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 <xemul@openvz.org>
   8  * Ethtool interface from: Eric W. Biederman <ebiederm@xmission.com>
   9  *
  10  */
  11
  12 #include <linux/netdevice.h>
  13 #include <linux/slab.h>
  14 #include <linux/ethtool.h>
  15 #include <linux/etherdevice.h>
  16 #include <linux/u64_stats_sync.h>
  17
  18 #include <net/rtnetlink.h>
  19 #include <net/dst.h>
  20 #include <net/xfrm.h>
  21 #include <net/xdp.h>
  22 #include <linux/veth.h>
  23 #include <linux/module.h>
  24 #include <linux/bpf.h>
  25 #include <linux/filter.h>
  26 #include <linux/ptr_ring.h>
  27 #include <linux/bpf_trace.h>
  28 #include <linux/net_tstamp.h>
  29
  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                                 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC))
 884                                         netif_receive_skb(skb);
 885                                 else
 886                                         napi_gro_receive(&rq->xdp_napi, skb);
 887                         }
 888                 }
 889                 done++;
 890         }
 891
 892         if (n_xdpf)
 893                 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats);
 894
 895         u64_stats_update_begin(&rq->stats.syncp);
 896         rq->stats.vs.xdp_redirect += stats->xdp_redirect;
 897         rq->stats.vs.xdp_bytes += stats->xdp_bytes;
 898         rq->stats.vs.xdp_drops += stats->xdp_drops;
 899         rq->stats.vs.rx_drops += stats->rx_drops;
 900         rq->stats.vs.xdp_packets += done;
 901         u64_stats_update_end(&rq->stats.syncp);
 902
 903         return done;
 904 }
 905
 906 static int veth_poll(struct napi_struct *napi, int budget)
 907 {
 908         struct veth_rq *rq =
 909                 container_of(napi, struct veth_rq, xdp_napi);
 910         struct veth_stats stats = {};
 911         struct veth_xdp_tx_bq bq;
 912         int done;
 913
 914         bq.count = 0;
 915
 916         xdp_set_return_frame_no_direct();
 917         done = veth_xdp_rcv(rq, budget, &bq, &stats);
 918
 919         if (done < budget && napi_complete_done(napi, done)) {
 920                 /* Write rx_notify_masked before reading ptr_ring */
 921                 smp_store_mb(rq->rx_notify_masked, false);
 922                 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
 923                         rq->rx_notify_masked = true;
 924                         napi_schedule(&rq->xdp_napi);
 925                 }
 926         }
 927
 928         if (stats.xdp_tx > 0)
 929                 veth_xdp_flush(rq, &bq);
 930         if (stats.xdp_redirect > 0)
 931                 xdp_do_flush();
 932         xdp_clear_return_frame_no_direct();
 933
 934         return done;
 935 }
 936
 937 static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
 938 {
 939         struct veth_priv *priv = netdev_priv(dev);
 940         int err, i;
 941
 942         for (i = start; i < end; i++) {
 943                 struct veth_rq *rq = &priv->rq[i];
 944
 945                 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
 946                 if (err)
 947                         goto err_xdp_ring;
 948         }
 949
 950         for (i = start; i < end; i++) {
 951                 struct veth_rq *rq = &priv->rq[i];
 952
 953                 napi_enable(&rq->xdp_napi);
 954                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
 955         }
 956
 957         return 0;
 958
 959 err_xdp_ring:
 960         for (i--; i >= start; i--)
 961                 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
 962
 963         return err;
 964 }
 965
 966 static int __veth_napi_enable(struct net_device *dev)
 967 {
 968         return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
 969 }
 970
 971 static void veth_napi_del_range(struct net_device *dev, int start, int end)
 972 {
 973         struct veth_priv *priv = netdev_priv(dev);
 974         int i;
 975
 976         for (i = start; i < end; i++) {
 977                 struct veth_rq *rq = &priv->rq[i];
 978
 979                 rcu_assign_pointer(priv->rq[i].napi, NULL);
 980                 napi_disable(&rq->xdp_napi);
 981                 __netif_napi_del(&rq->xdp_napi);
 982         }
 983         synchronize_net();
 984
 985         for (i = start; i < end; i++) {
 986                 struct veth_rq *rq = &priv->rq[i];
 987
 988                 rq->rx_notify_masked = false;
 989                 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
 990         }
 991 }
 992
 993 static void veth_napi_del(struct net_device *dev)
 994 {
 995         veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
 996 }
 997
 998 static bool veth_gro_requested(const struct net_device *dev)
 999 {
1000         return !!(dev->wanted_features & NETIF_F_GRO);
1001 }
1002
1003 static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1004                                  bool napi_already_on)
1005 {
1006         struct veth_priv *priv = netdev_priv(dev);
1007         int err, i;
1008
1009         for (i = start; i < end; i++) {
1010                 struct veth_rq *rq = &priv->rq[i];
1011
1012                 if (!napi_already_on)
1013                         netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1014                 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1015                 if (err < 0)
1016                         goto err_rxq_reg;
1017
1018                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1019                                                  MEM_TYPE_PAGE_SHARED,
1020                                                  NULL);
1021                 if (err < 0)
1022                         goto err_reg_mem;
1023
1024                 /* Save original mem info as it can be overwritten */
1025                 rq->xdp_mem = rq->xdp_rxq.mem;
1026         }
1027         return 0;
1028
1029 err_reg_mem:
1030         xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1031 err_rxq_reg:
1032         for (i--; i >= start; i--) {
1033                 struct veth_rq *rq = &priv->rq[i];
1034
1035                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1036                 if (!napi_already_on)
1037                         netif_napi_del(&rq->xdp_napi);
1038         }
1039
1040         return err;
1041 }
1042
1043 static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1044                                    bool delete_napi)
1045 {
1046         struct veth_priv *priv = netdev_priv(dev);
1047         int i;
1048
1049         for (i = start; i < end; i++) {
1050                 struct veth_rq *rq = &priv->rq[i];
1051
1052                 rq->xdp_rxq.mem = rq->xdp_mem;
1053                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1054
1055                 if (delete_napi)
1056                         netif_napi_del(&rq->xdp_napi);
1057         }
1058 }
1059
1060 static int veth_enable_xdp(struct net_device *dev)
1061 {
1062         bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1063         struct veth_priv *priv = netdev_priv(dev);
1064         int err, i;
1065
1066         if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1067                 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1068                 if (err)
1069                         return err;
1070
1071                 if (!napi_already_on) {
1072                         err = __veth_napi_enable(dev);
1073                         if (err) {
1074                                 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1075                                 return err;
1076                         }
1077
1078                         if (!veth_gro_requested(dev)) {
1079                                 /* user-space did not require GRO, but adding XDP
1080                                  * is supposed to get GRO working
1081                                  */
1082                                 dev->features |= NETIF_F_GRO;
1083                                 netdev_features_change(dev);
1084                         }
1085                 }
1086         }
1087
1088         for (i = 0; i < dev->real_num_rx_queues; i++) {
1089                 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1090                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1091         }
1092
1093         return 0;
1094 }
1095
1096 static void veth_disable_xdp(struct net_device *dev)
1097 {
1098         struct veth_priv *priv = netdev_priv(dev);
1099         int i;
1100
1101         for (i = 0; i < dev->real_num_rx_queues; i++)
1102                 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1103
1104         if (!netif_running(dev) || !veth_gro_requested(dev)) {
1105                 veth_napi_del(dev);
1106
1107                 /* if user-space did not require GRO, since adding XDP
1108                  * enabled it, clear it now
1109                  */
1110                 if (!veth_gro_requested(dev) && netif_running(dev)) {
1111                         dev->features &= ~NETIF_F_GRO;
1112                         netdev_features_change(dev);
1113                 }
1114         }
1115
1116         veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1117 }
1118
1119 static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1120 {
1121         struct veth_priv *priv = netdev_priv(dev);
1122         int err, i;
1123
1124         for (i = start; i < end; i++) {
1125                 struct veth_rq *rq = &priv->rq[i];
1126
1127                 netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1128         }
1129
1130         err = __veth_napi_enable_range(dev, start, end);
1131         if (err) {
1132                 for (i = start; i < end; i++) {
1133                         struct veth_rq *rq = &priv->rq[i];
1134
1135                         netif_napi_del(&rq->xdp_napi);
1136                 }
1137                 return err;
1138         }
1139         return err;
1140 }
1141
1142 static int veth_napi_enable(struct net_device *dev)
1143 {
1144         return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1145 }
1146
1147 static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1148 {
1149         struct veth_priv *priv = netdev_priv(dev);
1150
1151         if (start >= end)
1152                 return;
1153
1154         if (priv->_xdp_prog) {
1155                 veth_napi_del_range(dev, start, end);
1156                 veth_disable_xdp_range(dev, start, end, false);
1157         } else if (veth_gro_requested(dev)) {
1158                 veth_napi_del_range(dev, start, end);
1159         }
1160 }
1161
1162 static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1163 {
1164         struct veth_priv *priv = netdev_priv(dev);
1165         int err;
1166
1167         if (start >= end)
1168                 return 0;
1169
1170         if (priv->_xdp_prog) {
1171                 /* these channels are freshly initialized, napi is not on there even
1172                  * when GRO is requeste
1173                  */
1174                 err = veth_enable_xdp_range(dev, start, end, false);
1175                 if (err)
1176                         return err;
1177
1178                 err = __veth_napi_enable_range(dev, start, end);
1179                 if (err) {
1180                         /* on error always delete the newly added napis */
1181                         veth_disable_xdp_range(dev, start, end, true);
1182                         return err;
1183                 }
1184         } else if (veth_gro_requested(dev)) {
1185                 return veth_napi_enable_range(dev, start, end);
1186         }
1187         return 0;
1188 }
1189
1190 static int veth_set_channels(struct net_device *dev,
1191                              struct ethtool_channels *ch)
1192 {
1193         struct veth_priv *priv = netdev_priv(dev);
1194         unsigned int old_rx_count, new_rx_count;
1195         struct veth_priv *peer_priv;
1196         struct net_device *peer;
1197         int err;
1198
1199         /* sanity check. Upper bounds are already enforced by the caller */
1200         if (!ch->rx_count || !ch->tx_count)
1201                 return -EINVAL;
1202
1203         /* avoid braking XDP, if that is enabled */
1204         peer = rtnl_dereference(priv->peer);
1205         peer_priv = peer ? netdev_priv(peer) : NULL;
1206         if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1207                 return -EINVAL;
1208
1209         if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1210                 return -EINVAL;
1211
1212         old_rx_count = dev->real_num_rx_queues;
1213         new_rx_count = ch->rx_count;
1214         if (netif_running(dev)) {
1215                 /* turn device off */
1216                 netif_carrier_off(dev);
1217                 if (peer)
1218                         netif_carrier_off(peer);
1219
1220                 /* try to allocate new resurces, as needed*/
1221                 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1222                 if (err)
1223                         goto out;
1224         }
1225
1226         err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1227         if (err)
1228                 goto revert;
1229
1230         err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1231         if (err) {
1232                 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1233
1234                 /* this error condition could happen only if rx and tx change
1235                  * in opposite directions (e.g. tx nr raises, rx nr decreases)
1236                  * and we can't do anything to fully restore the original
1237                  * status
1238                  */
1239                 if (err2)
1240                         pr_warn("Can't restore rx queues config %d -> %d %d",
1241                                 new_rx_count, old_rx_count, err2);
1242                 else
1243                         goto revert;
1244         }
1245
1246 out:
1247         if (netif_running(dev)) {
1248                 /* note that we need to swap the arguments WRT the enable part
1249                  * to identify the range we have to disable
1250                  */
1251                 veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1252                 netif_carrier_on(dev);
1253                 if (peer)
1254                         netif_carrier_on(peer);
1255         }
1256         return err;
1257
1258 revert:
1259         new_rx_count = old_rx_count;
1260         old_rx_count = ch->rx_count;
1261         goto out;
1262 }
1263
1264 static int veth_open(struct net_device *dev)
1265 {
1266         struct veth_priv *priv = netdev_priv(dev);
1267         struct net_device *peer = rtnl_dereference(priv->peer);
1268         int err;
1269
1270         if (!peer)
1271                 return -ENOTCONN;
1272
1273         if (priv->_xdp_prog) {
1274                 err = veth_enable_xdp(dev);
1275                 if (err)
1276                         return err;
1277         } else if (veth_gro_requested(dev)) {
1278                 err = veth_napi_enable(dev);
1279                 if (err)
1280                         return err;
1281         }
1282
1283         if (peer->flags & IFF_UP) {
1284                 netif_carrier_on(dev);
1285                 netif_carrier_on(peer);
1286         }
1287
1288         return 0;
1289 }
1290
1291 static int veth_close(struct net_device *dev)
1292 {
1293         struct veth_priv *priv = netdev_priv(dev);
1294         struct net_device *peer = rtnl_dereference(priv->peer);
1295
1296         netif_carrier_off(dev);
1297         if (peer)
1298                 netif_carrier_off(peer);
1299
1300         if (priv->_xdp_prog)
1301                 veth_disable_xdp(dev);
1302         else if (veth_gro_requested(dev))
1303                 veth_napi_del(dev);
1304
1305         return 0;
1306 }
1307
1308 static int is_valid_veth_mtu(int mtu)
1309 {
1310         return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1311 }
1312
1313 static int veth_alloc_queues(struct net_device *dev)
1314 {
1315         struct veth_priv *priv = netdev_priv(dev);
1316         int i;
1317
1318         priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL);
1319         if (!priv->rq)
1320                 return -ENOMEM;
1321
1322         for (i = 0; i < dev->num_rx_queues; i++) {
1323                 priv->rq[i].dev = dev;
1324                 u64_stats_init(&priv->rq[i].stats.syncp);
1325         }
1326
1327         return 0;
1328 }
1329
1330 static void veth_free_queues(struct net_device *dev)
1331 {
1332         struct veth_priv *priv = netdev_priv(dev);
1333
1334         kfree(priv->rq);
1335 }
1336
1337 static int veth_dev_init(struct net_device *dev)
1338 {
1339         int err;
1340
1341         dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
1342         if (!dev->lstats)
1343                 return -ENOMEM;
1344
1345         err = veth_alloc_queues(dev);
1346         if (err) {
1347                 free_percpu(dev->lstats);
1348                 return err;
1349         }
1350
1351         return 0;
1352 }
1353
1354 static void veth_dev_free(struct net_device *dev)
1355 {
1356         veth_free_queues(dev);
1357         free_percpu(dev->lstats);
1358 }
1359
1360 #ifdef CONFIG_NET_POLL_CONTROLLER
1361 static void veth_poll_controller(struct net_device *dev)
1362 {
1363         /* veth only receives frames when its peer sends one
1364          * Since it has nothing to do with disabling irqs, we are guaranteed
1365          * never to have pending data when we poll for it so
1366          * there is nothing to do here.
1367          *
1368          * We need this though so netpoll recognizes us as an interface that
1369          * supports polling, which enables bridge devices in virt setups to
1370          * still use netconsole
1371          */
1372 }
1373 #endif  /* CONFIG_NET_POLL_CONTROLLER */
1374
1375 static int veth_get_iflink(const struct net_device *dev)
1376 {
1377         struct veth_priv *priv = netdev_priv(dev);
1378         struct net_device *peer;
1379         int iflink;
1380
1381         rcu_read_lock();
1382         peer = rcu_dereference(priv->peer);
1383         iflink = peer ? peer->ifindex : 0;
1384         rcu_read_unlock();
1385
1386         return iflink;
1387 }
1388
1389 static netdev_features_t veth_fix_features(struct net_device *dev,
1390                                            netdev_features_t features)
1391 {
1392         struct veth_priv *priv = netdev_priv(dev);
1393         struct net_device *peer;
1394
1395         peer = rtnl_dereference(priv->peer);
1396         if (peer) {
1397                 struct veth_priv *peer_priv = netdev_priv(peer);
1398
1399                 if (peer_priv->_xdp_prog)
1400                         features &= ~NETIF_F_GSO_SOFTWARE;
1401         }
1402         if (priv->_xdp_prog)
1403                 features |= NETIF_F_GRO;
1404
1405         return features;
1406 }
1407
1408 static int veth_set_features(struct net_device *dev,
1409                              netdev_features_t features)
1410 {
1411         netdev_features_t changed = features ^ dev->features;
1412         struct veth_priv *priv = netdev_priv(dev);
1413         int err;
1414
1415         if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1416                 return 0;
1417
1418         if (features & NETIF_F_GRO) {
1419                 err = veth_napi_enable(dev);
1420                 if (err)
1421                         return err;
1422         } else {
1423                 veth_napi_del(dev);
1424         }
1425         return 0;
1426 }
1427
1428 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1429 {
1430         struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1431         struct net_device *peer;
1432
1433         if (new_hr < 0)
1434                 new_hr = 0;
1435
1436         rcu_read_lock();
1437         peer = rcu_dereference(priv->peer);
1438         if (unlikely(!peer))
1439                 goto out;
1440
1441         peer_priv = netdev_priv(peer);
1442         priv->requested_headroom = new_hr;
1443         new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1444         dev->needed_headroom = new_hr;
1445         peer->needed_headroom = new_hr;
1446
1447 out:
1448         rcu_read_unlock();
1449 }
1450
1451 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1452                         struct netlink_ext_ack *extack)
1453 {
1454         struct veth_priv *priv = netdev_priv(dev);
1455         struct bpf_prog *old_prog;
1456         struct net_device *peer;
1457         unsigned int max_mtu;
1458         int err;
1459
1460         old_prog = priv->_xdp_prog;
1461         priv->_xdp_prog = prog;
1462         peer = rtnl_dereference(priv->peer);
1463
1464         if (prog) {
1465                 if (!peer) {
1466                         NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1467                         err = -ENOTCONN;
1468                         goto err;
1469                 }
1470
1471                 max_mtu = PAGE_SIZE - VETH_XDP_HEADROOM -
1472                           peer->hard_header_len -
1473                           SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1474                 if (peer->mtu > max_mtu) {
1475                         NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1476                         err = -ERANGE;
1477                         goto err;
1478                 }
1479
1480                 if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1481                         NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1482                         err = -ENOSPC;
1483                         goto err;
1484                 }
1485
1486                 if (dev->flags & IFF_UP) {
1487                         err = veth_enable_xdp(dev);
1488                         if (err) {
1489                                 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1490                                 goto err;
1491                         }
1492                 }
1493
1494                 if (!old_prog) {
1495                         peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1496                         peer->max_mtu = max_mtu;
1497                 }
1498         }
1499
1500         if (old_prog) {
1501                 if (!prog) {
1502                         if (dev->flags & IFF_UP)
1503                                 veth_disable_xdp(dev);
1504
1505                         if (peer) {
1506                                 peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1507                                 peer->max_mtu = ETH_MAX_MTU;
1508                         }
1509                 }
1510                 bpf_prog_put(old_prog);
1511         }
1512
1513         if ((!!old_prog ^ !!prog) && peer)
1514                 netdev_update_features(peer);
1515
1516         return 0;
1517 err:
1518         priv->_xdp_prog = old_prog;
1519
1520         return err;
1521 }
1522
1523 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1524 {
1525         switch (xdp->command) {
1526         case XDP_SETUP_PROG:
1527                 return veth_xdp_set(dev, xdp->prog, xdp->extack);
1528         default:
1529                 return -EINVAL;
1530         }
1531 }
1532
1533 static const struct net_device_ops veth_netdev_ops = {
1534         .ndo_init            = veth_dev_init,
1535         .ndo_open            = veth_open,
1536         .ndo_stop            = veth_close,
1537         .ndo_start_xmit      = veth_xmit,
1538         .ndo_get_stats64     = veth_get_stats64,
1539         .ndo_set_rx_mode     = veth_set_multicast_list,
1540         .ndo_set_mac_address = eth_mac_addr,
1541 #ifdef CONFIG_NET_POLL_CONTROLLER
1542         .ndo_poll_controller    = veth_poll_controller,
1543 #endif
1544         .ndo_get_iflink         = veth_get_iflink,
1545         .ndo_fix_features       = veth_fix_features,
1546         .ndo_set_features       = veth_set_features,
1547         .ndo_features_check     = passthru_features_check,
1548         .ndo_set_rx_headroom    = veth_set_rx_headroom,
1549         .ndo_bpf                = veth_xdp,
1550         .ndo_xdp_xmit           = veth_ndo_xdp_xmit,
1551         .ndo_get_peer_dev       = veth_peer_dev,
1552 };
1553
1554 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1555                        NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1556                        NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1557                        NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1558                        NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1559
1560 static void veth_setup(struct net_device *dev)
1561 {
1562         ether_setup(dev);
1563
1564         dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1565         dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1566         dev->priv_flags |= IFF_NO_QUEUE;
1567         dev->priv_flags |= IFF_PHONY_HEADROOM;
1568
1569         dev->netdev_ops = &veth_netdev_ops;
1570         dev->ethtool_ops = &veth_ethtool_ops;
1571         dev->features |= NETIF_F_LLTX;
1572         dev->features |= VETH_FEATURES;
1573         dev->vlan_features = dev->features &
1574                              ~(NETIF_F_HW_VLAN_CTAG_TX |
1575                                NETIF_F_HW_VLAN_STAG_TX |
1576                                NETIF_F_HW_VLAN_CTAG_RX |
1577                                NETIF_F_HW_VLAN_STAG_RX);
1578         dev->needs_free_netdev = true;
1579         dev->priv_destructor = veth_dev_free;
1580         dev->max_mtu = ETH_MAX_MTU;
1581
1582         dev->hw_features = VETH_FEATURES;
1583         dev->hw_enc_features = VETH_FEATURES;
1584         dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1585 }
1586
1587 /*
1588  * netlink interface
1589  */
1590
1591 static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1592                          struct netlink_ext_ack *extack)
1593 {
1594         if (tb[IFLA_ADDRESS]) {
1595                 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1596                         return -EINVAL;
1597                 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1598                         return -EADDRNOTAVAIL;
1599         }
1600         if (tb[IFLA_MTU]) {
1601                 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1602                         return -EINVAL;
1603         }
1604         return 0;
1605 }
1606
1607 static struct rtnl_link_ops veth_link_ops;
1608
1609 static void veth_disable_gro(struct net_device *dev)
1610 {
1611         dev->features &= ~NETIF_F_GRO;
1612         dev->wanted_features &= ~NETIF_F_GRO;
1613         netdev_update_features(dev);
1614 }
1615
1616 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1617 {
1618         int err;
1619
1620         if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1621                 err = netif_set_real_num_tx_queues(dev, 1);
1622                 if (err)
1623                         return err;
1624         }
1625         if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1626                 err = netif_set_real_num_rx_queues(dev, 1);
1627                 if (err)
1628                         return err;
1629         }
1630         return 0;
1631 }
1632
1633 static int veth_newlink(struct net *src_net, struct net_device *dev,
1634                         struct nlattr *tb[], struct nlattr *data[],
1635                         struct netlink_ext_ack *extack)
1636 {
1637         int err;
1638         struct net_device *peer;
1639         struct veth_priv *priv;
1640         char ifname[IFNAMSIZ];
1641         struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1642         unsigned char name_assign_type;
1643         struct ifinfomsg *ifmp;
1644         struct net *net;
1645
1646         /*
1647          * create and register peer first
1648          */
1649         if (data != NULL && data[VETH_INFO_PEER] != NULL) {
1650                 struct nlattr *nla_peer;
1651
1652                 nla_peer = data[VETH_INFO_PEER];
1653                 ifmp = nla_data(nla_peer);
1654                 err = rtnl_nla_parse_ifla(peer_tb,
1655                                           nla_data(nla_peer) + sizeof(struct ifinfomsg),
1656                                           nla_len(nla_peer) - sizeof(struct ifinfomsg),
1657                                           NULL);
1658                 if (err < 0)
1659                         return err;
1660
1661                 err = veth_validate(peer_tb, NULL, extack);
1662                 if (err < 0)
1663                         return err;
1664
1665                 tbp = peer_tb;
1666         } else {
1667                 ifmp = NULL;
1668                 tbp = tb;
1669         }
1670
1671         if (ifmp && tbp[IFLA_IFNAME]) {
1672                 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1673                 name_assign_type = NET_NAME_USER;
1674         } else {
1675                 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1676                 name_assign_type = NET_NAME_ENUM;
1677         }
1678
1679         net = rtnl_link_get_net(src_net, tbp);
1680         if (IS_ERR(net))
1681                 return PTR_ERR(net);
1682
1683         peer = rtnl_create_link(net, ifname, name_assign_type,
1684                                 &veth_link_ops, tbp, extack);
1685         if (IS_ERR(peer)) {
1686                 put_net(net);
1687                 return PTR_ERR(peer);
1688         }
1689
1690         if (!ifmp || !tbp[IFLA_ADDRESS])
1691                 eth_hw_addr_random(peer);
1692
1693         if (ifmp && (dev->ifindex != 0))
1694                 peer->ifindex = ifmp->ifi_index;
1695
1696         peer->gso_max_size = dev->gso_max_size;
1697         peer->gso_max_segs = dev->gso_max_segs;
1698
1699         err = register_netdevice(peer);
1700         put_net(net);
1701         net = NULL;
1702         if (err < 0)
1703                 goto err_register_peer;
1704
1705         /* keep GRO disabled by default to be consistent with the established
1706          * veth behavior
1707          */
1708         veth_disable_gro(peer);
1709         netif_carrier_off(peer);
1710
1711         err = rtnl_configure_link(peer, ifmp);
1712         if (err < 0)
1713                 goto err_configure_peer;
1714
1715         /*
1716          * register dev last
1717          *
1718          * note, that since we've registered new device the dev's name
1719          * should be re-allocated
1720          */
1721
1722         if (tb[IFLA_ADDRESS] == NULL)
1723                 eth_hw_addr_random(dev);
1724
1725         if (tb[IFLA_IFNAME])
1726                 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1727         else
1728                 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1729
1730         err = register_netdevice(dev);
1731         if (err < 0)
1732                 goto err_register_dev;
1733
1734         netif_carrier_off(dev);
1735
1736         /*
1737          * tie the deviced together
1738          */
1739
1740         priv = netdev_priv(dev);
1741         rcu_assign_pointer(priv->peer, peer);
1742         err = veth_init_queues(dev, tb);
1743         if (err)
1744                 goto err_queues;
1745
1746         priv = netdev_priv(peer);
1747         rcu_assign_pointer(priv->peer, dev);
1748         err = veth_init_queues(peer, tb);
1749         if (err)
1750                 goto err_queues;
1751
1752         veth_disable_gro(dev);
1753         return 0;
1754
1755 err_queues:
1756         unregister_netdevice(dev);
1757 err_register_dev:
1758         /* nothing to do */
1759 err_configure_peer:
1760         unregister_netdevice(peer);
1761         return err;
1762
1763 err_register_peer:
1764         free_netdev(peer);
1765         return err;
1766 }
1767
1768 static void veth_dellink(struct net_device *dev, struct list_head *head)
1769 {
1770         struct veth_priv *priv;
1771         struct net_device *peer;
1772
1773         priv = netdev_priv(dev);
1774         peer = rtnl_dereference(priv->peer);
1775
1776         /* Note : dellink() is called from default_device_exit_batch(),
1777          * before a rcu_synchronize() point. The devices are guaranteed
1778          * not being freed before one RCU grace period.
1779          */
1780         RCU_INIT_POINTER(priv->peer, NULL);
1781         unregister_netdevice_queue(dev, head);
1782
1783         if (peer) {
1784                 priv = netdev_priv(peer);
1785                 RCU_INIT_POINTER(priv->peer, NULL);
1786                 unregister_netdevice_queue(peer, head);
1787         }
1788 }
1789
1790 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
1791         [VETH_INFO_PEER]        = { .len = sizeof(struct ifinfomsg) },
1792 };
1793
1794 static struct net *veth_get_link_net(const struct net_device *dev)
1795 {
1796         struct veth_priv *priv = netdev_priv(dev);
1797         struct net_device *peer = rtnl_dereference(priv->peer);
1798
1799         return peer ? dev_net(peer) : dev_net(dev);
1800 }
1801
1802 static unsigned int veth_get_num_queues(void)
1803 {
1804         /* enforce the same queue limit as rtnl_create_link */
1805         int queues = num_possible_cpus();
1806
1807         if (queues > 4096)
1808                 queues = 4096;
1809         return queues;
1810 }
1811
1812 static struct rtnl_link_ops veth_link_ops = {
1813         .kind           = DRV_NAME,
1814         .priv_size      = sizeof(struct veth_priv),
1815         .setup          = veth_setup,
1816         .validate       = veth_validate,
1817         .newlink        = veth_newlink,
1818         .dellink        = veth_dellink,
1819         .policy         = veth_policy,
1820         .maxtype        = VETH_INFO_MAX,
1821         .get_link_net   = veth_get_link_net,
1822         .get_num_tx_queues      = veth_get_num_queues,
1823         .get_num_rx_queues      = veth_get_num_queues,
1824 };
1825
1826 /*
1827  * init/fini
1828  */
1829
1830 static __init int veth_init(void)
1831 {
1832         return rtnl_link_register(&veth_link_ops);
1833 }
1834
1835 static __exit void veth_exit(void)
1836 {
1837         rtnl_link_unregister(&veth_link_ops);
1838 }
1839
1840 module_init(veth_init);
1841 module_exit(veth_exit);
1842
1843 MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
1844 MODULE_LICENSE("GPL v2");
1845 MODULE_ALIAS_RTNL_LINK(DRV_NAME);