1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * NET3 Protocol independent device support routines. 4 * 5 * Derived from the non IP parts of dev.c 1.0.19 6 * Authors: Ross Biro 7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 8 * Mark Evans, <evansmp@uhura.aston.ac.uk> 9 * 10 * Additional Authors: 11 * Florian la Roche <rzsfl@rz.uni-sb.de> 12 * Alan Cox <gw4pts@gw4pts.ampr.org> 13 * David Hinds <dahinds@users.sourceforge.net> 14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru> 15 * Adam Sulmicki <adam@cfar.umd.edu> 16 * Pekka Riikonen <priikone@poesidon.pspt.fi> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/ethtool_netlink.h> 96 #include <linux/skbuff.h> 97 #include <linux/kthread.h> 98 #include <linux/bpf.h> 99 #include <linux/bpf_trace.h> 100 #include <net/net_namespace.h> 101 #include <net/sock.h> 102 #include <net/busy_poll.h> 103 #include <linux/rtnetlink.h> 104 #include <linux/stat.h> 105 #include <net/dsa.h> 106 #include <net/dst.h> 107 #include <net/dst_metadata.h> 108 #include <net/gro.h> 109 #include <net/netdev_queues.h> 110 #include <net/pkt_sched.h> 111 #include <net/pkt_cls.h> 112 #include <net/checksum.h> 113 #include <net/xfrm.h> 114 #include <net/tcx.h> 115 #include <linux/highmem.h> 116 #include <linux/init.h> 117 #include <linux/module.h> 118 #include <linux/netpoll.h> 119 #include <linux/rcupdate.h> 120 #include <linux/delay.h> 121 #include <net/iw_handler.h> 122 #include <asm/current.h> 123 #include <linux/audit.h> 124 #include <linux/dmaengine.h> 125 #include <linux/err.h> 126 #include <linux/ctype.h> 127 #include <linux/if_arp.h> 128 #include <linux/if_vlan.h> 129 #include <linux/ip.h> 130 #include <net/ip.h> 131 #include <net/mpls.h> 132 #include <linux/ipv6.h> 133 #include <linux/in.h> 134 #include <linux/jhash.h> 135 #include <linux/random.h> 136 #include <trace/events/napi.h> 137 #include <trace/events/net.h> 138 #include <trace/events/skb.h> 139 #include <trace/events/qdisc.h> 140 #include <trace/events/xdp.h> 141 #include <linux/inetdevice.h> 142 #include <linux/cpu_rmap.h> 143 #include <linux/static_key.h> 144 #include <linux/hashtable.h> 145 #include <linux/vmalloc.h> 146 #include <linux/if_macvlan.h> 147 #include <linux/errqueue.h> 148 #include <linux/hrtimer.h> 149 #include <linux/netfilter_netdev.h> 150 #include <linux/crash_dump.h> 151 #include <linux/sctp.h> 152 #include <net/udp_tunnel.h> 153 #include <linux/net_namespace.h> 154 #include <linux/indirect_call_wrapper.h> 155 #include <net/devlink.h> 156 #include <linux/pm_runtime.h> 157 #include <linux/prandom.h> 158 #include <linux/once_lite.h> 159 #include <net/netdev_lock.h> 160 #include <net/netdev_rx_queue.h> 161 #include <net/page_pool/types.h> 162 #include <net/page_pool/helpers.h> 163 #include <net/page_pool/memory_provider.h> 164 #include <net/rps.h> 165 #include <linux/phy_link_topology.h> 166 167 #include "dev.h" 168 #include "devmem.h" 169 #include "net-sysfs.h" 170 171 static DEFINE_SPINLOCK(ptype_lock); 172 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 173 174 static int netif_rx_internal(struct sk_buff *skb); 175 static int call_netdevice_notifiers_extack(unsigned long val, 176 struct net_device *dev, 177 struct netlink_ext_ack *extack); 178 179 static DEFINE_MUTEX(ifalias_mutex); 180 181 /* protects napi_hash addition/deletion and napi_gen_id */ 182 static DEFINE_SPINLOCK(napi_hash_lock); 183 184 static unsigned int napi_gen_id = NR_CPUS; 185 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 186 187 static inline void dev_base_seq_inc(struct net *net) 188 { 189 unsigned int val = net->dev_base_seq + 1; 190 191 WRITE_ONCE(net->dev_base_seq, val ?: 1); 192 } 193 194 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 195 { 196 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 197 198 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 199 } 200 201 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 202 { 203 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 204 } 205 206 #ifndef CONFIG_PREEMPT_RT 207 208 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 209 210 static int __init setup_backlog_napi_threads(char *arg) 211 { 212 static_branch_enable(&use_backlog_threads_key); 213 return 0; 214 } 215 early_param("thread_backlog_napi", setup_backlog_napi_threads); 216 217 static bool use_backlog_threads(void) 218 { 219 return static_branch_unlikely(&use_backlog_threads_key); 220 } 221 222 #else 223 224 static bool use_backlog_threads(void) 225 { 226 return true; 227 } 228 229 #endif 230 231 static inline void backlog_lock_irq_save(struct softnet_data *sd, 232 unsigned long *flags) 233 { 234 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 235 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 236 } else { 237 local_irq_save(*flags); 238 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 239 spin_lock(&sd->input_pkt_queue.lock); 240 } 241 } 242 243 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 244 { 245 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 246 spin_lock_irq(&sd->input_pkt_queue.lock); 247 else 248 local_irq_disable(); 249 } 250 251 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 252 unsigned long flags) 253 { 254 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 255 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, flags); 256 } else { 257 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 258 spin_unlock(&sd->input_pkt_queue.lock); 259 local_irq_restore(flags); 260 } 261 } 262 263 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 264 { 265 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 266 spin_unlock_irq(&sd->input_pkt_queue.lock); 267 else 268 local_irq_enable(); 269 } 270 271 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 272 const char *name) 273 { 274 struct netdev_name_node *name_node; 275 276 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 277 if (!name_node) 278 return NULL; 279 INIT_HLIST_NODE(&name_node->hlist); 280 name_node->dev = dev; 281 name_node->name = name; 282 return name_node; 283 } 284 285 static struct netdev_name_node * 286 netdev_name_node_head_alloc(struct net_device *dev) 287 { 288 struct netdev_name_node *name_node; 289 290 name_node = netdev_name_node_alloc(dev, dev->name); 291 if (!name_node) 292 return NULL; 293 INIT_LIST_HEAD(&name_node->list); 294 return name_node; 295 } 296 297 static void netdev_name_node_free(struct netdev_name_node *name_node) 298 { 299 kfree(name_node); 300 } 301 302 static void netdev_name_node_add(struct net *net, 303 struct netdev_name_node *name_node) 304 { 305 hlist_add_head_rcu(&name_node->hlist, 306 dev_name_hash(net, name_node->name)); 307 } 308 309 static void netdev_name_node_del(struct netdev_name_node *name_node) 310 { 311 hlist_del_rcu(&name_node->hlist); 312 } 313 314 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 315 const char *name) 316 { 317 struct hlist_head *head = dev_name_hash(net, name); 318 struct netdev_name_node *name_node; 319 320 hlist_for_each_entry(name_node, head, hlist) 321 if (!strcmp(name_node->name, name)) 322 return name_node; 323 return NULL; 324 } 325 326 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 327 const char *name) 328 { 329 struct hlist_head *head = dev_name_hash(net, name); 330 struct netdev_name_node *name_node; 331 332 hlist_for_each_entry_rcu(name_node, head, hlist) 333 if (!strcmp(name_node->name, name)) 334 return name_node; 335 return NULL; 336 } 337 338 bool netdev_name_in_use(struct net *net, const char *name) 339 { 340 return netdev_name_node_lookup(net, name); 341 } 342 EXPORT_SYMBOL(netdev_name_in_use); 343 344 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 345 { 346 struct netdev_name_node *name_node; 347 struct net *net = dev_net(dev); 348 349 name_node = netdev_name_node_lookup(net, name); 350 if (name_node) 351 return -EEXIST; 352 name_node = netdev_name_node_alloc(dev, name); 353 if (!name_node) 354 return -ENOMEM; 355 netdev_name_node_add(net, name_node); 356 /* The node that holds dev->name acts as a head of per-device list. */ 357 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 358 359 return 0; 360 } 361 362 static void netdev_name_node_alt_free(struct rcu_head *head) 363 { 364 struct netdev_name_node *name_node = 365 container_of(head, struct netdev_name_node, rcu); 366 367 kfree(name_node->name); 368 netdev_name_node_free(name_node); 369 } 370 371 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 372 { 373 netdev_name_node_del(name_node); 374 list_del(&name_node->list); 375 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 376 } 377 378 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 379 { 380 struct netdev_name_node *name_node; 381 struct net *net = dev_net(dev); 382 383 name_node = netdev_name_node_lookup(net, name); 384 if (!name_node) 385 return -ENOENT; 386 /* lookup might have found our primary name or a name belonging 387 * to another device. 388 */ 389 if (name_node == dev->name_node || name_node->dev != dev) 390 return -EINVAL; 391 392 __netdev_name_node_alt_destroy(name_node); 393 return 0; 394 } 395 396 static void netdev_name_node_alt_flush(struct net_device *dev) 397 { 398 struct netdev_name_node *name_node, *tmp; 399 400 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 401 list_del(&name_node->list); 402 netdev_name_node_alt_free(&name_node->rcu); 403 } 404 } 405 406 /* Device list insertion */ 407 static void list_netdevice(struct net_device *dev) 408 { 409 struct netdev_name_node *name_node; 410 struct net *net = dev_net(dev); 411 412 ASSERT_RTNL(); 413 414 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 415 netdev_name_node_add(net, dev->name_node); 416 hlist_add_head_rcu(&dev->index_hlist, 417 dev_index_hash(net, dev->ifindex)); 418 419 netdev_for_each_altname(dev, name_node) 420 netdev_name_node_add(net, name_node); 421 422 /* We reserved the ifindex, this can't fail */ 423 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 424 425 dev_base_seq_inc(net); 426 } 427 428 /* Device list removal 429 * caller must respect a RCU grace period before freeing/reusing dev 430 */ 431 static void unlist_netdevice(struct net_device *dev) 432 { 433 struct netdev_name_node *name_node; 434 struct net *net = dev_net(dev); 435 436 ASSERT_RTNL(); 437 438 xa_erase(&net->dev_by_index, dev->ifindex); 439 440 netdev_for_each_altname(dev, name_node) 441 netdev_name_node_del(name_node); 442 443 /* Unlink dev from the device chain */ 444 list_del_rcu(&dev->dev_list); 445 netdev_name_node_del(dev->name_node); 446 hlist_del_rcu(&dev->index_hlist); 447 448 dev_base_seq_inc(dev_net(dev)); 449 } 450 451 /* 452 * Our notifier list 453 */ 454 455 static RAW_NOTIFIER_HEAD(netdev_chain); 456 457 /* 458 * Device drivers call our routines to queue packets here. We empty the 459 * queue in the local softnet handler. 460 */ 461 462 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 463 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 464 }; 465 EXPORT_PER_CPU_SYMBOL(softnet_data); 466 467 /* Page_pool has a lockless array/stack to alloc/recycle pages. 468 * PP consumers must pay attention to run APIs in the appropriate context 469 * (e.g. NAPI context). 470 */ 471 DEFINE_PER_CPU(struct page_pool_bh, system_page_pool) = { 472 .bh_lock = INIT_LOCAL_LOCK(bh_lock), 473 }; 474 475 #ifdef CONFIG_LOCKDEP 476 /* 477 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 478 * according to dev->type 479 */ 480 static const unsigned short netdev_lock_type[] = { 481 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 482 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 483 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 484 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 485 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 486 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 487 ARPHRD_CAN, ARPHRD_MCTP, 488 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 489 ARPHRD_RAWHDLC, ARPHRD_RAWIP, 490 ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 491 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 492 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 493 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 494 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 495 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 496 ARPHRD_IEEE80211_RADIOTAP, 497 ARPHRD_IEEE802154, ARPHRD_IEEE802154_MONITOR, 498 ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 499 ARPHRD_CAIF, ARPHRD_IP6GRE, ARPHRD_NETLINK, ARPHRD_6LOWPAN, 500 ARPHRD_VSOCKMON, 501 ARPHRD_VOID, ARPHRD_NONE}; 502 503 static const char *const netdev_lock_name[] = { 504 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 505 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 506 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 507 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 508 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 509 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 510 "_xmit_CAN", "_xmit_MCTP", 511 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 512 "_xmit_RAWHDLC", "_xmit_RAWIP", 513 "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 514 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 515 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 516 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 517 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 518 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 519 "_xmit_IEEE80211_RADIOTAP", 520 "_xmit_IEEE802154", "_xmit_IEEE802154_MONITOR", 521 "_xmit_PHONET", "_xmit_PHONET_PIPE", 522 "_xmit_CAIF", "_xmit_IP6GRE", "_xmit_NETLINK", "_xmit_6LOWPAN", 523 "_xmit_VSOCKMON", 524 "_xmit_VOID", "_xmit_NONE"}; 525 526 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 527 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 528 529 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 530 { 531 int i; 532 533 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 534 if (netdev_lock_type[i] == dev_type) 535 return i; 536 /* the last key is used by default */ 537 WARN_ONCE(1, "netdev_lock_pos() could not find dev_type=%u\n", dev_type); 538 return ARRAY_SIZE(netdev_lock_type) - 1; 539 } 540 541 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 542 unsigned short dev_type) 543 { 544 int i; 545 546 i = netdev_lock_pos(dev_type); 547 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 548 netdev_lock_name[i]); 549 } 550 551 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 552 { 553 int i; 554 555 i = netdev_lock_pos(dev->type); 556 lockdep_set_class_and_name(&dev->addr_list_lock, 557 &netdev_addr_lock_key[i], 558 netdev_lock_name[i]); 559 } 560 #else 561 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 562 unsigned short dev_type) 563 { 564 } 565 566 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 567 { 568 } 569 #endif 570 571 /******************************************************************************* 572 * 573 * Protocol management and registration routines 574 * 575 *******************************************************************************/ 576 577 578 /* 579 * Add a protocol ID to the list. Now that the input handler is 580 * smarter we can dispense with all the messy stuff that used to be 581 * here. 582 * 583 * BEWARE!!! Protocol handlers, mangling input packets, 584 * MUST BE last in hash buckets and checking protocol handlers 585 * MUST start from promiscuous ptype_all chain in net_bh. 586 * It is true now, do not change it. 587 * Explanation follows: if protocol handler, mangling packet, will 588 * be the first on list, it is not able to sense, that packet 589 * is cloned and should be copied-on-write, so that it will 590 * change it and subsequent readers will get broken packet. 591 * --ANK (980803) 592 */ 593 594 static inline struct list_head *ptype_head(const struct packet_type *pt) 595 { 596 if (pt->type == htons(ETH_P_ALL)) { 597 if (!pt->af_packet_net && !pt->dev) 598 return NULL; 599 600 return pt->dev ? &pt->dev->ptype_all : 601 &pt->af_packet_net->ptype_all; 602 } 603 604 if (pt->dev) 605 return &pt->dev->ptype_specific; 606 607 return pt->af_packet_net ? &pt->af_packet_net->ptype_specific : 608 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 609 } 610 611 /** 612 * dev_add_pack - add packet handler 613 * @pt: packet type declaration 614 * 615 * Add a protocol handler to the networking stack. The passed &packet_type 616 * is linked into kernel lists and may not be freed until it has been 617 * removed from the kernel lists. 618 * 619 * This call does not sleep therefore it can not 620 * guarantee all CPU's that are in middle of receiving packets 621 * will see the new packet type (until the next received packet). 622 */ 623 624 void dev_add_pack(struct packet_type *pt) 625 { 626 struct list_head *head = ptype_head(pt); 627 628 if (WARN_ON_ONCE(!head)) 629 return; 630 631 spin_lock(&ptype_lock); 632 list_add_rcu(&pt->list, head); 633 spin_unlock(&ptype_lock); 634 } 635 EXPORT_SYMBOL(dev_add_pack); 636 637 /** 638 * __dev_remove_pack - remove packet handler 639 * @pt: packet type declaration 640 * 641 * Remove a protocol handler that was previously added to the kernel 642 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 643 * from the kernel lists and can be freed or reused once this function 644 * returns. 645 * 646 * The packet type might still be in use by receivers 647 * and must not be freed until after all the CPU's have gone 648 * through a quiescent state. 649 */ 650 void __dev_remove_pack(struct packet_type *pt) 651 { 652 struct list_head *head = ptype_head(pt); 653 struct packet_type *pt1; 654 655 if (!head) 656 return; 657 658 spin_lock(&ptype_lock); 659 660 list_for_each_entry(pt1, head, list) { 661 if (pt == pt1) { 662 list_del_rcu(&pt->list); 663 goto out; 664 } 665 } 666 667 pr_warn("dev_remove_pack: %p not found\n", pt); 668 out: 669 spin_unlock(&ptype_lock); 670 } 671 EXPORT_SYMBOL(__dev_remove_pack); 672 673 /** 674 * dev_remove_pack - remove packet handler 675 * @pt: packet type declaration 676 * 677 * Remove a protocol handler that was previously added to the kernel 678 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 679 * from the kernel lists and can be freed or reused once this function 680 * returns. 681 * 682 * This call sleeps to guarantee that no CPU is looking at the packet 683 * type after return. 684 */ 685 void dev_remove_pack(struct packet_type *pt) 686 { 687 __dev_remove_pack(pt); 688 689 synchronize_net(); 690 } 691 EXPORT_SYMBOL(dev_remove_pack); 692 693 694 /******************************************************************************* 695 * 696 * Device Interface Subroutines 697 * 698 *******************************************************************************/ 699 700 /** 701 * dev_get_iflink - get 'iflink' value of a interface 702 * @dev: targeted interface 703 * 704 * Indicates the ifindex the interface is linked to. 705 * Physical interfaces have the same 'ifindex' and 'iflink' values. 706 */ 707 708 int dev_get_iflink(const struct net_device *dev) 709 { 710 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 711 return dev->netdev_ops->ndo_get_iflink(dev); 712 713 return READ_ONCE(dev->ifindex); 714 } 715 EXPORT_SYMBOL(dev_get_iflink); 716 717 /** 718 * dev_fill_metadata_dst - Retrieve tunnel egress information. 719 * @dev: targeted interface 720 * @skb: The packet. 721 * 722 * For better visibility of tunnel traffic OVS needs to retrieve 723 * egress tunnel information for a packet. Following API allows 724 * user to get this info. 725 */ 726 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 727 { 728 struct ip_tunnel_info *info; 729 730 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 731 return -EINVAL; 732 733 info = skb_tunnel_info_unclone(skb); 734 if (!info) 735 return -ENOMEM; 736 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 737 return -EINVAL; 738 739 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 740 } 741 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 742 743 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 744 { 745 int k = stack->num_paths++; 746 747 if (k >= NET_DEVICE_PATH_STACK_MAX) 748 return NULL; 749 750 return &stack->path[k]; 751 } 752 753 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 754 struct net_device_path_stack *stack) 755 { 756 const struct net_device *last_dev; 757 struct net_device_path_ctx ctx = { 758 .dev = dev, 759 }; 760 struct net_device_path *path; 761 int ret = 0; 762 763 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 764 stack->num_paths = 0; 765 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 766 last_dev = ctx.dev; 767 path = dev_fwd_path(stack); 768 if (!path) 769 return -1; 770 771 memset(path, 0, sizeof(struct net_device_path)); 772 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 773 if (ret < 0) 774 return -1; 775 776 if (WARN_ON_ONCE(last_dev == ctx.dev)) 777 return -1; 778 } 779 780 if (!ctx.dev) 781 return ret; 782 783 path = dev_fwd_path(stack); 784 if (!path) 785 return -1; 786 path->type = DEV_PATH_ETHERNET; 787 path->dev = ctx.dev; 788 789 return ret; 790 } 791 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 792 793 /* must be called under rcu_read_lock(), as we dont take a reference */ 794 static struct napi_struct *napi_by_id(unsigned int napi_id) 795 { 796 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 797 struct napi_struct *napi; 798 799 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 800 if (napi->napi_id == napi_id) 801 return napi; 802 803 return NULL; 804 } 805 806 /* must be called under rcu_read_lock(), as we dont take a reference */ 807 static struct napi_struct * 808 netdev_napi_by_id(struct net *net, unsigned int napi_id) 809 { 810 struct napi_struct *napi; 811 812 napi = napi_by_id(napi_id); 813 if (!napi) 814 return NULL; 815 816 if (WARN_ON_ONCE(!napi->dev)) 817 return NULL; 818 if (!net_eq(net, dev_net(napi->dev))) 819 return NULL; 820 821 return napi; 822 } 823 824 /** 825 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 826 * @net: the applicable net namespace 827 * @napi_id: ID of a NAPI of a target device 828 * 829 * Find a NAPI instance with @napi_id. Lock its device. 830 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 831 * netdev_unlock() must be called to release it. 832 * 833 * Return: pointer to NAPI, its device with lock held, NULL if not found. 834 */ 835 struct napi_struct * 836 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 837 { 838 struct napi_struct *napi; 839 struct net_device *dev; 840 841 rcu_read_lock(); 842 napi = netdev_napi_by_id(net, napi_id); 843 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 844 rcu_read_unlock(); 845 return NULL; 846 } 847 848 dev = napi->dev; 849 dev_hold(dev); 850 rcu_read_unlock(); 851 852 dev = __netdev_put_lock(dev, net); 853 if (!dev) 854 return NULL; 855 856 rcu_read_lock(); 857 napi = netdev_napi_by_id(net, napi_id); 858 if (napi && napi->dev != dev) 859 napi = NULL; 860 rcu_read_unlock(); 861 862 if (!napi) 863 netdev_unlock(dev); 864 return napi; 865 } 866 867 /** 868 * __dev_get_by_name - find a device by its name 869 * @net: the applicable net namespace 870 * @name: name to find 871 * 872 * Find an interface by name. Must be called under RTNL semaphore. 873 * If the name is found a pointer to the device is returned. 874 * If the name is not found then %NULL is returned. The 875 * reference counters are not incremented so the caller must be 876 * careful with locks. 877 */ 878 879 struct net_device *__dev_get_by_name(struct net *net, const char *name) 880 { 881 struct netdev_name_node *node_name; 882 883 node_name = netdev_name_node_lookup(net, name); 884 return node_name ? node_name->dev : NULL; 885 } 886 EXPORT_SYMBOL(__dev_get_by_name); 887 888 /** 889 * dev_get_by_name_rcu - find a device by its name 890 * @net: the applicable net namespace 891 * @name: name to find 892 * 893 * Find an interface by name. 894 * If the name is found a pointer to the device is returned. 895 * If the name is not found then %NULL is returned. 896 * The reference counters are not incremented so the caller must be 897 * careful with locks. The caller must hold RCU lock. 898 */ 899 900 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 901 { 902 struct netdev_name_node *node_name; 903 904 node_name = netdev_name_node_lookup_rcu(net, name); 905 return node_name ? node_name->dev : NULL; 906 } 907 EXPORT_SYMBOL(dev_get_by_name_rcu); 908 909 /* Deprecated for new users, call netdev_get_by_name() instead */ 910 struct net_device *dev_get_by_name(struct net *net, const char *name) 911 { 912 struct net_device *dev; 913 914 rcu_read_lock(); 915 dev = dev_get_by_name_rcu(net, name); 916 dev_hold(dev); 917 rcu_read_unlock(); 918 return dev; 919 } 920 EXPORT_SYMBOL(dev_get_by_name); 921 922 /** 923 * netdev_get_by_name() - find a device by its name 924 * @net: the applicable net namespace 925 * @name: name to find 926 * @tracker: tracking object for the acquired reference 927 * @gfp: allocation flags for the tracker 928 * 929 * Find an interface by name. This can be called from any 930 * context and does its own locking. The returned handle has 931 * the usage count incremented and the caller must use netdev_put() to 932 * release it when it is no longer needed. %NULL is returned if no 933 * matching device is found. 934 */ 935 struct net_device *netdev_get_by_name(struct net *net, const char *name, 936 netdevice_tracker *tracker, gfp_t gfp) 937 { 938 struct net_device *dev; 939 940 dev = dev_get_by_name(net, name); 941 if (dev) 942 netdev_tracker_alloc(dev, tracker, gfp); 943 return dev; 944 } 945 EXPORT_SYMBOL(netdev_get_by_name); 946 947 /** 948 * __dev_get_by_index - find a device by its ifindex 949 * @net: the applicable net namespace 950 * @ifindex: index of device 951 * 952 * Search for an interface by index. Returns %NULL if the device 953 * is not found or a pointer to the device. The device has not 954 * had its reference counter increased so the caller must be careful 955 * about locking. The caller must hold the RTNL semaphore. 956 */ 957 958 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 959 { 960 struct net_device *dev; 961 struct hlist_head *head = dev_index_hash(net, ifindex); 962 963 hlist_for_each_entry(dev, head, index_hlist) 964 if (dev->ifindex == ifindex) 965 return dev; 966 967 return NULL; 968 } 969 EXPORT_SYMBOL(__dev_get_by_index); 970 971 /** 972 * dev_get_by_index_rcu - find a device by its ifindex 973 * @net: the applicable net namespace 974 * @ifindex: index of device 975 * 976 * Search for an interface by index. Returns %NULL if the device 977 * is not found or a pointer to the device. The device has not 978 * had its reference counter increased so the caller must be careful 979 * about locking. The caller must hold RCU lock. 980 */ 981 982 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 983 { 984 struct net_device *dev; 985 struct hlist_head *head = dev_index_hash(net, ifindex); 986 987 hlist_for_each_entry_rcu(dev, head, index_hlist) 988 if (dev->ifindex == ifindex) 989 return dev; 990 991 return NULL; 992 } 993 EXPORT_SYMBOL(dev_get_by_index_rcu); 994 995 /* Deprecated for new users, call netdev_get_by_index() instead */ 996 struct net_device *dev_get_by_index(struct net *net, int ifindex) 997 { 998 struct net_device *dev; 999 1000 rcu_read_lock(); 1001 dev = dev_get_by_index_rcu(net, ifindex); 1002 dev_hold(dev); 1003 rcu_read_unlock(); 1004 return dev; 1005 } 1006 EXPORT_SYMBOL(dev_get_by_index); 1007 1008 /** 1009 * netdev_get_by_index() - find a device by its ifindex 1010 * @net: the applicable net namespace 1011 * @ifindex: index of device 1012 * @tracker: tracking object for the acquired reference 1013 * @gfp: allocation flags for the tracker 1014 * 1015 * Search for an interface by index. Returns NULL if the device 1016 * is not found or a pointer to the device. The device returned has 1017 * had a reference added and the pointer is safe until the user calls 1018 * netdev_put() to indicate they have finished with it. 1019 */ 1020 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 1021 netdevice_tracker *tracker, gfp_t gfp) 1022 { 1023 struct net_device *dev; 1024 1025 dev = dev_get_by_index(net, ifindex); 1026 if (dev) 1027 netdev_tracker_alloc(dev, tracker, gfp); 1028 return dev; 1029 } 1030 EXPORT_SYMBOL(netdev_get_by_index); 1031 1032 /** 1033 * dev_get_by_napi_id - find a device by napi_id 1034 * @napi_id: ID of the NAPI struct 1035 * 1036 * Search for an interface by NAPI ID. Returns %NULL if the device 1037 * is not found or a pointer to the device. The device has not had 1038 * its reference counter increased so the caller must be careful 1039 * about locking. The caller must hold RCU lock. 1040 */ 1041 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1042 { 1043 struct napi_struct *napi; 1044 1045 WARN_ON_ONCE(!rcu_read_lock_held()); 1046 1047 if (!napi_id_valid(napi_id)) 1048 return NULL; 1049 1050 napi = napi_by_id(napi_id); 1051 1052 return napi ? napi->dev : NULL; 1053 } 1054 1055 /* Release the held reference on the net_device, and if the net_device 1056 * is still registered try to lock the instance lock. If device is being 1057 * unregistered NULL will be returned (but the reference has been released, 1058 * either way!) 1059 * 1060 * This helper is intended for locking net_device after it has been looked up 1061 * using a lockless lookup helper. Lock prevents the instance from going away. 1062 */ 1063 struct net_device *__netdev_put_lock(struct net_device *dev, struct net *net) 1064 { 1065 netdev_lock(dev); 1066 if (dev->reg_state > NETREG_REGISTERED || 1067 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1068 netdev_unlock(dev); 1069 dev_put(dev); 1070 return NULL; 1071 } 1072 dev_put(dev); 1073 return dev; 1074 } 1075 1076 static struct net_device * 1077 __netdev_put_lock_ops_compat(struct net_device *dev, struct net *net) 1078 { 1079 netdev_lock_ops_compat(dev); 1080 if (dev->reg_state > NETREG_REGISTERED || 1081 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1082 netdev_unlock_ops_compat(dev); 1083 dev_put(dev); 1084 return NULL; 1085 } 1086 dev_put(dev); 1087 return dev; 1088 } 1089 1090 /** 1091 * netdev_get_by_index_lock() - find a device by its ifindex 1092 * @net: the applicable net namespace 1093 * @ifindex: index of device 1094 * 1095 * Search for an interface by index. If a valid device 1096 * with @ifindex is found it will be returned with netdev->lock held. 1097 * netdev_unlock() must be called to release it. 1098 * 1099 * Return: pointer to a device with lock held, NULL if not found. 1100 */ 1101 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1102 { 1103 struct net_device *dev; 1104 1105 dev = dev_get_by_index(net, ifindex); 1106 if (!dev) 1107 return NULL; 1108 1109 return __netdev_put_lock(dev, net); 1110 } 1111 1112 struct net_device * 1113 netdev_get_by_index_lock_ops_compat(struct net *net, int ifindex) 1114 { 1115 struct net_device *dev; 1116 1117 dev = dev_get_by_index(net, ifindex); 1118 if (!dev) 1119 return NULL; 1120 1121 return __netdev_put_lock_ops_compat(dev, net); 1122 } 1123 1124 struct net_device * 1125 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1126 unsigned long *index) 1127 { 1128 if (dev) 1129 netdev_unlock(dev); 1130 1131 do { 1132 rcu_read_lock(); 1133 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1134 if (!dev) { 1135 rcu_read_unlock(); 1136 return NULL; 1137 } 1138 dev_hold(dev); 1139 rcu_read_unlock(); 1140 1141 dev = __netdev_put_lock(dev, net); 1142 if (dev) 1143 return dev; 1144 1145 (*index)++; 1146 } while (true); 1147 } 1148 1149 struct net_device * 1150 netdev_xa_find_lock_ops_compat(struct net *net, struct net_device *dev, 1151 unsigned long *index) 1152 { 1153 if (dev) 1154 netdev_unlock_ops_compat(dev); 1155 1156 do { 1157 rcu_read_lock(); 1158 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1159 if (!dev) { 1160 rcu_read_unlock(); 1161 return NULL; 1162 } 1163 dev_hold(dev); 1164 rcu_read_unlock(); 1165 1166 dev = __netdev_put_lock_ops_compat(dev, net); 1167 if (dev) 1168 return dev; 1169 1170 (*index)++; 1171 } while (true); 1172 } 1173 1174 static DEFINE_SEQLOCK(netdev_rename_lock); 1175 1176 void netdev_copy_name(struct net_device *dev, char *name) 1177 { 1178 unsigned int seq; 1179 1180 do { 1181 seq = read_seqbegin(&netdev_rename_lock); 1182 strscpy(name, dev->name, IFNAMSIZ); 1183 } while (read_seqretry(&netdev_rename_lock, seq)); 1184 } 1185 EXPORT_IPV6_MOD_GPL(netdev_copy_name); 1186 1187 /** 1188 * netdev_get_name - get a netdevice name, knowing its ifindex. 1189 * @net: network namespace 1190 * @name: a pointer to the buffer where the name will be stored. 1191 * @ifindex: the ifindex of the interface to get the name from. 1192 */ 1193 int netdev_get_name(struct net *net, char *name, int ifindex) 1194 { 1195 struct net_device *dev; 1196 int ret; 1197 1198 rcu_read_lock(); 1199 1200 dev = dev_get_by_index_rcu(net, ifindex); 1201 if (!dev) { 1202 ret = -ENODEV; 1203 goto out; 1204 } 1205 1206 netdev_copy_name(dev, name); 1207 1208 ret = 0; 1209 out: 1210 rcu_read_unlock(); 1211 return ret; 1212 } 1213 1214 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1215 const char *ha) 1216 { 1217 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1218 } 1219 1220 /** 1221 * dev_getbyhwaddr_rcu - find a device by its hardware address 1222 * @net: the applicable net namespace 1223 * @type: media type of device 1224 * @ha: hardware address 1225 * 1226 * Search for an interface by MAC address. Returns NULL if the device 1227 * is not found or a pointer to the device. 1228 * The caller must hold RCU. 1229 * The returned device has not had its ref count increased 1230 * and the caller must therefore be careful about locking 1231 * 1232 */ 1233 1234 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1235 const char *ha) 1236 { 1237 struct net_device *dev; 1238 1239 for_each_netdev_rcu(net, dev) 1240 if (dev_addr_cmp(dev, type, ha)) 1241 return dev; 1242 1243 return NULL; 1244 } 1245 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1246 1247 /** 1248 * dev_getbyhwaddr() - find a device by its hardware address 1249 * @net: the applicable net namespace 1250 * @type: media type of device 1251 * @ha: hardware address 1252 * 1253 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1254 * rtnl_lock. 1255 * 1256 * Context: rtnl_lock() must be held. 1257 * Return: pointer to the net_device, or NULL if not found 1258 */ 1259 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1260 const char *ha) 1261 { 1262 struct net_device *dev; 1263 1264 ASSERT_RTNL(); 1265 for_each_netdev(net, dev) 1266 if (dev_addr_cmp(dev, type, ha)) 1267 return dev; 1268 1269 return NULL; 1270 } 1271 EXPORT_SYMBOL(dev_getbyhwaddr); 1272 1273 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1274 { 1275 struct net_device *dev, *ret = NULL; 1276 1277 rcu_read_lock(); 1278 for_each_netdev_rcu(net, dev) 1279 if (dev->type == type) { 1280 dev_hold(dev); 1281 ret = dev; 1282 break; 1283 } 1284 rcu_read_unlock(); 1285 return ret; 1286 } 1287 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1288 1289 /** 1290 * netdev_get_by_flags_rcu - find any device with given flags 1291 * @net: the applicable net namespace 1292 * @tracker: tracking object for the acquired reference 1293 * @if_flags: IFF_* values 1294 * @mask: bitmask of bits in if_flags to check 1295 * 1296 * Search for any interface with the given flags. 1297 * 1298 * Context: rcu_read_lock() must be held. 1299 * Returns: NULL if a device is not found or a pointer to the device. 1300 */ 1301 struct net_device *netdev_get_by_flags_rcu(struct net *net, netdevice_tracker *tracker, 1302 unsigned short if_flags, unsigned short mask) 1303 { 1304 struct net_device *dev; 1305 1306 for_each_netdev_rcu(net, dev) { 1307 if (((READ_ONCE(dev->flags) ^ if_flags) & mask) == 0) { 1308 netdev_hold(dev, tracker, GFP_ATOMIC); 1309 return dev; 1310 } 1311 } 1312 1313 return NULL; 1314 } 1315 EXPORT_IPV6_MOD(netdev_get_by_flags_rcu); 1316 1317 /** 1318 * dev_valid_name - check if name is okay for network device 1319 * @name: name string 1320 * 1321 * Network device names need to be valid file names to 1322 * allow sysfs to work. We also disallow any kind of 1323 * whitespace. 1324 */ 1325 bool dev_valid_name(const char *name) 1326 { 1327 if (*name == '\0') 1328 return false; 1329 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1330 return false; 1331 if (!strcmp(name, ".") || !strcmp(name, "..")) 1332 return false; 1333 1334 while (*name) { 1335 if (*name == '/' || *name == ':' || isspace(*name)) 1336 return false; 1337 name++; 1338 } 1339 return true; 1340 } 1341 EXPORT_SYMBOL(dev_valid_name); 1342 1343 /** 1344 * __dev_alloc_name - allocate a name for a device 1345 * @net: network namespace to allocate the device name in 1346 * @name: name format string 1347 * @res: result name string 1348 * 1349 * Passed a format string - eg "lt%d" it will try and find a suitable 1350 * id. It scans list of devices to build up a free map, then chooses 1351 * the first empty slot. The caller must hold the dev_base or rtnl lock 1352 * while allocating the name and adding the device in order to avoid 1353 * duplicates. 1354 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1355 * Returns the number of the unit assigned or a negative errno code. 1356 */ 1357 1358 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1359 { 1360 int i = 0; 1361 const char *p; 1362 const int max_netdevices = 8*PAGE_SIZE; 1363 unsigned long *inuse; 1364 struct net_device *d; 1365 char buf[IFNAMSIZ]; 1366 1367 /* Verify the string as this thing may have come from the user. 1368 * There must be one "%d" and no other "%" characters. 1369 */ 1370 p = strchr(name, '%'); 1371 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1372 return -EINVAL; 1373 1374 /* Use one page as a bit array of possible slots */ 1375 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1376 if (!inuse) 1377 return -ENOMEM; 1378 1379 for_each_netdev(net, d) { 1380 struct netdev_name_node *name_node; 1381 1382 netdev_for_each_altname(d, name_node) { 1383 if (!sscanf(name_node->name, name, &i)) 1384 continue; 1385 if (i < 0 || i >= max_netdevices) 1386 continue; 1387 1388 /* avoid cases where sscanf is not exact inverse of printf */ 1389 snprintf(buf, IFNAMSIZ, name, i); 1390 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1391 __set_bit(i, inuse); 1392 } 1393 if (!sscanf(d->name, name, &i)) 1394 continue; 1395 if (i < 0 || i >= max_netdevices) 1396 continue; 1397 1398 /* avoid cases where sscanf is not exact inverse of printf */ 1399 snprintf(buf, IFNAMSIZ, name, i); 1400 if (!strncmp(buf, d->name, IFNAMSIZ)) 1401 __set_bit(i, inuse); 1402 } 1403 1404 i = find_first_zero_bit(inuse, max_netdevices); 1405 bitmap_free(inuse); 1406 if (i == max_netdevices) 1407 return -ENFILE; 1408 1409 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1410 strscpy(buf, name, IFNAMSIZ); 1411 snprintf(res, IFNAMSIZ, buf, i); 1412 return i; 1413 } 1414 1415 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1416 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1417 const char *want_name, char *out_name, 1418 int dup_errno) 1419 { 1420 if (!dev_valid_name(want_name)) 1421 return -EINVAL; 1422 1423 if (strchr(want_name, '%')) 1424 return __dev_alloc_name(net, want_name, out_name); 1425 1426 if (netdev_name_in_use(net, want_name)) 1427 return -dup_errno; 1428 if (out_name != want_name) 1429 strscpy(out_name, want_name, IFNAMSIZ); 1430 return 0; 1431 } 1432 1433 /** 1434 * dev_alloc_name - allocate a name for a device 1435 * @dev: device 1436 * @name: name format string 1437 * 1438 * Passed a format string - eg "lt%d" it will try and find a suitable 1439 * id. It scans list of devices to build up a free map, then chooses 1440 * the first empty slot. The caller must hold the dev_base or rtnl lock 1441 * while allocating the name and adding the device in order to avoid 1442 * duplicates. 1443 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1444 * Returns the number of the unit assigned or a negative errno code. 1445 */ 1446 1447 int dev_alloc_name(struct net_device *dev, const char *name) 1448 { 1449 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1450 } 1451 EXPORT_SYMBOL(dev_alloc_name); 1452 1453 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1454 const char *name) 1455 { 1456 int ret; 1457 1458 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1459 return ret < 0 ? ret : 0; 1460 } 1461 1462 int netif_change_name(struct net_device *dev, const char *newname) 1463 { 1464 struct net *net = dev_net(dev); 1465 unsigned char old_assign_type; 1466 char oldname[IFNAMSIZ]; 1467 int err = 0; 1468 int ret; 1469 1470 ASSERT_RTNL_NET(net); 1471 1472 if (!strncmp(newname, dev->name, IFNAMSIZ)) 1473 return 0; 1474 1475 memcpy(oldname, dev->name, IFNAMSIZ); 1476 1477 write_seqlock_bh(&netdev_rename_lock); 1478 err = dev_get_valid_name(net, dev, newname); 1479 write_sequnlock_bh(&netdev_rename_lock); 1480 1481 if (err < 0) 1482 return err; 1483 1484 if (oldname[0] && !strchr(oldname, '%')) 1485 netdev_info(dev, "renamed from %s%s\n", oldname, 1486 dev->flags & IFF_UP ? " (while UP)" : ""); 1487 1488 old_assign_type = dev->name_assign_type; 1489 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1490 1491 rollback: 1492 ret = device_rename(&dev->dev, dev->name); 1493 if (ret) { 1494 write_seqlock_bh(&netdev_rename_lock); 1495 memcpy(dev->name, oldname, IFNAMSIZ); 1496 write_sequnlock_bh(&netdev_rename_lock); 1497 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1498 return ret; 1499 } 1500 1501 netdev_adjacent_rename_links(dev, oldname); 1502 1503 netdev_name_node_del(dev->name_node); 1504 1505 synchronize_net(); 1506 1507 netdev_name_node_add(net, dev->name_node); 1508 1509 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1510 ret = notifier_to_errno(ret); 1511 1512 if (ret) { 1513 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1514 if (err >= 0) { 1515 err = ret; 1516 write_seqlock_bh(&netdev_rename_lock); 1517 memcpy(dev->name, oldname, IFNAMSIZ); 1518 write_sequnlock_bh(&netdev_rename_lock); 1519 memcpy(oldname, newname, IFNAMSIZ); 1520 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1521 old_assign_type = NET_NAME_RENAMED; 1522 goto rollback; 1523 } else { 1524 netdev_err(dev, "name change rollback failed: %d\n", 1525 ret); 1526 } 1527 } 1528 1529 return err; 1530 } 1531 1532 int netif_set_alias(struct net_device *dev, const char *alias, size_t len) 1533 { 1534 struct dev_ifalias *new_alias = NULL; 1535 1536 if (len >= IFALIASZ) 1537 return -EINVAL; 1538 1539 if (len) { 1540 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1541 if (!new_alias) 1542 return -ENOMEM; 1543 1544 memcpy(new_alias->ifalias, alias, len); 1545 new_alias->ifalias[len] = 0; 1546 } 1547 1548 mutex_lock(&ifalias_mutex); 1549 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1550 mutex_is_locked(&ifalias_mutex)); 1551 mutex_unlock(&ifalias_mutex); 1552 1553 if (new_alias) 1554 kfree_rcu(new_alias, rcuhead); 1555 1556 return len; 1557 } 1558 1559 /** 1560 * dev_get_alias - get ifalias of a device 1561 * @dev: device 1562 * @name: buffer to store name of ifalias 1563 * @len: size of buffer 1564 * 1565 * get ifalias for a device. Caller must make sure dev cannot go 1566 * away, e.g. rcu read lock or own a reference count to device. 1567 */ 1568 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1569 { 1570 const struct dev_ifalias *alias; 1571 int ret = 0; 1572 1573 rcu_read_lock(); 1574 alias = rcu_dereference(dev->ifalias); 1575 if (alias) 1576 ret = snprintf(name, len, "%s", alias->ifalias); 1577 rcu_read_unlock(); 1578 1579 return ret; 1580 } 1581 1582 /** 1583 * netdev_features_change - device changes features 1584 * @dev: device to cause notification 1585 * 1586 * Called to indicate a device has changed features. 1587 */ 1588 void netdev_features_change(struct net_device *dev) 1589 { 1590 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1591 } 1592 EXPORT_SYMBOL(netdev_features_change); 1593 1594 void netif_state_change(struct net_device *dev) 1595 { 1596 netdev_ops_assert_locked_or_invisible(dev); 1597 1598 if (dev->flags & IFF_UP) { 1599 struct netdev_notifier_change_info change_info = { 1600 .info.dev = dev, 1601 }; 1602 1603 call_netdevice_notifiers_info(NETDEV_CHANGE, 1604 &change_info.info); 1605 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1606 } 1607 } 1608 1609 /** 1610 * __netdev_notify_peers - notify network peers about existence of @dev, 1611 * to be called when rtnl lock is already held. 1612 * @dev: network device 1613 * 1614 * Generate traffic such that interested network peers are aware of 1615 * @dev, such as by generating a gratuitous ARP. This may be used when 1616 * a device wants to inform the rest of the network about some sort of 1617 * reconfiguration such as a failover event or virtual machine 1618 * migration. 1619 */ 1620 void __netdev_notify_peers(struct net_device *dev) 1621 { 1622 ASSERT_RTNL(); 1623 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1624 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1625 } 1626 EXPORT_SYMBOL(__netdev_notify_peers); 1627 1628 /** 1629 * netdev_notify_peers - notify network peers about existence of @dev 1630 * @dev: network device 1631 * 1632 * Generate traffic such that interested network peers are aware of 1633 * @dev, such as by generating a gratuitous ARP. This may be used when 1634 * a device wants to inform the rest of the network about some sort of 1635 * reconfiguration such as a failover event or virtual machine 1636 * migration. 1637 */ 1638 void netdev_notify_peers(struct net_device *dev) 1639 { 1640 rtnl_lock(); 1641 __netdev_notify_peers(dev); 1642 rtnl_unlock(); 1643 } 1644 EXPORT_SYMBOL(netdev_notify_peers); 1645 1646 static int napi_threaded_poll(void *data); 1647 1648 static int napi_kthread_create(struct napi_struct *n) 1649 { 1650 int err = 0; 1651 1652 /* Create and wake up the kthread once to put it in 1653 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1654 * warning and work with loadavg. 1655 */ 1656 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1657 n->dev->name, n->napi_id); 1658 if (IS_ERR(n->thread)) { 1659 err = PTR_ERR(n->thread); 1660 pr_err("kthread_run failed with err %d\n", err); 1661 n->thread = NULL; 1662 } 1663 1664 return err; 1665 } 1666 1667 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1668 { 1669 const struct net_device_ops *ops = dev->netdev_ops; 1670 int ret; 1671 1672 ASSERT_RTNL(); 1673 dev_addr_check(dev); 1674 1675 if (!netif_device_present(dev)) { 1676 /* may be detached because parent is runtime-suspended */ 1677 if (dev->dev.parent) 1678 pm_runtime_resume(dev->dev.parent); 1679 if (!netif_device_present(dev)) 1680 return -ENODEV; 1681 } 1682 1683 /* Block netpoll from trying to do any rx path servicing. 1684 * If we don't do this there is a chance ndo_poll_controller 1685 * or ndo_poll may be running while we open the device 1686 */ 1687 netpoll_poll_disable(dev); 1688 1689 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1690 ret = notifier_to_errno(ret); 1691 if (ret) 1692 return ret; 1693 1694 set_bit(__LINK_STATE_START, &dev->state); 1695 1696 netdev_ops_assert_locked(dev); 1697 1698 if (ops->ndo_validate_addr) 1699 ret = ops->ndo_validate_addr(dev); 1700 1701 if (!ret && ops->ndo_open) 1702 ret = ops->ndo_open(dev); 1703 1704 netpoll_poll_enable(dev); 1705 1706 if (ret) 1707 clear_bit(__LINK_STATE_START, &dev->state); 1708 else { 1709 netif_set_up(dev, true); 1710 dev_set_rx_mode(dev); 1711 dev_activate(dev); 1712 add_device_randomness(dev->dev_addr, dev->addr_len); 1713 } 1714 1715 return ret; 1716 } 1717 1718 int netif_open(struct net_device *dev, struct netlink_ext_ack *extack) 1719 { 1720 int ret; 1721 1722 if (dev->flags & IFF_UP) 1723 return 0; 1724 1725 ret = __dev_open(dev, extack); 1726 if (ret < 0) 1727 return ret; 1728 1729 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1730 call_netdevice_notifiers(NETDEV_UP, dev); 1731 1732 return ret; 1733 } 1734 1735 static void __dev_close_many(struct list_head *head) 1736 { 1737 struct net_device *dev; 1738 1739 ASSERT_RTNL(); 1740 might_sleep(); 1741 1742 list_for_each_entry(dev, head, close_list) { 1743 /* Temporarily disable netpoll until the interface is down */ 1744 netpoll_poll_disable(dev); 1745 1746 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1747 1748 clear_bit(__LINK_STATE_START, &dev->state); 1749 1750 /* Synchronize to scheduled poll. We cannot touch poll list, it 1751 * can be even on different cpu. So just clear netif_running(). 1752 * 1753 * dev->stop() will invoke napi_disable() on all of it's 1754 * napi_struct instances on this device. 1755 */ 1756 smp_mb__after_atomic(); /* Commit netif_running(). */ 1757 } 1758 1759 dev_deactivate_many(head); 1760 1761 list_for_each_entry(dev, head, close_list) { 1762 const struct net_device_ops *ops = dev->netdev_ops; 1763 1764 /* 1765 * Call the device specific close. This cannot fail. 1766 * Only if device is UP 1767 * 1768 * We allow it to be called even after a DETACH hot-plug 1769 * event. 1770 */ 1771 1772 netdev_ops_assert_locked(dev); 1773 1774 if (ops->ndo_stop) 1775 ops->ndo_stop(dev); 1776 1777 netif_set_up(dev, false); 1778 netpoll_poll_enable(dev); 1779 } 1780 } 1781 1782 static void __dev_close(struct net_device *dev) 1783 { 1784 LIST_HEAD(single); 1785 1786 list_add(&dev->close_list, &single); 1787 __dev_close_many(&single); 1788 list_del(&single); 1789 } 1790 1791 void netif_close_many(struct list_head *head, bool unlink) 1792 { 1793 struct net_device *dev, *tmp; 1794 1795 /* Remove the devices that don't need to be closed */ 1796 list_for_each_entry_safe(dev, tmp, head, close_list) 1797 if (!(dev->flags & IFF_UP)) 1798 list_del_init(&dev->close_list); 1799 1800 __dev_close_many(head); 1801 1802 list_for_each_entry_safe(dev, tmp, head, close_list) { 1803 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1804 call_netdevice_notifiers(NETDEV_DOWN, dev); 1805 if (unlink) 1806 list_del_init(&dev->close_list); 1807 } 1808 } 1809 EXPORT_SYMBOL_NS_GPL(netif_close_many, "NETDEV_INTERNAL"); 1810 1811 void netif_close(struct net_device *dev) 1812 { 1813 if (dev->flags & IFF_UP) { 1814 LIST_HEAD(single); 1815 1816 list_add(&dev->close_list, &single); 1817 netif_close_many(&single, true); 1818 list_del(&single); 1819 } 1820 } 1821 EXPORT_SYMBOL(netif_close); 1822 1823 void netif_disable_lro(struct net_device *dev) 1824 { 1825 struct net_device *lower_dev; 1826 struct list_head *iter; 1827 1828 dev->wanted_features &= ~NETIF_F_LRO; 1829 netdev_update_features(dev); 1830 1831 if (unlikely(dev->features & NETIF_F_LRO)) 1832 netdev_WARN(dev, "failed to disable LRO!\n"); 1833 1834 netdev_for_each_lower_dev(dev, lower_dev, iter) { 1835 netdev_lock_ops(lower_dev); 1836 netif_disable_lro(lower_dev); 1837 netdev_unlock_ops(lower_dev); 1838 } 1839 } 1840 EXPORT_IPV6_MOD(netif_disable_lro); 1841 1842 /** 1843 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1844 * @dev: device 1845 * 1846 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1847 * called under RTNL. This is needed if Generic XDP is installed on 1848 * the device. 1849 */ 1850 static void dev_disable_gro_hw(struct net_device *dev) 1851 { 1852 dev->wanted_features &= ~NETIF_F_GRO_HW; 1853 netdev_update_features(dev); 1854 1855 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1856 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1857 } 1858 1859 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1860 { 1861 #define N(val) \ 1862 case NETDEV_##val: \ 1863 return "NETDEV_" __stringify(val); 1864 switch (cmd) { 1865 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1866 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1867 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1868 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1869 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1870 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1871 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1872 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1873 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1874 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1875 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1876 N(XDP_FEAT_CHANGE) 1877 } 1878 #undef N 1879 return "UNKNOWN_NETDEV_EVENT"; 1880 } 1881 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1882 1883 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1884 struct net_device *dev) 1885 { 1886 struct netdev_notifier_info info = { 1887 .dev = dev, 1888 }; 1889 1890 return nb->notifier_call(nb, val, &info); 1891 } 1892 1893 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1894 struct net_device *dev) 1895 { 1896 int err; 1897 1898 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1899 err = notifier_to_errno(err); 1900 if (err) 1901 return err; 1902 1903 if (!(dev->flags & IFF_UP)) 1904 return 0; 1905 1906 call_netdevice_notifier(nb, NETDEV_UP, dev); 1907 return 0; 1908 } 1909 1910 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1911 struct net_device *dev) 1912 { 1913 if (dev->flags & IFF_UP) { 1914 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1915 dev); 1916 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1917 } 1918 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1919 } 1920 1921 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1922 struct net *net) 1923 { 1924 struct net_device *dev; 1925 int err; 1926 1927 for_each_netdev(net, dev) { 1928 netdev_lock_ops(dev); 1929 err = call_netdevice_register_notifiers(nb, dev); 1930 netdev_unlock_ops(dev); 1931 if (err) 1932 goto rollback; 1933 } 1934 return 0; 1935 1936 rollback: 1937 for_each_netdev_continue_reverse(net, dev) 1938 call_netdevice_unregister_notifiers(nb, dev); 1939 return err; 1940 } 1941 1942 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1943 struct net *net) 1944 { 1945 struct net_device *dev; 1946 1947 for_each_netdev(net, dev) 1948 call_netdevice_unregister_notifiers(nb, dev); 1949 } 1950 1951 static int dev_boot_phase = 1; 1952 1953 /** 1954 * register_netdevice_notifier - register a network notifier block 1955 * @nb: notifier 1956 * 1957 * Register a notifier to be called when network device events occur. 1958 * The notifier passed is linked into the kernel structures and must 1959 * not be reused until it has been unregistered. A negative errno code 1960 * is returned on a failure. 1961 * 1962 * When registered all registration and up events are replayed 1963 * to the new notifier to allow device to have a race free 1964 * view of the network device list. 1965 */ 1966 1967 int register_netdevice_notifier(struct notifier_block *nb) 1968 { 1969 struct net *net; 1970 int err; 1971 1972 /* Close race with setup_net() and cleanup_net() */ 1973 down_write(&pernet_ops_rwsem); 1974 1975 /* When RTNL is removed, we need protection for netdev_chain. */ 1976 rtnl_lock(); 1977 1978 err = raw_notifier_chain_register(&netdev_chain, nb); 1979 if (err) 1980 goto unlock; 1981 if (dev_boot_phase) 1982 goto unlock; 1983 for_each_net(net) { 1984 __rtnl_net_lock(net); 1985 err = call_netdevice_register_net_notifiers(nb, net); 1986 __rtnl_net_unlock(net); 1987 if (err) 1988 goto rollback; 1989 } 1990 1991 unlock: 1992 rtnl_unlock(); 1993 up_write(&pernet_ops_rwsem); 1994 return err; 1995 1996 rollback: 1997 for_each_net_continue_reverse(net) { 1998 __rtnl_net_lock(net); 1999 call_netdevice_unregister_net_notifiers(nb, net); 2000 __rtnl_net_unlock(net); 2001 } 2002 2003 raw_notifier_chain_unregister(&netdev_chain, nb); 2004 goto unlock; 2005 } 2006 EXPORT_SYMBOL(register_netdevice_notifier); 2007 2008 /** 2009 * unregister_netdevice_notifier - unregister a network notifier block 2010 * @nb: notifier 2011 * 2012 * Unregister a notifier previously registered by 2013 * register_netdevice_notifier(). The notifier is unlinked into the 2014 * kernel structures and may then be reused. A negative errno code 2015 * is returned on a failure. 2016 * 2017 * After unregistering unregister and down device events are synthesized 2018 * for all devices on the device list to the removed notifier to remove 2019 * the need for special case cleanup code. 2020 */ 2021 2022 int unregister_netdevice_notifier(struct notifier_block *nb) 2023 { 2024 struct net *net; 2025 int err; 2026 2027 /* Close race with setup_net() and cleanup_net() */ 2028 down_write(&pernet_ops_rwsem); 2029 rtnl_lock(); 2030 err = raw_notifier_chain_unregister(&netdev_chain, nb); 2031 if (err) 2032 goto unlock; 2033 2034 for_each_net(net) { 2035 __rtnl_net_lock(net); 2036 call_netdevice_unregister_net_notifiers(nb, net); 2037 __rtnl_net_unlock(net); 2038 } 2039 2040 unlock: 2041 rtnl_unlock(); 2042 up_write(&pernet_ops_rwsem); 2043 return err; 2044 } 2045 EXPORT_SYMBOL(unregister_netdevice_notifier); 2046 2047 static int __register_netdevice_notifier_net(struct net *net, 2048 struct notifier_block *nb, 2049 bool ignore_call_fail) 2050 { 2051 int err; 2052 2053 err = raw_notifier_chain_register(&net->netdev_chain, nb); 2054 if (err) 2055 return err; 2056 if (dev_boot_phase) 2057 return 0; 2058 2059 err = call_netdevice_register_net_notifiers(nb, net); 2060 if (err && !ignore_call_fail) 2061 goto chain_unregister; 2062 2063 return 0; 2064 2065 chain_unregister: 2066 raw_notifier_chain_unregister(&net->netdev_chain, nb); 2067 return err; 2068 } 2069 2070 static int __unregister_netdevice_notifier_net(struct net *net, 2071 struct notifier_block *nb) 2072 { 2073 int err; 2074 2075 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 2076 if (err) 2077 return err; 2078 2079 call_netdevice_unregister_net_notifiers(nb, net); 2080 return 0; 2081 } 2082 2083 /** 2084 * register_netdevice_notifier_net - register a per-netns network notifier block 2085 * @net: network namespace 2086 * @nb: notifier 2087 * 2088 * Register a notifier to be called when network device events occur. 2089 * The notifier passed is linked into the kernel structures and must 2090 * not be reused until it has been unregistered. A negative errno code 2091 * is returned on a failure. 2092 * 2093 * When registered all registration and up events are replayed 2094 * to the new notifier to allow device to have a race free 2095 * view of the network device list. 2096 */ 2097 2098 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 2099 { 2100 int err; 2101 2102 rtnl_net_lock(net); 2103 err = __register_netdevice_notifier_net(net, nb, false); 2104 rtnl_net_unlock(net); 2105 2106 return err; 2107 } 2108 EXPORT_SYMBOL(register_netdevice_notifier_net); 2109 2110 /** 2111 * unregister_netdevice_notifier_net - unregister a per-netns 2112 * network notifier block 2113 * @net: network namespace 2114 * @nb: notifier 2115 * 2116 * Unregister a notifier previously registered by 2117 * register_netdevice_notifier_net(). The notifier is unlinked from the 2118 * kernel structures and may then be reused. A negative errno code 2119 * is returned on a failure. 2120 * 2121 * After unregistering unregister and down device events are synthesized 2122 * for all devices on the device list to the removed notifier to remove 2123 * the need for special case cleanup code. 2124 */ 2125 2126 int unregister_netdevice_notifier_net(struct net *net, 2127 struct notifier_block *nb) 2128 { 2129 int err; 2130 2131 rtnl_net_lock(net); 2132 err = __unregister_netdevice_notifier_net(net, nb); 2133 rtnl_net_unlock(net); 2134 2135 return err; 2136 } 2137 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 2138 2139 static void __move_netdevice_notifier_net(struct net *src_net, 2140 struct net *dst_net, 2141 struct notifier_block *nb) 2142 { 2143 __unregister_netdevice_notifier_net(src_net, nb); 2144 __register_netdevice_notifier_net(dst_net, nb, true); 2145 } 2146 2147 static void rtnl_net_dev_lock(struct net_device *dev) 2148 { 2149 bool again; 2150 2151 do { 2152 struct net *net; 2153 2154 again = false; 2155 2156 /* netns might be being dismantled. */ 2157 rcu_read_lock(); 2158 net = dev_net_rcu(dev); 2159 net_passive_inc(net); 2160 rcu_read_unlock(); 2161 2162 rtnl_net_lock(net); 2163 2164 #ifdef CONFIG_NET_NS 2165 /* dev might have been moved to another netns. */ 2166 if (!net_eq(net, rcu_access_pointer(dev->nd_net.net))) { 2167 rtnl_net_unlock(net); 2168 net_passive_dec(net); 2169 again = true; 2170 } 2171 #endif 2172 } while (again); 2173 } 2174 2175 static void rtnl_net_dev_unlock(struct net_device *dev) 2176 { 2177 struct net *net = dev_net(dev); 2178 2179 rtnl_net_unlock(net); 2180 net_passive_dec(net); 2181 } 2182 2183 int register_netdevice_notifier_dev_net(struct net_device *dev, 2184 struct notifier_block *nb, 2185 struct netdev_net_notifier *nn) 2186 { 2187 int err; 2188 2189 rtnl_net_dev_lock(dev); 2190 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 2191 if (!err) { 2192 nn->nb = nb; 2193 list_add(&nn->list, &dev->net_notifier_list); 2194 } 2195 rtnl_net_dev_unlock(dev); 2196 2197 return err; 2198 } 2199 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 2200 2201 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 2202 struct notifier_block *nb, 2203 struct netdev_net_notifier *nn) 2204 { 2205 int err; 2206 2207 rtnl_net_dev_lock(dev); 2208 list_del(&nn->list); 2209 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 2210 rtnl_net_dev_unlock(dev); 2211 2212 return err; 2213 } 2214 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 2215 2216 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 2217 struct net *net) 2218 { 2219 struct netdev_net_notifier *nn; 2220 2221 list_for_each_entry(nn, &dev->net_notifier_list, list) 2222 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 2223 } 2224 2225 /** 2226 * call_netdevice_notifiers_info - call all network notifier blocks 2227 * @val: value passed unmodified to notifier function 2228 * @info: notifier information data 2229 * 2230 * Call all network notifier blocks. Parameters and return value 2231 * are as for raw_notifier_call_chain(). 2232 */ 2233 2234 int call_netdevice_notifiers_info(unsigned long val, 2235 struct netdev_notifier_info *info) 2236 { 2237 struct net *net = dev_net(info->dev); 2238 int ret; 2239 2240 ASSERT_RTNL(); 2241 2242 /* Run per-netns notifier block chain first, then run the global one. 2243 * Hopefully, one day, the global one is going to be removed after 2244 * all notifier block registrators get converted to be per-netns. 2245 */ 2246 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 2247 if (ret & NOTIFY_STOP_MASK) 2248 return ret; 2249 return raw_notifier_call_chain(&netdev_chain, val, info); 2250 } 2251 2252 /** 2253 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2254 * for and rollback on error 2255 * @val_up: value passed unmodified to notifier function 2256 * @val_down: value passed unmodified to the notifier function when 2257 * recovering from an error on @val_up 2258 * @info: notifier information data 2259 * 2260 * Call all per-netns network notifier blocks, but not notifier blocks on 2261 * the global notifier chain. Parameters and return value are as for 2262 * raw_notifier_call_chain_robust(). 2263 */ 2264 2265 static int 2266 call_netdevice_notifiers_info_robust(unsigned long val_up, 2267 unsigned long val_down, 2268 struct netdev_notifier_info *info) 2269 { 2270 struct net *net = dev_net(info->dev); 2271 2272 ASSERT_RTNL(); 2273 2274 return raw_notifier_call_chain_robust(&net->netdev_chain, 2275 val_up, val_down, info); 2276 } 2277 2278 static int call_netdevice_notifiers_extack(unsigned long val, 2279 struct net_device *dev, 2280 struct netlink_ext_ack *extack) 2281 { 2282 struct netdev_notifier_info info = { 2283 .dev = dev, 2284 .extack = extack, 2285 }; 2286 2287 return call_netdevice_notifiers_info(val, &info); 2288 } 2289 2290 /** 2291 * call_netdevice_notifiers - call all network notifier blocks 2292 * @val: value passed unmodified to notifier function 2293 * @dev: net_device pointer passed unmodified to notifier function 2294 * 2295 * Call all network notifier blocks. Parameters and return value 2296 * are as for raw_notifier_call_chain(). 2297 */ 2298 2299 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2300 { 2301 return call_netdevice_notifiers_extack(val, dev, NULL); 2302 } 2303 EXPORT_SYMBOL(call_netdevice_notifiers); 2304 2305 /** 2306 * call_netdevice_notifiers_mtu - call all network notifier blocks 2307 * @val: value passed unmodified to notifier function 2308 * @dev: net_device pointer passed unmodified to notifier function 2309 * @arg: additional u32 argument passed to the notifier function 2310 * 2311 * Call all network notifier blocks. Parameters and return value 2312 * are as for raw_notifier_call_chain(). 2313 */ 2314 static int call_netdevice_notifiers_mtu(unsigned long val, 2315 struct net_device *dev, u32 arg) 2316 { 2317 struct netdev_notifier_info_ext info = { 2318 .info.dev = dev, 2319 .ext.mtu = arg, 2320 }; 2321 2322 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2323 2324 return call_netdevice_notifiers_info(val, &info.info); 2325 } 2326 2327 #ifdef CONFIG_NET_INGRESS 2328 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2329 2330 void net_inc_ingress_queue(void) 2331 { 2332 static_branch_inc(&ingress_needed_key); 2333 } 2334 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2335 2336 void net_dec_ingress_queue(void) 2337 { 2338 static_branch_dec(&ingress_needed_key); 2339 } 2340 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2341 #endif 2342 2343 #ifdef CONFIG_NET_EGRESS 2344 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2345 2346 void net_inc_egress_queue(void) 2347 { 2348 static_branch_inc(&egress_needed_key); 2349 } 2350 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2351 2352 void net_dec_egress_queue(void) 2353 { 2354 static_branch_dec(&egress_needed_key); 2355 } 2356 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2357 #endif 2358 2359 #ifdef CONFIG_NET_CLS_ACT 2360 DEFINE_STATIC_KEY_FALSE(tcf_sw_enabled_key); 2361 EXPORT_SYMBOL(tcf_sw_enabled_key); 2362 #endif 2363 2364 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2365 EXPORT_SYMBOL(netstamp_needed_key); 2366 #ifdef CONFIG_JUMP_LABEL 2367 static atomic_t netstamp_needed_deferred; 2368 static atomic_t netstamp_wanted; 2369 static void netstamp_clear(struct work_struct *work) 2370 { 2371 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2372 int wanted; 2373 2374 wanted = atomic_add_return(deferred, &netstamp_wanted); 2375 if (wanted > 0) 2376 static_branch_enable(&netstamp_needed_key); 2377 else 2378 static_branch_disable(&netstamp_needed_key); 2379 } 2380 static DECLARE_WORK(netstamp_work, netstamp_clear); 2381 #endif 2382 2383 void net_enable_timestamp(void) 2384 { 2385 #ifdef CONFIG_JUMP_LABEL 2386 int wanted = atomic_read(&netstamp_wanted); 2387 2388 while (wanted > 0) { 2389 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2390 return; 2391 } 2392 atomic_inc(&netstamp_needed_deferred); 2393 schedule_work(&netstamp_work); 2394 #else 2395 static_branch_inc(&netstamp_needed_key); 2396 #endif 2397 } 2398 EXPORT_SYMBOL(net_enable_timestamp); 2399 2400 void net_disable_timestamp(void) 2401 { 2402 #ifdef CONFIG_JUMP_LABEL 2403 int wanted = atomic_read(&netstamp_wanted); 2404 2405 while (wanted > 1) { 2406 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2407 return; 2408 } 2409 atomic_dec(&netstamp_needed_deferred); 2410 schedule_work(&netstamp_work); 2411 #else 2412 static_branch_dec(&netstamp_needed_key); 2413 #endif 2414 } 2415 EXPORT_SYMBOL(net_disable_timestamp); 2416 2417 static inline void net_timestamp_set(struct sk_buff *skb) 2418 { 2419 skb->tstamp = 0; 2420 skb->tstamp_type = SKB_CLOCK_REALTIME; 2421 if (static_branch_unlikely(&netstamp_needed_key)) 2422 skb->tstamp = ktime_get_real(); 2423 } 2424 2425 #define net_timestamp_check(COND, SKB) \ 2426 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2427 if ((COND) && !(SKB)->tstamp) \ 2428 (SKB)->tstamp = ktime_get_real(); \ 2429 } \ 2430 2431 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2432 { 2433 return __is_skb_forwardable(dev, skb, true); 2434 } 2435 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2436 2437 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2438 bool check_mtu) 2439 { 2440 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2441 2442 if (likely(!ret)) { 2443 skb->protocol = eth_type_trans(skb, dev); 2444 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2445 } 2446 2447 return ret; 2448 } 2449 2450 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2451 { 2452 return __dev_forward_skb2(dev, skb, true); 2453 } 2454 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2455 2456 /** 2457 * dev_forward_skb - loopback an skb to another netif 2458 * 2459 * @dev: destination network device 2460 * @skb: buffer to forward 2461 * 2462 * return values: 2463 * NET_RX_SUCCESS (no congestion) 2464 * NET_RX_DROP (packet was dropped, but freed) 2465 * 2466 * dev_forward_skb can be used for injecting an skb from the 2467 * start_xmit function of one device into the receive queue 2468 * of another device. 2469 * 2470 * The receiving device may be in another namespace, so 2471 * we have to clear all information in the skb that could 2472 * impact namespace isolation. 2473 */ 2474 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2475 { 2476 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2477 } 2478 EXPORT_SYMBOL_GPL(dev_forward_skb); 2479 2480 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2481 { 2482 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2483 } 2484 2485 static int deliver_skb(struct sk_buff *skb, 2486 struct packet_type *pt_prev, 2487 struct net_device *orig_dev) 2488 { 2489 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2490 return -ENOMEM; 2491 refcount_inc(&skb->users); 2492 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2493 } 2494 2495 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2496 struct packet_type **pt, 2497 struct net_device *orig_dev, 2498 __be16 type, 2499 struct list_head *ptype_list) 2500 { 2501 struct packet_type *ptype, *pt_prev = *pt; 2502 2503 list_for_each_entry_rcu(ptype, ptype_list, list) { 2504 if (ptype->type != type) 2505 continue; 2506 if (unlikely(pt_prev)) 2507 deliver_skb(skb, pt_prev, orig_dev); 2508 pt_prev = ptype; 2509 } 2510 *pt = pt_prev; 2511 } 2512 2513 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2514 { 2515 if (!ptype->af_packet_priv || !skb->sk) 2516 return false; 2517 2518 if (ptype->id_match) 2519 return ptype->id_match(ptype, skb->sk); 2520 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2521 return true; 2522 2523 return false; 2524 } 2525 2526 /** 2527 * dev_nit_active_rcu - return true if any network interface taps are in use 2528 * 2529 * The caller must hold the RCU lock 2530 * 2531 * @dev: network device to check for the presence of taps 2532 */ 2533 bool dev_nit_active_rcu(const struct net_device *dev) 2534 { 2535 /* Callers may hold either RCU or RCU BH lock */ 2536 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 2537 2538 return !list_empty(&dev_net(dev)->ptype_all) || 2539 !list_empty(&dev->ptype_all); 2540 } 2541 EXPORT_SYMBOL_GPL(dev_nit_active_rcu); 2542 2543 /* 2544 * Support routine. Sends outgoing frames to any network 2545 * taps currently in use. 2546 */ 2547 2548 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2549 { 2550 struct packet_type *ptype, *pt_prev = NULL; 2551 struct list_head *ptype_list; 2552 struct sk_buff *skb2 = NULL; 2553 2554 rcu_read_lock(); 2555 ptype_list = &dev_net_rcu(dev)->ptype_all; 2556 again: 2557 list_for_each_entry_rcu(ptype, ptype_list, list) { 2558 if (READ_ONCE(ptype->ignore_outgoing)) 2559 continue; 2560 2561 /* Never send packets back to the socket 2562 * they originated from - MvS (miquels@drinkel.ow.org) 2563 */ 2564 if (skb_loop_sk(ptype, skb)) 2565 continue; 2566 2567 if (unlikely(pt_prev)) { 2568 deliver_skb(skb2, pt_prev, skb->dev); 2569 pt_prev = ptype; 2570 continue; 2571 } 2572 2573 /* need to clone skb, done only once */ 2574 skb2 = skb_clone(skb, GFP_ATOMIC); 2575 if (!skb2) 2576 goto out_unlock; 2577 2578 net_timestamp_set(skb2); 2579 2580 /* skb->nh should be correctly 2581 * set by sender, so that the second statement is 2582 * just protection against buggy protocols. 2583 */ 2584 skb_reset_mac_header(skb2); 2585 2586 if (skb_network_header(skb2) < skb2->data || 2587 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2588 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2589 ntohs(skb2->protocol), 2590 dev->name); 2591 skb_reset_network_header(skb2); 2592 } 2593 2594 skb2->transport_header = skb2->network_header; 2595 skb2->pkt_type = PACKET_OUTGOING; 2596 pt_prev = ptype; 2597 } 2598 2599 if (ptype_list != &dev->ptype_all) { 2600 ptype_list = &dev->ptype_all; 2601 goto again; 2602 } 2603 out_unlock: 2604 if (pt_prev) { 2605 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2606 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2607 else 2608 kfree_skb(skb2); 2609 } 2610 rcu_read_unlock(); 2611 } 2612 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2613 2614 /** 2615 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2616 * @dev: Network device 2617 * @txq: number of queues available 2618 * 2619 * If real_num_tx_queues is changed the tc mappings may no longer be 2620 * valid. To resolve this verify the tc mapping remains valid and if 2621 * not NULL the mapping. With no priorities mapping to this 2622 * offset/count pair it will no longer be used. In the worst case TC0 2623 * is invalid nothing can be done so disable priority mappings. If is 2624 * expected that drivers will fix this mapping if they can before 2625 * calling netif_set_real_num_tx_queues. 2626 */ 2627 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2628 { 2629 int i; 2630 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2631 2632 /* If TC0 is invalidated disable TC mapping */ 2633 if (tc->offset + tc->count > txq) { 2634 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2635 dev->num_tc = 0; 2636 return; 2637 } 2638 2639 /* Invalidated prio to tc mappings set to TC0 */ 2640 for (i = 1; i < TC_BITMASK + 1; i++) { 2641 int q = netdev_get_prio_tc_map(dev, i); 2642 2643 tc = &dev->tc_to_txq[q]; 2644 if (tc->offset + tc->count > txq) { 2645 netdev_warn(dev, "Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n", 2646 i, q); 2647 netdev_set_prio_tc_map(dev, i, 0); 2648 } 2649 } 2650 } 2651 2652 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2653 { 2654 if (dev->num_tc) { 2655 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2656 int i; 2657 2658 /* walk through the TCs and see if it falls into any of them */ 2659 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2660 if ((txq - tc->offset) < tc->count) 2661 return i; 2662 } 2663 2664 /* didn't find it, just return -1 to indicate no match */ 2665 return -1; 2666 } 2667 2668 return 0; 2669 } 2670 EXPORT_SYMBOL(netdev_txq_to_tc); 2671 2672 #ifdef CONFIG_XPS 2673 static struct static_key xps_needed __read_mostly; 2674 static struct static_key xps_rxqs_needed __read_mostly; 2675 static DEFINE_MUTEX(xps_map_mutex); 2676 #define xmap_dereference(P) \ 2677 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2678 2679 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2680 struct xps_dev_maps *old_maps, int tci, u16 index) 2681 { 2682 struct xps_map *map = NULL; 2683 int pos; 2684 2685 map = xmap_dereference(dev_maps->attr_map[tci]); 2686 if (!map) 2687 return false; 2688 2689 for (pos = map->len; pos--;) { 2690 if (map->queues[pos] != index) 2691 continue; 2692 2693 if (map->len > 1) { 2694 map->queues[pos] = map->queues[--map->len]; 2695 break; 2696 } 2697 2698 if (old_maps) 2699 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2700 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2701 kfree_rcu(map, rcu); 2702 return false; 2703 } 2704 2705 return true; 2706 } 2707 2708 static bool remove_xps_queue_cpu(struct net_device *dev, 2709 struct xps_dev_maps *dev_maps, 2710 int cpu, u16 offset, u16 count) 2711 { 2712 int num_tc = dev_maps->num_tc; 2713 bool active = false; 2714 int tci; 2715 2716 for (tci = cpu * num_tc; num_tc--; tci++) { 2717 int i, j; 2718 2719 for (i = count, j = offset; i--; j++) { 2720 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2721 break; 2722 } 2723 2724 active |= i < 0; 2725 } 2726 2727 return active; 2728 } 2729 2730 static void reset_xps_maps(struct net_device *dev, 2731 struct xps_dev_maps *dev_maps, 2732 enum xps_map_type type) 2733 { 2734 static_key_slow_dec_cpuslocked(&xps_needed); 2735 if (type == XPS_RXQS) 2736 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2737 2738 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2739 2740 kfree_rcu(dev_maps, rcu); 2741 } 2742 2743 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2744 u16 offset, u16 count) 2745 { 2746 struct xps_dev_maps *dev_maps; 2747 bool active = false; 2748 int i, j; 2749 2750 dev_maps = xmap_dereference(dev->xps_maps[type]); 2751 if (!dev_maps) 2752 return; 2753 2754 for (j = 0; j < dev_maps->nr_ids; j++) 2755 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2756 if (!active) 2757 reset_xps_maps(dev, dev_maps, type); 2758 2759 if (type == XPS_CPUS) { 2760 for (i = offset + (count - 1); count--; i--) 2761 netdev_queue_numa_node_write( 2762 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2763 } 2764 } 2765 2766 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2767 u16 count) 2768 { 2769 if (!static_key_false(&xps_needed)) 2770 return; 2771 2772 cpus_read_lock(); 2773 mutex_lock(&xps_map_mutex); 2774 2775 if (static_key_false(&xps_rxqs_needed)) 2776 clean_xps_maps(dev, XPS_RXQS, offset, count); 2777 2778 clean_xps_maps(dev, XPS_CPUS, offset, count); 2779 2780 mutex_unlock(&xps_map_mutex); 2781 cpus_read_unlock(); 2782 } 2783 2784 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2785 { 2786 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2787 } 2788 2789 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2790 u16 index, bool is_rxqs_map) 2791 { 2792 struct xps_map *new_map; 2793 int alloc_len = XPS_MIN_MAP_ALLOC; 2794 int i, pos; 2795 2796 for (pos = 0; map && pos < map->len; pos++) { 2797 if (map->queues[pos] != index) 2798 continue; 2799 return map; 2800 } 2801 2802 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2803 if (map) { 2804 if (pos < map->alloc_len) 2805 return map; 2806 2807 alloc_len = map->alloc_len * 2; 2808 } 2809 2810 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2811 * map 2812 */ 2813 if (is_rxqs_map) 2814 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2815 else 2816 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2817 cpu_to_node(attr_index)); 2818 if (!new_map) 2819 return NULL; 2820 2821 for (i = 0; i < pos; i++) 2822 new_map->queues[i] = map->queues[i]; 2823 new_map->alloc_len = alloc_len; 2824 new_map->len = pos; 2825 2826 return new_map; 2827 } 2828 2829 /* Copy xps maps at a given index */ 2830 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2831 struct xps_dev_maps *new_dev_maps, int index, 2832 int tc, bool skip_tc) 2833 { 2834 int i, tci = index * dev_maps->num_tc; 2835 struct xps_map *map; 2836 2837 /* copy maps belonging to foreign traffic classes */ 2838 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2839 if (i == tc && skip_tc) 2840 continue; 2841 2842 /* fill in the new device map from the old device map */ 2843 map = xmap_dereference(dev_maps->attr_map[tci]); 2844 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2845 } 2846 } 2847 2848 /* Must be called under cpus_read_lock */ 2849 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2850 u16 index, enum xps_map_type type) 2851 { 2852 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2853 const unsigned long *online_mask = NULL; 2854 bool active = false, copy = false; 2855 int i, j, tci, numa_node_id = -2; 2856 int maps_sz, num_tc = 1, tc = 0; 2857 struct xps_map *map, *new_map; 2858 unsigned int nr_ids; 2859 2860 WARN_ON_ONCE(index >= dev->num_tx_queues); 2861 2862 if (dev->num_tc) { 2863 /* Do not allow XPS on subordinate device directly */ 2864 num_tc = dev->num_tc; 2865 if (num_tc < 0) 2866 return -EINVAL; 2867 2868 /* If queue belongs to subordinate dev use its map */ 2869 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2870 2871 tc = netdev_txq_to_tc(dev, index); 2872 if (tc < 0) 2873 return -EINVAL; 2874 } 2875 2876 mutex_lock(&xps_map_mutex); 2877 2878 dev_maps = xmap_dereference(dev->xps_maps[type]); 2879 if (type == XPS_RXQS) { 2880 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2881 nr_ids = dev->num_rx_queues; 2882 } else { 2883 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2884 if (num_possible_cpus() > 1) 2885 online_mask = cpumask_bits(cpu_online_mask); 2886 nr_ids = nr_cpu_ids; 2887 } 2888 2889 if (maps_sz < L1_CACHE_BYTES) 2890 maps_sz = L1_CACHE_BYTES; 2891 2892 /* The old dev_maps could be larger or smaller than the one we're 2893 * setting up now, as dev->num_tc or nr_ids could have been updated in 2894 * between. We could try to be smart, but let's be safe instead and only 2895 * copy foreign traffic classes if the two map sizes match. 2896 */ 2897 if (dev_maps && 2898 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2899 copy = true; 2900 2901 /* allocate memory for queue storage */ 2902 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2903 j < nr_ids;) { 2904 if (!new_dev_maps) { 2905 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2906 if (!new_dev_maps) { 2907 mutex_unlock(&xps_map_mutex); 2908 return -ENOMEM; 2909 } 2910 2911 new_dev_maps->nr_ids = nr_ids; 2912 new_dev_maps->num_tc = num_tc; 2913 } 2914 2915 tci = j * num_tc + tc; 2916 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2917 2918 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2919 if (!map) 2920 goto error; 2921 2922 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2923 } 2924 2925 if (!new_dev_maps) 2926 goto out_no_new_maps; 2927 2928 if (!dev_maps) { 2929 /* Increment static keys at most once per type */ 2930 static_key_slow_inc_cpuslocked(&xps_needed); 2931 if (type == XPS_RXQS) 2932 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2933 } 2934 2935 for (j = 0; j < nr_ids; j++) { 2936 bool skip_tc = false; 2937 2938 tci = j * num_tc + tc; 2939 if (netif_attr_test_mask(j, mask, nr_ids) && 2940 netif_attr_test_online(j, online_mask, nr_ids)) { 2941 /* add tx-queue to CPU/rx-queue maps */ 2942 int pos = 0; 2943 2944 skip_tc = true; 2945 2946 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2947 while ((pos < map->len) && (map->queues[pos] != index)) 2948 pos++; 2949 2950 if (pos == map->len) 2951 map->queues[map->len++] = index; 2952 #ifdef CONFIG_NUMA 2953 if (type == XPS_CPUS) { 2954 if (numa_node_id == -2) 2955 numa_node_id = cpu_to_node(j); 2956 else if (numa_node_id != cpu_to_node(j)) 2957 numa_node_id = -1; 2958 } 2959 #endif 2960 } 2961 2962 if (copy) 2963 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2964 skip_tc); 2965 } 2966 2967 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2968 2969 /* Cleanup old maps */ 2970 if (!dev_maps) 2971 goto out_no_old_maps; 2972 2973 for (j = 0; j < dev_maps->nr_ids; j++) { 2974 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2975 map = xmap_dereference(dev_maps->attr_map[tci]); 2976 if (!map) 2977 continue; 2978 2979 if (copy) { 2980 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2981 if (map == new_map) 2982 continue; 2983 } 2984 2985 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2986 kfree_rcu(map, rcu); 2987 } 2988 } 2989 2990 old_dev_maps = dev_maps; 2991 2992 out_no_old_maps: 2993 dev_maps = new_dev_maps; 2994 active = true; 2995 2996 out_no_new_maps: 2997 if (type == XPS_CPUS) 2998 /* update Tx queue numa node */ 2999 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 3000 (numa_node_id >= 0) ? 3001 numa_node_id : NUMA_NO_NODE); 3002 3003 if (!dev_maps) 3004 goto out_no_maps; 3005 3006 /* removes tx-queue from unused CPUs/rx-queues */ 3007 for (j = 0; j < dev_maps->nr_ids; j++) { 3008 tci = j * dev_maps->num_tc; 3009 3010 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 3011 if (i == tc && 3012 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 3013 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 3014 continue; 3015 3016 active |= remove_xps_queue(dev_maps, 3017 copy ? old_dev_maps : NULL, 3018 tci, index); 3019 } 3020 } 3021 3022 if (old_dev_maps) 3023 kfree_rcu(old_dev_maps, rcu); 3024 3025 /* free map if not active */ 3026 if (!active) 3027 reset_xps_maps(dev, dev_maps, type); 3028 3029 out_no_maps: 3030 mutex_unlock(&xps_map_mutex); 3031 3032 return 0; 3033 error: 3034 /* remove any maps that we added */ 3035 for (j = 0; j < nr_ids; j++) { 3036 for (i = num_tc, tci = j * num_tc; i--; tci++) { 3037 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 3038 map = copy ? 3039 xmap_dereference(dev_maps->attr_map[tci]) : 3040 NULL; 3041 if (new_map && new_map != map) 3042 kfree(new_map); 3043 } 3044 } 3045 3046 mutex_unlock(&xps_map_mutex); 3047 3048 kfree(new_dev_maps); 3049 return -ENOMEM; 3050 } 3051 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 3052 3053 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 3054 u16 index) 3055 { 3056 int ret; 3057 3058 cpus_read_lock(); 3059 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 3060 cpus_read_unlock(); 3061 3062 return ret; 3063 } 3064 EXPORT_SYMBOL(netif_set_xps_queue); 3065 3066 #endif 3067 static void netdev_unbind_all_sb_channels(struct net_device *dev) 3068 { 3069 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3070 3071 /* Unbind any subordinate channels */ 3072 while (txq-- != &dev->_tx[0]) { 3073 if (txq->sb_dev) 3074 netdev_unbind_sb_channel(dev, txq->sb_dev); 3075 } 3076 } 3077 3078 void netdev_reset_tc(struct net_device *dev) 3079 { 3080 #ifdef CONFIG_XPS 3081 netif_reset_xps_queues_gt(dev, 0); 3082 #endif 3083 netdev_unbind_all_sb_channels(dev); 3084 3085 /* Reset TC configuration of device */ 3086 dev->num_tc = 0; 3087 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 3088 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 3089 } 3090 EXPORT_SYMBOL(netdev_reset_tc); 3091 3092 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 3093 { 3094 if (tc >= dev->num_tc) 3095 return -EINVAL; 3096 3097 #ifdef CONFIG_XPS 3098 netif_reset_xps_queues(dev, offset, count); 3099 #endif 3100 dev->tc_to_txq[tc].count = count; 3101 dev->tc_to_txq[tc].offset = offset; 3102 return 0; 3103 } 3104 EXPORT_SYMBOL(netdev_set_tc_queue); 3105 3106 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 3107 { 3108 if (num_tc > TC_MAX_QUEUE) 3109 return -EINVAL; 3110 3111 #ifdef CONFIG_XPS 3112 netif_reset_xps_queues_gt(dev, 0); 3113 #endif 3114 netdev_unbind_all_sb_channels(dev); 3115 3116 dev->num_tc = num_tc; 3117 return 0; 3118 } 3119 EXPORT_SYMBOL(netdev_set_num_tc); 3120 3121 void netdev_unbind_sb_channel(struct net_device *dev, 3122 struct net_device *sb_dev) 3123 { 3124 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3125 3126 #ifdef CONFIG_XPS 3127 netif_reset_xps_queues_gt(sb_dev, 0); 3128 #endif 3129 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 3130 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 3131 3132 while (txq-- != &dev->_tx[0]) { 3133 if (txq->sb_dev == sb_dev) 3134 txq->sb_dev = NULL; 3135 } 3136 } 3137 EXPORT_SYMBOL(netdev_unbind_sb_channel); 3138 3139 int netdev_bind_sb_channel_queue(struct net_device *dev, 3140 struct net_device *sb_dev, 3141 u8 tc, u16 count, u16 offset) 3142 { 3143 /* Make certain the sb_dev and dev are already configured */ 3144 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 3145 return -EINVAL; 3146 3147 /* We cannot hand out queues we don't have */ 3148 if ((offset + count) > dev->real_num_tx_queues) 3149 return -EINVAL; 3150 3151 /* Record the mapping */ 3152 sb_dev->tc_to_txq[tc].count = count; 3153 sb_dev->tc_to_txq[tc].offset = offset; 3154 3155 /* Provide a way for Tx queue to find the tc_to_txq map or 3156 * XPS map for itself. 3157 */ 3158 while (count--) 3159 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 3160 3161 return 0; 3162 } 3163 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 3164 3165 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 3166 { 3167 /* Do not use a multiqueue device to represent a subordinate channel */ 3168 if (netif_is_multiqueue(dev)) 3169 return -ENODEV; 3170 3171 /* We allow channels 1 - 32767 to be used for subordinate channels. 3172 * Channel 0 is meant to be "native" mode and used only to represent 3173 * the main root device. We allow writing 0 to reset the device back 3174 * to normal mode after being used as a subordinate channel. 3175 */ 3176 if (channel > S16_MAX) 3177 return -EINVAL; 3178 3179 dev->num_tc = -channel; 3180 3181 return 0; 3182 } 3183 EXPORT_SYMBOL(netdev_set_sb_channel); 3184 3185 /* 3186 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 3187 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 3188 */ 3189 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 3190 { 3191 bool disabling; 3192 int rc; 3193 3194 disabling = txq < dev->real_num_tx_queues; 3195 3196 if (txq < 1 || txq > dev->num_tx_queues) 3197 return -EINVAL; 3198 3199 if (dev->reg_state == NETREG_REGISTERED || 3200 dev->reg_state == NETREG_UNREGISTERING) { 3201 netdev_ops_assert_locked(dev); 3202 3203 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 3204 txq); 3205 if (rc) 3206 return rc; 3207 3208 if (dev->num_tc) 3209 netif_setup_tc(dev, txq); 3210 3211 net_shaper_set_real_num_tx_queues(dev, txq); 3212 3213 dev_qdisc_change_real_num_tx(dev, txq); 3214 3215 dev->real_num_tx_queues = txq; 3216 3217 if (disabling) { 3218 synchronize_net(); 3219 qdisc_reset_all_tx_gt(dev, txq); 3220 #ifdef CONFIG_XPS 3221 netif_reset_xps_queues_gt(dev, txq); 3222 #endif 3223 } 3224 } else { 3225 dev->real_num_tx_queues = txq; 3226 } 3227 3228 return 0; 3229 } 3230 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 3231 3232 /** 3233 * netif_set_real_num_rx_queues - set actual number of RX queues used 3234 * @dev: Network device 3235 * @rxq: Actual number of RX queues 3236 * 3237 * This must be called either with the rtnl_lock held or before 3238 * registration of the net device. Returns 0 on success, or a 3239 * negative error code. If called before registration, it always 3240 * succeeds. 3241 */ 3242 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 3243 { 3244 int rc; 3245 3246 if (rxq < 1 || rxq > dev->num_rx_queues) 3247 return -EINVAL; 3248 3249 if (dev->reg_state == NETREG_REGISTERED) { 3250 netdev_ops_assert_locked(dev); 3251 3252 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 3253 rxq); 3254 if (rc) 3255 return rc; 3256 } 3257 3258 dev->real_num_rx_queues = rxq; 3259 return 0; 3260 } 3261 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3262 3263 /** 3264 * netif_set_real_num_queues - set actual number of RX and TX queues used 3265 * @dev: Network device 3266 * @txq: Actual number of TX queues 3267 * @rxq: Actual number of RX queues 3268 * 3269 * Set the real number of both TX and RX queues. 3270 * Does nothing if the number of queues is already correct. 3271 */ 3272 int netif_set_real_num_queues(struct net_device *dev, 3273 unsigned int txq, unsigned int rxq) 3274 { 3275 unsigned int old_rxq = dev->real_num_rx_queues; 3276 int err; 3277 3278 if (txq < 1 || txq > dev->num_tx_queues || 3279 rxq < 1 || rxq > dev->num_rx_queues) 3280 return -EINVAL; 3281 3282 /* Start from increases, so the error path only does decreases - 3283 * decreases can't fail. 3284 */ 3285 if (rxq > dev->real_num_rx_queues) { 3286 err = netif_set_real_num_rx_queues(dev, rxq); 3287 if (err) 3288 return err; 3289 } 3290 if (txq > dev->real_num_tx_queues) { 3291 err = netif_set_real_num_tx_queues(dev, txq); 3292 if (err) 3293 goto undo_rx; 3294 } 3295 if (rxq < dev->real_num_rx_queues) 3296 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3297 if (txq < dev->real_num_tx_queues) 3298 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3299 3300 return 0; 3301 undo_rx: 3302 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3303 return err; 3304 } 3305 EXPORT_SYMBOL(netif_set_real_num_queues); 3306 3307 /** 3308 * netif_set_tso_max_size() - set the max size of TSO frames supported 3309 * @dev: netdev to update 3310 * @size: max skb->len of a TSO frame 3311 * 3312 * Set the limit on the size of TSO super-frames the device can handle. 3313 * Unless explicitly set the stack will assume the value of 3314 * %GSO_LEGACY_MAX_SIZE. 3315 */ 3316 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3317 { 3318 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3319 if (size < READ_ONCE(dev->gso_max_size)) 3320 netif_set_gso_max_size(dev, size); 3321 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3322 netif_set_gso_ipv4_max_size(dev, size); 3323 } 3324 EXPORT_SYMBOL(netif_set_tso_max_size); 3325 3326 /** 3327 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3328 * @dev: netdev to update 3329 * @segs: max number of TCP segments 3330 * 3331 * Set the limit on the number of TCP segments the device can generate from 3332 * a single TSO super-frame. 3333 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3334 */ 3335 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3336 { 3337 dev->tso_max_segs = segs; 3338 if (segs < READ_ONCE(dev->gso_max_segs)) 3339 netif_set_gso_max_segs(dev, segs); 3340 } 3341 EXPORT_SYMBOL(netif_set_tso_max_segs); 3342 3343 /** 3344 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3345 * @to: netdev to update 3346 * @from: netdev from which to copy the limits 3347 */ 3348 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3349 { 3350 netif_set_tso_max_size(to, from->tso_max_size); 3351 netif_set_tso_max_segs(to, from->tso_max_segs); 3352 } 3353 EXPORT_SYMBOL(netif_inherit_tso_max); 3354 3355 /** 3356 * netif_get_num_default_rss_queues - default number of RSS queues 3357 * 3358 * Default value is the number of physical cores if there are only 1 or 2, or 3359 * divided by 2 if there are more. 3360 */ 3361 int netif_get_num_default_rss_queues(void) 3362 { 3363 cpumask_var_t cpus; 3364 int cpu, count = 0; 3365 3366 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3367 return 1; 3368 3369 cpumask_copy(cpus, cpu_online_mask); 3370 for_each_cpu(cpu, cpus) { 3371 ++count; 3372 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3373 } 3374 free_cpumask_var(cpus); 3375 3376 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3377 } 3378 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3379 3380 static void __netif_reschedule(struct Qdisc *q) 3381 { 3382 struct softnet_data *sd; 3383 unsigned long flags; 3384 3385 local_irq_save(flags); 3386 sd = this_cpu_ptr(&softnet_data); 3387 q->next_sched = NULL; 3388 *sd->output_queue_tailp = q; 3389 sd->output_queue_tailp = &q->next_sched; 3390 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3391 local_irq_restore(flags); 3392 } 3393 3394 void __netif_schedule(struct Qdisc *q) 3395 { 3396 /* If q->defer_list is not empty, at least one thread is 3397 * in __dev_xmit_skb() before llist_del_all(&q->defer_list). 3398 * This thread will attempt to run the queue. 3399 */ 3400 if (!llist_empty(&q->defer_list)) 3401 return; 3402 3403 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3404 __netif_reschedule(q); 3405 } 3406 EXPORT_SYMBOL(__netif_schedule); 3407 3408 struct dev_kfree_skb_cb { 3409 enum skb_drop_reason reason; 3410 }; 3411 3412 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3413 { 3414 return (struct dev_kfree_skb_cb *)skb->cb; 3415 } 3416 3417 void netif_schedule_queue(struct netdev_queue *txq) 3418 { 3419 rcu_read_lock(); 3420 if (!netif_xmit_stopped(txq)) { 3421 struct Qdisc *q = rcu_dereference(txq->qdisc); 3422 3423 __netif_schedule(q); 3424 } 3425 rcu_read_unlock(); 3426 } 3427 EXPORT_SYMBOL(netif_schedule_queue); 3428 3429 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3430 { 3431 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3432 struct Qdisc *q; 3433 3434 rcu_read_lock(); 3435 q = rcu_dereference(dev_queue->qdisc); 3436 __netif_schedule(q); 3437 rcu_read_unlock(); 3438 } 3439 } 3440 EXPORT_SYMBOL(netif_tx_wake_queue); 3441 3442 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3443 { 3444 unsigned long flags; 3445 3446 if (unlikely(!skb)) 3447 return; 3448 3449 if (likely(refcount_read(&skb->users) == 1)) { 3450 smp_rmb(); 3451 refcount_set(&skb->users, 0); 3452 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3453 return; 3454 } 3455 get_kfree_skb_cb(skb)->reason = reason; 3456 local_irq_save(flags); 3457 skb->next = __this_cpu_read(softnet_data.completion_queue); 3458 __this_cpu_write(softnet_data.completion_queue, skb); 3459 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3460 local_irq_restore(flags); 3461 } 3462 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3463 3464 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3465 { 3466 if (in_hardirq() || irqs_disabled()) 3467 dev_kfree_skb_irq_reason(skb, reason); 3468 else 3469 kfree_skb_reason(skb, reason); 3470 } 3471 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3472 3473 3474 /** 3475 * netif_device_detach - mark device as removed 3476 * @dev: network device 3477 * 3478 * Mark device as removed from system and therefore no longer available. 3479 */ 3480 void netif_device_detach(struct net_device *dev) 3481 { 3482 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3483 netif_running(dev)) { 3484 netif_tx_stop_all_queues(dev); 3485 } 3486 } 3487 EXPORT_SYMBOL(netif_device_detach); 3488 3489 /** 3490 * netif_device_attach - mark device as attached 3491 * @dev: network device 3492 * 3493 * Mark device as attached from system and restart if needed. 3494 */ 3495 void netif_device_attach(struct net_device *dev) 3496 { 3497 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3498 netif_running(dev)) { 3499 netif_tx_wake_all_queues(dev); 3500 netdev_watchdog_up(dev); 3501 } 3502 } 3503 EXPORT_SYMBOL(netif_device_attach); 3504 3505 /* 3506 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3507 * to be used as a distribution range. 3508 */ 3509 static u16 skb_tx_hash(const struct net_device *dev, 3510 const struct net_device *sb_dev, 3511 struct sk_buff *skb) 3512 { 3513 u32 hash; 3514 u16 qoffset = 0; 3515 u16 qcount = dev->real_num_tx_queues; 3516 3517 if (dev->num_tc) { 3518 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3519 3520 qoffset = sb_dev->tc_to_txq[tc].offset; 3521 qcount = sb_dev->tc_to_txq[tc].count; 3522 if (unlikely(!qcount)) { 3523 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3524 sb_dev->name, qoffset, tc); 3525 qoffset = 0; 3526 qcount = dev->real_num_tx_queues; 3527 } 3528 } 3529 3530 if (skb_rx_queue_recorded(skb)) { 3531 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3532 hash = skb_get_rx_queue(skb); 3533 if (hash >= qoffset) 3534 hash -= qoffset; 3535 while (unlikely(hash >= qcount)) 3536 hash -= qcount; 3537 return hash + qoffset; 3538 } 3539 3540 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3541 } 3542 3543 void skb_warn_bad_offload(const struct sk_buff *skb) 3544 { 3545 static const netdev_features_t null_features; 3546 struct net_device *dev = skb->dev; 3547 const char *name = ""; 3548 3549 if (!net_ratelimit()) 3550 return; 3551 3552 if (dev) { 3553 if (dev->dev.parent) 3554 name = dev_driver_string(dev->dev.parent); 3555 else 3556 name = netdev_name(dev); 3557 } 3558 skb_dump(KERN_WARNING, skb, false); 3559 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3560 name, dev ? &dev->features : &null_features, 3561 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3562 } 3563 3564 /* 3565 * Invalidate hardware checksum when packet is to be mangled, and 3566 * complete checksum manually on outgoing path. 3567 */ 3568 int skb_checksum_help(struct sk_buff *skb) 3569 { 3570 __wsum csum; 3571 int ret = 0, offset; 3572 3573 if (skb->ip_summed == CHECKSUM_COMPLETE) 3574 goto out_set_summed; 3575 3576 if (unlikely(skb_is_gso(skb))) { 3577 skb_warn_bad_offload(skb); 3578 return -EINVAL; 3579 } 3580 3581 if (!skb_frags_readable(skb)) { 3582 return -EFAULT; 3583 } 3584 3585 /* Before computing a checksum, we should make sure no frag could 3586 * be modified by an external entity : checksum could be wrong. 3587 */ 3588 if (skb_has_shared_frag(skb)) { 3589 ret = __skb_linearize(skb); 3590 if (ret) 3591 goto out; 3592 } 3593 3594 offset = skb_checksum_start_offset(skb); 3595 ret = -EINVAL; 3596 if (unlikely(offset >= skb_headlen(skb))) { 3597 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3598 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3599 offset, skb_headlen(skb)); 3600 goto out; 3601 } 3602 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3603 3604 offset += skb->csum_offset; 3605 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3606 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3607 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3608 offset + sizeof(__sum16), skb_headlen(skb)); 3609 goto out; 3610 } 3611 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3612 if (ret) 3613 goto out; 3614 3615 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3616 out_set_summed: 3617 skb->ip_summed = CHECKSUM_NONE; 3618 out: 3619 return ret; 3620 } 3621 EXPORT_SYMBOL(skb_checksum_help); 3622 3623 #ifdef CONFIG_NET_CRC32C 3624 int skb_crc32c_csum_help(struct sk_buff *skb) 3625 { 3626 u32 crc; 3627 int ret = 0, offset, start; 3628 3629 if (skb->ip_summed != CHECKSUM_PARTIAL) 3630 goto out; 3631 3632 if (unlikely(skb_is_gso(skb))) 3633 goto out; 3634 3635 /* Before computing a checksum, we should make sure no frag could 3636 * be modified by an external entity : checksum could be wrong. 3637 */ 3638 if (unlikely(skb_has_shared_frag(skb))) { 3639 ret = __skb_linearize(skb); 3640 if (ret) 3641 goto out; 3642 } 3643 start = skb_checksum_start_offset(skb); 3644 offset = start + offsetof(struct sctphdr, checksum); 3645 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3646 ret = -EINVAL; 3647 goto out; 3648 } 3649 3650 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3651 if (ret) 3652 goto out; 3653 3654 crc = ~skb_crc32c(skb, start, skb->len - start, ~0); 3655 *(__le32 *)(skb->data + offset) = cpu_to_le32(crc); 3656 skb_reset_csum_not_inet(skb); 3657 out: 3658 return ret; 3659 } 3660 EXPORT_SYMBOL(skb_crc32c_csum_help); 3661 #endif /* CONFIG_NET_CRC32C */ 3662 3663 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3664 { 3665 __be16 type = skb->protocol; 3666 3667 /* Tunnel gso handlers can set protocol to ethernet. */ 3668 if (type == htons(ETH_P_TEB)) { 3669 struct ethhdr *eth; 3670 3671 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3672 return 0; 3673 3674 eth = (struct ethhdr *)skb->data; 3675 type = eth->h_proto; 3676 } 3677 3678 return vlan_get_protocol_and_depth(skb, type, depth); 3679 } 3680 3681 3682 /* Take action when hardware reception checksum errors are detected. */ 3683 #ifdef CONFIG_BUG 3684 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3685 { 3686 netdev_err(dev, "hw csum failure\n"); 3687 skb_dump(KERN_ERR, skb, true); 3688 dump_stack(); 3689 } 3690 3691 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3692 { 3693 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3694 } 3695 EXPORT_SYMBOL(netdev_rx_csum_fault); 3696 #endif 3697 3698 /* XXX: check that highmem exists at all on the given machine. */ 3699 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3700 { 3701 #ifdef CONFIG_HIGHMEM 3702 int i; 3703 3704 if (!(dev->features & NETIF_F_HIGHDMA)) { 3705 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3706 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3707 struct page *page = skb_frag_page(frag); 3708 3709 if (page && PageHighMem(page)) 3710 return 1; 3711 } 3712 } 3713 #endif 3714 return 0; 3715 } 3716 3717 /* If MPLS offload request, verify we are testing hardware MPLS features 3718 * instead of standard features for the netdev. 3719 */ 3720 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3721 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3722 netdev_features_t features, 3723 __be16 type) 3724 { 3725 if (eth_p_mpls(type)) 3726 features &= skb->dev->mpls_features; 3727 3728 return features; 3729 } 3730 #else 3731 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3732 netdev_features_t features, 3733 __be16 type) 3734 { 3735 return features; 3736 } 3737 #endif 3738 3739 static netdev_features_t harmonize_features(struct sk_buff *skb, 3740 netdev_features_t features) 3741 { 3742 __be16 type; 3743 3744 type = skb_network_protocol(skb, NULL); 3745 features = net_mpls_features(skb, features, type); 3746 3747 if (skb->ip_summed != CHECKSUM_NONE && 3748 !can_checksum_protocol(features, type)) { 3749 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3750 } 3751 if (illegal_highdma(skb->dev, skb)) 3752 features &= ~NETIF_F_SG; 3753 3754 return features; 3755 } 3756 3757 netdev_features_t passthru_features_check(struct sk_buff *skb, 3758 struct net_device *dev, 3759 netdev_features_t features) 3760 { 3761 return features; 3762 } 3763 EXPORT_SYMBOL(passthru_features_check); 3764 3765 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3766 struct net_device *dev, 3767 netdev_features_t features) 3768 { 3769 return vlan_features_check(skb, features); 3770 } 3771 3772 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3773 struct net_device *dev, 3774 netdev_features_t features) 3775 { 3776 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3777 3778 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3779 return features & ~NETIF_F_GSO_MASK; 3780 3781 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3782 return features & ~NETIF_F_GSO_MASK; 3783 3784 if (!skb_shinfo(skb)->gso_type) { 3785 skb_warn_bad_offload(skb); 3786 return features & ~NETIF_F_GSO_MASK; 3787 } 3788 3789 /* Support for GSO partial features requires software 3790 * intervention before we can actually process the packets 3791 * so we need to strip support for any partial features now 3792 * and we can pull them back in after we have partially 3793 * segmented the frame. 3794 */ 3795 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3796 features &= ~dev->gso_partial_features; 3797 3798 /* Make sure to clear the IPv4 ID mangling feature if the IPv4 header 3799 * has the potential to be fragmented so that TSO does not generate 3800 * segments with the same ID. For encapsulated packets, the ID mangling 3801 * feature is guaranteed not to use the same ID for the outer IPv4 3802 * headers of the generated segments if the headers have the potential 3803 * to be fragmented, so there is no need to clear the IPv4 ID mangling 3804 * feature (see the section about NETIF_F_TSO_MANGLEID in 3805 * segmentation-offloads.rst). 3806 */ 3807 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3808 struct iphdr *iph = skb->encapsulation ? 3809 inner_ip_hdr(skb) : ip_hdr(skb); 3810 3811 if (!(iph->frag_off & htons(IP_DF))) 3812 features &= ~dev->mangleid_features; 3813 } 3814 3815 /* NETIF_F_IPV6_CSUM does not support IPv6 extension headers, 3816 * so neither does TSO that depends on it. 3817 */ 3818 if (features & NETIF_F_IPV6_CSUM && 3819 (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6 || 3820 (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4 && 3821 vlan_get_protocol(skb) == htons(ETH_P_IPV6))) && 3822 skb_transport_header_was_set(skb) && 3823 skb_network_header_len(skb) != sizeof(struct ipv6hdr)) 3824 features &= ~(NETIF_F_IPV6_CSUM | NETIF_F_TSO6 | NETIF_F_GSO_UDP_L4); 3825 3826 return features; 3827 } 3828 3829 netdev_features_t netif_skb_features(struct sk_buff *skb) 3830 { 3831 struct net_device *dev = skb->dev; 3832 netdev_features_t features = dev->features; 3833 3834 if (skb_is_gso(skb)) 3835 features = gso_features_check(skb, dev, features); 3836 3837 /* If encapsulation offload request, verify we are testing 3838 * hardware encapsulation features instead of standard 3839 * features for the netdev 3840 */ 3841 if (skb->encapsulation) 3842 features &= dev->hw_enc_features; 3843 3844 if (skb_vlan_tagged(skb)) 3845 features = netdev_intersect_features(features, 3846 dev->vlan_features | 3847 NETIF_F_HW_VLAN_CTAG_TX | 3848 NETIF_F_HW_VLAN_STAG_TX); 3849 3850 if (dev->netdev_ops->ndo_features_check) 3851 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3852 features); 3853 else 3854 features &= dflt_features_check(skb, dev, features); 3855 3856 return harmonize_features(skb, features); 3857 } 3858 EXPORT_SYMBOL(netif_skb_features); 3859 3860 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3861 struct netdev_queue *txq, bool more) 3862 { 3863 unsigned int len; 3864 int rc; 3865 3866 if (dev_nit_active_rcu(dev)) 3867 dev_queue_xmit_nit(skb, dev); 3868 3869 len = skb->len; 3870 trace_net_dev_start_xmit(skb, dev); 3871 rc = netdev_start_xmit(skb, dev, txq, more); 3872 trace_net_dev_xmit(skb, rc, dev, len); 3873 3874 return rc; 3875 } 3876 3877 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3878 struct netdev_queue *txq, int *ret) 3879 { 3880 struct sk_buff *skb = first; 3881 int rc = NETDEV_TX_OK; 3882 3883 while (skb) { 3884 struct sk_buff *next = skb->next; 3885 3886 skb_mark_not_on_list(skb); 3887 rc = xmit_one(skb, dev, txq, next != NULL); 3888 if (unlikely(!dev_xmit_complete(rc))) { 3889 skb->next = next; 3890 goto out; 3891 } 3892 3893 skb = next; 3894 if (netif_tx_queue_stopped(txq) && skb) { 3895 rc = NETDEV_TX_BUSY; 3896 break; 3897 } 3898 } 3899 3900 out: 3901 *ret = rc; 3902 return skb; 3903 } 3904 3905 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3906 netdev_features_t features) 3907 { 3908 if (skb_vlan_tag_present(skb) && 3909 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3910 skb = __vlan_hwaccel_push_inside(skb); 3911 return skb; 3912 } 3913 3914 int skb_csum_hwoffload_help(struct sk_buff *skb, 3915 const netdev_features_t features) 3916 { 3917 if (unlikely(skb_csum_is_sctp(skb))) 3918 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3919 skb_crc32c_csum_help(skb); 3920 3921 if (features & NETIF_F_HW_CSUM) 3922 return 0; 3923 3924 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3925 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3926 skb_network_header_len(skb) != sizeof(struct ipv6hdr)) 3927 goto sw_checksum; 3928 3929 switch (skb->csum_offset) { 3930 case offsetof(struct tcphdr, check): 3931 case offsetof(struct udphdr, check): 3932 return 0; 3933 } 3934 } 3935 3936 sw_checksum: 3937 return skb_checksum_help(skb); 3938 } 3939 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3940 3941 /* Checks if this SKB belongs to an HW offloaded socket 3942 * and whether any SW fallbacks are required based on dev. 3943 * Check decrypted mark in case skb_orphan() cleared socket. 3944 */ 3945 static struct sk_buff *sk_validate_xmit_skb(struct sk_buff *skb, 3946 struct net_device *dev) 3947 { 3948 #ifdef CONFIG_SOCK_VALIDATE_XMIT 3949 struct sk_buff *(*sk_validate)(struct sock *sk, struct net_device *dev, 3950 struct sk_buff *skb); 3951 struct sock *sk = skb->sk; 3952 3953 sk_validate = NULL; 3954 if (sk) { 3955 if (sk_fullsock(sk)) 3956 sk_validate = sk->sk_validate_xmit_skb; 3957 else if (sk_is_inet(sk) && sk->sk_state == TCP_TIME_WAIT) 3958 sk_validate = inet_twsk(sk)->tw_validate_xmit_skb; 3959 } 3960 3961 if (sk_validate) { 3962 skb = sk_validate(sk, dev, skb); 3963 } else if (unlikely(skb_is_decrypted(skb))) { 3964 pr_warn_ratelimited("unencrypted skb with no associated socket - dropping\n"); 3965 kfree_skb(skb); 3966 skb = NULL; 3967 } 3968 #endif 3969 3970 return skb; 3971 } 3972 3973 static struct sk_buff *validate_xmit_unreadable_skb(struct sk_buff *skb, 3974 struct net_device *dev) 3975 { 3976 struct skb_shared_info *shinfo; 3977 struct net_iov *niov; 3978 3979 if (likely(skb_frags_readable(skb))) 3980 goto out; 3981 3982 if (!dev->netmem_tx) 3983 goto out_free; 3984 3985 shinfo = skb_shinfo(skb); 3986 3987 if (shinfo->nr_frags > 0) { 3988 niov = netmem_to_net_iov(skb_frag_netmem(&shinfo->frags[0])); 3989 if (net_is_devmem_iov(niov) && 3990 net_devmem_iov_binding(niov)->dev != dev) 3991 goto out_free; 3992 } 3993 3994 out: 3995 return skb; 3996 3997 out_free: 3998 kfree_skb(skb); 3999 return NULL; 4000 } 4001 4002 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 4003 { 4004 netdev_features_t features; 4005 4006 skb = validate_xmit_unreadable_skb(skb, dev); 4007 if (unlikely(!skb)) 4008 goto out_null; 4009 4010 features = netif_skb_features(skb); 4011 skb = validate_xmit_vlan(skb, features); 4012 if (unlikely(!skb)) 4013 goto out_null; 4014 4015 skb = sk_validate_xmit_skb(skb, dev); 4016 if (unlikely(!skb)) 4017 goto out_null; 4018 4019 if (netif_needs_gso(skb, features)) { 4020 struct sk_buff *segs; 4021 4022 segs = skb_gso_segment(skb, features); 4023 if (IS_ERR(segs)) { 4024 goto out_kfree_skb; 4025 } else if (segs) { 4026 consume_skb(skb); 4027 skb = segs; 4028 } 4029 } else { 4030 if (skb_needs_linearize(skb, features) && 4031 __skb_linearize(skb)) 4032 goto out_kfree_skb; 4033 4034 /* If packet is not checksummed and device does not 4035 * support checksumming for this protocol, complete 4036 * checksumming here. 4037 */ 4038 if (skb->ip_summed == CHECKSUM_PARTIAL) { 4039 if (skb->encapsulation) 4040 skb_set_inner_transport_header(skb, 4041 skb_checksum_start_offset(skb)); 4042 else 4043 skb_set_transport_header(skb, 4044 skb_checksum_start_offset(skb)); 4045 if (skb_csum_hwoffload_help(skb, features)) 4046 goto out_kfree_skb; 4047 } 4048 } 4049 4050 skb = validate_xmit_xfrm(skb, features, again); 4051 4052 return skb; 4053 4054 out_kfree_skb: 4055 kfree_skb(skb); 4056 out_null: 4057 dev_core_stats_tx_dropped_inc(dev); 4058 return NULL; 4059 } 4060 4061 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 4062 { 4063 struct sk_buff *next, *head = NULL, *tail; 4064 4065 for (; skb != NULL; skb = next) { 4066 next = skb->next; 4067 skb_mark_not_on_list(skb); 4068 4069 /* in case skb won't be segmented, point to itself */ 4070 skb->prev = skb; 4071 4072 skb = validate_xmit_skb(skb, dev, again); 4073 if (!skb) 4074 continue; 4075 4076 if (!head) 4077 head = skb; 4078 else 4079 tail->next = skb; 4080 /* If skb was segmented, skb->prev points to 4081 * the last segment. If not, it still contains skb. 4082 */ 4083 tail = skb->prev; 4084 } 4085 return head; 4086 } 4087 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 4088 4089 static void qdisc_pkt_len_segs_init(struct sk_buff *skb) 4090 { 4091 struct skb_shared_info *shinfo = skb_shinfo(skb); 4092 u16 gso_segs; 4093 4094 qdisc_skb_cb(skb)->pkt_len = skb->len; 4095 if (!shinfo->gso_size) { 4096 qdisc_skb_cb(skb)->pkt_segs = 1; 4097 return; 4098 } 4099 4100 qdisc_skb_cb(skb)->pkt_segs = gso_segs = shinfo->gso_segs; 4101 4102 /* To get more precise estimation of bytes sent on wire, 4103 * we add to pkt_len the headers size of all segments 4104 */ 4105 if (skb_transport_header_was_set(skb)) { 4106 unsigned int hdr_len; 4107 4108 /* mac layer + network layer */ 4109 if (!skb->encapsulation) 4110 hdr_len = skb_transport_offset(skb); 4111 else 4112 hdr_len = skb_inner_transport_offset(skb); 4113 4114 /* + transport layer */ 4115 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 4116 const struct tcphdr *th; 4117 struct tcphdr _tcphdr; 4118 4119 th = skb_header_pointer(skb, hdr_len, 4120 sizeof(_tcphdr), &_tcphdr); 4121 if (likely(th)) 4122 hdr_len += __tcp_hdrlen(th); 4123 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 4124 struct udphdr _udphdr; 4125 4126 if (skb_header_pointer(skb, hdr_len, 4127 sizeof(_udphdr), &_udphdr)) 4128 hdr_len += sizeof(struct udphdr); 4129 } 4130 4131 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 4132 int payload = skb->len - hdr_len; 4133 4134 /* Malicious packet. */ 4135 if (payload <= 0) 4136 return; 4137 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 4138 shinfo->gso_segs = gso_segs; 4139 qdisc_skb_cb(skb)->pkt_segs = gso_segs; 4140 } 4141 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 4142 } 4143 } 4144 4145 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 4146 struct sk_buff **to_free, 4147 struct netdev_queue *txq) 4148 { 4149 int rc; 4150 4151 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 4152 if (rc == NET_XMIT_SUCCESS) 4153 trace_qdisc_enqueue(q, txq, skb); 4154 return rc; 4155 } 4156 4157 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 4158 struct net_device *dev, 4159 struct netdev_queue *txq) 4160 { 4161 struct sk_buff *next, *to_free = NULL, *to_free2 = NULL; 4162 spinlock_t *root_lock = qdisc_lock(q); 4163 struct llist_node *ll_list, *first_n; 4164 unsigned long defer_count = 0; 4165 int rc; 4166 4167 qdisc_calculate_pkt_len(skb, q); 4168 4169 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 4170 4171 if (q->flags & TCQ_F_NOLOCK) { 4172 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 4173 qdisc_run_begin(q)) { 4174 /* Retest nolock_qdisc_is_empty() within the protection 4175 * of q->seqlock to protect from racing with requeuing. 4176 */ 4177 if (unlikely(!nolock_qdisc_is_empty(q))) { 4178 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4179 __qdisc_run(q); 4180 to_free2 = qdisc_run_end(q); 4181 4182 goto free_skbs; 4183 } 4184 4185 qdisc_bstats_cpu_update(q, skb); 4186 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4187 !nolock_qdisc_is_empty(q)) 4188 __qdisc_run(q); 4189 4190 to_free2 = qdisc_run_end(q); 4191 rc = NET_XMIT_SUCCESS; 4192 goto free_skbs; 4193 } 4194 4195 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4196 to_free2 = qdisc_run(q); 4197 goto free_skbs; 4198 } 4199 4200 /* Open code llist_add(&skb->ll_node, &q->defer_list) + queue limit. 4201 * In the try_cmpxchg() loop, we want to increment q->defer_count 4202 * at most once to limit the number of skbs in defer_list. 4203 * We perform the defer_count increment only if the list is not empty, 4204 * because some arches have slow atomic_long_inc_return(). 4205 */ 4206 first_n = READ_ONCE(q->defer_list.first); 4207 do { 4208 if (first_n && !defer_count) { 4209 defer_count = atomic_long_inc_return(&q->defer_count); 4210 if (unlikely(defer_count > READ_ONCE(net_hotdata.qdisc_max_burst))) { 4211 kfree_skb_reason(skb, SKB_DROP_REASON_QDISC_BURST_DROP); 4212 return NET_XMIT_DROP; 4213 } 4214 } 4215 skb->ll_node.next = first_n; 4216 } while (!try_cmpxchg(&q->defer_list.first, &first_n, &skb->ll_node)); 4217 4218 /* If defer_list was not empty, we know the cpu which queued 4219 * the first skb will process the whole list for us. 4220 */ 4221 if (first_n) 4222 return NET_XMIT_SUCCESS; 4223 4224 spin_lock(root_lock); 4225 4226 ll_list = llist_del_all(&q->defer_list); 4227 /* There is a small race because we clear defer_count not atomically 4228 * with the prior llist_del_all(). This means defer_list could grow 4229 * over qdisc_max_burst. 4230 */ 4231 atomic_long_set(&q->defer_count, 0); 4232 4233 ll_list = llist_reverse_order(ll_list); 4234 4235 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4236 llist_for_each_entry_safe(skb, next, ll_list, ll_node) 4237 __qdisc_drop(skb, &to_free); 4238 rc = NET_XMIT_DROP; 4239 goto unlock; 4240 } 4241 if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4242 !llist_next(ll_list) && qdisc_run_begin(q)) { 4243 /* 4244 * This is a work-conserving queue; there are no old skbs 4245 * waiting to be sent out; and the qdisc is not running - 4246 * xmit the skb directly. 4247 */ 4248 4249 DEBUG_NET_WARN_ON_ONCE(skb != llist_entry(ll_list, 4250 struct sk_buff, 4251 ll_node)); 4252 qdisc_bstats_update(q, skb); 4253 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) 4254 __qdisc_run(q); 4255 to_free2 = qdisc_run_end(q); 4256 rc = NET_XMIT_SUCCESS; 4257 } else { 4258 int count = 0; 4259 4260 llist_for_each_entry_safe(skb, next, ll_list, ll_node) { 4261 if (next) { 4262 prefetch(next); 4263 prefetch(&next->priority); 4264 skb_mark_not_on_list(skb); 4265 } 4266 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4267 count++; 4268 } 4269 to_free2 = qdisc_run(q); 4270 if (count != 1) 4271 rc = NET_XMIT_SUCCESS; 4272 } 4273 unlock: 4274 spin_unlock(root_lock); 4275 4276 free_skbs: 4277 tcf_kfree_skb_list(to_free); 4278 tcf_kfree_skb_list(to_free2); 4279 return rc; 4280 } 4281 4282 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4283 static void skb_update_prio(struct sk_buff *skb) 4284 { 4285 const struct netprio_map *map; 4286 const struct sock *sk; 4287 unsigned int prioidx; 4288 4289 if (skb->priority) 4290 return; 4291 map = rcu_dereference_bh(skb->dev->priomap); 4292 if (!map) 4293 return; 4294 sk = skb_to_full_sk(skb); 4295 if (!sk) 4296 return; 4297 4298 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4299 4300 if (prioidx < map->priomap_len) 4301 skb->priority = map->priomap[prioidx]; 4302 } 4303 #else 4304 #define skb_update_prio(skb) 4305 #endif 4306 4307 /** 4308 * dev_loopback_xmit - loop back @skb 4309 * @net: network namespace this loopback is happening in 4310 * @sk: sk needed to be a netfilter okfn 4311 * @skb: buffer to transmit 4312 */ 4313 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4314 { 4315 skb_reset_mac_header(skb); 4316 __skb_pull(skb, skb_network_offset(skb)); 4317 skb->pkt_type = PACKET_LOOPBACK; 4318 if (skb->ip_summed == CHECKSUM_NONE) 4319 skb->ip_summed = CHECKSUM_UNNECESSARY; 4320 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4321 skb_dst_force(skb); 4322 netif_rx(skb); 4323 return 0; 4324 } 4325 EXPORT_SYMBOL(dev_loopback_xmit); 4326 4327 #ifdef CONFIG_NET_EGRESS 4328 static struct netdev_queue * 4329 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4330 { 4331 int qm = skb_get_queue_mapping(skb); 4332 4333 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4334 } 4335 4336 #ifndef CONFIG_PREEMPT_RT 4337 static bool netdev_xmit_txqueue_skipped(void) 4338 { 4339 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4340 } 4341 4342 void netdev_xmit_skip_txqueue(bool skip) 4343 { 4344 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4345 } 4346 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4347 4348 #else 4349 static bool netdev_xmit_txqueue_skipped(void) 4350 { 4351 return current->net_xmit.skip_txqueue; 4352 } 4353 4354 void netdev_xmit_skip_txqueue(bool skip) 4355 { 4356 current->net_xmit.skip_txqueue = skip; 4357 } 4358 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4359 #endif 4360 #endif /* CONFIG_NET_EGRESS */ 4361 4362 #ifdef CONFIG_NET_XGRESS 4363 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4364 enum skb_drop_reason *drop_reason) 4365 { 4366 int ret = TC_ACT_UNSPEC; 4367 #ifdef CONFIG_NET_CLS_ACT 4368 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4369 struct tcf_result res; 4370 4371 if (!miniq) 4372 return ret; 4373 4374 /* Global bypass */ 4375 if (!static_branch_likely(&tcf_sw_enabled_key)) 4376 return ret; 4377 4378 /* Block-wise bypass */ 4379 if (tcf_block_bypass_sw(miniq->block)) 4380 return ret; 4381 4382 tc_skb_cb(skb)->mru = 0; 4383 qdisc_skb_cb(skb)->post_ct = false; 4384 tcf_set_drop_reason(skb, *drop_reason); 4385 4386 mini_qdisc_bstats_cpu_update(miniq, skb); 4387 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4388 /* Only tcf related quirks below. */ 4389 switch (ret) { 4390 case TC_ACT_SHOT: 4391 *drop_reason = tcf_get_drop_reason(skb); 4392 mini_qdisc_qstats_cpu_drop(miniq); 4393 break; 4394 case TC_ACT_OK: 4395 case TC_ACT_RECLASSIFY: 4396 skb->tc_index = TC_H_MIN(res.classid); 4397 break; 4398 } 4399 #endif /* CONFIG_NET_CLS_ACT */ 4400 return ret; 4401 } 4402 4403 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4404 4405 void tcx_inc(void) 4406 { 4407 static_branch_inc(&tcx_needed_key); 4408 } 4409 4410 void tcx_dec(void) 4411 { 4412 static_branch_dec(&tcx_needed_key); 4413 } 4414 4415 static __always_inline enum tcx_action_base 4416 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4417 const bool needs_mac) 4418 { 4419 const struct bpf_mprog_fp *fp; 4420 const struct bpf_prog *prog; 4421 int ret = TCX_NEXT; 4422 4423 if (needs_mac) 4424 __skb_push(skb, skb->mac_len); 4425 bpf_mprog_foreach_prog(entry, fp, prog) { 4426 bpf_compute_data_pointers(skb); 4427 ret = bpf_prog_run(prog, skb); 4428 if (ret != TCX_NEXT) 4429 break; 4430 } 4431 if (needs_mac) 4432 __skb_pull(skb, skb->mac_len); 4433 return tcx_action_code(skb, ret); 4434 } 4435 4436 static __always_inline struct sk_buff * 4437 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4438 struct net_device *orig_dev, bool *another) 4439 { 4440 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4441 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4442 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4443 int sch_ret; 4444 4445 if (!entry) 4446 return skb; 4447 4448 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4449 if (unlikely(*pt_prev)) { 4450 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4451 *pt_prev = NULL; 4452 } 4453 4454 qdisc_pkt_len_segs_init(skb); 4455 tcx_set_ingress(skb, true); 4456 4457 if (static_branch_unlikely(&tcx_needed_key)) { 4458 sch_ret = tcx_run(entry, skb, true); 4459 if (sch_ret != TC_ACT_UNSPEC) 4460 goto ingress_verdict; 4461 } 4462 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4463 ingress_verdict: 4464 switch (sch_ret) { 4465 case TC_ACT_REDIRECT: 4466 /* skb_mac_header check was done by BPF, so we can safely 4467 * push the L2 header back before redirecting to another 4468 * netdev. 4469 */ 4470 __skb_push(skb, skb->mac_len); 4471 if (skb_do_redirect(skb) == -EAGAIN) { 4472 __skb_pull(skb, skb->mac_len); 4473 *another = true; 4474 break; 4475 } 4476 *ret = NET_RX_SUCCESS; 4477 bpf_net_ctx_clear(bpf_net_ctx); 4478 return NULL; 4479 case TC_ACT_SHOT: 4480 kfree_skb_reason(skb, drop_reason); 4481 *ret = NET_RX_DROP; 4482 bpf_net_ctx_clear(bpf_net_ctx); 4483 return NULL; 4484 /* used by tc_run */ 4485 case TC_ACT_STOLEN: 4486 case TC_ACT_QUEUED: 4487 case TC_ACT_TRAP: 4488 consume_skb(skb); 4489 fallthrough; 4490 case TC_ACT_CONSUMED: 4491 *ret = NET_RX_SUCCESS; 4492 bpf_net_ctx_clear(bpf_net_ctx); 4493 return NULL; 4494 } 4495 bpf_net_ctx_clear(bpf_net_ctx); 4496 4497 return skb; 4498 } 4499 4500 static __always_inline struct sk_buff * 4501 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4502 { 4503 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4504 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4505 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4506 int sch_ret; 4507 4508 if (!entry) 4509 return skb; 4510 4511 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4512 4513 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4514 * already set by the caller. 4515 */ 4516 if (static_branch_unlikely(&tcx_needed_key)) { 4517 sch_ret = tcx_run(entry, skb, false); 4518 if (sch_ret != TC_ACT_UNSPEC) 4519 goto egress_verdict; 4520 } 4521 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4522 egress_verdict: 4523 switch (sch_ret) { 4524 case TC_ACT_REDIRECT: 4525 /* No need to push/pop skb's mac_header here on egress! */ 4526 skb_do_redirect(skb); 4527 *ret = NET_XMIT_SUCCESS; 4528 bpf_net_ctx_clear(bpf_net_ctx); 4529 return NULL; 4530 case TC_ACT_SHOT: 4531 kfree_skb_reason(skb, drop_reason); 4532 *ret = NET_XMIT_DROP; 4533 bpf_net_ctx_clear(bpf_net_ctx); 4534 return NULL; 4535 /* used by tc_run */ 4536 case TC_ACT_STOLEN: 4537 case TC_ACT_QUEUED: 4538 case TC_ACT_TRAP: 4539 consume_skb(skb); 4540 fallthrough; 4541 case TC_ACT_CONSUMED: 4542 *ret = NET_XMIT_SUCCESS; 4543 bpf_net_ctx_clear(bpf_net_ctx); 4544 return NULL; 4545 } 4546 bpf_net_ctx_clear(bpf_net_ctx); 4547 4548 return skb; 4549 } 4550 #else 4551 static __always_inline struct sk_buff * 4552 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4553 struct net_device *orig_dev, bool *another) 4554 { 4555 return skb; 4556 } 4557 4558 static __always_inline struct sk_buff * 4559 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4560 { 4561 return skb; 4562 } 4563 #endif /* CONFIG_NET_XGRESS */ 4564 4565 #ifdef CONFIG_XPS 4566 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4567 struct xps_dev_maps *dev_maps, unsigned int tci) 4568 { 4569 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4570 struct xps_map *map; 4571 int queue_index = -1; 4572 4573 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4574 return queue_index; 4575 4576 tci *= dev_maps->num_tc; 4577 tci += tc; 4578 4579 map = rcu_dereference(dev_maps->attr_map[tci]); 4580 if (map) { 4581 if (map->len == 1) 4582 queue_index = map->queues[0]; 4583 else 4584 queue_index = map->queues[reciprocal_scale( 4585 skb_get_hash(skb), map->len)]; 4586 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4587 queue_index = -1; 4588 } 4589 return queue_index; 4590 } 4591 #endif 4592 4593 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4594 struct sk_buff *skb) 4595 { 4596 #ifdef CONFIG_XPS 4597 struct xps_dev_maps *dev_maps; 4598 struct sock *sk = skb->sk; 4599 int queue_index = -1; 4600 4601 if (!static_key_false(&xps_needed)) 4602 return -1; 4603 4604 rcu_read_lock(); 4605 if (!static_key_false(&xps_rxqs_needed)) 4606 goto get_cpus_map; 4607 4608 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4609 if (dev_maps) { 4610 int tci = sk_rx_queue_get(sk); 4611 4612 if (tci >= 0) 4613 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4614 tci); 4615 } 4616 4617 get_cpus_map: 4618 if (queue_index < 0) { 4619 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4620 if (dev_maps) { 4621 unsigned int tci = skb->sender_cpu - 1; 4622 4623 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4624 tci); 4625 } 4626 } 4627 rcu_read_unlock(); 4628 4629 return queue_index; 4630 #else 4631 return -1; 4632 #endif 4633 } 4634 4635 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4636 struct net_device *sb_dev) 4637 { 4638 return 0; 4639 } 4640 EXPORT_SYMBOL(dev_pick_tx_zero); 4641 4642 int sk_tx_queue_get(const struct sock *sk) 4643 { 4644 int resel, val; 4645 4646 if (!sk) 4647 return -1; 4648 /* Paired with WRITE_ONCE() in sk_tx_queue_clear() 4649 * and sk_tx_queue_set(). 4650 */ 4651 val = READ_ONCE(sk->sk_tx_queue_mapping); 4652 4653 if (val == NO_QUEUE_MAPPING) 4654 return -1; 4655 4656 if (!sk_fullsock(sk)) 4657 return val; 4658 4659 resel = READ_ONCE(sock_net(sk)->core.sysctl_txq_reselection); 4660 if (resel && time_is_before_jiffies( 4661 READ_ONCE(sk->sk_tx_queue_mapping_jiffies) + resel)) 4662 return -1; 4663 4664 return val; 4665 } 4666 EXPORT_SYMBOL(sk_tx_queue_get); 4667 4668 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4669 struct net_device *sb_dev) 4670 { 4671 struct sock *sk = skb->sk; 4672 int queue_index = sk_tx_queue_get(sk); 4673 4674 sb_dev = sb_dev ? : dev; 4675 4676 if (queue_index < 0 || skb->ooo_okay || 4677 queue_index >= dev->real_num_tx_queues) { 4678 int new_index = get_xps_queue(dev, sb_dev, skb); 4679 4680 if (new_index < 0) 4681 new_index = skb_tx_hash(dev, sb_dev, skb); 4682 4683 if (sk && sk_fullsock(sk) && 4684 rcu_access_pointer(sk->sk_dst_cache)) 4685 sk_tx_queue_set(sk, new_index); 4686 4687 queue_index = new_index; 4688 } 4689 4690 return queue_index; 4691 } 4692 EXPORT_SYMBOL(netdev_pick_tx); 4693 4694 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4695 struct sk_buff *skb, 4696 struct net_device *sb_dev) 4697 { 4698 int queue_index = 0; 4699 4700 #ifdef CONFIG_XPS 4701 u32 sender_cpu = skb->sender_cpu - 1; 4702 4703 if (sender_cpu >= (u32)NR_CPUS) 4704 skb->sender_cpu = raw_smp_processor_id() + 1; 4705 #endif 4706 4707 if (dev->real_num_tx_queues != 1) { 4708 const struct net_device_ops *ops = dev->netdev_ops; 4709 4710 if (ops->ndo_select_queue) 4711 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4712 else 4713 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4714 4715 queue_index = netdev_cap_txqueue(dev, queue_index); 4716 } 4717 4718 skb_set_queue_mapping(skb, queue_index); 4719 return netdev_get_tx_queue(dev, queue_index); 4720 } 4721 4722 /** 4723 * __dev_queue_xmit() - transmit a buffer 4724 * @skb: buffer to transmit 4725 * @sb_dev: suboordinate device used for L2 forwarding offload 4726 * 4727 * Queue a buffer for transmission to a network device. The caller must 4728 * have set the device and priority and built the buffer before calling 4729 * this function. The function can be called from an interrupt. 4730 * 4731 * When calling this method, interrupts MUST be enabled. This is because 4732 * the BH enable code must have IRQs enabled so that it will not deadlock. 4733 * 4734 * Regardless of the return value, the skb is consumed, so it is currently 4735 * difficult to retry a send to this method. (You can bump the ref count 4736 * before sending to hold a reference for retry if you are careful.) 4737 * 4738 * Return: 4739 * * 0 - buffer successfully transmitted 4740 * * positive qdisc return code - NET_XMIT_DROP etc. 4741 * * negative errno - other errors 4742 */ 4743 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4744 { 4745 struct net_device *dev = skb->dev; 4746 struct netdev_queue *txq = NULL; 4747 struct Qdisc *q; 4748 int rc = -ENOMEM; 4749 bool again = false; 4750 4751 skb_reset_mac_header(skb); 4752 skb_assert_len(skb); 4753 4754 if (unlikely(skb_shinfo(skb)->tx_flags & 4755 (SKBTX_SCHED_TSTAMP | SKBTX_BPF))) 4756 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4757 4758 /* Disable soft irqs for various locks below. Also 4759 * stops preemption for RCU. 4760 */ 4761 rcu_read_lock_bh(); 4762 4763 skb_update_prio(skb); 4764 4765 qdisc_pkt_len_segs_init(skb); 4766 tcx_set_ingress(skb, false); 4767 #ifdef CONFIG_NET_EGRESS 4768 if (static_branch_unlikely(&egress_needed_key)) { 4769 if (nf_hook_egress_active()) { 4770 skb = nf_hook_egress(skb, &rc, dev); 4771 if (!skb) 4772 goto out; 4773 } 4774 4775 netdev_xmit_skip_txqueue(false); 4776 4777 nf_skip_egress(skb, true); 4778 skb = sch_handle_egress(skb, &rc, dev); 4779 if (!skb) 4780 goto out; 4781 nf_skip_egress(skb, false); 4782 4783 if (netdev_xmit_txqueue_skipped()) 4784 txq = netdev_tx_queue_mapping(dev, skb); 4785 } 4786 #endif 4787 /* If device/qdisc don't need skb->dst, release it right now while 4788 * its hot in this cpu cache. 4789 */ 4790 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4791 skb_dst_drop(skb); 4792 else 4793 skb_dst_force(skb); 4794 4795 if (!txq) 4796 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4797 4798 q = rcu_dereference_bh(txq->qdisc); 4799 4800 trace_net_dev_queue(skb); 4801 if (q->enqueue) { 4802 rc = __dev_xmit_skb(skb, q, dev, txq); 4803 goto out; 4804 } 4805 4806 /* The device has no queue. Common case for software devices: 4807 * loopback, all the sorts of tunnels... 4808 4809 * Really, it is unlikely that netif_tx_lock protection is necessary 4810 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4811 * counters.) 4812 * However, it is possible, that they rely on protection 4813 * made by us here. 4814 4815 * Check this and shot the lock. It is not prone from deadlocks. 4816 *Either shot noqueue qdisc, it is even simpler 8) 4817 */ 4818 if (dev->flags & IFF_UP) { 4819 int cpu = smp_processor_id(); /* ok because BHs are off */ 4820 4821 /* Other cpus might concurrently change txq->xmit_lock_owner 4822 * to -1 or to their cpu id, but not to our id. 4823 */ 4824 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4825 if (dev_xmit_recursion()) 4826 goto recursion_alert; 4827 4828 skb = validate_xmit_skb(skb, dev, &again); 4829 if (!skb) 4830 goto out; 4831 4832 HARD_TX_LOCK(dev, txq, cpu); 4833 4834 if (!netif_xmit_stopped(txq)) { 4835 dev_xmit_recursion_inc(); 4836 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4837 dev_xmit_recursion_dec(); 4838 if (dev_xmit_complete(rc)) { 4839 HARD_TX_UNLOCK(dev, txq); 4840 goto out; 4841 } 4842 } 4843 HARD_TX_UNLOCK(dev, txq); 4844 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4845 dev->name); 4846 } else { 4847 /* Recursion is detected! It is possible, 4848 * unfortunately 4849 */ 4850 recursion_alert: 4851 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4852 dev->name); 4853 } 4854 } 4855 4856 rc = -ENETDOWN; 4857 rcu_read_unlock_bh(); 4858 4859 dev_core_stats_tx_dropped_inc(dev); 4860 kfree_skb_list(skb); 4861 return rc; 4862 out: 4863 rcu_read_unlock_bh(); 4864 return rc; 4865 } 4866 EXPORT_SYMBOL(__dev_queue_xmit); 4867 4868 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4869 { 4870 struct net_device *dev = skb->dev; 4871 struct sk_buff *orig_skb = skb; 4872 struct netdev_queue *txq; 4873 int ret = NETDEV_TX_BUSY; 4874 bool again = false; 4875 4876 if (unlikely(!netif_running(dev) || 4877 !netif_carrier_ok(dev))) 4878 goto drop; 4879 4880 skb = validate_xmit_skb_list(skb, dev, &again); 4881 if (skb != orig_skb) 4882 goto drop; 4883 4884 skb_set_queue_mapping(skb, queue_id); 4885 txq = skb_get_tx_queue(dev, skb); 4886 4887 local_bh_disable(); 4888 4889 dev_xmit_recursion_inc(); 4890 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4891 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4892 ret = netdev_start_xmit(skb, dev, txq, false); 4893 HARD_TX_UNLOCK(dev, txq); 4894 dev_xmit_recursion_dec(); 4895 4896 local_bh_enable(); 4897 return ret; 4898 drop: 4899 dev_core_stats_tx_dropped_inc(dev); 4900 kfree_skb_list(skb); 4901 return NET_XMIT_DROP; 4902 } 4903 EXPORT_SYMBOL(__dev_direct_xmit); 4904 4905 /************************************************************************* 4906 * Receiver routines 4907 *************************************************************************/ 4908 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4909 4910 int weight_p __read_mostly = 64; /* old backlog weight */ 4911 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4912 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4913 4914 /* Called with irq disabled */ 4915 static inline void ____napi_schedule(struct softnet_data *sd, 4916 struct napi_struct *napi) 4917 { 4918 struct task_struct *thread; 4919 4920 lockdep_assert_irqs_disabled(); 4921 4922 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4923 /* Paired with smp_mb__before_atomic() in 4924 * napi_enable()/netif_set_threaded(). 4925 * Use READ_ONCE() to guarantee a complete 4926 * read on napi->thread. Only call 4927 * wake_up_process() when it's not NULL. 4928 */ 4929 thread = READ_ONCE(napi->thread); 4930 if (thread) { 4931 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4932 goto use_local_napi; 4933 4934 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4935 wake_up_process(thread); 4936 return; 4937 } 4938 } 4939 4940 use_local_napi: 4941 DEBUG_NET_WARN_ON_ONCE(!list_empty(&napi->poll_list)); 4942 list_add_tail(&napi->poll_list, &sd->poll_list); 4943 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4944 /* If not called from net_rx_action() 4945 * we have to raise NET_RX_SOFTIRQ. 4946 */ 4947 if (!sd->in_net_rx_action) 4948 raise_softirq_irqoff(NET_RX_SOFTIRQ); 4949 } 4950 4951 #ifdef CONFIG_RPS 4952 4953 struct static_key_false rps_needed __read_mostly; 4954 EXPORT_SYMBOL(rps_needed); 4955 struct static_key_false rfs_needed __read_mostly; 4956 EXPORT_SYMBOL(rfs_needed); 4957 4958 static u32 rfs_slot(u32 hash, const struct rps_dev_flow_table *flow_table) 4959 { 4960 return hash_32(hash, flow_table->log); 4961 } 4962 4963 #ifdef CONFIG_RFS_ACCEL 4964 /** 4965 * rps_flow_is_active - check whether the flow is recently active. 4966 * @rflow: Specific flow to check activity. 4967 * @flow_table: per-queue flowtable that @rflow belongs to. 4968 * @cpu: CPU saved in @rflow. 4969 * 4970 * If the CPU has processed many packets since the flow's last activity 4971 * (beyond 10 times the table size), the flow is considered stale. 4972 * 4973 * Return: true if flow was recently active. 4974 */ 4975 static bool rps_flow_is_active(struct rps_dev_flow *rflow, 4976 struct rps_dev_flow_table *flow_table, 4977 unsigned int cpu) 4978 { 4979 unsigned int flow_last_active; 4980 unsigned int sd_input_head; 4981 4982 if (cpu >= nr_cpu_ids) 4983 return false; 4984 4985 sd_input_head = READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head); 4986 flow_last_active = READ_ONCE(rflow->last_qtail); 4987 4988 return (int)(sd_input_head - flow_last_active) < 4989 (int)(10 << flow_table->log); 4990 } 4991 #endif 4992 4993 static struct rps_dev_flow * 4994 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4995 struct rps_dev_flow *rflow, u16 next_cpu, u32 hash) 4996 { 4997 if (next_cpu < nr_cpu_ids) { 4998 u32 head; 4999 #ifdef CONFIG_RFS_ACCEL 5000 struct netdev_rx_queue *rxqueue; 5001 struct rps_dev_flow_table *flow_table; 5002 struct rps_dev_flow *old_rflow; 5003 struct rps_dev_flow *tmp_rflow; 5004 unsigned int tmp_cpu; 5005 u16 rxq_index; 5006 u32 flow_id; 5007 int rc; 5008 5009 /* Should we steer this flow to a different hardware queue? */ 5010 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 5011 !(dev->features & NETIF_F_NTUPLE)) 5012 goto out; 5013 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 5014 if (rxq_index == skb_get_rx_queue(skb)) 5015 goto out; 5016 5017 rxqueue = dev->_rx + rxq_index; 5018 flow_table = rcu_dereference(rxqueue->rps_flow_table); 5019 if (!flow_table) 5020 goto out; 5021 5022 flow_id = rfs_slot(hash, flow_table); 5023 tmp_rflow = &flow_table->flows[flow_id]; 5024 tmp_cpu = READ_ONCE(tmp_rflow->cpu); 5025 5026 if (READ_ONCE(tmp_rflow->filter) != RPS_NO_FILTER) { 5027 if (rps_flow_is_active(tmp_rflow, flow_table, 5028 tmp_cpu)) { 5029 if (hash != READ_ONCE(tmp_rflow->hash) || 5030 next_cpu == tmp_cpu) 5031 goto out; 5032 } 5033 } 5034 5035 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 5036 rxq_index, flow_id); 5037 if (rc < 0) 5038 goto out; 5039 5040 old_rflow = rflow; 5041 rflow = tmp_rflow; 5042 WRITE_ONCE(rflow->filter, rc); 5043 WRITE_ONCE(rflow->hash, hash); 5044 5045 if (old_rflow->filter == rc) 5046 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 5047 out: 5048 #endif 5049 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 5050 rps_input_queue_tail_save(&rflow->last_qtail, head); 5051 } 5052 5053 WRITE_ONCE(rflow->cpu, next_cpu); 5054 return rflow; 5055 } 5056 5057 /* 5058 * get_rps_cpu is called from netif_receive_skb and returns the target 5059 * CPU from the RPS map of the receiving queue for a given skb. 5060 * rcu_read_lock must be held on entry. 5061 */ 5062 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 5063 struct rps_dev_flow **rflowp) 5064 { 5065 const struct rps_sock_flow_table *sock_flow_table; 5066 struct netdev_rx_queue *rxqueue = dev->_rx; 5067 struct rps_dev_flow_table *flow_table; 5068 struct rps_map *map; 5069 int cpu = -1; 5070 u32 tcpu; 5071 u32 hash; 5072 5073 if (skb_rx_queue_recorded(skb)) { 5074 u16 index = skb_get_rx_queue(skb); 5075 5076 if (unlikely(index >= dev->real_num_rx_queues)) { 5077 WARN_ONCE(dev->real_num_rx_queues > 1, 5078 "%s received packet on queue %u, but number " 5079 "of RX queues is %u\n", 5080 dev->name, index, dev->real_num_rx_queues); 5081 goto done; 5082 } 5083 rxqueue += index; 5084 } 5085 5086 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 5087 5088 flow_table = rcu_dereference(rxqueue->rps_flow_table); 5089 map = rcu_dereference(rxqueue->rps_map); 5090 if (!flow_table && !map) 5091 goto done; 5092 5093 skb_reset_network_header(skb); 5094 hash = skb_get_hash(skb); 5095 if (!hash) 5096 goto done; 5097 5098 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 5099 if (flow_table && sock_flow_table) { 5100 struct rps_dev_flow *rflow; 5101 u32 next_cpu; 5102 u32 ident; 5103 5104 /* First check into global flow table if there is a match. 5105 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 5106 */ 5107 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 5108 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 5109 goto try_rps; 5110 5111 next_cpu = ident & net_hotdata.rps_cpu_mask; 5112 5113 /* OK, now we know there is a match, 5114 * we can look at the local (per receive queue) flow table 5115 */ 5116 rflow = &flow_table->flows[rfs_slot(hash, flow_table)]; 5117 tcpu = rflow->cpu; 5118 5119 /* 5120 * If the desired CPU (where last recvmsg was done) is 5121 * different from current CPU (one in the rx-queue flow 5122 * table entry), switch if one of the following holds: 5123 * - Current CPU is unset (>= nr_cpu_ids). 5124 * - Current CPU is offline. 5125 * - The current CPU's queue tail has advanced beyond the 5126 * last packet that was enqueued using this table entry. 5127 * This guarantees that all previous packets for the flow 5128 * have been dequeued, thus preserving in order delivery. 5129 */ 5130 if (unlikely(tcpu != next_cpu) && 5131 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 5132 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 5133 rflow->last_qtail)) >= 0)) { 5134 tcpu = next_cpu; 5135 rflow = set_rps_cpu(dev, skb, rflow, next_cpu, hash); 5136 } 5137 5138 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 5139 *rflowp = rflow; 5140 cpu = tcpu; 5141 goto done; 5142 } 5143 } 5144 5145 try_rps: 5146 5147 if (map) { 5148 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 5149 if (cpu_online(tcpu)) { 5150 cpu = tcpu; 5151 goto done; 5152 } 5153 } 5154 5155 done: 5156 return cpu; 5157 } 5158 5159 #ifdef CONFIG_RFS_ACCEL 5160 5161 /** 5162 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 5163 * @dev: Device on which the filter was set 5164 * @rxq_index: RX queue index 5165 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 5166 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 5167 * 5168 * Drivers that implement ndo_rx_flow_steer() should periodically call 5169 * this function for each installed filter and remove the filters for 5170 * which it returns %true. 5171 */ 5172 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 5173 u32 flow_id, u16 filter_id) 5174 { 5175 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 5176 struct rps_dev_flow_table *flow_table; 5177 struct rps_dev_flow *rflow; 5178 bool expire = true; 5179 5180 rcu_read_lock(); 5181 flow_table = rcu_dereference(rxqueue->rps_flow_table); 5182 if (flow_table && flow_id < (1UL << flow_table->log)) { 5183 unsigned int cpu; 5184 5185 rflow = &flow_table->flows[flow_id]; 5186 cpu = READ_ONCE(rflow->cpu); 5187 if (READ_ONCE(rflow->filter) == filter_id && 5188 rps_flow_is_active(rflow, flow_table, cpu)) 5189 expire = false; 5190 } 5191 rcu_read_unlock(); 5192 return expire; 5193 } 5194 EXPORT_SYMBOL(rps_may_expire_flow); 5195 5196 #endif /* CONFIG_RFS_ACCEL */ 5197 5198 /* Called from hardirq (IPI) context */ 5199 static void rps_trigger_softirq(void *data) 5200 { 5201 struct softnet_data *sd = data; 5202 5203 ____napi_schedule(sd, &sd->backlog); 5204 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5205 WRITE_ONCE(sd->received_rps, sd->received_rps + 1); 5206 } 5207 5208 #endif /* CONFIG_RPS */ 5209 5210 /* Called from hardirq (IPI) context */ 5211 static void trigger_rx_softirq(void *data) 5212 { 5213 struct softnet_data *sd = data; 5214 5215 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5216 smp_store_release(&sd->defer_ipi_scheduled, 0); 5217 } 5218 5219 /* 5220 * After we queued a packet into sd->input_pkt_queue, 5221 * we need to make sure this queue is serviced soon. 5222 * 5223 * - If this is another cpu queue, link it to our rps_ipi_list, 5224 * and make sure we will process rps_ipi_list from net_rx_action(). 5225 * 5226 * - If this is our own queue, NAPI schedule our backlog. 5227 * Note that this also raises NET_RX_SOFTIRQ. 5228 */ 5229 static void napi_schedule_rps(struct softnet_data *sd) 5230 { 5231 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 5232 5233 #ifdef CONFIG_RPS 5234 if (sd != mysd) { 5235 if (use_backlog_threads()) { 5236 __napi_schedule_irqoff(&sd->backlog); 5237 return; 5238 } 5239 5240 sd->rps_ipi_next = mysd->rps_ipi_list; 5241 mysd->rps_ipi_list = sd; 5242 5243 /* If not called from net_rx_action() or napi_threaded_poll() 5244 * we have to raise NET_RX_SOFTIRQ. 5245 */ 5246 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 5247 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5248 return; 5249 } 5250 #endif /* CONFIG_RPS */ 5251 __napi_schedule_irqoff(&mysd->backlog); 5252 } 5253 5254 void kick_defer_list_purge(unsigned int cpu) 5255 { 5256 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 5257 unsigned long flags; 5258 5259 if (use_backlog_threads()) { 5260 backlog_lock_irq_save(sd, &flags); 5261 5262 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 5263 __napi_schedule_irqoff(&sd->backlog); 5264 5265 backlog_unlock_irq_restore(sd, flags); 5266 5267 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 5268 smp_call_function_single_async(cpu, &sd->defer_csd); 5269 } 5270 } 5271 5272 #ifdef CONFIG_NET_FLOW_LIMIT 5273 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 5274 #endif 5275 5276 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen, 5277 int max_backlog) 5278 { 5279 #ifdef CONFIG_NET_FLOW_LIMIT 5280 unsigned int old_flow, new_flow; 5281 const struct softnet_data *sd; 5282 struct sd_flow_limit *fl; 5283 5284 if (likely(qlen < (max_backlog >> 1))) 5285 return false; 5286 5287 sd = this_cpu_ptr(&softnet_data); 5288 5289 rcu_read_lock(); 5290 fl = rcu_dereference(sd->flow_limit); 5291 if (fl) { 5292 new_flow = hash_32(skb_get_hash(skb), fl->log_buckets); 5293 old_flow = fl->history[fl->history_head]; 5294 fl->history[fl->history_head] = new_flow; 5295 5296 fl->history_head++; 5297 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5298 5299 if (likely(fl->buckets[old_flow])) 5300 fl->buckets[old_flow]--; 5301 5302 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5303 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5304 WRITE_ONCE(fl->count, fl->count + 1); 5305 rcu_read_unlock(); 5306 return true; 5307 } 5308 } 5309 rcu_read_unlock(); 5310 #endif 5311 return false; 5312 } 5313 5314 /* 5315 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5316 * queue (may be a remote CPU queue). 5317 */ 5318 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5319 unsigned int *qtail) 5320 { 5321 enum skb_drop_reason reason; 5322 struct softnet_data *sd; 5323 unsigned long flags; 5324 unsigned int qlen; 5325 int max_backlog; 5326 u32 tail; 5327 5328 reason = SKB_DROP_REASON_DEV_READY; 5329 if (unlikely(!netif_running(skb->dev))) 5330 goto bad_dev; 5331 5332 sd = &per_cpu(softnet_data, cpu); 5333 5334 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5335 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5336 if (unlikely(qlen > max_backlog) || 5337 skb_flow_limit(skb, qlen, max_backlog)) 5338 goto cpu_backlog_drop; 5339 backlog_lock_irq_save(sd, &flags); 5340 qlen = skb_queue_len(&sd->input_pkt_queue); 5341 if (likely(qlen <= max_backlog)) { 5342 if (!qlen) { 5343 /* Schedule NAPI for backlog device. We can use 5344 * non atomic operation as we own the queue lock. 5345 */ 5346 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5347 &sd->backlog.state)) 5348 napi_schedule_rps(sd); 5349 } 5350 __skb_queue_tail(&sd->input_pkt_queue, skb); 5351 tail = rps_input_queue_tail_incr(sd); 5352 backlog_unlock_irq_restore(sd, flags); 5353 5354 /* save the tail outside of the critical section */ 5355 rps_input_queue_tail_save(qtail, tail); 5356 return NET_RX_SUCCESS; 5357 } 5358 5359 backlog_unlock_irq_restore(sd, flags); 5360 5361 cpu_backlog_drop: 5362 reason = SKB_DROP_REASON_CPU_BACKLOG; 5363 numa_drop_add(&sd->drop_counters, 1); 5364 bad_dev: 5365 dev_core_stats_rx_dropped_inc(skb->dev); 5366 kfree_skb_reason(skb, reason); 5367 return NET_RX_DROP; 5368 } 5369 5370 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5371 { 5372 struct net_device *dev = skb->dev; 5373 struct netdev_rx_queue *rxqueue; 5374 5375 rxqueue = dev->_rx; 5376 5377 if (skb_rx_queue_recorded(skb)) { 5378 u16 index = skb_get_rx_queue(skb); 5379 5380 if (unlikely(index >= dev->real_num_rx_queues)) { 5381 WARN_ONCE(dev->real_num_rx_queues > 1, 5382 "%s received packet on queue %u, but number " 5383 "of RX queues is %u\n", 5384 dev->name, index, dev->real_num_rx_queues); 5385 5386 return rxqueue; /* Return first rxqueue */ 5387 } 5388 rxqueue += index; 5389 } 5390 return rxqueue; 5391 } 5392 5393 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5394 const struct bpf_prog *xdp_prog) 5395 { 5396 void *orig_data, *orig_data_end, *hard_start; 5397 struct netdev_rx_queue *rxqueue; 5398 bool orig_bcast, orig_host; 5399 u32 mac_len, frame_sz; 5400 __be16 orig_eth_type; 5401 struct ethhdr *eth; 5402 u32 metalen, act; 5403 int off; 5404 5405 /* The XDP program wants to see the packet starting at the MAC 5406 * header. 5407 */ 5408 mac_len = skb->data - skb_mac_header(skb); 5409 hard_start = skb->data - skb_headroom(skb); 5410 5411 /* SKB "head" area always have tailroom for skb_shared_info */ 5412 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5413 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5414 5415 rxqueue = netif_get_rxqueue(skb); 5416 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5417 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5418 skb_headlen(skb) + mac_len, true); 5419 if (skb_is_nonlinear(skb)) { 5420 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5421 xdp_buff_set_frags_flag(xdp); 5422 } else { 5423 xdp_buff_clear_frags_flag(xdp); 5424 } 5425 5426 orig_data_end = xdp->data_end; 5427 orig_data = xdp->data; 5428 eth = (struct ethhdr *)xdp->data; 5429 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5430 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5431 orig_eth_type = eth->h_proto; 5432 5433 act = bpf_prog_run_xdp(xdp_prog, xdp); 5434 5435 /* check if bpf_xdp_adjust_head was used */ 5436 off = xdp->data - orig_data; 5437 if (off) { 5438 if (off > 0) 5439 __skb_pull(skb, off); 5440 else if (off < 0) 5441 __skb_push(skb, -off); 5442 5443 skb->mac_header += off; 5444 skb_reset_network_header(skb); 5445 } 5446 5447 /* check if bpf_xdp_adjust_tail was used */ 5448 off = xdp->data_end - orig_data_end; 5449 if (off != 0) { 5450 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5451 skb->len += off; /* positive on grow, negative on shrink */ 5452 } 5453 5454 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5455 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 5456 */ 5457 if (xdp_buff_has_frags(xdp)) 5458 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5459 else 5460 skb->data_len = 0; 5461 5462 /* check if XDP changed eth hdr such SKB needs update */ 5463 eth = (struct ethhdr *)xdp->data; 5464 if ((orig_eth_type != eth->h_proto) || 5465 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5466 skb->dev->dev_addr)) || 5467 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5468 __skb_push(skb, ETH_HLEN); 5469 skb->pkt_type = PACKET_HOST; 5470 skb->protocol = eth_type_trans(skb, skb->dev); 5471 } 5472 5473 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5474 * before calling us again on redirect path. We do not call do_redirect 5475 * as we leave that up to the caller. 5476 * 5477 * Caller is responsible for managing lifetime of skb (i.e. calling 5478 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5479 */ 5480 switch (act) { 5481 case XDP_REDIRECT: 5482 case XDP_TX: 5483 __skb_push(skb, mac_len); 5484 break; 5485 case XDP_PASS: 5486 metalen = xdp->data - xdp->data_meta; 5487 if (metalen) 5488 skb_metadata_set(skb, metalen); 5489 break; 5490 } 5491 5492 return act; 5493 } 5494 5495 static int 5496 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5497 { 5498 struct sk_buff *skb = *pskb; 5499 int err, hroom, troom; 5500 5501 local_lock_nested_bh(&system_page_pool.bh_lock); 5502 err = skb_cow_data_for_xdp(this_cpu_read(system_page_pool.pool), pskb, prog); 5503 local_unlock_nested_bh(&system_page_pool.bh_lock); 5504 if (!err) 5505 return 0; 5506 5507 /* In case we have to go down the path and also linearize, 5508 * then lets do the pskb_expand_head() work just once here. 5509 */ 5510 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5511 troom = skb->tail + skb->data_len - skb->end; 5512 err = pskb_expand_head(skb, 5513 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5514 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5515 if (err) 5516 return err; 5517 5518 return skb_linearize(skb); 5519 } 5520 5521 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5522 struct xdp_buff *xdp, 5523 const struct bpf_prog *xdp_prog) 5524 { 5525 struct sk_buff *skb = *pskb; 5526 u32 mac_len, act = XDP_DROP; 5527 5528 /* Reinjected packets coming from act_mirred or similar should 5529 * not get XDP generic processing. 5530 */ 5531 if (skb_is_redirected(skb)) 5532 return XDP_PASS; 5533 5534 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5535 * bytes. This is the guarantee that also native XDP provides, 5536 * thus we need to do it here as well. 5537 */ 5538 mac_len = skb->data - skb_mac_header(skb); 5539 __skb_push(skb, mac_len); 5540 5541 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5542 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5543 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5544 goto do_drop; 5545 } 5546 5547 __skb_pull(*pskb, mac_len); 5548 5549 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5550 switch (act) { 5551 case XDP_REDIRECT: 5552 case XDP_TX: 5553 case XDP_PASS: 5554 break; 5555 default: 5556 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5557 fallthrough; 5558 case XDP_ABORTED: 5559 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5560 fallthrough; 5561 case XDP_DROP: 5562 do_drop: 5563 kfree_skb(*pskb); 5564 break; 5565 } 5566 5567 return act; 5568 } 5569 5570 /* When doing generic XDP we have to bypass the qdisc layer and the 5571 * network taps in order to match in-driver-XDP behavior. This also means 5572 * that XDP packets are able to starve other packets going through a qdisc, 5573 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5574 * queues, so they do not have this starvation issue. 5575 */ 5576 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5577 { 5578 struct net_device *dev = skb->dev; 5579 struct netdev_queue *txq; 5580 bool free_skb = true; 5581 int cpu, rc; 5582 5583 txq = netdev_core_pick_tx(dev, skb, NULL); 5584 cpu = smp_processor_id(); 5585 HARD_TX_LOCK(dev, txq, cpu); 5586 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5587 rc = netdev_start_xmit(skb, dev, txq, 0); 5588 if (dev_xmit_complete(rc)) 5589 free_skb = false; 5590 } 5591 HARD_TX_UNLOCK(dev, txq); 5592 if (free_skb) { 5593 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5594 dev_core_stats_tx_dropped_inc(dev); 5595 kfree_skb(skb); 5596 } 5597 } 5598 5599 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5600 5601 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5602 { 5603 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5604 5605 if (xdp_prog) { 5606 struct xdp_buff xdp; 5607 u32 act; 5608 int err; 5609 5610 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5611 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5612 if (act != XDP_PASS) { 5613 switch (act) { 5614 case XDP_REDIRECT: 5615 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5616 &xdp, xdp_prog); 5617 if (err) 5618 goto out_redir; 5619 break; 5620 case XDP_TX: 5621 generic_xdp_tx(*pskb, xdp_prog); 5622 break; 5623 } 5624 bpf_net_ctx_clear(bpf_net_ctx); 5625 return XDP_DROP; 5626 } 5627 bpf_net_ctx_clear(bpf_net_ctx); 5628 } 5629 return XDP_PASS; 5630 out_redir: 5631 bpf_net_ctx_clear(bpf_net_ctx); 5632 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5633 return XDP_DROP; 5634 } 5635 EXPORT_SYMBOL_GPL(do_xdp_generic); 5636 5637 static int netif_rx_internal(struct sk_buff *skb) 5638 { 5639 int ret; 5640 5641 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5642 5643 trace_netif_rx(skb); 5644 5645 #ifdef CONFIG_RPS 5646 if (static_branch_unlikely(&rps_needed)) { 5647 struct rps_dev_flow voidflow, *rflow = &voidflow; 5648 int cpu; 5649 5650 rcu_read_lock(); 5651 5652 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5653 if (cpu < 0) 5654 cpu = smp_processor_id(); 5655 5656 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5657 5658 rcu_read_unlock(); 5659 } else 5660 #endif 5661 { 5662 unsigned int qtail; 5663 5664 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5665 } 5666 return ret; 5667 } 5668 5669 /** 5670 * __netif_rx - Slightly optimized version of netif_rx 5671 * @skb: buffer to post 5672 * 5673 * This behaves as netif_rx except that it does not disable bottom halves. 5674 * As a result this function may only be invoked from the interrupt context 5675 * (either hard or soft interrupt). 5676 */ 5677 int __netif_rx(struct sk_buff *skb) 5678 { 5679 int ret; 5680 5681 lockdep_assert_once(hardirq_count() | softirq_count()); 5682 5683 trace_netif_rx_entry(skb); 5684 ret = netif_rx_internal(skb); 5685 trace_netif_rx_exit(ret); 5686 return ret; 5687 } 5688 EXPORT_SYMBOL(__netif_rx); 5689 5690 /** 5691 * netif_rx - post buffer to the network code 5692 * @skb: buffer to post 5693 * 5694 * This function receives a packet from a device driver and queues it for 5695 * the upper (protocol) levels to process via the backlog NAPI device. It 5696 * always succeeds. The buffer may be dropped during processing for 5697 * congestion control or by the protocol layers. 5698 * The network buffer is passed via the backlog NAPI device. Modern NIC 5699 * driver should use NAPI and GRO. 5700 * This function can used from interrupt and from process context. The 5701 * caller from process context must not disable interrupts before invoking 5702 * this function. 5703 * 5704 * return values: 5705 * NET_RX_SUCCESS (no congestion) 5706 * NET_RX_DROP (packet was dropped) 5707 * 5708 */ 5709 int netif_rx(struct sk_buff *skb) 5710 { 5711 bool need_bh_off = !(hardirq_count() | softirq_count()); 5712 int ret; 5713 5714 if (need_bh_off) 5715 local_bh_disable(); 5716 trace_netif_rx_entry(skb); 5717 ret = netif_rx_internal(skb); 5718 trace_netif_rx_exit(ret); 5719 if (need_bh_off) 5720 local_bh_enable(); 5721 return ret; 5722 } 5723 EXPORT_SYMBOL(netif_rx); 5724 5725 static __latent_entropy void net_tx_action(void) 5726 { 5727 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5728 5729 if (sd->completion_queue) { 5730 struct sk_buff *clist; 5731 5732 local_irq_disable(); 5733 clist = sd->completion_queue; 5734 sd->completion_queue = NULL; 5735 local_irq_enable(); 5736 5737 while (clist) { 5738 struct sk_buff *skb = clist; 5739 5740 clist = clist->next; 5741 5742 WARN_ON(refcount_read(&skb->users)); 5743 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5744 trace_consume_skb(skb, net_tx_action); 5745 else 5746 trace_kfree_skb(skb, net_tx_action, 5747 get_kfree_skb_cb(skb)->reason, NULL); 5748 5749 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5750 __kfree_skb(skb); 5751 else 5752 __napi_kfree_skb(skb, 5753 get_kfree_skb_cb(skb)->reason); 5754 } 5755 } 5756 5757 if (sd->output_queue) { 5758 struct Qdisc *head; 5759 5760 local_irq_disable(); 5761 head = sd->output_queue; 5762 sd->output_queue = NULL; 5763 sd->output_queue_tailp = &sd->output_queue; 5764 local_irq_enable(); 5765 5766 rcu_read_lock(); 5767 5768 while (head) { 5769 spinlock_t *root_lock = NULL; 5770 struct sk_buff *to_free; 5771 struct Qdisc *q = head; 5772 5773 head = head->next_sched; 5774 5775 /* We need to make sure head->next_sched is read 5776 * before clearing __QDISC_STATE_SCHED 5777 */ 5778 smp_mb__before_atomic(); 5779 5780 if (!(q->flags & TCQ_F_NOLOCK)) { 5781 root_lock = qdisc_lock(q); 5782 spin_lock(root_lock); 5783 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5784 &q->state))) { 5785 /* There is a synchronize_net() between 5786 * STATE_DEACTIVATED flag being set and 5787 * qdisc_reset()/some_qdisc_is_busy() in 5788 * dev_deactivate(), so we can safely bail out 5789 * early here to avoid data race between 5790 * qdisc_deactivate() and some_qdisc_is_busy() 5791 * for lockless qdisc. 5792 */ 5793 clear_bit(__QDISC_STATE_SCHED, &q->state); 5794 continue; 5795 } 5796 5797 clear_bit(__QDISC_STATE_SCHED, &q->state); 5798 to_free = qdisc_run(q); 5799 if (root_lock) 5800 spin_unlock(root_lock); 5801 tcf_kfree_skb_list(to_free); 5802 } 5803 5804 rcu_read_unlock(); 5805 } 5806 5807 xfrm_dev_backlog(sd); 5808 } 5809 5810 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5811 /* This hook is defined here for ATM LANE */ 5812 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5813 unsigned char *addr) __read_mostly; 5814 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5815 #endif 5816 5817 /** 5818 * netdev_is_rx_handler_busy - check if receive handler is registered 5819 * @dev: device to check 5820 * 5821 * Check if a receive handler is already registered for a given device. 5822 * Return true if there one. 5823 * 5824 * The caller must hold the rtnl_mutex. 5825 */ 5826 bool netdev_is_rx_handler_busy(struct net_device *dev) 5827 { 5828 ASSERT_RTNL(); 5829 return dev && rtnl_dereference(dev->rx_handler); 5830 } 5831 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5832 5833 /** 5834 * netdev_rx_handler_register - register receive handler 5835 * @dev: device to register a handler for 5836 * @rx_handler: receive handler to register 5837 * @rx_handler_data: data pointer that is used by rx handler 5838 * 5839 * Register a receive handler for a device. This handler will then be 5840 * called from __netif_receive_skb. A negative errno code is returned 5841 * on a failure. 5842 * 5843 * The caller must hold the rtnl_mutex. 5844 * 5845 * For a general description of rx_handler, see enum rx_handler_result. 5846 */ 5847 int netdev_rx_handler_register(struct net_device *dev, 5848 rx_handler_func_t *rx_handler, 5849 void *rx_handler_data) 5850 { 5851 if (netdev_is_rx_handler_busy(dev)) 5852 return -EBUSY; 5853 5854 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5855 return -EINVAL; 5856 5857 /* Note: rx_handler_data must be set before rx_handler */ 5858 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5859 rcu_assign_pointer(dev->rx_handler, rx_handler); 5860 5861 return 0; 5862 } 5863 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5864 5865 /** 5866 * netdev_rx_handler_unregister - unregister receive handler 5867 * @dev: device to unregister a handler from 5868 * 5869 * Unregister a receive handler from a device. 5870 * 5871 * The caller must hold the rtnl_mutex. 5872 */ 5873 void netdev_rx_handler_unregister(struct net_device *dev) 5874 { 5875 5876 ASSERT_RTNL(); 5877 RCU_INIT_POINTER(dev->rx_handler, NULL); 5878 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5879 * section has a guarantee to see a non NULL rx_handler_data 5880 * as well. 5881 */ 5882 synchronize_net(); 5883 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5884 } 5885 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5886 5887 /* 5888 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5889 * the special handling of PFMEMALLOC skbs. 5890 */ 5891 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5892 { 5893 switch (skb->protocol) { 5894 case htons(ETH_P_ARP): 5895 case htons(ETH_P_IP): 5896 case htons(ETH_P_IPV6): 5897 case htons(ETH_P_8021Q): 5898 case htons(ETH_P_8021AD): 5899 return true; 5900 default: 5901 return false; 5902 } 5903 } 5904 5905 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5906 int *ret, struct net_device *orig_dev) 5907 { 5908 if (nf_hook_ingress_active(skb)) { 5909 int ingress_retval; 5910 5911 if (unlikely(*pt_prev)) { 5912 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5913 *pt_prev = NULL; 5914 } 5915 5916 rcu_read_lock(); 5917 ingress_retval = nf_hook_ingress(skb); 5918 rcu_read_unlock(); 5919 return ingress_retval; 5920 } 5921 return 0; 5922 } 5923 5924 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5925 struct packet_type **ppt_prev) 5926 { 5927 enum skb_drop_reason drop_reason = SKB_DROP_REASON_UNHANDLED_PROTO; 5928 struct packet_type *ptype, *pt_prev; 5929 rx_handler_func_t *rx_handler; 5930 struct sk_buff *skb = *pskb; 5931 struct net_device *orig_dev; 5932 bool deliver_exact = false; 5933 int ret = NET_RX_DROP; 5934 __be16 type; 5935 5936 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5937 5938 trace_netif_receive_skb(skb); 5939 5940 orig_dev = skb->dev; 5941 5942 skb_reset_network_header(skb); 5943 #if !defined(CONFIG_DEBUG_NET) 5944 /* We plan to no longer reset the transport header here. 5945 * Give some time to fuzzers and dev build to catch bugs 5946 * in network stacks. 5947 */ 5948 if (!skb_transport_header_was_set(skb)) 5949 skb_reset_transport_header(skb); 5950 #endif 5951 skb_reset_mac_len(skb); 5952 5953 pt_prev = NULL; 5954 5955 another_round: 5956 skb->skb_iif = skb->dev->ifindex; 5957 5958 __this_cpu_inc(softnet_data.processed); 5959 5960 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5961 int ret2; 5962 5963 migrate_disable(); 5964 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5965 &skb); 5966 migrate_enable(); 5967 5968 if (ret2 != XDP_PASS) { 5969 ret = NET_RX_DROP; 5970 goto out; 5971 } 5972 } 5973 5974 if (eth_type_vlan(skb->protocol)) { 5975 skb = skb_vlan_untag(skb); 5976 if (unlikely(!skb)) 5977 goto out; 5978 } 5979 5980 if (skb_skip_tc_classify(skb)) 5981 goto skip_classify; 5982 5983 if (pfmemalloc) 5984 goto skip_taps; 5985 5986 list_for_each_entry_rcu(ptype, &dev_net_rcu(skb->dev)->ptype_all, 5987 list) { 5988 if (unlikely(pt_prev)) 5989 ret = deliver_skb(skb, pt_prev, orig_dev); 5990 pt_prev = ptype; 5991 } 5992 5993 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5994 if (unlikely(pt_prev)) 5995 ret = deliver_skb(skb, pt_prev, orig_dev); 5996 pt_prev = ptype; 5997 } 5998 5999 skip_taps: 6000 #ifdef CONFIG_NET_INGRESS 6001 if (static_branch_unlikely(&ingress_needed_key)) { 6002 bool another = false; 6003 6004 nf_skip_egress(skb, true); 6005 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 6006 &another); 6007 if (another) 6008 goto another_round; 6009 if (!skb) 6010 goto out; 6011 6012 nf_skip_egress(skb, false); 6013 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 6014 goto out; 6015 } 6016 #endif 6017 skb_reset_redirect(skb); 6018 skip_classify: 6019 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) { 6020 drop_reason = SKB_DROP_REASON_PFMEMALLOC; 6021 goto drop; 6022 } 6023 6024 if (skb_vlan_tag_present(skb)) { 6025 if (unlikely(pt_prev)) { 6026 ret = deliver_skb(skb, pt_prev, orig_dev); 6027 pt_prev = NULL; 6028 } 6029 if (vlan_do_receive(&skb)) 6030 goto another_round; 6031 else if (unlikely(!skb)) 6032 goto out; 6033 } 6034 6035 rx_handler = rcu_dereference(skb->dev->rx_handler); 6036 if (rx_handler) { 6037 if (unlikely(pt_prev)) { 6038 ret = deliver_skb(skb, pt_prev, orig_dev); 6039 pt_prev = NULL; 6040 } 6041 switch (rx_handler(&skb)) { 6042 case RX_HANDLER_CONSUMED: 6043 ret = NET_RX_SUCCESS; 6044 goto out; 6045 case RX_HANDLER_ANOTHER: 6046 goto another_round; 6047 case RX_HANDLER_EXACT: 6048 deliver_exact = true; 6049 break; 6050 case RX_HANDLER_PASS: 6051 break; 6052 default: 6053 BUG(); 6054 } 6055 } 6056 6057 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 6058 check_vlan_id: 6059 if (skb_vlan_tag_get_id(skb)) { 6060 /* Vlan id is non 0 and vlan_do_receive() above couldn't 6061 * find vlan device. 6062 */ 6063 skb->pkt_type = PACKET_OTHERHOST; 6064 } else if (eth_type_vlan(skb->protocol)) { 6065 /* Outer header is 802.1P with vlan 0, inner header is 6066 * 802.1Q or 802.1AD and vlan_do_receive() above could 6067 * not find vlan dev for vlan id 0. 6068 */ 6069 __vlan_hwaccel_clear_tag(skb); 6070 skb = skb_vlan_untag(skb); 6071 if (unlikely(!skb)) 6072 goto out; 6073 if (vlan_do_receive(&skb)) 6074 /* After stripping off 802.1P header with vlan 0 6075 * vlan dev is found for inner header. 6076 */ 6077 goto another_round; 6078 else if (unlikely(!skb)) 6079 goto out; 6080 else 6081 /* We have stripped outer 802.1P vlan 0 header. 6082 * But could not find vlan dev. 6083 * check again for vlan id to set OTHERHOST. 6084 */ 6085 goto check_vlan_id; 6086 } 6087 /* Note: we might in the future use prio bits 6088 * and set skb->priority like in vlan_do_receive() 6089 * For the time being, just ignore Priority Code Point 6090 */ 6091 __vlan_hwaccel_clear_tag(skb); 6092 } 6093 6094 type = skb->protocol; 6095 6096 /* deliver only exact match when indicated */ 6097 if (likely(!deliver_exact)) { 6098 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6099 &ptype_base[ntohs(type) & 6100 PTYPE_HASH_MASK]); 6101 6102 /* orig_dev and skb->dev could belong to different netns; 6103 * Even in such case we need to traverse only the list 6104 * coming from skb->dev, as the ptype owner (packet socket) 6105 * will use dev_net(skb->dev) to do namespace filtering. 6106 */ 6107 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6108 &dev_net_rcu(skb->dev)->ptype_specific); 6109 } 6110 6111 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6112 &orig_dev->ptype_specific); 6113 6114 if (unlikely(skb->dev != orig_dev)) { 6115 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6116 &skb->dev->ptype_specific); 6117 } 6118 6119 if (pt_prev) { 6120 *ppt_prev = pt_prev; 6121 } else { 6122 drop: 6123 if (!deliver_exact) 6124 dev_core_stats_rx_dropped_inc(skb->dev); 6125 else 6126 dev_core_stats_rx_nohandler_inc(skb->dev); 6127 6128 kfree_skb_reason(skb, drop_reason); 6129 /* Jamal, now you will not able to escape explaining 6130 * me how you were going to use this. :-) 6131 */ 6132 ret = NET_RX_DROP; 6133 } 6134 6135 out: 6136 /* The invariant here is that if *ppt_prev is not NULL 6137 * then skb should also be non-NULL. 6138 * 6139 * Apparently *ppt_prev assignment above holds this invariant due to 6140 * skb dereferencing near it. 6141 */ 6142 *pskb = skb; 6143 return ret; 6144 } 6145 6146 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 6147 { 6148 struct net_device *orig_dev = skb->dev; 6149 struct packet_type *pt_prev = NULL; 6150 int ret; 6151 6152 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6153 if (pt_prev) 6154 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 6155 skb->dev, pt_prev, orig_dev); 6156 return ret; 6157 } 6158 6159 /** 6160 * netif_receive_skb_core - special purpose version of netif_receive_skb 6161 * @skb: buffer to process 6162 * 6163 * More direct receive version of netif_receive_skb(). It should 6164 * only be used by callers that have a need to skip RPS and Generic XDP. 6165 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 6166 * 6167 * This function may only be called from softirq context and interrupts 6168 * should be enabled. 6169 * 6170 * Return values (usually ignored): 6171 * NET_RX_SUCCESS: no congestion 6172 * NET_RX_DROP: packet was dropped 6173 */ 6174 int netif_receive_skb_core(struct sk_buff *skb) 6175 { 6176 int ret; 6177 6178 rcu_read_lock(); 6179 ret = __netif_receive_skb_one_core(skb, false); 6180 rcu_read_unlock(); 6181 6182 return ret; 6183 } 6184 EXPORT_SYMBOL(netif_receive_skb_core); 6185 6186 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 6187 struct packet_type *pt_prev, 6188 struct net_device *orig_dev) 6189 { 6190 struct sk_buff *skb, *next; 6191 6192 if (!pt_prev) 6193 return; 6194 if (list_empty(head)) 6195 return; 6196 if (pt_prev->list_func != NULL) 6197 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 6198 ip_list_rcv, head, pt_prev, orig_dev); 6199 else 6200 list_for_each_entry_safe(skb, next, head, list) { 6201 skb_list_del_init(skb); 6202 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 6203 } 6204 } 6205 6206 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 6207 { 6208 /* Fast-path assumptions: 6209 * - There is no RX handler. 6210 * - Only one packet_type matches. 6211 * If either of these fails, we will end up doing some per-packet 6212 * processing in-line, then handling the 'last ptype' for the whole 6213 * sublist. This can't cause out-of-order delivery to any single ptype, 6214 * because the 'last ptype' must be constant across the sublist, and all 6215 * other ptypes are handled per-packet. 6216 */ 6217 /* Current (common) ptype of sublist */ 6218 struct packet_type *pt_curr = NULL; 6219 /* Current (common) orig_dev of sublist */ 6220 struct net_device *od_curr = NULL; 6221 struct sk_buff *skb, *next; 6222 LIST_HEAD(sublist); 6223 6224 list_for_each_entry_safe(skb, next, head, list) { 6225 struct net_device *orig_dev = skb->dev; 6226 struct packet_type *pt_prev = NULL; 6227 6228 skb_list_del_init(skb); 6229 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6230 if (!pt_prev) 6231 continue; 6232 if (pt_curr != pt_prev || od_curr != orig_dev) { 6233 /* dispatch old sublist */ 6234 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6235 /* start new sublist */ 6236 INIT_LIST_HEAD(&sublist); 6237 pt_curr = pt_prev; 6238 od_curr = orig_dev; 6239 } 6240 list_add_tail(&skb->list, &sublist); 6241 } 6242 6243 /* dispatch final sublist */ 6244 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6245 } 6246 6247 static int __netif_receive_skb(struct sk_buff *skb) 6248 { 6249 int ret; 6250 6251 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 6252 unsigned int noreclaim_flag; 6253 6254 /* 6255 * PFMEMALLOC skbs are special, they should 6256 * - be delivered to SOCK_MEMALLOC sockets only 6257 * - stay away from userspace 6258 * - have bounded memory usage 6259 * 6260 * Use PF_MEMALLOC as this saves us from propagating the allocation 6261 * context down to all allocation sites. 6262 */ 6263 noreclaim_flag = memalloc_noreclaim_save(); 6264 ret = __netif_receive_skb_one_core(skb, true); 6265 memalloc_noreclaim_restore(noreclaim_flag); 6266 } else 6267 ret = __netif_receive_skb_one_core(skb, false); 6268 6269 return ret; 6270 } 6271 6272 static void __netif_receive_skb_list(struct list_head *head) 6273 { 6274 unsigned long noreclaim_flag = 0; 6275 struct sk_buff *skb, *next; 6276 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 6277 6278 list_for_each_entry_safe(skb, next, head, list) { 6279 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 6280 struct list_head sublist; 6281 6282 /* Handle the previous sublist */ 6283 list_cut_before(&sublist, head, &skb->list); 6284 if (!list_empty(&sublist)) 6285 __netif_receive_skb_list_core(&sublist, pfmemalloc); 6286 pfmemalloc = !pfmemalloc; 6287 /* See comments in __netif_receive_skb */ 6288 if (pfmemalloc) 6289 noreclaim_flag = memalloc_noreclaim_save(); 6290 else 6291 memalloc_noreclaim_restore(noreclaim_flag); 6292 } 6293 } 6294 /* Handle the remaining sublist */ 6295 if (!list_empty(head)) 6296 __netif_receive_skb_list_core(head, pfmemalloc); 6297 /* Restore pflags */ 6298 if (pfmemalloc) 6299 memalloc_noreclaim_restore(noreclaim_flag); 6300 } 6301 6302 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6303 { 6304 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6305 struct bpf_prog *new = xdp->prog; 6306 int ret = 0; 6307 6308 switch (xdp->command) { 6309 case XDP_SETUP_PROG: 6310 rcu_assign_pointer(dev->xdp_prog, new); 6311 if (old) 6312 bpf_prog_put(old); 6313 6314 if (old && !new) { 6315 static_branch_dec(&generic_xdp_needed_key); 6316 } else if (new && !old) { 6317 static_branch_inc(&generic_xdp_needed_key); 6318 netif_disable_lro(dev); 6319 dev_disable_gro_hw(dev); 6320 } 6321 break; 6322 6323 default: 6324 ret = -EINVAL; 6325 break; 6326 } 6327 6328 return ret; 6329 } 6330 6331 static int netif_receive_skb_internal(struct sk_buff *skb) 6332 { 6333 int ret; 6334 6335 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6336 6337 if (skb_defer_rx_timestamp(skb)) 6338 return NET_RX_SUCCESS; 6339 6340 rcu_read_lock(); 6341 #ifdef CONFIG_RPS 6342 if (static_branch_unlikely(&rps_needed)) { 6343 struct rps_dev_flow voidflow, *rflow = &voidflow; 6344 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6345 6346 if (cpu >= 0) { 6347 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6348 rcu_read_unlock(); 6349 return ret; 6350 } 6351 } 6352 #endif 6353 ret = __netif_receive_skb(skb); 6354 rcu_read_unlock(); 6355 return ret; 6356 } 6357 6358 void netif_receive_skb_list_internal(struct list_head *head) 6359 { 6360 struct sk_buff *skb, *next; 6361 LIST_HEAD(sublist); 6362 6363 list_for_each_entry_safe(skb, next, head, list) { 6364 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6365 skb); 6366 skb_list_del_init(skb); 6367 if (!skb_defer_rx_timestamp(skb)) 6368 list_add_tail(&skb->list, &sublist); 6369 } 6370 list_splice_init(&sublist, head); 6371 6372 rcu_read_lock(); 6373 #ifdef CONFIG_RPS 6374 if (static_branch_unlikely(&rps_needed)) { 6375 list_for_each_entry_safe(skb, next, head, list) { 6376 struct rps_dev_flow voidflow, *rflow = &voidflow; 6377 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6378 6379 if (cpu >= 0) { 6380 /* Will be handled, remove from list */ 6381 skb_list_del_init(skb); 6382 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6383 } 6384 } 6385 } 6386 #endif 6387 __netif_receive_skb_list(head); 6388 rcu_read_unlock(); 6389 } 6390 6391 /** 6392 * netif_receive_skb - process receive buffer from network 6393 * @skb: buffer to process 6394 * 6395 * netif_receive_skb() is the main receive data processing function. 6396 * It always succeeds. The buffer may be dropped during processing 6397 * for congestion control or by the protocol layers. 6398 * 6399 * This function may only be called from softirq context and interrupts 6400 * should be enabled. 6401 * 6402 * Return values (usually ignored): 6403 * NET_RX_SUCCESS: no congestion 6404 * NET_RX_DROP: packet was dropped 6405 */ 6406 int netif_receive_skb(struct sk_buff *skb) 6407 { 6408 int ret; 6409 6410 trace_netif_receive_skb_entry(skb); 6411 6412 ret = netif_receive_skb_internal(skb); 6413 trace_netif_receive_skb_exit(ret); 6414 6415 return ret; 6416 } 6417 EXPORT_SYMBOL(netif_receive_skb); 6418 6419 /** 6420 * netif_receive_skb_list - process many receive buffers from network 6421 * @head: list of skbs to process. 6422 * 6423 * Since return value of netif_receive_skb() is normally ignored, and 6424 * wouldn't be meaningful for a list, this function returns void. 6425 * 6426 * This function may only be called from softirq context and interrupts 6427 * should be enabled. 6428 */ 6429 void netif_receive_skb_list(struct list_head *head) 6430 { 6431 struct sk_buff *skb; 6432 6433 if (list_empty(head)) 6434 return; 6435 if (trace_netif_receive_skb_list_entry_enabled()) { 6436 list_for_each_entry(skb, head, list) 6437 trace_netif_receive_skb_list_entry(skb); 6438 } 6439 netif_receive_skb_list_internal(head); 6440 trace_netif_receive_skb_list_exit(0); 6441 } 6442 EXPORT_SYMBOL(netif_receive_skb_list); 6443 6444 /* Network device is going away, flush any packets still pending */ 6445 static void flush_backlog(struct work_struct *work) 6446 { 6447 struct sk_buff *skb, *tmp; 6448 struct sk_buff_head list; 6449 struct softnet_data *sd; 6450 6451 __skb_queue_head_init(&list); 6452 local_bh_disable(); 6453 sd = this_cpu_ptr(&softnet_data); 6454 6455 backlog_lock_irq_disable(sd); 6456 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6457 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6458 __skb_unlink(skb, &sd->input_pkt_queue); 6459 __skb_queue_tail(&list, skb); 6460 rps_input_queue_head_incr(sd); 6461 } 6462 } 6463 backlog_unlock_irq_enable(sd); 6464 6465 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6466 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6467 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6468 __skb_unlink(skb, &sd->process_queue); 6469 __skb_queue_tail(&list, skb); 6470 rps_input_queue_head_incr(sd); 6471 } 6472 } 6473 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6474 local_bh_enable(); 6475 6476 __skb_queue_purge_reason(&list, SKB_DROP_REASON_DEV_READY); 6477 } 6478 6479 static bool flush_required(int cpu) 6480 { 6481 #if IS_ENABLED(CONFIG_RPS) 6482 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6483 bool do_flush; 6484 6485 backlog_lock_irq_disable(sd); 6486 6487 /* as insertion into process_queue happens with the rps lock held, 6488 * process_queue access may race only with dequeue 6489 */ 6490 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6491 !skb_queue_empty_lockless(&sd->process_queue); 6492 backlog_unlock_irq_enable(sd); 6493 6494 return do_flush; 6495 #endif 6496 /* without RPS we can't safely check input_pkt_queue: during a 6497 * concurrent remote skb_queue_splice() we can detect as empty both 6498 * input_pkt_queue and process_queue even if the latter could end-up 6499 * containing a lot of packets. 6500 */ 6501 return true; 6502 } 6503 6504 struct flush_backlogs { 6505 cpumask_t flush_cpus; 6506 struct work_struct w[]; 6507 }; 6508 6509 static struct flush_backlogs *flush_backlogs_alloc(void) 6510 { 6511 return kmalloc(struct_size_t(struct flush_backlogs, w, nr_cpu_ids), 6512 GFP_KERNEL); 6513 } 6514 6515 static struct flush_backlogs *flush_backlogs_fallback; 6516 static DEFINE_MUTEX(flush_backlogs_mutex); 6517 6518 static void flush_all_backlogs(void) 6519 { 6520 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6521 unsigned int cpu; 6522 6523 if (!ptr) { 6524 mutex_lock(&flush_backlogs_mutex); 6525 ptr = flush_backlogs_fallback; 6526 } 6527 cpumask_clear(&ptr->flush_cpus); 6528 6529 cpus_read_lock(); 6530 6531 for_each_online_cpu(cpu) { 6532 if (flush_required(cpu)) { 6533 INIT_WORK(&ptr->w[cpu], flush_backlog); 6534 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6535 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6536 } 6537 } 6538 6539 /* we can have in flight packet[s] on the cpus we are not flushing, 6540 * synchronize_net() in unregister_netdevice_many() will take care of 6541 * them. 6542 */ 6543 for_each_cpu(cpu, &ptr->flush_cpus) 6544 flush_work(&ptr->w[cpu]); 6545 6546 cpus_read_unlock(); 6547 6548 if (ptr != flush_backlogs_fallback) 6549 kfree(ptr); 6550 else 6551 mutex_unlock(&flush_backlogs_mutex); 6552 } 6553 6554 static void net_rps_send_ipi(struct softnet_data *remsd) 6555 { 6556 #ifdef CONFIG_RPS 6557 while (remsd) { 6558 struct softnet_data *next = remsd->rps_ipi_next; 6559 6560 if (cpu_online(remsd->cpu)) 6561 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6562 remsd = next; 6563 } 6564 #endif 6565 } 6566 6567 /* 6568 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6569 * Note: called with local irq disabled, but exits with local irq enabled. 6570 */ 6571 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6572 { 6573 #ifdef CONFIG_RPS 6574 struct softnet_data *remsd = sd->rps_ipi_list; 6575 6576 if (!use_backlog_threads() && remsd) { 6577 sd->rps_ipi_list = NULL; 6578 6579 local_irq_enable(); 6580 6581 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6582 net_rps_send_ipi(remsd); 6583 } else 6584 #endif 6585 local_irq_enable(); 6586 } 6587 6588 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6589 { 6590 #ifdef CONFIG_RPS 6591 return !use_backlog_threads() && sd->rps_ipi_list; 6592 #else 6593 return false; 6594 #endif 6595 } 6596 6597 static int process_backlog(struct napi_struct *napi, int quota) 6598 { 6599 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6600 bool again = true; 6601 int work = 0; 6602 6603 /* Check if we have pending ipi, its better to send them now, 6604 * not waiting net_rx_action() end. 6605 */ 6606 if (sd_has_rps_ipi_waiting(sd)) { 6607 local_irq_disable(); 6608 net_rps_action_and_irq_enable(sd); 6609 } 6610 6611 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6612 while (again) { 6613 struct sk_buff *skb; 6614 6615 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6616 while ((skb = __skb_dequeue(&sd->process_queue))) { 6617 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6618 rcu_read_lock(); 6619 __netif_receive_skb(skb); 6620 rcu_read_unlock(); 6621 if (++work >= quota) { 6622 rps_input_queue_head_add(sd, work); 6623 return work; 6624 } 6625 6626 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6627 } 6628 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6629 6630 backlog_lock_irq_disable(sd); 6631 if (skb_queue_empty(&sd->input_pkt_queue)) { 6632 /* 6633 * Inline a custom version of __napi_complete(). 6634 * only current cpu owns and manipulates this napi, 6635 * and NAPI_STATE_SCHED is the only possible flag set 6636 * on backlog. 6637 * We can use a plain write instead of clear_bit(), 6638 * and we dont need an smp_mb() memory barrier. 6639 */ 6640 napi->state &= NAPIF_STATE_THREADED; 6641 again = false; 6642 } else { 6643 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6644 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6645 &sd->process_queue); 6646 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6647 } 6648 backlog_unlock_irq_enable(sd); 6649 } 6650 6651 if (work) 6652 rps_input_queue_head_add(sd, work); 6653 return work; 6654 } 6655 6656 /** 6657 * __napi_schedule - schedule for receive 6658 * @n: entry to schedule 6659 * 6660 * The entry's receive function will be scheduled to run. 6661 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6662 */ 6663 void __napi_schedule(struct napi_struct *n) 6664 { 6665 unsigned long flags; 6666 6667 local_irq_save(flags); 6668 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6669 local_irq_restore(flags); 6670 } 6671 EXPORT_SYMBOL(__napi_schedule); 6672 6673 /** 6674 * napi_schedule_prep - check if napi can be scheduled 6675 * @n: napi context 6676 * 6677 * Test if NAPI routine is already running, and if not mark 6678 * it as running. This is used as a condition variable to 6679 * insure only one NAPI poll instance runs. We also make 6680 * sure there is no pending NAPI disable. 6681 */ 6682 bool napi_schedule_prep(struct napi_struct *n) 6683 { 6684 unsigned long new, val = READ_ONCE(n->state); 6685 6686 do { 6687 if (unlikely(val & NAPIF_STATE_DISABLE)) 6688 return false; 6689 new = val | NAPIF_STATE_SCHED; 6690 6691 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6692 * This was suggested by Alexander Duyck, as compiler 6693 * emits better code than : 6694 * if (val & NAPIF_STATE_SCHED) 6695 * new |= NAPIF_STATE_MISSED; 6696 */ 6697 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6698 NAPIF_STATE_MISSED; 6699 } while (!try_cmpxchg(&n->state, &val, new)); 6700 6701 return !(val & NAPIF_STATE_SCHED); 6702 } 6703 EXPORT_SYMBOL(napi_schedule_prep); 6704 6705 /** 6706 * __napi_schedule_irqoff - schedule for receive 6707 * @n: entry to schedule 6708 * 6709 * Variant of __napi_schedule() assuming hard irqs are masked. 6710 * 6711 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6712 * because the interrupt disabled assumption might not be true 6713 * due to force-threaded interrupts and spinlock substitution. 6714 */ 6715 void __napi_schedule_irqoff(struct napi_struct *n) 6716 { 6717 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6718 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6719 else 6720 __napi_schedule(n); 6721 } 6722 EXPORT_SYMBOL(__napi_schedule_irqoff); 6723 6724 bool napi_complete_done(struct napi_struct *n, int work_done) 6725 { 6726 unsigned long flags, val, new, timeout = 0; 6727 bool ret = true; 6728 6729 /* 6730 * 1) Don't let napi dequeue from the cpu poll list 6731 * just in case its running on a different cpu. 6732 * 2) If we are busy polling, do nothing here, we have 6733 * the guarantee we will be called later. 6734 */ 6735 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6736 NAPIF_STATE_IN_BUSY_POLL))) 6737 return false; 6738 6739 if (work_done) { 6740 if (n->gro.bitmask) 6741 timeout = napi_get_gro_flush_timeout(n); 6742 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6743 } 6744 if (n->defer_hard_irqs_count > 0) { 6745 n->defer_hard_irqs_count--; 6746 timeout = napi_get_gro_flush_timeout(n); 6747 if (timeout) 6748 ret = false; 6749 } 6750 6751 /* 6752 * When the NAPI instance uses a timeout and keeps postponing 6753 * it, we need to bound somehow the time packets are kept in 6754 * the GRO layer. 6755 */ 6756 gro_flush_normal(&n->gro, !!timeout); 6757 6758 if (unlikely(!list_empty(&n->poll_list))) { 6759 /* If n->poll_list is not empty, we need to mask irqs */ 6760 local_irq_save(flags); 6761 list_del_init(&n->poll_list); 6762 local_irq_restore(flags); 6763 } 6764 WRITE_ONCE(n->list_owner, -1); 6765 6766 val = READ_ONCE(n->state); 6767 do { 6768 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6769 6770 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6771 NAPIF_STATE_SCHED_THREADED | 6772 NAPIF_STATE_PREFER_BUSY_POLL); 6773 6774 /* If STATE_MISSED was set, leave STATE_SCHED set, 6775 * because we will call napi->poll() one more time. 6776 * This C code was suggested by Alexander Duyck to help gcc. 6777 */ 6778 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6779 NAPIF_STATE_SCHED; 6780 } while (!try_cmpxchg(&n->state, &val, new)); 6781 6782 if (unlikely(val & NAPIF_STATE_MISSED)) { 6783 __napi_schedule(n); 6784 return false; 6785 } 6786 6787 if (timeout) 6788 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6789 HRTIMER_MODE_REL_PINNED); 6790 return ret; 6791 } 6792 EXPORT_SYMBOL(napi_complete_done); 6793 6794 static void skb_defer_free_flush(void) 6795 { 6796 struct llist_node *free_list; 6797 struct sk_buff *skb, *next; 6798 struct skb_defer_node *sdn; 6799 int node; 6800 6801 for_each_node(node) { 6802 sdn = this_cpu_ptr(net_hotdata.skb_defer_nodes) + node; 6803 6804 if (llist_empty(&sdn->defer_list)) 6805 continue; 6806 atomic_long_set(&sdn->defer_count, 0); 6807 free_list = llist_del_all(&sdn->defer_list); 6808 6809 llist_for_each_entry_safe(skb, next, free_list, ll_node) { 6810 prefetch(next); 6811 napi_consume_skb(skb, 1); 6812 } 6813 } 6814 } 6815 6816 #if defined(CONFIG_NET_RX_BUSY_POLL) 6817 6818 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6819 { 6820 if (!skip_schedule) { 6821 gro_normal_list(&napi->gro); 6822 __napi_schedule(napi); 6823 return; 6824 } 6825 6826 /* Flush too old packets. If HZ < 1000, flush all packets */ 6827 gro_flush_normal(&napi->gro, HZ >= 1000); 6828 6829 clear_bit(NAPI_STATE_SCHED, &napi->state); 6830 } 6831 6832 enum { 6833 NAPI_F_PREFER_BUSY_POLL = 1, 6834 NAPI_F_END_ON_RESCHED = 2, 6835 }; 6836 6837 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6838 unsigned flags, u16 budget) 6839 { 6840 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6841 bool skip_schedule = false; 6842 unsigned long timeout; 6843 int rc; 6844 6845 /* Busy polling means there is a high chance device driver hard irq 6846 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6847 * set in napi_schedule_prep(). 6848 * Since we are about to call napi->poll() once more, we can safely 6849 * clear NAPI_STATE_MISSED. 6850 * 6851 * Note: x86 could use a single "lock and ..." instruction 6852 * to perform these two clear_bit() 6853 */ 6854 clear_bit(NAPI_STATE_MISSED, &napi->state); 6855 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6856 6857 local_bh_disable(); 6858 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6859 6860 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6861 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6862 timeout = napi_get_gro_flush_timeout(napi); 6863 if (napi->defer_hard_irqs_count && timeout) { 6864 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6865 skip_schedule = true; 6866 } 6867 } 6868 6869 /* All we really want here is to re-enable device interrupts. 6870 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6871 */ 6872 rc = napi->poll(napi, budget); 6873 /* We can't gro_normal_list() here, because napi->poll() might have 6874 * rearmed the napi (napi_complete_done()) in which case it could 6875 * already be running on another CPU. 6876 */ 6877 trace_napi_poll(napi, rc, budget); 6878 netpoll_poll_unlock(have_poll_lock); 6879 if (rc == budget) 6880 __busy_poll_stop(napi, skip_schedule); 6881 bpf_net_ctx_clear(bpf_net_ctx); 6882 local_bh_enable(); 6883 } 6884 6885 static void __napi_busy_loop(unsigned int napi_id, 6886 bool (*loop_end)(void *, unsigned long), 6887 void *loop_end_arg, unsigned flags, u16 budget) 6888 { 6889 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6890 int (*napi_poll)(struct napi_struct *napi, int budget); 6891 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6892 void *have_poll_lock = NULL; 6893 struct napi_struct *napi; 6894 6895 WARN_ON_ONCE(!rcu_read_lock_held()); 6896 6897 restart: 6898 napi_poll = NULL; 6899 6900 napi = napi_by_id(napi_id); 6901 if (!napi) 6902 return; 6903 6904 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6905 preempt_disable(); 6906 for (;;) { 6907 int work = 0; 6908 6909 local_bh_disable(); 6910 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6911 if (!napi_poll) { 6912 unsigned long val = READ_ONCE(napi->state); 6913 6914 /* If multiple threads are competing for this napi, 6915 * we avoid dirtying napi->state as much as we can. 6916 */ 6917 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6918 NAPIF_STATE_IN_BUSY_POLL)) { 6919 if (flags & NAPI_F_PREFER_BUSY_POLL) 6920 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6921 goto count; 6922 } 6923 if (cmpxchg(&napi->state, val, 6924 val | NAPIF_STATE_IN_BUSY_POLL | 6925 NAPIF_STATE_SCHED) != val) { 6926 if (flags & NAPI_F_PREFER_BUSY_POLL) 6927 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6928 goto count; 6929 } 6930 have_poll_lock = netpoll_poll_lock(napi); 6931 napi_poll = napi->poll; 6932 } 6933 work = napi_poll(napi, budget); 6934 trace_napi_poll(napi, work, budget); 6935 gro_normal_list(&napi->gro); 6936 count: 6937 if (work > 0) 6938 __NET_ADD_STATS(dev_net(napi->dev), 6939 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6940 skb_defer_free_flush(); 6941 bpf_net_ctx_clear(bpf_net_ctx); 6942 local_bh_enable(); 6943 6944 if (!loop_end || loop_end(loop_end_arg, start_time)) 6945 break; 6946 6947 if (unlikely(need_resched())) { 6948 if (flags & NAPI_F_END_ON_RESCHED) 6949 break; 6950 if (napi_poll) 6951 busy_poll_stop(napi, have_poll_lock, flags, budget); 6952 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6953 preempt_enable(); 6954 rcu_read_unlock(); 6955 cond_resched(); 6956 rcu_read_lock(); 6957 if (loop_end(loop_end_arg, start_time)) 6958 return; 6959 goto restart; 6960 } 6961 cpu_relax(); 6962 } 6963 if (napi_poll) 6964 busy_poll_stop(napi, have_poll_lock, flags, budget); 6965 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6966 preempt_enable(); 6967 } 6968 6969 void napi_busy_loop_rcu(unsigned int napi_id, 6970 bool (*loop_end)(void *, unsigned long), 6971 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6972 { 6973 unsigned flags = NAPI_F_END_ON_RESCHED; 6974 6975 if (prefer_busy_poll) 6976 flags |= NAPI_F_PREFER_BUSY_POLL; 6977 6978 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6979 } 6980 6981 void napi_busy_loop(unsigned int napi_id, 6982 bool (*loop_end)(void *, unsigned long), 6983 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6984 { 6985 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6986 6987 rcu_read_lock(); 6988 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6989 rcu_read_unlock(); 6990 } 6991 EXPORT_SYMBOL(napi_busy_loop); 6992 6993 void napi_suspend_irqs(unsigned int napi_id) 6994 { 6995 struct napi_struct *napi; 6996 6997 rcu_read_lock(); 6998 napi = napi_by_id(napi_id); 6999 if (napi) { 7000 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 7001 7002 if (timeout) 7003 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 7004 HRTIMER_MODE_REL_PINNED); 7005 } 7006 rcu_read_unlock(); 7007 } 7008 7009 void napi_resume_irqs(unsigned int napi_id) 7010 { 7011 struct napi_struct *napi; 7012 7013 rcu_read_lock(); 7014 napi = napi_by_id(napi_id); 7015 if (napi) { 7016 /* If irq_suspend_timeout is set to 0 between the call to 7017 * napi_suspend_irqs and now, the original value still 7018 * determines the safety timeout as intended and napi_watchdog 7019 * will resume irq processing. 7020 */ 7021 if (napi_get_irq_suspend_timeout(napi)) { 7022 local_bh_disable(); 7023 napi_schedule(napi); 7024 local_bh_enable(); 7025 } 7026 } 7027 rcu_read_unlock(); 7028 } 7029 7030 #endif /* CONFIG_NET_RX_BUSY_POLL */ 7031 7032 static void __napi_hash_add_with_id(struct napi_struct *napi, 7033 unsigned int napi_id) 7034 { 7035 napi->gro.cached_napi_id = napi_id; 7036 7037 WRITE_ONCE(napi->napi_id, napi_id); 7038 hlist_add_head_rcu(&napi->napi_hash_node, 7039 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 7040 } 7041 7042 static void napi_hash_add_with_id(struct napi_struct *napi, 7043 unsigned int napi_id) 7044 { 7045 unsigned long flags; 7046 7047 spin_lock_irqsave(&napi_hash_lock, flags); 7048 WARN_ON_ONCE(napi_by_id(napi_id)); 7049 __napi_hash_add_with_id(napi, napi_id); 7050 spin_unlock_irqrestore(&napi_hash_lock, flags); 7051 } 7052 7053 static void napi_hash_add(struct napi_struct *napi) 7054 { 7055 unsigned long flags; 7056 7057 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 7058 return; 7059 7060 spin_lock_irqsave(&napi_hash_lock, flags); 7061 7062 /* 0..NR_CPUS range is reserved for sender_cpu use */ 7063 do { 7064 if (unlikely(!napi_id_valid(++napi_gen_id))) 7065 napi_gen_id = MIN_NAPI_ID; 7066 } while (napi_by_id(napi_gen_id)); 7067 7068 __napi_hash_add_with_id(napi, napi_gen_id); 7069 7070 spin_unlock_irqrestore(&napi_hash_lock, flags); 7071 } 7072 7073 /* Warning : caller is responsible to make sure rcu grace period 7074 * is respected before freeing memory containing @napi 7075 */ 7076 static void napi_hash_del(struct napi_struct *napi) 7077 { 7078 unsigned long flags; 7079 7080 spin_lock_irqsave(&napi_hash_lock, flags); 7081 7082 hlist_del_init_rcu(&napi->napi_hash_node); 7083 7084 spin_unlock_irqrestore(&napi_hash_lock, flags); 7085 } 7086 7087 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 7088 { 7089 struct napi_struct *napi; 7090 7091 napi = container_of(timer, struct napi_struct, timer); 7092 7093 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 7094 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 7095 */ 7096 if (!napi_disable_pending(napi) && 7097 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 7098 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 7099 __napi_schedule_irqoff(napi); 7100 } 7101 7102 return HRTIMER_NORESTART; 7103 } 7104 7105 static void napi_stop_kthread(struct napi_struct *napi) 7106 { 7107 unsigned long val, new; 7108 7109 /* Wait until the napi STATE_THREADED is unset. */ 7110 while (true) { 7111 val = READ_ONCE(napi->state); 7112 7113 /* If napi kthread own this napi or the napi is idle, 7114 * STATE_THREADED can be unset here. 7115 */ 7116 if ((val & NAPIF_STATE_SCHED_THREADED) || 7117 !(val & NAPIF_STATE_SCHED)) { 7118 new = val & (~(NAPIF_STATE_THREADED | 7119 NAPIF_STATE_THREADED_BUSY_POLL)); 7120 } else { 7121 msleep(20); 7122 continue; 7123 } 7124 7125 if (try_cmpxchg(&napi->state, &val, new)) 7126 break; 7127 } 7128 7129 /* Once STATE_THREADED is unset, wait for SCHED_THREADED to be unset by 7130 * the kthread. 7131 */ 7132 while (true) { 7133 if (!test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) 7134 break; 7135 7136 msleep(20); 7137 } 7138 7139 kthread_stop(napi->thread); 7140 napi->thread = NULL; 7141 } 7142 7143 static void napi_set_threaded_state(struct napi_struct *napi, 7144 enum netdev_napi_threaded threaded_mode) 7145 { 7146 bool threaded = threaded_mode != NETDEV_NAPI_THREADED_DISABLED; 7147 bool busy_poll = threaded_mode == NETDEV_NAPI_THREADED_BUSY_POLL; 7148 7149 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 7150 assign_bit(NAPI_STATE_THREADED_BUSY_POLL, &napi->state, busy_poll); 7151 } 7152 7153 int napi_set_threaded(struct napi_struct *napi, 7154 enum netdev_napi_threaded threaded) 7155 { 7156 if (threaded) { 7157 if (!napi->thread) { 7158 int err = napi_kthread_create(napi); 7159 7160 if (err) 7161 return err; 7162 } 7163 } 7164 7165 if (napi->config) 7166 napi->config->threaded = threaded; 7167 7168 /* Setting/unsetting threaded mode on a napi might not immediately 7169 * take effect, if the current napi instance is actively being 7170 * polled. In this case, the switch between threaded mode and 7171 * softirq mode will happen in the next round of napi_schedule(). 7172 * This should not cause hiccups/stalls to the live traffic. 7173 */ 7174 if (!threaded && napi->thread) { 7175 napi_stop_kthread(napi); 7176 } else { 7177 /* Make sure kthread is created before THREADED bit is set. */ 7178 smp_mb__before_atomic(); 7179 napi_set_threaded_state(napi, threaded); 7180 } 7181 7182 return 0; 7183 } 7184 7185 int netif_set_threaded(struct net_device *dev, 7186 enum netdev_napi_threaded threaded) 7187 { 7188 struct napi_struct *napi; 7189 int i, err = 0; 7190 7191 netdev_assert_locked_or_invisible(dev); 7192 7193 if (threaded) { 7194 list_for_each_entry(napi, &dev->napi_list, dev_list) { 7195 if (!napi->thread) { 7196 err = napi_kthread_create(napi); 7197 if (err) { 7198 threaded = NETDEV_NAPI_THREADED_DISABLED; 7199 break; 7200 } 7201 } 7202 } 7203 } 7204 7205 WRITE_ONCE(dev->threaded, threaded); 7206 7207 /* The error should not occur as the kthreads are already created. */ 7208 list_for_each_entry(napi, &dev->napi_list, dev_list) 7209 WARN_ON_ONCE(napi_set_threaded(napi, threaded)); 7210 7211 /* Override the config for all NAPIs even if currently not listed */ 7212 for (i = 0; i < dev->num_napi_configs; i++) 7213 dev->napi_config[i].threaded = threaded; 7214 7215 return err; 7216 } 7217 7218 /** 7219 * netif_threaded_enable() - enable threaded NAPIs 7220 * @dev: net_device instance 7221 * 7222 * Enable threaded mode for the NAPI instances of the device. This may be useful 7223 * for devices where multiple NAPI instances get scheduled by a single 7224 * interrupt. Threaded NAPI allows moving the NAPI processing to cores other 7225 * than the core where IRQ is mapped. 7226 * 7227 * This function should be called before @dev is registered. 7228 */ 7229 void netif_threaded_enable(struct net_device *dev) 7230 { 7231 WARN_ON_ONCE(netif_set_threaded(dev, NETDEV_NAPI_THREADED_ENABLED)); 7232 } 7233 EXPORT_SYMBOL(netif_threaded_enable); 7234 7235 /** 7236 * netif_queue_set_napi - Associate queue with the napi 7237 * @dev: device to which NAPI and queue belong 7238 * @queue_index: Index of queue 7239 * @type: queue type as RX or TX 7240 * @napi: NAPI context, pass NULL to clear previously set NAPI 7241 * 7242 * Set queue with its corresponding napi context. This should be done after 7243 * registering the NAPI handler for the queue-vector and the queues have been 7244 * mapped to the corresponding interrupt vector. 7245 */ 7246 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 7247 enum netdev_queue_type type, struct napi_struct *napi) 7248 { 7249 struct netdev_rx_queue *rxq; 7250 struct netdev_queue *txq; 7251 7252 if (WARN_ON_ONCE(napi && !napi->dev)) 7253 return; 7254 netdev_ops_assert_locked_or_invisible(dev); 7255 7256 switch (type) { 7257 case NETDEV_QUEUE_TYPE_RX: 7258 rxq = __netif_get_rx_queue(dev, queue_index); 7259 rxq->napi = napi; 7260 return; 7261 case NETDEV_QUEUE_TYPE_TX: 7262 txq = netdev_get_tx_queue(dev, queue_index); 7263 txq->napi = napi; 7264 return; 7265 default: 7266 return; 7267 } 7268 } 7269 EXPORT_SYMBOL(netif_queue_set_napi); 7270 7271 static void 7272 netif_napi_irq_notify(struct irq_affinity_notify *notify, 7273 const cpumask_t *mask) 7274 { 7275 struct napi_struct *napi = 7276 container_of(notify, struct napi_struct, notify); 7277 #ifdef CONFIG_RFS_ACCEL 7278 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7279 int err; 7280 #endif 7281 7282 if (napi->config && napi->dev->irq_affinity_auto) 7283 cpumask_copy(&napi->config->affinity_mask, mask); 7284 7285 #ifdef CONFIG_RFS_ACCEL 7286 if (napi->dev->rx_cpu_rmap_auto) { 7287 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 7288 if (err) 7289 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 7290 err); 7291 } 7292 #endif 7293 } 7294 7295 #ifdef CONFIG_RFS_ACCEL 7296 static void netif_napi_affinity_release(struct kref *ref) 7297 { 7298 struct napi_struct *napi = 7299 container_of(ref, struct napi_struct, notify.kref); 7300 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7301 7302 netdev_assert_locked(napi->dev); 7303 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 7304 &napi->state)); 7305 7306 if (!napi->dev->rx_cpu_rmap_auto) 7307 return; 7308 rmap->obj[napi->napi_rmap_idx] = NULL; 7309 napi->napi_rmap_idx = -1; 7310 cpu_rmap_put(rmap); 7311 } 7312 7313 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7314 { 7315 if (dev->rx_cpu_rmap_auto) 7316 return 0; 7317 7318 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 7319 if (!dev->rx_cpu_rmap) 7320 return -ENOMEM; 7321 7322 dev->rx_cpu_rmap_auto = true; 7323 return 0; 7324 } 7325 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7326 7327 static void netif_del_cpu_rmap(struct net_device *dev) 7328 { 7329 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 7330 7331 if (!dev->rx_cpu_rmap_auto) 7332 return; 7333 7334 /* Free the rmap */ 7335 cpu_rmap_put(rmap); 7336 dev->rx_cpu_rmap = NULL; 7337 dev->rx_cpu_rmap_auto = false; 7338 } 7339 7340 #else 7341 static void netif_napi_affinity_release(struct kref *ref) 7342 { 7343 } 7344 7345 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7346 { 7347 return 0; 7348 } 7349 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7350 7351 static void netif_del_cpu_rmap(struct net_device *dev) 7352 { 7353 } 7354 #endif 7355 7356 void netif_set_affinity_auto(struct net_device *dev) 7357 { 7358 unsigned int i, maxqs, numa; 7359 7360 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 7361 numa = dev_to_node(&dev->dev); 7362 7363 for (i = 0; i < maxqs; i++) 7364 cpumask_set_cpu(cpumask_local_spread(i, numa), 7365 &dev->napi_config[i].affinity_mask); 7366 7367 dev->irq_affinity_auto = true; 7368 } 7369 EXPORT_SYMBOL(netif_set_affinity_auto); 7370 7371 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 7372 { 7373 int rc; 7374 7375 netdev_assert_locked_or_invisible(napi->dev); 7376 7377 if (napi->irq == irq) 7378 return; 7379 7380 /* Remove existing resources */ 7381 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7382 irq_set_affinity_notifier(napi->irq, NULL); 7383 7384 napi->irq = irq; 7385 if (irq < 0 || 7386 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7387 return; 7388 7389 /* Abort for buggy drivers */ 7390 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7391 return; 7392 7393 #ifdef CONFIG_RFS_ACCEL 7394 if (napi->dev->rx_cpu_rmap_auto) { 7395 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7396 if (rc < 0) 7397 return; 7398 7399 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7400 napi->napi_rmap_idx = rc; 7401 } 7402 #endif 7403 7404 /* Use core IRQ notifier */ 7405 napi->notify.notify = netif_napi_irq_notify; 7406 napi->notify.release = netif_napi_affinity_release; 7407 rc = irq_set_affinity_notifier(irq, &napi->notify); 7408 if (rc) { 7409 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7410 rc); 7411 goto put_rmap; 7412 } 7413 7414 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7415 return; 7416 7417 put_rmap: 7418 #ifdef CONFIG_RFS_ACCEL 7419 if (napi->dev->rx_cpu_rmap_auto) { 7420 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7421 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7422 napi->napi_rmap_idx = -1; 7423 } 7424 #endif 7425 napi->notify.notify = NULL; 7426 napi->notify.release = NULL; 7427 } 7428 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7429 7430 static void napi_restore_config(struct napi_struct *n) 7431 { 7432 n->defer_hard_irqs = n->config->defer_hard_irqs; 7433 n->gro_flush_timeout = n->config->gro_flush_timeout; 7434 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7435 7436 if (n->dev->irq_affinity_auto && 7437 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7438 irq_set_affinity(n->irq, &n->config->affinity_mask); 7439 7440 /* a NAPI ID might be stored in the config, if so use it. if not, use 7441 * napi_hash_add to generate one for us. 7442 */ 7443 if (n->config->napi_id) { 7444 napi_hash_add_with_id(n, n->config->napi_id); 7445 } else { 7446 napi_hash_add(n); 7447 n->config->napi_id = n->napi_id; 7448 } 7449 7450 WARN_ON_ONCE(napi_set_threaded(n, n->config->threaded)); 7451 } 7452 7453 static void napi_save_config(struct napi_struct *n) 7454 { 7455 n->config->defer_hard_irqs = n->defer_hard_irqs; 7456 n->config->gro_flush_timeout = n->gro_flush_timeout; 7457 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7458 napi_hash_del(n); 7459 } 7460 7461 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7462 * inherit an existing ID try to insert it at the right position. 7463 */ 7464 static void 7465 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7466 { 7467 unsigned int new_id, pos_id; 7468 struct list_head *higher; 7469 struct napi_struct *pos; 7470 7471 new_id = UINT_MAX; 7472 if (napi->config && napi->config->napi_id) 7473 new_id = napi->config->napi_id; 7474 7475 higher = &dev->napi_list; 7476 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7477 if (napi_id_valid(pos->napi_id)) 7478 pos_id = pos->napi_id; 7479 else if (pos->config) 7480 pos_id = pos->config->napi_id; 7481 else 7482 pos_id = UINT_MAX; 7483 7484 if (pos_id <= new_id) 7485 break; 7486 higher = &pos->dev_list; 7487 } 7488 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7489 } 7490 7491 /* Double check that napi_get_frags() allocates skbs with 7492 * skb->head being backed by slab, not a page fragment. 7493 * This is to make sure bug fixed in 3226b158e67c 7494 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7495 * does not accidentally come back. 7496 */ 7497 static void napi_get_frags_check(struct napi_struct *napi) 7498 { 7499 struct sk_buff *skb; 7500 7501 local_bh_disable(); 7502 skb = napi_get_frags(napi); 7503 WARN_ON_ONCE(skb && skb->head_frag); 7504 napi_free_frags(napi); 7505 local_bh_enable(); 7506 } 7507 7508 void netif_napi_add_weight_locked(struct net_device *dev, 7509 struct napi_struct *napi, 7510 int (*poll)(struct napi_struct *, int), 7511 int weight) 7512 { 7513 netdev_assert_locked(dev); 7514 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7515 return; 7516 7517 INIT_LIST_HEAD(&napi->poll_list); 7518 INIT_HLIST_NODE(&napi->napi_hash_node); 7519 hrtimer_setup(&napi->timer, napi_watchdog, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7520 gro_init(&napi->gro); 7521 napi->skb = NULL; 7522 napi->poll = poll; 7523 if (weight > NAPI_POLL_WEIGHT) 7524 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7525 weight); 7526 napi->weight = weight; 7527 napi->dev = dev; 7528 #ifdef CONFIG_NETPOLL 7529 napi->poll_owner = -1; 7530 #endif 7531 napi->list_owner = -1; 7532 set_bit(NAPI_STATE_SCHED, &napi->state); 7533 set_bit(NAPI_STATE_NPSVC, &napi->state); 7534 netif_napi_dev_list_add(dev, napi); 7535 7536 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7537 * configuration will be loaded in napi_enable 7538 */ 7539 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7540 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7541 7542 napi_get_frags_check(napi); 7543 /* Create kthread for this napi if dev->threaded is set. 7544 * Clear dev->threaded if kthread creation failed so that 7545 * threaded mode will not be enabled in napi_enable(). 7546 */ 7547 if (napi_get_threaded_config(dev, napi)) 7548 if (napi_kthread_create(napi)) 7549 dev->threaded = NETDEV_NAPI_THREADED_DISABLED; 7550 netif_napi_set_irq_locked(napi, -1); 7551 } 7552 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7553 7554 void napi_disable_locked(struct napi_struct *n) 7555 { 7556 unsigned long val, new; 7557 7558 might_sleep(); 7559 netdev_assert_locked(n->dev); 7560 7561 set_bit(NAPI_STATE_DISABLE, &n->state); 7562 7563 val = READ_ONCE(n->state); 7564 do { 7565 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7566 usleep_range(20, 200); 7567 val = READ_ONCE(n->state); 7568 } 7569 7570 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7571 new &= ~(NAPIF_STATE_THREADED | 7572 NAPIF_STATE_THREADED_BUSY_POLL | 7573 NAPIF_STATE_PREFER_BUSY_POLL); 7574 } while (!try_cmpxchg(&n->state, &val, new)); 7575 7576 hrtimer_cancel(&n->timer); 7577 7578 if (n->config) 7579 napi_save_config(n); 7580 else 7581 napi_hash_del(n); 7582 7583 clear_bit(NAPI_STATE_DISABLE, &n->state); 7584 } 7585 EXPORT_SYMBOL(napi_disable_locked); 7586 7587 /** 7588 * napi_disable() - prevent NAPI from scheduling 7589 * @n: NAPI context 7590 * 7591 * Stop NAPI from being scheduled on this context. 7592 * Waits till any outstanding processing completes. 7593 * Takes netdev_lock() for associated net_device. 7594 */ 7595 void napi_disable(struct napi_struct *n) 7596 { 7597 netdev_lock(n->dev); 7598 napi_disable_locked(n); 7599 netdev_unlock(n->dev); 7600 } 7601 EXPORT_SYMBOL(napi_disable); 7602 7603 void napi_enable_locked(struct napi_struct *n) 7604 { 7605 unsigned long new, val = READ_ONCE(n->state); 7606 7607 if (n->config) 7608 napi_restore_config(n); 7609 else 7610 napi_hash_add(n); 7611 7612 do { 7613 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7614 7615 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7616 if (n->dev->threaded && n->thread) 7617 new |= NAPIF_STATE_THREADED; 7618 } while (!try_cmpxchg(&n->state, &val, new)); 7619 } 7620 EXPORT_SYMBOL(napi_enable_locked); 7621 7622 /** 7623 * napi_enable() - enable NAPI scheduling 7624 * @n: NAPI context 7625 * 7626 * Enable scheduling of a NAPI instance. 7627 * Must be paired with napi_disable(). 7628 * Takes netdev_lock() for associated net_device. 7629 */ 7630 void napi_enable(struct napi_struct *n) 7631 { 7632 netdev_lock(n->dev); 7633 napi_enable_locked(n); 7634 netdev_unlock(n->dev); 7635 } 7636 EXPORT_SYMBOL(napi_enable); 7637 7638 /* Must be called in process context */ 7639 void __netif_napi_del_locked(struct napi_struct *napi) 7640 { 7641 netdev_assert_locked(napi->dev); 7642 7643 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7644 return; 7645 7646 /* Make sure NAPI is disabled (or was never enabled). */ 7647 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7648 7649 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7650 irq_set_affinity_notifier(napi->irq, NULL); 7651 7652 if (napi->config) { 7653 napi->index = -1; 7654 napi->config = NULL; 7655 } 7656 7657 list_del_rcu(&napi->dev_list); 7658 napi_free_frags(napi); 7659 7660 gro_cleanup(&napi->gro); 7661 7662 if (napi->thread) { 7663 kthread_stop(napi->thread); 7664 napi->thread = NULL; 7665 } 7666 } 7667 EXPORT_SYMBOL(__netif_napi_del_locked); 7668 7669 static int __napi_poll(struct napi_struct *n, bool *repoll) 7670 { 7671 int work, weight; 7672 7673 weight = n->weight; 7674 7675 /* This NAPI_STATE_SCHED test is for avoiding a race 7676 * with netpoll's poll_napi(). Only the entity which 7677 * obtains the lock and sees NAPI_STATE_SCHED set will 7678 * actually make the ->poll() call. Therefore we avoid 7679 * accidentally calling ->poll() when NAPI is not scheduled. 7680 */ 7681 work = 0; 7682 if (napi_is_scheduled(n)) { 7683 work = n->poll(n, weight); 7684 trace_napi_poll(n, work, weight); 7685 7686 xdp_do_check_flushed(n); 7687 } 7688 7689 if (unlikely(work > weight)) 7690 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7691 n->poll, work, weight); 7692 7693 if (likely(work < weight)) 7694 return work; 7695 7696 /* Drivers must not modify the NAPI state if they 7697 * consume the entire weight. In such cases this code 7698 * still "owns" the NAPI instance and therefore can 7699 * move the instance around on the list at-will. 7700 */ 7701 if (unlikely(napi_disable_pending(n))) { 7702 napi_complete(n); 7703 return work; 7704 } 7705 7706 /* The NAPI context has more processing work, but busy-polling 7707 * is preferred. Exit early. 7708 */ 7709 if (napi_prefer_busy_poll(n)) { 7710 if (napi_complete_done(n, work)) { 7711 /* If timeout is not set, we need to make sure 7712 * that the NAPI is re-scheduled. 7713 */ 7714 napi_schedule(n); 7715 } 7716 return work; 7717 } 7718 7719 /* Flush too old packets. If HZ < 1000, flush all packets */ 7720 gro_flush_normal(&n->gro, HZ >= 1000); 7721 7722 /* Some drivers may have called napi_schedule 7723 * prior to exhausting their budget. 7724 */ 7725 if (unlikely(!list_empty(&n->poll_list))) { 7726 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7727 n->dev ? n->dev->name : "backlog"); 7728 return work; 7729 } 7730 7731 *repoll = true; 7732 7733 return work; 7734 } 7735 7736 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7737 { 7738 bool do_repoll = false; 7739 void *have; 7740 int work; 7741 7742 list_del_init(&n->poll_list); 7743 7744 have = netpoll_poll_lock(n); 7745 7746 work = __napi_poll(n, &do_repoll); 7747 7748 if (do_repoll) { 7749 #if defined(CONFIG_DEBUG_NET) 7750 if (unlikely(!napi_is_scheduled(n))) 7751 pr_crit("repoll requested for device %s %ps but napi is not scheduled.\n", 7752 n->dev->name, n->poll); 7753 #endif 7754 list_add_tail(&n->poll_list, repoll); 7755 } 7756 netpoll_poll_unlock(have); 7757 7758 return work; 7759 } 7760 7761 static int napi_thread_wait(struct napi_struct *napi) 7762 { 7763 set_current_state(TASK_INTERRUPTIBLE); 7764 7765 while (!kthread_should_stop()) { 7766 /* Testing SCHED_THREADED bit here to make sure the current 7767 * kthread owns this napi and could poll on this napi. 7768 * Testing SCHED bit is not enough because SCHED bit might be 7769 * set by some other busy poll thread or by napi_disable(). 7770 */ 7771 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7772 WARN_ON(!list_empty(&napi->poll_list)); 7773 __set_current_state(TASK_RUNNING); 7774 return 0; 7775 } 7776 7777 schedule(); 7778 set_current_state(TASK_INTERRUPTIBLE); 7779 } 7780 __set_current_state(TASK_RUNNING); 7781 7782 return -1; 7783 } 7784 7785 static void napi_threaded_poll_loop(struct napi_struct *napi, bool busy_poll) 7786 { 7787 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7788 struct softnet_data *sd; 7789 unsigned long last_qs = jiffies; 7790 7791 for (;;) { 7792 bool repoll = false; 7793 void *have; 7794 7795 local_bh_disable(); 7796 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7797 7798 sd = this_cpu_ptr(&softnet_data); 7799 sd->in_napi_threaded_poll = true; 7800 7801 have = netpoll_poll_lock(napi); 7802 __napi_poll(napi, &repoll); 7803 netpoll_poll_unlock(have); 7804 7805 sd->in_napi_threaded_poll = false; 7806 barrier(); 7807 7808 if (sd_has_rps_ipi_waiting(sd)) { 7809 local_irq_disable(); 7810 net_rps_action_and_irq_enable(sd); 7811 } 7812 skb_defer_free_flush(); 7813 bpf_net_ctx_clear(bpf_net_ctx); 7814 7815 /* When busy poll is enabled, the old packets are not flushed in 7816 * napi_complete_done. So flush them here. 7817 */ 7818 if (busy_poll) 7819 gro_flush_normal(&napi->gro, HZ >= 1000); 7820 local_bh_enable(); 7821 7822 /* Call cond_resched here to avoid watchdog warnings. */ 7823 if (repoll || busy_poll) { 7824 rcu_softirq_qs_periodic(last_qs); 7825 cond_resched(); 7826 } 7827 7828 if (!repoll) 7829 break; 7830 } 7831 } 7832 7833 static int napi_threaded_poll(void *data) 7834 { 7835 struct napi_struct *napi = data; 7836 bool want_busy_poll; 7837 bool in_busy_poll; 7838 unsigned long val; 7839 7840 while (!napi_thread_wait(napi)) { 7841 val = READ_ONCE(napi->state); 7842 7843 want_busy_poll = val & NAPIF_STATE_THREADED_BUSY_POLL; 7844 in_busy_poll = val & NAPIF_STATE_IN_BUSY_POLL; 7845 7846 if (unlikely(val & NAPIF_STATE_DISABLE)) 7847 want_busy_poll = false; 7848 7849 if (want_busy_poll != in_busy_poll) 7850 assign_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state, 7851 want_busy_poll); 7852 7853 napi_threaded_poll_loop(napi, want_busy_poll); 7854 } 7855 7856 return 0; 7857 } 7858 7859 static __latent_entropy void net_rx_action(void) 7860 { 7861 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7862 unsigned long time_limit = jiffies + 7863 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7864 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7865 int budget = READ_ONCE(net_hotdata.netdev_budget); 7866 LIST_HEAD(list); 7867 LIST_HEAD(repoll); 7868 7869 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7870 start: 7871 sd->in_net_rx_action = true; 7872 local_irq_disable(); 7873 list_splice_init(&sd->poll_list, &list); 7874 local_irq_enable(); 7875 7876 for (;;) { 7877 struct napi_struct *n; 7878 7879 skb_defer_free_flush(); 7880 7881 if (list_empty(&list)) { 7882 if (list_empty(&repoll)) { 7883 sd->in_net_rx_action = false; 7884 barrier(); 7885 /* We need to check if ____napi_schedule() 7886 * had refilled poll_list while 7887 * sd->in_net_rx_action was true. 7888 */ 7889 if (!list_empty(&sd->poll_list)) 7890 goto start; 7891 if (!sd_has_rps_ipi_waiting(sd)) 7892 goto end; 7893 } 7894 break; 7895 } 7896 7897 n = list_first_entry(&list, struct napi_struct, poll_list); 7898 budget -= napi_poll(n, &repoll); 7899 7900 /* If softirq window is exhausted then punt. 7901 * Allow this to run for 2 jiffies since which will allow 7902 * an average latency of 1.5/HZ. 7903 */ 7904 if (unlikely(budget <= 0 || 7905 time_after_eq(jiffies, time_limit))) { 7906 /* Pairs with READ_ONCE() in softnet_seq_show() */ 7907 WRITE_ONCE(sd->time_squeeze, sd->time_squeeze + 1); 7908 break; 7909 } 7910 } 7911 7912 local_irq_disable(); 7913 7914 list_splice_tail_init(&sd->poll_list, &list); 7915 list_splice_tail(&repoll, &list); 7916 list_splice(&list, &sd->poll_list); 7917 if (!list_empty(&sd->poll_list)) 7918 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7919 else 7920 sd->in_net_rx_action = false; 7921 7922 net_rps_action_and_irq_enable(sd); 7923 end: 7924 bpf_net_ctx_clear(bpf_net_ctx); 7925 } 7926 7927 struct netdev_adjacent { 7928 struct net_device *dev; 7929 netdevice_tracker dev_tracker; 7930 7931 /* upper master flag, there can only be one master device per list */ 7932 bool master; 7933 7934 /* lookup ignore flag */ 7935 bool ignore; 7936 7937 /* counter for the number of times this device was added to us */ 7938 u16 ref_nr; 7939 7940 /* private field for the users */ 7941 void *private; 7942 7943 struct list_head list; 7944 struct rcu_head rcu; 7945 }; 7946 7947 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7948 struct list_head *adj_list) 7949 { 7950 struct netdev_adjacent *adj; 7951 7952 list_for_each_entry(adj, adj_list, list) { 7953 if (adj->dev == adj_dev) 7954 return adj; 7955 } 7956 return NULL; 7957 } 7958 7959 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7960 struct netdev_nested_priv *priv) 7961 { 7962 struct net_device *dev = (struct net_device *)priv->data; 7963 7964 return upper_dev == dev; 7965 } 7966 7967 /** 7968 * netdev_has_upper_dev - Check if device is linked to an upper device 7969 * @dev: device 7970 * @upper_dev: upper device to check 7971 * 7972 * Find out if a device is linked to specified upper device and return true 7973 * in case it is. Note that this checks only immediate upper device, 7974 * not through a complete stack of devices. The caller must hold the RTNL lock. 7975 */ 7976 bool netdev_has_upper_dev(struct net_device *dev, 7977 struct net_device *upper_dev) 7978 { 7979 struct netdev_nested_priv priv = { 7980 .data = (void *)upper_dev, 7981 }; 7982 7983 ASSERT_RTNL(); 7984 7985 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7986 &priv); 7987 } 7988 EXPORT_SYMBOL(netdev_has_upper_dev); 7989 7990 /** 7991 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7992 * @dev: device 7993 * @upper_dev: upper device to check 7994 * 7995 * Find out if a device is linked to specified upper device and return true 7996 * in case it is. Note that this checks the entire upper device chain. 7997 * The caller must hold rcu lock. 7998 */ 7999 8000 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 8001 struct net_device *upper_dev) 8002 { 8003 struct netdev_nested_priv priv = { 8004 .data = (void *)upper_dev, 8005 }; 8006 8007 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 8008 &priv); 8009 } 8010 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 8011 8012 /** 8013 * netdev_has_any_upper_dev - Check if device is linked to some device 8014 * @dev: device 8015 * 8016 * Find out if a device is linked to an upper device and return true in case 8017 * it is. The caller must hold the RTNL lock. 8018 */ 8019 bool netdev_has_any_upper_dev(struct net_device *dev) 8020 { 8021 ASSERT_RTNL(); 8022 8023 return !list_empty(&dev->adj_list.upper); 8024 } 8025 EXPORT_SYMBOL(netdev_has_any_upper_dev); 8026 8027 /** 8028 * netdev_master_upper_dev_get - Get master upper device 8029 * @dev: device 8030 * 8031 * Find a master upper device and return pointer to it or NULL in case 8032 * it's not there. The caller must hold the RTNL lock. 8033 */ 8034 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 8035 { 8036 struct netdev_adjacent *upper; 8037 8038 ASSERT_RTNL(); 8039 8040 if (list_empty(&dev->adj_list.upper)) 8041 return NULL; 8042 8043 upper = list_first_entry(&dev->adj_list.upper, 8044 struct netdev_adjacent, list); 8045 if (likely(upper->master)) 8046 return upper->dev; 8047 return NULL; 8048 } 8049 EXPORT_SYMBOL(netdev_master_upper_dev_get); 8050 8051 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 8052 { 8053 struct netdev_adjacent *upper; 8054 8055 ASSERT_RTNL(); 8056 8057 if (list_empty(&dev->adj_list.upper)) 8058 return NULL; 8059 8060 upper = list_first_entry(&dev->adj_list.upper, 8061 struct netdev_adjacent, list); 8062 if (likely(upper->master) && !upper->ignore) 8063 return upper->dev; 8064 return NULL; 8065 } 8066 8067 /** 8068 * netdev_has_any_lower_dev - Check if device is linked to some device 8069 * @dev: device 8070 * 8071 * Find out if a device is linked to a lower device and return true in case 8072 * it is. The caller must hold the RTNL lock. 8073 */ 8074 static bool netdev_has_any_lower_dev(struct net_device *dev) 8075 { 8076 ASSERT_RTNL(); 8077 8078 return !list_empty(&dev->adj_list.lower); 8079 } 8080 8081 void *netdev_adjacent_get_private(struct list_head *adj_list) 8082 { 8083 struct netdev_adjacent *adj; 8084 8085 adj = list_entry(adj_list, struct netdev_adjacent, list); 8086 8087 return adj->private; 8088 } 8089 EXPORT_SYMBOL(netdev_adjacent_get_private); 8090 8091 /** 8092 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 8093 * @dev: device 8094 * @iter: list_head ** of the current position 8095 * 8096 * Gets the next device from the dev's upper list, starting from iter 8097 * position. The caller must hold RCU read lock. 8098 */ 8099 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 8100 struct list_head **iter) 8101 { 8102 struct netdev_adjacent *upper; 8103 8104 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 8105 8106 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8107 8108 if (&upper->list == &dev->adj_list.upper) 8109 return NULL; 8110 8111 *iter = &upper->list; 8112 8113 return upper->dev; 8114 } 8115 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 8116 8117 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 8118 struct list_head **iter, 8119 bool *ignore) 8120 { 8121 struct netdev_adjacent *upper; 8122 8123 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 8124 8125 if (&upper->list == &dev->adj_list.upper) 8126 return NULL; 8127 8128 *iter = &upper->list; 8129 *ignore = upper->ignore; 8130 8131 return upper->dev; 8132 } 8133 8134 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 8135 struct list_head **iter) 8136 { 8137 struct netdev_adjacent *upper; 8138 8139 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 8140 8141 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8142 8143 if (&upper->list == &dev->adj_list.upper) 8144 return NULL; 8145 8146 *iter = &upper->list; 8147 8148 return upper->dev; 8149 } 8150 8151 static int __netdev_walk_all_upper_dev(struct net_device *dev, 8152 int (*fn)(struct net_device *dev, 8153 struct netdev_nested_priv *priv), 8154 struct netdev_nested_priv *priv) 8155 { 8156 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8157 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8158 int ret, cur = 0; 8159 bool ignore; 8160 8161 now = dev; 8162 iter = &dev->adj_list.upper; 8163 8164 while (1) { 8165 if (now != dev) { 8166 ret = fn(now, priv); 8167 if (ret) 8168 return ret; 8169 } 8170 8171 next = NULL; 8172 while (1) { 8173 udev = __netdev_next_upper_dev(now, &iter, &ignore); 8174 if (!udev) 8175 break; 8176 if (ignore) 8177 continue; 8178 8179 next = udev; 8180 niter = &udev->adj_list.upper; 8181 dev_stack[cur] = now; 8182 iter_stack[cur++] = iter; 8183 break; 8184 } 8185 8186 if (!next) { 8187 if (!cur) 8188 return 0; 8189 next = dev_stack[--cur]; 8190 niter = iter_stack[cur]; 8191 } 8192 8193 now = next; 8194 iter = niter; 8195 } 8196 8197 return 0; 8198 } 8199 8200 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 8201 int (*fn)(struct net_device *dev, 8202 struct netdev_nested_priv *priv), 8203 struct netdev_nested_priv *priv) 8204 { 8205 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8206 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8207 int ret, cur = 0; 8208 8209 now = dev; 8210 iter = &dev->adj_list.upper; 8211 8212 while (1) { 8213 if (now != dev) { 8214 ret = fn(now, priv); 8215 if (ret) 8216 return ret; 8217 } 8218 8219 next = NULL; 8220 while (1) { 8221 udev = netdev_next_upper_dev_rcu(now, &iter); 8222 if (!udev) 8223 break; 8224 8225 next = udev; 8226 niter = &udev->adj_list.upper; 8227 dev_stack[cur] = now; 8228 iter_stack[cur++] = iter; 8229 break; 8230 } 8231 8232 if (!next) { 8233 if (!cur) 8234 return 0; 8235 next = dev_stack[--cur]; 8236 niter = iter_stack[cur]; 8237 } 8238 8239 now = next; 8240 iter = niter; 8241 } 8242 8243 return 0; 8244 } 8245 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 8246 8247 static bool __netdev_has_upper_dev(struct net_device *dev, 8248 struct net_device *upper_dev) 8249 { 8250 struct netdev_nested_priv priv = { 8251 .flags = 0, 8252 .data = (void *)upper_dev, 8253 }; 8254 8255 ASSERT_RTNL(); 8256 8257 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 8258 &priv); 8259 } 8260 8261 /** 8262 * netdev_lower_get_next_private - Get the next ->private from the 8263 * lower neighbour list 8264 * @dev: device 8265 * @iter: list_head ** of the current position 8266 * 8267 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8268 * list, starting from iter position. The caller must hold either hold the 8269 * RTNL lock or its own locking that guarantees that the neighbour lower 8270 * list will remain unchanged. 8271 */ 8272 void *netdev_lower_get_next_private(struct net_device *dev, 8273 struct list_head **iter) 8274 { 8275 struct netdev_adjacent *lower; 8276 8277 lower = list_entry(*iter, struct netdev_adjacent, list); 8278 8279 if (&lower->list == &dev->adj_list.lower) 8280 return NULL; 8281 8282 *iter = lower->list.next; 8283 8284 return lower->private; 8285 } 8286 EXPORT_SYMBOL(netdev_lower_get_next_private); 8287 8288 /** 8289 * netdev_lower_get_next_private_rcu - Get the next ->private from the 8290 * lower neighbour list, RCU 8291 * variant 8292 * @dev: device 8293 * @iter: list_head ** of the current position 8294 * 8295 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8296 * list, starting from iter position. The caller must hold RCU read lock. 8297 */ 8298 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 8299 struct list_head **iter) 8300 { 8301 struct netdev_adjacent *lower; 8302 8303 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 8304 8305 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8306 8307 if (&lower->list == &dev->adj_list.lower) 8308 return NULL; 8309 8310 *iter = &lower->list; 8311 8312 return lower->private; 8313 } 8314 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 8315 8316 /** 8317 * netdev_lower_get_next - Get the next device from the lower neighbour 8318 * list 8319 * @dev: device 8320 * @iter: list_head ** of the current position 8321 * 8322 * Gets the next netdev_adjacent from the dev's lower neighbour 8323 * list, starting from iter position. The caller must hold RTNL lock or 8324 * its own locking that guarantees that the neighbour lower 8325 * list will remain unchanged. 8326 */ 8327 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 8328 { 8329 struct netdev_adjacent *lower; 8330 8331 lower = list_entry(*iter, struct netdev_adjacent, list); 8332 8333 if (&lower->list == &dev->adj_list.lower) 8334 return NULL; 8335 8336 *iter = lower->list.next; 8337 8338 return lower->dev; 8339 } 8340 EXPORT_SYMBOL(netdev_lower_get_next); 8341 8342 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 8343 struct list_head **iter) 8344 { 8345 struct netdev_adjacent *lower; 8346 8347 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8348 8349 if (&lower->list == &dev->adj_list.lower) 8350 return NULL; 8351 8352 *iter = &lower->list; 8353 8354 return lower->dev; 8355 } 8356 8357 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 8358 struct list_head **iter, 8359 bool *ignore) 8360 { 8361 struct netdev_adjacent *lower; 8362 8363 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8364 8365 if (&lower->list == &dev->adj_list.lower) 8366 return NULL; 8367 8368 *iter = &lower->list; 8369 *ignore = lower->ignore; 8370 8371 return lower->dev; 8372 } 8373 8374 int netdev_walk_all_lower_dev(struct net_device *dev, 8375 int (*fn)(struct net_device *dev, 8376 struct netdev_nested_priv *priv), 8377 struct netdev_nested_priv *priv) 8378 { 8379 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8380 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8381 int ret, cur = 0; 8382 8383 now = dev; 8384 iter = &dev->adj_list.lower; 8385 8386 while (1) { 8387 if (now != dev) { 8388 ret = fn(now, priv); 8389 if (ret) 8390 return ret; 8391 } 8392 8393 next = NULL; 8394 while (1) { 8395 ldev = netdev_next_lower_dev(now, &iter); 8396 if (!ldev) 8397 break; 8398 8399 next = ldev; 8400 niter = &ldev->adj_list.lower; 8401 dev_stack[cur] = now; 8402 iter_stack[cur++] = iter; 8403 break; 8404 } 8405 8406 if (!next) { 8407 if (!cur) 8408 return 0; 8409 next = dev_stack[--cur]; 8410 niter = iter_stack[cur]; 8411 } 8412 8413 now = next; 8414 iter = niter; 8415 } 8416 8417 return 0; 8418 } 8419 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8420 8421 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8422 int (*fn)(struct net_device *dev, 8423 struct netdev_nested_priv *priv), 8424 struct netdev_nested_priv *priv) 8425 { 8426 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8427 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8428 int ret, cur = 0; 8429 bool ignore; 8430 8431 now = dev; 8432 iter = &dev->adj_list.lower; 8433 8434 while (1) { 8435 if (now != dev) { 8436 ret = fn(now, priv); 8437 if (ret) 8438 return ret; 8439 } 8440 8441 next = NULL; 8442 while (1) { 8443 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8444 if (!ldev) 8445 break; 8446 if (ignore) 8447 continue; 8448 8449 next = ldev; 8450 niter = &ldev->adj_list.lower; 8451 dev_stack[cur] = now; 8452 iter_stack[cur++] = iter; 8453 break; 8454 } 8455 8456 if (!next) { 8457 if (!cur) 8458 return 0; 8459 next = dev_stack[--cur]; 8460 niter = iter_stack[cur]; 8461 } 8462 8463 now = next; 8464 iter = niter; 8465 } 8466 8467 return 0; 8468 } 8469 8470 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8471 struct list_head **iter) 8472 { 8473 struct netdev_adjacent *lower; 8474 8475 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8476 if (&lower->list == &dev->adj_list.lower) 8477 return NULL; 8478 8479 *iter = &lower->list; 8480 8481 return lower->dev; 8482 } 8483 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8484 8485 static u8 __netdev_upper_depth(struct net_device *dev) 8486 { 8487 struct net_device *udev; 8488 struct list_head *iter; 8489 u8 max_depth = 0; 8490 bool ignore; 8491 8492 for (iter = &dev->adj_list.upper, 8493 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8494 udev; 8495 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8496 if (ignore) 8497 continue; 8498 if (max_depth < udev->upper_level) 8499 max_depth = udev->upper_level; 8500 } 8501 8502 return max_depth; 8503 } 8504 8505 static u8 __netdev_lower_depth(struct net_device *dev) 8506 { 8507 struct net_device *ldev; 8508 struct list_head *iter; 8509 u8 max_depth = 0; 8510 bool ignore; 8511 8512 for (iter = &dev->adj_list.lower, 8513 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8514 ldev; 8515 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8516 if (ignore) 8517 continue; 8518 if (max_depth < ldev->lower_level) 8519 max_depth = ldev->lower_level; 8520 } 8521 8522 return max_depth; 8523 } 8524 8525 static int __netdev_update_upper_level(struct net_device *dev, 8526 struct netdev_nested_priv *__unused) 8527 { 8528 dev->upper_level = __netdev_upper_depth(dev) + 1; 8529 return 0; 8530 } 8531 8532 #ifdef CONFIG_LOCKDEP 8533 static LIST_HEAD(net_unlink_list); 8534 8535 static void net_unlink_todo(struct net_device *dev) 8536 { 8537 if (list_empty(&dev->unlink_list)) 8538 list_add_tail(&dev->unlink_list, &net_unlink_list); 8539 } 8540 #endif 8541 8542 static int __netdev_update_lower_level(struct net_device *dev, 8543 struct netdev_nested_priv *priv) 8544 { 8545 dev->lower_level = __netdev_lower_depth(dev) + 1; 8546 8547 #ifdef CONFIG_LOCKDEP 8548 if (!priv) 8549 return 0; 8550 8551 if (priv->flags & NESTED_SYNC_IMM) 8552 dev->nested_level = dev->lower_level - 1; 8553 if (priv->flags & NESTED_SYNC_TODO) 8554 net_unlink_todo(dev); 8555 #endif 8556 return 0; 8557 } 8558 8559 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8560 int (*fn)(struct net_device *dev, 8561 struct netdev_nested_priv *priv), 8562 struct netdev_nested_priv *priv) 8563 { 8564 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8565 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8566 int ret, cur = 0; 8567 8568 now = dev; 8569 iter = &dev->adj_list.lower; 8570 8571 while (1) { 8572 if (now != dev) { 8573 ret = fn(now, priv); 8574 if (ret) 8575 return ret; 8576 } 8577 8578 next = NULL; 8579 while (1) { 8580 ldev = netdev_next_lower_dev_rcu(now, &iter); 8581 if (!ldev) 8582 break; 8583 8584 next = ldev; 8585 niter = &ldev->adj_list.lower; 8586 dev_stack[cur] = now; 8587 iter_stack[cur++] = iter; 8588 break; 8589 } 8590 8591 if (!next) { 8592 if (!cur) 8593 return 0; 8594 next = dev_stack[--cur]; 8595 niter = iter_stack[cur]; 8596 } 8597 8598 now = next; 8599 iter = niter; 8600 } 8601 8602 return 0; 8603 } 8604 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8605 8606 /** 8607 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8608 * lower neighbour list, RCU 8609 * variant 8610 * @dev: device 8611 * 8612 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8613 * list. The caller must hold RCU read lock. 8614 */ 8615 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8616 { 8617 struct netdev_adjacent *lower; 8618 8619 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8620 struct netdev_adjacent, list); 8621 if (lower) 8622 return lower->private; 8623 return NULL; 8624 } 8625 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8626 8627 /** 8628 * netdev_master_upper_dev_get_rcu - Get master upper device 8629 * @dev: device 8630 * 8631 * Find a master upper device and return pointer to it or NULL in case 8632 * it's not there. The caller must hold the RCU read lock. 8633 */ 8634 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8635 { 8636 struct netdev_adjacent *upper; 8637 8638 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8639 struct netdev_adjacent, list); 8640 if (upper && likely(upper->master)) 8641 return upper->dev; 8642 return NULL; 8643 } 8644 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8645 8646 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8647 struct net_device *adj_dev, 8648 struct list_head *dev_list) 8649 { 8650 char linkname[IFNAMSIZ+7]; 8651 8652 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8653 "upper_%s" : "lower_%s", adj_dev->name); 8654 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8655 linkname); 8656 } 8657 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8658 char *name, 8659 struct list_head *dev_list) 8660 { 8661 char linkname[IFNAMSIZ+7]; 8662 8663 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8664 "upper_%s" : "lower_%s", name); 8665 sysfs_remove_link(&(dev->dev.kobj), linkname); 8666 } 8667 8668 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8669 struct net_device *adj_dev, 8670 struct list_head *dev_list) 8671 { 8672 return (dev_list == &dev->adj_list.upper || 8673 dev_list == &dev->adj_list.lower) && 8674 net_eq(dev_net(dev), dev_net(adj_dev)); 8675 } 8676 8677 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8678 struct net_device *adj_dev, 8679 struct list_head *dev_list, 8680 void *private, bool master) 8681 { 8682 struct netdev_adjacent *adj; 8683 int ret; 8684 8685 adj = __netdev_find_adj(adj_dev, dev_list); 8686 8687 if (adj) { 8688 adj->ref_nr += 1; 8689 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8690 dev->name, adj_dev->name, adj->ref_nr); 8691 8692 return 0; 8693 } 8694 8695 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 8696 if (!adj) 8697 return -ENOMEM; 8698 8699 adj->dev = adj_dev; 8700 adj->master = master; 8701 adj->ref_nr = 1; 8702 adj->private = private; 8703 adj->ignore = false; 8704 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8705 8706 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8707 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8708 8709 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8710 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8711 if (ret) 8712 goto free_adj; 8713 } 8714 8715 /* Ensure that master link is always the first item in list. */ 8716 if (master) { 8717 ret = sysfs_create_link(&(dev->dev.kobj), 8718 &(adj_dev->dev.kobj), "master"); 8719 if (ret) 8720 goto remove_symlinks; 8721 8722 list_add_rcu(&adj->list, dev_list); 8723 } else { 8724 list_add_tail_rcu(&adj->list, dev_list); 8725 } 8726 8727 return 0; 8728 8729 remove_symlinks: 8730 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8731 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8732 free_adj: 8733 netdev_put(adj_dev, &adj->dev_tracker); 8734 kfree(adj); 8735 8736 return ret; 8737 } 8738 8739 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8740 struct net_device *adj_dev, 8741 u16 ref_nr, 8742 struct list_head *dev_list) 8743 { 8744 struct netdev_adjacent *adj; 8745 8746 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8747 dev->name, adj_dev->name, ref_nr); 8748 8749 adj = __netdev_find_adj(adj_dev, dev_list); 8750 8751 if (!adj) { 8752 pr_err("Adjacency does not exist for device %s from %s\n", 8753 dev->name, adj_dev->name); 8754 WARN_ON(1); 8755 return; 8756 } 8757 8758 if (adj->ref_nr > ref_nr) { 8759 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8760 dev->name, adj_dev->name, ref_nr, 8761 adj->ref_nr - ref_nr); 8762 adj->ref_nr -= ref_nr; 8763 return; 8764 } 8765 8766 if (adj->master) 8767 sysfs_remove_link(&(dev->dev.kobj), "master"); 8768 8769 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8770 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8771 8772 list_del_rcu(&adj->list); 8773 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8774 adj_dev->name, dev->name, adj_dev->name); 8775 netdev_put(adj_dev, &adj->dev_tracker); 8776 kfree_rcu(adj, rcu); 8777 } 8778 8779 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8780 struct net_device *upper_dev, 8781 struct list_head *up_list, 8782 struct list_head *down_list, 8783 void *private, bool master) 8784 { 8785 int ret; 8786 8787 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8788 private, master); 8789 if (ret) 8790 return ret; 8791 8792 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8793 private, false); 8794 if (ret) { 8795 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8796 return ret; 8797 } 8798 8799 return 0; 8800 } 8801 8802 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8803 struct net_device *upper_dev, 8804 u16 ref_nr, 8805 struct list_head *up_list, 8806 struct list_head *down_list) 8807 { 8808 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8809 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8810 } 8811 8812 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8813 struct net_device *upper_dev, 8814 void *private, bool master) 8815 { 8816 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8817 &dev->adj_list.upper, 8818 &upper_dev->adj_list.lower, 8819 private, master); 8820 } 8821 8822 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8823 struct net_device *upper_dev) 8824 { 8825 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8826 &dev->adj_list.upper, 8827 &upper_dev->adj_list.lower); 8828 } 8829 8830 static int __netdev_upper_dev_link(struct net_device *dev, 8831 struct net_device *upper_dev, bool master, 8832 void *upper_priv, void *upper_info, 8833 struct netdev_nested_priv *priv, 8834 struct netlink_ext_ack *extack) 8835 { 8836 struct netdev_notifier_changeupper_info changeupper_info = { 8837 .info = { 8838 .dev = dev, 8839 .extack = extack, 8840 }, 8841 .upper_dev = upper_dev, 8842 .master = master, 8843 .linking = true, 8844 .upper_info = upper_info, 8845 }; 8846 struct net_device *master_dev; 8847 int ret = 0; 8848 8849 ASSERT_RTNL(); 8850 8851 if (dev == upper_dev) 8852 return -EBUSY; 8853 8854 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8855 if (__netdev_has_upper_dev(upper_dev, dev)) 8856 return -EBUSY; 8857 8858 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8859 return -EMLINK; 8860 8861 if (!master) { 8862 if (__netdev_has_upper_dev(dev, upper_dev)) 8863 return -EEXIST; 8864 } else { 8865 master_dev = __netdev_master_upper_dev_get(dev); 8866 if (master_dev) 8867 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8868 } 8869 8870 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8871 &changeupper_info.info); 8872 ret = notifier_to_errno(ret); 8873 if (ret) 8874 return ret; 8875 8876 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8877 master); 8878 if (ret) 8879 return ret; 8880 8881 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8882 &changeupper_info.info); 8883 ret = notifier_to_errno(ret); 8884 if (ret) 8885 goto rollback; 8886 8887 __netdev_update_upper_level(dev, NULL); 8888 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8889 8890 __netdev_update_lower_level(upper_dev, priv); 8891 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8892 priv); 8893 8894 return 0; 8895 8896 rollback: 8897 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8898 8899 return ret; 8900 } 8901 8902 /** 8903 * netdev_upper_dev_link - Add a link to the upper device 8904 * @dev: device 8905 * @upper_dev: new upper device 8906 * @extack: netlink extended ack 8907 * 8908 * Adds a link to device which is upper to this one. The caller must hold 8909 * the RTNL lock. On a failure a negative errno code is returned. 8910 * On success the reference counts are adjusted and the function 8911 * returns zero. 8912 */ 8913 int netdev_upper_dev_link(struct net_device *dev, 8914 struct net_device *upper_dev, 8915 struct netlink_ext_ack *extack) 8916 { 8917 struct netdev_nested_priv priv = { 8918 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8919 .data = NULL, 8920 }; 8921 8922 return __netdev_upper_dev_link(dev, upper_dev, false, 8923 NULL, NULL, &priv, extack); 8924 } 8925 EXPORT_SYMBOL(netdev_upper_dev_link); 8926 8927 /** 8928 * netdev_master_upper_dev_link - Add a master link to the upper device 8929 * @dev: device 8930 * @upper_dev: new upper device 8931 * @upper_priv: upper device private 8932 * @upper_info: upper info to be passed down via notifier 8933 * @extack: netlink extended ack 8934 * 8935 * Adds a link to device which is upper to this one. In this case, only 8936 * one master upper device can be linked, although other non-master devices 8937 * might be linked as well. The caller must hold the RTNL lock. 8938 * On a failure a negative errno code is returned. On success the reference 8939 * counts are adjusted and the function returns zero. 8940 */ 8941 int netdev_master_upper_dev_link(struct net_device *dev, 8942 struct net_device *upper_dev, 8943 void *upper_priv, void *upper_info, 8944 struct netlink_ext_ack *extack) 8945 { 8946 struct netdev_nested_priv priv = { 8947 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8948 .data = NULL, 8949 }; 8950 8951 return __netdev_upper_dev_link(dev, upper_dev, true, 8952 upper_priv, upper_info, &priv, extack); 8953 } 8954 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8955 8956 static void __netdev_upper_dev_unlink(struct net_device *dev, 8957 struct net_device *upper_dev, 8958 struct netdev_nested_priv *priv) 8959 { 8960 struct netdev_notifier_changeupper_info changeupper_info = { 8961 .info = { 8962 .dev = dev, 8963 }, 8964 .upper_dev = upper_dev, 8965 .linking = false, 8966 }; 8967 8968 ASSERT_RTNL(); 8969 8970 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8971 8972 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8973 &changeupper_info.info); 8974 8975 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8976 8977 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8978 &changeupper_info.info); 8979 8980 __netdev_update_upper_level(dev, NULL); 8981 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8982 8983 __netdev_update_lower_level(upper_dev, priv); 8984 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8985 priv); 8986 } 8987 8988 /** 8989 * netdev_upper_dev_unlink - Removes a link to upper device 8990 * @dev: device 8991 * @upper_dev: new upper device 8992 * 8993 * Removes a link to device which is upper to this one. The caller must hold 8994 * the RTNL lock. 8995 */ 8996 void netdev_upper_dev_unlink(struct net_device *dev, 8997 struct net_device *upper_dev) 8998 { 8999 struct netdev_nested_priv priv = { 9000 .flags = NESTED_SYNC_TODO, 9001 .data = NULL, 9002 }; 9003 9004 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 9005 } 9006 EXPORT_SYMBOL(netdev_upper_dev_unlink); 9007 9008 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 9009 struct net_device *lower_dev, 9010 bool val) 9011 { 9012 struct netdev_adjacent *adj; 9013 9014 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 9015 if (adj) 9016 adj->ignore = val; 9017 9018 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 9019 if (adj) 9020 adj->ignore = val; 9021 } 9022 9023 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 9024 struct net_device *lower_dev) 9025 { 9026 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 9027 } 9028 9029 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 9030 struct net_device *lower_dev) 9031 { 9032 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 9033 } 9034 9035 int netdev_adjacent_change_prepare(struct net_device *old_dev, 9036 struct net_device *new_dev, 9037 struct net_device *dev, 9038 struct netlink_ext_ack *extack) 9039 { 9040 struct netdev_nested_priv priv = { 9041 .flags = 0, 9042 .data = NULL, 9043 }; 9044 int err; 9045 9046 if (!new_dev) 9047 return 0; 9048 9049 if (old_dev && new_dev != old_dev) 9050 netdev_adjacent_dev_disable(dev, old_dev); 9051 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 9052 extack); 9053 if (err) { 9054 if (old_dev && new_dev != old_dev) 9055 netdev_adjacent_dev_enable(dev, old_dev); 9056 return err; 9057 } 9058 9059 return 0; 9060 } 9061 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 9062 9063 void netdev_adjacent_change_commit(struct net_device *old_dev, 9064 struct net_device *new_dev, 9065 struct net_device *dev) 9066 { 9067 struct netdev_nested_priv priv = { 9068 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 9069 .data = NULL, 9070 }; 9071 9072 if (!new_dev || !old_dev) 9073 return; 9074 9075 if (new_dev == old_dev) 9076 return; 9077 9078 netdev_adjacent_dev_enable(dev, old_dev); 9079 __netdev_upper_dev_unlink(old_dev, dev, &priv); 9080 } 9081 EXPORT_SYMBOL(netdev_adjacent_change_commit); 9082 9083 void netdev_adjacent_change_abort(struct net_device *old_dev, 9084 struct net_device *new_dev, 9085 struct net_device *dev) 9086 { 9087 struct netdev_nested_priv priv = { 9088 .flags = 0, 9089 .data = NULL, 9090 }; 9091 9092 if (!new_dev) 9093 return; 9094 9095 if (old_dev && new_dev != old_dev) 9096 netdev_adjacent_dev_enable(dev, old_dev); 9097 9098 __netdev_upper_dev_unlink(new_dev, dev, &priv); 9099 } 9100 EXPORT_SYMBOL(netdev_adjacent_change_abort); 9101 9102 /** 9103 * netdev_bonding_info_change - Dispatch event about slave change 9104 * @dev: device 9105 * @bonding_info: info to dispatch 9106 * 9107 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 9108 * The caller must hold the RTNL lock. 9109 */ 9110 void netdev_bonding_info_change(struct net_device *dev, 9111 struct netdev_bonding_info *bonding_info) 9112 { 9113 struct netdev_notifier_bonding_info info = { 9114 .info.dev = dev, 9115 }; 9116 9117 memcpy(&info.bonding_info, bonding_info, 9118 sizeof(struct netdev_bonding_info)); 9119 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 9120 &info.info); 9121 } 9122 EXPORT_SYMBOL(netdev_bonding_info_change); 9123 9124 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 9125 struct netlink_ext_ack *extack) 9126 { 9127 struct netdev_notifier_offload_xstats_info info = { 9128 .info.dev = dev, 9129 .info.extack = extack, 9130 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9131 }; 9132 int err; 9133 int rc; 9134 9135 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 9136 GFP_KERNEL); 9137 if (!dev->offload_xstats_l3) 9138 return -ENOMEM; 9139 9140 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 9141 NETDEV_OFFLOAD_XSTATS_DISABLE, 9142 &info.info); 9143 err = notifier_to_errno(rc); 9144 if (err) 9145 goto free_stats; 9146 9147 return 0; 9148 9149 free_stats: 9150 kfree(dev->offload_xstats_l3); 9151 dev->offload_xstats_l3 = NULL; 9152 return err; 9153 } 9154 9155 int netdev_offload_xstats_enable(struct net_device *dev, 9156 enum netdev_offload_xstats_type type, 9157 struct netlink_ext_ack *extack) 9158 { 9159 ASSERT_RTNL(); 9160 9161 if (netdev_offload_xstats_enabled(dev, type)) 9162 return -EALREADY; 9163 9164 switch (type) { 9165 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9166 return netdev_offload_xstats_enable_l3(dev, extack); 9167 } 9168 9169 WARN_ON(1); 9170 return -EINVAL; 9171 } 9172 EXPORT_SYMBOL(netdev_offload_xstats_enable); 9173 9174 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 9175 { 9176 struct netdev_notifier_offload_xstats_info info = { 9177 .info.dev = dev, 9178 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9179 }; 9180 9181 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 9182 &info.info); 9183 kfree(dev->offload_xstats_l3); 9184 dev->offload_xstats_l3 = NULL; 9185 } 9186 9187 int netdev_offload_xstats_disable(struct net_device *dev, 9188 enum netdev_offload_xstats_type type) 9189 { 9190 ASSERT_RTNL(); 9191 9192 if (!netdev_offload_xstats_enabled(dev, type)) 9193 return -EALREADY; 9194 9195 switch (type) { 9196 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9197 netdev_offload_xstats_disable_l3(dev); 9198 return 0; 9199 } 9200 9201 WARN_ON(1); 9202 return -EINVAL; 9203 } 9204 EXPORT_SYMBOL(netdev_offload_xstats_disable); 9205 9206 static void netdev_offload_xstats_disable_all(struct net_device *dev) 9207 { 9208 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 9209 } 9210 9211 static struct rtnl_hw_stats64 * 9212 netdev_offload_xstats_get_ptr(const struct net_device *dev, 9213 enum netdev_offload_xstats_type type) 9214 { 9215 switch (type) { 9216 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9217 return dev->offload_xstats_l3; 9218 } 9219 9220 WARN_ON(1); 9221 return NULL; 9222 } 9223 9224 bool netdev_offload_xstats_enabled(const struct net_device *dev, 9225 enum netdev_offload_xstats_type type) 9226 { 9227 ASSERT_RTNL(); 9228 9229 return netdev_offload_xstats_get_ptr(dev, type); 9230 } 9231 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 9232 9233 struct netdev_notifier_offload_xstats_ru { 9234 bool used; 9235 }; 9236 9237 struct netdev_notifier_offload_xstats_rd { 9238 struct rtnl_hw_stats64 stats; 9239 bool used; 9240 }; 9241 9242 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 9243 const struct rtnl_hw_stats64 *src) 9244 { 9245 dest->rx_packets += src->rx_packets; 9246 dest->tx_packets += src->tx_packets; 9247 dest->rx_bytes += src->rx_bytes; 9248 dest->tx_bytes += src->tx_bytes; 9249 dest->rx_errors += src->rx_errors; 9250 dest->tx_errors += src->tx_errors; 9251 dest->rx_dropped += src->rx_dropped; 9252 dest->tx_dropped += src->tx_dropped; 9253 dest->multicast += src->multicast; 9254 } 9255 9256 static int netdev_offload_xstats_get_used(struct net_device *dev, 9257 enum netdev_offload_xstats_type type, 9258 bool *p_used, 9259 struct netlink_ext_ack *extack) 9260 { 9261 struct netdev_notifier_offload_xstats_ru report_used = {}; 9262 struct netdev_notifier_offload_xstats_info info = { 9263 .info.dev = dev, 9264 .info.extack = extack, 9265 .type = type, 9266 .report_used = &report_used, 9267 }; 9268 int rc; 9269 9270 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 9271 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 9272 &info.info); 9273 *p_used = report_used.used; 9274 return notifier_to_errno(rc); 9275 } 9276 9277 static int netdev_offload_xstats_get_stats(struct net_device *dev, 9278 enum netdev_offload_xstats_type type, 9279 struct rtnl_hw_stats64 *p_stats, 9280 bool *p_used, 9281 struct netlink_ext_ack *extack) 9282 { 9283 struct netdev_notifier_offload_xstats_rd report_delta = {}; 9284 struct netdev_notifier_offload_xstats_info info = { 9285 .info.dev = dev, 9286 .info.extack = extack, 9287 .type = type, 9288 .report_delta = &report_delta, 9289 }; 9290 struct rtnl_hw_stats64 *stats; 9291 int rc; 9292 9293 stats = netdev_offload_xstats_get_ptr(dev, type); 9294 if (WARN_ON(!stats)) 9295 return -EINVAL; 9296 9297 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 9298 &info.info); 9299 9300 /* Cache whatever we got, even if there was an error, otherwise the 9301 * successful stats retrievals would get lost. 9302 */ 9303 netdev_hw_stats64_add(stats, &report_delta.stats); 9304 9305 if (p_stats) 9306 *p_stats = *stats; 9307 *p_used = report_delta.used; 9308 9309 return notifier_to_errno(rc); 9310 } 9311 9312 int netdev_offload_xstats_get(struct net_device *dev, 9313 enum netdev_offload_xstats_type type, 9314 struct rtnl_hw_stats64 *p_stats, bool *p_used, 9315 struct netlink_ext_ack *extack) 9316 { 9317 ASSERT_RTNL(); 9318 9319 if (p_stats) 9320 return netdev_offload_xstats_get_stats(dev, type, p_stats, 9321 p_used, extack); 9322 else 9323 return netdev_offload_xstats_get_used(dev, type, p_used, 9324 extack); 9325 } 9326 EXPORT_SYMBOL(netdev_offload_xstats_get); 9327 9328 void 9329 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 9330 const struct rtnl_hw_stats64 *stats) 9331 { 9332 report_delta->used = true; 9333 netdev_hw_stats64_add(&report_delta->stats, stats); 9334 } 9335 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 9336 9337 void 9338 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 9339 { 9340 report_used->used = true; 9341 } 9342 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 9343 9344 void netdev_offload_xstats_push_delta(struct net_device *dev, 9345 enum netdev_offload_xstats_type type, 9346 const struct rtnl_hw_stats64 *p_stats) 9347 { 9348 struct rtnl_hw_stats64 *stats; 9349 9350 ASSERT_RTNL(); 9351 9352 stats = netdev_offload_xstats_get_ptr(dev, type); 9353 if (WARN_ON(!stats)) 9354 return; 9355 9356 netdev_hw_stats64_add(stats, p_stats); 9357 } 9358 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 9359 9360 /** 9361 * netdev_get_xmit_slave - Get the xmit slave of master device 9362 * @dev: device 9363 * @skb: The packet 9364 * @all_slaves: assume all the slaves are active 9365 * 9366 * The reference counters are not incremented so the caller must be 9367 * careful with locks. The caller must hold RCU lock. 9368 * %NULL is returned if no slave is found. 9369 */ 9370 9371 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 9372 struct sk_buff *skb, 9373 bool all_slaves) 9374 { 9375 const struct net_device_ops *ops = dev->netdev_ops; 9376 9377 if (!ops->ndo_get_xmit_slave) 9378 return NULL; 9379 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 9380 } 9381 EXPORT_SYMBOL(netdev_get_xmit_slave); 9382 9383 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 9384 struct sock *sk) 9385 { 9386 const struct net_device_ops *ops = dev->netdev_ops; 9387 9388 if (!ops->ndo_sk_get_lower_dev) 9389 return NULL; 9390 return ops->ndo_sk_get_lower_dev(dev, sk); 9391 } 9392 9393 /** 9394 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 9395 * @dev: device 9396 * @sk: the socket 9397 * 9398 * %NULL is returned if no lower device is found. 9399 */ 9400 9401 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 9402 struct sock *sk) 9403 { 9404 struct net_device *lower; 9405 9406 lower = netdev_sk_get_lower_dev(dev, sk); 9407 while (lower) { 9408 dev = lower; 9409 lower = netdev_sk_get_lower_dev(dev, sk); 9410 } 9411 9412 return dev; 9413 } 9414 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 9415 9416 static void netdev_adjacent_add_links(struct net_device *dev) 9417 { 9418 struct netdev_adjacent *iter; 9419 9420 struct net *net = dev_net(dev); 9421 9422 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9423 if (!net_eq(net, dev_net(iter->dev))) 9424 continue; 9425 netdev_adjacent_sysfs_add(iter->dev, dev, 9426 &iter->dev->adj_list.lower); 9427 netdev_adjacent_sysfs_add(dev, iter->dev, 9428 &dev->adj_list.upper); 9429 } 9430 9431 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9432 if (!net_eq(net, dev_net(iter->dev))) 9433 continue; 9434 netdev_adjacent_sysfs_add(iter->dev, dev, 9435 &iter->dev->adj_list.upper); 9436 netdev_adjacent_sysfs_add(dev, iter->dev, 9437 &dev->adj_list.lower); 9438 } 9439 } 9440 9441 static void netdev_adjacent_del_links(struct net_device *dev) 9442 { 9443 struct netdev_adjacent *iter; 9444 9445 struct net *net = dev_net(dev); 9446 9447 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9448 if (!net_eq(net, dev_net(iter->dev))) 9449 continue; 9450 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9451 &iter->dev->adj_list.lower); 9452 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9453 &dev->adj_list.upper); 9454 } 9455 9456 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9457 if (!net_eq(net, dev_net(iter->dev))) 9458 continue; 9459 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9460 &iter->dev->adj_list.upper); 9461 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9462 &dev->adj_list.lower); 9463 } 9464 } 9465 9466 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9467 { 9468 struct netdev_adjacent *iter; 9469 9470 struct net *net = dev_net(dev); 9471 9472 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9473 if (!net_eq(net, dev_net(iter->dev))) 9474 continue; 9475 netdev_adjacent_sysfs_del(iter->dev, oldname, 9476 &iter->dev->adj_list.lower); 9477 netdev_adjacent_sysfs_add(iter->dev, dev, 9478 &iter->dev->adj_list.lower); 9479 } 9480 9481 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9482 if (!net_eq(net, dev_net(iter->dev))) 9483 continue; 9484 netdev_adjacent_sysfs_del(iter->dev, oldname, 9485 &iter->dev->adj_list.upper); 9486 netdev_adjacent_sysfs_add(iter->dev, dev, 9487 &iter->dev->adj_list.upper); 9488 } 9489 } 9490 9491 void *netdev_lower_dev_get_private(struct net_device *dev, 9492 struct net_device *lower_dev) 9493 { 9494 struct netdev_adjacent *lower; 9495 9496 if (!lower_dev) 9497 return NULL; 9498 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9499 if (!lower) 9500 return NULL; 9501 9502 return lower->private; 9503 } 9504 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9505 9506 9507 /** 9508 * netdev_lower_state_changed - Dispatch event about lower device state change 9509 * @lower_dev: device 9510 * @lower_state_info: state to dispatch 9511 * 9512 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9513 * The caller must hold the RTNL lock. 9514 */ 9515 void netdev_lower_state_changed(struct net_device *lower_dev, 9516 void *lower_state_info) 9517 { 9518 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9519 .info.dev = lower_dev, 9520 }; 9521 9522 ASSERT_RTNL(); 9523 changelowerstate_info.lower_state_info = lower_state_info; 9524 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9525 &changelowerstate_info.info); 9526 } 9527 EXPORT_SYMBOL(netdev_lower_state_changed); 9528 9529 static void dev_change_rx_flags(struct net_device *dev, int flags) 9530 { 9531 const struct net_device_ops *ops = dev->netdev_ops; 9532 9533 if (ops->ndo_change_rx_flags) 9534 ops->ndo_change_rx_flags(dev, flags); 9535 } 9536 9537 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9538 { 9539 unsigned int old_flags = dev->flags; 9540 unsigned int promiscuity, flags; 9541 kuid_t uid; 9542 kgid_t gid; 9543 9544 ASSERT_RTNL(); 9545 9546 promiscuity = dev->promiscuity + inc; 9547 if (promiscuity == 0) { 9548 /* 9549 * Avoid overflow. 9550 * If inc causes overflow, untouch promisc and return error. 9551 */ 9552 if (unlikely(inc > 0)) { 9553 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9554 return -EOVERFLOW; 9555 } 9556 flags = old_flags & ~IFF_PROMISC; 9557 } else { 9558 flags = old_flags | IFF_PROMISC; 9559 } 9560 WRITE_ONCE(dev->promiscuity, promiscuity); 9561 if (flags != old_flags) { 9562 WRITE_ONCE(dev->flags, flags); 9563 netdev_info(dev, "%s promiscuous mode\n", 9564 dev->flags & IFF_PROMISC ? "entered" : "left"); 9565 if (audit_enabled) { 9566 current_uid_gid(&uid, &gid); 9567 audit_log(audit_context(), GFP_ATOMIC, 9568 AUDIT_ANOM_PROMISCUOUS, 9569 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9570 dev->name, (dev->flags & IFF_PROMISC), 9571 (old_flags & IFF_PROMISC), 9572 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9573 from_kuid(&init_user_ns, uid), 9574 from_kgid(&init_user_ns, gid), 9575 audit_get_sessionid(current)); 9576 } 9577 9578 dev_change_rx_flags(dev, IFF_PROMISC); 9579 } 9580 if (notify) { 9581 /* The ops lock is only required to ensure consistent locking 9582 * for `NETDEV_CHANGE` notifiers. This function is sometimes 9583 * called without the lock, even for devices that are ops 9584 * locked, such as in `dev_uc_sync_multiple` when using 9585 * bonding or teaming. 9586 */ 9587 netdev_ops_assert_locked(dev); 9588 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9589 } 9590 return 0; 9591 } 9592 9593 int netif_set_promiscuity(struct net_device *dev, int inc) 9594 { 9595 unsigned int old_flags = dev->flags; 9596 int err; 9597 9598 err = __dev_set_promiscuity(dev, inc, true); 9599 if (err < 0) 9600 return err; 9601 if (dev->flags != old_flags) 9602 dev_set_rx_mode(dev); 9603 return err; 9604 } 9605 9606 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9607 { 9608 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9609 unsigned int allmulti, flags; 9610 9611 ASSERT_RTNL(); 9612 9613 allmulti = dev->allmulti + inc; 9614 if (allmulti == 0) { 9615 /* 9616 * Avoid overflow. 9617 * If inc causes overflow, untouch allmulti and return error. 9618 */ 9619 if (unlikely(inc > 0)) { 9620 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9621 return -EOVERFLOW; 9622 } 9623 flags = old_flags & ~IFF_ALLMULTI; 9624 } else { 9625 flags = old_flags | IFF_ALLMULTI; 9626 } 9627 WRITE_ONCE(dev->allmulti, allmulti); 9628 if (flags != old_flags) { 9629 WRITE_ONCE(dev->flags, flags); 9630 netdev_info(dev, "%s allmulticast mode\n", 9631 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9632 dev_change_rx_flags(dev, IFF_ALLMULTI); 9633 dev_set_rx_mode(dev); 9634 if (notify) 9635 __dev_notify_flags(dev, old_flags, 9636 dev->gflags ^ old_gflags, 0, NULL); 9637 } 9638 return 0; 9639 } 9640 9641 /* 9642 * Upload unicast and multicast address lists to device and 9643 * configure RX filtering. When the device doesn't support unicast 9644 * filtering it is put in promiscuous mode while unicast addresses 9645 * are present. 9646 */ 9647 void __dev_set_rx_mode(struct net_device *dev) 9648 { 9649 const struct net_device_ops *ops = dev->netdev_ops; 9650 9651 /* dev_open will call this function so the list will stay sane. */ 9652 if (!(dev->flags&IFF_UP)) 9653 return; 9654 9655 if (!netif_device_present(dev)) 9656 return; 9657 9658 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 9659 /* Unicast addresses changes may only happen under the rtnl, 9660 * therefore calling __dev_set_promiscuity here is safe. 9661 */ 9662 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 9663 __dev_set_promiscuity(dev, 1, false); 9664 dev->uc_promisc = true; 9665 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 9666 __dev_set_promiscuity(dev, -1, false); 9667 dev->uc_promisc = false; 9668 } 9669 } 9670 9671 if (ops->ndo_set_rx_mode) 9672 ops->ndo_set_rx_mode(dev); 9673 } 9674 9675 void dev_set_rx_mode(struct net_device *dev) 9676 { 9677 netif_addr_lock_bh(dev); 9678 __dev_set_rx_mode(dev); 9679 netif_addr_unlock_bh(dev); 9680 } 9681 9682 /** 9683 * netif_get_flags() - get flags reported to userspace 9684 * @dev: device 9685 * 9686 * Get the combination of flag bits exported through APIs to userspace. 9687 */ 9688 unsigned int netif_get_flags(const struct net_device *dev) 9689 { 9690 unsigned int flags; 9691 9692 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9693 IFF_ALLMULTI | 9694 IFF_RUNNING | 9695 IFF_LOWER_UP | 9696 IFF_DORMANT)) | 9697 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9698 IFF_ALLMULTI)); 9699 9700 if (netif_running(dev)) { 9701 if (netif_oper_up(dev)) 9702 flags |= IFF_RUNNING; 9703 if (netif_carrier_ok(dev)) 9704 flags |= IFF_LOWER_UP; 9705 if (netif_dormant(dev)) 9706 flags |= IFF_DORMANT; 9707 } 9708 9709 return flags; 9710 } 9711 EXPORT_SYMBOL(netif_get_flags); 9712 9713 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9714 struct netlink_ext_ack *extack) 9715 { 9716 unsigned int old_flags = dev->flags; 9717 int ret; 9718 9719 ASSERT_RTNL(); 9720 9721 /* 9722 * Set the flags on our device. 9723 */ 9724 9725 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9726 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9727 IFF_AUTOMEDIA)) | 9728 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9729 IFF_ALLMULTI)); 9730 9731 /* 9732 * Load in the correct multicast list now the flags have changed. 9733 */ 9734 9735 if ((old_flags ^ flags) & IFF_MULTICAST) 9736 dev_change_rx_flags(dev, IFF_MULTICAST); 9737 9738 dev_set_rx_mode(dev); 9739 9740 /* 9741 * Have we downed the interface. We handle IFF_UP ourselves 9742 * according to user attempts to set it, rather than blindly 9743 * setting it. 9744 */ 9745 9746 ret = 0; 9747 if ((old_flags ^ flags) & IFF_UP) { 9748 if (old_flags & IFF_UP) 9749 __dev_close(dev); 9750 else 9751 ret = __dev_open(dev, extack); 9752 } 9753 9754 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9755 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9756 old_flags = dev->flags; 9757 9758 dev->gflags ^= IFF_PROMISC; 9759 9760 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9761 if (dev->flags != old_flags) 9762 dev_set_rx_mode(dev); 9763 } 9764 9765 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9766 * is important. Some (broken) drivers set IFF_PROMISC, when 9767 * IFF_ALLMULTI is requested not asking us and not reporting. 9768 */ 9769 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9770 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9771 9772 dev->gflags ^= IFF_ALLMULTI; 9773 netif_set_allmulti(dev, inc, false); 9774 } 9775 9776 return ret; 9777 } 9778 9779 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9780 unsigned int gchanges, u32 portid, 9781 const struct nlmsghdr *nlh) 9782 { 9783 unsigned int changes = dev->flags ^ old_flags; 9784 9785 if (gchanges) 9786 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9787 9788 if (changes & IFF_UP) { 9789 if (dev->flags & IFF_UP) 9790 call_netdevice_notifiers(NETDEV_UP, dev); 9791 else 9792 call_netdevice_notifiers(NETDEV_DOWN, dev); 9793 } 9794 9795 if (dev->flags & IFF_UP && 9796 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9797 struct netdev_notifier_change_info change_info = { 9798 .info = { 9799 .dev = dev, 9800 }, 9801 .flags_changed = changes, 9802 }; 9803 9804 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9805 } 9806 } 9807 9808 int netif_change_flags(struct net_device *dev, unsigned int flags, 9809 struct netlink_ext_ack *extack) 9810 { 9811 int ret; 9812 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9813 9814 ret = __dev_change_flags(dev, flags, extack); 9815 if (ret < 0) 9816 return ret; 9817 9818 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9819 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9820 return ret; 9821 } 9822 9823 int __netif_set_mtu(struct net_device *dev, int new_mtu) 9824 { 9825 const struct net_device_ops *ops = dev->netdev_ops; 9826 9827 if (ops->ndo_change_mtu) 9828 return ops->ndo_change_mtu(dev, new_mtu); 9829 9830 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9831 WRITE_ONCE(dev->mtu, new_mtu); 9832 return 0; 9833 } 9834 EXPORT_SYMBOL_NS_GPL(__netif_set_mtu, "NETDEV_INTERNAL"); 9835 9836 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9837 struct netlink_ext_ack *extack) 9838 { 9839 /* MTU must be positive, and in range */ 9840 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9841 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9842 return -EINVAL; 9843 } 9844 9845 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9846 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9847 return -EINVAL; 9848 } 9849 return 0; 9850 } 9851 9852 /** 9853 * netif_set_mtu_ext() - Change maximum transfer unit 9854 * @dev: device 9855 * @new_mtu: new transfer unit 9856 * @extack: netlink extended ack 9857 * 9858 * Change the maximum transfer size of the network device. 9859 * 9860 * Return: 0 on success, -errno on failure. 9861 */ 9862 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9863 struct netlink_ext_ack *extack) 9864 { 9865 int err, orig_mtu; 9866 9867 netdev_ops_assert_locked(dev); 9868 9869 if (new_mtu == dev->mtu) 9870 return 0; 9871 9872 err = dev_validate_mtu(dev, new_mtu, extack); 9873 if (err) 9874 return err; 9875 9876 if (!netif_device_present(dev)) 9877 return -ENODEV; 9878 9879 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9880 err = notifier_to_errno(err); 9881 if (err) 9882 return err; 9883 9884 orig_mtu = dev->mtu; 9885 err = __netif_set_mtu(dev, new_mtu); 9886 9887 if (!err) { 9888 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9889 orig_mtu); 9890 err = notifier_to_errno(err); 9891 if (err) { 9892 /* setting mtu back and notifying everyone again, 9893 * so that they have a chance to revert changes. 9894 */ 9895 __netif_set_mtu(dev, orig_mtu); 9896 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9897 new_mtu); 9898 } 9899 } 9900 return err; 9901 } 9902 9903 int netif_set_mtu(struct net_device *dev, int new_mtu) 9904 { 9905 struct netlink_ext_ack extack; 9906 int err; 9907 9908 memset(&extack, 0, sizeof(extack)); 9909 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9910 if (err && extack._msg) 9911 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9912 return err; 9913 } 9914 EXPORT_SYMBOL(netif_set_mtu); 9915 9916 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9917 { 9918 unsigned int orig_len = dev->tx_queue_len; 9919 int res; 9920 9921 if (new_len != (unsigned int)new_len) 9922 return -ERANGE; 9923 9924 if (new_len != orig_len) { 9925 WRITE_ONCE(dev->tx_queue_len, new_len); 9926 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9927 res = notifier_to_errno(res); 9928 if (res) 9929 goto err_rollback; 9930 res = dev_qdisc_change_tx_queue_len(dev); 9931 if (res) 9932 goto err_rollback; 9933 } 9934 9935 return 0; 9936 9937 err_rollback: 9938 netdev_err(dev, "refused to change device tx_queue_len\n"); 9939 WRITE_ONCE(dev->tx_queue_len, orig_len); 9940 return res; 9941 } 9942 9943 void netif_set_group(struct net_device *dev, int new_group) 9944 { 9945 dev->group = new_group; 9946 } 9947 9948 /** 9949 * netif_pre_changeaddr_notify() - Call NETDEV_PRE_CHANGEADDR. 9950 * @dev: device 9951 * @addr: new address 9952 * @extack: netlink extended ack 9953 * 9954 * Return: 0 on success, -errno on failure. 9955 */ 9956 int netif_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9957 struct netlink_ext_ack *extack) 9958 { 9959 struct netdev_notifier_pre_changeaddr_info info = { 9960 .info.dev = dev, 9961 .info.extack = extack, 9962 .dev_addr = addr, 9963 }; 9964 int rc; 9965 9966 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9967 return notifier_to_errno(rc); 9968 } 9969 EXPORT_SYMBOL_NS_GPL(netif_pre_changeaddr_notify, "NETDEV_INTERNAL"); 9970 9971 int netif_set_mac_address(struct net_device *dev, struct sockaddr_storage *ss, 9972 struct netlink_ext_ack *extack) 9973 { 9974 const struct net_device_ops *ops = dev->netdev_ops; 9975 int err; 9976 9977 if (!ops->ndo_set_mac_address) 9978 return -EOPNOTSUPP; 9979 if (ss->ss_family != dev->type) 9980 return -EINVAL; 9981 if (!netif_device_present(dev)) 9982 return -ENODEV; 9983 err = netif_pre_changeaddr_notify(dev, ss->__data, extack); 9984 if (err) 9985 return err; 9986 if (memcmp(dev->dev_addr, ss->__data, dev->addr_len)) { 9987 err = ops->ndo_set_mac_address(dev, ss); 9988 if (err) 9989 return err; 9990 } 9991 dev->addr_assign_type = NET_ADDR_SET; 9992 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9993 add_device_randomness(dev->dev_addr, dev->addr_len); 9994 return 0; 9995 } 9996 9997 DECLARE_RWSEM(dev_addr_sem); 9998 9999 /* "sa" is a true struct sockaddr with limited "sa_data" member. */ 10000 int netif_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 10001 { 10002 size_t size = sizeof(sa->sa_data); 10003 struct net_device *dev; 10004 int ret = 0; 10005 10006 down_read(&dev_addr_sem); 10007 rcu_read_lock(); 10008 10009 dev = dev_get_by_name_rcu(net, dev_name); 10010 if (!dev) { 10011 ret = -ENODEV; 10012 goto unlock; 10013 } 10014 if (!dev->addr_len) 10015 memset(sa->sa_data, 0, size); 10016 else 10017 memcpy(sa->sa_data, dev->dev_addr, 10018 min_t(size_t, size, dev->addr_len)); 10019 sa->sa_family = dev->type; 10020 10021 unlock: 10022 rcu_read_unlock(); 10023 up_read(&dev_addr_sem); 10024 return ret; 10025 } 10026 EXPORT_SYMBOL_NS_GPL(netif_get_mac_address, "NETDEV_INTERNAL"); 10027 10028 int netif_change_carrier(struct net_device *dev, bool new_carrier) 10029 { 10030 const struct net_device_ops *ops = dev->netdev_ops; 10031 10032 if (!ops->ndo_change_carrier) 10033 return -EOPNOTSUPP; 10034 if (!netif_device_present(dev)) 10035 return -ENODEV; 10036 return ops->ndo_change_carrier(dev, new_carrier); 10037 } 10038 10039 /** 10040 * dev_get_phys_port_id - Get device physical port ID 10041 * @dev: device 10042 * @ppid: port ID 10043 * 10044 * Get device physical port ID 10045 */ 10046 int dev_get_phys_port_id(struct net_device *dev, 10047 struct netdev_phys_item_id *ppid) 10048 { 10049 const struct net_device_ops *ops = dev->netdev_ops; 10050 10051 if (!ops->ndo_get_phys_port_id) 10052 return -EOPNOTSUPP; 10053 return ops->ndo_get_phys_port_id(dev, ppid); 10054 } 10055 10056 /** 10057 * dev_get_phys_port_name - Get device physical port name 10058 * @dev: device 10059 * @name: port name 10060 * @len: limit of bytes to copy to name 10061 * 10062 * Get device physical port name 10063 */ 10064 int dev_get_phys_port_name(struct net_device *dev, 10065 char *name, size_t len) 10066 { 10067 const struct net_device_ops *ops = dev->netdev_ops; 10068 int err; 10069 10070 if (ops->ndo_get_phys_port_name) { 10071 err = ops->ndo_get_phys_port_name(dev, name, len); 10072 if (err != -EOPNOTSUPP) 10073 return err; 10074 } 10075 return devlink_compat_phys_port_name_get(dev, name, len); 10076 } 10077 10078 /** 10079 * netif_get_port_parent_id() - Get the device's port parent identifier 10080 * @dev: network device 10081 * @ppid: pointer to a storage for the port's parent identifier 10082 * @recurse: allow/disallow recursion to lower devices 10083 * 10084 * Get the devices's port parent identifier. 10085 * 10086 * Return: 0 on success, -errno on failure. 10087 */ 10088 int netif_get_port_parent_id(struct net_device *dev, 10089 struct netdev_phys_item_id *ppid, bool recurse) 10090 { 10091 const struct net_device_ops *ops = dev->netdev_ops; 10092 struct netdev_phys_item_id first = { }; 10093 struct net_device *lower_dev; 10094 struct list_head *iter; 10095 int err; 10096 10097 if (ops->ndo_get_port_parent_id) { 10098 err = ops->ndo_get_port_parent_id(dev, ppid); 10099 if (err != -EOPNOTSUPP) 10100 return err; 10101 } 10102 10103 err = devlink_compat_switch_id_get(dev, ppid); 10104 if (!recurse || err != -EOPNOTSUPP) 10105 return err; 10106 10107 netdev_for_each_lower_dev(dev, lower_dev, iter) { 10108 err = netif_get_port_parent_id(lower_dev, ppid, true); 10109 if (err) 10110 break; 10111 if (!first.id_len) 10112 first = *ppid; 10113 else if (memcmp(&first, ppid, sizeof(*ppid))) 10114 return -EOPNOTSUPP; 10115 } 10116 10117 return err; 10118 } 10119 EXPORT_SYMBOL(netif_get_port_parent_id); 10120 10121 /** 10122 * netdev_port_same_parent_id - Indicate if two network devices have 10123 * the same port parent identifier 10124 * @a: first network device 10125 * @b: second network device 10126 */ 10127 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 10128 { 10129 struct netdev_phys_item_id a_id = { }; 10130 struct netdev_phys_item_id b_id = { }; 10131 10132 if (netif_get_port_parent_id(a, &a_id, true) || 10133 netif_get_port_parent_id(b, &b_id, true)) 10134 return false; 10135 10136 return netdev_phys_item_id_same(&a_id, &b_id); 10137 } 10138 EXPORT_SYMBOL(netdev_port_same_parent_id); 10139 10140 int netif_change_proto_down(struct net_device *dev, bool proto_down) 10141 { 10142 if (!dev->change_proto_down) 10143 return -EOPNOTSUPP; 10144 if (!netif_device_present(dev)) 10145 return -ENODEV; 10146 if (proto_down) 10147 netif_carrier_off(dev); 10148 else 10149 netif_carrier_on(dev); 10150 WRITE_ONCE(dev->proto_down, proto_down); 10151 return 0; 10152 } 10153 10154 /** 10155 * netdev_change_proto_down_reason_locked - proto down reason 10156 * 10157 * @dev: device 10158 * @mask: proto down mask 10159 * @value: proto down value 10160 */ 10161 void netdev_change_proto_down_reason_locked(struct net_device *dev, 10162 unsigned long mask, u32 value) 10163 { 10164 u32 proto_down_reason; 10165 int b; 10166 10167 if (!mask) { 10168 proto_down_reason = value; 10169 } else { 10170 proto_down_reason = dev->proto_down_reason; 10171 for_each_set_bit(b, &mask, 32) { 10172 if (value & (1 << b)) 10173 proto_down_reason |= BIT(b); 10174 else 10175 proto_down_reason &= ~BIT(b); 10176 } 10177 } 10178 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 10179 } 10180 10181 struct bpf_xdp_link { 10182 struct bpf_link link; 10183 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 10184 int flags; 10185 }; 10186 10187 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 10188 { 10189 if (flags & XDP_FLAGS_HW_MODE) 10190 return XDP_MODE_HW; 10191 if (flags & XDP_FLAGS_DRV_MODE) 10192 return XDP_MODE_DRV; 10193 if (flags & XDP_FLAGS_SKB_MODE) 10194 return XDP_MODE_SKB; 10195 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 10196 } 10197 10198 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 10199 { 10200 switch (mode) { 10201 case XDP_MODE_SKB: 10202 return generic_xdp_install; 10203 case XDP_MODE_DRV: 10204 case XDP_MODE_HW: 10205 return dev->netdev_ops->ndo_bpf; 10206 default: 10207 return NULL; 10208 } 10209 } 10210 10211 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 10212 enum bpf_xdp_mode mode) 10213 { 10214 return dev->xdp_state[mode].link; 10215 } 10216 10217 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 10218 enum bpf_xdp_mode mode) 10219 { 10220 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 10221 10222 if (link) 10223 return link->link.prog; 10224 return dev->xdp_state[mode].prog; 10225 } 10226 10227 u8 dev_xdp_prog_count(struct net_device *dev) 10228 { 10229 u8 count = 0; 10230 int i; 10231 10232 for (i = 0; i < __MAX_XDP_MODE; i++) 10233 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 10234 count++; 10235 return count; 10236 } 10237 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 10238 10239 u8 dev_xdp_sb_prog_count(struct net_device *dev) 10240 { 10241 u8 count = 0; 10242 int i; 10243 10244 for (i = 0; i < __MAX_XDP_MODE; i++) 10245 if (dev->xdp_state[i].prog && 10246 !dev->xdp_state[i].prog->aux->xdp_has_frags) 10247 count++; 10248 return count; 10249 } 10250 10251 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 10252 { 10253 if (!dev->netdev_ops->ndo_bpf) 10254 return -EOPNOTSUPP; 10255 10256 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10257 bpf->command == XDP_SETUP_PROG && 10258 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 10259 NL_SET_ERR_MSG(bpf->extack, 10260 "unable to propagate XDP to device using tcp-data-split"); 10261 return -EBUSY; 10262 } 10263 10264 if (dev_get_min_mp_channel_count(dev)) { 10265 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 10266 return -EBUSY; 10267 } 10268 10269 return dev->netdev_ops->ndo_bpf(dev, bpf); 10270 } 10271 EXPORT_SYMBOL_GPL(netif_xdp_propagate); 10272 10273 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 10274 { 10275 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 10276 10277 return prog ? prog->aux->id : 0; 10278 } 10279 10280 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 10281 struct bpf_xdp_link *link) 10282 { 10283 dev->xdp_state[mode].link = link; 10284 dev->xdp_state[mode].prog = NULL; 10285 } 10286 10287 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 10288 struct bpf_prog *prog) 10289 { 10290 dev->xdp_state[mode].link = NULL; 10291 dev->xdp_state[mode].prog = prog; 10292 } 10293 10294 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 10295 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 10296 u32 flags, struct bpf_prog *prog) 10297 { 10298 struct netdev_bpf xdp; 10299 int err; 10300 10301 netdev_ops_assert_locked(dev); 10302 10303 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10304 prog && !prog->aux->xdp_has_frags) { 10305 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 10306 return -EBUSY; 10307 } 10308 10309 if (dev_get_min_mp_channel_count(dev)) { 10310 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 10311 return -EBUSY; 10312 } 10313 10314 memset(&xdp, 0, sizeof(xdp)); 10315 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 10316 xdp.extack = extack; 10317 xdp.flags = flags; 10318 xdp.prog = prog; 10319 10320 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 10321 * "moved" into driver), so they don't increment it on their own, but 10322 * they do decrement refcnt when program is detached or replaced. 10323 * Given net_device also owns link/prog, we need to bump refcnt here 10324 * to prevent drivers from underflowing it. 10325 */ 10326 if (prog) 10327 bpf_prog_inc(prog); 10328 err = bpf_op(dev, &xdp); 10329 if (err) { 10330 if (prog) 10331 bpf_prog_put(prog); 10332 return err; 10333 } 10334 10335 if (mode != XDP_MODE_HW) 10336 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 10337 10338 return 0; 10339 } 10340 10341 static void dev_xdp_uninstall(struct net_device *dev) 10342 { 10343 struct bpf_xdp_link *link; 10344 struct bpf_prog *prog; 10345 enum bpf_xdp_mode mode; 10346 bpf_op_t bpf_op; 10347 10348 ASSERT_RTNL(); 10349 10350 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 10351 prog = dev_xdp_prog(dev, mode); 10352 if (!prog) 10353 continue; 10354 10355 bpf_op = dev_xdp_bpf_op(dev, mode); 10356 if (!bpf_op) 10357 continue; 10358 10359 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10360 10361 /* auto-detach link from net device */ 10362 link = dev_xdp_link(dev, mode); 10363 if (link) 10364 link->dev = NULL; 10365 else 10366 bpf_prog_put(prog); 10367 10368 dev_xdp_set_link(dev, mode, NULL); 10369 } 10370 } 10371 10372 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 10373 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 10374 struct bpf_prog *old_prog, u32 flags) 10375 { 10376 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 10377 struct bpf_prog *cur_prog; 10378 struct net_device *upper; 10379 struct list_head *iter; 10380 enum bpf_xdp_mode mode; 10381 bpf_op_t bpf_op; 10382 int err; 10383 10384 ASSERT_RTNL(); 10385 10386 /* either link or prog attachment, never both */ 10387 if (link && (new_prog || old_prog)) 10388 return -EINVAL; 10389 /* link supports only XDP mode flags */ 10390 if (link && (flags & ~XDP_FLAGS_MODES)) { 10391 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 10392 return -EINVAL; 10393 } 10394 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 10395 if (num_modes > 1) { 10396 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 10397 return -EINVAL; 10398 } 10399 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 10400 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 10401 NL_SET_ERR_MSG(extack, 10402 "More than one program loaded, unset mode is ambiguous"); 10403 return -EINVAL; 10404 } 10405 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 10406 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 10407 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 10408 return -EINVAL; 10409 } 10410 10411 mode = dev_xdp_mode(dev, flags); 10412 /* can't replace attached link */ 10413 if (dev_xdp_link(dev, mode)) { 10414 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 10415 return -EBUSY; 10416 } 10417 10418 /* don't allow if an upper device already has a program */ 10419 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 10420 if (dev_xdp_prog_count(upper) > 0) { 10421 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10422 return -EEXIST; 10423 } 10424 } 10425 10426 cur_prog = dev_xdp_prog(dev, mode); 10427 /* can't replace attached prog with link */ 10428 if (link && cur_prog) { 10429 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10430 return -EBUSY; 10431 } 10432 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10433 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10434 return -EEXIST; 10435 } 10436 10437 /* put effective new program into new_prog */ 10438 if (link) 10439 new_prog = link->link.prog; 10440 10441 if (new_prog) { 10442 bool offload = mode == XDP_MODE_HW; 10443 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10444 ? XDP_MODE_DRV : XDP_MODE_SKB; 10445 10446 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10447 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10448 return -EBUSY; 10449 } 10450 if (!offload && dev_xdp_prog(dev, other_mode)) { 10451 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10452 return -EEXIST; 10453 } 10454 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10455 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10456 return -EINVAL; 10457 } 10458 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10459 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10460 return -EINVAL; 10461 } 10462 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10463 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10464 return -EINVAL; 10465 } 10466 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10467 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10468 return -EINVAL; 10469 } 10470 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10471 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10472 return -EINVAL; 10473 } 10474 } 10475 10476 /* don't call drivers if the effective program didn't change */ 10477 if (new_prog != cur_prog) { 10478 bpf_op = dev_xdp_bpf_op(dev, mode); 10479 if (!bpf_op) { 10480 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10481 return -EOPNOTSUPP; 10482 } 10483 10484 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10485 if (err) 10486 return err; 10487 } 10488 10489 if (link) 10490 dev_xdp_set_link(dev, mode, link); 10491 else 10492 dev_xdp_set_prog(dev, mode, new_prog); 10493 if (cur_prog) 10494 bpf_prog_put(cur_prog); 10495 10496 return 0; 10497 } 10498 10499 static int dev_xdp_attach_link(struct net_device *dev, 10500 struct netlink_ext_ack *extack, 10501 struct bpf_xdp_link *link) 10502 { 10503 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10504 } 10505 10506 static int dev_xdp_detach_link(struct net_device *dev, 10507 struct netlink_ext_ack *extack, 10508 struct bpf_xdp_link *link) 10509 { 10510 enum bpf_xdp_mode mode; 10511 bpf_op_t bpf_op; 10512 10513 ASSERT_RTNL(); 10514 10515 mode = dev_xdp_mode(dev, link->flags); 10516 if (dev_xdp_link(dev, mode) != link) 10517 return -EINVAL; 10518 10519 bpf_op = dev_xdp_bpf_op(dev, mode); 10520 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10521 dev_xdp_set_link(dev, mode, NULL); 10522 return 0; 10523 } 10524 10525 static void bpf_xdp_link_release(struct bpf_link *link) 10526 { 10527 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10528 10529 rtnl_lock(); 10530 10531 /* if racing with net_device's tear down, xdp_link->dev might be 10532 * already NULL, in which case link was already auto-detached 10533 */ 10534 if (xdp_link->dev) { 10535 netdev_lock_ops(xdp_link->dev); 10536 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10537 netdev_unlock_ops(xdp_link->dev); 10538 xdp_link->dev = NULL; 10539 } 10540 10541 rtnl_unlock(); 10542 } 10543 10544 static int bpf_xdp_link_detach(struct bpf_link *link) 10545 { 10546 bpf_xdp_link_release(link); 10547 return 0; 10548 } 10549 10550 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10551 { 10552 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10553 10554 kfree(xdp_link); 10555 } 10556 10557 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10558 struct seq_file *seq) 10559 { 10560 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10561 u32 ifindex = 0; 10562 10563 rtnl_lock(); 10564 if (xdp_link->dev) 10565 ifindex = xdp_link->dev->ifindex; 10566 rtnl_unlock(); 10567 10568 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10569 } 10570 10571 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10572 struct bpf_link_info *info) 10573 { 10574 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10575 u32 ifindex = 0; 10576 10577 rtnl_lock(); 10578 if (xdp_link->dev) 10579 ifindex = xdp_link->dev->ifindex; 10580 rtnl_unlock(); 10581 10582 info->xdp.ifindex = ifindex; 10583 return 0; 10584 } 10585 10586 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10587 struct bpf_prog *old_prog) 10588 { 10589 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10590 enum bpf_xdp_mode mode; 10591 bpf_op_t bpf_op; 10592 int err = 0; 10593 10594 rtnl_lock(); 10595 10596 /* link might have been auto-released already, so fail */ 10597 if (!xdp_link->dev) { 10598 err = -ENOLINK; 10599 goto out_unlock; 10600 } 10601 10602 if (old_prog && link->prog != old_prog) { 10603 err = -EPERM; 10604 goto out_unlock; 10605 } 10606 old_prog = link->prog; 10607 if (old_prog->type != new_prog->type || 10608 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10609 err = -EINVAL; 10610 goto out_unlock; 10611 } 10612 10613 if (old_prog == new_prog) { 10614 /* no-op, don't disturb drivers */ 10615 bpf_prog_put(new_prog); 10616 goto out_unlock; 10617 } 10618 10619 netdev_lock_ops(xdp_link->dev); 10620 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10621 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10622 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10623 xdp_link->flags, new_prog); 10624 netdev_unlock_ops(xdp_link->dev); 10625 if (err) 10626 goto out_unlock; 10627 10628 old_prog = xchg(&link->prog, new_prog); 10629 bpf_prog_put(old_prog); 10630 10631 out_unlock: 10632 rtnl_unlock(); 10633 return err; 10634 } 10635 10636 static const struct bpf_link_ops bpf_xdp_link_lops = { 10637 .release = bpf_xdp_link_release, 10638 .dealloc = bpf_xdp_link_dealloc, 10639 .detach = bpf_xdp_link_detach, 10640 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10641 .fill_link_info = bpf_xdp_link_fill_link_info, 10642 .update_prog = bpf_xdp_link_update, 10643 }; 10644 10645 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10646 { 10647 struct net *net = current->nsproxy->net_ns; 10648 struct bpf_link_primer link_primer; 10649 struct netlink_ext_ack extack = {}; 10650 struct bpf_xdp_link *link; 10651 struct net_device *dev; 10652 int err, fd; 10653 10654 rtnl_lock(); 10655 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10656 if (!dev) { 10657 rtnl_unlock(); 10658 return -EINVAL; 10659 } 10660 10661 link = kzalloc(sizeof(*link), GFP_USER); 10662 if (!link) { 10663 err = -ENOMEM; 10664 goto unlock; 10665 } 10666 10667 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog, 10668 attr->link_create.attach_type); 10669 link->dev = dev; 10670 link->flags = attr->link_create.flags; 10671 10672 err = bpf_link_prime(&link->link, &link_primer); 10673 if (err) { 10674 kfree(link); 10675 goto unlock; 10676 } 10677 10678 netdev_lock_ops(dev); 10679 err = dev_xdp_attach_link(dev, &extack, link); 10680 netdev_unlock_ops(dev); 10681 rtnl_unlock(); 10682 10683 if (err) { 10684 link->dev = NULL; 10685 bpf_link_cleanup(&link_primer); 10686 trace_bpf_xdp_link_attach_failed(extack._msg); 10687 goto out_put_dev; 10688 } 10689 10690 fd = bpf_link_settle(&link_primer); 10691 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10692 dev_put(dev); 10693 return fd; 10694 10695 unlock: 10696 rtnl_unlock(); 10697 10698 out_put_dev: 10699 dev_put(dev); 10700 return err; 10701 } 10702 10703 /** 10704 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10705 * @dev: device 10706 * @extack: netlink extended ack 10707 * @fd: new program fd or negative value to clear 10708 * @expected_fd: old program fd that userspace expects to replace or clear 10709 * @flags: xdp-related flags 10710 * 10711 * Set or clear a bpf program for a device 10712 */ 10713 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10714 int fd, int expected_fd, u32 flags) 10715 { 10716 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10717 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10718 int err; 10719 10720 ASSERT_RTNL(); 10721 10722 if (fd >= 0) { 10723 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10724 mode != XDP_MODE_SKB); 10725 if (IS_ERR(new_prog)) 10726 return PTR_ERR(new_prog); 10727 } 10728 10729 if (expected_fd >= 0) { 10730 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10731 mode != XDP_MODE_SKB); 10732 if (IS_ERR(old_prog)) { 10733 err = PTR_ERR(old_prog); 10734 old_prog = NULL; 10735 goto err_out; 10736 } 10737 } 10738 10739 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10740 10741 err_out: 10742 if (err && new_prog) 10743 bpf_prog_put(new_prog); 10744 if (old_prog) 10745 bpf_prog_put(old_prog); 10746 return err; 10747 } 10748 10749 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10750 { 10751 int i; 10752 10753 netdev_ops_assert_locked(dev); 10754 10755 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10756 if (dev->_rx[i].mp_params.mp_priv) 10757 /* The channel count is the idx plus 1. */ 10758 return i + 1; 10759 10760 return 0; 10761 } 10762 10763 /** 10764 * dev_index_reserve() - allocate an ifindex in a namespace 10765 * @net: the applicable net namespace 10766 * @ifindex: requested ifindex, pass %0 to get one allocated 10767 * 10768 * Allocate a ifindex for a new device. Caller must either use the ifindex 10769 * to store the device (via list_netdevice()) or call dev_index_release() 10770 * to give the index up. 10771 * 10772 * Return: a suitable unique value for a new device interface number or -errno. 10773 */ 10774 static int dev_index_reserve(struct net *net, u32 ifindex) 10775 { 10776 int err; 10777 10778 if (ifindex > INT_MAX) { 10779 DEBUG_NET_WARN_ON_ONCE(1); 10780 return -EINVAL; 10781 } 10782 10783 if (!ifindex) 10784 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10785 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10786 else 10787 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10788 if (err < 0) 10789 return err; 10790 10791 return ifindex; 10792 } 10793 10794 static void dev_index_release(struct net *net, int ifindex) 10795 { 10796 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10797 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10798 } 10799 10800 static bool from_cleanup_net(void) 10801 { 10802 #ifdef CONFIG_NET_NS 10803 return current == READ_ONCE(cleanup_net_task); 10804 #else 10805 return false; 10806 #endif 10807 } 10808 10809 /* Delayed registration/unregisteration */ 10810 LIST_HEAD(net_todo_list); 10811 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10812 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10813 10814 static void net_set_todo(struct net_device *dev) 10815 { 10816 list_add_tail(&dev->todo_list, &net_todo_list); 10817 } 10818 10819 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10820 struct net_device *upper, netdev_features_t features) 10821 { 10822 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10823 netdev_features_t feature; 10824 int feature_bit; 10825 10826 for_each_netdev_feature(upper_disables, feature_bit) { 10827 feature = __NETIF_F_BIT(feature_bit); 10828 if (!(upper->wanted_features & feature) 10829 && (features & feature)) { 10830 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10831 &feature, upper->name); 10832 features &= ~feature; 10833 } 10834 } 10835 10836 return features; 10837 } 10838 10839 static void netdev_sync_lower_features(struct net_device *upper, 10840 struct net_device *lower, netdev_features_t features) 10841 { 10842 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10843 netdev_features_t feature; 10844 int feature_bit; 10845 10846 for_each_netdev_feature(upper_disables, feature_bit) { 10847 feature = __NETIF_F_BIT(feature_bit); 10848 if (!(features & feature) && (lower->features & feature)) { 10849 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10850 &feature, lower->name); 10851 netdev_lock_ops(lower); 10852 lower->wanted_features &= ~feature; 10853 __netdev_update_features(lower); 10854 10855 if (unlikely(lower->features & feature)) 10856 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10857 &feature, lower->name); 10858 else 10859 netdev_features_change(lower); 10860 netdev_unlock_ops(lower); 10861 } 10862 } 10863 } 10864 10865 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10866 { 10867 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10868 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10869 bool hw_csum = features & NETIF_F_HW_CSUM; 10870 10871 return ip_csum || hw_csum; 10872 } 10873 10874 static netdev_features_t netdev_fix_features(struct net_device *dev, 10875 netdev_features_t features) 10876 { 10877 /* Fix illegal checksum combinations */ 10878 if ((features & NETIF_F_HW_CSUM) && 10879 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10880 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10881 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10882 } 10883 10884 /* TSO requires that SG is present as well. */ 10885 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10886 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10887 features &= ~NETIF_F_ALL_TSO; 10888 } 10889 10890 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10891 !(features & NETIF_F_IP_CSUM)) { 10892 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10893 features &= ~NETIF_F_TSO; 10894 features &= ~NETIF_F_TSO_ECN; 10895 } 10896 10897 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10898 !(features & NETIF_F_IPV6_CSUM)) { 10899 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10900 features &= ~NETIF_F_TSO6; 10901 } 10902 10903 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10904 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10905 features &= ~NETIF_F_TSO_MANGLEID; 10906 10907 /* TSO ECN requires that TSO is present as well. */ 10908 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10909 features &= ~NETIF_F_TSO_ECN; 10910 10911 /* Software GSO depends on SG. */ 10912 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10913 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10914 features &= ~NETIF_F_GSO; 10915 } 10916 10917 /* GSO partial features require GSO partial be set */ 10918 if ((features & dev->gso_partial_features) && 10919 !(features & NETIF_F_GSO_PARTIAL)) { 10920 netdev_dbg(dev, 10921 "Dropping partially supported GSO features since no GSO partial.\n"); 10922 features &= ~dev->gso_partial_features; 10923 } 10924 10925 if (!(features & NETIF_F_RXCSUM)) { 10926 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10927 * successfully merged by hardware must also have the 10928 * checksum verified by hardware. If the user does not 10929 * want to enable RXCSUM, logically, we should disable GRO_HW. 10930 */ 10931 if (features & NETIF_F_GRO_HW) { 10932 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10933 features &= ~NETIF_F_GRO_HW; 10934 } 10935 } 10936 10937 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10938 if (features & NETIF_F_RXFCS) { 10939 if (features & NETIF_F_LRO) { 10940 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10941 features &= ~NETIF_F_LRO; 10942 } 10943 10944 if (features & NETIF_F_GRO_HW) { 10945 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10946 features &= ~NETIF_F_GRO_HW; 10947 } 10948 } 10949 10950 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10951 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10952 features &= ~NETIF_F_LRO; 10953 } 10954 10955 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10956 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10957 features &= ~NETIF_F_HW_TLS_TX; 10958 } 10959 10960 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10961 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10962 features &= ~NETIF_F_HW_TLS_RX; 10963 } 10964 10965 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10966 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10967 features &= ~NETIF_F_GSO_UDP_L4; 10968 } 10969 10970 return features; 10971 } 10972 10973 int __netdev_update_features(struct net_device *dev) 10974 { 10975 struct net_device *upper, *lower; 10976 netdev_features_t features; 10977 struct list_head *iter; 10978 int err = -1; 10979 10980 ASSERT_RTNL(); 10981 netdev_ops_assert_locked(dev); 10982 10983 features = netdev_get_wanted_features(dev); 10984 10985 if (dev->netdev_ops->ndo_fix_features) 10986 features = dev->netdev_ops->ndo_fix_features(dev, features); 10987 10988 /* driver might be less strict about feature dependencies */ 10989 features = netdev_fix_features(dev, features); 10990 10991 /* some features can't be enabled if they're off on an upper device */ 10992 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10993 features = netdev_sync_upper_features(dev, upper, features); 10994 10995 if (dev->features == features) 10996 goto sync_lower; 10997 10998 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10999 &dev->features, &features); 11000 11001 if (dev->netdev_ops->ndo_set_features) 11002 err = dev->netdev_ops->ndo_set_features(dev, features); 11003 else 11004 err = 0; 11005 11006 if (unlikely(err < 0)) { 11007 netdev_err(dev, 11008 "set_features() failed (%d); wanted %pNF, left %pNF\n", 11009 err, &features, &dev->features); 11010 /* return non-0 since some features might have changed and 11011 * it's better to fire a spurious notification than miss it 11012 */ 11013 return -1; 11014 } 11015 11016 sync_lower: 11017 /* some features must be disabled on lower devices when disabled 11018 * on an upper device (think: bonding master or bridge) 11019 */ 11020 netdev_for_each_lower_dev(dev, lower, iter) 11021 netdev_sync_lower_features(dev, lower, features); 11022 11023 if (!err) { 11024 netdev_features_t diff = features ^ dev->features; 11025 11026 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 11027 /* udp_tunnel_{get,drop}_rx_info both need 11028 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 11029 * device, or they won't do anything. 11030 * Thus we need to update dev->features 11031 * *before* calling udp_tunnel_get_rx_info, 11032 * but *after* calling udp_tunnel_drop_rx_info. 11033 */ 11034 udp_tunnel_nic_lock(dev); 11035 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 11036 dev->features = features; 11037 udp_tunnel_get_rx_info(dev); 11038 } else { 11039 udp_tunnel_drop_rx_info(dev); 11040 } 11041 udp_tunnel_nic_unlock(dev); 11042 } 11043 11044 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 11045 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 11046 dev->features = features; 11047 err |= vlan_get_rx_ctag_filter_info(dev); 11048 } else { 11049 vlan_drop_rx_ctag_filter_info(dev); 11050 } 11051 } 11052 11053 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 11054 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 11055 dev->features = features; 11056 err |= vlan_get_rx_stag_filter_info(dev); 11057 } else { 11058 vlan_drop_rx_stag_filter_info(dev); 11059 } 11060 } 11061 11062 dev->features = features; 11063 } 11064 11065 return err < 0 ? 0 : 1; 11066 } 11067 11068 /** 11069 * netdev_update_features - recalculate device features 11070 * @dev: the device to check 11071 * 11072 * Recalculate dev->features set and send notifications if it 11073 * has changed. Should be called after driver or hardware dependent 11074 * conditions might have changed that influence the features. 11075 */ 11076 void netdev_update_features(struct net_device *dev) 11077 { 11078 if (__netdev_update_features(dev)) 11079 netdev_features_change(dev); 11080 } 11081 EXPORT_SYMBOL(netdev_update_features); 11082 11083 /** 11084 * netdev_change_features - recalculate device features 11085 * @dev: the device to check 11086 * 11087 * Recalculate dev->features set and send notifications even 11088 * if they have not changed. Should be called instead of 11089 * netdev_update_features() if also dev->vlan_features might 11090 * have changed to allow the changes to be propagated to stacked 11091 * VLAN devices. 11092 */ 11093 void netdev_change_features(struct net_device *dev) 11094 { 11095 __netdev_update_features(dev); 11096 netdev_features_change(dev); 11097 } 11098 EXPORT_SYMBOL(netdev_change_features); 11099 11100 /** 11101 * netif_stacked_transfer_operstate - transfer operstate 11102 * @rootdev: the root or lower level device to transfer state from 11103 * @dev: the device to transfer operstate to 11104 * 11105 * Transfer operational state from root to device. This is normally 11106 * called when a stacking relationship exists between the root 11107 * device and the device(a leaf device). 11108 */ 11109 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 11110 struct net_device *dev) 11111 { 11112 if (rootdev->operstate == IF_OPER_DORMANT) 11113 netif_dormant_on(dev); 11114 else 11115 netif_dormant_off(dev); 11116 11117 if (rootdev->operstate == IF_OPER_TESTING) 11118 netif_testing_on(dev); 11119 else 11120 netif_testing_off(dev); 11121 11122 if (netif_carrier_ok(rootdev)) 11123 netif_carrier_on(dev); 11124 else 11125 netif_carrier_off(dev); 11126 } 11127 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 11128 11129 static int netif_alloc_rx_queues(struct net_device *dev) 11130 { 11131 unsigned int i, count = dev->num_rx_queues; 11132 struct netdev_rx_queue *rx; 11133 size_t sz = count * sizeof(*rx); 11134 int err = 0; 11135 11136 BUG_ON(count < 1); 11137 11138 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11139 if (!rx) 11140 return -ENOMEM; 11141 11142 dev->_rx = rx; 11143 11144 for (i = 0; i < count; i++) { 11145 rx[i].dev = dev; 11146 11147 /* XDP RX-queue setup */ 11148 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 11149 if (err < 0) 11150 goto err_rxq_info; 11151 } 11152 return 0; 11153 11154 err_rxq_info: 11155 /* Rollback successful reg's and free other resources */ 11156 while (i--) 11157 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 11158 kvfree(dev->_rx); 11159 dev->_rx = NULL; 11160 return err; 11161 } 11162 11163 static void netif_free_rx_queues(struct net_device *dev) 11164 { 11165 unsigned int i, count = dev->num_rx_queues; 11166 11167 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 11168 if (!dev->_rx) 11169 return; 11170 11171 for (i = 0; i < count; i++) 11172 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 11173 11174 kvfree(dev->_rx); 11175 } 11176 11177 static void netdev_init_one_queue(struct net_device *dev, 11178 struct netdev_queue *queue, void *_unused) 11179 { 11180 /* Initialize queue lock */ 11181 spin_lock_init(&queue->_xmit_lock); 11182 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 11183 queue->xmit_lock_owner = -1; 11184 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 11185 queue->dev = dev; 11186 #ifdef CONFIG_BQL 11187 dql_init(&queue->dql, HZ); 11188 #endif 11189 } 11190 11191 static void netif_free_tx_queues(struct net_device *dev) 11192 { 11193 kvfree(dev->_tx); 11194 } 11195 11196 static int netif_alloc_netdev_queues(struct net_device *dev) 11197 { 11198 unsigned int count = dev->num_tx_queues; 11199 struct netdev_queue *tx; 11200 size_t sz = count * sizeof(*tx); 11201 11202 if (count < 1 || count > 0xffff) 11203 return -EINVAL; 11204 11205 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11206 if (!tx) 11207 return -ENOMEM; 11208 11209 dev->_tx = tx; 11210 11211 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 11212 spin_lock_init(&dev->tx_global_lock); 11213 11214 return 0; 11215 } 11216 11217 void netif_tx_stop_all_queues(struct net_device *dev) 11218 { 11219 unsigned int i; 11220 11221 for (i = 0; i < dev->num_tx_queues; i++) { 11222 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 11223 11224 netif_tx_stop_queue(txq); 11225 } 11226 } 11227 EXPORT_SYMBOL(netif_tx_stop_all_queues); 11228 11229 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 11230 { 11231 void __percpu *v; 11232 11233 /* Drivers implementing ndo_get_peer_dev must support tstat 11234 * accounting, so that skb_do_redirect() can bump the dev's 11235 * RX stats upon network namespace switch. 11236 */ 11237 if (dev->netdev_ops->ndo_get_peer_dev && 11238 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 11239 return -EOPNOTSUPP; 11240 11241 switch (dev->pcpu_stat_type) { 11242 case NETDEV_PCPU_STAT_NONE: 11243 return 0; 11244 case NETDEV_PCPU_STAT_LSTATS: 11245 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 11246 break; 11247 case NETDEV_PCPU_STAT_TSTATS: 11248 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 11249 break; 11250 case NETDEV_PCPU_STAT_DSTATS: 11251 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 11252 break; 11253 default: 11254 return -EINVAL; 11255 } 11256 11257 return v ? 0 : -ENOMEM; 11258 } 11259 11260 static void netdev_do_free_pcpu_stats(struct net_device *dev) 11261 { 11262 switch (dev->pcpu_stat_type) { 11263 case NETDEV_PCPU_STAT_NONE: 11264 return; 11265 case NETDEV_PCPU_STAT_LSTATS: 11266 free_percpu(dev->lstats); 11267 break; 11268 case NETDEV_PCPU_STAT_TSTATS: 11269 free_percpu(dev->tstats); 11270 break; 11271 case NETDEV_PCPU_STAT_DSTATS: 11272 free_percpu(dev->dstats); 11273 break; 11274 } 11275 } 11276 11277 static void netdev_free_phy_link_topology(struct net_device *dev) 11278 { 11279 struct phy_link_topology *topo = dev->link_topo; 11280 11281 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 11282 xa_destroy(&topo->phys); 11283 kfree(topo); 11284 dev->link_topo = NULL; 11285 } 11286 } 11287 11288 /** 11289 * register_netdevice() - register a network device 11290 * @dev: device to register 11291 * 11292 * Take a prepared network device structure and make it externally accessible. 11293 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 11294 * Callers must hold the rtnl lock - you may want register_netdev() 11295 * instead of this. 11296 */ 11297 int register_netdevice(struct net_device *dev) 11298 { 11299 int ret; 11300 struct net *net = dev_net(dev); 11301 11302 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 11303 NETDEV_FEATURE_COUNT); 11304 BUG_ON(dev_boot_phase); 11305 ASSERT_RTNL(); 11306 11307 might_sleep(); 11308 11309 /* When net_device's are persistent, this will be fatal. */ 11310 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 11311 BUG_ON(!net); 11312 11313 ret = ethtool_check_ops(dev->ethtool_ops); 11314 if (ret) 11315 return ret; 11316 11317 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 11318 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 11319 mutex_init(&dev->ethtool->rss_lock); 11320 11321 spin_lock_init(&dev->addr_list_lock); 11322 netdev_set_addr_lockdep_class(dev); 11323 11324 ret = dev_get_valid_name(net, dev, dev->name); 11325 if (ret < 0) 11326 goto out; 11327 11328 ret = -ENOMEM; 11329 dev->name_node = netdev_name_node_head_alloc(dev); 11330 if (!dev->name_node) 11331 goto out; 11332 11333 /* Init, if this function is available */ 11334 if (dev->netdev_ops->ndo_init) { 11335 ret = dev->netdev_ops->ndo_init(dev); 11336 if (ret) { 11337 if (ret > 0) 11338 ret = -EIO; 11339 goto err_free_name; 11340 } 11341 } 11342 11343 if (((dev->hw_features | dev->features) & 11344 NETIF_F_HW_VLAN_CTAG_FILTER) && 11345 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 11346 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 11347 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 11348 ret = -EINVAL; 11349 goto err_uninit; 11350 } 11351 11352 ret = netdev_do_alloc_pcpu_stats(dev); 11353 if (ret) 11354 goto err_uninit; 11355 11356 ret = dev_index_reserve(net, dev->ifindex); 11357 if (ret < 0) 11358 goto err_free_pcpu; 11359 dev->ifindex = ret; 11360 11361 /* Transfer changeable features to wanted_features and enable 11362 * software offloads (GSO and GRO). 11363 */ 11364 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 11365 dev->features |= NETIF_F_SOFT_FEATURES; 11366 11367 if (dev->udp_tunnel_nic_info) { 11368 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11369 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11370 } 11371 11372 dev->wanted_features = dev->features & dev->hw_features; 11373 11374 if (!(dev->flags & IFF_LOOPBACK)) 11375 dev->hw_features |= NETIF_F_NOCACHE_COPY; 11376 11377 /* If IPv4 TCP segmentation offload is supported we should also 11378 * allow the device to enable segmenting the frame with the option 11379 * of ignoring a static IP ID value. This doesn't enable the 11380 * feature itself but allows the user to enable it later. 11381 */ 11382 if (dev->hw_features & NETIF_F_TSO) 11383 dev->hw_features |= NETIF_F_TSO_MANGLEID; 11384 if (dev->vlan_features & NETIF_F_TSO) 11385 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 11386 if (dev->mpls_features & NETIF_F_TSO) 11387 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 11388 if (dev->hw_enc_features & NETIF_F_TSO) 11389 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 11390 11391 /* TSO_MANGLEID belongs in mangleid_features by definition */ 11392 dev->mangleid_features |= NETIF_F_TSO_MANGLEID; 11393 11394 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 11395 */ 11396 dev->vlan_features |= NETIF_F_HIGHDMA; 11397 11398 /* Make NETIF_F_SG inheritable to tunnel devices. 11399 */ 11400 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 11401 11402 /* Make NETIF_F_SG inheritable to MPLS. 11403 */ 11404 dev->mpls_features |= NETIF_F_SG; 11405 11406 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 11407 ret = notifier_to_errno(ret); 11408 if (ret) 11409 goto err_ifindex_release; 11410 11411 ret = netdev_register_kobject(dev); 11412 11413 netdev_lock(dev); 11414 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 11415 netdev_unlock(dev); 11416 11417 if (ret) 11418 goto err_uninit_notify; 11419 11420 netdev_lock_ops(dev); 11421 __netdev_update_features(dev); 11422 netdev_unlock_ops(dev); 11423 11424 /* 11425 * Default initial state at registry is that the 11426 * device is present. 11427 */ 11428 11429 set_bit(__LINK_STATE_PRESENT, &dev->state); 11430 11431 linkwatch_init_dev(dev); 11432 11433 dev_init_scheduler(dev); 11434 11435 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11436 list_netdevice(dev); 11437 11438 add_device_randomness(dev->dev_addr, dev->addr_len); 11439 11440 /* If the device has permanent device address, driver should 11441 * set dev_addr and also addr_assign_type should be set to 11442 * NET_ADDR_PERM (default value). 11443 */ 11444 if (dev->addr_assign_type == NET_ADDR_PERM) 11445 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11446 11447 /* Notify protocols, that a new device appeared. */ 11448 netdev_lock_ops(dev); 11449 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11450 netdev_unlock_ops(dev); 11451 ret = notifier_to_errno(ret); 11452 if (ret) { 11453 /* Expect explicit free_netdev() on failure */ 11454 dev->needs_free_netdev = false; 11455 unregister_netdevice_queue(dev, NULL); 11456 goto out; 11457 } 11458 /* 11459 * Prevent userspace races by waiting until the network 11460 * device is fully setup before sending notifications. 11461 */ 11462 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 11463 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11464 11465 out: 11466 return ret; 11467 11468 err_uninit_notify: 11469 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11470 err_ifindex_release: 11471 dev_index_release(net, dev->ifindex); 11472 err_free_pcpu: 11473 netdev_do_free_pcpu_stats(dev); 11474 err_uninit: 11475 if (dev->netdev_ops->ndo_uninit) 11476 dev->netdev_ops->ndo_uninit(dev); 11477 if (dev->priv_destructor) 11478 dev->priv_destructor(dev); 11479 err_free_name: 11480 netdev_name_node_free(dev->name_node); 11481 goto out; 11482 } 11483 EXPORT_SYMBOL(register_netdevice); 11484 11485 /* Initialize the core of a dummy net device. 11486 * The setup steps dummy netdevs need which normal netdevs get by going 11487 * through register_netdevice(). 11488 */ 11489 static void init_dummy_netdev(struct net_device *dev) 11490 { 11491 /* make sure we BUG if trying to hit standard 11492 * register/unregister code path 11493 */ 11494 dev->reg_state = NETREG_DUMMY; 11495 11496 /* a dummy interface is started by default */ 11497 set_bit(__LINK_STATE_PRESENT, &dev->state); 11498 set_bit(__LINK_STATE_START, &dev->state); 11499 11500 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11501 * because users of this 'device' dont need to change 11502 * its refcount. 11503 */ 11504 } 11505 11506 /** 11507 * register_netdev - register a network device 11508 * @dev: device to register 11509 * 11510 * Take a completed network device structure and add it to the kernel 11511 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11512 * chain. 0 is returned on success. A negative errno code is returned 11513 * on a failure to set up the device, or if the name is a duplicate. 11514 * 11515 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11516 * and expands the device name if you passed a format string to 11517 * alloc_netdev. 11518 */ 11519 int register_netdev(struct net_device *dev) 11520 { 11521 struct net *net = dev_net(dev); 11522 int err; 11523 11524 if (rtnl_net_lock_killable(net)) 11525 return -EINTR; 11526 11527 err = register_netdevice(dev); 11528 11529 rtnl_net_unlock(net); 11530 11531 return err; 11532 } 11533 EXPORT_SYMBOL(register_netdev); 11534 11535 int netdev_refcnt_read(const struct net_device *dev) 11536 { 11537 #ifdef CONFIG_PCPU_DEV_REFCNT 11538 int i, refcnt = 0; 11539 11540 for_each_possible_cpu(i) 11541 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11542 return refcnt; 11543 #else 11544 return refcount_read(&dev->dev_refcnt); 11545 #endif 11546 } 11547 EXPORT_SYMBOL(netdev_refcnt_read); 11548 11549 int netdev_unregister_timeout_secs __read_mostly = 10; 11550 11551 #define WAIT_REFS_MIN_MSECS 1 11552 #define WAIT_REFS_MAX_MSECS 250 11553 /** 11554 * netdev_wait_allrefs_any - wait until all references are gone. 11555 * @list: list of net_devices to wait on 11556 * 11557 * This is called when unregistering network devices. 11558 * 11559 * Any protocol or device that holds a reference should register 11560 * for netdevice notification, and cleanup and put back the 11561 * reference if they receive an UNREGISTER event. 11562 * We can get stuck here if buggy protocols don't correctly 11563 * call dev_put. 11564 */ 11565 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11566 { 11567 unsigned long rebroadcast_time, warning_time; 11568 struct net_device *dev; 11569 int wait = 0; 11570 11571 rebroadcast_time = warning_time = jiffies; 11572 11573 list_for_each_entry(dev, list, todo_list) 11574 if (netdev_refcnt_read(dev) == 1) 11575 return dev; 11576 11577 while (true) { 11578 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11579 rtnl_lock(); 11580 11581 /* Rebroadcast unregister notification */ 11582 list_for_each_entry(dev, list, todo_list) 11583 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11584 11585 __rtnl_unlock(); 11586 rcu_barrier(); 11587 rtnl_lock(); 11588 11589 list_for_each_entry(dev, list, todo_list) 11590 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11591 &dev->state)) { 11592 /* We must not have linkwatch events 11593 * pending on unregister. If this 11594 * happens, we simply run the queue 11595 * unscheduled, resulting in a noop 11596 * for this device. 11597 */ 11598 linkwatch_run_queue(); 11599 break; 11600 } 11601 11602 __rtnl_unlock(); 11603 11604 rebroadcast_time = jiffies; 11605 } 11606 11607 rcu_barrier(); 11608 11609 if (!wait) { 11610 wait = WAIT_REFS_MIN_MSECS; 11611 } else { 11612 msleep(wait); 11613 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11614 } 11615 11616 list_for_each_entry(dev, list, todo_list) 11617 if (netdev_refcnt_read(dev) == 1) 11618 return dev; 11619 11620 if (time_after(jiffies, warning_time + 11621 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11622 list_for_each_entry(dev, list, todo_list) { 11623 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11624 dev->name, netdev_refcnt_read(dev)); 11625 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11626 } 11627 11628 warning_time = jiffies; 11629 } 11630 } 11631 } 11632 11633 /* The sequence is: 11634 * 11635 * rtnl_lock(); 11636 * ... 11637 * register_netdevice(x1); 11638 * register_netdevice(x2); 11639 * ... 11640 * unregister_netdevice(y1); 11641 * unregister_netdevice(y2); 11642 * ... 11643 * rtnl_unlock(); 11644 * free_netdev(y1); 11645 * free_netdev(y2); 11646 * 11647 * We are invoked by rtnl_unlock(). 11648 * This allows us to deal with problems: 11649 * 1) We can delete sysfs objects which invoke hotplug 11650 * without deadlocking with linkwatch via keventd. 11651 * 2) Since we run with the RTNL semaphore not held, we can sleep 11652 * safely in order to wait for the netdev refcnt to drop to zero. 11653 * 11654 * We must not return until all unregister events added during 11655 * the interval the lock was held have been completed. 11656 */ 11657 void netdev_run_todo(void) 11658 { 11659 struct net_device *dev, *tmp; 11660 struct list_head list; 11661 int cnt; 11662 #ifdef CONFIG_LOCKDEP 11663 struct list_head unlink_list; 11664 11665 list_replace_init(&net_unlink_list, &unlink_list); 11666 11667 while (!list_empty(&unlink_list)) { 11668 dev = list_first_entry(&unlink_list, struct net_device, 11669 unlink_list); 11670 list_del_init(&dev->unlink_list); 11671 dev->nested_level = dev->lower_level - 1; 11672 } 11673 #endif 11674 11675 /* Snapshot list, allow later requests */ 11676 list_replace_init(&net_todo_list, &list); 11677 11678 __rtnl_unlock(); 11679 11680 /* Wait for rcu callbacks to finish before next phase */ 11681 if (!list_empty(&list)) 11682 rcu_barrier(); 11683 11684 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11685 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11686 netdev_WARN(dev, "run_todo but not unregistering\n"); 11687 list_del(&dev->todo_list); 11688 continue; 11689 } 11690 11691 netdev_lock(dev); 11692 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11693 netdev_unlock(dev); 11694 linkwatch_sync_dev(dev); 11695 } 11696 11697 cnt = 0; 11698 while (!list_empty(&list)) { 11699 dev = netdev_wait_allrefs_any(&list); 11700 list_del(&dev->todo_list); 11701 11702 /* paranoia */ 11703 BUG_ON(netdev_refcnt_read(dev) != 1); 11704 BUG_ON(!list_empty(&dev->ptype_all)); 11705 BUG_ON(!list_empty(&dev->ptype_specific)); 11706 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11707 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11708 11709 netdev_do_free_pcpu_stats(dev); 11710 if (dev->priv_destructor) 11711 dev->priv_destructor(dev); 11712 if (dev->needs_free_netdev) 11713 free_netdev(dev); 11714 11715 cnt++; 11716 11717 /* Free network device */ 11718 kobject_put(&dev->dev.kobj); 11719 } 11720 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11721 wake_up(&netdev_unregistering_wq); 11722 } 11723 11724 /* Collate per-cpu network dstats statistics 11725 * 11726 * Read per-cpu network statistics from dev->dstats and populate the related 11727 * fields in @s. 11728 */ 11729 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11730 const struct pcpu_dstats __percpu *dstats) 11731 { 11732 int cpu; 11733 11734 for_each_possible_cpu(cpu) { 11735 u64 rx_packets, rx_bytes, rx_drops; 11736 u64 tx_packets, tx_bytes, tx_drops; 11737 const struct pcpu_dstats *stats; 11738 unsigned int start; 11739 11740 stats = per_cpu_ptr(dstats, cpu); 11741 do { 11742 start = u64_stats_fetch_begin(&stats->syncp); 11743 rx_packets = u64_stats_read(&stats->rx_packets); 11744 rx_bytes = u64_stats_read(&stats->rx_bytes); 11745 rx_drops = u64_stats_read(&stats->rx_drops); 11746 tx_packets = u64_stats_read(&stats->tx_packets); 11747 tx_bytes = u64_stats_read(&stats->tx_bytes); 11748 tx_drops = u64_stats_read(&stats->tx_drops); 11749 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11750 11751 s->rx_packets += rx_packets; 11752 s->rx_bytes += rx_bytes; 11753 s->rx_dropped += rx_drops; 11754 s->tx_packets += tx_packets; 11755 s->tx_bytes += tx_bytes; 11756 s->tx_dropped += tx_drops; 11757 } 11758 } 11759 11760 /* ndo_get_stats64 implementation for dtstats-based accounting. 11761 * 11762 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11763 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11764 */ 11765 static void dev_get_dstats64(const struct net_device *dev, 11766 struct rtnl_link_stats64 *s) 11767 { 11768 netdev_stats_to_stats64(s, &dev->stats); 11769 dev_fetch_dstats(s, dev->dstats); 11770 } 11771 11772 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11773 * all the same fields in the same order as net_device_stats, with only 11774 * the type differing, but rtnl_link_stats64 may have additional fields 11775 * at the end for newer counters. 11776 */ 11777 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11778 const struct net_device_stats *netdev_stats) 11779 { 11780 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11781 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11782 u64 *dst = (u64 *)stats64; 11783 11784 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11785 for (i = 0; i < n; i++) 11786 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11787 /* zero out counters that only exist in rtnl_link_stats64 */ 11788 memset((char *)stats64 + n * sizeof(u64), 0, 11789 sizeof(*stats64) - n * sizeof(u64)); 11790 } 11791 EXPORT_SYMBOL(netdev_stats_to_stats64); 11792 11793 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11794 struct net_device *dev) 11795 { 11796 struct net_device_core_stats __percpu *p; 11797 11798 p = alloc_percpu_gfp(struct net_device_core_stats, 11799 GFP_ATOMIC | __GFP_NOWARN); 11800 11801 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11802 free_percpu(p); 11803 11804 /* This READ_ONCE() pairs with the cmpxchg() above */ 11805 return READ_ONCE(dev->core_stats); 11806 } 11807 11808 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11809 { 11810 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11811 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11812 unsigned long __percpu *field; 11813 11814 if (unlikely(!p)) { 11815 p = netdev_core_stats_alloc(dev); 11816 if (!p) 11817 return; 11818 } 11819 11820 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11821 this_cpu_inc(*field); 11822 } 11823 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11824 11825 /** 11826 * dev_get_stats - get network device statistics 11827 * @dev: device to get statistics from 11828 * @storage: place to store stats 11829 * 11830 * Get network statistics from device. Return @storage. 11831 * The device driver may provide its own method by setting 11832 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11833 * otherwise the internal statistics structure is used. 11834 */ 11835 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11836 struct rtnl_link_stats64 *storage) 11837 { 11838 const struct net_device_ops *ops = dev->netdev_ops; 11839 const struct net_device_core_stats __percpu *p; 11840 11841 /* 11842 * IPv{4,6} and udp tunnels share common stat helpers and use 11843 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11844 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11845 */ 11846 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11847 offsetof(struct pcpu_dstats, rx_bytes)); 11848 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11849 offsetof(struct pcpu_dstats, rx_packets)); 11850 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11851 offsetof(struct pcpu_dstats, tx_bytes)); 11852 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11853 offsetof(struct pcpu_dstats, tx_packets)); 11854 11855 if (ops->ndo_get_stats64) { 11856 memset(storage, 0, sizeof(*storage)); 11857 ops->ndo_get_stats64(dev, storage); 11858 } else if (ops->ndo_get_stats) { 11859 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11860 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11861 dev_get_tstats64(dev, storage); 11862 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11863 dev_get_dstats64(dev, storage); 11864 } else { 11865 netdev_stats_to_stats64(storage, &dev->stats); 11866 } 11867 11868 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11869 p = READ_ONCE(dev->core_stats); 11870 if (p) { 11871 const struct net_device_core_stats *core_stats; 11872 int i; 11873 11874 for_each_possible_cpu(i) { 11875 core_stats = per_cpu_ptr(p, i); 11876 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11877 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11878 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11879 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11880 } 11881 } 11882 return storage; 11883 } 11884 EXPORT_SYMBOL(dev_get_stats); 11885 11886 /** 11887 * dev_fetch_sw_netstats - get per-cpu network device statistics 11888 * @s: place to store stats 11889 * @netstats: per-cpu network stats to read from 11890 * 11891 * Read per-cpu network statistics and populate the related fields in @s. 11892 */ 11893 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11894 const struct pcpu_sw_netstats __percpu *netstats) 11895 { 11896 int cpu; 11897 11898 for_each_possible_cpu(cpu) { 11899 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11900 const struct pcpu_sw_netstats *stats; 11901 unsigned int start; 11902 11903 stats = per_cpu_ptr(netstats, cpu); 11904 do { 11905 start = u64_stats_fetch_begin(&stats->syncp); 11906 rx_packets = u64_stats_read(&stats->rx_packets); 11907 rx_bytes = u64_stats_read(&stats->rx_bytes); 11908 tx_packets = u64_stats_read(&stats->tx_packets); 11909 tx_bytes = u64_stats_read(&stats->tx_bytes); 11910 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11911 11912 s->rx_packets += rx_packets; 11913 s->rx_bytes += rx_bytes; 11914 s->tx_packets += tx_packets; 11915 s->tx_bytes += tx_bytes; 11916 } 11917 } 11918 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11919 11920 /** 11921 * dev_get_tstats64 - ndo_get_stats64 implementation 11922 * @dev: device to get statistics from 11923 * @s: place to store stats 11924 * 11925 * Populate @s from dev->stats and dev->tstats. Can be used as 11926 * ndo_get_stats64() callback. 11927 */ 11928 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11929 { 11930 netdev_stats_to_stats64(s, &dev->stats); 11931 dev_fetch_sw_netstats(s, dev->tstats); 11932 } 11933 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11934 11935 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11936 { 11937 struct netdev_queue *queue = dev_ingress_queue(dev); 11938 11939 #ifdef CONFIG_NET_CLS_ACT 11940 if (queue) 11941 return queue; 11942 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11943 if (!queue) 11944 return NULL; 11945 netdev_init_one_queue(dev, queue, NULL); 11946 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11947 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11948 rcu_assign_pointer(dev->ingress_queue, queue); 11949 #endif 11950 return queue; 11951 } 11952 11953 static const struct ethtool_ops default_ethtool_ops; 11954 11955 void netdev_set_default_ethtool_ops(struct net_device *dev, 11956 const struct ethtool_ops *ops) 11957 { 11958 if (dev->ethtool_ops == &default_ethtool_ops) 11959 dev->ethtool_ops = ops; 11960 } 11961 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11962 11963 /** 11964 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11965 * @dev: netdev to enable the IRQ coalescing on 11966 * 11967 * Sets a conservative default for SW IRQ coalescing. Users can use 11968 * sysfs attributes to override the default values. 11969 */ 11970 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11971 { 11972 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11973 11974 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11975 netdev_set_gro_flush_timeout(dev, 20000); 11976 netdev_set_defer_hard_irqs(dev, 1); 11977 } 11978 } 11979 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11980 11981 /** 11982 * alloc_netdev_mqs - allocate network device 11983 * @sizeof_priv: size of private data to allocate space for 11984 * @name: device name format string 11985 * @name_assign_type: origin of device name 11986 * @setup: callback to initialize device 11987 * @txqs: the number of TX subqueues to allocate 11988 * @rxqs: the number of RX subqueues to allocate 11989 * 11990 * Allocates a struct net_device with private data area for driver use 11991 * and performs basic initialization. Also allocates subqueue structs 11992 * for each queue on the device. 11993 */ 11994 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11995 unsigned char name_assign_type, 11996 void (*setup)(struct net_device *), 11997 unsigned int txqs, unsigned int rxqs) 11998 { 11999 struct net_device *dev; 12000 size_t napi_config_sz; 12001 unsigned int maxqs; 12002 12003 BUG_ON(strlen(name) >= sizeof(dev->name)); 12004 12005 if (txqs < 1) { 12006 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 12007 return NULL; 12008 } 12009 12010 if (rxqs < 1) { 12011 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 12012 return NULL; 12013 } 12014 12015 maxqs = max(txqs, rxqs); 12016 12017 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 12018 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 12019 if (!dev) 12020 return NULL; 12021 12022 dev->priv_len = sizeof_priv; 12023 12024 ref_tracker_dir_init(&dev->refcnt_tracker, 128, "netdev"); 12025 #ifdef CONFIG_PCPU_DEV_REFCNT 12026 dev->pcpu_refcnt = alloc_percpu(int); 12027 if (!dev->pcpu_refcnt) 12028 goto free_dev; 12029 __dev_hold(dev); 12030 #else 12031 refcount_set(&dev->dev_refcnt, 1); 12032 #endif 12033 12034 if (dev_addr_init(dev)) 12035 goto free_pcpu; 12036 12037 dev_mc_init(dev); 12038 dev_uc_init(dev); 12039 12040 dev_net_set(dev, &init_net); 12041 12042 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 12043 dev->xdp_zc_max_segs = 1; 12044 dev->gso_max_segs = GSO_MAX_SEGS; 12045 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 12046 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 12047 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 12048 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 12049 dev->tso_max_segs = TSO_MAX_SEGS; 12050 dev->upper_level = 1; 12051 dev->lower_level = 1; 12052 #ifdef CONFIG_LOCKDEP 12053 dev->nested_level = 0; 12054 INIT_LIST_HEAD(&dev->unlink_list); 12055 #endif 12056 12057 INIT_LIST_HEAD(&dev->napi_list); 12058 INIT_LIST_HEAD(&dev->unreg_list); 12059 INIT_LIST_HEAD(&dev->close_list); 12060 INIT_LIST_HEAD(&dev->link_watch_list); 12061 INIT_LIST_HEAD(&dev->adj_list.upper); 12062 INIT_LIST_HEAD(&dev->adj_list.lower); 12063 INIT_LIST_HEAD(&dev->ptype_all); 12064 INIT_LIST_HEAD(&dev->ptype_specific); 12065 INIT_LIST_HEAD(&dev->net_notifier_list); 12066 #ifdef CONFIG_NET_SCHED 12067 hash_init(dev->qdisc_hash); 12068 #endif 12069 12070 mutex_init(&dev->lock); 12071 12072 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12073 setup(dev); 12074 12075 if (!dev->tx_queue_len) { 12076 dev->priv_flags |= IFF_NO_QUEUE; 12077 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 12078 } 12079 12080 dev->num_tx_queues = txqs; 12081 dev->real_num_tx_queues = txqs; 12082 if (netif_alloc_netdev_queues(dev)) 12083 goto free_all; 12084 12085 dev->num_rx_queues = rxqs; 12086 dev->real_num_rx_queues = rxqs; 12087 if (netif_alloc_rx_queues(dev)) 12088 goto free_all; 12089 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 12090 if (!dev->ethtool) 12091 goto free_all; 12092 12093 dev->cfg = kzalloc(sizeof(*dev->cfg), GFP_KERNEL_ACCOUNT); 12094 if (!dev->cfg) 12095 goto free_all; 12096 dev->cfg_pending = dev->cfg; 12097 12098 dev->num_napi_configs = maxqs; 12099 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 12100 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 12101 if (!dev->napi_config) 12102 goto free_all; 12103 12104 strscpy(dev->name, name); 12105 dev->name_assign_type = name_assign_type; 12106 dev->group = INIT_NETDEV_GROUP; 12107 if (!dev->ethtool_ops) 12108 dev->ethtool_ops = &default_ethtool_ops; 12109 12110 nf_hook_netdev_init(dev); 12111 12112 return dev; 12113 12114 free_all: 12115 free_netdev(dev); 12116 return NULL; 12117 12118 free_pcpu: 12119 #ifdef CONFIG_PCPU_DEV_REFCNT 12120 free_percpu(dev->pcpu_refcnt); 12121 free_dev: 12122 #endif 12123 kvfree(dev); 12124 return NULL; 12125 } 12126 EXPORT_SYMBOL(alloc_netdev_mqs); 12127 12128 static void netdev_napi_exit(struct net_device *dev) 12129 { 12130 if (!list_empty(&dev->napi_list)) { 12131 struct napi_struct *p, *n; 12132 12133 netdev_lock(dev); 12134 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 12135 __netif_napi_del_locked(p); 12136 netdev_unlock(dev); 12137 12138 synchronize_net(); 12139 } 12140 12141 kvfree(dev->napi_config); 12142 } 12143 12144 /** 12145 * free_netdev - free network device 12146 * @dev: device 12147 * 12148 * This function does the last stage of destroying an allocated device 12149 * interface. The reference to the device object is released. If this 12150 * is the last reference then it will be freed.Must be called in process 12151 * context. 12152 */ 12153 void free_netdev(struct net_device *dev) 12154 { 12155 might_sleep(); 12156 12157 /* When called immediately after register_netdevice() failed the unwind 12158 * handling may still be dismantling the device. Handle that case by 12159 * deferring the free. 12160 */ 12161 if (dev->reg_state == NETREG_UNREGISTERING) { 12162 ASSERT_RTNL(); 12163 dev->needs_free_netdev = true; 12164 return; 12165 } 12166 12167 WARN_ON(dev->cfg != dev->cfg_pending); 12168 kfree(dev->cfg); 12169 kfree(dev->ethtool); 12170 netif_free_tx_queues(dev); 12171 netif_free_rx_queues(dev); 12172 12173 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 12174 12175 /* Flush device addresses */ 12176 dev_addr_flush(dev); 12177 12178 netdev_napi_exit(dev); 12179 12180 netif_del_cpu_rmap(dev); 12181 12182 ref_tracker_dir_exit(&dev->refcnt_tracker); 12183 #ifdef CONFIG_PCPU_DEV_REFCNT 12184 free_percpu(dev->pcpu_refcnt); 12185 dev->pcpu_refcnt = NULL; 12186 #endif 12187 free_percpu(dev->core_stats); 12188 dev->core_stats = NULL; 12189 free_percpu(dev->xdp_bulkq); 12190 dev->xdp_bulkq = NULL; 12191 12192 netdev_free_phy_link_topology(dev); 12193 12194 mutex_destroy(&dev->lock); 12195 12196 /* Compatibility with error handling in drivers */ 12197 if (dev->reg_state == NETREG_UNINITIALIZED || 12198 dev->reg_state == NETREG_DUMMY) { 12199 kvfree(dev); 12200 return; 12201 } 12202 12203 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 12204 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 12205 12206 /* will free via device release */ 12207 put_device(&dev->dev); 12208 } 12209 EXPORT_SYMBOL(free_netdev); 12210 12211 /** 12212 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 12213 * @sizeof_priv: size of private data to allocate space for 12214 * 12215 * Return: the allocated net_device on success, NULL otherwise 12216 */ 12217 struct net_device *alloc_netdev_dummy(int sizeof_priv) 12218 { 12219 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 12220 init_dummy_netdev); 12221 } 12222 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 12223 12224 /** 12225 * synchronize_net - Synchronize with packet receive processing 12226 * 12227 * Wait for packets currently being received to be done. 12228 * Does not block later packets from starting. 12229 */ 12230 void synchronize_net(void) 12231 { 12232 might_sleep(); 12233 if (from_cleanup_net() || rtnl_is_locked()) 12234 synchronize_rcu_expedited(); 12235 else 12236 synchronize_rcu(); 12237 } 12238 EXPORT_SYMBOL(synchronize_net); 12239 12240 static void netdev_rss_contexts_free(struct net_device *dev) 12241 { 12242 struct ethtool_rxfh_context *ctx; 12243 unsigned long context; 12244 12245 mutex_lock(&dev->ethtool->rss_lock); 12246 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 12247 xa_erase(&dev->ethtool->rss_ctx, context); 12248 dev->ethtool_ops->remove_rxfh_context(dev, ctx, context, NULL); 12249 kfree(ctx); 12250 } 12251 xa_destroy(&dev->ethtool->rss_ctx); 12252 mutex_unlock(&dev->ethtool->rss_lock); 12253 } 12254 12255 /** 12256 * unregister_netdevice_queue - remove device from the kernel 12257 * @dev: device 12258 * @head: list 12259 * 12260 * This function shuts down a device interface and removes it 12261 * from the kernel tables. 12262 * If head not NULL, device is queued to be unregistered later. 12263 * 12264 * Callers must hold the rtnl semaphore. You may want 12265 * unregister_netdev() instead of this. 12266 */ 12267 12268 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 12269 { 12270 ASSERT_RTNL(); 12271 12272 if (head) { 12273 list_move_tail(&dev->unreg_list, head); 12274 } else { 12275 LIST_HEAD(single); 12276 12277 list_add(&dev->unreg_list, &single); 12278 unregister_netdevice_many(&single); 12279 } 12280 } 12281 EXPORT_SYMBOL(unregister_netdevice_queue); 12282 12283 static void dev_memory_provider_uninstall(struct net_device *dev) 12284 { 12285 unsigned int i; 12286 12287 for (i = 0; i < dev->real_num_rx_queues; i++) { 12288 struct netdev_rx_queue *rxq = &dev->_rx[i]; 12289 struct pp_memory_provider_params *p = &rxq->mp_params; 12290 12291 if (p->mp_ops && p->mp_ops->uninstall) 12292 p->mp_ops->uninstall(rxq->mp_params.mp_priv, rxq); 12293 } 12294 } 12295 12296 /* devices must be UP and netdev_lock()'d */ 12297 static void netif_close_many_and_unlock(struct list_head *close_head) 12298 { 12299 struct net_device *dev, *tmp; 12300 12301 netif_close_many(close_head, false); 12302 12303 /* ... now unlock them */ 12304 list_for_each_entry_safe(dev, tmp, close_head, close_list) { 12305 netdev_unlock(dev); 12306 list_del_init(&dev->close_list); 12307 } 12308 } 12309 12310 static void netif_close_many_and_unlock_cond(struct list_head *close_head) 12311 { 12312 #ifdef CONFIG_LOCKDEP 12313 /* We can only track up to MAX_LOCK_DEPTH locks per task. 12314 * 12315 * Reserve half the available slots for additional locks possibly 12316 * taken by notifiers and (soft)irqs. 12317 */ 12318 unsigned int limit = MAX_LOCK_DEPTH / 2; 12319 12320 if (lockdep_depth(current) > limit) 12321 netif_close_many_and_unlock(close_head); 12322 #endif 12323 } 12324 12325 void unregister_netdevice_many_notify(struct list_head *head, 12326 u32 portid, const struct nlmsghdr *nlh) 12327 { 12328 struct net_device *dev, *tmp; 12329 LIST_HEAD(close_head); 12330 int cnt = 0; 12331 12332 BUG_ON(dev_boot_phase); 12333 ASSERT_RTNL(); 12334 12335 if (list_empty(head)) 12336 return; 12337 12338 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 12339 /* Some devices call without registering 12340 * for initialization unwind. Remove those 12341 * devices and proceed with the remaining. 12342 */ 12343 if (dev->reg_state == NETREG_UNINITIALIZED) { 12344 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 12345 dev->name, dev); 12346 12347 WARN_ON(1); 12348 list_del(&dev->unreg_list); 12349 continue; 12350 } 12351 dev->dismantle = true; 12352 BUG_ON(dev->reg_state != NETREG_REGISTERED); 12353 } 12354 12355 /* If device is running, close it first. Start with ops locked... */ 12356 list_for_each_entry(dev, head, unreg_list) { 12357 if (!(dev->flags & IFF_UP)) 12358 continue; 12359 if (netdev_need_ops_lock(dev)) { 12360 list_add_tail(&dev->close_list, &close_head); 12361 netdev_lock(dev); 12362 } 12363 netif_close_many_and_unlock_cond(&close_head); 12364 } 12365 netif_close_many_and_unlock(&close_head); 12366 /* ... now go over the rest. */ 12367 list_for_each_entry(dev, head, unreg_list) { 12368 if (!netdev_need_ops_lock(dev)) 12369 list_add_tail(&dev->close_list, &close_head); 12370 } 12371 netif_close_many(&close_head, true); 12372 12373 list_for_each_entry(dev, head, unreg_list) { 12374 /* And unlink it from device chain. */ 12375 unlist_netdevice(dev); 12376 netdev_lock(dev); 12377 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 12378 netdev_unlock(dev); 12379 } 12380 flush_all_backlogs(); 12381 12382 synchronize_net(); 12383 12384 list_for_each_entry(dev, head, unreg_list) { 12385 struct sk_buff *skb = NULL; 12386 12387 /* Shutdown queueing discipline. */ 12388 netdev_lock_ops(dev); 12389 dev_shutdown(dev); 12390 dev_tcx_uninstall(dev); 12391 dev_xdp_uninstall(dev); 12392 dev_memory_provider_uninstall(dev); 12393 netdev_unlock_ops(dev); 12394 bpf_dev_bound_netdev_unregister(dev); 12395 12396 netdev_offload_xstats_disable_all(dev); 12397 12398 /* Notify protocols, that we are about to destroy 12399 * this device. They should clean all the things. 12400 */ 12401 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12402 12403 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 12404 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 12405 GFP_KERNEL, NULL, 0, 12406 portid, nlh); 12407 12408 /* 12409 * Flush the unicast and multicast chains 12410 */ 12411 dev_uc_flush(dev); 12412 dev_mc_flush(dev); 12413 12414 netdev_name_node_alt_flush(dev); 12415 netdev_name_node_free(dev->name_node); 12416 12417 netdev_rss_contexts_free(dev); 12418 12419 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 12420 12421 if (dev->netdev_ops->ndo_uninit) 12422 dev->netdev_ops->ndo_uninit(dev); 12423 12424 mutex_destroy(&dev->ethtool->rss_lock); 12425 12426 net_shaper_flush_netdev(dev); 12427 12428 if (skb) 12429 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 12430 12431 /* Notifier chain MUST detach us all upper devices. */ 12432 WARN_ON(netdev_has_any_upper_dev(dev)); 12433 WARN_ON(netdev_has_any_lower_dev(dev)); 12434 12435 /* Remove entries from kobject tree */ 12436 netdev_unregister_kobject(dev); 12437 #ifdef CONFIG_XPS 12438 /* Remove XPS queueing entries */ 12439 netif_reset_xps_queues_gt(dev, 0); 12440 #endif 12441 } 12442 12443 synchronize_net(); 12444 12445 list_for_each_entry(dev, head, unreg_list) { 12446 netdev_put(dev, &dev->dev_registered_tracker); 12447 net_set_todo(dev); 12448 cnt++; 12449 } 12450 atomic_add(cnt, &dev_unreg_count); 12451 12452 list_del(head); 12453 } 12454 12455 /** 12456 * unregister_netdevice_many - unregister many devices 12457 * @head: list of devices 12458 * 12459 * Note: As most callers use a stack allocated list_head, 12460 * we force a list_del() to make sure stack won't be corrupted later. 12461 */ 12462 void unregister_netdevice_many(struct list_head *head) 12463 { 12464 unregister_netdevice_many_notify(head, 0, NULL); 12465 } 12466 EXPORT_SYMBOL(unregister_netdevice_many); 12467 12468 /** 12469 * unregister_netdev - remove device from the kernel 12470 * @dev: device 12471 * 12472 * This function shuts down a device interface and removes it 12473 * from the kernel tables. 12474 * 12475 * This is just a wrapper for unregister_netdevice that takes 12476 * the rtnl semaphore. In general you want to use this and not 12477 * unregister_netdevice. 12478 */ 12479 void unregister_netdev(struct net_device *dev) 12480 { 12481 rtnl_net_dev_lock(dev); 12482 unregister_netdevice(dev); 12483 rtnl_net_dev_unlock(dev); 12484 } 12485 EXPORT_SYMBOL(unregister_netdev); 12486 12487 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 12488 const char *pat, int new_ifindex, 12489 struct netlink_ext_ack *extack) 12490 { 12491 struct netdev_name_node *name_node; 12492 struct net *net_old = dev_net(dev); 12493 char new_name[IFNAMSIZ] = {}; 12494 int err, new_nsid; 12495 12496 ASSERT_RTNL(); 12497 12498 /* Don't allow namespace local devices to be moved. */ 12499 err = -EINVAL; 12500 if (dev->netns_immutable) { 12501 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12502 goto out; 12503 } 12504 12505 /* Ensure the device has been registered */ 12506 if (dev->reg_state != NETREG_REGISTERED) { 12507 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12508 goto out; 12509 } 12510 12511 /* Get out if there is nothing todo */ 12512 err = 0; 12513 if (net_eq(net_old, net)) 12514 goto out; 12515 12516 /* Pick the destination device name, and ensure 12517 * we can use it in the destination network namespace. 12518 */ 12519 err = -EEXIST; 12520 if (netdev_name_in_use(net, dev->name)) { 12521 /* We get here if we can't use the current device name */ 12522 if (!pat) { 12523 NL_SET_ERR_MSG(extack, 12524 "An interface with the same name exists in the target netns"); 12525 goto out; 12526 } 12527 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12528 if (err < 0) { 12529 NL_SET_ERR_MSG_FMT(extack, 12530 "Unable to use '%s' for the new interface name in the target netns", 12531 pat); 12532 goto out; 12533 } 12534 } 12535 /* Check that none of the altnames conflicts. */ 12536 err = -EEXIST; 12537 netdev_for_each_altname(dev, name_node) { 12538 if (netdev_name_in_use(net, name_node->name)) { 12539 NL_SET_ERR_MSG_FMT(extack, 12540 "An interface with the altname %s exists in the target netns", 12541 name_node->name); 12542 goto out; 12543 } 12544 } 12545 12546 /* Check that new_ifindex isn't used yet. */ 12547 if (new_ifindex) { 12548 err = dev_index_reserve(net, new_ifindex); 12549 if (err < 0) { 12550 NL_SET_ERR_MSG_FMT(extack, 12551 "The ifindex %d is not available in the target netns", 12552 new_ifindex); 12553 goto out; 12554 } 12555 } else { 12556 /* If there is an ifindex conflict assign a new one */ 12557 err = dev_index_reserve(net, dev->ifindex); 12558 if (err == -EBUSY) 12559 err = dev_index_reserve(net, 0); 12560 if (err < 0) { 12561 NL_SET_ERR_MSG(extack, 12562 "Unable to allocate a new ifindex in the target netns"); 12563 goto out; 12564 } 12565 new_ifindex = err; 12566 } 12567 12568 /* 12569 * And now a mini version of register_netdevice unregister_netdevice. 12570 */ 12571 12572 netdev_lock_ops(dev); 12573 /* If device is running close it first. */ 12574 netif_close(dev); 12575 /* And unlink it from device chain */ 12576 unlist_netdevice(dev); 12577 12578 if (!netdev_need_ops_lock(dev)) 12579 netdev_lock(dev); 12580 dev->moving_ns = true; 12581 netdev_unlock(dev); 12582 12583 synchronize_net(); 12584 12585 /* Shutdown queueing discipline. */ 12586 netdev_lock_ops(dev); 12587 dev_shutdown(dev); 12588 netdev_unlock_ops(dev); 12589 12590 /* Notify protocols, that we are about to destroy 12591 * this device. They should clean all the things. 12592 * 12593 * Note that dev->reg_state stays at NETREG_REGISTERED. 12594 * This is wanted because this way 8021q and macvlan know 12595 * the device is just moving and can keep their slaves up. 12596 */ 12597 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12598 rcu_barrier(); 12599 12600 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12601 12602 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12603 new_ifindex); 12604 12605 /* 12606 * Flush the unicast and multicast chains 12607 */ 12608 dev_uc_flush(dev); 12609 dev_mc_flush(dev); 12610 12611 /* Send a netdev-removed uevent to the old namespace */ 12612 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12613 netdev_adjacent_del_links(dev); 12614 12615 /* Move per-net netdevice notifiers that are following the netdevice */ 12616 move_netdevice_notifiers_dev_net(dev, net); 12617 12618 /* Actually switch the network namespace */ 12619 netdev_lock(dev); 12620 dev_net_set(dev, net); 12621 netdev_unlock(dev); 12622 dev->ifindex = new_ifindex; 12623 12624 if (new_name[0]) { 12625 /* Rename the netdev to prepared name */ 12626 write_seqlock_bh(&netdev_rename_lock); 12627 strscpy(dev->name, new_name, IFNAMSIZ); 12628 write_sequnlock_bh(&netdev_rename_lock); 12629 } 12630 12631 /* Fixup kobjects */ 12632 dev_set_uevent_suppress(&dev->dev, 1); 12633 err = device_rename(&dev->dev, dev->name); 12634 dev_set_uevent_suppress(&dev->dev, 0); 12635 WARN_ON(err); 12636 12637 /* Send a netdev-add uevent to the new namespace */ 12638 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12639 netdev_adjacent_add_links(dev); 12640 12641 /* Adapt owner in case owning user namespace of target network 12642 * namespace is different from the original one. 12643 */ 12644 err = netdev_change_owner(dev, net_old, net); 12645 WARN_ON(err); 12646 12647 netdev_lock(dev); 12648 dev->moving_ns = false; 12649 if (!netdev_need_ops_lock(dev)) 12650 netdev_unlock(dev); 12651 12652 /* Add the device back in the hashes */ 12653 list_netdevice(dev); 12654 /* Notify protocols, that a new device appeared. */ 12655 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12656 netdev_unlock_ops(dev); 12657 12658 /* 12659 * Prevent userspace races by waiting until the network 12660 * device is fully setup before sending notifications. 12661 */ 12662 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12663 12664 synchronize_net(); 12665 err = 0; 12666 out: 12667 return err; 12668 } 12669 12670 static int dev_cpu_dead(unsigned int oldcpu) 12671 { 12672 struct sk_buff **list_skb; 12673 struct sk_buff *skb; 12674 unsigned int cpu; 12675 struct softnet_data *sd, *oldsd, *remsd = NULL; 12676 12677 local_irq_disable(); 12678 cpu = smp_processor_id(); 12679 sd = &per_cpu(softnet_data, cpu); 12680 oldsd = &per_cpu(softnet_data, oldcpu); 12681 12682 /* Find end of our completion_queue. */ 12683 list_skb = &sd->completion_queue; 12684 while (*list_skb) 12685 list_skb = &(*list_skb)->next; 12686 /* Append completion queue from offline CPU. */ 12687 *list_skb = oldsd->completion_queue; 12688 oldsd->completion_queue = NULL; 12689 12690 /* Append output queue from offline CPU. */ 12691 if (oldsd->output_queue) { 12692 *sd->output_queue_tailp = oldsd->output_queue; 12693 sd->output_queue_tailp = oldsd->output_queue_tailp; 12694 oldsd->output_queue = NULL; 12695 oldsd->output_queue_tailp = &oldsd->output_queue; 12696 } 12697 /* Append NAPI poll list from offline CPU, with one exception : 12698 * process_backlog() must be called by cpu owning percpu backlog. 12699 * We properly handle process_queue & input_pkt_queue later. 12700 */ 12701 while (!list_empty(&oldsd->poll_list)) { 12702 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12703 struct napi_struct, 12704 poll_list); 12705 12706 list_del_init(&napi->poll_list); 12707 if (napi->poll == process_backlog) 12708 napi->state &= NAPIF_STATE_THREADED; 12709 else 12710 ____napi_schedule(sd, napi); 12711 } 12712 12713 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12714 local_irq_enable(); 12715 12716 if (!use_backlog_threads()) { 12717 #ifdef CONFIG_RPS 12718 remsd = oldsd->rps_ipi_list; 12719 oldsd->rps_ipi_list = NULL; 12720 #endif 12721 /* send out pending IPI's on offline CPU */ 12722 net_rps_send_ipi(remsd); 12723 } 12724 12725 /* Process offline CPU's input_pkt_queue */ 12726 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12727 netif_rx(skb); 12728 rps_input_queue_head_incr(oldsd); 12729 } 12730 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12731 netif_rx(skb); 12732 rps_input_queue_head_incr(oldsd); 12733 } 12734 12735 return 0; 12736 } 12737 12738 /** 12739 * netdev_increment_features - increment feature set by one 12740 * @all: current feature set 12741 * @one: new feature set 12742 * @mask: mask feature set 12743 * 12744 * Computes a new feature set after adding a device with feature set 12745 * @one to the master device with current feature set @all. Will not 12746 * enable anything that is off in @mask. Returns the new feature set. 12747 */ 12748 netdev_features_t netdev_increment_features(netdev_features_t all, 12749 netdev_features_t one, netdev_features_t mask) 12750 { 12751 if (mask & NETIF_F_HW_CSUM) 12752 mask |= NETIF_F_CSUM_MASK; 12753 mask |= NETIF_F_VLAN_CHALLENGED; 12754 12755 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12756 all &= one | ~NETIF_F_ALL_FOR_ALL; 12757 12758 /* If one device supports hw checksumming, set for all. */ 12759 if (all & NETIF_F_HW_CSUM) 12760 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12761 12762 return all; 12763 } 12764 EXPORT_SYMBOL(netdev_increment_features); 12765 12766 /** 12767 * netdev_compute_master_upper_features - compute feature from lowers 12768 * @dev: the upper device 12769 * @update_header: whether to update upper device's header_len/headroom/tailroom 12770 * 12771 * Recompute the upper device's feature based on all lower devices. 12772 */ 12773 void netdev_compute_master_upper_features(struct net_device *dev, bool update_header) 12774 { 12775 unsigned int dst_release_flag = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12776 netdev_features_t gso_partial_features = MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES; 12777 netdev_features_t xfrm_features = MASTER_UPPER_DEV_XFRM_FEATURES; 12778 netdev_features_t mpls_features = MASTER_UPPER_DEV_MPLS_FEATURES; 12779 netdev_features_t vlan_features = MASTER_UPPER_DEV_VLAN_FEATURES; 12780 netdev_features_t enc_features = MASTER_UPPER_DEV_ENC_FEATURES; 12781 unsigned short max_header_len = ETH_HLEN; 12782 unsigned int tso_max_size = TSO_MAX_SIZE; 12783 unsigned short max_headroom = 0; 12784 unsigned short max_tailroom = 0; 12785 u16 tso_max_segs = TSO_MAX_SEGS; 12786 struct net_device *lower_dev; 12787 struct list_head *iter; 12788 12789 mpls_features = netdev_base_features(mpls_features); 12790 vlan_features = netdev_base_features(vlan_features); 12791 enc_features = netdev_base_features(enc_features); 12792 12793 netdev_for_each_lower_dev(dev, lower_dev, iter) { 12794 gso_partial_features = netdev_increment_features(gso_partial_features, 12795 lower_dev->gso_partial_features, 12796 MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES); 12797 12798 vlan_features = netdev_increment_features(vlan_features, 12799 lower_dev->vlan_features, 12800 MASTER_UPPER_DEV_VLAN_FEATURES); 12801 12802 enc_features = netdev_increment_features(enc_features, 12803 lower_dev->hw_enc_features, 12804 MASTER_UPPER_DEV_ENC_FEATURES); 12805 12806 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 12807 xfrm_features = netdev_increment_features(xfrm_features, 12808 lower_dev->hw_enc_features, 12809 MASTER_UPPER_DEV_XFRM_FEATURES); 12810 12811 mpls_features = netdev_increment_features(mpls_features, 12812 lower_dev->mpls_features, 12813 MASTER_UPPER_DEV_MPLS_FEATURES); 12814 12815 dst_release_flag &= lower_dev->priv_flags; 12816 12817 if (update_header) { 12818 max_header_len = max(max_header_len, lower_dev->hard_header_len); 12819 max_headroom = max(max_headroom, lower_dev->needed_headroom); 12820 max_tailroom = max(max_tailroom, lower_dev->needed_tailroom); 12821 } 12822 12823 tso_max_size = min(tso_max_size, lower_dev->tso_max_size); 12824 tso_max_segs = min(tso_max_segs, lower_dev->tso_max_segs); 12825 } 12826 12827 dev->gso_partial_features = gso_partial_features; 12828 dev->vlan_features = vlan_features; 12829 dev->hw_enc_features = enc_features | NETIF_F_GSO_ENCAP_ALL | 12830 NETIF_F_HW_VLAN_CTAG_TX | 12831 NETIF_F_HW_VLAN_STAG_TX; 12832 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 12833 dev->hw_enc_features |= xfrm_features; 12834 dev->mpls_features = mpls_features; 12835 12836 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE; 12837 if ((dev->priv_flags & IFF_XMIT_DST_RELEASE_PERM) && 12838 dst_release_flag == (IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM)) 12839 dev->priv_flags |= IFF_XMIT_DST_RELEASE; 12840 12841 if (update_header) { 12842 dev->hard_header_len = max_header_len; 12843 dev->needed_headroom = max_headroom; 12844 dev->needed_tailroom = max_tailroom; 12845 } 12846 12847 netif_set_tso_max_segs(dev, tso_max_segs); 12848 netif_set_tso_max_size(dev, tso_max_size); 12849 12850 netdev_change_features(dev); 12851 } 12852 EXPORT_SYMBOL(netdev_compute_master_upper_features); 12853 12854 static struct hlist_head * __net_init netdev_create_hash(void) 12855 { 12856 int i; 12857 struct hlist_head *hash; 12858 12859 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 12860 if (hash != NULL) 12861 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12862 INIT_HLIST_HEAD(&hash[i]); 12863 12864 return hash; 12865 } 12866 12867 /* Initialize per network namespace state */ 12868 static int __net_init netdev_init(struct net *net) 12869 { 12870 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12871 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 12872 12873 INIT_LIST_HEAD(&net->dev_base_head); 12874 12875 net->dev_name_head = netdev_create_hash(); 12876 if (net->dev_name_head == NULL) 12877 goto err_name; 12878 12879 net->dev_index_head = netdev_create_hash(); 12880 if (net->dev_index_head == NULL) 12881 goto err_idx; 12882 12883 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12884 12885 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12886 12887 return 0; 12888 12889 err_idx: 12890 kfree(net->dev_name_head); 12891 err_name: 12892 return -ENOMEM; 12893 } 12894 12895 /** 12896 * netdev_drivername - network driver for the device 12897 * @dev: network device 12898 * 12899 * Determine network driver for device. 12900 */ 12901 const char *netdev_drivername(const struct net_device *dev) 12902 { 12903 const struct device_driver *driver; 12904 const struct device *parent; 12905 const char *empty = ""; 12906 12907 parent = dev->dev.parent; 12908 if (!parent) 12909 return empty; 12910 12911 driver = parent->driver; 12912 if (driver && driver->name) 12913 return driver->name; 12914 return empty; 12915 } 12916 12917 static void __netdev_printk(const char *level, const struct net_device *dev, 12918 struct va_format *vaf) 12919 { 12920 if (dev && dev->dev.parent) { 12921 dev_printk_emit(level[1] - '0', 12922 dev->dev.parent, 12923 "%s %s %s%s: %pV", 12924 dev_driver_string(dev->dev.parent), 12925 dev_name(dev->dev.parent), 12926 netdev_name(dev), netdev_reg_state(dev), 12927 vaf); 12928 } else if (dev) { 12929 printk("%s%s%s: %pV", 12930 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12931 } else { 12932 printk("%s(NULL net_device): %pV", level, vaf); 12933 } 12934 } 12935 12936 void netdev_printk(const char *level, const struct net_device *dev, 12937 const char *format, ...) 12938 { 12939 struct va_format vaf; 12940 va_list args; 12941 12942 va_start(args, format); 12943 12944 vaf.fmt = format; 12945 vaf.va = &args; 12946 12947 __netdev_printk(level, dev, &vaf); 12948 12949 va_end(args); 12950 } 12951 EXPORT_SYMBOL(netdev_printk); 12952 12953 #define define_netdev_printk_level(func, level) \ 12954 void func(const struct net_device *dev, const char *fmt, ...) \ 12955 { \ 12956 struct va_format vaf; \ 12957 va_list args; \ 12958 \ 12959 va_start(args, fmt); \ 12960 \ 12961 vaf.fmt = fmt; \ 12962 vaf.va = &args; \ 12963 \ 12964 __netdev_printk(level, dev, &vaf); \ 12965 \ 12966 va_end(args); \ 12967 } \ 12968 EXPORT_SYMBOL(func); 12969 12970 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12971 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12972 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12973 define_netdev_printk_level(netdev_err, KERN_ERR); 12974 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12975 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12976 define_netdev_printk_level(netdev_info, KERN_INFO); 12977 12978 static void __net_exit netdev_exit(struct net *net) 12979 { 12980 kfree(net->dev_name_head); 12981 kfree(net->dev_index_head); 12982 xa_destroy(&net->dev_by_index); 12983 if (net != &init_net) 12984 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 12985 } 12986 12987 static struct pernet_operations __net_initdata netdev_net_ops = { 12988 .init = netdev_init, 12989 .exit = netdev_exit, 12990 }; 12991 12992 static void __net_exit default_device_exit_net(struct net *net) 12993 { 12994 struct netdev_name_node *name_node, *tmp; 12995 struct net_device *dev, *aux; 12996 /* 12997 * Push all migratable network devices back to the 12998 * initial network namespace 12999 */ 13000 ASSERT_RTNL(); 13001 for_each_netdev_safe(net, dev, aux) { 13002 int err; 13003 char fb_name[IFNAMSIZ]; 13004 13005 /* Ignore unmoveable devices (i.e. loopback) */ 13006 if (dev->netns_immutable) 13007 continue; 13008 13009 /* Leave virtual devices for the generic cleanup */ 13010 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 13011 continue; 13012 13013 /* Push remaining network devices to init_net */ 13014 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 13015 if (netdev_name_in_use(&init_net, fb_name)) 13016 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 13017 13018 netdev_for_each_altname_safe(dev, name_node, tmp) 13019 if (netdev_name_in_use(&init_net, name_node->name)) 13020 __netdev_name_node_alt_destroy(name_node); 13021 13022 err = dev_change_net_namespace(dev, &init_net, fb_name); 13023 if (err) { 13024 pr_emerg("%s: failed to move %s to init_net: %d\n", 13025 __func__, dev->name, err); 13026 BUG(); 13027 } 13028 } 13029 } 13030 13031 static void __net_exit default_device_exit_batch(struct list_head *net_list) 13032 { 13033 /* At exit all network devices most be removed from a network 13034 * namespace. Do this in the reverse order of registration. 13035 * Do this across as many network namespaces as possible to 13036 * improve batching efficiency. 13037 */ 13038 struct net_device *dev; 13039 struct net *net; 13040 LIST_HEAD(dev_kill_list); 13041 13042 rtnl_lock(); 13043 list_for_each_entry(net, net_list, exit_list) { 13044 default_device_exit_net(net); 13045 cond_resched(); 13046 } 13047 13048 list_for_each_entry(net, net_list, exit_list) { 13049 for_each_netdev_reverse(net, dev) { 13050 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 13051 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 13052 else 13053 unregister_netdevice_queue(dev, &dev_kill_list); 13054 } 13055 } 13056 unregister_netdevice_many(&dev_kill_list); 13057 rtnl_unlock(); 13058 } 13059 13060 static struct pernet_operations __net_initdata default_device_ops = { 13061 .exit_batch = default_device_exit_batch, 13062 }; 13063 13064 static void __init net_dev_struct_check(void) 13065 { 13066 /* TX read-mostly hotpath */ 13067 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 13068 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 13069 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 13070 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 13071 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 13072 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 13073 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 13074 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 13075 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 13076 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 13077 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 13078 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 13079 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 13080 #ifdef CONFIG_XPS 13081 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 13082 #endif 13083 #ifdef CONFIG_NETFILTER_EGRESS 13084 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 13085 #endif 13086 #ifdef CONFIG_NET_XGRESS 13087 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 13088 #endif 13089 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 13090 13091 /* TXRX read-mostly hotpath */ 13092 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 13093 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 13094 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 13095 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 13096 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 13097 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 13098 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 13099 13100 /* RX read-mostly hotpath */ 13101 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 13102 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 13103 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 13104 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 13105 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 13106 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 13107 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 13108 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 13109 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 13110 #ifdef CONFIG_NETPOLL 13111 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 13112 #endif 13113 #ifdef CONFIG_NET_XGRESS 13114 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 13115 #endif 13116 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 13117 } 13118 13119 /* 13120 * Initialize the DEV module. At boot time this walks the device list and 13121 * unhooks any devices that fail to initialise (normally hardware not 13122 * present) and leaves us with a valid list of present and active devices. 13123 * 13124 */ 13125 13126 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 13127 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 13128 13129 static int net_page_pool_create(int cpuid) 13130 { 13131 #if IS_ENABLED(CONFIG_PAGE_POOL) 13132 struct page_pool_params page_pool_params = { 13133 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 13134 .flags = PP_FLAG_SYSTEM_POOL, 13135 .nid = cpu_to_mem(cpuid), 13136 }; 13137 struct page_pool *pp_ptr; 13138 int err; 13139 13140 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 13141 if (IS_ERR(pp_ptr)) 13142 return -ENOMEM; 13143 13144 err = xdp_reg_page_pool(pp_ptr); 13145 if (err) { 13146 page_pool_destroy(pp_ptr); 13147 return err; 13148 } 13149 13150 per_cpu(system_page_pool.pool, cpuid) = pp_ptr; 13151 #endif 13152 return 0; 13153 } 13154 13155 static int backlog_napi_should_run(unsigned int cpu) 13156 { 13157 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13158 struct napi_struct *napi = &sd->backlog; 13159 13160 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 13161 } 13162 13163 static void run_backlog_napi(unsigned int cpu) 13164 { 13165 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13166 13167 napi_threaded_poll_loop(&sd->backlog, false); 13168 } 13169 13170 static void backlog_napi_setup(unsigned int cpu) 13171 { 13172 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13173 struct napi_struct *napi = &sd->backlog; 13174 13175 napi->thread = this_cpu_read(backlog_napi); 13176 set_bit(NAPI_STATE_THREADED, &napi->state); 13177 } 13178 13179 static struct smp_hotplug_thread backlog_threads = { 13180 .store = &backlog_napi, 13181 .thread_should_run = backlog_napi_should_run, 13182 .thread_fn = run_backlog_napi, 13183 .thread_comm = "backlog_napi/%u", 13184 .setup = backlog_napi_setup, 13185 }; 13186 13187 /* 13188 * This is called single threaded during boot, so no need 13189 * to take the rtnl semaphore. 13190 */ 13191 static int __init net_dev_init(void) 13192 { 13193 int i, rc = -ENOMEM; 13194 13195 BUG_ON(!dev_boot_phase); 13196 13197 net_dev_struct_check(); 13198 13199 if (dev_proc_init()) 13200 goto out; 13201 13202 if (netdev_kobject_init()) 13203 goto out; 13204 13205 for (i = 0; i < PTYPE_HASH_SIZE; i++) 13206 INIT_LIST_HEAD(&ptype_base[i]); 13207 13208 if (register_pernet_subsys(&netdev_net_ops)) 13209 goto out; 13210 13211 /* 13212 * Initialise the packet receive queues. 13213 */ 13214 13215 flush_backlogs_fallback = flush_backlogs_alloc(); 13216 if (!flush_backlogs_fallback) 13217 goto out; 13218 13219 for_each_possible_cpu(i) { 13220 struct softnet_data *sd = &per_cpu(softnet_data, i); 13221 13222 skb_queue_head_init(&sd->input_pkt_queue); 13223 skb_queue_head_init(&sd->process_queue); 13224 #ifdef CONFIG_XFRM_OFFLOAD 13225 skb_queue_head_init(&sd->xfrm_backlog); 13226 #endif 13227 INIT_LIST_HEAD(&sd->poll_list); 13228 sd->output_queue_tailp = &sd->output_queue; 13229 #ifdef CONFIG_RPS 13230 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 13231 sd->cpu = i; 13232 #endif 13233 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 13234 13235 gro_init(&sd->backlog.gro); 13236 sd->backlog.poll = process_backlog; 13237 sd->backlog.weight = weight_p; 13238 INIT_LIST_HEAD(&sd->backlog.poll_list); 13239 13240 if (net_page_pool_create(i)) 13241 goto out; 13242 } 13243 net_hotdata.skb_defer_nodes = 13244 __alloc_percpu(sizeof(struct skb_defer_node) * nr_node_ids, 13245 __alignof__(struct skb_defer_node)); 13246 if (!net_hotdata.skb_defer_nodes) 13247 goto out; 13248 if (use_backlog_threads()) 13249 smpboot_register_percpu_thread(&backlog_threads); 13250 13251 dev_boot_phase = 0; 13252 13253 /* The loopback device is special if any other network devices 13254 * is present in a network namespace the loopback device must 13255 * be present. Since we now dynamically allocate and free the 13256 * loopback device ensure this invariant is maintained by 13257 * keeping the loopback device as the first device on the 13258 * list of network devices. Ensuring the loopback devices 13259 * is the first device that appears and the last network device 13260 * that disappears. 13261 */ 13262 if (register_pernet_device(&loopback_net_ops)) 13263 goto out; 13264 13265 if (register_pernet_device(&default_device_ops)) 13266 goto out; 13267 13268 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 13269 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 13270 13271 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 13272 NULL, dev_cpu_dead); 13273 WARN_ON(rc < 0); 13274 rc = 0; 13275 13276 /* avoid static key IPIs to isolated CPUs */ 13277 if (housekeeping_enabled(HK_TYPE_MISC)) 13278 net_enable_timestamp(); 13279 out: 13280 if (rc < 0) { 13281 for_each_possible_cpu(i) { 13282 struct page_pool *pp_ptr; 13283 13284 pp_ptr = per_cpu(system_page_pool.pool, i); 13285 if (!pp_ptr) 13286 continue; 13287 13288 xdp_unreg_page_pool(pp_ptr); 13289 page_pool_destroy(pp_ptr); 13290 per_cpu(system_page_pool.pool, i) = NULL; 13291 } 13292 } 13293 13294 return rc; 13295 } 13296 13297 subsys_initcall(net_dev_init); 13298