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/netdev_lock.h>
  21 #include <net/xfrm.h>
  22 #include <net/xdp.h>
  23 #include <linux/veth.h>
  24 #include <linux/module.h>
  25 #include <linux/bpf.h>
  26 #include <linux/filter.h>
  27 #include <linux/ptr_ring.h>
  28 #include <linux/bpf_trace.h>
  29 #include <linux/net_tstamp.h>
  30 #include <linux/skbuff_ref.h>
  31 #include <net/page_pool/helpers.h>
  32
  33 #define DRV_NAME        "veth"
  34 #define DRV_VERSION     "1.0"
  35
  36 #define VETH_XDP_FLAG           BIT(0)
  37 #define VETH_RING_SIZE          256
  38 #define VETH_XDP_HEADROOM       (XDP_PACKET_HEADROOM + NET_IP_ALIGN)
  39
  40 #define VETH_XDP_TX_BULK_SIZE   16
  41 #define VETH_XDP_BATCH          16
  42
  43 struct veth_stats {
  44         u64     rx_drops;
  45         /* xdp */
  46         u64     xdp_packets;
  47         u64     xdp_bytes;
  48         u64     xdp_redirect;
  49         u64     xdp_drops;
  50         u64     xdp_tx;
  51         u64     xdp_tx_err;
  52         u64     peer_tq_xdp_xmit;
  53         u64     peer_tq_xdp_xmit_err;
  54 };
  55
  56 struct veth_rq_stats {
  57         struct veth_stats       vs;
  58         struct u64_stats_sync   syncp;
  59 };
  60
  61 struct veth_rq {
  62         struct napi_struct      xdp_napi;
  63         struct napi_struct __rcu *napi; /* points to xdp_napi when the latter is initialized */
  64         struct net_device       *dev;
  65         struct bpf_prog __rcu   *xdp_prog;
  66         struct xdp_mem_info     xdp_mem;
  67         struct veth_rq_stats    stats;
  68         bool                    rx_notify_masked;
  69         struct ptr_ring         xdp_ring;
  70         struct xdp_rxq_info     xdp_rxq;
  71         struct page_pool        *page_pool;
  72 };
  73
  74 struct veth_priv {
  75         struct net_device __rcu *peer;
  76         atomic64_t              dropped;
  77         struct bpf_prog         *_xdp_prog;
  78         struct veth_rq          *rq;
  79         unsigned int            requested_headroom;
  80 };
  81
  82 struct veth_xdp_tx_bq {
  83         struct xdp_frame *q[VETH_XDP_TX_BULK_SIZE];
  84         unsigned int count;
  85 };
  86
  87 /*
  88  * ethtool interface
  89  */
  90
  91 struct veth_q_stat_desc {
  92         char    desc[ETH_GSTRING_LEN];
  93         size_t  offset;
  94 };
  95
  96 #define VETH_RQ_STAT(m) offsetof(struct veth_stats, m)
  97
  98 static const struct veth_q_stat_desc veth_rq_stats_desc[] = {
  99         { "xdp_packets",        VETH_RQ_STAT(xdp_packets) },
 100         { "xdp_bytes",          VETH_RQ_STAT(xdp_bytes) },
 101         { "drops",              VETH_RQ_STAT(rx_drops) },
 102         { "xdp_redirect",       VETH_RQ_STAT(xdp_redirect) },
 103         { "xdp_drops",          VETH_RQ_STAT(xdp_drops) },
 104         { "xdp_tx",             VETH_RQ_STAT(xdp_tx) },
 105         { "xdp_tx_errors",      VETH_RQ_STAT(xdp_tx_err) },
 106 };
 107
 108 #define VETH_RQ_STATS_LEN       ARRAY_SIZE(veth_rq_stats_desc)
 109
 110 static const struct veth_q_stat_desc veth_tq_stats_desc[] = {
 111         { "xdp_xmit",           VETH_RQ_STAT(peer_tq_xdp_xmit) },
 112         { "xdp_xmit_errors",    VETH_RQ_STAT(peer_tq_xdp_xmit_err) },
 113 };
 114
 115 #define VETH_TQ_STATS_LEN       ARRAY_SIZE(veth_tq_stats_desc)
 116
 117 static struct {
 118         const char string[ETH_GSTRING_LEN];
 119 } ethtool_stats_keys[] = {
 120         { "peer_ifindex" },
 121 };
 122
 123 struct veth_xdp_buff {
 124         struct xdp_buff xdp;
 125         struct sk_buff *skb;
 126 };
 127
 128 static int veth_get_link_ksettings(struct net_device *dev,
 129                                    struct ethtool_link_ksettings *cmd)
 130 {
 131         cmd->base.speed         = SPEED_10000;
 132         cmd->base.duplex        = DUPLEX_FULL;
 133         cmd->base.port          = PORT_TP;
 134         cmd->base.autoneg       = AUTONEG_DISABLE;
 135         return 0;
 136 }
 137
 138 static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info)
 139 {
 140         strscpy(info->driver, DRV_NAME, sizeof(info->driver));
 141         strscpy(info->version, DRV_VERSION, sizeof(info->version));
 142 }
 143
 144 static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
 145 {
 146         u8 *p = buf;
 147         int i, j;
 148
 149         switch(stringset) {
 150         case ETH_SS_STATS:
 151                 memcpy(p, &ethtool_stats_keys, sizeof(ethtool_stats_keys));
 152                 p += sizeof(ethtool_stats_keys);
 153                 for (i = 0; i < dev->real_num_rx_queues; i++)
 154                         for (j = 0; j < VETH_RQ_STATS_LEN; j++)
 155                                 ethtool_sprintf(&p, "rx_queue_%u_%.18s",
 156                                                 i, veth_rq_stats_desc[j].desc);
 157
 158                 for (i = 0; i < dev->real_num_tx_queues; i++)
 159                         for (j = 0; j < VETH_TQ_STATS_LEN; j++)
 160                                 ethtool_sprintf(&p, "tx_queue_%u_%.18s",
 161                                                 i, veth_tq_stats_desc[j].desc);
 162
 163                 page_pool_ethtool_stats_get_strings(p);
 164                 break;
 165         }
 166 }
 167
 168 static int veth_get_sset_count(struct net_device *dev, int sset)
 169 {
 170         switch (sset) {
 171         case ETH_SS_STATS:
 172                 return ARRAY_SIZE(ethtool_stats_keys) +
 173                        VETH_RQ_STATS_LEN * dev->real_num_rx_queues +
 174                        VETH_TQ_STATS_LEN * dev->real_num_tx_queues +
 175                        page_pool_ethtool_stats_get_count();
 176         default:
 177                 return -EOPNOTSUPP;
 178         }
 179 }
 180
 181 static void veth_get_page_pool_stats(struct net_device *dev, u64 *data)
 182 {
 183 #ifdef CONFIG_PAGE_POOL_STATS
 184         struct veth_priv *priv = netdev_priv(dev);
 185         struct page_pool_stats pp_stats = {};
 186         int i;
 187
 188         for (i = 0; i < dev->real_num_rx_queues; i++) {
 189                 if (!priv->rq[i].page_pool)
 190                         continue;
 191                 page_pool_get_stats(priv->rq[i].page_pool, &pp_stats);
 192         }
 193         page_pool_ethtool_stats_get(data, &pp_stats);
 194 #endif /* CONFIG_PAGE_POOL_STATS */
 195 }
 196
 197 static void veth_get_ethtool_stats(struct net_device *dev,
 198                 struct ethtool_stats *stats, u64 *data)
 199 {
 200         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 201         struct net_device *peer = rtnl_dereference(priv->peer);
 202         int i, j, idx, pp_idx;
 203
 204         data[0] = peer ? peer->ifindex : 0;
 205         idx = 1;
 206         for (i = 0; i < dev->real_num_rx_queues; i++) {
 207                 const struct veth_rq_stats *rq_stats = &priv->rq[i].stats;
 208                 const void *stats_base = (void *)&rq_stats->vs;
 209                 unsigned int start;
 210                 size_t offset;
 211
 212                 do {
 213                         start = u64_stats_fetch_begin(&rq_stats->syncp);
 214                         for (j = 0; j < VETH_RQ_STATS_LEN; j++) {
 215                                 offset = veth_rq_stats_desc[j].offset;
 216                                 data[idx + j] = *(u64 *)(stats_base + offset);
 217                         }
 218                 } while (u64_stats_fetch_retry(&rq_stats->syncp, start));
 219                 idx += VETH_RQ_STATS_LEN;
 220         }
 221         pp_idx = idx;
 222
 223         if (!peer)
 224                 goto page_pool_stats;
 225
 226         rcv_priv = netdev_priv(peer);
 227         for (i = 0; i < peer->real_num_rx_queues; i++) {
 228                 const struct veth_rq_stats *rq_stats = &rcv_priv->rq[i].stats;
 229                 const void *base = (void *)&rq_stats->vs;
 230                 unsigned int start, tx_idx = idx;
 231                 size_t offset;
 232
 233                 tx_idx += (i % dev->real_num_tx_queues) * VETH_TQ_STATS_LEN;
 234                 do {
 235                         start = u64_stats_fetch_begin(&rq_stats->syncp);
 236                         for (j = 0; j < VETH_TQ_STATS_LEN; j++) {
 237                                 offset = veth_tq_stats_desc[j].offset;
 238                                 data[tx_idx + j] += *(u64 *)(base + offset);
 239                         }
 240                 } while (u64_stats_fetch_retry(&rq_stats->syncp, start));
 241         }
 242         pp_idx = idx + dev->real_num_tx_queues * VETH_TQ_STATS_LEN;
 243
 244 page_pool_stats:
 245         veth_get_page_pool_stats(dev, &data[pp_idx]);
 246 }
 247
 248 static void veth_get_channels(struct net_device *dev,
 249                               struct ethtool_channels *channels)
 250 {
 251         channels->tx_count = dev->real_num_tx_queues;
 252         channels->rx_count = dev->real_num_rx_queues;
 253         channels->max_tx = dev->num_tx_queues;
 254         channels->max_rx = dev->num_rx_queues;
 255 }
 256
 257 static int veth_set_channels(struct net_device *dev,
 258                              struct ethtool_channels *ch);
 259
 260 static const struct ethtool_ops veth_ethtool_ops = {
 261         .get_drvinfo            = veth_get_drvinfo,
 262         .get_link               = ethtool_op_get_link,
 263         .get_strings            = veth_get_strings,
 264         .get_sset_count         = veth_get_sset_count,
 265         .get_ethtool_stats      = veth_get_ethtool_stats,
 266         .get_link_ksettings     = veth_get_link_ksettings,
 267         .get_ts_info            = ethtool_op_get_ts_info,
 268         .get_channels           = veth_get_channels,
 269         .set_channels           = veth_set_channels,
 270 };
 271
 272 /* general routines */
 273
 274 static bool veth_is_xdp_frame(void *ptr)
 275 {
 276         return (unsigned long)ptr & VETH_XDP_FLAG;
 277 }
 278
 279 static struct xdp_frame *veth_ptr_to_xdp(void *ptr)
 280 {
 281         return (void *)((unsigned long)ptr & ~VETH_XDP_FLAG);
 282 }
 283
 284 static void *veth_xdp_to_ptr(struct xdp_frame *xdp)
 285 {
 286         return (void *)((unsigned long)xdp | VETH_XDP_FLAG);
 287 }
 288
 289 static void veth_ptr_free(void *ptr)
 290 {
 291         if (veth_is_xdp_frame(ptr))
 292                 xdp_return_frame(veth_ptr_to_xdp(ptr));
 293         else
 294                 kfree_skb(ptr);
 295 }
 296
 297 static void __veth_xdp_flush(struct veth_rq *rq)
 298 {
 299         /* Write ptr_ring before reading rx_notify_masked */
 300         smp_mb();
 301         if (!READ_ONCE(rq->rx_notify_masked) &&
 302             napi_schedule_prep(&rq->xdp_napi)) {
 303                 WRITE_ONCE(rq->rx_notify_masked, true);
 304                 __napi_schedule(&rq->xdp_napi);
 305         }
 306 }
 307
 308 static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb)
 309 {
 310         if (unlikely(ptr_ring_produce(&rq->xdp_ring, skb)))
 311                 return NETDEV_TX_BUSY; /* signal qdisc layer */
 312
 313         return NET_RX_SUCCESS; /* same as NETDEV_TX_OK */
 314 }
 315
 316 static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb,
 317                             struct veth_rq *rq, bool xdp)
 318 {
 319         return __dev_forward_skb(dev, skb) ?: xdp ?
 320                 veth_xdp_rx(rq, skb) :
 321                 __netif_rx(skb);
 322 }
 323
 324 /* return true if the specified skb has chances of GRO aggregation
 325  * Don't strive for accuracy, but try to avoid GRO overhead in the most
 326  * common scenarios.
 327  * When XDP is enabled, all traffic is considered eligible, as the xmit
 328  * device has TSO off.
 329  * When TSO is enabled on the xmit device, we are likely interested only
 330  * in UDP aggregation, explicitly check for that if the skb is suspected
 331  * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets -
 332  * to belong to locally generated UDP traffic.
 333  */
 334 static bool veth_skb_is_eligible_for_gro(const struct net_device *dev,
 335                                          const struct net_device *rcv,
 336                                          const struct sk_buff *skb)
 337 {
 338         return !(dev->features & NETIF_F_ALL_TSO) ||
 339                 (skb->destructor == sock_wfree &&
 340                  rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD));
 341 }
 342
 343 static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 344 {
 345         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 346         struct veth_rq *rq = NULL;
 347         struct netdev_queue *txq;
 348         struct net_device *rcv;
 349         int length = skb->len;
 350         bool use_napi = false;
 351         int ret, rxq;
 352
 353         rcu_read_lock();
 354         rcv = rcu_dereference(priv->peer);
 355         if (unlikely(!rcv) || !pskb_may_pull(skb, ETH_HLEN)) {
 356                 kfree_skb(skb);
 357                 goto drop;
 358         }
 359
 360         rcv_priv = netdev_priv(rcv);
 361         rxq = skb_get_queue_mapping(skb);
 362         if (rxq < rcv->real_num_rx_queues) {
 363                 rq = &rcv_priv->rq[rxq];
 364
 365                 /* The napi pointer is available when an XDP program is
 366                  * attached or when GRO is enabled
 367                  * Don't bother with napi/GRO if the skb can't be aggregated
 368                  */
 369                 use_napi = rcu_access_pointer(rq->napi) &&
 370                            veth_skb_is_eligible_for_gro(dev, rcv, skb);
 371         }
 372
 373         skb_tx_timestamp(skb);
 374
 375         ret = veth_forward_skb(rcv, skb, rq, use_napi);
 376         switch (ret) {
 377         case NET_RX_SUCCESS: /* same as NETDEV_TX_OK */
 378                 if (!use_napi)
 379                         dev_sw_netstats_tx_add(dev, 1, length);
 380                 else
 381                         __veth_xdp_flush(rq);
 382                 break;
 383         case NETDEV_TX_BUSY:
 384                 /* If a qdisc is attached to our virtual device, returning
 385                  * NETDEV_TX_BUSY is allowed.
 386                  */
 387                 txq = netdev_get_tx_queue(dev, rxq);
 388
 389                 if (qdisc_txq_has_no_queue(txq)) {
 390                         dev_kfree_skb_any(skb);
 391                         goto drop;
 392                 }
 393                 /* Restore Eth hdr pulled by dev_forward_skb/eth_type_trans */
 394                 __skb_push(skb, ETH_HLEN);
 395                 /* Depend on prior success packets started NAPI consumer via
 396                  * __veth_xdp_flush(). Cancel TXQ stop if consumer stopped,
 397                  * paired with empty check in veth_poll().
 398                  */
 399                 netif_tx_stop_queue(txq);
 400                 smp_mb__after_atomic();
 401                 if (unlikely(__ptr_ring_empty(&rq->xdp_ring)))
 402                         netif_tx_wake_queue(txq);
 403                 break;
 404         case NET_RX_DROP: /* same as NET_XMIT_DROP */
 405 drop:
 406                 atomic64_inc(&priv->dropped);
 407                 ret = NET_XMIT_DROP;
 408                 break;
 409         default:
 410                 net_crit_ratelimited("%s(%s): Invalid return code(%d)",
 411                                      __func__, dev->name, ret);
 412         }
 413         rcu_read_unlock();
 414
 415         return ret;
 416 }
 417
 418 static void veth_stats_rx(struct veth_stats *result, struct net_device *dev)
 419 {
 420         struct veth_priv *priv = netdev_priv(dev);
 421         int i;
 422
 423         result->peer_tq_xdp_xmit_err = 0;
 424         result->xdp_packets = 0;
 425         result->xdp_tx_err = 0;
 426         result->xdp_bytes = 0;
 427         result->rx_drops = 0;
 428         for (i = 0; i < dev->num_rx_queues; i++) {
 429                 u64 packets, bytes, drops, xdp_tx_err, peer_tq_xdp_xmit_err;
 430                 struct veth_rq_stats *stats = &priv->rq[i].stats;
 431                 unsigned int start;
 432
 433                 do {
 434                         start = u64_stats_fetch_begin(&stats->syncp);
 435                         peer_tq_xdp_xmit_err = stats->vs.peer_tq_xdp_xmit_err;
 436                         xdp_tx_err = stats->vs.xdp_tx_err;
 437                         packets = stats->vs.xdp_packets;
 438                         bytes = stats->vs.xdp_bytes;
 439                         drops = stats->vs.rx_drops;
 440                 } while (u64_stats_fetch_retry(&stats->syncp, start));
 441                 result->peer_tq_xdp_xmit_err += peer_tq_xdp_xmit_err;
 442                 result->xdp_tx_err += xdp_tx_err;
 443                 result->xdp_packets += packets;
 444                 result->xdp_bytes += bytes;
 445                 result->rx_drops += drops;
 446         }
 447 }
 448
 449 static void veth_get_stats64(struct net_device *dev,
 450                              struct rtnl_link_stats64 *tot)
 451 {
 452         struct veth_priv *priv = netdev_priv(dev);
 453         struct net_device *peer;
 454         struct veth_stats rx;
 455
 456         tot->tx_dropped = atomic64_read(&priv->dropped);
 457         dev_fetch_sw_netstats(tot, dev->tstats);
 458
 459         veth_stats_rx(&rx, dev);
 460         tot->tx_dropped += rx.xdp_tx_err;
 461         tot->rx_dropped = rx.rx_drops + rx.peer_tq_xdp_xmit_err;
 462         tot->rx_bytes += rx.xdp_bytes;
 463         tot->rx_packets += rx.xdp_packets;
 464
 465         rcu_read_lock();
 466         peer = rcu_dereference(priv->peer);
 467         if (peer) {
 468                 struct rtnl_link_stats64 tot_peer = {};
 469
 470                 dev_fetch_sw_netstats(&tot_peer, peer->tstats);
 471                 tot->rx_bytes += tot_peer.tx_bytes;
 472                 tot->rx_packets += tot_peer.tx_packets;
 473
 474                 veth_stats_rx(&rx, peer);
 475                 tot->tx_dropped += rx.peer_tq_xdp_xmit_err;
 476                 tot->rx_dropped += rx.xdp_tx_err;
 477                 tot->tx_bytes += rx.xdp_bytes;
 478                 tot->tx_packets += rx.xdp_packets;
 479         }
 480         rcu_read_unlock();
 481 }
 482
 483 /* fake multicast ability */
 484 static void veth_set_multicast_list(struct net_device *dev)
 485 {
 486 }
 487
 488 static int veth_select_rxq(struct net_device *dev)
 489 {
 490         return smp_processor_id() % dev->real_num_rx_queues;
 491 }
 492
 493 static struct net_device *veth_peer_dev(struct net_device *dev)
 494 {
 495         struct veth_priv *priv = netdev_priv(dev);
 496
 497         /* Callers must be under RCU read side. */
 498         return rcu_dereference(priv->peer);
 499 }
 500
 501 static int veth_xdp_xmit(struct net_device *dev, int n,
 502                          struct xdp_frame **frames,
 503                          u32 flags, bool ndo_xmit)
 504 {
 505         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 506         int i, ret = -ENXIO, nxmit = 0;
 507         struct net_device *rcv;
 508         unsigned int max_len;
 509         struct veth_rq *rq;
 510
 511         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
 512                 return -EINVAL;
 513
 514         rcu_read_lock();
 515         rcv = rcu_dereference(priv->peer);
 516         if (unlikely(!rcv))
 517                 goto out;
 518
 519         rcv_priv = netdev_priv(rcv);
 520         rq = &rcv_priv->rq[veth_select_rxq(rcv)];
 521         /* The napi pointer is set if NAPI is enabled, which ensures that
 522          * xdp_ring is initialized on receive side and the peer device is up.
 523          */
 524         if (!rcu_access_pointer(rq->napi))
 525                 goto out;
 526
 527         max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN;
 528
 529         spin_lock(&rq->xdp_ring.producer_lock);
 530         for (i = 0; i < n; i++) {
 531                 struct xdp_frame *frame = frames[i];
 532                 void *ptr = veth_xdp_to_ptr(frame);
 533
 534                 if (unlikely(xdp_get_frame_len(frame) > max_len ||
 535                              __ptr_ring_produce(&rq->xdp_ring, ptr)))
 536                         break;
 537                 nxmit++;
 538         }
 539         spin_unlock(&rq->xdp_ring.producer_lock);
 540
 541         if (flags & XDP_XMIT_FLUSH)
 542                 __veth_xdp_flush(rq);
 543
 544         ret = nxmit;
 545         if (ndo_xmit) {
 546                 u64_stats_update_begin(&rq->stats.syncp);
 547                 rq->stats.vs.peer_tq_xdp_xmit += nxmit;
 548                 rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit;
 549                 u64_stats_update_end(&rq->stats.syncp);
 550         }
 551
 552 out:
 553         rcu_read_unlock();
 554
 555         return ret;
 556 }
 557
 558 static int veth_ndo_xdp_xmit(struct net_device *dev, int n,
 559                              struct xdp_frame **frames, u32 flags)
 560 {
 561         int err;
 562
 563         err = veth_xdp_xmit(dev, n, frames, flags, true);
 564         if (err < 0) {
 565                 struct veth_priv *priv = netdev_priv(dev);
 566
 567                 atomic64_add(n, &priv->dropped);
 568         }
 569
 570         return err;
 571 }
 572
 573 static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
 574 {
 575         int sent, i, err = 0, drops;
 576
 577         sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false);
 578         if (sent < 0) {
 579                 err = sent;
 580                 sent = 0;
 581         }
 582
 583         for (i = sent; unlikely(i < bq->count); i++)
 584                 xdp_return_frame(bq->q[i]);
 585
 586         drops = bq->count - sent;
 587         trace_xdp_bulk_tx(rq->dev, sent, drops, err);
 588
 589         u64_stats_update_begin(&rq->stats.syncp);
 590         rq->stats.vs.xdp_tx += sent;
 591         rq->stats.vs.xdp_tx_err += drops;
 592         u64_stats_update_end(&rq->stats.syncp);
 593
 594         bq->count = 0;
 595 }
 596
 597 static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
 598 {
 599         struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev);
 600         struct net_device *rcv;
 601         struct veth_rq *rcv_rq;
 602
 603         rcu_read_lock();
 604         veth_xdp_flush_bq(rq, bq);
 605         rcv = rcu_dereference(priv->peer);
 606         if (unlikely(!rcv))
 607                 goto out;
 608
 609         rcv_priv = netdev_priv(rcv);
 610         rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)];
 611         /* xdp_ring is initialized on receive side? */
 612         if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog)))
 613                 goto out;
 614
 615         __veth_xdp_flush(rcv_rq);
 616 out:
 617         rcu_read_unlock();
 618 }
 619
 620 static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp,
 621                        struct veth_xdp_tx_bq *bq)
 622 {
 623         struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp);
 624
 625         if (unlikely(!frame))
 626                 return -EOVERFLOW;
 627
 628         if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE))
 629                 veth_xdp_flush_bq(rq, bq);
 630
 631         bq->q[bq->count++] = frame;
 632
 633         return 0;
 634 }
 635
 636 static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq,
 637                                           struct xdp_frame *frame,
 638                                           struct veth_xdp_tx_bq *bq,
 639                                           struct veth_stats *stats)
 640 {
 641         struct xdp_frame orig_frame;
 642         struct bpf_prog *xdp_prog;
 643
 644         rcu_read_lock();
 645         xdp_prog = rcu_dereference(rq->xdp_prog);
 646         if (likely(xdp_prog)) {
 647                 struct veth_xdp_buff vxbuf;
 648                 struct xdp_buff *xdp = &vxbuf.xdp;
 649                 u32 act;
 650
 651                 xdp_convert_frame_to_buff(frame, xdp);
 652                 xdp->rxq = &rq->xdp_rxq;
 653                 vxbuf.skb = NULL;
 654
 655                 act = bpf_prog_run_xdp(xdp_prog, xdp);
 656
 657                 switch (act) {
 658                 case XDP_PASS:
 659                         if (xdp_update_frame_from_buff(xdp, frame))
 660                                 goto err_xdp;
 661                         break;
 662                 case XDP_TX:
 663                         orig_frame = *frame;
 664                         xdp->rxq->mem.type = frame->mem_type;
 665                         if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) {
 666                                 trace_xdp_exception(rq->dev, xdp_prog, act);
 667                                 frame = &orig_frame;
 668                                 stats->rx_drops++;
 669                                 goto err_xdp;
 670                         }
 671                         stats->xdp_tx++;
 672                         rcu_read_unlock();
 673                         goto xdp_xmit;
 674                 case XDP_REDIRECT:
 675                         orig_frame = *frame;
 676                         xdp->rxq->mem.type = frame->mem_type;
 677                         if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) {
 678                                 frame = &orig_frame;
 679                                 stats->rx_drops++;
 680                                 goto err_xdp;
 681                         }
 682                         stats->xdp_redirect++;
 683                         rcu_read_unlock();
 684                         goto xdp_xmit;
 685                 default:
 686                         bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
 687                         fallthrough;
 688                 case XDP_ABORTED:
 689                         trace_xdp_exception(rq->dev, xdp_prog, act);
 690                         fallthrough;
 691                 case XDP_DROP:
 692                         stats->xdp_drops++;
 693                         goto err_xdp;
 694                 }
 695         }
 696         rcu_read_unlock();
 697
 698         return frame;
 699 err_xdp:
 700         rcu_read_unlock();
 701         xdp_return_frame(frame);
 702 xdp_xmit:
 703         return NULL;
 704 }
 705
 706 /* frames array contains VETH_XDP_BATCH at most */
 707 static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames,
 708                                   int n_xdpf, struct veth_xdp_tx_bq *bq,
 709                                   struct veth_stats *stats)
 710 {
 711         void *skbs[VETH_XDP_BATCH];
 712         int i;
 713
 714         if (unlikely(!napi_skb_cache_get_bulk(skbs, n_xdpf))) {
 715                 for (i = 0; i < n_xdpf; i++)
 716                         xdp_return_frame(frames[i]);
 717                 stats->rx_drops += n_xdpf;
 718
 719                 return;
 720         }
 721
 722         for (i = 0; i < n_xdpf; i++) {
 723                 struct sk_buff *skb = skbs[i];
 724
 725                 skb = __xdp_build_skb_from_frame(frames[i], skb,
 726                                                  rq->dev);
 727                 if (!skb) {
 728                         xdp_return_frame(frames[i]);
 729                         stats->rx_drops++;
 730                         continue;
 731                 }
 732                 napi_gro_receive(&rq->xdp_napi, skb);
 733         }
 734 }
 735
 736 static void veth_xdp_get(struct xdp_buff *xdp)
 737 {
 738         struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
 739         int i;
 740
 741         get_page(virt_to_page(xdp->data));
 742         if (likely(!xdp_buff_has_frags(xdp)))
 743                 return;
 744
 745         for (i = 0; i < sinfo->nr_frags; i++)
 746                 __skb_frag_ref(&sinfo->frags[i]);
 747 }
 748
 749 static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
 750                                         struct xdp_buff *xdp,
 751                                         struct sk_buff **pskb)
 752 {
 753         struct sk_buff *skb = *pskb;
 754         u32 frame_sz;
 755
 756         if (skb_shared(skb) || skb_head_is_locked(skb) ||
 757             skb_shinfo(skb)->nr_frags ||
 758             skb_headroom(skb) < XDP_PACKET_HEADROOM) {
 759                 if (skb_pp_cow_data(rq->page_pool, pskb, XDP_PACKET_HEADROOM))
 760                         goto drop;
 761
 762                 skb = *pskb;
 763         }
 764
 765         /* SKB "head" area always have tailroom for skb_shared_info */
 766         frame_sz = skb_end_pointer(skb) - skb->head;
 767         frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 768         xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
 769         xdp_prepare_buff(xdp, skb->head, skb_headroom(skb),
 770                          skb_headlen(skb), true);
 771
 772         if (skb_is_nonlinear(skb)) {
 773                 skb_shinfo(skb)->xdp_frags_size = skb->data_len;
 774                 xdp_buff_set_frags_flag(xdp);
 775         } else {
 776                 xdp_buff_clear_frags_flag(xdp);
 777         }
 778         *pskb = skb;
 779
 780         return 0;
 781 drop:
 782         consume_skb(skb);
 783         *pskb = NULL;
 784
 785         return -ENOMEM;
 786 }
 787
 788 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq,
 789                                         struct sk_buff *skb,
 790                                         struct veth_xdp_tx_bq *bq,
 791                                         struct veth_stats *stats)
 792 {
 793         void *orig_data, *orig_data_end;
 794         struct bpf_prog *xdp_prog;
 795         struct veth_xdp_buff vxbuf;
 796         struct xdp_buff *xdp = &vxbuf.xdp;
 797         u32 act, metalen;
 798         int off;
 799
 800         skb_prepare_for_gro(skb);
 801
 802         rcu_read_lock();
 803         xdp_prog = rcu_dereference(rq->xdp_prog);
 804         if (unlikely(!xdp_prog)) {
 805                 rcu_read_unlock();
 806                 goto out;
 807         }
 808
 809         __skb_push(skb, skb->data - skb_mac_header(skb));
 810         if (veth_convert_skb_to_xdp_buff(rq, xdp, &skb))
 811                 goto drop;
 812         vxbuf.skb = skb;
 813
 814         orig_data = xdp->data;
 815         orig_data_end = xdp->data_end;
 816
 817         act = bpf_prog_run_xdp(xdp_prog, xdp);
 818
 819         switch (act) {
 820         case XDP_PASS:
 821                 break;
 822         case XDP_TX:
 823                 veth_xdp_get(xdp);
 824                 consume_skb(skb);
 825                 xdp->rxq->mem = rq->xdp_mem;
 826                 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) {
 827                         trace_xdp_exception(rq->dev, xdp_prog, act);
 828                         stats->rx_drops++;
 829                         goto err_xdp;
 830                 }
 831                 stats->xdp_tx++;
 832                 rcu_read_unlock();
 833                 goto xdp_xmit;
 834         case XDP_REDIRECT:
 835                 veth_xdp_get(xdp);
 836                 consume_skb(skb);
 837                 xdp->rxq->mem = rq->xdp_mem;
 838                 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) {
 839                         stats->rx_drops++;
 840                         goto err_xdp;
 841                 }
 842                 stats->xdp_redirect++;
 843                 rcu_read_unlock();
 844                 goto xdp_xmit;
 845         default:
 846                 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
 847                 fallthrough;
 848         case XDP_ABORTED:
 849                 trace_xdp_exception(rq->dev, xdp_prog, act);
 850                 fallthrough;
 851         case XDP_DROP:
 852                 stats->xdp_drops++;
 853                 goto xdp_drop;
 854         }
 855         rcu_read_unlock();
 856
 857         /* check if bpf_xdp_adjust_head was used */
 858         off = orig_data - xdp->data;
 859         if (off > 0)
 860                 __skb_push(skb, off);
 861         else if (off < 0)
 862                 __skb_pull(skb, -off);
 863
 864         skb_reset_mac_header(skb);
 865
 866         /* check if bpf_xdp_adjust_tail was used */
 867         off = xdp->data_end - orig_data_end;
 868         if (off != 0)
 869                 __skb_put(skb, off); /* positive on grow, negative on shrink */
 870
 871         /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers
 872          * (e.g. bpf_xdp_adjust_tail), we need to update data_len here.
 873          */
 874         if (xdp_buff_has_frags(xdp))
 875                 skb->data_len = skb_shinfo(skb)->xdp_frags_size;
 876         else
 877                 skb->data_len = 0;
 878
 879         skb->protocol = eth_type_trans(skb, rq->dev);
 880
 881         metalen = xdp->data - xdp->data_meta;
 882         if (metalen)
 883                 skb_metadata_set(skb, metalen);
 884 out:
 885         return skb;
 886 drop:
 887         stats->rx_drops++;
 888 xdp_drop:
 889         rcu_read_unlock();
 890         kfree_skb(skb);
 891         return NULL;
 892 err_xdp:
 893         rcu_read_unlock();
 894         xdp_return_buff(xdp);
 895 xdp_xmit:
 896         return NULL;
 897 }
 898
 899 static int veth_xdp_rcv(struct veth_rq *rq, int budget,
 900                         struct veth_xdp_tx_bq *bq,
 901                         struct veth_stats *stats)
 902 {
 903         struct veth_priv *priv = netdev_priv(rq->dev);
 904         int queue_idx = rq->xdp_rxq.queue_index;
 905         struct netdev_queue *peer_txq;
 906         struct net_device *peer_dev;
 907         int i, done = 0, n_xdpf = 0;
 908         void *xdpf[VETH_XDP_BATCH];
 909
 910         /* NAPI functions as RCU section */
 911         peer_dev = rcu_dereference_check(priv->peer, rcu_read_lock_bh_held());
 912         peer_txq = peer_dev ? netdev_get_tx_queue(peer_dev, queue_idx) : NULL;
 913
 914         for (i = 0; i < budget; i++) {
 915                 void *ptr = __ptr_ring_consume(&rq->xdp_ring);
 916
 917                 if (!ptr)
 918                         break;
 919
 920                 if (veth_is_xdp_frame(ptr)) {
 921                         /* ndo_xdp_xmit */
 922                         struct xdp_frame *frame = veth_ptr_to_xdp(ptr);
 923
 924                         stats->xdp_bytes += xdp_get_frame_len(frame);
 925                         frame = veth_xdp_rcv_one(rq, frame, bq, stats);
 926                         if (frame) {
 927                                 /* XDP_PASS */
 928                                 xdpf[n_xdpf++] = frame;
 929                                 if (n_xdpf == VETH_XDP_BATCH) {
 930                                         veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf,
 931                                                               bq, stats);
 932                                         n_xdpf = 0;
 933                                 }
 934                         }
 935                 } else {
 936                         /* ndo_start_xmit */
 937                         struct sk_buff *skb = ptr;
 938
 939                         stats->xdp_bytes += skb->len;
 940                         skb = veth_xdp_rcv_skb(rq, skb, bq, stats);
 941                         if (skb) {
 942                                 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC))
 943                                         netif_receive_skb(skb);
 944                                 else
 945                                         napi_gro_receive(&rq->xdp_napi, skb);
 946                         }
 947                 }
 948                 done++;
 949         }
 950
 951         if (n_xdpf)
 952                 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats);
 953
 954         u64_stats_update_begin(&rq->stats.syncp);
 955         rq->stats.vs.xdp_redirect += stats->xdp_redirect;
 956         rq->stats.vs.xdp_bytes += stats->xdp_bytes;
 957         rq->stats.vs.xdp_drops += stats->xdp_drops;
 958         rq->stats.vs.rx_drops += stats->rx_drops;
 959         rq->stats.vs.xdp_packets += done;
 960         u64_stats_update_end(&rq->stats.syncp);
 961
 962         if (peer_txq && unlikely(netif_tx_queue_stopped(peer_txq)))
 963                 netif_tx_wake_queue(peer_txq);
 964
 965         return done;
 966 }
 967
 968 static int veth_poll(struct napi_struct *napi, int budget)
 969 {
 970         struct veth_rq *rq =
 971                 container_of(napi, struct veth_rq, xdp_napi);
 972         struct veth_stats stats = {};
 973         struct veth_xdp_tx_bq bq;
 974         int done;
 975
 976         bq.count = 0;
 977
 978         xdp_set_return_frame_no_direct();
 979         done = veth_xdp_rcv(rq, budget, &bq, &stats);
 980
 981         if (stats.xdp_redirect > 0)
 982                 xdp_do_flush();
 983
 984         if (done < budget && napi_complete_done(napi, done)) {
 985                 /* Write rx_notify_masked before reading ptr_ring */
 986                 smp_store_mb(rq->rx_notify_masked, false);
 987                 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
 988                         if (napi_schedule_prep(&rq->xdp_napi)) {
 989                                 WRITE_ONCE(rq->rx_notify_masked, true);
 990                                 __napi_schedule(&rq->xdp_napi);
 991                         }
 992                 }
 993         }
 994
 995         if (stats.xdp_tx > 0)
 996                 veth_xdp_flush(rq, &bq);
 997         xdp_clear_return_frame_no_direct();
 998
 999         return done;
1000 }
1001
1002 static int veth_create_page_pool(struct veth_rq *rq)
1003 {
1004         struct page_pool_params pp_params = {
1005                 .order = 0,
1006                 .pool_size = VETH_RING_SIZE,
1007                 .nid = NUMA_NO_NODE,
1008                 .dev = &rq->dev->dev,
1009         };
1010
1011         rq->page_pool = page_pool_create(&pp_params);
1012         if (IS_ERR(rq->page_pool)) {
1013                 int err = PTR_ERR(rq->page_pool);
1014
1015                 rq->page_pool = NULL;
1016                 return err;
1017         }
1018
1019         return 0;
1020 }
1021
1022 static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
1023 {
1024         struct veth_priv *priv = netdev_priv(dev);
1025         int err, i;
1026
1027         for (i = start; i < end; i++) {
1028                 err = veth_create_page_pool(&priv->rq[i]);
1029                 if (err)
1030                         goto err_page_pool;
1031         }
1032
1033         for (i = start; i < end; i++) {
1034                 struct veth_rq *rq = &priv->rq[i];
1035
1036                 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
1037                 if (err)
1038                         goto err_xdp_ring;
1039         }
1040
1041         for (i = start; i < end; i++) {
1042                 struct veth_rq *rq = &priv->rq[i];
1043
1044                 napi_enable(&rq->xdp_napi);
1045                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1046         }
1047
1048         return 0;
1049
1050 err_xdp_ring:
1051         for (i--; i >= start; i--)
1052                 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
1053         i = end;
1054 err_page_pool:
1055         for (i--; i >= start; i--) {
1056                 page_pool_destroy(priv->rq[i].page_pool);
1057                 priv->rq[i].page_pool = NULL;
1058         }
1059
1060         return err;
1061 }
1062
1063 static int __veth_napi_enable(struct net_device *dev)
1064 {
1065         return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1066 }
1067
1068 static void veth_napi_del_range(struct net_device *dev, int start, int end)
1069 {
1070         struct veth_priv *priv = netdev_priv(dev);
1071         int i;
1072
1073         for (i = start; i < end; i++) {
1074                 struct veth_rq *rq = &priv->rq[i];
1075
1076                 rcu_assign_pointer(priv->rq[i].napi, NULL);
1077                 napi_disable(&rq->xdp_napi);
1078                 __netif_napi_del(&rq->xdp_napi);
1079         }
1080         synchronize_net();
1081
1082         for (i = start; i < end; i++) {
1083                 struct veth_rq *rq = &priv->rq[i];
1084
1085                 rq->rx_notify_masked = false;
1086                 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
1087         }
1088
1089         for (i = start; i < end; i++) {
1090                 page_pool_destroy(priv->rq[i].page_pool);
1091                 priv->rq[i].page_pool = NULL;
1092         }
1093 }
1094
1095 static void veth_napi_del(struct net_device *dev)
1096 {
1097         veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
1098 }
1099
1100 static bool veth_gro_requested(const struct net_device *dev)
1101 {
1102         return !!(dev->wanted_features & NETIF_F_GRO);
1103 }
1104
1105 static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1106                                  bool napi_already_on)
1107 {
1108         struct veth_priv *priv = netdev_priv(dev);
1109         int err, i;
1110
1111         for (i = start; i < end; i++) {
1112                 struct veth_rq *rq = &priv->rq[i];
1113
1114                 if (!napi_already_on)
1115                         netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1116                 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1117                 if (err < 0)
1118                         goto err_rxq_reg;
1119
1120                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1121                                                  MEM_TYPE_PAGE_SHARED,
1122                                                  NULL);
1123                 if (err < 0)
1124                         goto err_reg_mem;
1125
1126                 /* Save original mem info as it can be overwritten */
1127                 rq->xdp_mem = rq->xdp_rxq.mem;
1128         }
1129         return 0;
1130
1131 err_reg_mem:
1132         xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1133 err_rxq_reg:
1134         for (i--; i >= start; i--) {
1135                 struct veth_rq *rq = &priv->rq[i];
1136
1137                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1138                 if (!napi_already_on)
1139                         netif_napi_del(&rq->xdp_napi);
1140         }
1141
1142         return err;
1143 }
1144
1145 static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1146                                    bool delete_napi)
1147 {
1148         struct veth_priv *priv = netdev_priv(dev);
1149         int i;
1150
1151         for (i = start; i < end; i++) {
1152                 struct veth_rq *rq = &priv->rq[i];
1153
1154                 rq->xdp_rxq.mem = rq->xdp_mem;
1155                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1156
1157                 if (delete_napi)
1158                         netif_napi_del(&rq->xdp_napi);
1159         }
1160 }
1161
1162 static int veth_enable_xdp(struct net_device *dev)
1163 {
1164         bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1165         struct veth_priv *priv = netdev_priv(dev);
1166         int err, i;
1167
1168         if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1169                 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1170                 if (err)
1171                         return err;
1172
1173                 if (!napi_already_on) {
1174                         err = __veth_napi_enable(dev);
1175                         if (err) {
1176                                 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1177                                 return err;
1178                         }
1179                 }
1180         }
1181
1182         for (i = 0; i < dev->real_num_rx_queues; i++) {
1183                 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1184                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1185         }
1186
1187         return 0;
1188 }
1189
1190 static void veth_disable_xdp(struct net_device *dev)
1191 {
1192         struct veth_priv *priv = netdev_priv(dev);
1193         int i;
1194
1195         for (i = 0; i < dev->real_num_rx_queues; i++)
1196                 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1197
1198         if (!netif_running(dev) || !veth_gro_requested(dev))
1199                 veth_napi_del(dev);
1200
1201         veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1202 }
1203
1204 static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1205 {
1206         struct veth_priv *priv = netdev_priv(dev);
1207         int err, i;
1208
1209         for (i = start; i < end; i++) {
1210                 struct veth_rq *rq = &priv->rq[i];
1211
1212                 netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1213         }
1214
1215         err = __veth_napi_enable_range(dev, start, end);
1216         if (err) {
1217                 for (i = start; i < end; i++) {
1218                         struct veth_rq *rq = &priv->rq[i];
1219
1220                         netif_napi_del(&rq->xdp_napi);
1221                 }
1222                 return err;
1223         }
1224         return err;
1225 }
1226
1227 static int veth_napi_enable(struct net_device *dev)
1228 {
1229         return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1230 }
1231
1232 static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1233 {
1234         struct veth_priv *priv = netdev_priv(dev);
1235
1236         if (start >= end)
1237                 return;
1238
1239         if (priv->_xdp_prog) {
1240                 veth_napi_del_range(dev, start, end);
1241                 veth_disable_xdp_range(dev, start, end, false);
1242         } else if (veth_gro_requested(dev)) {
1243                 veth_napi_del_range(dev, start, end);
1244         }
1245 }
1246
1247 static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1248 {
1249         struct veth_priv *priv = netdev_priv(dev);
1250         int err;
1251
1252         if (start >= end)
1253                 return 0;
1254
1255         if (priv->_xdp_prog) {
1256                 /* these channels are freshly initialized, napi is not on there even
1257                  * when GRO is requeste
1258                  */
1259                 err = veth_enable_xdp_range(dev, start, end, false);
1260                 if (err)
1261                         return err;
1262
1263                 err = __veth_napi_enable_range(dev, start, end);
1264                 if (err) {
1265                         /* on error always delete the newly added napis */
1266                         veth_disable_xdp_range(dev, start, end, true);
1267                         return err;
1268                 }
1269         } else if (veth_gro_requested(dev)) {
1270                 return veth_napi_enable_range(dev, start, end);
1271         }
1272         return 0;
1273 }
1274
1275 static void veth_set_xdp_features(struct net_device *dev)
1276 {
1277         struct veth_priv *priv = netdev_priv(dev);
1278         struct net_device *peer;
1279
1280         peer = rtnl_dereference(priv->peer);
1281         if (peer && peer->real_num_tx_queues <= dev->real_num_rx_queues) {
1282                 struct veth_priv *priv_peer = netdev_priv(peer);
1283                 xdp_features_t val = NETDEV_XDP_ACT_BASIC |
1284                                      NETDEV_XDP_ACT_REDIRECT |
1285                                      NETDEV_XDP_ACT_RX_SG;
1286
1287                 if (priv_peer->_xdp_prog || veth_gro_requested(peer))
1288                         val |= NETDEV_XDP_ACT_NDO_XMIT |
1289                                NETDEV_XDP_ACT_NDO_XMIT_SG;
1290                 xdp_set_features_flag(dev, val);
1291         } else {
1292                 xdp_clear_features_flag(dev);
1293         }
1294 }
1295
1296 static int veth_set_channels(struct net_device *dev,
1297                              struct ethtool_channels *ch)
1298 {
1299         struct veth_priv *priv = netdev_priv(dev);
1300         unsigned int old_rx_count, new_rx_count;
1301         struct veth_priv *peer_priv;
1302         struct net_device *peer;
1303         int err;
1304
1305         /* sanity check. Upper bounds are already enforced by the caller */
1306         if (!ch->rx_count || !ch->tx_count)
1307                 return -EINVAL;
1308
1309         /* avoid braking XDP, if that is enabled */
1310         peer = rtnl_dereference(priv->peer);
1311         peer_priv = peer ? netdev_priv(peer) : NULL;
1312         if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1313                 return -EINVAL;
1314
1315         if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1316                 return -EINVAL;
1317
1318         old_rx_count = dev->real_num_rx_queues;
1319         new_rx_count = ch->rx_count;
1320         if (netif_running(dev)) {
1321                 /* turn device off */
1322                 netif_carrier_off(dev);
1323                 if (peer)
1324                         netif_carrier_off(peer);
1325
1326                 /* try to allocate new resurces, as needed*/
1327                 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1328                 if (err)
1329                         goto out;
1330         }
1331
1332         err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1333         if (err)
1334                 goto revert;
1335
1336         err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1337         if (err) {
1338                 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1339
1340                 /* this error condition could happen only if rx and tx change
1341                  * in opposite directions (e.g. tx nr raises, rx nr decreases)
1342                  * and we can't do anything to fully restore the original
1343                  * status
1344                  */
1345                 if (err2)
1346                         pr_warn("Can't restore rx queues config %d -> %d %d",
1347                                 new_rx_count, old_rx_count, err2);
1348                 else
1349                         goto revert;
1350         }
1351
1352 out:
1353         if (netif_running(dev)) {
1354                 /* note that we need to swap the arguments WRT the enable part
1355                  * to identify the range we have to disable
1356                  */
1357                 veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1358                 netif_carrier_on(dev);
1359                 if (peer)
1360                         netif_carrier_on(peer);
1361         }
1362
1363         /* update XDP supported features */
1364         veth_set_xdp_features(dev);
1365         if (peer)
1366                 veth_set_xdp_features(peer);
1367
1368         return err;
1369
1370 revert:
1371         new_rx_count = old_rx_count;
1372         old_rx_count = ch->rx_count;
1373         goto out;
1374 }
1375
1376 static int veth_open(struct net_device *dev)
1377 {
1378         struct veth_priv *priv = netdev_priv(dev);
1379         struct net_device *peer = rtnl_dereference(priv->peer);
1380         int err;
1381
1382         if (!peer)
1383                 return -ENOTCONN;
1384
1385         if (priv->_xdp_prog) {
1386                 err = veth_enable_xdp(dev);
1387                 if (err)
1388                         return err;
1389         } else if (veth_gro_requested(dev)) {
1390                 err = veth_napi_enable(dev);
1391                 if (err)
1392                         return err;
1393         }
1394
1395         if (peer->flags & IFF_UP) {
1396                 netif_carrier_on(dev);
1397                 netif_carrier_on(peer);
1398         }
1399
1400         veth_set_xdp_features(dev);
1401
1402         return 0;
1403 }
1404
1405 static int veth_close(struct net_device *dev)
1406 {
1407         struct veth_priv *priv = netdev_priv(dev);
1408         struct net_device *peer = rtnl_dereference(priv->peer);
1409
1410         netif_carrier_off(dev);
1411         if (peer)
1412                 netif_carrier_off(peer);
1413
1414         if (priv->_xdp_prog)
1415                 veth_disable_xdp(dev);
1416         else if (veth_gro_requested(dev))
1417                 veth_napi_del(dev);
1418
1419         return 0;
1420 }
1421
1422 static int is_valid_veth_mtu(int mtu)
1423 {
1424         return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1425 }
1426
1427 static int veth_alloc_queues(struct net_device *dev)
1428 {
1429         struct veth_priv *priv = netdev_priv(dev);
1430         int i;
1431
1432         priv->rq = kvcalloc(dev->num_rx_queues, sizeof(*priv->rq),
1433                             GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
1434         if (!priv->rq)
1435                 return -ENOMEM;
1436
1437         for (i = 0; i < dev->num_rx_queues; i++) {
1438                 priv->rq[i].dev = dev;
1439                 u64_stats_init(&priv->rq[i].stats.syncp);
1440         }
1441
1442         return 0;
1443 }
1444
1445 static void veth_free_queues(struct net_device *dev)
1446 {
1447         struct veth_priv *priv = netdev_priv(dev);
1448
1449         kvfree(priv->rq);
1450 }
1451
1452 static int veth_dev_init(struct net_device *dev)
1453 {
1454         netdev_lockdep_set_classes(dev);
1455         return veth_alloc_queues(dev);
1456 }
1457
1458 static void veth_dev_free(struct net_device *dev)
1459 {
1460         veth_free_queues(dev);
1461 }
1462
1463 #ifdef CONFIG_NET_POLL_CONTROLLER
1464 static void veth_poll_controller(struct net_device *dev)
1465 {
1466         /* veth only receives frames when its peer sends one
1467          * Since it has nothing to do with disabling irqs, we are guaranteed
1468          * never to have pending data when we poll for it so
1469          * there is nothing to do here.
1470          *
1471          * We need this though so netpoll recognizes us as an interface that
1472          * supports polling, which enables bridge devices in virt setups to
1473          * still use netconsole
1474          */
1475 }
1476 #endif  /* CONFIG_NET_POLL_CONTROLLER */
1477
1478 static int veth_get_iflink(const struct net_device *dev)
1479 {
1480         struct veth_priv *priv = netdev_priv(dev);
1481         struct net_device *peer;
1482         int iflink;
1483
1484         rcu_read_lock();
1485         peer = rcu_dereference(priv->peer);
1486         iflink = peer ? READ_ONCE(peer->ifindex) : 0;
1487         rcu_read_unlock();
1488
1489         return iflink;
1490 }
1491
1492 static netdev_features_t veth_fix_features(struct net_device *dev,
1493                                            netdev_features_t features)
1494 {
1495         struct veth_priv *priv = netdev_priv(dev);
1496         struct net_device *peer;
1497
1498         peer = rtnl_dereference(priv->peer);
1499         if (peer) {
1500                 struct veth_priv *peer_priv = netdev_priv(peer);
1501
1502                 if (peer_priv->_xdp_prog)
1503                         features &= ~NETIF_F_GSO_SOFTWARE;
1504         }
1505
1506         return features;
1507 }
1508
1509 static int veth_set_features(struct net_device *dev,
1510                              netdev_features_t features)
1511 {
1512         netdev_features_t changed = features ^ dev->features;
1513         struct veth_priv *priv = netdev_priv(dev);
1514         struct net_device *peer;
1515         int err;
1516
1517         if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1518                 return 0;
1519
1520         peer = rtnl_dereference(priv->peer);
1521         if (features & NETIF_F_GRO) {
1522                 err = veth_napi_enable(dev);
1523                 if (err)
1524                         return err;
1525
1526                 if (peer)
1527                         xdp_features_set_redirect_target(peer, true);
1528         } else {
1529                 if (peer)
1530                         xdp_features_clear_redirect_target(peer);
1531                 veth_napi_del(dev);
1532         }
1533         return 0;
1534 }
1535
1536 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1537 {
1538         struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1539         struct net_device *peer;
1540
1541         if (new_hr < 0)
1542                 new_hr = 0;
1543
1544         rcu_read_lock();
1545         peer = rcu_dereference(priv->peer);
1546         if (unlikely(!peer))
1547                 goto out;
1548
1549         peer_priv = netdev_priv(peer);
1550         priv->requested_headroom = new_hr;
1551         new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1552         dev->needed_headroom = new_hr;
1553         peer->needed_headroom = new_hr;
1554
1555 out:
1556         rcu_read_unlock();
1557 }
1558
1559 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1560                         struct netlink_ext_ack *extack)
1561 {
1562         struct veth_priv *priv = netdev_priv(dev);
1563         struct bpf_prog *old_prog;
1564         struct net_device *peer;
1565         unsigned int max_mtu;
1566         int err;
1567
1568         old_prog = priv->_xdp_prog;
1569         priv->_xdp_prog = prog;
1570         peer = rtnl_dereference(priv->peer);
1571
1572         if (prog) {
1573                 if (!peer) {
1574                         NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1575                         err = -ENOTCONN;
1576                         goto err;
1577                 }
1578
1579                 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) -
1580                           peer->hard_header_len;
1581                 /* Allow increasing the max_mtu if the program supports
1582                  * XDP fragments.
1583                  */
1584                 if (prog->aux->xdp_has_frags)
1585                         max_mtu += PAGE_SIZE * MAX_SKB_FRAGS;
1586
1587                 if (peer->mtu > max_mtu) {
1588                         NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1589                         err = -ERANGE;
1590                         goto err;
1591                 }
1592
1593                 if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1594                         NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1595                         err = -ENOSPC;
1596                         goto err;
1597                 }
1598
1599                 if (dev->flags & IFF_UP) {
1600                         err = veth_enable_xdp(dev);
1601                         if (err) {
1602                                 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1603                                 goto err;
1604                         }
1605                 }
1606
1607                 if (!old_prog) {
1608                         peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1609                         peer->max_mtu = max_mtu;
1610                 }
1611
1612                 xdp_features_set_redirect_target(peer, true);
1613         }
1614
1615         if (old_prog) {
1616                 if (!prog) {
1617                         if (peer && !veth_gro_requested(dev))
1618                                 xdp_features_clear_redirect_target(peer);
1619
1620                         if (dev->flags & IFF_UP)
1621                                 veth_disable_xdp(dev);
1622
1623                         if (peer) {
1624                                 peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1625                                 peer->max_mtu = ETH_MAX_MTU;
1626                         }
1627                 }
1628                 bpf_prog_put(old_prog);
1629         }
1630
1631         if ((!!old_prog ^ !!prog) && peer)
1632                 netdev_update_features(peer);
1633
1634         return 0;
1635 err:
1636         priv->_xdp_prog = old_prog;
1637
1638         return err;
1639 }
1640
1641 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1642 {
1643         switch (xdp->command) {
1644         case XDP_SETUP_PROG:
1645                 return veth_xdp_set(dev, xdp->prog, xdp->extack);
1646         default:
1647                 return -EINVAL;
1648         }
1649 }
1650
1651 static int veth_xdp_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp)
1652 {
1653         struct veth_xdp_buff *_ctx = (void *)ctx;
1654
1655         if (!_ctx->skb)
1656                 return -ENODATA;
1657
1658         *timestamp = skb_hwtstamps(_ctx->skb)->hwtstamp;
1659         return 0;
1660 }
1661
1662 static int veth_xdp_rx_hash(const struct xdp_md *ctx, u32 *hash,
1663                             enum xdp_rss_hash_type *rss_type)
1664 {
1665         struct veth_xdp_buff *_ctx = (void *)ctx;
1666         struct sk_buff *skb = _ctx->skb;
1667
1668         if (!skb)
1669                 return -ENODATA;
1670
1671         *hash = skb_get_hash(skb);
1672         *rss_type = skb->l4_hash ? XDP_RSS_TYPE_L4_ANY : XDP_RSS_TYPE_NONE;
1673
1674         return 0;
1675 }
1676
1677 static int veth_xdp_rx_vlan_tag(const struct xdp_md *ctx, __be16 *vlan_proto,
1678                                 u16 *vlan_tci)
1679 {
1680         const struct veth_xdp_buff *_ctx = (void *)ctx;
1681         const struct sk_buff *skb = _ctx->skb;
1682         int err;
1683
1684         if (!skb)
1685                 return -ENODATA;
1686
1687         err = __vlan_hwaccel_get_tag(skb, vlan_tci);
1688         if (err)
1689                 return err;
1690
1691         *vlan_proto = skb->vlan_proto;
1692         return err;
1693 }
1694
1695 static const struct net_device_ops veth_netdev_ops = {
1696         .ndo_init            = veth_dev_init,
1697         .ndo_open            = veth_open,
1698         .ndo_stop            = veth_close,
1699         .ndo_start_xmit      = veth_xmit,
1700         .ndo_get_stats64     = veth_get_stats64,
1701         .ndo_set_rx_mode     = veth_set_multicast_list,
1702         .ndo_set_mac_address = eth_mac_addr,
1703 #ifdef CONFIG_NET_POLL_CONTROLLER
1704         .ndo_poll_controller    = veth_poll_controller,
1705 #endif
1706         .ndo_get_iflink         = veth_get_iflink,
1707         .ndo_fix_features       = veth_fix_features,
1708         .ndo_set_features       = veth_set_features,
1709         .ndo_features_check     = passthru_features_check,
1710         .ndo_set_rx_headroom    = veth_set_rx_headroom,
1711         .ndo_bpf                = veth_xdp,
1712         .ndo_xdp_xmit           = veth_ndo_xdp_xmit,
1713         .ndo_get_peer_dev       = veth_peer_dev,
1714 };
1715
1716 static const struct xdp_metadata_ops veth_xdp_metadata_ops = {
1717         .xmo_rx_timestamp               = veth_xdp_rx_timestamp,
1718         .xmo_rx_hash                    = veth_xdp_rx_hash,
1719         .xmo_rx_vlan_tag                = veth_xdp_rx_vlan_tag,
1720 };
1721
1722 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1723                        NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1724                        NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1725                        NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1726                        NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1727
1728 static void veth_setup(struct net_device *dev)
1729 {
1730         ether_setup(dev);
1731
1732         dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1733         dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1734         dev->priv_flags |= IFF_NO_QUEUE;
1735         dev->priv_flags |= IFF_PHONY_HEADROOM;
1736         dev->priv_flags |= IFF_DISABLE_NETPOLL;
1737         dev->lltx = true;
1738
1739         dev->netdev_ops = &veth_netdev_ops;
1740         dev->xdp_metadata_ops = &veth_xdp_metadata_ops;
1741         dev->ethtool_ops = &veth_ethtool_ops;
1742         dev->features |= VETH_FEATURES;
1743         dev->vlan_features = dev->features &
1744                              ~(NETIF_F_HW_VLAN_CTAG_TX |
1745                                NETIF_F_HW_VLAN_STAG_TX |
1746                                NETIF_F_HW_VLAN_CTAG_RX |
1747                                NETIF_F_HW_VLAN_STAG_RX);
1748         dev->needs_free_netdev = true;
1749         dev->priv_destructor = veth_dev_free;
1750         dev->pcpu_stat_type = NETDEV_PCPU_STAT_TSTATS;
1751         dev->max_mtu = ETH_MAX_MTU;
1752
1753         dev->hw_features = VETH_FEATURES;
1754         dev->hw_enc_features = VETH_FEATURES;
1755         dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1756         netif_set_tso_max_size(dev, GSO_MAX_SIZE);
1757 }
1758
1759 /*
1760  * netlink interface
1761  */
1762
1763 static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1764                          struct netlink_ext_ack *extack)
1765 {
1766         if (tb[IFLA_ADDRESS]) {
1767                 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1768                         return -EINVAL;
1769                 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1770                         return -EADDRNOTAVAIL;
1771         }
1772         if (tb[IFLA_MTU]) {
1773                 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1774                         return -EINVAL;
1775         }
1776         return 0;
1777 }
1778
1779 static struct rtnl_link_ops veth_link_ops;
1780
1781 static void veth_disable_gro(struct net_device *dev)
1782 {
1783         dev->features &= ~NETIF_F_GRO;
1784         dev->wanted_features &= ~NETIF_F_GRO;
1785         netdev_update_features(dev);
1786 }
1787
1788 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1789 {
1790         int err;
1791
1792         if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1793                 err = netif_set_real_num_tx_queues(dev, 1);
1794                 if (err)
1795                         return err;
1796         }
1797         if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1798                 err = netif_set_real_num_rx_queues(dev, 1);
1799                 if (err)
1800                         return err;
1801         }
1802         return 0;
1803 }
1804
1805 static int veth_newlink(struct net_device *dev,
1806                         struct rtnl_newlink_params *params,
1807                         struct netlink_ext_ack *extack)
1808 {
1809         struct net *peer_net = rtnl_newlink_peer_net(params);
1810         struct nlattr **data = params->data;
1811         struct nlattr **tb = params->tb;
1812         int err;
1813         struct net_device *peer;
1814         struct veth_priv *priv;
1815         char ifname[IFNAMSIZ];
1816         struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1817         unsigned char name_assign_type;
1818         struct ifinfomsg *ifmp;
1819
1820         /*
1821          * create and register peer first
1822          */
1823         if (data && data[VETH_INFO_PEER]) {
1824                 struct nlattr *nla_peer = data[VETH_INFO_PEER];
1825
1826                 ifmp = nla_data(nla_peer);
1827                 rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack);
1828                 tbp = peer_tb;
1829         } else {
1830                 ifmp = NULL;
1831                 tbp = tb;
1832         }
1833
1834         if (ifmp && tbp[IFLA_IFNAME]) {
1835                 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1836                 name_assign_type = NET_NAME_USER;
1837         } else {
1838                 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1839                 name_assign_type = NET_NAME_ENUM;
1840         }
1841
1842         peer = rtnl_create_link(peer_net, ifname, name_assign_type,
1843                                 &veth_link_ops, tbp, extack);
1844         if (IS_ERR(peer))
1845                 return PTR_ERR(peer);
1846
1847         if (!ifmp || !tbp[IFLA_ADDRESS])
1848                 eth_hw_addr_random(peer);
1849
1850         if (ifmp && (dev->ifindex != 0))
1851                 peer->ifindex = ifmp->ifi_index;
1852
1853         netif_inherit_tso_max(peer, dev);
1854
1855         err = register_netdevice(peer);
1856         if (err < 0)
1857                 goto err_register_peer;
1858
1859         /* keep GRO disabled by default to be consistent with the established
1860          * veth behavior
1861          */
1862         veth_disable_gro(peer);
1863         netif_carrier_off(peer);
1864
1865         err = rtnl_configure_link(peer, ifmp, 0, NULL);
1866         if (err < 0)
1867                 goto err_configure_peer;
1868
1869         /*
1870          * register dev last
1871          *
1872          * note, that since we've registered new device the dev's name
1873          * should be re-allocated
1874          */
1875
1876         if (tb[IFLA_ADDRESS] == NULL)
1877                 eth_hw_addr_random(dev);
1878
1879         if (tb[IFLA_IFNAME])
1880                 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1881         else
1882                 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1883
1884         err = register_netdevice(dev);
1885         if (err < 0)
1886                 goto err_register_dev;
1887
1888         netif_carrier_off(dev);
1889
1890         /*
1891          * tie the deviced together
1892          */
1893
1894         priv = netdev_priv(dev);
1895         rcu_assign_pointer(priv->peer, peer);
1896         err = veth_init_queues(dev, tb);
1897         if (err)
1898                 goto err_queues;
1899
1900         priv = netdev_priv(peer);
1901         rcu_assign_pointer(priv->peer, dev);
1902         err = veth_init_queues(peer, tb);
1903         if (err)
1904                 goto err_queues;
1905
1906         veth_disable_gro(dev);
1907         /* update XDP supported features */
1908         veth_set_xdp_features(dev);
1909         veth_set_xdp_features(peer);
1910
1911         return 0;
1912
1913 err_queues:
1914         unregister_netdevice(dev);
1915 err_register_dev:
1916         /* nothing to do */
1917 err_configure_peer:
1918         unregister_netdevice(peer);
1919         return err;
1920
1921 err_register_peer:
1922         free_netdev(peer);
1923         return err;
1924 }
1925
1926 static void veth_dellink(struct net_device *dev, struct list_head *head)
1927 {
1928         struct veth_priv *priv;
1929         struct net_device *peer;
1930
1931         priv = netdev_priv(dev);
1932         peer = rtnl_dereference(priv->peer);
1933
1934         /* Note : dellink() is called from default_device_exit_batch(),
1935          * before a rcu_synchronize() point. The devices are guaranteed
1936          * not being freed before one RCU grace period.
1937          */
1938         RCU_INIT_POINTER(priv->peer, NULL);
1939         unregister_netdevice_queue(dev, head);
1940
1941         if (peer) {
1942                 priv = netdev_priv(peer);
1943                 RCU_INIT_POINTER(priv->peer, NULL);
1944                 unregister_netdevice_queue(peer, head);
1945         }
1946 }
1947
1948 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
1949         [VETH_INFO_PEER]        = { .len = sizeof(struct ifinfomsg) },
1950 };
1951
1952 static struct net *veth_get_link_net(const struct net_device *dev)
1953 {
1954         struct veth_priv *priv = netdev_priv(dev);
1955         struct net_device *peer = rtnl_dereference(priv->peer);
1956
1957         return peer ? dev_net(peer) : dev_net(dev);
1958 }
1959
1960 static unsigned int veth_get_num_queues(void)
1961 {
1962         /* enforce the same queue limit as rtnl_create_link */
1963         int queues = num_possible_cpus();
1964
1965         if (queues > 4096)
1966                 queues = 4096;
1967         return queues;
1968 }
1969
1970 static struct rtnl_link_ops veth_link_ops = {
1971         .kind           = DRV_NAME,
1972         .priv_size      = sizeof(struct veth_priv),
1973         .setup          = veth_setup,
1974         .validate       = veth_validate,
1975         .newlink        = veth_newlink,
1976         .dellink        = veth_dellink,
1977         .policy         = veth_policy,
1978         .peer_type      = VETH_INFO_PEER,
1979         .maxtype        = VETH_INFO_MAX,
1980         .get_link_net   = veth_get_link_net,
1981         .get_num_tx_queues      = veth_get_num_queues,
1982         .get_num_rx_queues      = veth_get_num_queues,
1983 };
1984
1985 /*
1986  * init/fini
1987  */
1988
1989 static __init int veth_init(void)
1990 {
1991         return rtnl_link_register(&veth_link_ops);
1992 }
1993
1994 static __exit void veth_exit(void)
1995 {
1996         rtnl_link_unregister(&veth_link_ops);
1997 }
1998
1999 module_init(veth_init);
2000 module_exit(veth_exit);
2001
2002 MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
2003 MODULE_LICENSE("GPL v2");
2004 MODULE_ALIAS_RTNL_LINK(DRV_NAME);