budget, &bq, &stats); 1011 1012 if (stats.xdp_redirect > 0) 1013 xdp_do_flush(); 1014 1015 if (done < budget && napi_complete_done(napi, done)) { 1016 /* Write rx_notify_masked before reading ptr_ring */ 1017 smp_store_mb(rq->rx_notify_masked, false); 1018 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) { 1019 if (napi_schedule_prep(&rq->xdp_napi)) { 1020 WRITE_ONCE(rq->rx_notify_masked, true); 1021 __napi_schedule(&rq->xdp_napi); 1022 } 1023 } 1024 } 1025 1026 if (stats.xdp_tx > 0) 1027 veth_xdp_flush(rq, &bq); 1028 xdp_clear_return_frame_no_direct(); 1029 1030 return done; 1031 } 1032 1033 static int veth_create_page_pool(struct veth_rq *rq) 1034 { 1035 struct page_pool_params pp_params = { 1036 .order = 0, 1037 .pool_size = VETH_RING_SIZE, 1038 .nid = NUMA_NO_NODE, 1039 .dev = &rq->dev->dev, 1040 }; 1041 1042 rq->page_pool = page_pool_create(&pp_params); 1043 if (IS_ERR(rq->page_pool)) { 1044 int err = PTR_ERR(rq->page_pool); 1045 1046 rq->page_pool = NULL; 1047 return err; 1048 } 1049 1050 return 0; 1051 } 1052 1053 static int __veth_napi_enable_range(struct net_device *dev, int start, int end) 1054 { 1055 struct veth_priv *priv = netdev_priv(dev); 1056 int err, i; 1057 1058 for (i = start; i < end; i++) { 1059 err = veth_create_page_pool(&priv->rq[i]); 1060 if (err) 1061 goto err_page_pool; 1062 } 1063 1064 for (i = start; i < end; i++) { 1065 struct veth_rq *rq = &priv->rq[i]; 1066 1067 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL); 1068 if (err) 1069 goto err_xdp_ring; 1070 } 1071 1072 for (i = start; i < end; i++) { 1073 struct veth_rq *rq = &priv->rq[i]; 1074 1075 napi_enable(&rq->xdp_napi); 1076 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi); 1077 } 1078 1079 return 0; 1080 1081 err_xdp_ring: 1082 for (i--; i >= start; i--) 1083 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free); 1084 i = end; 1085 err_page_pool: 1086 for (i--; i >= start; i--) { 1087 page_pool_destroy(priv->rq[i].page_pool); 1088 priv->rq[i].page_pool = NULL; 1089 } 1090 1091 return err; 1092 } 1093 1094 static int __veth_napi_enable(struct net_device *dev) 1095 { 1096 return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues); 1097 } 1098 1099 static void veth_napi_del_range(struct net_device *dev, int start, int end) 1100 { 1101 struct veth_priv *priv = netdev_priv(dev); 1102 int i; 1103 1104 for (i = start; i < end; i++) { 1105 struct veth_rq *rq = &priv->rq[i]; 1106 1107 rcu_assign_pointer(priv->rq[i].napi, NULL); 1108 napi_disable(&rq->xdp_napi); 1109 __netif_napi_del(&rq->xdp_napi); 1110 } 1111 synchronize_net(); 1112 1113 for (i = start; i < end; i++) { 1114 struct veth_rq *rq = &priv->rq[i]; 1115 1116 rq->rx_notify_masked = false; 1117 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free); 1118 } 1119 1120 for (i = start; i < end; i++) { 1121 page_pool_destroy(priv->rq[i].page_pool); 1122 priv->rq[i].page_pool = NULL; 1123 } 1124 } 1125 1126 static void veth_napi_del(struct net_device *dev) 1127 { 1128 veth_napi_del_range(dev, 0, dev->real_num_rx_queues); 1129 } 1130 1131 static bool veth_gro_requested(const struct net_device *dev) 1132 { 1133 return !!(dev->wanted_features & NETIF_F_GRO); 1134 } 1135 1136 static int veth_enable_xdp_range(struct net_device *dev, int start, int end, 1137 bool napi_already_on) 1138 { 1139 struct veth_priv *priv = netdev_priv(dev); 1140 int err, i; 1141 1142 for (i = start; i < end; i++) { 1143 struct veth_rq *rq = &priv->rq[i]; 1144 1145 if (!napi_already_on) 1146 netif_napi_add(dev, &rq->xdp_napi, veth_poll); 1147 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id); 1148 if (err < 0) 1149 goto err_rxq_reg; 1150 1151 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq, 1152 MEM_TYPE_PAGE_SHARED, 1153 NULL); 1154 if (err < 0) 1155 goto err_reg_mem; 1156 1157 /* Save original mem info as it can be overwritten */ 1158 rq->xdp_mem = rq->xdp_rxq.mem; 1159 } 1160 return 0; 1161 1162 err_reg_mem: 1163 xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq); 1164 err_rxq_reg: 1165 for (i--; i >= start; i--) { 1166 struct veth_rq *rq = &priv->rq[i]; 1167 1168 xdp_rxq_info_unreg(&rq->xdp_rxq); 1169 if (!napi_already_on) 1170 netif_napi_del(&rq->xdp_napi); 1171 } 1172 1173 return err; 1174 } 1175 1176 static void veth_disable_xdp_range(struct net_device *dev, int start, int end, 1177 bool delete_napi) 1178 { 1179 struct veth_priv *priv = netdev_priv(dev); 1180 int i; 1181 1182 for (i = start; i < end; i++) { 1183 struct veth_rq *rq = &priv->rq[i]; 1184 1185 rq->xdp_rxq.mem = rq->xdp_mem; 1186 xdp_rxq_info_unreg(&rq->xdp_rxq); 1187 1188 if (delete_napi) 1189 netif_napi_del(&rq->xdp_napi); 1190 } 1191 } 1192 1193 static int veth_enable_xdp(struct net_device *dev) 1194 { 1195 bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP); 1196 struct veth_priv *priv = netdev_priv(dev); 1197 int err, i; 1198 1199 if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) { 1200 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on); 1201 if (err) 1202 return err; 1203 1204 if (!napi_already_on) { 1205 err = __veth_napi_enable(dev); 1206 if (err) { 1207 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true); 1208 return err; 1209 } 1210 1211 if (!veth_gro_requested(dev)) { 1212 /* user-space did not require GRO, but adding XDP 1213 * is supposed to get GRO working 1214 */ 1215 dev->features |= NETIF_F_GRO; 1216 netdev_features_change(dev); 1217 } 1218 } 1219 } 1220 1221 for (i = 0; i < dev->real_num_rx_queues; i++) { 1222 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog); 1223 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi); 1224 } 1225 1226 return 0; 1227 } 1228 1229 static void veth_disable_xdp(struct net_device *dev) 1230 { 1231 struct veth_priv *priv = netdev_priv(dev); 1232 int i; 1233 1234 for (i = 0; i < dev->real_num_rx_queues; i++) 1235 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL); 1236 1237 if (!netif_running(dev) || !veth_gro_requested(dev)) { 1238 veth_napi_del(dev); 1239 1240 /* if user-space did not require GRO, since adding XDP 1241 * enabled it, clear it now 1242 */ 1243 if (!veth_gro_requested(dev) && netif_running(dev)) { 1244 dev->features &= ~NETIF_F_GRO; 1245 netdev_features_change(dev); 1246 } 1247 } 1248 1249 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false); 1250 } 1251 1252 static int veth_napi_enable_range(struct net_device *dev, int start, int end) 1253 { 1254 struct veth_priv *priv = netdev_priv(dev); 1255 int err, i; 1256 1257 for (i = start; i < end; i++) { 1258 struct veth_rq *rq = &priv->rq[i]; 1259 1260 netif_napi_add(dev, &rq->xdp_napi, veth_poll); 1261 } 1262 1263 err = __veth_napi_enable_range(dev, start, end); 1264 if (err) { 1265 for (i = start; i < end; i++) { 1266 struct veth_rq *rq = &priv->rq[i]; 1267 1268 netif_napi_del(&rq->xdp_napi); 1269 } 1270 return err; 1271 } 1272 return err; 1273 } 1274 1275 static int veth_napi_enable(struct net_device *dev) 1276 { 1277 return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues); 1278 } 1279 1280 static void veth_disable_range_safe(struct net_device *dev, int start, int end) 1281 { 1282 struct veth_priv *priv = netdev_priv(dev); 1283 1284 if (start >= end) 1285 return; 1286 1287 if (priv->_xdp_prog) { 1288 veth_napi_del_range(dev, start, end); 1289 veth_disable_xdp_range(dev, start, end, false); 1290 } else if (veth_gro_requested(dev)) { 1291 veth_napi_del_range(dev, start, end); 1292 } 1293 } 1294 1295 static int veth_enable_range_safe(struct net_device *dev, int start, int end) 1296 { 1297 struct veth_priv *priv = netdev_priv(dev); 1298 int err; 1299 1300 if (start >= end) 1301 return 0; 1302 1303 if (priv->_xdp_prog) { 1304 /* these channels are freshly initialized, napi is not on there even 1305 * when GRO is requeste 1306 */ 1307 err = veth_enable_xdp_range(dev, start, end, false); 1308 if (err) 1309 return err; 1310 1311 err = __veth_napi_enable_range(dev, start, end); 1312 if (err) { 1313 /* on error always delete the newly added napis */ 1314 veth_disable_xdp_range(dev, start, end, true); 1315 return err; 1316 } 1317 } else if (veth_gro_requested(dev)) { 1318 return veth_napi_enable_range(dev, start, end); 1319 } 1320 return 0; 1321 } 1322 1323 static void veth_set_xdp_features(struct net_device *dev) 1324 { 1325 struct veth_priv *priv = netdev_priv(dev); 1326 struct net_device *peer; 1327 1328 peer = rtnl_dereference(priv->peer); 1329 if (peer && peer->real_num_tx_queues <= dev->real_num_rx_queues) { 1330 struct veth_priv *priv_peer = netdev_priv(peer); 1331 xdp_features_t val = NETDEV_XDP_ACT_BASIC | 1332 NETDEV_XDP_ACT_REDIRECT | 1333 NETDEV_XDP_ACT_RX_SG; 1334 1335 if (priv_peer->_xdp_prog || veth_gro_requested(peer)) 1336 val |= NETDEV_XDP_ACT_NDO_XMIT | 1337 NETDEV_XDP_ACT_NDO_XMIT_SG; 1338 xdp_set_features_flag(dev, val); 1339 } else { 1340 xdp_clear_features_flag(dev); 1341 } 1342 } 1343 1344 static int veth_set_channels(struct net_device *dev, 1345 struct ethtool_channels *ch) 1346 { 1347 struct veth_priv *priv = netdev_priv(dev); 1348 unsigned int old_rx_count, new_rx_count; 1349 struct veth_priv *peer_priv; 1350 struct net_device *peer; 1351 int err; 1352 1353 /* sanity check. Upper bounds are already enforced by the caller */ 1354 if (!ch->rx_count || !ch->tx_count) 1355 return -EINVAL; 1356 1357 /* avoid braking XDP, if that is enabled */ 1358 peer = rtnl_dereference(priv->peer); 1359 peer_priv = peer ? netdev_priv(peer) : NULL; 1360 if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues) 1361 return -EINVAL; 1362 1363 if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues) 1364 return -EINVAL; 1365 1366 old_rx_count = dev->real_num_rx_queues; 1367 new_rx_count = ch->rx_count; 1368 if (netif_running(dev)) { 1369 /* turn device off */ 1370 netif_carrier_off(dev); 1371 if (peer) 1372 netif_carrier_off(peer); 1373 1374 /* try to allocate new resurces, as needed*/ 1375 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count); 1376 if (err) 1377 goto out; 1378 } 1379 1380 err = netif_set_real_num_rx_queues(dev, ch->rx_count); 1381 if (err) 1382 goto revert; 1383 1384 err = netif_set_real_num_tx_queues(dev, ch->tx_count); 1385 if (err) { 1386 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count); 1387 1388 /* this error condition could happen only if rx and tx change 1389 * in opposite directions (e.g. tx nr raises, rx nr decreases) 1390 * and we can't do anything to fully restore the original 1391 * status 1392 */ 1393 if (err2) 1394 pr_warn("Can't restore rx queues config %d -> %d %d", 1395 new_rx_count, old_rx_count, err2); 1396 else 1397 goto revert; 1398 } 1399 1400 out: 1401 if (netif_running(dev)) { 1402 /* note that we need to swap the arguments WRT the enable part 1403 * to identify the range we have to disable 1404 */ 1405 veth_disable_range_safe(dev, new_rx_count, old_rx_count); 1406 netif_carrier_on(dev); 1407 if (peer) 1408 netif_carrier_on(peer); 1409 } 1410 1411 /* update XDP supported features */ 1412 veth_set_xdp_features(dev); 1413 if (peer) 1414 veth_set_xdp_features(peer); 1415 1416 return err; 1417 1418 revert: 1419 new_rx_count = old_rx_count; 1420 old_rx_count = ch->rx_count; 1421 goto out; 1422 } 1423 1424 static int veth_open(struct net_device *dev) 1425 { 1426 struct veth_priv *priv = netdev_priv(dev); 1427 struct net_device *peer = rtnl_dereference(priv->peer); 1428 int err; 1429 1430 if (!peer) 1431 return -ENOTCONN; 1432 1433 if (priv->_xdp_prog) { 1434 err = veth_enable_xdp(dev); 1435 if (err) 1436 return err; 1437 } else if (veth_gro_requested(dev)) { 1438 err = veth_napi_enable(dev); 1439 if (err) 1440 return err; 1441 } 1442 1443 if (peer->flags & IFF_UP) { 1444 netif_carrier_on(dev); 1445 netif_carrier_on(peer); 1446 } 1447 1448 veth_set_xdp_features(dev); 1449 1450 return 0; 1451 } 1452 1453 static int veth_close(struct net_device *dev) 1454 { 1455 struct veth_priv *priv = netdev_priv(dev); 1456 struct net_device *peer = rtnl_dereference(priv->peer); 1457 1458 netif_carrier_off(dev); 1459 if (peer) 1460 netif_carrier_off(peer); 1461 1462 if (priv->_xdp_prog) 1463 veth_disable_xdp(dev); 1464 else if (veth_gro_requested(dev)) 1465 veth_napi_del(dev); 1466 1467 return 0; 1468 } 1469 1470 static int is_valid_veth_mtu(int mtu) 1471 { 1472 return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU; 1473 } 1474 1475 static int veth_alloc_queues(struct net_device *dev) 1476 { 1477 struct veth_priv *priv = netdev_priv(dev); 1478 int i; 1479 1480 priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL_ACCOUNT); 1481 if (!priv->rq) 1482 return -ENOMEM; 1483 1484 for (i = 0; i < dev->num_rx_queues; i++) { 1485 priv->rq[i].dev = dev; 1486 u64_stats_init(&priv->rq[i].stats.syncp); 1487 } 1488 1489 return 0; 1490 } 1491 1492 static void veth_free_queues(struct net_device *dev) 1493 { 1494 struct veth_priv *priv = netdev_priv(dev); 1495 1496 kfree(priv->rq); 1497 } 1498 1499 static int veth_dev_init(struct net_device *dev) 1500 { 1501 return veth_alloc_queues(dev); 1502 } 1503 1504 static void veth_dev_free(struct net_device *dev) 1505 { 1506 veth_free_queues(dev); 1507 } 1508 1509 #ifdef CONFIG_NET_POLL_CONTROLLER 1510 static void veth_poll_controller(struct net_device *dev) 1511 { 1512 /* veth only receives frames when its peer sends one 1513 * Since it has nothing to do with disabling irqs, we are guaranteed 1514 * never to have pending data when we poll for it so 1515 * there is nothing to do here. 1516 * 1517 * We need this though so netpoll recognizes us as an interface that 1518 * supports polling, which enables bridge devices in virt setups to 1519 * still use netconsole 1520 */ 1521 } 1522 #endif /* CONFIG_NET_POLL_CONTROLLER */ 1523 1524 static int veth_get_iflink(const struct net_device *dev) 1525 { 1526 struct veth_priv *priv = netdev_priv(dev); 1527 struct net_device *peer; 1528 int iflink; 1529 1530 rcu_read_lock(); 1531 peer = rcu_dereference(priv->peer); 1532 iflink = peer ? peer->ifindex : 0; 1533 rcu_read_unlock(); 1534 1535 return iflink; 1536 } 1537 1538 static netdev_features_t veth_fix_features(struct net_device *dev, 1539 netdev_features_t features) 1540 { 1541 struct veth_priv *priv = netdev_priv(dev); 1542 struct net_device *peer; 1543 1544 peer = rtnl_dereference(priv->peer); 1545 if (peer) { 1546 struct veth_priv *peer_priv = netdev_priv(peer); 1547 1548 if (peer_priv->_xdp_prog) 1549 features &= ~NETIF_F_GSO_SOFTWARE; 1550 } 1551 if (priv->_xdp_prog) 1552 features |= NETIF_F_GRO; 1553 1554 return features; 1555 } 1556 1557 static int veth_set_features(struct net_device *dev, 1558 netdev_features_t features) 1559 { 1560 netdev_features_t changed = features ^ dev->features; 1561 struct veth_priv *priv = netdev_priv(dev); 1562 struct net_device *peer; 1563 int err; 1564 1565 if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog) 1566 return 0; 1567 1568 peer = rtnl_dereference(priv->peer); 1569 if (features & NETIF_F_GRO) { 1570 err = veth_napi_enable(dev); 1571 if (err) 1572 return err; 1573 1574 if (peer) 1575 xdp_features_set_redirect_target(peer, true); 1576 } else { 1577 if (peer) 1578 xdp_features_clear_redirect_target(peer); 1579 veth_napi_del(dev); 1580 } 1581 return 0; 1582 } 1583 1584 static void veth_set_rx_headroom(struct net_device *dev, int new_hr) 1585 { 1586 struct veth_priv *peer_priv, *priv = netdev_priv(dev); 1587 struct net_device *peer; 1588 1589 if (new_hr < 0) 1590 new_hr = 0; 1591 1592 rcu_read_lock(); 1593 peer = rcu_dereference(priv->peer); 1594 if (unlikely(!peer)) 1595 goto out; 1596 1597 peer_priv = netdev_priv(peer); 1598 priv->requested_headroom = new_hr; 1599 new_hr = max(priv->requested_headroom, peer_priv->requested_headroom); 1600 dev->needed_headroom = new_hr; 1601 peer->needed_headroom = new_hr; 1602 1603 out: 1604 rcu_read_unlock(); 1605 } 1606 1607 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog, 1608 struct netlink_ext_ack *extack) 1609 { 1610 struct veth_priv *priv = netdev_priv(dev); 1611 struct bpf_prog *old_prog; 1612 struct net_device *peer; 1613 unsigned int max_mtu; 1614 int err; 1615 1616 old_prog = priv->_xdp_prog; 1617 priv->_xdp_prog = prog; 1618 peer = rtnl_dereference(priv->peer); 1619 1620 if (prog) { 1621 if (!peer) { 1622 NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached"); 1623 err = -ENOTCONN; 1624 goto err; 1625 } 1626 1627 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) - 1628 peer->hard_header_len; 1629 /* Allow increasing the max_mtu if the program supports 1630 * XDP fragments. 1631 */ 1632 if (prog->aux->xdp_has_frags) 1633 max_mtu += PAGE_SIZE * MAX_SKB_FRAGS; 1634 1635 if (peer->mtu > max_mtu) { 1636 NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP"); 1637 err = -ERANGE; 1638 goto err; 1639 } 1640 1641 if (dev->real_num_rx_queues < peer->real_num_tx_queues) { 1642 NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues"); 1643 err = -ENOSPC; 1644 goto err; 1645 } 1646 1647 if (dev->flags & IFF_UP) { 1648 err = veth_enable_xdp(dev); 1649 if (err) { 1650 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed"); 1651 goto err; 1652 } 1653 } 1654 1655 if (!old_prog) { 1656 peer->hw_features &= ~NETIF_F_GSO_SOFTWARE; 1657 peer->max_mtu = max_mtu; 1658 } 1659 1660 xdp_features_set_redirect_target(peer, true); 1661 } 1662 1663 if (old_prog) { 1664 if (!prog) { 1665 if (peer && !veth_gro_requested(dev)) 1666 xdp_features_clear_redirect_target(peer); 1667 1668 if (dev->flags & IFF_UP) 1669 veth_disable_xdp(dev); 1670 1671 if (peer) { 1672 peer->hw_features |= NETIF_F_GSO_SOFTWARE; 1673 peer->max_mtu = ETH_MAX_MTU; 1674 } 1675 } 1676 bpf_prog_put(old_prog); 1677 } 1678 1679 if ((!!old_prog ^ !!prog) && peer) 1680 netdev_update_features(peer); 1681 1682 return 0; 1683 err: 1684 priv->_xdp_prog = old_prog; 1685 1686 return err; 1687 } 1688 1689 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp) 1690 { 1691 switch (xdp->command) { 1692 case XDP_SETUP_PROG: 1693 return veth_xdp_set(dev, xdp->prog, xdp->extack); 1694 default: 1695 return -EINVAL; 1696 } 1697 } 1698 1699 static int veth_xdp_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp) 1700 { 1701 struct veth_xdp_buff *_ctx = (void *)ctx; 1702 1703 if (!_ctx->skb) 1704 return -ENODATA; 1705 1706 *timestamp = skb_hwtstamps(_ctx->skb)->hwtstamp; 1707 return 0; 1708 } 1709 1710 static int veth_xdp_rx_hash(const struct xdp_md *ctx, u32 *hash, 1711 enum xdp_rss_hash_type *rss_type) 1712 { 1713 struct veth_xdp_buff *_ctx = (void *)ctx; 1714 struct sk_buff *skb = _ctx->skb; 1715 1716 if (!skb) 1717 return -ENODATA; 1718 1719 *hash = skb_get_hash(skb); 1720 *rss_type = skb->l4_hash ? XDP_RSS_TYPE_L4_ANY : XDP_RSS_TYPE_NONE; 1721 1722 return 0; 1723 } 1724 1725 static const struct net_device_ops veth_netdev_ops = { 1726 .ndo_init = veth_dev_init, 1727 .ndo_open = veth_open, 1728 .ndo_stop = veth_close, 1729 .ndo_start_xmit = veth_xmit, 1730 .ndo_get_stats64 = veth_get_stats64, 1731 .ndo_set_rx_mode = veth_set_multicast_list, 1732 .ndo_set_mac_address = eth_mac_addr, 1733 #ifdef CONFIG_NET_POLL_CONTROLLER 1734 .ndo_poll_controller = veth_poll_controller, 1735 #endif 1736 .ndo_get_iflink = veth_get_iflink, 1737 .ndo_fix_features = veth_fix_features, 1738 .ndo_set_features = veth_set_features, 1739 .ndo_features_check = passthru_features_check, 1740 .ndo_set_rx_headroom = veth_set_rx_headroom, 1741 .ndo_bpf = veth_xdp, 1742 .ndo_xdp_xmit = veth_ndo_xdp_xmit, 1743 .ndo_get_peer_dev = veth_peer_dev, 1744 }; 1745 1746 static const struct xdp_metadata_ops veth_xdp_metadata_ops = { 1747 .xmo_rx_timestamp = veth_xdp_rx_timestamp, 1748 .xmo_rx_hash = veth_xdp_rx_hash, 1749 }; 1750 1751 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \ 1752 NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \ 1753 NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \ 1754 NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \ 1755 NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX ) 1756 1757 static void veth_setup(struct net_device *dev) 1758 { 1759 ether_setup(dev); 1760 1761 dev->priv_flags &= ~IFF_TX_SKB_SHARING; 1762 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE; 1763 dev->priv_flags |= IFF_NO_QUEUE; 1764 dev->priv_flags |= IFF_PHONY_HEADROOM; 1765 1766 dev->netdev_ops = &veth_netdev_ops; 1767 dev->xdp_metadata_ops = &veth_xdp_metadata_ops; 1768 dev->ethtool_ops = &veth_ethtool_ops; 1769 dev->features |= NETIF_F_LLTX; 1770 dev->features |= VETH_FEATURES; 1771 dev->vlan_features = dev->features & 1772 ~(NETIF_F_HW_VLAN_CTAG_TX | 1773 NETIF_F_HW_VLAN_STAG_TX | 1774 NETIF_F_HW_VLAN_CTAG_RX | 1775 NETIF_F_HW_VLAN_STAG_RX); 1776 dev->needs_free_netdev = true; 1777 dev->priv_destructor = veth_dev_free; 1778 dev->pcpu_stat_type = NETDEV_PCPU_STAT_TSTATS; 1779 dev->max_mtu = ETH_MAX_MTU; 1780 1781 dev->hw_features = VETH_FEATURES; 1782 dev->hw_enc_features = VETH_FEATURES; 1783 dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE; 1784 netif_set_tso_max_size(dev, GSO_MAX_SIZE); 1785 } 1786 1787 /* 1788 * netlink interface 1789 */ 1790 1791 static int veth_validate(struct nlattr *tb[], struct nlattr *data[], 1792 struct netlink_ext_ack *extack) 1793 { 1794 if (tb[IFLA_ADDRESS]) { 1795 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) 1796 return -EINVAL; 1797 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) 1798 return -EADDRNOTAVAIL; 1799 } 1800 if (tb[IFLA_MTU]) { 1801 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU]))) 1802 return -EINVAL; 1803 } 1804 return 0; 1805 } 1806 1807 static struct rtnl_link_ops veth_link_ops; 1808 1809 static void veth_disable_gro(struct net_device *dev) 1810 { 1811 dev->features &= ~NETIF_F_GRO; 1812 dev->wanted_features &= ~NETIF_F_GRO; 1813 netdev_update_features(dev); 1814 } 1815 1816 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[]) 1817 { 1818 int err; 1819 1820 if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) { 1821 err = netif_set_real_num_tx_queues(dev, 1); 1822 if (err) 1823 return err; 1824 } 1825 if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) { 1826 err = netif_set_real_num_rx_queues(dev, 1); 1827 if (err) 1828 return err; 1829 } 1830 return 0; 1831 } 1832 1833 static int veth_newlink(struct net *src_net, struct net_device *dev, 1834 struct nlattr *tb[], struct nlattr *data[], 1835 struct netlink_ext_ack *extack) 1836 { 1837 int err; 1838 struct net_device *peer; 1839 struct veth_priv *priv; 1840 char ifname[IFNAMSIZ]; 1841 struct nlattr *peer_tb[IFLA_MAX + 1], **tbp; 1842 unsigned char name_assign_type; 1843 struct ifinfomsg *ifmp; 1844 struct net *net; 1845 1846 /* 1847 * create and register peer first 1848 */ 1849 if (data != NULL && data[VETH_INFO_PEER] != NULL) { 1850 struct nlattr *nla_peer; 1851 1852 nla_peer = data[VETH_INFO_PEER]; 1853 ifmp = nla_data(nla_peer); 1854 err = rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack); 1855 if (err < 0) 1856 return err; 1857 1858 err = veth_validate(peer_tb, NULL, extack); 1859 if (err < 0) 1860 return err; 1861 1862 tbp = peer_tb; 1863 } else { 1864 ifmp = NULL; 1865 tbp = tb; 1866 } 1867 1868 if (ifmp && tbp[IFLA_IFNAME]) { 1869 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ); 1870 name_assign_type = NET_NAME_USER; 1871 } else { 1872 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d"); 1873 name_assign_type = NET_NAME_ENUM; 1874 } 1875 1876 net = rtnl_link_get_net(src_net, tbp); 1877 if (IS_ERR(net)) 1878 return PTR_ERR(net); 1879 1880 peer = rtnl_create_link(net, ifname, name_assign_type, 1881 &veth_link_ops, tbp, extack); 1882 if (IS_ERR(peer)) { 1883 put_net(net); 1884 return PTR_ERR(peer); 1885 } 1886 1887 if (!ifmp || !tbp[IFLA_ADDRESS]) 1888 eth_hw_addr_random(peer); 1889 1890 if (ifmp && (dev->ifindex != 0)) 1891 peer->ifindex = ifmp->ifi_index; 1892 1893 netif_inherit_tso_max(peer, dev); 1894 1895 err = register_netdevice(peer); 1896 put_net(net); 1897 net = NULL; 1898 if (err < 0) 1899 goto err_register_peer; 1900 1901 /* keep GRO disabled by default to be consistent with the established 1902 * veth behavior 1903 */ 1904 veth_disable_gro(peer); 1905 netif_carrier_off(peer); 1906 1907 err = rtnl_configure_link(peer, ifmp, 0, NULL); 1908 if (err < 0) 1909 goto err_configure_peer; 1910 1911 /* 1912 * register dev last 1913 * 1914 * note, that since we've registered new device the dev's name 1915 * should be re-allocated 1916 */ 1917 1918 if (tb[IFLA_ADDRESS] == NULL) 1919 eth_hw_addr_random(dev); 1920 1921 if (tb[IFLA_IFNAME]) 1922 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ); 1923 else 1924 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d"); 1925 1926 err = register_netdevice(dev); 1927 if (err < 0) 1928 goto err_register_dev; 1929 1930 netif_carrier_off(dev); 1931 1932 /* 1933 * tie the deviced together 1934 */ 1935 1936 priv = netdev_priv(dev); 1937 rcu_assign_pointer(priv->peer, peer); 1938 err = veth_init_queues(dev, tb); 1939 if (err) 1940 goto err_queues; 1941 1942 priv = netdev_priv(peer); 1943 rcu_assign_pointer(priv->peer, dev); 1944 err = veth_init_queues(peer, tb); 1945 if (err) 1946 goto err_queues; 1947 1948 veth_disable_gro(dev); 1949 /* update XDP supported features */ 1950 veth_set_xdp_features(dev); 1951 veth_set_xdp_features(peer); 1952 1953 return 0; 1954 1955 err_queues: 1956 unregister_netdevice(dev); 1957 err_register_dev: 1958 /* nothing to do */ 1959 err_configure_peer: 1960 unregister_netdevice(peer); 1961 return err; 1962 1963 err_register_peer: 1964 free_netdev(peer); 1965 return err; 1966 } 1967 1968 static void veth_dellink(struct net_device *dev, struct list_head *head) 1969 { 1970 struct veth_priv *priv; 1971 struct net_device *peer; 1972 1973 priv = netdev_priv(dev); 1974 peer = rtnl_dereference(priv->peer); 1975 1976 /* Note : dellink() is called from default_device_exit_batch(), 1977 * before a rcu_synchronize() point. The devices are guaranteed 1978 * not being freed before one RCU grace period. 1979 */ 1980 RCU_INIT_POINTER(priv->peer, NULL); 1981 unregister_netdevice_queue(dev, head); 1982 1983 if (peer) { 1984 priv = netdev_priv(peer); 1985 RCU_INIT_POINTER(priv->peer, NULL); 1986 unregister_netdevice_queue(peer, head); 1987 } 1988 } 1989 1990 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = { 1991 [VETH_INFO_PEER] = { .len = sizeof(struct ifinfomsg) }, 1992 }; 1993 1994 static struct net *veth_get_link_net(const struct net_device *dev) 1995 { 1996 struct veth_priv *priv = netdev_priv(dev); 1997 struct net_device *peer = rtnl_dereference(priv->peer); 1998 1999 return peer ? dev_net(peer) : dev_net(dev); 2000 } 2001 2002 static unsigned int veth_get_num_queues(void) 2003 { 2004 /* enforce the same queue limit as rtnl_create_link */ 2005 int queues = num_possible_cpus(); 2006 2007 if (queues > 4096) 2008 queues = 4096; 2009 return queues; 2010 } 2011 2012 static struct rtnl_link_ops veth_link_ops = { 2013 .kind = DRV_NAME, 2014 .priv_size = sizeof(struct veth_priv), 2015 .setup = veth_setup, 2016 .validate = veth_validate, 2017 .newlink = veth_newlink, 2018 .dellink = veth_dellink, 2019 .policy = veth_policy, 2020 .maxtype = VETH_INFO_MAX, 2021 .get_link_net = veth_get_link_net, 2022 .get_num_tx_queues = veth_get_num_queues, 2023 .get_num_rx_queues = veth_get_num_queues, 2024 }; 2025 2026 /* 2027 * init/fini 2028 */ 2029 2030 static __init int veth_init(void) 2031 { 2032 return rtnl_link_register(&veth_link_ops); 2033 } 2034 2035 static __exit void veth_exit(void) 2036 { 2037 rtnl_link_unregister(&veth_link_ops); 2038 } 2039 2040 module_init(veth_init); 2041 module_exit(veth_exit); 2042 2043 MODULE_DESCRIPTION("Virtual Ethernet Tunnel"); 2044 MODULE_LICENSE("GPL v2"); 2045 MODULE_ALIAS_RTNL_LINK(DRV_NAME); 2046