| 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
| 2 | /* drivers/net/ifb.c: |
| 3 | |
| 4 | The purpose of this driver is to provide a device that allows |
| 5 | for sharing of resources: |
| 6 | |
| 7 | 1) qdiscs/policies that are per device as opposed to system wide. |
| 8 | ifb allows for a device which can be redirected to thus providing |
| 9 | an impression of sharing. |
| 10 | |
| 11 | 2) Allows for queueing incoming traffic for shaping instead of |
| 12 | dropping. |
| 13 | |
| 14 | The original concept is based on what is known as the IMQ |
| 15 | driver initially written by Martin Devera, later rewritten |
| 16 | by Patrick McHardy and then maintained by Andre Correa. |
| 17 | |
| 18 | You need the tc action mirror or redirect to feed this device |
| 19 | packets. |
| 20 | |
| 21 | |
| 22 | Authors: Jamal Hadi Salim (2005) |
| 23 | |
| 24 | */ |
| 25 | |
| 26 | |
| 27 | #include <linux/module.h> |
| 28 | #include <linux/kernel.h> |
| 29 | #include <linux/netdevice.h> |
| 30 | #include <linux/ethtool.h> |
| 31 | #include <linux/etherdevice.h> |
| 32 | #include <linux/init.h> |
| 33 | #include <linux/interrupt.h> |
| 34 | #include <linux/moduleparam.h> |
| 35 | #include <linux/netfilter_netdev.h> |
| 36 | #include <net/pkt_sched.h> |
| 37 | #include <net/net_namespace.h> |
| 38 | |
| 39 | #define TX_Q_LIMIT 32 |
| 40 | |
| 41 | struct ifb_q_stats { |
| 42 | u64 packets; |
| 43 | u64 bytes; |
| 44 | struct u64_stats_sync sync; |
| 45 | }; |
| 46 | |
| 47 | struct ifb_q_private { |
| 48 | struct net_device *dev; |
| 49 | struct tasklet_struct ifb_tasklet; |
| 50 | int tasklet_pending; |
| 51 | int txqnum; |
| 52 | struct sk_buff_head rq; |
| 53 | struct sk_buff_head tq; |
| 54 | struct ifb_q_stats rx_stats; |
| 55 | struct ifb_q_stats tx_stats; |
| 56 | } ____cacheline_aligned_in_smp; |
| 57 | |
| 58 | struct ifb_dev_private { |
| 59 | struct ifb_q_private *tx_private; |
| 60 | }; |
| 61 | |
| 62 | /* For ethtools stats. */ |
| 63 | struct ifb_q_stats_desc { |
| 64 | char desc[ETH_GSTRING_LEN]; |
| 65 | size_t offset; |
| 66 | }; |
| 67 | |
| 68 | #define IFB_Q_STAT(m) offsetof(struct ifb_q_stats, m) |
| 69 | |
| 70 | static const struct ifb_q_stats_desc ifb_q_stats_desc[] = { |
| 71 | { "packets", IFB_Q_STAT(packets) }, |
| 72 | { "bytes", IFB_Q_STAT(bytes) }, |
| 73 | }; |
| 74 | |
| 75 | #define IFB_Q_STATS_LEN ARRAY_SIZE(ifb_q_stats_desc) |
| 76 | |
| 77 | static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev); |
| 78 | static int ifb_open(struct net_device *dev); |
| 79 | static int ifb_close(struct net_device *dev); |
| 80 | |
| 81 | static void ifb_update_q_stats(struct ifb_q_stats *stats, int len) |
| 82 | { |
| 83 | u64_stats_update_begin(&stats->sync); |
| 84 | stats->packets++; |
| 85 | stats->bytes += len; |
| 86 | u64_stats_update_end(&stats->sync); |
| 87 | } |
| 88 | |
| 89 | static void ifb_ri_tasklet(struct tasklet_struct *t) |
| 90 | { |
| 91 | struct ifb_q_private *txp = from_tasklet(txp, t, ifb_tasklet); |
| 92 | struct netdev_queue *txq; |
| 93 | struct sk_buff *skb; |
| 94 | |
| 95 | txq = netdev_get_tx_queue(txp->dev, txp->txqnum); |
| 96 | skb = skb_peek(&txp->tq); |
| 97 | if (!skb) { |
| 98 | if (!__netif_tx_trylock(txq)) |
| 99 | goto resched; |
| 100 | skb_queue_splice_tail_init(&txp->rq, &txp->tq); |
| 101 | __netif_tx_unlock(txq); |
| 102 | } |
| 103 | |
| 104 | while ((skb = __skb_dequeue(&txp->tq)) != NULL) { |
| 105 | /* Skip tc and netfilter to prevent redirection loop. */ |
| 106 | skb->redirected = 0; |
| 107 | #ifdef CONFIG_NET_CLS_ACT |
| 108 | skb->tc_skip_classify = 1; |
| 109 | #endif |
| 110 | nf_skip_egress(skb, true); |
| 111 | |
| 112 | ifb_update_q_stats(&txp->tx_stats, skb->len); |
| 113 | |
| 114 | rcu_read_lock(); |
| 115 | skb->dev = dev_get_by_index_rcu(dev_net(txp->dev), skb->skb_iif); |
| 116 | if (!skb->dev) { |
| 117 | rcu_read_unlock(); |
| 118 | dev_kfree_skb(skb); |
| 119 | txp->dev->stats.tx_dropped++; |
| 120 | if (skb_queue_len(&txp->tq) != 0) |
| 121 | goto resched; |
| 122 | break; |
| 123 | } |
| 124 | rcu_read_unlock(); |
| 125 | skb->skb_iif = txp->dev->ifindex; |
| 126 | |
| 127 | if (!skb->from_ingress) { |
| 128 | dev_queue_xmit(skb); |
| 129 | } else { |
| 130 | skb_pull_rcsum(skb, skb->mac_len); |
| 131 | netif_receive_skb(skb); |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | if (__netif_tx_trylock(txq)) { |
| 136 | skb = skb_peek(&txp->rq); |
| 137 | if (!skb) { |
| 138 | txp->tasklet_pending = 0; |
| 139 | if (netif_tx_queue_stopped(txq)) |
| 140 | netif_tx_wake_queue(txq); |
| 141 | } else { |
| 142 | __netif_tx_unlock(txq); |
| 143 | goto resched; |
| 144 | } |
| 145 | __netif_tx_unlock(txq); |
| 146 | } else { |
| 147 | resched: |
| 148 | txp->tasklet_pending = 1; |
| 149 | tasklet_schedule(&txp->ifb_tasklet); |
| 150 | } |
| 151 | |
| 152 | } |
| 153 | |
| 154 | static void ifb_stats64(struct net_device *dev, |
| 155 | struct rtnl_link_stats64 *stats) |
| 156 | { |
| 157 | struct ifb_dev_private *dp = netdev_priv(dev); |
| 158 | struct ifb_q_private *txp = dp->tx_private; |
| 159 | unsigned int start; |
| 160 | u64 packets, bytes; |
| 161 | int i; |
| 162 | |
| 163 | for (i = 0; i < dev->num_tx_queues; i++,txp++) { |
| 164 | do { |
| 165 | start = u64_stats_fetch_begin(&txp->rx_stats.sync); |
| 166 | packets = txp->rx_stats.packets; |
| 167 | bytes = txp->rx_stats.bytes; |
| 168 | } while (u64_stats_fetch_retry(&txp->rx_stats.sync, start)); |
| 169 | stats->rx_packets += packets; |
| 170 | stats->rx_bytes += bytes; |
| 171 | |
| 172 | do { |
| 173 | start = u64_stats_fetch_begin(&txp->tx_stats.sync); |
| 174 | packets = txp->tx_stats.packets; |
| 175 | bytes = txp->tx_stats.bytes; |
| 176 | } while (u64_stats_fetch_retry(&txp->tx_stats.sync, start)); |
| 177 | stats->tx_packets += packets; |
| 178 | stats->tx_bytes += bytes; |
| 179 | } |
| 180 | stats->rx_dropped = dev->stats.rx_dropped; |
| 181 | stats->tx_dropped = dev->stats.tx_dropped; |
| 182 | } |
| 183 | |
| 184 | static int ifb_dev_init(struct net_device *dev) |
| 185 | { |
| 186 | struct ifb_dev_private *dp = netdev_priv(dev); |
| 187 | struct ifb_q_private *txp; |
| 188 | int i; |
| 189 | |
| 190 | txp = kcalloc(dev->num_tx_queues, sizeof(*txp), GFP_KERNEL); |
| 191 | if (!txp) |
| 192 | return -ENOMEM; |
| 193 | dp->tx_private = txp; |
| 194 | for (i = 0; i < dev->num_tx_queues; i++,txp++) { |
| 195 | txp->txqnum = i; |
| 196 | txp->dev = dev; |
| 197 | __skb_queue_head_init(&txp->rq); |
| 198 | __skb_queue_head_init(&txp->tq); |
| 199 | u64_stats_init(&txp->rx_stats.sync); |
| 200 | u64_stats_init(&txp->tx_stats.sync); |
| 201 | tasklet_setup(&txp->ifb_tasklet, ifb_ri_tasklet); |
| 202 | netif_tx_start_queue(netdev_get_tx_queue(dev, i)); |
| 203 | } |
| 204 | return 0; |
| 205 | } |
| 206 | |
| 207 | static void ifb_get_strings(struct net_device *dev, u32 stringset, u8 *buf) |
| 208 | { |
| 209 | u8 *p = buf; |
| 210 | int i, j; |
| 211 | |
| 212 | switch (stringset) { |
| 213 | case ETH_SS_STATS: |
| 214 | for (i = 0; i < dev->real_num_rx_queues; i++) |
| 215 | for (j = 0; j < IFB_Q_STATS_LEN; j++) |
| 216 | ethtool_sprintf(&p, "rx_queue_%u_%.18s", |
| 217 | i, ifb_q_stats_desc[j].desc); |
| 218 | |
| 219 | for (i = 0; i < dev->real_num_tx_queues; i++) |
| 220 | for (j = 0; j < IFB_Q_STATS_LEN; j++) |
| 221 | ethtool_sprintf(&p, "tx_queue_%u_%.18s", |
| 222 | i, ifb_q_stats_desc[j].desc); |
| 223 | |
| 224 | break; |
| 225 | } |
| 226 | } |
| 227 | |
| 228 | static int ifb_get_sset_count(struct net_device *dev, int sset) |
| 229 | { |
| 230 | switch (sset) { |
| 231 | case ETH_SS_STATS: |
| 232 | return IFB_Q_STATS_LEN * (dev->real_num_rx_queues + |
| 233 | dev->real_num_tx_queues); |
| 234 | default: |
| 235 | return -EOPNOTSUPP; |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | static void ifb_fill_stats_data(u64 **data, |
| 240 | struct ifb_q_stats *q_stats) |
| 241 | { |
| 242 | void *stats_base = (void *)q_stats; |
| 243 | unsigned int start; |
| 244 | size_t offset; |
| 245 | int j; |
| 246 | |
| 247 | do { |
| 248 | start = u64_stats_fetch_begin(&q_stats->sync); |
| 249 | for (j = 0; j < IFB_Q_STATS_LEN; j++) { |
| 250 | offset = ifb_q_stats_desc[j].offset; |
| 251 | (*data)[j] = *(u64 *)(stats_base + offset); |
| 252 | } |
| 253 | } while (u64_stats_fetch_retry(&q_stats->sync, start)); |
| 254 | |
| 255 | *data += IFB_Q_STATS_LEN; |
| 256 | } |
| 257 | |
| 258 | static void ifb_get_ethtool_stats(struct net_device *dev, |
| 259 | struct ethtool_stats *stats, u64 *data) |
| 260 | { |
| 261 | struct ifb_dev_private *dp = netdev_priv(dev); |
| 262 | struct ifb_q_private *txp; |
| 263 | int i; |
| 264 | |
| 265 | for (i = 0; i < dev->real_num_rx_queues; i++) { |
| 266 | txp = dp->tx_private + i; |
| 267 | ifb_fill_stats_data(&data, &txp->rx_stats); |
| 268 | } |
| 269 | |
| 270 | for (i = 0; i < dev->real_num_tx_queues; i++) { |
| 271 | txp = dp->tx_private + i; |
| 272 | ifb_fill_stats_data(&data, &txp->tx_stats); |
| 273 | } |
| 274 | } |
| 275 | |
| 276 | static const struct net_device_ops ifb_netdev_ops = { |
| 277 | .ndo_open = ifb_open, |
| 278 | .ndo_stop = ifb_close, |
| 279 | .ndo_get_stats64 = ifb_stats64, |
| 280 | .ndo_start_xmit = ifb_xmit, |
| 281 | .ndo_validate_addr = eth_validate_addr, |
| 282 | .ndo_init = ifb_dev_init, |
| 283 | }; |
| 284 | |
| 285 | static const struct ethtool_ops ifb_ethtool_ops = { |
| 286 | .get_strings = ifb_get_strings, |
| 287 | .get_sset_count = ifb_get_sset_count, |
| 288 | .get_ethtool_stats = ifb_get_ethtool_stats, |
| 289 | }; |
| 290 | |
| 291 | #define IFB_FEATURES (NETIF_F_HW_CSUM | NETIF_F_SG | NETIF_F_FRAGLIST | \ |
| 292 | NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \ |
| 293 | NETIF_F_HIGHDMA | NETIF_F_HW_VLAN_CTAG_TX | \ |
| 294 | NETIF_F_HW_VLAN_STAG_TX) |
| 295 | |
| 296 | static void ifb_dev_free(struct net_device *dev) |
| 297 | { |
| 298 | struct ifb_dev_private *dp = netdev_priv(dev); |
| 299 | struct ifb_q_private *txp = dp->tx_private; |
| 300 | int i; |
| 301 | |
| 302 | for (i = 0; i < dev->num_tx_queues; i++,txp++) { |
| 303 | tasklet_kill(&txp->ifb_tasklet); |
| 304 | __skb_queue_purge(&txp->rq); |
| 305 | __skb_queue_purge(&txp->tq); |
| 306 | } |
| 307 | kfree(dp->tx_private); |
| 308 | } |
| 309 | |
| 310 | static void ifb_setup(struct net_device *dev) |
| 311 | { |
| 312 | /* Initialize the device structure. */ |
| 313 | dev->netdev_ops = &ifb_netdev_ops; |
| 314 | dev->ethtool_ops = &ifb_ethtool_ops; |
| 315 | |
| 316 | /* Fill in device structure with ethernet-generic values. */ |
| 317 | ether_setup(dev); |
| 318 | dev->tx_queue_len = TX_Q_LIMIT; |
| 319 | |
| 320 | dev->features |= IFB_FEATURES; |
| 321 | dev->hw_features |= dev->features; |
| 322 | dev->hw_enc_features |= dev->features; |
| 323 | dev->vlan_features |= IFB_FEATURES & ~(NETIF_F_HW_VLAN_CTAG_TX | |
| 324 | NETIF_F_HW_VLAN_STAG_TX); |
| 325 | |
| 326 | dev->flags |= IFF_NOARP; |
| 327 | dev->flags &= ~IFF_MULTICAST; |
| 328 | dev->priv_flags &= ~IFF_TX_SKB_SHARING; |
| 329 | netif_keep_dst(dev); |
| 330 | eth_hw_addr_random(dev); |
| 331 | dev->needs_free_netdev = true; |
| 332 | dev->priv_destructor = ifb_dev_free; |
| 333 | |
| 334 | dev->min_mtu = 0; |
| 335 | dev->max_mtu = 0; |
| 336 | } |
| 337 | |
| 338 | static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev) |
| 339 | { |
| 340 | struct ifb_dev_private *dp = netdev_priv(dev); |
| 341 | struct ifb_q_private *txp = dp->tx_private + skb_get_queue_mapping(skb); |
| 342 | |
| 343 | ifb_update_q_stats(&txp->rx_stats, skb->len); |
| 344 | |
| 345 | if (!skb->redirected || !skb->skb_iif) { |
| 346 | dev_kfree_skb(skb); |
| 347 | dev->stats.rx_dropped++; |
| 348 | return NETDEV_TX_OK; |
| 349 | } |
| 350 | |
| 351 | if (skb_queue_len(&txp->rq) >= dev->tx_queue_len) |
| 352 | netif_tx_stop_queue(netdev_get_tx_queue(dev, txp->txqnum)); |
| 353 | |
| 354 | __skb_queue_tail(&txp->rq, skb); |
| 355 | if (!txp->tasklet_pending) { |
| 356 | txp->tasklet_pending = 1; |
| 357 | tasklet_schedule(&txp->ifb_tasklet); |
| 358 | } |
| 359 | |
| 360 | return NETDEV_TX_OK; |
| 361 | } |
| 362 | |
| 363 | static int ifb_close(struct net_device *dev) |
| 364 | { |
| 365 | netif_tx_stop_all_queues(dev); |
| 366 | return 0; |
| 367 | } |
| 368 | |
| 369 | static int ifb_open(struct net_device *dev) |
| 370 | { |
| 371 | netif_tx_start_all_queues(dev); |
| 372 | return 0; |
| 373 | } |
| 374 | |
| 375 | static int ifb_validate(struct nlattr *tb[], struct nlattr *data[], |
| 376 | struct netlink_ext_ack *extack) |
| 377 | { |
| 378 | if (tb[IFLA_ADDRESS]) { |
| 379 | if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) |
| 380 | return -EINVAL; |
| 381 | if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) |
| 382 | return -EADDRNOTAVAIL; |
| 383 | } |
| 384 | return 0; |
| 385 | } |
| 386 | |
| 387 | static struct rtnl_link_ops ifb_link_ops __read_mostly = { |
| 388 | .kind = "ifb", |
| 389 | .priv_size = sizeof(struct ifb_dev_private), |
| 390 | .setup = ifb_setup, |
| 391 | .validate = ifb_validate, |
| 392 | }; |
| 393 | |
| 394 | /* Number of ifb devices to be set up by this module. |
| 395 | * Note that these legacy devices have one queue. |
| 396 | * Prefer something like : ip link add ifb10 numtxqueues 8 type ifb |
| 397 | */ |
| 398 | static int numifbs = 2; |
| 399 | module_param(numifbs, int, 0); |
| 400 | MODULE_PARM_DESC(numifbs, "Number of ifb devices"); |
| 401 | |
| 402 | static int __init ifb_init_one(int index) |
| 403 | { |
| 404 | struct net_device *dev_ifb; |
| 405 | int err; |
| 406 | |
| 407 | dev_ifb = alloc_netdev(sizeof(struct ifb_dev_private), "ifb%d", |
| 408 | NET_NAME_UNKNOWN, ifb_setup); |
| 409 | |
| 410 | if (!dev_ifb) |
| 411 | return -ENOMEM; |
| 412 | |
| 413 | dev_ifb->rtnl_link_ops = &ifb_link_ops; |
| 414 | err = register_netdevice(dev_ifb); |
| 415 | if (err < 0) |
| 416 | goto err; |
| 417 | |
| 418 | return 0; |
| 419 | |
| 420 | err: |
| 421 | free_netdev(dev_ifb); |
| 422 | return err; |
| 423 | } |
| 424 | |
| 425 | static int __init ifb_init_module(void) |
| 426 | { |
| 427 | int i, err; |
| 428 | |
| 429 | down_write(&pernet_ops_rwsem); |
| 430 | rtnl_lock(); |
| 431 | err = __rtnl_link_register(&ifb_link_ops); |
| 432 | if (err < 0) |
| 433 | goto out; |
| 434 | |
| 435 | for (i = 0; i < numifbs && !err; i++) { |
| 436 | err = ifb_init_one(i); |
| 437 | cond_resched(); |
| 438 | } |
| 439 | if (err) |
| 440 | __rtnl_link_unregister(&ifb_link_ops); |
| 441 | |
| 442 | out: |
| 443 | rtnl_unlock(); |
| 444 | up_write(&pernet_ops_rwsem); |
| 445 | |
| 446 | return err; |
| 447 | } |
| 448 | |
| 449 | static void __exit ifb_cleanup_module(void) |
| 450 | { |
| 451 | rtnl_link_unregister(&ifb_link_ops); |
| 452 | } |
| 453 | |
| 454 | module_init(ifb_init_module); |
| 455 | module_exit(ifb_cleanup_module); |
| 456 | MODULE_LICENSE("GPL"); |
| 457 | MODULE_AUTHOR("Jamal Hadi Salim"); |
| 458 | MODULE_ALIAS_RTNL_LINK("ifb"); |