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/netdev_lock.h> 21 #include <net/xfrm.h> 22 #include <net/xdp.h> 23 #include <linux/veth.h> 24 #include <linux/module.h> 25 #include <linux/bpf.h> 26 #include <linux/filter.h> 27 #include <linux/ptr_ring.h> 28 #include <linux/bpf_trace.h> 29 #include <linux/net_tstamp.h> 30 #include <linux/skbuff_ref.h> 31 #include <net/page_pool/helpers.h> 32 33 #define DRV_NAME "veth" 34 #define DRV_VERSION "1.0" 35 36 #define VETH_XDP_FLAG BIT(0) 37 #define VETH_RING_SIZE 256 38 #define VETH_XDP_HEADROOM (XDP_PACKET_HEADROOM + NET_IP_ALIGN) 39 40 #define VETH_XDP_TX_BULK_SIZE 16 41 #define VETH_XDP_BATCH 16 42 43 struct veth_stats { 44 u64 rx_drops; 45 /* xdp */ 46 u64 xdp_packets; 47 u64 xdp_bytes; 48 u64 xdp_redirect; 49 u64 xdp_drops; 50 u64 xdp_tx; 51 u64 xdp_tx_err; 52 u64 peer_tq_xdp_xmit; 53 u64 peer_tq_xdp_xmit_err; 54 }; 55 56 struct veth_rq_stats { 57 struct veth_stats vs; 58 struct u64_stats_sync syncp; 59 }; 60 61 struct veth_rq { 62 struct napi_struct xdp_napi; 63 struct napi_struct __rcu *napi; /* points to xdp_napi when the latter is initialized */ 64 struct net_device *dev; 65 struct bpf_prog __rcu *xdp_prog; 66 struct xdp_mem_info xdp_mem; 67 struct veth_rq_stats stats; 68 bool rx_notify_masked; 69 struct ptr_ring xdp_ring; 70 struct xdp_rxq_info xdp_rxq; 71 struct page_pool *page_pool; 72 }; 73 74 struct veth_priv { 75 struct net_device __rcu *peer; 76 atomic64_t dropped; 77 struct bpf_prog *_xdp_prog; 78 struct veth_rq *rq; 79 unsigned int requested_headroom; 80 }; 81 82 struct veth_xdp_tx_bq { 83 struct xdp_frame *q[VETH_XDP_TX_BULK_SIZE]; 84 unsigned int count; 85 }; 86 87 /* 88 * ethtool interface 89 */ 90 91 struct veth_q_stat_desc { 92 char desc[ETH_GSTRING_LEN]; 93 size_t offset; 94 }; 95 96 #define VETH_RQ_STAT(m) offsetof(struct veth_stats, m) 97 98 static const struct veth_q_stat_desc veth_rq_stats_desc[] = { 99 { "xdp_packets", VETH_RQ_STAT(xdp_packets) }, 100 { "xdp_bytes", VETH_RQ_STAT(xdp_bytes) }, 101 { "drops", VETH_RQ_STAT(rx_drops) }, 102 { "xdp_redirect", VETH_RQ_STAT(xdp_redirect) }, 103 { "xdp_drops", VETH_RQ_STAT(xdp_drops) }, 104 { "xdp_tx", VETH_RQ_STAT(xdp_tx) }, 105 { "xdp_tx_errors", VETH_RQ_STAT(xdp_tx_err) }, 106 }; 107 108 #define VETH_RQ_STATS_LEN ARRAY_SIZE(veth_rq_stats_desc) 109 110 static const struct veth_q_stat_desc veth_tq_stats_desc[] = { 111 { "xdp_xmit", VETH_RQ_STAT(peer_tq_xdp_xmit) }, 112 { "xdp_xmit_errors", VETH_RQ_STAT(peer_tq_xdp_xmit_err) }, 113 }; 114 115 #define VETH_TQ_STATS_LEN ARRAY_SIZE(veth_tq_stats_desc) 116 117 static struct { 118 const char string[ETH_GSTRING_LEN]; 119 } ethtool_stats_keys[] = { 120 { "peer_ifindex" }, 121 }; 122 123 struct veth_xdp_buff { 124 struct xdp_buff xdp; 125 struct sk_buff *skb; 126 }; 127 128 static int veth_get_link_ksettings(struct net_device *dev, 129 struct ethtool_link_ksettings *cmd) 130 { 131 cmd->base.speed = SPEED_10000; 132 cmd->base.duplex = DUPLEX_FULL; 133 cmd->base.port = PORT_TP; 134 cmd->base.autoneg = AUTONEG_DISABLE; 135 return 0; 136 } 137 138 static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info) 139 { 140 strscpy(info->driver, DRV_NAME, sizeof(info->driver)); 141 strscpy(info->version, DRV_VERSION, sizeof(info->version)); 142 } 143 144 static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf) 145 { 146 u8 *p = buf; 147 int i, j; 148 149 switch(stringset) { 150 case ETH_SS_STATS: 151 memcpy(p, ðtool_stats_keys, sizeof(ethtool_stats_keys)); 152 p += sizeof(ethtool_stats_keys); 153 for (i = 0; i < dev->real_num_rx_queues; i++) 154 for (j = 0; j < VETH_RQ_STATS_LEN; j++) 155 ethtool_sprintf(&p, "rx_queue_%u_%.18s", 156 i, veth_rq_stats_desc[j].desc); 157 158 for (i = 0; i < dev->real_num_tx_queues; i++) 159 for (j = 0; j < VETH_TQ_STATS_LEN; j++) 160 ethtool_sprintf(&p, "tx_queue_%u_%.18s", 161 i, veth_tq_stats_desc[j].desc); 162 163 page_pool_ethtool_stats_get_strings(p); 164 break; 165 } 166 } 167 168 static int veth_get_sset_count(struct net_device *dev, int sset) 169 { 170 switch (sset) { 171 case ETH_SS_STATS: 172 return ARRAY_SIZE(ethtool_stats_keys) + 173 VETH_RQ_STATS_LEN * dev->real_num_rx_queues + 174 VETH_TQ_STATS_LEN * dev->real_num_tx_queues + 175 page_pool_ethtool_stats_get_count(); 176 default: 177 return -EOPNOTSUPP; 178 } 179 } 180 181 static void veth_get_page_pool_stats(struct net_device *dev, u64 *data) 182 { 183 #ifdef CONFIG_PAGE_POOL_STATS 184 struct veth_priv *priv = netdev_priv(dev); 185 struct page_pool_stats pp_stats = {}; 186 int i; 187 188 for (i = 0; i < dev->real_num_rx_queues; i++) { 189 if (!priv->rq[i].page_pool) 190 continue; 191 page_pool_get_stats(priv->rq[i].page_pool, &pp_stats); 192 } 193 page_pool_ethtool_stats_get(data, &pp_stats); 194 #endif /* CONFIG_PAGE_POOL_STATS */ 195 } 196 197 static void veth_get_ethtool_stats(struct net_device *dev, 198 struct ethtool_stats *stats, u64 *data) 199 { 200 struct veth_priv *rcv_priv, *priv = netdev_priv(dev); 201 struct net_device *peer = rtnl_dereference(priv->peer); 202 int i, j, idx, pp_idx; 203 204 data[0] = peer ? peer->ifindex : 0; 205 idx = 1; 206 for (i = 0; i < dev->real_num_rx_queues; i++) { 207 const struct veth_rq_stats *rq_stats = &priv->rq[i].stats; 208 const void *stats_base = (void *)&rq_stats->vs; 209 unsigned int start; 210 size_t offset; 211 212 do { 213 start = u64_stats_fetch_begin(&rq_stats->syncp); 214 for (j = 0; j < VETH_RQ_STATS_LEN; j++) { 215 offset = veth_rq_stats_desc[j].offset; 216 data[idx + j] = *(u64 *)(stats_base + offset); 217 } 218 } while (u64_stats_fetch_retry(&rq_stats->syncp, start)); 219 idx += VETH_RQ_STATS_LEN; 220 } 221 pp_idx = idx; 222 223 if (!peer) 224 goto page_pool_stats; 225 226 rcv_priv = netdev_priv(peer); 227 for (i = 0; i < peer->real_num_rx_queues; i++) { 228 const struct veth_rq_stats *rq_stats = &rcv_priv->rq[i].stats; 229 const void *base = (void *)&rq_stats->vs; 230 unsigned int start, tx_idx = idx; 231 size_t offset; 232 233 tx_idx += (i % dev->real_num_tx_queues) * VETH_TQ_STATS_LEN; 234 do { 235 start = u64_stats_fetch_begin(&rq_stats->syncp); 236 for (j = 0; j < VETH_TQ_STATS_LEN; j++) { 237 offset = veth_tq_stats_desc[j].offset; 238 data[tx_idx + j] += *(u64 *)(base + offset); 239 } 240 } while (u64_stats_fetch_retry(&rq_stats->syncp, start)); 241 } 242 pp_idx = idx + dev->real_num_tx_queues * VETH_TQ_STATS_LEN; 243 244 page_pool_stats: 245 veth_get_page_pool_stats(dev, &data[pp_idx]); 246 } 247 248 static void veth_get_channels(struct net_device *dev, 249 struct ethtool_channels *channels) 250 { 251 channels->tx_count = dev->real_num_tx_queues; 252 channels->rx_count = dev->real_num_rx_queues; 253 channels->max_tx = dev->num_tx_queues; 254 channels->max_rx = dev->num_rx_queues; 255 } 256 257 static int veth_set_channels(struct net_device *dev, 258 struct ethtool_channels *ch); 259 260 static const struct ethtool_ops veth_ethtool_ops = { 261 .get_drvinfo = veth_get_drvinfo, 262 .get_link = ethtool_op_get_link, 263 .get_strings = veth_get_strings, 264 .get_sset_count = veth_get_sset_count, 265 .get_ethtool_stats = veth_get_ethtool_stats, 266 .get_link_ksettings = veth_get_link_ksettings, 267 .get_ts_info = ethtool_op_get_ts_info, 268 .get_channels = veth_get_channels, 269 .set_channels = veth_set_channels, 270 }; 271 272 /* general routines */ 273 274 static bool veth_is_xdp_frame(void *ptr) 275 { 276 return (unsigned long)ptr & VETH_XDP_FLAG; 277 } 278 279 static struct xdp_frame *veth_ptr_to_xdp(void *ptr) 280 { 281 return (void *)((unsigned long)ptr & ~VETH_XDP_FLAG); 282 } 283 284 static void *veth_xdp_to_ptr(struct xdp_frame *xdp) 285 { 286 return (void *)((unsigned long)xdp | VETH_XDP_FLAG); 287 } 288 289 static void veth_ptr_free(void *ptr) 290 { 291 if (veth_is_xdp_frame(ptr)) 292 xdp_return_frame(veth_ptr_to_xdp(ptr)); 293 else 294 kfree_skb(ptr); 295 } 296 297 static void __veth_xdp_flush(struct veth_rq *rq) 298 { 299 /* Write ptr_ring before reading rx_notify_masked */ 300 smp_mb(); 301 if (!READ_ONCE(rq->rx_notify_masked) && 302 napi_schedule_prep(&rq->xdp_napi)) { 303 WRITE_ONCE(rq->rx_notify_masked, true); 304 __napi_schedule(&rq->xdp_napi); 305 } 306 } 307 308 static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb) 309 { 310 if (unlikely(ptr_ring_produce(&rq->xdp_ring, skb))) 311 return NETDEV_TX_BUSY; /* signal qdisc layer */ 312 313 return NET_RX_SUCCESS; /* same as NETDEV_TX_OK */ 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 struct netdev_queue *txq; 348 struct net_device *rcv; 349 int length = skb->len; 350 bool use_napi = false; 351 int ret, 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); 374 375 ret = veth_forward_skb(rcv, skb, rq, use_napi); 376 switch (ret) { 377 case NET_RX_SUCCESS: /* same as NETDEV_TX_OK */ 378 if (!use_napi) 379 dev_sw_netstats_tx_add(dev, 1, length); 380 else 381 __veth_xdp_flush(rq); 382 break; 383 case NETDEV_TX_BUSY: 384 /* If a qdisc is attached to our virtual device, returning 385 * NETDEV_TX_BUSY is allowed. 386 */ 387 txq = netdev_get_tx_queue(dev, rxq); 388 389 if (qdisc_txq_has_no_queue(txq)) { 390 dev_kfree_skb_any(skb); 391 goto drop; 392 } 393 /* Restore Eth hdr pulled by dev_forward_skb/eth_type_trans */ 394 __skb_push(skb, ETH_HLEN); 395 netif_tx_stop_queue(txq); 396 /* Makes sure NAPI peer consumer runs. Consumer is responsible 397 * for starting txq again, until then ndo_start_xmit (this 398 * function) will not be invoked by the netstack again. 399 */ 400 __veth_xdp_flush(rq); 401 break; 402 case NET_RX_DROP: /* same as NET_XMIT_DROP */ 403 drop: 404 atomic64_inc(&priv->dropped); 405 ret = NET_XMIT_DROP; 406 break; 407 default: 408 net_crit_ratelimited("%s(%s): Invalid return code(%d)", 409 __func__, dev->name, ret); 410 } 411 rcu_read_unlock(); 412 413 return ret; 414 } 415 416 static void veth_stats_rx(struct veth_stats *result, struct net_device *dev) 417 { 418 struct veth_priv *priv = netdev_priv(dev); 419 int i; 420 421 result->peer_tq_xdp_xmit_err = 0; 422 result->xdp_packets = 0; 423 result->xdp_tx_err = 0; 424 result->xdp_bytes = 0; 425 result->rx_drops = 0; 426 for (i = 0; i < dev->num_rx_queues; i++) { 427 u64 packets, bytes, drops, xdp_tx_err, peer_tq_xdp_xmit_err; 428 struct veth_rq_stats *stats = &priv->rq[i].stats; 429 unsigned int start; 430 431 do { 432 start = u64_stats_fetch_begin(&stats->syncp); 433 peer_tq_xdp_xmit_err = stats->vs.peer_tq_xdp_xmit_err; 434 xdp_tx_err = stats->vs.xdp_tx_err; 435 packets = stats->vs.xdp_packets; 436 bytes = stats->vs.xdp_bytes; 437 drops = stats->vs.rx_drops; 438 } while (u64_stats_fetch_retry(&stats->syncp, start)); 439 result->peer_tq_xdp_xmit_err += peer_tq_xdp_xmit_err; 440 result->xdp_tx_err += xdp_tx_err; 441 result->xdp_packets += packets; 442 result->xdp_bytes += bytes; 443 result->rx_drops += drops; 444 } 445 } 446 447 static void veth_get_stats64(struct net_device *dev, 448 struct rtnl_link_stats64 *tot) 449 { 450 struct veth_priv *priv = netdev_priv(dev); 451 struct net_device *peer; 452 struct veth_stats rx; 453 454 tot->tx_dropped = atomic64_read(&priv->dropped); 455 dev_fetch_sw_netstats(tot, dev->tstats); 456 457 veth_stats_rx(&rx, dev); 458 tot->tx_dropped += rx.xdp_tx_err; 459 tot->rx_dropped = rx.rx_drops + rx.peer_tq_xdp_xmit_err; 460 tot->rx_bytes += rx.xdp_bytes; 461 tot->rx_packets += rx.xdp_packets; 462 463 rcu_read_lock(); 464 peer = rcu_dereference(priv->peer); 465 if (peer) { 466 struct rtnl_link_stats64 tot_peer = {}; 467 468 dev_fetch_sw_netstats(&tot_peer, peer->tstats); 469 tot->rx_bytes += tot_peer.tx_bytes; 470 tot->rx_packets += tot_peer.tx_packets; 471 472 veth_stats_rx(&rx, peer); 473 tot->tx_dropped += rx.peer_tq_xdp_xmit_err; 474 tot->rx_dropped += rx.xdp_tx_err; 475 tot->tx_bytes += rx.xdp_bytes; 476 tot->tx_packets += rx.xdp_packets; 477 } 478 rcu_read_unlock(); 479 } 480 481 /* fake multicast ability */ 482 static void veth_set_multicast_list(struct net_device *dev) 483 { 484 } 485 486 static int veth_select_rxq(struct net_device *dev) 487 { 488 return smp_processor_id() % dev->real_num_rx_queues; 489 } 490 491 static struct net_device *veth_peer_dev(struct net_device *dev) 492 { 493 struct veth_priv *priv = netdev_priv(dev); 494 495 /* Callers must be under RCU read side. */ 496 return rcu_dereference(priv->peer); 497 } 498 499 static int veth_xdp_xmit(struct net_device *dev, int n, 500 struct xdp_frame **frames, 501 u32 flags, bool ndo_xmit) 502 { 503 struct veth_priv *rcv_priv, *priv = netdev_priv(dev); 504 int i, ret = -ENXIO, nxmit = 0; 505 struct net_device *rcv; 506 unsigned int max_len; 507 struct veth_rq *rq; 508 509 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) 510 return -EINVAL; 511 512 rcu_read_lock(); 513 rcv = rcu_dereference(priv->peer); 514 if (unlikely(!rcv)) 515 goto out; 516 517 rcv_priv = netdev_priv(rcv); 518 rq = &rcv_priv->rq[veth_select_rxq(rcv)]; 519 /* The napi pointer is set if NAPI is enabled, which ensures that 520 * xdp_ring is initialized on receive side and the peer device is up. 521 */ 522 if (!rcu_access_pointer(rq->napi)) 523 goto out; 524 525 max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN; 526 527 spin_lock(&rq->xdp_ring.producer_lock); 528 for (i = 0; i < n; i++) { 529 struct xdp_frame *frame = frames[i]; 530 void *ptr = veth_xdp_to_ptr(frame); 531 532 if (unlikely(xdp_get_frame_len(frame) > max_len || 533 __ptr_ring_produce(&rq->xdp_ring, ptr))) 534 break; 535 nxmit++; 536 } 537 spin_unlock(&rq->xdp_ring.producer_lock); 538 539 if (flags & XDP_XMIT_FLUSH) 540 __veth_xdp_flush(rq); 541 542 ret = nxmit; 543 if (ndo_xmit) { 544 u64_stats_update_begin(&rq->stats.syncp); 545 rq->stats.vs.peer_tq_xdp_xmit += nxmit; 546 rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit; 547 u64_stats_update_end(&rq->stats.syncp); 548 } 549 550 out: 551 rcu_read_unlock(); 552 553 return ret; 554 } 555 556 static int veth_ndo_xdp_xmit(struct net_device *dev, int n, 557 struct xdp_frame **frames, u32 flags) 558 { 559 int err; 560 561 err = veth_xdp_xmit(dev, n, frames, flags, true); 562 if (err < 0) { 563 struct veth_priv *priv = netdev_priv(dev); 564 565 atomic64_add(n, &priv->dropped); 566 } 567 568 return err; 569 } 570 571 static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq) 572 { 573 int sent, i, err = 0, drops; 574 575 sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false); 576 if (sent < 0) { 577 err = sent; 578 sent = 0; 579 } 580 581 for (i = sent; unlikely(i < bq->count); i++) 582 xdp_return_frame(bq->q[i]); 583 584 drops = bq->count - sent; 585 trace_xdp_bulk_tx(rq->dev, sent, drops, err); 586 587 u64_stats_update_begin(&rq->stats.syncp); 588 rq->stats.vs.xdp_tx += sent; 589 rq->stats.vs.xdp_tx_err += drops; 590 u64_stats_update_end(&rq->stats.syncp); 591 592 bq->count = 0; 593 } 594 595 static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq) 596 { 597 struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev); 598 struct net_device *rcv; 599 struct veth_rq *rcv_rq; 600 601 rcu_read_lock(); 602 veth_xdp_flush_bq(rq, bq); 603 rcv = rcu_dereference(priv->peer); 604 if (unlikely(!rcv)) 605 goto out; 606 607 rcv_priv = netdev_priv(rcv); 608 rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)]; 609 /* xdp_ring is initialized on receive side? */ 610 if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog))) 611 goto out; 612 613 __veth_xdp_flush(rcv_rq); 614 out: 615 rcu_read_unlock(); 616 } 617 618 static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp, 619 struct veth_xdp_tx_bq *bq) 620 { 621 struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp); 622 623 if (unlikely(!frame)) 624 return -EOVERFLOW; 625 626 if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE)) 627 veth_xdp_flush_bq(rq, bq); 628 629 bq->q[bq->count++] = frame; 630 631 return 0; 632 } 633 634 static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq, 635 struct xdp_frame *frame, 636 struct veth_xdp_tx_bq *bq, 637 struct veth_stats *stats) 638 { 639 struct xdp_frame orig_frame; 640 struct bpf_prog *xdp_prog; 641 642 rcu_read_lock(); 643 xdp_prog = rcu_dereference(rq->xdp_prog); 644 if (likely(xdp_prog)) { 645 struct veth_xdp_buff vxbuf; 646 struct xdp_buff *xdp = &vxbuf.xdp; 647 u32 act; 648 649 xdp_convert_frame_to_buff(frame, xdp); 650 xdp->rxq = &rq->xdp_rxq; 651 vxbuf.skb = NULL; 652 653 act = bpf_prog_run_xdp(xdp_prog, xdp); 654 655 switch (act) { 656 case XDP_PASS: 657 if (xdp_update_frame_from_buff(xdp, frame)) 658 goto err_xdp; 659 break; 660 case XDP_TX: 661 orig_frame = *frame; 662 xdp->rxq->mem.type = frame->mem_type; 663 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) { 664 trace_xdp_exception(rq->dev, xdp_prog, act); 665 frame = &orig_frame; 666 stats->rx_drops++; 667 goto err_xdp; 668 } 669 stats->xdp_tx++; 670 rcu_read_unlock(); 671 goto xdp_xmit; 672 case XDP_REDIRECT: 673 orig_frame = *frame; 674 xdp->rxq->mem.type = frame->mem_type; 675 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) { 676 frame = &orig_frame; 677 stats->rx_drops++; 678 goto err_xdp; 679 } 680 stats->xdp_redirect++; 681 rcu_read_unlock(); 682 goto xdp_xmit; 683 default: 684 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act); 685 fallthrough; 686 case XDP_ABORTED: 687 trace_xdp_exception(rq->dev, xdp_prog, act); 688 fallthrough; 689 case XDP_DROP: 690 stats->xdp_drops++; 691 goto err_xdp; 692 } 693 } 694 rcu_read_unlock(); 695 696 return frame; 697 err_xdp: 698 rcu_read_unlock(); 699 xdp_return_frame(frame); 700 xdp_xmit: 701 return NULL; 702 } 703 704 /* frames array contains VETH_XDP_BATCH at most */ 705 static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames, 706 int n_xdpf, struct veth_xdp_tx_bq *bq, 707 struct veth_stats *stats) 708 { 709 void *skbs[VETH_XDP_BATCH]; 710 int i; 711 712 if (unlikely(!napi_skb_cache_get_bulk(skbs, n_xdpf))) { 713 for (i = 0; i < n_xdpf; i++) 714 xdp_return_frame(frames[i]); 715 stats->rx_drops += n_xdpf; 716 717 return; 718 } 719 720 for (i = 0; i < n_xdpf; i++) { 721 struct sk_buff *skb = skbs[i]; 722 723 skb = __xdp_build_skb_from_frame(frames[i], skb, 724 rq->dev); 725 if (!skb) { 726 xdp_return_frame(frames[i]); 727 stats->rx_drops++; 728 continue; 729 } 730 napi_gro_receive(&rq->xdp_napi, skb); 731 } 732 } 733 734 static void veth_xdp_get(struct xdp_buff *xdp) 735 { 736 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp); 737 int i; 738 739 get_page(virt_to_page(xdp->data)); 740 if (likely(!xdp_buff_has_frags(xdp))) 741 return; 742 743 for (i = 0; i < sinfo->nr_frags; i++) 744 __skb_frag_ref(&sinfo->frags[i]); 745 } 746 747 static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq, 748 struct xdp_buff *xdp, 749 struct sk_buff **pskb) 750 { 751 struct sk_buff *skb = *pskb; 752 u32 frame_sz; 753 754 if (skb_shared(skb) || skb_head_is_locked(skb) || 755 skb_shinfo(skb)->nr_frags || 756 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 757 if (skb_pp_cow_data(rq->page_pool, pskb, XDP_PACKET_HEADROOM)) 758 goto drop; 759 760 skb = *pskb; 761 } 762 763 /* SKB "head" area always have tailroom for skb_shared_info */ 764 frame_sz = skb_end_pointer(skb) - skb->head; 765 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 766 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq); 767 xdp_prepare_buff(xdp, skb->head, skb_headroom(skb), 768 skb_headlen(skb), true); 769 770 if (skb_is_nonlinear(skb)) { 771 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 772 xdp_buff_set_frags_flag(xdp); 773 } else { 774 xdp_buff_clear_frags_flag(xdp); 775 } 776 *pskb = skb; 777 778 return 0; 779 drop: 780 consume_skb(skb); 781 *pskb = NULL; 782 783 return -ENOMEM; 784 } 785 786 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq, 787 struct sk_buff *skb, 788 struct veth_xdp_tx_bq *bq, 789 struct veth_stats *stats) 790 { 791 void *orig_data, *orig_data_end; 792 struct bpf_prog *xdp_prog; 793 struct veth_xdp_buff vxbuf; 794 struct xdp_buff *xdp = &vxbuf.xdp; 795 u32 act, metalen; 796 int off; 797 798 skb_prepare_for_gro(skb); 799 800 rcu_read_lock(); 801 xdp_prog = rcu_dereference(rq->xdp_prog); 802 if (unlikely(!xdp_prog)) { 803 rcu_read_unlock(); 804 goto out; 805 } 806 807 __skb_push(skb, skb->data - skb_mac_header(skb)); 808 if (veth_convert_skb_to_xdp_buff(rq, xdp, &skb)) 809 goto drop; 810 vxbuf.skb = skb; 811 812 orig_data = xdp->data; 813 orig_data_end = xdp->data_end; 814 815 act = bpf_prog_run_xdp(xdp_prog, xdp); 816 817 switch (act) { 818 case XDP_PASS: 819 break; 820 case XDP_TX: 821 veth_xdp_get(xdp); 822 consume_skb(skb); 823 xdp->rxq->mem = rq->xdp_mem; 824 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) { 825 trace_xdp_exception(rq->dev, xdp_prog, act); 826 stats->rx_drops++; 827 goto err_xdp; 828 } 829 stats->xdp_tx++; 830 rcu_read_unlock(); 831 goto xdp_xmit; 832 case XDP_REDIRECT: 833 veth_xdp_get(xdp); 834 consume_skb(skb); 835 xdp->rxq->mem = rq->xdp_mem; 836 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) { 837 stats->rx_drops++; 838 goto err_xdp; 839 } 840 stats->xdp_redirect++; 841 rcu_read_unlock(); 842 goto xdp_xmit; 843 default: 844 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act); 845 fallthrough; 846 case XDP_ABORTED: 847 trace_xdp_exception(rq->dev, xdp_prog, act); 848 fallthrough; 849 case XDP_DROP: 850 stats->xdp_drops++; 851 goto xdp_drop; 852 } 853 rcu_read_unlock(); 854 855 /* check if bpf_xdp_adjust_head was used */ 856 off = orig_data - xdp->data; 857 if (off > 0) 858 __skb_push(skb, off); 859 else if (off < 0) 860 __skb_pull(skb, -off); 861 862 skb_reset_mac_header(skb); 863 864 /* check if bpf_xdp_adjust_tail was used */ 865 off = xdp->data_end - orig_data_end; 866 if (off != 0) 867 __skb_put(skb, off); /* positive on grow, negative on shrink */ 868 869 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 870 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 871 */ 872 if (xdp_buff_has_frags(xdp)) 873 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 874 else 875 skb->data_len = 0; 876 877 skb->protocol = eth_type_trans(skb, rq->dev); 878 879 metalen = xdp->data - xdp->data_meta; 880 if (metalen) 881 skb_metadata_set(skb, metalen); 882 out: 883 return skb; 884 drop: 885 stats->rx_drops++; 886 xdp_drop: 887 rcu_read_unlock(); 888 kfree_skb(skb); 889 return NULL; 890 err_xdp: 891 rcu_read_unlock(); 892 xdp_return_buff(xdp); 893 xdp_xmit: 894 return NULL; 895 } 896 897 static int veth_xdp_rcv(struct veth_rq *rq, int budget, 898 struct veth_xdp_tx_bq *bq, 899 struct veth_stats *stats) 900 { 901 int i, done = 0, n_xdpf = 0; 902 void *xdpf[VETH_XDP_BATCH]; 903 904 for (i = 0; i < budget; i++) { 905 void *ptr = __ptr_ring_consume(&rq->xdp_ring); 906 907 if (!ptr) 908 break; 909 910 if (veth_is_xdp_frame(ptr)) { 911 /* ndo_xdp_xmit */ 912 struct xdp_frame *frame = veth_ptr_to_xdp(ptr); 913 914 stats->xdp_bytes += xdp_get_frame_len(frame); 915 frame = veth_xdp_rcv_one(rq, frame, bq, stats); 916 if (frame) { 917 /* XDP_PASS */ 918 xdpf[n_xdpf++] = frame; 919 if (n_xdpf == VETH_XDP_BATCH) { 920 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, 921 bq, stats); 922 n_xdpf = 0; 923 } 924 } 925 } else { 926 /* ndo_start_xmit */ 927 struct sk_buff *skb = ptr; 928 929 stats->xdp_bytes += skb->len; 930 skb = veth_xdp_rcv_skb(rq, skb, bq, stats); 931 if (skb) { 932 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC)) 933 netif_receive_skb(skb); 934 else 935 napi_gro_receive(&rq->xdp_napi, skb); 936 } 937 } 938 done++; 939 } 940 941 if (n_xdpf) 942 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats); 943 944 u64_stats_update_begin(&rq->stats.syncp); 945 rq->stats.vs.xdp_redirect += stats->xdp_redirect; 946 rq->stats.vs.xdp_bytes += stats->xdp_bytes; 947 rq->stats.vs.xdp_drops += stats->xdp_drops; 948 rq->stats.vs.rx_drops += stats->rx_drops; 949 rq->stats.vs.xdp_packets += done; 950 u64_stats_update_end(&rq->stats.syncp); 951 952 return done; 953 } 954 955 static int veth_poll(struct napi_struct *napi, int budget) 956 { 957 struct veth_rq *rq = 958 container_of(napi, struct veth_rq, xdp_napi); 959 struct veth_priv *priv = netdev_priv(rq->dev); 960 int queue_idx = rq->xdp_rxq.queue_index; 961 struct netdev_queue *peer_txq; 962 struct veth_stats stats = {}; 963 struct net_device *peer_dev; 964 struct veth_xdp_tx_bq bq; 965 int done; 966 967 bq.count = 0; 968 969 /* NAPI functions as RCU section */ 970 peer_dev = rcu_dereference_check(priv->peer, rcu_read_lock_bh_held()); 971 peer_txq = peer_dev ? netdev_get_tx_queue(peer_dev, queue_idx) : NULL; 972 973 xdp_set_return_frame_no_direct(); 974 done = veth_xdp_rcv(rq, budget, &bq, &stats); 975 976 if (stats.xdp_redirect > 0) 977 xdp_do_flush(); 978 if (stats.xdp_tx > 0) 979 veth_xdp_flush(rq, &bq); 980 xdp_clear_return_frame_no_direct(); 981 982 if (done < budget && napi_complete_done(napi, done)) { 983 /* Write rx_notify_masked before reading ptr_ring */ 984 smp_store_mb(rq->rx_notify_masked, false); 985 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) { 986 if (napi_schedule_prep(&rq->xdp_napi)) { 987 WRITE_ONCE(rq->rx_notify_masked, true); 988 __napi_schedule(&rq->xdp_napi); 989 } 990 } 991 } 992 993 /* Release backpressure per NAPI poll */ 994 smp_rmb(); /* Paired with netif_tx_stop_queue set_bit */ 995 if (peer_txq && netif_tx_queue_stopped(peer_txq)) { 996 txq_trans_cond_update(peer_txq); 997 netif_tx_wake_queue(peer_txq); 998 } 999 1000 return done; 1001 } 1002 1003 static int veth_create_page_pool(struct veth_rq *rq) 1004 { 1005 struct page_pool_params pp_params = { 1006 .order = 0, 1007 .pool_size = VETH_RING_SIZE, 1008 .nid = NUMA_NO_NODE, 1009 .dev = &rq->dev->dev, 1010 }; 1011 1012 rq->page_pool = page_pool_create(&pp_params); 1013 if (IS_ERR(rq->page_pool)) { 1014 int err = PTR_ERR(rq->page_pool); 1015 1016 rq->page_pool = NULL; 1017 return err; 1018 } 1019 1020 return 0; 1021 } 1022 1023 static int __veth_napi_enable_range(struct net_device *dev, int start, int end) 1024 { 1025 struct veth_priv *priv = netdev_priv(dev); 1026 int err, i; 1027 1028 for (i = start; i < end; i++) { 1029 err = veth_create_page_pool(&priv->rq[i]); 1030 if (err) 1031 goto err_page_pool; 1032 } 1033 1034 for (i = start; i < end; i++) { 1035 struct veth_rq *rq = &priv->rq[i]; 1036 1037 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL); 1038 if (err) 1039 goto err_xdp_ring; 1040 } 1041 1042 for (i = start; i < end; i++) { 1043 struct veth_rq *rq = &priv->rq[i]; 1044 1045 napi_enable(&rq->xdp_napi); 1046 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi); 1047 } 1048 1049 return 0; 1050 1051 err_xdp_ring: 1052 for (i--; i >= start; i--) 1053 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free); 1054 i = end; 1055 err_page_pool: 1056 for (i--; i >= start; i--) { 1057 page_pool_destroy(priv->rq[i].page_pool); 1058 priv->rq[i].page_pool = NULL; 1059 } 1060 1061 return err; 1062 } 1063 1064 static int __veth_napi_enable(struct net_device *dev) 1065 { 1066 return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues); 1067 } 1068 1069 static void veth_napi_del_range(struct net_device *dev, int start, int end) 1070 { 1071 struct veth_priv *priv = netdev_priv(dev); 1072 int i; 1073 1074 for (i = start; i < end; i++) { 1075 struct veth_rq *rq = &priv->rq[i]; 1076 1077 rcu_assign_pointer(priv->rq[i].napi, NULL); 1078 napi_disable(&rq->xdp_napi); 1079 __netif_napi_del(&rq->xdp_napi); 1080 } 1081 synchronize_net(); 1082 1083 for (i = start; i < end; i++) { 1084 struct veth_rq *rq = &priv->rq[i]; 1085 1086 rq->rx_notify_masked = false; 1087 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free); 1088 } 1089 1090 for (i = start; i < end; i++) { 1091 page_pool_destroy(priv->rq[i].page_pool); 1092 priv->rq[i].page_pool = NULL; 1093 } 1094 } 1095 1096 static void veth_napi_del(struct net_device *dev) 1097 { 1098 veth_napi_del_range(dev, 0, dev->real_num_rx_queues); 1099 } 1100 1101 static bool veth_gro_requested(const struct net_device *dev) 1102 { 1103 return !!(dev->wanted_features & NETIF_F_GRO); 1104 } 1105 1106 static int veth_enable_xdp_range(struct net_device *dev, int start, int end, 1107 bool napi_already_on) 1108 { 1109 struct veth_priv *priv = netdev_priv(dev); 1110 int err, i; 1111 1112 for (i = start; i < end; i++) { 1113 struct veth_rq *rq = &priv->rq[i]; 1114 1115 if (!napi_already_on) 1116 netif_napi_add(dev, &rq->xdp_napi, veth_poll); 1117 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id); 1118 if (err < 0) 1119 goto err_rxq_reg; 1120 1121 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq, 1122 MEM_TYPE_PAGE_SHARED, 1123 NULL); 1124 if (err < 0) 1125 goto err_reg_mem; 1126 1127 /* Save original mem info as it can be overwritten */ 1128 rq->xdp_mem = rq->xdp_rxq.mem; 1129 } 1130 return 0; 1131 1132 err_reg_mem: 1133 xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq); 1134 err_rxq_reg: 1135 for (i--; i >= start; i--) { 1136 struct veth_rq *rq = &priv->rq[i]; 1137 1138 xdp_rxq_info_unreg(&rq->xdp_rxq); 1139 if (!napi_already_on) 1140 netif_napi_del(&rq->xdp_napi); 1141 } 1142 1143 return err; 1144 } 1145 1146 static void veth_disable_xdp_range(struct net_device *dev, int start, int end, 1147 bool delete_napi) 1148 { 1149 struct veth_priv *priv = netdev_priv(dev); 1150 int i; 1151 1152 for (i = start; i < end; i++) { 1153 struct veth_rq *rq = &priv->rq[i]; 1154 1155 rq->xdp_rxq.mem = rq->xdp_mem; 1156 xdp_rxq_info_unreg(&rq->xdp_rxq); 1157 1158 if (delete_napi) 1159 netif_napi_del(&rq->xdp_napi); 1160 } 1161 } 1162 1163 static int veth_enable_xdp(struct net_device *dev) 1164 { 1165 bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP); 1166 struct veth_priv *priv = netdev_priv(dev); 1167 int err, i; 1168 1169 if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) { 1170 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on); 1171 if (err) 1172 return err; 1173 1174 if (!napi_already_on) { 1175 err = __veth_napi_enable(dev); 1176 if (err) { 1177 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true); 1178 return err; 1179 } 1180 } 1181 } 1182 1183 for (i = 0; i < dev->real_num_rx_queues; i++) { 1184 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog); 1185 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi); 1186 } 1187 1188 return 0; 1189 } 1190 1191 static void veth_disable_xdp(struct net_device *dev) 1192 { 1193 struct veth_priv *priv = netdev_priv(dev); 1194 int i; 1195 1196 for (i = 0; i < dev->real_num_rx_queues; i++) 1197 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL); 1198 1199 if (!netif_running(dev) || !veth_gro_requested(dev)) 1200 veth_napi_del(dev); 1201 1202 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false); 1203 } 1204 1205 static int veth_napi_enable_range(struct net_device *dev, int start, int end) 1206 { 1207 struct veth_priv *priv = netdev_priv(dev); 1208 int err, i; 1209 1210 for (i = start; i < end; i++) { 1211 struct veth_rq *rq = &priv->rq[i]; 1212 1213 netif_napi_add(dev, &rq->xdp_napi, veth_poll); 1214 } 1215 1216 err = __veth_napi_enable_range(dev, start, end); 1217 if (err) { 1218 for (i = start; i < end; i++) { 1219 struct veth_rq *rq = &priv->rq[i]; 1220 1221 netif_napi_del(&rq->xdp_napi); 1222 } 1223 return err; 1224 } 1225 return err; 1226 } 1227 1228 static int veth_napi_enable(struct net_device *dev) 1229 { 1230 return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues); 1231 } 1232 1233 static void veth_disable_range_safe(struct net_device *dev, int start, int end) 1234 { 1235 struct veth_priv *priv = netdev_priv(dev); 1236 1237 if (start >= end) 1238 return; 1239 1240 if (priv->_xdp_prog) { 1241 veth_napi_del_range(dev, start, end); 1242 veth_disable_xdp_range(dev, start, end, false); 1243 } else if (veth_gro_requested(dev)) { 1244 veth_napi_del_range(dev, start, end); 1245 } 1246 } 1247 1248 static int veth_enable_range_safe(struct net_device *dev, int start, int end) 1249 { 1250 struct veth_priv *priv = netdev_priv(dev); 1251 int err; 1252 1253 if (start >= end) 1254 return 0; 1255 1256 if (priv->_xdp_prog) { 1257 /* these channels are freshly initialized, napi is not on there even 1258 * when GRO is requeste 1259 */ 1260 err = veth_enable_xdp_range(dev, start, end, false); 1261 if (err) 1262 return err; 1263 1264 err = __veth_napi_enable_range(dev, start, end); 1265 if (err) { 1266 /* on error always delete the newly added napis */ 1267 veth_disable_xdp_range(dev, start, end, true); 1268 return err; 1269 } 1270 } else if (veth_gro_requested(dev)) { 1271 return veth_napi_enable_range(dev, start, end); 1272 } 1273 return 0; 1274 } 1275 1276 static void veth_set_xdp_features(struct net_device *dev) 1277 { 1278 struct veth_priv *priv = netdev_priv(dev); 1279 struct net_device *peer; 1280 1281 peer = rtnl_dereference(priv->peer); 1282 if (peer && peer->real_num_tx_queues <= dev->real_num_rx_queues) { 1283 struct veth_priv *priv_peer = netdev_priv(peer); 1284 xdp_features_t val = NETDEV_XDP_ACT_BASIC | 1285 NETDEV_XDP_ACT_REDIRECT | 1286 NETDEV_XDP_ACT_RX_SG; 1287 1288 if (priv_peer->_xdp_prog || veth_gro_requested(peer)) 1289 val |= NETDEV_XDP_ACT_NDO_XMIT | 1290 NETDEV_XDP_ACT_NDO_XMIT_SG; 1291 xdp_set_features_flag(dev, val); 1292 } else { 1293 xdp_clear_features_flag(dev); 1294 } 1295 } 1296 1297 static int veth_set_channels(struct net_device *dev, 1298 struct ethtool_channels *ch) 1299 { 1300 struct veth_priv *priv = netdev_priv(dev); 1301 unsigned int old_rx_count, new_rx_count; 1302 struct veth_priv *peer_priv; 1303 struct net_device *peer; 1304 int err; 1305 1306 /* sanity check. Upper bounds are already enforced by the caller */ 1307 if (!ch->rx_count || !ch->tx_count) 1308 return -EINVAL; 1309 1310 /* avoid braking XDP, if that is enabled */ 1311 peer = rtnl_dereference(priv->peer); 1312 peer_priv = peer ? netdev_priv(peer) : NULL; 1313 if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues) 1314 return -EINVAL; 1315 1316 if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues) 1317 return -EINVAL; 1318 1319 old_rx_count = dev->real_num_rx_queues; 1320 new_rx_count = ch->rx_count; 1321 if (netif_running(dev)) { 1322 /* turn device off */ 1323 netif_carrier_off(dev); 1324 if (peer) 1325 netif_carrier_off(peer); 1326 1327 /* try to allocate new resurces, as needed*/ 1328 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count); 1329 if (err) 1330 goto out; 1331 } 1332 1333 err = netif_set_real_num_rx_queues(dev, ch->rx_count); 1334 if (err) 1335 goto revert; 1336 1337 err = netif_set_real_num_tx_queues(dev, ch->tx_count); 1338 if (err) { 1339 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count); 1340 1341 /* this error condition could happen only if rx and tx change 1342 * in opposite directions (e.g. tx nr raises, rx nr decreases) 1343 * and we can't do anything to fully restore the original 1344 * status 1345 */ 1346 if (err2) 1347 pr_warn("Can't restore rx queues config %d -> %d %d", 1348 new_rx_count, old_rx_count, err2); 1349 else 1350 goto revert; 1351 } 1352 1353 out: 1354 if (netif_running(dev)) { 1355 /* note that we need to swap the arguments WRT the enable part 1356 * to identify the range we have to disable 1357 */ 1358 veth_disable_range_safe(dev, new_rx_count, old_rx_count); 1359 netif_carrier_on(dev); 1360 if (peer) 1361 netif_carrier_on(peer); 1362 } 1363 1364 /* update XDP supported features */ 1365 veth_set_xdp_features(dev); 1366 if (peer) 1367 veth_set_xdp_features(peer); 1368 1369 return err; 1370 1371 revert: 1372 new_rx_count = old_rx_count; 1373 old_rx_count = ch->rx_count; 1374 goto out; 1375 } 1376 1377 static int veth_open(struct net_device *dev) 1378 { 1379 struct veth_priv *priv = netdev_priv(dev); 1380 struct net_device *peer = rtnl_dereference(priv->peer); 1381 int err; 1382 1383 if (!peer) 1384 return -ENOTCONN; 1385 1386 if (priv->_xdp_prog) { 1387 err = veth_enable_xdp(dev); 1388 if (err) 1389 return err; 1390 } else if (veth_gro_requested(dev)) { 1391 err = veth_napi_enable(dev); 1392 if (err) 1393 return err; 1394 } 1395 1396 if (peer->flags & IFF_UP) { 1397 netif_carrier_on(dev); 1398 netif_carrier_on(peer); 1399 } 1400 1401 veth_set_xdp_features(dev); 1402 1403 return 0; 1404 } 1405 1406 static int veth_close(struct net_device *dev) 1407 { 1408 struct veth_priv *priv = netdev_priv(dev); 1409 struct net_device *peer = rtnl_dereference(priv->peer); 1410 1411 netif_carrier_off(dev); 1412 if (peer) 1413 netif_carrier_off(peer); 1414 1415 if (priv->_xdp_prog) 1416 veth_disable_xdp(dev); 1417 else if (veth_gro_requested(dev)) 1418 veth_napi_del(dev); 1419 1420 return 0; 1421 } 1422 1423 static int is_valid_veth_mtu(int mtu) 1424 { 1425 return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU; 1426 } 1427 1428 static int veth_alloc_queues(struct net_device *dev) 1429 { 1430 struct veth_priv *priv = netdev_priv(dev); 1431 int i; 1432 1433 priv->rq = kvcalloc(dev->num_rx_queues, sizeof(*priv->rq), 1434 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 1435 if (!priv->rq) 1436 return -ENOMEM; 1437 1438 for (i = 0; i < dev->num_rx_queues; i++) { 1439 priv->rq[i].dev = dev; 1440 u64_stats_init(&priv->rq[i].stats.syncp); 1441 } 1442 1443 return 0; 1444 } 1445 1446 static void veth_free_queues(struct net_device *dev) 1447 { 1448 struct veth_priv *priv = netdev_priv(dev); 1449 1450 kvfree(priv->rq); 1451 } 1452 1453 static int veth_dev_init(struct net_device *dev) 1454 { 1455 netdev_lockdep_set_classes(dev); 1456 return veth_alloc_queues(dev); 1457 } 1458 1459 static void veth_dev_free(struct net_device *dev) 1460 { 1461 veth_free_queues(dev); 1462 } 1463 1464 #ifdef CONFIG_NET_POLL_CONTROLLER 1465 static void veth_poll_controller(struct net_device *dev) 1466 { 1467 /* veth only receives frames when its peer sends one 1468 * Since it has nothing to do with disabling irqs, we are guaranteed 1469 * never to have pending data when we poll for it so 1470 * there is nothing to do here. 1471 * 1472 * We need this though so netpoll recognizes us as an interface that 1473 * supports polling, which enables bridge devices in virt setups to 1474 * still use netconsole 1475 */ 1476 } 1477 #endif /* CONFIG_NET_POLL_CONTROLLER */ 1478 1479 static int veth_get_iflink(const struct net_device *dev) 1480 { 1481 struct veth_priv *priv = netdev_priv(dev); 1482 struct net_device *peer; 1483 int iflink; 1484 1485 rcu_read_lock(); 1486 peer = rcu_dereference(priv->peer); 1487 iflink = peer ? READ_ONCE(peer->ifindex) : 0; 1488 rcu_read_unlock(); 1489 1490 return iflink; 1491 } 1492 1493 static netdev_features_t veth_fix_features(struct net_device *dev, 1494 netdev_features_t features) 1495 { 1496 struct veth_priv *priv = netdev_priv(dev); 1497 struct net_device *peer; 1498 1499 peer = rtnl_dereference(priv->peer); 1500 if (peer) { 1501 struct veth_priv *peer_priv = netdev_priv(peer); 1502 1503 if (peer_priv->_xdp_prog) 1504 features &= ~NETIF_F_GSO_SOFTWARE; 1505 } 1506 1507 return features; 1508 } 1509 1510 static int veth_set_features(struct net_device *dev, 1511 netdev_features_t features) 1512 { 1513 netdev_features_t changed = features ^ dev->features; 1514 struct veth_priv *priv = netdev_priv(dev); 1515 struct net_device *peer; 1516 int err; 1517 1518 if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog) 1519 return 0; 1520 1521 peer = rtnl_dereference(priv->peer); 1522 if (features & NETIF_F_GRO) { 1523 err = veth_napi_enable(dev); 1524 if (err) 1525 return err; 1526 1527 if (peer) 1528 xdp_features_set_redirect_target(peer, true); 1529 } else { 1530 if (peer) 1531 xdp_features_clear_redirect_target(peer); 1532 veth_napi_del(dev); 1533 } 1534 return 0; 1535 } 1536 1537 static void veth_set_rx_headroom(struct net_device *dev, int new_hr) 1538 { 1539 struct veth_priv *peer_priv, *priv = netdev_priv(dev); 1540 struct net_device *peer; 1541 1542 if (new_hr < 0) 1543 new_hr = 0; 1544 1545 rcu_read_lock(); 1546 peer = rcu_dereference(priv->peer); 1547 if (unlikely(!peer)) 1548 goto out; 1549 1550 peer_priv = netdev_priv(peer); 1551 priv->requested_headroom = new_hr; 1552 new_hr = max(priv->requested_headroom, peer_priv->requested_headroom); 1553 dev->needed_headroom = new_hr; 1554 peer->needed_headroom = new_hr; 1555 1556 out: 1557 rcu_read_unlock(); 1558 } 1559 1560 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog, 1561 struct netlink_ext_ack *extack) 1562 { 1563 struct veth_priv *priv = netdev_priv(dev); 1564 struct bpf_prog *old_prog; 1565 struct net_device *peer; 1566 unsigned int max_mtu; 1567 int err; 1568 1569 old_prog = priv->_xdp_prog; 1570 priv->_xdp_prog = prog; 1571 peer = rtnl_dereference(priv->peer); 1572 1573 if (prog) { 1574 if (!peer) { 1575 NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached"); 1576 err = -ENOTCONN; 1577 goto err; 1578 } 1579 1580 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) - 1581 peer->hard_header_len; 1582 /* Allow increasing the max_mtu if the program supports 1583 * XDP fragments. 1584 */ 1585 if (prog->aux->xdp_has_frags) 1586 max_mtu += PAGE_SIZE * MAX_SKB_FRAGS; 1587 1588 if (peer->mtu > max_mtu) { 1589 NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP"); 1590 err = -ERANGE; 1591 goto err; 1592 } 1593 1594 if (dev->real_num_rx_queues < peer->real_num_tx_queues) { 1595 NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues"); 1596 err = -ENOSPC; 1597 goto err; 1598 } 1599 1600 if (dev->flags & IFF_UP) { 1601 err = veth_enable_xdp(dev); 1602 if (err) { 1603 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed"); 1604 goto err; 1605 } 1606 } 1607 1608 if (!old_prog) { 1609 peer->hw_features &= ~NETIF_F_GSO_SOFTWARE; 1610 peer->max_mtu = max_mtu; 1611 } 1612 1613 xdp_features_set_redirect_target(peer, true); 1614 } 1615 1616 if (old_prog) { 1617 if (!prog) { 1618 if (peer && !veth_gro_requested(dev)) 1619 xdp_features_clear_redirect_target(peer); 1620 1621 if (dev->flags & IFF_UP) 1622 veth_disable_xdp(dev); 1623 1624 if (peer) { 1625 peer->hw_features |= NETIF_F_GSO_SOFTWARE; 1626 peer->max_mtu = ETH_MAX_MTU; 1627 } 1628 } 1629 bpf_prog_put(old_prog); 1630 } 1631 1632 if ((!!old_prog ^ !!prog) && peer) 1633 netdev_update_features(peer); 1634 1635 return 0; 1636 err: 1637 priv->_xdp_prog = old_prog; 1638 1639 return err; 1640 } 1641 1642 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp) 1643 { 1644 switch (xdp->command) { 1645 case XDP_SETUP_PROG: 1646 return veth_xdp_set(dev, xdp->prog, xdp->extack); 1647 default: 1648 return -EINVAL; 1649 } 1650 } 1651 1652 static int veth_xdp_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp) 1653 { 1654 struct veth_xdp_buff *_ctx = (void *)ctx; 1655 1656 if (!_ctx->skb) 1657 return -ENODATA; 1658 1659 *timestamp = skb_hwtstamps(_ctx->skb)->hwtstamp; 1660 return 0; 1661 } 1662 1663 static int veth_xdp_rx_hash(const struct xdp_md *ctx, u32 *hash, 1664 enum xdp_rss_hash_type *rss_type) 1665 { 1666 struct veth_xdp_buff *_ctx = (void *)ctx; 1667 struct sk_buff *skb = _ctx->skb; 1668 1669 if (!skb) 1670 return -ENODATA; 1671 1672 *hash = skb_get_hash(skb); 1673 *rss_type = skb->l4_hash ? XDP_RSS_TYPE_L4_ANY : XDP_RSS_TYPE_NONE; 1674 1675 return 0; 1676 } 1677 1678 static int veth_xdp_rx_vlan_tag(const struct xdp_md *ctx, __be16 *vlan_proto, 1679 u16 *vlan_tci) 1680 { 1681 const struct veth_xdp_buff *_ctx = (void *)ctx; 1682 const struct sk_buff *skb = _ctx->skb; 1683 int err; 1684 1685 if (!skb) 1686 return -ENODATA; 1687 1688 err = __vlan_hwaccel_get_tag(skb, vlan_tci); 1689 if (err) 1690 return err; 1691 1692 *vlan_proto = skb->vlan_proto; 1693 return err; 1694 } 1695 1696 static const struct net_device_ops veth_netdev_ops = { 1697 .ndo_init = veth_dev_init, 1698 .ndo_open = veth_open, 1699 .ndo_stop = veth_close, 1700 .ndo_start_xmit = veth_xmit, 1701 .ndo_get_stats64 = veth_get_stats64, 1702 .ndo_set_rx_mode = veth_set_multicast_list, 1703 .ndo_set_mac_address = eth_mac_addr, 1704 #ifdef CONFIG_NET_POLL_CONTROLLER 1705 .ndo_poll_controller = veth_poll_controller, 1706 #endif 1707 .ndo_get_iflink = veth_get_iflink, 1708 .ndo_fix_features = veth_fix_features, 1709 .ndo_set_features = veth_set_features, 1710 .ndo_features_check = passthru_features_check, 1711 .ndo_set_rx_headroom = veth_set_rx_headroom, 1712 .ndo_bpf = veth_xdp, 1713 .ndo_xdp_xmit = veth_ndo_xdp_xmit, 1714 .ndo_get_peer_dev = veth_peer_dev, 1715 }; 1716 1717 static const struct xdp_metadata_ops veth_xdp_metadata_ops = { 1718 .xmo_rx_timestamp = veth_xdp_rx_timestamp, 1719 .xmo_rx_hash = veth_xdp_rx_hash, 1720 .xmo_rx_vlan_tag = veth_xdp_rx_vlan_tag, 1721 }; 1722 1723 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \ 1724 NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \ 1725 NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \ 1726 NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \ 1727 NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX ) 1728 1729 static void veth_setup(struct net_device *dev) 1730 { 1731 ether_setup(dev); 1732 1733 dev->priv_flags &= ~IFF_TX_SKB_SHARING; 1734 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE; 1735 dev->priv_flags |= IFF_NO_QUEUE; 1736 dev->priv_flags |= IFF_PHONY_HEADROOM; 1737 dev->priv_flags |= IFF_DISABLE_NETPOLL; 1738 dev->lltx = true; 1739 1740 dev->netdev_ops = &veth_netdev_ops; 1741 dev->xdp_metadata_ops = &veth_xdp_metadata_ops; 1742 dev->ethtool_ops = &veth_ethtool_ops; 1743 dev->features |= VETH_FEATURES; 1744 dev->vlan_features = dev->features & 1745 ~(NETIF_F_HW_VLAN_CTAG_TX | 1746 NETIF_F_HW_VLAN_STAG_TX | 1747 NETIF_F_HW_VLAN_CTAG_RX | 1748 NETIF_F_HW_VLAN_STAG_RX); 1749 dev->needs_free_netdev = true; 1750 dev->priv_destructor = veth_dev_free; 1751 dev->pcpu_stat_type = NETDEV_PCPU_STAT_TSTATS; 1752 dev->max_mtu = ETH_MAX_MTU; 1753 1754 dev->hw_features = VETH_FEATURES; 1755 dev->hw_enc_features = VETH_FEATURES; 1756 dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE; 1757 netif_set_tso_max_size(dev, GSO_MAX_SIZE); 1758 } 1759 1760 /* 1761 * netlink interface 1762 */ 1763 1764 static int veth_validate(struct nlattr *tb[], struct nlattr *data[], 1765 struct netlink_ext_ack *extack) 1766 { 1767 if (tb[IFLA_ADDRESS]) { 1768 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) 1769 return -EINVAL; 1770 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) 1771 return -EADDRNOTAVAIL; 1772 } 1773 if (tb[IFLA_MTU]) { 1774 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU]))) 1775 return -EINVAL; 1776 } 1777 return 0; 1778 } 1779 1780 static struct rtnl_link_ops veth_link_ops; 1781 1782 static void veth_disable_gro(struct net_device *dev) 1783 { 1784 dev->features &= ~NETIF_F_GRO; 1785 dev->wanted_features &= ~NETIF_F_GRO; 1786 netdev_update_features(dev); 1787 } 1788 1789 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[]) 1790 { 1791 int err; 1792 1793 if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) { 1794 err = netif_set_real_num_tx_queues(dev, 1); 1795 if (err) 1796 return err; 1797 } 1798 if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) { 1799 err = netif_set_real_num_rx_queues(dev, 1); 1800 if (err) 1801 return err; 1802 } 1803 return 0; 1804 } 1805 1806 static int veth_newlink(struct net_device *dev, 1807 struct rtnl_newlink_params *params, 1808 struct netlink_ext_ack *extack) 1809 { 1810 struct net *peer_net = rtnl_newlink_peer_net(params); 1811 struct nlattr **data = params->data; 1812 struct nlattr **tb = params->tb; 1813 int err; 1814 struct net_device *peer; 1815 struct veth_priv *priv; 1816 char ifname[IFNAMSIZ]; 1817 struct nlattr *peer_tb[IFLA_MAX + 1], **tbp; 1818 unsigned char name_assign_type; 1819 struct ifinfomsg *ifmp; 1820 1821 /* 1822 * create and register peer first 1823 */ 1824 if (data && data[VETH_INFO_PEER]) { 1825 struct nlattr *nla_peer = data[VETH_INFO_PEER]; 1826 1827 ifmp = nla_data(nla_peer); 1828 rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack); 1829 tbp = peer_tb; 1830 } else { 1831 ifmp = NULL; 1832 tbp = tb; 1833 } 1834 1835 if (ifmp && tbp[IFLA_IFNAME]) { 1836 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ); 1837 name_assign_type = NET_NAME_USER; 1838 } else { 1839 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d"); 1840 name_assign_type = NET_NAME_ENUM; 1841 } 1842 1843 peer = rtnl_create_link(peer_net, ifname, name_assign_type, 1844 &veth_link_ops, tbp, extack); 1845 if (IS_ERR(peer)) 1846 return PTR_ERR(peer); 1847 1848 if (!ifmp || !tbp[IFLA_ADDRESS]) 1849 eth_hw_addr_random(peer); 1850 1851 if (ifmp && (dev->ifindex != 0)) 1852 peer->ifindex = ifmp->ifi_index; 1853 1854 netif_inherit_tso_max(peer, dev); 1855 1856 err = register_netdevice(peer); 1857 if (err < 0) 1858 goto err_register_peer; 1859 1860 /* keep GRO disabled by default to be consistent with the established 1861 * veth behavior 1862 */ 1863 veth_disable_gro(peer); 1864 netif_carrier_off(peer); 1865 1866 err = rtnl_configure_link(peer, ifmp, 0, NULL); 1867 if (err < 0) 1868 goto err_configure_peer; 1869 1870 /* 1871 * register dev last 1872 * 1873 * note, that since we've registered new device the dev's name 1874 * should be re-allocated 1875 */ 1876 1877 if (tb[IFLA_ADDRESS] == NULL) 1878 eth_hw_addr_random(dev); 1879 1880 if (tb[IFLA_IFNAME]) 1881 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ); 1882 else 1883 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d"); 1884 1885 err = register_netdevice(dev); 1886 if (err < 0) 1887 goto err_register_dev; 1888 1889 netif_carrier_off(dev); 1890 1891 /* 1892 * tie the deviced together 1893 */ 1894 1895 priv = netdev_priv(dev); 1896 rcu_assign_pointer(priv->peer, peer); 1897 err = veth_init_queues(dev, tb); 1898 if (err) 1899 goto err_queues; 1900 1901 priv = netdev_priv(peer); 1902 rcu_assign_pointer(priv->peer, dev); 1903 err = veth_init_queues(peer, tb); 1904 if (err) 1905 goto err_queues; 1906 1907 veth_disable_gro(dev); 1908 /* update XDP supported features */ 1909 veth_set_xdp_features(dev); 1910 veth_set_xdp_features(peer); 1911 1912 return 0; 1913 1914 err_queues: 1915 unregister_netdevice(dev); 1916 err_register_dev: 1917 /* nothing to do */ 1918 err_configure_peer: 1919 unregister_netdevice(peer); 1920 return err; 1921 1922 err_register_peer: 1923 free_netdev(peer); 1924 return err; 1925 } 1926 1927 static void veth_dellink(struct net_device *dev, struct list_head *head) 1928 { 1929 struct veth_priv *priv; 1930 struct net_device *peer; 1931 1932 priv = netdev_priv(dev); 1933 peer = rtnl_dereference(priv->peer); 1934 1935 /* Note : dellink() is called from default_device_exit_batch(), 1936 * before a rcu_synchronize() point. The devices are guaranteed 1937 * not being freed before one RCU grace period. 1938 */ 1939 RCU_INIT_POINTER(priv->peer, NULL); 1940 unregister_netdevice_queue(dev, head); 1941 1942 if (peer) { 1943 priv = netdev_priv(peer); 1944 RCU_INIT_POINTER(priv->peer, NULL); 1945 unregister_netdevice_queue(peer, head); 1946 } 1947 } 1948 1949 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = { 1950 [VETH_INFO_PEER] = { .len = sizeof(struct ifinfomsg) }, 1951 }; 1952 1953 static struct net *veth_get_link_net(const struct net_device *dev) 1954 { 1955 struct veth_priv *priv = netdev_priv(dev); 1956 struct net_device *peer = rtnl_dereference(priv->peer); 1957 1958 return peer ? dev_net(peer) : dev_net(dev); 1959 } 1960 1961 static unsigned int veth_get_num_queues(void) 1962 { 1963 /* enforce the same queue limit as rtnl_create_link */ 1964 int queues = num_possible_cpus(); 1965 1966 if (queues > 4096) 1967 queues = 4096; 1968 return queues; 1969 } 1970 1971 static struct rtnl_link_ops veth_link_ops = { 1972 .kind = DRV_NAME, 1973 .priv_size = sizeof(struct veth_priv), 1974 .setup = veth_setup, 1975 .validate = veth_validate, 1976 .newlink = veth_newlink, 1977 .dellink = veth_dellink, 1978 .policy = veth_policy, 1979 .peer_type = VETH_INFO_PEER, 1980 .maxtype = VETH_INFO_MAX, 1981 .get_link_net = veth_get_link_net, 1982 .get_num_tx_queues = veth_get_num_queues, 1983 .get_num_rx_queues = veth_get_num_queues, 1984 }; 1985 1986 /* 1987 * init/fini 1988 */ 1989 1990 static __init int veth_init(void) 1991 { 1992 return rtnl_link_register(&veth_link_ops); 1993 } 1994 1995 static __exit void veth_exit(void) 1996 { 1997 rtnl_link_unregister(&veth_link_ops); 1998 } 1999 2000 module_init(veth_init); 2001 module_exit(veth_exit); 2002 2003 MODULE_DESCRIPTION("Virtual Ethernet Tunnel"); 2004 MODULE_LICENSE("GPL v2"); 2005 MODULE_ALIAS_RTNL_LINK(DRV_NAME); 2006