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 (!READ_ONCE(rq->rx_notify_masked) &&
 276             napi_schedule_prep(&rq->xdp_napi)) {
 277                 WRITE_ONCE(rq->rx_notify_masked, true);
 278                 __napi_schedule(&rq->xdp_napi);
 279         }
 280 }
 281
 282 static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb)
 283 {
 284         if (unlikely(ptr_ring_produce(&rq->xdp_ring, skb))) {
 285                 dev_kfree_skb_any(skb);
 286                 return NET_RX_DROP;
 287         }
 288
 289         return NET_RX_SUCCESS;
 290 }
 291
 292 static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb,
 293                             struct veth_rq *rq, bool xdp)
 294 {
 295         return __dev_forward_skb(dev, skb) ?: xdp ?
 296                 veth_xdp_rx(rq, skb) :
 297                 netif_rx(skb);
 298 }
 299
 300 /* return true if the specified skb has chances of GRO aggregation
 301  * Don't strive for accuracy, but try to avoid GRO overhead in the most
 302  * common scenarios.
 303  * When XDP is enabled, all traffic is considered eligible, as the xmit
 304  * device has TSO off.
 305  * When TSO is enabled on the xmit device, we are likely interested only
 306  * in UDP aggregation, explicitly check for that if the skb is suspected
 307  * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets -
 308  * to belong to locally generated UDP traffic.
 309  */
 310 static bool veth_skb_is_eligible_for_gro(const struct net_device *dev,
 311                                          const struct net_device *rcv,
 312                                          const struct sk_buff *skb)
 313 {
 314         return !(dev->features & NETIF_F_ALL_TSO) ||
 315                 (skb->destructor == sock_wfree &&
 316                  rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD));
 317 }
 318
 319 static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 320 {
 321         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 322         struct veth_rq *rq = NULL;
 323         struct net_device *rcv;
 324         int length = skb->len;
 325         bool use_napi = false;
 326         int rxq;
 327
 328         rcu_read_lock();
 329         rcv = rcu_dereference(priv->peer);
 330         if (unlikely(!rcv) || !pskb_may_pull(skb, ETH_HLEN)) {
 331                 kfree_skb(skb);
 332                 goto drop;
 333         }
 334
 335         rcv_priv = netdev_priv(rcv);
 336         rxq = skb_get_queue_mapping(skb);
 337         if (rxq < rcv->real_num_rx_queues) {
 338                 rq = &rcv_priv->rq[rxq];
 339
 340                 /* The napi pointer is available when an XDP program is
 341                  * attached or when GRO is enabled
 342                  * Don't bother with napi/GRO if the skb can't be aggregated
 343                  */
 344                 use_napi = rcu_access_pointer(rq->napi) &&
 345                            veth_skb_is_eligible_for_gro(dev, rcv, skb);
 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 (stats.xdp_redirect > 0)
 920                 xdp_do_flush();
 921
 922         if (done < budget && napi_complete_done(napi, done)) {
 923                 /* Write rx_notify_masked before reading ptr_ring */
 924                 smp_store_mb(rq->rx_notify_masked, false);
 925                 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
 926                         if (napi_schedule_prep(&rq->xdp_napi)) {
 927                                 WRITE_ONCE(rq->rx_notify_masked, true);
 928                                 __napi_schedule(&rq->xdp_napi);
 929                         }
 930                 }
 931         }
 932
 933         if (stats.xdp_tx > 0)
 934                 veth_xdp_flush(rq, &bq);
 935         xdp_clear_return_frame_no_direct();
 936
 937         return done;
 938 }
 939
 940 static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
 941 {
 942         struct veth_priv *priv = netdev_priv(dev);
 943         int err, i;
 944
 945         for (i = start; i < end; i++) {
 946                 struct veth_rq *rq = &priv->rq[i];
 947
 948                 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
 949                 if (err)
 950                         goto err_xdp_ring;
 951         }
 952
 953         for (i = start; i < end; i++) {
 954                 struct veth_rq *rq = &priv->rq[i];
 955
 956                 napi_enable(&rq->xdp_napi);
 957                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
 958         }
 959
 960         return 0;
 961
 962 err_xdp_ring:
 963         for (i--; i >= start; i--)
 964                 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
 965
 966         return err;
 967 }
 968
 969 static int __veth_napi_enable(struct net_device *dev)
 970 {
 971         return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
 972 }
 973
 974 static void veth_napi_del_range(struct net_device *dev, int start, int end)
 975 {
 976         struct veth_priv *priv = netdev_priv(dev);
 977         int i;
 978
 979         for (i = start; i < end; i++) {
 980                 struct veth_rq *rq = &priv->rq[i];
 981
 982                 rcu_assign_pointer(priv->rq[i].napi, NULL);
 983                 napi_disable(&rq->xdp_napi);
 984                 __netif_napi_del(&rq->xdp_napi);
 985         }
 986         synchronize_net();
 987
 988         for (i = start; i < end; i++) {
 989                 struct veth_rq *rq = &priv->rq[i];
 990
 991                 rq->rx_notify_masked = false;
 992                 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
 993         }
 994 }
 995
 996 static void veth_napi_del(struct net_device *dev)
 997 {
 998         veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
 999 }
1000
1001 static bool veth_gro_requested(const struct net_device *dev)
1002 {
1003         return !!(dev->wanted_features & NETIF_F_GRO);
1004 }
1005
1006 static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1007                                  bool napi_already_on)
1008 {
1009         struct veth_priv *priv = netdev_priv(dev);
1010         int err, i;
1011
1012         for (i = start; i < end; i++) {
1013                 struct veth_rq *rq = &priv->rq[i];
1014
1015                 if (!napi_already_on)
1016                         netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1017                 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1018                 if (err < 0)
1019                         goto err_rxq_reg;
1020
1021                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1022                                                  MEM_TYPE_PAGE_SHARED,
1023                                                  NULL);
1024                 if (err < 0)
1025                         goto err_reg_mem;
1026
1027                 /* Save original mem info as it can be overwritten */
1028                 rq->xdp_mem = rq->xdp_rxq.mem;
1029         }
1030         return 0;
1031
1032 err_reg_mem:
1033         xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1034 err_rxq_reg:
1035         for (i--; i >= start; i--) {
1036                 struct veth_rq *rq = &priv->rq[i];
1037
1038                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1039                 if (!napi_already_on)
1040                         netif_napi_del(&rq->xdp_napi);
1041         }
1042
1043         return err;
1044 }
1045
1046 static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1047                                    bool delete_napi)
1048 {
1049         struct veth_priv *priv = netdev_priv(dev);
1050         int i;
1051
1052         for (i = start; i < end; i++) {
1053                 struct veth_rq *rq = &priv->rq[i];
1054
1055                 rq->xdp_rxq.mem = rq->xdp_mem;
1056                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1057
1058                 if (delete_napi)
1059                         netif_napi_del(&rq->xdp_napi);
1060         }
1061 }
1062
1063 static int veth_enable_xdp(struct net_device *dev)
1064 {
1065         bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1066         struct veth_priv *priv = netdev_priv(dev);
1067         int err, i;
1068
1069         if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1070                 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1071                 if (err)
1072                         return err;
1073
1074                 if (!napi_already_on) {
1075                         err = __veth_napi_enable(dev);
1076                         if (err) {
1077                                 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1078                                 return err;
1079                         }
1080
1081                         if (!veth_gro_requested(dev)) {
1082                                 /* user-space did not require GRO, but adding XDP
1083                                  * is supposed to get GRO working
1084                                  */
1085                                 dev->features |= NETIF_F_GRO;
1086                                 netdev_features_change(dev);
1087                         }
1088                 }
1089         }
1090
1091         for (i = 0; i < dev->real_num_rx_queues; i++) {
1092                 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1093                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1094         }
1095
1096         return 0;
1097 }
1098
1099 static void veth_disable_xdp(struct net_device *dev)
1100 {
1101         struct veth_priv *priv = netdev_priv(dev);
1102         int i;
1103
1104         for (i = 0; i < dev->real_num_rx_queues; i++)
1105                 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1106
1107         if (!netif_running(dev) || !veth_gro_requested(dev)) {
1108                 veth_napi_del(dev);
1109
1110                 /* if user-space did not require GRO, since adding XDP
1111                  * enabled it, clear it now
1112                  */
1113                 if (!veth_gro_requested(dev) && netif_running(dev)) {
1114                         dev->features &= ~NETIF_F_GRO;
1115                         netdev_features_change(dev);
1116                 }
1117         }
1118
1119         veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1120 }
1121
1122 static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1123 {
1124         struct veth_priv *priv = netdev_priv(dev);
1125         int err, i;
1126
1127         for (i = start; i < end; i++) {
1128                 struct veth_rq *rq = &priv->rq[i];
1129
1130                 netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1131         }
1132
1133         err = __veth_napi_enable_range(dev, start, end);
1134         if (err) {
1135                 for (i = start; i < end; i++) {
1136                         struct veth_rq *rq = &priv->rq[i];
1137
1138                         netif_napi_del(&rq->xdp_napi);
1139                 }
1140                 return err;
1141         }
1142         return err;
1143 }
1144
1145 static int veth_napi_enable(struct net_device *dev)
1146 {
1147         return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1148 }
1149
1150 static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1151 {
1152         struct veth_priv *priv = netdev_priv(dev);
1153
1154         if (start >= end)
1155                 return;
1156
1157         if (priv->_xdp_prog) {
1158                 veth_napi_del_range(dev, start, end);
1159                 veth_disable_xdp_range(dev, start, end, false);
1160         } else if (veth_gro_requested(dev)) {
1161                 veth_napi_del_range(dev, start, end);
1162         }
1163 }
1164
1165 static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1166 {
1167         struct veth_priv *priv = netdev_priv(dev);
1168         int err;
1169
1170         if (start >= end)
1171                 return 0;
1172
1173         if (priv->_xdp_prog) {
1174                 /* these channels are freshly initialized, napi is not on there even
1175                  * when GRO is requeste
1176                  */
1177                 err = veth_enable_xdp_range(dev, start, end, false);
1178                 if (err)
1179                         return err;
1180
1181                 err = __veth_napi_enable_range(dev, start, end);
1182                 if (err) {
1183                         /* on error always delete the newly added napis */
1184                         veth_disable_xdp_range(dev, start, end, true);
1185                         return err;
1186                 }
1187         } else if (veth_gro_requested(dev)) {
1188                 return veth_napi_enable_range(dev, start, end);
1189         }
1190         return 0;
1191 }
1192
1193 static int veth_set_channels(struct net_device *dev,
1194                              struct ethtool_channels *ch)
1195 {
1196         struct veth_priv *priv = netdev_priv(dev);
1197         unsigned int old_rx_count, new_rx_count;
1198         struct veth_priv *peer_priv;
1199         struct net_device *peer;
1200         int err;
1201
1202         /* sanity check. Upper bounds are already enforced by the caller */
1203         if (!ch->rx_count || !ch->tx_count)
1204                 return -EINVAL;
1205
1206         /* avoid braking XDP, if that is enabled */
1207         peer = rtnl_dereference(priv->peer);
1208         peer_priv = peer ? netdev_priv(peer) : NULL;
1209         if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1210                 return -EINVAL;
1211
1212         if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1213                 return -EINVAL;
1214
1215         old_rx_count = dev->real_num_rx_queues;
1216         new_rx_count = ch->rx_count;
1217         if (netif_running(dev)) {
1218                 /* turn device off */
1219                 netif_carrier_off(dev);
1220                 if (peer)
1221                         netif_carrier_off(peer);
1222
1223                 /* try to allocate new resurces, as needed*/
1224                 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1225                 if (err)
1226                         goto out;
1227         }
1228
1229         err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1230         if (err)
1231                 goto revert;
1232
1233         err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1234         if (err) {
1235                 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1236
1237                 /* this error condition could happen only if rx and tx change
1238                  * in opposite directions (e.g. tx nr raises, rx nr decreases)
1239                  * and we can't do anything to fully restore the original
1240                  * status
1241                  */
1242                 if (err2)
1243                         pr_warn("Can't restore rx queues config %d -> %d %d",
1244                                 new_rx_count, old_rx_count, err2);
1245                 else
1246                         goto revert;
1247         }
1248
1249 out:
1250         if (netif_running(dev)) {
1251                 /* note that we need to swap the arguments WRT the enable part
1252                  * to identify the range we have to disable
1253                  */
1254                 veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1255                 netif_carrier_on(dev);
1256                 if (peer)
1257                         netif_carrier_on(peer);
1258         }
1259         return err;
1260
1261 revert:
1262         new_rx_count = old_rx_count;
1263         old_rx_count = ch->rx_count;
1264         goto out;
1265 }
1266
1267 static int veth_open(struct net_device *dev)
1268 {
1269         struct veth_priv *priv = netdev_priv(dev);
1270         struct net_device *peer = rtnl_dereference(priv->peer);
1271         int err;
1272
1273         if (!peer)
1274                 return -ENOTCONN;
1275
1276         if (priv->_xdp_prog) {
1277                 err = veth_enable_xdp(dev);
1278                 if (err)
1279                         return err;
1280         } else if (veth_gro_requested(dev)) {
1281                 err = veth_napi_enable(dev);
1282                 if (err)
1283                         return err;
1284         }
1285
1286         if (peer->flags & IFF_UP) {
1287                 netif_carrier_on(dev);
1288                 netif_carrier_on(peer);
1289         }
1290
1291         return 0;
1292 }
1293
1294 static int veth_close(struct net_device *dev)
1295 {
1296         struct veth_priv *priv = netdev_priv(dev);
1297         struct net_device *peer = rtnl_dereference(priv->peer);
1298
1299         netif_carrier_off(dev);
1300         if (peer)
1301                 netif_carrier_off(peer);
1302
1303         if (priv->_xdp_prog)
1304                 veth_disable_xdp(dev);
1305         else if (veth_gro_requested(dev))
1306                 veth_napi_del(dev);
1307
1308         return 0;
1309 }
1310
1311 static int is_valid_veth_mtu(int mtu)
1312 {
1313         return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1314 }
1315
1316 static int veth_alloc_queues(struct net_device *dev)
1317 {
1318         struct veth_priv *priv = netdev_priv(dev);
1319         int i;
1320
1321         priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL);
1322         if (!priv->rq)
1323                 return -ENOMEM;
1324
1325         for (i = 0; i < dev->num_rx_queues; i++) {
1326                 priv->rq[i].dev = dev;
1327                 u64_stats_init(&priv->rq[i].stats.syncp);
1328         }
1329
1330         return 0;
1331 }
1332
1333 static void veth_free_queues(struct net_device *dev)
1334 {
1335         struct veth_priv *priv = netdev_priv(dev);
1336
1337         kfree(priv->rq);
1338 }
1339
1340 static int veth_dev_init(struct net_device *dev)
1341 {
1342         int err;
1343
1344         dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
1345         if (!dev->lstats)
1346                 return -ENOMEM;
1347
1348         err = veth_alloc_queues(dev);
1349         if (err) {
1350                 free_percpu(dev->lstats);
1351                 return err;
1352         }
1353
1354         return 0;
1355 }
1356
1357 static void veth_dev_free(struct net_device *dev)
1358 {
1359         veth_free_queues(dev);
1360         free_percpu(dev->lstats);
1361 }
1362
1363 #ifdef CONFIG_NET_POLL_CONTROLLER
1364 static void veth_poll_controller(struct net_device *dev)
1365 {
1366         /* veth only receives frames when its peer sends one
1367          * Since it has nothing to do with disabling irqs, we are guaranteed
1368          * never to have pending data when we poll for it so
1369          * there is nothing to do here.
1370          *
1371          * We need this though so netpoll recognizes us as an interface that
1372          * supports polling, which enables bridge devices in virt setups to
1373          * still use netconsole
1374          */
1375 }
1376 #endif  /* CONFIG_NET_POLL_CONTROLLER */
1377
1378 static int veth_get_iflink(const struct net_device *dev)
1379 {
1380         struct veth_priv *priv = netdev_priv(dev);
1381         struct net_device *peer;
1382         int iflink;
1383
1384         rcu_read_lock();
1385         peer = rcu_dereference(priv->peer);
1386         iflink = peer ? peer->ifindex : 0;
1387         rcu_read_unlock();
1388
1389         return iflink;
1390 }
1391
1392 static netdev_features_t veth_fix_features(struct net_device *dev,
1393                                            netdev_features_t features)
1394 {
1395         struct veth_priv *priv = netdev_priv(dev);
1396         struct net_device *peer;
1397
1398         peer = rtnl_dereference(priv->peer);
1399         if (peer) {
1400                 struct veth_priv *peer_priv = netdev_priv(peer);
1401
1402                 if (peer_priv->_xdp_prog)
1403                         features &= ~NETIF_F_GSO_SOFTWARE;
1404         }
1405         if (priv->_xdp_prog)
1406                 features |= NETIF_F_GRO;
1407
1408         return features;
1409 }
1410
1411 static int veth_set_features(struct net_device *dev,
1412                              netdev_features_t features)
1413 {
1414         netdev_features_t changed = features ^ dev->features;
1415         struct veth_priv *priv = netdev_priv(dev);
1416         int err;
1417
1418         if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1419                 return 0;
1420
1421         if (features & NETIF_F_GRO) {
1422                 err = veth_napi_enable(dev);
1423                 if (err)
1424                         return err;
1425         } else {
1426                 veth_napi_del(dev);
1427         }
1428         return 0;
1429 }
1430
1431 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1432 {
1433         struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1434         struct net_device *peer;
1435
1436         if (new_hr < 0)
1437                 new_hr = 0;
1438
1439         rcu_read_lock();
1440         peer = rcu_dereference(priv->peer);
1441         if (unlikely(!peer))
1442                 goto out;
1443
1444         peer_priv = netdev_priv(peer);
1445         priv->requested_headroom = new_hr;
1446         new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1447         dev->needed_headroom = new_hr;
1448         peer->needed_headroom = new_hr;
1449
1450 out:
1451         rcu_read_unlock();
1452 }
1453
1454 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1455                         struct netlink_ext_ack *extack)
1456 {
1457         struct veth_priv *priv = netdev_priv(dev);
1458         struct bpf_prog *old_prog;
1459         struct net_device *peer;
1460         unsigned int max_mtu;
1461         int err;
1462
1463         old_prog = priv->_xdp_prog;
1464         priv->_xdp_prog = prog;
1465         peer = rtnl_dereference(priv->peer);
1466
1467         if (prog) {
1468                 if (!peer) {
1469                         NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1470                         err = -ENOTCONN;
1471                         goto err;
1472                 }
1473
1474                 max_mtu = PAGE_SIZE - VETH_XDP_HEADROOM -
1475                           peer->hard_header_len -
1476                           SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1477                 if (peer->mtu > max_mtu) {
1478                         NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1479                         err = -ERANGE;
1480                         goto err;
1481                 }
1482
1483                 if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1484                         NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1485                         err = -ENOSPC;
1486                         goto err;
1487                 }
1488
1489                 if (dev->flags & IFF_UP) {
1490                         err = veth_enable_xdp(dev);
1491                         if (err) {
1492                                 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1493                                 goto err;
1494                         }
1495                 }
1496
1497                 if (!old_prog) {
1498                         peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1499                         peer->max_mtu = max_mtu;
1500                 }
1501         }
1502
1503         if (old_prog) {
1504                 if (!prog) {
1505                         if (dev->flags & IFF_UP)
1506                                 veth_disable_xdp(dev);
1507
1508                         if (peer) {
1509                                 peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1510                                 peer->max_mtu = ETH_MAX_MTU;
1511                         }
1512                 }
1513                 bpf_prog_put(old_prog);
1514         }
1515
1516         if ((!!old_prog ^ !!prog) && peer)
1517                 netdev_update_features(peer);
1518
1519         return 0;
1520 err:
1521         priv->_xdp_prog = old_prog;
1522
1523         return err;
1524 }
1525
1526 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1527 {
1528         switch (xdp->command) {
1529         case XDP_SETUP_PROG:
1530                 return veth_xdp_set(dev, xdp->prog, xdp->extack);
1531         default:
1532                 return -EINVAL;
1533         }
1534 }
1535
1536 static const struct net_device_ops veth_netdev_ops = {
1537         .ndo_init            = veth_dev_init,
1538         .ndo_open            = veth_open,
1539         .ndo_stop            = veth_close,
1540         .ndo_start_xmit      = veth_xmit,
1541         .ndo_get_stats64     = veth_get_stats64,
1542         .ndo_set_rx_mode     = veth_set_multicast_list,
1543         .ndo_set_mac_address = eth_mac_addr,
1544 #ifdef CONFIG_NET_POLL_CONTROLLER
1545         .ndo_poll_controller    = veth_poll_controller,
1546 #endif
1547         .ndo_get_iflink         = veth_get_iflink,
1548         .ndo_fix_features       = veth_fix_features,
1549         .ndo_set_features       = veth_set_features,
1550         .ndo_features_check     = passthru_features_check,
1551         .ndo_set_rx_headroom    = veth_set_rx_headroom,
1552         .ndo_bpf                = veth_xdp,
1553         .ndo_xdp_xmit           = veth_ndo_xdp_xmit,
1554         .ndo_get_peer_dev       = veth_peer_dev,
1555 };
1556
1557 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1558                        NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1559                        NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1560                        NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1561                        NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1562
1563 static void veth_setup(struct net_device *dev)
1564 {
1565         ether_setup(dev);
1566
1567         dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1568         dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1569         dev->priv_flags |= IFF_NO_QUEUE;
1570         dev->priv_flags |= IFF_PHONY_HEADROOM;
1571
1572         dev->netdev_ops = &veth_netdev_ops;
1573         dev->ethtool_ops = &veth_ethtool_ops;
1574         dev->features |= NETIF_F_LLTX;
1575         dev->features |= VETH_FEATURES;
1576         dev->vlan_features = dev->features &
1577                              ~(NETIF_F_HW_VLAN_CTAG_TX |
1578                                NETIF_F_HW_VLAN_STAG_TX |
1579                                NETIF_F_HW_VLAN_CTAG_RX |
1580                                NETIF_F_HW_VLAN_STAG_RX);
1581         dev->needs_free_netdev = true;
1582         dev->priv_destructor = veth_dev_free;
1583         dev->max_mtu = ETH_MAX_MTU;
1584
1585         dev->hw_features = VETH_FEATURES;
1586         dev->hw_enc_features = VETH_FEATURES;
1587         dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1588 }
1589
1590 /*
1591  * netlink interface
1592  */
1593
1594 static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1595                          struct netlink_ext_ack *extack)
1596 {
1597         if (tb[IFLA_ADDRESS]) {
1598                 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1599                         return -EINVAL;
1600                 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1601                         return -EADDRNOTAVAIL;
1602         }
1603         if (tb[IFLA_MTU]) {
1604                 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1605                         return -EINVAL;
1606         }
1607         return 0;
1608 }
1609
1610 static struct rtnl_link_ops veth_link_ops;
1611
1612 static void veth_disable_gro(struct net_device *dev)
1613 {
1614         dev->features &= ~NETIF_F_GRO;
1615         dev->wanted_features &= ~NETIF_F_GRO;
1616         netdev_update_features(dev);
1617 }
1618
1619 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1620 {
1621         int err;
1622
1623         if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1624                 err = netif_set_real_num_tx_queues(dev, 1);
1625                 if (err)
1626                         return err;
1627         }
1628         if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1629                 err = netif_set_real_num_rx_queues(dev, 1);
1630                 if (err)
1631                         return err;
1632         }
1633         return 0;
1634 }
1635
1636 static int veth_newlink(struct net *src_net, struct net_device *dev,
1637                         struct nlattr *tb[], struct nlattr *data[],
1638                         struct netlink_ext_ack *extack)
1639 {
1640         int err;
1641         struct net_device *peer;
1642         struct veth_priv *priv;
1643         char ifname[IFNAMSIZ];
1644         struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1645         unsigned char name_assign_type;
1646         struct ifinfomsg *ifmp;
1647         struct net *net;
1648
1649         /*
1650          * create and register peer first
1651          */
1652         if (data != NULL && data[VETH_INFO_PEER] != NULL) {
1653                 struct nlattr *nla_peer;
1654
1655                 nla_peer = data[VETH_INFO_PEER];
1656                 ifmp = nla_data(nla_peer);
1657                 err = rtnl_nla_parse_ifla(peer_tb,
1658                                           nla_data(nla_peer) + sizeof(struct ifinfomsg),
1659                                           nla_len(nla_peer) - sizeof(struct ifinfomsg),
1660                                           NULL);
1661                 if (err < 0)
1662                         return err;
1663
1664                 err = veth_validate(peer_tb, NULL, extack);
1665                 if (err < 0)
1666                         return err;
1667
1668                 tbp = peer_tb;
1669         } else {
1670                 ifmp = NULL;
1671                 tbp = tb;
1672         }
1673
1674         if (ifmp && tbp[IFLA_IFNAME]) {
1675                 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1676                 name_assign_type = NET_NAME_USER;
1677         } else {
1678                 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1679                 name_assign_type = NET_NAME_ENUM;
1680         }
1681
1682         net = rtnl_link_get_net(src_net, tbp);
1683         if (IS_ERR(net))
1684                 return PTR_ERR(net);
1685
1686         peer = rtnl_create_link(net, ifname, name_assign_type,
1687                                 &veth_link_ops, tbp, extack);
1688         if (IS_ERR(peer)) {
1689                 put_net(net);
1690                 return PTR_ERR(peer);
1691         }
1692
1693         if (!ifmp || !tbp[IFLA_ADDRESS])
1694                 eth_hw_addr_random(peer);
1695
1696         if (ifmp && (dev->ifindex != 0))
1697                 peer->ifindex = ifmp->ifi_index;
1698
1699         peer->gso_max_size = dev->gso_max_size;
1700         peer->gso_max_segs = dev->gso_max_segs;
1701
1702         err = register_netdevice(peer);
1703         put_net(net);
1704         net = NULL;
1705         if (err < 0)
1706                 goto err_register_peer;
1707
1708         /* keep GRO disabled by default to be consistent with the established
1709          * veth behavior
1710          */
1711         veth_disable_gro(peer);
1712         netif_carrier_off(peer);
1713
1714         err = rtnl_configure_link(peer, ifmp);
1715         if (err < 0)
1716                 goto err_configure_peer;
1717
1718         /*
1719          * register dev last
1720          *
1721          * note, that since we've registered new device the dev's name
1722          * should be re-allocated
1723          */
1724
1725         if (tb[IFLA_ADDRESS] == NULL)
1726                 eth_hw_addr_random(dev);
1727
1728         if (tb[IFLA_IFNAME])
1729                 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1730         else
1731                 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1732
1733         err = register_netdevice(dev);
1734         if (err < 0)
1735                 goto err_register_dev;
1736
1737         netif_carrier_off(dev);
1738
1739         /*
1740          * tie the deviced together
1741          */
1742
1743         priv = netdev_priv(dev);
1744         rcu_assign_pointer(priv->peer, peer);
1745         err = veth_init_queues(dev, tb);
1746         if (err)
1747                 goto err_queues;
1748
1749         priv = netdev_priv(peer);
1750         rcu_assign_pointer(priv->peer, dev);
1751         err = veth_init_queues(peer, tb);
1752         if (err)
1753                 goto err_queues;
1754
1755         veth_disable_gro(dev);
1756         return 0;
1757
1758 err_queues:
1759         unregister_netdevice(dev);
1760 err_register_dev:
1761         /* nothing to do */
1762 err_configure_peer:
1763         unregister_netdevice(peer);
1764         return err;
1765
1766 err_register_peer:
1767         free_netdev(peer);
1768         return err;
1769 }
1770
1771 static void veth_dellink(struct net_device *dev, struct list_head *head)
1772 {
1773         struct veth_priv *priv;
1774         struct net_device *peer;
1775
1776         priv = netdev_priv(dev);
1777         peer = rtnl_dereference(priv->peer);
1778
1779         /* Note : dellink() is called from default_device_exit_batch(),
1780          * before a rcu_synchronize() point. The devices are guaranteed
1781          * not being freed before one RCU grace period.
1782          */
1783         RCU_INIT_POINTER(priv->peer, NULL);
1784         unregister_netdevice_queue(dev, head);
1785
1786         if (peer) {
1787                 priv = netdev_priv(peer);
1788                 RCU_INIT_POINTER(priv->peer, NULL);
1789                 unregister_netdevice_queue(peer, head);
1790         }
1791 }
1792
1793 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
1794         [VETH_INFO_PEER]        = { .len = sizeof(struct ifinfomsg) },
1795 };
1796
1797 static struct net *veth_get_link_net(const struct net_device *dev)
1798 {
1799         struct veth_priv *priv = netdev_priv(dev);
1800         struct net_device *peer = rtnl_dereference(priv->peer);
1801
1802         return peer ? dev_net(peer) : dev_net(dev);
1803 }
1804
1805 static unsigned int veth_get_num_queues(void)
1806 {
1807         /* enforce the same queue limit as rtnl_create_link */
1808         int queues = num_possible_cpus();
1809
1810         if (queues > 4096)
1811                 queues = 4096;
1812         return queues;
1813 }
1814
1815 static struct rtnl_link_ops veth_link_ops = {
1816         .kind           = DRV_NAME,
1817         .priv_size      = sizeof(struct veth_priv),
1818         .setup          = veth_setup,
1819         .validate       = veth_validate,
1820         .newlink        = veth_newlink,
1821         .dellink        = veth_dellink,
1822         .policy         = veth_policy,
1823         .maxtype        = VETH_INFO_MAX,
1824         .get_link_net   = veth_get_link_net,
1825         .get_num_tx_queues      = veth_get_num_queues,
1826         .get_num_rx_queues      = veth_get_num_queues,
1827 };
1828
1829 /*
1830  * init/fini
1831  */
1832
1833 static __init int veth_init(void)
1834 {
1835         return rtnl_link_register(&veth_link_ops);
1836 }
1837
1838 static __exit void veth_exit(void)
1839 {
1840         rtnl_link_unregister(&veth_link_ops);
1841 }
1842
1843 module_init(veth_init);
1844 module_exit(veth_exit);
1845
1846 MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
1847 MODULE_LICENSE("GPL v2");
1848 MODULE_ALIAS_RTNL_LINK(DRV_NAME);