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.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                 pp_idx = tx_idx + VETH_TQ_STATS_LEN;
 240         }
 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         struct net_device *rcv;
 348         int length = skb->len;
 349         bool use_napi = false;
 350         int rxq;
 351
 352         rcu_read_lock();
 353         rcv = rcu_dereference(priv->peer);
 354         if (unlikely(!rcv) || !pskb_may_pull(skb, ETH_HLEN)) {
 355                 kfree_skb(skb);
 356                 goto drop;
 357         }
 358
 359         rcv_priv = netdev_priv(rcv);
 360         rxq = skb_get_queue_mapping(skb);
 361         if (rxq < rcv->real_num_rx_queues) {
 362                 rq = &rcv_priv->rq[rxq];
 363
 364                 /* The napi pointer is available when an XDP program is
 365                  * attached or when GRO is enabled
 366                  * Don't bother with napi/GRO if the skb can't be aggregated
 367                  */
 368                 use_napi = rcu_access_pointer(rq->napi) &&
 369                            veth_skb_is_eligible_for_gro(dev, rcv, skb);
 370         }
 371
 372         skb_tx_timestamp(skb);
 373         if (likely(veth_forward_skb(rcv, skb, rq, use_napi) == NET_RX_SUCCESS)) {
 374                 if (!use_napi)
 375                         dev_lstats_add(dev, length);
 376         } else {
 377 drop:
 378                 atomic64_inc(&priv->dropped);
 379         }
 380
 381         if (use_napi)
 382                 __veth_xdp_flush(rq);
 383
 384         rcu_read_unlock();
 385
 386         return NETDEV_TX_OK;
 387 }
 388
 389 static u64 veth_stats_tx(struct net_device *dev, u64 *packets, u64 *bytes)
 390 {
 391         struct veth_priv *priv = netdev_priv(dev);
 392
 393         dev_lstats_read(dev, packets, bytes);
 394         return atomic64_read(&priv->dropped);
 395 }
 396
 397 static void veth_stats_rx(struct veth_stats *result, struct net_device *dev)
 398 {
 399         struct veth_priv *priv = netdev_priv(dev);
 400         int i;
 401
 402         result->peer_tq_xdp_xmit_err = 0;
 403         result->xdp_packets = 0;
 404         result->xdp_tx_err = 0;
 405         result->xdp_bytes = 0;
 406         result->rx_drops = 0;
 407         for (i = 0; i < dev->num_rx_queues; i++) {
 408                 u64 packets, bytes, drops, xdp_tx_err, peer_tq_xdp_xmit_err;
 409                 struct veth_rq_stats *stats = &priv->rq[i].stats;
 410                 unsigned int start;
 411
 412                 do {
 413                         start = u64_stats_fetch_begin(&stats->syncp);
 414                         peer_tq_xdp_xmit_err = stats->vs.peer_tq_xdp_xmit_err;
 415                         xdp_tx_err = stats->vs.xdp_tx_err;
 416                         packets = stats->vs.xdp_packets;
 417                         bytes = stats->vs.xdp_bytes;
 418                         drops = stats->vs.rx_drops;
 419                 } while (u64_stats_fetch_retry(&stats->syncp, start));
 420                 result->peer_tq_xdp_xmit_err += peer_tq_xdp_xmit_err;
 421                 result->xdp_tx_err += xdp_tx_err;
 422                 result->xdp_packets += packets;
 423                 result->xdp_bytes += bytes;
 424                 result->rx_drops += drops;
 425         }
 426 }
 427
 428 static void veth_get_stats64(struct net_device *dev,
 429                              struct rtnl_link_stats64 *tot)
 430 {
 431         struct veth_priv *priv = netdev_priv(dev);
 432         struct net_device *peer;
 433         struct veth_stats rx;
 434         u64 packets, bytes;
 435
 436         tot->tx_dropped = veth_stats_tx(dev, &packets, &bytes);
 437         tot->tx_bytes = bytes;
 438         tot->tx_packets = packets;
 439
 440         veth_stats_rx(&rx, dev);
 441         tot->tx_dropped += rx.xdp_tx_err;
 442         tot->rx_dropped = rx.rx_drops + rx.peer_tq_xdp_xmit_err;
 443         tot->rx_bytes = rx.xdp_bytes;
 444         tot->rx_packets = rx.xdp_packets;
 445
 446         rcu_read_lock();
 447         peer = rcu_dereference(priv->peer);
 448         if (peer) {
 449                 veth_stats_tx(peer, &packets, &bytes);
 450                 tot->rx_bytes += bytes;
 451                 tot->rx_packets += packets;
 452
 453                 veth_stats_rx(&rx, peer);
 454                 tot->tx_dropped += rx.peer_tq_xdp_xmit_err;
 455                 tot->rx_dropped += rx.xdp_tx_err;
 456                 tot->tx_bytes += rx.xdp_bytes;
 457                 tot->tx_packets += rx.xdp_packets;
 458         }
 459         rcu_read_unlock();
 460 }
 461
 462 /* fake multicast ability */
 463 static void veth_set_multicast_list(struct net_device *dev)
 464 {
 465 }
 466
 467 static int veth_select_rxq(struct net_device *dev)
 468 {
 469         return smp_processor_id() % dev->real_num_rx_queues;
 470 }
 471
 472 static struct net_device *veth_peer_dev(struct net_device *dev)
 473 {
 474         struct veth_priv *priv = netdev_priv(dev);
 475
 476         /* Callers must be under RCU read side. */
 477         return rcu_dereference(priv->peer);
 478 }
 479
 480 static int veth_xdp_xmit(struct net_device *dev, int n,
 481                          struct xdp_frame **frames,
 482                          u32 flags, bool ndo_xmit)
 483 {
 484         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 485         int i, ret = -ENXIO, nxmit = 0;
 486         struct net_device *rcv;
 487         unsigned int max_len;
 488         struct veth_rq *rq;
 489
 490         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
 491                 return -EINVAL;
 492
 493         rcu_read_lock();
 494         rcv = rcu_dereference(priv->peer);
 495         if (unlikely(!rcv))
 496                 goto out;
 497
 498         rcv_priv = netdev_priv(rcv);
 499         rq = &rcv_priv->rq[veth_select_rxq(rcv)];
 500         /* The napi pointer is set if NAPI is enabled, which ensures that
 501          * xdp_ring is initialized on receive side and the peer device is up.
 502          */
 503         if (!rcu_access_pointer(rq->napi))
 504                 goto out;
 505
 506         max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN;
 507
 508         spin_lock(&rq->xdp_ring.producer_lock);
 509         for (i = 0; i < n; i++) {
 510                 struct xdp_frame *frame = frames[i];
 511                 void *ptr = veth_xdp_to_ptr(frame);
 512
 513                 if (unlikely(xdp_get_frame_len(frame) > max_len ||
 514                              __ptr_ring_produce(&rq->xdp_ring, ptr)))
 515                         break;
 516                 nxmit++;
 517         }
 518         spin_unlock(&rq->xdp_ring.producer_lock);
 519
 520         if (flags & XDP_XMIT_FLUSH)
 521                 __veth_xdp_flush(rq);
 522
 523         ret = nxmit;
 524         if (ndo_xmit) {
 525                 u64_stats_update_begin(&rq->stats.syncp);
 526                 rq->stats.vs.peer_tq_xdp_xmit += nxmit;
 527                 rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit;
 528                 u64_stats_update_end(&rq->stats.syncp);
 529         }
 530
 531 out:
 532         rcu_read_unlock();
 533
 534         return ret;
 535 }
 536
 537 static int veth_ndo_xdp_xmit(struct net_device *dev, int n,
 538                              struct xdp_frame **frames, u32 flags)
 539 {
 540         int err;
 541
 542         err = veth_xdp_xmit(dev, n, frames, flags, true);
 543         if (err < 0) {
 544                 struct veth_priv *priv = netdev_priv(dev);
 545
 546                 atomic64_add(n, &priv->dropped);
 547         }
 548
 549         return err;
 550 }
 551
 552 static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
 553 {
 554         int sent, i, err = 0, drops;
 555
 556         sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false);
 557         if (sent < 0) {
 558                 err = sent;
 559                 sent = 0;
 560         }
 561
 562         for (i = sent; unlikely(i < bq->count); i++)
 563                 xdp_return_frame(bq->q[i]);
 564
 565         drops = bq->count - sent;
 566         trace_xdp_bulk_tx(rq->dev, sent, drops, err);
 567
 568         u64_stats_update_begin(&rq->stats.syncp);
 569         rq->stats.vs.xdp_tx += sent;
 570         rq->stats.vs.xdp_tx_err += drops;
 571         u64_stats_update_end(&rq->stats.syncp);
 572
 573         bq->count = 0;
 574 }
 575
 576 static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
 577 {
 578         struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev);
 579         struct net_device *rcv;
 580         struct veth_rq *rcv_rq;
 581
 582         rcu_read_lock();
 583         veth_xdp_flush_bq(rq, bq);
 584         rcv = rcu_dereference(priv->peer);
 585         if (unlikely(!rcv))
 586                 goto out;
 587
 588         rcv_priv = netdev_priv(rcv);
 589         rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)];
 590         /* xdp_ring is initialized on receive side? */
 591         if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog)))
 592                 goto out;
 593
 594         __veth_xdp_flush(rcv_rq);
 595 out:
 596         rcu_read_unlock();
 597 }
 598
 599 static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp,
 600                        struct veth_xdp_tx_bq *bq)
 601 {
 602         struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp);
 603
 604         if (unlikely(!frame))
 605                 return -EOVERFLOW;
 606
 607         if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE))
 608                 veth_xdp_flush_bq(rq, bq);
 609
 610         bq->q[bq->count++] = frame;
 611
 612         return 0;
 613 }
 614
 615 static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq,
 616                                           struct xdp_frame *frame,
 617                                           struct veth_xdp_tx_bq *bq,
 618                                           struct veth_stats *stats)
 619 {
 620         struct xdp_frame orig_frame;
 621         struct bpf_prog *xdp_prog;
 622
 623         rcu_read_lock();
 624         xdp_prog = rcu_dereference(rq->xdp_prog);
 625         if (likely(xdp_prog)) {
 626                 struct veth_xdp_buff vxbuf;
 627                 struct xdp_buff *xdp = &vxbuf.xdp;
 628                 u32 act;
 629
 630                 xdp_convert_frame_to_buff(frame, xdp);
 631                 xdp->rxq = &rq->xdp_rxq;
 632                 vxbuf.skb = NULL;
 633
 634                 act = bpf_prog_run_xdp(xdp_prog, xdp);
 635
 636                 switch (act) {
 637                 case XDP_PASS:
 638                         if (xdp_update_frame_from_buff(xdp, frame))
 639                                 goto err_xdp;
 640                         break;
 641                 case XDP_TX:
 642                         orig_frame = *frame;
 643                         xdp->rxq->mem = frame->mem;
 644                         if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) {
 645                                 trace_xdp_exception(rq->dev, xdp_prog, act);
 646                                 frame = &orig_frame;
 647                                 stats->rx_drops++;
 648                                 goto err_xdp;
 649                         }
 650                         stats->xdp_tx++;
 651                         rcu_read_unlock();
 652                         goto xdp_xmit;
 653                 case XDP_REDIRECT:
 654                         orig_frame = *frame;
 655                         xdp->rxq->mem = frame->mem;
 656                         if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) {
 657                                 frame = &orig_frame;
 658                                 stats->rx_drops++;
 659                                 goto err_xdp;
 660                         }
 661                         stats->xdp_redirect++;
 662                         rcu_read_unlock();
 663                         goto xdp_xmit;
 664                 default:
 665                         bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
 666                         fallthrough;
 667                 case XDP_ABORTED:
 668                         trace_xdp_exception(rq->dev, xdp_prog, act);
 669                         fallthrough;
 670                 case XDP_DROP:
 671                         stats->xdp_drops++;
 672                         goto err_xdp;
 673                 }
 674         }
 675         rcu_read_unlock();
 676
 677         return frame;
 678 err_xdp:
 679         rcu_read_unlock();
 680         xdp_return_frame(frame);
 681 xdp_xmit:
 682         return NULL;
 683 }
 684
 685 /* frames array contains VETH_XDP_BATCH at most */
 686 static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames,
 687                                   int n_xdpf, struct veth_xdp_tx_bq *bq,
 688                                   struct veth_stats *stats)
 689 {
 690         void *skbs[VETH_XDP_BATCH];
 691         int i;
 692
 693         if (xdp_alloc_skb_bulk(skbs, n_xdpf,
 694                                GFP_ATOMIC | __GFP_ZERO) < 0) {
 695                 for (i = 0; i < n_xdpf; i++)
 696                         xdp_return_frame(frames[i]);
 697                 stats->rx_drops += n_xdpf;
 698
 699                 return;
 700         }
 701
 702         for (i = 0; i < n_xdpf; i++) {
 703                 struct sk_buff *skb = skbs[i];
 704
 705                 skb = __xdp_build_skb_from_frame(frames[i], skb,
 706                                                  rq->dev);
 707                 if (!skb) {
 708                         xdp_return_frame(frames[i]);
 709                         stats->rx_drops++;
 710                         continue;
 711                 }
 712                 napi_gro_receive(&rq->xdp_napi, skb);
 713         }
 714 }
 715
 716 static void veth_xdp_get(struct xdp_buff *xdp)
 717 {
 718         struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
 719         int i;
 720
 721         get_page(virt_to_page(xdp->data));
 722         if (likely(!xdp_buff_has_frags(xdp)))
 723                 return;
 724
 725         for (i = 0; i < sinfo->nr_frags; i++)
 726                 __skb_frag_ref(&sinfo->frags[i]);
 727 }
 728
 729 static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
 730                                         struct xdp_buff *xdp,
 731                                         struct sk_buff **pskb)
 732 {
 733         struct sk_buff *skb = *pskb;
 734         u32 frame_sz;
 735
 736         if (skb_shared(skb) || skb_head_is_locked(skb) ||
 737             skb_shinfo(skb)->nr_frags ||
 738             skb_headroom(skb) < XDP_PACKET_HEADROOM) {
 739                 u32 size, len, max_head_size, off;
 740                 struct sk_buff *nskb;
 741                 struct page *page;
 742                 int i, head_off;
 743
 744                 /* We need a private copy of the skb and data buffers since
 745                  * the ebpf program can modify it. We segment the original skb
 746                  * into order-0 pages without linearize it.
 747                  *
 748                  * Make sure we have enough space for linear and paged area
 749                  */
 750                 max_head_size = SKB_WITH_OVERHEAD(PAGE_SIZE -
 751                                                   VETH_XDP_HEADROOM);
 752                 if (skb->len > PAGE_SIZE * MAX_SKB_FRAGS + max_head_size)
 753                         goto drop;
 754
 755                 /* Allocate skb head */
 756                 page = page_pool_dev_alloc_pages(rq->page_pool);
 757                 if (!page)
 758                         goto drop;
 759
 760                 nskb = napi_build_skb(page_address(page), PAGE_SIZE);
 761                 if (!nskb) {
 762                         page_pool_put_full_page(rq->page_pool, page, true);
 763                         goto drop;
 764                 }
 765
 766                 skb_reserve(nskb, VETH_XDP_HEADROOM);
 767                 skb_copy_header(nskb, skb);
 768                 skb_mark_for_recycle(nskb);
 769
 770                 size = min_t(u32, skb->len, max_head_size);
 771                 if (skb_copy_bits(skb, 0, nskb->data, size)) {
 772                         consume_skb(nskb);
 773                         goto drop;
 774                 }
 775                 skb_put(nskb, size);
 776
 777                 head_off = skb_headroom(nskb) - skb_headroom(skb);
 778                 skb_headers_offset_update(nskb, head_off);
 779
 780                 /* Allocate paged area of new skb */
 781                 off = size;
 782                 len = skb->len - off;
 783
 784                 for (i = 0; i < MAX_SKB_FRAGS && off < skb->len; i++) {
 785                         page = page_pool_dev_alloc_pages(rq->page_pool);
 786                         if (!page) {
 787                                 consume_skb(nskb);
 788                                 goto drop;
 789                         }
 790
 791                         size = min_t(u32, len, PAGE_SIZE);
 792                         skb_add_rx_frag(nskb, i, page, 0, size, PAGE_SIZE);
 793                         if (skb_copy_bits(skb, off, page_address(page),
 794                                           size)) {
 795                                 consume_skb(nskb);
 796                                 goto drop;
 797                         }
 798
 799                         len -= size;
 800                         off += size;
 801                 }
 802
 803                 consume_skb(skb);
 804                 skb = nskb;
 805         }
 806
 807         /* SKB "head" area always have tailroom for skb_shared_info */
 808         frame_sz = skb_end_pointer(skb) - skb->head;
 809         frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 810         xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
 811         xdp_prepare_buff(xdp, skb->head, skb_headroom(skb),
 812                          skb_headlen(skb), true);
 813
 814         if (skb_is_nonlinear(skb)) {
 815                 skb_shinfo(skb)->xdp_frags_size = skb->data_len;
 816                 xdp_buff_set_frags_flag(xdp);
 817         } else {
 818                 xdp_buff_clear_frags_flag(xdp);
 819         }
 820         *pskb = skb;
 821
 822         return 0;
 823 drop:
 824         consume_skb(skb);
 825         *pskb = NULL;
 826
 827         return -ENOMEM;
 828 }
 829
 830 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq,
 831                                         struct sk_buff *skb,
 832                                         struct veth_xdp_tx_bq *bq,
 833                                         struct veth_stats *stats)
 834 {
 835         void *orig_data, *orig_data_end;
 836         struct bpf_prog *xdp_prog;
 837         struct veth_xdp_buff vxbuf;
 838         struct xdp_buff *xdp = &vxbuf.xdp;
 839         u32 act, metalen;
 840         int off;
 841
 842         skb_prepare_for_gro(skb);
 843
 844         rcu_read_lock();
 845         xdp_prog = rcu_dereference(rq->xdp_prog);
 846         if (unlikely(!xdp_prog)) {
 847                 rcu_read_unlock();
 848                 goto out;
 849         }
 850
 851         __skb_push(skb, skb->data - skb_mac_header(skb));
 852         if (veth_convert_skb_to_xdp_buff(rq, xdp, &skb))
 853                 goto drop;
 854         vxbuf.skb = skb;
 855
 856         orig_data = xdp->data;
 857         orig_data_end = xdp->data_end;
 858
 859         act = bpf_prog_run_xdp(xdp_prog, xdp);
 860
 861         switch (act) {
 862         case XDP_PASS:
 863                 break;
 864         case XDP_TX:
 865                 veth_xdp_get(xdp);
 866                 consume_skb(skb);
 867                 xdp->rxq->mem = rq->xdp_mem;
 868                 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) {
 869                         trace_xdp_exception(rq->dev, xdp_prog, act);
 870                         stats->rx_drops++;
 871                         goto err_xdp;
 872                 }
 873                 stats->xdp_tx++;
 874                 rcu_read_unlock();
 875                 goto xdp_xmit;
 876         case XDP_REDIRECT:
 877                 veth_xdp_get(xdp);
 878                 consume_skb(skb);
 879                 xdp->rxq->mem = rq->xdp_mem;
 880                 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) {
 881                         stats->rx_drops++;
 882                         goto err_xdp;
 883                 }
 884                 stats->xdp_redirect++;
 885                 rcu_read_unlock();
 886                 goto xdp_xmit;
 887         default:
 888                 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
 889                 fallthrough;
 890         case XDP_ABORTED:
 891                 trace_xdp_exception(rq->dev, xdp_prog, act);
 892                 fallthrough;
 893         case XDP_DROP:
 894                 stats->xdp_drops++;
 895                 goto xdp_drop;
 896         }
 897         rcu_read_unlock();
 898
 899         /* check if bpf_xdp_adjust_head was used */
 900         off = orig_data - xdp->data;
 901         if (off > 0)
 902                 __skb_push(skb, off);
 903         else if (off < 0)
 904                 __skb_pull(skb, -off);
 905
 906         skb_reset_mac_header(skb);
 907
 908         /* check if bpf_xdp_adjust_tail was used */
 909         off = xdp->data_end - orig_data_end;
 910         if (off != 0)
 911                 __skb_put(skb, off); /* positive on grow, negative on shrink */
 912
 913         /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers
 914          * (e.g. bpf_xdp_adjust_tail), we need to update data_len here.
 915          */
 916         if (xdp_buff_has_frags(xdp))
 917                 skb->data_len = skb_shinfo(skb)->xdp_frags_size;
 918         else
 919                 skb->data_len = 0;
 920
 921         skb->protocol = eth_type_trans(skb, rq->dev);
 922
 923         metalen = xdp->data - xdp->data_meta;
 924         if (metalen)
 925                 skb_metadata_set(skb, metalen);
 926 out:
 927         return skb;
 928 drop:
 929         stats->rx_drops++;
 930 xdp_drop:
 931         rcu_read_unlock();
 932         kfree_skb(skb);
 933         return NULL;
 934 err_xdp:
 935         rcu_read_unlock();
 936         xdp_return_buff(xdp);
 937 xdp_xmit:
 938         return NULL;
 939 }
 940
 941 static int veth_xdp_rcv(struct veth_rq *rq, int budget,
 942                         struct veth_xdp_tx_bq *bq,
 943                         struct veth_stats *stats)
 944 {
 945         int i, done = 0, n_xdpf = 0;
 946         void *xdpf[VETH_XDP_BATCH];
 947
 948         for (i = 0; i < budget; i++) {
 949                 void *ptr = __ptr_ring_consume(&rq->xdp_ring);
 950
 951                 if (!ptr)
 952                         break;
 953
 954                 if (veth_is_xdp_frame(ptr)) {
 955                         /* ndo_xdp_xmit */
 956                         struct xdp_frame *frame = veth_ptr_to_xdp(ptr);
 957
 958                         stats->xdp_bytes += xdp_get_frame_len(frame);
 959                         frame = veth_xdp_rcv_one(rq, frame, bq, stats);
 960                         if (frame) {
 961                                 /* XDP_PASS */
 962                                 xdpf[n_xdpf++] = frame;
 963                                 if (n_xdpf == VETH_XDP_BATCH) {
 964                                         veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf,
 965                                                               bq, stats);
 966                                         n_xdpf = 0;
 967                                 }
 968                         }
 969                 } else {
 970                         /* ndo_start_xmit */
 971                         struct sk_buff *skb = ptr;
 972
 973                         stats->xdp_bytes += skb->len;
 974                         skb = veth_xdp_rcv_skb(rq, skb, bq, stats);
 975                         if (skb) {
 976                                 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC))
 977                                         netif_receive_skb(skb);
 978                                 else
 979                                         napi_gro_receive(&rq->xdp_napi, skb);
 980                         }
 981                 }
 982                 done++;
 983         }
 984
 985         if (n_xdpf)
 986                 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats);
 987
 988         u64_stats_update_begin(&rq->stats.syncp);
 989         rq->stats.vs.xdp_redirect += stats->xdp_redirect;
 990         rq->stats.vs.xdp_bytes += stats->xdp_bytes;
 991         rq->stats.vs.xdp_drops += stats->xdp_drops;
 992         rq->stats.vs.rx_drops += stats->rx_drops;
 993         rq->stats.vs.xdp_packets += done;
 994         u64_stats_update_end(&rq->stats.syncp);
 995
 996         return done;
 997 }
 998
 999 static int veth_poll(struct napi_struct *napi, int budget)
1000 {
1001         struct veth_rq *rq =
1002                 container_of(napi, struct veth_rq, xdp_napi);
1003         struct veth_stats stats = {};
1004         struct veth_xdp_tx_bq bq;
1005         int done;
1006
1007         bq.count = 0;
1008
1009         xdp_set_return_frame_no_direct();
1010         done = veth_xdp_rcv(rq, budget, &bq, &stats);
1011
1012         if (stats.xdp_redirect > 0)
1013                 xdp_do_flush();
1014
1015         if (done < budget && napi_complete_done(napi, done)) {
1016                 /* Write rx_notify_masked before reading ptr_ring */
1017                 smp_store_mb(rq->rx_notify_masked, false);
1018                 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
1019                         if (napi_schedule_prep(&rq->xdp_napi)) {
1020                                 WRITE_ONCE(rq->rx_notify_masked, true);
1021                                 __napi_schedule(&rq->xdp_napi);
1022                         }
1023                 }
1024         }
1025
1026         if (stats.xdp_tx > 0)
1027                 veth_xdp_flush(rq, &bq);
1028         xdp_clear_return_frame_no_direct();
1029
1030         return done;
1031 }
1032
1033 static int veth_create_page_pool(struct veth_rq *rq)
1034 {
1035         struct page_pool_params pp_params = {
1036                 .order = 0,
1037                 .pool_size = VETH_RING_SIZE,
1038                 .nid = NUMA_NO_NODE,
1039                 .dev = &rq->dev->dev,
1040         };
1041
1042         rq->page_pool = page_pool_create(&pp_params);
1043         if (IS_ERR(rq->page_pool)) {
1044                 int err = PTR_ERR(rq->page_pool);
1045
1046                 rq->page_pool = NULL;
1047                 return err;
1048         }
1049
1050         return 0;
1051 }
1052
1053 static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
1054 {
1055         struct veth_priv *priv = netdev_priv(dev);
1056         int err, i;
1057
1058         for (i = start; i < end; i++) {
1059                 err = veth_create_page_pool(&priv->rq[i]);
1060                 if (err)
1061                         goto err_page_pool;
1062         }
1063
1064         for (i = start; i < end; i++) {
1065                 struct veth_rq *rq = &priv->rq[i];
1066
1067                 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
1068                 if (err)
1069                         goto err_xdp_ring;
1070         }
1071
1072         for (i = start; i < end; i++) {
1073                 struct veth_rq *rq = &priv->rq[i];
1074
1075                 napi_enable(&rq->xdp_napi);
1076                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1077         }
1078
1079         return 0;
1080
1081 err_xdp_ring:
1082         for (i--; i >= start; i--)
1083                 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
1084         i = end;
1085 err_page_pool:
1086         for (i--; i >= start; i--) {
1087                 page_pool_destroy(priv->rq[i].page_pool);
1088                 priv->rq[i].page_pool = NULL;
1089         }
1090
1091         return err;
1092 }
1093
1094 static int __veth_napi_enable(struct net_device *dev)
1095 {
1096         return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1097 }
1098
1099 static void veth_napi_del_range(struct net_device *dev, int start, int end)
1100 {
1101         struct veth_priv *priv = netdev_priv(dev);
1102         int i;
1103
1104         for (i = start; i < end; i++) {
1105                 struct veth_rq *rq = &priv->rq[i];
1106
1107                 rcu_assign_pointer(priv->rq[i].napi, NULL);
1108                 napi_disable(&rq->xdp_napi);
1109                 __netif_napi_del(&rq->xdp_napi);
1110         }
1111         synchronize_net();
1112
1113         for (i = start; i < end; i++) {
1114                 struct veth_rq *rq = &priv->rq[i];
1115
1116                 rq->rx_notify_masked = false;
1117                 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
1118         }
1119
1120         for (i = start; i < end; i++) {
1121                 page_pool_destroy(priv->rq[i].page_pool);
1122                 priv->rq[i].page_pool = NULL;
1123         }
1124 }
1125
1126 static void veth_napi_del(struct net_device *dev)
1127 {
1128         veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
1129 }
1130
1131 static bool veth_gro_requested(const struct net_device *dev)
1132 {
1133         return !!(dev->wanted_features & NETIF_F_GRO);
1134 }
1135
1136 static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1137                                  bool napi_already_on)
1138 {
1139         struct veth_priv *priv = netdev_priv(dev);
1140         int err, i;
1141
1142         for (i = start; i < end; i++) {
1143                 struct veth_rq *rq = &priv->rq[i];
1144
1145                 if (!napi_already_on)
1146                         netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1147                 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1148                 if (err < 0)
1149                         goto err_rxq_reg;
1150
1151                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1152                                                  MEM_TYPE_PAGE_SHARED,
1153                                                  NULL);
1154                 if (err < 0)
1155                         goto err_reg_mem;
1156
1157                 /* Save original mem info as it can be overwritten */
1158                 rq->xdp_mem = rq->xdp_rxq.mem;
1159         }
1160         return 0;
1161
1162 err_reg_mem:
1163         xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1164 err_rxq_reg:
1165         for (i--; i >= start; i--) {
1166                 struct veth_rq *rq = &priv->rq[i];
1167
1168                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1169                 if (!napi_already_on)
1170                         netif_napi_del(&rq->xdp_napi);
1171         }
1172
1173         return err;
1174 }
1175
1176 static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1177                                    bool delete_napi)
1178 {
1179         struct veth_priv *priv = netdev_priv(dev);
1180         int i;
1181
1182         for (i = start; i < end; i++) {
1183                 struct veth_rq *rq = &priv->rq[i];
1184
1185                 rq->xdp_rxq.mem = rq->xdp_mem;
1186                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1187
1188                 if (delete_napi)
1189                         netif_napi_del(&rq->xdp_napi);
1190         }
1191 }
1192
1193 static int veth_enable_xdp(struct net_device *dev)
1194 {
1195         bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1196         struct veth_priv *priv = netdev_priv(dev);
1197         int err, i;
1198
1199         if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1200                 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1201                 if (err)
1202                         return err;
1203
1204                 if (!napi_already_on) {
1205                         err = __veth_napi_enable(dev);
1206                         if (err) {
1207                                 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1208                                 return err;
1209                         }
1210
1211                         if (!veth_gro_requested(dev)) {
1212                                 /* user-space did not require GRO, but adding XDP
1213                                  * is supposed to get GRO working
1214                                  */
1215                                 dev->features |= NETIF_F_GRO;
1216                                 netdev_features_change(dev);
1217                         }
1218                 }
1219         }
1220
1221         for (i = 0; i < dev->real_num_rx_queues; i++) {
1222                 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1223                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1224         }
1225
1226         return 0;
1227 }
1228
1229 static void veth_disable_xdp(struct net_device *dev)
1230 {
1231         struct veth_priv *priv = netdev_priv(dev);
1232         int i;
1233
1234         for (i = 0; i < dev->real_num_rx_queues; i++)
1235                 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1236
1237         if (!netif_running(dev) || !veth_gro_requested(dev)) {
1238                 veth_napi_del(dev);
1239
1240                 /* if user-space did not require GRO, since adding XDP
1241                  * enabled it, clear it now
1242                  */
1243                 if (!veth_gro_requested(dev) && netif_running(dev)) {
1244                         dev->features &= ~NETIF_F_GRO;
1245                         netdev_features_change(dev);
1246                 }
1247         }
1248
1249         veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1250 }
1251
1252 static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1253 {
1254         struct veth_priv *priv = netdev_priv(dev);
1255         int err, i;
1256
1257         for (i = start; i < end; i++) {
1258                 struct veth_rq *rq = &priv->rq[i];
1259
1260                 netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1261         }
1262
1263         err = __veth_napi_enable_range(dev, start, end);
1264         if (err) {
1265                 for (i = start; i < end; i++) {
1266                         struct veth_rq *rq = &priv->rq[i];
1267
1268                         netif_napi_del(&rq->xdp_napi);
1269                 }
1270                 return err;
1271         }
1272         return err;
1273 }
1274
1275 static int veth_napi_enable(struct net_device *dev)
1276 {
1277         return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1278 }
1279
1280 static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1281 {
1282         struct veth_priv *priv = netdev_priv(dev);
1283
1284         if (start >= end)
1285                 return;
1286
1287         if (priv->_xdp_prog) {
1288                 veth_napi_del_range(dev, start, end);
1289                 veth_disable_xdp_range(dev, start, end, false);
1290         } else if (veth_gro_requested(dev)) {
1291                 veth_napi_del_range(dev, start, end);
1292         }
1293 }
1294
1295 static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1296 {
1297         struct veth_priv *priv = netdev_priv(dev);
1298         int err;
1299
1300         if (start >= end)
1301                 return 0;
1302
1303         if (priv->_xdp_prog) {
1304                 /* these channels are freshly initialized, napi is not on there even
1305                  * when GRO is requeste
1306                  */
1307                 err = veth_enable_xdp_range(dev, start, end, false);
1308                 if (err)
1309                         return err;
1310
1311                 err = __veth_napi_enable_range(dev, start, end);
1312                 if (err) {
1313                         /* on error always delete the newly added napis */
1314                         veth_disable_xdp_range(dev, start, end, true);
1315                         return err;
1316                 }
1317         } else if (veth_gro_requested(dev)) {
1318                 return veth_napi_enable_range(dev, start, end);
1319         }
1320         return 0;
1321 }
1322
1323 static void veth_set_xdp_features(struct net_device *dev)
1324 {
1325         struct veth_priv *priv = netdev_priv(dev);
1326         struct net_device *peer;
1327
1328         peer = rtnl_dereference(priv->peer);
1329         if (peer && peer->real_num_tx_queues <= dev->real_num_rx_queues) {
1330                 struct veth_priv *priv_peer = netdev_priv(peer);
1331                 xdp_features_t val = NETDEV_XDP_ACT_BASIC |
1332                                      NETDEV_XDP_ACT_REDIRECT |
1333                                      NETDEV_XDP_ACT_RX_SG;
1334
1335                 if (priv_peer->_xdp_prog || veth_gro_requested(peer))
1336                         val |= NETDEV_XDP_ACT_NDO_XMIT |
1337                                NETDEV_XDP_ACT_NDO_XMIT_SG;
1338                 xdp_set_features_flag(dev, val);
1339         } else {
1340                 xdp_clear_features_flag(dev);
1341         }
1342 }
1343
1344 static int veth_set_channels(struct net_device *dev,
1345                              struct ethtool_channels *ch)
1346 {
1347         struct veth_priv *priv = netdev_priv(dev);
1348         unsigned int old_rx_count, new_rx_count;
1349         struct veth_priv *peer_priv;
1350         struct net_device *peer;
1351         int err;
1352
1353         /* sanity check. Upper bounds are already enforced by the caller */
1354         if (!ch->rx_count || !ch->tx_count)
1355                 return -EINVAL;
1356
1357         /* avoid braking XDP, if that is enabled */
1358         peer = rtnl_dereference(priv->peer);
1359         peer_priv = peer ? netdev_priv(peer) : NULL;
1360         if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1361                 return -EINVAL;
1362
1363         if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1364                 return -EINVAL;
1365
1366         old_rx_count = dev->real_num_rx_queues;
1367         new_rx_count = ch->rx_count;
1368         if (netif_running(dev)) {
1369                 /* turn device off */
1370                 netif_carrier_off(dev);
1371                 if (peer)
1372                         netif_carrier_off(peer);
1373
1374                 /* try to allocate new resurces, as needed*/
1375                 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1376                 if (err)
1377                         goto out;
1378         }
1379
1380         err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1381         if (err)
1382                 goto revert;
1383
1384         err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1385         if (err) {
1386                 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1387
1388                 /* this error condition could happen only if rx and tx change
1389                  * in opposite directions (e.g. tx nr raises, rx nr decreases)
1390                  * and we can't do anything to fully restore the original
1391                  * status
1392                  */
1393                 if (err2)
1394                         pr_warn("Can't restore rx queues config %d -> %d %d",
1395                                 new_rx_count, old_rx_count, err2);
1396                 else
1397                         goto revert;
1398         }
1399
1400 out:
1401         if (netif_running(dev)) {
1402                 /* note that we need to swap the arguments WRT the enable part
1403                  * to identify the range we have to disable
1404                  */
1405                 veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1406                 netif_carrier_on(dev);
1407                 if (peer)
1408                         netif_carrier_on(peer);
1409         }
1410
1411         /* update XDP supported features */
1412         veth_set_xdp_features(dev);
1413         if (peer)
1414                 veth_set_xdp_features(peer);
1415
1416         return err;
1417
1418 revert:
1419         new_rx_count = old_rx_count;
1420         old_rx_count = ch->rx_count;
1421         goto out;
1422 }
1423
1424 static int veth_open(struct net_device *dev)
1425 {
1426         struct veth_priv *priv = netdev_priv(dev);
1427         struct net_device *peer = rtnl_dereference(priv->peer);
1428         int err;
1429
1430         if (!peer)
1431                 return -ENOTCONN;
1432
1433         if (priv->_xdp_prog) {
1434                 err = veth_enable_xdp(dev);
1435                 if (err)
1436                         return err;
1437         } else if (veth_gro_requested(dev)) {
1438                 err = veth_napi_enable(dev);
1439                 if (err)
1440                         return err;
1441         }
1442
1443         if (peer->flags & IFF_UP) {
1444                 netif_carrier_on(dev);
1445                 netif_carrier_on(peer);
1446         }
1447
1448         return 0;
1449 }
1450
1451 static int veth_close(struct net_device *dev)
1452 {
1453         struct veth_priv *priv = netdev_priv(dev);
1454         struct net_device *peer = rtnl_dereference(priv->peer);
1455
1456         netif_carrier_off(dev);
1457         if (peer)
1458                 netif_carrier_off(peer);
1459
1460         if (priv->_xdp_prog)
1461                 veth_disable_xdp(dev);
1462         else if (veth_gro_requested(dev))
1463                 veth_napi_del(dev);
1464
1465         return 0;
1466 }
1467
1468 static int is_valid_veth_mtu(int mtu)
1469 {
1470         return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1471 }
1472
1473 static int veth_alloc_queues(struct net_device *dev)
1474 {
1475         struct veth_priv *priv = netdev_priv(dev);
1476         int i;
1477
1478         priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL_ACCOUNT);
1479         if (!priv->rq)
1480                 return -ENOMEM;
1481
1482         for (i = 0; i < dev->num_rx_queues; i++) {
1483                 priv->rq[i].dev = dev;
1484                 u64_stats_init(&priv->rq[i].stats.syncp);
1485         }
1486
1487         return 0;
1488 }
1489
1490 static void veth_free_queues(struct net_device *dev)
1491 {
1492         struct veth_priv *priv = netdev_priv(dev);
1493
1494         kfree(priv->rq);
1495 }
1496
1497 static int veth_dev_init(struct net_device *dev)
1498 {
1499         int err;
1500
1501         dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
1502         if (!dev->lstats)
1503                 return -ENOMEM;
1504
1505         err = veth_alloc_queues(dev);
1506         if (err) {
1507                 free_percpu(dev->lstats);
1508                 return err;
1509         }
1510
1511         return 0;
1512 }
1513
1514 static void veth_dev_free(struct net_device *dev)
1515 {
1516         veth_free_queues(dev);
1517         free_percpu(dev->lstats);
1518 }
1519
1520 #ifdef CONFIG_NET_POLL_CONTROLLER
1521 static void veth_poll_controller(struct net_device *dev)
1522 {
1523         /* veth only receives frames when its peer sends one
1524          * Since it has nothing to do with disabling irqs, we are guaranteed
1525          * never to have pending data when we poll for it so
1526          * there is nothing to do here.
1527          *
1528          * We need this though so netpoll recognizes us as an interface that
1529          * supports polling, which enables bridge devices in virt setups to
1530          * still use netconsole
1531          */
1532 }
1533 #endif  /* CONFIG_NET_POLL_CONTROLLER */
1534
1535 static int veth_get_iflink(const struct net_device *dev)
1536 {
1537         struct veth_priv *priv = netdev_priv(dev);
1538         struct net_device *peer;
1539         int iflink;
1540
1541         rcu_read_lock();
1542         peer = rcu_dereference(priv->peer);
1543         iflink = peer ? peer->ifindex : 0;
1544         rcu_read_unlock();
1545
1546         return iflink;
1547 }
1548
1549 static netdev_features_t veth_fix_features(struct net_device *dev,
1550                                            netdev_features_t features)
1551 {
1552         struct veth_priv *priv = netdev_priv(dev);
1553         struct net_device *peer;
1554
1555         peer = rtnl_dereference(priv->peer);
1556         if (peer) {
1557                 struct veth_priv *peer_priv = netdev_priv(peer);
1558
1559                 if (peer_priv->_xdp_prog)
1560                         features &= ~NETIF_F_GSO_SOFTWARE;
1561         }
1562         if (priv->_xdp_prog)
1563                 features |= NETIF_F_GRO;
1564
1565         return features;
1566 }
1567
1568 static int veth_set_features(struct net_device *dev,
1569                              netdev_features_t features)
1570 {
1571         netdev_features_t changed = features ^ dev->features;
1572         struct veth_priv *priv = netdev_priv(dev);
1573         struct net_device *peer;
1574         int err;
1575
1576         if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1577                 return 0;
1578
1579         peer = rtnl_dereference(priv->peer);
1580         if (features & NETIF_F_GRO) {
1581                 err = veth_napi_enable(dev);
1582                 if (err)
1583                         return err;
1584
1585                 if (peer)
1586                         xdp_features_set_redirect_target(peer, true);
1587         } else {
1588                 if (peer)
1589                         xdp_features_clear_redirect_target(peer);
1590                 veth_napi_del(dev);
1591         }
1592         return 0;
1593 }
1594
1595 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1596 {
1597         struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1598         struct net_device *peer;
1599
1600         if (new_hr < 0)
1601                 new_hr = 0;
1602
1603         rcu_read_lock();
1604         peer = rcu_dereference(priv->peer);
1605         if (unlikely(!peer))
1606                 goto out;
1607
1608         peer_priv = netdev_priv(peer);
1609         priv->requested_headroom = new_hr;
1610         new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1611         dev->needed_headroom = new_hr;
1612         peer->needed_headroom = new_hr;
1613
1614 out:
1615         rcu_read_unlock();
1616 }
1617
1618 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1619                         struct netlink_ext_ack *extack)
1620 {
1621         struct veth_priv *priv = netdev_priv(dev);
1622         struct bpf_prog *old_prog;
1623         struct net_device *peer;
1624         unsigned int max_mtu;
1625         int err;
1626
1627         old_prog = priv->_xdp_prog;
1628         priv->_xdp_prog = prog;
1629         peer = rtnl_dereference(priv->peer);
1630
1631         if (prog) {
1632                 if (!peer) {
1633                         NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1634                         err = -ENOTCONN;
1635                         goto err;
1636                 }
1637
1638                 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) -
1639                           peer->hard_header_len;
1640                 /* Allow increasing the max_mtu if the program supports
1641                  * XDP fragments.
1642                  */
1643                 if (prog->aux->xdp_has_frags)
1644                         max_mtu += PAGE_SIZE * MAX_SKB_FRAGS;
1645
1646                 if (peer->mtu > max_mtu) {
1647                         NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1648                         err = -ERANGE;
1649                         goto err;
1650                 }
1651
1652                 if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1653                         NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1654                         err = -ENOSPC;
1655                         goto err;
1656                 }
1657
1658                 if (dev->flags & IFF_UP) {
1659                         err = veth_enable_xdp(dev);
1660                         if (err) {
1661                                 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1662                                 goto err;
1663                         }
1664                 }
1665
1666                 if (!old_prog) {
1667                         peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1668                         peer->max_mtu = max_mtu;
1669                 }
1670
1671                 xdp_features_set_redirect_target(peer, true);
1672         }
1673
1674         if (old_prog) {
1675                 if (!prog) {
1676                         if (peer && !veth_gro_requested(dev))
1677                                 xdp_features_clear_redirect_target(peer);
1678
1679                         if (dev->flags & IFF_UP)
1680                                 veth_disable_xdp(dev);
1681
1682                         if (peer) {
1683                                 peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1684                                 peer->max_mtu = ETH_MAX_MTU;
1685                         }
1686                 }
1687                 bpf_prog_put(old_prog);
1688         }
1689
1690         if ((!!old_prog ^ !!prog) && peer)
1691                 netdev_update_features(peer);
1692
1693         return 0;
1694 err:
1695         priv->_xdp_prog = old_prog;
1696
1697         return err;
1698 }
1699
1700 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1701 {
1702         switch (xdp->command) {
1703         case XDP_SETUP_PROG:
1704                 return veth_xdp_set(dev, xdp->prog, xdp->extack);
1705         default:
1706                 return -EINVAL;
1707         }
1708 }
1709
1710 static int veth_xdp_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp)
1711 {
1712         struct veth_xdp_buff *_ctx = (void *)ctx;
1713
1714         if (!_ctx->skb)
1715                 return -ENODATA;
1716
1717         *timestamp = skb_hwtstamps(_ctx->skb)->hwtstamp;
1718         return 0;
1719 }
1720
1721 static int veth_xdp_rx_hash(const struct xdp_md *ctx, u32 *hash,
1722                             enum xdp_rss_hash_type *rss_type)
1723 {
1724         struct veth_xdp_buff *_ctx = (void *)ctx;
1725         struct sk_buff *skb = _ctx->skb;
1726
1727         if (!skb)
1728                 return -ENODATA;
1729
1730         *hash = skb_get_hash(skb);
1731         *rss_type = skb->l4_hash ? XDP_RSS_TYPE_L4_ANY : XDP_RSS_TYPE_NONE;
1732
1733         return 0;
1734 }
1735
1736 static const struct net_device_ops veth_netdev_ops = {
1737         .ndo_init            = veth_dev_init,
1738         .ndo_open            = veth_open,
1739         .ndo_stop            = veth_close,
1740         .ndo_start_xmit      = veth_xmit,
1741         .ndo_get_stats64     = veth_get_stats64,
1742         .ndo_set_rx_mode     = veth_set_multicast_list,
1743         .ndo_set_mac_address = eth_mac_addr,
1744 #ifdef CONFIG_NET_POLL_CONTROLLER
1745         .ndo_poll_controller    = veth_poll_controller,
1746 #endif
1747         .ndo_get_iflink         = veth_get_iflink,
1748         .ndo_fix_features       = veth_fix_features,
1749         .ndo_set_features       = veth_set_features,
1750         .ndo_features_check     = passthru_features_check,
1751         .ndo_set_rx_headroom    = veth_set_rx_headroom,
1752         .ndo_bpf                = veth_xdp,
1753         .ndo_xdp_xmit           = veth_ndo_xdp_xmit,
1754         .ndo_get_peer_dev       = veth_peer_dev,
1755 };
1756
1757 static const struct xdp_metadata_ops veth_xdp_metadata_ops = {
1758         .xmo_rx_timestamp               = veth_xdp_rx_timestamp,
1759         .xmo_rx_hash                    = veth_xdp_rx_hash,
1760 };
1761
1762 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1763                        NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1764                        NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1765                        NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1766                        NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1767
1768 static void veth_setup(struct net_device *dev)
1769 {
1770         ether_setup(dev);
1771
1772         dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1773         dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1774         dev->priv_flags |= IFF_NO_QUEUE;
1775         dev->priv_flags |= IFF_PHONY_HEADROOM;
1776
1777         dev->netdev_ops = &veth_netdev_ops;
1778         dev->xdp_metadata_ops = &veth_xdp_metadata_ops;
1779         dev->ethtool_ops = &veth_ethtool_ops;
1780         dev->features |= NETIF_F_LLTX;
1781         dev->features |= VETH_FEATURES;
1782         dev->vlan_features = dev->features &
1783                              ~(NETIF_F_HW_VLAN_CTAG_TX |
1784                                NETIF_F_HW_VLAN_STAG_TX |
1785                                NETIF_F_HW_VLAN_CTAG_RX |
1786                                NETIF_F_HW_VLAN_STAG_RX);
1787         dev->needs_free_netdev = true;
1788         dev->priv_destructor = veth_dev_free;
1789         dev->max_mtu = ETH_MAX_MTU;
1790
1791         dev->hw_features = VETH_FEATURES;
1792         dev->hw_enc_features = VETH_FEATURES;
1793         dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1794         netif_set_tso_max_size(dev, GSO_MAX_SIZE);
1795 }
1796
1797 /*
1798  * netlink interface
1799  */
1800
1801 static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1802                          struct netlink_ext_ack *extack)
1803 {
1804         if (tb[IFLA_ADDRESS]) {
1805                 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1806                         return -EINVAL;
1807                 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1808                         return -EADDRNOTAVAIL;
1809         }
1810         if (tb[IFLA_MTU]) {
1811                 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1812                         return -EINVAL;
1813         }
1814         return 0;
1815 }
1816
1817 static struct rtnl_link_ops veth_link_ops;
1818
1819 static void veth_disable_gro(struct net_device *dev)
1820 {
1821         dev->features &= ~NETIF_F_GRO;
1822         dev->wanted_features &= ~NETIF_F_GRO;
1823         netdev_update_features(dev);
1824 }
1825
1826 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1827 {
1828         int err;
1829
1830         if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1831                 err = netif_set_real_num_tx_queues(dev, 1);
1832                 if (err)
1833                         return err;
1834         }
1835         if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1836                 err = netif_set_real_num_rx_queues(dev, 1);
1837                 if (err)
1838                         return err;
1839         }
1840         return 0;
1841 }
1842
1843 static int veth_newlink(struct net *src_net, struct net_device *dev,
1844                         struct nlattr *tb[], struct nlattr *data[],
1845                         struct netlink_ext_ack *extack)
1846 {
1847         int err;
1848         struct net_device *peer;
1849         struct veth_priv *priv;
1850         char ifname[IFNAMSIZ];
1851         struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1852         unsigned char name_assign_type;
1853         struct ifinfomsg *ifmp;
1854         struct net *net;
1855
1856         /*
1857          * create and register peer first
1858          */
1859         if (data != NULL && data[VETH_INFO_PEER] != NULL) {
1860                 struct nlattr *nla_peer;
1861
1862                 nla_peer = data[VETH_INFO_PEER];
1863                 ifmp = nla_data(nla_peer);
1864                 err = rtnl_nla_parse_ifla(peer_tb,
1865                                           nla_data(nla_peer) + sizeof(struct ifinfomsg),
1866                                           nla_len(nla_peer) - sizeof(struct ifinfomsg),
1867                                           NULL);
1868                 if (err < 0)
1869                         return err;
1870
1871                 err = veth_validate(peer_tb, NULL, extack);
1872                 if (err < 0)
1873                         return err;
1874
1875                 tbp = peer_tb;
1876         } else {
1877                 ifmp = NULL;
1878                 tbp = tb;
1879         }
1880
1881         if (ifmp && tbp[IFLA_IFNAME]) {
1882                 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1883                 name_assign_type = NET_NAME_USER;
1884         } else {
1885                 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1886                 name_assign_type = NET_NAME_ENUM;
1887         }
1888
1889         net = rtnl_link_get_net(src_net, tbp);
1890         if (IS_ERR(net))
1891                 return PTR_ERR(net);
1892
1893         peer = rtnl_create_link(net, ifname, name_assign_type,
1894                                 &veth_link_ops, tbp, extack);
1895         if (IS_ERR(peer)) {
1896                 put_net(net);
1897                 return PTR_ERR(peer);
1898         }
1899
1900         if (!ifmp || !tbp[IFLA_ADDRESS])
1901                 eth_hw_addr_random(peer);
1902
1903         if (ifmp && (dev->ifindex != 0))
1904                 peer->ifindex = ifmp->ifi_index;
1905
1906         netif_inherit_tso_max(peer, dev);
1907
1908         err = register_netdevice(peer);
1909         put_net(net);
1910         net = NULL;
1911         if (err < 0)
1912                 goto err_register_peer;
1913
1914         /* keep GRO disabled by default to be consistent with the established
1915          * veth behavior
1916          */
1917         veth_disable_gro(peer);
1918         netif_carrier_off(peer);
1919
1920         err = rtnl_configure_link(peer, ifmp, 0, NULL);
1921         if (err < 0)
1922                 goto err_configure_peer;
1923
1924         /*
1925          * register dev last
1926          *
1927          * note, that since we've registered new device the dev's name
1928          * should be re-allocated
1929          */
1930
1931         if (tb[IFLA_ADDRESS] == NULL)
1932                 eth_hw_addr_random(dev);
1933
1934         if (tb[IFLA_IFNAME])
1935                 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1936         else
1937                 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1938
1939         err = register_netdevice(dev);
1940         if (err < 0)
1941                 goto err_register_dev;
1942
1943         netif_carrier_off(dev);
1944
1945         /*
1946          * tie the deviced together
1947          */
1948
1949         priv = netdev_priv(dev);
1950         rcu_assign_pointer(priv->peer, peer);
1951         err = veth_init_queues(dev, tb);
1952         if (err)
1953                 goto err_queues;
1954
1955         priv = netdev_priv(peer);
1956         rcu_assign_pointer(priv->peer, dev);
1957         err = veth_init_queues(peer, tb);
1958         if (err)
1959                 goto err_queues;
1960
1961         veth_disable_gro(dev);
1962         /* update XDP supported features */
1963         veth_set_xdp_features(dev);
1964         veth_set_xdp_features(peer);
1965
1966         return 0;
1967
1968 err_queues:
1969         unregister_netdevice(dev);
1970 err_register_dev:
1971         /* nothing to do */
1972 err_configure_peer:
1973         unregister_netdevice(peer);
1974         return err;
1975
1976 err_register_peer:
1977         free_netdev(peer);
1978         return err;
1979 }
1980
1981 static void veth_dellink(struct net_device *dev, struct list_head *head)
1982 {
1983         struct veth_priv *priv;
1984         struct net_device *peer;
1985
1986         priv = netdev_priv(dev);
1987         peer = rtnl_dereference(priv->peer);
1988
1989         /* Note : dellink() is called from default_device_exit_batch(),
1990          * before a rcu_synchronize() point. The devices are guaranteed
1991          * not being freed before one RCU grace period.
1992          */
1993         RCU_INIT_POINTER(priv->peer, NULL);
1994         unregister_netdevice_queue(dev, head);
1995
1996         if (peer) {
1997                 priv = netdev_priv(peer);
1998                 RCU_INIT_POINTER(priv->peer, NULL);
1999                 unregister_netdevice_queue(peer, head);
2000         }
2001 }
2002
2003 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
2004         [VETH_INFO_PEER]        = { .len = sizeof(struct ifinfomsg) },
2005 };
2006
2007 static struct net *veth_get_link_net(const struct net_device *dev)
2008 {
2009         struct veth_priv *priv = netdev_priv(dev);
2010         struct net_device *peer = rtnl_dereference(priv->peer);
2011
2012         return peer ? dev_net(peer) : dev_net(dev);
2013 }
2014
2015 static unsigned int veth_get_num_queues(void)
2016 {
2017         /* enforce the same queue limit as rtnl_create_link */
2018         int queues = num_possible_cpus();
2019
2020         if (queues > 4096)
2021                 queues = 4096;
2022         return queues;
2023 }
2024
2025 static struct rtnl_link_ops veth_link_ops = {
2026         .kind           = DRV_NAME,
2027         .priv_size      = sizeof(struct veth_priv),
2028         .setup          = veth_setup,
2029         .validate       = veth_validate,
2030         .newlink        = veth_newlink,
2031         .dellink        = veth_dellink,
2032         .policy         = veth_policy,
2033         .maxtype        = VETH_INFO_MAX,
2034         .get_link_net   = veth_get_link_net,
2035         .get_num_tx_queues      = veth_get_num_queues,
2036         .get_num_rx_queues      = veth_get_num_queues,
2037 };
2038
2039 /*
2040  * init/fini
2041  */
2042
2043 static __init int veth_init(void)
2044 {
2045         return rtnl_link_register(&veth_link_ops);
2046 }
2047
2048 static __exit void veth_exit(void)
2049 {
2050         rtnl_link_unregister(&veth_link_ops);
2051 }
2052
2053 module_init(veth_init);
2054 module_exit(veth_exit);
2055
2056 MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
2057 MODULE_LICENSE("GPL v2");
2058 MODULE_ALIAS_RTNL_LINK(DRV_NAME);