drivers/net/veth.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  drivers/net/veth.c
   4  *
   5  *  Copyright (C) 2007 OpenVZ http://openvz.org, SWsoft Inc
   6  *
   7  * Author: Pavel Emelianov <xemul@openvz.org>
   8  * Ethtool interface from: Eric W. Biederman <ebiederm@xmission.com>
   9  *
  10  */
  11
  12 #include <linux/netdevice.h>
  13 #include <linux/slab.h>
  14 #include <linux/ethtool.h>
  15 #include <linux/etherdevice.h>
  16 #include <linux/u64_stats_sync.h>
  17
  18 #include <net/rtnetlink.h>
  19 #include <net/dst.h>
  20 #include <net/xfrm.h>
  21 #include <net/xdp.h>
  22 #include <linux/veth.h>
  23 #include <linux/module.h>
  24 #include <linux/bpf.h>
  25 #include <linux/filter.h>
  26 #include <linux/ptr_ring.h>
  27 #include <linux/bpf_trace.h>
  28 #include <linux/net_tstamp.h>
  29
  30 #define DRV_NAME        "veth"
  31 #define DRV_VERSION     "1.0"
  32
  33 #define VETH_XDP_FLAG           BIT(0)
  34 #define VETH_RING_SIZE          256
  35 #define VETH_XDP_HEADROOM       (XDP_PACKET_HEADROOM + NET_IP_ALIGN)
  36
  37 #define VETH_XDP_TX_BULK_SIZE   16
  38 #define VETH_XDP_BATCH          16
  39
  40 struct veth_stats {
  41         u64     rx_drops;
  42         /* xdp */
  43         u64     xdp_packets;
  44         u64     xdp_bytes;
  45         u64     xdp_redirect;
  46         u64     xdp_drops;
  47         u64     xdp_tx;
  48         u64     xdp_tx_err;
  49         u64     peer_tq_xdp_xmit;
  50         u64     peer_tq_xdp_xmit_err;
  51 };
  52
  53 struct veth_rq_stats {
  54         struct veth_stats       vs;
  55         struct u64_stats_sync   syncp;
  56 };
  57
  58 struct veth_rq {
  59         struct napi_struct      xdp_napi;
  60         struct napi_struct __rcu *napi; /* points to xdp_napi when the latter is initialized */
  61         struct net_device       *dev;
  62         struct bpf_prog __rcu   *xdp_prog;
  63         struct xdp_mem_info     xdp_mem;
  64         struct veth_rq_stats    stats;
  65         bool                    rx_notify_masked;
  66         struct ptr_ring         xdp_ring;
  67         struct xdp_rxq_info     xdp_rxq;
  68 };
  69
  70 struct veth_priv {
  71         struct net_device __rcu *peer;
  72         atomic64_t              dropped;
  73         struct bpf_prog         *_xdp_prog;
  74         struct veth_rq          *rq;
  75         unsigned int            requested_headroom;
  76 };
  77
  78 struct veth_xdp_tx_bq {
  79         struct xdp_frame *q[VETH_XDP_TX_BULK_SIZE];
  80         unsigned int count;
  81 };
  82
  83 /*
  84  * ethtool interface
  85  */
  86
  87 struct veth_q_stat_desc {
  88         char    desc[ETH_GSTRING_LEN];
  89         size_t  offset;
  90 };
  91
  92 #define VETH_RQ_STAT(m) offsetof(struct veth_stats, m)
  93
  94 static const struct veth_q_stat_desc veth_rq_stats_desc[] = {
  95         { "xdp_packets",        VETH_RQ_STAT(xdp_packets) },
  96         { "xdp_bytes",          VETH_RQ_STAT(xdp_bytes) },
  97         { "drops",              VETH_RQ_STAT(rx_drops) },
  98         { "xdp_redirect",       VETH_RQ_STAT(xdp_redirect) },
  99         { "xdp_drops",          VETH_RQ_STAT(xdp_drops) },
 100         { "xdp_tx",             VETH_RQ_STAT(xdp_tx) },
 101         { "xdp_tx_errors",      VETH_RQ_STAT(xdp_tx_err) },
 102 };
 103
 104 #define VETH_RQ_STATS_LEN       ARRAY_SIZE(veth_rq_stats_desc)
 105
 106 static const struct veth_q_stat_desc veth_tq_stats_desc[] = {
 107         { "xdp_xmit",           VETH_RQ_STAT(peer_tq_xdp_xmit) },
 108         { "xdp_xmit_errors",    VETH_RQ_STAT(peer_tq_xdp_xmit_err) },
 109 };
 110
 111 #define VETH_TQ_STATS_LEN       ARRAY_SIZE(veth_tq_stats_desc)
 112
 113 static struct {
 114         const char string[ETH_GSTRING_LEN];
 115 } ethtool_stats_keys[] = {
 116         { "peer_ifindex" },
 117 };
 118
 119 static int veth_get_link_ksettings(struct net_device *dev,
 120                                    struct ethtool_link_ksettings *cmd)
 121 {
 122         cmd->base.speed         = SPEED_10000;
 123         cmd->base.duplex        = DUPLEX_FULL;
 124         cmd->base.port          = PORT_TP;
 125         cmd->base.autoneg       = AUTONEG_DISABLE;
 126         return 0;
 127 }
 128
 129 static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info)
 130 {
 131         strlcpy(info->driver, DRV_NAME, sizeof(info->driver));
 132         strlcpy(info->version, DRV_VERSION, sizeof(info->version));
 133 }
 134
 135 static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
 136 {
 137         u8 *p = buf;
 138         int i, j;
 139
 140         switch(stringset) {
 141         case ETH_SS_STATS:
 142                 memcpy(p, &ethtool_stats_keys, sizeof(ethtool_stats_keys));
 143                 p += sizeof(ethtool_stats_keys);
 144                 for (i = 0; i < dev->real_num_rx_queues; i++)
 145                         for (j = 0; j < VETH_RQ_STATS_LEN; j++)
 146                                 ethtool_sprintf(&p, "rx_queue_%u_%.18s",
 147                                                 i, veth_rq_stats_desc[j].desc);
 148
 149                 for (i = 0; i < dev->real_num_tx_queues; i++)
 150                         for (j = 0; j < VETH_TQ_STATS_LEN; j++)
 151                                 ethtool_sprintf(&p, "tx_queue_%u_%.18s",
 152                                                 i, veth_tq_stats_desc[j].desc);
 153                 break;
 154         }
 155 }
 156
 157 static int veth_get_sset_count(struct net_device *dev, int sset)
 158 {
 159         switch (sset) {
 160         case ETH_SS_STATS:
 161                 return ARRAY_SIZE(ethtool_stats_keys) +
 162                        VETH_RQ_STATS_LEN * dev->real_num_rx_queues +
 163                        VETH_TQ_STATS_LEN * dev->real_num_tx_queues;
 164         default:
 165                 return -EOPNOTSUPP;
 166         }
 167 }
 168
 169 static void veth_get_ethtool_stats(struct net_device *dev,
 170                 struct ethtool_stats *stats, u64 *data)
 171 {
 172         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 173         struct net_device *peer = rtnl_dereference(priv->peer);
 174         int i, j, idx;
 175
 176         data[0] = peer ? peer->ifindex : 0;
 177         idx = 1;
 178         for (i = 0; i < dev->real_num_rx_queues; i++) {
 179                 const struct veth_rq_stats *rq_stats = &priv->rq[i].stats;
 180                 const void *stats_base = (void *)&rq_stats->vs;
 181                 unsigned int start;
 182                 size_t offset;
 183
 184                 do {
 185                         start = u64_stats_fetch_begin_irq(&rq_stats->syncp);
 186                         for (j = 0; j < VETH_RQ_STATS_LEN; j++) {
 187                                 offset = veth_rq_stats_desc[j].offset;
 188                                 data[idx + j] = *(u64 *)(stats_base + offset);
 189                         }
 190                 } while (u64_stats_fetch_retry_irq(&rq_stats->syncp, start));
 191                 idx += VETH_RQ_STATS_LEN;
 192         }
 193
 194         if (!peer)
 195                 return;
 196
 197         rcv_priv = netdev_priv(peer);
 198         for (i = 0; i < peer->real_num_rx_queues; i++) {
 199                 const struct veth_rq_stats *rq_stats = &rcv_priv->rq[i].stats;
 200                 const void *base = (void *)&rq_stats->vs;
 201                 unsigned int start, tx_idx = idx;
 202                 size_t offset;
 203
 204                 tx_idx += (i % dev->real_num_tx_queues) * VETH_TQ_STATS_LEN;
 205                 do {
 206                         start = u64_stats_fetch_begin_irq(&rq_stats->syncp);
 207                         for (j = 0; j < VETH_TQ_STATS_LEN; j++) {
 208                                 offset = veth_tq_stats_desc[j].offset;
 209                                 data[tx_idx + j] += *(u64 *)(base + offset);
 210                         }
 211                 } while (u64_stats_fetch_retry_irq(&rq_stats->syncp, start));
 212         }
 213 }
 214
 215 static void veth_get_channels(struct net_device *dev,
 216                               struct ethtool_channels *channels)
 217 {
 218         channels->tx_count = dev->real_num_tx_queues;
 219         channels->rx_count = dev->real_num_rx_queues;
 220         channels->max_tx = dev->num_tx_queues;
 221         channels->max_rx = dev->num_rx_queues;
 222 }
 223
 224 static int veth_set_channels(struct net_device *dev,
 225                              struct ethtool_channels *ch);
 226
 227 static const struct ethtool_ops veth_ethtool_ops = {
 228         .get_drvinfo            = veth_get_drvinfo,
 229         .get_link               = ethtool_op_get_link,
 230         .get_strings            = veth_get_strings,
 231         .get_sset_count         = veth_get_sset_count,
 232         .get_ethtool_stats      = veth_get_ethtool_stats,
 233         .get_link_ksettings     = veth_get_link_ksettings,
 234         .get_ts_info            = ethtool_op_get_ts_info,
 235         .get_channels           = veth_get_channels,
 236         .set_channels           = veth_set_channels,
 237 };
 238
 239 /* general routines */
 240
 241 static bool veth_is_xdp_frame(void *ptr)
 242 {
 243         return (unsigned long)ptr & VETH_XDP_FLAG;
 244 }
 245
 246 static struct xdp_frame *veth_ptr_to_xdp(void *ptr)
 247 {
 248         return (void *)((unsigned long)ptr & ~VETH_XDP_FLAG);
 249 }
 250
 251 static void *veth_xdp_to_ptr(struct xdp_frame *xdp)
 252 {
 253         return (void *)((unsigned long)xdp | VETH_XDP_FLAG);
 254 }
 255
 256 static void veth_ptr_free(void *ptr)
 257 {
 258         if (veth_is_xdp_frame(ptr))
 259                 xdp_return_frame(veth_ptr_to_xdp(ptr));
 260         else
 261                 kfree_skb(ptr);
 262 }
 263
 264 static void __veth_xdp_flush(struct veth_rq *rq)
 265 {
 266         /* Write ptr_ring before reading rx_notify_masked */
 267         smp_mb();
 268         if (!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 netdev_queue *queue = NULL;
 316         struct veth_rq *rq = NULL;
 317         struct net_device *rcv;
 318         int length = skb->len;
 319         bool use_napi = false;
 320         int rxq;
 321
 322         rcu_read_lock();
 323         rcv = rcu_dereference(priv->peer);
 324         if (unlikely(!rcv) || !pskb_may_pull(skb, ETH_HLEN)) {
 325                 kfree_skb(skb);
 326                 goto drop;
 327         }
 328
 329         rcv_priv = netdev_priv(rcv);
 330         rxq = skb_get_queue_mapping(skb);
 331         if (rxq < rcv->real_num_rx_queues) {
 332                 rq = &rcv_priv->rq[rxq];
 333                 queue = netdev_get_tx_queue(dev, rxq);
 334
 335                 /* The napi pointer is available when an XDP program is
 336                  * attached or when GRO is enabled
 337                  * Don't bother with napi/GRO if the skb can't be aggregated
 338                  */
 339                 use_napi = rcu_access_pointer(rq->napi) &&
 340                            veth_skb_is_eligible_for_gro(dev, rcv, skb);
 341         }
 342
 343         skb_tx_timestamp(skb);
 344         if (likely(veth_forward_skb(rcv, skb, rq, use_napi) == NET_RX_SUCCESS)) {
 345                 if (queue)
 346                         txq_trans_cond_update(queue);
 347                 if (!use_napi)
 348                         dev_lstats_add(dev, length);
 349         } else {
 350 drop:
 351                 atomic64_inc(&priv->dropped);
 352         }
 353
 354         if (use_napi)
 355                 __veth_xdp_flush(rq);
 356
 357         rcu_read_unlock();
 358
 359         return NETDEV_TX_OK;
 360 }
 361
 362 static u64 veth_stats_tx(struct net_device *dev, u64 *packets, u64 *bytes)
 363 {
 364         struct veth_priv *priv = netdev_priv(dev);
 365
 366         dev_lstats_read(dev, packets, bytes);
 367         return atomic64_read(&priv->dropped);
 368 }
 369
 370 static void veth_stats_rx(struct veth_stats *result, struct net_device *dev)
 371 {
 372         struct veth_priv *priv = netdev_priv(dev);
 373         int i;
 374
 375         result->peer_tq_xdp_xmit_err = 0;
 376         result->xdp_packets = 0;
 377         result->xdp_tx_err = 0;
 378         result->xdp_bytes = 0;
 379         result->rx_drops = 0;
 380         for (i = 0; i < dev->num_rx_queues; i++) {
 381                 u64 packets, bytes, drops, xdp_tx_err, peer_tq_xdp_xmit_err;
 382                 struct veth_rq_stats *stats = &priv->rq[i].stats;
 383                 unsigned int start;
 384
 385                 do {
 386                         start = u64_stats_fetch_begin_irq(&stats->syncp);
 387                         peer_tq_xdp_xmit_err = stats->vs.peer_tq_xdp_xmit_err;
 388                         xdp_tx_err = stats->vs.xdp_tx_err;
 389                         packets = stats->vs.xdp_packets;
 390                         bytes = stats->vs.xdp_bytes;
 391                         drops = stats->vs.rx_drops;
 392                 } while (u64_stats_fetch_retry_irq(&stats->syncp, start));
 393                 result->peer_tq_xdp_xmit_err += peer_tq_xdp_xmit_err;
 394                 result->xdp_tx_err += xdp_tx_err;
 395                 result->xdp_packets += packets;
 396                 result->xdp_bytes += bytes;
 397                 result->rx_drops += drops;
 398         }
 399 }
 400
 401 static void veth_get_stats64(struct net_device *dev,
 402                              struct rtnl_link_stats64 *tot)
 403 {
 404         struct veth_priv *priv = netdev_priv(dev);
 405         struct net_device *peer;
 406         struct veth_stats rx;
 407         u64 packets, bytes;
 408
 409         tot->tx_dropped = veth_stats_tx(dev, &packets, &bytes);
 410         tot->tx_bytes = bytes;
 411         tot->tx_packets = packets;
 412
 413         veth_stats_rx(&rx, dev);
 414         tot->tx_dropped += rx.xdp_tx_err;
 415         tot->rx_dropped = rx.rx_drops + rx.peer_tq_xdp_xmit_err;
 416         tot->rx_bytes = rx.xdp_bytes;
 417         tot->rx_packets = rx.xdp_packets;
 418
 419         rcu_read_lock();
 420         peer = rcu_dereference(priv->peer);
 421         if (peer) {
 422                 veth_stats_tx(peer, &packets, &bytes);
 423                 tot->rx_bytes += bytes;
 424                 tot->rx_packets += packets;
 425
 426                 veth_stats_rx(&rx, peer);
 427                 tot->tx_dropped += rx.peer_tq_xdp_xmit_err;
 428                 tot->rx_dropped += rx.xdp_tx_err;
 429                 tot->tx_bytes += rx.xdp_bytes;
 430                 tot->tx_packets += rx.xdp_packets;
 431         }
 432         rcu_read_unlock();
 433 }
 434
 435 /* fake multicast ability */
 436 static void veth_set_multicast_list(struct net_device *dev)
 437 {
 438 }
 439
 440 static int veth_select_rxq(struct net_device *dev)
 441 {
 442         return smp_processor_id() % dev->real_num_rx_queues;
 443 }
 444
 445 static struct net_device *veth_peer_dev(struct net_device *dev)
 446 {
 447         struct veth_priv *priv = netdev_priv(dev);
 448
 449         /* Callers must be under RCU read side. */
 450         return rcu_dereference(priv->peer);
 451 }
 452
 453 static int veth_xdp_xmit(struct net_device *dev, int n,
 454                          struct xdp_frame **frames,
 455                          u32 flags, bool ndo_xmit)
 456 {
 457         struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 458         int i, ret = -ENXIO, nxmit = 0;
 459         struct net_device *rcv;
 460         unsigned int max_len;
 461         struct veth_rq *rq;
 462
 463         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
 464                 return -EINVAL;
 465
 466         rcu_read_lock();
 467         rcv = rcu_dereference(priv->peer);
 468         if (unlikely(!rcv))
 469                 goto out;
 470
 471         rcv_priv = netdev_priv(rcv);
 472         rq = &rcv_priv->rq[veth_select_rxq(rcv)];
 473         /* The napi pointer is set if NAPI is enabled, which ensures that
 474          * xdp_ring is initialized on receive side and the peer device is up.
 475          */
 476         if (!rcu_access_pointer(rq->napi))
 477                 goto out;
 478
 479         max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN;
 480
 481         spin_lock(&rq->xdp_ring.producer_lock);
 482         for (i = 0; i < n; i++) {
 483                 struct xdp_frame *frame = frames[i];
 484                 void *ptr = veth_xdp_to_ptr(frame);
 485
 486                 if (unlikely(xdp_get_frame_len(frame) > max_len ||
 487                              __ptr_ring_produce(&rq->xdp_ring, ptr)))
 488                         break;
 489                 nxmit++;
 490         }
 491         spin_unlock(&rq->xdp_ring.producer_lock);
 492
 493         if (flags & XDP_XMIT_FLUSH)
 494                 __veth_xdp_flush(rq);
 495
 496         ret = nxmit;
 497         if (ndo_xmit) {
 498                 u64_stats_update_begin(&rq->stats.syncp);
 499                 rq->stats.vs.peer_tq_xdp_xmit += nxmit;
 500                 rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit;
 501                 u64_stats_update_end(&rq->stats.syncp);
 502         }
 503
 504 out:
 505         rcu_read_unlock();
 506
 507         return ret;
 508 }
 509
 510 static int veth_ndo_xdp_xmit(struct net_device *dev, int n,
 511                              struct xdp_frame **frames, u32 flags)
 512 {
 513         int err;
 514
 515         err = veth_xdp_xmit(dev, n, frames, flags, true);
 516         if (err < 0) {
 517                 struct veth_priv *priv = netdev_priv(dev);
 518
 519                 atomic64_add(n, &priv->dropped);
 520         }
 521
 522         return err;
 523 }
 524
 525 static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
 526 {
 527         int sent, i, err = 0, drops;
 528
 529         sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false);
 530         if (sent < 0) {
 531                 err = sent;
 532                 sent = 0;
 533         }
 534
 535         for (i = sent; unlikely(i < bq->count); i++)
 536                 xdp_return_frame(bq->q[i]);
 537
 538         drops = bq->count - sent;
 539         trace_xdp_bulk_tx(rq->dev, sent, drops, err);
 540
 541         u64_stats_update_begin(&rq->stats.syncp);
 542         rq->stats.vs.xdp_tx += sent;
 543         rq->stats.vs.xdp_tx_err += drops;
 544         u64_stats_update_end(&rq->stats.syncp);
 545
 546         bq->count = 0;
 547 }
 548
 549 static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
 550 {
 551         struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev);
 552         struct net_device *rcv;
 553         struct veth_rq *rcv_rq;
 554
 555         rcu_read_lock();
 556         veth_xdp_flush_bq(rq, bq);
 557         rcv = rcu_dereference(priv->peer);
 558         if (unlikely(!rcv))
 559                 goto out;
 560
 561         rcv_priv = netdev_priv(rcv);
 562         rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)];
 563         /* xdp_ring is initialized on receive side? */
 564         if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog)))
 565                 goto out;
 566
 567         __veth_xdp_flush(rcv_rq);
 568 out:
 569         rcu_read_unlock();
 570 }
 571
 572 static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp,
 573                        struct veth_xdp_tx_bq *bq)
 574 {
 575         struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp);
 576
 577         if (unlikely(!frame))
 578                 return -EOVERFLOW;
 579
 580         if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE))
 581                 veth_xdp_flush_bq(rq, bq);
 582
 583         bq->q[bq->count++] = frame;
 584
 585         return 0;
 586 }
 587
 588 static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq,
 589                                           struct xdp_frame *frame,
 590                                           struct veth_xdp_tx_bq *bq,
 591                                           struct veth_stats *stats)
 592 {
 593         struct xdp_frame orig_frame;
 594         struct bpf_prog *xdp_prog;
 595
 596         rcu_read_lock();
 597         xdp_prog = rcu_dereference(rq->xdp_prog);
 598         if (likely(xdp_prog)) {
 599                 struct xdp_buff xdp;
 600                 u32 act;
 601
 602                 xdp_convert_frame_to_buff(frame, &xdp);
 603                 xdp.rxq = &rq->xdp_rxq;
 604
 605                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 606
 607                 switch (act) {
 608                 case XDP_PASS:
 609                         if (xdp_update_frame_from_buff(&xdp, frame))
 610                                 goto err_xdp;
 611                         break;
 612                 case XDP_TX:
 613                         orig_frame = *frame;
 614                         xdp.rxq->mem = frame->mem;
 615                         if (unlikely(veth_xdp_tx(rq, &xdp, bq) < 0)) {
 616                                 trace_xdp_exception(rq->dev, xdp_prog, act);
 617                                 frame = &orig_frame;
 618                                 stats->rx_drops++;
 619                                 goto err_xdp;
 620                         }
 621                         stats->xdp_tx++;
 622                         rcu_read_unlock();
 623                         goto xdp_xmit;
 624                 case XDP_REDIRECT:
 625                         orig_frame = *frame;
 626                         xdp.rxq->mem = frame->mem;
 627                         if (xdp_do_redirect(rq->dev, &xdp, xdp_prog)) {
 628                                 frame = &orig_frame;
 629                                 stats->rx_drops++;
 630                                 goto err_xdp;
 631                         }
 632                         stats->xdp_redirect++;
 633                         rcu_read_unlock();
 634                         goto xdp_xmit;
 635                 default:
 636                         bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
 637                         fallthrough;
 638                 case XDP_ABORTED:
 639                         trace_xdp_exception(rq->dev, xdp_prog, act);
 640                         fallthrough;
 641                 case XDP_DROP:
 642                         stats->xdp_drops++;
 643                         goto err_xdp;
 644                 }
 645         }
 646         rcu_read_unlock();
 647
 648         return frame;
 649 err_xdp:
 650         rcu_read_unlock();
 651         xdp_return_frame(frame);
 652 xdp_xmit:
 653         return NULL;
 654 }
 655
 656 /* frames array contains VETH_XDP_BATCH at most */
 657 static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames,
 658                                   int n_xdpf, struct veth_xdp_tx_bq *bq,
 659                                   struct veth_stats *stats)
 660 {
 661         void *skbs[VETH_XDP_BATCH];
 662         int i;
 663
 664         if (xdp_alloc_skb_bulk(skbs, n_xdpf,
 665                                GFP_ATOMIC | __GFP_ZERO) < 0) {
 666                 for (i = 0; i < n_xdpf; i++)
 667                         xdp_return_frame(frames[i]);
 668                 stats->rx_drops += n_xdpf;
 669
 670                 return;
 671         }
 672
 673         for (i = 0; i < n_xdpf; i++) {
 674                 struct sk_buff *skb = skbs[i];
 675
 676                 skb = __xdp_build_skb_from_frame(frames[i], skb,
 677                                                  rq->dev);
 678                 if (!skb) {
 679                         xdp_return_frame(frames[i]);
 680                         stats->rx_drops++;
 681                         continue;
 682                 }
 683                 napi_gro_receive(&rq->xdp_napi, skb);
 684         }
 685 }
 686
 687 static void veth_xdp_get(struct xdp_buff *xdp)
 688 {
 689         struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
 690         int i;
 691
 692         get_page(virt_to_page(xdp->data));
 693         if (likely(!xdp_buff_has_frags(xdp)))
 694                 return;
 695
 696         for (i = 0; i < sinfo->nr_frags; i++)
 697                 __skb_frag_ref(&sinfo->frags[i]);
 698 }
 699
 700 static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
 701                                         struct xdp_buff *xdp,
 702                                         struct sk_buff **pskb)
 703 {
 704         struct sk_buff *skb = *pskb;
 705         u32 frame_sz;
 706
 707         if (skb_shared(skb) || skb_head_is_locked(skb) ||
 708             skb_shinfo(skb)->nr_frags) {
 709                 u32 size, len, max_head_size, off;
 710                 struct sk_buff *nskb;
 711                 struct page *page;
 712                 int i, head_off;
 713
 714                 /* We need a private copy of the skb and data buffers since
 715                  * the ebpf program can modify it. We segment the original skb
 716                  * into order-0 pages without linearize it.
 717                  *
 718                  * Make sure we have enough space for linear and paged area
 719                  */
 720                 max_head_size = SKB_WITH_OVERHEAD(PAGE_SIZE -
 721                                                   VETH_XDP_HEADROOM);
 722                 if (skb->len > PAGE_SIZE * MAX_SKB_FRAGS + max_head_size)
 723                         goto drop;
 724
 725                 /* Allocate skb head */
 726                 page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
 727                 if (!page)
 728                         goto drop;
 729
 730                 nskb = build_skb(page_address(page), PAGE_SIZE);
 731                 if (!nskb) {
 732                         put_page(page);
 733                         goto drop;
 734                 }
 735
 736                 skb_reserve(nskb, VETH_XDP_HEADROOM);
 737                 size = min_t(u32, skb->len, max_head_size);
 738                 if (skb_copy_bits(skb, 0, nskb->data, size)) {
 739                         consume_skb(nskb);
 740                         goto drop;
 741                 }
 742                 skb_put(nskb, size);
 743
 744                 skb_copy_header(nskb, skb);
 745                 head_off = skb_headroom(nskb) - skb_headroom(skb);
 746                 skb_headers_offset_update(nskb, head_off);
 747
 748                 /* Allocate paged area of new skb */
 749                 off = size;
 750                 len = skb->len - off;
 751
 752                 for (i = 0; i < MAX_SKB_FRAGS && off < skb->len; i++) {
 753                         page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
 754                         if (!page) {
 755                                 consume_skb(nskb);
 756                                 goto drop;
 757                         }
 758
 759                         size = min_t(u32, len, PAGE_SIZE);
 760                         skb_add_rx_frag(nskb, i, page, 0, size, PAGE_SIZE);
 761                         if (skb_copy_bits(skb, off, page_address(page),
 762                                           size)) {
 763                                 consume_skb(nskb);
 764                                 goto drop;
 765                         }
 766
 767                         len -= size;
 768                         off += size;
 769                 }
 770
 771                 consume_skb(skb);
 772                 skb = nskb;
 773         } else if (skb_headroom(skb) < XDP_PACKET_HEADROOM &&
 774                    pskb_expand_head(skb, VETH_XDP_HEADROOM, 0, GFP_ATOMIC)) {
 775                 goto drop;
 776         }
 777
 778         /* SKB "head" area always have tailroom for skb_shared_info */
 779         frame_sz = skb_end_pointer(skb) - skb->head;
 780         frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 781         xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
 782         xdp_prepare_buff(xdp, skb->head, skb_headroom(skb),
 783                          skb_headlen(skb), true);
 784
 785         if (skb_is_nonlinear(skb)) {
 786                 skb_shinfo(skb)->xdp_frags_size = skb->data_len;
 787                 xdp_buff_set_frags_flag(xdp);
 788         } else {
 789                 xdp_buff_clear_frags_flag(xdp);
 790         }
 791         *pskb = skb;
 792
 793         return 0;
 794 drop:
 795         consume_skb(skb);
 796         *pskb = NULL;
 797
 798         return -ENOMEM;
 799 }
 800
 801 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq,
 802                                         struct sk_buff *skb,
 803                                         struct veth_xdp_tx_bq *bq,
 804                                         struct veth_stats *stats)
 805 {
 806         void *orig_data, *orig_data_end;
 807         struct bpf_prog *xdp_prog;
 808         struct xdp_buff xdp;
 809         u32 act, metalen;
 810         int off;
 811
 812         skb_prepare_for_gro(skb);
 813
 814         rcu_read_lock();
 815         xdp_prog = rcu_dereference(rq->xdp_prog);
 816         if (unlikely(!xdp_prog)) {
 817                 rcu_read_unlock();
 818                 goto out;
 819         }
 820
 821         __skb_push(skb, skb->data - skb_mac_header(skb));
 822         if (veth_convert_skb_to_xdp_buff(rq, &xdp, &skb))
 823                 goto drop;
 824
 825         orig_data = xdp.data;
 826         orig_data_end = xdp.data_end;
 827
 828         act = bpf_prog_run_xdp(xdp_prog, &xdp);
 829
 830         switch (act) {
 831         case XDP_PASS:
 832                 break;
 833         case XDP_TX:
 834                 veth_xdp_get(&xdp);
 835                 consume_skb(skb);
 836                 xdp.rxq->mem = rq->xdp_mem;
 837                 if (unlikely(veth_xdp_tx(rq, &xdp, bq) < 0)) {
 838                         trace_xdp_exception(rq->dev, xdp_prog, act);
 839                         stats->rx_drops++;
 840                         goto err_xdp;
 841                 }
 842                 stats->xdp_tx++;
 843                 rcu_read_unlock();
 844                 goto xdp_xmit;
 845         case XDP_REDIRECT:
 846                 veth_xdp_get(&xdp);
 847                 consume_skb(skb);
 848                 xdp.rxq->mem = rq->xdp_mem;
 849                 if (xdp_do_redirect(rq->dev, &xdp, xdp_prog)) {
 850                         stats->rx_drops++;
 851                         goto err_xdp;
 852                 }
 853                 stats->xdp_redirect++;
 854                 rcu_read_unlock();
 855                 goto xdp_xmit;
 856         default:
 857                 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
 858                 fallthrough;
 859         case XDP_ABORTED:
 860                 trace_xdp_exception(rq->dev, xdp_prog, act);
 861                 fallthrough;
 862         case XDP_DROP:
 863                 stats->xdp_drops++;
 864                 goto xdp_drop;
 865         }
 866         rcu_read_unlock();
 867
 868         /* check if bpf_xdp_adjust_head was used */
 869         off = orig_data - xdp.data;
 870         if (off > 0)
 871                 __skb_push(skb, off);
 872         else if (off < 0)
 873                 __skb_pull(skb, -off);
 874
 875         skb_reset_mac_header(skb);
 876
 877         /* check if bpf_xdp_adjust_tail was used */
 878         off = xdp.data_end - orig_data_end;
 879         if (off != 0)
 880                 __skb_put(skb, off); /* positive on grow, negative on shrink */
 881
 882         /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers
 883          * (e.g. bpf_xdp_adjust_tail), we need to update data_len here.
 884          */
 885         if (xdp_buff_has_frags(&xdp))
 886                 skb->data_len = skb_shinfo(skb)->xdp_frags_size;
 887         else
 888                 skb->data_len = 0;
 889
 890         skb->protocol = eth_type_trans(skb, rq->dev);
 891
 892         metalen = xdp.data - xdp.data_meta;
 893         if (metalen)
 894                 skb_metadata_set(skb, metalen);
 895 out:
 896         return skb;
 897 drop:
 898         stats->rx_drops++;
 899 xdp_drop:
 900         rcu_read_unlock();
 901         kfree_skb(skb);
 902         return NULL;
 903 err_xdp:
 904         rcu_read_unlock();
 905         xdp_return_buff(&xdp);
 906 xdp_xmit:
 907         return NULL;
 908 }
 909
 910 static int veth_xdp_rcv(struct veth_rq *rq, int budget,
 911                         struct veth_xdp_tx_bq *bq,
 912                         struct veth_stats *stats)
 913 {
 914         int i, done = 0, n_xdpf = 0;
 915         void *xdpf[VETH_XDP_BATCH];
 916
 917         for (i = 0; i < budget; i++) {
 918                 void *ptr = __ptr_ring_consume(&rq->xdp_ring);
 919
 920                 if (!ptr)
 921                         break;
 922
 923                 if (veth_is_xdp_frame(ptr)) {
 924                         /* ndo_xdp_xmit */
 925                         struct xdp_frame *frame = veth_ptr_to_xdp(ptr);
 926
 927                         stats->xdp_bytes += xdp_get_frame_len(frame);
 928                         frame = veth_xdp_rcv_one(rq, frame, bq, stats);
 929                         if (frame) {
 930                                 /* XDP_PASS */
 931                                 xdpf[n_xdpf++] = frame;
 932                                 if (n_xdpf == VETH_XDP_BATCH) {
 933                                         veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf,
 934                                                               bq, stats);
 935                                         n_xdpf = 0;
 936                                 }
 937                         }
 938                 } else {
 939                         /* ndo_start_xmit */
 940                         struct sk_buff *skb = ptr;
 941
 942                         stats->xdp_bytes += skb->len;
 943                         skb = veth_xdp_rcv_skb(rq, skb, bq, stats);
 944                         if (skb) {
 945                                 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC))
 946                                         netif_receive_skb(skb);
 947                                 else
 948                                         napi_gro_receive(&rq->xdp_napi, skb);
 949                         }
 950                 }
 951                 done++;
 952         }
 953
 954         if (n_xdpf)
 955                 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats);
 956
 957         u64_stats_update_begin(&rq->stats.syncp);
 958         rq->stats.vs.xdp_redirect += stats->xdp_redirect;
 959         rq->stats.vs.xdp_bytes += stats->xdp_bytes;
 960         rq->stats.vs.xdp_drops += stats->xdp_drops;
 961         rq->stats.vs.rx_drops += stats->rx_drops;
 962         rq->stats.vs.xdp_packets += done;
 963         u64_stats_update_end(&rq->stats.syncp);
 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 (done < budget && napi_complete_done(napi, done)) {
 982                 /* Write rx_notify_masked before reading ptr_ring */
 983                 smp_store_mb(rq->rx_notify_masked, false);
 984                 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
 985                         if (napi_schedule_prep(&rq->xdp_napi)) {
 986                                 WRITE_ONCE(rq->rx_notify_masked, true);
 987                                 __napi_schedule(&rq->xdp_napi);
 988                         }
 989                 }
 990         }
 991
 992         if (stats.xdp_tx > 0)
 993                 veth_xdp_flush(rq, &bq);
 994         if (stats.xdp_redirect > 0)
 995                 xdp_do_flush();
 996         xdp_clear_return_frame_no_direct();
 997
 998         return done;
 999 }
1000
1001 static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
1002 {
1003         struct veth_priv *priv = netdev_priv(dev);
1004         int err, i;
1005
1006         for (i = start; i < end; i++) {
1007                 struct veth_rq *rq = &priv->rq[i];
1008
1009                 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
1010                 if (err)
1011                         goto err_xdp_ring;
1012         }
1013
1014         for (i = start; i < end; i++) {
1015                 struct veth_rq *rq = &priv->rq[i];
1016
1017                 napi_enable(&rq->xdp_napi);
1018                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1019         }
1020
1021         return 0;
1022
1023 err_xdp_ring:
1024         for (i--; i >= start; i--)
1025                 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
1026
1027         return err;
1028 }
1029
1030 static int __veth_napi_enable(struct net_device *dev)
1031 {
1032         return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1033 }
1034
1035 static void veth_napi_del_range(struct net_device *dev, int start, int end)
1036 {
1037         struct veth_priv *priv = netdev_priv(dev);
1038         int i;
1039
1040         for (i = start; i < end; i++) {
1041                 struct veth_rq *rq = &priv->rq[i];
1042
1043                 rcu_assign_pointer(priv->rq[i].napi, NULL);
1044                 napi_disable(&rq->xdp_napi);
1045                 __netif_napi_del(&rq->xdp_napi);
1046         }
1047         synchronize_net();
1048
1049         for (i = start; i < end; i++) {
1050                 struct veth_rq *rq = &priv->rq[i];
1051
1052                 rq->rx_notify_masked = false;
1053                 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
1054         }
1055 }
1056
1057 static void veth_napi_del(struct net_device *dev)
1058 {
1059         veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
1060 }
1061
1062 static bool veth_gro_requested(const struct net_device *dev)
1063 {
1064         return !!(dev->wanted_features & NETIF_F_GRO);
1065 }
1066
1067 static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1068                                  bool napi_already_on)
1069 {
1070         struct veth_priv *priv = netdev_priv(dev);
1071         int err, i;
1072
1073         for (i = start; i < end; i++) {
1074                 struct veth_rq *rq = &priv->rq[i];
1075
1076                 if (!napi_already_on)
1077                         netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1078                 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1079                 if (err < 0)
1080                         goto err_rxq_reg;
1081
1082                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1083                                                  MEM_TYPE_PAGE_SHARED,
1084                                                  NULL);
1085                 if (err < 0)
1086                         goto err_reg_mem;
1087
1088                 /* Save original mem info as it can be overwritten */
1089                 rq->xdp_mem = rq->xdp_rxq.mem;
1090         }
1091         return 0;
1092
1093 err_reg_mem:
1094         xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1095 err_rxq_reg:
1096         for (i--; i >= start; i--) {
1097                 struct veth_rq *rq = &priv->rq[i];
1098
1099                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1100                 if (!napi_already_on)
1101                         netif_napi_del(&rq->xdp_napi);
1102         }
1103
1104         return err;
1105 }
1106
1107 static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1108                                    bool delete_napi)
1109 {
1110         struct veth_priv *priv = netdev_priv(dev);
1111         int i;
1112
1113         for (i = start; i < end; i++) {
1114                 struct veth_rq *rq = &priv->rq[i];
1115
1116                 rq->xdp_rxq.mem = rq->xdp_mem;
1117                 xdp_rxq_info_unreg(&rq->xdp_rxq);
1118
1119                 if (delete_napi)
1120                         netif_napi_del(&rq->xdp_napi);
1121         }
1122 }
1123
1124 static int veth_enable_xdp(struct net_device *dev)
1125 {
1126         bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1127         struct veth_priv *priv = netdev_priv(dev);
1128         int err, i;
1129
1130         if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1131                 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1132                 if (err)
1133                         return err;
1134
1135                 if (!napi_already_on) {
1136                         err = __veth_napi_enable(dev);
1137                         if (err) {
1138                                 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1139                                 return err;
1140                         }
1141
1142                         if (!veth_gro_requested(dev)) {
1143                                 /* user-space did not require GRO, but adding XDP
1144                                  * is supposed to get GRO working
1145                                  */
1146                                 dev->features |= NETIF_F_GRO;
1147                                 netdev_features_change(dev);
1148                         }
1149                 }
1150         }
1151
1152         for (i = 0; i < dev->real_num_rx_queues; i++) {
1153                 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1154                 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1155         }
1156
1157         return 0;
1158 }
1159
1160 static void veth_disable_xdp(struct net_device *dev)
1161 {
1162         struct veth_priv *priv = netdev_priv(dev);
1163         int i;
1164
1165         for (i = 0; i < dev->real_num_rx_queues; i++)
1166                 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1167
1168         if (!netif_running(dev) || !veth_gro_requested(dev)) {
1169                 veth_napi_del(dev);
1170
1171                 /* if user-space did not require GRO, since adding XDP
1172                  * enabled it, clear it now
1173                  */
1174                 if (!veth_gro_requested(dev) && netif_running(dev)) {
1175                         dev->features &= ~NETIF_F_GRO;
1176                         netdev_features_change(dev);
1177                 }
1178         }
1179
1180         veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1181 }
1182
1183 static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1184 {
1185         struct veth_priv *priv = netdev_priv(dev);
1186         int err, i;
1187
1188         for (i = start; i < end; i++) {
1189                 struct veth_rq *rq = &priv->rq[i];
1190
1191                 netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1192         }
1193
1194         err = __veth_napi_enable_range(dev, start, end);
1195         if (err) {
1196                 for (i = start; i < end; i++) {
1197                         struct veth_rq *rq = &priv->rq[i];
1198
1199                         netif_napi_del(&rq->xdp_napi);
1200                 }
1201                 return err;
1202         }
1203         return err;
1204 }
1205
1206 static int veth_napi_enable(struct net_device *dev)
1207 {
1208         return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1209 }
1210
1211 static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1212 {
1213         struct veth_priv *priv = netdev_priv(dev);
1214
1215         if (start >= end)
1216                 return;
1217
1218         if (priv->_xdp_prog) {
1219                 veth_napi_del_range(dev, start, end);
1220                 veth_disable_xdp_range(dev, start, end, false);
1221         } else if (veth_gro_requested(dev)) {
1222                 veth_napi_del_range(dev, start, end);
1223         }
1224 }
1225
1226 static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1227 {
1228         struct veth_priv *priv = netdev_priv(dev);
1229         int err;
1230
1231         if (start >= end)
1232                 return 0;
1233
1234         if (priv->_xdp_prog) {
1235                 /* these channels are freshly initialized, napi is not on there even
1236                  * when GRO is requeste
1237                  */
1238                 err = veth_enable_xdp_range(dev, start, end, false);
1239                 if (err)
1240                         return err;
1241
1242                 err = __veth_napi_enable_range(dev, start, end);
1243                 if (err) {
1244                         /* on error always delete the newly added napis */
1245                         veth_disable_xdp_range(dev, start, end, true);
1246                         return err;
1247                 }
1248         } else if (veth_gro_requested(dev)) {
1249                 return veth_napi_enable_range(dev, start, end);
1250         }
1251         return 0;
1252 }
1253
1254 static int veth_set_channels(struct net_device *dev,
1255                              struct ethtool_channels *ch)
1256 {
1257         struct veth_priv *priv = netdev_priv(dev);
1258         unsigned int old_rx_count, new_rx_count;
1259         struct veth_priv *peer_priv;
1260         struct net_device *peer;
1261         int err;
1262
1263         /* sanity check. Upper bounds are already enforced by the caller */
1264         if (!ch->rx_count || !ch->tx_count)
1265                 return -EINVAL;
1266
1267         /* avoid braking XDP, if that is enabled */
1268         peer = rtnl_dereference(priv->peer);
1269         peer_priv = peer ? netdev_priv(peer) : NULL;
1270         if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1271                 return -EINVAL;
1272
1273         if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1274                 return -EINVAL;
1275
1276         old_rx_count = dev->real_num_rx_queues;
1277         new_rx_count = ch->rx_count;
1278         if (netif_running(dev)) {
1279                 /* turn device off */
1280                 netif_carrier_off(dev);
1281                 if (peer)
1282                         netif_carrier_off(peer);
1283
1284                 /* try to allocate new resurces, as needed*/
1285                 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1286                 if (err)
1287                         goto out;
1288         }
1289
1290         err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1291         if (err)
1292                 goto revert;
1293
1294         err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1295         if (err) {
1296                 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1297
1298                 /* this error condition could happen only if rx and tx change
1299                  * in opposite directions (e.g. tx nr raises, rx nr decreases)
1300                  * and we can't do anything to fully restore the original
1301                  * status
1302                  */
1303                 if (err2)
1304                         pr_warn("Can't restore rx queues config %d -> %d %d",
1305                                 new_rx_count, old_rx_count, err2);
1306                 else
1307                         goto revert;
1308         }
1309
1310 out:
1311         if (netif_running(dev)) {
1312                 /* note that we need to swap the arguments WRT the enable part
1313                  * to identify the range we have to disable
1314                  */
1315                 veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1316                 netif_carrier_on(dev);
1317                 if (peer)
1318                         netif_carrier_on(peer);
1319         }
1320         return err;
1321
1322 revert:
1323         new_rx_count = old_rx_count;
1324         old_rx_count = ch->rx_count;
1325         goto out;
1326 }
1327
1328 static int veth_open(struct net_device *dev)
1329 {
1330         struct veth_priv *priv = netdev_priv(dev);
1331         struct net_device *peer = rtnl_dereference(priv->peer);
1332         int err;
1333
1334         if (!peer)
1335                 return -ENOTCONN;
1336
1337         if (priv->_xdp_prog) {
1338                 err = veth_enable_xdp(dev);
1339                 if (err)
1340                         return err;
1341         } else if (veth_gro_requested(dev)) {
1342                 err = veth_napi_enable(dev);
1343                 if (err)
1344                         return err;
1345         }
1346
1347         if (peer->flags & IFF_UP) {
1348                 netif_carrier_on(dev);
1349                 netif_carrier_on(peer);
1350         }
1351
1352         return 0;
1353 }
1354
1355 static int veth_close(struct net_device *dev)
1356 {
1357         struct veth_priv *priv = netdev_priv(dev);
1358         struct net_device *peer = rtnl_dereference(priv->peer);
1359
1360         netif_carrier_off(dev);
1361         if (peer)
1362                 netif_carrier_off(peer);
1363
1364         if (priv->_xdp_prog)
1365                 veth_disable_xdp(dev);
1366         else if (veth_gro_requested(dev))
1367                 veth_napi_del(dev);
1368
1369         return 0;
1370 }
1371
1372 static int is_valid_veth_mtu(int mtu)
1373 {
1374         return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1375 }
1376
1377 static int veth_alloc_queues(struct net_device *dev)
1378 {
1379         struct veth_priv *priv = netdev_priv(dev);
1380         int i;
1381
1382         priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL_ACCOUNT);
1383         if (!priv->rq)
1384                 return -ENOMEM;
1385
1386         for (i = 0; i < dev->num_rx_queues; i++) {
1387                 priv->rq[i].dev = dev;
1388                 u64_stats_init(&priv->rq[i].stats.syncp);
1389         }
1390
1391         return 0;
1392 }
1393
1394 static void veth_free_queues(struct net_device *dev)
1395 {
1396         struct veth_priv *priv = netdev_priv(dev);
1397
1398         kfree(priv->rq);
1399 }
1400
1401 static int veth_dev_init(struct net_device *dev)
1402 {
1403         int err;
1404
1405         dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
1406         if (!dev->lstats)
1407                 return -ENOMEM;
1408
1409         err = veth_alloc_queues(dev);
1410         if (err) {
1411                 free_percpu(dev->lstats);
1412                 return err;
1413         }
1414
1415         return 0;
1416 }
1417
1418 static void veth_dev_free(struct net_device *dev)
1419 {
1420         veth_free_queues(dev);
1421         free_percpu(dev->lstats);
1422 }
1423
1424 #ifdef CONFIG_NET_POLL_CONTROLLER
1425 static void veth_poll_controller(struct net_device *dev)
1426 {
1427         /* veth only receives frames when its peer sends one
1428          * Since it has nothing to do with disabling irqs, we are guaranteed
1429          * never to have pending data when we poll for it so
1430          * there is nothing to do here.
1431          *
1432          * We need this though so netpoll recognizes us as an interface that
1433          * supports polling, which enables bridge devices in virt setups to
1434          * still use netconsole
1435          */
1436 }
1437 #endif  /* CONFIG_NET_POLL_CONTROLLER */
1438
1439 static int veth_get_iflink(const struct net_device *dev)
1440 {
1441         struct veth_priv *priv = netdev_priv(dev);
1442         struct net_device *peer;
1443         int iflink;
1444
1445         rcu_read_lock();
1446         peer = rcu_dereference(priv->peer);
1447         iflink = peer ? peer->ifindex : 0;
1448         rcu_read_unlock();
1449
1450         return iflink;
1451 }
1452
1453 static netdev_features_t veth_fix_features(struct net_device *dev,
1454                                            netdev_features_t features)
1455 {
1456         struct veth_priv *priv = netdev_priv(dev);
1457         struct net_device *peer;
1458
1459         peer = rtnl_dereference(priv->peer);
1460         if (peer) {
1461                 struct veth_priv *peer_priv = netdev_priv(peer);
1462
1463                 if (peer_priv->_xdp_prog)
1464                         features &= ~NETIF_F_GSO_SOFTWARE;
1465         }
1466         if (priv->_xdp_prog)
1467                 features |= NETIF_F_GRO;
1468
1469         return features;
1470 }
1471
1472 static int veth_set_features(struct net_device *dev,
1473                              netdev_features_t features)
1474 {
1475         netdev_features_t changed = features ^ dev->features;
1476         struct veth_priv *priv = netdev_priv(dev);
1477         int err;
1478
1479         if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1480                 return 0;
1481
1482         if (features & NETIF_F_GRO) {
1483                 err = veth_napi_enable(dev);
1484                 if (err)
1485                         return err;
1486         } else {
1487                 veth_napi_del(dev);
1488         }
1489         return 0;
1490 }
1491
1492 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1493 {
1494         struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1495         struct net_device *peer;
1496
1497         if (new_hr < 0)
1498                 new_hr = 0;
1499
1500         rcu_read_lock();
1501         peer = rcu_dereference(priv->peer);
1502         if (unlikely(!peer))
1503                 goto out;
1504
1505         peer_priv = netdev_priv(peer);
1506         priv->requested_headroom = new_hr;
1507         new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1508         dev->needed_headroom = new_hr;
1509         peer->needed_headroom = new_hr;
1510
1511 out:
1512         rcu_read_unlock();
1513 }
1514
1515 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1516                         struct netlink_ext_ack *extack)
1517 {
1518         struct veth_priv *priv = netdev_priv(dev);
1519         struct bpf_prog *old_prog;
1520         struct net_device *peer;
1521         unsigned int max_mtu;
1522         int err;
1523
1524         old_prog = priv->_xdp_prog;
1525         priv->_xdp_prog = prog;
1526         peer = rtnl_dereference(priv->peer);
1527
1528         if (prog) {
1529                 if (!peer) {
1530                         NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1531                         err = -ENOTCONN;
1532                         goto err;
1533                 }
1534
1535                 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) -
1536                           peer->hard_header_len;
1537                 /* Allow increasing the max_mtu if the program supports
1538                  * XDP fragments.
1539                  */
1540                 if (prog->aux->xdp_has_frags)
1541                         max_mtu += PAGE_SIZE * MAX_SKB_FRAGS;
1542
1543                 if (peer->mtu > max_mtu) {
1544                         NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1545                         err = -ERANGE;
1546                         goto err;
1547                 }
1548
1549                 if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1550                         NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1551                         err = -ENOSPC;
1552                         goto err;
1553                 }
1554
1555                 if (dev->flags & IFF_UP) {
1556                         err = veth_enable_xdp(dev);
1557                         if (err) {
1558                                 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1559                                 goto err;
1560                         }
1561                 }
1562
1563                 if (!old_prog) {
1564                         peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1565                         peer->max_mtu = max_mtu;
1566                 }
1567         }
1568
1569         if (old_prog) {
1570                 if (!prog) {
1571                         if (dev->flags & IFF_UP)
1572                                 veth_disable_xdp(dev);
1573
1574                         if (peer) {
1575                                 peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1576                                 peer->max_mtu = ETH_MAX_MTU;
1577                         }
1578                 }
1579                 bpf_prog_put(old_prog);
1580         }
1581
1582         if ((!!old_prog ^ !!prog) && peer)
1583                 netdev_update_features(peer);
1584
1585         return 0;
1586 err:
1587         priv->_xdp_prog = old_prog;
1588
1589         return err;
1590 }
1591
1592 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1593 {
1594         switch (xdp->command) {
1595         case XDP_SETUP_PROG:
1596                 return veth_xdp_set(dev, xdp->prog, xdp->extack);
1597         default:
1598                 return -EINVAL;
1599         }
1600 }
1601
1602 static const struct net_device_ops veth_netdev_ops = {
1603         .ndo_init            = veth_dev_init,
1604         .ndo_open            = veth_open,
1605         .ndo_stop            = veth_close,
1606         .ndo_start_xmit      = veth_xmit,
1607         .ndo_get_stats64     = veth_get_stats64,
1608         .ndo_set_rx_mode     = veth_set_multicast_list,
1609         .ndo_set_mac_address = eth_mac_addr,
1610 #ifdef CONFIG_NET_POLL_CONTROLLER
1611         .ndo_poll_controller    = veth_poll_controller,
1612 #endif
1613         .ndo_get_iflink         = veth_get_iflink,
1614         .ndo_fix_features       = veth_fix_features,
1615         .ndo_set_features       = veth_set_features,
1616         .ndo_features_check     = passthru_features_check,
1617         .ndo_set_rx_headroom    = veth_set_rx_headroom,
1618         .ndo_bpf                = veth_xdp,
1619         .ndo_xdp_xmit           = veth_ndo_xdp_xmit,
1620         .ndo_get_peer_dev       = veth_peer_dev,
1621 };
1622
1623 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1624                        NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1625                        NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1626                        NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1627                        NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1628
1629 static void veth_setup(struct net_device *dev)
1630 {
1631         ether_setup(dev);
1632
1633         dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1634         dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1635         dev->priv_flags |= IFF_NO_QUEUE;
1636         dev->priv_flags |= IFF_PHONY_HEADROOM;
1637
1638         dev->netdev_ops = &veth_netdev_ops;
1639         dev->ethtool_ops = &veth_ethtool_ops;
1640         dev->features |= NETIF_F_LLTX;
1641         dev->features |= VETH_FEATURES;
1642         dev->vlan_features = dev->features &
1643                              ~(NETIF_F_HW_VLAN_CTAG_TX |
1644                                NETIF_F_HW_VLAN_STAG_TX |
1645                                NETIF_F_HW_VLAN_CTAG_RX |
1646                                NETIF_F_HW_VLAN_STAG_RX);
1647         dev->needs_free_netdev = true;
1648         dev->priv_destructor = veth_dev_free;
1649         dev->max_mtu = ETH_MAX_MTU;
1650
1651         dev->hw_features = VETH_FEATURES;
1652         dev->hw_enc_features = VETH_FEATURES;
1653         dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1654         netif_set_tso_max_size(dev, GSO_MAX_SIZE);
1655 }
1656
1657 /*
1658  * netlink interface
1659  */
1660
1661 static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1662                          struct netlink_ext_ack *extack)
1663 {
1664         if (tb[IFLA_ADDRESS]) {
1665                 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1666                         return -EINVAL;
1667                 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1668                         return -EADDRNOTAVAIL;
1669         }
1670         if (tb[IFLA_MTU]) {
1671                 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1672                         return -EINVAL;
1673         }
1674         return 0;
1675 }
1676
1677 static struct rtnl_link_ops veth_link_ops;
1678
1679 static void veth_disable_gro(struct net_device *dev)
1680 {
1681         dev->features &= ~NETIF_F_GRO;
1682         dev->wanted_features &= ~NETIF_F_GRO;
1683         netdev_update_features(dev);
1684 }
1685
1686 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1687 {
1688         int err;
1689
1690         if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1691                 err = netif_set_real_num_tx_queues(dev, 1);
1692                 if (err)
1693                         return err;
1694         }
1695         if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1696                 err = netif_set_real_num_rx_queues(dev, 1);
1697                 if (err)
1698                         return err;
1699         }
1700         return 0;
1701 }
1702
1703 static int veth_newlink(struct net *src_net, struct net_device *dev,
1704                         struct nlattr *tb[], struct nlattr *data[],
1705                         struct netlink_ext_ack *extack)
1706 {
1707         int err;
1708         struct net_device *peer;
1709         struct veth_priv *priv;
1710         char ifname[IFNAMSIZ];
1711         struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1712         unsigned char name_assign_type;
1713         struct ifinfomsg *ifmp;
1714         struct net *net;
1715
1716         /*
1717          * create and register peer first
1718          */
1719         if (data != NULL && data[VETH_INFO_PEER] != NULL) {
1720                 struct nlattr *nla_peer;
1721
1722                 nla_peer = data[VETH_INFO_PEER];
1723                 ifmp = nla_data(nla_peer);
1724                 err = rtnl_nla_parse_ifla(peer_tb,
1725                                           nla_data(nla_peer) + sizeof(struct ifinfomsg),
1726                                           nla_len(nla_peer) - sizeof(struct ifinfomsg),
1727                                           NULL);
1728                 if (err < 0)
1729                         return err;
1730
1731                 err = veth_validate(peer_tb, NULL, extack);
1732                 if (err < 0)
1733                         return err;
1734
1735                 tbp = peer_tb;
1736         } else {
1737                 ifmp = NULL;
1738                 tbp = tb;
1739         }
1740
1741         if (ifmp && tbp[IFLA_IFNAME]) {
1742                 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1743                 name_assign_type = NET_NAME_USER;
1744         } else {
1745                 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1746                 name_assign_type = NET_NAME_ENUM;
1747         }
1748
1749         net = rtnl_link_get_net(src_net, tbp);
1750         if (IS_ERR(net))
1751                 return PTR_ERR(net);
1752
1753         peer = rtnl_create_link(net, ifname, name_assign_type,
1754                                 &veth_link_ops, tbp, extack);
1755         if (IS_ERR(peer)) {
1756                 put_net(net);
1757                 return PTR_ERR(peer);
1758         }
1759
1760         if (!ifmp || !tbp[IFLA_ADDRESS])
1761                 eth_hw_addr_random(peer);
1762
1763         if (ifmp && (dev->ifindex != 0))
1764                 peer->ifindex = ifmp->ifi_index;
1765
1766         netif_inherit_tso_max(peer, dev);
1767
1768         err = register_netdevice(peer);
1769         put_net(net);
1770         net = NULL;
1771         if (err < 0)
1772                 goto err_register_peer;
1773
1774         /* keep GRO disabled by default to be consistent with the established
1775          * veth behavior
1776          */
1777         veth_disable_gro(peer);
1778         netif_carrier_off(peer);
1779
1780         err = rtnl_configure_link(peer, ifmp);
1781         if (err < 0)
1782                 goto err_configure_peer;
1783
1784         /*
1785          * register dev last
1786          *
1787          * note, that since we've registered new device the dev's name
1788          * should be re-allocated
1789          */
1790
1791         if (tb[IFLA_ADDRESS] == NULL)
1792                 eth_hw_addr_random(dev);
1793
1794         if (tb[IFLA_IFNAME])
1795                 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1796         else
1797                 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1798
1799         err = register_netdevice(dev);
1800         if (err < 0)
1801                 goto err_register_dev;
1802
1803         netif_carrier_off(dev);
1804
1805         /*
1806          * tie the deviced together
1807          */
1808
1809         priv = netdev_priv(dev);
1810         rcu_assign_pointer(priv->peer, peer);
1811         err = veth_init_queues(dev, tb);
1812         if (err)
1813                 goto err_queues;
1814
1815         priv = netdev_priv(peer);
1816         rcu_assign_pointer(priv->peer, dev);
1817         err = veth_init_queues(peer, tb);
1818         if (err)
1819                 goto err_queues;
1820
1821         veth_disable_gro(dev);
1822         return 0;
1823
1824 err_queues:
1825         unregister_netdevice(dev);
1826 err_register_dev:
1827         /* nothing to do */
1828 err_configure_peer:
1829         unregister_netdevice(peer);
1830         return err;
1831
1832 err_register_peer:
1833         free_netdev(peer);
1834         return err;
1835 }
1836
1837 static void veth_dellink(struct net_device *dev, struct list_head *head)
1838 {
1839         struct veth_priv *priv;
1840         struct net_device *peer;
1841
1842         priv = netdev_priv(dev);
1843         peer = rtnl_dereference(priv->peer);
1844
1845         /* Note : dellink() is called from default_device_exit_batch(),
1846          * before a rcu_synchronize() point. The devices are guaranteed
1847          * not being freed before one RCU grace period.
1848          */
1849         RCU_INIT_POINTER(priv->peer, NULL);
1850         unregister_netdevice_queue(dev, head);
1851
1852         if (peer) {
1853                 priv = netdev_priv(peer);
1854                 RCU_INIT_POINTER(priv->peer, NULL);
1855                 unregister_netdevice_queue(peer, head);
1856         }
1857 }
1858
1859 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
1860         [VETH_INFO_PEER]        = { .len = sizeof(struct ifinfomsg) },
1861 };
1862
1863 static struct net *veth_get_link_net(const struct net_device *dev)
1864 {
1865         struct veth_priv *priv = netdev_priv(dev);
1866         struct net_device *peer = rtnl_dereference(priv->peer);
1867
1868         return peer ? dev_net(peer) : dev_net(dev);
1869 }
1870
1871 static unsigned int veth_get_num_queues(void)
1872 {
1873         /* enforce the same queue limit as rtnl_create_link */
1874         int queues = num_possible_cpus();
1875
1876         if (queues > 4096)
1877                 queues = 4096;
1878         return queues;
1879 }
1880
1881 static struct rtnl_link_ops veth_link_ops = {
1882         .kind           = DRV_NAME,
1883         .priv_size      = sizeof(struct veth_priv),
1884         .setup          = veth_setup,
1885         .validate       = veth_validate,
1886         .newlink        = veth_newlink,
1887         .dellink        = veth_dellink,
1888         .policy         = veth_policy,
1889         .maxtype        = VETH_INFO_MAX,
1890         .get_link_net   = veth_get_link_net,
1891         .get_num_tx_queues      = veth_get_num_queues,
1892         .get_num_rx_queues      = veth_get_num_queues,
1893 };
1894
1895 /*
1896  * init/fini
1897  */
1898
1899 static __init int veth_init(void)
1900 {
1901         return rtnl_link_register(&veth_link_ops);
1902 }
1903
1904 static __exit void veth_exit(void)
1905 {
1906         rtnl_link_unregister(&veth_link_ops);
1907 }
1908
1909 module_init(veth_init);
1910 module_exit(veth_exit);
1911
1912 MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
1913 MODULE_LICENSE("GPL v2");
1914 MODULE_ALIAS_RTNL_LINK(DRV_NAME);