drivers/net/veth.c

   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 <[email protected]>
   8  * Ethtool interface from: Eric W. Biederman <[email protected]>
   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, &ethtool_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,
 794                                           page_address(page) + page_offset,
 795                                           size)) {
 796                                 consume_skb(nskb);
 797                                 goto drop;
 798                         }
 799
 800                         len -= size;
 801                         off += size;
 802                 }
 803
 804                 consume_skb(skb);
 805                 skb = nskb;
 806         }
 807
 808         /* SKB "head" area always have tailroom for skb_shared_info */
 809         frame_sz = skb_end_pointer(skb) - skb->head;
 810         frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 811         xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
 812         xdp_prepare_buff(xdp, skb->head, skb_headroom(skb),
 813                          skb_headlen(skb), true);
 814
 815         if (skb_is_nonlinear(skb)) {
 816                 skb_shinfo(skb)->xdp_frags_size = skb->data_len;
 817                 xdp_buff_set_frags_flag(xdp);
 818         } else {
 819                 xdp_buff_clear_frags_flag(xdp);
 820         }
 821         *pskb = skb;
 822
 823         return 0;
 824 drop:
 825         consume_skb(skb);
 826         *pskb = NULL;
 827
 828         return -ENOMEM;
 829 }
 830
 831 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq,
 832                                         struct sk_buff *skb,
 833                                         struct veth_xdp_tx_bq *bq,
 834                                         struct veth_stats *stats)
 835 {
 836         void *orig_data, *orig_data_end;
 837         struct bpf_prog *xdp_prog;
 838         struct veth_xdp_buff vxbuf;
 839         struct xdp_buff *xdp = &vxbuf.xdp;
 840         u32 act, metalen;
 841         int off;
 842
 843         skb_prepare_for_gro(skb);
 844
 845         rcu_read_lock();
 846         xdp_prog = rcu_dereference(rq->xdp_prog);
 847         if (unlikely(!xdp_prog)) {
 848                 rcu_read_unlock();
 849                 goto out;
 850         }
 851
 852         __skb_push(skb, skb->data - skb_mac_header(skb));
 853         if (veth_convert_skb_to_xdp_buff(rq, xdp, &skb))
 854                 goto drop;
 855         vxbuf.skb = skb;
 856
 857         orig_data = xdp->data;
 858         orig_data_end = xdp->data_end;
 859
 860         act = bpf_prog_run_xdp(xdp_prog, xdp);
 861
 862         switch (act) {
 863         case XDP_PASS:
 864                 break;
 865         case XDP_TX:
 866                 veth_xdp_get(xdp);
 867                 consume_skb(skb);
 868                 xdp->rxq->mem = rq->xdp_mem;
 869                 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) {
 870                         trace_xdp_exception(rq->dev, xdp_prog, act);
 871                         stats->rx_drops++;
 872                         goto err_xdp;
 873                 }
 874                 stats->xdp_tx++;
 875                 rcu_read_unlock();
 876                 goto xdp_xmit;
 877         case XDP_REDIRECT:
 878                 veth_xdp_get(xdp);
 879                 consume_skb(skb);
 880                 xdp->rxq->mem = rq->xdp_mem;
 881                 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) {
 882                         stats->rx_drops++;
 883                         goto err_xdp;
 884                 }
 885                 stats->xdp_redirect++;
 886                 rcu_read_unlock();
 887                 goto xdp_xmit;
 888         default:
 889                 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
 890                 fallthrough;
 891         case XDP_ABORTED:
 892                 trace_xdp_exception(rq->dev, xdp_prog, act);
 893                 fallthrough;
 894         case XDP_DROP:
 895                 stats->xdp_drops++;
 896                 goto xdp_drop;
 897         }
 898         rcu_read_unlock();
 899
 900         /* check if bpf_xdp_adjust_head was used */
 901         off = orig_data - xdp->data;
 902         if (off > 0)
 903                 __skb_push(skb, off);
 904         else if (off < 0)
 905                 __skb_pull(skb, -off);
 906
 907         skb_reset_mac_header(skb);
 908
 909         /* check if bpf_xdp_adjust_tail was used */
 910         off = xdp->data_end - orig_data_end;
 911         if (off != 0)
 912                 __skb_put(skb, off); /* positive on grow, negative on shrink */
 913
 914         /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers
 915          * (e.g. bpf_xdp_adjust_tail), we need to update data_len here.
 916          */
 917         if (xdp_buff_has_frags(xdp))
 918                 skb->data_len = skb_shinfo(skb)->xdp_frags_size;
 919         else
 920                 skb->data_len = 0;
 921
 922         skb->protocol = eth_type_trans(skb, rq->dev);
 923
 924         metalen = xdp->data - xdp->data_meta;
 925         if (metalen)
 926                 skb_metadata_set(skb, metalen);
 927 out:
 928         return skb;
 929 drop:
 930         stats->rx_drops++;
 931 xdp_drop:
 932         rcu_read_unlock();
 933         kfree_skb(skb);
 934         return NULL;
 935 err_xdp:
 936         rcu_read_unlock();
 937         xdp_return_buff(xdp);
 938 xdp_xmit:
 939         return NULL;
 940 }
 941
 942 static int veth_xdp_rcv(struct veth_rq *rq, int budget,
 943                         struct veth_xdp_tx_bq *bq,
 944                         struct veth_stats *stats)
 945 {
 946         int i, done = 0, n_xdpf = 0;
 947         void *xdpf[VETH_XDP_BATCH];
 948
 949         for (i = 0; i < budget; i++) {
 950                 void *ptr = __ptr_ring_consume(&rq->xdp_ring);
 951
 952                 if (!ptr)
 953                         break;
 954
 955                 if (veth_is_xdp_frame(ptr)) {
 956                         /* ndo_xdp_xmit */
 957                         struct xdp_frame *frame = veth_ptr_to_xdp(ptr);
 958
 959                         stats->xdp_bytes += xdp_get_frame_len(frame);
 960                         frame = veth_xdp_rcv_one(rq, frame, bq, stats);
 961                         if (frame) {
 962                                 /* XDP_PASS */
 963                                 xdpf[n_xdpf++] = frame;
 964                                 if (n_xdpf == VETH_XDP_BATCH) {
 965                                         veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf,
 966                                                               bq, stats);
 967                                         n_xdpf = 0;
 968                                 }
 969                         }
 970                 } else {
 971                         /* ndo_start_xmit */
 972                         struct sk_buff *skb = ptr;
 973
 974                         stats->xdp_bytes += skb->len;
 975                         skb = veth_xdp_rcv_skb(rq, skb, bq, stats);
 976                         if (skb) {
 977                                 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC))
 978                                         netif_receive_skb(skb);
 979                                 else
 980                                         napi_gro_receive(&rq->xdp_napi, skb);
 981                         }
 982                 }
 983                 done++;
 984         }
 985
 986         if (n_xdpf)
 987                 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats);
 988
 989         u64_stats_update_begin(&rq->stats.syncp);
 990         rq->stats.vs.xdp_redirect += stats->xdp_redirect;
 991         rq->stats.vs.xdp_bytes += stats->xdp_bytes;
 992         rq->stats.vs.xdp_drops += stats->xdp_drops;
 993         rq->stats.vs.rx_drops += stats->rx_drops;
 994         rq->stats.vs.xdp_packets += done;
 995         u64_stats_update_end(&rq->stats.syncp);
 996
 997         return done;
 998 }
 999
1000 static int veth_poll(struct napi_struct *napi, int budget)
1001 {
1002         struct veth_rq *rq =
1003                 container_of(napi, struct veth_rq, xdp_napi);
1004         struct veth_stats stats = {};
1005         struct veth_xdp_tx_bq bq;
1006         int done;
1007
1008         bq.count = 0;
1009
1010         xdp_set_return_frame_no_direct();
1011         done = veth_xdp_rcv(rq, budget, &bq, &stats);
1012
1013         if (stats.xdp_redirect > 0)
1014                 xdp_do_flush();
1015
1016         if (done < budget && napi_complete_done(napi, done)) {
1017                 /* Write rx_notify_masked before reading ptr_ring */
1018                 smp_store_mb(rq->rx_notify_masked, false);
1019                 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
1020                         if (napi_schedule_prep(&rq->xdp_napi)) {
1021                                 WRITE_ONCE(rq->rx_notify_masked, true);
1022                                 __napi_schedule(&rq->xdp_napi);
1023                         }
1024                 }
1025         }
1026
1027         if (stats.xdp_tx > 0)
1028                 veth_xdp_flush(rq, &bq);
1029         xdp_clear_return_frame_no_direct();
1030
1031         return done;
1032 }
1033
1034 static int veth_create_page_pool(struct veth_rq *rq)
1035 {
1036         struct page_pool_params pp_params = {
1037                 .order = 0,
1038                 .pool_size = VETH_RING_SIZE,
1039                 .nid = NUMA_NO_NODE,
1040                 .dev = &rq->dev->dev,
1041         };
1042
1043         rq->page_pool = page_pool_create(&pp_params);
1044         if (IS_ERR(rq->page_pool)) {
1045                 int err = PTR_ERR(rq->page_pool);
1046
1047                 rq->page_pool = NULL;
1048                 return err;
1049         }
1050
1051         return 0;
1052 }
1053
1054 static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
1055 {
1056         struct veth_priv *priv = netdev_priv(dev);
1057         int err, i;
1058
1059         for (i = start; i < end; i++) {
1060                 err = veth_create_page_pool(&priv->rq[i]);
1061                 if (err)
1062                         goto err_page_pool;
1063         }
1064
1065         for (i = start; i < end; i++) {
1066                 struct veth_rq *rq = &priv->rq[i];
1067
1068                 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
1069                 if (err)
1070                         goto err_xdp_ring;
1071         }
1072
1073         for (i = start; i < end; i++) {
1074                 struct veth_rq *rq = &priv->rq[i];
1075
1076                 napi_enable(&rq->xdp_napi);
1077                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1078         }
1079
1080         return 0;
1081
1082 err_xdp_ring:
1083         for (i--; i >= start; i--)
1084                 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
1085         i = end;
1086 err_page_pool:
1087         for (i--; i >= start; i--) {
1088                 page_pool_destroy(priv->rq[i].page_pool);
1089                 priv->rq[i].page_pool = NULL;
1090         }
1091
1092         return err;
1093 }
1094
1095 static int __veth_napi_enable(struct net_device *dev)
1096 {
1097         return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1098 }
1099
1100 static void veth_napi_del_range(struct net_device *dev, int start, int end)
1101 {
1102         struct veth_priv *priv = netdev_priv(dev);
1103         int i;
1104
1105         for (i = start; i < end; i++) {
1106                 struct veth_rq *rq = &priv->rq[i];
1107
1108                 rcu_assign_pointer(priv->rq[i].napi, NULL);
1109                 napi_disable(&rq->xdp_napi);
1110                 __netif_napi_del(&rq->xdp_napi);
1111         }
1112         synchronize_net();
1113
1114         for (i = start; i < end; i++) {
1115                 struct veth_rq *rq = &priv->rq[i];
1116
1117                 rq->rx_notify_masked = false;
1118                 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
1119         }
1120
1121         for (i = start; i < end; i++) {
1122                 page_pool_destroy(priv->rq[i].page_pool);
1123                 priv->rq[i].page_pool = NULL;
1124         }
1125 }
1126
1127 static void veth_napi_del(struct net_device *dev)
1128 {
1129         veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
1130 }
1131
1132 static bool veth_gro_requested(const struct net_device *dev)
1133 {
1134         return !!(dev->wanted_features & NETIF_F_GRO);
1135 }
1136
1137 static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1138                                  bool napi_already_on)
1139 {
1140         struct veth_priv *priv = netdev_priv(dev);
1141         int err, i;
1142
1143         for (i = start; i < end; i++) {
1144                 struct veth_rq *rq = &priv->rq[i];
1145
1146                 if (!napi_already_on)
1147                         netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1148                 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1149                 if (err < 0)
1150                         goto err_rxq_reg;
1151
1152                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1153                                                  MEM_TYPE_PAGE_SHARED,
1154                                                  NULL);
1155                 if (err < 0)
1156                         goto err_reg_mem;
1157
1158                 /* Save original mem info as it can be overwritten */
1159                 rq->xdp_mem = rq->xdp_rxq.mem;
1160         }
1161         return 0;
1162
1163 err_reg_mem:
1164         xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1165 err_rxq_reg:
1166         for (i--; i >= start; i--) {
1167                 struct veth_rq *rq = &priv->rq[i];
1168
1169                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1170                 if (!napi_already_on)
1171                         netif_napi_del(&rq->xdp_napi);
1172         }
1173
1174         return err;
1175 }
1176
1177 static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1178                                    bool delete_napi)
1179 {
1180         struct veth_priv *priv = netdev_priv(dev);
1181         int i;
1182
1183         for (i = start; i < end; i++) {
1184                 struct veth_rq *rq = &priv->rq[i];
1185
1186                 rq->xdp_rxq.mem = rq->xdp_mem;
1187                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1188
1189                 if (delete_napi)
1190                         netif_napi_del(&rq->xdp_napi);
1191         }
1192 }
1193
1194 static int veth_enable_xdp(struct net_device *dev)
1195 {
1196         bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1197         struct veth_priv *priv = netdev_priv(dev);
1198         int err, i;
1199
1200         if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1201                 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1202                 if (err)
1203                         return err;
1204
1205                 if (!napi_already_on) {
1206                         err = __veth_napi_enable(dev);
1207                         if (err) {
1208                                 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1209                                 return err;
1210                         }
1211
1212                         if (!veth_gro_requested(dev)) {
1213                                 /* user-space did not require GRO, but adding XDP
1214                                  * is supposed to get GRO working
1215                                  */
1216                                 dev->features |= NETIF_F_GRO;
1217                                 netdev_features_change(dev);
1218                         }
1219                 }
1220         }
1221
1222         for (i = 0; i < dev->real_num_rx_queues; i++) {
1223                 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1224                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1225         }
1226
1227         return 0;
1228 }
1229
1230 static void veth_disable_xdp(struct net_device *dev)
1231 {
1232         struct veth_priv *priv = netdev_priv(dev);
1233         int i;
1234
1235         for (i = 0; i < dev->real_num_rx_queues; i++)
1236                 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1237
1238         if (!netif_running(dev) || !veth_gro_requested(dev)) {
1239                 veth_napi_del(dev);
1240
1241                 /* if user-space did not require GRO, since adding XDP
1242                  * enabled it, clear it now
1243                  */
1244                 if (!veth_gro_requested(dev) && netif_running(dev)) {
1245                         dev->features &= ~NETIF_F_GRO;
1246                         netdev_features_change(dev);
1247                 }
1248         }
1249
1250         veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1251 }
1252
1253 static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1254 {
1255         struct veth_priv *priv = netdev_priv(dev);
1256         int err, i;
1257
1258         for (i = start; i < end; i++) {
1259                 struct veth_rq *rq = &priv->rq[i];
1260
1261                 netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1262         }
1263
1264         err = __veth_napi_enable_range(dev, start, end);
1265         if (err) {
1266                 for (i = start; i < end; i++) {
1267                         struct veth_rq *rq = &priv->rq[i];
1268
1269                         netif_napi_del(&rq->xdp_napi);
1270                 }
1271                 return err;
1272         }
1273         return err;
1274 }
1275
1276 static int veth_napi_enable(struct net_device *dev)
1277 {
1278         return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1279 }
1280
1281 static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1282 {
1283         struct veth_priv *priv = netdev_priv(dev);
1284
1285         if (start >= end)
1286                 return;
1287
1288         if (priv->_xdp_prog) {
1289                 veth_napi_del_range(dev, start, end);
1290                 veth_disable_xdp_range(dev, start, end, false);
1291         } else if (veth_gro_requested(dev)) {
1292                 veth_napi_del_range(dev, start, end);
1293         }
1294 }
1295
1296 static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1297 {
1298         struct veth_priv *priv = netdev_priv(dev);
1299         int err;
1300
1301         if (start >= end)
1302                 return 0;
1303
1304         if (priv->_xdp_prog) {
1305                 /* these channels are freshly initialized, napi is not on there even
1306                  * when GRO is requeste
1307                  */
1308                 err = veth_enable_xdp_range(dev, start, end, false);
1309                 if (err)
1310                         return err;
1311
1312                 err = __veth_napi_enable_range(dev, start, end);
1313                 if (err) {
1314                         /* on error always delete the newly added napis */
1315                         veth_disable_xdp_range(dev, start, end, true);
1316                         return err;
1317                 }
1318         } else if (veth_gro_requested(dev)) {
1319                 return veth_napi_enable_range(dev, start, end);
1320         }
1321         return 0;
1322 }
1323
1324 static void veth_set_xdp_features(struct net_device *dev)
1325 {
1326         struct veth_priv *priv = netdev_priv(dev);
1327         struct net_device *peer;
1328
1329         peer = rtnl_dereference(priv->peer);
1330         if (peer && peer->real_num_tx_queues <= dev->real_num_rx_queues) {
1331                 struct veth_priv *priv_peer = netdev_priv(peer);
1332                 xdp_features_t val = NETDEV_XDP_ACT_BASIC |
1333                                      NETDEV_XDP_ACT_REDIRECT |
1334                                      NETDEV_XDP_ACT_RX_SG;
1335
1336                 if (priv_peer->_xdp_prog || veth_gro_requested(peer))
1337                         val |= NETDEV_XDP_ACT_NDO_XMIT |
1338                                NETDEV_XDP_ACT_NDO_XMIT_SG;
1339                 xdp_set_features_flag(dev, val);
1340         } else {
1341                 xdp_clear_features_flag(dev);
1342         }
1343 }
1344
1345 static int veth_set_channels(struct net_device *dev,
1346                              struct ethtool_channels *ch)
1347 {
1348         struct veth_priv *priv = netdev_priv(dev);
1349         unsigned int old_rx_count, new_rx_count;
1350         struct veth_priv *peer_priv;
1351         struct net_device *peer;
1352         int err;
1353
1354         /* sanity check. Upper bounds are already enforced by the caller */
1355         if (!ch->rx_count || !ch->tx_count)
1356                 return -EINVAL;
1357
1358         /* avoid braking XDP, if that is enabled */
1359         peer = rtnl_dereference(priv->peer);
1360         peer_priv = peer ? netdev_priv(peer) : NULL;
1361         if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1362                 return -EINVAL;
1363
1364         if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1365                 return -EINVAL;
1366
1367         old_rx_count = dev->real_num_rx_queues;
1368         new_rx_count = ch->rx_count;
1369         if (netif_running(dev)) {
1370                 /* turn device off */
1371                 netif_carrier_off(dev);
1372                 if (peer)
1373                         netif_carrier_off(peer);
1374
1375                 /* try to allocate new resurces, as needed*/
1376                 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1377                 if (err)
1378                         goto out;
1379         }
1380
1381         err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1382         if (err)
1383                 goto revert;
1384
1385         err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1386         if (err) {
1387                 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1388
1389                 /* this error condition could happen only if rx and tx change
1390                  * in opposite directions (e.g. tx nr raises, rx nr decreases)
1391                  * and we can't do anything to fully restore the original
1392                  * status
1393                  */
1394                 if (err2)
1395                         pr_warn("Can't restore rx queues config %d -> %d %d",
1396                                 new_rx_count, old_rx_count, err2);
1397                 else
1398                         goto revert;
1399         }
1400
1401 out:
1402         if (netif_running(dev)) {
1403                 /* note that we need to swap the arguments WRT the enable part
1404                  * to identify the range we have to disable
1405                  */
1406                 veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1407                 netif_carrier_on(dev);
1408                 if (peer)
1409                         netif_carrier_on(peer);
1410         }
1411
1412         /* update XDP supported features */
1413         veth_set_xdp_features(dev);
1414         if (peer)
1415                 veth_set_xdp_features(peer);
1416
1417         return err;
1418
1419 revert:
1420         new_rx_count = old_rx_count;
1421         old_rx_count = ch->rx_count;
1422         goto out;
1423 }
1424
1425 static int veth_open(struct net_device *dev)
1426 {
1427         struct veth_priv *priv = netdev_priv(dev);
1428         struct net_device *peer = rtnl_dereference(priv->peer);
1429         int err;
1430
1431         if (!peer)
1432                 return -ENOTCONN;
1433
1434         if (priv->_xdp_prog) {
1435                 err = veth_enable_xdp(dev);
1436                 if (err)
1437                         return err;
1438         } else if (veth_gro_requested(dev)) {
1439                 err = veth_napi_enable(dev);
1440                 if (err)
1441                         return err;
1442         }
1443
1444         if (peer->flags & IFF_UP) {
1445                 netif_carrier_on(dev);
1446                 netif_carrier_on(peer);
1447         }
1448
1449         veth_set_xdp_features(dev);
1450
1451         return 0;
1452 }
1453
1454 static int veth_close(struct net_device *dev)
1455 {
1456         struct veth_priv *priv = netdev_priv(dev);
1457         struct net_device *peer = rtnl_dereference(priv->peer);
1458
1459         netif_carrier_off(dev);
1460         if (peer)
1461                 netif_carrier_off(peer);
1462
1463         if (priv->_xdp_prog)
1464                 veth_disable_xdp(dev);
1465         else if (veth_gro_requested(dev))
1466                 veth_napi_del(dev);
1467
1468         return 0;
1469 }
1470
1471 static int is_valid_veth_mtu(int mtu)
1472 {
1473         return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1474 }
1475
1476 static int veth_alloc_queues(struct net_device *dev)
1477 {
1478         struct veth_priv *priv = netdev_priv(dev);
1479         int i;
1480
1481         priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL_ACCOUNT);
1482         if (!priv->rq)
1483                 return -ENOMEM;
1484
1485         for (i = 0; i < dev->num_rx_queues; i++) {
1486                 priv->rq[i].dev = dev;
1487                 u64_stats_init(&priv->rq[i].stats.syncp);
1488         }
1489
1490         return 0;
1491 }
1492
1493 static void veth_free_queues(struct net_device *dev)
1494 {
1495         struct veth_priv *priv = netdev_priv(dev);
1496
1497         kfree(priv->rq);
1498 }
1499
1500 static int veth_dev_init(struct net_device *dev)
1501 {
1502         return veth_alloc_queues(dev);
1503 }
1504
1505 static void veth_dev_free(struct net_device *dev)
1506 {
1507         veth_free_queues(dev);
1508 }
1509
1510 #ifdef CONFIG_NET_POLL_CONTROLLER
1511 static void veth_poll_controller(struct net_device *dev)
1512 {
1513         /* veth only receives frames when its peer sends one
1514          * Since it has nothing to do with disabling irqs, we are guaranteed
1515          * never to have pending data when we poll for it so
1516          * there is nothing to do here.
1517          *
1518          * We need this though so netpoll recognizes us as an interface that
1519          * supports polling, which enables bridge devices in virt setups to
1520          * still use netconsole
1521          */
1522 }
1523 #endif  /* CONFIG_NET_POLL_CONTROLLER */
1524
1525 static int veth_get_iflink(const struct net_device *dev)
1526 {
1527         struct veth_priv *priv = netdev_priv(dev);
1528         struct net_device *peer;
1529         int iflink;
1530
1531         rcu_read_lock();
1532         peer = rcu_dereference(priv->peer);
1533         iflink = peer ? peer->ifindex : 0;
1534         rcu_read_unlock();
1535
1536         return iflink;
1537 }
1538
1539 static netdev_features_t veth_fix_features(struct net_device *dev,
1540                                            netdev_features_t features)
1541 {
1542         struct veth_priv *priv = netdev_priv(dev);
1543         struct net_device *peer;
1544
1545         peer = rtnl_dereference(priv->peer);
1546         if (peer) {
1547                 struct veth_priv *peer_priv = netdev_priv(peer);
1548
1549                 if (peer_priv->_xdp_prog)
1550                         features &= ~NETIF_F_GSO_SOFTWARE;
1551         }
1552         if (priv->_xdp_prog)
1553                 features |= NETIF_F_GRO;
1554
1555         return features;
1556 }
1557
1558 static int veth_set_features(struct net_device *dev,
1559                              netdev_features_t features)
1560 {
1561         netdev_features_t changed = features ^ dev->features;
1562         struct veth_priv *priv = netdev_priv(dev);
1563         struct net_device *peer;
1564         int err;
1565
1566         if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1567                 return 0;
1568
1569         peer = rtnl_dereference(priv->peer);
1570         if (features & NETIF_F_GRO) {
1571                 err = veth_napi_enable(dev);
1572                 if (err)
1573                         return err;
1574
1575                 if (peer)
1576                         xdp_features_set_redirect_target(peer, true);
1577         } else {
1578                 if (peer)
1579                         xdp_features_clear_redirect_target(peer);
1580                 veth_napi_del(dev);
1581         }
1582         return 0;
1583 }
1584
1585 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1586 {
1587         struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1588         struct net_device *peer;
1589
1590         if (new_hr < 0)
1591                 new_hr = 0;
1592
1593         rcu_read_lock();
1594         peer = rcu_dereference(priv->peer);
1595         if (unlikely(!peer))
1596                 goto out;
1597
1598         peer_priv = netdev_priv(peer);
1599         priv->requested_headroom = new_hr;
1600         new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1601         dev->needed_headroom = new_hr;
1602         peer->needed_headroom = new_hr;
1603
1604 out:
1605         rcu_read_unlock();
1606 }
1607
1608 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1609                         struct netlink_ext_ack *extack)
1610 {
1611         struct veth_priv *priv = netdev_priv(dev);
1612         struct bpf_prog *old_prog;
1613         struct net_device *peer;
1614         unsigned int max_mtu;
1615         int err;
1616
1617         old_prog = priv->_xdp_prog;
1618         priv->_xdp_prog = prog;
1619         peer = rtnl_dereference(priv->peer);
1620
1621         if (prog) {
1622                 if (!peer) {
1623                         NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1624                         err = -ENOTCONN;
1625                         goto err;
1626                 }
1627
1628                 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) -
1629                           peer->hard_header_len;
1630                 /* Allow increasing the max_mtu if the program supports
1631                  * XDP fragments.
1632                  */
1633                 if (prog->aux->xdp_has_frags)
1634                         max_mtu += PAGE_SIZE * MAX_SKB_FRAGS;
1635
1636                 if (peer->mtu > max_mtu) {
1637                         NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1638                         err = -ERANGE;
1639                         goto err;
1640                 }
1641
1642                 if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1643                         NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1644                         err = -ENOSPC;
1645                         goto err;
1646                 }
1647
1648                 if (dev->flags & IFF_UP) {
1649                         err = veth_enable_xdp(dev);
1650                         if (err) {
1651                                 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1652                                 goto err;
1653                         }
1654                 }
1655
1656                 if (!old_prog) {
1657                         peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1658                         peer->max_mtu = max_mtu;
1659                 }
1660
1661                 xdp_features_set_redirect_target(peer, true);
1662         }
1663
1664         if (old_prog) {
1665                 if (!prog) {
1666                         if (peer && !veth_gro_requested(dev))
1667                                 xdp_features_clear_redirect_target(peer);
1668
1669                         if (dev->flags & IFF_UP)
1670                                 veth_disable_xdp(dev);
1671
1672                         if (peer) {
1673                                 peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1674                                 peer->max_mtu = ETH_MAX_MTU;
1675                         }
1676                 }
1677                 bpf_prog_put(old_prog);
1678         }
1679
1680         if ((!!old_prog ^ !!prog) && peer)
1681                 netdev_update_features(peer);
1682
1683         return 0;
1684 err:
1685         priv->_xdp_prog = old_prog;
1686
1687         return err;
1688 }
1689
1690 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1691 {
1692         switch (xdp->command) {
1693         case XDP_SETUP_PROG:
1694                 return veth_xdp_set(dev, xdp->prog, xdp->extack);
1695         default:
1696                 return -EINVAL;
1697         }
1698 }
1699
1700 static int veth_xdp_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp)
1701 {
1702         struct veth_xdp_buff *_ctx = (void *)ctx;
1703
1704         if (!_ctx->skb)
1705                 return -ENODATA;
1706
1707         *timestamp = skb_hwtstamps(_ctx->skb)->hwtstamp;
1708         return 0;
1709 }
1710
1711 static int veth_xdp_rx_hash(const struct xdp_md *ctx, u32 *hash,
1712                             enum xdp_rss_hash_type *rss_type)
1713 {
1714         struct veth_xdp_buff *_ctx = (void *)ctx;
1715         struct sk_buff *skb = _ctx->skb;
1716
1717         if (!skb)
1718                 return -ENODATA;
1719
1720         *hash = skb_get_hash(skb);
1721         *rss_type = skb->l4_hash ? XDP_RSS_TYPE_L4_ANY : XDP_RSS_TYPE_NONE;
1722
1723         return 0;
1724 }
1725
1726 static int veth_xdp_rx_vlan_tag(const struct xdp_md *ctx, __be16 *vlan_proto,
1727                                 u16 *vlan_tci)
1728 {
1729         const struct veth_xdp_buff *_ctx = (void *)ctx;
1730         const struct sk_buff *skb = _ctx->skb;
1731         int err;
1732
1733         if (!skb)
1734                 return -ENODATA;
1735
1736         err = __vlan_hwaccel_get_tag(skb, vlan_tci);
1737         if (err)
1738                 return err;
1739
1740         *vlan_proto = skb->vlan_proto;
1741         return err;
1742 }
1743
1744 static const struct net_device_ops veth_netdev_ops = {
1745         .ndo_init            = veth_dev_init,
1746         .ndo_open            = veth_open,
1747         .ndo_stop            = veth_close,
1748         .ndo_start_xmit      = veth_xmit,
1749         .ndo_get_stats64     = veth_get_stats64,
1750         .ndo_set_rx_mode     = veth_set_multicast_list,
1751         .ndo_set_mac_address = eth_mac_addr,
1752 #ifdef CONFIG_NET_POLL_CONTROLLER
1753         .ndo_poll_controller    = veth_poll_controller,
1754 #endif
1755         .ndo_get_iflink         = veth_get_iflink,
1756         .ndo_fix_features       = veth_fix_features,
1757         .ndo_set_features       = veth_set_features,
1758         .ndo_features_check     = passthru_features_check,
1759         .ndo_set_rx_headroom    = veth_set_rx_headroom,
1760         .ndo_bpf                = veth_xdp,
1761         .ndo_xdp_xmit           = veth_ndo_xdp_xmit,
1762         .ndo_get_peer_dev       = veth_peer_dev,
1763 };
1764
1765 static const struct xdp_metadata_ops veth_xdp_metadata_ops = {
1766         .xmo_rx_timestamp               = veth_xdp_rx_timestamp,
1767         .xmo_rx_hash                    = veth_xdp_rx_hash,
1768         .xmo_rx_vlan_tag                = veth_xdp_rx_vlan_tag,
1769 };
1770
1771 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1772                        NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1773                        NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1774                        NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1775                        NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1776
1777 static void veth_setup(struct net_device *dev)
1778 {
1779         ether_setup(dev);
1780
1781         dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1782         dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1783         dev->priv_flags |= IFF_NO_QUEUE;
1784         dev->priv_flags |= IFF_PHONY_HEADROOM;
1785
1786         dev->netdev_ops = &veth_netdev_ops;
1787         dev->xdp_metadata_ops = &veth_xdp_metadata_ops;
1788         dev->ethtool_ops = &veth_ethtool_ops;
1789         dev->features |= NETIF_F_LLTX;
1790         dev->features |= VETH_FEATURES;
1791         dev->vlan_features = dev->features &
1792                              ~(NETIF_F_HW_VLAN_CTAG_TX |
1793                                NETIF_F_HW_VLAN_STAG_TX |
1794                                NETIF_F_HW_VLAN_CTAG_RX |
1795                                NETIF_F_HW_VLAN_STAG_RX);
1796         dev->needs_free_netdev = true;
1797         dev->priv_destructor = veth_dev_free;
1798         dev->pcpu_stat_type = NETDEV_PCPU_STAT_TSTATS;
1799         dev->max_mtu = ETH_MAX_MTU;
1800
1801         dev->hw_features = VETH_FEATURES;
1802         dev->hw_enc_features = VETH_FEATURES;
1803         dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1804         netif_set_tso_max_size(dev, GSO_MAX_SIZE);
1805 }
1806
1807 /*
1808  * netlink interface
1809  */
1810
1811 static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1812                          struct netlink_ext_ack *extack)
1813 {
1814         if (tb[IFLA_ADDRESS]) {
1815                 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1816                         return -EINVAL;
1817                 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1818                         return -EADDRNOTAVAIL;
1819         }
1820         if (tb[IFLA_MTU]) {
1821                 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1822                         return -EINVAL;
1823         }
1824         return 0;
1825 }
1826
1827 static struct rtnl_link_ops veth_link_ops;
1828
1829 static void veth_disable_gro(struct net_device *dev)
1830 {
1831         dev->features &= ~NETIF_F_GRO;
1832         dev->wanted_features &= ~NETIF_F_GRO;
1833         netdev_update_features(dev);
1834 }
1835
1836 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1837 {
1838         int err;
1839
1840         if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1841                 err = netif_set_real_num_tx_queues(dev, 1);
1842                 if (err)
1843                         return err;
1844         }
1845         if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1846                 err = netif_set_real_num_rx_queues(dev, 1);
1847                 if (err)
1848                         return err;
1849         }
1850         return 0;
1851 }
1852
1853 static int veth_newlink(struct net *src_net, struct net_device *dev,
1854                         struct nlattr *tb[], struct nlattr *data[],
1855                         struct netlink_ext_ack *extack)
1856 {
1857         int err;
1858         struct net_device *peer;
1859         struct veth_priv *priv;
1860         char ifname[IFNAMSIZ];
1861         struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1862         unsigned char name_assign_type;
1863         struct ifinfomsg *ifmp;
1864         struct net *net;
1865
1866         /*
1867          * create and register peer first
1868          */
1869         if (data != NULL && data[VETH_INFO_PEER] != NULL) {
1870                 struct nlattr *nla_peer;
1871
1872                 nla_peer = data[VETH_INFO_PEER];
1873                 ifmp = nla_data(nla_peer);
1874                 err = rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack);
1875                 if (err < 0)
1876                         return err;
1877
1878                 err = veth_validate(peer_tb, NULL, extack);
1879                 if (err < 0)
1880                         return err;
1881
1882                 tbp = peer_tb;
1883         } else {
1884                 ifmp = NULL;
1885                 tbp = tb;
1886         }
1887
1888         if (ifmp && tbp[IFLA_IFNAME]) {
1889                 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1890                 name_assign_type = NET_NAME_USER;
1891         } else {
1892                 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1893                 name_assign_type = NET_NAME_ENUM;
1894         }
1895
1896         net = rtnl_link_get_net(src_net, tbp);
1897         if (IS_ERR(net))
1898                 return PTR_ERR(net);
1899
1900         peer = rtnl_create_link(net, ifname, name_assign_type,
1901                                 &veth_link_ops, tbp, extack);
1902         if (IS_ERR(peer)) {
1903                 put_net(net);
1904                 return PTR_ERR(peer);
1905         }
1906
1907         if (!ifmp || !tbp[IFLA_ADDRESS])
1908                 eth_hw_addr_random(peer);
1909
1910         if (ifmp && (dev->ifindex != 0))
1911                 peer->ifindex = ifmp->ifi_index;
1912
1913         netif_inherit_tso_max(peer, dev);
1914
1915         err = register_netdevice(peer);
1916         put_net(net);
1917         net = NULL;
1918         if (err < 0)
1919                 goto err_register_peer;
1920
1921         /* keep GRO disabled by default to be consistent with the established
1922          * veth behavior
1923          */
1924         veth_disable_gro(peer);
1925         netif_carrier_off(peer);
1926
1927         err = rtnl_configure_link(peer, ifmp, 0, NULL);
1928         if (err < 0)
1929                 goto err_configure_peer;
1930
1931         /*
1932          * register dev last
1933          *
1934          * note, that since we've registered new device the dev's name
1935          * should be re-allocated
1936          */
1937
1938         if (tb[IFLA_ADDRESS] == NULL)
1939                 eth_hw_addr_random(dev);
1940
1941         if (tb[IFLA_IFNAME])
1942                 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1943         else
1944                 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1945
1946         err = register_netdevice(dev);
1947         if (err < 0)
1948                 goto err_register_dev;
1949
1950         netif_carrier_off(dev);
1951
1952         /*
1953          * tie the deviced together
1954          */
1955
1956         priv = netdev_priv(dev);
1957         rcu_assign_pointer(priv->peer, peer);
1958         err = veth_init_queues(dev, tb);
1959         if (err)
1960                 goto err_queues;
1961
1962         priv = netdev_priv(peer);
1963         rcu_assign_pointer(priv->peer, dev);
1964         err = veth_init_queues(peer, tb);
1965         if (err)
1966                 goto err_queues;
1967
1968         veth_disable_gro(dev);
1969         /* update XDP supported features */
1970         veth_set_xdp_features(dev);
1971         veth_set_xdp_features(peer);
1972
1973         return 0;
1974
1975 err_queues:
1976         unregister_netdevice(dev);
1977 err_register_dev:
1978         /* nothing to do */
1979 err_configure_peer:
1980         unregister_netdevice(peer);
1981         return err;
1982
1983 err_register_peer:
1984         free_netdev(peer);
1985         return err;
1986 }
1987
1988 static void veth_dellink(struct net_device *dev, struct list_head *head)
1989 {
1990         struct veth_priv *priv;
1991         struct net_device *peer;
1992
1993         priv = netdev_priv(dev);
1994         peer = rtnl_dereference(priv->peer);
1995
1996         /* Note : dellink() is called from default_device_exit_batch(),
1997          * before a rcu_synchronize() point. The devices are guaranteed
1998          * not being freed before one RCU grace period.
1999          */
2000         RCU_INIT_POINTER(priv->peer, NULL);
2001         unregister_netdevice_queue(dev, head);
2002
2003         if (peer) {
2004                 priv = netdev_priv(peer);
2005                 RCU_INIT_POINTER(priv->peer, NULL);
2006                 unregister_netdevice_queue(peer, head);
2007         }
2008 }
2009
2010 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
2011         [VETH_INFO_PEER]        = { .len = sizeof(struct ifinfomsg) },
2012 };
2013
2014 static struct net *veth_get_link_net(const struct net_device *dev)
2015 {
2016         struct veth_priv *priv = netdev_priv(dev);
2017         struct net_device *peer = rtnl_dereference(priv->peer);
2018
2019         return peer ? dev_net(peer) : dev_net(dev);
2020 }
2021
2022 static unsigned int veth_get_num_queues(void)
2023 {
2024         /* enforce the same queue limit as rtnl_create_link */
2025         int queues = num_possible_cpus();
2026
2027         if (queues > 4096)
2028                 queues = 4096;
2029         return queues;
2030 }
2031
2032 static struct rtnl_link_ops veth_link_ops = {
2033         .kind           = DRV_NAME,
2034         .priv_size      = sizeof(struct veth_priv),
2035         .setup          = veth_setup,
2036         .validate       = veth_validate,
2037         .newlink        = veth_newlink,
2038         .dellink        = veth_dellink,
2039         .policy         = veth_policy,
2040         .maxtype        = VETH_INFO_MAX,
2041         .get_link_net   = veth_get_link_net,
2042         .get_num_tx_queues      = veth_get_num_queues,
2043         .get_num_rx_queues      = veth_get_num_queues,
2044 };
2045
2046 /*
2047  * init/fini
2048  */
2049
2050 static __init int veth_init(void)
2051 {
2052         return rtnl_link_register(&veth_link_ops);
2053 }
2054
2055 static __exit void veth_exit(void)
2056 {
2057         rtnl_link_unregister(&veth_link_ops);
2058 }
2059
2060 module_init(veth_init);
2061 module_exit(veth_exit);
2062
2063 MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
2064 MODULE_LICENSE("GPL v2");
2065 MODULE_ALIAS_RTNL_LINK(DRV_NAME);