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
  30 #define DRV_NAME        "veth"
  31 #define DRV_VERSION     "1.0"
  32
  33 #define VETH_XDP_FLAG           BIT(0)
  34 #define VETH_RING_SIZE          256
  35 #define VETH_XDP_HEADROOM       (XDP_PACKET_HEADROOM + NET_IP_ALIGN)
  36
  37 #define VETH_XDP_TX_BULK_SIZE   16
  38 #define VETH_XDP_BATCH          16
  39
  40 struct veth_stats {
  41         u64     rx_drops;
  42         /* xdp */
  43         u64     xdp_packets;
  44         u64     xdp_bytes;
  45         u64     xdp_redirect;
  46         u64     xdp_drops;
  47         u64     xdp_tx;
  48         u64     xdp_tx_err;
  49         u64     peer_tq_xdp_xmit;
  50         u64     peer_tq_xdp_xmit_err;
  51 };
  52
  53 struct veth_rq_stats {
  54         struct veth_stats       vs;
  55         struct u64_stats_sync   syncp;
  56 };
  57
  58 struct veth_rq {
  59         struct napi_struct      xdp_napi;
  60         struct napi_struct __rcu *napi; /* points to xdp_napi when the latter is initialized */
  61         struct net_device       *dev;
  62         struct bpf_prog __rcu   *xdp_prog;
  63         struct xdp_mem_info     xdp_mem;
  64         struct veth_rq_stats    stats;
  65         bool                    rx_notify_masked;
  66         struct ptr_ring         xdp_ring;
  67         struct xdp_rxq_info     xdp_rxq;
  68 };
  69
  70 struct veth_priv {
  71         struct net_device __rcu *peer;
  72         atomic64_t              dropped;
  73         struct bpf_prog         *_xdp_prog;
  74         struct veth_rq          *rq;
  75         unsigned int            requested_headroom;
  76 };
  77
  78 struct veth_xdp_tx_bq {
  79         struct xdp_frame *q[VETH_XDP_TX_BULK_SIZE];
  80         unsigned int count;
  81 };
  82
  83 /*
  84  * ethtool interface
  85  */
  86
  87 struct veth_q_stat_desc {
  88         char    desc[ETH_GSTRING_LEN];
  89         size_t  offset;
  90 };
  91
  92 #define VETH_RQ_STAT(m) offsetof(struct veth_stats, m)
  93
  94 static const struct veth_q_stat_desc veth_rq_stats_desc[] = {
  95         { "xdp_packets",        VETH_RQ_STAT(xdp_packets) },
  96         { "xdp_bytes",          VETH_RQ_STAT(xdp_bytes) },
  97         { "drops",              VETH_RQ_STAT(rx_drops) },
  98         { "xdp_redirect",       VETH_RQ_STAT(xdp_redirect) },
  99         { "xdp_drops",          VETH_RQ_STAT(xdp_drops) },
 100         { "xdp_tx",             VETH_RQ_STAT(xdp_tx) },
 101         { "xdp_tx_errors",      VETH_RQ_STAT(xdp_tx_err) },
 102 };
 103
 104 #define VETH_RQ_STATS_LEN       ARRAY_SIZE(veth_rq_stats_desc)
 105
 106 static const struct veth_q_stat_desc veth_tq_stats_desc[] = {
 107         { "xdp_xmit",           VETH_RQ_STAT(peer_tq_xdp_xmit) },
 108         { "xdp_xmit_errors",    VETH_RQ_STAT(peer_tq_xdp_xmit_err) },
 109 };
 110
 111 #define VETH_TQ_STATS_LEN       ARRAY_SIZE(veth_tq_stats_desc)
 112
 113 static struct {
 114         const char string[ETH_GSTRING_LEN];
 115 } ethtool_stats_keys[] = {
 116         { "peer_ifindex" },
 117 };
 118
 119 static int veth_get_link_ksettings(struct net_device *dev,
 120                                    struct ethtool_link_ksettings *cmd)
 121 {
 122         cmd->base.speed         = SPEED_10000;
 123         cmd->base.duplex        = DUPLEX_FULL;
 124         cmd->base.port          = PORT_TP;
 125         cmd->base.autoneg       = AUTONEG_DISABLE;
 126         return 0;
 127 }
 128
 129 static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info)
 130 {
 131         strlcpy(info->driver, DRV_NAME, sizeof(info->driver));
 132         strlcpy(info->version, DRV_VERSION, sizeof(info->version));
 133 }
 134
 135 static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
 136 {
 137         char *p = (char *)buf;
 138         int i, j;
 139
 140         switch(stringset) {
 141         case ETH_SS_STATS:
 142                 memcpy(p, &ethtool_stats_keys, sizeof(ethtool_stats_keys));
 143                 p += sizeof(ethtool_stats_keys);
 144                 for (i = 0; i < dev->real_num_rx_queues; i++) {
 145                         for (j = 0; j < VETH_RQ_STATS_LEN; j++) {
 146                                 snprintf(p, ETH_GSTRING_LEN,
 147                                          "rx_queue_%u_%.18s",
 148                                          i, veth_rq_stats_desc[j].desc);
 149                                 p += ETH_GSTRING_LEN;
 150                         }
 151                 }
 152                 for (i = 0; i < dev->real_num_tx_queues; i++) {
 153                         for (j = 0; j < VETH_TQ_STATS_LEN; j++) {
 154                                 snprintf(p, ETH_GSTRING_LEN,
 155                                          "tx_queue_%u_%.18s",
 156                                          i, veth_tq_stats_desc[j].desc);
 157                                 p += ETH_GSTRING_LEN;
 158                         }
 159                 }
 160                 break;
 161         }
 162 }
 163
 164 static int veth_get_sset_count(struct net_device *dev, int sset)
 165 {
 166         switch (sset) {
 167         case ETH_SS_STATS:
 168                 return ARRAY_SIZE(ethtool_stats_keys) +
 169                        VETH_RQ_STATS_LEN * dev->real_num_rx_queues +
 170                        VETH_TQ_STATS_LEN * dev->real_num_tx_queues;
 171         default:
 172                 return -EOPNOTSUPP;
 173         }
 174 }
 175
 176 static void veth_get_ethtool_stats(struct net_device *dev,
 177                 struct ethtool_stats *stats, u64 *data)
 178 {
 179         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 180         struct net_device *peer = rtnl_dereference(priv->peer);
 181         int i, j, idx;
 182
 183         data[0] = peer ? peer->ifindex : 0;
 184         idx = 1;
 185         for (i = 0; i < dev->real_num_rx_queues; i++) {
 186                 const struct veth_rq_stats *rq_stats = &priv->rq[i].stats;
 187                 const void *stats_base = (void *)&rq_stats->vs;
 188                 unsigned int start;
 189                 size_t offset;
 190
 191                 do {
 192                         start = u64_stats_fetch_begin_irq(&rq_stats->syncp);
 193                         for (j = 0; j < VETH_RQ_STATS_LEN; j++) {
 194                                 offset = veth_rq_stats_desc[j].offset;
 195                                 data[idx + j] = *(u64 *)(stats_base + offset);
 196                         }
 197                 } while (u64_stats_fetch_retry_irq(&rq_stats->syncp, start));
 198                 idx += VETH_RQ_STATS_LEN;
 199         }
 200
 201         if (!peer)
 202                 return;
 203
 204         rcv_priv = netdev_priv(peer);
 205         for (i = 0; i < peer->real_num_rx_queues; i++) {
 206                 const struct veth_rq_stats *rq_stats = &rcv_priv->rq[i].stats;
 207                 const void *base = (void *)&rq_stats->vs;
 208                 unsigned int start, tx_idx = idx;
 209                 size_t offset;
 210
 211                 tx_idx += (i % dev->real_num_tx_queues) * VETH_TQ_STATS_LEN;
 212                 do {
 213                         start = u64_stats_fetch_begin_irq(&rq_stats->syncp);
 214                         for (j = 0; j < VETH_TQ_STATS_LEN; j++) {
 215                                 offset = veth_tq_stats_desc[j].offset;
 216                                 data[tx_idx + j] += *(u64 *)(base + offset);
 217                         }
 218                 } while (u64_stats_fetch_retry_irq(&rq_stats->syncp, start));
 219         }
 220 }
 221
 222 static void veth_get_channels(struct net_device *dev,
 223                               struct ethtool_channels *channels)
 224 {
 225         channels->tx_count = dev->real_num_tx_queues;
 226         channels->rx_count = dev->real_num_rx_queues;
 227         channels->max_tx = dev->num_tx_queues;
 228         channels->max_rx = dev->num_rx_queues;
 229 }
 230
 231 static int veth_set_channels(struct net_device *dev,
 232                              struct ethtool_channels *ch);
 233
 234 static const struct ethtool_ops veth_ethtool_ops = {
 235         .get_drvinfo            = veth_get_drvinfo,
 236         .get_link               = ethtool_op_get_link,
 237         .get_strings            = veth_get_strings,
 238         .get_sset_count         = veth_get_sset_count,
 239         .get_ethtool_stats      = veth_get_ethtool_stats,
 240         .get_link_ksettings     = veth_get_link_ksettings,
 241         .get_ts_info            = ethtool_op_get_ts_info,
 242         .get_channels           = veth_get_channels,
 243         .set_channels           = veth_set_channels,
 244 };
 245
 246 /* general routines */
 247
 248 static bool veth_is_xdp_frame(void *ptr)
 249 {
 250         return (unsigned long)ptr & VETH_XDP_FLAG;
 251 }
 252
 253 static struct xdp_frame *veth_ptr_to_xdp(void *ptr)
 254 {
 255         return (void *)((unsigned long)ptr & ~VETH_XDP_FLAG);
 256 }
 257
 258 static void *veth_xdp_to_ptr(struct xdp_frame *xdp)
 259 {
 260         return (void *)((unsigned long)xdp | VETH_XDP_FLAG);
 261 }
 262
 263 static void veth_ptr_free(void *ptr)
 264 {
 265         if (veth_is_xdp_frame(ptr))
 266                 xdp_return_frame(veth_ptr_to_xdp(ptr));
 267         else
 268                 kfree_skb(ptr);
 269 }
 270
 271 static void __veth_xdp_flush(struct veth_rq *rq)
 272 {
 273         /* Write ptr_ring before reading rx_notify_masked */
 274         smp_mb();
 275         if (!rq->rx_notify_masked) {
 276                 rq->rx_notify_masked = true;
 277                 napi_schedule(&rq->xdp_napi);
 278         }
 279 }
 280
 281 static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb)
 282 {
 283         if (unlikely(ptr_ring_produce(&rq->xdp_ring, skb))) {
 284                 dev_kfree_skb_any(skb);
 285                 return NET_RX_DROP;
 286         }
 287
 288         return NET_RX_SUCCESS;
 289 }
 290
 291 static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb,
 292                             struct veth_rq *rq, bool xdp)
 293 {
 294         return __dev_forward_skb(dev, skb) ?: xdp ?
 295                 veth_xdp_rx(rq, skb) :
 296                 netif_rx(skb);
 297 }
 298
 299 /* return true if the specified skb has chances of GRO aggregation
 300  * Don't strive for accuracy, but try to avoid GRO overhead in the most
 301  * common scenarios.
 302  * When XDP is enabled, all traffic is considered eligible, as the xmit
 303  * device has TSO off.
 304  * When TSO is enabled on the xmit device, we are likely interested only
 305  * in UDP aggregation, explicitly check for that if the skb is suspected
 306  * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets -
 307  * to belong to locally generated UDP traffic.
 308  */
 309 static bool veth_skb_is_eligible_for_gro(const struct net_device *dev,
 310                                          const struct net_device *rcv,
 311                                          const struct sk_buff *skb)
 312 {
 313         return !(dev->features & NETIF_F_ALL_TSO) ||
 314                 (skb->destructor == sock_wfree &&
 315                  rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD));
 316 }
 317
 318 static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 319 {
 320         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 321         struct veth_rq *rq = NULL;
 322         struct net_device *rcv;
 323         int length = skb->len;
 324         bool use_napi = false;
 325         int rxq;
 326
 327         rcu_read_lock();
 328         rcv = rcu_dereference(priv->peer);
 329         if (unlikely(!rcv)) {
 330                 kfree_skb(skb);
 331                 goto drop;
 332         }
 333
 334         rcv_priv = netdev_priv(rcv);
 335         rxq = skb_get_queue_mapping(skb);
 336         if (rxq < rcv->real_num_rx_queues) {
 337                 rq = &rcv_priv->rq[rxq];
 338
 339                 /* The napi pointer is available when an XDP program is
 340                  * attached or when GRO is enabled
 341                  * Don't bother with napi/GRO if the skb can't be aggregated
 342                  */
 343                 use_napi = rcu_access_pointer(rq->napi) &&
 344                            veth_skb_is_eligible_for_gro(dev, rcv, skb);
 345                 skb_record_rx_queue(skb, rxq);
 346         }
 347
 348         skb_tx_timestamp(skb);
 349         if (likely(veth_forward_skb(rcv, skb, rq, use_napi) == NET_RX_SUCCESS)) {
 350                 if (!use_napi)
 351                         dev_lstats_add(dev, length);
 352         } else {
 353 drop:
 354                 atomic64_inc(&priv->dropped);
 355         }
 356
 357         if (use_napi)
 358                 __veth_xdp_flush(rq);
 359
 360         rcu_read_unlock();
 361
 362         return NETDEV_TX_OK;
 363 }
 364
 365 static u64 veth_stats_tx(struct net_device *dev, u64 *packets, u64 *bytes)
 366 {
 367         struct veth_priv *priv = netdev_priv(dev);
 368
 369         dev_lstats_read(dev, packets, bytes);
 370         return atomic64_read(&priv->dropped);
 371 }
 372
 373 static void veth_stats_rx(struct veth_stats *result, struct net_device *dev)
 374 {
 375         struct veth_priv *priv = netdev_priv(dev);
 376         int i;
 377
 378         result->peer_tq_xdp_xmit_err = 0;
 379         result->xdp_packets = 0;
 380         result->xdp_tx_err = 0;
 381         result->xdp_bytes = 0;
 382         result->rx_drops = 0;
 383         for (i = 0; i < dev->num_rx_queues; i++) {
 384                 u64 packets, bytes, drops, xdp_tx_err, peer_tq_xdp_xmit_err;
 385                 struct veth_rq_stats *stats = &priv->rq[i].stats;
 386                 unsigned int start;
 387
 388                 do {
 389                         start = u64_stats_fetch_begin_irq(&stats->syncp);
 390                         peer_tq_xdp_xmit_err = stats->vs.peer_tq_xdp_xmit_err;
 391                         xdp_tx_err = stats->vs.xdp_tx_err;
 392                         packets = stats->vs.xdp_packets;
 393                         bytes = stats->vs.xdp_bytes;
 394                         drops = stats->vs.rx_drops;
 395                 } while (u64_stats_fetch_retry_irq(&stats->syncp, start));
 396                 result->peer_tq_xdp_xmit_err += peer_tq_xdp_xmit_err;
 397                 result->xdp_tx_err += xdp_tx_err;
 398                 result->xdp_packets += packets;
 399                 result->xdp_bytes += bytes;
 400                 result->rx_drops += drops;
 401         }
 402 }
 403
 404 static void veth_get_stats64(struct net_device *dev,
 405                              struct rtnl_link_stats64 *tot)
 406 {
 407         struct veth_priv *priv = netdev_priv(dev);
 408         struct net_device *peer;
 409         struct veth_stats rx;
 410         u64 packets, bytes;
 411
 412         tot->tx_dropped = veth_stats_tx(dev, &packets, &bytes);
 413         tot->tx_bytes = bytes;
 414         tot->tx_packets = packets;
 415
 416         veth_stats_rx(&rx, dev);
 417         tot->tx_dropped += rx.xdp_tx_err;
 418         tot->rx_dropped = rx.rx_drops + rx.peer_tq_xdp_xmit_err;
 419         tot->rx_bytes = rx.xdp_bytes;
 420         tot->rx_packets = rx.xdp_packets;
 421
 422         rcu_read_lock();
 423         peer = rcu_dereference(priv->peer);
 424         if (peer) {
 425                 veth_stats_tx(peer, &packets, &bytes);
 426                 tot->rx_bytes += bytes;
 427                 tot->rx_packets += packets;
 428
 429                 veth_stats_rx(&rx, peer);
 430                 tot->tx_dropped += rx.peer_tq_xdp_xmit_err;
 431                 tot->rx_dropped += rx.xdp_tx_err;
 432                 tot->tx_bytes += rx.xdp_bytes;
 433                 tot->tx_packets += rx.xdp_packets;
 434         }
 435         rcu_read_unlock();
 436 }
 437
 438 /* fake multicast ability */
 439 static void veth_set_multicast_list(struct net_device *dev)
 440 {
 441 }
 442
 443 static struct sk_buff *veth_build_skb(void *head, int headroom, int len,
 444                                       int buflen)
 445 {
 446         struct sk_buff *skb;
 447
 448         skb = build_skb(head, buflen);
 449         if (!skb)
 450                 return NULL;
 451
 452         skb_reserve(skb, headroom);
 453         skb_put(skb, len);
 454
 455         return skb;
 456 }
 457
 458 static int veth_select_rxq(struct net_device *dev)
 459 {
 460         return smp_processor_id() % dev->real_num_rx_queues;
 461 }
 462
 463 static struct net_device *veth_peer_dev(struct net_device *dev)
 464 {
 465         struct veth_priv *priv = netdev_priv(dev);
 466
 467         /* Callers must be under RCU read side. */
 468         return rcu_dereference(priv->peer);
 469 }
 470
 471 static int veth_xdp_xmit(struct net_device *dev, int n,
 472                          struct xdp_frame **frames,
 473                          u32 flags, bool ndo_xmit)
 474 {
 475         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 476         int i, ret = -ENXIO, nxmit = 0;
 477         struct net_device *rcv;
 478         unsigned int max_len;
 479         struct veth_rq *rq;
 480
 481         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
 482                 return -EINVAL;
 483
 484         rcu_read_lock();
 485         rcv = rcu_dereference(priv->peer);
 486         if (unlikely(!rcv))
 487                 goto out;
 488
 489         rcv_priv = netdev_priv(rcv);
 490         rq = &rcv_priv->rq[veth_select_rxq(rcv)];
 491         /* The napi pointer is set if NAPI is enabled, which ensures that
 492          * xdp_ring is initialized on receive side and the peer device is up.
 493          */
 494         if (!rcu_access_pointer(rq->napi))
 495                 goto out;
 496
 497         max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN;
 498
 499         spin_lock(&rq->xdp_ring.producer_lock);
 500         for (i = 0; i < n; i++) {
 501                 struct xdp_frame *frame = frames[i];
 502                 void *ptr = veth_xdp_to_ptr(frame);
 503
 504                 if (unlikely(frame->len > max_len ||
 505                              __ptr_ring_produce(&rq->xdp_ring, ptr)))
 506                         break;
 507                 nxmit++;
 508         }
 509         spin_unlock(&rq->xdp_ring.producer_lock);
 510
 511         if (flags & XDP_XMIT_FLUSH)
 512                 __veth_xdp_flush(rq);
 513
 514         ret = nxmit;
 515         if (ndo_xmit) {
 516                 u64_stats_update_begin(&rq->stats.syncp);
 517                 rq->stats.vs.peer_tq_xdp_xmit += nxmit;
 518                 rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit;
 519                 u64_stats_update_end(&rq->stats.syncp);
 520         }
 521
 522 out:
 523         rcu_read_unlock();
 524
 525         return ret;
 526 }
 527
 528 static int veth_ndo_xdp_xmit(struct net_device *dev, int n,
 529                              struct xdp_frame **frames, u32 flags)
 530 {
 531         int err;
 532
 533         err = veth_xdp_xmit(dev, n, frames, flags, true);
 534         if (err < 0) {
 535                 struct veth_priv *priv = netdev_priv(dev);
 536
 537                 atomic64_add(n, &priv->dropped);
 538         }
 539
 540         return err;
 541 }
 542
 543 static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
 544 {
 545         int sent, i, err = 0, drops;
 546
 547         sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false);
 548         if (sent < 0) {
 549                 err = sent;
 550                 sent = 0;
 551         }
 552
 553         for (i = sent; unlikely(i < bq->count); i++)
 554                 xdp_return_frame(bq->q[i]);
 555
 556         drops = bq->count - sent;
 557         trace_xdp_bulk_tx(rq->dev, sent, drops, err);
 558
 559         u64_stats_update_begin(&rq->stats.syncp);
 560         rq->stats.vs.xdp_tx += sent;
 561         rq->stats.vs.xdp_tx_err += drops;
 562         u64_stats_update_end(&rq->stats.syncp);
 563
 564         bq->count = 0;
 565 }
 566
 567 static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
 568 {
 569         struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev);
 570         struct net_device *rcv;
 571         struct veth_rq *rcv_rq;
 572
 573         rcu_read_lock();
 574         veth_xdp_flush_bq(rq, bq);
 575         rcv = rcu_dereference(priv->peer);
 576         if (unlikely(!rcv))
 577                 goto out;
 578
 579         rcv_priv = netdev_priv(rcv);
 580         rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)];
 581         /* xdp_ring is initialized on receive side? */
 582         if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog)))
 583                 goto out;
 584
 585         __veth_xdp_flush(rcv_rq);
 586 out:
 587         rcu_read_unlock();
 588 }
 589
 590 static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp,
 591                        struct veth_xdp_tx_bq *bq)
 592 {
 593         struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp);
 594
 595         if (unlikely(!frame))
 596                 return -EOVERFLOW;
 597
 598         if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE))
 599                 veth_xdp_flush_bq(rq, bq);
 600
 601         bq->q[bq->count++] = frame;
 602
 603         return 0;
 604 }
 605
 606 static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq,
 607                                           struct xdp_frame *frame,
 608                                           struct veth_xdp_tx_bq *bq,
 609                                           struct veth_stats *stats)
 610 {
 611         struct xdp_frame orig_frame;
 612         struct bpf_prog *xdp_prog;
 613
 614         rcu_read_lock();
 615         xdp_prog = rcu_dereference(rq->xdp_prog);
 616         if (likely(xdp_prog)) {
 617                 struct xdp_buff xdp;
 618                 u32 act;
 619
 620                 xdp_convert_frame_to_buff(frame, &xdp);
 621                 xdp.rxq = &rq->xdp_rxq;
 622
 623                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 624
 625                 switch (act) {
 626                 case XDP_PASS:
 627                         if (xdp_update_frame_from_buff(&xdp, frame))
 628                                 goto err_xdp;
 629                         break;
 630                 case XDP_TX:
 631                         orig_frame = *frame;
 632                         xdp.rxq->mem = frame->mem;
 633                         if (unlikely(veth_xdp_tx(rq, &xdp, bq) < 0)) {
 634                                 trace_xdp_exception(rq->dev, xdp_prog, act);
 635                                 frame = &orig_frame;
 636                                 stats->rx_drops++;
 637                                 goto err_xdp;
 638                         }
 639                         stats->xdp_tx++;
 640                         rcu_read_unlock();
 641                         goto xdp_xmit;
 642                 case XDP_REDIRECT:
 643                         orig_frame = *frame;
 644                         xdp.rxq->mem = frame->mem;
 645                         if (xdp_do_redirect(rq->dev, &xdp, xdp_prog)) {
 646                                 frame = &orig_frame;
 647                                 stats->rx_drops++;
 648                                 goto err_xdp;
 649                         }
 650                         stats->xdp_redirect++;
 651                         rcu_read_unlock();
 652                         goto xdp_xmit;
 653                 default:
 654                         bpf_warn_invalid_xdp_action(act);
 655                         fallthrough;
 656                 case XDP_ABORTED:
 657                         trace_xdp_exception(rq->dev, xdp_prog, act);
 658                         fallthrough;
 659                 case XDP_DROP:
 660                         stats->xdp_drops++;
 661                         goto err_xdp;
 662                 }
 663         }
 664         rcu_read_unlock();
 665
 666         return frame;
 667 err_xdp:
 668         rcu_read_unlock();
 669         xdp_return_frame(frame);
 670 xdp_xmit:
 671         return NULL;
 672 }
 673
 674 /* frames array contains VETH_XDP_BATCH at most */
 675 static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames,
 676                                   int n_xdpf, struct veth_xdp_tx_bq *bq,
 677                                   struct veth_stats *stats)
 678 {
 679         void *skbs[VETH_XDP_BATCH];
 680         int i;
 681
 682         if (xdp_alloc_skb_bulk(skbs, n_xdpf,
 683                                GFP_ATOMIC | __GFP_ZERO) < 0) {
 684                 for (i = 0; i < n_xdpf; i++)
 685                         xdp_return_frame(frames[i]);
 686                 stats->rx_drops += n_xdpf;
 687
 688                 return;
 689         }
 690
 691         for (i = 0; i < n_xdpf; i++) {
 692                 struct sk_buff *skb = skbs[i];
 693
 694                 skb = __xdp_build_skb_from_frame(frames[i], skb,
 695                                                  rq->dev);
 696                 if (!skb) {
 697                         xdp_return_frame(frames[i]);
 698                         stats->rx_drops++;
 699                         continue;
 700                 }
 701                 napi_gro_receive(&rq->xdp_napi, skb);
 702         }
 703 }
 704
 705 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq,
 706                                         struct sk_buff *skb,
 707                                         struct veth_xdp_tx_bq *bq,
 708                                         struct veth_stats *stats)
 709 {
 710         u32 pktlen, headroom, act, metalen, frame_sz;
 711         void *orig_data, *orig_data_end;
 712         struct bpf_prog *xdp_prog;
 713         int mac_len, delta, off;
 714         struct xdp_buff xdp;
 715
 716         skb_prepare_for_gro(skb);
 717
 718         rcu_read_lock();
 719         xdp_prog = rcu_dereference(rq->xdp_prog);
 720         if (unlikely(!xdp_prog)) {
 721                 rcu_read_unlock();
 722                 goto out;
 723         }
 724
 725         mac_len = skb->data - skb_mac_header(skb);
 726         pktlen = skb->len + mac_len;
 727         headroom = skb_headroom(skb) - mac_len;
 728
 729         if (skb_shared(skb) || skb_head_is_locked(skb) ||
 730             skb_is_nonlinear(skb) || headroom < XDP_PACKET_HEADROOM) {
 731                 struct sk_buff *nskb;
 732                 int size, head_off;
 733                 void *head, *start;
 734                 struct page *page;
 735
 736                 size = SKB_DATA_ALIGN(VETH_XDP_HEADROOM + pktlen) +
 737                        SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 738                 if (size > PAGE_SIZE)
 739                         goto drop;
 740
 741                 page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
 742                 if (!page)
 743                         goto drop;
 744
 745                 head = page_address(page);
 746                 start = head + VETH_XDP_HEADROOM;
 747                 if (skb_copy_bits(skb, -mac_len, start, pktlen)) {
 748                         page_frag_free(head);
 749                         goto drop;
 750                 }
 751
 752                 nskb = veth_build_skb(head, VETH_XDP_HEADROOM + mac_len,
 753                                       skb->len, PAGE_SIZE);
 754                 if (!nskb) {
 755                         page_frag_free(head);
 756                         goto drop;
 757                 }
 758
 759                 skb_copy_header(nskb, skb);
 760                 head_off = skb_headroom(nskb) - skb_headroom(skb);
 761                 skb_headers_offset_update(nskb, head_off);
 762                 consume_skb(skb);
 763                 skb = nskb;
 764         }
 765
 766         /* SKB "head" area always have tailroom for skb_shared_info */
 767         frame_sz = skb_end_pointer(skb) - skb->head;
 768         frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 769         xdp_init_buff(&xdp, frame_sz, &rq->xdp_rxq);
 770         xdp_prepare_buff(&xdp, skb->head, skb->mac_header, pktlen, true);
 771
 772         orig_data = xdp.data;
 773         orig_data_end = xdp.data_end;
 774
 775         act = bpf_prog_run_xdp(xdp_prog, &xdp);
 776
 777         switch (act) {
 778         case XDP_PASS:
 779                 break;
 780         case XDP_TX:
 781                 get_page(virt_to_page(xdp.data));
 782                 consume_skb(skb);
 783                 xdp.rxq->mem = rq->xdp_mem;
 784                 if (unlikely(veth_xdp_tx(rq, &xdp, bq) < 0)) {
 785                         trace_xdp_exception(rq->dev, xdp_prog, act);
 786                         stats->rx_drops++;
 787                         goto err_xdp;
 788                 }
 789                 stats->xdp_tx++;
 790                 rcu_read_unlock();
 791                 goto xdp_xmit;
 792         case XDP_REDIRECT:
 793                 get_page(virt_to_page(xdp.data));
 794                 consume_skb(skb);
 795                 xdp.rxq->mem = rq->xdp_mem;
 796                 if (xdp_do_redirect(rq->dev, &xdp, xdp_prog)) {
 797                         stats->rx_drops++;
 798                         goto err_xdp;
 799                 }
 800                 stats->xdp_redirect++;
 801                 rcu_read_unlock();
 802                 goto xdp_xmit;
 803         default:
 804                 bpf_warn_invalid_xdp_action(act);
 805                 fallthrough;
 806         case XDP_ABORTED:
 807                 trace_xdp_exception(rq->dev, xdp_prog, act);
 808                 fallthrough;
 809         case XDP_DROP:
 810                 stats->xdp_drops++;
 811                 goto xdp_drop;
 812         }
 813         rcu_read_unlock();
 814
 815         /* check if bpf_xdp_adjust_head was used */
 816         delta = orig_data - xdp.data;
 817         off = mac_len + delta;
 818         if (off > 0)
 819                 __skb_push(skb, off);
 820         else if (off < 0)
 821                 __skb_pull(skb, -off);
 822         skb->mac_header -= delta;
 823
 824         /* check if bpf_xdp_adjust_tail was used */
 825         off = xdp.data_end - orig_data_end;
 826         if (off != 0)
 827                 __skb_put(skb, off); /* positive on grow, negative on shrink */
 828         skb->protocol = eth_type_trans(skb, rq->dev);
 829
 830         metalen = xdp.data - xdp.data_meta;
 831         if (metalen)
 832                 skb_metadata_set(skb, metalen);
 833 out:
 834         return skb;
 835 drop:
 836         stats->rx_drops++;
 837 xdp_drop:
 838         rcu_read_unlock();
 839         kfree_skb(skb);
 840         return NULL;
 841 err_xdp:
 842         rcu_read_unlock();
 843         page_frag_free(xdp.data);
 844 xdp_xmit:
 845         return NULL;
 846 }
 847
 848 static int veth_xdp_rcv(struct veth_rq *rq, int budget,
 849                         struct veth_xdp_tx_bq *bq,
 850                         struct veth_stats *stats)
 851 {
 852         int i, done = 0, n_xdpf = 0;
 853         void *xdpf[VETH_XDP_BATCH];
 854
 855         for (i = 0; i < budget; i++) {
 856                 void *ptr = __ptr_ring_consume(&rq->xdp_ring);
 857
 858                 if (!ptr)
 859                         break;
 860
 861                 if (veth_is_xdp_frame(ptr)) {
 862                         /* ndo_xdp_xmit */
 863                         struct xdp_frame *frame = veth_ptr_to_xdp(ptr);
 864
 865                         stats->xdp_bytes += frame->len;
 866                         frame = veth_xdp_rcv_one(rq, frame, bq, stats);
 867                         if (frame) {
 868                                 /* XDP_PASS */
 869                                 xdpf[n_xdpf++] = frame;
 870                                 if (n_xdpf == VETH_XDP_BATCH) {
 871                                         veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf,
 872                                                               bq, stats);
 873                                         n_xdpf = 0;
 874                                 }
 875                         }
 876                 } else {
 877                         /* ndo_start_xmit */
 878                         struct sk_buff *skb = ptr;
 879
 880                         stats->xdp_bytes += skb->len;
 881                         skb = veth_xdp_rcv_skb(rq, skb, bq, stats);
 882                         if (skb)
 883                                 napi_gro_receive(&rq->xdp_napi, skb);
 884                 }
 885                 done++;
 886         }
 887
 888         if (n_xdpf)
 889                 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats);
 890
 891         u64_stats_update_begin(&rq->stats.syncp);
 892         rq->stats.vs.xdp_redirect += stats->xdp_redirect;
 893         rq->stats.vs.xdp_bytes += stats->xdp_bytes;
 894         rq->stats.vs.xdp_drops += stats->xdp_drops;
 895         rq->stats.vs.rx_drops += stats->rx_drops;
 896         rq->stats.vs.xdp_packets += done;
 897         u64_stats_update_end(&rq->stats.syncp);
 898
 899         return done;
 900 }
 901
 902 static int veth_poll(struct napi_struct *napi, int budget)
 903 {
 904         struct veth_rq *rq =
 905                 container_of(napi, struct veth_rq, xdp_napi);
 906         struct veth_stats stats = {};
 907         struct veth_xdp_tx_bq bq;
 908         int done;
 909
 910         bq.count = 0;
 911
 912         xdp_set_return_frame_no_direct();
 913         done = veth_xdp_rcv(rq, budget, &bq, &stats);
 914
 915         if (done < budget && napi_complete_done(napi, done)) {
 916                 /* Write rx_notify_masked before reading ptr_ring */
 917                 smp_store_mb(rq->rx_notify_masked, false);
 918                 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
 919                         rq->rx_notify_masked = true;
 920                         napi_schedule(&rq->xdp_napi);
 921                 }
 922         }
 923
 924         if (stats.xdp_tx > 0)
 925                 veth_xdp_flush(rq, &bq);
 926         if (stats.xdp_redirect > 0)
 927                 xdp_do_flush();
 928         xdp_clear_return_frame_no_direct();
 929
 930         return done;
 931 }
 932
 933 static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
 934 {
 935         struct veth_priv *priv = netdev_priv(dev);
 936         int err, i;
 937
 938         for (i = start; i < end; i++) {
 939                 struct veth_rq *rq = &priv->rq[i];
 940
 941                 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
 942                 if (err)
 943                         goto err_xdp_ring;
 944         }
 945
 946         for (i = start; i < end; i++) {
 947                 struct veth_rq *rq = &priv->rq[i];
 948
 949                 napi_enable(&rq->xdp_napi);
 950                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
 951         }
 952
 953         return 0;
 954
 955 err_xdp_ring:
 956         for (i--; i >= start; i--)
 957                 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
 958
 959         return err;
 960 }
 961
 962 static int __veth_napi_enable(struct net_device *dev)
 963 {
 964         return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
 965 }
 966
 967 static void veth_napi_del_range(struct net_device *dev, int start, int end)
 968 {
 969         struct veth_priv *priv = netdev_priv(dev);
 970         int i;
 971
 972         for (i = start; i < end; i++) {
 973                 struct veth_rq *rq = &priv->rq[i];
 974
 975                 rcu_assign_pointer(priv->rq[i].napi, NULL);
 976                 napi_disable(&rq->xdp_napi);
 977                 __netif_napi_del(&rq->xdp_napi);
 978         }
 979         synchronize_net();
 980
 981         for (i = start; i < end; i++) {
 982                 struct veth_rq *rq = &priv->rq[i];
 983
 984                 rq->rx_notify_masked = false;
 985                 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
 986         }
 987 }
 988
 989 static void veth_napi_del(struct net_device *dev)
 990 {
 991         veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
 992 }
 993
 994 static bool veth_gro_requested(const struct net_device *dev)
 995 {
 996         return !!(dev->wanted_features & NETIF_F_GRO);
 997 }
 998
 999 static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1000                                  bool napi_already_on)
1001 {
1002         struct veth_priv *priv = netdev_priv(dev);
1003         int err, i;
1004
1005         for (i = start; i < end; i++) {
1006                 struct veth_rq *rq = &priv->rq[i];
1007
1008                 if (!napi_already_on)
1009                         netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1010                 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1011                 if (err < 0)
1012                         goto err_rxq_reg;
1013
1014                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1015                                                  MEM_TYPE_PAGE_SHARED,
1016                                                  NULL);
1017                 if (err < 0)
1018                         goto err_reg_mem;
1019
1020                 /* Save original mem info as it can be overwritten */
1021                 rq->xdp_mem = rq->xdp_rxq.mem;
1022         }
1023         return 0;
1024
1025 err_reg_mem:
1026         xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1027 err_rxq_reg:
1028         for (i--; i >= start; i--) {
1029                 struct veth_rq *rq = &priv->rq[i];
1030
1031                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1032                 if (!napi_already_on)
1033                         netif_napi_del(&rq->xdp_napi);
1034         }
1035
1036         return err;
1037 }
1038
1039 static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1040                                    bool delete_napi)
1041 {
1042         struct veth_priv *priv = netdev_priv(dev);
1043         int i;
1044
1045         for (i = start; i < end; i++) {
1046                 struct veth_rq *rq = &priv->rq[i];
1047
1048                 rq->xdp_rxq.mem = rq->xdp_mem;
1049                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1050
1051                 if (delete_napi)
1052                         netif_napi_del(&rq->xdp_napi);
1053         }
1054 }
1055
1056 static int veth_enable_xdp(struct net_device *dev)
1057 {
1058         bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1059         struct veth_priv *priv = netdev_priv(dev);
1060         int err, i;
1061
1062         if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1063                 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1064                 if (err)
1065                         return err;
1066
1067                 if (!napi_already_on) {
1068                         err = __veth_napi_enable(dev);
1069                         if (err) {
1070                                 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1071                                 return err;
1072                         }
1073
1074                         if (!veth_gro_requested(dev)) {
1075                                 /* user-space did not require GRO, but adding XDP
1076                                  * is supposed to get GRO working
1077                                  */
1078                                 dev->features |= NETIF_F_GRO;
1079                                 netdev_features_change(dev);
1080                         }
1081                 }
1082         }
1083
1084         for (i = 0; i < dev->real_num_rx_queues; i++) {
1085                 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1086                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1087         }
1088
1089         return 0;
1090 }
1091
1092 static void veth_disable_xdp(struct net_device *dev)
1093 {
1094         struct veth_priv *priv = netdev_priv(dev);
1095         int i;
1096
1097         for (i = 0; i < dev->real_num_rx_queues; i++)
1098                 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1099
1100         if (!netif_running(dev) || !veth_gro_requested(dev)) {
1101                 veth_napi_del(dev);
1102
1103                 /* if user-space did not require GRO, since adding XDP
1104                  * enabled it, clear it now
1105                  */
1106                 if (!veth_gro_requested(dev) && netif_running(dev)) {
1107                         dev->features &= ~NETIF_F_GRO;
1108                         netdev_features_change(dev);
1109                 }
1110         }
1111
1112         veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1113 }
1114
1115 static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1116 {
1117         struct veth_priv *priv = netdev_priv(dev);
1118         int err, i;
1119
1120         for (i = start; i < end; i++) {
1121                 struct veth_rq *rq = &priv->rq[i];
1122
1123                 netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1124         }
1125
1126         err = __veth_napi_enable_range(dev, start, end);
1127         if (err) {
1128                 for (i = start; i < end; i++) {
1129                         struct veth_rq *rq = &priv->rq[i];
1130
1131                         netif_napi_del(&rq->xdp_napi);
1132                 }
1133                 return err;
1134         }
1135         return err;
1136 }
1137
1138 static int veth_napi_enable(struct net_device *dev)
1139 {
1140         return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1141 }
1142
1143 static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1144 {
1145         struct veth_priv *priv = netdev_priv(dev);
1146
1147         if (start >= end)
1148                 return;
1149
1150         if (priv->_xdp_prog) {
1151                 veth_napi_del_range(dev, start, end);
1152                 veth_disable_xdp_range(dev, start, end, false);
1153         } else if (veth_gro_requested(dev)) {
1154                 veth_napi_del_range(dev, start, end);
1155         }
1156 }
1157
1158 static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1159 {
1160         struct veth_priv *priv = netdev_priv(dev);
1161         int err;
1162
1163         if (start >= end)
1164                 return 0;
1165
1166         if (priv->_xdp_prog) {
1167                 /* these channels are freshly initialized, napi is not on there even
1168                  * when GRO is requeste
1169                  */
1170                 err = veth_enable_xdp_range(dev, start, end, false);
1171                 if (err)
1172                         return err;
1173
1174                 err = __veth_napi_enable_range(dev, start, end);
1175                 if (err) {
1176                         /* on error always delete the newly added napis */
1177                         veth_disable_xdp_range(dev, start, end, true);
1178                         return err;
1179                 }
1180         } else if (veth_gro_requested(dev)) {
1181                 return veth_napi_enable_range(dev, start, end);
1182         }
1183         return 0;
1184 }
1185
1186 static int veth_set_channels(struct net_device *dev,
1187                              struct ethtool_channels *ch)
1188 {
1189         struct veth_priv *priv = netdev_priv(dev);
1190         unsigned int old_rx_count, new_rx_count;
1191         struct veth_priv *peer_priv;
1192         struct net_device *peer;
1193         int err;
1194
1195         /* sanity check. Upper bounds are already enforced by the caller */
1196         if (!ch->rx_count || !ch->tx_count)
1197                 return -EINVAL;
1198
1199         /* avoid braking XDP, if that is enabled */
1200         peer = rtnl_dereference(priv->peer);
1201         peer_priv = peer ? netdev_priv(peer) : NULL;
1202         if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1203                 return -EINVAL;
1204
1205         if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1206                 return -EINVAL;
1207
1208         old_rx_count = dev->real_num_rx_queues;
1209         new_rx_count = ch->rx_count;
1210         if (netif_running(dev)) {
1211                 /* turn device off */
1212                 netif_carrier_off(dev);
1213                 if (peer)
1214                         netif_carrier_off(peer);
1215
1216                 /* try to allocate new resurces, as needed*/
1217                 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1218                 if (err)
1219                         goto out;
1220         }
1221
1222         err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1223         if (err)
1224                 goto revert;
1225
1226         err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1227         if (err) {
1228                 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1229
1230                 /* this error condition could happen only if rx and tx change
1231                  * in opposite directions (e.g. tx nr raises, rx nr decreases)
1232                  * and we can't do anything to fully restore the original
1233                  * status
1234                  */
1235                 if (err2)
1236                         pr_warn("Can't restore rx queues config %d -> %d %d",
1237                                 new_rx_count, old_rx_count, err2);
1238                 else
1239                         goto revert;
1240         }
1241
1242 out:
1243         if (netif_running(dev)) {
1244                 /* note that we need to swap the arguments WRT the enable part
1245                  * to identify the range we have to disable
1246                  */
1247                 veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1248                 netif_carrier_on(dev);
1249                 if (peer)
1250                         netif_carrier_on(peer);
1251         }
1252         return err;
1253
1254 revert:
1255         new_rx_count = old_rx_count;
1256         old_rx_count = ch->rx_count;
1257         goto out;
1258 }
1259
1260 static int veth_open(struct net_device *dev)
1261 {
1262         struct veth_priv *priv = netdev_priv(dev);
1263         struct net_device *peer = rtnl_dereference(priv->peer);
1264         int err;
1265
1266         if (!peer)
1267                 return -ENOTCONN;
1268
1269         if (priv->_xdp_prog) {
1270                 err = veth_enable_xdp(dev);
1271                 if (err)
1272                         return err;
1273         } else if (veth_gro_requested(dev)) {
1274                 err = veth_napi_enable(dev);
1275                 if (err)
1276                         return err;
1277         }
1278
1279         if (peer->flags & IFF_UP) {
1280                 netif_carrier_on(dev);
1281                 netif_carrier_on(peer);
1282         }
1283
1284         return 0;
1285 }
1286
1287 static int veth_close(struct net_device *dev)
1288 {
1289         struct veth_priv *priv = netdev_priv(dev);
1290         struct net_device *peer = rtnl_dereference(priv->peer);
1291
1292         netif_carrier_off(dev);
1293         if (peer)
1294                 netif_carrier_off(peer);
1295
1296         if (priv->_xdp_prog)
1297                 veth_disable_xdp(dev);
1298         else if (veth_gro_requested(dev))
1299                 veth_napi_del(dev);
1300
1301         return 0;
1302 }
1303
1304 static int is_valid_veth_mtu(int mtu)
1305 {
1306         return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1307 }
1308
1309 static int veth_alloc_queues(struct net_device *dev)
1310 {
1311         struct veth_priv *priv = netdev_priv(dev);
1312         int i;
1313
1314         priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL);
1315         if (!priv->rq)
1316                 return -ENOMEM;
1317
1318         for (i = 0; i < dev->num_rx_queues; i++) {
1319                 priv->rq[i].dev = dev;
1320                 u64_stats_init(&priv->rq[i].stats.syncp);
1321         }
1322
1323         return 0;
1324 }
1325
1326 static void veth_free_queues(struct net_device *dev)
1327 {
1328         struct veth_priv *priv = netdev_priv(dev);
1329
1330         kfree(priv->rq);
1331 }
1332
1333 static int veth_dev_init(struct net_device *dev)
1334 {
1335         int err;
1336
1337         dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
1338         if (!dev->lstats)
1339                 return -ENOMEM;
1340
1341         err = veth_alloc_queues(dev);
1342         if (err) {
1343                 free_percpu(dev->lstats);
1344                 return err;
1345         }
1346
1347         return 0;
1348 }
1349
1350 static void veth_dev_free(struct net_device *dev)
1351 {
1352         veth_free_queues(dev);
1353         free_percpu(dev->lstats);
1354 }
1355
1356 #ifdef CONFIG_NET_POLL_CONTROLLER
1357 static void veth_poll_controller(struct net_device *dev)
1358 {
1359         /* veth only receives frames when its peer sends one
1360          * Since it has nothing to do with disabling irqs, we are guaranteed
1361          * never to have pending data when we poll for it so
1362          * there is nothing to do here.
1363          *
1364          * We need this though so netpoll recognizes us as an interface that
1365          * supports polling, which enables bridge devices in virt setups to
1366          * still use netconsole
1367          */
1368 }
1369 #endif  /* CONFIG_NET_POLL_CONTROLLER */
1370
1371 static int veth_get_iflink(const struct net_device *dev)
1372 {
1373         struct veth_priv *priv = netdev_priv(dev);
1374         struct net_device *peer;
1375         int iflink;
1376
1377         rcu_read_lock();
1378         peer = rcu_dereference(priv->peer);
1379         iflink = peer ? peer->ifindex : 0;
1380         rcu_read_unlock();
1381
1382         return iflink;
1383 }
1384
1385 static netdev_features_t veth_fix_features(struct net_device *dev,
1386                                            netdev_features_t features)
1387 {
1388         struct veth_priv *priv = netdev_priv(dev);
1389         struct net_device *peer;
1390
1391         peer = rtnl_dereference(priv->peer);
1392         if (peer) {
1393                 struct veth_priv *peer_priv = netdev_priv(peer);
1394
1395                 if (peer_priv->_xdp_prog)
1396                         features &= ~NETIF_F_GSO_SOFTWARE;
1397         }
1398         if (priv->_xdp_prog)
1399                 features |= NETIF_F_GRO;
1400
1401         return features;
1402 }
1403
1404 static int veth_set_features(struct net_device *dev,
1405                              netdev_features_t features)
1406 {
1407         netdev_features_t changed = features ^ dev->features;
1408         struct veth_priv *priv = netdev_priv(dev);
1409         int err;
1410
1411         if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1412                 return 0;
1413
1414         if (features & NETIF_F_GRO) {
1415                 err = veth_napi_enable(dev);
1416                 if (err)
1417                         return err;
1418         } else {
1419                 veth_napi_del(dev);
1420         }
1421         return 0;
1422 }
1423
1424 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1425 {
1426         struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1427         struct net_device *peer;
1428
1429         if (new_hr < 0)
1430                 new_hr = 0;
1431
1432         rcu_read_lock();
1433         peer = rcu_dereference(priv->peer);
1434         if (unlikely(!peer))
1435                 goto out;
1436
1437         peer_priv = netdev_priv(peer);
1438         priv->requested_headroom = new_hr;
1439         new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1440         dev->needed_headroom = new_hr;
1441         peer->needed_headroom = new_hr;
1442
1443 out:
1444         rcu_read_unlock();
1445 }
1446
1447 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1448                         struct netlink_ext_ack *extack)
1449 {
1450         struct veth_priv *priv = netdev_priv(dev);
1451         struct bpf_prog *old_prog;
1452         struct net_device *peer;
1453         unsigned int max_mtu;
1454         int err;
1455
1456         old_prog = priv->_xdp_prog;
1457         priv->_xdp_prog = prog;
1458         peer = rtnl_dereference(priv->peer);
1459
1460         if (prog) {
1461                 if (!peer) {
1462                         NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1463                         err = -ENOTCONN;
1464                         goto err;
1465                 }
1466
1467                 max_mtu = PAGE_SIZE - VETH_XDP_HEADROOM -
1468                           peer->hard_header_len -
1469                           SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1470                 if (peer->mtu > max_mtu) {
1471                         NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1472                         err = -ERANGE;
1473                         goto err;
1474                 }
1475
1476                 if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1477                         NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1478                         err = -ENOSPC;
1479                         goto err;
1480                 }
1481
1482                 if (dev->flags & IFF_UP) {
1483                         err = veth_enable_xdp(dev);
1484                         if (err) {
1485                                 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1486                                 goto err;
1487                         }
1488                 }
1489
1490                 if (!old_prog) {
1491                         peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1492                         peer->max_mtu = max_mtu;
1493                 }
1494         }
1495
1496         if (old_prog) {
1497                 if (!prog) {
1498                         if (dev->flags & IFF_UP)
1499                                 veth_disable_xdp(dev);
1500
1501                         if (peer) {
1502                                 peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1503                                 peer->max_mtu = ETH_MAX_MTU;
1504                         }
1505                 }
1506                 bpf_prog_put(old_prog);
1507         }
1508
1509         if ((!!old_prog ^ !!prog) && peer)
1510                 netdev_update_features(peer);
1511
1512         return 0;
1513 err:
1514         priv->_xdp_prog = old_prog;
1515
1516         return err;
1517 }
1518
1519 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1520 {
1521         switch (xdp->command) {
1522         case XDP_SETUP_PROG:
1523                 return veth_xdp_set(dev, xdp->prog, xdp->extack);
1524         default:
1525                 return -EINVAL;
1526         }
1527 }
1528
1529 static const struct net_device_ops veth_netdev_ops = {
1530         .ndo_init            = veth_dev_init,
1531         .ndo_open            = veth_open,
1532         .ndo_stop            = veth_close,
1533         .ndo_start_xmit      = veth_xmit,
1534         .ndo_get_stats64     = veth_get_stats64,
1535         .ndo_set_rx_mode     = veth_set_multicast_list,
1536         .ndo_set_mac_address = eth_mac_addr,
1537 #ifdef CONFIG_NET_POLL_CONTROLLER
1538         .ndo_poll_controller    = veth_poll_controller,
1539 #endif
1540         .ndo_get_iflink         = veth_get_iflink,
1541         .ndo_fix_features       = veth_fix_features,
1542         .ndo_set_features       = veth_set_features,
1543         .ndo_features_check     = passthru_features_check,
1544         .ndo_set_rx_headroom    = veth_set_rx_headroom,
1545         .ndo_bpf                = veth_xdp,
1546         .ndo_xdp_xmit           = veth_ndo_xdp_xmit,
1547         .ndo_get_peer_dev       = veth_peer_dev,
1548 };
1549
1550 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1551                        NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1552                        NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1553                        NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1554                        NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1555
1556 static void veth_setup(struct net_device *dev)
1557 {
1558         ether_setup(dev);
1559
1560         dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1561         dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1562         dev->priv_flags |= IFF_NO_QUEUE;
1563         dev->priv_flags |= IFF_PHONY_HEADROOM;
1564
1565         dev->netdev_ops = &veth_netdev_ops;
1566         dev->ethtool_ops = &veth_ethtool_ops;
1567         dev->features |= NETIF_F_LLTX;
1568         dev->features |= VETH_FEATURES;
1569         dev->vlan_features = dev->features &
1570                              ~(NETIF_F_HW_VLAN_CTAG_TX |
1571                                NETIF_F_HW_VLAN_STAG_TX |
1572                                NETIF_F_HW_VLAN_CTAG_RX |
1573                                NETIF_F_HW_VLAN_STAG_RX);
1574         dev->needs_free_netdev = true;
1575         dev->priv_destructor = veth_dev_free;
1576         dev->max_mtu = ETH_MAX_MTU;
1577
1578         dev->hw_features = VETH_FEATURES;
1579         dev->hw_enc_features = VETH_FEATURES;
1580         dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1581 }
1582
1583 /*
1584  * netlink interface
1585  */
1586
1587 static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1588                          struct netlink_ext_ack *extack)
1589 {
1590         if (tb[IFLA_ADDRESS]) {
1591                 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1592                         return -EINVAL;
1593                 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1594                         return -EADDRNOTAVAIL;
1595         }
1596         if (tb[IFLA_MTU]) {
1597                 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1598                         return -EINVAL;
1599         }
1600         return 0;
1601 }
1602
1603 static struct rtnl_link_ops veth_link_ops;
1604
1605 static void veth_disable_gro(struct net_device *dev)
1606 {
1607         dev->features &= ~NETIF_F_GRO;
1608         dev->wanted_features &= ~NETIF_F_GRO;
1609         netdev_update_features(dev);
1610 }
1611
1612 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1613 {
1614         int err;
1615
1616         if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1617                 err = netif_set_real_num_tx_queues(dev, 1);
1618                 if (err)
1619                         return err;
1620         }
1621         if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1622                 err = netif_set_real_num_rx_queues(dev, 1);
1623                 if (err)
1624                         return err;
1625         }
1626         return 0;
1627 }
1628
1629 static int veth_newlink(struct net *src_net, struct net_device *dev,
1630                         struct nlattr *tb[], struct nlattr *data[],
1631                         struct netlink_ext_ack *extack)
1632 {
1633         int err;
1634         struct net_device *peer;
1635         struct veth_priv *priv;
1636         char ifname[IFNAMSIZ];
1637         struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1638         unsigned char name_assign_type;
1639         struct ifinfomsg *ifmp;
1640         struct net *net;
1641
1642         /*
1643          * create and register peer first
1644          */
1645         if (data != NULL && data[VETH_INFO_PEER] != NULL) {
1646                 struct nlattr *nla_peer;
1647
1648                 nla_peer = data[VETH_INFO_PEER];
1649                 ifmp = nla_data(nla_peer);
1650                 err = rtnl_nla_parse_ifla(peer_tb,
1651                                           nla_data(nla_peer) + sizeof(struct ifinfomsg),
1652                                           nla_len(nla_peer) - sizeof(struct ifinfomsg),
1653                                           NULL);
1654                 if (err < 0)
1655                         return err;
1656
1657                 err = veth_validate(peer_tb, NULL, extack);
1658                 if (err < 0)
1659                         return err;
1660
1661                 tbp = peer_tb;
1662         } else {
1663                 ifmp = NULL;
1664                 tbp = tb;
1665         }
1666
1667         if (ifmp && tbp[IFLA_IFNAME]) {
1668                 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1669                 name_assign_type = NET_NAME_USER;
1670         } else {
1671                 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1672                 name_assign_type = NET_NAME_ENUM;
1673         }
1674
1675         net = rtnl_link_get_net(src_net, tbp);
1676         if (IS_ERR(net))
1677                 return PTR_ERR(net);
1678
1679         peer = rtnl_create_link(net, ifname, name_assign_type,
1680                                 &veth_link_ops, tbp, extack);
1681         if (IS_ERR(peer)) {
1682                 put_net(net);
1683                 return PTR_ERR(peer);
1684         }
1685
1686         if (!ifmp || !tbp[IFLA_ADDRESS])
1687                 eth_hw_addr_random(peer);
1688
1689         if (ifmp && (dev->ifindex != 0))
1690                 peer->ifindex = ifmp->ifi_index;
1691
1692         peer->gso_max_size = dev->gso_max_size;
1693         peer->gso_max_segs = dev->gso_max_segs;
1694
1695         err = register_netdevice(peer);
1696         put_net(net);
1697         net = NULL;
1698         if (err < 0)
1699                 goto err_register_peer;
1700
1701         /* keep GRO disabled by default to be consistent with the established
1702          * veth behavior
1703          */
1704         veth_disable_gro(peer);
1705         netif_carrier_off(peer);
1706
1707         err = rtnl_configure_link(peer, ifmp);
1708         if (err < 0)
1709                 goto err_configure_peer;
1710
1711         /*
1712          * register dev last
1713          *
1714          * note, that since we've registered new device the dev's name
1715          * should be re-allocated
1716          */
1717
1718         if (tb[IFLA_ADDRESS] == NULL)
1719                 eth_hw_addr_random(dev);
1720
1721         if (tb[IFLA_IFNAME])
1722                 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1723         else
1724                 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1725
1726         err = register_netdevice(dev);
1727         if (err < 0)
1728                 goto err_register_dev;
1729
1730         netif_carrier_off(dev);
1731
1732         /*
1733          * tie the deviced together
1734          */
1735
1736         priv = netdev_priv(dev);
1737         rcu_assign_pointer(priv->peer, peer);
1738         err = veth_init_queues(dev, tb);
1739         if (err)
1740                 goto err_queues;
1741
1742         priv = netdev_priv(peer);
1743         rcu_assign_pointer(priv->peer, dev);
1744         err = veth_init_queues(peer, tb);
1745         if (err)
1746                 goto err_queues;
1747
1748         veth_disable_gro(dev);
1749         return 0;
1750
1751 err_queues:
1752         unregister_netdevice(dev);
1753 err_register_dev:
1754         /* nothing to do */
1755 err_configure_peer:
1756         unregister_netdevice(peer);
1757         return err;
1758
1759 err_register_peer:
1760         free_netdev(peer);
1761         return err;
1762 }
1763
1764 static void veth_dellink(struct net_device *dev, struct list_head *head)
1765 {
1766         struct veth_priv *priv;
1767         struct net_device *peer;
1768
1769         priv = netdev_priv(dev);
1770         peer = rtnl_dereference(priv->peer);
1771
1772         /* Note : dellink() is called from default_device_exit_batch(),
1773          * before a rcu_synchronize() point. The devices are guaranteed
1774          * not being freed before one RCU grace period.
1775          */
1776         RCU_INIT_POINTER(priv->peer, NULL);
1777         unregister_netdevice_queue(dev, head);
1778
1779         if (peer) {
1780                 priv = netdev_priv(peer);
1781                 RCU_INIT_POINTER(priv->peer, NULL);
1782                 unregister_netdevice_queue(peer, head);
1783         }
1784 }
1785
1786 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
1787         [VETH_INFO_PEER]        = { .len = sizeof(struct ifinfomsg) },
1788 };
1789
1790 static struct net *veth_get_link_net(const struct net_device *dev)
1791 {
1792         struct veth_priv *priv = netdev_priv(dev);
1793         struct net_device *peer = rtnl_dereference(priv->peer);
1794
1795         return peer ? dev_net(peer) : dev_net(dev);
1796 }
1797
1798 static unsigned int veth_get_num_queues(void)
1799 {
1800         /* enforce the same queue limit as rtnl_create_link */
1801         int queues = num_possible_cpus();
1802
1803         if (queues > 4096)
1804                 queues = 4096;
1805         return queues;
1806 }
1807
1808 static struct rtnl_link_ops veth_link_ops = {
1809         .kind           = DRV_NAME,
1810         .priv_size      = sizeof(struct veth_priv),
1811         .setup          = veth_setup,
1812         .validate       = veth_validate,
1813         .newlink        = veth_newlink,
1814         .dellink        = veth_dellink,
1815         .policy         = veth_policy,
1816         .maxtype        = VETH_INFO_MAX,
1817         .get_link_net   = veth_get_link_net,
1818         .get_num_tx_queues      = veth_get_num_queues,
1819         .get_num_rx_queues      = veth_get_num_queues,
1820 };
1821
1822 /*
1823  * init/fini
1824  */
1825
1826 static __init int veth_init(void)
1827 {
1828         return rtnl_link_register(&veth_link_ops);
1829 }
1830
1831 static __exit void veth_exit(void)
1832 {
1833         rtnl_link_unregister(&veth_link_ops);
1834 }
1835
1836 module_init(veth_init);
1837 module_exit(veth_exit);
1838
1839 MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
1840 MODULE_LICENSE("GPL v2");
1841 MODULE_ALIAS_RTNL_LINK(DRV_NAME);