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