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/skbuff.h> 96 #include <linux/kthread.h> 97 #include <linux/bpf.h> 98 #include <linux/bpf_trace.h> 99 #include <net/net_namespace.h> 100 #include <net/sock.h> 101 #include <net/busy_poll.h> 102 #include <linux/rtnetlink.h> 103 #include <linux/stat.h> 104 #include <net/dsa.h> 105 #include <net/dst.h> 106 #include <net/dst_metadata.h> 107 #include <net/gro.h> 108 #include <net/pkt_sched.h> 109 #include <net/pkt_cls.h> 110 #include <net/checksum.h> 111 #include <net/xfrm.h> 112 #include <net/tcx.h> 113 #include <linux/highmem.h> 114 #include <linux/init.h> 115 #include <linux/module.h> 116 #include <linux/netpoll.h> 117 #include <linux/rcupdate.h> 118 #include <linux/delay.h> 119 #include <net/iw_handler.h> 120 #include <asm/current.h> 121 #include <linux/audit.h> 122 #include <linux/dmaengine.h> 123 #include <linux/err.h> 124 #include <linux/ctype.h> 125 #include <linux/if_arp.h> 126 #include <linux/if_vlan.h> 127 #include <linux/ip.h> 128 #include <net/ip.h> 129 #include <net/mpls.h> 130 #include <linux/ipv6.h> 131 #include <linux/in.h> 132 #include <linux/jhash.h> 133 #include <linux/random.h> 134 #include <trace/events/napi.h> 135 #include <trace/events/net.h> 136 #include <trace/events/skb.h> 137 #include <trace/events/qdisc.h> 138 #include <trace/events/xdp.h> 139 #include <linux/inetdevice.h> 140 #include <linux/cpu_rmap.h> 141 #include <linux/static_key.h> 142 #include <linux/hashtable.h> 143 #include <linux/vmalloc.h> 144 #include <linux/if_macvlan.h> 145 #include <linux/errqueue.h> 146 #include <linux/hrtimer.h> 147 #include <linux/netfilter_netdev.h> 148 #include <linux/crash_dump.h> 149 #include <linux/sctp.h> 150 #include <net/udp_tunnel.h> 151 #include <linux/net_namespace.h> 152 #include <linux/indirect_call_wrapper.h> 153 #include <net/devlink.h> 154 #include <linux/pm_runtime.h> 155 #include <linux/prandom.h> 156 #include <linux/once_lite.h> 157 #include <net/netdev_rx_queue.h> 158 #include <net/page_pool/types.h> 159 #include <net/page_pool/helpers.h> 160 #include <net/rps.h> 161 162 #include "dev.h" 163 #include "net-sysfs.h" 164 165 static DEFINE_SPINLOCK(ptype_lock); 166 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 167 168 static int netif_rx_internal(struct sk_buff *skb); 169 static int call_netdevice_notifiers_extack(unsigned long val, 170 struct net_device *dev, 171 struct netlink_ext_ack *extack); 172 173 static DEFINE_MUTEX(ifalias_mutex); 174 175 /* protects napi_hash addition/deletion and napi_gen_id */ 176 static DEFINE_SPINLOCK(napi_hash_lock); 177 178 static unsigned int napi_gen_id = NR_CPUS; 179 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 180 181 static DECLARE_RWSEM(devnet_rename_sem); 182 183 static inline void dev_base_seq_inc(struct net *net) 184 { 185 unsigned int val = net->dev_base_seq + 1; 186 187 WRITE_ONCE(net->dev_base_seq, val ?: 1); 188 } 189 190 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 191 { 192 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 193 194 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 195 } 196 197 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 198 { 199 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 200 } 201 202 #ifndef CONFIG_PREEMPT_RT 203 204 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 205 206 static int __init setup_backlog_napi_threads(char *arg) 207 { 208 static_branch_enable(&use_backlog_threads_key); 209 return 0; 210 } 211 early_param("thread_backlog_napi", setup_backlog_napi_threads); 212 213 static bool use_backlog_threads(void) 214 { 215 return static_branch_unlikely(&use_backlog_threads_key); 216 } 217 218 #else 219 220 static bool use_backlog_threads(void) 221 { 222 return true; 223 } 224 225 #endif 226 227 static inline void backlog_lock_irq_save(struct softnet_data *sd, 228 unsigned long *flags) 229 { 230 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 231 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 232 else if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 233 local_irq_save(*flags); 234 } 235 236 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 237 { 238 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 239 spin_lock_irq(&sd->input_pkt_queue.lock); 240 else if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 241 local_irq_disable(); 242 } 243 244 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 245 unsigned long *flags) 246 { 247 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 248 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags); 249 else if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 250 local_irq_restore(*flags); 251 } 252 253 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 254 { 255 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 256 spin_unlock_irq(&sd->input_pkt_queue.lock); 257 else if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 258 local_irq_enable(); 259 } 260 261 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 262 const char *name) 263 { 264 struct netdev_name_node *name_node; 265 266 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 267 if (!name_node) 268 return NULL; 269 INIT_HLIST_NODE(&name_node->hlist); 270 name_node->dev = dev; 271 name_node->name = name; 272 return name_node; 273 } 274 275 static struct netdev_name_node * 276 netdev_name_node_head_alloc(struct net_device *dev) 277 { 278 struct netdev_name_node *name_node; 279 280 name_node = netdev_name_node_alloc(dev, dev->name); 281 if (!name_node) 282 return NULL; 283 INIT_LIST_HEAD(&name_node->list); 284 return name_node; 285 } 286 287 static void netdev_name_node_free(struct netdev_name_node *name_node) 288 { 289 kfree(name_node); 290 } 291 292 static void netdev_name_node_add(struct net *net, 293 struct netdev_name_node *name_node) 294 { 295 hlist_add_head_rcu(&name_node->hlist, 296 dev_name_hash(net, name_node->name)); 297 } 298 299 static void netdev_name_node_del(struct netdev_name_node *name_node) 300 { 301 hlist_del_rcu(&name_node->hlist); 302 } 303 304 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 305 const char *name) 306 { 307 struct hlist_head *head = dev_name_hash(net, name); 308 struct netdev_name_node *name_node; 309 310 hlist_for_each_entry(name_node, head, hlist) 311 if (!strcmp(name_node->name, name)) 312 return name_node; 313 return NULL; 314 } 315 316 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 317 const char *name) 318 { 319 struct hlist_head *head = dev_name_hash(net, name); 320 struct netdev_name_node *name_node; 321 322 hlist_for_each_entry_rcu(name_node, head, hlist) 323 if (!strcmp(name_node->name, name)) 324 return name_node; 325 return NULL; 326 } 327 328 bool netdev_name_in_use(struct net *net, const char *name) 329 { 330 return netdev_name_node_lookup(net, name); 331 } 332 EXPORT_SYMBOL(netdev_name_in_use); 333 334 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 335 { 336 struct netdev_name_node *name_node; 337 struct net *net = dev_net(dev); 338 339 name_node = netdev_name_node_lookup(net, name); 340 if (name_node) 341 return -EEXIST; 342 name_node = netdev_name_node_alloc(dev, name); 343 if (!name_node) 344 return -ENOMEM; 345 netdev_name_node_add(net, name_node); 346 /* The node that holds dev->name acts as a head of per-device list. */ 347 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 348 349 return 0; 350 } 351 352 static void netdev_name_node_alt_free(struct rcu_head *head) 353 { 354 struct netdev_name_node *name_node = 355 container_of(head, struct netdev_name_node, rcu); 356 357 kfree(name_node->name); 358 netdev_name_node_free(name_node); 359 } 360 361 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 362 { 363 netdev_name_node_del(name_node); 364 list_del(&name_node->list); 365 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 366 } 367 368 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 369 { 370 struct netdev_name_node *name_node; 371 struct net *net = dev_net(dev); 372 373 name_node = netdev_name_node_lookup(net, name); 374 if (!name_node) 375 return -ENOENT; 376 /* lookup might have found our primary name or a name belonging 377 * to another device. 378 */ 379 if (name_node == dev->name_node || name_node->dev != dev) 380 return -EINVAL; 381 382 __netdev_name_node_alt_destroy(name_node); 383 return 0; 384 } 385 386 static void netdev_name_node_alt_flush(struct net_device *dev) 387 { 388 struct netdev_name_node *name_node, *tmp; 389 390 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 391 list_del(&name_node->list); 392 netdev_name_node_alt_free(&name_node->rcu); 393 } 394 } 395 396 /* Device list insertion */ 397 static void list_netdevice(struct net_device *dev) 398 { 399 struct netdev_name_node *name_node; 400 struct net *net = dev_net(dev); 401 402 ASSERT_RTNL(); 403 404 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 405 netdev_name_node_add(net, dev->name_node); 406 hlist_add_head_rcu(&dev->index_hlist, 407 dev_index_hash(net, dev->ifindex)); 408 409 netdev_for_each_altname(dev, name_node) 410 netdev_name_node_add(net, name_node); 411 412 /* We reserved the ifindex, this can't fail */ 413 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 414 415 dev_base_seq_inc(net); 416 } 417 418 /* Device list removal 419 * caller must respect a RCU grace period before freeing/reusing dev 420 */ 421 static void unlist_netdevice(struct net_device *dev) 422 { 423 struct netdev_name_node *name_node; 424 struct net *net = dev_net(dev); 425 426 ASSERT_RTNL(); 427 428 xa_erase(&net->dev_by_index, dev->ifindex); 429 430 netdev_for_each_altname(dev, name_node) 431 netdev_name_node_del(name_node); 432 433 /* Unlink dev from the device chain */ 434 list_del_rcu(&dev->dev_list); 435 netdev_name_node_del(dev->name_node); 436 hlist_del_rcu(&dev->index_hlist); 437 438 dev_base_seq_inc(dev_net(dev)); 439 } 440 441 /* 442 * Our notifier list 443 */ 444 445 static RAW_NOTIFIER_HEAD(netdev_chain); 446 447 /* 448 * Device drivers call our routines to queue packets here. We empty the 449 * queue in the local softnet handler. 450 */ 451 452 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data); 453 EXPORT_PER_CPU_SYMBOL(softnet_data); 454 455 /* Page_pool has a lockless array/stack to alloc/recycle pages. 456 * PP consumers must pay attention to run APIs in the appropriate context 457 * (e.g. NAPI context). 458 */ 459 static DEFINE_PER_CPU_ALIGNED(struct page_pool *, system_page_pool); 460 461 #ifdef CONFIG_LOCKDEP 462 /* 463 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 464 * according to dev->type 465 */ 466 static const unsigned short netdev_lock_type[] = { 467 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 468 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 469 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 470 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 471 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 472 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 473 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 474 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 475 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 476 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 477 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 478 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 479 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 480 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 481 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE}; 482 483 static const char *const netdev_lock_name[] = { 484 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 485 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 486 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 487 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 488 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 489 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 490 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 491 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 492 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 493 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 494 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 495 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 496 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 497 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE", 498 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"}; 499 500 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 501 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 502 503 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 504 { 505 int i; 506 507 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 508 if (netdev_lock_type[i] == dev_type) 509 return i; 510 /* the last key is used by default */ 511 return ARRAY_SIZE(netdev_lock_type) - 1; 512 } 513 514 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 515 unsigned short dev_type) 516 { 517 int i; 518 519 i = netdev_lock_pos(dev_type); 520 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 521 netdev_lock_name[i]); 522 } 523 524 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 525 { 526 int i; 527 528 i = netdev_lock_pos(dev->type); 529 lockdep_set_class_and_name(&dev->addr_list_lock, 530 &netdev_addr_lock_key[i], 531 netdev_lock_name[i]); 532 } 533 #else 534 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 535 unsigned short dev_type) 536 { 537 } 538 539 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 540 { 541 } 542 #endif 543 544 /******************************************************************************* 545 * 546 * Protocol management and registration routines 547 * 548 *******************************************************************************/ 549 550 551 /* 552 * Add a protocol ID to the list. Now that the input handler is 553 * smarter we can dispense with all the messy stuff that used to be 554 * here. 555 * 556 * BEWARE!!! Protocol handlers, mangling input packets, 557 * MUST BE last in hash buckets and checking protocol handlers 558 * MUST start from promiscuous ptype_all chain in net_bh. 559 * It is true now, do not change it. 560 * Explanation follows: if protocol handler, mangling packet, will 561 * be the first on list, it is not able to sense, that packet 562 * is cloned and should be copied-on-write, so that it will 563 * change it and subsequent readers will get broken packet. 564 * --ANK (980803) 565 */ 566 567 static inline struct list_head *ptype_head(const struct packet_type *pt) 568 { 569 if (pt->type == htons(ETH_P_ALL)) 570 return pt->dev ? &pt->dev->ptype_all : &net_hotdata.ptype_all; 571 else 572 return pt->dev ? &pt->dev->ptype_specific : 573 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 574 } 575 576 /** 577 * dev_add_pack - add packet handler 578 * @pt: packet type declaration 579 * 580 * Add a protocol handler to the networking stack. The passed &packet_type 581 * is linked into kernel lists and may not be freed until it has been 582 * removed from the kernel lists. 583 * 584 * This call does not sleep therefore it can not 585 * guarantee all CPU's that are in middle of receiving packets 586 * will see the new packet type (until the next received packet). 587 */ 588 589 void dev_add_pack(struct packet_type *pt) 590 { 591 struct list_head *head = ptype_head(pt); 592 593 spin_lock(&ptype_lock); 594 list_add_rcu(&pt->list, head); 595 spin_unlock(&ptype_lock); 596 } 597 EXPORT_SYMBOL(dev_add_pack); 598 599 /** 600 * __dev_remove_pack - remove packet handler 601 * @pt: packet type declaration 602 * 603 * Remove a protocol handler that was previously added to the kernel 604 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 605 * from the kernel lists and can be freed or reused once this function 606 * returns. 607 * 608 * The packet type might still be in use by receivers 609 * and must not be freed until after all the CPU's have gone 610 * through a quiescent state. 611 */ 612 void __dev_remove_pack(struct packet_type *pt) 613 { 614 struct list_head *head = ptype_head(pt); 615 struct packet_type *pt1; 616 617 spin_lock(&ptype_lock); 618 619 list_for_each_entry(pt1, head, list) { 620 if (pt == pt1) { 621 list_del_rcu(&pt->list); 622 goto out; 623 } 624 } 625 626 pr_warn("dev_remove_pack: %p not found\n", pt); 627 out: 628 spin_unlock(&ptype_lock); 629 } 630 EXPORT_SYMBOL(__dev_remove_pack); 631 632 /** 633 * dev_remove_pack - remove packet handler 634 * @pt: packet type declaration 635 * 636 * Remove a protocol handler that was previously added to the kernel 637 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 638 * from the kernel lists and can be freed or reused once this function 639 * returns. 640 * 641 * This call sleeps to guarantee that no CPU is looking at the packet 642 * type after return. 643 */ 644 void dev_remove_pack(struct packet_type *pt) 645 { 646 __dev_remove_pack(pt); 647 648 synchronize_net(); 649 } 650 EXPORT_SYMBOL(dev_remove_pack); 651 652 653 /******************************************************************************* 654 * 655 * Device Interface Subroutines 656 * 657 *******************************************************************************/ 658 659 /** 660 * dev_get_iflink - get 'iflink' value of a interface 661 * @dev: targeted interface 662 * 663 * Indicates the ifindex the interface is linked to. 664 * Physical interfaces have the same 'ifindex' and 'iflink' values. 665 */ 666 667 int dev_get_iflink(const struct net_device *dev) 668 { 669 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 670 return dev->netdev_ops->ndo_get_iflink(dev); 671 672 return READ_ONCE(dev->ifindex); 673 } 674 EXPORT_SYMBOL(dev_get_iflink); 675 676 /** 677 * dev_fill_metadata_dst - Retrieve tunnel egress information. 678 * @dev: targeted interface 679 * @skb: The packet. 680 * 681 * For better visibility of tunnel traffic OVS needs to retrieve 682 * egress tunnel information for a packet. Following API allows 683 * user to get this info. 684 */ 685 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 686 { 687 struct ip_tunnel_info *info; 688 689 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 690 return -EINVAL; 691 692 info = skb_tunnel_info_unclone(skb); 693 if (!info) 694 return -ENOMEM; 695 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 696 return -EINVAL; 697 698 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 699 } 700 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 701 702 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 703 { 704 int k = stack->num_paths++; 705 706 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 707 return NULL; 708 709 return &stack->path[k]; 710 } 711 712 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 713 struct net_device_path_stack *stack) 714 { 715 const struct net_device *last_dev; 716 struct net_device_path_ctx ctx = { 717 .dev = dev, 718 }; 719 struct net_device_path *path; 720 int ret = 0; 721 722 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 723 stack->num_paths = 0; 724 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 725 last_dev = ctx.dev; 726 path = dev_fwd_path(stack); 727 if (!path) 728 return -1; 729 730 memset(path, 0, sizeof(struct net_device_path)); 731 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 732 if (ret < 0) 733 return -1; 734 735 if (WARN_ON_ONCE(last_dev == ctx.dev)) 736 return -1; 737 } 738 739 if (!ctx.dev) 740 return ret; 741 742 path = dev_fwd_path(stack); 743 if (!path) 744 return -1; 745 path->type = DEV_PATH_ETHERNET; 746 path->dev = ctx.dev; 747 748 return ret; 749 } 750 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 751 752 /** 753 * __dev_get_by_name - find a device by its name 754 * @net: the applicable net namespace 755 * @name: name to find 756 * 757 * Find an interface by name. Must be called under RTNL semaphore. 758 * If the name is found a pointer to the device is returned. 759 * If the name is not found then %NULL is returned. The 760 * reference counters are not incremented so the caller must be 761 * careful with locks. 762 */ 763 764 struct net_device *__dev_get_by_name(struct net *net, const char *name) 765 { 766 struct netdev_name_node *node_name; 767 768 node_name = netdev_name_node_lookup(net, name); 769 return node_name ? node_name->dev : NULL; 770 } 771 EXPORT_SYMBOL(__dev_get_by_name); 772 773 /** 774 * dev_get_by_name_rcu - find a device by its name 775 * @net: the applicable net namespace 776 * @name: name to find 777 * 778 * Find an interface by name. 779 * If the name is found a pointer to the device is returned. 780 * If the name is not found then %NULL is returned. 781 * The reference counters are not incremented so the caller must be 782 * careful with locks. The caller must hold RCU lock. 783 */ 784 785 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 786 { 787 struct netdev_name_node *node_name; 788 789 node_name = netdev_name_node_lookup_rcu(net, name); 790 return node_name ? node_name->dev : NULL; 791 } 792 EXPORT_SYMBOL(dev_get_by_name_rcu); 793 794 /* Deprecated for new users, call netdev_get_by_name() instead */ 795 struct net_device *dev_get_by_name(struct net *net, const char *name) 796 { 797 struct net_device *dev; 798 799 rcu_read_lock(); 800 dev = dev_get_by_name_rcu(net, name); 801 dev_hold(dev); 802 rcu_read_unlock(); 803 return dev; 804 } 805 EXPORT_SYMBOL(dev_get_by_name); 806 807 /** 808 * netdev_get_by_name() - find a device by its name 809 * @net: the applicable net namespace 810 * @name: name to find 811 * @tracker: tracking object for the acquired reference 812 * @gfp: allocation flags for the tracker 813 * 814 * Find an interface by name. This can be called from any 815 * context and does its own locking. The returned handle has 816 * the usage count incremented and the caller must use netdev_put() to 817 * release it when it is no longer needed. %NULL is returned if no 818 * matching device is found. 819 */ 820 struct net_device *netdev_get_by_name(struct net *net, const char *name, 821 netdevice_tracker *tracker, gfp_t gfp) 822 { 823 struct net_device *dev; 824 825 dev = dev_get_by_name(net, name); 826 if (dev) 827 netdev_tracker_alloc(dev, tracker, gfp); 828 return dev; 829 } 830 EXPORT_SYMBOL(netdev_get_by_name); 831 832 /** 833 * __dev_get_by_index - find a device by its ifindex 834 * @net: the applicable net namespace 835 * @ifindex: index of device 836 * 837 * Search for an interface by index. Returns %NULL if the device 838 * is not found or a pointer to the device. The device has not 839 * had its reference counter increased so the caller must be careful 840 * about locking. The caller must hold the RTNL semaphore. 841 */ 842 843 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 844 { 845 struct net_device *dev; 846 struct hlist_head *head = dev_index_hash(net, ifindex); 847 848 hlist_for_each_entry(dev, head, index_hlist) 849 if (dev->ifindex == ifindex) 850 return dev; 851 852 return NULL; 853 } 854 EXPORT_SYMBOL(__dev_get_by_index); 855 856 /** 857 * dev_get_by_index_rcu - find a device by its ifindex 858 * @net: the applicable net namespace 859 * @ifindex: index of device 860 * 861 * Search for an interface by index. Returns %NULL if the device 862 * is not found or a pointer to the device. The device has not 863 * had its reference counter increased so the caller must be careful 864 * about locking. The caller must hold RCU lock. 865 */ 866 867 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 868 { 869 struct net_device *dev; 870 struct hlist_head *head = dev_index_hash(net, ifindex); 871 872 hlist_for_each_entry_rcu(dev, head, index_hlist) 873 if (dev->ifindex == ifindex) 874 return dev; 875 876 return NULL; 877 } 878 EXPORT_SYMBOL(dev_get_by_index_rcu); 879 880 /* Deprecated for new users, call netdev_get_by_index() instead */ 881 struct net_device *dev_get_by_index(struct net *net, int ifindex) 882 { 883 struct net_device *dev; 884 885 rcu_read_lock(); 886 dev = dev_get_by_index_rcu(net, ifindex); 887 dev_hold(dev); 888 rcu_read_unlock(); 889 return dev; 890 } 891 EXPORT_SYMBOL(dev_get_by_index); 892 893 /** 894 * netdev_get_by_index() - find a device by its ifindex 895 * @net: the applicable net namespace 896 * @ifindex: index of device 897 * @tracker: tracking object for the acquired reference 898 * @gfp: allocation flags for the tracker 899 * 900 * Search for an interface by index. Returns NULL if the device 901 * is not found or a pointer to the device. The device returned has 902 * had a reference added and the pointer is safe until the user calls 903 * netdev_put() to indicate they have finished with it. 904 */ 905 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 906 netdevice_tracker *tracker, gfp_t gfp) 907 { 908 struct net_device *dev; 909 910 dev = dev_get_by_index(net, ifindex); 911 if (dev) 912 netdev_tracker_alloc(dev, tracker, gfp); 913 return dev; 914 } 915 EXPORT_SYMBOL(netdev_get_by_index); 916 917 /** 918 * dev_get_by_napi_id - find a device by napi_id 919 * @napi_id: ID of the NAPI struct 920 * 921 * Search for an interface by NAPI ID. Returns %NULL if the device 922 * is not found or a pointer to the device. The device has not had 923 * its reference counter increased so the caller must be careful 924 * about locking. The caller must hold RCU lock. 925 */ 926 927 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 928 { 929 struct napi_struct *napi; 930 931 WARN_ON_ONCE(!rcu_read_lock_held()); 932 933 if (napi_id < MIN_NAPI_ID) 934 return NULL; 935 936 napi = napi_by_id(napi_id); 937 938 return napi ? napi->dev : NULL; 939 } 940 EXPORT_SYMBOL(dev_get_by_napi_id); 941 942 /** 943 * netdev_get_name - get a netdevice name, knowing its ifindex. 944 * @net: network namespace 945 * @name: a pointer to the buffer where the name will be stored. 946 * @ifindex: the ifindex of the interface to get the name from. 947 */ 948 int netdev_get_name(struct net *net, char *name, int ifindex) 949 { 950 struct net_device *dev; 951 int ret; 952 953 down_read(&devnet_rename_sem); 954 rcu_read_lock(); 955 956 dev = dev_get_by_index_rcu(net, ifindex); 957 if (!dev) { 958 ret = -ENODEV; 959 goto out; 960 } 961 962 strcpy(name, dev->name); 963 964 ret = 0; 965 out: 966 rcu_read_unlock(); 967 up_read(&devnet_rename_sem); 968 return ret; 969 } 970 971 /** 972 * dev_getbyhwaddr_rcu - find a device by its hardware address 973 * @net: the applicable net namespace 974 * @type: media type of device 975 * @ha: hardware address 976 * 977 * Search for an interface by MAC address. Returns NULL if the device 978 * is not found or a pointer to the device. 979 * The caller must hold RCU or RTNL. 980 * The returned device has not had its ref count increased 981 * and the caller must therefore be careful about locking 982 * 983 */ 984 985 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 986 const char *ha) 987 { 988 struct net_device *dev; 989 990 for_each_netdev_rcu(net, dev) 991 if (dev->type == type && 992 !memcmp(dev->dev_addr, ha, dev->addr_len)) 993 return dev; 994 995 return NULL; 996 } 997 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 998 999 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1000 { 1001 struct net_device *dev, *ret = NULL; 1002 1003 rcu_read_lock(); 1004 for_each_netdev_rcu(net, dev) 1005 if (dev->type == type) { 1006 dev_hold(dev); 1007 ret = dev; 1008 break; 1009 } 1010 rcu_read_unlock(); 1011 return ret; 1012 } 1013 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1014 1015 /** 1016 * __dev_get_by_flags - find any device with given flags 1017 * @net: the applicable net namespace 1018 * @if_flags: IFF_* values 1019 * @mask: bitmask of bits in if_flags to check 1020 * 1021 * Search for any interface with the given flags. Returns NULL if a device 1022 * is not found or a pointer to the device. Must be called inside 1023 * rtnl_lock(), and result refcount is unchanged. 1024 */ 1025 1026 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags, 1027 unsigned short mask) 1028 { 1029 struct net_device *dev, *ret; 1030 1031 ASSERT_RTNL(); 1032 1033 ret = NULL; 1034 for_each_netdev(net, dev) { 1035 if (((dev->flags ^ if_flags) & mask) == 0) { 1036 ret = dev; 1037 break; 1038 } 1039 } 1040 return ret; 1041 } 1042 EXPORT_SYMBOL(__dev_get_by_flags); 1043 1044 /** 1045 * dev_valid_name - check if name is okay for network device 1046 * @name: name string 1047 * 1048 * Network device names need to be valid file names to 1049 * allow sysfs to work. We also disallow any kind of 1050 * whitespace. 1051 */ 1052 bool dev_valid_name(const char *name) 1053 { 1054 if (*name == '\0') 1055 return false; 1056 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1057 return false; 1058 if (!strcmp(name, ".") || !strcmp(name, "..")) 1059 return false; 1060 1061 while (*name) { 1062 if (*name == '/' || *name == ':' || isspace(*name)) 1063 return false; 1064 name++; 1065 } 1066 return true; 1067 } 1068 EXPORT_SYMBOL(dev_valid_name); 1069 1070 /** 1071 * __dev_alloc_name - allocate a name for a device 1072 * @net: network namespace to allocate the device name in 1073 * @name: name format string 1074 * @res: result name string 1075 * 1076 * Passed a format string - eg "lt%d" it will try and find a suitable 1077 * id. It scans list of devices to build up a free map, then chooses 1078 * the first empty slot. The caller must hold the dev_base or rtnl lock 1079 * while allocating the name and adding the device in order to avoid 1080 * duplicates. 1081 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1082 * Returns the number of the unit assigned or a negative errno code. 1083 */ 1084 1085 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1086 { 1087 int i = 0; 1088 const char *p; 1089 const int max_netdevices = 8*PAGE_SIZE; 1090 unsigned long *inuse; 1091 struct net_device *d; 1092 char buf[IFNAMSIZ]; 1093 1094 /* Verify the string as this thing may have come from the user. 1095 * There must be one "%d" and no other "%" characters. 1096 */ 1097 p = strchr(name, '%'); 1098 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1099 return -EINVAL; 1100 1101 /* Use one page as a bit array of possible slots */ 1102 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1103 if (!inuse) 1104 return -ENOMEM; 1105 1106 for_each_netdev(net, d) { 1107 struct netdev_name_node *name_node; 1108 1109 netdev_for_each_altname(d, name_node) { 1110 if (!sscanf(name_node->name, name, &i)) 1111 continue; 1112 if (i < 0 || i >= max_netdevices) 1113 continue; 1114 1115 /* avoid cases where sscanf is not exact inverse of printf */ 1116 snprintf(buf, IFNAMSIZ, name, i); 1117 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1118 __set_bit(i, inuse); 1119 } 1120 if (!sscanf(d->name, name, &i)) 1121 continue; 1122 if (i < 0 || i >= max_netdevices) 1123 continue; 1124 1125 /* avoid cases where sscanf is not exact inverse of printf */ 1126 snprintf(buf, IFNAMSIZ, name, i); 1127 if (!strncmp(buf, d->name, IFNAMSIZ)) 1128 __set_bit(i, inuse); 1129 } 1130 1131 i = find_first_zero_bit(inuse, max_netdevices); 1132 bitmap_free(inuse); 1133 if (i == max_netdevices) 1134 return -ENFILE; 1135 1136 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1137 strscpy(buf, name, IFNAMSIZ); 1138 snprintf(res, IFNAMSIZ, buf, i); 1139 return i; 1140 } 1141 1142 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1143 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1144 const char *want_name, char *out_name, 1145 int dup_errno) 1146 { 1147 if (!dev_valid_name(want_name)) 1148 return -EINVAL; 1149 1150 if (strchr(want_name, '%')) 1151 return __dev_alloc_name(net, want_name, out_name); 1152 1153 if (netdev_name_in_use(net, want_name)) 1154 return -dup_errno; 1155 if (out_name != want_name) 1156 strscpy(out_name, want_name, IFNAMSIZ); 1157 return 0; 1158 } 1159 1160 /** 1161 * dev_alloc_name - allocate a name for a device 1162 * @dev: device 1163 * @name: name format string 1164 * 1165 * Passed a format string - eg "lt%d" it will try and find a suitable 1166 * id. It scans list of devices to build up a free map, then chooses 1167 * the first empty slot. The caller must hold the dev_base or rtnl lock 1168 * while allocating the name and adding the device in order to avoid 1169 * duplicates. 1170 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1171 * Returns the number of the unit assigned or a negative errno code. 1172 */ 1173 1174 int dev_alloc_name(struct net_device *dev, const char *name) 1175 { 1176 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1177 } 1178 EXPORT_SYMBOL(dev_alloc_name); 1179 1180 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1181 const char *name) 1182 { 1183 int ret; 1184 1185 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1186 return ret < 0 ? ret : 0; 1187 } 1188 1189 /** 1190 * dev_change_name - change name of a device 1191 * @dev: device 1192 * @newname: name (or format string) must be at least IFNAMSIZ 1193 * 1194 * Change name of a device, can pass format strings "eth%d". 1195 * for wildcarding. 1196 */ 1197 int dev_change_name(struct net_device *dev, const char *newname) 1198 { 1199 unsigned char old_assign_type; 1200 char oldname[IFNAMSIZ]; 1201 int err = 0; 1202 int ret; 1203 struct net *net; 1204 1205 ASSERT_RTNL(); 1206 BUG_ON(!dev_net(dev)); 1207 1208 net = dev_net(dev); 1209 1210 down_write(&devnet_rename_sem); 1211 1212 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) { 1213 up_write(&devnet_rename_sem); 1214 return 0; 1215 } 1216 1217 memcpy(oldname, dev->name, IFNAMSIZ); 1218 1219 err = dev_get_valid_name(net, dev, newname); 1220 if (err < 0) { 1221 up_write(&devnet_rename_sem); 1222 return err; 1223 } 1224 1225 if (oldname[0] && !strchr(oldname, '%')) 1226 netdev_info(dev, "renamed from %s%s\n", oldname, 1227 dev->flags & IFF_UP ? " (while UP)" : ""); 1228 1229 old_assign_type = dev->name_assign_type; 1230 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1231 1232 rollback: 1233 ret = device_rename(&dev->dev, dev->name); 1234 if (ret) { 1235 memcpy(dev->name, oldname, IFNAMSIZ); 1236 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1237 up_write(&devnet_rename_sem); 1238 return ret; 1239 } 1240 1241 up_write(&devnet_rename_sem); 1242 1243 netdev_adjacent_rename_links(dev, oldname); 1244 1245 netdev_name_node_del(dev->name_node); 1246 1247 synchronize_net(); 1248 1249 netdev_name_node_add(net, dev->name_node); 1250 1251 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1252 ret = notifier_to_errno(ret); 1253 1254 if (ret) { 1255 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1256 if (err >= 0) { 1257 err = ret; 1258 down_write(&devnet_rename_sem); 1259 memcpy(dev->name, oldname, IFNAMSIZ); 1260 memcpy(oldname, newname, IFNAMSIZ); 1261 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1262 old_assign_type = NET_NAME_RENAMED; 1263 goto rollback; 1264 } else { 1265 netdev_err(dev, "name change rollback failed: %d\n", 1266 ret); 1267 } 1268 } 1269 1270 return err; 1271 } 1272 1273 /** 1274 * dev_set_alias - change ifalias of a device 1275 * @dev: device 1276 * @alias: name up to IFALIASZ 1277 * @len: limit of bytes to copy from info 1278 * 1279 * Set ifalias for a device, 1280 */ 1281 int dev_set_alias(struct net_device *dev, const char *alias, size_t len) 1282 { 1283 struct dev_ifalias *new_alias = NULL; 1284 1285 if (len >= IFALIASZ) 1286 return -EINVAL; 1287 1288 if (len) { 1289 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1290 if (!new_alias) 1291 return -ENOMEM; 1292 1293 memcpy(new_alias->ifalias, alias, len); 1294 new_alias->ifalias[len] = 0; 1295 } 1296 1297 mutex_lock(&ifalias_mutex); 1298 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1299 mutex_is_locked(&ifalias_mutex)); 1300 mutex_unlock(&ifalias_mutex); 1301 1302 if (new_alias) 1303 kfree_rcu(new_alias, rcuhead); 1304 1305 return len; 1306 } 1307 EXPORT_SYMBOL(dev_set_alias); 1308 1309 /** 1310 * dev_get_alias - get ifalias of a device 1311 * @dev: device 1312 * @name: buffer to store name of ifalias 1313 * @len: size of buffer 1314 * 1315 * get ifalias for a device. Caller must make sure dev cannot go 1316 * away, e.g. rcu read lock or own a reference count to device. 1317 */ 1318 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1319 { 1320 const struct dev_ifalias *alias; 1321 int ret = 0; 1322 1323 rcu_read_lock(); 1324 alias = rcu_dereference(dev->ifalias); 1325 if (alias) 1326 ret = snprintf(name, len, "%s", alias->ifalias); 1327 rcu_read_unlock(); 1328 1329 return ret; 1330 } 1331 1332 /** 1333 * netdev_features_change - device changes features 1334 * @dev: device to cause notification 1335 * 1336 * Called to indicate a device has changed features. 1337 */ 1338 void netdev_features_change(struct net_device *dev) 1339 { 1340 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1341 } 1342 EXPORT_SYMBOL(netdev_features_change); 1343 1344 /** 1345 * netdev_state_change - device changes state 1346 * @dev: device to cause notification 1347 * 1348 * Called to indicate a device has changed state. This function calls 1349 * the notifier chains for netdev_chain and sends a NEWLINK message 1350 * to the routing socket. 1351 */ 1352 void netdev_state_change(struct net_device *dev) 1353 { 1354 if (dev->flags & IFF_UP) { 1355 struct netdev_notifier_change_info change_info = { 1356 .info.dev = dev, 1357 }; 1358 1359 call_netdevice_notifiers_info(NETDEV_CHANGE, 1360 &change_info.info); 1361 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1362 } 1363 } 1364 EXPORT_SYMBOL(netdev_state_change); 1365 1366 /** 1367 * __netdev_notify_peers - notify network peers about existence of @dev, 1368 * to be called when rtnl lock is already held. 1369 * @dev: network device 1370 * 1371 * Generate traffic such that interested network peers are aware of 1372 * @dev, such as by generating a gratuitous ARP. This may be used when 1373 * a device wants to inform the rest of the network about some sort of 1374 * reconfiguration such as a failover event or virtual machine 1375 * migration. 1376 */ 1377 void __netdev_notify_peers(struct net_device *dev) 1378 { 1379 ASSERT_RTNL(); 1380 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1381 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1382 } 1383 EXPORT_SYMBOL(__netdev_notify_peers); 1384 1385 /** 1386 * netdev_notify_peers - notify network peers about existence of @dev 1387 * @dev: network device 1388 * 1389 * Generate traffic such that interested network peers are aware of 1390 * @dev, such as by generating a gratuitous ARP. This may be used when 1391 * a device wants to inform the rest of the network about some sort of 1392 * reconfiguration such as a failover event or virtual machine 1393 * migration. 1394 */ 1395 void netdev_notify_peers(struct net_device *dev) 1396 { 1397 rtnl_lock(); 1398 __netdev_notify_peers(dev); 1399 rtnl_unlock(); 1400 } 1401 EXPORT_SYMBOL(netdev_notify_peers); 1402 1403 static int napi_threaded_poll(void *data); 1404 1405 static int napi_kthread_create(struct napi_struct *n) 1406 { 1407 int err = 0; 1408 1409 /* Create and wake up the kthread once to put it in 1410 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1411 * warning and work with loadavg. 1412 */ 1413 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1414 n->dev->name, n->napi_id); 1415 if (IS_ERR(n->thread)) { 1416 err = PTR_ERR(n->thread); 1417 pr_err("kthread_run failed with err %d\n", err); 1418 n->thread = NULL; 1419 } 1420 1421 return err; 1422 } 1423 1424 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1425 { 1426 const struct net_device_ops *ops = dev->netdev_ops; 1427 int ret; 1428 1429 ASSERT_RTNL(); 1430 dev_addr_check(dev); 1431 1432 if (!netif_device_present(dev)) { 1433 /* may be detached because parent is runtime-suspended */ 1434 if (dev->dev.parent) 1435 pm_runtime_resume(dev->dev.parent); 1436 if (!netif_device_present(dev)) 1437 return -ENODEV; 1438 } 1439 1440 /* Block netpoll from trying to do any rx path servicing. 1441 * If we don't do this there is a chance ndo_poll_controller 1442 * or ndo_poll may be running while we open the device 1443 */ 1444 netpoll_poll_disable(dev); 1445 1446 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1447 ret = notifier_to_errno(ret); 1448 if (ret) 1449 return ret; 1450 1451 set_bit(__LINK_STATE_START, &dev->state); 1452 1453 if (ops->ndo_validate_addr) 1454 ret = ops->ndo_validate_addr(dev); 1455 1456 if (!ret && ops->ndo_open) 1457 ret = ops->ndo_open(dev); 1458 1459 netpoll_poll_enable(dev); 1460 1461 if (ret) 1462 clear_bit(__LINK_STATE_START, &dev->state); 1463 else { 1464 dev->flags |= IFF_UP; 1465 dev_set_rx_mode(dev); 1466 dev_activate(dev); 1467 add_device_randomness(dev->dev_addr, dev->addr_len); 1468 } 1469 1470 return ret; 1471 } 1472 1473 /** 1474 * dev_open - prepare an interface for use. 1475 * @dev: device to open 1476 * @extack: netlink extended ack 1477 * 1478 * Takes a device from down to up state. The device's private open 1479 * function is invoked and then the multicast lists are loaded. Finally 1480 * the device is moved into the up state and a %NETDEV_UP message is 1481 * sent to the netdev notifier chain. 1482 * 1483 * Calling this function on an active interface is a nop. On a failure 1484 * a negative errno code is returned. 1485 */ 1486 int dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1487 { 1488 int ret; 1489 1490 if (dev->flags & IFF_UP) 1491 return 0; 1492 1493 ret = __dev_open(dev, extack); 1494 if (ret < 0) 1495 return ret; 1496 1497 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1498 call_netdevice_notifiers(NETDEV_UP, dev); 1499 1500 return ret; 1501 } 1502 EXPORT_SYMBOL(dev_open); 1503 1504 static void __dev_close_many(struct list_head *head) 1505 { 1506 struct net_device *dev; 1507 1508 ASSERT_RTNL(); 1509 might_sleep(); 1510 1511 list_for_each_entry(dev, head, close_list) { 1512 /* Temporarily disable netpoll until the interface is down */ 1513 netpoll_poll_disable(dev); 1514 1515 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1516 1517 clear_bit(__LINK_STATE_START, &dev->state); 1518 1519 /* Synchronize to scheduled poll. We cannot touch poll list, it 1520 * can be even on different cpu. So just clear netif_running(). 1521 * 1522 * dev->stop() will invoke napi_disable() on all of it's 1523 * napi_struct instances on this device. 1524 */ 1525 smp_mb__after_atomic(); /* Commit netif_running(). */ 1526 } 1527 1528 dev_deactivate_many(head); 1529 1530 list_for_each_entry(dev, head, close_list) { 1531 const struct net_device_ops *ops = dev->netdev_ops; 1532 1533 /* 1534 * Call the device specific close. This cannot fail. 1535 * Only if device is UP 1536 * 1537 * We allow it to be called even after a DETACH hot-plug 1538 * event. 1539 */ 1540 if (ops->ndo_stop) 1541 ops->ndo_stop(dev); 1542 1543 dev->flags &= ~IFF_UP; 1544 netpoll_poll_enable(dev); 1545 } 1546 } 1547 1548 static void __dev_close(struct net_device *dev) 1549 { 1550 LIST_HEAD(single); 1551 1552 list_add(&dev->close_list, &single); 1553 __dev_close_many(&single); 1554 list_del(&single); 1555 } 1556 1557 void dev_close_many(struct list_head *head, bool unlink) 1558 { 1559 struct net_device *dev, *tmp; 1560 1561 /* Remove the devices that don't need to be closed */ 1562 list_for_each_entry_safe(dev, tmp, head, close_list) 1563 if (!(dev->flags & IFF_UP)) 1564 list_del_init(&dev->close_list); 1565 1566 __dev_close_many(head); 1567 1568 list_for_each_entry_safe(dev, tmp, head, close_list) { 1569 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1570 call_netdevice_notifiers(NETDEV_DOWN, dev); 1571 if (unlink) 1572 list_del_init(&dev->close_list); 1573 } 1574 } 1575 EXPORT_SYMBOL(dev_close_many); 1576 1577 /** 1578 * dev_close - shutdown an interface. 1579 * @dev: device to shutdown 1580 * 1581 * This function moves an active device into down state. A 1582 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device 1583 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier 1584 * chain. 1585 */ 1586 void dev_close(struct net_device *dev) 1587 { 1588 if (dev->flags & IFF_UP) { 1589 LIST_HEAD(single); 1590 1591 list_add(&dev->close_list, &single); 1592 dev_close_many(&single, true); 1593 list_del(&single); 1594 } 1595 } 1596 EXPORT_SYMBOL(dev_close); 1597 1598 1599 /** 1600 * dev_disable_lro - disable Large Receive Offload on a device 1601 * @dev: device 1602 * 1603 * Disable Large Receive Offload (LRO) on a net device. Must be 1604 * called under RTNL. This is needed if received packets may be 1605 * forwarded to another interface. 1606 */ 1607 void dev_disable_lro(struct net_device *dev) 1608 { 1609 struct net_device *lower_dev; 1610 struct list_head *iter; 1611 1612 dev->wanted_features &= ~NETIF_F_LRO; 1613 netdev_update_features(dev); 1614 1615 if (unlikely(dev->features & NETIF_F_LRO)) 1616 netdev_WARN(dev, "failed to disable LRO!\n"); 1617 1618 netdev_for_each_lower_dev(dev, lower_dev, iter) 1619 dev_disable_lro(lower_dev); 1620 } 1621 EXPORT_SYMBOL(dev_disable_lro); 1622 1623 /** 1624 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1625 * @dev: device 1626 * 1627 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1628 * called under RTNL. This is needed if Generic XDP is installed on 1629 * the device. 1630 */ 1631 static void dev_disable_gro_hw(struct net_device *dev) 1632 { 1633 dev->wanted_features &= ~NETIF_F_GRO_HW; 1634 netdev_update_features(dev); 1635 1636 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1637 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1638 } 1639 1640 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1641 { 1642 #define N(val) \ 1643 case NETDEV_##val: \ 1644 return "NETDEV_" __stringify(val); 1645 switch (cmd) { 1646 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1647 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1648 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1649 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1650 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1651 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1652 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1653 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1654 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1655 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1656 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1657 N(XDP_FEAT_CHANGE) 1658 } 1659 #undef N 1660 return "UNKNOWN_NETDEV_EVENT"; 1661 } 1662 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1663 1664 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1665 struct net_device *dev) 1666 { 1667 struct netdev_notifier_info info = { 1668 .dev = dev, 1669 }; 1670 1671 return nb->notifier_call(nb, val, &info); 1672 } 1673 1674 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1675 struct net_device *dev) 1676 { 1677 int err; 1678 1679 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1680 err = notifier_to_errno(err); 1681 if (err) 1682 return err; 1683 1684 if (!(dev->flags & IFF_UP)) 1685 return 0; 1686 1687 call_netdevice_notifier(nb, NETDEV_UP, dev); 1688 return 0; 1689 } 1690 1691 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1692 struct net_device *dev) 1693 { 1694 if (dev->flags & IFF_UP) { 1695 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1696 dev); 1697 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1698 } 1699 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1700 } 1701 1702 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1703 struct net *net) 1704 { 1705 struct net_device *dev; 1706 int err; 1707 1708 for_each_netdev(net, dev) { 1709 err = call_netdevice_register_notifiers(nb, dev); 1710 if (err) 1711 goto rollback; 1712 } 1713 return 0; 1714 1715 rollback: 1716 for_each_netdev_continue_reverse(net, dev) 1717 call_netdevice_unregister_notifiers(nb, dev); 1718 return err; 1719 } 1720 1721 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1722 struct net *net) 1723 { 1724 struct net_device *dev; 1725 1726 for_each_netdev(net, dev) 1727 call_netdevice_unregister_notifiers(nb, dev); 1728 } 1729 1730 static int dev_boot_phase = 1; 1731 1732 /** 1733 * register_netdevice_notifier - register a network notifier block 1734 * @nb: notifier 1735 * 1736 * Register a notifier to be called when network device events occur. 1737 * The notifier passed is linked into the kernel structures and must 1738 * not be reused until it has been unregistered. A negative errno code 1739 * is returned on a failure. 1740 * 1741 * When registered all registration and up events are replayed 1742 * to the new notifier to allow device to have a race free 1743 * view of the network device list. 1744 */ 1745 1746 int register_netdevice_notifier(struct notifier_block *nb) 1747 { 1748 struct net *net; 1749 int err; 1750 1751 /* Close race with setup_net() and cleanup_net() */ 1752 down_write(&pernet_ops_rwsem); 1753 rtnl_lock(); 1754 err = raw_notifier_chain_register(&netdev_chain, nb); 1755 if (err) 1756 goto unlock; 1757 if (dev_boot_phase) 1758 goto unlock; 1759 for_each_net(net) { 1760 err = call_netdevice_register_net_notifiers(nb, net); 1761 if (err) 1762 goto rollback; 1763 } 1764 1765 unlock: 1766 rtnl_unlock(); 1767 up_write(&pernet_ops_rwsem); 1768 return err; 1769 1770 rollback: 1771 for_each_net_continue_reverse(net) 1772 call_netdevice_unregister_net_notifiers(nb, net); 1773 1774 raw_notifier_chain_unregister(&netdev_chain, nb); 1775 goto unlock; 1776 } 1777 EXPORT_SYMBOL(register_netdevice_notifier); 1778 1779 /** 1780 * unregister_netdevice_notifier - unregister a network notifier block 1781 * @nb: notifier 1782 * 1783 * Unregister a notifier previously registered by 1784 * register_netdevice_notifier(). The notifier is unlinked into the 1785 * kernel structures and may then be reused. A negative errno code 1786 * is returned on a failure. 1787 * 1788 * After unregistering unregister and down device events are synthesized 1789 * for all devices on the device list to the removed notifier to remove 1790 * the need for special case cleanup code. 1791 */ 1792 1793 int unregister_netdevice_notifier(struct notifier_block *nb) 1794 { 1795 struct net *net; 1796 int err; 1797 1798 /* Close race with setup_net() and cleanup_net() */ 1799 down_write(&pernet_ops_rwsem); 1800 rtnl_lock(); 1801 err = raw_notifier_chain_unregister(&netdev_chain, nb); 1802 if (err) 1803 goto unlock; 1804 1805 for_each_net(net) 1806 call_netdevice_unregister_net_notifiers(nb, net); 1807 1808 unlock: 1809 rtnl_unlock(); 1810 up_write(&pernet_ops_rwsem); 1811 return err; 1812 } 1813 EXPORT_SYMBOL(unregister_netdevice_notifier); 1814 1815 static int __register_netdevice_notifier_net(struct net *net, 1816 struct notifier_block *nb, 1817 bool ignore_call_fail) 1818 { 1819 int err; 1820 1821 err = raw_notifier_chain_register(&net->netdev_chain, nb); 1822 if (err) 1823 return err; 1824 if (dev_boot_phase) 1825 return 0; 1826 1827 err = call_netdevice_register_net_notifiers(nb, net); 1828 if (err && !ignore_call_fail) 1829 goto chain_unregister; 1830 1831 return 0; 1832 1833 chain_unregister: 1834 raw_notifier_chain_unregister(&net->netdev_chain, nb); 1835 return err; 1836 } 1837 1838 static int __unregister_netdevice_notifier_net(struct net *net, 1839 struct notifier_block *nb) 1840 { 1841 int err; 1842 1843 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 1844 if (err) 1845 return err; 1846 1847 call_netdevice_unregister_net_notifiers(nb, net); 1848 return 0; 1849 } 1850 1851 /** 1852 * register_netdevice_notifier_net - register a per-netns network notifier block 1853 * @net: network namespace 1854 * @nb: notifier 1855 * 1856 * Register a notifier to be called when network device events occur. 1857 * The notifier passed is linked into the kernel structures and must 1858 * not be reused until it has been unregistered. A negative errno code 1859 * is returned on a failure. 1860 * 1861 * When registered all registration and up events are replayed 1862 * to the new notifier to allow device to have a race free 1863 * view of the network device list. 1864 */ 1865 1866 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 1867 { 1868 int err; 1869 1870 rtnl_lock(); 1871 err = __register_netdevice_notifier_net(net, nb, false); 1872 rtnl_unlock(); 1873 return err; 1874 } 1875 EXPORT_SYMBOL(register_netdevice_notifier_net); 1876 1877 /** 1878 * unregister_netdevice_notifier_net - unregister a per-netns 1879 * network notifier block 1880 * @net: network namespace 1881 * @nb: notifier 1882 * 1883 * Unregister a notifier previously registered by 1884 * register_netdevice_notifier_net(). The notifier is unlinked from the 1885 * kernel structures and may then be reused. A negative errno code 1886 * is returned on a failure. 1887 * 1888 * After unregistering unregister and down device events are synthesized 1889 * for all devices on the device list to the removed notifier to remove 1890 * the need for special case cleanup code. 1891 */ 1892 1893 int unregister_netdevice_notifier_net(struct net *net, 1894 struct notifier_block *nb) 1895 { 1896 int err; 1897 1898 rtnl_lock(); 1899 err = __unregister_netdevice_notifier_net(net, nb); 1900 rtnl_unlock(); 1901 return err; 1902 } 1903 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 1904 1905 static void __move_netdevice_notifier_net(struct net *src_net, 1906 struct net *dst_net, 1907 struct notifier_block *nb) 1908 { 1909 __unregister_netdevice_notifier_net(src_net, nb); 1910 __register_netdevice_notifier_net(dst_net, nb, true); 1911 } 1912 1913 int register_netdevice_notifier_dev_net(struct net_device *dev, 1914 struct notifier_block *nb, 1915 struct netdev_net_notifier *nn) 1916 { 1917 int err; 1918 1919 rtnl_lock(); 1920 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 1921 if (!err) { 1922 nn->nb = nb; 1923 list_add(&nn->list, &dev->net_notifier_list); 1924 } 1925 rtnl_unlock(); 1926 return err; 1927 } 1928 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 1929 1930 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 1931 struct notifier_block *nb, 1932 struct netdev_net_notifier *nn) 1933 { 1934 int err; 1935 1936 rtnl_lock(); 1937 list_del(&nn->list); 1938 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 1939 rtnl_unlock(); 1940 return err; 1941 } 1942 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 1943 1944 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 1945 struct net *net) 1946 { 1947 struct netdev_net_notifier *nn; 1948 1949 list_for_each_entry(nn, &dev->net_notifier_list, list) 1950 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 1951 } 1952 1953 /** 1954 * call_netdevice_notifiers_info - call all network notifier blocks 1955 * @val: value passed unmodified to notifier function 1956 * @info: notifier information data 1957 * 1958 * Call all network notifier blocks. Parameters and return value 1959 * are as for raw_notifier_call_chain(). 1960 */ 1961 1962 int call_netdevice_notifiers_info(unsigned long val, 1963 struct netdev_notifier_info *info) 1964 { 1965 struct net *net = dev_net(info->dev); 1966 int ret; 1967 1968 ASSERT_RTNL(); 1969 1970 /* Run per-netns notifier block chain first, then run the global one. 1971 * Hopefully, one day, the global one is going to be removed after 1972 * all notifier block registrators get converted to be per-netns. 1973 */ 1974 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 1975 if (ret & NOTIFY_STOP_MASK) 1976 return ret; 1977 return raw_notifier_call_chain(&netdev_chain, val, info); 1978 } 1979 1980 /** 1981 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 1982 * for and rollback on error 1983 * @val_up: value passed unmodified to notifier function 1984 * @val_down: value passed unmodified to the notifier function when 1985 * recovering from an error on @val_up 1986 * @info: notifier information data 1987 * 1988 * Call all per-netns network notifier blocks, but not notifier blocks on 1989 * the global notifier chain. Parameters and return value are as for 1990 * raw_notifier_call_chain_robust(). 1991 */ 1992 1993 static int 1994 call_netdevice_notifiers_info_robust(unsigned long val_up, 1995 unsigned long val_down, 1996 struct netdev_notifier_info *info) 1997 { 1998 struct net *net = dev_net(info->dev); 1999 2000 ASSERT_RTNL(); 2001 2002 return raw_notifier_call_chain_robust(&net->netdev_chain, 2003 val_up, val_down, info); 2004 } 2005 2006 static int call_netdevice_notifiers_extack(unsigned long val, 2007 struct net_device *dev, 2008 struct netlink_ext_ack *extack) 2009 { 2010 struct netdev_notifier_info info = { 2011 .dev = dev, 2012 .extack = extack, 2013 }; 2014 2015 return call_netdevice_notifiers_info(val, &info); 2016 } 2017 2018 /** 2019 * call_netdevice_notifiers - call all network notifier blocks 2020 * @val: value passed unmodified to notifier function 2021 * @dev: net_device pointer passed unmodified to notifier function 2022 * 2023 * Call all network notifier blocks. Parameters and return value 2024 * are as for raw_notifier_call_chain(). 2025 */ 2026 2027 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2028 { 2029 return call_netdevice_notifiers_extack(val, dev, NULL); 2030 } 2031 EXPORT_SYMBOL(call_netdevice_notifiers); 2032 2033 /** 2034 * call_netdevice_notifiers_mtu - call all network notifier blocks 2035 * @val: value passed unmodified to notifier function 2036 * @dev: net_device pointer passed unmodified to notifier function 2037 * @arg: additional u32 argument passed to the notifier function 2038 * 2039 * Call all network notifier blocks. Parameters and return value 2040 * are as for raw_notifier_call_chain(). 2041 */ 2042 static int call_netdevice_notifiers_mtu(unsigned long val, 2043 struct net_device *dev, u32 arg) 2044 { 2045 struct netdev_notifier_info_ext info = { 2046 .info.dev = dev, 2047 .ext.mtu = arg, 2048 }; 2049 2050 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2051 2052 return call_netdevice_notifiers_info(val, &info.info); 2053 } 2054 2055 #ifdef CONFIG_NET_INGRESS 2056 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2057 2058 void net_inc_ingress_queue(void) 2059 { 2060 static_branch_inc(&ingress_needed_key); 2061 } 2062 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2063 2064 void net_dec_ingress_queue(void) 2065 { 2066 static_branch_dec(&ingress_needed_key); 2067 } 2068 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2069 #endif 2070 2071 #ifdef CONFIG_NET_EGRESS 2072 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2073 2074 void net_inc_egress_queue(void) 2075 { 2076 static_branch_inc(&egress_needed_key); 2077 } 2078 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2079 2080 void net_dec_egress_queue(void) 2081 { 2082 static_branch_dec(&egress_needed_key); 2083 } 2084 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2085 #endif 2086 2087 #ifdef CONFIG_NET_CLS_ACT 2088 DEFINE_STATIC_KEY_FALSE(tcf_bypass_check_needed_key); 2089 EXPORT_SYMBOL(tcf_bypass_check_needed_key); 2090 #endif 2091 2092 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2093 EXPORT_SYMBOL(netstamp_needed_key); 2094 #ifdef CONFIG_JUMP_LABEL 2095 static atomic_t netstamp_needed_deferred; 2096 static atomic_t netstamp_wanted; 2097 static void netstamp_clear(struct work_struct *work) 2098 { 2099 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2100 int wanted; 2101 2102 wanted = atomic_add_return(deferred, &netstamp_wanted); 2103 if (wanted > 0) 2104 static_branch_enable(&netstamp_needed_key); 2105 else 2106 static_branch_disable(&netstamp_needed_key); 2107 } 2108 static DECLARE_WORK(netstamp_work, netstamp_clear); 2109 #endif 2110 2111 void net_enable_timestamp(void) 2112 { 2113 #ifdef CONFIG_JUMP_LABEL 2114 int wanted = atomic_read(&netstamp_wanted); 2115 2116 while (wanted > 0) { 2117 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2118 return; 2119 } 2120 atomic_inc(&netstamp_needed_deferred); 2121 schedule_work(&netstamp_work); 2122 #else 2123 static_branch_inc(&netstamp_needed_key); 2124 #endif 2125 } 2126 EXPORT_SYMBOL(net_enable_timestamp); 2127 2128 void net_disable_timestamp(void) 2129 { 2130 #ifdef CONFIG_JUMP_LABEL 2131 int wanted = atomic_read(&netstamp_wanted); 2132 2133 while (wanted > 1) { 2134 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2135 return; 2136 } 2137 atomic_dec(&netstamp_needed_deferred); 2138 schedule_work(&netstamp_work); 2139 #else 2140 static_branch_dec(&netstamp_needed_key); 2141 #endif 2142 } 2143 EXPORT_SYMBOL(net_disable_timestamp); 2144 2145 static inline void net_timestamp_set(struct sk_buff *skb) 2146 { 2147 skb->tstamp = 0; 2148 skb->mono_delivery_time = 0; 2149 if (static_branch_unlikely(&netstamp_needed_key)) 2150 skb->tstamp = ktime_get_real(); 2151 } 2152 2153 #define net_timestamp_check(COND, SKB) \ 2154 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2155 if ((COND) && !(SKB)->tstamp) \ 2156 (SKB)->tstamp = ktime_get_real(); \ 2157 } \ 2158 2159 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2160 { 2161 return __is_skb_forwardable(dev, skb, true); 2162 } 2163 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2164 2165 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2166 bool check_mtu) 2167 { 2168 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2169 2170 if (likely(!ret)) { 2171 skb->protocol = eth_type_trans(skb, dev); 2172 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2173 } 2174 2175 return ret; 2176 } 2177 2178 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2179 { 2180 return __dev_forward_skb2(dev, skb, true); 2181 } 2182 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2183 2184 /** 2185 * dev_forward_skb - loopback an skb to another netif 2186 * 2187 * @dev: destination network device 2188 * @skb: buffer to forward 2189 * 2190 * return values: 2191 * NET_RX_SUCCESS (no congestion) 2192 * NET_RX_DROP (packet was dropped, but freed) 2193 * 2194 * dev_forward_skb can be used for injecting an skb from the 2195 * start_xmit function of one device into the receive queue 2196 * of another device. 2197 * 2198 * The receiving device may be in another namespace, so 2199 * we have to clear all information in the skb that could 2200 * impact namespace isolation. 2201 */ 2202 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2203 { 2204 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2205 } 2206 EXPORT_SYMBOL_GPL(dev_forward_skb); 2207 2208 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2209 { 2210 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2211 } 2212 2213 static inline int deliver_skb(struct sk_buff *skb, 2214 struct packet_type *pt_prev, 2215 struct net_device *orig_dev) 2216 { 2217 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2218 return -ENOMEM; 2219 refcount_inc(&skb->users); 2220 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2221 } 2222 2223 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2224 struct packet_type **pt, 2225 struct net_device *orig_dev, 2226 __be16 type, 2227 struct list_head *ptype_list) 2228 { 2229 struct packet_type *ptype, *pt_prev = *pt; 2230 2231 list_for_each_entry_rcu(ptype, ptype_list, list) { 2232 if (ptype->type != type) 2233 continue; 2234 if (pt_prev) 2235 deliver_skb(skb, pt_prev, orig_dev); 2236 pt_prev = ptype; 2237 } 2238 *pt = pt_prev; 2239 } 2240 2241 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2242 { 2243 if (!ptype->af_packet_priv || !skb->sk) 2244 return false; 2245 2246 if (ptype->id_match) 2247 return ptype->id_match(ptype, skb->sk); 2248 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2249 return true; 2250 2251 return false; 2252 } 2253 2254 /** 2255 * dev_nit_active - return true if any network interface taps are in use 2256 * 2257 * @dev: network device to check for the presence of taps 2258 */ 2259 bool dev_nit_active(struct net_device *dev) 2260 { 2261 return !list_empty(&net_hotdata.ptype_all) || 2262 !list_empty(&dev->ptype_all); 2263 } 2264 EXPORT_SYMBOL_GPL(dev_nit_active); 2265 2266 /* 2267 * Support routine. Sends outgoing frames to any network 2268 * taps currently in use. 2269 */ 2270 2271 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2272 { 2273 struct list_head *ptype_list = &net_hotdata.ptype_all; 2274 struct packet_type *ptype, *pt_prev = NULL; 2275 struct sk_buff *skb2 = NULL; 2276 2277 rcu_read_lock(); 2278 again: 2279 list_for_each_entry_rcu(ptype, ptype_list, list) { 2280 if (READ_ONCE(ptype->ignore_outgoing)) 2281 continue; 2282 2283 /* Never send packets back to the socket 2284 * they originated from - MvS (miquels@drinkel.ow.org) 2285 */ 2286 if (skb_loop_sk(ptype, skb)) 2287 continue; 2288 2289 if (pt_prev) { 2290 deliver_skb(skb2, pt_prev, skb->dev); 2291 pt_prev = ptype; 2292 continue; 2293 } 2294 2295 /* need to clone skb, done only once */ 2296 skb2 = skb_clone(skb, GFP_ATOMIC); 2297 if (!skb2) 2298 goto out_unlock; 2299 2300 net_timestamp_set(skb2); 2301 2302 /* skb->nh should be correctly 2303 * set by sender, so that the second statement is 2304 * just protection against buggy protocols. 2305 */ 2306 skb_reset_mac_header(skb2); 2307 2308 if (skb_network_header(skb2) < skb2->data || 2309 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2310 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2311 ntohs(skb2->protocol), 2312 dev->name); 2313 skb_reset_network_header(skb2); 2314 } 2315 2316 skb2->transport_header = skb2->network_header; 2317 skb2->pkt_type = PACKET_OUTGOING; 2318 pt_prev = ptype; 2319 } 2320 2321 if (ptype_list == &net_hotdata.ptype_all) { 2322 ptype_list = &dev->ptype_all; 2323 goto again; 2324 } 2325 out_unlock: 2326 if (pt_prev) { 2327 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2328 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2329 else 2330 kfree_skb(skb2); 2331 } 2332 rcu_read_unlock(); 2333 } 2334 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2335 2336 /** 2337 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2338 * @dev: Network device 2339 * @txq: number of queues available 2340 * 2341 * If real_num_tx_queues is changed the tc mappings may no longer be 2342 * valid. To resolve this verify the tc mapping remains valid and if 2343 * not NULL the mapping. With no priorities mapping to this 2344 * offset/count pair it will no longer be used. In the worst case TC0 2345 * is invalid nothing can be done so disable priority mappings. If is 2346 * expected that drivers will fix this mapping if they can before 2347 * calling netif_set_real_num_tx_queues. 2348 */ 2349 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2350 { 2351 int i; 2352 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2353 2354 /* If TC0 is invalidated disable TC mapping */ 2355 if (tc->offset + tc->count > txq) { 2356 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2357 dev->num_tc = 0; 2358 return; 2359 } 2360 2361 /* Invalidated prio to tc mappings set to TC0 */ 2362 for (i = 1; i < TC_BITMASK + 1; i++) { 2363 int q = netdev_get_prio_tc_map(dev, i); 2364 2365 tc = &dev->tc_to_txq[q]; 2366 if (tc->offset + tc->count > txq) { 2367 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", 2368 i, q); 2369 netdev_set_prio_tc_map(dev, i, 0); 2370 } 2371 } 2372 } 2373 2374 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2375 { 2376 if (dev->num_tc) { 2377 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2378 int i; 2379 2380 /* walk through the TCs and see if it falls into any of them */ 2381 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2382 if ((txq - tc->offset) < tc->count) 2383 return i; 2384 } 2385 2386 /* didn't find it, just return -1 to indicate no match */ 2387 return -1; 2388 } 2389 2390 return 0; 2391 } 2392 EXPORT_SYMBOL(netdev_txq_to_tc); 2393 2394 #ifdef CONFIG_XPS 2395 static struct static_key xps_needed __read_mostly; 2396 static struct static_key xps_rxqs_needed __read_mostly; 2397 static DEFINE_MUTEX(xps_map_mutex); 2398 #define xmap_dereference(P) \ 2399 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2400 2401 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2402 struct xps_dev_maps *old_maps, int tci, u16 index) 2403 { 2404 struct xps_map *map = NULL; 2405 int pos; 2406 2407 map = xmap_dereference(dev_maps->attr_map[tci]); 2408 if (!map) 2409 return false; 2410 2411 for (pos = map->len; pos--;) { 2412 if (map->queues[pos] != index) 2413 continue; 2414 2415 if (map->len > 1) { 2416 map->queues[pos] = map->queues[--map->len]; 2417 break; 2418 } 2419 2420 if (old_maps) 2421 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2422 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2423 kfree_rcu(map, rcu); 2424 return false; 2425 } 2426 2427 return true; 2428 } 2429 2430 static bool remove_xps_queue_cpu(struct net_device *dev, 2431 struct xps_dev_maps *dev_maps, 2432 int cpu, u16 offset, u16 count) 2433 { 2434 int num_tc = dev_maps->num_tc; 2435 bool active = false; 2436 int tci; 2437 2438 for (tci = cpu * num_tc; num_tc--; tci++) { 2439 int i, j; 2440 2441 for (i = count, j = offset; i--; j++) { 2442 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2443 break; 2444 } 2445 2446 active |= i < 0; 2447 } 2448 2449 return active; 2450 } 2451 2452 static void reset_xps_maps(struct net_device *dev, 2453 struct xps_dev_maps *dev_maps, 2454 enum xps_map_type type) 2455 { 2456 static_key_slow_dec_cpuslocked(&xps_needed); 2457 if (type == XPS_RXQS) 2458 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2459 2460 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2461 2462 kfree_rcu(dev_maps, rcu); 2463 } 2464 2465 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2466 u16 offset, u16 count) 2467 { 2468 struct xps_dev_maps *dev_maps; 2469 bool active = false; 2470 int i, j; 2471 2472 dev_maps = xmap_dereference(dev->xps_maps[type]); 2473 if (!dev_maps) 2474 return; 2475 2476 for (j = 0; j < dev_maps->nr_ids; j++) 2477 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2478 if (!active) 2479 reset_xps_maps(dev, dev_maps, type); 2480 2481 if (type == XPS_CPUS) { 2482 for (i = offset + (count - 1); count--; i--) 2483 netdev_queue_numa_node_write( 2484 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2485 } 2486 } 2487 2488 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2489 u16 count) 2490 { 2491 if (!static_key_false(&xps_needed)) 2492 return; 2493 2494 cpus_read_lock(); 2495 mutex_lock(&xps_map_mutex); 2496 2497 if (static_key_false(&xps_rxqs_needed)) 2498 clean_xps_maps(dev, XPS_RXQS, offset, count); 2499 2500 clean_xps_maps(dev, XPS_CPUS, offset, count); 2501 2502 mutex_unlock(&xps_map_mutex); 2503 cpus_read_unlock(); 2504 } 2505 2506 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2507 { 2508 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2509 } 2510 2511 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2512 u16 index, bool is_rxqs_map) 2513 { 2514 struct xps_map *new_map; 2515 int alloc_len = XPS_MIN_MAP_ALLOC; 2516 int i, pos; 2517 2518 for (pos = 0; map && pos < map->len; pos++) { 2519 if (map->queues[pos] != index) 2520 continue; 2521 return map; 2522 } 2523 2524 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2525 if (map) { 2526 if (pos < map->alloc_len) 2527 return map; 2528 2529 alloc_len = map->alloc_len * 2; 2530 } 2531 2532 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2533 * map 2534 */ 2535 if (is_rxqs_map) 2536 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2537 else 2538 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2539 cpu_to_node(attr_index)); 2540 if (!new_map) 2541 return NULL; 2542 2543 for (i = 0; i < pos; i++) 2544 new_map->queues[i] = map->queues[i]; 2545 new_map->alloc_len = alloc_len; 2546 new_map->len = pos; 2547 2548 return new_map; 2549 } 2550 2551 /* Copy xps maps at a given index */ 2552 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2553 struct xps_dev_maps *new_dev_maps, int index, 2554 int tc, bool skip_tc) 2555 { 2556 int i, tci = index * dev_maps->num_tc; 2557 struct xps_map *map; 2558 2559 /* copy maps belonging to foreign traffic classes */ 2560 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2561 if (i == tc && skip_tc) 2562 continue; 2563 2564 /* fill in the new device map from the old device map */ 2565 map = xmap_dereference(dev_maps->attr_map[tci]); 2566 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2567 } 2568 } 2569 2570 /* Must be called under cpus_read_lock */ 2571 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2572 u16 index, enum xps_map_type type) 2573 { 2574 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2575 const unsigned long *online_mask = NULL; 2576 bool active = false, copy = false; 2577 int i, j, tci, numa_node_id = -2; 2578 int maps_sz, num_tc = 1, tc = 0; 2579 struct xps_map *map, *new_map; 2580 unsigned int nr_ids; 2581 2582 WARN_ON_ONCE(index >= dev->num_tx_queues); 2583 2584 if (dev->num_tc) { 2585 /* Do not allow XPS on subordinate device directly */ 2586 num_tc = dev->num_tc; 2587 if (num_tc < 0) 2588 return -EINVAL; 2589 2590 /* If queue belongs to subordinate dev use its map */ 2591 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2592 2593 tc = netdev_txq_to_tc(dev, index); 2594 if (tc < 0) 2595 return -EINVAL; 2596 } 2597 2598 mutex_lock(&xps_map_mutex); 2599 2600 dev_maps = xmap_dereference(dev->xps_maps[type]); 2601 if (type == XPS_RXQS) { 2602 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2603 nr_ids = dev->num_rx_queues; 2604 } else { 2605 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2606 if (num_possible_cpus() > 1) 2607 online_mask = cpumask_bits(cpu_online_mask); 2608 nr_ids = nr_cpu_ids; 2609 } 2610 2611 if (maps_sz < L1_CACHE_BYTES) 2612 maps_sz = L1_CACHE_BYTES; 2613 2614 /* The old dev_maps could be larger or smaller than the one we're 2615 * setting up now, as dev->num_tc or nr_ids could have been updated in 2616 * between. We could try to be smart, but let's be safe instead and only 2617 * copy foreign traffic classes if the two map sizes match. 2618 */ 2619 if (dev_maps && 2620 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2621 copy = true; 2622 2623 /* allocate memory for queue storage */ 2624 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2625 j < nr_ids;) { 2626 if (!new_dev_maps) { 2627 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2628 if (!new_dev_maps) { 2629 mutex_unlock(&xps_map_mutex); 2630 return -ENOMEM; 2631 } 2632 2633 new_dev_maps->nr_ids = nr_ids; 2634 new_dev_maps->num_tc = num_tc; 2635 } 2636 2637 tci = j * num_tc + tc; 2638 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2639 2640 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2641 if (!map) 2642 goto error; 2643 2644 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2645 } 2646 2647 if (!new_dev_maps) 2648 goto out_no_new_maps; 2649 2650 if (!dev_maps) { 2651 /* Increment static keys at most once per type */ 2652 static_key_slow_inc_cpuslocked(&xps_needed); 2653 if (type == XPS_RXQS) 2654 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2655 } 2656 2657 for (j = 0; j < nr_ids; j++) { 2658 bool skip_tc = false; 2659 2660 tci = j * num_tc + tc; 2661 if (netif_attr_test_mask(j, mask, nr_ids) && 2662 netif_attr_test_online(j, online_mask, nr_ids)) { 2663 /* add tx-queue to CPU/rx-queue maps */ 2664 int pos = 0; 2665 2666 skip_tc = true; 2667 2668 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2669 while ((pos < map->len) && (map->queues[pos] != index)) 2670 pos++; 2671 2672 if (pos == map->len) 2673 map->queues[map->len++] = index; 2674 #ifdef CONFIG_NUMA 2675 if (type == XPS_CPUS) { 2676 if (numa_node_id == -2) 2677 numa_node_id = cpu_to_node(j); 2678 else if (numa_node_id != cpu_to_node(j)) 2679 numa_node_id = -1; 2680 } 2681 #endif 2682 } 2683 2684 if (copy) 2685 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2686 skip_tc); 2687 } 2688 2689 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2690 2691 /* Cleanup old maps */ 2692 if (!dev_maps) 2693 goto out_no_old_maps; 2694 2695 for (j = 0; j < dev_maps->nr_ids; j++) { 2696 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2697 map = xmap_dereference(dev_maps->attr_map[tci]); 2698 if (!map) 2699 continue; 2700 2701 if (copy) { 2702 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2703 if (map == new_map) 2704 continue; 2705 } 2706 2707 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2708 kfree_rcu(map, rcu); 2709 } 2710 } 2711 2712 old_dev_maps = dev_maps; 2713 2714 out_no_old_maps: 2715 dev_maps = new_dev_maps; 2716 active = true; 2717 2718 out_no_new_maps: 2719 if (type == XPS_CPUS) 2720 /* update Tx queue numa node */ 2721 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2722 (numa_node_id >= 0) ? 2723 numa_node_id : NUMA_NO_NODE); 2724 2725 if (!dev_maps) 2726 goto out_no_maps; 2727 2728 /* removes tx-queue from unused CPUs/rx-queues */ 2729 for (j = 0; j < dev_maps->nr_ids; j++) { 2730 tci = j * dev_maps->num_tc; 2731 2732 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2733 if (i == tc && 2734 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 2735 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 2736 continue; 2737 2738 active |= remove_xps_queue(dev_maps, 2739 copy ? old_dev_maps : NULL, 2740 tci, index); 2741 } 2742 } 2743 2744 if (old_dev_maps) 2745 kfree_rcu(old_dev_maps, rcu); 2746 2747 /* free map if not active */ 2748 if (!active) 2749 reset_xps_maps(dev, dev_maps, type); 2750 2751 out_no_maps: 2752 mutex_unlock(&xps_map_mutex); 2753 2754 return 0; 2755 error: 2756 /* remove any maps that we added */ 2757 for (j = 0; j < nr_ids; j++) { 2758 for (i = num_tc, tci = j * num_tc; i--; tci++) { 2759 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2760 map = copy ? 2761 xmap_dereference(dev_maps->attr_map[tci]) : 2762 NULL; 2763 if (new_map && new_map != map) 2764 kfree(new_map); 2765 } 2766 } 2767 2768 mutex_unlock(&xps_map_mutex); 2769 2770 kfree(new_dev_maps); 2771 return -ENOMEM; 2772 } 2773 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 2774 2775 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 2776 u16 index) 2777 { 2778 int ret; 2779 2780 cpus_read_lock(); 2781 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 2782 cpus_read_unlock(); 2783 2784 return ret; 2785 } 2786 EXPORT_SYMBOL(netif_set_xps_queue); 2787 2788 #endif 2789 static void netdev_unbind_all_sb_channels(struct net_device *dev) 2790 { 2791 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 2792 2793 /* Unbind any subordinate channels */ 2794 while (txq-- != &dev->_tx[0]) { 2795 if (txq->sb_dev) 2796 netdev_unbind_sb_channel(dev, txq->sb_dev); 2797 } 2798 } 2799 2800 void netdev_reset_tc(struct net_device *dev) 2801 { 2802 #ifdef CONFIG_XPS 2803 netif_reset_xps_queues_gt(dev, 0); 2804 #endif 2805 netdev_unbind_all_sb_channels(dev); 2806 2807 /* Reset TC configuration of device */ 2808 dev->num_tc = 0; 2809 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 2810 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 2811 } 2812 EXPORT_SYMBOL(netdev_reset_tc); 2813 2814 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 2815 { 2816 if (tc >= dev->num_tc) 2817 return -EINVAL; 2818 2819 #ifdef CONFIG_XPS 2820 netif_reset_xps_queues(dev, offset, count); 2821 #endif 2822 dev->tc_to_txq[tc].count = count; 2823 dev->tc_to_txq[tc].offset = offset; 2824 return 0; 2825 } 2826 EXPORT_SYMBOL(netdev_set_tc_queue); 2827 2828 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 2829 { 2830 if (num_tc > TC_MAX_QUEUE) 2831 return -EINVAL; 2832 2833 #ifdef CONFIG_XPS 2834 netif_reset_xps_queues_gt(dev, 0); 2835 #endif 2836 netdev_unbind_all_sb_channels(dev); 2837 2838 dev->num_tc = num_tc; 2839 return 0; 2840 } 2841 EXPORT_SYMBOL(netdev_set_num_tc); 2842 2843 void netdev_unbind_sb_channel(struct net_device *dev, 2844 struct net_device *sb_dev) 2845 { 2846 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 2847 2848 #ifdef CONFIG_XPS 2849 netif_reset_xps_queues_gt(sb_dev, 0); 2850 #endif 2851 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 2852 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 2853 2854 while (txq-- != &dev->_tx[0]) { 2855 if (txq->sb_dev == sb_dev) 2856 txq->sb_dev = NULL; 2857 } 2858 } 2859 EXPORT_SYMBOL(netdev_unbind_sb_channel); 2860 2861 int netdev_bind_sb_channel_queue(struct net_device *dev, 2862 struct net_device *sb_dev, 2863 u8 tc, u16 count, u16 offset) 2864 { 2865 /* Make certain the sb_dev and dev are already configured */ 2866 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 2867 return -EINVAL; 2868 2869 /* We cannot hand out queues we don't have */ 2870 if ((offset + count) > dev->real_num_tx_queues) 2871 return -EINVAL; 2872 2873 /* Record the mapping */ 2874 sb_dev->tc_to_txq[tc].count = count; 2875 sb_dev->tc_to_txq[tc].offset = offset; 2876 2877 /* Provide a way for Tx queue to find the tc_to_txq map or 2878 * XPS map for itself. 2879 */ 2880 while (count--) 2881 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 2882 2883 return 0; 2884 } 2885 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 2886 2887 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 2888 { 2889 /* Do not use a multiqueue device to represent a subordinate channel */ 2890 if (netif_is_multiqueue(dev)) 2891 return -ENODEV; 2892 2893 /* We allow channels 1 - 32767 to be used for subordinate channels. 2894 * Channel 0 is meant to be "native" mode and used only to represent 2895 * the main root device. We allow writing 0 to reset the device back 2896 * to normal mode after being used as a subordinate channel. 2897 */ 2898 if (channel > S16_MAX) 2899 return -EINVAL; 2900 2901 dev->num_tc = -channel; 2902 2903 return 0; 2904 } 2905 EXPORT_SYMBOL(netdev_set_sb_channel); 2906 2907 /* 2908 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 2909 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 2910 */ 2911 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 2912 { 2913 bool disabling; 2914 int rc; 2915 2916 disabling = txq < dev->real_num_tx_queues; 2917 2918 if (txq < 1 || txq > dev->num_tx_queues) 2919 return -EINVAL; 2920 2921 if (dev->reg_state == NETREG_REGISTERED || 2922 dev->reg_state == NETREG_UNREGISTERING) { 2923 ASSERT_RTNL(); 2924 2925 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 2926 txq); 2927 if (rc) 2928 return rc; 2929 2930 if (dev->num_tc) 2931 netif_setup_tc(dev, txq); 2932 2933 dev_qdisc_change_real_num_tx(dev, txq); 2934 2935 dev->real_num_tx_queues = txq; 2936 2937 if (disabling) { 2938 synchronize_net(); 2939 qdisc_reset_all_tx_gt(dev, txq); 2940 #ifdef CONFIG_XPS 2941 netif_reset_xps_queues_gt(dev, txq); 2942 #endif 2943 } 2944 } else { 2945 dev->real_num_tx_queues = txq; 2946 } 2947 2948 return 0; 2949 } 2950 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 2951 2952 #ifdef CONFIG_SYSFS 2953 /** 2954 * netif_set_real_num_rx_queues - set actual number of RX queues used 2955 * @dev: Network device 2956 * @rxq: Actual number of RX queues 2957 * 2958 * This must be called either with the rtnl_lock held or before 2959 * registration of the net device. Returns 0 on success, or a 2960 * negative error code. If called before registration, it always 2961 * succeeds. 2962 */ 2963 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 2964 { 2965 int rc; 2966 2967 if (rxq < 1 || rxq > dev->num_rx_queues) 2968 return -EINVAL; 2969 2970 if (dev->reg_state == NETREG_REGISTERED) { 2971 ASSERT_RTNL(); 2972 2973 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 2974 rxq); 2975 if (rc) 2976 return rc; 2977 } 2978 2979 dev->real_num_rx_queues = rxq; 2980 return 0; 2981 } 2982 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 2983 #endif 2984 2985 /** 2986 * netif_set_real_num_queues - set actual number of RX and TX queues used 2987 * @dev: Network device 2988 * @txq: Actual number of TX queues 2989 * @rxq: Actual number of RX queues 2990 * 2991 * Set the real number of both TX and RX queues. 2992 * Does nothing if the number of queues is already correct. 2993 */ 2994 int netif_set_real_num_queues(struct net_device *dev, 2995 unsigned int txq, unsigned int rxq) 2996 { 2997 unsigned int old_rxq = dev->real_num_rx_queues; 2998 int err; 2999 3000 if (txq < 1 || txq > dev->num_tx_queues || 3001 rxq < 1 || rxq > dev->num_rx_queues) 3002 return -EINVAL; 3003 3004 /* Start from increases, so the error path only does decreases - 3005 * decreases can't fail. 3006 */ 3007 if (rxq > dev->real_num_rx_queues) { 3008 err = netif_set_real_num_rx_queues(dev, rxq); 3009 if (err) 3010 return err; 3011 } 3012 if (txq > dev->real_num_tx_queues) { 3013 err = netif_set_real_num_tx_queues(dev, txq); 3014 if (err) 3015 goto undo_rx; 3016 } 3017 if (rxq < dev->real_num_rx_queues) 3018 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3019 if (txq < dev->real_num_tx_queues) 3020 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3021 3022 return 0; 3023 undo_rx: 3024 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3025 return err; 3026 } 3027 EXPORT_SYMBOL(netif_set_real_num_queues); 3028 3029 /** 3030 * netif_set_tso_max_size() - set the max size of TSO frames supported 3031 * @dev: netdev to update 3032 * @size: max skb->len of a TSO frame 3033 * 3034 * Set the limit on the size of TSO super-frames the device can handle. 3035 * Unless explicitly set the stack will assume the value of 3036 * %GSO_LEGACY_MAX_SIZE. 3037 */ 3038 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3039 { 3040 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3041 if (size < READ_ONCE(dev->gso_max_size)) 3042 netif_set_gso_max_size(dev, size); 3043 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3044 netif_set_gso_ipv4_max_size(dev, size); 3045 } 3046 EXPORT_SYMBOL(netif_set_tso_max_size); 3047 3048 /** 3049 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3050 * @dev: netdev to update 3051 * @segs: max number of TCP segments 3052 * 3053 * Set the limit on the number of TCP segments the device can generate from 3054 * a single TSO super-frame. 3055 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3056 */ 3057 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3058 { 3059 dev->tso_max_segs = segs; 3060 if (segs < READ_ONCE(dev->gso_max_segs)) 3061 netif_set_gso_max_segs(dev, segs); 3062 } 3063 EXPORT_SYMBOL(netif_set_tso_max_segs); 3064 3065 /** 3066 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3067 * @to: netdev to update 3068 * @from: netdev from which to copy the limits 3069 */ 3070 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3071 { 3072 netif_set_tso_max_size(to, from->tso_max_size); 3073 netif_set_tso_max_segs(to, from->tso_max_segs); 3074 } 3075 EXPORT_SYMBOL(netif_inherit_tso_max); 3076 3077 /** 3078 * netif_get_num_default_rss_queues - default number of RSS queues 3079 * 3080 * Default value is the number of physical cores if there are only 1 or 2, or 3081 * divided by 2 if there are more. 3082 */ 3083 int netif_get_num_default_rss_queues(void) 3084 { 3085 cpumask_var_t cpus; 3086 int cpu, count = 0; 3087 3088 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3089 return 1; 3090 3091 cpumask_copy(cpus, cpu_online_mask); 3092 for_each_cpu(cpu, cpus) { 3093 ++count; 3094 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3095 } 3096 free_cpumask_var(cpus); 3097 3098 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3099 } 3100 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3101 3102 static void __netif_reschedule(struct Qdisc *q) 3103 { 3104 struct softnet_data *sd; 3105 unsigned long flags; 3106 3107 local_irq_save(flags); 3108 sd = this_cpu_ptr(&softnet_data); 3109 q->next_sched = NULL; 3110 *sd->output_queue_tailp = q; 3111 sd->output_queue_tailp = &q->next_sched; 3112 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3113 local_irq_restore(flags); 3114 } 3115 3116 void __netif_schedule(struct Qdisc *q) 3117 { 3118 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3119 __netif_reschedule(q); 3120 } 3121 EXPORT_SYMBOL(__netif_schedule); 3122 3123 struct dev_kfree_skb_cb { 3124 enum skb_drop_reason reason; 3125 }; 3126 3127 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3128 { 3129 return (struct dev_kfree_skb_cb *)skb->cb; 3130 } 3131 3132 void netif_schedule_queue(struct netdev_queue *txq) 3133 { 3134 rcu_read_lock(); 3135 if (!netif_xmit_stopped(txq)) { 3136 struct Qdisc *q = rcu_dereference(txq->qdisc); 3137 3138 __netif_schedule(q); 3139 } 3140 rcu_read_unlock(); 3141 } 3142 EXPORT_SYMBOL(netif_schedule_queue); 3143 3144 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3145 { 3146 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3147 struct Qdisc *q; 3148 3149 rcu_read_lock(); 3150 q = rcu_dereference(dev_queue->qdisc); 3151 __netif_schedule(q); 3152 rcu_read_unlock(); 3153 } 3154 } 3155 EXPORT_SYMBOL(netif_tx_wake_queue); 3156 3157 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3158 { 3159 unsigned long flags; 3160 3161 if (unlikely(!skb)) 3162 return; 3163 3164 if (likely(refcount_read(&skb->users) == 1)) { 3165 smp_rmb(); 3166 refcount_set(&skb->users, 0); 3167 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3168 return; 3169 } 3170 get_kfree_skb_cb(skb)->reason = reason; 3171 local_irq_save(flags); 3172 skb->next = __this_cpu_read(softnet_data.completion_queue); 3173 __this_cpu_write(softnet_data.completion_queue, skb); 3174 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3175 local_irq_restore(flags); 3176 } 3177 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3178 3179 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3180 { 3181 if (in_hardirq() || irqs_disabled()) 3182 dev_kfree_skb_irq_reason(skb, reason); 3183 else 3184 kfree_skb_reason(skb, reason); 3185 } 3186 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3187 3188 3189 /** 3190 * netif_device_detach - mark device as removed 3191 * @dev: network device 3192 * 3193 * Mark device as removed from system and therefore no longer available. 3194 */ 3195 void netif_device_detach(struct net_device *dev) 3196 { 3197 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3198 netif_running(dev)) { 3199 netif_tx_stop_all_queues(dev); 3200 } 3201 } 3202 EXPORT_SYMBOL(netif_device_detach); 3203 3204 /** 3205 * netif_device_attach - mark device as attached 3206 * @dev: network device 3207 * 3208 * Mark device as attached from system and restart if needed. 3209 */ 3210 void netif_device_attach(struct net_device *dev) 3211 { 3212 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3213 netif_running(dev)) { 3214 netif_tx_wake_all_queues(dev); 3215 __netdev_watchdog_up(dev); 3216 } 3217 } 3218 EXPORT_SYMBOL(netif_device_attach); 3219 3220 /* 3221 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3222 * to be used as a distribution range. 3223 */ 3224 static u16 skb_tx_hash(const struct net_device *dev, 3225 const struct net_device *sb_dev, 3226 struct sk_buff *skb) 3227 { 3228 u32 hash; 3229 u16 qoffset = 0; 3230 u16 qcount = dev->real_num_tx_queues; 3231 3232 if (dev->num_tc) { 3233 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3234 3235 qoffset = sb_dev->tc_to_txq[tc].offset; 3236 qcount = sb_dev->tc_to_txq[tc].count; 3237 if (unlikely(!qcount)) { 3238 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3239 sb_dev->name, qoffset, tc); 3240 qoffset = 0; 3241 qcount = dev->real_num_tx_queues; 3242 } 3243 } 3244 3245 if (skb_rx_queue_recorded(skb)) { 3246 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3247 hash = skb_get_rx_queue(skb); 3248 if (hash >= qoffset) 3249 hash -= qoffset; 3250 while (unlikely(hash >= qcount)) 3251 hash -= qcount; 3252 return hash + qoffset; 3253 } 3254 3255 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3256 } 3257 3258 void skb_warn_bad_offload(const struct sk_buff *skb) 3259 { 3260 static const netdev_features_t null_features; 3261 struct net_device *dev = skb->dev; 3262 const char *name = ""; 3263 3264 if (!net_ratelimit()) 3265 return; 3266 3267 if (dev) { 3268 if (dev->dev.parent) 3269 name = dev_driver_string(dev->dev.parent); 3270 else 3271 name = netdev_name(dev); 3272 } 3273 skb_dump(KERN_WARNING, skb, false); 3274 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3275 name, dev ? &dev->features : &null_features, 3276 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3277 } 3278 3279 /* 3280 * Invalidate hardware checksum when packet is to be mangled, and 3281 * complete checksum manually on outgoing path. 3282 */ 3283 int skb_checksum_help(struct sk_buff *skb) 3284 { 3285 __wsum csum; 3286 int ret = 0, offset; 3287 3288 if (skb->ip_summed == CHECKSUM_COMPLETE) 3289 goto out_set_summed; 3290 3291 if (unlikely(skb_is_gso(skb))) { 3292 skb_warn_bad_offload(skb); 3293 return -EINVAL; 3294 } 3295 3296 /* Before computing a checksum, we should make sure no frag could 3297 * be modified by an external entity : checksum could be wrong. 3298 */ 3299 if (skb_has_shared_frag(skb)) { 3300 ret = __skb_linearize(skb); 3301 if (ret) 3302 goto out; 3303 } 3304 3305 offset = skb_checksum_start_offset(skb); 3306 ret = -EINVAL; 3307 if (unlikely(offset >= skb_headlen(skb))) { 3308 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3309 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3310 offset, skb_headlen(skb)); 3311 goto out; 3312 } 3313 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3314 3315 offset += skb->csum_offset; 3316 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3317 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3318 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3319 offset + sizeof(__sum16), skb_headlen(skb)); 3320 goto out; 3321 } 3322 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3323 if (ret) 3324 goto out; 3325 3326 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3327 out_set_summed: 3328 skb->ip_summed = CHECKSUM_NONE; 3329 out: 3330 return ret; 3331 } 3332 EXPORT_SYMBOL(skb_checksum_help); 3333 3334 int skb_crc32c_csum_help(struct sk_buff *skb) 3335 { 3336 __le32 crc32c_csum; 3337 int ret = 0, offset, start; 3338 3339 if (skb->ip_summed != CHECKSUM_PARTIAL) 3340 goto out; 3341 3342 if (unlikely(skb_is_gso(skb))) 3343 goto out; 3344 3345 /* Before computing a checksum, we should make sure no frag could 3346 * be modified by an external entity : checksum could be wrong. 3347 */ 3348 if (unlikely(skb_has_shared_frag(skb))) { 3349 ret = __skb_linearize(skb); 3350 if (ret) 3351 goto out; 3352 } 3353 start = skb_checksum_start_offset(skb); 3354 offset = start + offsetof(struct sctphdr, checksum); 3355 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3356 ret = -EINVAL; 3357 goto out; 3358 } 3359 3360 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3361 if (ret) 3362 goto out; 3363 3364 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start, 3365 skb->len - start, ~(__u32)0, 3366 crc32c_csum_stub)); 3367 *(__le32 *)(skb->data + offset) = crc32c_csum; 3368 skb_reset_csum_not_inet(skb); 3369 out: 3370 return ret; 3371 } 3372 3373 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3374 { 3375 __be16 type = skb->protocol; 3376 3377 /* Tunnel gso handlers can set protocol to ethernet. */ 3378 if (type == htons(ETH_P_TEB)) { 3379 struct ethhdr *eth; 3380 3381 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3382 return 0; 3383 3384 eth = (struct ethhdr *)skb->data; 3385 type = eth->h_proto; 3386 } 3387 3388 return vlan_get_protocol_and_depth(skb, type, depth); 3389 } 3390 3391 3392 /* Take action when hardware reception checksum errors are detected. */ 3393 #ifdef CONFIG_BUG 3394 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3395 { 3396 netdev_err(dev, "hw csum failure\n"); 3397 skb_dump(KERN_ERR, skb, true); 3398 dump_stack(); 3399 } 3400 3401 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3402 { 3403 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3404 } 3405 EXPORT_SYMBOL(netdev_rx_csum_fault); 3406 #endif 3407 3408 /* XXX: check that highmem exists at all on the given machine. */ 3409 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3410 { 3411 #ifdef CONFIG_HIGHMEM 3412 int i; 3413 3414 if (!(dev->features & NETIF_F_HIGHDMA)) { 3415 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3416 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3417 3418 if (PageHighMem(skb_frag_page(frag))) 3419 return 1; 3420 } 3421 } 3422 #endif 3423 return 0; 3424 } 3425 3426 /* If MPLS offload request, verify we are testing hardware MPLS features 3427 * instead of standard features for the netdev. 3428 */ 3429 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3430 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3431 netdev_features_t features, 3432 __be16 type) 3433 { 3434 if (eth_p_mpls(type)) 3435 features &= skb->dev->mpls_features; 3436 3437 return features; 3438 } 3439 #else 3440 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3441 netdev_features_t features, 3442 __be16 type) 3443 { 3444 return features; 3445 } 3446 #endif 3447 3448 static netdev_features_t harmonize_features(struct sk_buff *skb, 3449 netdev_features_t features) 3450 { 3451 __be16 type; 3452 3453 type = skb_network_protocol(skb, NULL); 3454 features = net_mpls_features(skb, features, type); 3455 3456 if (skb->ip_summed != CHECKSUM_NONE && 3457 !can_checksum_protocol(features, type)) { 3458 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3459 } 3460 if (illegal_highdma(skb->dev, skb)) 3461 features &= ~NETIF_F_SG; 3462 3463 return features; 3464 } 3465 3466 netdev_features_t passthru_features_check(struct sk_buff *skb, 3467 struct net_device *dev, 3468 netdev_features_t features) 3469 { 3470 return features; 3471 } 3472 EXPORT_SYMBOL(passthru_features_check); 3473 3474 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3475 struct net_device *dev, 3476 netdev_features_t features) 3477 { 3478 return vlan_features_check(skb, features); 3479 } 3480 3481 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3482 struct net_device *dev, 3483 netdev_features_t features) 3484 { 3485 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3486 3487 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3488 return features & ~NETIF_F_GSO_MASK; 3489 3490 if (unlikely(skb->len >= READ_ONCE(dev->gso_max_size))) 3491 return features & ~NETIF_F_GSO_MASK; 3492 3493 if (!skb_shinfo(skb)->gso_type) { 3494 skb_warn_bad_offload(skb); 3495 return features & ~NETIF_F_GSO_MASK; 3496 } 3497 3498 /* Support for GSO partial features requires software 3499 * intervention before we can actually process the packets 3500 * so we need to strip support for any partial features now 3501 * and we can pull them back in after we have partially 3502 * segmented the frame. 3503 */ 3504 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3505 features &= ~dev->gso_partial_features; 3506 3507 /* Make sure to clear the IPv4 ID mangling feature if the 3508 * IPv4 header has the potential to be fragmented. 3509 */ 3510 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3511 struct iphdr *iph = skb->encapsulation ? 3512 inner_ip_hdr(skb) : ip_hdr(skb); 3513 3514 if (!(iph->frag_off & htons(IP_DF))) 3515 features &= ~NETIF_F_TSO_MANGLEID; 3516 } 3517 3518 return features; 3519 } 3520 3521 netdev_features_t netif_skb_features(struct sk_buff *skb) 3522 { 3523 struct net_device *dev = skb->dev; 3524 netdev_features_t features = dev->features; 3525 3526 if (skb_is_gso(skb)) 3527 features = gso_features_check(skb, dev, features); 3528 3529 /* If encapsulation offload request, verify we are testing 3530 * hardware encapsulation features instead of standard 3531 * features for the netdev 3532 */ 3533 if (skb->encapsulation) 3534 features &= dev->hw_enc_features; 3535 3536 if (skb_vlan_tagged(skb)) 3537 features = netdev_intersect_features(features, 3538 dev->vlan_features | 3539 NETIF_F_HW_VLAN_CTAG_TX | 3540 NETIF_F_HW_VLAN_STAG_TX); 3541 3542 if (dev->netdev_ops->ndo_features_check) 3543 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3544 features); 3545 else 3546 features &= dflt_features_check(skb, dev, features); 3547 3548 return harmonize_features(skb, features); 3549 } 3550 EXPORT_SYMBOL(netif_skb_features); 3551 3552 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3553 struct netdev_queue *txq, bool more) 3554 { 3555 unsigned int len; 3556 int rc; 3557 3558 if (dev_nit_active(dev)) 3559 dev_queue_xmit_nit(skb, dev); 3560 3561 len = skb->len; 3562 trace_net_dev_start_xmit(skb, dev); 3563 rc = netdev_start_xmit(skb, dev, txq, more); 3564 trace_net_dev_xmit(skb, rc, dev, len); 3565 3566 return rc; 3567 } 3568 3569 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3570 struct netdev_queue *txq, int *ret) 3571 { 3572 struct sk_buff *skb = first; 3573 int rc = NETDEV_TX_OK; 3574 3575 while (skb) { 3576 struct sk_buff *next = skb->next; 3577 3578 skb_mark_not_on_list(skb); 3579 rc = xmit_one(skb, dev, txq, next != NULL); 3580 if (unlikely(!dev_xmit_complete(rc))) { 3581 skb->next = next; 3582 goto out; 3583 } 3584 3585 skb = next; 3586 if (netif_tx_queue_stopped(txq) && skb) { 3587 rc = NETDEV_TX_BUSY; 3588 break; 3589 } 3590 } 3591 3592 out: 3593 *ret = rc; 3594 return skb; 3595 } 3596 3597 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3598 netdev_features_t features) 3599 { 3600 if (skb_vlan_tag_present(skb) && 3601 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3602 skb = __vlan_hwaccel_push_inside(skb); 3603 return skb; 3604 } 3605 3606 int skb_csum_hwoffload_help(struct sk_buff *skb, 3607 const netdev_features_t features) 3608 { 3609 if (unlikely(skb_csum_is_sctp(skb))) 3610 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3611 skb_crc32c_csum_help(skb); 3612 3613 if (features & NETIF_F_HW_CSUM) 3614 return 0; 3615 3616 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3617 switch (skb->csum_offset) { 3618 case offsetof(struct tcphdr, check): 3619 case offsetof(struct udphdr, check): 3620 return 0; 3621 } 3622 } 3623 3624 return skb_checksum_help(skb); 3625 } 3626 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3627 3628 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3629 { 3630 netdev_features_t features; 3631 3632 features = netif_skb_features(skb); 3633 skb = validate_xmit_vlan(skb, features); 3634 if (unlikely(!skb)) 3635 goto out_null; 3636 3637 skb = sk_validate_xmit_skb(skb, dev); 3638 if (unlikely(!skb)) 3639 goto out_null; 3640 3641 if (netif_needs_gso(skb, features)) { 3642 struct sk_buff *segs; 3643 3644 segs = skb_gso_segment(skb, features); 3645 if (IS_ERR(segs)) { 3646 goto out_kfree_skb; 3647 } else if (segs) { 3648 consume_skb(skb); 3649 skb = segs; 3650 } 3651 } else { 3652 if (skb_needs_linearize(skb, features) && 3653 __skb_linearize(skb)) 3654 goto out_kfree_skb; 3655 3656 /* If packet is not checksummed and device does not 3657 * support checksumming for this protocol, complete 3658 * checksumming here. 3659 */ 3660 if (skb->ip_summed == CHECKSUM_PARTIAL) { 3661 if (skb->encapsulation) 3662 skb_set_inner_transport_header(skb, 3663 skb_checksum_start_offset(skb)); 3664 else 3665 skb_set_transport_header(skb, 3666 skb_checksum_start_offset(skb)); 3667 if (skb_csum_hwoffload_help(skb, features)) 3668 goto out_kfree_skb; 3669 } 3670 } 3671 3672 skb = validate_xmit_xfrm(skb, features, again); 3673 3674 return skb; 3675 3676 out_kfree_skb: 3677 kfree_skb(skb); 3678 out_null: 3679 dev_core_stats_tx_dropped_inc(dev); 3680 return NULL; 3681 } 3682 3683 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 3684 { 3685 struct sk_buff *next, *head = NULL, *tail; 3686 3687 for (; skb != NULL; skb = next) { 3688 next = skb->next; 3689 skb_mark_not_on_list(skb); 3690 3691 /* in case skb wont be segmented, point to itself */ 3692 skb->prev = skb; 3693 3694 skb = validate_xmit_skb(skb, dev, again); 3695 if (!skb) 3696 continue; 3697 3698 if (!head) 3699 head = skb; 3700 else 3701 tail->next = skb; 3702 /* If skb was segmented, skb->prev points to 3703 * the last segment. If not, it still contains skb. 3704 */ 3705 tail = skb->prev; 3706 } 3707 return head; 3708 } 3709 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 3710 3711 static void qdisc_pkt_len_init(struct sk_buff *skb) 3712 { 3713 const struct skb_shared_info *shinfo = skb_shinfo(skb); 3714 3715 qdisc_skb_cb(skb)->pkt_len = skb->len; 3716 3717 /* To get more precise estimation of bytes sent on wire, 3718 * we add to pkt_len the headers size of all segments 3719 */ 3720 if (shinfo->gso_size && skb_transport_header_was_set(skb)) { 3721 u16 gso_segs = shinfo->gso_segs; 3722 unsigned int hdr_len; 3723 3724 /* mac layer + network layer */ 3725 hdr_len = skb_transport_offset(skb); 3726 3727 /* + transport layer */ 3728 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 3729 const struct tcphdr *th; 3730 struct tcphdr _tcphdr; 3731 3732 th = skb_header_pointer(skb, hdr_len, 3733 sizeof(_tcphdr), &_tcphdr); 3734 if (likely(th)) 3735 hdr_len += __tcp_hdrlen(th); 3736 } else { 3737 struct udphdr _udphdr; 3738 3739 if (skb_header_pointer(skb, hdr_len, 3740 sizeof(_udphdr), &_udphdr)) 3741 hdr_len += sizeof(struct udphdr); 3742 } 3743 3744 if (shinfo->gso_type & SKB_GSO_DODGY) 3745 gso_segs = DIV_ROUND_UP(skb->len - hdr_len, 3746 shinfo->gso_size); 3747 3748 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 3749 } 3750 } 3751 3752 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 3753 struct sk_buff **to_free, 3754 struct netdev_queue *txq) 3755 { 3756 int rc; 3757 3758 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 3759 if (rc == NET_XMIT_SUCCESS) 3760 trace_qdisc_enqueue(q, txq, skb); 3761 return rc; 3762 } 3763 3764 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 3765 struct net_device *dev, 3766 struct netdev_queue *txq) 3767 { 3768 spinlock_t *root_lock = qdisc_lock(q); 3769 struct sk_buff *to_free = NULL; 3770 bool contended; 3771 int rc; 3772 3773 qdisc_calculate_pkt_len(skb, q); 3774 3775 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 3776 3777 if (q->flags & TCQ_F_NOLOCK) { 3778 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 3779 qdisc_run_begin(q)) { 3780 /* Retest nolock_qdisc_is_empty() within the protection 3781 * of q->seqlock to protect from racing with requeuing. 3782 */ 3783 if (unlikely(!nolock_qdisc_is_empty(q))) { 3784 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3785 __qdisc_run(q); 3786 qdisc_run_end(q); 3787 3788 goto no_lock_out; 3789 } 3790 3791 qdisc_bstats_cpu_update(q, skb); 3792 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 3793 !nolock_qdisc_is_empty(q)) 3794 __qdisc_run(q); 3795 3796 qdisc_run_end(q); 3797 return NET_XMIT_SUCCESS; 3798 } 3799 3800 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3801 qdisc_run(q); 3802 3803 no_lock_out: 3804 if (unlikely(to_free)) 3805 kfree_skb_list_reason(to_free, 3806 tcf_get_drop_reason(to_free)); 3807 return rc; 3808 } 3809 3810 /* 3811 * Heuristic to force contended enqueues to serialize on a 3812 * separate lock before trying to get qdisc main lock. 3813 * This permits qdisc->running owner to get the lock more 3814 * often and dequeue packets faster. 3815 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 3816 * and then other tasks will only enqueue packets. The packets will be 3817 * sent after the qdisc owner is scheduled again. To prevent this 3818 * scenario the task always serialize on the lock. 3819 */ 3820 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 3821 if (unlikely(contended)) 3822 spin_lock(&q->busylock); 3823 3824 spin_lock(root_lock); 3825 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 3826 __qdisc_drop(skb, &to_free); 3827 rc = NET_XMIT_DROP; 3828 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 3829 qdisc_run_begin(q)) { 3830 /* 3831 * This is a work-conserving queue; there are no old skbs 3832 * waiting to be sent out; and the qdisc is not running - 3833 * xmit the skb directly. 3834 */ 3835 3836 qdisc_bstats_update(q, skb); 3837 3838 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 3839 if (unlikely(contended)) { 3840 spin_unlock(&q->busylock); 3841 contended = false; 3842 } 3843 __qdisc_run(q); 3844 } 3845 3846 qdisc_run_end(q); 3847 rc = NET_XMIT_SUCCESS; 3848 } else { 3849 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3850 if (qdisc_run_begin(q)) { 3851 if (unlikely(contended)) { 3852 spin_unlock(&q->busylock); 3853 contended = false; 3854 } 3855 __qdisc_run(q); 3856 qdisc_run_end(q); 3857 } 3858 } 3859 spin_unlock(root_lock); 3860 if (unlikely(to_free)) 3861 kfree_skb_list_reason(to_free, 3862 tcf_get_drop_reason(to_free)); 3863 if (unlikely(contended)) 3864 spin_unlock(&q->busylock); 3865 return rc; 3866 } 3867 3868 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 3869 static void skb_update_prio(struct sk_buff *skb) 3870 { 3871 const struct netprio_map *map; 3872 const struct sock *sk; 3873 unsigned int prioidx; 3874 3875 if (skb->priority) 3876 return; 3877 map = rcu_dereference_bh(skb->dev->priomap); 3878 if (!map) 3879 return; 3880 sk = skb_to_full_sk(skb); 3881 if (!sk) 3882 return; 3883 3884 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 3885 3886 if (prioidx < map->priomap_len) 3887 skb->priority = map->priomap[prioidx]; 3888 } 3889 #else 3890 #define skb_update_prio(skb) 3891 #endif 3892 3893 /** 3894 * dev_loopback_xmit - loop back @skb 3895 * @net: network namespace this loopback is happening in 3896 * @sk: sk needed to be a netfilter okfn 3897 * @skb: buffer to transmit 3898 */ 3899 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 3900 { 3901 skb_reset_mac_header(skb); 3902 __skb_pull(skb, skb_network_offset(skb)); 3903 skb->pkt_type = PACKET_LOOPBACK; 3904 if (skb->ip_summed == CHECKSUM_NONE) 3905 skb->ip_summed = CHECKSUM_UNNECESSARY; 3906 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 3907 skb_dst_force(skb); 3908 netif_rx(skb); 3909 return 0; 3910 } 3911 EXPORT_SYMBOL(dev_loopback_xmit); 3912 3913 #ifdef CONFIG_NET_EGRESS 3914 static struct netdev_queue * 3915 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 3916 { 3917 int qm = skb_get_queue_mapping(skb); 3918 3919 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 3920 } 3921 3922 static bool netdev_xmit_txqueue_skipped(void) 3923 { 3924 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 3925 } 3926 3927 void netdev_xmit_skip_txqueue(bool skip) 3928 { 3929 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 3930 } 3931 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 3932 #endif /* CONFIG_NET_EGRESS */ 3933 3934 #ifdef CONFIG_NET_XGRESS 3935 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 3936 enum skb_drop_reason *drop_reason) 3937 { 3938 int ret = TC_ACT_UNSPEC; 3939 #ifdef CONFIG_NET_CLS_ACT 3940 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 3941 struct tcf_result res; 3942 3943 if (!miniq) 3944 return ret; 3945 3946 if (static_branch_unlikely(&tcf_bypass_check_needed_key)) { 3947 if (tcf_block_bypass_sw(miniq->block)) 3948 return ret; 3949 } 3950 3951 tc_skb_cb(skb)->mru = 0; 3952 tc_skb_cb(skb)->post_ct = false; 3953 tcf_set_drop_reason(skb, *drop_reason); 3954 3955 mini_qdisc_bstats_cpu_update(miniq, skb); 3956 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 3957 /* Only tcf related quirks below. */ 3958 switch (ret) { 3959 case TC_ACT_SHOT: 3960 *drop_reason = tcf_get_drop_reason(skb); 3961 mini_qdisc_qstats_cpu_drop(miniq); 3962 break; 3963 case TC_ACT_OK: 3964 case TC_ACT_RECLASSIFY: 3965 skb->tc_index = TC_H_MIN(res.classid); 3966 break; 3967 } 3968 #endif /* CONFIG_NET_CLS_ACT */ 3969 return ret; 3970 } 3971 3972 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 3973 3974 void tcx_inc(void) 3975 { 3976 static_branch_inc(&tcx_needed_key); 3977 } 3978 3979 void tcx_dec(void) 3980 { 3981 static_branch_dec(&tcx_needed_key); 3982 } 3983 3984 static __always_inline enum tcx_action_base 3985 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 3986 const bool needs_mac) 3987 { 3988 const struct bpf_mprog_fp *fp; 3989 const struct bpf_prog *prog; 3990 int ret = TCX_NEXT; 3991 3992 if (needs_mac) 3993 __skb_push(skb, skb->mac_len); 3994 bpf_mprog_foreach_prog(entry, fp, prog) { 3995 bpf_compute_data_pointers(skb); 3996 ret = bpf_prog_run(prog, skb); 3997 if (ret != TCX_NEXT) 3998 break; 3999 } 4000 if (needs_mac) 4001 __skb_pull(skb, skb->mac_len); 4002 return tcx_action_code(skb, ret); 4003 } 4004 4005 static __always_inline struct sk_buff * 4006 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4007 struct net_device *orig_dev, bool *another) 4008 { 4009 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4010 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4011 int sch_ret; 4012 4013 if (!entry) 4014 return skb; 4015 if (*pt_prev) { 4016 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4017 *pt_prev = NULL; 4018 } 4019 4020 qdisc_skb_cb(skb)->pkt_len = skb->len; 4021 tcx_set_ingress(skb, true); 4022 4023 if (static_branch_unlikely(&tcx_needed_key)) { 4024 sch_ret = tcx_run(entry, skb, true); 4025 if (sch_ret != TC_ACT_UNSPEC) 4026 goto ingress_verdict; 4027 } 4028 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4029 ingress_verdict: 4030 switch (sch_ret) { 4031 case TC_ACT_REDIRECT: 4032 /* skb_mac_header check was done by BPF, so we can safely 4033 * push the L2 header back before redirecting to another 4034 * netdev. 4035 */ 4036 __skb_push(skb, skb->mac_len); 4037 if (skb_do_redirect(skb) == -EAGAIN) { 4038 __skb_pull(skb, skb->mac_len); 4039 *another = true; 4040 break; 4041 } 4042 *ret = NET_RX_SUCCESS; 4043 return NULL; 4044 case TC_ACT_SHOT: 4045 kfree_skb_reason(skb, drop_reason); 4046 *ret = NET_RX_DROP; 4047 return NULL; 4048 /* used by tc_run */ 4049 case TC_ACT_STOLEN: 4050 case TC_ACT_QUEUED: 4051 case TC_ACT_TRAP: 4052 consume_skb(skb); 4053 fallthrough; 4054 case TC_ACT_CONSUMED: 4055 *ret = NET_RX_SUCCESS; 4056 return NULL; 4057 } 4058 4059 return skb; 4060 } 4061 4062 static __always_inline struct sk_buff * 4063 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4064 { 4065 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4066 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4067 int sch_ret; 4068 4069 if (!entry) 4070 return skb; 4071 4072 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4073 * already set by the caller. 4074 */ 4075 if (static_branch_unlikely(&tcx_needed_key)) { 4076 sch_ret = tcx_run(entry, skb, false); 4077 if (sch_ret != TC_ACT_UNSPEC) 4078 goto egress_verdict; 4079 } 4080 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4081 egress_verdict: 4082 switch (sch_ret) { 4083 case TC_ACT_REDIRECT: 4084 /* No need to push/pop skb's mac_header here on egress! */ 4085 skb_do_redirect(skb); 4086 *ret = NET_XMIT_SUCCESS; 4087 return NULL; 4088 case TC_ACT_SHOT: 4089 kfree_skb_reason(skb, drop_reason); 4090 *ret = NET_XMIT_DROP; 4091 return NULL; 4092 /* used by tc_run */ 4093 case TC_ACT_STOLEN: 4094 case TC_ACT_QUEUED: 4095 case TC_ACT_TRAP: 4096 consume_skb(skb); 4097 fallthrough; 4098 case TC_ACT_CONSUMED: 4099 *ret = NET_XMIT_SUCCESS; 4100 return NULL; 4101 } 4102 4103 return skb; 4104 } 4105 #else 4106 static __always_inline struct sk_buff * 4107 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4108 struct net_device *orig_dev, bool *another) 4109 { 4110 return skb; 4111 } 4112 4113 static __always_inline struct sk_buff * 4114 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4115 { 4116 return skb; 4117 } 4118 #endif /* CONFIG_NET_XGRESS */ 4119 4120 #ifdef CONFIG_XPS 4121 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4122 struct xps_dev_maps *dev_maps, unsigned int tci) 4123 { 4124 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4125 struct xps_map *map; 4126 int queue_index = -1; 4127 4128 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4129 return queue_index; 4130 4131 tci *= dev_maps->num_tc; 4132 tci += tc; 4133 4134 map = rcu_dereference(dev_maps->attr_map[tci]); 4135 if (map) { 4136 if (map->len == 1) 4137 queue_index = map->queues[0]; 4138 else 4139 queue_index = map->queues[reciprocal_scale( 4140 skb_get_hash(skb), map->len)]; 4141 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4142 queue_index = -1; 4143 } 4144 return queue_index; 4145 } 4146 #endif 4147 4148 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4149 struct sk_buff *skb) 4150 { 4151 #ifdef CONFIG_XPS 4152 struct xps_dev_maps *dev_maps; 4153 struct sock *sk = skb->sk; 4154 int queue_index = -1; 4155 4156 if (!static_key_false(&xps_needed)) 4157 return -1; 4158 4159 rcu_read_lock(); 4160 if (!static_key_false(&xps_rxqs_needed)) 4161 goto get_cpus_map; 4162 4163 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4164 if (dev_maps) { 4165 int tci = sk_rx_queue_get(sk); 4166 4167 if (tci >= 0) 4168 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4169 tci); 4170 } 4171 4172 get_cpus_map: 4173 if (queue_index < 0) { 4174 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4175 if (dev_maps) { 4176 unsigned int tci = skb->sender_cpu - 1; 4177 4178 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4179 tci); 4180 } 4181 } 4182 rcu_read_unlock(); 4183 4184 return queue_index; 4185 #else 4186 return -1; 4187 #endif 4188 } 4189 4190 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4191 struct net_device *sb_dev) 4192 { 4193 return 0; 4194 } 4195 EXPORT_SYMBOL(dev_pick_tx_zero); 4196 4197 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb, 4198 struct net_device *sb_dev) 4199 { 4200 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues; 4201 } 4202 EXPORT_SYMBOL(dev_pick_tx_cpu_id); 4203 4204 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4205 struct net_device *sb_dev) 4206 { 4207 struct sock *sk = skb->sk; 4208 int queue_index = sk_tx_queue_get(sk); 4209 4210 sb_dev = sb_dev ? : dev; 4211 4212 if (queue_index < 0 || skb->ooo_okay || 4213 queue_index >= dev->real_num_tx_queues) { 4214 int new_index = get_xps_queue(dev, sb_dev, skb); 4215 4216 if (new_index < 0) 4217 new_index = skb_tx_hash(dev, sb_dev, skb); 4218 4219 if (queue_index != new_index && sk && 4220 sk_fullsock(sk) && 4221 rcu_access_pointer(sk->sk_dst_cache)) 4222 sk_tx_queue_set(sk, new_index); 4223 4224 queue_index = new_index; 4225 } 4226 4227 return queue_index; 4228 } 4229 EXPORT_SYMBOL(netdev_pick_tx); 4230 4231 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4232 struct sk_buff *skb, 4233 struct net_device *sb_dev) 4234 { 4235 int queue_index = 0; 4236 4237 #ifdef CONFIG_XPS 4238 u32 sender_cpu = skb->sender_cpu - 1; 4239 4240 if (sender_cpu >= (u32)NR_CPUS) 4241 skb->sender_cpu = raw_smp_processor_id() + 1; 4242 #endif 4243 4244 if (dev->real_num_tx_queues != 1) { 4245 const struct net_device_ops *ops = dev->netdev_ops; 4246 4247 if (ops->ndo_select_queue) 4248 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4249 else 4250 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4251 4252 queue_index = netdev_cap_txqueue(dev, queue_index); 4253 } 4254 4255 skb_set_queue_mapping(skb, queue_index); 4256 return netdev_get_tx_queue(dev, queue_index); 4257 } 4258 4259 /** 4260 * __dev_queue_xmit() - transmit a buffer 4261 * @skb: buffer to transmit 4262 * @sb_dev: suboordinate device used for L2 forwarding offload 4263 * 4264 * Queue a buffer for transmission to a network device. The caller must 4265 * have set the device and priority and built the buffer before calling 4266 * this function. The function can be called from an interrupt. 4267 * 4268 * When calling this method, interrupts MUST be enabled. This is because 4269 * the BH enable code must have IRQs enabled so that it will not deadlock. 4270 * 4271 * Regardless of the return value, the skb is consumed, so it is currently 4272 * difficult to retry a send to this method. (You can bump the ref count 4273 * before sending to hold a reference for retry if you are careful.) 4274 * 4275 * Return: 4276 * * 0 - buffer successfully transmitted 4277 * * positive qdisc return code - NET_XMIT_DROP etc. 4278 * * negative errno - other errors 4279 */ 4280 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4281 { 4282 struct net_device *dev = skb->dev; 4283 struct netdev_queue *txq = NULL; 4284 struct Qdisc *q; 4285 int rc = -ENOMEM; 4286 bool again = false; 4287 4288 skb_reset_mac_header(skb); 4289 skb_assert_len(skb); 4290 4291 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP)) 4292 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4293 4294 /* Disable soft irqs for various locks below. Also 4295 * stops preemption for RCU. 4296 */ 4297 rcu_read_lock_bh(); 4298 4299 skb_update_prio(skb); 4300 4301 qdisc_pkt_len_init(skb); 4302 tcx_set_ingress(skb, false); 4303 #ifdef CONFIG_NET_EGRESS 4304 if (static_branch_unlikely(&egress_needed_key)) { 4305 if (nf_hook_egress_active()) { 4306 skb = nf_hook_egress(skb, &rc, dev); 4307 if (!skb) 4308 goto out; 4309 } 4310 4311 netdev_xmit_skip_txqueue(false); 4312 4313 nf_skip_egress(skb, true); 4314 skb = sch_handle_egress(skb, &rc, dev); 4315 if (!skb) 4316 goto out; 4317 nf_skip_egress(skb, false); 4318 4319 if (netdev_xmit_txqueue_skipped()) 4320 txq = netdev_tx_queue_mapping(dev, skb); 4321 } 4322 #endif 4323 /* If device/qdisc don't need skb->dst, release it right now while 4324 * its hot in this cpu cache. 4325 */ 4326 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4327 skb_dst_drop(skb); 4328 else 4329 skb_dst_force(skb); 4330 4331 if (!txq) 4332 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4333 4334 q = rcu_dereference_bh(txq->qdisc); 4335 4336 trace_net_dev_queue(skb); 4337 if (q->enqueue) { 4338 rc = __dev_xmit_skb(skb, q, dev, txq); 4339 goto out; 4340 } 4341 4342 /* The device has no queue. Common case for software devices: 4343 * loopback, all the sorts of tunnels... 4344 4345 * Really, it is unlikely that netif_tx_lock protection is necessary 4346 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4347 * counters.) 4348 * However, it is possible, that they rely on protection 4349 * made by us here. 4350 4351 * Check this and shot the lock. It is not prone from deadlocks. 4352 *Either shot noqueue qdisc, it is even simpler 8) 4353 */ 4354 if (dev->flags & IFF_UP) { 4355 int cpu = smp_processor_id(); /* ok because BHs are off */ 4356 4357 /* Other cpus might concurrently change txq->xmit_lock_owner 4358 * to -1 or to their cpu id, but not to our id. 4359 */ 4360 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4361 if (dev_xmit_recursion()) 4362 goto recursion_alert; 4363 4364 skb = validate_xmit_skb(skb, dev, &again); 4365 if (!skb) 4366 goto out; 4367 4368 HARD_TX_LOCK(dev, txq, cpu); 4369 4370 if (!netif_xmit_stopped(txq)) { 4371 dev_xmit_recursion_inc(); 4372 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4373 dev_xmit_recursion_dec(); 4374 if (dev_xmit_complete(rc)) { 4375 HARD_TX_UNLOCK(dev, txq); 4376 goto out; 4377 } 4378 } 4379 HARD_TX_UNLOCK(dev, txq); 4380 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4381 dev->name); 4382 } else { 4383 /* Recursion is detected! It is possible, 4384 * unfortunately 4385 */ 4386 recursion_alert: 4387 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4388 dev->name); 4389 } 4390 } 4391 4392 rc = -ENETDOWN; 4393 rcu_read_unlock_bh(); 4394 4395 dev_core_stats_tx_dropped_inc(dev); 4396 kfree_skb_list(skb); 4397 return rc; 4398 out: 4399 rcu_read_unlock_bh(); 4400 return rc; 4401 } 4402 EXPORT_SYMBOL(__dev_queue_xmit); 4403 4404 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4405 { 4406 struct net_device *dev = skb->dev; 4407 struct sk_buff *orig_skb = skb; 4408 struct netdev_queue *txq; 4409 int ret = NETDEV_TX_BUSY; 4410 bool again = false; 4411 4412 if (unlikely(!netif_running(dev) || 4413 !netif_carrier_ok(dev))) 4414 goto drop; 4415 4416 skb = validate_xmit_skb_list(skb, dev, &again); 4417 if (skb != orig_skb) 4418 goto drop; 4419 4420 skb_set_queue_mapping(skb, queue_id); 4421 txq = skb_get_tx_queue(dev, skb); 4422 4423 local_bh_disable(); 4424 4425 dev_xmit_recursion_inc(); 4426 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4427 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4428 ret = netdev_start_xmit(skb, dev, txq, false); 4429 HARD_TX_UNLOCK(dev, txq); 4430 dev_xmit_recursion_dec(); 4431 4432 local_bh_enable(); 4433 return ret; 4434 drop: 4435 dev_core_stats_tx_dropped_inc(dev); 4436 kfree_skb_list(skb); 4437 return NET_XMIT_DROP; 4438 } 4439 EXPORT_SYMBOL(__dev_direct_xmit); 4440 4441 /************************************************************************* 4442 * Receiver routines 4443 *************************************************************************/ 4444 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4445 4446 unsigned int sysctl_skb_defer_max __read_mostly = 64; 4447 int weight_p __read_mostly = 64; /* old backlog weight */ 4448 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4449 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4450 4451 /* Called with irq disabled */ 4452 static inline void ____napi_schedule(struct softnet_data *sd, 4453 struct napi_struct *napi) 4454 { 4455 struct task_struct *thread; 4456 4457 lockdep_assert_irqs_disabled(); 4458 4459 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4460 /* Paired with smp_mb__before_atomic() in 4461 * napi_enable()/dev_set_threaded(). 4462 * Use READ_ONCE() to guarantee a complete 4463 * read on napi->thread. Only call 4464 * wake_up_process() when it's not NULL. 4465 */ 4466 thread = READ_ONCE(napi->thread); 4467 if (thread) { 4468 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4469 goto use_local_napi; 4470 4471 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4472 wake_up_process(thread); 4473 return; 4474 } 4475 } 4476 4477 use_local_napi: 4478 list_add_tail(&napi->poll_list, &sd->poll_list); 4479 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4480 /* If not called from net_rx_action() 4481 * we have to raise NET_RX_SOFTIRQ. 4482 */ 4483 if (!sd->in_net_rx_action) 4484 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4485 } 4486 4487 #ifdef CONFIG_RPS 4488 4489 struct static_key_false rps_needed __read_mostly; 4490 EXPORT_SYMBOL(rps_needed); 4491 struct static_key_false rfs_needed __read_mostly; 4492 EXPORT_SYMBOL(rfs_needed); 4493 4494 static struct rps_dev_flow * 4495 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4496 struct rps_dev_flow *rflow, u16 next_cpu) 4497 { 4498 if (next_cpu < nr_cpu_ids) { 4499 #ifdef CONFIG_RFS_ACCEL 4500 struct netdev_rx_queue *rxqueue; 4501 struct rps_dev_flow_table *flow_table; 4502 struct rps_dev_flow *old_rflow; 4503 u32 flow_id; 4504 u16 rxq_index; 4505 int rc; 4506 4507 /* Should we steer this flow to a different hardware queue? */ 4508 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4509 !(dev->features & NETIF_F_NTUPLE)) 4510 goto out; 4511 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4512 if (rxq_index == skb_get_rx_queue(skb)) 4513 goto out; 4514 4515 rxqueue = dev->_rx + rxq_index; 4516 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4517 if (!flow_table) 4518 goto out; 4519 flow_id = skb_get_hash(skb) & flow_table->mask; 4520 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4521 rxq_index, flow_id); 4522 if (rc < 0) 4523 goto out; 4524 old_rflow = rflow; 4525 rflow = &flow_table->flows[flow_id]; 4526 rflow->filter = rc; 4527 if (old_rflow->filter == rflow->filter) 4528 old_rflow->filter = RPS_NO_FILTER; 4529 out: 4530 #endif 4531 rflow->last_qtail = 4532 READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 4533 } 4534 4535 rflow->cpu = next_cpu; 4536 return rflow; 4537 } 4538 4539 /* 4540 * get_rps_cpu is called from netif_receive_skb and returns the target 4541 * CPU from the RPS map of the receiving queue for a given skb. 4542 * rcu_read_lock must be held on entry. 4543 */ 4544 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4545 struct rps_dev_flow **rflowp) 4546 { 4547 const struct rps_sock_flow_table *sock_flow_table; 4548 struct netdev_rx_queue *rxqueue = dev->_rx; 4549 struct rps_dev_flow_table *flow_table; 4550 struct rps_map *map; 4551 int cpu = -1; 4552 u32 tcpu; 4553 u32 hash; 4554 4555 if (skb_rx_queue_recorded(skb)) { 4556 u16 index = skb_get_rx_queue(skb); 4557 4558 if (unlikely(index >= dev->real_num_rx_queues)) { 4559 WARN_ONCE(dev->real_num_rx_queues > 1, 4560 "%s received packet on queue %u, but number " 4561 "of RX queues is %u\n", 4562 dev->name, index, dev->real_num_rx_queues); 4563 goto done; 4564 } 4565 rxqueue += index; 4566 } 4567 4568 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4569 4570 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4571 map = rcu_dereference(rxqueue->rps_map); 4572 if (!flow_table && !map) 4573 goto done; 4574 4575 skb_reset_network_header(skb); 4576 hash = skb_get_hash(skb); 4577 if (!hash) 4578 goto done; 4579 4580 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4581 if (flow_table && sock_flow_table) { 4582 struct rps_dev_flow *rflow; 4583 u32 next_cpu; 4584 u32 ident; 4585 4586 /* First check into global flow table if there is a match. 4587 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4588 */ 4589 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4590 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4591 goto try_rps; 4592 4593 next_cpu = ident & net_hotdata.rps_cpu_mask; 4594 4595 /* OK, now we know there is a match, 4596 * we can look at the local (per receive queue) flow table 4597 */ 4598 rflow = &flow_table->flows[hash & flow_table->mask]; 4599 tcpu = rflow->cpu; 4600 4601 /* 4602 * If the desired CPU (where last recvmsg was done) is 4603 * different from current CPU (one in the rx-queue flow 4604 * table entry), switch if one of the following holds: 4605 * - Current CPU is unset (>= nr_cpu_ids). 4606 * - Current CPU is offline. 4607 * - The current CPU's queue tail has advanced beyond the 4608 * last packet that was enqueued using this table entry. 4609 * This guarantees that all previous packets for the flow 4610 * have been dequeued, thus preserving in order delivery. 4611 */ 4612 if (unlikely(tcpu != next_cpu) && 4613 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4614 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 4615 READ_ONCE(rflow->last_qtail))) >= 0)) { 4616 tcpu = next_cpu; 4617 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4618 } 4619 4620 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4621 *rflowp = rflow; 4622 cpu = tcpu; 4623 goto done; 4624 } 4625 } 4626 4627 try_rps: 4628 4629 if (map) { 4630 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4631 if (cpu_online(tcpu)) { 4632 cpu = tcpu; 4633 goto done; 4634 } 4635 } 4636 4637 done: 4638 return cpu; 4639 } 4640 4641 #ifdef CONFIG_RFS_ACCEL 4642 4643 /** 4644 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4645 * @dev: Device on which the filter was set 4646 * @rxq_index: RX queue index 4647 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4648 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4649 * 4650 * Drivers that implement ndo_rx_flow_steer() should periodically call 4651 * this function for each installed filter and remove the filters for 4652 * which it returns %true. 4653 */ 4654 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4655 u32 flow_id, u16 filter_id) 4656 { 4657 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4658 struct rps_dev_flow_table *flow_table; 4659 struct rps_dev_flow *rflow; 4660 bool expire = true; 4661 unsigned int cpu; 4662 4663 rcu_read_lock(); 4664 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4665 if (flow_table && flow_id <= flow_table->mask) { 4666 rflow = &flow_table->flows[flow_id]; 4667 cpu = READ_ONCE(rflow->cpu); 4668 if (rflow->filter == filter_id && cpu < nr_cpu_ids && 4669 ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - 4670 READ_ONCE(rflow->last_qtail)) < 4671 (int)(10 * flow_table->mask))) 4672 expire = false; 4673 } 4674 rcu_read_unlock(); 4675 return expire; 4676 } 4677 EXPORT_SYMBOL(rps_may_expire_flow); 4678 4679 #endif /* CONFIG_RFS_ACCEL */ 4680 4681 /* Called from hardirq (IPI) context */ 4682 static void rps_trigger_softirq(void *data) 4683 { 4684 struct softnet_data *sd = data; 4685 4686 ____napi_schedule(sd, &sd->backlog); 4687 sd->received_rps++; 4688 } 4689 4690 #endif /* CONFIG_RPS */ 4691 4692 /* Called from hardirq (IPI) context */ 4693 static void trigger_rx_softirq(void *data) 4694 { 4695 struct softnet_data *sd = data; 4696 4697 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4698 smp_store_release(&sd->defer_ipi_scheduled, 0); 4699 } 4700 4701 /* 4702 * After we queued a packet into sd->input_pkt_queue, 4703 * we need to make sure this queue is serviced soon. 4704 * 4705 * - If this is another cpu queue, link it to our rps_ipi_list, 4706 * and make sure we will process rps_ipi_list from net_rx_action(). 4707 * 4708 * - If this is our own queue, NAPI schedule our backlog. 4709 * Note that this also raises NET_RX_SOFTIRQ. 4710 */ 4711 static void napi_schedule_rps(struct softnet_data *sd) 4712 { 4713 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 4714 4715 #ifdef CONFIG_RPS 4716 if (sd != mysd) { 4717 if (use_backlog_threads()) { 4718 __napi_schedule_irqoff(&sd->backlog); 4719 return; 4720 } 4721 4722 sd->rps_ipi_next = mysd->rps_ipi_list; 4723 mysd->rps_ipi_list = sd; 4724 4725 /* If not called from net_rx_action() or napi_threaded_poll() 4726 * we have to raise NET_RX_SOFTIRQ. 4727 */ 4728 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 4729 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4730 return; 4731 } 4732 #endif /* CONFIG_RPS */ 4733 __napi_schedule_irqoff(&mysd->backlog); 4734 } 4735 4736 void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu) 4737 { 4738 unsigned long flags; 4739 4740 if (use_backlog_threads()) { 4741 backlog_lock_irq_save(sd, &flags); 4742 4743 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 4744 __napi_schedule_irqoff(&sd->backlog); 4745 4746 backlog_unlock_irq_restore(sd, &flags); 4747 4748 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 4749 smp_call_function_single_async(cpu, &sd->defer_csd); 4750 } 4751 } 4752 4753 #ifdef CONFIG_NET_FLOW_LIMIT 4754 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 4755 #endif 4756 4757 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 4758 { 4759 #ifdef CONFIG_NET_FLOW_LIMIT 4760 struct sd_flow_limit *fl; 4761 struct softnet_data *sd; 4762 unsigned int old_flow, new_flow; 4763 4764 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 4765 return false; 4766 4767 sd = this_cpu_ptr(&softnet_data); 4768 4769 rcu_read_lock(); 4770 fl = rcu_dereference(sd->flow_limit); 4771 if (fl) { 4772 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1); 4773 old_flow = fl->history[fl->history_head]; 4774 fl->history[fl->history_head] = new_flow; 4775 4776 fl->history_head++; 4777 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 4778 4779 if (likely(fl->buckets[old_flow])) 4780 fl->buckets[old_flow]--; 4781 4782 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 4783 fl->count++; 4784 rcu_read_unlock(); 4785 return true; 4786 } 4787 } 4788 rcu_read_unlock(); 4789 #endif 4790 return false; 4791 } 4792 4793 /* 4794 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 4795 * queue (may be a remote CPU queue). 4796 */ 4797 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 4798 unsigned int *qtail) 4799 { 4800 enum skb_drop_reason reason; 4801 struct softnet_data *sd; 4802 unsigned long flags; 4803 unsigned int qlen; 4804 int max_backlog; 4805 u32 tail; 4806 4807 reason = SKB_DROP_REASON_DEV_READY; 4808 if (!netif_running(skb->dev)) 4809 goto bad_dev; 4810 4811 reason = SKB_DROP_REASON_CPU_BACKLOG; 4812 sd = &per_cpu(softnet_data, cpu); 4813 4814 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 4815 max_backlog = READ_ONCE(net_hotdata.max_backlog); 4816 if (unlikely(qlen > max_backlog)) 4817 goto cpu_backlog_drop; 4818 backlog_lock_irq_save(sd, &flags); 4819 qlen = skb_queue_len(&sd->input_pkt_queue); 4820 if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) { 4821 if (!qlen) { 4822 /* Schedule NAPI for backlog device. We can use 4823 * non atomic operation as we own the queue lock. 4824 */ 4825 if (!__test_and_set_bit(NAPI_STATE_SCHED, 4826 &sd->backlog.state)) 4827 napi_schedule_rps(sd); 4828 } 4829 __skb_queue_tail(&sd->input_pkt_queue, skb); 4830 tail = rps_input_queue_tail_incr(sd); 4831 backlog_unlock_irq_restore(sd, &flags); 4832 4833 /* save the tail outside of the critical section */ 4834 rps_input_queue_tail_save(qtail, tail); 4835 return NET_RX_SUCCESS; 4836 } 4837 4838 backlog_unlock_irq_restore(sd, &flags); 4839 4840 cpu_backlog_drop: 4841 atomic_inc(&sd->dropped); 4842 bad_dev: 4843 dev_core_stats_rx_dropped_inc(skb->dev); 4844 kfree_skb_reason(skb, reason); 4845 return NET_RX_DROP; 4846 } 4847 4848 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 4849 { 4850 struct net_device *dev = skb->dev; 4851 struct netdev_rx_queue *rxqueue; 4852 4853 rxqueue = dev->_rx; 4854 4855 if (skb_rx_queue_recorded(skb)) { 4856 u16 index = skb_get_rx_queue(skb); 4857 4858 if (unlikely(index >= dev->real_num_rx_queues)) { 4859 WARN_ONCE(dev->real_num_rx_queues > 1, 4860 "%s received packet on queue %u, but number " 4861 "of RX queues is %u\n", 4862 dev->name, index, dev->real_num_rx_queues); 4863 4864 return rxqueue; /* Return first rxqueue */ 4865 } 4866 rxqueue += index; 4867 } 4868 return rxqueue; 4869 } 4870 4871 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 4872 struct bpf_prog *xdp_prog) 4873 { 4874 void *orig_data, *orig_data_end, *hard_start; 4875 struct netdev_rx_queue *rxqueue; 4876 bool orig_bcast, orig_host; 4877 u32 mac_len, frame_sz; 4878 __be16 orig_eth_type; 4879 struct ethhdr *eth; 4880 u32 metalen, act; 4881 int off; 4882 4883 /* The XDP program wants to see the packet starting at the MAC 4884 * header. 4885 */ 4886 mac_len = skb->data - skb_mac_header(skb); 4887 hard_start = skb->data - skb_headroom(skb); 4888 4889 /* SKB "head" area always have tailroom for skb_shared_info */ 4890 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 4891 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 4892 4893 rxqueue = netif_get_rxqueue(skb); 4894 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 4895 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 4896 skb_headlen(skb) + mac_len, true); 4897 if (skb_is_nonlinear(skb)) { 4898 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 4899 xdp_buff_set_frags_flag(xdp); 4900 } else { 4901 xdp_buff_clear_frags_flag(xdp); 4902 } 4903 4904 orig_data_end = xdp->data_end; 4905 orig_data = xdp->data; 4906 eth = (struct ethhdr *)xdp->data; 4907 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 4908 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 4909 orig_eth_type = eth->h_proto; 4910 4911 act = bpf_prog_run_xdp(xdp_prog, xdp); 4912 4913 /* check if bpf_xdp_adjust_head was used */ 4914 off = xdp->data - orig_data; 4915 if (off) { 4916 if (off > 0) 4917 __skb_pull(skb, off); 4918 else if (off < 0) 4919 __skb_push(skb, -off); 4920 4921 skb->mac_header += off; 4922 skb_reset_network_header(skb); 4923 } 4924 4925 /* check if bpf_xdp_adjust_tail was used */ 4926 off = xdp->data_end - orig_data_end; 4927 if (off != 0) { 4928 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 4929 skb->len += off; /* positive on grow, negative on shrink */ 4930 } 4931 4932 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 4933 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 4934 */ 4935 if (xdp_buff_has_frags(xdp)) 4936 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 4937 else 4938 skb->data_len = 0; 4939 4940 /* check if XDP changed eth hdr such SKB needs update */ 4941 eth = (struct ethhdr *)xdp->data; 4942 if ((orig_eth_type != eth->h_proto) || 4943 (orig_host != ether_addr_equal_64bits(eth->h_dest, 4944 skb->dev->dev_addr)) || 4945 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 4946 __skb_push(skb, ETH_HLEN); 4947 skb->pkt_type = PACKET_HOST; 4948 skb->protocol = eth_type_trans(skb, skb->dev); 4949 } 4950 4951 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 4952 * before calling us again on redirect path. We do not call do_redirect 4953 * as we leave that up to the caller. 4954 * 4955 * Caller is responsible for managing lifetime of skb (i.e. calling 4956 * kfree_skb in response to actions it cannot handle/XDP_DROP). 4957 */ 4958 switch (act) { 4959 case XDP_REDIRECT: 4960 case XDP_TX: 4961 __skb_push(skb, mac_len); 4962 break; 4963 case XDP_PASS: 4964 metalen = xdp->data - xdp->data_meta; 4965 if (metalen) 4966 skb_metadata_set(skb, metalen); 4967 break; 4968 } 4969 4970 return act; 4971 } 4972 4973 static int 4974 netif_skb_check_for_xdp(struct sk_buff **pskb, struct bpf_prog *prog) 4975 { 4976 struct sk_buff *skb = *pskb; 4977 int err, hroom, troom; 4978 4979 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 4980 return 0; 4981 4982 /* In case we have to go down the path and also linearize, 4983 * then lets do the pskb_expand_head() work just once here. 4984 */ 4985 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 4986 troom = skb->tail + skb->data_len - skb->end; 4987 err = pskb_expand_head(skb, 4988 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 4989 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 4990 if (err) 4991 return err; 4992 4993 return skb_linearize(skb); 4994 } 4995 4996 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 4997 struct xdp_buff *xdp, 4998 struct bpf_prog *xdp_prog) 4999 { 5000 struct sk_buff *skb = *pskb; 5001 u32 mac_len, act = XDP_DROP; 5002 5003 /* Reinjected packets coming from act_mirred or similar should 5004 * not get XDP generic processing. 5005 */ 5006 if (skb_is_redirected(skb)) 5007 return XDP_PASS; 5008 5009 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5010 * bytes. This is the guarantee that also native XDP provides, 5011 * thus we need to do it here as well. 5012 */ 5013 mac_len = skb->data - skb_mac_header(skb); 5014 __skb_push(skb, mac_len); 5015 5016 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5017 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5018 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5019 goto do_drop; 5020 } 5021 5022 __skb_pull(*pskb, mac_len); 5023 5024 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5025 switch (act) { 5026 case XDP_REDIRECT: 5027 case XDP_TX: 5028 case XDP_PASS: 5029 break; 5030 default: 5031 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5032 fallthrough; 5033 case XDP_ABORTED: 5034 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5035 fallthrough; 5036 case XDP_DROP: 5037 do_drop: 5038 kfree_skb(*pskb); 5039 break; 5040 } 5041 5042 return act; 5043 } 5044 5045 /* When doing generic XDP we have to bypass the qdisc layer and the 5046 * network taps in order to match in-driver-XDP behavior. This also means 5047 * that XDP packets are able to starve other packets going through a qdisc, 5048 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5049 * queues, so they do not have this starvation issue. 5050 */ 5051 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog) 5052 { 5053 struct net_device *dev = skb->dev; 5054 struct netdev_queue *txq; 5055 bool free_skb = true; 5056 int cpu, rc; 5057 5058 txq = netdev_core_pick_tx(dev, skb, NULL); 5059 cpu = smp_processor_id(); 5060 HARD_TX_LOCK(dev, txq, cpu); 5061 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5062 rc = netdev_start_xmit(skb, dev, txq, 0); 5063 if (dev_xmit_complete(rc)) 5064 free_skb = false; 5065 } 5066 HARD_TX_UNLOCK(dev, txq); 5067 if (free_skb) { 5068 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5069 dev_core_stats_tx_dropped_inc(dev); 5070 kfree_skb(skb); 5071 } 5072 } 5073 5074 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5075 5076 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5077 { 5078 if (xdp_prog) { 5079 struct xdp_buff xdp; 5080 u32 act; 5081 int err; 5082 5083 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5084 if (act != XDP_PASS) { 5085 switch (act) { 5086 case XDP_REDIRECT: 5087 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5088 &xdp, xdp_prog); 5089 if (err) 5090 goto out_redir; 5091 break; 5092 case XDP_TX: 5093 generic_xdp_tx(*pskb, xdp_prog); 5094 break; 5095 } 5096 return XDP_DROP; 5097 } 5098 } 5099 return XDP_PASS; 5100 out_redir: 5101 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5102 return XDP_DROP; 5103 } 5104 EXPORT_SYMBOL_GPL(do_xdp_generic); 5105 5106 static int netif_rx_internal(struct sk_buff *skb) 5107 { 5108 int ret; 5109 5110 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5111 5112 trace_netif_rx(skb); 5113 5114 #ifdef CONFIG_RPS 5115 if (static_branch_unlikely(&rps_needed)) { 5116 struct rps_dev_flow voidflow, *rflow = &voidflow; 5117 int cpu; 5118 5119 rcu_read_lock(); 5120 5121 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5122 if (cpu < 0) 5123 cpu = smp_processor_id(); 5124 5125 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5126 5127 rcu_read_unlock(); 5128 } else 5129 #endif 5130 { 5131 unsigned int qtail; 5132 5133 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5134 } 5135 return ret; 5136 } 5137 5138 /** 5139 * __netif_rx - Slightly optimized version of netif_rx 5140 * @skb: buffer to post 5141 * 5142 * This behaves as netif_rx except that it does not disable bottom halves. 5143 * As a result this function may only be invoked from the interrupt context 5144 * (either hard or soft interrupt). 5145 */ 5146 int __netif_rx(struct sk_buff *skb) 5147 { 5148 int ret; 5149 5150 lockdep_assert_once(hardirq_count() | softirq_count()); 5151 5152 trace_netif_rx_entry(skb); 5153 ret = netif_rx_internal(skb); 5154 trace_netif_rx_exit(ret); 5155 return ret; 5156 } 5157 EXPORT_SYMBOL(__netif_rx); 5158 5159 /** 5160 * netif_rx - post buffer to the network code 5161 * @skb: buffer to post 5162 * 5163 * This function receives a packet from a device driver and queues it for 5164 * the upper (protocol) levels to process via the backlog NAPI device. It 5165 * always succeeds. The buffer may be dropped during processing for 5166 * congestion control or by the protocol layers. 5167 * The network buffer is passed via the backlog NAPI device. Modern NIC 5168 * driver should use NAPI and GRO. 5169 * This function can used from interrupt and from process context. The 5170 * caller from process context must not disable interrupts before invoking 5171 * this function. 5172 * 5173 * return values: 5174 * NET_RX_SUCCESS (no congestion) 5175 * NET_RX_DROP (packet was dropped) 5176 * 5177 */ 5178 int netif_rx(struct sk_buff *skb) 5179 { 5180 bool need_bh_off = !(hardirq_count() | softirq_count()); 5181 int ret; 5182 5183 if (need_bh_off) 5184 local_bh_disable(); 5185 trace_netif_rx_entry(skb); 5186 ret = netif_rx_internal(skb); 5187 trace_netif_rx_exit(ret); 5188 if (need_bh_off) 5189 local_bh_enable(); 5190 return ret; 5191 } 5192 EXPORT_SYMBOL(netif_rx); 5193 5194 static __latent_entropy void net_tx_action(struct softirq_action *h) 5195 { 5196 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5197 5198 if (sd->completion_queue) { 5199 struct sk_buff *clist; 5200 5201 local_irq_disable(); 5202 clist = sd->completion_queue; 5203 sd->completion_queue = NULL; 5204 local_irq_enable(); 5205 5206 while (clist) { 5207 struct sk_buff *skb = clist; 5208 5209 clist = clist->next; 5210 5211 WARN_ON(refcount_read(&skb->users)); 5212 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5213 trace_consume_skb(skb, net_tx_action); 5214 else 5215 trace_kfree_skb(skb, net_tx_action, 5216 get_kfree_skb_cb(skb)->reason); 5217 5218 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5219 __kfree_skb(skb); 5220 else 5221 __napi_kfree_skb(skb, 5222 get_kfree_skb_cb(skb)->reason); 5223 } 5224 } 5225 5226 if (sd->output_queue) { 5227 struct Qdisc *head; 5228 5229 local_irq_disable(); 5230 head = sd->output_queue; 5231 sd->output_queue = NULL; 5232 sd->output_queue_tailp = &sd->output_queue; 5233 local_irq_enable(); 5234 5235 rcu_read_lock(); 5236 5237 while (head) { 5238 struct Qdisc *q = head; 5239 spinlock_t *root_lock = NULL; 5240 5241 head = head->next_sched; 5242 5243 /* We need to make sure head->next_sched is read 5244 * before clearing __QDISC_STATE_SCHED 5245 */ 5246 smp_mb__before_atomic(); 5247 5248 if (!(q->flags & TCQ_F_NOLOCK)) { 5249 root_lock = qdisc_lock(q); 5250 spin_lock(root_lock); 5251 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5252 &q->state))) { 5253 /* There is a synchronize_net() between 5254 * STATE_DEACTIVATED flag being set and 5255 * qdisc_reset()/some_qdisc_is_busy() in 5256 * dev_deactivate(), so we can safely bail out 5257 * early here to avoid data race between 5258 * qdisc_deactivate() and some_qdisc_is_busy() 5259 * for lockless qdisc. 5260 */ 5261 clear_bit(__QDISC_STATE_SCHED, &q->state); 5262 continue; 5263 } 5264 5265 clear_bit(__QDISC_STATE_SCHED, &q->state); 5266 qdisc_run(q); 5267 if (root_lock) 5268 spin_unlock(root_lock); 5269 } 5270 5271 rcu_read_unlock(); 5272 } 5273 5274 xfrm_dev_backlog(sd); 5275 } 5276 5277 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5278 /* This hook is defined here for ATM LANE */ 5279 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5280 unsigned char *addr) __read_mostly; 5281 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5282 #endif 5283 5284 /** 5285 * netdev_is_rx_handler_busy - check if receive handler is registered 5286 * @dev: device to check 5287 * 5288 * Check if a receive handler is already registered for a given device. 5289 * Return true if there one. 5290 * 5291 * The caller must hold the rtnl_mutex. 5292 */ 5293 bool netdev_is_rx_handler_busy(struct net_device *dev) 5294 { 5295 ASSERT_RTNL(); 5296 return dev && rtnl_dereference(dev->rx_handler); 5297 } 5298 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5299 5300 /** 5301 * netdev_rx_handler_register - register receive handler 5302 * @dev: device to register a handler for 5303 * @rx_handler: receive handler to register 5304 * @rx_handler_data: data pointer that is used by rx handler 5305 * 5306 * Register a receive handler for a device. This handler will then be 5307 * called from __netif_receive_skb. A negative errno code is returned 5308 * on a failure. 5309 * 5310 * The caller must hold the rtnl_mutex. 5311 * 5312 * For a general description of rx_handler, see enum rx_handler_result. 5313 */ 5314 int netdev_rx_handler_register(struct net_device *dev, 5315 rx_handler_func_t *rx_handler, 5316 void *rx_handler_data) 5317 { 5318 if (netdev_is_rx_handler_busy(dev)) 5319 return -EBUSY; 5320 5321 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5322 return -EINVAL; 5323 5324 /* Note: rx_handler_data must be set before rx_handler */ 5325 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5326 rcu_assign_pointer(dev->rx_handler, rx_handler); 5327 5328 return 0; 5329 } 5330 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5331 5332 /** 5333 * netdev_rx_handler_unregister - unregister receive handler 5334 * @dev: device to unregister a handler from 5335 * 5336 * Unregister a receive handler from a device. 5337 * 5338 * The caller must hold the rtnl_mutex. 5339 */ 5340 void netdev_rx_handler_unregister(struct net_device *dev) 5341 { 5342 5343 ASSERT_RTNL(); 5344 RCU_INIT_POINTER(dev->rx_handler, NULL); 5345 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5346 * section has a guarantee to see a non NULL rx_handler_data 5347 * as well. 5348 */ 5349 synchronize_net(); 5350 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5351 } 5352 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5353 5354 /* 5355 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5356 * the special handling of PFMEMALLOC skbs. 5357 */ 5358 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5359 { 5360 switch (skb->protocol) { 5361 case htons(ETH_P_ARP): 5362 case htons(ETH_P_IP): 5363 case htons(ETH_P_IPV6): 5364 case htons(ETH_P_8021Q): 5365 case htons(ETH_P_8021AD): 5366 return true; 5367 default: 5368 return false; 5369 } 5370 } 5371 5372 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5373 int *ret, struct net_device *orig_dev) 5374 { 5375 if (nf_hook_ingress_active(skb)) { 5376 int ingress_retval; 5377 5378 if (*pt_prev) { 5379 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5380 *pt_prev = NULL; 5381 } 5382 5383 rcu_read_lock(); 5384 ingress_retval = nf_hook_ingress(skb); 5385 rcu_read_unlock(); 5386 return ingress_retval; 5387 } 5388 return 0; 5389 } 5390 5391 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5392 struct packet_type **ppt_prev) 5393 { 5394 struct packet_type *ptype, *pt_prev; 5395 rx_handler_func_t *rx_handler; 5396 struct sk_buff *skb = *pskb; 5397 struct net_device *orig_dev; 5398 bool deliver_exact = false; 5399 int ret = NET_RX_DROP; 5400 __be16 type; 5401 5402 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5403 5404 trace_netif_receive_skb(skb); 5405 5406 orig_dev = skb->dev; 5407 5408 skb_reset_network_header(skb); 5409 if (!skb_transport_header_was_set(skb)) 5410 skb_reset_transport_header(skb); 5411 skb_reset_mac_len(skb); 5412 5413 pt_prev = NULL; 5414 5415 another_round: 5416 skb->skb_iif = skb->dev->ifindex; 5417 5418 __this_cpu_inc(softnet_data.processed); 5419 5420 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5421 int ret2; 5422 5423 migrate_disable(); 5424 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5425 &skb); 5426 migrate_enable(); 5427 5428 if (ret2 != XDP_PASS) { 5429 ret = NET_RX_DROP; 5430 goto out; 5431 } 5432 } 5433 5434 if (eth_type_vlan(skb->protocol)) { 5435 skb = skb_vlan_untag(skb); 5436 if (unlikely(!skb)) 5437 goto out; 5438 } 5439 5440 if (skb_skip_tc_classify(skb)) 5441 goto skip_classify; 5442 5443 if (pfmemalloc) 5444 goto skip_taps; 5445 5446 list_for_each_entry_rcu(ptype, &net_hotdata.ptype_all, list) { 5447 if (pt_prev) 5448 ret = deliver_skb(skb, pt_prev, orig_dev); 5449 pt_prev = ptype; 5450 } 5451 5452 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5453 if (pt_prev) 5454 ret = deliver_skb(skb, pt_prev, orig_dev); 5455 pt_prev = ptype; 5456 } 5457 5458 skip_taps: 5459 #ifdef CONFIG_NET_INGRESS 5460 if (static_branch_unlikely(&ingress_needed_key)) { 5461 bool another = false; 5462 5463 nf_skip_egress(skb, true); 5464 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5465 &another); 5466 if (another) 5467 goto another_round; 5468 if (!skb) 5469 goto out; 5470 5471 nf_skip_egress(skb, false); 5472 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5473 goto out; 5474 } 5475 #endif 5476 skb_reset_redirect(skb); 5477 skip_classify: 5478 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5479 goto drop; 5480 5481 if (skb_vlan_tag_present(skb)) { 5482 if (pt_prev) { 5483 ret = deliver_skb(skb, pt_prev, orig_dev); 5484 pt_prev = NULL; 5485 } 5486 if (vlan_do_receive(&skb)) 5487 goto another_round; 5488 else if (unlikely(!skb)) 5489 goto out; 5490 } 5491 5492 rx_handler = rcu_dereference(skb->dev->rx_handler); 5493 if (rx_handler) { 5494 if (pt_prev) { 5495 ret = deliver_skb(skb, pt_prev, orig_dev); 5496 pt_prev = NULL; 5497 } 5498 switch (rx_handler(&skb)) { 5499 case RX_HANDLER_CONSUMED: 5500 ret = NET_RX_SUCCESS; 5501 goto out; 5502 case RX_HANDLER_ANOTHER: 5503 goto another_round; 5504 case RX_HANDLER_EXACT: 5505 deliver_exact = true; 5506 break; 5507 case RX_HANDLER_PASS: 5508 break; 5509 default: 5510 BUG(); 5511 } 5512 } 5513 5514 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5515 check_vlan_id: 5516 if (skb_vlan_tag_get_id(skb)) { 5517 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5518 * find vlan device. 5519 */ 5520 skb->pkt_type = PACKET_OTHERHOST; 5521 } else if (eth_type_vlan(skb->protocol)) { 5522 /* Outer header is 802.1P with vlan 0, inner header is 5523 * 802.1Q or 802.1AD and vlan_do_receive() above could 5524 * not find vlan dev for vlan id 0. 5525 */ 5526 __vlan_hwaccel_clear_tag(skb); 5527 skb = skb_vlan_untag(skb); 5528 if (unlikely(!skb)) 5529 goto out; 5530 if (vlan_do_receive(&skb)) 5531 /* After stripping off 802.1P header with vlan 0 5532 * vlan dev is found for inner header. 5533 */ 5534 goto another_round; 5535 else if (unlikely(!skb)) 5536 goto out; 5537 else 5538 /* We have stripped outer 802.1P vlan 0 header. 5539 * But could not find vlan dev. 5540 * check again for vlan id to set OTHERHOST. 5541 */ 5542 goto check_vlan_id; 5543 } 5544 /* Note: we might in the future use prio bits 5545 * and set skb->priority like in vlan_do_receive() 5546 * For the time being, just ignore Priority Code Point 5547 */ 5548 __vlan_hwaccel_clear_tag(skb); 5549 } 5550 5551 type = skb->protocol; 5552 5553 /* deliver only exact match when indicated */ 5554 if (likely(!deliver_exact)) { 5555 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5556 &ptype_base[ntohs(type) & 5557 PTYPE_HASH_MASK]); 5558 } 5559 5560 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5561 &orig_dev->ptype_specific); 5562 5563 if (unlikely(skb->dev != orig_dev)) { 5564 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5565 &skb->dev->ptype_specific); 5566 } 5567 5568 if (pt_prev) { 5569 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5570 goto drop; 5571 *ppt_prev = pt_prev; 5572 } else { 5573 drop: 5574 if (!deliver_exact) 5575 dev_core_stats_rx_dropped_inc(skb->dev); 5576 else 5577 dev_core_stats_rx_nohandler_inc(skb->dev); 5578 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5579 /* Jamal, now you will not able to escape explaining 5580 * me how you were going to use this. :-) 5581 */ 5582 ret = NET_RX_DROP; 5583 } 5584 5585 out: 5586 /* The invariant here is that if *ppt_prev is not NULL 5587 * then skb should also be non-NULL. 5588 * 5589 * Apparently *ppt_prev assignment above holds this invariant due to 5590 * skb dereferencing near it. 5591 */ 5592 *pskb = skb; 5593 return ret; 5594 } 5595 5596 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5597 { 5598 struct net_device *orig_dev = skb->dev; 5599 struct packet_type *pt_prev = NULL; 5600 int ret; 5601 5602 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5603 if (pt_prev) 5604 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5605 skb->dev, pt_prev, orig_dev); 5606 return ret; 5607 } 5608 5609 /** 5610 * netif_receive_skb_core - special purpose version of netif_receive_skb 5611 * @skb: buffer to process 5612 * 5613 * More direct receive version of netif_receive_skb(). It should 5614 * only be used by callers that have a need to skip RPS and Generic XDP. 5615 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5616 * 5617 * This function may only be called from softirq context and interrupts 5618 * should be enabled. 5619 * 5620 * Return values (usually ignored): 5621 * NET_RX_SUCCESS: no congestion 5622 * NET_RX_DROP: packet was dropped 5623 */ 5624 int netif_receive_skb_core(struct sk_buff *skb) 5625 { 5626 int ret; 5627 5628 rcu_read_lock(); 5629 ret = __netif_receive_skb_one_core(skb, false); 5630 rcu_read_unlock(); 5631 5632 return ret; 5633 } 5634 EXPORT_SYMBOL(netif_receive_skb_core); 5635 5636 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5637 struct packet_type *pt_prev, 5638 struct net_device *orig_dev) 5639 { 5640 struct sk_buff *skb, *next; 5641 5642 if (!pt_prev) 5643 return; 5644 if (list_empty(head)) 5645 return; 5646 if (pt_prev->list_func != NULL) 5647 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5648 ip_list_rcv, head, pt_prev, orig_dev); 5649 else 5650 list_for_each_entry_safe(skb, next, head, list) { 5651 skb_list_del_init(skb); 5652 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5653 } 5654 } 5655 5656 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5657 { 5658 /* Fast-path assumptions: 5659 * - There is no RX handler. 5660 * - Only one packet_type matches. 5661 * If either of these fails, we will end up doing some per-packet 5662 * processing in-line, then handling the 'last ptype' for the whole 5663 * sublist. This can't cause out-of-order delivery to any single ptype, 5664 * because the 'last ptype' must be constant across the sublist, and all 5665 * other ptypes are handled per-packet. 5666 */ 5667 /* Current (common) ptype of sublist */ 5668 struct packet_type *pt_curr = NULL; 5669 /* Current (common) orig_dev of sublist */ 5670 struct net_device *od_curr = NULL; 5671 struct list_head sublist; 5672 struct sk_buff *skb, *next; 5673 5674 INIT_LIST_HEAD(&sublist); 5675 list_for_each_entry_safe(skb, next, head, list) { 5676 struct net_device *orig_dev = skb->dev; 5677 struct packet_type *pt_prev = NULL; 5678 5679 skb_list_del_init(skb); 5680 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5681 if (!pt_prev) 5682 continue; 5683 if (pt_curr != pt_prev || od_curr != orig_dev) { 5684 /* dispatch old sublist */ 5685 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5686 /* start new sublist */ 5687 INIT_LIST_HEAD(&sublist); 5688 pt_curr = pt_prev; 5689 od_curr = orig_dev; 5690 } 5691 list_add_tail(&skb->list, &sublist); 5692 } 5693 5694 /* dispatch final sublist */ 5695 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5696 } 5697 5698 static int __netif_receive_skb(struct sk_buff *skb) 5699 { 5700 int ret; 5701 5702 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 5703 unsigned int noreclaim_flag; 5704 5705 /* 5706 * PFMEMALLOC skbs are special, they should 5707 * - be delivered to SOCK_MEMALLOC sockets only 5708 * - stay away from userspace 5709 * - have bounded memory usage 5710 * 5711 * Use PF_MEMALLOC as this saves us from propagating the allocation 5712 * context down to all allocation sites. 5713 */ 5714 noreclaim_flag = memalloc_noreclaim_save(); 5715 ret = __netif_receive_skb_one_core(skb, true); 5716 memalloc_noreclaim_restore(noreclaim_flag); 5717 } else 5718 ret = __netif_receive_skb_one_core(skb, false); 5719 5720 return ret; 5721 } 5722 5723 static void __netif_receive_skb_list(struct list_head *head) 5724 { 5725 unsigned long noreclaim_flag = 0; 5726 struct sk_buff *skb, *next; 5727 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 5728 5729 list_for_each_entry_safe(skb, next, head, list) { 5730 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 5731 struct list_head sublist; 5732 5733 /* Handle the previous sublist */ 5734 list_cut_before(&sublist, head, &skb->list); 5735 if (!list_empty(&sublist)) 5736 __netif_receive_skb_list_core(&sublist, pfmemalloc); 5737 pfmemalloc = !pfmemalloc; 5738 /* See comments in __netif_receive_skb */ 5739 if (pfmemalloc) 5740 noreclaim_flag = memalloc_noreclaim_save(); 5741 else 5742 memalloc_noreclaim_restore(noreclaim_flag); 5743 } 5744 } 5745 /* Handle the remaining sublist */ 5746 if (!list_empty(head)) 5747 __netif_receive_skb_list_core(head, pfmemalloc); 5748 /* Restore pflags */ 5749 if (pfmemalloc) 5750 memalloc_noreclaim_restore(noreclaim_flag); 5751 } 5752 5753 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 5754 { 5755 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 5756 struct bpf_prog *new = xdp->prog; 5757 int ret = 0; 5758 5759 switch (xdp->command) { 5760 case XDP_SETUP_PROG: 5761 rcu_assign_pointer(dev->xdp_prog, new); 5762 if (old) 5763 bpf_prog_put(old); 5764 5765 if (old && !new) { 5766 static_branch_dec(&generic_xdp_needed_key); 5767 } else if (new && !old) { 5768 static_branch_inc(&generic_xdp_needed_key); 5769 dev_disable_lro(dev); 5770 dev_disable_gro_hw(dev); 5771 } 5772 break; 5773 5774 default: 5775 ret = -EINVAL; 5776 break; 5777 } 5778 5779 return ret; 5780 } 5781 5782 static int netif_receive_skb_internal(struct sk_buff *skb) 5783 { 5784 int ret; 5785 5786 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5787 5788 if (skb_defer_rx_timestamp(skb)) 5789 return NET_RX_SUCCESS; 5790 5791 rcu_read_lock(); 5792 #ifdef CONFIG_RPS 5793 if (static_branch_unlikely(&rps_needed)) { 5794 struct rps_dev_flow voidflow, *rflow = &voidflow; 5795 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 5796 5797 if (cpu >= 0) { 5798 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5799 rcu_read_unlock(); 5800 return ret; 5801 } 5802 } 5803 #endif 5804 ret = __netif_receive_skb(skb); 5805 rcu_read_unlock(); 5806 return ret; 5807 } 5808 5809 void netif_receive_skb_list_internal(struct list_head *head) 5810 { 5811 struct sk_buff *skb, *next; 5812 struct list_head sublist; 5813 5814 INIT_LIST_HEAD(&sublist); 5815 list_for_each_entry_safe(skb, next, head, list) { 5816 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 5817 skb); 5818 skb_list_del_init(skb); 5819 if (!skb_defer_rx_timestamp(skb)) 5820 list_add_tail(&skb->list, &sublist); 5821 } 5822 list_splice_init(&sublist, head); 5823 5824 rcu_read_lock(); 5825 #ifdef CONFIG_RPS 5826 if (static_branch_unlikely(&rps_needed)) { 5827 list_for_each_entry_safe(skb, next, head, list) { 5828 struct rps_dev_flow voidflow, *rflow = &voidflow; 5829 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 5830 5831 if (cpu >= 0) { 5832 /* Will be handled, remove from list */ 5833 skb_list_del_init(skb); 5834 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5835 } 5836 } 5837 } 5838 #endif 5839 __netif_receive_skb_list(head); 5840 rcu_read_unlock(); 5841 } 5842 5843 /** 5844 * netif_receive_skb - process receive buffer from network 5845 * @skb: buffer to process 5846 * 5847 * netif_receive_skb() is the main receive data processing function. 5848 * It always succeeds. The buffer may be dropped during processing 5849 * for congestion control or by the protocol layers. 5850 * 5851 * This function may only be called from softirq context and interrupts 5852 * should be enabled. 5853 * 5854 * Return values (usually ignored): 5855 * NET_RX_SUCCESS: no congestion 5856 * NET_RX_DROP: packet was dropped 5857 */ 5858 int netif_receive_skb(struct sk_buff *skb) 5859 { 5860 int ret; 5861 5862 trace_netif_receive_skb_entry(skb); 5863 5864 ret = netif_receive_skb_internal(skb); 5865 trace_netif_receive_skb_exit(ret); 5866 5867 return ret; 5868 } 5869 EXPORT_SYMBOL(netif_receive_skb); 5870 5871 /** 5872 * netif_receive_skb_list - process many receive buffers from network 5873 * @head: list of skbs to process. 5874 * 5875 * Since return value of netif_receive_skb() is normally ignored, and 5876 * wouldn't be meaningful for a list, this function returns void. 5877 * 5878 * This function may only be called from softirq context and interrupts 5879 * should be enabled. 5880 */ 5881 void netif_receive_skb_list(struct list_head *head) 5882 { 5883 struct sk_buff *skb; 5884 5885 if (list_empty(head)) 5886 return; 5887 if (trace_netif_receive_skb_list_entry_enabled()) { 5888 list_for_each_entry(skb, head, list) 5889 trace_netif_receive_skb_list_entry(skb); 5890 } 5891 netif_receive_skb_list_internal(head); 5892 trace_netif_receive_skb_list_exit(0); 5893 } 5894 EXPORT_SYMBOL(netif_receive_skb_list); 5895 5896 static DEFINE_PER_CPU(struct work_struct, flush_works); 5897 5898 /* Network device is going away, flush any packets still pending */ 5899 static void flush_backlog(struct work_struct *work) 5900 { 5901 struct sk_buff *skb, *tmp; 5902 struct softnet_data *sd; 5903 5904 local_bh_disable(); 5905 sd = this_cpu_ptr(&softnet_data); 5906 5907 backlog_lock_irq_disable(sd); 5908 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 5909 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 5910 __skb_unlink(skb, &sd->input_pkt_queue); 5911 dev_kfree_skb_irq(skb); 5912 rps_input_queue_head_incr(sd); 5913 } 5914 } 5915 backlog_unlock_irq_enable(sd); 5916 5917 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 5918 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 5919 __skb_unlink(skb, &sd->process_queue); 5920 kfree_skb(skb); 5921 rps_input_queue_head_incr(sd); 5922 } 5923 } 5924 local_bh_enable(); 5925 } 5926 5927 static bool flush_required(int cpu) 5928 { 5929 #if IS_ENABLED(CONFIG_RPS) 5930 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 5931 bool do_flush; 5932 5933 backlog_lock_irq_disable(sd); 5934 5935 /* as insertion into process_queue happens with the rps lock held, 5936 * process_queue access may race only with dequeue 5937 */ 5938 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 5939 !skb_queue_empty_lockless(&sd->process_queue); 5940 backlog_unlock_irq_enable(sd); 5941 5942 return do_flush; 5943 #endif 5944 /* without RPS we can't safely check input_pkt_queue: during a 5945 * concurrent remote skb_queue_splice() we can detect as empty both 5946 * input_pkt_queue and process_queue even if the latter could end-up 5947 * containing a lot of packets. 5948 */ 5949 return true; 5950 } 5951 5952 static void flush_all_backlogs(void) 5953 { 5954 static cpumask_t flush_cpus; 5955 unsigned int cpu; 5956 5957 /* since we are under rtnl lock protection we can use static data 5958 * for the cpumask and avoid allocating on stack the possibly 5959 * large mask 5960 */ 5961 ASSERT_RTNL(); 5962 5963 cpus_read_lock(); 5964 5965 cpumask_clear(&flush_cpus); 5966 for_each_online_cpu(cpu) { 5967 if (flush_required(cpu)) { 5968 queue_work_on(cpu, system_highpri_wq, 5969 per_cpu_ptr(&flush_works, cpu)); 5970 cpumask_set_cpu(cpu, &flush_cpus); 5971 } 5972 } 5973 5974 /* we can have in flight packet[s] on the cpus we are not flushing, 5975 * synchronize_net() in unregister_netdevice_many() will take care of 5976 * them 5977 */ 5978 for_each_cpu(cpu, &flush_cpus) 5979 flush_work(per_cpu_ptr(&flush_works, cpu)); 5980 5981 cpus_read_unlock(); 5982 } 5983 5984 static void net_rps_send_ipi(struct softnet_data *remsd) 5985 { 5986 #ifdef CONFIG_RPS 5987 while (remsd) { 5988 struct softnet_data *next = remsd->rps_ipi_next; 5989 5990 if (cpu_online(remsd->cpu)) 5991 smp_call_function_single_async(remsd->cpu, &remsd->csd); 5992 remsd = next; 5993 } 5994 #endif 5995 } 5996 5997 /* 5998 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 5999 * Note: called with local irq disabled, but exits with local irq enabled. 6000 */ 6001 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6002 { 6003 #ifdef CONFIG_RPS 6004 struct softnet_data *remsd = sd->rps_ipi_list; 6005 6006 if (!use_backlog_threads() && remsd) { 6007 sd->rps_ipi_list = NULL; 6008 6009 local_irq_enable(); 6010 6011 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6012 net_rps_send_ipi(remsd); 6013 } else 6014 #endif 6015 local_irq_enable(); 6016 } 6017 6018 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6019 { 6020 #ifdef CONFIG_RPS 6021 return !use_backlog_threads() && sd->rps_ipi_list; 6022 #else 6023 return false; 6024 #endif 6025 } 6026 6027 static int process_backlog(struct napi_struct *napi, int quota) 6028 { 6029 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6030 bool again = true; 6031 int work = 0; 6032 6033 /* Check if we have pending ipi, its better to send them now, 6034 * not waiting net_rx_action() end. 6035 */ 6036 if (sd_has_rps_ipi_waiting(sd)) { 6037 local_irq_disable(); 6038 net_rps_action_and_irq_enable(sd); 6039 } 6040 6041 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6042 while (again) { 6043 struct sk_buff *skb; 6044 6045 while ((skb = __skb_dequeue(&sd->process_queue))) { 6046 rcu_read_lock(); 6047 __netif_receive_skb(skb); 6048 rcu_read_unlock(); 6049 if (++work >= quota) { 6050 rps_input_queue_head_add(sd, work); 6051 return work; 6052 } 6053 6054 } 6055 6056 backlog_lock_irq_disable(sd); 6057 if (skb_queue_empty(&sd->input_pkt_queue)) { 6058 /* 6059 * Inline a custom version of __napi_complete(). 6060 * only current cpu owns and manipulates this napi, 6061 * and NAPI_STATE_SCHED is the only possible flag set 6062 * on backlog. 6063 * We can use a plain write instead of clear_bit(), 6064 * and we dont need an smp_mb() memory barrier. 6065 */ 6066 napi->state &= NAPIF_STATE_THREADED; 6067 again = false; 6068 } else { 6069 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6070 &sd->process_queue); 6071 } 6072 backlog_unlock_irq_enable(sd); 6073 } 6074 6075 if (work) 6076 rps_input_queue_head_add(sd, work); 6077 return work; 6078 } 6079 6080 /** 6081 * __napi_schedule - schedule for receive 6082 * @n: entry to schedule 6083 * 6084 * The entry's receive function will be scheduled to run. 6085 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6086 */ 6087 void __napi_schedule(struct napi_struct *n) 6088 { 6089 unsigned long flags; 6090 6091 local_irq_save(flags); 6092 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6093 local_irq_restore(flags); 6094 } 6095 EXPORT_SYMBOL(__napi_schedule); 6096 6097 /** 6098 * napi_schedule_prep - check if napi can be scheduled 6099 * @n: napi context 6100 * 6101 * Test if NAPI routine is already running, and if not mark 6102 * it as running. This is used as a condition variable to 6103 * insure only one NAPI poll instance runs. We also make 6104 * sure there is no pending NAPI disable. 6105 */ 6106 bool napi_schedule_prep(struct napi_struct *n) 6107 { 6108 unsigned long new, val = READ_ONCE(n->state); 6109 6110 do { 6111 if (unlikely(val & NAPIF_STATE_DISABLE)) 6112 return false; 6113 new = val | NAPIF_STATE_SCHED; 6114 6115 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6116 * This was suggested by Alexander Duyck, as compiler 6117 * emits better code than : 6118 * if (val & NAPIF_STATE_SCHED) 6119 * new |= NAPIF_STATE_MISSED; 6120 */ 6121 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6122 NAPIF_STATE_MISSED; 6123 } while (!try_cmpxchg(&n->state, &val, new)); 6124 6125 return !(val & NAPIF_STATE_SCHED); 6126 } 6127 EXPORT_SYMBOL(napi_schedule_prep); 6128 6129 /** 6130 * __napi_schedule_irqoff - schedule for receive 6131 * @n: entry to schedule 6132 * 6133 * Variant of __napi_schedule() assuming hard irqs are masked. 6134 * 6135 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6136 * because the interrupt disabled assumption might not be true 6137 * due to force-threaded interrupts and spinlock substitution. 6138 */ 6139 void __napi_schedule_irqoff(struct napi_struct *n) 6140 { 6141 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6142 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6143 else 6144 __napi_schedule(n); 6145 } 6146 EXPORT_SYMBOL(__napi_schedule_irqoff); 6147 6148 bool napi_complete_done(struct napi_struct *n, int work_done) 6149 { 6150 unsigned long flags, val, new, timeout = 0; 6151 bool ret = true; 6152 6153 /* 6154 * 1) Don't let napi dequeue from the cpu poll list 6155 * just in case its running on a different cpu. 6156 * 2) If we are busy polling, do nothing here, we have 6157 * the guarantee we will be called later. 6158 */ 6159 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6160 NAPIF_STATE_IN_BUSY_POLL))) 6161 return false; 6162 6163 if (work_done) { 6164 if (n->gro_bitmask) 6165 timeout = READ_ONCE(n->dev->gro_flush_timeout); 6166 n->defer_hard_irqs_count = READ_ONCE(n->dev->napi_defer_hard_irqs); 6167 } 6168 if (n->defer_hard_irqs_count > 0) { 6169 n->defer_hard_irqs_count--; 6170 timeout = READ_ONCE(n->dev->gro_flush_timeout); 6171 if (timeout) 6172 ret = false; 6173 } 6174 if (n->gro_bitmask) { 6175 /* When the NAPI instance uses a timeout and keeps postponing 6176 * it, we need to bound somehow the time packets are kept in 6177 * the GRO layer 6178 */ 6179 napi_gro_flush(n, !!timeout); 6180 } 6181 6182 gro_normal_list(n); 6183 6184 if (unlikely(!list_empty(&n->poll_list))) { 6185 /* If n->poll_list is not empty, we need to mask irqs */ 6186 local_irq_save(flags); 6187 list_del_init(&n->poll_list); 6188 local_irq_restore(flags); 6189 } 6190 WRITE_ONCE(n->list_owner, -1); 6191 6192 val = READ_ONCE(n->state); 6193 do { 6194 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6195 6196 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6197 NAPIF_STATE_SCHED_THREADED | 6198 NAPIF_STATE_PREFER_BUSY_POLL); 6199 6200 /* If STATE_MISSED was set, leave STATE_SCHED set, 6201 * because we will call napi->poll() one more time. 6202 * This C code was suggested by Alexander Duyck to help gcc. 6203 */ 6204 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6205 NAPIF_STATE_SCHED; 6206 } while (!try_cmpxchg(&n->state, &val, new)); 6207 6208 if (unlikely(val & NAPIF_STATE_MISSED)) { 6209 __napi_schedule(n); 6210 return false; 6211 } 6212 6213 if (timeout) 6214 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6215 HRTIMER_MODE_REL_PINNED); 6216 return ret; 6217 } 6218 EXPORT_SYMBOL(napi_complete_done); 6219 6220 /* must be called under rcu_read_lock(), as we dont take a reference */ 6221 struct napi_struct *napi_by_id(unsigned int napi_id) 6222 { 6223 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 6224 struct napi_struct *napi; 6225 6226 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 6227 if (napi->napi_id == napi_id) 6228 return napi; 6229 6230 return NULL; 6231 } 6232 6233 static void skb_defer_free_flush(struct softnet_data *sd) 6234 { 6235 struct sk_buff *skb, *next; 6236 6237 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6238 if (!READ_ONCE(sd->defer_list)) 6239 return; 6240 6241 spin_lock(&sd->defer_lock); 6242 skb = sd->defer_list; 6243 sd->defer_list = NULL; 6244 sd->defer_count = 0; 6245 spin_unlock(&sd->defer_lock); 6246 6247 while (skb != NULL) { 6248 next = skb->next; 6249 napi_consume_skb(skb, 1); 6250 skb = next; 6251 } 6252 } 6253 6254 #if defined(CONFIG_NET_RX_BUSY_POLL) 6255 6256 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6257 { 6258 if (!skip_schedule) { 6259 gro_normal_list(napi); 6260 __napi_schedule(napi); 6261 return; 6262 } 6263 6264 if (napi->gro_bitmask) { 6265 /* flush too old packets 6266 * If HZ < 1000, flush all packets. 6267 */ 6268 napi_gro_flush(napi, HZ >= 1000); 6269 } 6270 6271 gro_normal_list(napi); 6272 clear_bit(NAPI_STATE_SCHED, &napi->state); 6273 } 6274 6275 enum { 6276 NAPI_F_PREFER_BUSY_POLL = 1, 6277 NAPI_F_END_ON_RESCHED = 2, 6278 }; 6279 6280 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6281 unsigned flags, u16 budget) 6282 { 6283 bool skip_schedule = false; 6284 unsigned long timeout; 6285 int rc; 6286 6287 /* Busy polling means there is a high chance device driver hard irq 6288 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6289 * set in napi_schedule_prep(). 6290 * Since we are about to call napi->poll() once more, we can safely 6291 * clear NAPI_STATE_MISSED. 6292 * 6293 * Note: x86 could use a single "lock and ..." instruction 6294 * to perform these two clear_bit() 6295 */ 6296 clear_bit(NAPI_STATE_MISSED, &napi->state); 6297 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6298 6299 local_bh_disable(); 6300 6301 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6302 napi->defer_hard_irqs_count = READ_ONCE(napi->dev->napi_defer_hard_irqs); 6303 timeout = READ_ONCE(napi->dev->gro_flush_timeout); 6304 if (napi->defer_hard_irqs_count && timeout) { 6305 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6306 skip_schedule = true; 6307 } 6308 } 6309 6310 /* All we really want here is to re-enable device interrupts. 6311 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6312 */ 6313 rc = napi->poll(napi, budget); 6314 /* We can't gro_normal_list() here, because napi->poll() might have 6315 * rearmed the napi (napi_complete_done()) in which case it could 6316 * already be running on another CPU. 6317 */ 6318 trace_napi_poll(napi, rc, budget); 6319 netpoll_poll_unlock(have_poll_lock); 6320 if (rc == budget) 6321 __busy_poll_stop(napi, skip_schedule); 6322 local_bh_enable(); 6323 } 6324 6325 static void __napi_busy_loop(unsigned int napi_id, 6326 bool (*loop_end)(void *, unsigned long), 6327 void *loop_end_arg, unsigned flags, u16 budget) 6328 { 6329 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6330 int (*napi_poll)(struct napi_struct *napi, int budget); 6331 void *have_poll_lock = NULL; 6332 struct napi_struct *napi; 6333 6334 WARN_ON_ONCE(!rcu_read_lock_held()); 6335 6336 restart: 6337 napi_poll = NULL; 6338 6339 napi = napi_by_id(napi_id); 6340 if (!napi) 6341 return; 6342 6343 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6344 preempt_disable(); 6345 for (;;) { 6346 int work = 0; 6347 6348 local_bh_disable(); 6349 if (!napi_poll) { 6350 unsigned long val = READ_ONCE(napi->state); 6351 6352 /* If multiple threads are competing for this napi, 6353 * we avoid dirtying napi->state as much as we can. 6354 */ 6355 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6356 NAPIF_STATE_IN_BUSY_POLL)) { 6357 if (flags & NAPI_F_PREFER_BUSY_POLL) 6358 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6359 goto count; 6360 } 6361 if (cmpxchg(&napi->state, val, 6362 val | NAPIF_STATE_IN_BUSY_POLL | 6363 NAPIF_STATE_SCHED) != val) { 6364 if (flags & NAPI_F_PREFER_BUSY_POLL) 6365 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6366 goto count; 6367 } 6368 have_poll_lock = netpoll_poll_lock(napi); 6369 napi_poll = napi->poll; 6370 } 6371 work = napi_poll(napi, budget); 6372 trace_napi_poll(napi, work, budget); 6373 gro_normal_list(napi); 6374 count: 6375 if (work > 0) 6376 __NET_ADD_STATS(dev_net(napi->dev), 6377 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6378 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6379 local_bh_enable(); 6380 6381 if (!loop_end || loop_end(loop_end_arg, start_time)) 6382 break; 6383 6384 if (unlikely(need_resched())) { 6385 if (flags & NAPI_F_END_ON_RESCHED) 6386 break; 6387 if (napi_poll) 6388 busy_poll_stop(napi, have_poll_lock, flags, budget); 6389 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6390 preempt_enable(); 6391 rcu_read_unlock(); 6392 cond_resched(); 6393 rcu_read_lock(); 6394 if (loop_end(loop_end_arg, start_time)) 6395 return; 6396 goto restart; 6397 } 6398 cpu_relax(); 6399 } 6400 if (napi_poll) 6401 busy_poll_stop(napi, have_poll_lock, flags, budget); 6402 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6403 preempt_enable(); 6404 } 6405 6406 void napi_busy_loop_rcu(unsigned int napi_id, 6407 bool (*loop_end)(void *, unsigned long), 6408 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6409 { 6410 unsigned flags = NAPI_F_END_ON_RESCHED; 6411 6412 if (prefer_busy_poll) 6413 flags |= NAPI_F_PREFER_BUSY_POLL; 6414 6415 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6416 } 6417 6418 void napi_busy_loop(unsigned int napi_id, 6419 bool (*loop_end)(void *, unsigned long), 6420 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6421 { 6422 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6423 6424 rcu_read_lock(); 6425 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6426 rcu_read_unlock(); 6427 } 6428 EXPORT_SYMBOL(napi_busy_loop); 6429 6430 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6431 6432 static void napi_hash_add(struct napi_struct *napi) 6433 { 6434 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6435 return; 6436 6437 spin_lock(&napi_hash_lock); 6438 6439 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6440 do { 6441 if (unlikely(++napi_gen_id < MIN_NAPI_ID)) 6442 napi_gen_id = MIN_NAPI_ID; 6443 } while (napi_by_id(napi_gen_id)); 6444 napi->napi_id = napi_gen_id; 6445 6446 hlist_add_head_rcu(&napi->napi_hash_node, 6447 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6448 6449 spin_unlock(&napi_hash_lock); 6450 } 6451 6452 /* Warning : caller is responsible to make sure rcu grace period 6453 * is respected before freeing memory containing @napi 6454 */ 6455 static void napi_hash_del(struct napi_struct *napi) 6456 { 6457 spin_lock(&napi_hash_lock); 6458 6459 hlist_del_init_rcu(&napi->napi_hash_node); 6460 6461 spin_unlock(&napi_hash_lock); 6462 } 6463 6464 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6465 { 6466 struct napi_struct *napi; 6467 6468 napi = container_of(timer, struct napi_struct, timer); 6469 6470 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6471 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6472 */ 6473 if (!napi_disable_pending(napi) && 6474 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6475 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6476 __napi_schedule_irqoff(napi); 6477 } 6478 6479 return HRTIMER_NORESTART; 6480 } 6481 6482 static void init_gro_hash(struct napi_struct *napi) 6483 { 6484 int i; 6485 6486 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6487 INIT_LIST_HEAD(&napi->gro_hash[i].list); 6488 napi->gro_hash[i].count = 0; 6489 } 6490 napi->gro_bitmask = 0; 6491 } 6492 6493 int dev_set_threaded(struct net_device *dev, bool threaded) 6494 { 6495 struct napi_struct *napi; 6496 int err = 0; 6497 6498 if (dev->threaded == threaded) 6499 return 0; 6500 6501 if (threaded) { 6502 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6503 if (!napi->thread) { 6504 err = napi_kthread_create(napi); 6505 if (err) { 6506 threaded = false; 6507 break; 6508 } 6509 } 6510 } 6511 } 6512 6513 dev->threaded = threaded; 6514 6515 /* Make sure kthread is created before THREADED bit 6516 * is set. 6517 */ 6518 smp_mb__before_atomic(); 6519 6520 /* Setting/unsetting threaded mode on a napi might not immediately 6521 * take effect, if the current napi instance is actively being 6522 * polled. In this case, the switch between threaded mode and 6523 * softirq mode will happen in the next round of napi_schedule(). 6524 * This should not cause hiccups/stalls to the live traffic. 6525 */ 6526 list_for_each_entry(napi, &dev->napi_list, dev_list) 6527 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6528 6529 return err; 6530 } 6531 EXPORT_SYMBOL(dev_set_threaded); 6532 6533 /** 6534 * netif_queue_set_napi - Associate queue with the napi 6535 * @dev: device to which NAPI and queue belong 6536 * @queue_index: Index of queue 6537 * @type: queue type as RX or TX 6538 * @napi: NAPI context, pass NULL to clear previously set NAPI 6539 * 6540 * Set queue with its corresponding napi context. This should be done after 6541 * registering the NAPI handler for the queue-vector and the queues have been 6542 * mapped to the corresponding interrupt vector. 6543 */ 6544 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6545 enum netdev_queue_type type, struct napi_struct *napi) 6546 { 6547 struct netdev_rx_queue *rxq; 6548 struct netdev_queue *txq; 6549 6550 if (WARN_ON_ONCE(napi && !napi->dev)) 6551 return; 6552 if (dev->reg_state >= NETREG_REGISTERED) 6553 ASSERT_RTNL(); 6554 6555 switch (type) { 6556 case NETDEV_QUEUE_TYPE_RX: 6557 rxq = __netif_get_rx_queue(dev, queue_index); 6558 rxq->napi = napi; 6559 return; 6560 case NETDEV_QUEUE_TYPE_TX: 6561 txq = netdev_get_tx_queue(dev, queue_index); 6562 txq->napi = napi; 6563 return; 6564 default: 6565 return; 6566 } 6567 } 6568 EXPORT_SYMBOL(netif_queue_set_napi); 6569 6570 void netif_napi_add_weight(struct net_device *dev, struct napi_struct *napi, 6571 int (*poll)(struct napi_struct *, int), int weight) 6572 { 6573 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 6574 return; 6575 6576 INIT_LIST_HEAD(&napi->poll_list); 6577 INIT_HLIST_NODE(&napi->napi_hash_node); 6578 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 6579 napi->timer.function = napi_watchdog; 6580 init_gro_hash(napi); 6581 napi->skb = NULL; 6582 INIT_LIST_HEAD(&napi->rx_list); 6583 napi->rx_count = 0; 6584 napi->poll = poll; 6585 if (weight > NAPI_POLL_WEIGHT) 6586 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 6587 weight); 6588 napi->weight = weight; 6589 napi->dev = dev; 6590 #ifdef CONFIG_NETPOLL 6591 napi->poll_owner = -1; 6592 #endif 6593 napi->list_owner = -1; 6594 set_bit(NAPI_STATE_SCHED, &napi->state); 6595 set_bit(NAPI_STATE_NPSVC, &napi->state); 6596 list_add_rcu(&napi->dev_list, &dev->napi_list); 6597 napi_hash_add(napi); 6598 napi_get_frags_check(napi); 6599 /* Create kthread for this napi if dev->threaded is set. 6600 * Clear dev->threaded if kthread creation failed so that 6601 * threaded mode will not be enabled in napi_enable(). 6602 */ 6603 if (dev->threaded && napi_kthread_create(napi)) 6604 dev->threaded = 0; 6605 netif_napi_set_irq(napi, -1); 6606 } 6607 EXPORT_SYMBOL(netif_napi_add_weight); 6608 6609 void napi_disable(struct napi_struct *n) 6610 { 6611 unsigned long val, new; 6612 6613 might_sleep(); 6614 set_bit(NAPI_STATE_DISABLE, &n->state); 6615 6616 val = READ_ONCE(n->state); 6617 do { 6618 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 6619 usleep_range(20, 200); 6620 val = READ_ONCE(n->state); 6621 } 6622 6623 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 6624 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 6625 } while (!try_cmpxchg(&n->state, &val, new)); 6626 6627 hrtimer_cancel(&n->timer); 6628 6629 clear_bit(NAPI_STATE_DISABLE, &n->state); 6630 } 6631 EXPORT_SYMBOL(napi_disable); 6632 6633 /** 6634 * napi_enable - enable NAPI scheduling 6635 * @n: NAPI context 6636 * 6637 * Resume NAPI from being scheduled on this context. 6638 * Must be paired with napi_disable. 6639 */ 6640 void napi_enable(struct napi_struct *n) 6641 { 6642 unsigned long new, val = READ_ONCE(n->state); 6643 6644 do { 6645 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 6646 6647 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 6648 if (n->dev->threaded && n->thread) 6649 new |= NAPIF_STATE_THREADED; 6650 } while (!try_cmpxchg(&n->state, &val, new)); 6651 } 6652 EXPORT_SYMBOL(napi_enable); 6653 6654 static void flush_gro_hash(struct napi_struct *napi) 6655 { 6656 int i; 6657 6658 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6659 struct sk_buff *skb, *n; 6660 6661 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list) 6662 kfree_skb(skb); 6663 napi->gro_hash[i].count = 0; 6664 } 6665 } 6666 6667 /* Must be called in process context */ 6668 void __netif_napi_del(struct napi_struct *napi) 6669 { 6670 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 6671 return; 6672 6673 napi_hash_del(napi); 6674 list_del_rcu(&napi->dev_list); 6675 napi_free_frags(napi); 6676 6677 flush_gro_hash(napi); 6678 napi->gro_bitmask = 0; 6679 6680 if (napi->thread) { 6681 kthread_stop(napi->thread); 6682 napi->thread = NULL; 6683 } 6684 } 6685 EXPORT_SYMBOL(__netif_napi_del); 6686 6687 static int __napi_poll(struct napi_struct *n, bool *repoll) 6688 { 6689 int work, weight; 6690 6691 weight = n->weight; 6692 6693 /* This NAPI_STATE_SCHED test is for avoiding a race 6694 * with netpoll's poll_napi(). Only the entity which 6695 * obtains the lock and sees NAPI_STATE_SCHED set will 6696 * actually make the ->poll() call. Therefore we avoid 6697 * accidentally calling ->poll() when NAPI is not scheduled. 6698 */ 6699 work = 0; 6700 if (napi_is_scheduled(n)) { 6701 work = n->poll(n, weight); 6702 trace_napi_poll(n, work, weight); 6703 6704 xdp_do_check_flushed(n); 6705 } 6706 6707 if (unlikely(work > weight)) 6708 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 6709 n->poll, work, weight); 6710 6711 if (likely(work < weight)) 6712 return work; 6713 6714 /* Drivers must not modify the NAPI state if they 6715 * consume the entire weight. In such cases this code 6716 * still "owns" the NAPI instance and therefore can 6717 * move the instance around on the list at-will. 6718 */ 6719 if (unlikely(napi_disable_pending(n))) { 6720 napi_complete(n); 6721 return work; 6722 } 6723 6724 /* The NAPI context has more processing work, but busy-polling 6725 * is preferred. Exit early. 6726 */ 6727 if (napi_prefer_busy_poll(n)) { 6728 if (napi_complete_done(n, work)) { 6729 /* If timeout is not set, we need to make sure 6730 * that the NAPI is re-scheduled. 6731 */ 6732 napi_schedule(n); 6733 } 6734 return work; 6735 } 6736 6737 if (n->gro_bitmask) { 6738 /* flush too old packets 6739 * If HZ < 1000, flush all packets. 6740 */ 6741 napi_gro_flush(n, HZ >= 1000); 6742 } 6743 6744 gro_normal_list(n); 6745 6746 /* Some drivers may have called napi_schedule 6747 * prior to exhausting their budget. 6748 */ 6749 if (unlikely(!list_empty(&n->poll_list))) { 6750 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 6751 n->dev ? n->dev->name : "backlog"); 6752 return work; 6753 } 6754 6755 *repoll = true; 6756 6757 return work; 6758 } 6759 6760 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 6761 { 6762 bool do_repoll = false; 6763 void *have; 6764 int work; 6765 6766 list_del_init(&n->poll_list); 6767 6768 have = netpoll_poll_lock(n); 6769 6770 work = __napi_poll(n, &do_repoll); 6771 6772 if (do_repoll) 6773 list_add_tail(&n->poll_list, repoll); 6774 6775 netpoll_poll_unlock(have); 6776 6777 return work; 6778 } 6779 6780 static int napi_thread_wait(struct napi_struct *napi) 6781 { 6782 set_current_state(TASK_INTERRUPTIBLE); 6783 6784 while (!kthread_should_stop()) { 6785 /* Testing SCHED_THREADED bit here to make sure the current 6786 * kthread owns this napi and could poll on this napi. 6787 * Testing SCHED bit is not enough because SCHED bit might be 6788 * set by some other busy poll thread or by napi_disable(). 6789 */ 6790 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 6791 WARN_ON(!list_empty(&napi->poll_list)); 6792 __set_current_state(TASK_RUNNING); 6793 return 0; 6794 } 6795 6796 schedule(); 6797 set_current_state(TASK_INTERRUPTIBLE); 6798 } 6799 __set_current_state(TASK_RUNNING); 6800 6801 return -1; 6802 } 6803 6804 static void napi_threaded_poll_loop(struct napi_struct *napi) 6805 { 6806 struct softnet_data *sd; 6807 unsigned long last_qs = jiffies; 6808 6809 for (;;) { 6810 bool repoll = false; 6811 void *have; 6812 6813 local_bh_disable(); 6814 sd = this_cpu_ptr(&softnet_data); 6815 sd->in_napi_threaded_poll = true; 6816 6817 have = netpoll_poll_lock(napi); 6818 __napi_poll(napi, &repoll); 6819 netpoll_poll_unlock(have); 6820 6821 sd->in_napi_threaded_poll = false; 6822 barrier(); 6823 6824 if (sd_has_rps_ipi_waiting(sd)) { 6825 local_irq_disable(); 6826 net_rps_action_and_irq_enable(sd); 6827 } 6828 skb_defer_free_flush(sd); 6829 local_bh_enable(); 6830 6831 if (!repoll) 6832 break; 6833 6834 rcu_softirq_qs_periodic(last_qs); 6835 cond_resched(); 6836 } 6837 } 6838 6839 static int napi_threaded_poll(void *data) 6840 { 6841 struct napi_struct *napi = data; 6842 6843 while (!napi_thread_wait(napi)) 6844 napi_threaded_poll_loop(napi); 6845 6846 return 0; 6847 } 6848 6849 static __latent_entropy void net_rx_action(struct softirq_action *h) 6850 { 6851 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 6852 unsigned long time_limit = jiffies + 6853 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 6854 int budget = READ_ONCE(net_hotdata.netdev_budget); 6855 LIST_HEAD(list); 6856 LIST_HEAD(repoll); 6857 6858 start: 6859 sd->in_net_rx_action = true; 6860 local_irq_disable(); 6861 list_splice_init(&sd->poll_list, &list); 6862 local_irq_enable(); 6863 6864 for (;;) { 6865 struct napi_struct *n; 6866 6867 skb_defer_free_flush(sd); 6868 6869 if (list_empty(&list)) { 6870 if (list_empty(&repoll)) { 6871 sd->in_net_rx_action = false; 6872 barrier(); 6873 /* We need to check if ____napi_schedule() 6874 * had refilled poll_list while 6875 * sd->in_net_rx_action was true. 6876 */ 6877 if (!list_empty(&sd->poll_list)) 6878 goto start; 6879 if (!sd_has_rps_ipi_waiting(sd)) 6880 goto end; 6881 } 6882 break; 6883 } 6884 6885 n = list_first_entry(&list, struct napi_struct, poll_list); 6886 budget -= napi_poll(n, &repoll); 6887 6888 /* If softirq window is exhausted then punt. 6889 * Allow this to run for 2 jiffies since which will allow 6890 * an average latency of 1.5/HZ. 6891 */ 6892 if (unlikely(budget <= 0 || 6893 time_after_eq(jiffies, time_limit))) { 6894 sd->time_squeeze++; 6895 break; 6896 } 6897 } 6898 6899 local_irq_disable(); 6900 6901 list_splice_tail_init(&sd->poll_list, &list); 6902 list_splice_tail(&repoll, &list); 6903 list_splice(&list, &sd->poll_list); 6904 if (!list_empty(&sd->poll_list)) 6905 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 6906 else 6907 sd->in_net_rx_action = false; 6908 6909 net_rps_action_and_irq_enable(sd); 6910 end:; 6911 } 6912 6913 struct netdev_adjacent { 6914 struct net_device *dev; 6915 netdevice_tracker dev_tracker; 6916 6917 /* upper master flag, there can only be one master device per list */ 6918 bool master; 6919 6920 /* lookup ignore flag */ 6921 bool ignore; 6922 6923 /* counter for the number of times this device was added to us */ 6924 u16 ref_nr; 6925 6926 /* private field for the users */ 6927 void *private; 6928 6929 struct list_head list; 6930 struct rcu_head rcu; 6931 }; 6932 6933 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 6934 struct list_head *adj_list) 6935 { 6936 struct netdev_adjacent *adj; 6937 6938 list_for_each_entry(adj, adj_list, list) { 6939 if (adj->dev == adj_dev) 6940 return adj; 6941 } 6942 return NULL; 6943 } 6944 6945 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 6946 struct netdev_nested_priv *priv) 6947 { 6948 struct net_device *dev = (struct net_device *)priv->data; 6949 6950 return upper_dev == dev; 6951 } 6952 6953 /** 6954 * netdev_has_upper_dev - Check if device is linked to an upper device 6955 * @dev: device 6956 * @upper_dev: upper device to check 6957 * 6958 * Find out if a device is linked to specified upper device and return true 6959 * in case it is. Note that this checks only immediate upper device, 6960 * not through a complete stack of devices. The caller must hold the RTNL lock. 6961 */ 6962 bool netdev_has_upper_dev(struct net_device *dev, 6963 struct net_device *upper_dev) 6964 { 6965 struct netdev_nested_priv priv = { 6966 .data = (void *)upper_dev, 6967 }; 6968 6969 ASSERT_RTNL(); 6970 6971 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 6972 &priv); 6973 } 6974 EXPORT_SYMBOL(netdev_has_upper_dev); 6975 6976 /** 6977 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 6978 * @dev: device 6979 * @upper_dev: upper device to check 6980 * 6981 * Find out if a device is linked to specified upper device and return true 6982 * in case it is. Note that this checks the entire upper device chain. 6983 * The caller must hold rcu lock. 6984 */ 6985 6986 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 6987 struct net_device *upper_dev) 6988 { 6989 struct netdev_nested_priv priv = { 6990 .data = (void *)upper_dev, 6991 }; 6992 6993 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 6994 &priv); 6995 } 6996 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 6997 6998 /** 6999 * netdev_has_any_upper_dev - Check if device is linked to some device 7000 * @dev: device 7001 * 7002 * Find out if a device is linked to an upper device and return true in case 7003 * it is. The caller must hold the RTNL lock. 7004 */ 7005 bool netdev_has_any_upper_dev(struct net_device *dev) 7006 { 7007 ASSERT_RTNL(); 7008 7009 return !list_empty(&dev->adj_list.upper); 7010 } 7011 EXPORT_SYMBOL(netdev_has_any_upper_dev); 7012 7013 /** 7014 * netdev_master_upper_dev_get - Get master upper device 7015 * @dev: device 7016 * 7017 * Find a master upper device and return pointer to it or NULL in case 7018 * it's not there. The caller must hold the RTNL lock. 7019 */ 7020 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 7021 { 7022 struct netdev_adjacent *upper; 7023 7024 ASSERT_RTNL(); 7025 7026 if (list_empty(&dev->adj_list.upper)) 7027 return NULL; 7028 7029 upper = list_first_entry(&dev->adj_list.upper, 7030 struct netdev_adjacent, list); 7031 if (likely(upper->master)) 7032 return upper->dev; 7033 return NULL; 7034 } 7035 EXPORT_SYMBOL(netdev_master_upper_dev_get); 7036 7037 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 7038 { 7039 struct netdev_adjacent *upper; 7040 7041 ASSERT_RTNL(); 7042 7043 if (list_empty(&dev->adj_list.upper)) 7044 return NULL; 7045 7046 upper = list_first_entry(&dev->adj_list.upper, 7047 struct netdev_adjacent, list); 7048 if (likely(upper->master) && !upper->ignore) 7049 return upper->dev; 7050 return NULL; 7051 } 7052 7053 /** 7054 * netdev_has_any_lower_dev - Check if device is linked to some device 7055 * @dev: device 7056 * 7057 * Find out if a device is linked to a lower device and return true in case 7058 * it is. The caller must hold the RTNL lock. 7059 */ 7060 static bool netdev_has_any_lower_dev(struct net_device *dev) 7061 { 7062 ASSERT_RTNL(); 7063 7064 return !list_empty(&dev->adj_list.lower); 7065 } 7066 7067 void *netdev_adjacent_get_private(struct list_head *adj_list) 7068 { 7069 struct netdev_adjacent *adj; 7070 7071 adj = list_entry(adj_list, struct netdev_adjacent, list); 7072 7073 return adj->private; 7074 } 7075 EXPORT_SYMBOL(netdev_adjacent_get_private); 7076 7077 /** 7078 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7079 * @dev: device 7080 * @iter: list_head ** of the current position 7081 * 7082 * Gets the next device from the dev's upper list, starting from iter 7083 * position. The caller must hold RCU read lock. 7084 */ 7085 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7086 struct list_head **iter) 7087 { 7088 struct netdev_adjacent *upper; 7089 7090 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7091 7092 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7093 7094 if (&upper->list == &dev->adj_list.upper) 7095 return NULL; 7096 7097 *iter = &upper->list; 7098 7099 return upper->dev; 7100 } 7101 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7102 7103 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7104 struct list_head **iter, 7105 bool *ignore) 7106 { 7107 struct netdev_adjacent *upper; 7108 7109 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7110 7111 if (&upper->list == &dev->adj_list.upper) 7112 return NULL; 7113 7114 *iter = &upper->list; 7115 *ignore = upper->ignore; 7116 7117 return upper->dev; 7118 } 7119 7120 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7121 struct list_head **iter) 7122 { 7123 struct netdev_adjacent *upper; 7124 7125 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7126 7127 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7128 7129 if (&upper->list == &dev->adj_list.upper) 7130 return NULL; 7131 7132 *iter = &upper->list; 7133 7134 return upper->dev; 7135 } 7136 7137 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7138 int (*fn)(struct net_device *dev, 7139 struct netdev_nested_priv *priv), 7140 struct netdev_nested_priv *priv) 7141 { 7142 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7143 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7144 int ret, cur = 0; 7145 bool ignore; 7146 7147 now = dev; 7148 iter = &dev->adj_list.upper; 7149 7150 while (1) { 7151 if (now != dev) { 7152 ret = fn(now, priv); 7153 if (ret) 7154 return ret; 7155 } 7156 7157 next = NULL; 7158 while (1) { 7159 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7160 if (!udev) 7161 break; 7162 if (ignore) 7163 continue; 7164 7165 next = udev; 7166 niter = &udev->adj_list.upper; 7167 dev_stack[cur] = now; 7168 iter_stack[cur++] = iter; 7169 break; 7170 } 7171 7172 if (!next) { 7173 if (!cur) 7174 return 0; 7175 next = dev_stack[--cur]; 7176 niter = iter_stack[cur]; 7177 } 7178 7179 now = next; 7180 iter = niter; 7181 } 7182 7183 return 0; 7184 } 7185 7186 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7187 int (*fn)(struct net_device *dev, 7188 struct netdev_nested_priv *priv), 7189 struct netdev_nested_priv *priv) 7190 { 7191 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7192 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7193 int ret, cur = 0; 7194 7195 now = dev; 7196 iter = &dev->adj_list.upper; 7197 7198 while (1) { 7199 if (now != dev) { 7200 ret = fn(now, priv); 7201 if (ret) 7202 return ret; 7203 } 7204 7205 next = NULL; 7206 while (1) { 7207 udev = netdev_next_upper_dev_rcu(now, &iter); 7208 if (!udev) 7209 break; 7210 7211 next = udev; 7212 niter = &udev->adj_list.upper; 7213 dev_stack[cur] = now; 7214 iter_stack[cur++] = iter; 7215 break; 7216 } 7217 7218 if (!next) { 7219 if (!cur) 7220 return 0; 7221 next = dev_stack[--cur]; 7222 niter = iter_stack[cur]; 7223 } 7224 7225 now = next; 7226 iter = niter; 7227 } 7228 7229 return 0; 7230 } 7231 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7232 7233 static bool __netdev_has_upper_dev(struct net_device *dev, 7234 struct net_device *upper_dev) 7235 { 7236 struct netdev_nested_priv priv = { 7237 .flags = 0, 7238 .data = (void *)upper_dev, 7239 }; 7240 7241 ASSERT_RTNL(); 7242 7243 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7244 &priv); 7245 } 7246 7247 /** 7248 * netdev_lower_get_next_private - Get the next ->private from the 7249 * lower neighbour list 7250 * @dev: device 7251 * @iter: list_head ** of the current position 7252 * 7253 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7254 * list, starting from iter position. The caller must hold either hold the 7255 * RTNL lock or its own locking that guarantees that the neighbour lower 7256 * list will remain unchanged. 7257 */ 7258 void *netdev_lower_get_next_private(struct net_device *dev, 7259 struct list_head **iter) 7260 { 7261 struct netdev_adjacent *lower; 7262 7263 lower = list_entry(*iter, struct netdev_adjacent, list); 7264 7265 if (&lower->list == &dev->adj_list.lower) 7266 return NULL; 7267 7268 *iter = lower->list.next; 7269 7270 return lower->private; 7271 } 7272 EXPORT_SYMBOL(netdev_lower_get_next_private); 7273 7274 /** 7275 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7276 * lower neighbour list, RCU 7277 * variant 7278 * @dev: device 7279 * @iter: list_head ** of the current position 7280 * 7281 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7282 * list, starting from iter position. The caller must hold RCU read lock. 7283 */ 7284 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7285 struct list_head **iter) 7286 { 7287 struct netdev_adjacent *lower; 7288 7289 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7290 7291 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7292 7293 if (&lower->list == &dev->adj_list.lower) 7294 return NULL; 7295 7296 *iter = &lower->list; 7297 7298 return lower->private; 7299 } 7300 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7301 7302 /** 7303 * netdev_lower_get_next - Get the next device from the lower neighbour 7304 * list 7305 * @dev: device 7306 * @iter: list_head ** of the current position 7307 * 7308 * Gets the next netdev_adjacent from the dev's lower neighbour 7309 * list, starting from iter position. The caller must hold RTNL lock or 7310 * its own locking that guarantees that the neighbour lower 7311 * list will remain unchanged. 7312 */ 7313 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7314 { 7315 struct netdev_adjacent *lower; 7316 7317 lower = list_entry(*iter, struct netdev_adjacent, list); 7318 7319 if (&lower->list == &dev->adj_list.lower) 7320 return NULL; 7321 7322 *iter = lower->list.next; 7323 7324 return lower->dev; 7325 } 7326 EXPORT_SYMBOL(netdev_lower_get_next); 7327 7328 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 7329 struct list_head **iter) 7330 { 7331 struct netdev_adjacent *lower; 7332 7333 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7334 7335 if (&lower->list == &dev->adj_list.lower) 7336 return NULL; 7337 7338 *iter = &lower->list; 7339 7340 return lower->dev; 7341 } 7342 7343 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 7344 struct list_head **iter, 7345 bool *ignore) 7346 { 7347 struct netdev_adjacent *lower; 7348 7349 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7350 7351 if (&lower->list == &dev->adj_list.lower) 7352 return NULL; 7353 7354 *iter = &lower->list; 7355 *ignore = lower->ignore; 7356 7357 return lower->dev; 7358 } 7359 7360 int netdev_walk_all_lower_dev(struct net_device *dev, 7361 int (*fn)(struct net_device *dev, 7362 struct netdev_nested_priv *priv), 7363 struct netdev_nested_priv *priv) 7364 { 7365 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7366 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7367 int ret, cur = 0; 7368 7369 now = dev; 7370 iter = &dev->adj_list.lower; 7371 7372 while (1) { 7373 if (now != dev) { 7374 ret = fn(now, priv); 7375 if (ret) 7376 return ret; 7377 } 7378 7379 next = NULL; 7380 while (1) { 7381 ldev = netdev_next_lower_dev(now, &iter); 7382 if (!ldev) 7383 break; 7384 7385 next = ldev; 7386 niter = &ldev->adj_list.lower; 7387 dev_stack[cur] = now; 7388 iter_stack[cur++] = iter; 7389 break; 7390 } 7391 7392 if (!next) { 7393 if (!cur) 7394 return 0; 7395 next = dev_stack[--cur]; 7396 niter = iter_stack[cur]; 7397 } 7398 7399 now = next; 7400 iter = niter; 7401 } 7402 7403 return 0; 7404 } 7405 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 7406 7407 static int __netdev_walk_all_lower_dev(struct net_device *dev, 7408 int (*fn)(struct net_device *dev, 7409 struct netdev_nested_priv *priv), 7410 struct netdev_nested_priv *priv) 7411 { 7412 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7413 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7414 int ret, cur = 0; 7415 bool ignore; 7416 7417 now = dev; 7418 iter = &dev->adj_list.lower; 7419 7420 while (1) { 7421 if (now != dev) { 7422 ret = fn(now, priv); 7423 if (ret) 7424 return ret; 7425 } 7426 7427 next = NULL; 7428 while (1) { 7429 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 7430 if (!ldev) 7431 break; 7432 if (ignore) 7433 continue; 7434 7435 next = ldev; 7436 niter = &ldev->adj_list.lower; 7437 dev_stack[cur] = now; 7438 iter_stack[cur++] = iter; 7439 break; 7440 } 7441 7442 if (!next) { 7443 if (!cur) 7444 return 0; 7445 next = dev_stack[--cur]; 7446 niter = iter_stack[cur]; 7447 } 7448 7449 now = next; 7450 iter = niter; 7451 } 7452 7453 return 0; 7454 } 7455 7456 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 7457 struct list_head **iter) 7458 { 7459 struct netdev_adjacent *lower; 7460 7461 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7462 if (&lower->list == &dev->adj_list.lower) 7463 return NULL; 7464 7465 *iter = &lower->list; 7466 7467 return lower->dev; 7468 } 7469 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 7470 7471 static u8 __netdev_upper_depth(struct net_device *dev) 7472 { 7473 struct net_device *udev; 7474 struct list_head *iter; 7475 u8 max_depth = 0; 7476 bool ignore; 7477 7478 for (iter = &dev->adj_list.upper, 7479 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 7480 udev; 7481 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 7482 if (ignore) 7483 continue; 7484 if (max_depth < udev->upper_level) 7485 max_depth = udev->upper_level; 7486 } 7487 7488 return max_depth; 7489 } 7490 7491 static u8 __netdev_lower_depth(struct net_device *dev) 7492 { 7493 struct net_device *ldev; 7494 struct list_head *iter; 7495 u8 max_depth = 0; 7496 bool ignore; 7497 7498 for (iter = &dev->adj_list.lower, 7499 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 7500 ldev; 7501 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 7502 if (ignore) 7503 continue; 7504 if (max_depth < ldev->lower_level) 7505 max_depth = ldev->lower_level; 7506 } 7507 7508 return max_depth; 7509 } 7510 7511 static int __netdev_update_upper_level(struct net_device *dev, 7512 struct netdev_nested_priv *__unused) 7513 { 7514 dev->upper_level = __netdev_upper_depth(dev) + 1; 7515 return 0; 7516 } 7517 7518 #ifdef CONFIG_LOCKDEP 7519 static LIST_HEAD(net_unlink_list); 7520 7521 static void net_unlink_todo(struct net_device *dev) 7522 { 7523 if (list_empty(&dev->unlink_list)) 7524 list_add_tail(&dev->unlink_list, &net_unlink_list); 7525 } 7526 #endif 7527 7528 static int __netdev_update_lower_level(struct net_device *dev, 7529 struct netdev_nested_priv *priv) 7530 { 7531 dev->lower_level = __netdev_lower_depth(dev) + 1; 7532 7533 #ifdef CONFIG_LOCKDEP 7534 if (!priv) 7535 return 0; 7536 7537 if (priv->flags & NESTED_SYNC_IMM) 7538 dev->nested_level = dev->lower_level - 1; 7539 if (priv->flags & NESTED_SYNC_TODO) 7540 net_unlink_todo(dev); 7541 #endif 7542 return 0; 7543 } 7544 7545 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 7546 int (*fn)(struct net_device *dev, 7547 struct netdev_nested_priv *priv), 7548 struct netdev_nested_priv *priv) 7549 { 7550 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7551 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7552 int ret, cur = 0; 7553 7554 now = dev; 7555 iter = &dev->adj_list.lower; 7556 7557 while (1) { 7558 if (now != dev) { 7559 ret = fn(now, priv); 7560 if (ret) 7561 return ret; 7562 } 7563 7564 next = NULL; 7565 while (1) { 7566 ldev = netdev_next_lower_dev_rcu(now, &iter); 7567 if (!ldev) 7568 break; 7569 7570 next = ldev; 7571 niter = &ldev->adj_list.lower; 7572 dev_stack[cur] = now; 7573 iter_stack[cur++] = iter; 7574 break; 7575 } 7576 7577 if (!next) { 7578 if (!cur) 7579 return 0; 7580 next = dev_stack[--cur]; 7581 niter = iter_stack[cur]; 7582 } 7583 7584 now = next; 7585 iter = niter; 7586 } 7587 7588 return 0; 7589 } 7590 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 7591 7592 /** 7593 * netdev_lower_get_first_private_rcu - Get the first ->private from the 7594 * lower neighbour list, RCU 7595 * variant 7596 * @dev: device 7597 * 7598 * Gets the first netdev_adjacent->private from the dev's lower neighbour 7599 * list. The caller must hold RCU read lock. 7600 */ 7601 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 7602 { 7603 struct netdev_adjacent *lower; 7604 7605 lower = list_first_or_null_rcu(&dev->adj_list.lower, 7606 struct netdev_adjacent, list); 7607 if (lower) 7608 return lower->private; 7609 return NULL; 7610 } 7611 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 7612 7613 /** 7614 * netdev_master_upper_dev_get_rcu - Get master upper device 7615 * @dev: device 7616 * 7617 * Find a master upper device and return pointer to it or NULL in case 7618 * it's not there. The caller must hold the RCU read lock. 7619 */ 7620 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 7621 { 7622 struct netdev_adjacent *upper; 7623 7624 upper = list_first_or_null_rcu(&dev->adj_list.upper, 7625 struct netdev_adjacent, list); 7626 if (upper && likely(upper->master)) 7627 return upper->dev; 7628 return NULL; 7629 } 7630 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 7631 7632 static int netdev_adjacent_sysfs_add(struct net_device *dev, 7633 struct net_device *adj_dev, 7634 struct list_head *dev_list) 7635 { 7636 char linkname[IFNAMSIZ+7]; 7637 7638 sprintf(linkname, dev_list == &dev->adj_list.upper ? 7639 "upper_%s" : "lower_%s", adj_dev->name); 7640 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 7641 linkname); 7642 } 7643 static void netdev_adjacent_sysfs_del(struct net_device *dev, 7644 char *name, 7645 struct list_head *dev_list) 7646 { 7647 char linkname[IFNAMSIZ+7]; 7648 7649 sprintf(linkname, dev_list == &dev->adj_list.upper ? 7650 "upper_%s" : "lower_%s", name); 7651 sysfs_remove_link(&(dev->dev.kobj), linkname); 7652 } 7653 7654 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 7655 struct net_device *adj_dev, 7656 struct list_head *dev_list) 7657 { 7658 return (dev_list == &dev->adj_list.upper || 7659 dev_list == &dev->adj_list.lower) && 7660 net_eq(dev_net(dev), dev_net(adj_dev)); 7661 } 7662 7663 static int __netdev_adjacent_dev_insert(struct net_device *dev, 7664 struct net_device *adj_dev, 7665 struct list_head *dev_list, 7666 void *private, bool master) 7667 { 7668 struct netdev_adjacent *adj; 7669 int ret; 7670 7671 adj = __netdev_find_adj(adj_dev, dev_list); 7672 7673 if (adj) { 7674 adj->ref_nr += 1; 7675 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 7676 dev->name, adj_dev->name, adj->ref_nr); 7677 7678 return 0; 7679 } 7680 7681 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 7682 if (!adj) 7683 return -ENOMEM; 7684 7685 adj->dev = adj_dev; 7686 adj->master = master; 7687 adj->ref_nr = 1; 7688 adj->private = private; 7689 adj->ignore = false; 7690 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 7691 7692 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 7693 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 7694 7695 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 7696 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 7697 if (ret) 7698 goto free_adj; 7699 } 7700 7701 /* Ensure that master link is always the first item in list. */ 7702 if (master) { 7703 ret = sysfs_create_link(&(dev->dev.kobj), 7704 &(adj_dev->dev.kobj), "master"); 7705 if (ret) 7706 goto remove_symlinks; 7707 7708 list_add_rcu(&adj->list, dev_list); 7709 } else { 7710 list_add_tail_rcu(&adj->list, dev_list); 7711 } 7712 7713 return 0; 7714 7715 remove_symlinks: 7716 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 7717 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 7718 free_adj: 7719 netdev_put(adj_dev, &adj->dev_tracker); 7720 kfree(adj); 7721 7722 return ret; 7723 } 7724 7725 static void __netdev_adjacent_dev_remove(struct net_device *dev, 7726 struct net_device *adj_dev, 7727 u16 ref_nr, 7728 struct list_head *dev_list) 7729 { 7730 struct netdev_adjacent *adj; 7731 7732 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 7733 dev->name, adj_dev->name, ref_nr); 7734 7735 adj = __netdev_find_adj(adj_dev, dev_list); 7736 7737 if (!adj) { 7738 pr_err("Adjacency does not exist for device %s from %s\n", 7739 dev->name, adj_dev->name); 7740 WARN_ON(1); 7741 return; 7742 } 7743 7744 if (adj->ref_nr > ref_nr) { 7745 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 7746 dev->name, adj_dev->name, ref_nr, 7747 adj->ref_nr - ref_nr); 7748 adj->ref_nr -= ref_nr; 7749 return; 7750 } 7751 7752 if (adj->master) 7753 sysfs_remove_link(&(dev->dev.kobj), "master"); 7754 7755 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 7756 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 7757 7758 list_del_rcu(&adj->list); 7759 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 7760 adj_dev->name, dev->name, adj_dev->name); 7761 netdev_put(adj_dev, &adj->dev_tracker); 7762 kfree_rcu(adj, rcu); 7763 } 7764 7765 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 7766 struct net_device *upper_dev, 7767 struct list_head *up_list, 7768 struct list_head *down_list, 7769 void *private, bool master) 7770 { 7771 int ret; 7772 7773 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 7774 private, master); 7775 if (ret) 7776 return ret; 7777 7778 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 7779 private, false); 7780 if (ret) { 7781 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 7782 return ret; 7783 } 7784 7785 return 0; 7786 } 7787 7788 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 7789 struct net_device *upper_dev, 7790 u16 ref_nr, 7791 struct list_head *up_list, 7792 struct list_head *down_list) 7793 { 7794 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 7795 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 7796 } 7797 7798 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 7799 struct net_device *upper_dev, 7800 void *private, bool master) 7801 { 7802 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 7803 &dev->adj_list.upper, 7804 &upper_dev->adj_list.lower, 7805 private, master); 7806 } 7807 7808 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 7809 struct net_device *upper_dev) 7810 { 7811 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 7812 &dev->adj_list.upper, 7813 &upper_dev->adj_list.lower); 7814 } 7815 7816 static int __netdev_upper_dev_link(struct net_device *dev, 7817 struct net_device *upper_dev, bool master, 7818 void *upper_priv, void *upper_info, 7819 struct netdev_nested_priv *priv, 7820 struct netlink_ext_ack *extack) 7821 { 7822 struct netdev_notifier_changeupper_info changeupper_info = { 7823 .info = { 7824 .dev = dev, 7825 .extack = extack, 7826 }, 7827 .upper_dev = upper_dev, 7828 .master = master, 7829 .linking = true, 7830 .upper_info = upper_info, 7831 }; 7832 struct net_device *master_dev; 7833 int ret = 0; 7834 7835 ASSERT_RTNL(); 7836 7837 if (dev == upper_dev) 7838 return -EBUSY; 7839 7840 /* To prevent loops, check if dev is not upper device to upper_dev. */ 7841 if (__netdev_has_upper_dev(upper_dev, dev)) 7842 return -EBUSY; 7843 7844 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 7845 return -EMLINK; 7846 7847 if (!master) { 7848 if (__netdev_has_upper_dev(dev, upper_dev)) 7849 return -EEXIST; 7850 } else { 7851 master_dev = __netdev_master_upper_dev_get(dev); 7852 if (master_dev) 7853 return master_dev == upper_dev ? -EEXIST : -EBUSY; 7854 } 7855 7856 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 7857 &changeupper_info.info); 7858 ret = notifier_to_errno(ret); 7859 if (ret) 7860 return ret; 7861 7862 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 7863 master); 7864 if (ret) 7865 return ret; 7866 7867 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 7868 &changeupper_info.info); 7869 ret = notifier_to_errno(ret); 7870 if (ret) 7871 goto rollback; 7872 7873 __netdev_update_upper_level(dev, NULL); 7874 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 7875 7876 __netdev_update_lower_level(upper_dev, priv); 7877 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 7878 priv); 7879 7880 return 0; 7881 7882 rollback: 7883 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 7884 7885 return ret; 7886 } 7887 7888 /** 7889 * netdev_upper_dev_link - Add a link to the upper device 7890 * @dev: device 7891 * @upper_dev: new upper device 7892 * @extack: netlink extended ack 7893 * 7894 * Adds a link to device which is upper to this one. The caller must hold 7895 * the RTNL lock. On a failure a negative errno code is returned. 7896 * On success the reference counts are adjusted and the function 7897 * returns zero. 7898 */ 7899 int netdev_upper_dev_link(struct net_device *dev, 7900 struct net_device *upper_dev, 7901 struct netlink_ext_ack *extack) 7902 { 7903 struct netdev_nested_priv priv = { 7904 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 7905 .data = NULL, 7906 }; 7907 7908 return __netdev_upper_dev_link(dev, upper_dev, false, 7909 NULL, NULL, &priv, extack); 7910 } 7911 EXPORT_SYMBOL(netdev_upper_dev_link); 7912 7913 /** 7914 * netdev_master_upper_dev_link - Add a master link to the upper device 7915 * @dev: device 7916 * @upper_dev: new upper device 7917 * @upper_priv: upper device private 7918 * @upper_info: upper info to be passed down via notifier 7919 * @extack: netlink extended ack 7920 * 7921 * Adds a link to device which is upper to this one. In this case, only 7922 * one master upper device can be linked, although other non-master devices 7923 * might be linked as well. The caller must hold the RTNL lock. 7924 * On a failure a negative errno code is returned. On success the reference 7925 * counts are adjusted and the function returns zero. 7926 */ 7927 int netdev_master_upper_dev_link(struct net_device *dev, 7928 struct net_device *upper_dev, 7929 void *upper_priv, void *upper_info, 7930 struct netlink_ext_ack *extack) 7931 { 7932 struct netdev_nested_priv priv = { 7933 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 7934 .data = NULL, 7935 }; 7936 7937 return __netdev_upper_dev_link(dev, upper_dev, true, 7938 upper_priv, upper_info, &priv, extack); 7939 } 7940 EXPORT_SYMBOL(netdev_master_upper_dev_link); 7941 7942 static void __netdev_upper_dev_unlink(struct net_device *dev, 7943 struct net_device *upper_dev, 7944 struct netdev_nested_priv *priv) 7945 { 7946 struct netdev_notifier_changeupper_info changeupper_info = { 7947 .info = { 7948 .dev = dev, 7949 }, 7950 .upper_dev = upper_dev, 7951 .linking = false, 7952 }; 7953 7954 ASSERT_RTNL(); 7955 7956 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 7957 7958 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 7959 &changeupper_info.info); 7960 7961 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 7962 7963 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 7964 &changeupper_info.info); 7965 7966 __netdev_update_upper_level(dev, NULL); 7967 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 7968 7969 __netdev_update_lower_level(upper_dev, priv); 7970 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 7971 priv); 7972 } 7973 7974 /** 7975 * netdev_upper_dev_unlink - Removes a link to upper device 7976 * @dev: device 7977 * @upper_dev: new upper device 7978 * 7979 * Removes a link to device which is upper to this one. The caller must hold 7980 * the RTNL lock. 7981 */ 7982 void netdev_upper_dev_unlink(struct net_device *dev, 7983 struct net_device *upper_dev) 7984 { 7985 struct netdev_nested_priv priv = { 7986 .flags = NESTED_SYNC_TODO, 7987 .data = NULL, 7988 }; 7989 7990 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 7991 } 7992 EXPORT_SYMBOL(netdev_upper_dev_unlink); 7993 7994 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 7995 struct net_device *lower_dev, 7996 bool val) 7997 { 7998 struct netdev_adjacent *adj; 7999 8000 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 8001 if (adj) 8002 adj->ignore = val; 8003 8004 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 8005 if (adj) 8006 adj->ignore = val; 8007 } 8008 8009 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 8010 struct net_device *lower_dev) 8011 { 8012 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 8013 } 8014 8015 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 8016 struct net_device *lower_dev) 8017 { 8018 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 8019 } 8020 8021 int netdev_adjacent_change_prepare(struct net_device *old_dev, 8022 struct net_device *new_dev, 8023 struct net_device *dev, 8024 struct netlink_ext_ack *extack) 8025 { 8026 struct netdev_nested_priv priv = { 8027 .flags = 0, 8028 .data = NULL, 8029 }; 8030 int err; 8031 8032 if (!new_dev) 8033 return 0; 8034 8035 if (old_dev && new_dev != old_dev) 8036 netdev_adjacent_dev_disable(dev, old_dev); 8037 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 8038 extack); 8039 if (err) { 8040 if (old_dev && new_dev != old_dev) 8041 netdev_adjacent_dev_enable(dev, old_dev); 8042 return err; 8043 } 8044 8045 return 0; 8046 } 8047 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 8048 8049 void netdev_adjacent_change_commit(struct net_device *old_dev, 8050 struct net_device *new_dev, 8051 struct net_device *dev) 8052 { 8053 struct netdev_nested_priv priv = { 8054 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8055 .data = NULL, 8056 }; 8057 8058 if (!new_dev || !old_dev) 8059 return; 8060 8061 if (new_dev == old_dev) 8062 return; 8063 8064 netdev_adjacent_dev_enable(dev, old_dev); 8065 __netdev_upper_dev_unlink(old_dev, dev, &priv); 8066 } 8067 EXPORT_SYMBOL(netdev_adjacent_change_commit); 8068 8069 void netdev_adjacent_change_abort(struct net_device *old_dev, 8070 struct net_device *new_dev, 8071 struct net_device *dev) 8072 { 8073 struct netdev_nested_priv priv = { 8074 .flags = 0, 8075 .data = NULL, 8076 }; 8077 8078 if (!new_dev) 8079 return; 8080 8081 if (old_dev && new_dev != old_dev) 8082 netdev_adjacent_dev_enable(dev, old_dev); 8083 8084 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8085 } 8086 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8087 8088 /** 8089 * netdev_bonding_info_change - Dispatch event about slave change 8090 * @dev: device 8091 * @bonding_info: info to dispatch 8092 * 8093 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8094 * The caller must hold the RTNL lock. 8095 */ 8096 void netdev_bonding_info_change(struct net_device *dev, 8097 struct netdev_bonding_info *bonding_info) 8098 { 8099 struct netdev_notifier_bonding_info info = { 8100 .info.dev = dev, 8101 }; 8102 8103 memcpy(&info.bonding_info, bonding_info, 8104 sizeof(struct netdev_bonding_info)); 8105 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8106 &info.info); 8107 } 8108 EXPORT_SYMBOL(netdev_bonding_info_change); 8109 8110 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8111 struct netlink_ext_ack *extack) 8112 { 8113 struct netdev_notifier_offload_xstats_info info = { 8114 .info.dev = dev, 8115 .info.extack = extack, 8116 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8117 }; 8118 int err; 8119 int rc; 8120 8121 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8122 GFP_KERNEL); 8123 if (!dev->offload_xstats_l3) 8124 return -ENOMEM; 8125 8126 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8127 NETDEV_OFFLOAD_XSTATS_DISABLE, 8128 &info.info); 8129 err = notifier_to_errno(rc); 8130 if (err) 8131 goto free_stats; 8132 8133 return 0; 8134 8135 free_stats: 8136 kfree(dev->offload_xstats_l3); 8137 dev->offload_xstats_l3 = NULL; 8138 return err; 8139 } 8140 8141 int netdev_offload_xstats_enable(struct net_device *dev, 8142 enum netdev_offload_xstats_type type, 8143 struct netlink_ext_ack *extack) 8144 { 8145 ASSERT_RTNL(); 8146 8147 if (netdev_offload_xstats_enabled(dev, type)) 8148 return -EALREADY; 8149 8150 switch (type) { 8151 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8152 return netdev_offload_xstats_enable_l3(dev, extack); 8153 } 8154 8155 WARN_ON(1); 8156 return -EINVAL; 8157 } 8158 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8159 8160 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8161 { 8162 struct netdev_notifier_offload_xstats_info info = { 8163 .info.dev = dev, 8164 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8165 }; 8166 8167 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8168 &info.info); 8169 kfree(dev->offload_xstats_l3); 8170 dev->offload_xstats_l3 = NULL; 8171 } 8172 8173 int netdev_offload_xstats_disable(struct net_device *dev, 8174 enum netdev_offload_xstats_type type) 8175 { 8176 ASSERT_RTNL(); 8177 8178 if (!netdev_offload_xstats_enabled(dev, type)) 8179 return -EALREADY; 8180 8181 switch (type) { 8182 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8183 netdev_offload_xstats_disable_l3(dev); 8184 return 0; 8185 } 8186 8187 WARN_ON(1); 8188 return -EINVAL; 8189 } 8190 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8191 8192 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8193 { 8194 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8195 } 8196 8197 static struct rtnl_hw_stats64 * 8198 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8199 enum netdev_offload_xstats_type type) 8200 { 8201 switch (type) { 8202 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8203 return dev->offload_xstats_l3; 8204 } 8205 8206 WARN_ON(1); 8207 return NULL; 8208 } 8209 8210 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8211 enum netdev_offload_xstats_type type) 8212 { 8213 ASSERT_RTNL(); 8214 8215 return netdev_offload_xstats_get_ptr(dev, type); 8216 } 8217 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8218 8219 struct netdev_notifier_offload_xstats_ru { 8220 bool used; 8221 }; 8222 8223 struct netdev_notifier_offload_xstats_rd { 8224 struct rtnl_hw_stats64 stats; 8225 bool used; 8226 }; 8227 8228 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8229 const struct rtnl_hw_stats64 *src) 8230 { 8231 dest->rx_packets += src->rx_packets; 8232 dest->tx_packets += src->tx_packets; 8233 dest->rx_bytes += src->rx_bytes; 8234 dest->tx_bytes += src->tx_bytes; 8235 dest->rx_errors += src->rx_errors; 8236 dest->tx_errors += src->tx_errors; 8237 dest->rx_dropped += src->rx_dropped; 8238 dest->tx_dropped += src->tx_dropped; 8239 dest->multicast += src->multicast; 8240 } 8241 8242 static int netdev_offload_xstats_get_used(struct net_device *dev, 8243 enum netdev_offload_xstats_type type, 8244 bool *p_used, 8245 struct netlink_ext_ack *extack) 8246 { 8247 struct netdev_notifier_offload_xstats_ru report_used = {}; 8248 struct netdev_notifier_offload_xstats_info info = { 8249 .info.dev = dev, 8250 .info.extack = extack, 8251 .type = type, 8252 .report_used = &report_used, 8253 }; 8254 int rc; 8255 8256 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8257 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8258 &info.info); 8259 *p_used = report_used.used; 8260 return notifier_to_errno(rc); 8261 } 8262 8263 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8264 enum netdev_offload_xstats_type type, 8265 struct rtnl_hw_stats64 *p_stats, 8266 bool *p_used, 8267 struct netlink_ext_ack *extack) 8268 { 8269 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8270 struct netdev_notifier_offload_xstats_info info = { 8271 .info.dev = dev, 8272 .info.extack = extack, 8273 .type = type, 8274 .report_delta = &report_delta, 8275 }; 8276 struct rtnl_hw_stats64 *stats; 8277 int rc; 8278 8279 stats = netdev_offload_xstats_get_ptr(dev, type); 8280 if (WARN_ON(!stats)) 8281 return -EINVAL; 8282 8283 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8284 &info.info); 8285 8286 /* Cache whatever we got, even if there was an error, otherwise the 8287 * successful stats retrievals would get lost. 8288 */ 8289 netdev_hw_stats64_add(stats, &report_delta.stats); 8290 8291 if (p_stats) 8292 *p_stats = *stats; 8293 *p_used = report_delta.used; 8294 8295 return notifier_to_errno(rc); 8296 } 8297 8298 int netdev_offload_xstats_get(struct net_device *dev, 8299 enum netdev_offload_xstats_type type, 8300 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8301 struct netlink_ext_ack *extack) 8302 { 8303 ASSERT_RTNL(); 8304 8305 if (p_stats) 8306 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8307 p_used, extack); 8308 else 8309 return netdev_offload_xstats_get_used(dev, type, p_used, 8310 extack); 8311 } 8312 EXPORT_SYMBOL(netdev_offload_xstats_get); 8313 8314 void 8315 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8316 const struct rtnl_hw_stats64 *stats) 8317 { 8318 report_delta->used = true; 8319 netdev_hw_stats64_add(&report_delta->stats, stats); 8320 } 8321 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 8322 8323 void 8324 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 8325 { 8326 report_used->used = true; 8327 } 8328 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 8329 8330 void netdev_offload_xstats_push_delta(struct net_device *dev, 8331 enum netdev_offload_xstats_type type, 8332 const struct rtnl_hw_stats64 *p_stats) 8333 { 8334 struct rtnl_hw_stats64 *stats; 8335 8336 ASSERT_RTNL(); 8337 8338 stats = netdev_offload_xstats_get_ptr(dev, type); 8339 if (WARN_ON(!stats)) 8340 return; 8341 8342 netdev_hw_stats64_add(stats, p_stats); 8343 } 8344 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 8345 8346 /** 8347 * netdev_get_xmit_slave - Get the xmit slave of master device 8348 * @dev: device 8349 * @skb: The packet 8350 * @all_slaves: assume all the slaves are active 8351 * 8352 * The reference counters are not incremented so the caller must be 8353 * careful with locks. The caller must hold RCU lock. 8354 * %NULL is returned if no slave is found. 8355 */ 8356 8357 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 8358 struct sk_buff *skb, 8359 bool all_slaves) 8360 { 8361 const struct net_device_ops *ops = dev->netdev_ops; 8362 8363 if (!ops->ndo_get_xmit_slave) 8364 return NULL; 8365 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 8366 } 8367 EXPORT_SYMBOL(netdev_get_xmit_slave); 8368 8369 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 8370 struct sock *sk) 8371 { 8372 const struct net_device_ops *ops = dev->netdev_ops; 8373 8374 if (!ops->ndo_sk_get_lower_dev) 8375 return NULL; 8376 return ops->ndo_sk_get_lower_dev(dev, sk); 8377 } 8378 8379 /** 8380 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 8381 * @dev: device 8382 * @sk: the socket 8383 * 8384 * %NULL is returned if no lower device is found. 8385 */ 8386 8387 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 8388 struct sock *sk) 8389 { 8390 struct net_device *lower; 8391 8392 lower = netdev_sk_get_lower_dev(dev, sk); 8393 while (lower) { 8394 dev = lower; 8395 lower = netdev_sk_get_lower_dev(dev, sk); 8396 } 8397 8398 return dev; 8399 } 8400 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 8401 8402 static void netdev_adjacent_add_links(struct net_device *dev) 8403 { 8404 struct netdev_adjacent *iter; 8405 8406 struct net *net = dev_net(dev); 8407 8408 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8409 if (!net_eq(net, dev_net(iter->dev))) 8410 continue; 8411 netdev_adjacent_sysfs_add(iter->dev, dev, 8412 &iter->dev->adj_list.lower); 8413 netdev_adjacent_sysfs_add(dev, iter->dev, 8414 &dev->adj_list.upper); 8415 } 8416 8417 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8418 if (!net_eq(net, dev_net(iter->dev))) 8419 continue; 8420 netdev_adjacent_sysfs_add(iter->dev, dev, 8421 &iter->dev->adj_list.upper); 8422 netdev_adjacent_sysfs_add(dev, iter->dev, 8423 &dev->adj_list.lower); 8424 } 8425 } 8426 8427 static void netdev_adjacent_del_links(struct net_device *dev) 8428 { 8429 struct netdev_adjacent *iter; 8430 8431 struct net *net = dev_net(dev); 8432 8433 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8434 if (!net_eq(net, dev_net(iter->dev))) 8435 continue; 8436 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8437 &iter->dev->adj_list.lower); 8438 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8439 &dev->adj_list.upper); 8440 } 8441 8442 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8443 if (!net_eq(net, dev_net(iter->dev))) 8444 continue; 8445 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8446 &iter->dev->adj_list.upper); 8447 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8448 &dev->adj_list.lower); 8449 } 8450 } 8451 8452 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 8453 { 8454 struct netdev_adjacent *iter; 8455 8456 struct net *net = dev_net(dev); 8457 8458 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8459 if (!net_eq(net, dev_net(iter->dev))) 8460 continue; 8461 netdev_adjacent_sysfs_del(iter->dev, oldname, 8462 &iter->dev->adj_list.lower); 8463 netdev_adjacent_sysfs_add(iter->dev, dev, 8464 &iter->dev->adj_list.lower); 8465 } 8466 8467 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8468 if (!net_eq(net, dev_net(iter->dev))) 8469 continue; 8470 netdev_adjacent_sysfs_del(iter->dev, oldname, 8471 &iter->dev->adj_list.upper); 8472 netdev_adjacent_sysfs_add(iter->dev, dev, 8473 &iter->dev->adj_list.upper); 8474 } 8475 } 8476 8477 void *netdev_lower_dev_get_private(struct net_device *dev, 8478 struct net_device *lower_dev) 8479 { 8480 struct netdev_adjacent *lower; 8481 8482 if (!lower_dev) 8483 return NULL; 8484 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 8485 if (!lower) 8486 return NULL; 8487 8488 return lower->private; 8489 } 8490 EXPORT_SYMBOL(netdev_lower_dev_get_private); 8491 8492 8493 /** 8494 * netdev_lower_state_changed - Dispatch event about lower device state change 8495 * @lower_dev: device 8496 * @lower_state_info: state to dispatch 8497 * 8498 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 8499 * The caller must hold the RTNL lock. 8500 */ 8501 void netdev_lower_state_changed(struct net_device *lower_dev, 8502 void *lower_state_info) 8503 { 8504 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 8505 .info.dev = lower_dev, 8506 }; 8507 8508 ASSERT_RTNL(); 8509 changelowerstate_info.lower_state_info = lower_state_info; 8510 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 8511 &changelowerstate_info.info); 8512 } 8513 EXPORT_SYMBOL(netdev_lower_state_changed); 8514 8515 static void dev_change_rx_flags(struct net_device *dev, int flags) 8516 { 8517 const struct net_device_ops *ops = dev->netdev_ops; 8518 8519 if (ops->ndo_change_rx_flags) 8520 ops->ndo_change_rx_flags(dev, flags); 8521 } 8522 8523 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 8524 { 8525 unsigned int old_flags = dev->flags; 8526 kuid_t uid; 8527 kgid_t gid; 8528 8529 ASSERT_RTNL(); 8530 8531 dev->flags |= IFF_PROMISC; 8532 dev->promiscuity += inc; 8533 if (dev->promiscuity == 0) { 8534 /* 8535 * Avoid overflow. 8536 * If inc causes overflow, untouch promisc and return error. 8537 */ 8538 if (inc < 0) 8539 dev->flags &= ~IFF_PROMISC; 8540 else { 8541 dev->promiscuity -= inc; 8542 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 8543 return -EOVERFLOW; 8544 } 8545 } 8546 if (dev->flags != old_flags) { 8547 netdev_info(dev, "%s promiscuous mode\n", 8548 dev->flags & IFF_PROMISC ? "entered" : "left"); 8549 if (audit_enabled) { 8550 current_uid_gid(&uid, &gid); 8551 audit_log(audit_context(), GFP_ATOMIC, 8552 AUDIT_ANOM_PROMISCUOUS, 8553 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 8554 dev->name, (dev->flags & IFF_PROMISC), 8555 (old_flags & IFF_PROMISC), 8556 from_kuid(&init_user_ns, audit_get_loginuid(current)), 8557 from_kuid(&init_user_ns, uid), 8558 from_kgid(&init_user_ns, gid), 8559 audit_get_sessionid(current)); 8560 } 8561 8562 dev_change_rx_flags(dev, IFF_PROMISC); 8563 } 8564 if (notify) 8565 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 8566 return 0; 8567 } 8568 8569 /** 8570 * dev_set_promiscuity - update promiscuity count on a device 8571 * @dev: device 8572 * @inc: modifier 8573 * 8574 * Add or remove promiscuity from a device. While the count in the device 8575 * remains above zero the interface remains promiscuous. Once it hits zero 8576 * the device reverts back to normal filtering operation. A negative inc 8577 * value is used to drop promiscuity on the device. 8578 * Return 0 if successful or a negative errno code on error. 8579 */ 8580 int dev_set_promiscuity(struct net_device *dev, int inc) 8581 { 8582 unsigned int old_flags = dev->flags; 8583 int err; 8584 8585 err = __dev_set_promiscuity(dev, inc, true); 8586 if (err < 0) 8587 return err; 8588 if (dev->flags != old_flags) 8589 dev_set_rx_mode(dev); 8590 return err; 8591 } 8592 EXPORT_SYMBOL(dev_set_promiscuity); 8593 8594 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify) 8595 { 8596 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 8597 8598 ASSERT_RTNL(); 8599 8600 dev->flags |= IFF_ALLMULTI; 8601 dev->allmulti += inc; 8602 if (dev->allmulti == 0) { 8603 /* 8604 * Avoid overflow. 8605 * If inc causes overflow, untouch allmulti and return error. 8606 */ 8607 if (inc < 0) 8608 dev->flags &= ~IFF_ALLMULTI; 8609 else { 8610 dev->allmulti -= inc; 8611 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 8612 return -EOVERFLOW; 8613 } 8614 } 8615 if (dev->flags ^ old_flags) { 8616 netdev_info(dev, "%s allmulticast mode\n", 8617 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 8618 dev_change_rx_flags(dev, IFF_ALLMULTI); 8619 dev_set_rx_mode(dev); 8620 if (notify) 8621 __dev_notify_flags(dev, old_flags, 8622 dev->gflags ^ old_gflags, 0, NULL); 8623 } 8624 return 0; 8625 } 8626 8627 /** 8628 * dev_set_allmulti - update allmulti count on a device 8629 * @dev: device 8630 * @inc: modifier 8631 * 8632 * Add or remove reception of all multicast frames to a device. While the 8633 * count in the device remains above zero the interface remains listening 8634 * to all interfaces. Once it hits zero the device reverts back to normal 8635 * filtering operation. A negative @inc value is used to drop the counter 8636 * when releasing a resource needing all multicasts. 8637 * Return 0 if successful or a negative errno code on error. 8638 */ 8639 8640 int dev_set_allmulti(struct net_device *dev, int inc) 8641 { 8642 return __dev_set_allmulti(dev, inc, true); 8643 } 8644 EXPORT_SYMBOL(dev_set_allmulti); 8645 8646 /* 8647 * Upload unicast and multicast address lists to device and 8648 * configure RX filtering. When the device doesn't support unicast 8649 * filtering it is put in promiscuous mode while unicast addresses 8650 * are present. 8651 */ 8652 void __dev_set_rx_mode(struct net_device *dev) 8653 { 8654 const struct net_device_ops *ops = dev->netdev_ops; 8655 8656 /* dev_open will call this function so the list will stay sane. */ 8657 if (!(dev->flags&IFF_UP)) 8658 return; 8659 8660 if (!netif_device_present(dev)) 8661 return; 8662 8663 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 8664 /* Unicast addresses changes may only happen under the rtnl, 8665 * therefore calling __dev_set_promiscuity here is safe. 8666 */ 8667 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 8668 __dev_set_promiscuity(dev, 1, false); 8669 dev->uc_promisc = true; 8670 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 8671 __dev_set_promiscuity(dev, -1, false); 8672 dev->uc_promisc = false; 8673 } 8674 } 8675 8676 if (ops->ndo_set_rx_mode) 8677 ops->ndo_set_rx_mode(dev); 8678 } 8679 8680 void dev_set_rx_mode(struct net_device *dev) 8681 { 8682 netif_addr_lock_bh(dev); 8683 __dev_set_rx_mode(dev); 8684 netif_addr_unlock_bh(dev); 8685 } 8686 8687 /** 8688 * dev_get_flags - get flags reported to userspace 8689 * @dev: device 8690 * 8691 * Get the combination of flag bits exported through APIs to userspace. 8692 */ 8693 unsigned int dev_get_flags(const struct net_device *dev) 8694 { 8695 unsigned int flags; 8696 8697 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 8698 IFF_ALLMULTI | 8699 IFF_RUNNING | 8700 IFF_LOWER_UP | 8701 IFF_DORMANT)) | 8702 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 8703 IFF_ALLMULTI)); 8704 8705 if (netif_running(dev)) { 8706 if (netif_oper_up(dev)) 8707 flags |= IFF_RUNNING; 8708 if (netif_carrier_ok(dev)) 8709 flags |= IFF_LOWER_UP; 8710 if (netif_dormant(dev)) 8711 flags |= IFF_DORMANT; 8712 } 8713 8714 return flags; 8715 } 8716 EXPORT_SYMBOL(dev_get_flags); 8717 8718 int __dev_change_flags(struct net_device *dev, unsigned int flags, 8719 struct netlink_ext_ack *extack) 8720 { 8721 unsigned int old_flags = dev->flags; 8722 int ret; 8723 8724 ASSERT_RTNL(); 8725 8726 /* 8727 * Set the flags on our device. 8728 */ 8729 8730 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 8731 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 8732 IFF_AUTOMEDIA)) | 8733 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 8734 IFF_ALLMULTI)); 8735 8736 /* 8737 * Load in the correct multicast list now the flags have changed. 8738 */ 8739 8740 if ((old_flags ^ flags) & IFF_MULTICAST) 8741 dev_change_rx_flags(dev, IFF_MULTICAST); 8742 8743 dev_set_rx_mode(dev); 8744 8745 /* 8746 * Have we downed the interface. We handle IFF_UP ourselves 8747 * according to user attempts to set it, rather than blindly 8748 * setting it. 8749 */ 8750 8751 ret = 0; 8752 if ((old_flags ^ flags) & IFF_UP) { 8753 if (old_flags & IFF_UP) 8754 __dev_close(dev); 8755 else 8756 ret = __dev_open(dev, extack); 8757 } 8758 8759 if ((flags ^ dev->gflags) & IFF_PROMISC) { 8760 int inc = (flags & IFF_PROMISC) ? 1 : -1; 8761 unsigned int old_flags = dev->flags; 8762 8763 dev->gflags ^= IFF_PROMISC; 8764 8765 if (__dev_set_promiscuity(dev, inc, false) >= 0) 8766 if (dev->flags != old_flags) 8767 dev_set_rx_mode(dev); 8768 } 8769 8770 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 8771 * is important. Some (broken) drivers set IFF_PROMISC, when 8772 * IFF_ALLMULTI is requested not asking us and not reporting. 8773 */ 8774 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 8775 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 8776 8777 dev->gflags ^= IFF_ALLMULTI; 8778 __dev_set_allmulti(dev, inc, false); 8779 } 8780 8781 return ret; 8782 } 8783 8784 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 8785 unsigned int gchanges, u32 portid, 8786 const struct nlmsghdr *nlh) 8787 { 8788 unsigned int changes = dev->flags ^ old_flags; 8789 8790 if (gchanges) 8791 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 8792 8793 if (changes & IFF_UP) { 8794 if (dev->flags & IFF_UP) 8795 call_netdevice_notifiers(NETDEV_UP, dev); 8796 else 8797 call_netdevice_notifiers(NETDEV_DOWN, dev); 8798 } 8799 8800 if (dev->flags & IFF_UP && 8801 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 8802 struct netdev_notifier_change_info change_info = { 8803 .info = { 8804 .dev = dev, 8805 }, 8806 .flags_changed = changes, 8807 }; 8808 8809 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 8810 } 8811 } 8812 8813 /** 8814 * dev_change_flags - change device settings 8815 * @dev: device 8816 * @flags: device state flags 8817 * @extack: netlink extended ack 8818 * 8819 * Change settings on device based state flags. The flags are 8820 * in the userspace exported format. 8821 */ 8822 int dev_change_flags(struct net_device *dev, unsigned int flags, 8823 struct netlink_ext_ack *extack) 8824 { 8825 int ret; 8826 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 8827 8828 ret = __dev_change_flags(dev, flags, extack); 8829 if (ret < 0) 8830 return ret; 8831 8832 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 8833 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 8834 return ret; 8835 } 8836 EXPORT_SYMBOL(dev_change_flags); 8837 8838 int __dev_set_mtu(struct net_device *dev, int new_mtu) 8839 { 8840 const struct net_device_ops *ops = dev->netdev_ops; 8841 8842 if (ops->ndo_change_mtu) 8843 return ops->ndo_change_mtu(dev, new_mtu); 8844 8845 /* Pairs with all the lockless reads of dev->mtu in the stack */ 8846 WRITE_ONCE(dev->mtu, new_mtu); 8847 return 0; 8848 } 8849 EXPORT_SYMBOL(__dev_set_mtu); 8850 8851 int dev_validate_mtu(struct net_device *dev, int new_mtu, 8852 struct netlink_ext_ack *extack) 8853 { 8854 /* MTU must be positive, and in range */ 8855 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 8856 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 8857 return -EINVAL; 8858 } 8859 8860 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 8861 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 8862 return -EINVAL; 8863 } 8864 return 0; 8865 } 8866 8867 /** 8868 * dev_set_mtu_ext - Change maximum transfer unit 8869 * @dev: device 8870 * @new_mtu: new transfer unit 8871 * @extack: netlink extended ack 8872 * 8873 * Change the maximum transfer size of the network device. 8874 */ 8875 int dev_set_mtu_ext(struct net_device *dev, int new_mtu, 8876 struct netlink_ext_ack *extack) 8877 { 8878 int err, orig_mtu; 8879 8880 if (new_mtu == dev->mtu) 8881 return 0; 8882 8883 err = dev_validate_mtu(dev, new_mtu, extack); 8884 if (err) 8885 return err; 8886 8887 if (!netif_device_present(dev)) 8888 return -ENODEV; 8889 8890 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 8891 err = notifier_to_errno(err); 8892 if (err) 8893 return err; 8894 8895 orig_mtu = dev->mtu; 8896 err = __dev_set_mtu(dev, new_mtu); 8897 8898 if (!err) { 8899 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 8900 orig_mtu); 8901 err = notifier_to_errno(err); 8902 if (err) { 8903 /* setting mtu back and notifying everyone again, 8904 * so that they have a chance to revert changes. 8905 */ 8906 __dev_set_mtu(dev, orig_mtu); 8907 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 8908 new_mtu); 8909 } 8910 } 8911 return err; 8912 } 8913 8914 int dev_set_mtu(struct net_device *dev, int new_mtu) 8915 { 8916 struct netlink_ext_ack extack; 8917 int err; 8918 8919 memset(&extack, 0, sizeof(extack)); 8920 err = dev_set_mtu_ext(dev, new_mtu, &extack); 8921 if (err && extack._msg) 8922 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 8923 return err; 8924 } 8925 EXPORT_SYMBOL(dev_set_mtu); 8926 8927 /** 8928 * dev_change_tx_queue_len - Change TX queue length of a netdevice 8929 * @dev: device 8930 * @new_len: new tx queue length 8931 */ 8932 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 8933 { 8934 unsigned int orig_len = dev->tx_queue_len; 8935 int res; 8936 8937 if (new_len != (unsigned int)new_len) 8938 return -ERANGE; 8939 8940 if (new_len != orig_len) { 8941 dev->tx_queue_len = new_len; 8942 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 8943 res = notifier_to_errno(res); 8944 if (res) 8945 goto err_rollback; 8946 res = dev_qdisc_change_tx_queue_len(dev); 8947 if (res) 8948 goto err_rollback; 8949 } 8950 8951 return 0; 8952 8953 err_rollback: 8954 netdev_err(dev, "refused to change device tx_queue_len\n"); 8955 dev->tx_queue_len = orig_len; 8956 return res; 8957 } 8958 8959 /** 8960 * dev_set_group - Change group this device belongs to 8961 * @dev: device 8962 * @new_group: group this device should belong to 8963 */ 8964 void dev_set_group(struct net_device *dev, int new_group) 8965 { 8966 dev->group = new_group; 8967 } 8968 8969 /** 8970 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 8971 * @dev: device 8972 * @addr: new address 8973 * @extack: netlink extended ack 8974 */ 8975 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 8976 struct netlink_ext_ack *extack) 8977 { 8978 struct netdev_notifier_pre_changeaddr_info info = { 8979 .info.dev = dev, 8980 .info.extack = extack, 8981 .dev_addr = addr, 8982 }; 8983 int rc; 8984 8985 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 8986 return notifier_to_errno(rc); 8987 } 8988 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 8989 8990 /** 8991 * dev_set_mac_address - Change Media Access Control Address 8992 * @dev: device 8993 * @sa: new address 8994 * @extack: netlink extended ack 8995 * 8996 * Change the hardware (MAC) address of the device 8997 */ 8998 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa, 8999 struct netlink_ext_ack *extack) 9000 { 9001 const struct net_device_ops *ops = dev->netdev_ops; 9002 int err; 9003 9004 if (!ops->ndo_set_mac_address) 9005 return -EOPNOTSUPP; 9006 if (sa->sa_family != dev->type) 9007 return -EINVAL; 9008 if (!netif_device_present(dev)) 9009 return -ENODEV; 9010 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 9011 if (err) 9012 return err; 9013 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 9014 err = ops->ndo_set_mac_address(dev, sa); 9015 if (err) 9016 return err; 9017 } 9018 dev->addr_assign_type = NET_ADDR_SET; 9019 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9020 add_device_randomness(dev->dev_addr, dev->addr_len); 9021 return 0; 9022 } 9023 EXPORT_SYMBOL(dev_set_mac_address); 9024 9025 DECLARE_RWSEM(dev_addr_sem); 9026 9027 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa, 9028 struct netlink_ext_ack *extack) 9029 { 9030 int ret; 9031 9032 down_write(&dev_addr_sem); 9033 ret = dev_set_mac_address(dev, sa, extack); 9034 up_write(&dev_addr_sem); 9035 return ret; 9036 } 9037 EXPORT_SYMBOL(dev_set_mac_address_user); 9038 9039 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9040 { 9041 size_t size = sizeof(sa->sa_data_min); 9042 struct net_device *dev; 9043 int ret = 0; 9044 9045 down_read(&dev_addr_sem); 9046 rcu_read_lock(); 9047 9048 dev = dev_get_by_name_rcu(net, dev_name); 9049 if (!dev) { 9050 ret = -ENODEV; 9051 goto unlock; 9052 } 9053 if (!dev->addr_len) 9054 memset(sa->sa_data, 0, size); 9055 else 9056 memcpy(sa->sa_data, dev->dev_addr, 9057 min_t(size_t, size, dev->addr_len)); 9058 sa->sa_family = dev->type; 9059 9060 unlock: 9061 rcu_read_unlock(); 9062 up_read(&dev_addr_sem); 9063 return ret; 9064 } 9065 EXPORT_SYMBOL(dev_get_mac_address); 9066 9067 /** 9068 * dev_change_carrier - Change device carrier 9069 * @dev: device 9070 * @new_carrier: new value 9071 * 9072 * Change device carrier 9073 */ 9074 int dev_change_carrier(struct net_device *dev, bool new_carrier) 9075 { 9076 const struct net_device_ops *ops = dev->netdev_ops; 9077 9078 if (!ops->ndo_change_carrier) 9079 return -EOPNOTSUPP; 9080 if (!netif_device_present(dev)) 9081 return -ENODEV; 9082 return ops->ndo_change_carrier(dev, new_carrier); 9083 } 9084 9085 /** 9086 * dev_get_phys_port_id - Get device physical port ID 9087 * @dev: device 9088 * @ppid: port ID 9089 * 9090 * Get device physical port ID 9091 */ 9092 int dev_get_phys_port_id(struct net_device *dev, 9093 struct netdev_phys_item_id *ppid) 9094 { 9095 const struct net_device_ops *ops = dev->netdev_ops; 9096 9097 if (!ops->ndo_get_phys_port_id) 9098 return -EOPNOTSUPP; 9099 return ops->ndo_get_phys_port_id(dev, ppid); 9100 } 9101 9102 /** 9103 * dev_get_phys_port_name - Get device physical port name 9104 * @dev: device 9105 * @name: port name 9106 * @len: limit of bytes to copy to name 9107 * 9108 * Get device physical port name 9109 */ 9110 int dev_get_phys_port_name(struct net_device *dev, 9111 char *name, size_t len) 9112 { 9113 const struct net_device_ops *ops = dev->netdev_ops; 9114 int err; 9115 9116 if (ops->ndo_get_phys_port_name) { 9117 err = ops->ndo_get_phys_port_name(dev, name, len); 9118 if (err != -EOPNOTSUPP) 9119 return err; 9120 } 9121 return devlink_compat_phys_port_name_get(dev, name, len); 9122 } 9123 9124 /** 9125 * dev_get_port_parent_id - Get the device's port parent identifier 9126 * @dev: network device 9127 * @ppid: pointer to a storage for the port's parent identifier 9128 * @recurse: allow/disallow recursion to lower devices 9129 * 9130 * Get the devices's port parent identifier 9131 */ 9132 int dev_get_port_parent_id(struct net_device *dev, 9133 struct netdev_phys_item_id *ppid, 9134 bool recurse) 9135 { 9136 const struct net_device_ops *ops = dev->netdev_ops; 9137 struct netdev_phys_item_id first = { }; 9138 struct net_device *lower_dev; 9139 struct list_head *iter; 9140 int err; 9141 9142 if (ops->ndo_get_port_parent_id) { 9143 err = ops->ndo_get_port_parent_id(dev, ppid); 9144 if (err != -EOPNOTSUPP) 9145 return err; 9146 } 9147 9148 err = devlink_compat_switch_id_get(dev, ppid); 9149 if (!recurse || err != -EOPNOTSUPP) 9150 return err; 9151 9152 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9153 err = dev_get_port_parent_id(lower_dev, ppid, true); 9154 if (err) 9155 break; 9156 if (!first.id_len) 9157 first = *ppid; 9158 else if (memcmp(&first, ppid, sizeof(*ppid))) 9159 return -EOPNOTSUPP; 9160 } 9161 9162 return err; 9163 } 9164 EXPORT_SYMBOL(dev_get_port_parent_id); 9165 9166 /** 9167 * netdev_port_same_parent_id - Indicate if two network devices have 9168 * the same port parent identifier 9169 * @a: first network device 9170 * @b: second network device 9171 */ 9172 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9173 { 9174 struct netdev_phys_item_id a_id = { }; 9175 struct netdev_phys_item_id b_id = { }; 9176 9177 if (dev_get_port_parent_id(a, &a_id, true) || 9178 dev_get_port_parent_id(b, &b_id, true)) 9179 return false; 9180 9181 return netdev_phys_item_id_same(&a_id, &b_id); 9182 } 9183 EXPORT_SYMBOL(netdev_port_same_parent_id); 9184 9185 /** 9186 * dev_change_proto_down - set carrier according to proto_down. 9187 * 9188 * @dev: device 9189 * @proto_down: new value 9190 */ 9191 int dev_change_proto_down(struct net_device *dev, bool proto_down) 9192 { 9193 if (!(dev->priv_flags & IFF_CHANGE_PROTO_DOWN)) 9194 return -EOPNOTSUPP; 9195 if (!netif_device_present(dev)) 9196 return -ENODEV; 9197 if (proto_down) 9198 netif_carrier_off(dev); 9199 else 9200 netif_carrier_on(dev); 9201 dev->proto_down = proto_down; 9202 return 0; 9203 } 9204 9205 /** 9206 * dev_change_proto_down_reason - proto down reason 9207 * 9208 * @dev: device 9209 * @mask: proto down mask 9210 * @value: proto down value 9211 */ 9212 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask, 9213 u32 value) 9214 { 9215 int b; 9216 9217 if (!mask) { 9218 dev->proto_down_reason = value; 9219 } else { 9220 for_each_set_bit(b, &mask, 32) { 9221 if (value & (1 << b)) 9222 dev->proto_down_reason |= BIT(b); 9223 else 9224 dev->proto_down_reason &= ~BIT(b); 9225 } 9226 } 9227 } 9228 9229 struct bpf_xdp_link { 9230 struct bpf_link link; 9231 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9232 int flags; 9233 }; 9234 9235 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9236 { 9237 if (flags & XDP_FLAGS_HW_MODE) 9238 return XDP_MODE_HW; 9239 if (flags & XDP_FLAGS_DRV_MODE) 9240 return XDP_MODE_DRV; 9241 if (flags & XDP_FLAGS_SKB_MODE) 9242 return XDP_MODE_SKB; 9243 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9244 } 9245 9246 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9247 { 9248 switch (mode) { 9249 case XDP_MODE_SKB: 9250 return generic_xdp_install; 9251 case XDP_MODE_DRV: 9252 case XDP_MODE_HW: 9253 return dev->netdev_ops->ndo_bpf; 9254 default: 9255 return NULL; 9256 } 9257 } 9258 9259 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9260 enum bpf_xdp_mode mode) 9261 { 9262 return dev->xdp_state[mode].link; 9263 } 9264 9265 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9266 enum bpf_xdp_mode mode) 9267 { 9268 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9269 9270 if (link) 9271 return link->link.prog; 9272 return dev->xdp_state[mode].prog; 9273 } 9274 9275 u8 dev_xdp_prog_count(struct net_device *dev) 9276 { 9277 u8 count = 0; 9278 int i; 9279 9280 for (i = 0; i < __MAX_XDP_MODE; i++) 9281 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9282 count++; 9283 return count; 9284 } 9285 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9286 9287 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9288 { 9289 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9290 9291 return prog ? prog->aux->id : 0; 9292 } 9293 9294 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9295 struct bpf_xdp_link *link) 9296 { 9297 dev->xdp_state[mode].link = link; 9298 dev->xdp_state[mode].prog = NULL; 9299 } 9300 9301 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9302 struct bpf_prog *prog) 9303 { 9304 dev->xdp_state[mode].link = NULL; 9305 dev->xdp_state[mode].prog = prog; 9306 } 9307 9308 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9309 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9310 u32 flags, struct bpf_prog *prog) 9311 { 9312 struct netdev_bpf xdp; 9313 int err; 9314 9315 memset(&xdp, 0, sizeof(xdp)); 9316 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9317 xdp.extack = extack; 9318 xdp.flags = flags; 9319 xdp.prog = prog; 9320 9321 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9322 * "moved" into driver), so they don't increment it on their own, but 9323 * they do decrement refcnt when program is detached or replaced. 9324 * Given net_device also owns link/prog, we need to bump refcnt here 9325 * to prevent drivers from underflowing it. 9326 */ 9327 if (prog) 9328 bpf_prog_inc(prog); 9329 err = bpf_op(dev, &xdp); 9330 if (err) { 9331 if (prog) 9332 bpf_prog_put(prog); 9333 return err; 9334 } 9335 9336 if (mode != XDP_MODE_HW) 9337 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9338 9339 return 0; 9340 } 9341 9342 static void dev_xdp_uninstall(struct net_device *dev) 9343 { 9344 struct bpf_xdp_link *link; 9345 struct bpf_prog *prog; 9346 enum bpf_xdp_mode mode; 9347 bpf_op_t bpf_op; 9348 9349 ASSERT_RTNL(); 9350 9351 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 9352 prog = dev_xdp_prog(dev, mode); 9353 if (!prog) 9354 continue; 9355 9356 bpf_op = dev_xdp_bpf_op(dev, mode); 9357 if (!bpf_op) 9358 continue; 9359 9360 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9361 9362 /* auto-detach link from net device */ 9363 link = dev_xdp_link(dev, mode); 9364 if (link) 9365 link->dev = NULL; 9366 else 9367 bpf_prog_put(prog); 9368 9369 dev_xdp_set_link(dev, mode, NULL); 9370 } 9371 } 9372 9373 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 9374 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 9375 struct bpf_prog *old_prog, u32 flags) 9376 { 9377 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 9378 struct bpf_prog *cur_prog; 9379 struct net_device *upper; 9380 struct list_head *iter; 9381 enum bpf_xdp_mode mode; 9382 bpf_op_t bpf_op; 9383 int err; 9384 9385 ASSERT_RTNL(); 9386 9387 /* either link or prog attachment, never both */ 9388 if (link && (new_prog || old_prog)) 9389 return -EINVAL; 9390 /* link supports only XDP mode flags */ 9391 if (link && (flags & ~XDP_FLAGS_MODES)) { 9392 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 9393 return -EINVAL; 9394 } 9395 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 9396 if (num_modes > 1) { 9397 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 9398 return -EINVAL; 9399 } 9400 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 9401 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 9402 NL_SET_ERR_MSG(extack, 9403 "More than one program loaded, unset mode is ambiguous"); 9404 return -EINVAL; 9405 } 9406 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 9407 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 9408 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 9409 return -EINVAL; 9410 } 9411 9412 mode = dev_xdp_mode(dev, flags); 9413 /* can't replace attached link */ 9414 if (dev_xdp_link(dev, mode)) { 9415 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 9416 return -EBUSY; 9417 } 9418 9419 /* don't allow if an upper device already has a program */ 9420 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 9421 if (dev_xdp_prog_count(upper) > 0) { 9422 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 9423 return -EEXIST; 9424 } 9425 } 9426 9427 cur_prog = dev_xdp_prog(dev, mode); 9428 /* can't replace attached prog with link */ 9429 if (link && cur_prog) { 9430 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 9431 return -EBUSY; 9432 } 9433 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 9434 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 9435 return -EEXIST; 9436 } 9437 9438 /* put effective new program into new_prog */ 9439 if (link) 9440 new_prog = link->link.prog; 9441 9442 if (new_prog) { 9443 bool offload = mode == XDP_MODE_HW; 9444 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 9445 ? XDP_MODE_DRV : XDP_MODE_SKB; 9446 9447 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 9448 NL_SET_ERR_MSG(extack, "XDP program already attached"); 9449 return -EBUSY; 9450 } 9451 if (!offload && dev_xdp_prog(dev, other_mode)) { 9452 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 9453 return -EEXIST; 9454 } 9455 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 9456 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 9457 return -EINVAL; 9458 } 9459 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 9460 NL_SET_ERR_MSG(extack, "Program bound to different device"); 9461 return -EINVAL; 9462 } 9463 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 9464 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 9465 return -EINVAL; 9466 } 9467 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 9468 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 9469 return -EINVAL; 9470 } 9471 } 9472 9473 /* don't call drivers if the effective program didn't change */ 9474 if (new_prog != cur_prog) { 9475 bpf_op = dev_xdp_bpf_op(dev, mode); 9476 if (!bpf_op) { 9477 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 9478 return -EOPNOTSUPP; 9479 } 9480 9481 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 9482 if (err) 9483 return err; 9484 } 9485 9486 if (link) 9487 dev_xdp_set_link(dev, mode, link); 9488 else 9489 dev_xdp_set_prog(dev, mode, new_prog); 9490 if (cur_prog) 9491 bpf_prog_put(cur_prog); 9492 9493 return 0; 9494 } 9495 9496 static int dev_xdp_attach_link(struct net_device *dev, 9497 struct netlink_ext_ack *extack, 9498 struct bpf_xdp_link *link) 9499 { 9500 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 9501 } 9502 9503 static int dev_xdp_detach_link(struct net_device *dev, 9504 struct netlink_ext_ack *extack, 9505 struct bpf_xdp_link *link) 9506 { 9507 enum bpf_xdp_mode mode; 9508 bpf_op_t bpf_op; 9509 9510 ASSERT_RTNL(); 9511 9512 mode = dev_xdp_mode(dev, link->flags); 9513 if (dev_xdp_link(dev, mode) != link) 9514 return -EINVAL; 9515 9516 bpf_op = dev_xdp_bpf_op(dev, mode); 9517 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9518 dev_xdp_set_link(dev, mode, NULL); 9519 return 0; 9520 } 9521 9522 static void bpf_xdp_link_release(struct bpf_link *link) 9523 { 9524 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9525 9526 rtnl_lock(); 9527 9528 /* if racing with net_device's tear down, xdp_link->dev might be 9529 * already NULL, in which case link was already auto-detached 9530 */ 9531 if (xdp_link->dev) { 9532 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 9533 xdp_link->dev = NULL; 9534 } 9535 9536 rtnl_unlock(); 9537 } 9538 9539 static int bpf_xdp_link_detach(struct bpf_link *link) 9540 { 9541 bpf_xdp_link_release(link); 9542 return 0; 9543 } 9544 9545 static void bpf_xdp_link_dealloc(struct bpf_link *link) 9546 { 9547 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9548 9549 kfree(xdp_link); 9550 } 9551 9552 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 9553 struct seq_file *seq) 9554 { 9555 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9556 u32 ifindex = 0; 9557 9558 rtnl_lock(); 9559 if (xdp_link->dev) 9560 ifindex = xdp_link->dev->ifindex; 9561 rtnl_unlock(); 9562 9563 seq_printf(seq, "ifindex:\t%u\n", ifindex); 9564 } 9565 9566 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 9567 struct bpf_link_info *info) 9568 { 9569 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9570 u32 ifindex = 0; 9571 9572 rtnl_lock(); 9573 if (xdp_link->dev) 9574 ifindex = xdp_link->dev->ifindex; 9575 rtnl_unlock(); 9576 9577 info->xdp.ifindex = ifindex; 9578 return 0; 9579 } 9580 9581 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 9582 struct bpf_prog *old_prog) 9583 { 9584 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9585 enum bpf_xdp_mode mode; 9586 bpf_op_t bpf_op; 9587 int err = 0; 9588 9589 rtnl_lock(); 9590 9591 /* link might have been auto-released already, so fail */ 9592 if (!xdp_link->dev) { 9593 err = -ENOLINK; 9594 goto out_unlock; 9595 } 9596 9597 if (old_prog && link->prog != old_prog) { 9598 err = -EPERM; 9599 goto out_unlock; 9600 } 9601 old_prog = link->prog; 9602 if (old_prog->type != new_prog->type || 9603 old_prog->expected_attach_type != new_prog->expected_attach_type) { 9604 err = -EINVAL; 9605 goto out_unlock; 9606 } 9607 9608 if (old_prog == new_prog) { 9609 /* no-op, don't disturb drivers */ 9610 bpf_prog_put(new_prog); 9611 goto out_unlock; 9612 } 9613 9614 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 9615 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 9616 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 9617 xdp_link->flags, new_prog); 9618 if (err) 9619 goto out_unlock; 9620 9621 old_prog = xchg(&link->prog, new_prog); 9622 bpf_prog_put(old_prog); 9623 9624 out_unlock: 9625 rtnl_unlock(); 9626 return err; 9627 } 9628 9629 static const struct bpf_link_ops bpf_xdp_link_lops = { 9630 .release = bpf_xdp_link_release, 9631 .dealloc = bpf_xdp_link_dealloc, 9632 .detach = bpf_xdp_link_detach, 9633 .show_fdinfo = bpf_xdp_link_show_fdinfo, 9634 .fill_link_info = bpf_xdp_link_fill_link_info, 9635 .update_prog = bpf_xdp_link_update, 9636 }; 9637 9638 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 9639 { 9640 struct net *net = current->nsproxy->net_ns; 9641 struct bpf_link_primer link_primer; 9642 struct netlink_ext_ack extack = {}; 9643 struct bpf_xdp_link *link; 9644 struct net_device *dev; 9645 int err, fd; 9646 9647 rtnl_lock(); 9648 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 9649 if (!dev) { 9650 rtnl_unlock(); 9651 return -EINVAL; 9652 } 9653 9654 link = kzalloc(sizeof(*link), GFP_USER); 9655 if (!link) { 9656 err = -ENOMEM; 9657 goto unlock; 9658 } 9659 9660 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 9661 link->dev = dev; 9662 link->flags = attr->link_create.flags; 9663 9664 err = bpf_link_prime(&link->link, &link_primer); 9665 if (err) { 9666 kfree(link); 9667 goto unlock; 9668 } 9669 9670 err = dev_xdp_attach_link(dev, &extack, link); 9671 rtnl_unlock(); 9672 9673 if (err) { 9674 link->dev = NULL; 9675 bpf_link_cleanup(&link_primer); 9676 trace_bpf_xdp_link_attach_failed(extack._msg); 9677 goto out_put_dev; 9678 } 9679 9680 fd = bpf_link_settle(&link_primer); 9681 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 9682 dev_put(dev); 9683 return fd; 9684 9685 unlock: 9686 rtnl_unlock(); 9687 9688 out_put_dev: 9689 dev_put(dev); 9690 return err; 9691 } 9692 9693 /** 9694 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 9695 * @dev: device 9696 * @extack: netlink extended ack 9697 * @fd: new program fd or negative value to clear 9698 * @expected_fd: old program fd that userspace expects to replace or clear 9699 * @flags: xdp-related flags 9700 * 9701 * Set or clear a bpf program for a device 9702 */ 9703 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 9704 int fd, int expected_fd, u32 flags) 9705 { 9706 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 9707 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 9708 int err; 9709 9710 ASSERT_RTNL(); 9711 9712 if (fd >= 0) { 9713 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 9714 mode != XDP_MODE_SKB); 9715 if (IS_ERR(new_prog)) 9716 return PTR_ERR(new_prog); 9717 } 9718 9719 if (expected_fd >= 0) { 9720 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 9721 mode != XDP_MODE_SKB); 9722 if (IS_ERR(old_prog)) { 9723 err = PTR_ERR(old_prog); 9724 old_prog = NULL; 9725 goto err_out; 9726 } 9727 } 9728 9729 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 9730 9731 err_out: 9732 if (err && new_prog) 9733 bpf_prog_put(new_prog); 9734 if (old_prog) 9735 bpf_prog_put(old_prog); 9736 return err; 9737 } 9738 9739 /** 9740 * dev_index_reserve() - allocate an ifindex in a namespace 9741 * @net: the applicable net namespace 9742 * @ifindex: requested ifindex, pass %0 to get one allocated 9743 * 9744 * Allocate a ifindex for a new device. Caller must either use the ifindex 9745 * to store the device (via list_netdevice()) or call dev_index_release() 9746 * to give the index up. 9747 * 9748 * Return: a suitable unique value for a new device interface number or -errno. 9749 */ 9750 static int dev_index_reserve(struct net *net, u32 ifindex) 9751 { 9752 int err; 9753 9754 if (ifindex > INT_MAX) { 9755 DEBUG_NET_WARN_ON_ONCE(1); 9756 return -EINVAL; 9757 } 9758 9759 if (!ifindex) 9760 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 9761 xa_limit_31b, &net->ifindex, GFP_KERNEL); 9762 else 9763 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 9764 if (err < 0) 9765 return err; 9766 9767 return ifindex; 9768 } 9769 9770 static void dev_index_release(struct net *net, int ifindex) 9771 { 9772 /* Expect only unused indexes, unlist_netdevice() removes the used */ 9773 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 9774 } 9775 9776 /* Delayed registration/unregisteration */ 9777 LIST_HEAD(net_todo_list); 9778 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 9779 atomic_t dev_unreg_count = ATOMIC_INIT(0); 9780 9781 static void net_set_todo(struct net_device *dev) 9782 { 9783 list_add_tail(&dev->todo_list, &net_todo_list); 9784 } 9785 9786 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 9787 struct net_device *upper, netdev_features_t features) 9788 { 9789 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 9790 netdev_features_t feature; 9791 int feature_bit; 9792 9793 for_each_netdev_feature(upper_disables, feature_bit) { 9794 feature = __NETIF_F_BIT(feature_bit); 9795 if (!(upper->wanted_features & feature) 9796 && (features & feature)) { 9797 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 9798 &feature, upper->name); 9799 features &= ~feature; 9800 } 9801 } 9802 9803 return features; 9804 } 9805 9806 static void netdev_sync_lower_features(struct net_device *upper, 9807 struct net_device *lower, netdev_features_t features) 9808 { 9809 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 9810 netdev_features_t feature; 9811 int feature_bit; 9812 9813 for_each_netdev_feature(upper_disables, feature_bit) { 9814 feature = __NETIF_F_BIT(feature_bit); 9815 if (!(features & feature) && (lower->features & feature)) { 9816 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 9817 &feature, lower->name); 9818 lower->wanted_features &= ~feature; 9819 __netdev_update_features(lower); 9820 9821 if (unlikely(lower->features & feature)) 9822 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 9823 &feature, lower->name); 9824 else 9825 netdev_features_change(lower); 9826 } 9827 } 9828 } 9829 9830 static netdev_features_t netdev_fix_features(struct net_device *dev, 9831 netdev_features_t features) 9832 { 9833 /* Fix illegal checksum combinations */ 9834 if ((features & NETIF_F_HW_CSUM) && 9835 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 9836 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 9837 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 9838 } 9839 9840 /* TSO requires that SG is present as well. */ 9841 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 9842 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 9843 features &= ~NETIF_F_ALL_TSO; 9844 } 9845 9846 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 9847 !(features & NETIF_F_IP_CSUM)) { 9848 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 9849 features &= ~NETIF_F_TSO; 9850 features &= ~NETIF_F_TSO_ECN; 9851 } 9852 9853 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 9854 !(features & NETIF_F_IPV6_CSUM)) { 9855 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 9856 features &= ~NETIF_F_TSO6; 9857 } 9858 9859 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 9860 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 9861 features &= ~NETIF_F_TSO_MANGLEID; 9862 9863 /* TSO ECN requires that TSO is present as well. */ 9864 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 9865 features &= ~NETIF_F_TSO_ECN; 9866 9867 /* Software GSO depends on SG. */ 9868 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 9869 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 9870 features &= ~NETIF_F_GSO; 9871 } 9872 9873 /* GSO partial features require GSO partial be set */ 9874 if ((features & dev->gso_partial_features) && 9875 !(features & NETIF_F_GSO_PARTIAL)) { 9876 netdev_dbg(dev, 9877 "Dropping partially supported GSO features since no GSO partial.\n"); 9878 features &= ~dev->gso_partial_features; 9879 } 9880 9881 if (!(features & NETIF_F_RXCSUM)) { 9882 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 9883 * successfully merged by hardware must also have the 9884 * checksum verified by hardware. If the user does not 9885 * want to enable RXCSUM, logically, we should disable GRO_HW. 9886 */ 9887 if (features & NETIF_F_GRO_HW) { 9888 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 9889 features &= ~NETIF_F_GRO_HW; 9890 } 9891 } 9892 9893 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 9894 if (features & NETIF_F_RXFCS) { 9895 if (features & NETIF_F_LRO) { 9896 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 9897 features &= ~NETIF_F_LRO; 9898 } 9899 9900 if (features & NETIF_F_GRO_HW) { 9901 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 9902 features &= ~NETIF_F_GRO_HW; 9903 } 9904 } 9905 9906 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 9907 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 9908 features &= ~NETIF_F_LRO; 9909 } 9910 9911 if (features & NETIF_F_HW_TLS_TX) { 9912 bool ip_csum = (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) == 9913 (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM); 9914 bool hw_csum = features & NETIF_F_HW_CSUM; 9915 9916 if (!ip_csum && !hw_csum) { 9917 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 9918 features &= ~NETIF_F_HW_TLS_TX; 9919 } 9920 } 9921 9922 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 9923 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 9924 features &= ~NETIF_F_HW_TLS_RX; 9925 } 9926 9927 return features; 9928 } 9929 9930 int __netdev_update_features(struct net_device *dev) 9931 { 9932 struct net_device *upper, *lower; 9933 netdev_features_t features; 9934 struct list_head *iter; 9935 int err = -1; 9936 9937 ASSERT_RTNL(); 9938 9939 features = netdev_get_wanted_features(dev); 9940 9941 if (dev->netdev_ops->ndo_fix_features) 9942 features = dev->netdev_ops->ndo_fix_features(dev, features); 9943 9944 /* driver might be less strict about feature dependencies */ 9945 features = netdev_fix_features(dev, features); 9946 9947 /* some features can't be enabled if they're off on an upper device */ 9948 netdev_for_each_upper_dev_rcu(dev, upper, iter) 9949 features = netdev_sync_upper_features(dev, upper, features); 9950 9951 if (dev->features == features) 9952 goto sync_lower; 9953 9954 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 9955 &dev->features, &features); 9956 9957 if (dev->netdev_ops->ndo_set_features) 9958 err = dev->netdev_ops->ndo_set_features(dev, features); 9959 else 9960 err = 0; 9961 9962 if (unlikely(err < 0)) { 9963 netdev_err(dev, 9964 "set_features() failed (%d); wanted %pNF, left %pNF\n", 9965 err, &features, &dev->features); 9966 /* return non-0 since some features might have changed and 9967 * it's better to fire a spurious notification than miss it 9968 */ 9969 return -1; 9970 } 9971 9972 sync_lower: 9973 /* some features must be disabled on lower devices when disabled 9974 * on an upper device (think: bonding master or bridge) 9975 */ 9976 netdev_for_each_lower_dev(dev, lower, iter) 9977 netdev_sync_lower_features(dev, lower, features); 9978 9979 if (!err) { 9980 netdev_features_t diff = features ^ dev->features; 9981 9982 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 9983 /* udp_tunnel_{get,drop}_rx_info both need 9984 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 9985 * device, or they won't do anything. 9986 * Thus we need to update dev->features 9987 * *before* calling udp_tunnel_get_rx_info, 9988 * but *after* calling udp_tunnel_drop_rx_info. 9989 */ 9990 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 9991 dev->features = features; 9992 udp_tunnel_get_rx_info(dev); 9993 } else { 9994 udp_tunnel_drop_rx_info(dev); 9995 } 9996 } 9997 9998 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 9999 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 10000 dev->features = features; 10001 err |= vlan_get_rx_ctag_filter_info(dev); 10002 } else { 10003 vlan_drop_rx_ctag_filter_info(dev); 10004 } 10005 } 10006 10007 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 10008 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 10009 dev->features = features; 10010 err |= vlan_get_rx_stag_filter_info(dev); 10011 } else { 10012 vlan_drop_rx_stag_filter_info(dev); 10013 } 10014 } 10015 10016 dev->features = features; 10017 } 10018 10019 return err < 0 ? 0 : 1; 10020 } 10021 10022 /** 10023 * netdev_update_features - recalculate device features 10024 * @dev: the device to check 10025 * 10026 * Recalculate dev->features set and send notifications if it 10027 * has changed. Should be called after driver or hardware dependent 10028 * conditions might have changed that influence the features. 10029 */ 10030 void netdev_update_features(struct net_device *dev) 10031 { 10032 if (__netdev_update_features(dev)) 10033 netdev_features_change(dev); 10034 } 10035 EXPORT_SYMBOL(netdev_update_features); 10036 10037 /** 10038 * netdev_change_features - recalculate device features 10039 * @dev: the device to check 10040 * 10041 * Recalculate dev->features set and send notifications even 10042 * if they have not changed. Should be called instead of 10043 * netdev_update_features() if also dev->vlan_features might 10044 * have changed to allow the changes to be propagated to stacked 10045 * VLAN devices. 10046 */ 10047 void netdev_change_features(struct net_device *dev) 10048 { 10049 __netdev_update_features(dev); 10050 netdev_features_change(dev); 10051 } 10052 EXPORT_SYMBOL(netdev_change_features); 10053 10054 /** 10055 * netif_stacked_transfer_operstate - transfer operstate 10056 * @rootdev: the root or lower level device to transfer state from 10057 * @dev: the device to transfer operstate to 10058 * 10059 * Transfer operational state from root to device. This is normally 10060 * called when a stacking relationship exists between the root 10061 * device and the device(a leaf device). 10062 */ 10063 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 10064 struct net_device *dev) 10065 { 10066 if (rootdev->operstate == IF_OPER_DORMANT) 10067 netif_dormant_on(dev); 10068 else 10069 netif_dormant_off(dev); 10070 10071 if (rootdev->operstate == IF_OPER_TESTING) 10072 netif_testing_on(dev); 10073 else 10074 netif_testing_off(dev); 10075 10076 if (netif_carrier_ok(rootdev)) 10077 netif_carrier_on(dev); 10078 else 10079 netif_carrier_off(dev); 10080 } 10081 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10082 10083 static int netif_alloc_rx_queues(struct net_device *dev) 10084 { 10085 unsigned int i, count = dev->num_rx_queues; 10086 struct netdev_rx_queue *rx; 10087 size_t sz = count * sizeof(*rx); 10088 int err = 0; 10089 10090 BUG_ON(count < 1); 10091 10092 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10093 if (!rx) 10094 return -ENOMEM; 10095 10096 dev->_rx = rx; 10097 10098 for (i = 0; i < count; i++) { 10099 rx[i].dev = dev; 10100 10101 /* XDP RX-queue setup */ 10102 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10103 if (err < 0) 10104 goto err_rxq_info; 10105 } 10106 return 0; 10107 10108 err_rxq_info: 10109 /* Rollback successful reg's and free other resources */ 10110 while (i--) 10111 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10112 kvfree(dev->_rx); 10113 dev->_rx = NULL; 10114 return err; 10115 } 10116 10117 static void netif_free_rx_queues(struct net_device *dev) 10118 { 10119 unsigned int i, count = dev->num_rx_queues; 10120 10121 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10122 if (!dev->_rx) 10123 return; 10124 10125 for (i = 0; i < count; i++) 10126 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10127 10128 kvfree(dev->_rx); 10129 } 10130 10131 static void netdev_init_one_queue(struct net_device *dev, 10132 struct netdev_queue *queue, void *_unused) 10133 { 10134 /* Initialize queue lock */ 10135 spin_lock_init(&queue->_xmit_lock); 10136 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10137 queue->xmit_lock_owner = -1; 10138 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10139 queue->dev = dev; 10140 #ifdef CONFIG_BQL 10141 dql_init(&queue->dql, HZ); 10142 #endif 10143 } 10144 10145 static void netif_free_tx_queues(struct net_device *dev) 10146 { 10147 kvfree(dev->_tx); 10148 } 10149 10150 static int netif_alloc_netdev_queues(struct net_device *dev) 10151 { 10152 unsigned int count = dev->num_tx_queues; 10153 struct netdev_queue *tx; 10154 size_t sz = count * sizeof(*tx); 10155 10156 if (count < 1 || count > 0xffff) 10157 return -EINVAL; 10158 10159 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10160 if (!tx) 10161 return -ENOMEM; 10162 10163 dev->_tx = tx; 10164 10165 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10166 spin_lock_init(&dev->tx_global_lock); 10167 10168 return 0; 10169 } 10170 10171 void netif_tx_stop_all_queues(struct net_device *dev) 10172 { 10173 unsigned int i; 10174 10175 for (i = 0; i < dev->num_tx_queues; i++) { 10176 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10177 10178 netif_tx_stop_queue(txq); 10179 } 10180 } 10181 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10182 10183 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10184 { 10185 void __percpu *v; 10186 10187 /* Drivers implementing ndo_get_peer_dev must support tstat 10188 * accounting, so that skb_do_redirect() can bump the dev's 10189 * RX stats upon network namespace switch. 10190 */ 10191 if (dev->netdev_ops->ndo_get_peer_dev && 10192 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10193 return -EOPNOTSUPP; 10194 10195 switch (dev->pcpu_stat_type) { 10196 case NETDEV_PCPU_STAT_NONE: 10197 return 0; 10198 case NETDEV_PCPU_STAT_LSTATS: 10199 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10200 break; 10201 case NETDEV_PCPU_STAT_TSTATS: 10202 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10203 break; 10204 case NETDEV_PCPU_STAT_DSTATS: 10205 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10206 break; 10207 default: 10208 return -EINVAL; 10209 } 10210 10211 return v ? 0 : -ENOMEM; 10212 } 10213 10214 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10215 { 10216 switch (dev->pcpu_stat_type) { 10217 case NETDEV_PCPU_STAT_NONE: 10218 return; 10219 case NETDEV_PCPU_STAT_LSTATS: 10220 free_percpu(dev->lstats); 10221 break; 10222 case NETDEV_PCPU_STAT_TSTATS: 10223 free_percpu(dev->tstats); 10224 break; 10225 case NETDEV_PCPU_STAT_DSTATS: 10226 free_percpu(dev->dstats); 10227 break; 10228 } 10229 } 10230 10231 /** 10232 * register_netdevice() - register a network device 10233 * @dev: device to register 10234 * 10235 * Take a prepared network device structure and make it externally accessible. 10236 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10237 * Callers must hold the rtnl lock - you may want register_netdev() 10238 * instead of this. 10239 */ 10240 int register_netdevice(struct net_device *dev) 10241 { 10242 int ret; 10243 struct net *net = dev_net(dev); 10244 10245 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10246 NETDEV_FEATURE_COUNT); 10247 BUG_ON(dev_boot_phase); 10248 ASSERT_RTNL(); 10249 10250 might_sleep(); 10251 10252 /* When net_device's are persistent, this will be fatal. */ 10253 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10254 BUG_ON(!net); 10255 10256 ret = ethtool_check_ops(dev->ethtool_ops); 10257 if (ret) 10258 return ret; 10259 10260 spin_lock_init(&dev->addr_list_lock); 10261 netdev_set_addr_lockdep_class(dev); 10262 10263 ret = dev_get_valid_name(net, dev, dev->name); 10264 if (ret < 0) 10265 goto out; 10266 10267 ret = -ENOMEM; 10268 dev->name_node = netdev_name_node_head_alloc(dev); 10269 if (!dev->name_node) 10270 goto out; 10271 10272 /* Init, if this function is available */ 10273 if (dev->netdev_ops->ndo_init) { 10274 ret = dev->netdev_ops->ndo_init(dev); 10275 if (ret) { 10276 if (ret > 0) 10277 ret = -EIO; 10278 goto err_free_name; 10279 } 10280 } 10281 10282 if (((dev->hw_features | dev->features) & 10283 NETIF_F_HW_VLAN_CTAG_FILTER) && 10284 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 10285 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 10286 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 10287 ret = -EINVAL; 10288 goto err_uninit; 10289 } 10290 10291 ret = netdev_do_alloc_pcpu_stats(dev); 10292 if (ret) 10293 goto err_uninit; 10294 10295 ret = dev_index_reserve(net, dev->ifindex); 10296 if (ret < 0) 10297 goto err_free_pcpu; 10298 dev->ifindex = ret; 10299 10300 /* Transfer changeable features to wanted_features and enable 10301 * software offloads (GSO and GRO). 10302 */ 10303 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 10304 dev->features |= NETIF_F_SOFT_FEATURES; 10305 10306 if (dev->udp_tunnel_nic_info) { 10307 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10308 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10309 } 10310 10311 dev->wanted_features = dev->features & dev->hw_features; 10312 10313 if (!(dev->flags & IFF_LOOPBACK)) 10314 dev->hw_features |= NETIF_F_NOCACHE_COPY; 10315 10316 /* If IPv4 TCP segmentation offload is supported we should also 10317 * allow the device to enable segmenting the frame with the option 10318 * of ignoring a static IP ID value. This doesn't enable the 10319 * feature itself but allows the user to enable it later. 10320 */ 10321 if (dev->hw_features & NETIF_F_TSO) 10322 dev->hw_features |= NETIF_F_TSO_MANGLEID; 10323 if (dev->vlan_features & NETIF_F_TSO) 10324 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 10325 if (dev->mpls_features & NETIF_F_TSO) 10326 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 10327 if (dev->hw_enc_features & NETIF_F_TSO) 10328 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 10329 10330 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 10331 */ 10332 dev->vlan_features |= NETIF_F_HIGHDMA; 10333 10334 /* Make NETIF_F_SG inheritable to tunnel devices. 10335 */ 10336 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 10337 10338 /* Make NETIF_F_SG inheritable to MPLS. 10339 */ 10340 dev->mpls_features |= NETIF_F_SG; 10341 10342 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 10343 ret = notifier_to_errno(ret); 10344 if (ret) 10345 goto err_ifindex_release; 10346 10347 ret = netdev_register_kobject(dev); 10348 10349 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 10350 10351 if (ret) 10352 goto err_uninit_notify; 10353 10354 __netdev_update_features(dev); 10355 10356 /* 10357 * Default initial state at registry is that the 10358 * device is present. 10359 */ 10360 10361 set_bit(__LINK_STATE_PRESENT, &dev->state); 10362 10363 linkwatch_init_dev(dev); 10364 10365 dev_init_scheduler(dev); 10366 10367 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 10368 list_netdevice(dev); 10369 10370 add_device_randomness(dev->dev_addr, dev->addr_len); 10371 10372 /* If the device has permanent device address, driver should 10373 * set dev_addr and also addr_assign_type should be set to 10374 * NET_ADDR_PERM (default value). 10375 */ 10376 if (dev->addr_assign_type == NET_ADDR_PERM) 10377 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 10378 10379 /* Notify protocols, that a new device appeared. */ 10380 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 10381 ret = notifier_to_errno(ret); 10382 if (ret) { 10383 /* Expect explicit free_netdev() on failure */ 10384 dev->needs_free_netdev = false; 10385 unregister_netdevice_queue(dev, NULL); 10386 goto out; 10387 } 10388 /* 10389 * Prevent userspace races by waiting until the network 10390 * device is fully setup before sending notifications. 10391 */ 10392 if (!dev->rtnl_link_ops || 10393 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 10394 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 10395 10396 out: 10397 return ret; 10398 10399 err_uninit_notify: 10400 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 10401 err_ifindex_release: 10402 dev_index_release(net, dev->ifindex); 10403 err_free_pcpu: 10404 netdev_do_free_pcpu_stats(dev); 10405 err_uninit: 10406 if (dev->netdev_ops->ndo_uninit) 10407 dev->netdev_ops->ndo_uninit(dev); 10408 if (dev->priv_destructor) 10409 dev->priv_destructor(dev); 10410 err_free_name: 10411 netdev_name_node_free(dev->name_node); 10412 goto out; 10413 } 10414 EXPORT_SYMBOL(register_netdevice); 10415 10416 /** 10417 * init_dummy_netdev - init a dummy network device for NAPI 10418 * @dev: device to init 10419 * 10420 * This takes a network device structure and initialize the minimum 10421 * amount of fields so it can be used to schedule NAPI polls without 10422 * registering a full blown interface. This is to be used by drivers 10423 * that need to tie several hardware interfaces to a single NAPI 10424 * poll scheduler due to HW limitations. 10425 */ 10426 void init_dummy_netdev(struct net_device *dev) 10427 { 10428 /* Clear everything. Note we don't initialize spinlocks 10429 * are they aren't supposed to be taken by any of the 10430 * NAPI code and this dummy netdev is supposed to be 10431 * only ever used for NAPI polls 10432 */ 10433 memset(dev, 0, sizeof(struct net_device)); 10434 10435 /* make sure we BUG if trying to hit standard 10436 * register/unregister code path 10437 */ 10438 dev->reg_state = NETREG_DUMMY; 10439 10440 /* NAPI wants this */ 10441 INIT_LIST_HEAD(&dev->napi_list); 10442 10443 /* a dummy interface is started by default */ 10444 set_bit(__LINK_STATE_PRESENT, &dev->state); 10445 set_bit(__LINK_STATE_START, &dev->state); 10446 10447 /* napi_busy_loop stats accounting wants this */ 10448 dev_net_set(dev, &init_net); 10449 10450 /* Note : We dont allocate pcpu_refcnt for dummy devices, 10451 * because users of this 'device' dont need to change 10452 * its refcount. 10453 */ 10454 } 10455 EXPORT_SYMBOL_GPL(init_dummy_netdev); 10456 10457 10458 /** 10459 * register_netdev - register a network device 10460 * @dev: device to register 10461 * 10462 * Take a completed network device structure and add it to the kernel 10463 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 10464 * chain. 0 is returned on success. A negative errno code is returned 10465 * on a failure to set up the device, or if the name is a duplicate. 10466 * 10467 * This is a wrapper around register_netdevice that takes the rtnl semaphore 10468 * and expands the device name if you passed a format string to 10469 * alloc_netdev. 10470 */ 10471 int register_netdev(struct net_device *dev) 10472 { 10473 int err; 10474 10475 if (rtnl_lock_killable()) 10476 return -EINTR; 10477 err = register_netdevice(dev); 10478 rtnl_unlock(); 10479 return err; 10480 } 10481 EXPORT_SYMBOL(register_netdev); 10482 10483 int netdev_refcnt_read(const struct net_device *dev) 10484 { 10485 #ifdef CONFIG_PCPU_DEV_REFCNT 10486 int i, refcnt = 0; 10487 10488 for_each_possible_cpu(i) 10489 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 10490 return refcnt; 10491 #else 10492 return refcount_read(&dev->dev_refcnt); 10493 #endif 10494 } 10495 EXPORT_SYMBOL(netdev_refcnt_read); 10496 10497 int netdev_unregister_timeout_secs __read_mostly = 10; 10498 10499 #define WAIT_REFS_MIN_MSECS 1 10500 #define WAIT_REFS_MAX_MSECS 250 10501 /** 10502 * netdev_wait_allrefs_any - wait until all references are gone. 10503 * @list: list of net_devices to wait on 10504 * 10505 * This is called when unregistering network devices. 10506 * 10507 * Any protocol or device that holds a reference should register 10508 * for netdevice notification, and cleanup and put back the 10509 * reference if they receive an UNREGISTER event. 10510 * We can get stuck here if buggy protocols don't correctly 10511 * call dev_put. 10512 */ 10513 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 10514 { 10515 unsigned long rebroadcast_time, warning_time; 10516 struct net_device *dev; 10517 int wait = 0; 10518 10519 rebroadcast_time = warning_time = jiffies; 10520 10521 list_for_each_entry(dev, list, todo_list) 10522 if (netdev_refcnt_read(dev) == 1) 10523 return dev; 10524 10525 while (true) { 10526 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 10527 rtnl_lock(); 10528 10529 /* Rebroadcast unregister notification */ 10530 list_for_each_entry(dev, list, todo_list) 10531 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 10532 10533 __rtnl_unlock(); 10534 rcu_barrier(); 10535 rtnl_lock(); 10536 10537 list_for_each_entry(dev, list, todo_list) 10538 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 10539 &dev->state)) { 10540 /* We must not have linkwatch events 10541 * pending on unregister. If this 10542 * happens, we simply run the queue 10543 * unscheduled, resulting in a noop 10544 * for this device. 10545 */ 10546 linkwatch_run_queue(); 10547 break; 10548 } 10549 10550 __rtnl_unlock(); 10551 10552 rebroadcast_time = jiffies; 10553 } 10554 10555 if (!wait) { 10556 rcu_barrier(); 10557 wait = WAIT_REFS_MIN_MSECS; 10558 } else { 10559 msleep(wait); 10560 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 10561 } 10562 10563 list_for_each_entry(dev, list, todo_list) 10564 if (netdev_refcnt_read(dev) == 1) 10565 return dev; 10566 10567 if (time_after(jiffies, warning_time + 10568 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 10569 list_for_each_entry(dev, list, todo_list) { 10570 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 10571 dev->name, netdev_refcnt_read(dev)); 10572 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 10573 } 10574 10575 warning_time = jiffies; 10576 } 10577 } 10578 } 10579 10580 /* The sequence is: 10581 * 10582 * rtnl_lock(); 10583 * ... 10584 * register_netdevice(x1); 10585 * register_netdevice(x2); 10586 * ... 10587 * unregister_netdevice(y1); 10588 * unregister_netdevice(y2); 10589 * ... 10590 * rtnl_unlock(); 10591 * free_netdev(y1); 10592 * free_netdev(y2); 10593 * 10594 * We are invoked by rtnl_unlock(). 10595 * This allows us to deal with problems: 10596 * 1) We can delete sysfs objects which invoke hotplug 10597 * without deadlocking with linkwatch via keventd. 10598 * 2) Since we run with the RTNL semaphore not held, we can sleep 10599 * safely in order to wait for the netdev refcnt to drop to zero. 10600 * 10601 * We must not return until all unregister events added during 10602 * the interval the lock was held have been completed. 10603 */ 10604 void netdev_run_todo(void) 10605 { 10606 struct net_device *dev, *tmp; 10607 struct list_head list; 10608 int cnt; 10609 #ifdef CONFIG_LOCKDEP 10610 struct list_head unlink_list; 10611 10612 list_replace_init(&net_unlink_list, &unlink_list); 10613 10614 while (!list_empty(&unlink_list)) { 10615 struct net_device *dev = list_first_entry(&unlink_list, 10616 struct net_device, 10617 unlink_list); 10618 list_del_init(&dev->unlink_list); 10619 dev->nested_level = dev->lower_level - 1; 10620 } 10621 #endif 10622 10623 /* Snapshot list, allow later requests */ 10624 list_replace_init(&net_todo_list, &list); 10625 10626 __rtnl_unlock(); 10627 10628 /* Wait for rcu callbacks to finish before next phase */ 10629 if (!list_empty(&list)) 10630 rcu_barrier(); 10631 10632 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 10633 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 10634 netdev_WARN(dev, "run_todo but not unregistering\n"); 10635 list_del(&dev->todo_list); 10636 continue; 10637 } 10638 10639 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 10640 linkwatch_sync_dev(dev); 10641 } 10642 10643 cnt = 0; 10644 while (!list_empty(&list)) { 10645 dev = netdev_wait_allrefs_any(&list); 10646 list_del(&dev->todo_list); 10647 10648 /* paranoia */ 10649 BUG_ON(netdev_refcnt_read(dev) != 1); 10650 BUG_ON(!list_empty(&dev->ptype_all)); 10651 BUG_ON(!list_empty(&dev->ptype_specific)); 10652 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 10653 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 10654 10655 netdev_do_free_pcpu_stats(dev); 10656 if (dev->priv_destructor) 10657 dev->priv_destructor(dev); 10658 if (dev->needs_free_netdev) 10659 free_netdev(dev); 10660 10661 cnt++; 10662 10663 /* Free network device */ 10664 kobject_put(&dev->dev.kobj); 10665 } 10666 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 10667 wake_up(&netdev_unregistering_wq); 10668 } 10669 10670 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 10671 * all the same fields in the same order as net_device_stats, with only 10672 * the type differing, but rtnl_link_stats64 may have additional fields 10673 * at the end for newer counters. 10674 */ 10675 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 10676 const struct net_device_stats *netdev_stats) 10677 { 10678 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 10679 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 10680 u64 *dst = (u64 *)stats64; 10681 10682 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 10683 for (i = 0; i < n; i++) 10684 dst[i] = (unsigned long)atomic_long_read(&src[i]); 10685 /* zero out counters that only exist in rtnl_link_stats64 */ 10686 memset((char *)stats64 + n * sizeof(u64), 0, 10687 sizeof(*stats64) - n * sizeof(u64)); 10688 } 10689 EXPORT_SYMBOL(netdev_stats_to_stats64); 10690 10691 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 10692 struct net_device *dev) 10693 { 10694 struct net_device_core_stats __percpu *p; 10695 10696 p = alloc_percpu_gfp(struct net_device_core_stats, 10697 GFP_ATOMIC | __GFP_NOWARN); 10698 10699 if (p && cmpxchg(&dev->core_stats, NULL, p)) 10700 free_percpu(p); 10701 10702 /* This READ_ONCE() pairs with the cmpxchg() above */ 10703 return READ_ONCE(dev->core_stats); 10704 } 10705 10706 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 10707 { 10708 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 10709 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 10710 unsigned long __percpu *field; 10711 10712 if (unlikely(!p)) { 10713 p = netdev_core_stats_alloc(dev); 10714 if (!p) 10715 return; 10716 } 10717 10718 field = (__force unsigned long __percpu *)((__force void *)p + offset); 10719 this_cpu_inc(*field); 10720 } 10721 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 10722 10723 /** 10724 * dev_get_stats - get network device statistics 10725 * @dev: device to get statistics from 10726 * @storage: place to store stats 10727 * 10728 * Get network statistics from device. Return @storage. 10729 * The device driver may provide its own method by setting 10730 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 10731 * otherwise the internal statistics structure is used. 10732 */ 10733 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 10734 struct rtnl_link_stats64 *storage) 10735 { 10736 const struct net_device_ops *ops = dev->netdev_ops; 10737 const struct net_device_core_stats __percpu *p; 10738 10739 if (ops->ndo_get_stats64) { 10740 memset(storage, 0, sizeof(*storage)); 10741 ops->ndo_get_stats64(dev, storage); 10742 } else if (ops->ndo_get_stats) { 10743 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 10744 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 10745 dev_get_tstats64(dev, storage); 10746 } else { 10747 netdev_stats_to_stats64(storage, &dev->stats); 10748 } 10749 10750 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 10751 p = READ_ONCE(dev->core_stats); 10752 if (p) { 10753 const struct net_device_core_stats *core_stats; 10754 int i; 10755 10756 for_each_possible_cpu(i) { 10757 core_stats = per_cpu_ptr(p, i); 10758 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 10759 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 10760 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 10761 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 10762 } 10763 } 10764 return storage; 10765 } 10766 EXPORT_SYMBOL(dev_get_stats); 10767 10768 /** 10769 * dev_fetch_sw_netstats - get per-cpu network device statistics 10770 * @s: place to store stats 10771 * @netstats: per-cpu network stats to read from 10772 * 10773 * Read per-cpu network statistics and populate the related fields in @s. 10774 */ 10775 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 10776 const struct pcpu_sw_netstats __percpu *netstats) 10777 { 10778 int cpu; 10779 10780 for_each_possible_cpu(cpu) { 10781 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 10782 const struct pcpu_sw_netstats *stats; 10783 unsigned int start; 10784 10785 stats = per_cpu_ptr(netstats, cpu); 10786 do { 10787 start = u64_stats_fetch_begin(&stats->syncp); 10788 rx_packets = u64_stats_read(&stats->rx_packets); 10789 rx_bytes = u64_stats_read(&stats->rx_bytes); 10790 tx_packets = u64_stats_read(&stats->tx_packets); 10791 tx_bytes = u64_stats_read(&stats->tx_bytes); 10792 } while (u64_stats_fetch_retry(&stats->syncp, start)); 10793 10794 s->rx_packets += rx_packets; 10795 s->rx_bytes += rx_bytes; 10796 s->tx_packets += tx_packets; 10797 s->tx_bytes += tx_bytes; 10798 } 10799 } 10800 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 10801 10802 /** 10803 * dev_get_tstats64 - ndo_get_stats64 implementation 10804 * @dev: device to get statistics from 10805 * @s: place to store stats 10806 * 10807 * Populate @s from dev->stats and dev->tstats. Can be used as 10808 * ndo_get_stats64() callback. 10809 */ 10810 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 10811 { 10812 netdev_stats_to_stats64(s, &dev->stats); 10813 dev_fetch_sw_netstats(s, dev->tstats); 10814 } 10815 EXPORT_SYMBOL_GPL(dev_get_tstats64); 10816 10817 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 10818 { 10819 struct netdev_queue *queue = dev_ingress_queue(dev); 10820 10821 #ifdef CONFIG_NET_CLS_ACT 10822 if (queue) 10823 return queue; 10824 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 10825 if (!queue) 10826 return NULL; 10827 netdev_init_one_queue(dev, queue, NULL); 10828 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 10829 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 10830 rcu_assign_pointer(dev->ingress_queue, queue); 10831 #endif 10832 return queue; 10833 } 10834 10835 static const struct ethtool_ops default_ethtool_ops; 10836 10837 void netdev_set_default_ethtool_ops(struct net_device *dev, 10838 const struct ethtool_ops *ops) 10839 { 10840 if (dev->ethtool_ops == &default_ethtool_ops) 10841 dev->ethtool_ops = ops; 10842 } 10843 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 10844 10845 /** 10846 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 10847 * @dev: netdev to enable the IRQ coalescing on 10848 * 10849 * Sets a conservative default for SW IRQ coalescing. Users can use 10850 * sysfs attributes to override the default values. 10851 */ 10852 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 10853 { 10854 WARN_ON(dev->reg_state == NETREG_REGISTERED); 10855 10856 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 10857 dev->gro_flush_timeout = 20000; 10858 dev->napi_defer_hard_irqs = 1; 10859 } 10860 } 10861 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 10862 10863 void netdev_freemem(struct net_device *dev) 10864 { 10865 char *addr = (char *)dev - dev->padded; 10866 10867 kvfree(addr); 10868 } 10869 10870 /** 10871 * alloc_netdev_mqs - allocate network device 10872 * @sizeof_priv: size of private data to allocate space for 10873 * @name: device name format string 10874 * @name_assign_type: origin of device name 10875 * @setup: callback to initialize device 10876 * @txqs: the number of TX subqueues to allocate 10877 * @rxqs: the number of RX subqueues to allocate 10878 * 10879 * Allocates a struct net_device with private data area for driver use 10880 * and performs basic initialization. Also allocates subqueue structs 10881 * for each queue on the device. 10882 */ 10883 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 10884 unsigned char name_assign_type, 10885 void (*setup)(struct net_device *), 10886 unsigned int txqs, unsigned int rxqs) 10887 { 10888 struct net_device *dev; 10889 unsigned int alloc_size; 10890 struct net_device *p; 10891 10892 BUG_ON(strlen(name) >= sizeof(dev->name)); 10893 10894 if (txqs < 1) { 10895 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 10896 return NULL; 10897 } 10898 10899 if (rxqs < 1) { 10900 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 10901 return NULL; 10902 } 10903 10904 alloc_size = sizeof(struct net_device); 10905 if (sizeof_priv) { 10906 /* ensure 32-byte alignment of private area */ 10907 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN); 10908 alloc_size += sizeof_priv; 10909 } 10910 /* ensure 32-byte alignment of whole construct */ 10911 alloc_size += NETDEV_ALIGN - 1; 10912 10913 p = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10914 if (!p) 10915 return NULL; 10916 10917 dev = PTR_ALIGN(p, NETDEV_ALIGN); 10918 dev->padded = (char *)dev - (char *)p; 10919 10920 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 10921 #ifdef CONFIG_PCPU_DEV_REFCNT 10922 dev->pcpu_refcnt = alloc_percpu(int); 10923 if (!dev->pcpu_refcnt) 10924 goto free_dev; 10925 __dev_hold(dev); 10926 #else 10927 refcount_set(&dev->dev_refcnt, 1); 10928 #endif 10929 10930 if (dev_addr_init(dev)) 10931 goto free_pcpu; 10932 10933 dev_mc_init(dev); 10934 dev_uc_init(dev); 10935 10936 dev_net_set(dev, &init_net); 10937 10938 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 10939 dev->xdp_zc_max_segs = 1; 10940 dev->gso_max_segs = GSO_MAX_SEGS; 10941 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 10942 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 10943 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 10944 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 10945 dev->tso_max_segs = TSO_MAX_SEGS; 10946 dev->upper_level = 1; 10947 dev->lower_level = 1; 10948 #ifdef CONFIG_LOCKDEP 10949 dev->nested_level = 0; 10950 INIT_LIST_HEAD(&dev->unlink_list); 10951 #endif 10952 10953 INIT_LIST_HEAD(&dev->napi_list); 10954 INIT_LIST_HEAD(&dev->unreg_list); 10955 INIT_LIST_HEAD(&dev->close_list); 10956 INIT_LIST_HEAD(&dev->link_watch_list); 10957 INIT_LIST_HEAD(&dev->adj_list.upper); 10958 INIT_LIST_HEAD(&dev->adj_list.lower); 10959 INIT_LIST_HEAD(&dev->ptype_all); 10960 INIT_LIST_HEAD(&dev->ptype_specific); 10961 INIT_LIST_HEAD(&dev->net_notifier_list); 10962 #ifdef CONFIG_NET_SCHED 10963 hash_init(dev->qdisc_hash); 10964 #endif 10965 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 10966 setup(dev); 10967 10968 if (!dev->tx_queue_len) { 10969 dev->priv_flags |= IFF_NO_QUEUE; 10970 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 10971 } 10972 10973 dev->num_tx_queues = txqs; 10974 dev->real_num_tx_queues = txqs; 10975 if (netif_alloc_netdev_queues(dev)) 10976 goto free_all; 10977 10978 dev->num_rx_queues = rxqs; 10979 dev->real_num_rx_queues = rxqs; 10980 if (netif_alloc_rx_queues(dev)) 10981 goto free_all; 10982 10983 strcpy(dev->name, name); 10984 dev->name_assign_type = name_assign_type; 10985 dev->group = INIT_NETDEV_GROUP; 10986 if (!dev->ethtool_ops) 10987 dev->ethtool_ops = &default_ethtool_ops; 10988 10989 nf_hook_netdev_init(dev); 10990 10991 return dev; 10992 10993 free_all: 10994 free_netdev(dev); 10995 return NULL; 10996 10997 free_pcpu: 10998 #ifdef CONFIG_PCPU_DEV_REFCNT 10999 free_percpu(dev->pcpu_refcnt); 11000 free_dev: 11001 #endif 11002 netdev_freemem(dev); 11003 return NULL; 11004 } 11005 EXPORT_SYMBOL(alloc_netdev_mqs); 11006 11007 /** 11008 * free_netdev - free network device 11009 * @dev: device 11010 * 11011 * This function does the last stage of destroying an allocated device 11012 * interface. The reference to the device object is released. If this 11013 * is the last reference then it will be freed.Must be called in process 11014 * context. 11015 */ 11016 void free_netdev(struct net_device *dev) 11017 { 11018 struct napi_struct *p, *n; 11019 11020 might_sleep(); 11021 11022 /* When called immediately after register_netdevice() failed the unwind 11023 * handling may still be dismantling the device. Handle that case by 11024 * deferring the free. 11025 */ 11026 if (dev->reg_state == NETREG_UNREGISTERING) { 11027 ASSERT_RTNL(); 11028 dev->needs_free_netdev = true; 11029 return; 11030 } 11031 11032 netif_free_tx_queues(dev); 11033 netif_free_rx_queues(dev); 11034 11035 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 11036 11037 /* Flush device addresses */ 11038 dev_addr_flush(dev); 11039 11040 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 11041 netif_napi_del(p); 11042 11043 ref_tracker_dir_exit(&dev->refcnt_tracker); 11044 #ifdef CONFIG_PCPU_DEV_REFCNT 11045 free_percpu(dev->pcpu_refcnt); 11046 dev->pcpu_refcnt = NULL; 11047 #endif 11048 free_percpu(dev->core_stats); 11049 dev->core_stats = NULL; 11050 free_percpu(dev->xdp_bulkq); 11051 dev->xdp_bulkq = NULL; 11052 11053 /* Compatibility with error handling in drivers */ 11054 if (dev->reg_state == NETREG_UNINITIALIZED) { 11055 netdev_freemem(dev); 11056 return; 11057 } 11058 11059 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 11060 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 11061 11062 /* will free via device release */ 11063 put_device(&dev->dev); 11064 } 11065 EXPORT_SYMBOL(free_netdev); 11066 11067 /** 11068 * synchronize_net - Synchronize with packet receive processing 11069 * 11070 * Wait for packets currently being received to be done. 11071 * Does not block later packets from starting. 11072 */ 11073 void synchronize_net(void) 11074 { 11075 might_sleep(); 11076 if (rtnl_is_locked()) 11077 synchronize_rcu_expedited(); 11078 else 11079 synchronize_rcu(); 11080 } 11081 EXPORT_SYMBOL(synchronize_net); 11082 11083 /** 11084 * unregister_netdevice_queue - remove device from the kernel 11085 * @dev: device 11086 * @head: list 11087 * 11088 * This function shuts down a device interface and removes it 11089 * from the kernel tables. 11090 * If head not NULL, device is queued to be unregistered later. 11091 * 11092 * Callers must hold the rtnl semaphore. You may want 11093 * unregister_netdev() instead of this. 11094 */ 11095 11096 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11097 { 11098 ASSERT_RTNL(); 11099 11100 if (head) { 11101 list_move_tail(&dev->unreg_list, head); 11102 } else { 11103 LIST_HEAD(single); 11104 11105 list_add(&dev->unreg_list, &single); 11106 unregister_netdevice_many(&single); 11107 } 11108 } 11109 EXPORT_SYMBOL(unregister_netdevice_queue); 11110 11111 void unregister_netdevice_many_notify(struct list_head *head, 11112 u32 portid, const struct nlmsghdr *nlh) 11113 { 11114 struct net_device *dev, *tmp; 11115 LIST_HEAD(close_head); 11116 int cnt = 0; 11117 11118 BUG_ON(dev_boot_phase); 11119 ASSERT_RTNL(); 11120 11121 if (list_empty(head)) 11122 return; 11123 11124 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11125 /* Some devices call without registering 11126 * for initialization unwind. Remove those 11127 * devices and proceed with the remaining. 11128 */ 11129 if (dev->reg_state == NETREG_UNINITIALIZED) { 11130 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11131 dev->name, dev); 11132 11133 WARN_ON(1); 11134 list_del(&dev->unreg_list); 11135 continue; 11136 } 11137 dev->dismantle = true; 11138 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11139 } 11140 11141 /* If device is running, close it first. */ 11142 list_for_each_entry(dev, head, unreg_list) 11143 list_add_tail(&dev->close_list, &close_head); 11144 dev_close_many(&close_head, true); 11145 11146 list_for_each_entry(dev, head, unreg_list) { 11147 /* And unlink it from device chain. */ 11148 unlist_netdevice(dev); 11149 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 11150 } 11151 flush_all_backlogs(); 11152 11153 synchronize_net(); 11154 11155 list_for_each_entry(dev, head, unreg_list) { 11156 struct sk_buff *skb = NULL; 11157 11158 /* Shutdown queueing discipline. */ 11159 dev_shutdown(dev); 11160 dev_tcx_uninstall(dev); 11161 dev_xdp_uninstall(dev); 11162 bpf_dev_bound_netdev_unregister(dev); 11163 11164 netdev_offload_xstats_disable_all(dev); 11165 11166 /* Notify protocols, that we are about to destroy 11167 * this device. They should clean all the things. 11168 */ 11169 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11170 11171 if (!dev->rtnl_link_ops || 11172 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11173 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 11174 GFP_KERNEL, NULL, 0, 11175 portid, nlh); 11176 11177 /* 11178 * Flush the unicast and multicast chains 11179 */ 11180 dev_uc_flush(dev); 11181 dev_mc_flush(dev); 11182 11183 netdev_name_node_alt_flush(dev); 11184 netdev_name_node_free(dev->name_node); 11185 11186 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11187 11188 if (dev->netdev_ops->ndo_uninit) 11189 dev->netdev_ops->ndo_uninit(dev); 11190 11191 if (skb) 11192 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 11193 11194 /* Notifier chain MUST detach us all upper devices. */ 11195 WARN_ON(netdev_has_any_upper_dev(dev)); 11196 WARN_ON(netdev_has_any_lower_dev(dev)); 11197 11198 /* Remove entries from kobject tree */ 11199 netdev_unregister_kobject(dev); 11200 #ifdef CONFIG_XPS 11201 /* Remove XPS queueing entries */ 11202 netif_reset_xps_queues_gt(dev, 0); 11203 #endif 11204 } 11205 11206 synchronize_net(); 11207 11208 list_for_each_entry(dev, head, unreg_list) { 11209 netdev_put(dev, &dev->dev_registered_tracker); 11210 net_set_todo(dev); 11211 cnt++; 11212 } 11213 atomic_add(cnt, &dev_unreg_count); 11214 11215 list_del(head); 11216 } 11217 11218 /** 11219 * unregister_netdevice_many - unregister many devices 11220 * @head: list of devices 11221 * 11222 * Note: As most callers use a stack allocated list_head, 11223 * we force a list_del() to make sure stack wont be corrupted later. 11224 */ 11225 void unregister_netdevice_many(struct list_head *head) 11226 { 11227 unregister_netdevice_many_notify(head, 0, NULL); 11228 } 11229 EXPORT_SYMBOL(unregister_netdevice_many); 11230 11231 /** 11232 * unregister_netdev - remove device from the kernel 11233 * @dev: device 11234 * 11235 * This function shuts down a device interface and removes it 11236 * from the kernel tables. 11237 * 11238 * This is just a wrapper for unregister_netdevice that takes 11239 * the rtnl semaphore. In general you want to use this and not 11240 * unregister_netdevice. 11241 */ 11242 void unregister_netdev(struct net_device *dev) 11243 { 11244 rtnl_lock(); 11245 unregister_netdevice(dev); 11246 rtnl_unlock(); 11247 } 11248 EXPORT_SYMBOL(unregister_netdev); 11249 11250 /** 11251 * __dev_change_net_namespace - move device to different nethost namespace 11252 * @dev: device 11253 * @net: network namespace 11254 * @pat: If not NULL name pattern to try if the current device name 11255 * is already taken in the destination network namespace. 11256 * @new_ifindex: If not zero, specifies device index in the target 11257 * namespace. 11258 * 11259 * This function shuts down a device interface and moves it 11260 * to a new network namespace. On success 0 is returned, on 11261 * a failure a netagive errno code is returned. 11262 * 11263 * Callers must hold the rtnl semaphore. 11264 */ 11265 11266 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 11267 const char *pat, int new_ifindex) 11268 { 11269 struct netdev_name_node *name_node; 11270 struct net *net_old = dev_net(dev); 11271 char new_name[IFNAMSIZ] = {}; 11272 int err, new_nsid; 11273 11274 ASSERT_RTNL(); 11275 11276 /* Don't allow namespace local devices to be moved. */ 11277 err = -EINVAL; 11278 if (dev->features & NETIF_F_NETNS_LOCAL) 11279 goto out; 11280 11281 /* Ensure the device has been registrered */ 11282 if (dev->reg_state != NETREG_REGISTERED) 11283 goto out; 11284 11285 /* Get out if there is nothing todo */ 11286 err = 0; 11287 if (net_eq(net_old, net)) 11288 goto out; 11289 11290 /* Pick the destination device name, and ensure 11291 * we can use it in the destination network namespace. 11292 */ 11293 err = -EEXIST; 11294 if (netdev_name_in_use(net, dev->name)) { 11295 /* We get here if we can't use the current device name */ 11296 if (!pat) 11297 goto out; 11298 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 11299 if (err < 0) 11300 goto out; 11301 } 11302 /* Check that none of the altnames conflicts. */ 11303 err = -EEXIST; 11304 netdev_for_each_altname(dev, name_node) 11305 if (netdev_name_in_use(net, name_node->name)) 11306 goto out; 11307 11308 /* Check that new_ifindex isn't used yet. */ 11309 if (new_ifindex) { 11310 err = dev_index_reserve(net, new_ifindex); 11311 if (err < 0) 11312 goto out; 11313 } else { 11314 /* If there is an ifindex conflict assign a new one */ 11315 err = dev_index_reserve(net, dev->ifindex); 11316 if (err == -EBUSY) 11317 err = dev_index_reserve(net, 0); 11318 if (err < 0) 11319 goto out; 11320 new_ifindex = err; 11321 } 11322 11323 /* 11324 * And now a mini version of register_netdevice unregister_netdevice. 11325 */ 11326 11327 /* If device is running close it first. */ 11328 dev_close(dev); 11329 11330 /* And unlink it from device chain */ 11331 unlist_netdevice(dev); 11332 11333 synchronize_net(); 11334 11335 /* Shutdown queueing discipline. */ 11336 dev_shutdown(dev); 11337 11338 /* Notify protocols, that we are about to destroy 11339 * this device. They should clean all the things. 11340 * 11341 * Note that dev->reg_state stays at NETREG_REGISTERED. 11342 * This is wanted because this way 8021q and macvlan know 11343 * the device is just moving and can keep their slaves up. 11344 */ 11345 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11346 rcu_barrier(); 11347 11348 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 11349 11350 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 11351 new_ifindex); 11352 11353 /* 11354 * Flush the unicast and multicast chains 11355 */ 11356 dev_uc_flush(dev); 11357 dev_mc_flush(dev); 11358 11359 /* Send a netdev-removed uevent to the old namespace */ 11360 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 11361 netdev_adjacent_del_links(dev); 11362 11363 /* Move per-net netdevice notifiers that are following the netdevice */ 11364 move_netdevice_notifiers_dev_net(dev, net); 11365 11366 /* Actually switch the network namespace */ 11367 dev_net_set(dev, net); 11368 dev->ifindex = new_ifindex; 11369 11370 if (new_name[0]) /* Rename the netdev to prepared name */ 11371 strscpy(dev->name, new_name, IFNAMSIZ); 11372 11373 /* Fixup kobjects */ 11374 dev_set_uevent_suppress(&dev->dev, 1); 11375 err = device_rename(&dev->dev, dev->name); 11376 dev_set_uevent_suppress(&dev->dev, 0); 11377 WARN_ON(err); 11378 11379 /* Send a netdev-add uevent to the new namespace */ 11380 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 11381 netdev_adjacent_add_links(dev); 11382 11383 /* Adapt owner in case owning user namespace of target network 11384 * namespace is different from the original one. 11385 */ 11386 err = netdev_change_owner(dev, net_old, net); 11387 WARN_ON(err); 11388 11389 /* Add the device back in the hashes */ 11390 list_netdevice(dev); 11391 11392 /* Notify protocols, that a new device appeared. */ 11393 call_netdevice_notifiers(NETDEV_REGISTER, dev); 11394 11395 /* 11396 * Prevent userspace races by waiting until the network 11397 * device is fully setup before sending notifications. 11398 */ 11399 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11400 11401 synchronize_net(); 11402 err = 0; 11403 out: 11404 return err; 11405 } 11406 EXPORT_SYMBOL_GPL(__dev_change_net_namespace); 11407 11408 static int dev_cpu_dead(unsigned int oldcpu) 11409 { 11410 struct sk_buff **list_skb; 11411 struct sk_buff *skb; 11412 unsigned int cpu; 11413 struct softnet_data *sd, *oldsd, *remsd = NULL; 11414 11415 local_irq_disable(); 11416 cpu = smp_processor_id(); 11417 sd = &per_cpu(softnet_data, cpu); 11418 oldsd = &per_cpu(softnet_data, oldcpu); 11419 11420 /* Find end of our completion_queue. */ 11421 list_skb = &sd->completion_queue; 11422 while (*list_skb) 11423 list_skb = &(*list_skb)->next; 11424 /* Append completion queue from offline CPU. */ 11425 *list_skb = oldsd->completion_queue; 11426 oldsd->completion_queue = NULL; 11427 11428 /* Append output queue from offline CPU. */ 11429 if (oldsd->output_queue) { 11430 *sd->output_queue_tailp = oldsd->output_queue; 11431 sd->output_queue_tailp = oldsd->output_queue_tailp; 11432 oldsd->output_queue = NULL; 11433 oldsd->output_queue_tailp = &oldsd->output_queue; 11434 } 11435 /* Append NAPI poll list from offline CPU, with one exception : 11436 * process_backlog() must be called by cpu owning percpu backlog. 11437 * We properly handle process_queue & input_pkt_queue later. 11438 */ 11439 while (!list_empty(&oldsd->poll_list)) { 11440 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 11441 struct napi_struct, 11442 poll_list); 11443 11444 list_del_init(&napi->poll_list); 11445 if (napi->poll == process_backlog) 11446 napi->state &= NAPIF_STATE_THREADED; 11447 else 11448 ____napi_schedule(sd, napi); 11449 } 11450 11451 raise_softirq_irqoff(NET_TX_SOFTIRQ); 11452 local_irq_enable(); 11453 11454 if (!use_backlog_threads()) { 11455 #ifdef CONFIG_RPS 11456 remsd = oldsd->rps_ipi_list; 11457 oldsd->rps_ipi_list = NULL; 11458 #endif 11459 /* send out pending IPI's on offline CPU */ 11460 net_rps_send_ipi(remsd); 11461 } 11462 11463 /* Process offline CPU's input_pkt_queue */ 11464 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 11465 netif_rx(skb); 11466 rps_input_queue_head_incr(oldsd); 11467 } 11468 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 11469 netif_rx(skb); 11470 rps_input_queue_head_incr(oldsd); 11471 } 11472 11473 return 0; 11474 } 11475 11476 /** 11477 * netdev_increment_features - increment feature set by one 11478 * @all: current feature set 11479 * @one: new feature set 11480 * @mask: mask feature set 11481 * 11482 * Computes a new feature set after adding a device with feature set 11483 * @one to the master device with current feature set @all. Will not 11484 * enable anything that is off in @mask. Returns the new feature set. 11485 */ 11486 netdev_features_t netdev_increment_features(netdev_features_t all, 11487 netdev_features_t one, netdev_features_t mask) 11488 { 11489 if (mask & NETIF_F_HW_CSUM) 11490 mask |= NETIF_F_CSUM_MASK; 11491 mask |= NETIF_F_VLAN_CHALLENGED; 11492 11493 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 11494 all &= one | ~NETIF_F_ALL_FOR_ALL; 11495 11496 /* If one device supports hw checksumming, set for all. */ 11497 if (all & NETIF_F_HW_CSUM) 11498 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 11499 11500 return all; 11501 } 11502 EXPORT_SYMBOL(netdev_increment_features); 11503 11504 static struct hlist_head * __net_init netdev_create_hash(void) 11505 { 11506 int i; 11507 struct hlist_head *hash; 11508 11509 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 11510 if (hash != NULL) 11511 for (i = 0; i < NETDEV_HASHENTRIES; i++) 11512 INIT_HLIST_HEAD(&hash[i]); 11513 11514 return hash; 11515 } 11516 11517 /* Initialize per network namespace state */ 11518 static int __net_init netdev_init(struct net *net) 11519 { 11520 BUILD_BUG_ON(GRO_HASH_BUCKETS > 11521 8 * sizeof_field(struct napi_struct, gro_bitmask)); 11522 11523 INIT_LIST_HEAD(&net->dev_base_head); 11524 11525 net->dev_name_head = netdev_create_hash(); 11526 if (net->dev_name_head == NULL) 11527 goto err_name; 11528 11529 net->dev_index_head = netdev_create_hash(); 11530 if (net->dev_index_head == NULL) 11531 goto err_idx; 11532 11533 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 11534 11535 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 11536 11537 return 0; 11538 11539 err_idx: 11540 kfree(net->dev_name_head); 11541 err_name: 11542 return -ENOMEM; 11543 } 11544 11545 /** 11546 * netdev_drivername - network driver for the device 11547 * @dev: network device 11548 * 11549 * Determine network driver for device. 11550 */ 11551 const char *netdev_drivername(const struct net_device *dev) 11552 { 11553 const struct device_driver *driver; 11554 const struct device *parent; 11555 const char *empty = ""; 11556 11557 parent = dev->dev.parent; 11558 if (!parent) 11559 return empty; 11560 11561 driver = parent->driver; 11562 if (driver && driver->name) 11563 return driver->name; 11564 return empty; 11565 } 11566 11567 static void __netdev_printk(const char *level, const struct net_device *dev, 11568 struct va_format *vaf) 11569 { 11570 if (dev && dev->dev.parent) { 11571 dev_printk_emit(level[1] - '0', 11572 dev->dev.parent, 11573 "%s %s %s%s: %pV", 11574 dev_driver_string(dev->dev.parent), 11575 dev_name(dev->dev.parent), 11576 netdev_name(dev), netdev_reg_state(dev), 11577 vaf); 11578 } else if (dev) { 11579 printk("%s%s%s: %pV", 11580 level, netdev_name(dev), netdev_reg_state(dev), vaf); 11581 } else { 11582 printk("%s(NULL net_device): %pV", level, vaf); 11583 } 11584 } 11585 11586 void netdev_printk(const char *level, const struct net_device *dev, 11587 const char *format, ...) 11588 { 11589 struct va_format vaf; 11590 va_list args; 11591 11592 va_start(args, format); 11593 11594 vaf.fmt = format; 11595 vaf.va = &args; 11596 11597 __netdev_printk(level, dev, &vaf); 11598 11599 va_end(args); 11600 } 11601 EXPORT_SYMBOL(netdev_printk); 11602 11603 #define define_netdev_printk_level(func, level) \ 11604 void func(const struct net_device *dev, const char *fmt, ...) \ 11605 { \ 11606 struct va_format vaf; \ 11607 va_list args; \ 11608 \ 11609 va_start(args, fmt); \ 11610 \ 11611 vaf.fmt = fmt; \ 11612 vaf.va = &args; \ 11613 \ 11614 __netdev_printk(level, dev, &vaf); \ 11615 \ 11616 va_end(args); \ 11617 } \ 11618 EXPORT_SYMBOL(func); 11619 11620 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 11621 define_netdev_printk_level(netdev_alert, KERN_ALERT); 11622 define_netdev_printk_level(netdev_crit, KERN_CRIT); 11623 define_netdev_printk_level(netdev_err, KERN_ERR); 11624 define_netdev_printk_level(netdev_warn, KERN_WARNING); 11625 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 11626 define_netdev_printk_level(netdev_info, KERN_INFO); 11627 11628 static void __net_exit netdev_exit(struct net *net) 11629 { 11630 kfree(net->dev_name_head); 11631 kfree(net->dev_index_head); 11632 xa_destroy(&net->dev_by_index); 11633 if (net != &init_net) 11634 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 11635 } 11636 11637 static struct pernet_operations __net_initdata netdev_net_ops = { 11638 .init = netdev_init, 11639 .exit = netdev_exit, 11640 }; 11641 11642 static void __net_exit default_device_exit_net(struct net *net) 11643 { 11644 struct netdev_name_node *name_node, *tmp; 11645 struct net_device *dev, *aux; 11646 /* 11647 * Push all migratable network devices back to the 11648 * initial network namespace 11649 */ 11650 ASSERT_RTNL(); 11651 for_each_netdev_safe(net, dev, aux) { 11652 int err; 11653 char fb_name[IFNAMSIZ]; 11654 11655 /* Ignore unmoveable devices (i.e. loopback) */ 11656 if (dev->features & NETIF_F_NETNS_LOCAL) 11657 continue; 11658 11659 /* Leave virtual devices for the generic cleanup */ 11660 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 11661 continue; 11662 11663 /* Push remaining network devices to init_net */ 11664 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 11665 if (netdev_name_in_use(&init_net, fb_name)) 11666 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 11667 11668 netdev_for_each_altname_safe(dev, name_node, tmp) 11669 if (netdev_name_in_use(&init_net, name_node->name)) 11670 __netdev_name_node_alt_destroy(name_node); 11671 11672 err = dev_change_net_namespace(dev, &init_net, fb_name); 11673 if (err) { 11674 pr_emerg("%s: failed to move %s to init_net: %d\n", 11675 __func__, dev->name, err); 11676 BUG(); 11677 } 11678 } 11679 } 11680 11681 static void __net_exit default_device_exit_batch(struct list_head *net_list) 11682 { 11683 /* At exit all network devices most be removed from a network 11684 * namespace. Do this in the reverse order of registration. 11685 * Do this across as many network namespaces as possible to 11686 * improve batching efficiency. 11687 */ 11688 struct net_device *dev; 11689 struct net *net; 11690 LIST_HEAD(dev_kill_list); 11691 11692 rtnl_lock(); 11693 list_for_each_entry(net, net_list, exit_list) { 11694 default_device_exit_net(net); 11695 cond_resched(); 11696 } 11697 11698 list_for_each_entry(net, net_list, exit_list) { 11699 for_each_netdev_reverse(net, dev) { 11700 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 11701 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 11702 else 11703 unregister_netdevice_queue(dev, &dev_kill_list); 11704 } 11705 } 11706 unregister_netdevice_many(&dev_kill_list); 11707 rtnl_unlock(); 11708 } 11709 11710 static struct pernet_operations __net_initdata default_device_ops = { 11711 .exit_batch = default_device_exit_batch, 11712 }; 11713 11714 static void __init net_dev_struct_check(void) 11715 { 11716 /* TX read-mostly hotpath */ 11717 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags); 11718 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 11719 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 11720 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 11721 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 11722 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 11723 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 11724 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 11725 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 11726 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 11727 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 11728 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 11729 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 11730 #ifdef CONFIG_XPS 11731 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 11732 #endif 11733 #ifdef CONFIG_NETFILTER_EGRESS 11734 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 11735 #endif 11736 #ifdef CONFIG_NET_XGRESS 11737 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 11738 #endif 11739 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 11740 11741 /* TXRX read-mostly hotpath */ 11742 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 11743 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 11744 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 11745 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 11746 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 11747 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 11748 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 11749 11750 /* RX read-mostly hotpath */ 11751 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 11752 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 11753 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 11754 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 11755 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_flush_timeout); 11756 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, napi_defer_hard_irqs); 11757 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 11758 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 11759 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 11760 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 11761 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 11762 #ifdef CONFIG_NETPOLL 11763 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 11764 #endif 11765 #ifdef CONFIG_NET_XGRESS 11766 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 11767 #endif 11768 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 104); 11769 } 11770 11771 /* 11772 * Initialize the DEV module. At boot time this walks the device list and 11773 * unhooks any devices that fail to initialise (normally hardware not 11774 * present) and leaves us with a valid list of present and active devices. 11775 * 11776 */ 11777 11778 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 11779 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 11780 11781 static int net_page_pool_create(int cpuid) 11782 { 11783 #if IS_ENABLED(CONFIG_PAGE_POOL) 11784 struct page_pool_params page_pool_params = { 11785 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 11786 .flags = PP_FLAG_SYSTEM_POOL, 11787 .nid = cpu_to_mem(cpuid), 11788 }; 11789 struct page_pool *pp_ptr; 11790 11791 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 11792 if (IS_ERR(pp_ptr)) 11793 return -ENOMEM; 11794 11795 per_cpu(system_page_pool, cpuid) = pp_ptr; 11796 #endif 11797 return 0; 11798 } 11799 11800 static int backlog_napi_should_run(unsigned int cpu) 11801 { 11802 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 11803 struct napi_struct *napi = &sd->backlog; 11804 11805 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 11806 } 11807 11808 static void run_backlog_napi(unsigned int cpu) 11809 { 11810 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 11811 11812 napi_threaded_poll_loop(&sd->backlog); 11813 } 11814 11815 static void backlog_napi_setup(unsigned int cpu) 11816 { 11817 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 11818 struct napi_struct *napi = &sd->backlog; 11819 11820 napi->thread = this_cpu_read(backlog_napi); 11821 set_bit(NAPI_STATE_THREADED, &napi->state); 11822 } 11823 11824 static struct smp_hotplug_thread backlog_threads = { 11825 .store = &backlog_napi, 11826 .thread_should_run = backlog_napi_should_run, 11827 .thread_fn = run_backlog_napi, 11828 .thread_comm = "backlog_napi/%u", 11829 .setup = backlog_napi_setup, 11830 }; 11831 11832 /* 11833 * This is called single threaded during boot, so no need 11834 * to take the rtnl semaphore. 11835 */ 11836 static int __init net_dev_init(void) 11837 { 11838 int i, rc = -ENOMEM; 11839 11840 BUG_ON(!dev_boot_phase); 11841 11842 net_dev_struct_check(); 11843 11844 if (dev_proc_init()) 11845 goto out; 11846 11847 if (netdev_kobject_init()) 11848 goto out; 11849 11850 for (i = 0; i < PTYPE_HASH_SIZE; i++) 11851 INIT_LIST_HEAD(&ptype_base[i]); 11852 11853 if (register_pernet_subsys(&netdev_net_ops)) 11854 goto out; 11855 11856 /* 11857 * Initialise the packet receive queues. 11858 */ 11859 11860 for_each_possible_cpu(i) { 11861 struct work_struct *flush = per_cpu_ptr(&flush_works, i); 11862 struct softnet_data *sd = &per_cpu(softnet_data, i); 11863 11864 INIT_WORK(flush, flush_backlog); 11865 11866 skb_queue_head_init(&sd->input_pkt_queue); 11867 skb_queue_head_init(&sd->process_queue); 11868 #ifdef CONFIG_XFRM_OFFLOAD 11869 skb_queue_head_init(&sd->xfrm_backlog); 11870 #endif 11871 INIT_LIST_HEAD(&sd->poll_list); 11872 sd->output_queue_tailp = &sd->output_queue; 11873 #ifdef CONFIG_RPS 11874 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 11875 sd->cpu = i; 11876 #endif 11877 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 11878 spin_lock_init(&sd->defer_lock); 11879 11880 init_gro_hash(&sd->backlog); 11881 sd->backlog.poll = process_backlog; 11882 sd->backlog.weight = weight_p; 11883 INIT_LIST_HEAD(&sd->backlog.poll_list); 11884 11885 if (net_page_pool_create(i)) 11886 goto out; 11887 } 11888 if (use_backlog_threads()) 11889 smpboot_register_percpu_thread(&backlog_threads); 11890 11891 dev_boot_phase = 0; 11892 11893 /* The loopback device is special if any other network devices 11894 * is present in a network namespace the loopback device must 11895 * be present. Since we now dynamically allocate and free the 11896 * loopback device ensure this invariant is maintained by 11897 * keeping the loopback device as the first device on the 11898 * list of network devices. Ensuring the loopback devices 11899 * is the first device that appears and the last network device 11900 * that disappears. 11901 */ 11902 if (register_pernet_device(&loopback_net_ops)) 11903 goto out; 11904 11905 if (register_pernet_device(&default_device_ops)) 11906 goto out; 11907 11908 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 11909 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 11910 11911 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 11912 NULL, dev_cpu_dead); 11913 WARN_ON(rc < 0); 11914 rc = 0; 11915 11916 /* avoid static key IPIs to isolated CPUs */ 11917 if (housekeeping_enabled(HK_TYPE_MISC)) 11918 net_enable_timestamp(); 11919 out: 11920 if (rc < 0) { 11921 for_each_possible_cpu(i) { 11922 struct page_pool *pp_ptr; 11923 11924 pp_ptr = per_cpu(system_page_pool, i); 11925 if (!pp_ptr) 11926 continue; 11927 11928 page_pool_destroy(pp_ptr); 11929 per_cpu(system_page_pool, i) = NULL; 11930 } 11931 } 11932 11933 return rc; 11934 } 11935 11936 subsys_initcall(net_dev_init); 11937