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         strscpy(info->driver, DRV_NAME, sizeof(info->driver));
 132         strscpy(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 (!READ_ONCE(rq->rx_notify_masked) &&
 269             napi_schedule_prep(&rq->xdp_napi)) {
 270                 WRITE_ONCE(rq->rx_notify_masked, true);
 271                 __napi_schedule(&rq->xdp_napi);
 272         }
 273 }
 274
 275 static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb)
 276 {
 277         if (unlikely(ptr_ring_produce(&rq->xdp_ring, skb))) {
 278                 dev_kfree_skb_any(skb);
 279                 return NET_RX_DROP;
 280         }
 281
 282         return NET_RX_SUCCESS;
 283 }
 284
 285 static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb,
 286                             struct veth_rq *rq, bool xdp)
 287 {
 288         return __dev_forward_skb(dev, skb) ?: xdp ?
 289                 veth_xdp_rx(rq, skb) :
 290                 __netif_rx(skb);
 291 }
 292
 293 /* return true if the specified skb has chances of GRO aggregation
 294  * Don't strive for accuracy, but try to avoid GRO overhead in the most
 295  * common scenarios.
 296  * When XDP is enabled, all traffic is considered eligible, as the xmit
 297  * device has TSO off.
 298  * When TSO is enabled on the xmit device, we are likely interested only
 299  * in UDP aggregation, explicitly check for that if the skb is suspected
 300  * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets -
 301  * to belong to locally generated UDP traffic.
 302  */
 303 static bool veth_skb_is_eligible_for_gro(const struct net_device *dev,
 304                                          const struct net_device *rcv,
 305                                          const struct sk_buff *skb)
 306 {
 307         return !(dev->features & NETIF_F_ALL_TSO) ||
 308                 (skb->destructor == sock_wfree &&
 309                  rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD));
 310 }
 311
 312 static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 313 {
 314         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 315         struct veth_rq *rq = NULL;
 316         struct net_device *rcv;
 317         int length = skb->len;
 318         bool use_napi = false;
 319         int rxq;
 320
 321         rcu_read_lock();
 322         rcv = rcu_dereference(priv->peer);
 323         if (unlikely(!rcv) || !pskb_may_pull(skb, ETH_HLEN)) {
 324                 kfree_skb(skb);
 325                 goto drop;
 326         }
 327
 328         rcv_priv = netdev_priv(rcv);
 329         rxq = skb_get_queue_mapping(skb);
 330         if (rxq < rcv->real_num_rx_queues) {
 331                 rq = &rcv_priv->rq[rxq];
 332
 333                 /* The napi pointer is available when an XDP program is
 334                  * attached or when GRO is enabled
 335                  * Don't bother with napi/GRO if the skb can't be aggregated
 336                  */
 337                 use_napi = rcu_access_pointer(rq->napi) &&
 338                            veth_skb_is_eligible_for_gro(dev, rcv, skb);
 339         }
 340
 341         skb_tx_timestamp(skb);
 342         if (likely(veth_forward_skb(rcv, skb, rq, use_napi) == NET_RX_SUCCESS)) {
 343                 if (!use_napi)
 344                         dev_lstats_add(dev, length);
 345         } else {
 346 drop:
 347                 atomic64_inc(&priv->dropped);
 348         }
 349
 350         if (use_napi)
 351                 __veth_xdp_flush(rq);
 352
 353         rcu_read_unlock();
 354
 355         return NETDEV_TX_OK;
 356 }
 357
 358 static u64 veth_stats_tx(struct net_device *dev, u64 *packets, u64 *bytes)
 359 {
 360         struct veth_priv *priv = netdev_priv(dev);
 361
 362         dev_lstats_read(dev, packets, bytes);
 363         return atomic64_read(&priv->dropped);
 364 }
 365
 366 static void veth_stats_rx(struct veth_stats *result, struct net_device *dev)
 367 {
 368         struct veth_priv *priv = netdev_priv(dev);
 369         int i;
 370
 371         result->peer_tq_xdp_xmit_err = 0;
 372         result->xdp_packets = 0;
 373         result->xdp_tx_err = 0;
 374         result->xdp_bytes = 0;
 375         result->rx_drops = 0;
 376         for (i = 0; i < dev->num_rx_queues; i++) {
 377                 u64 packets, bytes, drops, xdp_tx_err, peer_tq_xdp_xmit_err;
 378                 struct veth_rq_stats *stats = &priv->rq[i].stats;
 379                 unsigned int start;
 380
 381                 do {
 382                         start = u64_stats_fetch_begin_irq(&stats->syncp);
 383                         peer_tq_xdp_xmit_err = stats->vs.peer_tq_xdp_xmit_err;
 384                         xdp_tx_err = stats->vs.xdp_tx_err;
 385                         packets = stats->vs.xdp_packets;
 386                         bytes = stats->vs.xdp_bytes;
 387                         drops = stats->vs.rx_drops;
 388                 } while (u64_stats_fetch_retry_irq(&stats->syncp, start));
 389                 result->peer_tq_xdp_xmit_err += peer_tq_xdp_xmit_err;
 390                 result->xdp_tx_err += xdp_tx_err;
 391                 result->xdp_packets += packets;
 392                 result->xdp_bytes += bytes;
 393                 result->rx_drops += drops;
 394         }
 395 }
 396
 397 static void veth_get_stats64(struct net_device *dev,
 398                              struct rtnl_link_stats64 *tot)
 399 {
 400         struct veth_priv *priv = netdev_priv(dev);
 401         struct net_device *peer;
 402         struct veth_stats rx;
 403         u64 packets, bytes;
 404
 405         tot->tx_dropped = veth_stats_tx(dev, &packets, &bytes);
 406         tot->tx_bytes = bytes;
 407         tot->tx_packets = packets;
 408
 409         veth_stats_rx(&rx, dev);
 410         tot->tx_dropped += rx.xdp_tx_err;
 411         tot->rx_dropped = rx.rx_drops + rx.peer_tq_xdp_xmit_err;
 412         tot->rx_bytes = rx.xdp_bytes;
 413         tot->rx_packets = rx.xdp_packets;
 414
 415         rcu_read_lock();
 416         peer = rcu_dereference(priv->peer);
 417         if (peer) {
 418                 veth_stats_tx(peer, &packets, &bytes);
 419                 tot->rx_bytes += bytes;
 420                 tot->rx_packets += packets;
 421
 422                 veth_stats_rx(&rx, peer);
 423                 tot->tx_dropped += rx.peer_tq_xdp_xmit_err;
 424                 tot->rx_dropped += rx.xdp_tx_err;
 425                 tot->tx_bytes += rx.xdp_bytes;
 426                 tot->tx_packets += rx.xdp_packets;
 427         }
 428         rcu_read_unlock();
 429 }
 430
 431 /* fake multicast ability */
 432 static void veth_set_multicast_list(struct net_device *dev)
 433 {
 434 }
 435
 436 static int veth_select_rxq(struct net_device *dev)
 437 {
 438         return smp_processor_id() % dev->real_num_rx_queues;
 439 }
 440
 441 static struct net_device *veth_peer_dev(struct net_device *dev)
 442 {
 443         struct veth_priv *priv = netdev_priv(dev);
 444
 445         /* Callers must be under RCU read side. */
 446         return rcu_dereference(priv->peer);
 447 }
 448
 449 static int veth_xdp_xmit(struct net_device *dev, int n,
 450                          struct xdp_frame **frames,
 451                          u32 flags, bool ndo_xmit)
 452 {
 453         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 454         int i, ret = -ENXIO, nxmit = 0;
 455         struct net_device *rcv;
 456         unsigned int max_len;
 457         struct veth_rq *rq;
 458
 459         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
 460                 return -EINVAL;
 461
 462         rcu_read_lock();
 463         rcv = rcu_dereference(priv->peer);
 464         if (unlikely(!rcv))
 465                 goto out;
 466
 467         rcv_priv = netdev_priv(rcv);
 468         rq = &rcv_priv->rq[veth_select_rxq(rcv)];
 469         /* The napi pointer is set if NAPI is enabled, which ensures that
 470          * xdp_ring is initialized on receive side and the peer device is up.
 471          */
 472         if (!rcu_access_pointer(rq->napi))
 473                 goto out;
 474
 475         max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN;
 476
 477         spin_lock(&rq->xdp_ring.producer_lock);
 478         for (i = 0; i < n; i++) {
 479                 struct xdp_frame *frame = frames[i];
 480                 void *ptr = veth_xdp_to_ptr(frame);
 481
 482                 if (unlikely(xdp_get_frame_len(frame) > max_len ||
 483                              __ptr_ring_produce(&rq->xdp_ring, ptr)))
 484                         break;
 485                 nxmit++;
 486         }
 487         spin_unlock(&rq->xdp_ring.producer_lock);
 488
 489         if (flags & XDP_XMIT_FLUSH)
 490                 __veth_xdp_flush(rq);
 491
 492         ret = nxmit;
 493         if (ndo_xmit) {
 494                 u64_stats_update_begin(&rq->stats.syncp);
 495                 rq->stats.vs.peer_tq_xdp_xmit += nxmit;
 496                 rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit;
 497                 u64_stats_update_end(&rq->stats.syncp);
 498         }
 499
 500 out:
 501         rcu_read_unlock();
 502
 503         return ret;
 504 }
 505
 506 static int veth_ndo_xdp_xmit(struct net_device *dev, int n,
 507                              struct xdp_frame **frames, u32 flags)
 508 {
 509         int err;
 510
 511         err = veth_xdp_xmit(dev, n, frames, flags, true);
 512         if (err < 0) {
 513                 struct veth_priv *priv = netdev_priv(dev);
 514
 515                 atomic64_add(n, &priv->dropped);
 516         }
 517
 518         return err;
 519 }
 520
 521 static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
 522 {
 523         int sent, i, err = 0, drops;
 524
 525         sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false);
 526         if (sent < 0) {
 527                 err = sent;
 528                 sent = 0;
 529         }
 530
 531         for (i = sent; unlikely(i < bq->count); i++)
 532                 xdp_return_frame(bq->q[i]);
 533
 534         drops = bq->count - sent;
 535         trace_xdp_bulk_tx(rq->dev, sent, drops, err);
 536
 537         u64_stats_update_begin(&rq->stats.syncp);
 538         rq->stats.vs.xdp_tx += sent;
 539         rq->stats.vs.xdp_tx_err += drops;
 540         u64_stats_update_end(&rq->stats.syncp);
 541
 542         bq->count = 0;
 543 }
 544
 545 static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
 546 {
 547         struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev);
 548         struct net_device *rcv;
 549         struct veth_rq *rcv_rq;
 550
 551         rcu_read_lock();
 552         veth_xdp_flush_bq(rq, bq);
 553         rcv = rcu_dereference(priv->peer);
 554         if (unlikely(!rcv))
 555                 goto out;
 556
 557         rcv_priv = netdev_priv(rcv);
 558         rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)];
 559         /* xdp_ring is initialized on receive side? */
 560         if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog)))
 561                 goto out;
 562
 563         __veth_xdp_flush(rcv_rq);
 564 out:
 565         rcu_read_unlock();
 566 }
 567
 568 static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp,
 569                        struct veth_xdp_tx_bq *bq)
 570 {
 571         struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp);
 572
 573         if (unlikely(!frame))
 574                 return -EOVERFLOW;
 575
 576         if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE))
 577                 veth_xdp_flush_bq(rq, bq);
 578
 579         bq->q[bq->count++] = frame;
 580
 581         return 0;
 582 }
 583
 584 static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq,
 585                                           struct xdp_frame *frame,
 586                                           struct veth_xdp_tx_bq *bq,
 587                                           struct veth_stats *stats)
 588 {
 589         struct xdp_frame orig_frame;
 590         struct bpf_prog *xdp_prog;
 591
 592         rcu_read_lock();
 593         xdp_prog = rcu_dereference(rq->xdp_prog);
 594         if (likely(xdp_prog)) {
 595                 struct xdp_buff xdp;
 596                 u32 act;
 597
 598                 xdp_convert_frame_to_buff(frame, &xdp);
 599                 xdp.rxq = &rq->xdp_rxq;
 600
 601                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 602
 603                 switch (act) {
 604                 case XDP_PASS:
 605                         if (xdp_update_frame_from_buff(&xdp, frame))
 606                                 goto err_xdp;
 607                         break;
 608                 case XDP_TX:
 609                         orig_frame = *frame;
 610                         xdp.rxq->mem = frame->mem;
 611                         if (unlikely(veth_xdp_tx(rq, &xdp, bq) < 0)) {
 612                                 trace_xdp_exception(rq->dev, xdp_prog, act);
 613                                 frame = &orig_frame;
 614                                 stats->rx_drops++;
 615                                 goto err_xdp;
 616                         }
 617                         stats->xdp_tx++;
 618                         rcu_read_unlock();
 619                         goto xdp_xmit;
 620                 case XDP_REDIRECT:
 621                         orig_frame = *frame;
 622                         xdp.rxq->mem = frame->mem;
 623                         if (xdp_do_redirect(rq->dev, &xdp, xdp_prog)) {
 624                                 frame = &orig_frame;
 625                                 stats->rx_drops++;
 626                                 goto err_xdp;
 627                         }
 628                         stats->xdp_redirect++;
 629                         rcu_read_unlock();
 630                         goto xdp_xmit;
 631                 default:
 632                         bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
 633                         fallthrough;
 634                 case XDP_ABORTED:
 635                         trace_xdp_exception(rq->dev, xdp_prog, act);
 636                         fallthrough;
 637                 case XDP_DROP:
 638                         stats->xdp_drops++;
 639                         goto err_xdp;
 640                 }
 641         }
 642         rcu_read_unlock();
 643
 644         return frame;
 645 err_xdp:
 646         rcu_read_unlock();
 647         xdp_return_frame(frame);
 648 xdp_xmit:
 649         return NULL;
 650 }
 651
 652 /* frames array contains VETH_XDP_BATCH at most */
 653 static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames,
 654                                   int n_xdpf, struct veth_xdp_tx_bq *bq,
 655                                   struct veth_stats *stats)
 656 {
 657         void *skbs[VETH_XDP_BATCH];
 658         int i;
 659
 660         if (xdp_alloc_skb_bulk(skbs, n_xdpf,
 661                                GFP_ATOMIC | __GFP_ZERO) < 0) {
 662                 for (i = 0; i < n_xdpf; i++)
 663                         xdp_return_frame(frames[i]);
 664                 stats->rx_drops += n_xdpf;
 665
 666                 return;
 667         }
 668
 669         for (i = 0; i < n_xdpf; i++) {
 670                 struct sk_buff *skb = skbs[i];
 671
 672                 skb = __xdp_build_skb_from_frame(frames[i], skb,
 673                                                  rq->dev);
 674                 if (!skb) {
 675                         xdp_return_frame(frames[i]);
 676                         stats->rx_drops++;
 677                         continue;
 678                 }
 679                 napi_gro_receive(&rq->xdp_napi, skb);
 680         }
 681 }
 682
 683 static void veth_xdp_get(struct xdp_buff *xdp)
 684 {
 685         struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
 686         int i;
 687
 688         get_page(virt_to_page(xdp->data));
 689         if (likely(!xdp_buff_has_frags(xdp)))
 690                 return;
 691
 692         for (i = 0; i < sinfo->nr_frags; i++)
 693                 __skb_frag_ref(&sinfo->frags[i]);
 694 }
 695
 696 static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
 697                                         struct xdp_buff *xdp,
 698                                         struct sk_buff **pskb)
 699 {
 700         struct sk_buff *skb = *pskb;
 701         u32 frame_sz;
 702
 703         if (skb_shared(skb) || skb_head_is_locked(skb) ||
 704             skb_shinfo(skb)->nr_frags ||
 705             skb_headroom(skb) < XDP_PACKET_HEADROOM) {
 706                 u32 size, len, max_head_size, off;
 707                 struct sk_buff *nskb;
 708                 struct page *page;
 709                 int i, head_off;
 710
 711                 /* We need a private copy of the skb and data buffers since
 712                  * the ebpf program can modify it. We segment the original skb
 713                  * into order-0 pages without linearize it.
 714                  *
 715                  * Make sure we have enough space for linear and paged area
 716                  */
 717                 max_head_size = SKB_WITH_OVERHEAD(PAGE_SIZE -
 718                                                   VETH_XDP_HEADROOM);
 719                 if (skb->len > PAGE_SIZE * MAX_SKB_FRAGS + max_head_size)
 720                         goto drop;
 721
 722                 /* Allocate skb head */
 723                 page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
 724                 if (!page)
 725                         goto drop;
 726
 727                 nskb = build_skb(page_address(page), PAGE_SIZE);
 728                 if (!nskb) {
 729                         put_page(page);
 730                         goto drop;
 731                 }
 732
 733                 skb_reserve(nskb, VETH_XDP_HEADROOM);
 734                 size = min_t(u32, skb->len, max_head_size);
 735                 if (skb_copy_bits(skb, 0, nskb->data, size)) {
 736                         consume_skb(nskb);
 737                         goto drop;
 738                 }
 739                 skb_put(nskb, size);
 740
 741                 skb_copy_header(nskb, skb);
 742                 head_off = skb_headroom(nskb) - skb_headroom(skb);
 743                 skb_headers_offset_update(nskb, head_off);
 744
 745                 /* Allocate paged area of new skb */
 746                 off = size;
 747                 len = skb->len - off;
 748
 749                 for (i = 0; i < MAX_SKB_FRAGS && off < skb->len; i++) {
 750                         page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
 751                         if (!page) {
 752                                 consume_skb(nskb);
 753                                 goto drop;
 754                         }
 755
 756                         size = min_t(u32, len, PAGE_SIZE);
 757                         skb_add_rx_frag(nskb, i, page, 0, size, PAGE_SIZE);
 758                         if (skb_copy_bits(skb, off, page_address(page),
 759                                           size)) {
 760                                 consume_skb(nskb);
 761                                 goto drop;
 762                         }
 763
 764                         len -= size;
 765                         off += size;
 766                 }
 767
 768                 consume_skb(skb);
 769                 skb = nskb;
 770         }
 771
 772         /* SKB "head" area always have tailroom for skb_shared_info */
 773         frame_sz = skb_end_pointer(skb) - skb->head;
 774         frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 775         xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
 776         xdp_prepare_buff(xdp, skb->head, skb_headroom(skb),
 777                          skb_headlen(skb), true);
 778
 779         if (skb_is_nonlinear(skb)) {
 780                 skb_shinfo(skb)->xdp_frags_size = skb->data_len;
 781                 xdp_buff_set_frags_flag(xdp);
 782         } else {
 783                 xdp_buff_clear_frags_flag(xdp);
 784         }
 785         *pskb = skb;
 786
 787         return 0;
 788 drop:
 789         consume_skb(skb);
 790         *pskb = NULL;
 791
 792         return -ENOMEM;
 793 }
 794
 795 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq,
 796                                         struct sk_buff *skb,
 797                                         struct veth_xdp_tx_bq *bq,
 798                                         struct veth_stats *stats)
 799 {
 800         void *orig_data, *orig_data_end;
 801         struct bpf_prog *xdp_prog;
 802         struct xdp_buff xdp;
 803         u32 act, metalen;
 804         int off;
 805
 806         skb_prepare_for_gro(skb);
 807
 808         rcu_read_lock();
 809         xdp_prog = rcu_dereference(rq->xdp_prog);
 810         if (unlikely(!xdp_prog)) {
 811                 rcu_read_unlock();
 812                 goto out;
 813         }
 814
 815         __skb_push(skb, skb->data - skb_mac_header(skb));
 816         if (veth_convert_skb_to_xdp_buff(rq, &xdp, &skb))
 817                 goto drop;
 818
 819         orig_data = xdp.data;
 820         orig_data_end = xdp.data_end;
 821
 822         act = bpf_prog_run_xdp(xdp_prog, &xdp);
 823
 824         switch (act) {
 825         case XDP_PASS:
 826                 break;
 827         case XDP_TX:
 828                 veth_xdp_get(&xdp);
 829                 consume_skb(skb);
 830                 xdp.rxq->mem = rq->xdp_mem;
 831                 if (unlikely(veth_xdp_tx(rq, &xdp, bq) < 0)) {
 832                         trace_xdp_exception(rq->dev, xdp_prog, act);
 833                         stats->rx_drops++;
 834                         goto err_xdp;
 835                 }
 836                 stats->xdp_tx++;
 837                 rcu_read_unlock();
 838                 goto xdp_xmit;
 839         case XDP_REDIRECT:
 840                 veth_xdp_get(&xdp);
 841                 consume_skb(skb);
 842                 xdp.rxq->mem = rq->xdp_mem;
 843                 if (xdp_do_redirect(rq->dev, &xdp, xdp_prog)) {
 844                         stats->rx_drops++;
 845                         goto err_xdp;
 846                 }
 847                 stats->xdp_redirect++;
 848                 rcu_read_unlock();
 849                 goto xdp_xmit;
 850         default:
 851                 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
 852                 fallthrough;
 853         case XDP_ABORTED:
 854                 trace_xdp_exception(rq->dev, xdp_prog, act);
 855                 fallthrough;
 856         case XDP_DROP:
 857                 stats->xdp_drops++;
 858                 goto xdp_drop;
 859         }
 860         rcu_read_unlock();
 861
 862         /* check if bpf_xdp_adjust_head was used */
 863         off = orig_data - xdp.data;
 864         if (off > 0)
 865                 __skb_push(skb, off);
 866         else if (off < 0)
 867                 __skb_pull(skb, -off);
 868
 869         skb_reset_mac_header(skb);
 870
 871         /* check if bpf_xdp_adjust_tail was used */
 872         off = xdp.data_end - orig_data_end;
 873         if (off != 0)
 874                 __skb_put(skb, off); /* positive on grow, negative on shrink */
 875
 876         /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers
 877          * (e.g. bpf_xdp_adjust_tail), we need to update data_len here.
 878          */
 879         if (xdp_buff_has_frags(&xdp))
 880                 skb->data_len = skb_shinfo(skb)->xdp_frags_size;
 881         else
 882                 skb->data_len = 0;
 883
 884         skb->protocol = eth_type_trans(skb, rq->dev);
 885
 886         metalen = xdp.data - xdp.data_meta;
 887         if (metalen)
 888                 skb_metadata_set(skb, metalen);
 889 out:
 890         return skb;
 891 drop:
 892         stats->rx_drops++;
 893 xdp_drop:
 894         rcu_read_unlock();
 895         kfree_skb(skb);
 896         return NULL;
 897 err_xdp:
 898         rcu_read_unlock();
 899         xdp_return_buff(&xdp);
 900 xdp_xmit:
 901         return NULL;
 902 }
 903
 904 static int veth_xdp_rcv(struct veth_rq *rq, int budget,
 905                         struct veth_xdp_tx_bq *bq,
 906                         struct veth_stats *stats)
 907 {
 908         int i, done = 0, n_xdpf = 0;
 909         void *xdpf[VETH_XDP_BATCH];
 910
 911         for (i = 0; i < budget; i++) {
 912                 void *ptr = __ptr_ring_consume(&rq->xdp_ring);
 913
 914                 if (!ptr)
 915                         break;
 916
 917                 if (veth_is_xdp_frame(ptr)) {
 918                         /* ndo_xdp_xmit */
 919                         struct xdp_frame *frame = veth_ptr_to_xdp(ptr);
 920
 921                         stats->xdp_bytes += xdp_get_frame_len(frame);
 922                         frame = veth_xdp_rcv_one(rq, frame, bq, stats);
 923                         if (frame) {
 924                                 /* XDP_PASS */
 925                                 xdpf[n_xdpf++] = frame;
 926                                 if (n_xdpf == VETH_XDP_BATCH) {
 927                                         veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf,
 928                                                               bq, stats);
 929                                         n_xdpf = 0;
 930                                 }
 931                         }
 932                 } else {
 933                         /* ndo_start_xmit */
 934                         struct sk_buff *skb = ptr;
 935
 936                         stats->xdp_bytes += skb->len;
 937                         skb = veth_xdp_rcv_skb(rq, skb, bq, stats);
 938                         if (skb) {
 939                                 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC))
 940                                         netif_receive_skb(skb);
 941                                 else
 942                                         napi_gro_receive(&rq->xdp_napi, skb);
 943                         }
 944                 }
 945                 done++;
 946         }
 947
 948         if (n_xdpf)
 949                 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats);
 950
 951         u64_stats_update_begin(&rq->stats.syncp);
 952         rq->stats.vs.xdp_redirect += stats->xdp_redirect;
 953         rq->stats.vs.xdp_bytes += stats->xdp_bytes;
 954         rq->stats.vs.xdp_drops += stats->xdp_drops;
 955         rq->stats.vs.rx_drops += stats->rx_drops;
 956         rq->stats.vs.xdp_packets += done;
 957         u64_stats_update_end(&rq->stats.syncp);
 958
 959         return done;
 960 }
 961
 962 static int veth_poll(struct napi_struct *napi, int budget)
 963 {
 964         struct veth_rq *rq =
 965                 container_of(napi, struct veth_rq, xdp_napi);
 966         struct veth_stats stats = {};
 967         struct veth_xdp_tx_bq bq;
 968         int done;
 969
 970         bq.count = 0;
 971
 972         xdp_set_return_frame_no_direct();
 973         done = veth_xdp_rcv(rq, budget, &bq, &stats);
 974
 975         if (stats.xdp_redirect > 0)
 976                 xdp_do_flush();
 977
 978         if (done < budget && napi_complete_done(napi, done)) {
 979                 /* Write rx_notify_masked before reading ptr_ring */
 980                 smp_store_mb(rq->rx_notify_masked, false);
 981                 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
 982                         if (napi_schedule_prep(&rq->xdp_napi)) {
 983                                 WRITE_ONCE(rq->rx_notify_masked, true);
 984                                 __napi_schedule(&rq->xdp_napi);
 985                         }
 986                 }
 987         }
 988
 989         if (stats.xdp_tx > 0)
 990                 veth_xdp_flush(rq, &bq);
 991         xdp_clear_return_frame_no_direct();
 992
 993         return done;
 994 }
 995
 996 static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
 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                 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
1005                 if (err)
1006                         goto err_xdp_ring;
1007         }
1008
1009         for (i = start; i < end; i++) {
1010                 struct veth_rq *rq = &priv->rq[i];
1011
1012                 napi_enable(&rq->xdp_napi);
1013                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1014         }
1015
1016         return 0;
1017
1018 err_xdp_ring:
1019         for (i--; i >= start; i--)
1020                 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
1021
1022         return err;
1023 }
1024
1025 static int __veth_napi_enable(struct net_device *dev)
1026 {
1027         return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1028 }
1029
1030 static void veth_napi_del_range(struct net_device *dev, int start, int end)
1031 {
1032         struct veth_priv *priv = netdev_priv(dev);
1033         int i;
1034
1035         for (i = start; i < end; i++) {
1036                 struct veth_rq *rq = &priv->rq[i];
1037
1038                 rcu_assign_pointer(priv->rq[i].napi, NULL);
1039                 napi_disable(&rq->xdp_napi);
1040                 __netif_napi_del(&rq->xdp_napi);
1041         }
1042         synchronize_net();
1043
1044         for (i = start; i < end; i++) {
1045                 struct veth_rq *rq = &priv->rq[i];
1046
1047                 rq->rx_notify_masked = false;
1048                 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
1049         }
1050 }
1051
1052 static void veth_napi_del(struct net_device *dev)
1053 {
1054         veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
1055 }
1056
1057 static bool veth_gro_requested(const struct net_device *dev)
1058 {
1059         return !!(dev->wanted_features & NETIF_F_GRO);
1060 }
1061
1062 static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1063                                  bool napi_already_on)
1064 {
1065         struct veth_priv *priv = netdev_priv(dev);
1066         int err, i;
1067
1068         for (i = start; i < end; i++) {
1069                 struct veth_rq *rq = &priv->rq[i];
1070
1071                 if (!napi_already_on)
1072                         netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1073                 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1074                 if (err < 0)
1075                         goto err_rxq_reg;
1076
1077                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1078                                                  MEM_TYPE_PAGE_SHARED,
1079                                                  NULL);
1080                 if (err < 0)
1081                         goto err_reg_mem;
1082
1083                 /* Save original mem info as it can be overwritten */
1084                 rq->xdp_mem = rq->xdp_rxq.mem;
1085         }
1086         return 0;
1087
1088 err_reg_mem:
1089         xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1090 err_rxq_reg:
1091         for (i--; i >= start; i--) {
1092                 struct veth_rq *rq = &priv->rq[i];
1093
1094                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1095                 if (!napi_already_on)
1096                         netif_napi_del(&rq->xdp_napi);
1097         }
1098
1099         return err;
1100 }
1101
1102 static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1103                                    bool delete_napi)
1104 {
1105         struct veth_priv *priv = netdev_priv(dev);
1106         int i;
1107
1108         for (i = start; i < end; i++) {
1109                 struct veth_rq *rq = &priv->rq[i];
1110
1111                 rq->xdp_rxq.mem = rq->xdp_mem;
1112                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1113
1114                 if (delete_napi)
1115                         netif_napi_del(&rq->xdp_napi);
1116         }
1117 }
1118
1119 static int veth_enable_xdp(struct net_device *dev)
1120 {
1121         bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1122         struct veth_priv *priv = netdev_priv(dev);
1123         int err, i;
1124
1125         if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1126                 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1127                 if (err)
1128                         return err;
1129
1130                 if (!napi_already_on) {
1131                         err = __veth_napi_enable(dev);
1132                         if (err) {
1133                                 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1134                                 return err;
1135                         }
1136
1137                         if (!veth_gro_requested(dev)) {
1138                                 /* user-space did not require GRO, but adding XDP
1139                                  * is supposed to get GRO working
1140                                  */
1141                                 dev->features |= NETIF_F_GRO;
1142                                 netdev_features_change(dev);
1143                         }
1144                 }
1145         }
1146
1147         for (i = 0; i < dev->real_num_rx_queues; i++) {
1148                 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1149                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1150         }
1151
1152         return 0;
1153 }
1154
1155 static void veth_disable_xdp(struct net_device *dev)
1156 {
1157         struct veth_priv *priv = netdev_priv(dev);
1158         int i;
1159
1160         for (i = 0; i < dev->real_num_rx_queues; i++)
1161                 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1162
1163         if (!netif_running(dev) || !veth_gro_requested(dev)) {
1164                 veth_napi_del(dev);
1165
1166                 /* if user-space did not require GRO, since adding XDP
1167                  * enabled it, clear it now
1168                  */
1169                 if (!veth_gro_requested(dev) && netif_running(dev)) {
1170                         dev->features &= ~NETIF_F_GRO;
1171                         netdev_features_change(dev);
1172                 }
1173         }
1174
1175         veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1176 }
1177
1178 static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1179 {
1180         struct veth_priv *priv = netdev_priv(dev);
1181         int err, i;
1182
1183         for (i = start; i < end; i++) {
1184                 struct veth_rq *rq = &priv->rq[i];
1185
1186                 netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1187         }
1188
1189         err = __veth_napi_enable_range(dev, start, end);
1190         if (err) {
1191                 for (i = start; i < end; i++) {
1192                         struct veth_rq *rq = &priv->rq[i];
1193
1194                         netif_napi_del(&rq->xdp_napi);
1195                 }
1196                 return err;
1197         }
1198         return err;
1199 }
1200
1201 static int veth_napi_enable(struct net_device *dev)
1202 {
1203         return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1204 }
1205
1206 static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1207 {
1208         struct veth_priv *priv = netdev_priv(dev);
1209
1210         if (start >= end)
1211                 return;
1212
1213         if (priv->_xdp_prog) {
1214                 veth_napi_del_range(dev, start, end);
1215                 veth_disable_xdp_range(dev, start, end, false);
1216         } else if (veth_gro_requested(dev)) {
1217                 veth_napi_del_range(dev, start, end);
1218         }
1219 }
1220
1221 static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1222 {
1223         struct veth_priv *priv = netdev_priv(dev);
1224         int err;
1225
1226         if (start >= end)
1227                 return 0;
1228
1229         if (priv->_xdp_prog) {
1230                 /* these channels are freshly initialized, napi is not on there even
1231                  * when GRO is requeste
1232                  */
1233                 err = veth_enable_xdp_range(dev, start, end, false);
1234                 if (err)
1235                         return err;
1236
1237                 err = __veth_napi_enable_range(dev, start, end);
1238                 if (err) {
1239                         /* on error always delete the newly added napis */
1240                         veth_disable_xdp_range(dev, start, end, true);
1241                         return err;
1242                 }
1243         } else if (veth_gro_requested(dev)) {
1244                 return veth_napi_enable_range(dev, start, end);
1245         }
1246         return 0;
1247 }
1248
1249 static int veth_set_channels(struct net_device *dev,
1250                              struct ethtool_channels *ch)
1251 {
1252         struct veth_priv *priv = netdev_priv(dev);
1253         unsigned int old_rx_count, new_rx_count;
1254         struct veth_priv *peer_priv;
1255         struct net_device *peer;
1256         int err;
1257
1258         /* sanity check. Upper bounds are already enforced by the caller */
1259         if (!ch->rx_count || !ch->tx_count)
1260                 return -EINVAL;
1261
1262         /* avoid braking XDP, if that is enabled */
1263         peer = rtnl_dereference(priv->peer);
1264         peer_priv = peer ? netdev_priv(peer) : NULL;
1265         if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1266                 return -EINVAL;
1267
1268         if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1269                 return -EINVAL;
1270
1271         old_rx_count = dev->real_num_rx_queues;
1272         new_rx_count = ch->rx_count;
1273         if (netif_running(dev)) {
1274                 /* turn device off */
1275                 netif_carrier_off(dev);
1276                 if (peer)
1277                         netif_carrier_off(peer);
1278
1279                 /* try to allocate new resurces, as needed*/
1280                 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1281                 if (err)
1282                         goto out;
1283         }
1284
1285         err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1286         if (err)
1287                 goto revert;
1288
1289         err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1290         if (err) {
1291                 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1292
1293                 /* this error condition could happen only if rx and tx change
1294                  * in opposite directions (e.g. tx nr raises, rx nr decreases)
1295                  * and we can't do anything to fully restore the original
1296                  * status
1297                  */
1298                 if (err2)
1299                         pr_warn("Can't restore rx queues config %d -> %d %d",
1300                                 new_rx_count, old_rx_count, err2);
1301                 else
1302                         goto revert;
1303         }
1304
1305 out:
1306         if (netif_running(dev)) {
1307                 /* note that we need to swap the arguments WRT the enable part
1308                  * to identify the range we have to disable
1309                  */
1310                 veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1311                 netif_carrier_on(dev);
1312                 if (peer)
1313                         netif_carrier_on(peer);
1314         }
1315         return err;
1316
1317 revert:
1318         new_rx_count = old_rx_count;
1319         old_rx_count = ch->rx_count;
1320         goto out;
1321 }
1322
1323 static int veth_open(struct net_device *dev)
1324 {
1325         struct veth_priv *priv = netdev_priv(dev);
1326         struct net_device *peer = rtnl_dereference(priv->peer);
1327         int err;
1328
1329         if (!peer)
1330                 return -ENOTCONN;
1331
1332         if (priv->_xdp_prog) {
1333                 err = veth_enable_xdp(dev);
1334                 if (err)
1335                         return err;
1336         } else if (veth_gro_requested(dev)) {
1337                 err = veth_napi_enable(dev);
1338                 if (err)
1339                         return err;
1340         }
1341
1342         if (peer->flags & IFF_UP) {
1343                 netif_carrier_on(dev);
1344                 netif_carrier_on(peer);
1345         }
1346
1347         return 0;
1348 }
1349
1350 static int veth_close(struct net_device *dev)
1351 {
1352         struct veth_priv *priv = netdev_priv(dev);
1353         struct net_device *peer = rtnl_dereference(priv->peer);
1354
1355         netif_carrier_off(dev);
1356         if (peer)
1357                 netif_carrier_off(peer);
1358
1359         if (priv->_xdp_prog)
1360                 veth_disable_xdp(dev);
1361         else if (veth_gro_requested(dev))
1362                 veth_napi_del(dev);
1363
1364         return 0;
1365 }
1366
1367 static int is_valid_veth_mtu(int mtu)
1368 {
1369         return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1370 }
1371
1372 static int veth_alloc_queues(struct net_device *dev)
1373 {
1374         struct veth_priv *priv = netdev_priv(dev);
1375         int i;
1376
1377         priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL_ACCOUNT);
1378         if (!priv->rq)
1379                 return -ENOMEM;
1380
1381         for (i = 0; i < dev->num_rx_queues; i++) {
1382                 priv->rq[i].dev = dev;
1383                 u64_stats_init(&priv->rq[i].stats.syncp);
1384         }
1385
1386         return 0;
1387 }
1388
1389 static void veth_free_queues(struct net_device *dev)
1390 {
1391         struct veth_priv *priv = netdev_priv(dev);
1392
1393         kfree(priv->rq);
1394 }
1395
1396 static int veth_dev_init(struct net_device *dev)
1397 {
1398         int err;
1399
1400         dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
1401         if (!dev->lstats)
1402                 return -ENOMEM;
1403
1404         err = veth_alloc_queues(dev);
1405         if (err) {
1406                 free_percpu(dev->lstats);
1407                 return err;
1408         }
1409
1410         return 0;
1411 }
1412
1413 static void veth_dev_free(struct net_device *dev)
1414 {
1415         veth_free_queues(dev);
1416         free_percpu(dev->lstats);
1417 }
1418
1419 #ifdef CONFIG_NET_POLL_CONTROLLER
1420 static void veth_poll_controller(struct net_device *dev)
1421 {
1422         /* veth only receives frames when its peer sends one
1423          * Since it has nothing to do with disabling irqs, we are guaranteed
1424          * never to have pending data when we poll for it so
1425          * there is nothing to do here.
1426          *
1427          * We need this though so netpoll recognizes us as an interface that
1428          * supports polling, which enables bridge devices in virt setups to
1429          * still use netconsole
1430          */
1431 }
1432 #endif  /* CONFIG_NET_POLL_CONTROLLER */
1433
1434 static int veth_get_iflink(const struct net_device *dev)
1435 {
1436         struct veth_priv *priv = netdev_priv(dev);
1437         struct net_device *peer;
1438         int iflink;
1439
1440         rcu_read_lock();
1441         peer = rcu_dereference(priv->peer);
1442         iflink = peer ? peer->ifindex : 0;
1443         rcu_read_unlock();
1444
1445         return iflink;
1446 }
1447
1448 static netdev_features_t veth_fix_features(struct net_device *dev,
1449                                            netdev_features_t features)
1450 {
1451         struct veth_priv *priv = netdev_priv(dev);
1452         struct net_device *peer;
1453
1454         peer = rtnl_dereference(priv->peer);
1455         if (peer) {
1456                 struct veth_priv *peer_priv = netdev_priv(peer);
1457
1458                 if (peer_priv->_xdp_prog)
1459                         features &= ~NETIF_F_GSO_SOFTWARE;
1460         }
1461         if (priv->_xdp_prog)
1462                 features |= NETIF_F_GRO;
1463
1464         return features;
1465 }
1466
1467 static int veth_set_features(struct net_device *dev,
1468                              netdev_features_t features)
1469 {
1470         netdev_features_t changed = features ^ dev->features;
1471         struct veth_priv *priv = netdev_priv(dev);
1472         int err;
1473
1474         if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1475                 return 0;
1476
1477         if (features & NETIF_F_GRO) {
1478                 err = veth_napi_enable(dev);
1479                 if (err)
1480                         return err;
1481         } else {
1482                 veth_napi_del(dev);
1483         }
1484         return 0;
1485 }
1486
1487 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1488 {
1489         struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1490         struct net_device *peer;
1491
1492         if (new_hr < 0)
1493                 new_hr = 0;
1494
1495         rcu_read_lock();
1496         peer = rcu_dereference(priv->peer);
1497         if (unlikely(!peer))
1498                 goto out;
1499
1500         peer_priv = netdev_priv(peer);
1501         priv->requested_headroom = new_hr;
1502         new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1503         dev->needed_headroom = new_hr;
1504         peer->needed_headroom = new_hr;
1505
1506 out:
1507         rcu_read_unlock();
1508 }
1509
1510 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1511                         struct netlink_ext_ack *extack)
1512 {
1513         struct veth_priv *priv = netdev_priv(dev);
1514         struct bpf_prog *old_prog;
1515         struct net_device *peer;
1516         unsigned int max_mtu;
1517         int err;
1518
1519         old_prog = priv->_xdp_prog;
1520         priv->_xdp_prog = prog;
1521         peer = rtnl_dereference(priv->peer);
1522
1523         if (prog) {
1524                 if (!peer) {
1525                         NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1526                         err = -ENOTCONN;
1527                         goto err;
1528                 }
1529
1530                 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) -
1531                           peer->hard_header_len;
1532                 /* Allow increasing the max_mtu if the program supports
1533                  * XDP fragments.
1534                  */
1535                 if (prog->aux->xdp_has_frags)
1536                         max_mtu += PAGE_SIZE * MAX_SKB_FRAGS;
1537
1538                 if (peer->mtu > max_mtu) {
1539                         NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1540                         err = -ERANGE;
1541                         goto err;
1542                 }
1543
1544                 if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1545                         NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1546                         err = -ENOSPC;
1547                         goto err;
1548                 }
1549
1550                 if (dev->flags & IFF_UP) {
1551                         err = veth_enable_xdp(dev);
1552                         if (err) {
1553                                 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1554                                 goto err;
1555                         }
1556                 }
1557
1558                 if (!old_prog) {
1559                         peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1560                         peer->max_mtu = max_mtu;
1561                 }
1562         }
1563
1564         if (old_prog) {
1565                 if (!prog) {
1566                         if (dev->flags & IFF_UP)
1567                                 veth_disable_xdp(dev);
1568
1569                         if (peer) {
1570                                 peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1571                                 peer->max_mtu = ETH_MAX_MTU;
1572                         }
1573                 }
1574                 bpf_prog_put(old_prog);
1575         }
1576
1577         if ((!!old_prog ^ !!prog) && peer)
1578                 netdev_update_features(peer);
1579
1580         return 0;
1581 err:
1582         priv->_xdp_prog = old_prog;
1583
1584         return err;
1585 }
1586
1587 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1588 {
1589         switch (xdp->command) {
1590         case XDP_SETUP_PROG:
1591                 return veth_xdp_set(dev, xdp->prog, xdp->extack);
1592         default:
1593                 return -EINVAL;
1594         }
1595 }
1596
1597 static const struct net_device_ops veth_netdev_ops = {
1598         .ndo_init            = veth_dev_init,
1599         .ndo_open            = veth_open,
1600         .ndo_stop            = veth_close,
1601         .ndo_start_xmit      = veth_xmit,
1602         .ndo_get_stats64     = veth_get_stats64,
1603         .ndo_set_rx_mode     = veth_set_multicast_list,
1604         .ndo_set_mac_address = eth_mac_addr,
1605 #ifdef CONFIG_NET_POLL_CONTROLLER
1606         .ndo_poll_controller    = veth_poll_controller,
1607 #endif
1608         .ndo_get_iflink         = veth_get_iflink,
1609         .ndo_fix_features       = veth_fix_features,
1610         .ndo_set_features       = veth_set_features,
1611         .ndo_features_check     = passthru_features_check,
1612         .ndo_set_rx_headroom    = veth_set_rx_headroom,
1613         .ndo_bpf                = veth_xdp,
1614         .ndo_xdp_xmit           = veth_ndo_xdp_xmit,
1615         .ndo_get_peer_dev       = veth_peer_dev,
1616 };
1617
1618 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1619                        NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1620                        NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1621                        NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1622                        NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1623
1624 static void veth_setup(struct net_device *dev)
1625 {
1626         ether_setup(dev);
1627
1628         dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1629         dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1630         dev->priv_flags |= IFF_NO_QUEUE;
1631         dev->priv_flags |= IFF_PHONY_HEADROOM;
1632
1633         dev->netdev_ops = &veth_netdev_ops;
1634         dev->ethtool_ops = &veth_ethtool_ops;
1635         dev->features |= NETIF_F_LLTX;
1636         dev->features |= VETH_FEATURES;
1637         dev->vlan_features = dev->features &
1638                              ~(NETIF_F_HW_VLAN_CTAG_TX |
1639                                NETIF_F_HW_VLAN_STAG_TX |
1640                                NETIF_F_HW_VLAN_CTAG_RX |
1641                                NETIF_F_HW_VLAN_STAG_RX);
1642         dev->needs_free_netdev = true;
1643         dev->priv_destructor = veth_dev_free;
1644         dev->max_mtu = ETH_MAX_MTU;
1645
1646         dev->hw_features = VETH_FEATURES;
1647         dev->hw_enc_features = VETH_FEATURES;
1648         dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1649         netif_set_tso_max_size(dev, GSO_MAX_SIZE);
1650 }
1651
1652 /*
1653  * netlink interface
1654  */
1655
1656 static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1657                          struct netlink_ext_ack *extack)
1658 {
1659         if (tb[IFLA_ADDRESS]) {
1660                 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1661                         return -EINVAL;
1662                 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1663                         return -EADDRNOTAVAIL;
1664         }
1665         if (tb[IFLA_MTU]) {
1666                 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1667                         return -EINVAL;
1668         }
1669         return 0;
1670 }
1671
1672 static struct rtnl_link_ops veth_link_ops;
1673
1674 static void veth_disable_gro(struct net_device *dev)
1675 {
1676         dev->features &= ~NETIF_F_GRO;
1677         dev->wanted_features &= ~NETIF_F_GRO;
1678         netdev_update_features(dev);
1679 }
1680
1681 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1682 {
1683         int err;
1684
1685         if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1686                 err = netif_set_real_num_tx_queues(dev, 1);
1687                 if (err)
1688                         return err;
1689         }
1690         if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1691                 err = netif_set_real_num_rx_queues(dev, 1);
1692                 if (err)
1693                         return err;
1694         }
1695         return 0;
1696 }
1697
1698 static int veth_newlink(struct net *src_net, struct net_device *dev,
1699                         struct nlattr *tb[], struct nlattr *data[],
1700                         struct netlink_ext_ack *extack)
1701 {
1702         int err;
1703         struct net_device *peer;
1704         struct veth_priv *priv;
1705         char ifname[IFNAMSIZ];
1706         struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1707         unsigned char name_assign_type;
1708         struct ifinfomsg *ifmp;
1709         struct net *net;
1710
1711         /*
1712          * create and register peer first
1713          */
1714         if (data != NULL && data[VETH_INFO_PEER] != NULL) {
1715                 struct nlattr *nla_peer;
1716
1717                 nla_peer = data[VETH_INFO_PEER];
1718                 ifmp = nla_data(nla_peer);
1719                 err = rtnl_nla_parse_ifla(peer_tb,
1720                                           nla_data(nla_peer) + sizeof(struct ifinfomsg),
1721                                           nla_len(nla_peer) - sizeof(struct ifinfomsg),
1722                                           NULL);
1723                 if (err < 0)
1724                         return err;
1725
1726                 err = veth_validate(peer_tb, NULL, extack);
1727                 if (err < 0)
1728                         return err;
1729
1730                 tbp = peer_tb;
1731         } else {
1732                 ifmp = NULL;
1733                 tbp = tb;
1734         }
1735
1736         if (ifmp && tbp[IFLA_IFNAME]) {
1737                 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1738                 name_assign_type = NET_NAME_USER;
1739         } else {
1740                 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1741                 name_assign_type = NET_NAME_ENUM;
1742         }
1743
1744         net = rtnl_link_get_net(src_net, tbp);
1745         if (IS_ERR(net))
1746                 return PTR_ERR(net);
1747
1748         peer = rtnl_create_link(net, ifname, name_assign_type,
1749                                 &veth_link_ops, tbp, extack);
1750         if (IS_ERR(peer)) {
1751                 put_net(net);
1752                 return PTR_ERR(peer);
1753         }
1754
1755         if (!ifmp || !tbp[IFLA_ADDRESS])
1756                 eth_hw_addr_random(peer);
1757
1758         if (ifmp && (dev->ifindex != 0))
1759                 peer->ifindex = ifmp->ifi_index;
1760
1761         netif_inherit_tso_max(peer, dev);
1762
1763         err = register_netdevice(peer);
1764         put_net(net);
1765         net = NULL;
1766         if (err < 0)
1767                 goto err_register_peer;
1768
1769         /* keep GRO disabled by default to be consistent with the established
1770          * veth behavior
1771          */
1772         veth_disable_gro(peer);
1773         netif_carrier_off(peer);
1774
1775         err = rtnl_configure_link(peer, ifmp);
1776         if (err < 0)
1777                 goto err_configure_peer;
1778
1779         /*
1780          * register dev last
1781          *
1782          * note, that since we've registered new device the dev's name
1783          * should be re-allocated
1784          */
1785
1786         if (tb[IFLA_ADDRESS] == NULL)
1787                 eth_hw_addr_random(dev);
1788
1789         if (tb[IFLA_IFNAME])
1790                 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1791         else
1792                 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1793
1794         err = register_netdevice(dev);
1795         if (err < 0)
1796                 goto err_register_dev;
1797
1798         netif_carrier_off(dev);
1799
1800         /*
1801          * tie the deviced together
1802          */
1803
1804         priv = netdev_priv(dev);
1805         rcu_assign_pointer(priv->peer, peer);
1806         err = veth_init_queues(dev, tb);
1807         if (err)
1808                 goto err_queues;
1809
1810         priv = netdev_priv(peer);
1811         rcu_assign_pointer(priv->peer, dev);
1812         err = veth_init_queues(peer, tb);
1813         if (err)
1814                 goto err_queues;
1815
1816         veth_disable_gro(dev);
1817         return 0;
1818
1819 err_queues:
1820         unregister_netdevice(dev);
1821 err_register_dev:
1822         /* nothing to do */
1823 err_configure_peer:
1824         unregister_netdevice(peer);
1825         return err;
1826
1827 err_register_peer:
1828         free_netdev(peer);
1829         return err;
1830 }
1831
1832 static void veth_dellink(struct net_device *dev, struct list_head *head)
1833 {
1834         struct veth_priv *priv;
1835         struct net_device *peer;
1836
1837         priv = netdev_priv(dev);
1838         peer = rtnl_dereference(priv->peer);
1839
1840         /* Note : dellink() is called from default_device_exit_batch(),
1841          * before a rcu_synchronize() point. The devices are guaranteed
1842          * not being freed before one RCU grace period.
1843          */
1844         RCU_INIT_POINTER(priv->peer, NULL);
1845         unregister_netdevice_queue(dev, head);
1846
1847         if (peer) {
1848                 priv = netdev_priv(peer);
1849                 RCU_INIT_POINTER(priv->peer, NULL);
1850                 unregister_netdevice_queue(peer, head);
1851         }
1852 }
1853
1854 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
1855         [VETH_INFO_PEER]        = { .len = sizeof(struct ifinfomsg) },
1856 };
1857
1858 static struct net *veth_get_link_net(const struct net_device *dev)
1859 {
1860         struct veth_priv *priv = netdev_priv(dev);
1861         struct net_device *peer = rtnl_dereference(priv->peer);
1862
1863         return peer ? dev_net(peer) : dev_net(dev);
1864 }
1865
1866 static unsigned int veth_get_num_queues(void)
1867 {
1868         /* enforce the same queue limit as rtnl_create_link */
1869         int queues = num_possible_cpus();
1870
1871         if (queues > 4096)
1872                 queues = 4096;
1873         return queues;
1874 }
1875
1876 static struct rtnl_link_ops veth_link_ops = {
1877         .kind           = DRV_NAME,
1878         .priv_size      = sizeof(struct veth_priv),
1879         .setup          = veth_setup,
1880         .validate       = veth_validate,
1881         .newlink        = veth_newlink,
1882         .dellink        = veth_dellink,
1883         .policy         = veth_policy,
1884         .maxtype        = VETH_INFO_MAX,
1885         .get_link_net   = veth_get_link_net,
1886         .get_num_tx_queues      = veth_get_num_queues,
1887         .get_num_rx_queues      = veth_get_num_queues,
1888 };
1889
1890 /*
1891  * init/fini
1892  */
1893
1894 static __init int veth_init(void)
1895 {
1896         return rtnl_link_register(&veth_link_ops);
1897 }
1898
1899 static __exit void veth_exit(void)
1900 {
1901         rtnl_link_unregister(&veth_link_ops);
1902 }
1903
1904 module_init(veth_init);
1905 module_exit(veth_exit);
1906
1907 MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
1908 MODULE_LICENSE("GPL v2");
1909 MODULE_ALIAS_RTNL_LINK(DRV_NAME);