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 #include <linux/phy_link_topology.h> 162 163 #include "dev.h" 164 #include "devmem.h" 165 #include "net-sysfs.h" 166 167 static DEFINE_SPINLOCK(ptype_lock); 168 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 169 170 static int netif_rx_internal(struct sk_buff *skb); 171 static int call_netdevice_notifiers_extack(unsigned long val, 172 struct net_device *dev, 173 struct netlink_ext_ack *extack); 174 175 static DEFINE_MUTEX(ifalias_mutex); 176 177 /* protects napi_hash addition/deletion and napi_gen_id */ 178 static DEFINE_SPINLOCK(napi_hash_lock); 179 180 static unsigned int napi_gen_id = NR_CPUS; 181 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 182 183 static DECLARE_RWSEM(devnet_rename_sem); 184 185 static inline void dev_base_seq_inc(struct net *net) 186 { 187 unsigned int val = net->dev_base_seq + 1; 188 189 WRITE_ONCE(net->dev_base_seq, val ?: 1); 190 } 191 192 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 193 { 194 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 195 196 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 197 } 198 199 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 200 { 201 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 202 } 203 204 #ifndef CONFIG_PREEMPT_RT 205 206 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 207 208 static int __init setup_backlog_napi_threads(char *arg) 209 { 210 static_branch_enable(&use_backlog_threads_key); 211 return 0; 212 } 213 early_param("thread_backlog_napi", setup_backlog_napi_threads); 214 215 static bool use_backlog_threads(void) 216 { 217 return static_branch_unlikely(&use_backlog_threads_key); 218 } 219 220 #else 221 222 static bool use_backlog_threads(void) 223 { 224 return true; 225 } 226 227 #endif 228 229 static inline void backlog_lock_irq_save(struct softnet_data *sd, 230 unsigned long *flags) 231 { 232 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 233 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 234 else 235 local_irq_save(*flags); 236 } 237 238 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 239 { 240 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 241 spin_lock_irq(&sd->input_pkt_queue.lock); 242 else 243 local_irq_disable(); 244 } 245 246 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 247 unsigned long *flags) 248 { 249 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 250 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags); 251 else 252 local_irq_restore(*flags); 253 } 254 255 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 256 { 257 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 258 spin_unlock_irq(&sd->input_pkt_queue.lock); 259 else 260 local_irq_enable(); 261 } 262 263 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 264 const char *name) 265 { 266 struct netdev_name_node *name_node; 267 268 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 269 if (!name_node) 270 return NULL; 271 INIT_HLIST_NODE(&name_node->hlist); 272 name_node->dev = dev; 273 name_node->name = name; 274 return name_node; 275 } 276 277 static struct netdev_name_node * 278 netdev_name_node_head_alloc(struct net_device *dev) 279 { 280 struct netdev_name_node *name_node; 281 282 name_node = netdev_name_node_alloc(dev, dev->name); 283 if (!name_node) 284 return NULL; 285 INIT_LIST_HEAD(&name_node->list); 286 return name_node; 287 } 288 289 static void netdev_name_node_free(struct netdev_name_node *name_node) 290 { 291 kfree(name_node); 292 } 293 294 static void netdev_name_node_add(struct net *net, 295 struct netdev_name_node *name_node) 296 { 297 hlist_add_head_rcu(&name_node->hlist, 298 dev_name_hash(net, name_node->name)); 299 } 300 301 static void netdev_name_node_del(struct netdev_name_node *name_node) 302 { 303 hlist_del_rcu(&name_node->hlist); 304 } 305 306 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 307 const char *name) 308 { 309 struct hlist_head *head = dev_name_hash(net, name); 310 struct netdev_name_node *name_node; 311 312 hlist_for_each_entry(name_node, head, hlist) 313 if (!strcmp(name_node->name, name)) 314 return name_node; 315 return NULL; 316 } 317 318 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 319 const char *name) 320 { 321 struct hlist_head *head = dev_name_hash(net, name); 322 struct netdev_name_node *name_node; 323 324 hlist_for_each_entry_rcu(name_node, head, hlist) 325 if (!strcmp(name_node->name, name)) 326 return name_node; 327 return NULL; 328 } 329 330 bool netdev_name_in_use(struct net *net, const char *name) 331 { 332 return netdev_name_node_lookup(net, name); 333 } 334 EXPORT_SYMBOL(netdev_name_in_use); 335 336 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 337 { 338 struct netdev_name_node *name_node; 339 struct net *net = dev_net(dev); 340 341 name_node = netdev_name_node_lookup(net, name); 342 if (name_node) 343 return -EEXIST; 344 name_node = netdev_name_node_alloc(dev, name); 345 if (!name_node) 346 return -ENOMEM; 347 netdev_name_node_add(net, name_node); 348 /* The node that holds dev->name acts as a head of per-device list. */ 349 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 350 351 return 0; 352 } 353 354 static void netdev_name_node_alt_free(struct rcu_head *head) 355 { 356 struct netdev_name_node *name_node = 357 container_of(head, struct netdev_name_node, rcu); 358 359 kfree(name_node->name); 360 netdev_name_node_free(name_node); 361 } 362 363 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 364 { 365 netdev_name_node_del(name_node); 366 list_del(&name_node->list); 367 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 368 } 369 370 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 371 { 372 struct netdev_name_node *name_node; 373 struct net *net = dev_net(dev); 374 375 name_node = netdev_name_node_lookup(net, name); 376 if (!name_node) 377 return -ENOENT; 378 /* lookup might have found our primary name or a name belonging 379 * to another device. 380 */ 381 if (name_node == dev->name_node || name_node->dev != dev) 382 return -EINVAL; 383 384 __netdev_name_node_alt_destroy(name_node); 385 return 0; 386 } 387 388 static void netdev_name_node_alt_flush(struct net_device *dev) 389 { 390 struct netdev_name_node *name_node, *tmp; 391 392 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 393 list_del(&name_node->list); 394 netdev_name_node_alt_free(&name_node->rcu); 395 } 396 } 397 398 /* Device list insertion */ 399 static void list_netdevice(struct net_device *dev) 400 { 401 struct netdev_name_node *name_node; 402 struct net *net = dev_net(dev); 403 404 ASSERT_RTNL(); 405 406 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 407 netdev_name_node_add(net, dev->name_node); 408 hlist_add_head_rcu(&dev->index_hlist, 409 dev_index_hash(net, dev->ifindex)); 410 411 netdev_for_each_altname(dev, name_node) 412 netdev_name_node_add(net, name_node); 413 414 /* We reserved the ifindex, this can't fail */ 415 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 416 417 dev_base_seq_inc(net); 418 } 419 420 /* Device list removal 421 * caller must respect a RCU grace period before freeing/reusing dev 422 */ 423 static void unlist_netdevice(struct net_device *dev) 424 { 425 struct netdev_name_node *name_node; 426 struct net *net = dev_net(dev); 427 428 ASSERT_RTNL(); 429 430 xa_erase(&net->dev_by_index, dev->ifindex); 431 432 netdev_for_each_altname(dev, name_node) 433 netdev_name_node_del(name_node); 434 435 /* Unlink dev from the device chain */ 436 list_del_rcu(&dev->dev_list); 437 netdev_name_node_del(dev->name_node); 438 hlist_del_rcu(&dev->index_hlist); 439 440 dev_base_seq_inc(dev_net(dev)); 441 } 442 443 /* 444 * Our notifier list 445 */ 446 447 static RAW_NOTIFIER_HEAD(netdev_chain); 448 449 /* 450 * Device drivers call our routines to queue packets here. We empty the 451 * queue in the local softnet handler. 452 */ 453 454 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 455 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 456 }; 457 EXPORT_PER_CPU_SYMBOL(softnet_data); 458 459 /* Page_pool has a lockless array/stack to alloc/recycle pages. 460 * PP consumers must pay attention to run APIs in the appropriate context 461 * (e.g. NAPI context). 462 */ 463 static DEFINE_PER_CPU(struct page_pool *, system_page_pool); 464 465 #ifdef CONFIG_LOCKDEP 466 /* 467 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 468 * according to dev->type 469 */ 470 static const unsigned short netdev_lock_type[] = { 471 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 472 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 473 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 474 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 475 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 476 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 477 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 478 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 479 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 480 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 481 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 482 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 483 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 484 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 485 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE}; 486 487 static const char *const netdev_lock_name[] = { 488 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 489 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 490 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 491 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 492 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 493 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 494 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 495 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 496 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 497 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 498 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 499 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 500 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 501 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE", 502 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"}; 503 504 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 505 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 506 507 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 508 { 509 int i; 510 511 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 512 if (netdev_lock_type[i] == dev_type) 513 return i; 514 /* the last key is used by default */ 515 return ARRAY_SIZE(netdev_lock_type) - 1; 516 } 517 518 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 519 unsigned short dev_type) 520 { 521 int i; 522 523 i = netdev_lock_pos(dev_type); 524 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 525 netdev_lock_name[i]); 526 } 527 528 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 529 { 530 int i; 531 532 i = netdev_lock_pos(dev->type); 533 lockdep_set_class_and_name(&dev->addr_list_lock, 534 &netdev_addr_lock_key[i], 535 netdev_lock_name[i]); 536 } 537 #else 538 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 539 unsigned short dev_type) 540 { 541 } 542 543 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 544 { 545 } 546 #endif 547 548 /******************************************************************************* 549 * 550 * Protocol management and registration routines 551 * 552 *******************************************************************************/ 553 554 555 /* 556 * Add a protocol ID to the list. Now that the input handler is 557 * smarter we can dispense with all the messy stuff that used to be 558 * here. 559 * 560 * BEWARE!!! Protocol handlers, mangling input packets, 561 * MUST BE last in hash buckets and checking protocol handlers 562 * MUST start from promiscuous ptype_all chain in net_bh. 563 * It is true now, do not change it. 564 * Explanation follows: if protocol handler, mangling packet, will 565 * be the first on list, it is not able to sense, that packet 566 * is cloned and should be copied-on-write, so that it will 567 * change it and subsequent readers will get broken packet. 568 * --ANK (980803) 569 */ 570 571 static inline struct list_head *ptype_head(const struct packet_type *pt) 572 { 573 if (pt->type == htons(ETH_P_ALL)) 574 return pt->dev ? &pt->dev->ptype_all : &net_hotdata.ptype_all; 575 else 576 return pt->dev ? &pt->dev->ptype_specific : 577 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 578 } 579 580 /** 581 * dev_add_pack - add packet handler 582 * @pt: packet type declaration 583 * 584 * Add a protocol handler to the networking stack. The passed &packet_type 585 * is linked into kernel lists and may not be freed until it has been 586 * removed from the kernel lists. 587 * 588 * This call does not sleep therefore it can not 589 * guarantee all CPU's that are in middle of receiving packets 590 * will see the new packet type (until the next received packet). 591 */ 592 593 void dev_add_pack(struct packet_type *pt) 594 { 595 struct list_head *head = ptype_head(pt); 596 597 spin_lock(&ptype_lock); 598 list_add_rcu(&pt->list, head); 599 spin_unlock(&ptype_lock); 600 } 601 EXPORT_SYMBOL(dev_add_pack); 602 603 /** 604 * __dev_remove_pack - remove packet handler 605 * @pt: packet type declaration 606 * 607 * Remove a protocol handler that was previously added to the kernel 608 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 609 * from the kernel lists and can be freed or reused once this function 610 * returns. 611 * 612 * The packet type might still be in use by receivers 613 * and must not be freed until after all the CPU's have gone 614 * through a quiescent state. 615 */ 616 void __dev_remove_pack(struct packet_type *pt) 617 { 618 struct list_head *head = ptype_head(pt); 619 struct packet_type *pt1; 620 621 spin_lock(&ptype_lock); 622 623 list_for_each_entry(pt1, head, list) { 624 if (pt == pt1) { 625 list_del_rcu(&pt->list); 626 goto out; 627 } 628 } 629 630 pr_warn("dev_remove_pack: %p not found\n", pt); 631 out: 632 spin_unlock(&ptype_lock); 633 } 634 EXPORT_SYMBOL(__dev_remove_pack); 635 636 /** 637 * dev_remove_pack - remove packet handler 638 * @pt: packet type declaration 639 * 640 * Remove a protocol handler that was previously added to the kernel 641 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 642 * from the kernel lists and can be freed or reused once this function 643 * returns. 644 * 645 * This call sleeps to guarantee that no CPU is looking at the packet 646 * type after return. 647 */ 648 void dev_remove_pack(struct packet_type *pt) 649 { 650 __dev_remove_pack(pt); 651 652 synchronize_net(); 653 } 654 EXPORT_SYMBOL(dev_remove_pack); 655 656 657 /******************************************************************************* 658 * 659 * Device Interface Subroutines 660 * 661 *******************************************************************************/ 662 663 /** 664 * dev_get_iflink - get 'iflink' value of a interface 665 * @dev: targeted interface 666 * 667 * Indicates the ifindex the interface is linked to. 668 * Physical interfaces have the same 'ifindex' and 'iflink' values. 669 */ 670 671 int dev_get_iflink(const struct net_device *dev) 672 { 673 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 674 return dev->netdev_ops->ndo_get_iflink(dev); 675 676 return READ_ONCE(dev->ifindex); 677 } 678 EXPORT_SYMBOL(dev_get_iflink); 679 680 /** 681 * dev_fill_metadata_dst - Retrieve tunnel egress information. 682 * @dev: targeted interface 683 * @skb: The packet. 684 * 685 * For better visibility of tunnel traffic OVS needs to retrieve 686 * egress tunnel information for a packet. Following API allows 687 * user to get this info. 688 */ 689 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 690 { 691 struct ip_tunnel_info *info; 692 693 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 694 return -EINVAL; 695 696 info = skb_tunnel_info_unclone(skb); 697 if (!info) 698 return -ENOMEM; 699 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 700 return -EINVAL; 701 702 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 703 } 704 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 705 706 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 707 { 708 int k = stack->num_paths++; 709 710 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 711 return NULL; 712 713 return &stack->path[k]; 714 } 715 716 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 717 struct net_device_path_stack *stack) 718 { 719 const struct net_device *last_dev; 720 struct net_device_path_ctx ctx = { 721 .dev = dev, 722 }; 723 struct net_device_path *path; 724 int ret = 0; 725 726 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 727 stack->num_paths = 0; 728 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 729 last_dev = ctx.dev; 730 path = dev_fwd_path(stack); 731 if (!path) 732 return -1; 733 734 memset(path, 0, sizeof(struct net_device_path)); 735 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 736 if (ret < 0) 737 return -1; 738 739 if (WARN_ON_ONCE(last_dev == ctx.dev)) 740 return -1; 741 } 742 743 if (!ctx.dev) 744 return ret; 745 746 path = dev_fwd_path(stack); 747 if (!path) 748 return -1; 749 path->type = DEV_PATH_ETHERNET; 750 path->dev = ctx.dev; 751 752 return ret; 753 } 754 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 755 756 /** 757 * __dev_get_by_name - find a device by its name 758 * @net: the applicable net namespace 759 * @name: name to find 760 * 761 * Find an interface by name. Must be called under RTNL semaphore. 762 * If the name is found a pointer to the device is returned. 763 * If the name is not found then %NULL is returned. The 764 * reference counters are not incremented so the caller must be 765 * careful with locks. 766 */ 767 768 struct net_device *__dev_get_by_name(struct net *net, const char *name) 769 { 770 struct netdev_name_node *node_name; 771 772 node_name = netdev_name_node_lookup(net, name); 773 return node_name ? node_name->dev : NULL; 774 } 775 EXPORT_SYMBOL(__dev_get_by_name); 776 777 /** 778 * dev_get_by_name_rcu - find a device by its name 779 * @net: the applicable net namespace 780 * @name: name to find 781 * 782 * Find an interface by name. 783 * If the name is found a pointer to the device is returned. 784 * If the name is not found then %NULL is returned. 785 * The reference counters are not incremented so the caller must be 786 * careful with locks. The caller must hold RCU lock. 787 */ 788 789 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 790 { 791 struct netdev_name_node *node_name; 792 793 node_name = netdev_name_node_lookup_rcu(net, name); 794 return node_name ? node_name->dev : NULL; 795 } 796 EXPORT_SYMBOL(dev_get_by_name_rcu); 797 798 /* Deprecated for new users, call netdev_get_by_name() instead */ 799 struct net_device *dev_get_by_name(struct net *net, const char *name) 800 { 801 struct net_device *dev; 802 803 rcu_read_lock(); 804 dev = dev_get_by_name_rcu(net, name); 805 dev_hold(dev); 806 rcu_read_unlock(); 807 return dev; 808 } 809 EXPORT_SYMBOL(dev_get_by_name); 810 811 /** 812 * netdev_get_by_name() - find a device by its name 813 * @net: the applicable net namespace 814 * @name: name to find 815 * @tracker: tracking object for the acquired reference 816 * @gfp: allocation flags for the tracker 817 * 818 * Find an interface by name. This can be called from any 819 * context and does its own locking. The returned handle has 820 * the usage count incremented and the caller must use netdev_put() to 821 * release it when it is no longer needed. %NULL is returned if no 822 * matching device is found. 823 */ 824 struct net_device *netdev_get_by_name(struct net *net, const char *name, 825 netdevice_tracker *tracker, gfp_t gfp) 826 { 827 struct net_device *dev; 828 829 dev = dev_get_by_name(net, name); 830 if (dev) 831 netdev_tracker_alloc(dev, tracker, gfp); 832 return dev; 833 } 834 EXPORT_SYMBOL(netdev_get_by_name); 835 836 /** 837 * __dev_get_by_index - find a device by its ifindex 838 * @net: the applicable net namespace 839 * @ifindex: index of device 840 * 841 * Search for an interface by index. Returns %NULL if the device 842 * is not found or a pointer to the device. The device has not 843 * had its reference counter increased so the caller must be careful 844 * about locking. The caller must hold the RTNL semaphore. 845 */ 846 847 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 848 { 849 struct net_device *dev; 850 struct hlist_head *head = dev_index_hash(net, ifindex); 851 852 hlist_for_each_entry(dev, head, index_hlist) 853 if (dev->ifindex == ifindex) 854 return dev; 855 856 return NULL; 857 } 858 EXPORT_SYMBOL(__dev_get_by_index); 859 860 /** 861 * dev_get_by_index_rcu - find a device by its ifindex 862 * @net: the applicable net namespace 863 * @ifindex: index of device 864 * 865 * Search for an interface by index. Returns %NULL if the device 866 * is not found or a pointer to the device. The device has not 867 * had its reference counter increased so the caller must be careful 868 * about locking. The caller must hold RCU lock. 869 */ 870 871 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 872 { 873 struct net_device *dev; 874 struct hlist_head *head = dev_index_hash(net, ifindex); 875 876 hlist_for_each_entry_rcu(dev, head, index_hlist) 877 if (dev->ifindex == ifindex) 878 return dev; 879 880 return NULL; 881 } 882 EXPORT_SYMBOL(dev_get_by_index_rcu); 883 884 /* Deprecated for new users, call netdev_get_by_index() instead */ 885 struct net_device *dev_get_by_index(struct net *net, int ifindex) 886 { 887 struct net_device *dev; 888 889 rcu_read_lock(); 890 dev = dev_get_by_index_rcu(net, ifindex); 891 dev_hold(dev); 892 rcu_read_unlock(); 893 return dev; 894 } 895 EXPORT_SYMBOL(dev_get_by_index); 896 897 /** 898 * netdev_get_by_index() - find a device by its ifindex 899 * @net: the applicable net namespace 900 * @ifindex: index of device 901 * @tracker: tracking object for the acquired reference 902 * @gfp: allocation flags for the tracker 903 * 904 * Search for an interface by index. Returns NULL if the device 905 * is not found or a pointer to the device. The device returned has 906 * had a reference added and the pointer is safe until the user calls 907 * netdev_put() to indicate they have finished with it. 908 */ 909 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 910 netdevice_tracker *tracker, gfp_t gfp) 911 { 912 struct net_device *dev; 913 914 dev = dev_get_by_index(net, ifindex); 915 if (dev) 916 netdev_tracker_alloc(dev, tracker, gfp); 917 return dev; 918 } 919 EXPORT_SYMBOL(netdev_get_by_index); 920 921 /** 922 * dev_get_by_napi_id - find a device by napi_id 923 * @napi_id: ID of the NAPI struct 924 * 925 * Search for an interface by NAPI ID. Returns %NULL if the device 926 * is not found or a pointer to the device. The device has not had 927 * its reference counter increased so the caller must be careful 928 * about locking. The caller must hold RCU lock. 929 */ 930 931 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 932 { 933 struct napi_struct *napi; 934 935 WARN_ON_ONCE(!rcu_read_lock_held()); 936 937 if (napi_id < MIN_NAPI_ID) 938 return NULL; 939 940 napi = napi_by_id(napi_id); 941 942 return napi ? napi->dev : NULL; 943 } 944 EXPORT_SYMBOL(dev_get_by_napi_id); 945 946 static DEFINE_SEQLOCK(netdev_rename_lock); 947 948 void netdev_copy_name(struct net_device *dev, char *name) 949 { 950 unsigned int seq; 951 952 do { 953 seq = read_seqbegin(&netdev_rename_lock); 954 strscpy(name, dev->name, IFNAMSIZ); 955 } while (read_seqretry(&netdev_rename_lock, seq)); 956 } 957 958 /** 959 * netdev_get_name - get a netdevice name, knowing its ifindex. 960 * @net: network namespace 961 * @name: a pointer to the buffer where the name will be stored. 962 * @ifindex: the ifindex of the interface to get the name from. 963 */ 964 int netdev_get_name(struct net *net, char *name, int ifindex) 965 { 966 struct net_device *dev; 967 int ret; 968 969 rcu_read_lock(); 970 971 dev = dev_get_by_index_rcu(net, ifindex); 972 if (!dev) { 973 ret = -ENODEV; 974 goto out; 975 } 976 977 netdev_copy_name(dev, name); 978 979 ret = 0; 980 out: 981 rcu_read_unlock(); 982 return ret; 983 } 984 985 /** 986 * dev_getbyhwaddr_rcu - find a device by its hardware address 987 * @net: the applicable net namespace 988 * @type: media type of device 989 * @ha: hardware address 990 * 991 * Search for an interface by MAC address. Returns NULL if the device 992 * is not found or a pointer to the device. 993 * The caller must hold RCU or RTNL. 994 * The returned device has not had its ref count increased 995 * and the caller must therefore be careful about locking 996 * 997 */ 998 999 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1000 const char *ha) 1001 { 1002 struct net_device *dev; 1003 1004 for_each_netdev_rcu(net, dev) 1005 if (dev->type == type && 1006 !memcmp(dev->dev_addr, ha, dev->addr_len)) 1007 return dev; 1008 1009 return NULL; 1010 } 1011 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1012 1013 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1014 { 1015 struct net_device *dev, *ret = NULL; 1016 1017 rcu_read_lock(); 1018 for_each_netdev_rcu(net, dev) 1019 if (dev->type == type) { 1020 dev_hold(dev); 1021 ret = dev; 1022 break; 1023 } 1024 rcu_read_unlock(); 1025 return ret; 1026 } 1027 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1028 1029 /** 1030 * __dev_get_by_flags - find any device with given flags 1031 * @net: the applicable net namespace 1032 * @if_flags: IFF_* values 1033 * @mask: bitmask of bits in if_flags to check 1034 * 1035 * Search for any interface with the given flags. Returns NULL if a device 1036 * is not found or a pointer to the device. Must be called inside 1037 * rtnl_lock(), and result refcount is unchanged. 1038 */ 1039 1040 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags, 1041 unsigned short mask) 1042 { 1043 struct net_device *dev, *ret; 1044 1045 ASSERT_RTNL(); 1046 1047 ret = NULL; 1048 for_each_netdev(net, dev) { 1049 if (((dev->flags ^ if_flags) & mask) == 0) { 1050 ret = dev; 1051 break; 1052 } 1053 } 1054 return ret; 1055 } 1056 EXPORT_SYMBOL(__dev_get_by_flags); 1057 1058 /** 1059 * dev_valid_name - check if name is okay for network device 1060 * @name: name string 1061 * 1062 * Network device names need to be valid file names to 1063 * allow sysfs to work. We also disallow any kind of 1064 * whitespace. 1065 */ 1066 bool dev_valid_name(const char *name) 1067 { 1068 if (*name == '\0') 1069 return false; 1070 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1071 return false; 1072 if (!strcmp(name, ".") || !strcmp(name, "..")) 1073 return false; 1074 1075 while (*name) { 1076 if (*name == '/' || *name == ':' || isspace(*name)) 1077 return false; 1078 name++; 1079 } 1080 return true; 1081 } 1082 EXPORT_SYMBOL(dev_valid_name); 1083 1084 /** 1085 * __dev_alloc_name - allocate a name for a device 1086 * @net: network namespace to allocate the device name in 1087 * @name: name format string 1088 * @res: result name string 1089 * 1090 * Passed a format string - eg "lt%d" it will try and find a suitable 1091 * id. It scans list of devices to build up a free map, then chooses 1092 * the first empty slot. The caller must hold the dev_base or rtnl lock 1093 * while allocating the name and adding the device in order to avoid 1094 * duplicates. 1095 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1096 * Returns the number of the unit assigned or a negative errno code. 1097 */ 1098 1099 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1100 { 1101 int i = 0; 1102 const char *p; 1103 const int max_netdevices = 8*PAGE_SIZE; 1104 unsigned long *inuse; 1105 struct net_device *d; 1106 char buf[IFNAMSIZ]; 1107 1108 /* Verify the string as this thing may have come from the user. 1109 * There must be one "%d" and no other "%" characters. 1110 */ 1111 p = strchr(name, '%'); 1112 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1113 return -EINVAL; 1114 1115 /* Use one page as a bit array of possible slots */ 1116 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1117 if (!inuse) 1118 return -ENOMEM; 1119 1120 for_each_netdev(net, d) { 1121 struct netdev_name_node *name_node; 1122 1123 netdev_for_each_altname(d, name_node) { 1124 if (!sscanf(name_node->name, name, &i)) 1125 continue; 1126 if (i < 0 || i >= max_netdevices) 1127 continue; 1128 1129 /* avoid cases where sscanf is not exact inverse of printf */ 1130 snprintf(buf, IFNAMSIZ, name, i); 1131 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1132 __set_bit(i, inuse); 1133 } 1134 if (!sscanf(d->name, name, &i)) 1135 continue; 1136 if (i < 0 || i >= max_netdevices) 1137 continue; 1138 1139 /* avoid cases where sscanf is not exact inverse of printf */ 1140 snprintf(buf, IFNAMSIZ, name, i); 1141 if (!strncmp(buf, d->name, IFNAMSIZ)) 1142 __set_bit(i, inuse); 1143 } 1144 1145 i = find_first_zero_bit(inuse, max_netdevices); 1146 bitmap_free(inuse); 1147 if (i == max_netdevices) 1148 return -ENFILE; 1149 1150 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1151 strscpy(buf, name, IFNAMSIZ); 1152 snprintf(res, IFNAMSIZ, buf, i); 1153 return i; 1154 } 1155 1156 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1157 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1158 const char *want_name, char *out_name, 1159 int dup_errno) 1160 { 1161 if (!dev_valid_name(want_name)) 1162 return -EINVAL; 1163 1164 if (strchr(want_name, '%')) 1165 return __dev_alloc_name(net, want_name, out_name); 1166 1167 if (netdev_name_in_use(net, want_name)) 1168 return -dup_errno; 1169 if (out_name != want_name) 1170 strscpy(out_name, want_name, IFNAMSIZ); 1171 return 0; 1172 } 1173 1174 /** 1175 * dev_alloc_name - allocate a name for a device 1176 * @dev: device 1177 * @name: name format string 1178 * 1179 * Passed a format string - eg "lt%d" it will try and find a suitable 1180 * id. It scans list of devices to build up a free map, then chooses 1181 * the first empty slot. The caller must hold the dev_base or rtnl lock 1182 * while allocating the name and adding the device in order to avoid 1183 * duplicates. 1184 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1185 * Returns the number of the unit assigned or a negative errno code. 1186 */ 1187 1188 int dev_alloc_name(struct net_device *dev, const char *name) 1189 { 1190 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1191 } 1192 EXPORT_SYMBOL(dev_alloc_name); 1193 1194 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1195 const char *name) 1196 { 1197 int ret; 1198 1199 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1200 return ret < 0 ? ret : 0; 1201 } 1202 1203 /** 1204 * dev_change_name - change name of a device 1205 * @dev: device 1206 * @newname: name (or format string) must be at least IFNAMSIZ 1207 * 1208 * Change name of a device, can pass format strings "eth%d". 1209 * for wildcarding. 1210 */ 1211 int dev_change_name(struct net_device *dev, const char *newname) 1212 { 1213 unsigned char old_assign_type; 1214 char oldname[IFNAMSIZ]; 1215 int err = 0; 1216 int ret; 1217 struct net *net; 1218 1219 ASSERT_RTNL(); 1220 BUG_ON(!dev_net(dev)); 1221 1222 net = dev_net(dev); 1223 1224 down_write(&devnet_rename_sem); 1225 1226 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) { 1227 up_write(&devnet_rename_sem); 1228 return 0; 1229 } 1230 1231 memcpy(oldname, dev->name, IFNAMSIZ); 1232 1233 write_seqlock_bh(&netdev_rename_lock); 1234 err = dev_get_valid_name(net, dev, newname); 1235 write_sequnlock_bh(&netdev_rename_lock); 1236 1237 if (err < 0) { 1238 up_write(&devnet_rename_sem); 1239 return err; 1240 } 1241 1242 if (oldname[0] && !strchr(oldname, '%')) 1243 netdev_info(dev, "renamed from %s%s\n", oldname, 1244 dev->flags & IFF_UP ? " (while UP)" : ""); 1245 1246 old_assign_type = dev->name_assign_type; 1247 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1248 1249 rollback: 1250 ret = device_rename(&dev->dev, dev->name); 1251 if (ret) { 1252 memcpy(dev->name, oldname, IFNAMSIZ); 1253 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1254 up_write(&devnet_rename_sem); 1255 return ret; 1256 } 1257 1258 up_write(&devnet_rename_sem); 1259 1260 netdev_adjacent_rename_links(dev, oldname); 1261 1262 netdev_name_node_del(dev->name_node); 1263 1264 synchronize_net(); 1265 1266 netdev_name_node_add(net, dev->name_node); 1267 1268 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1269 ret = notifier_to_errno(ret); 1270 1271 if (ret) { 1272 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1273 if (err >= 0) { 1274 err = ret; 1275 down_write(&devnet_rename_sem); 1276 write_seqlock_bh(&netdev_rename_lock); 1277 memcpy(dev->name, oldname, IFNAMSIZ); 1278 write_sequnlock_bh(&netdev_rename_lock); 1279 memcpy(oldname, newname, IFNAMSIZ); 1280 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1281 old_assign_type = NET_NAME_RENAMED; 1282 goto rollback; 1283 } else { 1284 netdev_err(dev, "name change rollback failed: %d\n", 1285 ret); 1286 } 1287 } 1288 1289 return err; 1290 } 1291 1292 /** 1293 * dev_set_alias - change ifalias of a device 1294 * @dev: device 1295 * @alias: name up to IFALIASZ 1296 * @len: limit of bytes to copy from info 1297 * 1298 * Set ifalias for a device, 1299 */ 1300 int dev_set_alias(struct net_device *dev, const char *alias, size_t len) 1301 { 1302 struct dev_ifalias *new_alias = NULL; 1303 1304 if (len >= IFALIASZ) 1305 return -EINVAL; 1306 1307 if (len) { 1308 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1309 if (!new_alias) 1310 return -ENOMEM; 1311 1312 memcpy(new_alias->ifalias, alias, len); 1313 new_alias->ifalias[len] = 0; 1314 } 1315 1316 mutex_lock(&ifalias_mutex); 1317 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1318 mutex_is_locked(&ifalias_mutex)); 1319 mutex_unlock(&ifalias_mutex); 1320 1321 if (new_alias) 1322 kfree_rcu(new_alias, rcuhead); 1323 1324 return len; 1325 } 1326 EXPORT_SYMBOL(dev_set_alias); 1327 1328 /** 1329 * dev_get_alias - get ifalias of a device 1330 * @dev: device 1331 * @name: buffer to store name of ifalias 1332 * @len: size of buffer 1333 * 1334 * get ifalias for a device. Caller must make sure dev cannot go 1335 * away, e.g. rcu read lock or own a reference count to device. 1336 */ 1337 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1338 { 1339 const struct dev_ifalias *alias; 1340 int ret = 0; 1341 1342 rcu_read_lock(); 1343 alias = rcu_dereference(dev->ifalias); 1344 if (alias) 1345 ret = snprintf(name, len, "%s", alias->ifalias); 1346 rcu_read_unlock(); 1347 1348 return ret; 1349 } 1350 1351 /** 1352 * netdev_features_change - device changes features 1353 * @dev: device to cause notification 1354 * 1355 * Called to indicate a device has changed features. 1356 */ 1357 void netdev_features_change(struct net_device *dev) 1358 { 1359 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1360 } 1361 EXPORT_SYMBOL(netdev_features_change); 1362 1363 /** 1364 * netdev_state_change - device changes state 1365 * @dev: device to cause notification 1366 * 1367 * Called to indicate a device has changed state. This function calls 1368 * the notifier chains for netdev_chain and sends a NEWLINK message 1369 * to the routing socket. 1370 */ 1371 void netdev_state_change(struct net_device *dev) 1372 { 1373 if (dev->flags & IFF_UP) { 1374 struct netdev_notifier_change_info change_info = { 1375 .info.dev = dev, 1376 }; 1377 1378 call_netdevice_notifiers_info(NETDEV_CHANGE, 1379 &change_info.info); 1380 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1381 } 1382 } 1383 EXPORT_SYMBOL(netdev_state_change); 1384 1385 /** 1386 * __netdev_notify_peers - notify network peers about existence of @dev, 1387 * to be called when rtnl lock is already held. 1388 * @dev: network device 1389 * 1390 * Generate traffic such that interested network peers are aware of 1391 * @dev, such as by generating a gratuitous ARP. This may be used when 1392 * a device wants to inform the rest of the network about some sort of 1393 * reconfiguration such as a failover event or virtual machine 1394 * migration. 1395 */ 1396 void __netdev_notify_peers(struct net_device *dev) 1397 { 1398 ASSERT_RTNL(); 1399 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1400 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1401 } 1402 EXPORT_SYMBOL(__netdev_notify_peers); 1403 1404 /** 1405 * netdev_notify_peers - notify network peers about existence of @dev 1406 * @dev: network device 1407 * 1408 * Generate traffic such that interested network peers are aware of 1409 * @dev, such as by generating a gratuitous ARP. This may be used when 1410 * a device wants to inform the rest of the network about some sort of 1411 * reconfiguration such as a failover event or virtual machine 1412 * migration. 1413 */ 1414 void netdev_notify_peers(struct net_device *dev) 1415 { 1416 rtnl_lock(); 1417 __netdev_notify_peers(dev); 1418 rtnl_unlock(); 1419 } 1420 EXPORT_SYMBOL(netdev_notify_peers); 1421 1422 static int napi_threaded_poll(void *data); 1423 1424 static int napi_kthread_create(struct napi_struct *n) 1425 { 1426 int err = 0; 1427 1428 /* Create and wake up the kthread once to put it in 1429 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1430 * warning and work with loadavg. 1431 */ 1432 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1433 n->dev->name, n->napi_id); 1434 if (IS_ERR(n->thread)) { 1435 err = PTR_ERR(n->thread); 1436 pr_err("kthread_run failed with err %d\n", err); 1437 n->thread = NULL; 1438 } 1439 1440 return err; 1441 } 1442 1443 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1444 { 1445 const struct net_device_ops *ops = dev->netdev_ops; 1446 int ret; 1447 1448 ASSERT_RTNL(); 1449 dev_addr_check(dev); 1450 1451 if (!netif_device_present(dev)) { 1452 /* may be detached because parent is runtime-suspended */ 1453 if (dev->dev.parent) 1454 pm_runtime_resume(dev->dev.parent); 1455 if (!netif_device_present(dev)) 1456 return -ENODEV; 1457 } 1458 1459 /* Block netpoll from trying to do any rx path servicing. 1460 * If we don't do this there is a chance ndo_poll_controller 1461 * or ndo_poll may be running while we open the device 1462 */ 1463 netpoll_poll_disable(dev); 1464 1465 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1466 ret = notifier_to_errno(ret); 1467 if (ret) 1468 return ret; 1469 1470 set_bit(__LINK_STATE_START, &dev->state); 1471 1472 if (ops->ndo_validate_addr) 1473 ret = ops->ndo_validate_addr(dev); 1474 1475 if (!ret && ops->ndo_open) 1476 ret = ops->ndo_open(dev); 1477 1478 netpoll_poll_enable(dev); 1479 1480 if (ret) 1481 clear_bit(__LINK_STATE_START, &dev->state); 1482 else { 1483 dev->flags |= IFF_UP; 1484 dev_set_rx_mode(dev); 1485 dev_activate(dev); 1486 add_device_randomness(dev->dev_addr, dev->addr_len); 1487 } 1488 1489 return ret; 1490 } 1491 1492 /** 1493 * dev_open - prepare an interface for use. 1494 * @dev: device to open 1495 * @extack: netlink extended ack 1496 * 1497 * Takes a device from down to up state. The device's private open 1498 * function is invoked and then the multicast lists are loaded. Finally 1499 * the device is moved into the up state and a %NETDEV_UP message is 1500 * sent to the netdev notifier chain. 1501 * 1502 * Calling this function on an active interface is a nop. On a failure 1503 * a negative errno code is returned. 1504 */ 1505 int dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1506 { 1507 int ret; 1508 1509 if (dev->flags & IFF_UP) 1510 return 0; 1511 1512 ret = __dev_open(dev, extack); 1513 if (ret < 0) 1514 return ret; 1515 1516 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1517 call_netdevice_notifiers(NETDEV_UP, dev); 1518 1519 return ret; 1520 } 1521 EXPORT_SYMBOL(dev_open); 1522 1523 static void __dev_close_many(struct list_head *head) 1524 { 1525 struct net_device *dev; 1526 1527 ASSERT_RTNL(); 1528 might_sleep(); 1529 1530 list_for_each_entry(dev, head, close_list) { 1531 /* Temporarily disable netpoll until the interface is down */ 1532 netpoll_poll_disable(dev); 1533 1534 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1535 1536 clear_bit(__LINK_STATE_START, &dev->state); 1537 1538 /* Synchronize to scheduled poll. We cannot touch poll list, it 1539 * can be even on different cpu. So just clear netif_running(). 1540 * 1541 * dev->stop() will invoke napi_disable() on all of it's 1542 * napi_struct instances on this device. 1543 */ 1544 smp_mb__after_atomic(); /* Commit netif_running(). */ 1545 } 1546 1547 dev_deactivate_many(head); 1548 1549 list_for_each_entry(dev, head, close_list) { 1550 const struct net_device_ops *ops = dev->netdev_ops; 1551 1552 /* 1553 * Call the device specific close. This cannot fail. 1554 * Only if device is UP 1555 * 1556 * We allow it to be called even after a DETACH hot-plug 1557 * event. 1558 */ 1559 if (ops->ndo_stop) 1560 ops->ndo_stop(dev); 1561 1562 dev->flags &= ~IFF_UP; 1563 netpoll_poll_enable(dev); 1564 } 1565 } 1566 1567 static void __dev_close(struct net_device *dev) 1568 { 1569 LIST_HEAD(single); 1570 1571 list_add(&dev->close_list, &single); 1572 __dev_close_many(&single); 1573 list_del(&single); 1574 } 1575 1576 void dev_close_many(struct list_head *head, bool unlink) 1577 { 1578 struct net_device *dev, *tmp; 1579 1580 /* Remove the devices that don't need to be closed */ 1581 list_for_each_entry_safe(dev, tmp, head, close_list) 1582 if (!(dev->flags & IFF_UP)) 1583 list_del_init(&dev->close_list); 1584 1585 __dev_close_many(head); 1586 1587 list_for_each_entry_safe(dev, tmp, head, close_list) { 1588 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1589 call_netdevice_notifiers(NETDEV_DOWN, dev); 1590 if (unlink) 1591 list_del_init(&dev->close_list); 1592 } 1593 } 1594 EXPORT_SYMBOL(dev_close_many); 1595 1596 /** 1597 * dev_close - shutdown an interface. 1598 * @dev: device to shutdown 1599 * 1600 * This function moves an active device into down state. A 1601 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device 1602 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier 1603 * chain. 1604 */ 1605 void dev_close(struct net_device *dev) 1606 { 1607 if (dev->flags & IFF_UP) { 1608 LIST_HEAD(single); 1609 1610 list_add(&dev->close_list, &single); 1611 dev_close_many(&single, true); 1612 list_del(&single); 1613 } 1614 } 1615 EXPORT_SYMBOL(dev_close); 1616 1617 1618 /** 1619 * dev_disable_lro - disable Large Receive Offload on a device 1620 * @dev: device 1621 * 1622 * Disable Large Receive Offload (LRO) on a net device. Must be 1623 * called under RTNL. This is needed if received packets may be 1624 * forwarded to another interface. 1625 */ 1626 void dev_disable_lro(struct net_device *dev) 1627 { 1628 struct net_device *lower_dev; 1629 struct list_head *iter; 1630 1631 dev->wanted_features &= ~NETIF_F_LRO; 1632 netdev_update_features(dev); 1633 1634 if (unlikely(dev->features & NETIF_F_LRO)) 1635 netdev_WARN(dev, "failed to disable LRO!\n"); 1636 1637 netdev_for_each_lower_dev(dev, lower_dev, iter) 1638 dev_disable_lro(lower_dev); 1639 } 1640 EXPORT_SYMBOL(dev_disable_lro); 1641 1642 /** 1643 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1644 * @dev: device 1645 * 1646 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1647 * called under RTNL. This is needed if Generic XDP is installed on 1648 * the device. 1649 */ 1650 static void dev_disable_gro_hw(struct net_device *dev) 1651 { 1652 dev->wanted_features &= ~NETIF_F_GRO_HW; 1653 netdev_update_features(dev); 1654 1655 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1656 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1657 } 1658 1659 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1660 { 1661 #define N(val) \ 1662 case NETDEV_##val: \ 1663 return "NETDEV_" __stringify(val); 1664 switch (cmd) { 1665 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1666 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1667 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1668 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1669 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1670 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1671 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1672 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1673 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1674 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1675 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1676 N(XDP_FEAT_CHANGE) 1677 } 1678 #undef N 1679 return "UNKNOWN_NETDEV_EVENT"; 1680 } 1681 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1682 1683 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1684 struct net_device *dev) 1685 { 1686 struct netdev_notifier_info info = { 1687 .dev = dev, 1688 }; 1689 1690 return nb->notifier_call(nb, val, &info); 1691 } 1692 1693 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1694 struct net_device *dev) 1695 { 1696 int err; 1697 1698 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1699 err = notifier_to_errno(err); 1700 if (err) 1701 return err; 1702 1703 if (!(dev->flags & IFF_UP)) 1704 return 0; 1705 1706 call_netdevice_notifier(nb, NETDEV_UP, dev); 1707 return 0; 1708 } 1709 1710 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1711 struct net_device *dev) 1712 { 1713 if (dev->flags & IFF_UP) { 1714 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1715 dev); 1716 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1717 } 1718 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1719 } 1720 1721 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1722 struct net *net) 1723 { 1724 struct net_device *dev; 1725 int err; 1726 1727 for_each_netdev(net, dev) { 1728 err = call_netdevice_register_notifiers(nb, dev); 1729 if (err) 1730 goto rollback; 1731 } 1732 return 0; 1733 1734 rollback: 1735 for_each_netdev_continue_reverse(net, dev) 1736 call_netdevice_unregister_notifiers(nb, dev); 1737 return err; 1738 } 1739 1740 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1741 struct net *net) 1742 { 1743 struct net_device *dev; 1744 1745 for_each_netdev(net, dev) 1746 call_netdevice_unregister_notifiers(nb, dev); 1747 } 1748 1749 static int dev_boot_phase = 1; 1750 1751 /** 1752 * register_netdevice_notifier - register a network notifier block 1753 * @nb: notifier 1754 * 1755 * Register a notifier to be called when network device events occur. 1756 * The notifier passed is linked into the kernel structures and must 1757 * not be reused until it has been unregistered. A negative errno code 1758 * is returned on a failure. 1759 * 1760 * When registered all registration and up events are replayed 1761 * to the new notifier to allow device to have a race free 1762 * view of the network device list. 1763 */ 1764 1765 int register_netdevice_notifier(struct notifier_block *nb) 1766 { 1767 struct net *net; 1768 int err; 1769 1770 /* Close race with setup_net() and cleanup_net() */ 1771 down_write(&pernet_ops_rwsem); 1772 rtnl_lock(); 1773 err = raw_notifier_chain_register(&netdev_chain, nb); 1774 if (err) 1775 goto unlock; 1776 if (dev_boot_phase) 1777 goto unlock; 1778 for_each_net(net) { 1779 err = call_netdevice_register_net_notifiers(nb, net); 1780 if (err) 1781 goto rollback; 1782 } 1783 1784 unlock: 1785 rtnl_unlock(); 1786 up_write(&pernet_ops_rwsem); 1787 return err; 1788 1789 rollback: 1790 for_each_net_continue_reverse(net) 1791 call_netdevice_unregister_net_notifiers(nb, net); 1792 1793 raw_notifier_chain_unregister(&netdev_chain, nb); 1794 goto unlock; 1795 } 1796 EXPORT_SYMBOL(register_netdevice_notifier); 1797 1798 /** 1799 * unregister_netdevice_notifier - unregister a network notifier block 1800 * @nb: notifier 1801 * 1802 * Unregister a notifier previously registered by 1803 * register_netdevice_notifier(). The notifier is unlinked into the 1804 * kernel structures and may then be reused. A negative errno code 1805 * is returned on a failure. 1806 * 1807 * After unregistering unregister and down device events are synthesized 1808 * for all devices on the device list to the removed notifier to remove 1809 * the need for special case cleanup code. 1810 */ 1811 1812 int unregister_netdevice_notifier(struct notifier_block *nb) 1813 { 1814 struct net *net; 1815 int err; 1816 1817 /* Close race with setup_net() and cleanup_net() */ 1818 down_write(&pernet_ops_rwsem); 1819 rtnl_lock(); 1820 err = raw_notifier_chain_unregister(&netdev_chain, nb); 1821 if (err) 1822 goto unlock; 1823 1824 for_each_net(net) 1825 call_netdevice_unregister_net_notifiers(nb, net); 1826 1827 unlock: 1828 rtnl_unlock(); 1829 up_write(&pernet_ops_rwsem); 1830 return err; 1831 } 1832 EXPORT_SYMBOL(unregister_netdevice_notifier); 1833 1834 static int __register_netdevice_notifier_net(struct net *net, 1835 struct notifier_block *nb, 1836 bool ignore_call_fail) 1837 { 1838 int err; 1839 1840 err = raw_notifier_chain_register(&net->netdev_chain, nb); 1841 if (err) 1842 return err; 1843 if (dev_boot_phase) 1844 return 0; 1845 1846 err = call_netdevice_register_net_notifiers(nb, net); 1847 if (err && !ignore_call_fail) 1848 goto chain_unregister; 1849 1850 return 0; 1851 1852 chain_unregister: 1853 raw_notifier_chain_unregister(&net->netdev_chain, nb); 1854 return err; 1855 } 1856 1857 static int __unregister_netdevice_notifier_net(struct net *net, 1858 struct notifier_block *nb) 1859 { 1860 int err; 1861 1862 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 1863 if (err) 1864 return err; 1865 1866 call_netdevice_unregister_net_notifiers(nb, net); 1867 return 0; 1868 } 1869 1870 /** 1871 * register_netdevice_notifier_net - register a per-netns network notifier block 1872 * @net: network namespace 1873 * @nb: notifier 1874 * 1875 * Register a notifier to be called when network device events occur. 1876 * The notifier passed is linked into the kernel structures and must 1877 * not be reused until it has been unregistered. A negative errno code 1878 * is returned on a failure. 1879 * 1880 * When registered all registration and up events are replayed 1881 * to the new notifier to allow device to have a race free 1882 * view of the network device list. 1883 */ 1884 1885 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 1886 { 1887 int err; 1888 1889 rtnl_lock(); 1890 err = __register_netdevice_notifier_net(net, nb, false); 1891 rtnl_unlock(); 1892 return err; 1893 } 1894 EXPORT_SYMBOL(register_netdevice_notifier_net); 1895 1896 /** 1897 * unregister_netdevice_notifier_net - unregister a per-netns 1898 * network notifier block 1899 * @net: network namespace 1900 * @nb: notifier 1901 * 1902 * Unregister a notifier previously registered by 1903 * register_netdevice_notifier_net(). The notifier is unlinked from the 1904 * kernel structures and may then be reused. A negative errno code 1905 * is returned on a failure. 1906 * 1907 * After unregistering unregister and down device events are synthesized 1908 * for all devices on the device list to the removed notifier to remove 1909 * the need for special case cleanup code. 1910 */ 1911 1912 int unregister_netdevice_notifier_net(struct net *net, 1913 struct notifier_block *nb) 1914 { 1915 int err; 1916 1917 rtnl_lock(); 1918 err = __unregister_netdevice_notifier_net(net, nb); 1919 rtnl_unlock(); 1920 return err; 1921 } 1922 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 1923 1924 static void __move_netdevice_notifier_net(struct net *src_net, 1925 struct net *dst_net, 1926 struct notifier_block *nb) 1927 { 1928 __unregister_netdevice_notifier_net(src_net, nb); 1929 __register_netdevice_notifier_net(dst_net, nb, true); 1930 } 1931 1932 int register_netdevice_notifier_dev_net(struct net_device *dev, 1933 struct notifier_block *nb, 1934 struct netdev_net_notifier *nn) 1935 { 1936 int err; 1937 1938 rtnl_lock(); 1939 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 1940 if (!err) { 1941 nn->nb = nb; 1942 list_add(&nn->list, &dev->net_notifier_list); 1943 } 1944 rtnl_unlock(); 1945 return err; 1946 } 1947 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 1948 1949 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 1950 struct notifier_block *nb, 1951 struct netdev_net_notifier *nn) 1952 { 1953 int err; 1954 1955 rtnl_lock(); 1956 list_del(&nn->list); 1957 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 1958 rtnl_unlock(); 1959 return err; 1960 } 1961 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 1962 1963 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 1964 struct net *net) 1965 { 1966 struct netdev_net_notifier *nn; 1967 1968 list_for_each_entry(nn, &dev->net_notifier_list, list) 1969 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 1970 } 1971 1972 /** 1973 * call_netdevice_notifiers_info - call all network notifier blocks 1974 * @val: value passed unmodified to notifier function 1975 * @info: notifier information data 1976 * 1977 * Call all network notifier blocks. Parameters and return value 1978 * are as for raw_notifier_call_chain(). 1979 */ 1980 1981 int call_netdevice_notifiers_info(unsigned long val, 1982 struct netdev_notifier_info *info) 1983 { 1984 struct net *net = dev_net(info->dev); 1985 int ret; 1986 1987 ASSERT_RTNL(); 1988 1989 /* Run per-netns notifier block chain first, then run the global one. 1990 * Hopefully, one day, the global one is going to be removed after 1991 * all notifier block registrators get converted to be per-netns. 1992 */ 1993 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 1994 if (ret & NOTIFY_STOP_MASK) 1995 return ret; 1996 return raw_notifier_call_chain(&netdev_chain, val, info); 1997 } 1998 1999 /** 2000 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2001 * for and rollback on error 2002 * @val_up: value passed unmodified to notifier function 2003 * @val_down: value passed unmodified to the notifier function when 2004 * recovering from an error on @val_up 2005 * @info: notifier information data 2006 * 2007 * Call all per-netns network notifier blocks, but not notifier blocks on 2008 * the global notifier chain. Parameters and return value are as for 2009 * raw_notifier_call_chain_robust(). 2010 */ 2011 2012 static int 2013 call_netdevice_notifiers_info_robust(unsigned long val_up, 2014 unsigned long val_down, 2015 struct netdev_notifier_info *info) 2016 { 2017 struct net *net = dev_net(info->dev); 2018 2019 ASSERT_RTNL(); 2020 2021 return raw_notifier_call_chain_robust(&net->netdev_chain, 2022 val_up, val_down, info); 2023 } 2024 2025 static int call_netdevice_notifiers_extack(unsigned long val, 2026 struct net_device *dev, 2027 struct netlink_ext_ack *extack) 2028 { 2029 struct netdev_notifier_info info = { 2030 .dev = dev, 2031 .extack = extack, 2032 }; 2033 2034 return call_netdevice_notifiers_info(val, &info); 2035 } 2036 2037 /** 2038 * call_netdevice_notifiers - call all network notifier blocks 2039 * @val: value passed unmodified to notifier function 2040 * @dev: net_device pointer passed unmodified to notifier function 2041 * 2042 * Call all network notifier blocks. Parameters and return value 2043 * are as for raw_notifier_call_chain(). 2044 */ 2045 2046 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2047 { 2048 return call_netdevice_notifiers_extack(val, dev, NULL); 2049 } 2050 EXPORT_SYMBOL(call_netdevice_notifiers); 2051 2052 /** 2053 * call_netdevice_notifiers_mtu - call all network notifier blocks 2054 * @val: value passed unmodified to notifier function 2055 * @dev: net_device pointer passed unmodified to notifier function 2056 * @arg: additional u32 argument passed to the notifier function 2057 * 2058 * Call all network notifier blocks. Parameters and return value 2059 * are as for raw_notifier_call_chain(). 2060 */ 2061 static int call_netdevice_notifiers_mtu(unsigned long val, 2062 struct net_device *dev, u32 arg) 2063 { 2064 struct netdev_notifier_info_ext info = { 2065 .info.dev = dev, 2066 .ext.mtu = arg, 2067 }; 2068 2069 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2070 2071 return call_netdevice_notifiers_info(val, &info.info); 2072 } 2073 2074 #ifdef CONFIG_NET_INGRESS 2075 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2076 2077 void net_inc_ingress_queue(void) 2078 { 2079 static_branch_inc(&ingress_needed_key); 2080 } 2081 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2082 2083 void net_dec_ingress_queue(void) 2084 { 2085 static_branch_dec(&ingress_needed_key); 2086 } 2087 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2088 #endif 2089 2090 #ifdef CONFIG_NET_EGRESS 2091 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2092 2093 void net_inc_egress_queue(void) 2094 { 2095 static_branch_inc(&egress_needed_key); 2096 } 2097 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2098 2099 void net_dec_egress_queue(void) 2100 { 2101 static_branch_dec(&egress_needed_key); 2102 } 2103 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2104 #endif 2105 2106 #ifdef CONFIG_NET_CLS_ACT 2107 DEFINE_STATIC_KEY_FALSE(tcf_bypass_check_needed_key); 2108 EXPORT_SYMBOL(tcf_bypass_check_needed_key); 2109 #endif 2110 2111 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2112 EXPORT_SYMBOL(netstamp_needed_key); 2113 #ifdef CONFIG_JUMP_LABEL 2114 static atomic_t netstamp_needed_deferred; 2115 static atomic_t netstamp_wanted; 2116 static void netstamp_clear(struct work_struct *work) 2117 { 2118 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2119 int wanted; 2120 2121 wanted = atomic_add_return(deferred, &netstamp_wanted); 2122 if (wanted > 0) 2123 static_branch_enable(&netstamp_needed_key); 2124 else 2125 static_branch_disable(&netstamp_needed_key); 2126 } 2127 static DECLARE_WORK(netstamp_work, netstamp_clear); 2128 #endif 2129 2130 void net_enable_timestamp(void) 2131 { 2132 #ifdef CONFIG_JUMP_LABEL 2133 int wanted = atomic_read(&netstamp_wanted); 2134 2135 while (wanted > 0) { 2136 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2137 return; 2138 } 2139 atomic_inc(&netstamp_needed_deferred); 2140 schedule_work(&netstamp_work); 2141 #else 2142 static_branch_inc(&netstamp_needed_key); 2143 #endif 2144 } 2145 EXPORT_SYMBOL(net_enable_timestamp); 2146 2147 void net_disable_timestamp(void) 2148 { 2149 #ifdef CONFIG_JUMP_LABEL 2150 int wanted = atomic_read(&netstamp_wanted); 2151 2152 while (wanted > 1) { 2153 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2154 return; 2155 } 2156 atomic_dec(&netstamp_needed_deferred); 2157 schedule_work(&netstamp_work); 2158 #else 2159 static_branch_dec(&netstamp_needed_key); 2160 #endif 2161 } 2162 EXPORT_SYMBOL(net_disable_timestamp); 2163 2164 static inline void net_timestamp_set(struct sk_buff *skb) 2165 { 2166 skb->tstamp = 0; 2167 skb->tstamp_type = SKB_CLOCK_REALTIME; 2168 if (static_branch_unlikely(&netstamp_needed_key)) 2169 skb->tstamp = ktime_get_real(); 2170 } 2171 2172 #define net_timestamp_check(COND, SKB) \ 2173 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2174 if ((COND) && !(SKB)->tstamp) \ 2175 (SKB)->tstamp = ktime_get_real(); \ 2176 } \ 2177 2178 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2179 { 2180 return __is_skb_forwardable(dev, skb, true); 2181 } 2182 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2183 2184 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2185 bool check_mtu) 2186 { 2187 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2188 2189 if (likely(!ret)) { 2190 skb->protocol = eth_type_trans(skb, dev); 2191 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2192 } 2193 2194 return ret; 2195 } 2196 2197 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2198 { 2199 return __dev_forward_skb2(dev, skb, true); 2200 } 2201 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2202 2203 /** 2204 * dev_forward_skb - loopback an skb to another netif 2205 * 2206 * @dev: destination network device 2207 * @skb: buffer to forward 2208 * 2209 * return values: 2210 * NET_RX_SUCCESS (no congestion) 2211 * NET_RX_DROP (packet was dropped, but freed) 2212 * 2213 * dev_forward_skb can be used for injecting an skb from the 2214 * start_xmit function of one device into the receive queue 2215 * of another device. 2216 * 2217 * The receiving device may be in another namespace, so 2218 * we have to clear all information in the skb that could 2219 * impact namespace isolation. 2220 */ 2221 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2222 { 2223 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2224 } 2225 EXPORT_SYMBOL_GPL(dev_forward_skb); 2226 2227 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2228 { 2229 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2230 } 2231 2232 static inline int deliver_skb(struct sk_buff *skb, 2233 struct packet_type *pt_prev, 2234 struct net_device *orig_dev) 2235 { 2236 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2237 return -ENOMEM; 2238 refcount_inc(&skb->users); 2239 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2240 } 2241 2242 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2243 struct packet_type **pt, 2244 struct net_device *orig_dev, 2245 __be16 type, 2246 struct list_head *ptype_list) 2247 { 2248 struct packet_type *ptype, *pt_prev = *pt; 2249 2250 list_for_each_entry_rcu(ptype, ptype_list, list) { 2251 if (ptype->type != type) 2252 continue; 2253 if (pt_prev) 2254 deliver_skb(skb, pt_prev, orig_dev); 2255 pt_prev = ptype; 2256 } 2257 *pt = pt_prev; 2258 } 2259 2260 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2261 { 2262 if (!ptype->af_packet_priv || !skb->sk) 2263 return false; 2264 2265 if (ptype->id_match) 2266 return ptype->id_match(ptype, skb->sk); 2267 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2268 return true; 2269 2270 return false; 2271 } 2272 2273 /** 2274 * dev_nit_active - return true if any network interface taps are in use 2275 * 2276 * @dev: network device to check for the presence of taps 2277 */ 2278 bool dev_nit_active(struct net_device *dev) 2279 { 2280 return !list_empty(&net_hotdata.ptype_all) || 2281 !list_empty(&dev->ptype_all); 2282 } 2283 EXPORT_SYMBOL_GPL(dev_nit_active); 2284 2285 /* 2286 * Support routine. Sends outgoing frames to any network 2287 * taps currently in use. 2288 */ 2289 2290 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2291 { 2292 struct list_head *ptype_list = &net_hotdata.ptype_all; 2293 struct packet_type *ptype, *pt_prev = NULL; 2294 struct sk_buff *skb2 = NULL; 2295 2296 rcu_read_lock(); 2297 again: 2298 list_for_each_entry_rcu(ptype, ptype_list, list) { 2299 if (READ_ONCE(ptype->ignore_outgoing)) 2300 continue; 2301 2302 /* Never send packets back to the socket 2303 * they originated from - MvS (miquels@drinkel.ow.org) 2304 */ 2305 if (skb_loop_sk(ptype, skb)) 2306 continue; 2307 2308 if (pt_prev) { 2309 deliver_skb(skb2, pt_prev, skb->dev); 2310 pt_prev = ptype; 2311 continue; 2312 } 2313 2314 /* need to clone skb, done only once */ 2315 skb2 = skb_clone(skb, GFP_ATOMIC); 2316 if (!skb2) 2317 goto out_unlock; 2318 2319 net_timestamp_set(skb2); 2320 2321 /* skb->nh should be correctly 2322 * set by sender, so that the second statement is 2323 * just protection against buggy protocols. 2324 */ 2325 skb_reset_mac_header(skb2); 2326 2327 if (skb_network_header(skb2) < skb2->data || 2328 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2329 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2330 ntohs(skb2->protocol), 2331 dev->name); 2332 skb_reset_network_header(skb2); 2333 } 2334 2335 skb2->transport_header = skb2->network_header; 2336 skb2->pkt_type = PACKET_OUTGOING; 2337 pt_prev = ptype; 2338 } 2339 2340 if (ptype_list == &net_hotdata.ptype_all) { 2341 ptype_list = &dev->ptype_all; 2342 goto again; 2343 } 2344 out_unlock: 2345 if (pt_prev) { 2346 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2347 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2348 else 2349 kfree_skb(skb2); 2350 } 2351 rcu_read_unlock(); 2352 } 2353 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2354 2355 /** 2356 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2357 * @dev: Network device 2358 * @txq: number of queues available 2359 * 2360 * If real_num_tx_queues is changed the tc mappings may no longer be 2361 * valid. To resolve this verify the tc mapping remains valid and if 2362 * not NULL the mapping. With no priorities mapping to this 2363 * offset/count pair it will no longer be used. In the worst case TC0 2364 * is invalid nothing can be done so disable priority mappings. If is 2365 * expected that drivers will fix this mapping if they can before 2366 * calling netif_set_real_num_tx_queues. 2367 */ 2368 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2369 { 2370 int i; 2371 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2372 2373 /* If TC0 is invalidated disable TC mapping */ 2374 if (tc->offset + tc->count > txq) { 2375 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2376 dev->num_tc = 0; 2377 return; 2378 } 2379 2380 /* Invalidated prio to tc mappings set to TC0 */ 2381 for (i = 1; i < TC_BITMASK + 1; i++) { 2382 int q = netdev_get_prio_tc_map(dev, i); 2383 2384 tc = &dev->tc_to_txq[q]; 2385 if (tc->offset + tc->count > txq) { 2386 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", 2387 i, q); 2388 netdev_set_prio_tc_map(dev, i, 0); 2389 } 2390 } 2391 } 2392 2393 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2394 { 2395 if (dev->num_tc) { 2396 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2397 int i; 2398 2399 /* walk through the TCs and see if it falls into any of them */ 2400 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2401 if ((txq - tc->offset) < tc->count) 2402 return i; 2403 } 2404 2405 /* didn't find it, just return -1 to indicate no match */ 2406 return -1; 2407 } 2408 2409 return 0; 2410 } 2411 EXPORT_SYMBOL(netdev_txq_to_tc); 2412 2413 #ifdef CONFIG_XPS 2414 static struct static_key xps_needed __read_mostly; 2415 static struct static_key xps_rxqs_needed __read_mostly; 2416 static DEFINE_MUTEX(xps_map_mutex); 2417 #define xmap_dereference(P) \ 2418 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2419 2420 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2421 struct xps_dev_maps *old_maps, int tci, u16 index) 2422 { 2423 struct xps_map *map = NULL; 2424 int pos; 2425 2426 map = xmap_dereference(dev_maps->attr_map[tci]); 2427 if (!map) 2428 return false; 2429 2430 for (pos = map->len; pos--;) { 2431 if (map->queues[pos] != index) 2432 continue; 2433 2434 if (map->len > 1) { 2435 map->queues[pos] = map->queues[--map->len]; 2436 break; 2437 } 2438 2439 if (old_maps) 2440 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2441 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2442 kfree_rcu(map, rcu); 2443 return false; 2444 } 2445 2446 return true; 2447 } 2448 2449 static bool remove_xps_queue_cpu(struct net_device *dev, 2450 struct xps_dev_maps *dev_maps, 2451 int cpu, u16 offset, u16 count) 2452 { 2453 int num_tc = dev_maps->num_tc; 2454 bool active = false; 2455 int tci; 2456 2457 for (tci = cpu * num_tc; num_tc--; tci++) { 2458 int i, j; 2459 2460 for (i = count, j = offset; i--; j++) { 2461 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2462 break; 2463 } 2464 2465 active |= i < 0; 2466 } 2467 2468 return active; 2469 } 2470 2471 static void reset_xps_maps(struct net_device *dev, 2472 struct xps_dev_maps *dev_maps, 2473 enum xps_map_type type) 2474 { 2475 static_key_slow_dec_cpuslocked(&xps_needed); 2476 if (type == XPS_RXQS) 2477 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2478 2479 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2480 2481 kfree_rcu(dev_maps, rcu); 2482 } 2483 2484 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2485 u16 offset, u16 count) 2486 { 2487 struct xps_dev_maps *dev_maps; 2488 bool active = false; 2489 int i, j; 2490 2491 dev_maps = xmap_dereference(dev->xps_maps[type]); 2492 if (!dev_maps) 2493 return; 2494 2495 for (j = 0; j < dev_maps->nr_ids; j++) 2496 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2497 if (!active) 2498 reset_xps_maps(dev, dev_maps, type); 2499 2500 if (type == XPS_CPUS) { 2501 for (i = offset + (count - 1); count--; i--) 2502 netdev_queue_numa_node_write( 2503 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2504 } 2505 } 2506 2507 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2508 u16 count) 2509 { 2510 if (!static_key_false(&xps_needed)) 2511 return; 2512 2513 cpus_read_lock(); 2514 mutex_lock(&xps_map_mutex); 2515 2516 if (static_key_false(&xps_rxqs_needed)) 2517 clean_xps_maps(dev, XPS_RXQS, offset, count); 2518 2519 clean_xps_maps(dev, XPS_CPUS, offset, count); 2520 2521 mutex_unlock(&xps_map_mutex); 2522 cpus_read_unlock(); 2523 } 2524 2525 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2526 { 2527 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2528 } 2529 2530 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2531 u16 index, bool is_rxqs_map) 2532 { 2533 struct xps_map *new_map; 2534 int alloc_len = XPS_MIN_MAP_ALLOC; 2535 int i, pos; 2536 2537 for (pos = 0; map && pos < map->len; pos++) { 2538 if (map->queues[pos] != index) 2539 continue; 2540 return map; 2541 } 2542 2543 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2544 if (map) { 2545 if (pos < map->alloc_len) 2546 return map; 2547 2548 alloc_len = map->alloc_len * 2; 2549 } 2550 2551 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2552 * map 2553 */ 2554 if (is_rxqs_map) 2555 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2556 else 2557 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2558 cpu_to_node(attr_index)); 2559 if (!new_map) 2560 return NULL; 2561 2562 for (i = 0; i < pos; i++) 2563 new_map->queues[i] = map->queues[i]; 2564 new_map->alloc_len = alloc_len; 2565 new_map->len = pos; 2566 2567 return new_map; 2568 } 2569 2570 /* Copy xps maps at a given index */ 2571 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2572 struct xps_dev_maps *new_dev_maps, int index, 2573 int tc, bool skip_tc) 2574 { 2575 int i, tci = index * dev_maps->num_tc; 2576 struct xps_map *map; 2577 2578 /* copy maps belonging to foreign traffic classes */ 2579 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2580 if (i == tc && skip_tc) 2581 continue; 2582 2583 /* fill in the new device map from the old device map */ 2584 map = xmap_dereference(dev_maps->attr_map[tci]); 2585 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2586 } 2587 } 2588 2589 /* Must be called under cpus_read_lock */ 2590 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2591 u16 index, enum xps_map_type type) 2592 { 2593 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2594 const unsigned long *online_mask = NULL; 2595 bool active = false, copy = false; 2596 int i, j, tci, numa_node_id = -2; 2597 int maps_sz, num_tc = 1, tc = 0; 2598 struct xps_map *map, *new_map; 2599 unsigned int nr_ids; 2600 2601 WARN_ON_ONCE(index >= dev->num_tx_queues); 2602 2603 if (dev->num_tc) { 2604 /* Do not allow XPS on subordinate device directly */ 2605 num_tc = dev->num_tc; 2606 if (num_tc < 0) 2607 return -EINVAL; 2608 2609 /* If queue belongs to subordinate dev use its map */ 2610 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2611 2612 tc = netdev_txq_to_tc(dev, index); 2613 if (tc < 0) 2614 return -EINVAL; 2615 } 2616 2617 mutex_lock(&xps_map_mutex); 2618 2619 dev_maps = xmap_dereference(dev->xps_maps[type]); 2620 if (type == XPS_RXQS) { 2621 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2622 nr_ids = dev->num_rx_queues; 2623 } else { 2624 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2625 if (num_possible_cpus() > 1) 2626 online_mask = cpumask_bits(cpu_online_mask); 2627 nr_ids = nr_cpu_ids; 2628 } 2629 2630 if (maps_sz < L1_CACHE_BYTES) 2631 maps_sz = L1_CACHE_BYTES; 2632 2633 /* The old dev_maps could be larger or smaller than the one we're 2634 * setting up now, as dev->num_tc or nr_ids could have been updated in 2635 * between. We could try to be smart, but let's be safe instead and only 2636 * copy foreign traffic classes if the two map sizes match. 2637 */ 2638 if (dev_maps && 2639 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2640 copy = true; 2641 2642 /* allocate memory for queue storage */ 2643 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2644 j < nr_ids;) { 2645 if (!new_dev_maps) { 2646 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2647 if (!new_dev_maps) { 2648 mutex_unlock(&xps_map_mutex); 2649 return -ENOMEM; 2650 } 2651 2652 new_dev_maps->nr_ids = nr_ids; 2653 new_dev_maps->num_tc = num_tc; 2654 } 2655 2656 tci = j * num_tc + tc; 2657 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2658 2659 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2660 if (!map) 2661 goto error; 2662 2663 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2664 } 2665 2666 if (!new_dev_maps) 2667 goto out_no_new_maps; 2668 2669 if (!dev_maps) { 2670 /* Increment static keys at most once per type */ 2671 static_key_slow_inc_cpuslocked(&xps_needed); 2672 if (type == XPS_RXQS) 2673 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2674 } 2675 2676 for (j = 0; j < nr_ids; j++) { 2677 bool skip_tc = false; 2678 2679 tci = j * num_tc + tc; 2680 if (netif_attr_test_mask(j, mask, nr_ids) && 2681 netif_attr_test_online(j, online_mask, nr_ids)) { 2682 /* add tx-queue to CPU/rx-queue maps */ 2683 int pos = 0; 2684 2685 skip_tc = true; 2686 2687 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2688 while ((pos < map->len) && (map->queues[pos] != index)) 2689 pos++; 2690 2691 if (pos == map->len) 2692 map->queues[map->len++] = index; 2693 #ifdef CONFIG_NUMA 2694 if (type == XPS_CPUS) { 2695 if (numa_node_id == -2) 2696 numa_node_id = cpu_to_node(j); 2697 else if (numa_node_id != cpu_to_node(j)) 2698 numa_node_id = -1; 2699 } 2700 #endif 2701 } 2702 2703 if (copy) 2704 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2705 skip_tc); 2706 } 2707 2708 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2709 2710 /* Cleanup old maps */ 2711 if (!dev_maps) 2712 goto out_no_old_maps; 2713 2714 for (j = 0; j < dev_maps->nr_ids; j++) { 2715 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2716 map = xmap_dereference(dev_maps->attr_map[tci]); 2717 if (!map) 2718 continue; 2719 2720 if (copy) { 2721 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2722 if (map == new_map) 2723 continue; 2724 } 2725 2726 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2727 kfree_rcu(map, rcu); 2728 } 2729 } 2730 2731 old_dev_maps = dev_maps; 2732 2733 out_no_old_maps: 2734 dev_maps = new_dev_maps; 2735 active = true; 2736 2737 out_no_new_maps: 2738 if (type == XPS_CPUS) 2739 /* update Tx queue numa node */ 2740 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2741 (numa_node_id >= 0) ? 2742 numa_node_id : NUMA_NO_NODE); 2743 2744 if (!dev_maps) 2745 goto out_no_maps; 2746 2747 /* removes tx-queue from unused CPUs/rx-queues */ 2748 for (j = 0; j < dev_maps->nr_ids; j++) { 2749 tci = j * dev_maps->num_tc; 2750 2751 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2752 if (i == tc && 2753 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 2754 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 2755 continue; 2756 2757 active |= remove_xps_queue(dev_maps, 2758 copy ? old_dev_maps : NULL, 2759 tci, index); 2760 } 2761 } 2762 2763 if (old_dev_maps) 2764 kfree_rcu(old_dev_maps, rcu); 2765 2766 /* free map if not active */ 2767 if (!active) 2768 reset_xps_maps(dev, dev_maps, type); 2769 2770 out_no_maps: 2771 mutex_unlock(&xps_map_mutex); 2772 2773 return 0; 2774 error: 2775 /* remove any maps that we added */ 2776 for (j = 0; j < nr_ids; j++) { 2777 for (i = num_tc, tci = j * num_tc; i--; tci++) { 2778 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2779 map = copy ? 2780 xmap_dereference(dev_maps->attr_map[tci]) : 2781 NULL; 2782 if (new_map && new_map != map) 2783 kfree(new_map); 2784 } 2785 } 2786 2787 mutex_unlock(&xps_map_mutex); 2788 2789 kfree(new_dev_maps); 2790 return -ENOMEM; 2791 } 2792 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 2793 2794 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 2795 u16 index) 2796 { 2797 int ret; 2798 2799 cpus_read_lock(); 2800 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 2801 cpus_read_unlock(); 2802 2803 return ret; 2804 } 2805 EXPORT_SYMBOL(netif_set_xps_queue); 2806 2807 #endif 2808 static void netdev_unbind_all_sb_channels(struct net_device *dev) 2809 { 2810 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 2811 2812 /* Unbind any subordinate channels */ 2813 while (txq-- != &dev->_tx[0]) { 2814 if (txq->sb_dev) 2815 netdev_unbind_sb_channel(dev, txq->sb_dev); 2816 } 2817 } 2818 2819 void netdev_reset_tc(struct net_device *dev) 2820 { 2821 #ifdef CONFIG_XPS 2822 netif_reset_xps_queues_gt(dev, 0); 2823 #endif 2824 netdev_unbind_all_sb_channels(dev); 2825 2826 /* Reset TC configuration of device */ 2827 dev->num_tc = 0; 2828 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 2829 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 2830 } 2831 EXPORT_SYMBOL(netdev_reset_tc); 2832 2833 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 2834 { 2835 if (tc >= dev->num_tc) 2836 return -EINVAL; 2837 2838 #ifdef CONFIG_XPS 2839 netif_reset_xps_queues(dev, offset, count); 2840 #endif 2841 dev->tc_to_txq[tc].count = count; 2842 dev->tc_to_txq[tc].offset = offset; 2843 return 0; 2844 } 2845 EXPORT_SYMBOL(netdev_set_tc_queue); 2846 2847 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 2848 { 2849 if (num_tc > TC_MAX_QUEUE) 2850 return -EINVAL; 2851 2852 #ifdef CONFIG_XPS 2853 netif_reset_xps_queues_gt(dev, 0); 2854 #endif 2855 netdev_unbind_all_sb_channels(dev); 2856 2857 dev->num_tc = num_tc; 2858 return 0; 2859 } 2860 EXPORT_SYMBOL(netdev_set_num_tc); 2861 2862 void netdev_unbind_sb_channel(struct net_device *dev, 2863 struct net_device *sb_dev) 2864 { 2865 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 2866 2867 #ifdef CONFIG_XPS 2868 netif_reset_xps_queues_gt(sb_dev, 0); 2869 #endif 2870 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 2871 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 2872 2873 while (txq-- != &dev->_tx[0]) { 2874 if (txq->sb_dev == sb_dev) 2875 txq->sb_dev = NULL; 2876 } 2877 } 2878 EXPORT_SYMBOL(netdev_unbind_sb_channel); 2879 2880 int netdev_bind_sb_channel_queue(struct net_device *dev, 2881 struct net_device *sb_dev, 2882 u8 tc, u16 count, u16 offset) 2883 { 2884 /* Make certain the sb_dev and dev are already configured */ 2885 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 2886 return -EINVAL; 2887 2888 /* We cannot hand out queues we don't have */ 2889 if ((offset + count) > dev->real_num_tx_queues) 2890 return -EINVAL; 2891 2892 /* Record the mapping */ 2893 sb_dev->tc_to_txq[tc].count = count; 2894 sb_dev->tc_to_txq[tc].offset = offset; 2895 2896 /* Provide a way for Tx queue to find the tc_to_txq map or 2897 * XPS map for itself. 2898 */ 2899 while (count--) 2900 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 2901 2902 return 0; 2903 } 2904 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 2905 2906 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 2907 { 2908 /* Do not use a multiqueue device to represent a subordinate channel */ 2909 if (netif_is_multiqueue(dev)) 2910 return -ENODEV; 2911 2912 /* We allow channels 1 - 32767 to be used for subordinate channels. 2913 * Channel 0 is meant to be "native" mode and used only to represent 2914 * the main root device. We allow writing 0 to reset the device back 2915 * to normal mode after being used as a subordinate channel. 2916 */ 2917 if (channel > S16_MAX) 2918 return -EINVAL; 2919 2920 dev->num_tc = -channel; 2921 2922 return 0; 2923 } 2924 EXPORT_SYMBOL(netdev_set_sb_channel); 2925 2926 /* 2927 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 2928 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 2929 */ 2930 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 2931 { 2932 bool disabling; 2933 int rc; 2934 2935 disabling = txq < dev->real_num_tx_queues; 2936 2937 if (txq < 1 || txq > dev->num_tx_queues) 2938 return -EINVAL; 2939 2940 if (dev->reg_state == NETREG_REGISTERED || 2941 dev->reg_state == NETREG_UNREGISTERING) { 2942 ASSERT_RTNL(); 2943 2944 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 2945 txq); 2946 if (rc) 2947 return rc; 2948 2949 if (dev->num_tc) 2950 netif_setup_tc(dev, txq); 2951 2952 net_shaper_set_real_num_tx_queues(dev, txq); 2953 2954 dev_qdisc_change_real_num_tx(dev, txq); 2955 2956 dev->real_num_tx_queues = txq; 2957 2958 if (disabling) { 2959 synchronize_net(); 2960 qdisc_reset_all_tx_gt(dev, txq); 2961 #ifdef CONFIG_XPS 2962 netif_reset_xps_queues_gt(dev, txq); 2963 #endif 2964 } 2965 } else { 2966 dev->real_num_tx_queues = txq; 2967 } 2968 2969 return 0; 2970 } 2971 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 2972 2973 #ifdef CONFIG_SYSFS 2974 /** 2975 * netif_set_real_num_rx_queues - set actual number of RX queues used 2976 * @dev: Network device 2977 * @rxq: Actual number of RX queues 2978 * 2979 * This must be called either with the rtnl_lock held or before 2980 * registration of the net device. Returns 0 on success, or a 2981 * negative error code. If called before registration, it always 2982 * succeeds. 2983 */ 2984 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 2985 { 2986 int rc; 2987 2988 if (rxq < 1 || rxq > dev->num_rx_queues) 2989 return -EINVAL; 2990 2991 if (dev->reg_state == NETREG_REGISTERED) { 2992 ASSERT_RTNL(); 2993 2994 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 2995 rxq); 2996 if (rc) 2997 return rc; 2998 } 2999 3000 dev->real_num_rx_queues = rxq; 3001 return 0; 3002 } 3003 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3004 #endif 3005 3006 /** 3007 * netif_set_real_num_queues - set actual number of RX and TX queues used 3008 * @dev: Network device 3009 * @txq: Actual number of TX queues 3010 * @rxq: Actual number of RX queues 3011 * 3012 * Set the real number of both TX and RX queues. 3013 * Does nothing if the number of queues is already correct. 3014 */ 3015 int netif_set_real_num_queues(struct net_device *dev, 3016 unsigned int txq, unsigned int rxq) 3017 { 3018 unsigned int old_rxq = dev->real_num_rx_queues; 3019 int err; 3020 3021 if (txq < 1 || txq > dev->num_tx_queues || 3022 rxq < 1 || rxq > dev->num_rx_queues) 3023 return -EINVAL; 3024 3025 /* Start from increases, so the error path only does decreases - 3026 * decreases can't fail. 3027 */ 3028 if (rxq > dev->real_num_rx_queues) { 3029 err = netif_set_real_num_rx_queues(dev, rxq); 3030 if (err) 3031 return err; 3032 } 3033 if (txq > dev->real_num_tx_queues) { 3034 err = netif_set_real_num_tx_queues(dev, txq); 3035 if (err) 3036 goto undo_rx; 3037 } 3038 if (rxq < dev->real_num_rx_queues) 3039 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3040 if (txq < dev->real_num_tx_queues) 3041 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3042 3043 return 0; 3044 undo_rx: 3045 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3046 return err; 3047 } 3048 EXPORT_SYMBOL(netif_set_real_num_queues); 3049 3050 /** 3051 * netif_set_tso_max_size() - set the max size of TSO frames supported 3052 * @dev: netdev to update 3053 * @size: max skb->len of a TSO frame 3054 * 3055 * Set the limit on the size of TSO super-frames the device can handle. 3056 * Unless explicitly set the stack will assume the value of 3057 * %GSO_LEGACY_MAX_SIZE. 3058 */ 3059 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3060 { 3061 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3062 if (size < READ_ONCE(dev->gso_max_size)) 3063 netif_set_gso_max_size(dev, size); 3064 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3065 netif_set_gso_ipv4_max_size(dev, size); 3066 } 3067 EXPORT_SYMBOL(netif_set_tso_max_size); 3068 3069 /** 3070 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3071 * @dev: netdev to update 3072 * @segs: max number of TCP segments 3073 * 3074 * Set the limit on the number of TCP segments the device can generate from 3075 * a single TSO super-frame. 3076 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3077 */ 3078 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3079 { 3080 dev->tso_max_segs = segs; 3081 if (segs < READ_ONCE(dev->gso_max_segs)) 3082 netif_set_gso_max_segs(dev, segs); 3083 } 3084 EXPORT_SYMBOL(netif_set_tso_max_segs); 3085 3086 /** 3087 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3088 * @to: netdev to update 3089 * @from: netdev from which to copy the limits 3090 */ 3091 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3092 { 3093 netif_set_tso_max_size(to, from->tso_max_size); 3094 netif_set_tso_max_segs(to, from->tso_max_segs); 3095 } 3096 EXPORT_SYMBOL(netif_inherit_tso_max); 3097 3098 /** 3099 * netif_get_num_default_rss_queues - default number of RSS queues 3100 * 3101 * Default value is the number of physical cores if there are only 1 or 2, or 3102 * divided by 2 if there are more. 3103 */ 3104 int netif_get_num_default_rss_queues(void) 3105 { 3106 cpumask_var_t cpus; 3107 int cpu, count = 0; 3108 3109 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3110 return 1; 3111 3112 cpumask_copy(cpus, cpu_online_mask); 3113 for_each_cpu(cpu, cpus) { 3114 ++count; 3115 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3116 } 3117 free_cpumask_var(cpus); 3118 3119 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3120 } 3121 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3122 3123 static void __netif_reschedule(struct Qdisc *q) 3124 { 3125 struct softnet_data *sd; 3126 unsigned long flags; 3127 3128 local_irq_save(flags); 3129 sd = this_cpu_ptr(&softnet_data); 3130 q->next_sched = NULL; 3131 *sd->output_queue_tailp = q; 3132 sd->output_queue_tailp = &q->next_sched; 3133 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3134 local_irq_restore(flags); 3135 } 3136 3137 void __netif_schedule(struct Qdisc *q) 3138 { 3139 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3140 __netif_reschedule(q); 3141 } 3142 EXPORT_SYMBOL(__netif_schedule); 3143 3144 struct dev_kfree_skb_cb { 3145 enum skb_drop_reason reason; 3146 }; 3147 3148 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3149 { 3150 return (struct dev_kfree_skb_cb *)skb->cb; 3151 } 3152 3153 void netif_schedule_queue(struct netdev_queue *txq) 3154 { 3155 rcu_read_lock(); 3156 if (!netif_xmit_stopped(txq)) { 3157 struct Qdisc *q = rcu_dereference(txq->qdisc); 3158 3159 __netif_schedule(q); 3160 } 3161 rcu_read_unlock(); 3162 } 3163 EXPORT_SYMBOL(netif_schedule_queue); 3164 3165 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3166 { 3167 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3168 struct Qdisc *q; 3169 3170 rcu_read_lock(); 3171 q = rcu_dereference(dev_queue->qdisc); 3172 __netif_schedule(q); 3173 rcu_read_unlock(); 3174 } 3175 } 3176 EXPORT_SYMBOL(netif_tx_wake_queue); 3177 3178 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3179 { 3180 unsigned long flags; 3181 3182 if (unlikely(!skb)) 3183 return; 3184 3185 if (likely(refcount_read(&skb->users) == 1)) { 3186 smp_rmb(); 3187 refcount_set(&skb->users, 0); 3188 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3189 return; 3190 } 3191 get_kfree_skb_cb(skb)->reason = reason; 3192 local_irq_save(flags); 3193 skb->next = __this_cpu_read(softnet_data.completion_queue); 3194 __this_cpu_write(softnet_data.completion_queue, skb); 3195 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3196 local_irq_restore(flags); 3197 } 3198 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3199 3200 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3201 { 3202 if (in_hardirq() || irqs_disabled()) 3203 dev_kfree_skb_irq_reason(skb, reason); 3204 else 3205 kfree_skb_reason(skb, reason); 3206 } 3207 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3208 3209 3210 /** 3211 * netif_device_detach - mark device as removed 3212 * @dev: network device 3213 * 3214 * Mark device as removed from system and therefore no longer available. 3215 */ 3216 void netif_device_detach(struct net_device *dev) 3217 { 3218 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3219 netif_running(dev)) { 3220 netif_tx_stop_all_queues(dev); 3221 } 3222 } 3223 EXPORT_SYMBOL(netif_device_detach); 3224 3225 /** 3226 * netif_device_attach - mark device as attached 3227 * @dev: network device 3228 * 3229 * Mark device as attached from system and restart if needed. 3230 */ 3231 void netif_device_attach(struct net_device *dev) 3232 { 3233 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3234 netif_running(dev)) { 3235 netif_tx_wake_all_queues(dev); 3236 __netdev_watchdog_up(dev); 3237 } 3238 } 3239 EXPORT_SYMBOL(netif_device_attach); 3240 3241 /* 3242 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3243 * to be used as a distribution range. 3244 */ 3245 static u16 skb_tx_hash(const struct net_device *dev, 3246 const struct net_device *sb_dev, 3247 struct sk_buff *skb) 3248 { 3249 u32 hash; 3250 u16 qoffset = 0; 3251 u16 qcount = dev->real_num_tx_queues; 3252 3253 if (dev->num_tc) { 3254 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3255 3256 qoffset = sb_dev->tc_to_txq[tc].offset; 3257 qcount = sb_dev->tc_to_txq[tc].count; 3258 if (unlikely(!qcount)) { 3259 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3260 sb_dev->name, qoffset, tc); 3261 qoffset = 0; 3262 qcount = dev->real_num_tx_queues; 3263 } 3264 } 3265 3266 if (skb_rx_queue_recorded(skb)) { 3267 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3268 hash = skb_get_rx_queue(skb); 3269 if (hash >= qoffset) 3270 hash -= qoffset; 3271 while (unlikely(hash >= qcount)) 3272 hash -= qcount; 3273 return hash + qoffset; 3274 } 3275 3276 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3277 } 3278 3279 void skb_warn_bad_offload(const struct sk_buff *skb) 3280 { 3281 static const netdev_features_t null_features; 3282 struct net_device *dev = skb->dev; 3283 const char *name = ""; 3284 3285 if (!net_ratelimit()) 3286 return; 3287 3288 if (dev) { 3289 if (dev->dev.parent) 3290 name = dev_driver_string(dev->dev.parent); 3291 else 3292 name = netdev_name(dev); 3293 } 3294 skb_dump(KERN_WARNING, skb, false); 3295 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3296 name, dev ? &dev->features : &null_features, 3297 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3298 } 3299 3300 /* 3301 * Invalidate hardware checksum when packet is to be mangled, and 3302 * complete checksum manually on outgoing path. 3303 */ 3304 int skb_checksum_help(struct sk_buff *skb) 3305 { 3306 __wsum csum; 3307 int ret = 0, offset; 3308 3309 if (skb->ip_summed == CHECKSUM_COMPLETE) 3310 goto out_set_summed; 3311 3312 if (unlikely(skb_is_gso(skb))) { 3313 skb_warn_bad_offload(skb); 3314 return -EINVAL; 3315 } 3316 3317 if (!skb_frags_readable(skb)) { 3318 return -EFAULT; 3319 } 3320 3321 /* Before computing a checksum, we should make sure no frag could 3322 * be modified by an external entity : checksum could be wrong. 3323 */ 3324 if (skb_has_shared_frag(skb)) { 3325 ret = __skb_linearize(skb); 3326 if (ret) 3327 goto out; 3328 } 3329 3330 offset = skb_checksum_start_offset(skb); 3331 ret = -EINVAL; 3332 if (unlikely(offset >= skb_headlen(skb))) { 3333 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3334 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3335 offset, skb_headlen(skb)); 3336 goto out; 3337 } 3338 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3339 3340 offset += skb->csum_offset; 3341 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3342 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3343 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3344 offset + sizeof(__sum16), skb_headlen(skb)); 3345 goto out; 3346 } 3347 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3348 if (ret) 3349 goto out; 3350 3351 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3352 out_set_summed: 3353 skb->ip_summed = CHECKSUM_NONE; 3354 out: 3355 return ret; 3356 } 3357 EXPORT_SYMBOL(skb_checksum_help); 3358 3359 int skb_crc32c_csum_help(struct sk_buff *skb) 3360 { 3361 __le32 crc32c_csum; 3362 int ret = 0, offset, start; 3363 3364 if (skb->ip_summed != CHECKSUM_PARTIAL) 3365 goto out; 3366 3367 if (unlikely(skb_is_gso(skb))) 3368 goto out; 3369 3370 /* Before computing a checksum, we should make sure no frag could 3371 * be modified by an external entity : checksum could be wrong. 3372 */ 3373 if (unlikely(skb_has_shared_frag(skb))) { 3374 ret = __skb_linearize(skb); 3375 if (ret) 3376 goto out; 3377 } 3378 start = skb_checksum_start_offset(skb); 3379 offset = start + offsetof(struct sctphdr, checksum); 3380 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3381 ret = -EINVAL; 3382 goto out; 3383 } 3384 3385 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3386 if (ret) 3387 goto out; 3388 3389 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start, 3390 skb->len - start, ~(__u32)0, 3391 crc32c_csum_stub)); 3392 *(__le32 *)(skb->data + offset) = crc32c_csum; 3393 skb_reset_csum_not_inet(skb); 3394 out: 3395 return ret; 3396 } 3397 EXPORT_SYMBOL(skb_crc32c_csum_help); 3398 3399 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3400 { 3401 __be16 type = skb->protocol; 3402 3403 /* Tunnel gso handlers can set protocol to ethernet. */ 3404 if (type == htons(ETH_P_TEB)) { 3405 struct ethhdr *eth; 3406 3407 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3408 return 0; 3409 3410 eth = (struct ethhdr *)skb->data; 3411 type = eth->h_proto; 3412 } 3413 3414 return vlan_get_protocol_and_depth(skb, type, depth); 3415 } 3416 3417 3418 /* Take action when hardware reception checksum errors are detected. */ 3419 #ifdef CONFIG_BUG 3420 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3421 { 3422 netdev_err(dev, "hw csum failure\n"); 3423 skb_dump(KERN_ERR, skb, true); 3424 dump_stack(); 3425 } 3426 3427 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3428 { 3429 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3430 } 3431 EXPORT_SYMBOL(netdev_rx_csum_fault); 3432 #endif 3433 3434 /* XXX: check that highmem exists at all on the given machine. */ 3435 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3436 { 3437 #ifdef CONFIG_HIGHMEM 3438 int i; 3439 3440 if (!(dev->features & NETIF_F_HIGHDMA)) { 3441 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3442 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3443 struct page *page = skb_frag_page(frag); 3444 3445 if (page && PageHighMem(page)) 3446 return 1; 3447 } 3448 } 3449 #endif 3450 return 0; 3451 } 3452 3453 /* If MPLS offload request, verify we are testing hardware MPLS features 3454 * instead of standard features for the netdev. 3455 */ 3456 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3457 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3458 netdev_features_t features, 3459 __be16 type) 3460 { 3461 if (eth_p_mpls(type)) 3462 features &= skb->dev->mpls_features; 3463 3464 return features; 3465 } 3466 #else 3467 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3468 netdev_features_t features, 3469 __be16 type) 3470 { 3471 return features; 3472 } 3473 #endif 3474 3475 static netdev_features_t harmonize_features(struct sk_buff *skb, 3476 netdev_features_t features) 3477 { 3478 __be16 type; 3479 3480 type = skb_network_protocol(skb, NULL); 3481 features = net_mpls_features(skb, features, type); 3482 3483 if (skb->ip_summed != CHECKSUM_NONE && 3484 !can_checksum_protocol(features, type)) { 3485 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3486 } 3487 if (illegal_highdma(skb->dev, skb)) 3488 features &= ~NETIF_F_SG; 3489 3490 return features; 3491 } 3492 3493 netdev_features_t passthru_features_check(struct sk_buff *skb, 3494 struct net_device *dev, 3495 netdev_features_t features) 3496 { 3497 return features; 3498 } 3499 EXPORT_SYMBOL(passthru_features_check); 3500 3501 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3502 struct net_device *dev, 3503 netdev_features_t features) 3504 { 3505 return vlan_features_check(skb, features); 3506 } 3507 3508 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3509 struct net_device *dev, 3510 netdev_features_t features) 3511 { 3512 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3513 3514 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3515 return features & ~NETIF_F_GSO_MASK; 3516 3517 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3518 return features & ~NETIF_F_GSO_MASK; 3519 3520 if (!skb_shinfo(skb)->gso_type) { 3521 skb_warn_bad_offload(skb); 3522 return features & ~NETIF_F_GSO_MASK; 3523 } 3524 3525 /* Support for GSO partial features requires software 3526 * intervention before we can actually process the packets 3527 * so we need to strip support for any partial features now 3528 * and we can pull them back in after we have partially 3529 * segmented the frame. 3530 */ 3531 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3532 features &= ~dev->gso_partial_features; 3533 3534 /* Make sure to clear the IPv4 ID mangling feature if the 3535 * IPv4 header has the potential to be fragmented. 3536 */ 3537 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3538 struct iphdr *iph = skb->encapsulation ? 3539 inner_ip_hdr(skb) : ip_hdr(skb); 3540 3541 if (!(iph->frag_off & htons(IP_DF))) 3542 features &= ~NETIF_F_TSO_MANGLEID; 3543 } 3544 3545 return features; 3546 } 3547 3548 netdev_features_t netif_skb_features(struct sk_buff *skb) 3549 { 3550 struct net_device *dev = skb->dev; 3551 netdev_features_t features = dev->features; 3552 3553 if (skb_is_gso(skb)) 3554 features = gso_features_check(skb, dev, features); 3555 3556 /* If encapsulation offload request, verify we are testing 3557 * hardware encapsulation features instead of standard 3558 * features for the netdev 3559 */ 3560 if (skb->encapsulation) 3561 features &= dev->hw_enc_features; 3562 3563 if (skb_vlan_tagged(skb)) 3564 features = netdev_intersect_features(features, 3565 dev->vlan_features | 3566 NETIF_F_HW_VLAN_CTAG_TX | 3567 NETIF_F_HW_VLAN_STAG_TX); 3568 3569 if (dev->netdev_ops->ndo_features_check) 3570 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3571 features); 3572 else 3573 features &= dflt_features_check(skb, dev, features); 3574 3575 return harmonize_features(skb, features); 3576 } 3577 EXPORT_SYMBOL(netif_skb_features); 3578 3579 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3580 struct netdev_queue *txq, bool more) 3581 { 3582 unsigned int len; 3583 int rc; 3584 3585 if (dev_nit_active(dev)) 3586 dev_queue_xmit_nit(skb, dev); 3587 3588 len = skb->len; 3589 trace_net_dev_start_xmit(skb, dev); 3590 rc = netdev_start_xmit(skb, dev, txq, more); 3591 trace_net_dev_xmit(skb, rc, dev, len); 3592 3593 return rc; 3594 } 3595 3596 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3597 struct netdev_queue *txq, int *ret) 3598 { 3599 struct sk_buff *skb = first; 3600 int rc = NETDEV_TX_OK; 3601 3602 while (skb) { 3603 struct sk_buff *next = skb->next; 3604 3605 skb_mark_not_on_list(skb); 3606 rc = xmit_one(skb, dev, txq, next != NULL); 3607 if (unlikely(!dev_xmit_complete(rc))) { 3608 skb->next = next; 3609 goto out; 3610 } 3611 3612 skb = next; 3613 if (netif_tx_queue_stopped(txq) && skb) { 3614 rc = NETDEV_TX_BUSY; 3615 break; 3616 } 3617 } 3618 3619 out: 3620 *ret = rc; 3621 return skb; 3622 } 3623 3624 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3625 netdev_features_t features) 3626 { 3627 if (skb_vlan_tag_present(skb) && 3628 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3629 skb = __vlan_hwaccel_push_inside(skb); 3630 return skb; 3631 } 3632 3633 int skb_csum_hwoffload_help(struct sk_buff *skb, 3634 const netdev_features_t features) 3635 { 3636 if (unlikely(skb_csum_is_sctp(skb))) 3637 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3638 skb_crc32c_csum_help(skb); 3639 3640 if (features & NETIF_F_HW_CSUM) 3641 return 0; 3642 3643 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3644 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3645 skb_network_header_len(skb) != sizeof(struct ipv6hdr)) 3646 goto sw_checksum; 3647 switch (skb->csum_offset) { 3648 case offsetof(struct tcphdr, check): 3649 case offsetof(struct udphdr, check): 3650 return 0; 3651 } 3652 } 3653 3654 sw_checksum: 3655 return skb_checksum_help(skb); 3656 } 3657 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3658 3659 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3660 { 3661 netdev_features_t features; 3662 3663 features = netif_skb_features(skb); 3664 skb = validate_xmit_vlan(skb, features); 3665 if (unlikely(!skb)) 3666 goto out_null; 3667 3668 skb = sk_validate_xmit_skb(skb, dev); 3669 if (unlikely(!skb)) 3670 goto out_null; 3671 3672 if (netif_needs_gso(skb, features)) { 3673 struct sk_buff *segs; 3674 3675 segs = skb_gso_segment(skb, features); 3676 if (IS_ERR(segs)) { 3677 goto out_kfree_skb; 3678 } else if (segs) { 3679 consume_skb(skb); 3680 skb = segs; 3681 } 3682 } else { 3683 if (skb_needs_linearize(skb, features) && 3684 __skb_linearize(skb)) 3685 goto out_kfree_skb; 3686 3687 /* If packet is not checksummed and device does not 3688 * support checksumming for this protocol, complete 3689 * checksumming here. 3690 */ 3691 if (skb->ip_summed == CHECKSUM_PARTIAL) { 3692 if (skb->encapsulation) 3693 skb_set_inner_transport_header(skb, 3694 skb_checksum_start_offset(skb)); 3695 else 3696 skb_set_transport_header(skb, 3697 skb_checksum_start_offset(skb)); 3698 if (skb_csum_hwoffload_help(skb, features)) 3699 goto out_kfree_skb; 3700 } 3701 } 3702 3703 skb = validate_xmit_xfrm(skb, features, again); 3704 3705 return skb; 3706 3707 out_kfree_skb: 3708 kfree_skb(skb); 3709 out_null: 3710 dev_core_stats_tx_dropped_inc(dev); 3711 return NULL; 3712 } 3713 3714 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 3715 { 3716 struct sk_buff *next, *head = NULL, *tail; 3717 3718 for (; skb != NULL; skb = next) { 3719 next = skb->next; 3720 skb_mark_not_on_list(skb); 3721 3722 /* in case skb won't be segmented, point to itself */ 3723 skb->prev = skb; 3724 3725 skb = validate_xmit_skb(skb, dev, again); 3726 if (!skb) 3727 continue; 3728 3729 if (!head) 3730 head = skb; 3731 else 3732 tail->next = skb; 3733 /* If skb was segmented, skb->prev points to 3734 * the last segment. If not, it still contains skb. 3735 */ 3736 tail = skb->prev; 3737 } 3738 return head; 3739 } 3740 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 3741 3742 static void qdisc_pkt_len_init(struct sk_buff *skb) 3743 { 3744 const struct skb_shared_info *shinfo = skb_shinfo(skb); 3745 3746 qdisc_skb_cb(skb)->pkt_len = skb->len; 3747 3748 /* To get more precise estimation of bytes sent on wire, 3749 * we add to pkt_len the headers size of all segments 3750 */ 3751 if (shinfo->gso_size && skb_transport_header_was_set(skb)) { 3752 u16 gso_segs = shinfo->gso_segs; 3753 unsigned int hdr_len; 3754 3755 /* mac layer + network layer */ 3756 hdr_len = skb_transport_offset(skb); 3757 3758 /* + transport layer */ 3759 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 3760 const struct tcphdr *th; 3761 struct tcphdr _tcphdr; 3762 3763 th = skb_header_pointer(skb, hdr_len, 3764 sizeof(_tcphdr), &_tcphdr); 3765 if (likely(th)) 3766 hdr_len += __tcp_hdrlen(th); 3767 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 3768 struct udphdr _udphdr; 3769 3770 if (skb_header_pointer(skb, hdr_len, 3771 sizeof(_udphdr), &_udphdr)) 3772 hdr_len += sizeof(struct udphdr); 3773 } 3774 3775 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 3776 int payload = skb->len - hdr_len; 3777 3778 /* Malicious packet. */ 3779 if (payload <= 0) 3780 return; 3781 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 3782 } 3783 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 3784 } 3785 } 3786 3787 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 3788 struct sk_buff **to_free, 3789 struct netdev_queue *txq) 3790 { 3791 int rc; 3792 3793 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 3794 if (rc == NET_XMIT_SUCCESS) 3795 trace_qdisc_enqueue(q, txq, skb); 3796 return rc; 3797 } 3798 3799 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 3800 struct net_device *dev, 3801 struct netdev_queue *txq) 3802 { 3803 spinlock_t *root_lock = qdisc_lock(q); 3804 struct sk_buff *to_free = NULL; 3805 bool contended; 3806 int rc; 3807 3808 qdisc_calculate_pkt_len(skb, q); 3809 3810 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 3811 3812 if (q->flags & TCQ_F_NOLOCK) { 3813 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 3814 qdisc_run_begin(q)) { 3815 /* Retest nolock_qdisc_is_empty() within the protection 3816 * of q->seqlock to protect from racing with requeuing. 3817 */ 3818 if (unlikely(!nolock_qdisc_is_empty(q))) { 3819 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3820 __qdisc_run(q); 3821 qdisc_run_end(q); 3822 3823 goto no_lock_out; 3824 } 3825 3826 qdisc_bstats_cpu_update(q, skb); 3827 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 3828 !nolock_qdisc_is_empty(q)) 3829 __qdisc_run(q); 3830 3831 qdisc_run_end(q); 3832 return NET_XMIT_SUCCESS; 3833 } 3834 3835 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3836 qdisc_run(q); 3837 3838 no_lock_out: 3839 if (unlikely(to_free)) 3840 kfree_skb_list_reason(to_free, 3841 tcf_get_drop_reason(to_free)); 3842 return rc; 3843 } 3844 3845 if (unlikely(READ_ONCE(q->owner) == smp_processor_id())) { 3846 kfree_skb_reason(skb, SKB_DROP_REASON_TC_RECLASSIFY_LOOP); 3847 return NET_XMIT_DROP; 3848 } 3849 /* 3850 * Heuristic to force contended enqueues to serialize on a 3851 * separate lock before trying to get qdisc main lock. 3852 * This permits qdisc->running owner to get the lock more 3853 * often and dequeue packets faster. 3854 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 3855 * and then other tasks will only enqueue packets. The packets will be 3856 * sent after the qdisc owner is scheduled again. To prevent this 3857 * scenario the task always serialize on the lock. 3858 */ 3859 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 3860 if (unlikely(contended)) 3861 spin_lock(&q->busylock); 3862 3863 spin_lock(root_lock); 3864 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 3865 __qdisc_drop(skb, &to_free); 3866 rc = NET_XMIT_DROP; 3867 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 3868 qdisc_run_begin(q)) { 3869 /* 3870 * This is a work-conserving queue; there are no old skbs 3871 * waiting to be sent out; and the qdisc is not running - 3872 * xmit the skb directly. 3873 */ 3874 3875 qdisc_bstats_update(q, skb); 3876 3877 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 3878 if (unlikely(contended)) { 3879 spin_unlock(&q->busylock); 3880 contended = false; 3881 } 3882 __qdisc_run(q); 3883 } 3884 3885 qdisc_run_end(q); 3886 rc = NET_XMIT_SUCCESS; 3887 } else { 3888 WRITE_ONCE(q->owner, smp_processor_id()); 3889 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3890 WRITE_ONCE(q->owner, -1); 3891 if (qdisc_run_begin(q)) { 3892 if (unlikely(contended)) { 3893 spin_unlock(&q->busylock); 3894 contended = false; 3895 } 3896 __qdisc_run(q); 3897 qdisc_run_end(q); 3898 } 3899 } 3900 spin_unlock(root_lock); 3901 if (unlikely(to_free)) 3902 kfree_skb_list_reason(to_free, 3903 tcf_get_drop_reason(to_free)); 3904 if (unlikely(contended)) 3905 spin_unlock(&q->busylock); 3906 return rc; 3907 } 3908 3909 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 3910 static void skb_update_prio(struct sk_buff *skb) 3911 { 3912 const struct netprio_map *map; 3913 const struct sock *sk; 3914 unsigned int prioidx; 3915 3916 if (skb->priority) 3917 return; 3918 map = rcu_dereference_bh(skb->dev->priomap); 3919 if (!map) 3920 return; 3921 sk = skb_to_full_sk(skb); 3922 if (!sk) 3923 return; 3924 3925 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 3926 3927 if (prioidx < map->priomap_len) 3928 skb->priority = map->priomap[prioidx]; 3929 } 3930 #else 3931 #define skb_update_prio(skb) 3932 #endif 3933 3934 /** 3935 * dev_loopback_xmit - loop back @skb 3936 * @net: network namespace this loopback is happening in 3937 * @sk: sk needed to be a netfilter okfn 3938 * @skb: buffer to transmit 3939 */ 3940 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 3941 { 3942 skb_reset_mac_header(skb); 3943 __skb_pull(skb, skb_network_offset(skb)); 3944 skb->pkt_type = PACKET_LOOPBACK; 3945 if (skb->ip_summed == CHECKSUM_NONE) 3946 skb->ip_summed = CHECKSUM_UNNECESSARY; 3947 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 3948 skb_dst_force(skb); 3949 netif_rx(skb); 3950 return 0; 3951 } 3952 EXPORT_SYMBOL(dev_loopback_xmit); 3953 3954 #ifdef CONFIG_NET_EGRESS 3955 static struct netdev_queue * 3956 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 3957 { 3958 int qm = skb_get_queue_mapping(skb); 3959 3960 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 3961 } 3962 3963 #ifndef CONFIG_PREEMPT_RT 3964 static bool netdev_xmit_txqueue_skipped(void) 3965 { 3966 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 3967 } 3968 3969 void netdev_xmit_skip_txqueue(bool skip) 3970 { 3971 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 3972 } 3973 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 3974 3975 #else 3976 static bool netdev_xmit_txqueue_skipped(void) 3977 { 3978 return current->net_xmit.skip_txqueue; 3979 } 3980 3981 void netdev_xmit_skip_txqueue(bool skip) 3982 { 3983 current->net_xmit.skip_txqueue = skip; 3984 } 3985 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 3986 #endif 3987 #endif /* CONFIG_NET_EGRESS */ 3988 3989 #ifdef CONFIG_NET_XGRESS 3990 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 3991 enum skb_drop_reason *drop_reason) 3992 { 3993 int ret = TC_ACT_UNSPEC; 3994 #ifdef CONFIG_NET_CLS_ACT 3995 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 3996 struct tcf_result res; 3997 3998 if (!miniq) 3999 return ret; 4000 4001 if (static_branch_unlikely(&tcf_bypass_check_needed_key)) { 4002 if (tcf_block_bypass_sw(miniq->block)) 4003 return ret; 4004 } 4005 4006 tc_skb_cb(skb)->mru = 0; 4007 tc_skb_cb(skb)->post_ct = false; 4008 tcf_set_drop_reason(skb, *drop_reason); 4009 4010 mini_qdisc_bstats_cpu_update(miniq, skb); 4011 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4012 /* Only tcf related quirks below. */ 4013 switch (ret) { 4014 case TC_ACT_SHOT: 4015 *drop_reason = tcf_get_drop_reason(skb); 4016 mini_qdisc_qstats_cpu_drop(miniq); 4017 break; 4018 case TC_ACT_OK: 4019 case TC_ACT_RECLASSIFY: 4020 skb->tc_index = TC_H_MIN(res.classid); 4021 break; 4022 } 4023 #endif /* CONFIG_NET_CLS_ACT */ 4024 return ret; 4025 } 4026 4027 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4028 4029 void tcx_inc(void) 4030 { 4031 static_branch_inc(&tcx_needed_key); 4032 } 4033 4034 void tcx_dec(void) 4035 { 4036 static_branch_dec(&tcx_needed_key); 4037 } 4038 4039 static __always_inline enum tcx_action_base 4040 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4041 const bool needs_mac) 4042 { 4043 const struct bpf_mprog_fp *fp; 4044 const struct bpf_prog *prog; 4045 int ret = TCX_NEXT; 4046 4047 if (needs_mac) 4048 __skb_push(skb, skb->mac_len); 4049 bpf_mprog_foreach_prog(entry, fp, prog) { 4050 bpf_compute_data_pointers(skb); 4051 ret = bpf_prog_run(prog, skb); 4052 if (ret != TCX_NEXT) 4053 break; 4054 } 4055 if (needs_mac) 4056 __skb_pull(skb, skb->mac_len); 4057 return tcx_action_code(skb, ret); 4058 } 4059 4060 static __always_inline struct sk_buff * 4061 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4062 struct net_device *orig_dev, bool *another) 4063 { 4064 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4065 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4066 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4067 int sch_ret; 4068 4069 if (!entry) 4070 return skb; 4071 4072 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4073 if (*pt_prev) { 4074 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4075 *pt_prev = NULL; 4076 } 4077 4078 qdisc_skb_cb(skb)->pkt_len = skb->len; 4079 tcx_set_ingress(skb, true); 4080 4081 if (static_branch_unlikely(&tcx_needed_key)) { 4082 sch_ret = tcx_run(entry, skb, true); 4083 if (sch_ret != TC_ACT_UNSPEC) 4084 goto ingress_verdict; 4085 } 4086 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4087 ingress_verdict: 4088 switch (sch_ret) { 4089 case TC_ACT_REDIRECT: 4090 /* skb_mac_header check was done by BPF, so we can safely 4091 * push the L2 header back before redirecting to another 4092 * netdev. 4093 */ 4094 __skb_push(skb, skb->mac_len); 4095 if (skb_do_redirect(skb) == -EAGAIN) { 4096 __skb_pull(skb, skb->mac_len); 4097 *another = true; 4098 break; 4099 } 4100 *ret = NET_RX_SUCCESS; 4101 bpf_net_ctx_clear(bpf_net_ctx); 4102 return NULL; 4103 case TC_ACT_SHOT: 4104 kfree_skb_reason(skb, drop_reason); 4105 *ret = NET_RX_DROP; 4106 bpf_net_ctx_clear(bpf_net_ctx); 4107 return NULL; 4108 /* used by tc_run */ 4109 case TC_ACT_STOLEN: 4110 case TC_ACT_QUEUED: 4111 case TC_ACT_TRAP: 4112 consume_skb(skb); 4113 fallthrough; 4114 case TC_ACT_CONSUMED: 4115 *ret = NET_RX_SUCCESS; 4116 bpf_net_ctx_clear(bpf_net_ctx); 4117 return NULL; 4118 } 4119 bpf_net_ctx_clear(bpf_net_ctx); 4120 4121 return skb; 4122 } 4123 4124 static __always_inline struct sk_buff * 4125 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4126 { 4127 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4128 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4129 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4130 int sch_ret; 4131 4132 if (!entry) 4133 return skb; 4134 4135 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4136 4137 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4138 * already set by the caller. 4139 */ 4140 if (static_branch_unlikely(&tcx_needed_key)) { 4141 sch_ret = tcx_run(entry, skb, false); 4142 if (sch_ret != TC_ACT_UNSPEC) 4143 goto egress_verdict; 4144 } 4145 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4146 egress_verdict: 4147 switch (sch_ret) { 4148 case TC_ACT_REDIRECT: 4149 /* No need to push/pop skb's mac_header here on egress! */ 4150 skb_do_redirect(skb); 4151 *ret = NET_XMIT_SUCCESS; 4152 bpf_net_ctx_clear(bpf_net_ctx); 4153 return NULL; 4154 case TC_ACT_SHOT: 4155 kfree_skb_reason(skb, drop_reason); 4156 *ret = NET_XMIT_DROP; 4157 bpf_net_ctx_clear(bpf_net_ctx); 4158 return NULL; 4159 /* used by tc_run */ 4160 case TC_ACT_STOLEN: 4161 case TC_ACT_QUEUED: 4162 case TC_ACT_TRAP: 4163 consume_skb(skb); 4164 fallthrough; 4165 case TC_ACT_CONSUMED: 4166 *ret = NET_XMIT_SUCCESS; 4167 bpf_net_ctx_clear(bpf_net_ctx); 4168 return NULL; 4169 } 4170 bpf_net_ctx_clear(bpf_net_ctx); 4171 4172 return skb; 4173 } 4174 #else 4175 static __always_inline struct sk_buff * 4176 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4177 struct net_device *orig_dev, bool *another) 4178 { 4179 return skb; 4180 } 4181 4182 static __always_inline struct sk_buff * 4183 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4184 { 4185 return skb; 4186 } 4187 #endif /* CONFIG_NET_XGRESS */ 4188 4189 #ifdef CONFIG_XPS 4190 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4191 struct xps_dev_maps *dev_maps, unsigned int tci) 4192 { 4193 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4194 struct xps_map *map; 4195 int queue_index = -1; 4196 4197 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4198 return queue_index; 4199 4200 tci *= dev_maps->num_tc; 4201 tci += tc; 4202 4203 map = rcu_dereference(dev_maps->attr_map[tci]); 4204 if (map) { 4205 if (map->len == 1) 4206 queue_index = map->queues[0]; 4207 else 4208 queue_index = map->queues[reciprocal_scale( 4209 skb_get_hash(skb), map->len)]; 4210 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4211 queue_index = -1; 4212 } 4213 return queue_index; 4214 } 4215 #endif 4216 4217 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4218 struct sk_buff *skb) 4219 { 4220 #ifdef CONFIG_XPS 4221 struct xps_dev_maps *dev_maps; 4222 struct sock *sk = skb->sk; 4223 int queue_index = -1; 4224 4225 if (!static_key_false(&xps_needed)) 4226 return -1; 4227 4228 rcu_read_lock(); 4229 if (!static_key_false(&xps_rxqs_needed)) 4230 goto get_cpus_map; 4231 4232 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4233 if (dev_maps) { 4234 int tci = sk_rx_queue_get(sk); 4235 4236 if (tci >= 0) 4237 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4238 tci); 4239 } 4240 4241 get_cpus_map: 4242 if (queue_index < 0) { 4243 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4244 if (dev_maps) { 4245 unsigned int tci = skb->sender_cpu - 1; 4246 4247 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4248 tci); 4249 } 4250 } 4251 rcu_read_unlock(); 4252 4253 return queue_index; 4254 #else 4255 return -1; 4256 #endif 4257 } 4258 4259 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4260 struct net_device *sb_dev) 4261 { 4262 return 0; 4263 } 4264 EXPORT_SYMBOL(dev_pick_tx_zero); 4265 4266 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4267 struct net_device *sb_dev) 4268 { 4269 struct sock *sk = skb->sk; 4270 int queue_index = sk_tx_queue_get(sk); 4271 4272 sb_dev = sb_dev ? : dev; 4273 4274 if (queue_index < 0 || skb->ooo_okay || 4275 queue_index >= dev->real_num_tx_queues) { 4276 int new_index = get_xps_queue(dev, sb_dev, skb); 4277 4278 if (new_index < 0) 4279 new_index = skb_tx_hash(dev, sb_dev, skb); 4280 4281 if (queue_index != new_index && sk && 4282 sk_fullsock(sk) && 4283 rcu_access_pointer(sk->sk_dst_cache)) 4284 sk_tx_queue_set(sk, new_index); 4285 4286 queue_index = new_index; 4287 } 4288 4289 return queue_index; 4290 } 4291 EXPORT_SYMBOL(netdev_pick_tx); 4292 4293 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4294 struct sk_buff *skb, 4295 struct net_device *sb_dev) 4296 { 4297 int queue_index = 0; 4298 4299 #ifdef CONFIG_XPS 4300 u32 sender_cpu = skb->sender_cpu - 1; 4301 4302 if (sender_cpu >= (u32)NR_CPUS) 4303 skb->sender_cpu = raw_smp_processor_id() + 1; 4304 #endif 4305 4306 if (dev->real_num_tx_queues != 1) { 4307 const struct net_device_ops *ops = dev->netdev_ops; 4308 4309 if (ops->ndo_select_queue) 4310 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4311 else 4312 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4313 4314 queue_index = netdev_cap_txqueue(dev, queue_index); 4315 } 4316 4317 skb_set_queue_mapping(skb, queue_index); 4318 return netdev_get_tx_queue(dev, queue_index); 4319 } 4320 4321 /** 4322 * __dev_queue_xmit() - transmit a buffer 4323 * @skb: buffer to transmit 4324 * @sb_dev: suboordinate device used for L2 forwarding offload 4325 * 4326 * Queue a buffer for transmission to a network device. The caller must 4327 * have set the device and priority and built the buffer before calling 4328 * this function. The function can be called from an interrupt. 4329 * 4330 * When calling this method, interrupts MUST be enabled. This is because 4331 * the BH enable code must have IRQs enabled so that it will not deadlock. 4332 * 4333 * Regardless of the return value, the skb is consumed, so it is currently 4334 * difficult to retry a send to this method. (You can bump the ref count 4335 * before sending to hold a reference for retry if you are careful.) 4336 * 4337 * Return: 4338 * * 0 - buffer successfully transmitted 4339 * * positive qdisc return code - NET_XMIT_DROP etc. 4340 * * negative errno - other errors 4341 */ 4342 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4343 { 4344 struct net_device *dev = skb->dev; 4345 struct netdev_queue *txq = NULL; 4346 struct Qdisc *q; 4347 int rc = -ENOMEM; 4348 bool again = false; 4349 4350 skb_reset_mac_header(skb); 4351 skb_assert_len(skb); 4352 4353 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP)) 4354 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4355 4356 /* Disable soft irqs for various locks below. Also 4357 * stops preemption for RCU. 4358 */ 4359 rcu_read_lock_bh(); 4360 4361 skb_update_prio(skb); 4362 4363 qdisc_pkt_len_init(skb); 4364 tcx_set_ingress(skb, false); 4365 #ifdef CONFIG_NET_EGRESS 4366 if (static_branch_unlikely(&egress_needed_key)) { 4367 if (nf_hook_egress_active()) { 4368 skb = nf_hook_egress(skb, &rc, dev); 4369 if (!skb) 4370 goto out; 4371 } 4372 4373 netdev_xmit_skip_txqueue(false); 4374 4375 nf_skip_egress(skb, true); 4376 skb = sch_handle_egress(skb, &rc, dev); 4377 if (!skb) 4378 goto out; 4379 nf_skip_egress(skb, false); 4380 4381 if (netdev_xmit_txqueue_skipped()) 4382 txq = netdev_tx_queue_mapping(dev, skb); 4383 } 4384 #endif 4385 /* If device/qdisc don't need skb->dst, release it right now while 4386 * its hot in this cpu cache. 4387 */ 4388 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4389 skb_dst_drop(skb); 4390 else 4391 skb_dst_force(skb); 4392 4393 if (!txq) 4394 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4395 4396 q = rcu_dereference_bh(txq->qdisc); 4397 4398 trace_net_dev_queue(skb); 4399 if (q->enqueue) { 4400 rc = __dev_xmit_skb(skb, q, dev, txq); 4401 goto out; 4402 } 4403 4404 /* The device has no queue. Common case for software devices: 4405 * loopback, all the sorts of tunnels... 4406 4407 * Really, it is unlikely that netif_tx_lock protection is necessary 4408 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4409 * counters.) 4410 * However, it is possible, that they rely on protection 4411 * made by us here. 4412 4413 * Check this and shot the lock. It is not prone from deadlocks. 4414 *Either shot noqueue qdisc, it is even simpler 8) 4415 */ 4416 if (dev->flags & IFF_UP) { 4417 int cpu = smp_processor_id(); /* ok because BHs are off */ 4418 4419 /* Other cpus might concurrently change txq->xmit_lock_owner 4420 * to -1 or to their cpu id, but not to our id. 4421 */ 4422 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4423 if (dev_xmit_recursion()) 4424 goto recursion_alert; 4425 4426 skb = validate_xmit_skb(skb, dev, &again); 4427 if (!skb) 4428 goto out; 4429 4430 HARD_TX_LOCK(dev, txq, cpu); 4431 4432 if (!netif_xmit_stopped(txq)) { 4433 dev_xmit_recursion_inc(); 4434 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4435 dev_xmit_recursion_dec(); 4436 if (dev_xmit_complete(rc)) { 4437 HARD_TX_UNLOCK(dev, txq); 4438 goto out; 4439 } 4440 } 4441 HARD_TX_UNLOCK(dev, txq); 4442 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4443 dev->name); 4444 } else { 4445 /* Recursion is detected! It is possible, 4446 * unfortunately 4447 */ 4448 recursion_alert: 4449 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4450 dev->name); 4451 } 4452 } 4453 4454 rc = -ENETDOWN; 4455 rcu_read_unlock_bh(); 4456 4457 dev_core_stats_tx_dropped_inc(dev); 4458 kfree_skb_list(skb); 4459 return rc; 4460 out: 4461 rcu_read_unlock_bh(); 4462 return rc; 4463 } 4464 EXPORT_SYMBOL(__dev_queue_xmit); 4465 4466 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4467 { 4468 struct net_device *dev = skb->dev; 4469 struct sk_buff *orig_skb = skb; 4470 struct netdev_queue *txq; 4471 int ret = NETDEV_TX_BUSY; 4472 bool again = false; 4473 4474 if (unlikely(!netif_running(dev) || 4475 !netif_carrier_ok(dev))) 4476 goto drop; 4477 4478 skb = validate_xmit_skb_list(skb, dev, &again); 4479 if (skb != orig_skb) 4480 goto drop; 4481 4482 skb_set_queue_mapping(skb, queue_id); 4483 txq = skb_get_tx_queue(dev, skb); 4484 4485 local_bh_disable(); 4486 4487 dev_xmit_recursion_inc(); 4488 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4489 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4490 ret = netdev_start_xmit(skb, dev, txq, false); 4491 HARD_TX_UNLOCK(dev, txq); 4492 dev_xmit_recursion_dec(); 4493 4494 local_bh_enable(); 4495 return ret; 4496 drop: 4497 dev_core_stats_tx_dropped_inc(dev); 4498 kfree_skb_list(skb); 4499 return NET_XMIT_DROP; 4500 } 4501 EXPORT_SYMBOL(__dev_direct_xmit); 4502 4503 /************************************************************************* 4504 * Receiver routines 4505 *************************************************************************/ 4506 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4507 4508 int weight_p __read_mostly = 64; /* old backlog weight */ 4509 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4510 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4511 4512 /* Called with irq disabled */ 4513 static inline void ____napi_schedule(struct softnet_data *sd, 4514 struct napi_struct *napi) 4515 { 4516 struct task_struct *thread; 4517 4518 lockdep_assert_irqs_disabled(); 4519 4520 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4521 /* Paired with smp_mb__before_atomic() in 4522 * napi_enable()/dev_set_threaded(). 4523 * Use READ_ONCE() to guarantee a complete 4524 * read on napi->thread. Only call 4525 * wake_up_process() when it's not NULL. 4526 */ 4527 thread = READ_ONCE(napi->thread); 4528 if (thread) { 4529 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4530 goto use_local_napi; 4531 4532 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4533 wake_up_process(thread); 4534 return; 4535 } 4536 } 4537 4538 use_local_napi: 4539 list_add_tail(&napi->poll_list, &sd->poll_list); 4540 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4541 /* If not called from net_rx_action() 4542 * we have to raise NET_RX_SOFTIRQ. 4543 */ 4544 if (!sd->in_net_rx_action) 4545 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4546 } 4547 4548 #ifdef CONFIG_RPS 4549 4550 struct static_key_false rps_needed __read_mostly; 4551 EXPORT_SYMBOL(rps_needed); 4552 struct static_key_false rfs_needed __read_mostly; 4553 EXPORT_SYMBOL(rfs_needed); 4554 4555 static struct rps_dev_flow * 4556 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4557 struct rps_dev_flow *rflow, u16 next_cpu) 4558 { 4559 if (next_cpu < nr_cpu_ids) { 4560 u32 head; 4561 #ifdef CONFIG_RFS_ACCEL 4562 struct netdev_rx_queue *rxqueue; 4563 struct rps_dev_flow_table *flow_table; 4564 struct rps_dev_flow *old_rflow; 4565 u16 rxq_index; 4566 u32 flow_id; 4567 int rc; 4568 4569 /* Should we steer this flow to a different hardware queue? */ 4570 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4571 !(dev->features & NETIF_F_NTUPLE)) 4572 goto out; 4573 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4574 if (rxq_index == skb_get_rx_queue(skb)) 4575 goto out; 4576 4577 rxqueue = dev->_rx + rxq_index; 4578 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4579 if (!flow_table) 4580 goto out; 4581 flow_id = skb_get_hash(skb) & flow_table->mask; 4582 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4583 rxq_index, flow_id); 4584 if (rc < 0) 4585 goto out; 4586 old_rflow = rflow; 4587 rflow = &flow_table->flows[flow_id]; 4588 WRITE_ONCE(rflow->filter, rc); 4589 if (old_rflow->filter == rc) 4590 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 4591 out: 4592 #endif 4593 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 4594 rps_input_queue_tail_save(&rflow->last_qtail, head); 4595 } 4596 4597 WRITE_ONCE(rflow->cpu, next_cpu); 4598 return rflow; 4599 } 4600 4601 /* 4602 * get_rps_cpu is called from netif_receive_skb and returns the target 4603 * CPU from the RPS map of the receiving queue for a given skb. 4604 * rcu_read_lock must be held on entry. 4605 */ 4606 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4607 struct rps_dev_flow **rflowp) 4608 { 4609 const struct rps_sock_flow_table *sock_flow_table; 4610 struct netdev_rx_queue *rxqueue = dev->_rx; 4611 struct rps_dev_flow_table *flow_table; 4612 struct rps_map *map; 4613 int cpu = -1; 4614 u32 tcpu; 4615 u32 hash; 4616 4617 if (skb_rx_queue_recorded(skb)) { 4618 u16 index = skb_get_rx_queue(skb); 4619 4620 if (unlikely(index >= dev->real_num_rx_queues)) { 4621 WARN_ONCE(dev->real_num_rx_queues > 1, 4622 "%s received packet on queue %u, but number " 4623 "of RX queues is %u\n", 4624 dev->name, index, dev->real_num_rx_queues); 4625 goto done; 4626 } 4627 rxqueue += index; 4628 } 4629 4630 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4631 4632 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4633 map = rcu_dereference(rxqueue->rps_map); 4634 if (!flow_table && !map) 4635 goto done; 4636 4637 skb_reset_network_header(skb); 4638 hash = skb_get_hash(skb); 4639 if (!hash) 4640 goto done; 4641 4642 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4643 if (flow_table && sock_flow_table) { 4644 struct rps_dev_flow *rflow; 4645 u32 next_cpu; 4646 u32 ident; 4647 4648 /* First check into global flow table if there is a match. 4649 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4650 */ 4651 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4652 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4653 goto try_rps; 4654 4655 next_cpu = ident & net_hotdata.rps_cpu_mask; 4656 4657 /* OK, now we know there is a match, 4658 * we can look at the local (per receive queue) flow table 4659 */ 4660 rflow = &flow_table->flows[hash & flow_table->mask]; 4661 tcpu = rflow->cpu; 4662 4663 /* 4664 * If the desired CPU (where last recvmsg was done) is 4665 * different from current CPU (one in the rx-queue flow 4666 * table entry), switch if one of the following holds: 4667 * - Current CPU is unset (>= nr_cpu_ids). 4668 * - Current CPU is offline. 4669 * - The current CPU's queue tail has advanced beyond the 4670 * last packet that was enqueued using this table entry. 4671 * This guarantees that all previous packets for the flow 4672 * have been dequeued, thus preserving in order delivery. 4673 */ 4674 if (unlikely(tcpu != next_cpu) && 4675 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4676 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 4677 rflow->last_qtail)) >= 0)) { 4678 tcpu = next_cpu; 4679 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4680 } 4681 4682 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4683 *rflowp = rflow; 4684 cpu = tcpu; 4685 goto done; 4686 } 4687 } 4688 4689 try_rps: 4690 4691 if (map) { 4692 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4693 if (cpu_online(tcpu)) { 4694 cpu = tcpu; 4695 goto done; 4696 } 4697 } 4698 4699 done: 4700 return cpu; 4701 } 4702 4703 #ifdef CONFIG_RFS_ACCEL 4704 4705 /** 4706 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4707 * @dev: Device on which the filter was set 4708 * @rxq_index: RX queue index 4709 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4710 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4711 * 4712 * Drivers that implement ndo_rx_flow_steer() should periodically call 4713 * this function for each installed filter and remove the filters for 4714 * which it returns %true. 4715 */ 4716 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4717 u32 flow_id, u16 filter_id) 4718 { 4719 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4720 struct rps_dev_flow_table *flow_table; 4721 struct rps_dev_flow *rflow; 4722 bool expire = true; 4723 unsigned int cpu; 4724 4725 rcu_read_lock(); 4726 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4727 if (flow_table && flow_id <= flow_table->mask) { 4728 rflow = &flow_table->flows[flow_id]; 4729 cpu = READ_ONCE(rflow->cpu); 4730 if (READ_ONCE(rflow->filter) == filter_id && cpu < nr_cpu_ids && 4731 ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - 4732 READ_ONCE(rflow->last_qtail)) < 4733 (int)(10 * flow_table->mask))) 4734 expire = false; 4735 } 4736 rcu_read_unlock(); 4737 return expire; 4738 } 4739 EXPORT_SYMBOL(rps_may_expire_flow); 4740 4741 #endif /* CONFIG_RFS_ACCEL */ 4742 4743 /* Called from hardirq (IPI) context */ 4744 static void rps_trigger_softirq(void *data) 4745 { 4746 struct softnet_data *sd = data; 4747 4748 ____napi_schedule(sd, &sd->backlog); 4749 sd->received_rps++; 4750 } 4751 4752 #endif /* CONFIG_RPS */ 4753 4754 /* Called from hardirq (IPI) context */ 4755 static void trigger_rx_softirq(void *data) 4756 { 4757 struct softnet_data *sd = data; 4758 4759 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4760 smp_store_release(&sd->defer_ipi_scheduled, 0); 4761 } 4762 4763 /* 4764 * After we queued a packet into sd->input_pkt_queue, 4765 * we need to make sure this queue is serviced soon. 4766 * 4767 * - If this is another cpu queue, link it to our rps_ipi_list, 4768 * and make sure we will process rps_ipi_list from net_rx_action(). 4769 * 4770 * - If this is our own queue, NAPI schedule our backlog. 4771 * Note that this also raises NET_RX_SOFTIRQ. 4772 */ 4773 static void napi_schedule_rps(struct softnet_data *sd) 4774 { 4775 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 4776 4777 #ifdef CONFIG_RPS 4778 if (sd != mysd) { 4779 if (use_backlog_threads()) { 4780 __napi_schedule_irqoff(&sd->backlog); 4781 return; 4782 } 4783 4784 sd->rps_ipi_next = mysd->rps_ipi_list; 4785 mysd->rps_ipi_list = sd; 4786 4787 /* If not called from net_rx_action() or napi_threaded_poll() 4788 * we have to raise NET_RX_SOFTIRQ. 4789 */ 4790 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 4791 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4792 return; 4793 } 4794 #endif /* CONFIG_RPS */ 4795 __napi_schedule_irqoff(&mysd->backlog); 4796 } 4797 4798 void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu) 4799 { 4800 unsigned long flags; 4801 4802 if (use_backlog_threads()) { 4803 backlog_lock_irq_save(sd, &flags); 4804 4805 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 4806 __napi_schedule_irqoff(&sd->backlog); 4807 4808 backlog_unlock_irq_restore(sd, &flags); 4809 4810 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 4811 smp_call_function_single_async(cpu, &sd->defer_csd); 4812 } 4813 } 4814 4815 #ifdef CONFIG_NET_FLOW_LIMIT 4816 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 4817 #endif 4818 4819 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 4820 { 4821 #ifdef CONFIG_NET_FLOW_LIMIT 4822 struct sd_flow_limit *fl; 4823 struct softnet_data *sd; 4824 unsigned int old_flow, new_flow; 4825 4826 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 4827 return false; 4828 4829 sd = this_cpu_ptr(&softnet_data); 4830 4831 rcu_read_lock(); 4832 fl = rcu_dereference(sd->flow_limit); 4833 if (fl) { 4834 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1); 4835 old_flow = fl->history[fl->history_head]; 4836 fl->history[fl->history_head] = new_flow; 4837 4838 fl->history_head++; 4839 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 4840 4841 if (likely(fl->buckets[old_flow])) 4842 fl->buckets[old_flow]--; 4843 4844 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 4845 fl->count++; 4846 rcu_read_unlock(); 4847 return true; 4848 } 4849 } 4850 rcu_read_unlock(); 4851 #endif 4852 return false; 4853 } 4854 4855 /* 4856 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 4857 * queue (may be a remote CPU queue). 4858 */ 4859 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 4860 unsigned int *qtail) 4861 { 4862 enum skb_drop_reason reason; 4863 struct softnet_data *sd; 4864 unsigned long flags; 4865 unsigned int qlen; 4866 int max_backlog; 4867 u32 tail; 4868 4869 reason = SKB_DROP_REASON_DEV_READY; 4870 if (!netif_running(skb->dev)) 4871 goto bad_dev; 4872 4873 reason = SKB_DROP_REASON_CPU_BACKLOG; 4874 sd = &per_cpu(softnet_data, cpu); 4875 4876 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 4877 max_backlog = READ_ONCE(net_hotdata.max_backlog); 4878 if (unlikely(qlen > max_backlog)) 4879 goto cpu_backlog_drop; 4880 backlog_lock_irq_save(sd, &flags); 4881 qlen = skb_queue_len(&sd->input_pkt_queue); 4882 if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) { 4883 if (!qlen) { 4884 /* Schedule NAPI for backlog device. We can use 4885 * non atomic operation as we own the queue lock. 4886 */ 4887 if (!__test_and_set_bit(NAPI_STATE_SCHED, 4888 &sd->backlog.state)) 4889 napi_schedule_rps(sd); 4890 } 4891 __skb_queue_tail(&sd->input_pkt_queue, skb); 4892 tail = rps_input_queue_tail_incr(sd); 4893 backlog_unlock_irq_restore(sd, &flags); 4894 4895 /* save the tail outside of the critical section */ 4896 rps_input_queue_tail_save(qtail, tail); 4897 return NET_RX_SUCCESS; 4898 } 4899 4900 backlog_unlock_irq_restore(sd, &flags); 4901 4902 cpu_backlog_drop: 4903 atomic_inc(&sd->dropped); 4904 bad_dev: 4905 dev_core_stats_rx_dropped_inc(skb->dev); 4906 kfree_skb_reason(skb, reason); 4907 return NET_RX_DROP; 4908 } 4909 4910 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 4911 { 4912 struct net_device *dev = skb->dev; 4913 struct netdev_rx_queue *rxqueue; 4914 4915 rxqueue = dev->_rx; 4916 4917 if (skb_rx_queue_recorded(skb)) { 4918 u16 index = skb_get_rx_queue(skb); 4919 4920 if (unlikely(index >= dev->real_num_rx_queues)) { 4921 WARN_ONCE(dev->real_num_rx_queues > 1, 4922 "%s received packet on queue %u, but number " 4923 "of RX queues is %u\n", 4924 dev->name, index, dev->real_num_rx_queues); 4925 4926 return rxqueue; /* Return first rxqueue */ 4927 } 4928 rxqueue += index; 4929 } 4930 return rxqueue; 4931 } 4932 4933 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 4934 struct bpf_prog *xdp_prog) 4935 { 4936 void *orig_data, *orig_data_end, *hard_start; 4937 struct netdev_rx_queue *rxqueue; 4938 bool orig_bcast, orig_host; 4939 u32 mac_len, frame_sz; 4940 __be16 orig_eth_type; 4941 struct ethhdr *eth; 4942 u32 metalen, act; 4943 int off; 4944 4945 /* The XDP program wants to see the packet starting at the MAC 4946 * header. 4947 */ 4948 mac_len = skb->data - skb_mac_header(skb); 4949 hard_start = skb->data - skb_headroom(skb); 4950 4951 /* SKB "head" area always have tailroom for skb_shared_info */ 4952 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 4953 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 4954 4955 rxqueue = netif_get_rxqueue(skb); 4956 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 4957 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 4958 skb_headlen(skb) + mac_len, true); 4959 if (skb_is_nonlinear(skb)) { 4960 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 4961 xdp_buff_set_frags_flag(xdp); 4962 } else { 4963 xdp_buff_clear_frags_flag(xdp); 4964 } 4965 4966 orig_data_end = xdp->data_end; 4967 orig_data = xdp->data; 4968 eth = (struct ethhdr *)xdp->data; 4969 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 4970 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 4971 orig_eth_type = eth->h_proto; 4972 4973 act = bpf_prog_run_xdp(xdp_prog, xdp); 4974 4975 /* check if bpf_xdp_adjust_head was used */ 4976 off = xdp->data - orig_data; 4977 if (off) { 4978 if (off > 0) 4979 __skb_pull(skb, off); 4980 else if (off < 0) 4981 __skb_push(skb, -off); 4982 4983 skb->mac_header += off; 4984 skb_reset_network_header(skb); 4985 } 4986 4987 /* check if bpf_xdp_adjust_tail was used */ 4988 off = xdp->data_end - orig_data_end; 4989 if (off != 0) { 4990 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 4991 skb->len += off; /* positive on grow, negative on shrink */ 4992 } 4993 4994 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 4995 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 4996 */ 4997 if (xdp_buff_has_frags(xdp)) 4998 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 4999 else 5000 skb->data_len = 0; 5001 5002 /* check if XDP changed eth hdr such SKB needs update */ 5003 eth = (struct ethhdr *)xdp->data; 5004 if ((orig_eth_type != eth->h_proto) || 5005 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5006 skb->dev->dev_addr)) || 5007 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5008 __skb_push(skb, ETH_HLEN); 5009 skb->pkt_type = PACKET_HOST; 5010 skb->protocol = eth_type_trans(skb, skb->dev); 5011 } 5012 5013 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5014 * before calling us again on redirect path. We do not call do_redirect 5015 * as we leave that up to the caller. 5016 * 5017 * Caller is responsible for managing lifetime of skb (i.e. calling 5018 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5019 */ 5020 switch (act) { 5021 case XDP_REDIRECT: 5022 case XDP_TX: 5023 __skb_push(skb, mac_len); 5024 break; 5025 case XDP_PASS: 5026 metalen = xdp->data - xdp->data_meta; 5027 if (metalen) 5028 skb_metadata_set(skb, metalen); 5029 break; 5030 } 5031 5032 return act; 5033 } 5034 5035 static int 5036 netif_skb_check_for_xdp(struct sk_buff **pskb, struct bpf_prog *prog) 5037 { 5038 struct sk_buff *skb = *pskb; 5039 int err, hroom, troom; 5040 5041 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 5042 return 0; 5043 5044 /* In case we have to go down the path and also linearize, 5045 * then lets do the pskb_expand_head() work just once here. 5046 */ 5047 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5048 troom = skb->tail + skb->data_len - skb->end; 5049 err = pskb_expand_head(skb, 5050 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5051 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5052 if (err) 5053 return err; 5054 5055 return skb_linearize(skb); 5056 } 5057 5058 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5059 struct xdp_buff *xdp, 5060 struct bpf_prog *xdp_prog) 5061 { 5062 struct sk_buff *skb = *pskb; 5063 u32 mac_len, act = XDP_DROP; 5064 5065 /* Reinjected packets coming from act_mirred or similar should 5066 * not get XDP generic processing. 5067 */ 5068 if (skb_is_redirected(skb)) 5069 return XDP_PASS; 5070 5071 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5072 * bytes. This is the guarantee that also native XDP provides, 5073 * thus we need to do it here as well. 5074 */ 5075 mac_len = skb->data - skb_mac_header(skb); 5076 __skb_push(skb, mac_len); 5077 5078 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5079 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5080 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5081 goto do_drop; 5082 } 5083 5084 __skb_pull(*pskb, mac_len); 5085 5086 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5087 switch (act) { 5088 case XDP_REDIRECT: 5089 case XDP_TX: 5090 case XDP_PASS: 5091 break; 5092 default: 5093 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5094 fallthrough; 5095 case XDP_ABORTED: 5096 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5097 fallthrough; 5098 case XDP_DROP: 5099 do_drop: 5100 kfree_skb(*pskb); 5101 break; 5102 } 5103 5104 return act; 5105 } 5106 5107 /* When doing generic XDP we have to bypass the qdisc layer and the 5108 * network taps in order to match in-driver-XDP behavior. This also means 5109 * that XDP packets are able to starve other packets going through a qdisc, 5110 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5111 * queues, so they do not have this starvation issue. 5112 */ 5113 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog) 5114 { 5115 struct net_device *dev = skb->dev; 5116 struct netdev_queue *txq; 5117 bool free_skb = true; 5118 int cpu, rc; 5119 5120 txq = netdev_core_pick_tx(dev, skb, NULL); 5121 cpu = smp_processor_id(); 5122 HARD_TX_LOCK(dev, txq, cpu); 5123 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5124 rc = netdev_start_xmit(skb, dev, txq, 0); 5125 if (dev_xmit_complete(rc)) 5126 free_skb = false; 5127 } 5128 HARD_TX_UNLOCK(dev, txq); 5129 if (free_skb) { 5130 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5131 dev_core_stats_tx_dropped_inc(dev); 5132 kfree_skb(skb); 5133 } 5134 } 5135 5136 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5137 5138 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5139 { 5140 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5141 5142 if (xdp_prog) { 5143 struct xdp_buff xdp; 5144 u32 act; 5145 int err; 5146 5147 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5148 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5149 if (act != XDP_PASS) { 5150 switch (act) { 5151 case XDP_REDIRECT: 5152 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5153 &xdp, xdp_prog); 5154 if (err) 5155 goto out_redir; 5156 break; 5157 case XDP_TX: 5158 generic_xdp_tx(*pskb, xdp_prog); 5159 break; 5160 } 5161 bpf_net_ctx_clear(bpf_net_ctx); 5162 return XDP_DROP; 5163 } 5164 bpf_net_ctx_clear(bpf_net_ctx); 5165 } 5166 return XDP_PASS; 5167 out_redir: 5168 bpf_net_ctx_clear(bpf_net_ctx); 5169 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5170 return XDP_DROP; 5171 } 5172 EXPORT_SYMBOL_GPL(do_xdp_generic); 5173 5174 static int netif_rx_internal(struct sk_buff *skb) 5175 { 5176 int ret; 5177 5178 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5179 5180 trace_netif_rx(skb); 5181 5182 #ifdef CONFIG_RPS 5183 if (static_branch_unlikely(&rps_needed)) { 5184 struct rps_dev_flow voidflow, *rflow = &voidflow; 5185 int cpu; 5186 5187 rcu_read_lock(); 5188 5189 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5190 if (cpu < 0) 5191 cpu = smp_processor_id(); 5192 5193 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5194 5195 rcu_read_unlock(); 5196 } else 5197 #endif 5198 { 5199 unsigned int qtail; 5200 5201 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5202 } 5203 return ret; 5204 } 5205 5206 /** 5207 * __netif_rx - Slightly optimized version of netif_rx 5208 * @skb: buffer to post 5209 * 5210 * This behaves as netif_rx except that it does not disable bottom halves. 5211 * As a result this function may only be invoked from the interrupt context 5212 * (either hard or soft interrupt). 5213 */ 5214 int __netif_rx(struct sk_buff *skb) 5215 { 5216 int ret; 5217 5218 lockdep_assert_once(hardirq_count() | softirq_count()); 5219 5220 trace_netif_rx_entry(skb); 5221 ret = netif_rx_internal(skb); 5222 trace_netif_rx_exit(ret); 5223 return ret; 5224 } 5225 EXPORT_SYMBOL(__netif_rx); 5226 5227 /** 5228 * netif_rx - post buffer to the network code 5229 * @skb: buffer to post 5230 * 5231 * This function receives a packet from a device driver and queues it for 5232 * the upper (protocol) levels to process via the backlog NAPI device. It 5233 * always succeeds. The buffer may be dropped during processing for 5234 * congestion control or by the protocol layers. 5235 * The network buffer is passed via the backlog NAPI device. Modern NIC 5236 * driver should use NAPI and GRO. 5237 * This function can used from interrupt and from process context. The 5238 * caller from process context must not disable interrupts before invoking 5239 * this function. 5240 * 5241 * return values: 5242 * NET_RX_SUCCESS (no congestion) 5243 * NET_RX_DROP (packet was dropped) 5244 * 5245 */ 5246 int netif_rx(struct sk_buff *skb) 5247 { 5248 bool need_bh_off = !(hardirq_count() | softirq_count()); 5249 int ret; 5250 5251 if (need_bh_off) 5252 local_bh_disable(); 5253 trace_netif_rx_entry(skb); 5254 ret = netif_rx_internal(skb); 5255 trace_netif_rx_exit(ret); 5256 if (need_bh_off) 5257 local_bh_enable(); 5258 return ret; 5259 } 5260 EXPORT_SYMBOL(netif_rx); 5261 5262 static __latent_entropy void net_tx_action(void) 5263 { 5264 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5265 5266 if (sd->completion_queue) { 5267 struct sk_buff *clist; 5268 5269 local_irq_disable(); 5270 clist = sd->completion_queue; 5271 sd->completion_queue = NULL; 5272 local_irq_enable(); 5273 5274 while (clist) { 5275 struct sk_buff *skb = clist; 5276 5277 clist = clist->next; 5278 5279 WARN_ON(refcount_read(&skb->users)); 5280 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5281 trace_consume_skb(skb, net_tx_action); 5282 else 5283 trace_kfree_skb(skb, net_tx_action, 5284 get_kfree_skb_cb(skb)->reason, NULL); 5285 5286 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5287 __kfree_skb(skb); 5288 else 5289 __napi_kfree_skb(skb, 5290 get_kfree_skb_cb(skb)->reason); 5291 } 5292 } 5293 5294 if (sd->output_queue) { 5295 struct Qdisc *head; 5296 5297 local_irq_disable(); 5298 head = sd->output_queue; 5299 sd->output_queue = NULL; 5300 sd->output_queue_tailp = &sd->output_queue; 5301 local_irq_enable(); 5302 5303 rcu_read_lock(); 5304 5305 while (head) { 5306 struct Qdisc *q = head; 5307 spinlock_t *root_lock = NULL; 5308 5309 head = head->next_sched; 5310 5311 /* We need to make sure head->next_sched is read 5312 * before clearing __QDISC_STATE_SCHED 5313 */ 5314 smp_mb__before_atomic(); 5315 5316 if (!(q->flags & TCQ_F_NOLOCK)) { 5317 root_lock = qdisc_lock(q); 5318 spin_lock(root_lock); 5319 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5320 &q->state))) { 5321 /* There is a synchronize_net() between 5322 * STATE_DEACTIVATED flag being set and 5323 * qdisc_reset()/some_qdisc_is_busy() in 5324 * dev_deactivate(), so we can safely bail out 5325 * early here to avoid data race between 5326 * qdisc_deactivate() and some_qdisc_is_busy() 5327 * for lockless qdisc. 5328 */ 5329 clear_bit(__QDISC_STATE_SCHED, &q->state); 5330 continue; 5331 } 5332 5333 clear_bit(__QDISC_STATE_SCHED, &q->state); 5334 qdisc_run(q); 5335 if (root_lock) 5336 spin_unlock(root_lock); 5337 } 5338 5339 rcu_read_unlock(); 5340 } 5341 5342 xfrm_dev_backlog(sd); 5343 } 5344 5345 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5346 /* This hook is defined here for ATM LANE */ 5347 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5348 unsigned char *addr) __read_mostly; 5349 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5350 #endif 5351 5352 /** 5353 * netdev_is_rx_handler_busy - check if receive handler is registered 5354 * @dev: device to check 5355 * 5356 * Check if a receive handler is already registered for a given device. 5357 * Return true if there one. 5358 * 5359 * The caller must hold the rtnl_mutex. 5360 */ 5361 bool netdev_is_rx_handler_busy(struct net_device *dev) 5362 { 5363 ASSERT_RTNL(); 5364 return dev && rtnl_dereference(dev->rx_handler); 5365 } 5366 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5367 5368 /** 5369 * netdev_rx_handler_register - register receive handler 5370 * @dev: device to register a handler for 5371 * @rx_handler: receive handler to register 5372 * @rx_handler_data: data pointer that is used by rx handler 5373 * 5374 * Register a receive handler for a device. This handler will then be 5375 * called from __netif_receive_skb. A negative errno code is returned 5376 * on a failure. 5377 * 5378 * The caller must hold the rtnl_mutex. 5379 * 5380 * For a general description of rx_handler, see enum rx_handler_result. 5381 */ 5382 int netdev_rx_handler_register(struct net_device *dev, 5383 rx_handler_func_t *rx_handler, 5384 void *rx_handler_data) 5385 { 5386 if (netdev_is_rx_handler_busy(dev)) 5387 return -EBUSY; 5388 5389 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5390 return -EINVAL; 5391 5392 /* Note: rx_handler_data must be set before rx_handler */ 5393 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5394 rcu_assign_pointer(dev->rx_handler, rx_handler); 5395 5396 return 0; 5397 } 5398 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5399 5400 /** 5401 * netdev_rx_handler_unregister - unregister receive handler 5402 * @dev: device to unregister a handler from 5403 * 5404 * Unregister a receive handler from a device. 5405 * 5406 * The caller must hold the rtnl_mutex. 5407 */ 5408 void netdev_rx_handler_unregister(struct net_device *dev) 5409 { 5410 5411 ASSERT_RTNL(); 5412 RCU_INIT_POINTER(dev->rx_handler, NULL); 5413 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5414 * section has a guarantee to see a non NULL rx_handler_data 5415 * as well. 5416 */ 5417 synchronize_net(); 5418 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5419 } 5420 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5421 5422 /* 5423 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5424 * the special handling of PFMEMALLOC skbs. 5425 */ 5426 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5427 { 5428 switch (skb->protocol) { 5429 case htons(ETH_P_ARP): 5430 case htons(ETH_P_IP): 5431 case htons(ETH_P_IPV6): 5432 case htons(ETH_P_8021Q): 5433 case htons(ETH_P_8021AD): 5434 return true; 5435 default: 5436 return false; 5437 } 5438 } 5439 5440 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5441 int *ret, struct net_device *orig_dev) 5442 { 5443 if (nf_hook_ingress_active(skb)) { 5444 int ingress_retval; 5445 5446 if (*pt_prev) { 5447 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5448 *pt_prev = NULL; 5449 } 5450 5451 rcu_read_lock(); 5452 ingress_retval = nf_hook_ingress(skb); 5453 rcu_read_unlock(); 5454 return ingress_retval; 5455 } 5456 return 0; 5457 } 5458 5459 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5460 struct packet_type **ppt_prev) 5461 { 5462 struct packet_type *ptype, *pt_prev; 5463 rx_handler_func_t *rx_handler; 5464 struct sk_buff *skb = *pskb; 5465 struct net_device *orig_dev; 5466 bool deliver_exact = false; 5467 int ret = NET_RX_DROP; 5468 __be16 type; 5469 5470 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5471 5472 trace_netif_receive_skb(skb); 5473 5474 orig_dev = skb->dev; 5475 5476 skb_reset_network_header(skb); 5477 if (!skb_transport_header_was_set(skb)) 5478 skb_reset_transport_header(skb); 5479 skb_reset_mac_len(skb); 5480 5481 pt_prev = NULL; 5482 5483 another_round: 5484 skb->skb_iif = skb->dev->ifindex; 5485 5486 __this_cpu_inc(softnet_data.processed); 5487 5488 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5489 int ret2; 5490 5491 migrate_disable(); 5492 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5493 &skb); 5494 migrate_enable(); 5495 5496 if (ret2 != XDP_PASS) { 5497 ret = NET_RX_DROP; 5498 goto out; 5499 } 5500 } 5501 5502 if (eth_type_vlan(skb->protocol)) { 5503 skb = skb_vlan_untag(skb); 5504 if (unlikely(!skb)) 5505 goto out; 5506 } 5507 5508 if (skb_skip_tc_classify(skb)) 5509 goto skip_classify; 5510 5511 if (pfmemalloc) 5512 goto skip_taps; 5513 5514 list_for_each_entry_rcu(ptype, &net_hotdata.ptype_all, list) { 5515 if (pt_prev) 5516 ret = deliver_skb(skb, pt_prev, orig_dev); 5517 pt_prev = ptype; 5518 } 5519 5520 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5521 if (pt_prev) 5522 ret = deliver_skb(skb, pt_prev, orig_dev); 5523 pt_prev = ptype; 5524 } 5525 5526 skip_taps: 5527 #ifdef CONFIG_NET_INGRESS 5528 if (static_branch_unlikely(&ingress_needed_key)) { 5529 bool another = false; 5530 5531 nf_skip_egress(skb, true); 5532 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5533 &another); 5534 if (another) 5535 goto another_round; 5536 if (!skb) 5537 goto out; 5538 5539 nf_skip_egress(skb, false); 5540 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5541 goto out; 5542 } 5543 #endif 5544 skb_reset_redirect(skb); 5545 skip_classify: 5546 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5547 goto drop; 5548 5549 if (skb_vlan_tag_present(skb)) { 5550 if (pt_prev) { 5551 ret = deliver_skb(skb, pt_prev, orig_dev); 5552 pt_prev = NULL; 5553 } 5554 if (vlan_do_receive(&skb)) 5555 goto another_round; 5556 else if (unlikely(!skb)) 5557 goto out; 5558 } 5559 5560 rx_handler = rcu_dereference(skb->dev->rx_handler); 5561 if (rx_handler) { 5562 if (pt_prev) { 5563 ret = deliver_skb(skb, pt_prev, orig_dev); 5564 pt_prev = NULL; 5565 } 5566 switch (rx_handler(&skb)) { 5567 case RX_HANDLER_CONSUMED: 5568 ret = NET_RX_SUCCESS; 5569 goto out; 5570 case RX_HANDLER_ANOTHER: 5571 goto another_round; 5572 case RX_HANDLER_EXACT: 5573 deliver_exact = true; 5574 break; 5575 case RX_HANDLER_PASS: 5576 break; 5577 default: 5578 BUG(); 5579 } 5580 } 5581 5582 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5583 check_vlan_id: 5584 if (skb_vlan_tag_get_id(skb)) { 5585 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5586 * find vlan device. 5587 */ 5588 skb->pkt_type = PACKET_OTHERHOST; 5589 } else if (eth_type_vlan(skb->protocol)) { 5590 /* Outer header is 802.1P with vlan 0, inner header is 5591 * 802.1Q or 802.1AD and vlan_do_receive() above could 5592 * not find vlan dev for vlan id 0. 5593 */ 5594 __vlan_hwaccel_clear_tag(skb); 5595 skb = skb_vlan_untag(skb); 5596 if (unlikely(!skb)) 5597 goto out; 5598 if (vlan_do_receive(&skb)) 5599 /* After stripping off 802.1P header with vlan 0 5600 * vlan dev is found for inner header. 5601 */ 5602 goto another_round; 5603 else if (unlikely(!skb)) 5604 goto out; 5605 else 5606 /* We have stripped outer 802.1P vlan 0 header. 5607 * But could not find vlan dev. 5608 * check again for vlan id to set OTHERHOST. 5609 */ 5610 goto check_vlan_id; 5611 } 5612 /* Note: we might in the future use prio bits 5613 * and set skb->priority like in vlan_do_receive() 5614 * For the time being, just ignore Priority Code Point 5615 */ 5616 __vlan_hwaccel_clear_tag(skb); 5617 } 5618 5619 type = skb->protocol; 5620 5621 /* deliver only exact match when indicated */ 5622 if (likely(!deliver_exact)) { 5623 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5624 &ptype_base[ntohs(type) & 5625 PTYPE_HASH_MASK]); 5626 } 5627 5628 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5629 &orig_dev->ptype_specific); 5630 5631 if (unlikely(skb->dev != orig_dev)) { 5632 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5633 &skb->dev->ptype_specific); 5634 } 5635 5636 if (pt_prev) { 5637 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5638 goto drop; 5639 *ppt_prev = pt_prev; 5640 } else { 5641 drop: 5642 if (!deliver_exact) 5643 dev_core_stats_rx_dropped_inc(skb->dev); 5644 else 5645 dev_core_stats_rx_nohandler_inc(skb->dev); 5646 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5647 /* Jamal, now you will not able to escape explaining 5648 * me how you were going to use this. :-) 5649 */ 5650 ret = NET_RX_DROP; 5651 } 5652 5653 out: 5654 /* The invariant here is that if *ppt_prev is not NULL 5655 * then skb should also be non-NULL. 5656 * 5657 * Apparently *ppt_prev assignment above holds this invariant due to 5658 * skb dereferencing near it. 5659 */ 5660 *pskb = skb; 5661 return ret; 5662 } 5663 5664 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5665 { 5666 struct net_device *orig_dev = skb->dev; 5667 struct packet_type *pt_prev = NULL; 5668 int ret; 5669 5670 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5671 if (pt_prev) 5672 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5673 skb->dev, pt_prev, orig_dev); 5674 return ret; 5675 } 5676 5677 /** 5678 * netif_receive_skb_core - special purpose version of netif_receive_skb 5679 * @skb: buffer to process 5680 * 5681 * More direct receive version of netif_receive_skb(). It should 5682 * only be used by callers that have a need to skip RPS and Generic XDP. 5683 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5684 * 5685 * This function may only be called from softirq context and interrupts 5686 * should be enabled. 5687 * 5688 * Return values (usually ignored): 5689 * NET_RX_SUCCESS: no congestion 5690 * NET_RX_DROP: packet was dropped 5691 */ 5692 int netif_receive_skb_core(struct sk_buff *skb) 5693 { 5694 int ret; 5695 5696 rcu_read_lock(); 5697 ret = __netif_receive_skb_one_core(skb, false); 5698 rcu_read_unlock(); 5699 5700 return ret; 5701 } 5702 EXPORT_SYMBOL(netif_receive_skb_core); 5703 5704 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5705 struct packet_type *pt_prev, 5706 struct net_device *orig_dev) 5707 { 5708 struct sk_buff *skb, *next; 5709 5710 if (!pt_prev) 5711 return; 5712 if (list_empty(head)) 5713 return; 5714 if (pt_prev->list_func != NULL) 5715 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5716 ip_list_rcv, head, pt_prev, orig_dev); 5717 else 5718 list_for_each_entry_safe(skb, next, head, list) { 5719 skb_list_del_init(skb); 5720 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5721 } 5722 } 5723 5724 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5725 { 5726 /* Fast-path assumptions: 5727 * - There is no RX handler. 5728 * - Only one packet_type matches. 5729 * If either of these fails, we will end up doing some per-packet 5730 * processing in-line, then handling the 'last ptype' for the whole 5731 * sublist. This can't cause out-of-order delivery to any single ptype, 5732 * because the 'last ptype' must be constant across the sublist, and all 5733 * other ptypes are handled per-packet. 5734 */ 5735 /* Current (common) ptype of sublist */ 5736 struct packet_type *pt_curr = NULL; 5737 /* Current (common) orig_dev of sublist */ 5738 struct net_device *od_curr = NULL; 5739 struct sk_buff *skb, *next; 5740 LIST_HEAD(sublist); 5741 5742 list_for_each_entry_safe(skb, next, head, list) { 5743 struct net_device *orig_dev = skb->dev; 5744 struct packet_type *pt_prev = NULL; 5745 5746 skb_list_del_init(skb); 5747 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5748 if (!pt_prev) 5749 continue; 5750 if (pt_curr != pt_prev || od_curr != orig_dev) { 5751 /* dispatch old sublist */ 5752 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5753 /* start new sublist */ 5754 INIT_LIST_HEAD(&sublist); 5755 pt_curr = pt_prev; 5756 od_curr = orig_dev; 5757 } 5758 list_add_tail(&skb->list, &sublist); 5759 } 5760 5761 /* dispatch final sublist */ 5762 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5763 } 5764 5765 static int __netif_receive_skb(struct sk_buff *skb) 5766 { 5767 int ret; 5768 5769 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 5770 unsigned int noreclaim_flag; 5771 5772 /* 5773 * PFMEMALLOC skbs are special, they should 5774 * - be delivered to SOCK_MEMALLOC sockets only 5775 * - stay away from userspace 5776 * - have bounded memory usage 5777 * 5778 * Use PF_MEMALLOC as this saves us from propagating the allocation 5779 * context down to all allocation sites. 5780 */ 5781 noreclaim_flag = memalloc_noreclaim_save(); 5782 ret = __netif_receive_skb_one_core(skb, true); 5783 memalloc_noreclaim_restore(noreclaim_flag); 5784 } else 5785 ret = __netif_receive_skb_one_core(skb, false); 5786 5787 return ret; 5788 } 5789 5790 static void __netif_receive_skb_list(struct list_head *head) 5791 { 5792 unsigned long noreclaim_flag = 0; 5793 struct sk_buff *skb, *next; 5794 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 5795 5796 list_for_each_entry_safe(skb, next, head, list) { 5797 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 5798 struct list_head sublist; 5799 5800 /* Handle the previous sublist */ 5801 list_cut_before(&sublist, head, &skb->list); 5802 if (!list_empty(&sublist)) 5803 __netif_receive_skb_list_core(&sublist, pfmemalloc); 5804 pfmemalloc = !pfmemalloc; 5805 /* See comments in __netif_receive_skb */ 5806 if (pfmemalloc) 5807 noreclaim_flag = memalloc_noreclaim_save(); 5808 else 5809 memalloc_noreclaim_restore(noreclaim_flag); 5810 } 5811 } 5812 /* Handle the remaining sublist */ 5813 if (!list_empty(head)) 5814 __netif_receive_skb_list_core(head, pfmemalloc); 5815 /* Restore pflags */ 5816 if (pfmemalloc) 5817 memalloc_noreclaim_restore(noreclaim_flag); 5818 } 5819 5820 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 5821 { 5822 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 5823 struct bpf_prog *new = xdp->prog; 5824 int ret = 0; 5825 5826 switch (xdp->command) { 5827 case XDP_SETUP_PROG: 5828 rcu_assign_pointer(dev->xdp_prog, new); 5829 if (old) 5830 bpf_prog_put(old); 5831 5832 if (old && !new) { 5833 static_branch_dec(&generic_xdp_needed_key); 5834 } else if (new && !old) { 5835 static_branch_inc(&generic_xdp_needed_key); 5836 dev_disable_lro(dev); 5837 dev_disable_gro_hw(dev); 5838 } 5839 break; 5840 5841 default: 5842 ret = -EINVAL; 5843 break; 5844 } 5845 5846 return ret; 5847 } 5848 5849 static int netif_receive_skb_internal(struct sk_buff *skb) 5850 { 5851 int ret; 5852 5853 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5854 5855 if (skb_defer_rx_timestamp(skb)) 5856 return NET_RX_SUCCESS; 5857 5858 rcu_read_lock(); 5859 #ifdef CONFIG_RPS 5860 if (static_branch_unlikely(&rps_needed)) { 5861 struct rps_dev_flow voidflow, *rflow = &voidflow; 5862 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 5863 5864 if (cpu >= 0) { 5865 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5866 rcu_read_unlock(); 5867 return ret; 5868 } 5869 } 5870 #endif 5871 ret = __netif_receive_skb(skb); 5872 rcu_read_unlock(); 5873 return ret; 5874 } 5875 5876 void netif_receive_skb_list_internal(struct list_head *head) 5877 { 5878 struct sk_buff *skb, *next; 5879 LIST_HEAD(sublist); 5880 5881 list_for_each_entry_safe(skb, next, head, list) { 5882 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 5883 skb); 5884 skb_list_del_init(skb); 5885 if (!skb_defer_rx_timestamp(skb)) 5886 list_add_tail(&skb->list, &sublist); 5887 } 5888 list_splice_init(&sublist, head); 5889 5890 rcu_read_lock(); 5891 #ifdef CONFIG_RPS 5892 if (static_branch_unlikely(&rps_needed)) { 5893 list_for_each_entry_safe(skb, next, head, list) { 5894 struct rps_dev_flow voidflow, *rflow = &voidflow; 5895 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 5896 5897 if (cpu >= 0) { 5898 /* Will be handled, remove from list */ 5899 skb_list_del_init(skb); 5900 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5901 } 5902 } 5903 } 5904 #endif 5905 __netif_receive_skb_list(head); 5906 rcu_read_unlock(); 5907 } 5908 5909 /** 5910 * netif_receive_skb - process receive buffer from network 5911 * @skb: buffer to process 5912 * 5913 * netif_receive_skb() is the main receive data processing function. 5914 * It always succeeds. The buffer may be dropped during processing 5915 * for congestion control or by the protocol layers. 5916 * 5917 * This function may only be called from softirq context and interrupts 5918 * should be enabled. 5919 * 5920 * Return values (usually ignored): 5921 * NET_RX_SUCCESS: no congestion 5922 * NET_RX_DROP: packet was dropped 5923 */ 5924 int netif_receive_skb(struct sk_buff *skb) 5925 { 5926 int ret; 5927 5928 trace_netif_receive_skb_entry(skb); 5929 5930 ret = netif_receive_skb_internal(skb); 5931 trace_netif_receive_skb_exit(ret); 5932 5933 return ret; 5934 } 5935 EXPORT_SYMBOL(netif_receive_skb); 5936 5937 /** 5938 * netif_receive_skb_list - process many receive buffers from network 5939 * @head: list of skbs to process. 5940 * 5941 * Since return value of netif_receive_skb() is normally ignored, and 5942 * wouldn't be meaningful for a list, this function returns void. 5943 * 5944 * This function may only be called from softirq context and interrupts 5945 * should be enabled. 5946 */ 5947 void netif_receive_skb_list(struct list_head *head) 5948 { 5949 struct sk_buff *skb; 5950 5951 if (list_empty(head)) 5952 return; 5953 if (trace_netif_receive_skb_list_entry_enabled()) { 5954 list_for_each_entry(skb, head, list) 5955 trace_netif_receive_skb_list_entry(skb); 5956 } 5957 netif_receive_skb_list_internal(head); 5958 trace_netif_receive_skb_list_exit(0); 5959 } 5960 EXPORT_SYMBOL(netif_receive_skb_list); 5961 5962 static DEFINE_PER_CPU(struct work_struct, flush_works); 5963 5964 /* Network device is going away, flush any packets still pending */ 5965 static void flush_backlog(struct work_struct *work) 5966 { 5967 struct sk_buff *skb, *tmp; 5968 struct softnet_data *sd; 5969 5970 local_bh_disable(); 5971 sd = this_cpu_ptr(&softnet_data); 5972 5973 backlog_lock_irq_disable(sd); 5974 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 5975 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 5976 __skb_unlink(skb, &sd->input_pkt_queue); 5977 dev_kfree_skb_irq(skb); 5978 rps_input_queue_head_incr(sd); 5979 } 5980 } 5981 backlog_unlock_irq_enable(sd); 5982 5983 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 5984 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 5985 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 5986 __skb_unlink(skb, &sd->process_queue); 5987 kfree_skb(skb); 5988 rps_input_queue_head_incr(sd); 5989 } 5990 } 5991 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 5992 local_bh_enable(); 5993 } 5994 5995 static bool flush_required(int cpu) 5996 { 5997 #if IS_ENABLED(CONFIG_RPS) 5998 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 5999 bool do_flush; 6000 6001 backlog_lock_irq_disable(sd); 6002 6003 /* as insertion into process_queue happens with the rps lock held, 6004 * process_queue access may race only with dequeue 6005 */ 6006 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6007 !skb_queue_empty_lockless(&sd->process_queue); 6008 backlog_unlock_irq_enable(sd); 6009 6010 return do_flush; 6011 #endif 6012 /* without RPS we can't safely check input_pkt_queue: during a 6013 * concurrent remote skb_queue_splice() we can detect as empty both 6014 * input_pkt_queue and process_queue even if the latter could end-up 6015 * containing a lot of packets. 6016 */ 6017 return true; 6018 } 6019 6020 static void flush_all_backlogs(void) 6021 { 6022 static cpumask_t flush_cpus; 6023 unsigned int cpu; 6024 6025 /* since we are under rtnl lock protection we can use static data 6026 * for the cpumask and avoid allocating on stack the possibly 6027 * large mask 6028 */ 6029 ASSERT_RTNL(); 6030 6031 cpus_read_lock(); 6032 6033 cpumask_clear(&flush_cpus); 6034 for_each_online_cpu(cpu) { 6035 if (flush_required(cpu)) { 6036 queue_work_on(cpu, system_highpri_wq, 6037 per_cpu_ptr(&flush_works, cpu)); 6038 cpumask_set_cpu(cpu, &flush_cpus); 6039 } 6040 } 6041 6042 /* we can have in flight packet[s] on the cpus we are not flushing, 6043 * synchronize_net() in unregister_netdevice_many() will take care of 6044 * them 6045 */ 6046 for_each_cpu(cpu, &flush_cpus) 6047 flush_work(per_cpu_ptr(&flush_works, cpu)); 6048 6049 cpus_read_unlock(); 6050 } 6051 6052 static void net_rps_send_ipi(struct softnet_data *remsd) 6053 { 6054 #ifdef CONFIG_RPS 6055 while (remsd) { 6056 struct softnet_data *next = remsd->rps_ipi_next; 6057 6058 if (cpu_online(remsd->cpu)) 6059 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6060 remsd = next; 6061 } 6062 #endif 6063 } 6064 6065 /* 6066 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6067 * Note: called with local irq disabled, but exits with local irq enabled. 6068 */ 6069 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6070 { 6071 #ifdef CONFIG_RPS 6072 struct softnet_data *remsd = sd->rps_ipi_list; 6073 6074 if (!use_backlog_threads() && remsd) { 6075 sd->rps_ipi_list = NULL; 6076 6077 local_irq_enable(); 6078 6079 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6080 net_rps_send_ipi(remsd); 6081 } else 6082 #endif 6083 local_irq_enable(); 6084 } 6085 6086 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6087 { 6088 #ifdef CONFIG_RPS 6089 return !use_backlog_threads() && sd->rps_ipi_list; 6090 #else 6091 return false; 6092 #endif 6093 } 6094 6095 static int process_backlog(struct napi_struct *napi, int quota) 6096 { 6097 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6098 bool again = true; 6099 int work = 0; 6100 6101 /* Check if we have pending ipi, its better to send them now, 6102 * not waiting net_rx_action() end. 6103 */ 6104 if (sd_has_rps_ipi_waiting(sd)) { 6105 local_irq_disable(); 6106 net_rps_action_and_irq_enable(sd); 6107 } 6108 6109 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6110 while (again) { 6111 struct sk_buff *skb; 6112 6113 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6114 while ((skb = __skb_dequeue(&sd->process_queue))) { 6115 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6116 rcu_read_lock(); 6117 __netif_receive_skb(skb); 6118 rcu_read_unlock(); 6119 if (++work >= quota) { 6120 rps_input_queue_head_add(sd, work); 6121 return work; 6122 } 6123 6124 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6125 } 6126 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6127 6128 backlog_lock_irq_disable(sd); 6129 if (skb_queue_empty(&sd->input_pkt_queue)) { 6130 /* 6131 * Inline a custom version of __napi_complete(). 6132 * only current cpu owns and manipulates this napi, 6133 * and NAPI_STATE_SCHED is the only possible flag set 6134 * on backlog. 6135 * We can use a plain write instead of clear_bit(), 6136 * and we dont need an smp_mb() memory barrier. 6137 */ 6138 napi->state &= NAPIF_STATE_THREADED; 6139 again = false; 6140 } else { 6141 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6142 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6143 &sd->process_queue); 6144 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6145 } 6146 backlog_unlock_irq_enable(sd); 6147 } 6148 6149 if (work) 6150 rps_input_queue_head_add(sd, work); 6151 return work; 6152 } 6153 6154 /** 6155 * __napi_schedule - schedule for receive 6156 * @n: entry to schedule 6157 * 6158 * The entry's receive function will be scheduled to run. 6159 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6160 */ 6161 void __napi_schedule(struct napi_struct *n) 6162 { 6163 unsigned long flags; 6164 6165 local_irq_save(flags); 6166 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6167 local_irq_restore(flags); 6168 } 6169 EXPORT_SYMBOL(__napi_schedule); 6170 6171 /** 6172 * napi_schedule_prep - check if napi can be scheduled 6173 * @n: napi context 6174 * 6175 * Test if NAPI routine is already running, and if not mark 6176 * it as running. This is used as a condition variable to 6177 * insure only one NAPI poll instance runs. We also make 6178 * sure there is no pending NAPI disable. 6179 */ 6180 bool napi_schedule_prep(struct napi_struct *n) 6181 { 6182 unsigned long new, val = READ_ONCE(n->state); 6183 6184 do { 6185 if (unlikely(val & NAPIF_STATE_DISABLE)) 6186 return false; 6187 new = val | NAPIF_STATE_SCHED; 6188 6189 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6190 * This was suggested by Alexander Duyck, as compiler 6191 * emits better code than : 6192 * if (val & NAPIF_STATE_SCHED) 6193 * new |= NAPIF_STATE_MISSED; 6194 */ 6195 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6196 NAPIF_STATE_MISSED; 6197 } while (!try_cmpxchg(&n->state, &val, new)); 6198 6199 return !(val & NAPIF_STATE_SCHED); 6200 } 6201 EXPORT_SYMBOL(napi_schedule_prep); 6202 6203 /** 6204 * __napi_schedule_irqoff - schedule for receive 6205 * @n: entry to schedule 6206 * 6207 * Variant of __napi_schedule() assuming hard irqs are masked. 6208 * 6209 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6210 * because the interrupt disabled assumption might not be true 6211 * due to force-threaded interrupts and spinlock substitution. 6212 */ 6213 void __napi_schedule_irqoff(struct napi_struct *n) 6214 { 6215 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6216 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6217 else 6218 __napi_schedule(n); 6219 } 6220 EXPORT_SYMBOL(__napi_schedule_irqoff); 6221 6222 bool napi_complete_done(struct napi_struct *n, int work_done) 6223 { 6224 unsigned long flags, val, new, timeout = 0; 6225 bool ret = true; 6226 6227 /* 6228 * 1) Don't let napi dequeue from the cpu poll list 6229 * just in case its running on a different cpu. 6230 * 2) If we are busy polling, do nothing here, we have 6231 * the guarantee we will be called later. 6232 */ 6233 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6234 NAPIF_STATE_IN_BUSY_POLL))) 6235 return false; 6236 6237 if (work_done) { 6238 if (n->gro_bitmask) 6239 timeout = napi_get_gro_flush_timeout(n); 6240 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6241 } 6242 if (n->defer_hard_irqs_count > 0) { 6243 n->defer_hard_irqs_count--; 6244 timeout = napi_get_gro_flush_timeout(n); 6245 if (timeout) 6246 ret = false; 6247 } 6248 if (n->gro_bitmask) { 6249 /* When the NAPI instance uses a timeout and keeps postponing 6250 * it, we need to bound somehow the time packets are kept in 6251 * the GRO layer 6252 */ 6253 napi_gro_flush(n, !!timeout); 6254 } 6255 6256 gro_normal_list(n); 6257 6258 if (unlikely(!list_empty(&n->poll_list))) { 6259 /* If n->poll_list is not empty, we need to mask irqs */ 6260 local_irq_save(flags); 6261 list_del_init(&n->poll_list); 6262 local_irq_restore(flags); 6263 } 6264 WRITE_ONCE(n->list_owner, -1); 6265 6266 val = READ_ONCE(n->state); 6267 do { 6268 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6269 6270 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6271 NAPIF_STATE_SCHED_THREADED | 6272 NAPIF_STATE_PREFER_BUSY_POLL); 6273 6274 /* If STATE_MISSED was set, leave STATE_SCHED set, 6275 * because we will call napi->poll() one more time. 6276 * This C code was suggested by Alexander Duyck to help gcc. 6277 */ 6278 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6279 NAPIF_STATE_SCHED; 6280 } while (!try_cmpxchg(&n->state, &val, new)); 6281 6282 if (unlikely(val & NAPIF_STATE_MISSED)) { 6283 __napi_schedule(n); 6284 return false; 6285 } 6286 6287 if (timeout) 6288 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6289 HRTIMER_MODE_REL_PINNED); 6290 return ret; 6291 } 6292 EXPORT_SYMBOL(napi_complete_done); 6293 6294 /* must be called under rcu_read_lock(), as we dont take a reference */ 6295 struct napi_struct *napi_by_id(unsigned int napi_id) 6296 { 6297 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 6298 struct napi_struct *napi; 6299 6300 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 6301 if (napi->napi_id == napi_id) 6302 return napi; 6303 6304 return NULL; 6305 } 6306 6307 static void skb_defer_free_flush(struct softnet_data *sd) 6308 { 6309 struct sk_buff *skb, *next; 6310 6311 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6312 if (!READ_ONCE(sd->defer_list)) 6313 return; 6314 6315 spin_lock(&sd->defer_lock); 6316 skb = sd->defer_list; 6317 sd->defer_list = NULL; 6318 sd->defer_count = 0; 6319 spin_unlock(&sd->defer_lock); 6320 6321 while (skb != NULL) { 6322 next = skb->next; 6323 napi_consume_skb(skb, 1); 6324 skb = next; 6325 } 6326 } 6327 6328 #if defined(CONFIG_NET_RX_BUSY_POLL) 6329 6330 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6331 { 6332 if (!skip_schedule) { 6333 gro_normal_list(napi); 6334 __napi_schedule(napi); 6335 return; 6336 } 6337 6338 if (napi->gro_bitmask) { 6339 /* flush too old packets 6340 * If HZ < 1000, flush all packets. 6341 */ 6342 napi_gro_flush(napi, HZ >= 1000); 6343 } 6344 6345 gro_normal_list(napi); 6346 clear_bit(NAPI_STATE_SCHED, &napi->state); 6347 } 6348 6349 enum { 6350 NAPI_F_PREFER_BUSY_POLL = 1, 6351 NAPI_F_END_ON_RESCHED = 2, 6352 }; 6353 6354 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6355 unsigned flags, u16 budget) 6356 { 6357 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6358 bool skip_schedule = false; 6359 unsigned long timeout; 6360 int rc; 6361 6362 /* Busy polling means there is a high chance device driver hard irq 6363 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6364 * set in napi_schedule_prep(). 6365 * Since we are about to call napi->poll() once more, we can safely 6366 * clear NAPI_STATE_MISSED. 6367 * 6368 * Note: x86 could use a single "lock and ..." instruction 6369 * to perform these two clear_bit() 6370 */ 6371 clear_bit(NAPI_STATE_MISSED, &napi->state); 6372 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6373 6374 local_bh_disable(); 6375 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6376 6377 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6378 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6379 timeout = napi_get_gro_flush_timeout(napi); 6380 if (napi->defer_hard_irqs_count && timeout) { 6381 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6382 skip_schedule = true; 6383 } 6384 } 6385 6386 /* All we really want here is to re-enable device interrupts. 6387 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6388 */ 6389 rc = napi->poll(napi, budget); 6390 /* We can't gro_normal_list() here, because napi->poll() might have 6391 * rearmed the napi (napi_complete_done()) in which case it could 6392 * already be running on another CPU. 6393 */ 6394 trace_napi_poll(napi, rc, budget); 6395 netpoll_poll_unlock(have_poll_lock); 6396 if (rc == budget) 6397 __busy_poll_stop(napi, skip_schedule); 6398 bpf_net_ctx_clear(bpf_net_ctx); 6399 local_bh_enable(); 6400 } 6401 6402 static void __napi_busy_loop(unsigned int napi_id, 6403 bool (*loop_end)(void *, unsigned long), 6404 void *loop_end_arg, unsigned flags, u16 budget) 6405 { 6406 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6407 int (*napi_poll)(struct napi_struct *napi, int budget); 6408 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6409 void *have_poll_lock = NULL; 6410 struct napi_struct *napi; 6411 6412 WARN_ON_ONCE(!rcu_read_lock_held()); 6413 6414 restart: 6415 napi_poll = NULL; 6416 6417 napi = napi_by_id(napi_id); 6418 if (!napi) 6419 return; 6420 6421 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6422 preempt_disable(); 6423 for (;;) { 6424 int work = 0; 6425 6426 local_bh_disable(); 6427 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6428 if (!napi_poll) { 6429 unsigned long val = READ_ONCE(napi->state); 6430 6431 /* If multiple threads are competing for this napi, 6432 * we avoid dirtying napi->state as much as we can. 6433 */ 6434 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6435 NAPIF_STATE_IN_BUSY_POLL)) { 6436 if (flags & NAPI_F_PREFER_BUSY_POLL) 6437 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6438 goto count; 6439 } 6440 if (cmpxchg(&napi->state, val, 6441 val | NAPIF_STATE_IN_BUSY_POLL | 6442 NAPIF_STATE_SCHED) != val) { 6443 if (flags & NAPI_F_PREFER_BUSY_POLL) 6444 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6445 goto count; 6446 } 6447 have_poll_lock = netpoll_poll_lock(napi); 6448 napi_poll = napi->poll; 6449 } 6450 work = napi_poll(napi, budget); 6451 trace_napi_poll(napi, work, budget); 6452 gro_normal_list(napi); 6453 count: 6454 if (work > 0) 6455 __NET_ADD_STATS(dev_net(napi->dev), 6456 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6457 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6458 bpf_net_ctx_clear(bpf_net_ctx); 6459 local_bh_enable(); 6460 6461 if (!loop_end || loop_end(loop_end_arg, start_time)) 6462 break; 6463 6464 if (unlikely(need_resched())) { 6465 if (flags & NAPI_F_END_ON_RESCHED) 6466 break; 6467 if (napi_poll) 6468 busy_poll_stop(napi, have_poll_lock, flags, budget); 6469 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6470 preempt_enable(); 6471 rcu_read_unlock(); 6472 cond_resched(); 6473 rcu_read_lock(); 6474 if (loop_end(loop_end_arg, start_time)) 6475 return; 6476 goto restart; 6477 } 6478 cpu_relax(); 6479 } 6480 if (napi_poll) 6481 busy_poll_stop(napi, have_poll_lock, flags, budget); 6482 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6483 preempt_enable(); 6484 } 6485 6486 void napi_busy_loop_rcu(unsigned int napi_id, 6487 bool (*loop_end)(void *, unsigned long), 6488 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6489 { 6490 unsigned flags = NAPI_F_END_ON_RESCHED; 6491 6492 if (prefer_busy_poll) 6493 flags |= NAPI_F_PREFER_BUSY_POLL; 6494 6495 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6496 } 6497 6498 void napi_busy_loop(unsigned int napi_id, 6499 bool (*loop_end)(void *, unsigned long), 6500 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6501 { 6502 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6503 6504 rcu_read_lock(); 6505 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6506 rcu_read_unlock(); 6507 } 6508 EXPORT_SYMBOL(napi_busy_loop); 6509 6510 void napi_suspend_irqs(unsigned int napi_id) 6511 { 6512 struct napi_struct *napi; 6513 6514 rcu_read_lock(); 6515 napi = napi_by_id(napi_id); 6516 if (napi) { 6517 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 6518 6519 if (timeout) 6520 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 6521 HRTIMER_MODE_REL_PINNED); 6522 } 6523 rcu_read_unlock(); 6524 } 6525 6526 void napi_resume_irqs(unsigned int napi_id) 6527 { 6528 struct napi_struct *napi; 6529 6530 rcu_read_lock(); 6531 napi = napi_by_id(napi_id); 6532 if (napi) { 6533 /* If irq_suspend_timeout is set to 0 between the call to 6534 * napi_suspend_irqs and now, the original value still 6535 * determines the safety timeout as intended and napi_watchdog 6536 * will resume irq processing. 6537 */ 6538 if (napi_get_irq_suspend_timeout(napi)) { 6539 local_bh_disable(); 6540 napi_schedule(napi); 6541 local_bh_enable(); 6542 } 6543 } 6544 rcu_read_unlock(); 6545 } 6546 6547 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6548 6549 static void __napi_hash_add_with_id(struct napi_struct *napi, 6550 unsigned int napi_id) 6551 { 6552 napi->napi_id = napi_id; 6553 hlist_add_head_rcu(&napi->napi_hash_node, 6554 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6555 } 6556 6557 static void napi_hash_add_with_id(struct napi_struct *napi, 6558 unsigned int napi_id) 6559 { 6560 spin_lock(&napi_hash_lock); 6561 WARN_ON_ONCE(napi_by_id(napi_id)); 6562 __napi_hash_add_with_id(napi, napi_id); 6563 spin_unlock(&napi_hash_lock); 6564 } 6565 6566 static void napi_hash_add(struct napi_struct *napi) 6567 { 6568 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6569 return; 6570 6571 spin_lock(&napi_hash_lock); 6572 6573 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6574 do { 6575 if (unlikely(++napi_gen_id < MIN_NAPI_ID)) 6576 napi_gen_id = MIN_NAPI_ID; 6577 } while (napi_by_id(napi_gen_id)); 6578 6579 __napi_hash_add_with_id(napi, napi_gen_id); 6580 6581 spin_unlock(&napi_hash_lock); 6582 } 6583 6584 /* Warning : caller is responsible to make sure rcu grace period 6585 * is respected before freeing memory containing @napi 6586 */ 6587 static void napi_hash_del(struct napi_struct *napi) 6588 { 6589 spin_lock(&napi_hash_lock); 6590 6591 hlist_del_init_rcu(&napi->napi_hash_node); 6592 6593 spin_unlock(&napi_hash_lock); 6594 } 6595 6596 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6597 { 6598 struct napi_struct *napi; 6599 6600 napi = container_of(timer, struct napi_struct, timer); 6601 6602 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6603 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6604 */ 6605 if (!napi_disable_pending(napi) && 6606 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6607 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6608 __napi_schedule_irqoff(napi); 6609 } 6610 6611 return HRTIMER_NORESTART; 6612 } 6613 6614 static void init_gro_hash(struct napi_struct *napi) 6615 { 6616 int i; 6617 6618 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6619 INIT_LIST_HEAD(&napi->gro_hash[i].list); 6620 napi->gro_hash[i].count = 0; 6621 } 6622 napi->gro_bitmask = 0; 6623 } 6624 6625 int dev_set_threaded(struct net_device *dev, bool threaded) 6626 { 6627 struct napi_struct *napi; 6628 int err = 0; 6629 6630 if (dev->threaded == threaded) 6631 return 0; 6632 6633 if (threaded) { 6634 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6635 if (!napi->thread) { 6636 err = napi_kthread_create(napi); 6637 if (err) { 6638 threaded = false; 6639 break; 6640 } 6641 } 6642 } 6643 } 6644 6645 WRITE_ONCE(dev->threaded, threaded); 6646 6647 /* Make sure kthread is created before THREADED bit 6648 * is set. 6649 */ 6650 smp_mb__before_atomic(); 6651 6652 /* Setting/unsetting threaded mode on a napi might not immediately 6653 * take effect, if the current napi instance is actively being 6654 * polled. In this case, the switch between threaded mode and 6655 * softirq mode will happen in the next round of napi_schedule(). 6656 * This should not cause hiccups/stalls to the live traffic. 6657 */ 6658 list_for_each_entry(napi, &dev->napi_list, dev_list) 6659 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6660 6661 return err; 6662 } 6663 EXPORT_SYMBOL(dev_set_threaded); 6664 6665 /** 6666 * netif_queue_set_napi - Associate queue with the napi 6667 * @dev: device to which NAPI and queue belong 6668 * @queue_index: Index of queue 6669 * @type: queue type as RX or TX 6670 * @napi: NAPI context, pass NULL to clear previously set NAPI 6671 * 6672 * Set queue with its corresponding napi context. This should be done after 6673 * registering the NAPI handler for the queue-vector and the queues have been 6674 * mapped to the corresponding interrupt vector. 6675 */ 6676 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6677 enum netdev_queue_type type, struct napi_struct *napi) 6678 { 6679 struct netdev_rx_queue *rxq; 6680 struct netdev_queue *txq; 6681 6682 if (WARN_ON_ONCE(napi && !napi->dev)) 6683 return; 6684 if (dev->reg_state >= NETREG_REGISTERED) 6685 ASSERT_RTNL(); 6686 6687 switch (type) { 6688 case NETDEV_QUEUE_TYPE_RX: 6689 rxq = __netif_get_rx_queue(dev, queue_index); 6690 rxq->napi = napi; 6691 return; 6692 case NETDEV_QUEUE_TYPE_TX: 6693 txq = netdev_get_tx_queue(dev, queue_index); 6694 txq->napi = napi; 6695 return; 6696 default: 6697 return; 6698 } 6699 } 6700 EXPORT_SYMBOL(netif_queue_set_napi); 6701 6702 static void napi_restore_config(struct napi_struct *n) 6703 { 6704 n->defer_hard_irqs = n->config->defer_hard_irqs; 6705 n->gro_flush_timeout = n->config->gro_flush_timeout; 6706 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 6707 /* a NAPI ID might be stored in the config, if so use it. if not, use 6708 * napi_hash_add to generate one for us. It will be saved to the config 6709 * in napi_disable. 6710 */ 6711 if (n->config->napi_id) 6712 napi_hash_add_with_id(n, n->config->napi_id); 6713 else 6714 napi_hash_add(n); 6715 } 6716 6717 static void napi_save_config(struct napi_struct *n) 6718 { 6719 n->config->defer_hard_irqs = n->defer_hard_irqs; 6720 n->config->gro_flush_timeout = n->gro_flush_timeout; 6721 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 6722 n->config->napi_id = n->napi_id; 6723 napi_hash_del(n); 6724 } 6725 6726 void netif_napi_add_weight(struct net_device *dev, struct napi_struct *napi, 6727 int (*poll)(struct napi_struct *, int), int weight) 6728 { 6729 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 6730 return; 6731 6732 INIT_LIST_HEAD(&napi->poll_list); 6733 INIT_HLIST_NODE(&napi->napi_hash_node); 6734 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 6735 napi->timer.function = napi_watchdog; 6736 init_gro_hash(napi); 6737 napi->skb = NULL; 6738 INIT_LIST_HEAD(&napi->rx_list); 6739 napi->rx_count = 0; 6740 napi->poll = poll; 6741 if (weight > NAPI_POLL_WEIGHT) 6742 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 6743 weight); 6744 napi->weight = weight; 6745 napi->dev = dev; 6746 #ifdef CONFIG_NETPOLL 6747 napi->poll_owner = -1; 6748 #endif 6749 napi->list_owner = -1; 6750 set_bit(NAPI_STATE_SCHED, &napi->state); 6751 set_bit(NAPI_STATE_NPSVC, &napi->state); 6752 list_add_rcu(&napi->dev_list, &dev->napi_list); 6753 6754 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 6755 * configuration will be loaded in napi_enable 6756 */ 6757 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 6758 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 6759 6760 napi_get_frags_check(napi); 6761 /* Create kthread for this napi if dev->threaded is set. 6762 * Clear dev->threaded if kthread creation failed so that 6763 * threaded mode will not be enabled in napi_enable(). 6764 */ 6765 if (dev->threaded && napi_kthread_create(napi)) 6766 dev->threaded = false; 6767 netif_napi_set_irq(napi, -1); 6768 } 6769 EXPORT_SYMBOL(netif_napi_add_weight); 6770 6771 void napi_disable(struct napi_struct *n) 6772 { 6773 unsigned long val, new; 6774 6775 might_sleep(); 6776 set_bit(NAPI_STATE_DISABLE, &n->state); 6777 6778 val = READ_ONCE(n->state); 6779 do { 6780 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 6781 usleep_range(20, 200); 6782 val = READ_ONCE(n->state); 6783 } 6784 6785 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 6786 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 6787 } while (!try_cmpxchg(&n->state, &val, new)); 6788 6789 hrtimer_cancel(&n->timer); 6790 6791 if (n->config) 6792 napi_save_config(n); 6793 else 6794 napi_hash_del(n); 6795 6796 clear_bit(NAPI_STATE_DISABLE, &n->state); 6797 } 6798 EXPORT_SYMBOL(napi_disable); 6799 6800 /** 6801 * napi_enable - enable NAPI scheduling 6802 * @n: NAPI context 6803 * 6804 * Resume NAPI from being scheduled on this context. 6805 * Must be paired with napi_disable. 6806 */ 6807 void napi_enable(struct napi_struct *n) 6808 { 6809 unsigned long new, val = READ_ONCE(n->state); 6810 6811 if (n->config) 6812 napi_restore_config(n); 6813 else 6814 napi_hash_add(n); 6815 6816 do { 6817 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 6818 6819 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 6820 if (n->dev->threaded && n->thread) 6821 new |= NAPIF_STATE_THREADED; 6822 } while (!try_cmpxchg(&n->state, &val, new)); 6823 } 6824 EXPORT_SYMBOL(napi_enable); 6825 6826 static void flush_gro_hash(struct napi_struct *napi) 6827 { 6828 int i; 6829 6830 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6831 struct sk_buff *skb, *n; 6832 6833 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list) 6834 kfree_skb(skb); 6835 napi->gro_hash[i].count = 0; 6836 } 6837 } 6838 6839 /* Must be called in process context */ 6840 void __netif_napi_del(struct napi_struct *napi) 6841 { 6842 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 6843 return; 6844 6845 if (napi->config) { 6846 napi->index = -1; 6847 napi->config = NULL; 6848 } 6849 6850 list_del_rcu(&napi->dev_list); 6851 napi_free_frags(napi); 6852 6853 flush_gro_hash(napi); 6854 napi->gro_bitmask = 0; 6855 6856 if (napi->thread) { 6857 kthread_stop(napi->thread); 6858 napi->thread = NULL; 6859 } 6860 } 6861 EXPORT_SYMBOL(__netif_napi_del); 6862 6863 static int __napi_poll(struct napi_struct *n, bool *repoll) 6864 { 6865 int work, weight; 6866 6867 weight = n->weight; 6868 6869 /* This NAPI_STATE_SCHED test is for avoiding a race 6870 * with netpoll's poll_napi(). Only the entity which 6871 * obtains the lock and sees NAPI_STATE_SCHED set will 6872 * actually make the ->poll() call. Therefore we avoid 6873 * accidentally calling ->poll() when NAPI is not scheduled. 6874 */ 6875 work = 0; 6876 if (napi_is_scheduled(n)) { 6877 work = n->poll(n, weight); 6878 trace_napi_poll(n, work, weight); 6879 6880 xdp_do_check_flushed(n); 6881 } 6882 6883 if (unlikely(work > weight)) 6884 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 6885 n->poll, work, weight); 6886 6887 if (likely(work < weight)) 6888 return work; 6889 6890 /* Drivers must not modify the NAPI state if they 6891 * consume the entire weight. In such cases this code 6892 * still "owns" the NAPI instance and therefore can 6893 * move the instance around on the list at-will. 6894 */ 6895 if (unlikely(napi_disable_pending(n))) { 6896 napi_complete(n); 6897 return work; 6898 } 6899 6900 /* The NAPI context has more processing work, but busy-polling 6901 * is preferred. Exit early. 6902 */ 6903 if (napi_prefer_busy_poll(n)) { 6904 if (napi_complete_done(n, work)) { 6905 /* If timeout is not set, we need to make sure 6906 * that the NAPI is re-scheduled. 6907 */ 6908 napi_schedule(n); 6909 } 6910 return work; 6911 } 6912 6913 if (n->gro_bitmask) { 6914 /* flush too old packets 6915 * If HZ < 1000, flush all packets. 6916 */ 6917 napi_gro_flush(n, HZ >= 1000); 6918 } 6919 6920 gro_normal_list(n); 6921 6922 /* Some drivers may have called napi_schedule 6923 * prior to exhausting their budget. 6924 */ 6925 if (unlikely(!list_empty(&n->poll_list))) { 6926 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 6927 n->dev ? n->dev->name : "backlog"); 6928 return work; 6929 } 6930 6931 *repoll = true; 6932 6933 return work; 6934 } 6935 6936 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 6937 { 6938 bool do_repoll = false; 6939 void *have; 6940 int work; 6941 6942 list_del_init(&n->poll_list); 6943 6944 have = netpoll_poll_lock(n); 6945 6946 work = __napi_poll(n, &do_repoll); 6947 6948 if (do_repoll) 6949 list_add_tail(&n->poll_list, repoll); 6950 6951 netpoll_poll_unlock(have); 6952 6953 return work; 6954 } 6955 6956 static int napi_thread_wait(struct napi_struct *napi) 6957 { 6958 set_current_state(TASK_INTERRUPTIBLE); 6959 6960 while (!kthread_should_stop()) { 6961 /* Testing SCHED_THREADED bit here to make sure the current 6962 * kthread owns this napi and could poll on this napi. 6963 * Testing SCHED bit is not enough because SCHED bit might be 6964 * set by some other busy poll thread or by napi_disable(). 6965 */ 6966 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 6967 WARN_ON(!list_empty(&napi->poll_list)); 6968 __set_current_state(TASK_RUNNING); 6969 return 0; 6970 } 6971 6972 schedule(); 6973 set_current_state(TASK_INTERRUPTIBLE); 6974 } 6975 __set_current_state(TASK_RUNNING); 6976 6977 return -1; 6978 } 6979 6980 static void napi_threaded_poll_loop(struct napi_struct *napi) 6981 { 6982 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6983 struct softnet_data *sd; 6984 unsigned long last_qs = jiffies; 6985 6986 for (;;) { 6987 bool repoll = false; 6988 void *have; 6989 6990 local_bh_disable(); 6991 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6992 6993 sd = this_cpu_ptr(&softnet_data); 6994 sd->in_napi_threaded_poll = true; 6995 6996 have = netpoll_poll_lock(napi); 6997 __napi_poll(napi, &repoll); 6998 netpoll_poll_unlock(have); 6999 7000 sd->in_napi_threaded_poll = false; 7001 barrier(); 7002 7003 if (sd_has_rps_ipi_waiting(sd)) { 7004 local_irq_disable(); 7005 net_rps_action_and_irq_enable(sd); 7006 } 7007 skb_defer_free_flush(sd); 7008 bpf_net_ctx_clear(bpf_net_ctx); 7009 local_bh_enable(); 7010 7011 if (!repoll) 7012 break; 7013 7014 rcu_softirq_qs_periodic(last_qs); 7015 cond_resched(); 7016 } 7017 } 7018 7019 static int napi_threaded_poll(void *data) 7020 { 7021 struct napi_struct *napi = data; 7022 7023 while (!napi_thread_wait(napi)) 7024 napi_threaded_poll_loop(napi); 7025 7026 return 0; 7027 } 7028 7029 static __latent_entropy void net_rx_action(void) 7030 { 7031 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7032 unsigned long time_limit = jiffies + 7033 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7034 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7035 int budget = READ_ONCE(net_hotdata.netdev_budget); 7036 LIST_HEAD(list); 7037 LIST_HEAD(repoll); 7038 7039 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7040 start: 7041 sd->in_net_rx_action = true; 7042 local_irq_disable(); 7043 list_splice_init(&sd->poll_list, &list); 7044 local_irq_enable(); 7045 7046 for (;;) { 7047 struct napi_struct *n; 7048 7049 skb_defer_free_flush(sd); 7050 7051 if (list_empty(&list)) { 7052 if (list_empty(&repoll)) { 7053 sd->in_net_rx_action = false; 7054 barrier(); 7055 /* We need to check if ____napi_schedule() 7056 * had refilled poll_list while 7057 * sd->in_net_rx_action was true. 7058 */ 7059 if (!list_empty(&sd->poll_list)) 7060 goto start; 7061 if (!sd_has_rps_ipi_waiting(sd)) 7062 goto end; 7063 } 7064 break; 7065 } 7066 7067 n = list_first_entry(&list, struct napi_struct, poll_list); 7068 budget -= napi_poll(n, &repoll); 7069 7070 /* If softirq window is exhausted then punt. 7071 * Allow this to run for 2 jiffies since which will allow 7072 * an average latency of 1.5/HZ. 7073 */ 7074 if (unlikely(budget <= 0 || 7075 time_after_eq(jiffies, time_limit))) { 7076 sd->time_squeeze++; 7077 break; 7078 } 7079 } 7080 7081 local_irq_disable(); 7082 7083 list_splice_tail_init(&sd->poll_list, &list); 7084 list_splice_tail(&repoll, &list); 7085 list_splice(&list, &sd->poll_list); 7086 if (!list_empty(&sd->poll_list)) 7087 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7088 else 7089 sd->in_net_rx_action = false; 7090 7091 net_rps_action_and_irq_enable(sd); 7092 end: 7093 bpf_net_ctx_clear(bpf_net_ctx); 7094 } 7095 7096 struct netdev_adjacent { 7097 struct net_device *dev; 7098 netdevice_tracker dev_tracker; 7099 7100 /* upper master flag, there can only be one master device per list */ 7101 bool master; 7102 7103 /* lookup ignore flag */ 7104 bool ignore; 7105 7106 /* counter for the number of times this device was added to us */ 7107 u16 ref_nr; 7108 7109 /* private field for the users */ 7110 void *private; 7111 7112 struct list_head list; 7113 struct rcu_head rcu; 7114 }; 7115 7116 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7117 struct list_head *adj_list) 7118 { 7119 struct netdev_adjacent *adj; 7120 7121 list_for_each_entry(adj, adj_list, list) { 7122 if (adj->dev == adj_dev) 7123 return adj; 7124 } 7125 return NULL; 7126 } 7127 7128 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7129 struct netdev_nested_priv *priv) 7130 { 7131 struct net_device *dev = (struct net_device *)priv->data; 7132 7133 return upper_dev == dev; 7134 } 7135 7136 /** 7137 * netdev_has_upper_dev - Check if device is linked to an upper device 7138 * @dev: device 7139 * @upper_dev: upper device to check 7140 * 7141 * Find out if a device is linked to specified upper device and return true 7142 * in case it is. Note that this checks only immediate upper device, 7143 * not through a complete stack of devices. The caller must hold the RTNL lock. 7144 */ 7145 bool netdev_has_upper_dev(struct net_device *dev, 7146 struct net_device *upper_dev) 7147 { 7148 struct netdev_nested_priv priv = { 7149 .data = (void *)upper_dev, 7150 }; 7151 7152 ASSERT_RTNL(); 7153 7154 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7155 &priv); 7156 } 7157 EXPORT_SYMBOL(netdev_has_upper_dev); 7158 7159 /** 7160 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7161 * @dev: device 7162 * @upper_dev: upper device to check 7163 * 7164 * Find out if a device is linked to specified upper device and return true 7165 * in case it is. Note that this checks the entire upper device chain. 7166 * The caller must hold rcu lock. 7167 */ 7168 7169 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 7170 struct net_device *upper_dev) 7171 { 7172 struct netdev_nested_priv priv = { 7173 .data = (void *)upper_dev, 7174 }; 7175 7176 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7177 &priv); 7178 } 7179 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 7180 7181 /** 7182 * netdev_has_any_upper_dev - Check if device is linked to some device 7183 * @dev: device 7184 * 7185 * Find out if a device is linked to an upper device and return true in case 7186 * it is. The caller must hold the RTNL lock. 7187 */ 7188 bool netdev_has_any_upper_dev(struct net_device *dev) 7189 { 7190 ASSERT_RTNL(); 7191 7192 return !list_empty(&dev->adj_list.upper); 7193 } 7194 EXPORT_SYMBOL(netdev_has_any_upper_dev); 7195 7196 /** 7197 * netdev_master_upper_dev_get - Get master upper device 7198 * @dev: device 7199 * 7200 * Find a master upper device and return pointer to it or NULL in case 7201 * it's not there. The caller must hold the RTNL lock. 7202 */ 7203 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 7204 { 7205 struct netdev_adjacent *upper; 7206 7207 ASSERT_RTNL(); 7208 7209 if (list_empty(&dev->adj_list.upper)) 7210 return NULL; 7211 7212 upper = list_first_entry(&dev->adj_list.upper, 7213 struct netdev_adjacent, list); 7214 if (likely(upper->master)) 7215 return upper->dev; 7216 return NULL; 7217 } 7218 EXPORT_SYMBOL(netdev_master_upper_dev_get); 7219 7220 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 7221 { 7222 struct netdev_adjacent *upper; 7223 7224 ASSERT_RTNL(); 7225 7226 if (list_empty(&dev->adj_list.upper)) 7227 return NULL; 7228 7229 upper = list_first_entry(&dev->adj_list.upper, 7230 struct netdev_adjacent, list); 7231 if (likely(upper->master) && !upper->ignore) 7232 return upper->dev; 7233 return NULL; 7234 } 7235 7236 /** 7237 * netdev_has_any_lower_dev - Check if device is linked to some device 7238 * @dev: device 7239 * 7240 * Find out if a device is linked to a lower device and return true in case 7241 * it is. The caller must hold the RTNL lock. 7242 */ 7243 static bool netdev_has_any_lower_dev(struct net_device *dev) 7244 { 7245 ASSERT_RTNL(); 7246 7247 return !list_empty(&dev->adj_list.lower); 7248 } 7249 7250 void *netdev_adjacent_get_private(struct list_head *adj_list) 7251 { 7252 struct netdev_adjacent *adj; 7253 7254 adj = list_entry(adj_list, struct netdev_adjacent, list); 7255 7256 return adj->private; 7257 } 7258 EXPORT_SYMBOL(netdev_adjacent_get_private); 7259 7260 /** 7261 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7262 * @dev: device 7263 * @iter: list_head ** of the current position 7264 * 7265 * Gets the next device from the dev's upper list, starting from iter 7266 * position. The caller must hold RCU read lock. 7267 */ 7268 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7269 struct list_head **iter) 7270 { 7271 struct netdev_adjacent *upper; 7272 7273 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7274 7275 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7276 7277 if (&upper->list == &dev->adj_list.upper) 7278 return NULL; 7279 7280 *iter = &upper->list; 7281 7282 return upper->dev; 7283 } 7284 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7285 7286 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7287 struct list_head **iter, 7288 bool *ignore) 7289 { 7290 struct netdev_adjacent *upper; 7291 7292 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7293 7294 if (&upper->list == &dev->adj_list.upper) 7295 return NULL; 7296 7297 *iter = &upper->list; 7298 *ignore = upper->ignore; 7299 7300 return upper->dev; 7301 } 7302 7303 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7304 struct list_head **iter) 7305 { 7306 struct netdev_adjacent *upper; 7307 7308 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7309 7310 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7311 7312 if (&upper->list == &dev->adj_list.upper) 7313 return NULL; 7314 7315 *iter = &upper->list; 7316 7317 return upper->dev; 7318 } 7319 7320 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7321 int (*fn)(struct net_device *dev, 7322 struct netdev_nested_priv *priv), 7323 struct netdev_nested_priv *priv) 7324 { 7325 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7326 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7327 int ret, cur = 0; 7328 bool ignore; 7329 7330 now = dev; 7331 iter = &dev->adj_list.upper; 7332 7333 while (1) { 7334 if (now != dev) { 7335 ret = fn(now, priv); 7336 if (ret) 7337 return ret; 7338 } 7339 7340 next = NULL; 7341 while (1) { 7342 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7343 if (!udev) 7344 break; 7345 if (ignore) 7346 continue; 7347 7348 next = udev; 7349 niter = &udev->adj_list.upper; 7350 dev_stack[cur] = now; 7351 iter_stack[cur++] = iter; 7352 break; 7353 } 7354 7355 if (!next) { 7356 if (!cur) 7357 return 0; 7358 next = dev_stack[--cur]; 7359 niter = iter_stack[cur]; 7360 } 7361 7362 now = next; 7363 iter = niter; 7364 } 7365 7366 return 0; 7367 } 7368 7369 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7370 int (*fn)(struct net_device *dev, 7371 struct netdev_nested_priv *priv), 7372 struct netdev_nested_priv *priv) 7373 { 7374 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7375 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7376 int ret, cur = 0; 7377 7378 now = dev; 7379 iter = &dev->adj_list.upper; 7380 7381 while (1) { 7382 if (now != dev) { 7383 ret = fn(now, priv); 7384 if (ret) 7385 return ret; 7386 } 7387 7388 next = NULL; 7389 while (1) { 7390 udev = netdev_next_upper_dev_rcu(now, &iter); 7391 if (!udev) 7392 break; 7393 7394 next = udev; 7395 niter = &udev->adj_list.upper; 7396 dev_stack[cur] = now; 7397 iter_stack[cur++] = iter; 7398 break; 7399 } 7400 7401 if (!next) { 7402 if (!cur) 7403 return 0; 7404 next = dev_stack[--cur]; 7405 niter = iter_stack[cur]; 7406 } 7407 7408 now = next; 7409 iter = niter; 7410 } 7411 7412 return 0; 7413 } 7414 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7415 7416 static bool __netdev_has_upper_dev(struct net_device *dev, 7417 struct net_device *upper_dev) 7418 { 7419 struct netdev_nested_priv priv = { 7420 .flags = 0, 7421 .data = (void *)upper_dev, 7422 }; 7423 7424 ASSERT_RTNL(); 7425 7426 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7427 &priv); 7428 } 7429 7430 /** 7431 * netdev_lower_get_next_private - Get the next ->private from the 7432 * lower neighbour list 7433 * @dev: device 7434 * @iter: list_head ** of the current position 7435 * 7436 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7437 * list, starting from iter position. The caller must hold either hold the 7438 * RTNL lock or its own locking that guarantees that the neighbour lower 7439 * list will remain unchanged. 7440 */ 7441 void *netdev_lower_get_next_private(struct net_device *dev, 7442 struct list_head **iter) 7443 { 7444 struct netdev_adjacent *lower; 7445 7446 lower = list_entry(*iter, struct netdev_adjacent, list); 7447 7448 if (&lower->list == &dev->adj_list.lower) 7449 return NULL; 7450 7451 *iter = lower->list.next; 7452 7453 return lower->private; 7454 } 7455 EXPORT_SYMBOL(netdev_lower_get_next_private); 7456 7457 /** 7458 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7459 * lower neighbour list, RCU 7460 * variant 7461 * @dev: device 7462 * @iter: list_head ** of the current position 7463 * 7464 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7465 * list, starting from iter position. The caller must hold RCU read lock. 7466 */ 7467 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7468 struct list_head **iter) 7469 { 7470 struct netdev_adjacent *lower; 7471 7472 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7473 7474 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7475 7476 if (&lower->list == &dev->adj_list.lower) 7477 return NULL; 7478 7479 *iter = &lower->list; 7480 7481 return lower->private; 7482 } 7483 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7484 7485 /** 7486 * netdev_lower_get_next - Get the next device from the lower neighbour 7487 * list 7488 * @dev: device 7489 * @iter: list_head ** of the current position 7490 * 7491 * Gets the next netdev_adjacent from the dev's lower neighbour 7492 * list, starting from iter position. The caller must hold RTNL lock or 7493 * its own locking that guarantees that the neighbour lower 7494 * list will remain unchanged. 7495 */ 7496 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7497 { 7498 struct netdev_adjacent *lower; 7499 7500 lower = list_entry(*iter, struct netdev_adjacent, list); 7501 7502 if (&lower->list == &dev->adj_list.lower) 7503 return NULL; 7504 7505 *iter = lower->list.next; 7506 7507 return lower->dev; 7508 } 7509 EXPORT_SYMBOL(netdev_lower_get_next); 7510 7511 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 7512 struct list_head **iter) 7513 { 7514 struct netdev_adjacent *lower; 7515 7516 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7517 7518 if (&lower->list == &dev->adj_list.lower) 7519 return NULL; 7520 7521 *iter = &lower->list; 7522 7523 return lower->dev; 7524 } 7525 7526 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 7527 struct list_head **iter, 7528 bool *ignore) 7529 { 7530 struct netdev_adjacent *lower; 7531 7532 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7533 7534 if (&lower->list == &dev->adj_list.lower) 7535 return NULL; 7536 7537 *iter = &lower->list; 7538 *ignore = lower->ignore; 7539 7540 return lower->dev; 7541 } 7542 7543 int netdev_walk_all_lower_dev(struct net_device *dev, 7544 int (*fn)(struct net_device *dev, 7545 struct netdev_nested_priv *priv), 7546 struct netdev_nested_priv *priv) 7547 { 7548 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7549 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7550 int ret, cur = 0; 7551 7552 now = dev; 7553 iter = &dev->adj_list.lower; 7554 7555 while (1) { 7556 if (now != dev) { 7557 ret = fn(now, priv); 7558 if (ret) 7559 return ret; 7560 } 7561 7562 next = NULL; 7563 while (1) { 7564 ldev = netdev_next_lower_dev(now, &iter); 7565 if (!ldev) 7566 break; 7567 7568 next = ldev; 7569 niter = &ldev->adj_list.lower; 7570 dev_stack[cur] = now; 7571 iter_stack[cur++] = iter; 7572 break; 7573 } 7574 7575 if (!next) { 7576 if (!cur) 7577 return 0; 7578 next = dev_stack[--cur]; 7579 niter = iter_stack[cur]; 7580 } 7581 7582 now = next; 7583 iter = niter; 7584 } 7585 7586 return 0; 7587 } 7588 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 7589 7590 static int __netdev_walk_all_lower_dev(struct net_device *dev, 7591 int (*fn)(struct net_device *dev, 7592 struct netdev_nested_priv *priv), 7593 struct netdev_nested_priv *priv) 7594 { 7595 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7596 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7597 int ret, cur = 0; 7598 bool ignore; 7599 7600 now = dev; 7601 iter = &dev->adj_list.lower; 7602 7603 while (1) { 7604 if (now != dev) { 7605 ret = fn(now, priv); 7606 if (ret) 7607 return ret; 7608 } 7609 7610 next = NULL; 7611 while (1) { 7612 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 7613 if (!ldev) 7614 break; 7615 if (ignore) 7616 continue; 7617 7618 next = ldev; 7619 niter = &ldev->adj_list.lower; 7620 dev_stack[cur] = now; 7621 iter_stack[cur++] = iter; 7622 break; 7623 } 7624 7625 if (!next) { 7626 if (!cur) 7627 return 0; 7628 next = dev_stack[--cur]; 7629 niter = iter_stack[cur]; 7630 } 7631 7632 now = next; 7633 iter = niter; 7634 } 7635 7636 return 0; 7637 } 7638 7639 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 7640 struct list_head **iter) 7641 { 7642 struct netdev_adjacent *lower; 7643 7644 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7645 if (&lower->list == &dev->adj_list.lower) 7646 return NULL; 7647 7648 *iter = &lower->list; 7649 7650 return lower->dev; 7651 } 7652 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 7653 7654 static u8 __netdev_upper_depth(struct net_device *dev) 7655 { 7656 struct net_device *udev; 7657 struct list_head *iter; 7658 u8 max_depth = 0; 7659 bool ignore; 7660 7661 for (iter = &dev->adj_list.upper, 7662 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 7663 udev; 7664 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 7665 if (ignore) 7666 continue; 7667 if (max_depth < udev->upper_level) 7668 max_depth = udev->upper_level; 7669 } 7670 7671 return max_depth; 7672 } 7673 7674 static u8 __netdev_lower_depth(struct net_device *dev) 7675 { 7676 struct net_device *ldev; 7677 struct list_head *iter; 7678 u8 max_depth = 0; 7679 bool ignore; 7680 7681 for (iter = &dev->adj_list.lower, 7682 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 7683 ldev; 7684 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 7685 if (ignore) 7686 continue; 7687 if (max_depth < ldev->lower_level) 7688 max_depth = ldev->lower_level; 7689 } 7690 7691 return max_depth; 7692 } 7693 7694 static int __netdev_update_upper_level(struct net_device *dev, 7695 struct netdev_nested_priv *__unused) 7696 { 7697 dev->upper_level = __netdev_upper_depth(dev) + 1; 7698 return 0; 7699 } 7700 7701 #ifdef CONFIG_LOCKDEP 7702 static LIST_HEAD(net_unlink_list); 7703 7704 static void net_unlink_todo(struct net_device *dev) 7705 { 7706 if (list_empty(&dev->unlink_list)) 7707 list_add_tail(&dev->unlink_list, &net_unlink_list); 7708 } 7709 #endif 7710 7711 static int __netdev_update_lower_level(struct net_device *dev, 7712 struct netdev_nested_priv *priv) 7713 { 7714 dev->lower_level = __netdev_lower_depth(dev) + 1; 7715 7716 #ifdef CONFIG_LOCKDEP 7717 if (!priv) 7718 return 0; 7719 7720 if (priv->flags & NESTED_SYNC_IMM) 7721 dev->nested_level = dev->lower_level - 1; 7722 if (priv->flags & NESTED_SYNC_TODO) 7723 net_unlink_todo(dev); 7724 #endif 7725 return 0; 7726 } 7727 7728 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 7729 int (*fn)(struct net_device *dev, 7730 struct netdev_nested_priv *priv), 7731 struct netdev_nested_priv *priv) 7732 { 7733 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7734 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7735 int ret, cur = 0; 7736 7737 now = dev; 7738 iter = &dev->adj_list.lower; 7739 7740 while (1) { 7741 if (now != dev) { 7742 ret = fn(now, priv); 7743 if (ret) 7744 return ret; 7745 } 7746 7747 next = NULL; 7748 while (1) { 7749 ldev = netdev_next_lower_dev_rcu(now, &iter); 7750 if (!ldev) 7751 break; 7752 7753 next = ldev; 7754 niter = &ldev->adj_list.lower; 7755 dev_stack[cur] = now; 7756 iter_stack[cur++] = iter; 7757 break; 7758 } 7759 7760 if (!next) { 7761 if (!cur) 7762 return 0; 7763 next = dev_stack[--cur]; 7764 niter = iter_stack[cur]; 7765 } 7766 7767 now = next; 7768 iter = niter; 7769 } 7770 7771 return 0; 7772 } 7773 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 7774 7775 /** 7776 * netdev_lower_get_first_private_rcu - Get the first ->private from the 7777 * lower neighbour list, RCU 7778 * variant 7779 * @dev: device 7780 * 7781 * Gets the first netdev_adjacent->private from the dev's lower neighbour 7782 * list. The caller must hold RCU read lock. 7783 */ 7784 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 7785 { 7786 struct netdev_adjacent *lower; 7787 7788 lower = list_first_or_null_rcu(&dev->adj_list.lower, 7789 struct netdev_adjacent, list); 7790 if (lower) 7791 return lower->private; 7792 return NULL; 7793 } 7794 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 7795 7796 /** 7797 * netdev_master_upper_dev_get_rcu - Get master upper device 7798 * @dev: device 7799 * 7800 * Find a master upper device and return pointer to it or NULL in case 7801 * it's not there. The caller must hold the RCU read lock. 7802 */ 7803 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 7804 { 7805 struct netdev_adjacent *upper; 7806 7807 upper = list_first_or_null_rcu(&dev->adj_list.upper, 7808 struct netdev_adjacent, list); 7809 if (upper && likely(upper->master)) 7810 return upper->dev; 7811 return NULL; 7812 } 7813 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 7814 7815 static int netdev_adjacent_sysfs_add(struct net_device *dev, 7816 struct net_device *adj_dev, 7817 struct list_head *dev_list) 7818 { 7819 char linkname[IFNAMSIZ+7]; 7820 7821 sprintf(linkname, dev_list == &dev->adj_list.upper ? 7822 "upper_%s" : "lower_%s", adj_dev->name); 7823 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 7824 linkname); 7825 } 7826 static void netdev_adjacent_sysfs_del(struct net_device *dev, 7827 char *name, 7828 struct list_head *dev_list) 7829 { 7830 char linkname[IFNAMSIZ+7]; 7831 7832 sprintf(linkname, dev_list == &dev->adj_list.upper ? 7833 "upper_%s" : "lower_%s", name); 7834 sysfs_remove_link(&(dev->dev.kobj), linkname); 7835 } 7836 7837 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 7838 struct net_device *adj_dev, 7839 struct list_head *dev_list) 7840 { 7841 return (dev_list == &dev->adj_list.upper || 7842 dev_list == &dev->adj_list.lower) && 7843 net_eq(dev_net(dev), dev_net(adj_dev)); 7844 } 7845 7846 static int __netdev_adjacent_dev_insert(struct net_device *dev, 7847 struct net_device *adj_dev, 7848 struct list_head *dev_list, 7849 void *private, bool master) 7850 { 7851 struct netdev_adjacent *adj; 7852 int ret; 7853 7854 adj = __netdev_find_adj(adj_dev, dev_list); 7855 7856 if (adj) { 7857 adj->ref_nr += 1; 7858 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 7859 dev->name, adj_dev->name, adj->ref_nr); 7860 7861 return 0; 7862 } 7863 7864 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 7865 if (!adj) 7866 return -ENOMEM; 7867 7868 adj->dev = adj_dev; 7869 adj->master = master; 7870 adj->ref_nr = 1; 7871 adj->private = private; 7872 adj->ignore = false; 7873 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 7874 7875 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 7876 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 7877 7878 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 7879 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 7880 if (ret) 7881 goto free_adj; 7882 } 7883 7884 /* Ensure that master link is always the first item in list. */ 7885 if (master) { 7886 ret = sysfs_create_link(&(dev->dev.kobj), 7887 &(adj_dev->dev.kobj), "master"); 7888 if (ret) 7889 goto remove_symlinks; 7890 7891 list_add_rcu(&adj->list, dev_list); 7892 } else { 7893 list_add_tail_rcu(&adj->list, dev_list); 7894 } 7895 7896 return 0; 7897 7898 remove_symlinks: 7899 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 7900 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 7901 free_adj: 7902 netdev_put(adj_dev, &adj->dev_tracker); 7903 kfree(adj); 7904 7905 return ret; 7906 } 7907 7908 static void __netdev_adjacent_dev_remove(struct net_device *dev, 7909 struct net_device *adj_dev, 7910 u16 ref_nr, 7911 struct list_head *dev_list) 7912 { 7913 struct netdev_adjacent *adj; 7914 7915 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 7916 dev->name, adj_dev->name, ref_nr); 7917 7918 adj = __netdev_find_adj(adj_dev, dev_list); 7919 7920 if (!adj) { 7921 pr_err("Adjacency does not exist for device %s from %s\n", 7922 dev->name, adj_dev->name); 7923 WARN_ON(1); 7924 return; 7925 } 7926 7927 if (adj->ref_nr > ref_nr) { 7928 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 7929 dev->name, adj_dev->name, ref_nr, 7930 adj->ref_nr - ref_nr); 7931 adj->ref_nr -= ref_nr; 7932 return; 7933 } 7934 7935 if (adj->master) 7936 sysfs_remove_link(&(dev->dev.kobj), "master"); 7937 7938 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 7939 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 7940 7941 list_del_rcu(&adj->list); 7942 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 7943 adj_dev->name, dev->name, adj_dev->name); 7944 netdev_put(adj_dev, &adj->dev_tracker); 7945 kfree_rcu(adj, rcu); 7946 } 7947 7948 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 7949 struct net_device *upper_dev, 7950 struct list_head *up_list, 7951 struct list_head *down_list, 7952 void *private, bool master) 7953 { 7954 int ret; 7955 7956 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 7957 private, master); 7958 if (ret) 7959 return ret; 7960 7961 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 7962 private, false); 7963 if (ret) { 7964 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 7965 return ret; 7966 } 7967 7968 return 0; 7969 } 7970 7971 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 7972 struct net_device *upper_dev, 7973 u16 ref_nr, 7974 struct list_head *up_list, 7975 struct list_head *down_list) 7976 { 7977 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 7978 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 7979 } 7980 7981 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 7982 struct net_device *upper_dev, 7983 void *private, bool master) 7984 { 7985 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 7986 &dev->adj_list.upper, 7987 &upper_dev->adj_list.lower, 7988 private, master); 7989 } 7990 7991 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 7992 struct net_device *upper_dev) 7993 { 7994 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 7995 &dev->adj_list.upper, 7996 &upper_dev->adj_list.lower); 7997 } 7998 7999 static int __netdev_upper_dev_link(struct net_device *dev, 8000 struct net_device *upper_dev, bool master, 8001 void *upper_priv, void *upper_info, 8002 struct netdev_nested_priv *priv, 8003 struct netlink_ext_ack *extack) 8004 { 8005 struct netdev_notifier_changeupper_info changeupper_info = { 8006 .info = { 8007 .dev = dev, 8008 .extack = extack, 8009 }, 8010 .upper_dev = upper_dev, 8011 .master = master, 8012 .linking = true, 8013 .upper_info = upper_info, 8014 }; 8015 struct net_device *master_dev; 8016 int ret = 0; 8017 8018 ASSERT_RTNL(); 8019 8020 if (dev == upper_dev) 8021 return -EBUSY; 8022 8023 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8024 if (__netdev_has_upper_dev(upper_dev, dev)) 8025 return -EBUSY; 8026 8027 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8028 return -EMLINK; 8029 8030 if (!master) { 8031 if (__netdev_has_upper_dev(dev, upper_dev)) 8032 return -EEXIST; 8033 } else { 8034 master_dev = __netdev_master_upper_dev_get(dev); 8035 if (master_dev) 8036 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8037 } 8038 8039 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8040 &changeupper_info.info); 8041 ret = notifier_to_errno(ret); 8042 if (ret) 8043 return ret; 8044 8045 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8046 master); 8047 if (ret) 8048 return ret; 8049 8050 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8051 &changeupper_info.info); 8052 ret = notifier_to_errno(ret); 8053 if (ret) 8054 goto rollback; 8055 8056 __netdev_update_upper_level(dev, NULL); 8057 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8058 8059 __netdev_update_lower_level(upper_dev, priv); 8060 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8061 priv); 8062 8063 return 0; 8064 8065 rollback: 8066 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8067 8068 return ret; 8069 } 8070 8071 /** 8072 * netdev_upper_dev_link - Add a link to the upper device 8073 * @dev: device 8074 * @upper_dev: new upper device 8075 * @extack: netlink extended ack 8076 * 8077 * Adds a link to device which is upper to this one. The caller must hold 8078 * the RTNL lock. On a failure a negative errno code is returned. 8079 * On success the reference counts are adjusted and the function 8080 * returns zero. 8081 */ 8082 int netdev_upper_dev_link(struct net_device *dev, 8083 struct net_device *upper_dev, 8084 struct netlink_ext_ack *extack) 8085 { 8086 struct netdev_nested_priv priv = { 8087 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8088 .data = NULL, 8089 }; 8090 8091 return __netdev_upper_dev_link(dev, upper_dev, false, 8092 NULL, NULL, &priv, extack); 8093 } 8094 EXPORT_SYMBOL(netdev_upper_dev_link); 8095 8096 /** 8097 * netdev_master_upper_dev_link - Add a master link to the upper device 8098 * @dev: device 8099 * @upper_dev: new upper device 8100 * @upper_priv: upper device private 8101 * @upper_info: upper info to be passed down via notifier 8102 * @extack: netlink extended ack 8103 * 8104 * Adds a link to device which is upper to this one. In this case, only 8105 * one master upper device can be linked, although other non-master devices 8106 * might be linked as well. The caller must hold the RTNL lock. 8107 * On a failure a negative errno code is returned. On success the reference 8108 * counts are adjusted and the function returns zero. 8109 */ 8110 int netdev_master_upper_dev_link(struct net_device *dev, 8111 struct net_device *upper_dev, 8112 void *upper_priv, void *upper_info, 8113 struct netlink_ext_ack *extack) 8114 { 8115 struct netdev_nested_priv priv = { 8116 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8117 .data = NULL, 8118 }; 8119 8120 return __netdev_upper_dev_link(dev, upper_dev, true, 8121 upper_priv, upper_info, &priv, extack); 8122 } 8123 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8124 8125 static void __netdev_upper_dev_unlink(struct net_device *dev, 8126 struct net_device *upper_dev, 8127 struct netdev_nested_priv *priv) 8128 { 8129 struct netdev_notifier_changeupper_info changeupper_info = { 8130 .info = { 8131 .dev = dev, 8132 }, 8133 .upper_dev = upper_dev, 8134 .linking = false, 8135 }; 8136 8137 ASSERT_RTNL(); 8138 8139 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8140 8141 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8142 &changeupper_info.info); 8143 8144 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8145 8146 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8147 &changeupper_info.info); 8148 8149 __netdev_update_upper_level(dev, NULL); 8150 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8151 8152 __netdev_update_lower_level(upper_dev, priv); 8153 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8154 priv); 8155 } 8156 8157 /** 8158 * netdev_upper_dev_unlink - Removes a link to upper device 8159 * @dev: device 8160 * @upper_dev: new upper device 8161 * 8162 * Removes a link to device which is upper to this one. The caller must hold 8163 * the RTNL lock. 8164 */ 8165 void netdev_upper_dev_unlink(struct net_device *dev, 8166 struct net_device *upper_dev) 8167 { 8168 struct netdev_nested_priv priv = { 8169 .flags = NESTED_SYNC_TODO, 8170 .data = NULL, 8171 }; 8172 8173 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 8174 } 8175 EXPORT_SYMBOL(netdev_upper_dev_unlink); 8176 8177 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 8178 struct net_device *lower_dev, 8179 bool val) 8180 { 8181 struct netdev_adjacent *adj; 8182 8183 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 8184 if (adj) 8185 adj->ignore = val; 8186 8187 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 8188 if (adj) 8189 adj->ignore = val; 8190 } 8191 8192 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 8193 struct net_device *lower_dev) 8194 { 8195 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 8196 } 8197 8198 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 8199 struct net_device *lower_dev) 8200 { 8201 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 8202 } 8203 8204 int netdev_adjacent_change_prepare(struct net_device *old_dev, 8205 struct net_device *new_dev, 8206 struct net_device *dev, 8207 struct netlink_ext_ack *extack) 8208 { 8209 struct netdev_nested_priv priv = { 8210 .flags = 0, 8211 .data = NULL, 8212 }; 8213 int err; 8214 8215 if (!new_dev) 8216 return 0; 8217 8218 if (old_dev && new_dev != old_dev) 8219 netdev_adjacent_dev_disable(dev, old_dev); 8220 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 8221 extack); 8222 if (err) { 8223 if (old_dev && new_dev != old_dev) 8224 netdev_adjacent_dev_enable(dev, old_dev); 8225 return err; 8226 } 8227 8228 return 0; 8229 } 8230 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 8231 8232 void netdev_adjacent_change_commit(struct net_device *old_dev, 8233 struct net_device *new_dev, 8234 struct net_device *dev) 8235 { 8236 struct netdev_nested_priv priv = { 8237 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8238 .data = NULL, 8239 }; 8240 8241 if (!new_dev || !old_dev) 8242 return; 8243 8244 if (new_dev == old_dev) 8245 return; 8246 8247 netdev_adjacent_dev_enable(dev, old_dev); 8248 __netdev_upper_dev_unlink(old_dev, dev, &priv); 8249 } 8250 EXPORT_SYMBOL(netdev_adjacent_change_commit); 8251 8252 void netdev_adjacent_change_abort(struct net_device *old_dev, 8253 struct net_device *new_dev, 8254 struct net_device *dev) 8255 { 8256 struct netdev_nested_priv priv = { 8257 .flags = 0, 8258 .data = NULL, 8259 }; 8260 8261 if (!new_dev) 8262 return; 8263 8264 if (old_dev && new_dev != old_dev) 8265 netdev_adjacent_dev_enable(dev, old_dev); 8266 8267 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8268 } 8269 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8270 8271 /** 8272 * netdev_bonding_info_change - Dispatch event about slave change 8273 * @dev: device 8274 * @bonding_info: info to dispatch 8275 * 8276 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8277 * The caller must hold the RTNL lock. 8278 */ 8279 void netdev_bonding_info_change(struct net_device *dev, 8280 struct netdev_bonding_info *bonding_info) 8281 { 8282 struct netdev_notifier_bonding_info info = { 8283 .info.dev = dev, 8284 }; 8285 8286 memcpy(&info.bonding_info, bonding_info, 8287 sizeof(struct netdev_bonding_info)); 8288 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8289 &info.info); 8290 } 8291 EXPORT_SYMBOL(netdev_bonding_info_change); 8292 8293 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8294 struct netlink_ext_ack *extack) 8295 { 8296 struct netdev_notifier_offload_xstats_info info = { 8297 .info.dev = dev, 8298 .info.extack = extack, 8299 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8300 }; 8301 int err; 8302 int rc; 8303 8304 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8305 GFP_KERNEL); 8306 if (!dev->offload_xstats_l3) 8307 return -ENOMEM; 8308 8309 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8310 NETDEV_OFFLOAD_XSTATS_DISABLE, 8311 &info.info); 8312 err = notifier_to_errno(rc); 8313 if (err) 8314 goto free_stats; 8315 8316 return 0; 8317 8318 free_stats: 8319 kfree(dev->offload_xstats_l3); 8320 dev->offload_xstats_l3 = NULL; 8321 return err; 8322 } 8323 8324 int netdev_offload_xstats_enable(struct net_device *dev, 8325 enum netdev_offload_xstats_type type, 8326 struct netlink_ext_ack *extack) 8327 { 8328 ASSERT_RTNL(); 8329 8330 if (netdev_offload_xstats_enabled(dev, type)) 8331 return -EALREADY; 8332 8333 switch (type) { 8334 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8335 return netdev_offload_xstats_enable_l3(dev, extack); 8336 } 8337 8338 WARN_ON(1); 8339 return -EINVAL; 8340 } 8341 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8342 8343 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8344 { 8345 struct netdev_notifier_offload_xstats_info info = { 8346 .info.dev = dev, 8347 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8348 }; 8349 8350 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8351 &info.info); 8352 kfree(dev->offload_xstats_l3); 8353 dev->offload_xstats_l3 = NULL; 8354 } 8355 8356 int netdev_offload_xstats_disable(struct net_device *dev, 8357 enum netdev_offload_xstats_type type) 8358 { 8359 ASSERT_RTNL(); 8360 8361 if (!netdev_offload_xstats_enabled(dev, type)) 8362 return -EALREADY; 8363 8364 switch (type) { 8365 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8366 netdev_offload_xstats_disable_l3(dev); 8367 return 0; 8368 } 8369 8370 WARN_ON(1); 8371 return -EINVAL; 8372 } 8373 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8374 8375 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8376 { 8377 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8378 } 8379 8380 static struct rtnl_hw_stats64 * 8381 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8382 enum netdev_offload_xstats_type type) 8383 { 8384 switch (type) { 8385 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8386 return dev->offload_xstats_l3; 8387 } 8388 8389 WARN_ON(1); 8390 return NULL; 8391 } 8392 8393 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8394 enum netdev_offload_xstats_type type) 8395 { 8396 ASSERT_RTNL(); 8397 8398 return netdev_offload_xstats_get_ptr(dev, type); 8399 } 8400 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8401 8402 struct netdev_notifier_offload_xstats_ru { 8403 bool used; 8404 }; 8405 8406 struct netdev_notifier_offload_xstats_rd { 8407 struct rtnl_hw_stats64 stats; 8408 bool used; 8409 }; 8410 8411 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8412 const struct rtnl_hw_stats64 *src) 8413 { 8414 dest->rx_packets += src->rx_packets; 8415 dest->tx_packets += src->tx_packets; 8416 dest->rx_bytes += src->rx_bytes; 8417 dest->tx_bytes += src->tx_bytes; 8418 dest->rx_errors += src->rx_errors; 8419 dest->tx_errors += src->tx_errors; 8420 dest->rx_dropped += src->rx_dropped; 8421 dest->tx_dropped += src->tx_dropped; 8422 dest->multicast += src->multicast; 8423 } 8424 8425 static int netdev_offload_xstats_get_used(struct net_device *dev, 8426 enum netdev_offload_xstats_type type, 8427 bool *p_used, 8428 struct netlink_ext_ack *extack) 8429 { 8430 struct netdev_notifier_offload_xstats_ru report_used = {}; 8431 struct netdev_notifier_offload_xstats_info info = { 8432 .info.dev = dev, 8433 .info.extack = extack, 8434 .type = type, 8435 .report_used = &report_used, 8436 }; 8437 int rc; 8438 8439 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8440 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8441 &info.info); 8442 *p_used = report_used.used; 8443 return notifier_to_errno(rc); 8444 } 8445 8446 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8447 enum netdev_offload_xstats_type type, 8448 struct rtnl_hw_stats64 *p_stats, 8449 bool *p_used, 8450 struct netlink_ext_ack *extack) 8451 { 8452 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8453 struct netdev_notifier_offload_xstats_info info = { 8454 .info.dev = dev, 8455 .info.extack = extack, 8456 .type = type, 8457 .report_delta = &report_delta, 8458 }; 8459 struct rtnl_hw_stats64 *stats; 8460 int rc; 8461 8462 stats = netdev_offload_xstats_get_ptr(dev, type); 8463 if (WARN_ON(!stats)) 8464 return -EINVAL; 8465 8466 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8467 &info.info); 8468 8469 /* Cache whatever we got, even if there was an error, otherwise the 8470 * successful stats retrievals would get lost. 8471 */ 8472 netdev_hw_stats64_add(stats, &report_delta.stats); 8473 8474 if (p_stats) 8475 *p_stats = *stats; 8476 *p_used = report_delta.used; 8477 8478 return notifier_to_errno(rc); 8479 } 8480 8481 int netdev_offload_xstats_get(struct net_device *dev, 8482 enum netdev_offload_xstats_type type, 8483 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8484 struct netlink_ext_ack *extack) 8485 { 8486 ASSERT_RTNL(); 8487 8488 if (p_stats) 8489 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8490 p_used, extack); 8491 else 8492 return netdev_offload_xstats_get_used(dev, type, p_used, 8493 extack); 8494 } 8495 EXPORT_SYMBOL(netdev_offload_xstats_get); 8496 8497 void 8498 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8499 const struct rtnl_hw_stats64 *stats) 8500 { 8501 report_delta->used = true; 8502 netdev_hw_stats64_add(&report_delta->stats, stats); 8503 } 8504 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 8505 8506 void 8507 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 8508 { 8509 report_used->used = true; 8510 } 8511 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 8512 8513 void netdev_offload_xstats_push_delta(struct net_device *dev, 8514 enum netdev_offload_xstats_type type, 8515 const struct rtnl_hw_stats64 *p_stats) 8516 { 8517 struct rtnl_hw_stats64 *stats; 8518 8519 ASSERT_RTNL(); 8520 8521 stats = netdev_offload_xstats_get_ptr(dev, type); 8522 if (WARN_ON(!stats)) 8523 return; 8524 8525 netdev_hw_stats64_add(stats, p_stats); 8526 } 8527 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 8528 8529 /** 8530 * netdev_get_xmit_slave - Get the xmit slave of master device 8531 * @dev: device 8532 * @skb: The packet 8533 * @all_slaves: assume all the slaves are active 8534 * 8535 * The reference counters are not incremented so the caller must be 8536 * careful with locks. The caller must hold RCU lock. 8537 * %NULL is returned if no slave is found. 8538 */ 8539 8540 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 8541 struct sk_buff *skb, 8542 bool all_slaves) 8543 { 8544 const struct net_device_ops *ops = dev->netdev_ops; 8545 8546 if (!ops->ndo_get_xmit_slave) 8547 return NULL; 8548 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 8549 } 8550 EXPORT_SYMBOL(netdev_get_xmit_slave); 8551 8552 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 8553 struct sock *sk) 8554 { 8555 const struct net_device_ops *ops = dev->netdev_ops; 8556 8557 if (!ops->ndo_sk_get_lower_dev) 8558 return NULL; 8559 return ops->ndo_sk_get_lower_dev(dev, sk); 8560 } 8561 8562 /** 8563 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 8564 * @dev: device 8565 * @sk: the socket 8566 * 8567 * %NULL is returned if no lower device is found. 8568 */ 8569 8570 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 8571 struct sock *sk) 8572 { 8573 struct net_device *lower; 8574 8575 lower = netdev_sk_get_lower_dev(dev, sk); 8576 while (lower) { 8577 dev = lower; 8578 lower = netdev_sk_get_lower_dev(dev, sk); 8579 } 8580 8581 return dev; 8582 } 8583 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 8584 8585 static void netdev_adjacent_add_links(struct net_device *dev) 8586 { 8587 struct netdev_adjacent *iter; 8588 8589 struct net *net = dev_net(dev); 8590 8591 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8592 if (!net_eq(net, dev_net(iter->dev))) 8593 continue; 8594 netdev_adjacent_sysfs_add(iter->dev, dev, 8595 &iter->dev->adj_list.lower); 8596 netdev_adjacent_sysfs_add(dev, iter->dev, 8597 &dev->adj_list.upper); 8598 } 8599 8600 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8601 if (!net_eq(net, dev_net(iter->dev))) 8602 continue; 8603 netdev_adjacent_sysfs_add(iter->dev, dev, 8604 &iter->dev->adj_list.upper); 8605 netdev_adjacent_sysfs_add(dev, iter->dev, 8606 &dev->adj_list.lower); 8607 } 8608 } 8609 8610 static void netdev_adjacent_del_links(struct net_device *dev) 8611 { 8612 struct netdev_adjacent *iter; 8613 8614 struct net *net = dev_net(dev); 8615 8616 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8617 if (!net_eq(net, dev_net(iter->dev))) 8618 continue; 8619 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8620 &iter->dev->adj_list.lower); 8621 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8622 &dev->adj_list.upper); 8623 } 8624 8625 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8626 if (!net_eq(net, dev_net(iter->dev))) 8627 continue; 8628 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8629 &iter->dev->adj_list.upper); 8630 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8631 &dev->adj_list.lower); 8632 } 8633 } 8634 8635 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 8636 { 8637 struct netdev_adjacent *iter; 8638 8639 struct net *net = dev_net(dev); 8640 8641 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8642 if (!net_eq(net, dev_net(iter->dev))) 8643 continue; 8644 netdev_adjacent_sysfs_del(iter->dev, oldname, 8645 &iter->dev->adj_list.lower); 8646 netdev_adjacent_sysfs_add(iter->dev, dev, 8647 &iter->dev->adj_list.lower); 8648 } 8649 8650 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8651 if (!net_eq(net, dev_net(iter->dev))) 8652 continue; 8653 netdev_adjacent_sysfs_del(iter->dev, oldname, 8654 &iter->dev->adj_list.upper); 8655 netdev_adjacent_sysfs_add(iter->dev, dev, 8656 &iter->dev->adj_list.upper); 8657 } 8658 } 8659 8660 void *netdev_lower_dev_get_private(struct net_device *dev, 8661 struct net_device *lower_dev) 8662 { 8663 struct netdev_adjacent *lower; 8664 8665 if (!lower_dev) 8666 return NULL; 8667 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 8668 if (!lower) 8669 return NULL; 8670 8671 return lower->private; 8672 } 8673 EXPORT_SYMBOL(netdev_lower_dev_get_private); 8674 8675 8676 /** 8677 * netdev_lower_state_changed - Dispatch event about lower device state change 8678 * @lower_dev: device 8679 * @lower_state_info: state to dispatch 8680 * 8681 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 8682 * The caller must hold the RTNL lock. 8683 */ 8684 void netdev_lower_state_changed(struct net_device *lower_dev, 8685 void *lower_state_info) 8686 { 8687 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 8688 .info.dev = lower_dev, 8689 }; 8690 8691 ASSERT_RTNL(); 8692 changelowerstate_info.lower_state_info = lower_state_info; 8693 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 8694 &changelowerstate_info.info); 8695 } 8696 EXPORT_SYMBOL(netdev_lower_state_changed); 8697 8698 static void dev_change_rx_flags(struct net_device *dev, int flags) 8699 { 8700 const struct net_device_ops *ops = dev->netdev_ops; 8701 8702 if (ops->ndo_change_rx_flags) 8703 ops->ndo_change_rx_flags(dev, flags); 8704 } 8705 8706 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 8707 { 8708 unsigned int old_flags = dev->flags; 8709 unsigned int promiscuity, flags; 8710 kuid_t uid; 8711 kgid_t gid; 8712 8713 ASSERT_RTNL(); 8714 8715 promiscuity = dev->promiscuity + inc; 8716 if (promiscuity == 0) { 8717 /* 8718 * Avoid overflow. 8719 * If inc causes overflow, untouch promisc and return error. 8720 */ 8721 if (unlikely(inc > 0)) { 8722 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 8723 return -EOVERFLOW; 8724 } 8725 flags = old_flags & ~IFF_PROMISC; 8726 } else { 8727 flags = old_flags | IFF_PROMISC; 8728 } 8729 WRITE_ONCE(dev->promiscuity, promiscuity); 8730 if (flags != old_flags) { 8731 WRITE_ONCE(dev->flags, flags); 8732 netdev_info(dev, "%s promiscuous mode\n", 8733 dev->flags & IFF_PROMISC ? "entered" : "left"); 8734 if (audit_enabled) { 8735 current_uid_gid(&uid, &gid); 8736 audit_log(audit_context(), GFP_ATOMIC, 8737 AUDIT_ANOM_PROMISCUOUS, 8738 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 8739 dev->name, (dev->flags & IFF_PROMISC), 8740 (old_flags & IFF_PROMISC), 8741 from_kuid(&init_user_ns, audit_get_loginuid(current)), 8742 from_kuid(&init_user_ns, uid), 8743 from_kgid(&init_user_ns, gid), 8744 audit_get_sessionid(current)); 8745 } 8746 8747 dev_change_rx_flags(dev, IFF_PROMISC); 8748 } 8749 if (notify) 8750 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 8751 return 0; 8752 } 8753 8754 /** 8755 * dev_set_promiscuity - update promiscuity count on a device 8756 * @dev: device 8757 * @inc: modifier 8758 * 8759 * Add or remove promiscuity from a device. While the count in the device 8760 * remains above zero the interface remains promiscuous. Once it hits zero 8761 * the device reverts back to normal filtering operation. A negative inc 8762 * value is used to drop promiscuity on the device. 8763 * Return 0 if successful or a negative errno code on error. 8764 */ 8765 int dev_set_promiscuity(struct net_device *dev, int inc) 8766 { 8767 unsigned int old_flags = dev->flags; 8768 int err; 8769 8770 err = __dev_set_promiscuity(dev, inc, true); 8771 if (err < 0) 8772 return err; 8773 if (dev->flags != old_flags) 8774 dev_set_rx_mode(dev); 8775 return err; 8776 } 8777 EXPORT_SYMBOL(dev_set_promiscuity); 8778 8779 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify) 8780 { 8781 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 8782 unsigned int allmulti, flags; 8783 8784 ASSERT_RTNL(); 8785 8786 allmulti = dev->allmulti + inc; 8787 if (allmulti == 0) { 8788 /* 8789 * Avoid overflow. 8790 * If inc causes overflow, untouch allmulti and return error. 8791 */ 8792 if (unlikely(inc > 0)) { 8793 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 8794 return -EOVERFLOW; 8795 } 8796 flags = old_flags & ~IFF_ALLMULTI; 8797 } else { 8798 flags = old_flags | IFF_ALLMULTI; 8799 } 8800 WRITE_ONCE(dev->allmulti, allmulti); 8801 if (flags != old_flags) { 8802 WRITE_ONCE(dev->flags, flags); 8803 netdev_info(dev, "%s allmulticast mode\n", 8804 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 8805 dev_change_rx_flags(dev, IFF_ALLMULTI); 8806 dev_set_rx_mode(dev); 8807 if (notify) 8808 __dev_notify_flags(dev, old_flags, 8809 dev->gflags ^ old_gflags, 0, NULL); 8810 } 8811 return 0; 8812 } 8813 8814 /** 8815 * dev_set_allmulti - update allmulti count on a device 8816 * @dev: device 8817 * @inc: modifier 8818 * 8819 * Add or remove reception of all multicast frames to a device. While the 8820 * count in the device remains above zero the interface remains listening 8821 * to all interfaces. Once it hits zero the device reverts back to normal 8822 * filtering operation. A negative @inc value is used to drop the counter 8823 * when releasing a resource needing all multicasts. 8824 * Return 0 if successful or a negative errno code on error. 8825 */ 8826 8827 int dev_set_allmulti(struct net_device *dev, int inc) 8828 { 8829 return __dev_set_allmulti(dev, inc, true); 8830 } 8831 EXPORT_SYMBOL(dev_set_allmulti); 8832 8833 /* 8834 * Upload unicast and multicast address lists to device and 8835 * configure RX filtering. When the device doesn't support unicast 8836 * filtering it is put in promiscuous mode while unicast addresses 8837 * are present. 8838 */ 8839 void __dev_set_rx_mode(struct net_device *dev) 8840 { 8841 const struct net_device_ops *ops = dev->netdev_ops; 8842 8843 /* dev_open will call this function so the list will stay sane. */ 8844 if (!(dev->flags&IFF_UP)) 8845 return; 8846 8847 if (!netif_device_present(dev)) 8848 return; 8849 8850 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 8851 /* Unicast addresses changes may only happen under the rtnl, 8852 * therefore calling __dev_set_promiscuity here is safe. 8853 */ 8854 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 8855 __dev_set_promiscuity(dev, 1, false); 8856 dev->uc_promisc = true; 8857 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 8858 __dev_set_promiscuity(dev, -1, false); 8859 dev->uc_promisc = false; 8860 } 8861 } 8862 8863 if (ops->ndo_set_rx_mode) 8864 ops->ndo_set_rx_mode(dev); 8865 } 8866 8867 void dev_set_rx_mode(struct net_device *dev) 8868 { 8869 netif_addr_lock_bh(dev); 8870 __dev_set_rx_mode(dev); 8871 netif_addr_unlock_bh(dev); 8872 } 8873 8874 /** 8875 * dev_get_flags - get flags reported to userspace 8876 * @dev: device 8877 * 8878 * Get the combination of flag bits exported through APIs to userspace. 8879 */ 8880 unsigned int dev_get_flags(const struct net_device *dev) 8881 { 8882 unsigned int flags; 8883 8884 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 8885 IFF_ALLMULTI | 8886 IFF_RUNNING | 8887 IFF_LOWER_UP | 8888 IFF_DORMANT)) | 8889 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 8890 IFF_ALLMULTI)); 8891 8892 if (netif_running(dev)) { 8893 if (netif_oper_up(dev)) 8894 flags |= IFF_RUNNING; 8895 if (netif_carrier_ok(dev)) 8896 flags |= IFF_LOWER_UP; 8897 if (netif_dormant(dev)) 8898 flags |= IFF_DORMANT; 8899 } 8900 8901 return flags; 8902 } 8903 EXPORT_SYMBOL(dev_get_flags); 8904 8905 int __dev_change_flags(struct net_device *dev, unsigned int flags, 8906 struct netlink_ext_ack *extack) 8907 { 8908 unsigned int old_flags = dev->flags; 8909 int ret; 8910 8911 ASSERT_RTNL(); 8912 8913 /* 8914 * Set the flags on our device. 8915 */ 8916 8917 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 8918 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 8919 IFF_AUTOMEDIA)) | 8920 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 8921 IFF_ALLMULTI)); 8922 8923 /* 8924 * Load in the correct multicast list now the flags have changed. 8925 */ 8926 8927 if ((old_flags ^ flags) & IFF_MULTICAST) 8928 dev_change_rx_flags(dev, IFF_MULTICAST); 8929 8930 dev_set_rx_mode(dev); 8931 8932 /* 8933 * Have we downed the interface. We handle IFF_UP ourselves 8934 * according to user attempts to set it, rather than blindly 8935 * setting it. 8936 */ 8937 8938 ret = 0; 8939 if ((old_flags ^ flags) & IFF_UP) { 8940 if (old_flags & IFF_UP) 8941 __dev_close(dev); 8942 else 8943 ret = __dev_open(dev, extack); 8944 } 8945 8946 if ((flags ^ dev->gflags) & IFF_PROMISC) { 8947 int inc = (flags & IFF_PROMISC) ? 1 : -1; 8948 unsigned int old_flags = dev->flags; 8949 8950 dev->gflags ^= IFF_PROMISC; 8951 8952 if (__dev_set_promiscuity(dev, inc, false) >= 0) 8953 if (dev->flags != old_flags) 8954 dev_set_rx_mode(dev); 8955 } 8956 8957 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 8958 * is important. Some (broken) drivers set IFF_PROMISC, when 8959 * IFF_ALLMULTI is requested not asking us and not reporting. 8960 */ 8961 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 8962 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 8963 8964 dev->gflags ^= IFF_ALLMULTI; 8965 __dev_set_allmulti(dev, inc, false); 8966 } 8967 8968 return ret; 8969 } 8970 8971 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 8972 unsigned int gchanges, u32 portid, 8973 const struct nlmsghdr *nlh) 8974 { 8975 unsigned int changes = dev->flags ^ old_flags; 8976 8977 if (gchanges) 8978 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 8979 8980 if (changes & IFF_UP) { 8981 if (dev->flags & IFF_UP) 8982 call_netdevice_notifiers(NETDEV_UP, dev); 8983 else 8984 call_netdevice_notifiers(NETDEV_DOWN, dev); 8985 } 8986 8987 if (dev->flags & IFF_UP && 8988 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 8989 struct netdev_notifier_change_info change_info = { 8990 .info = { 8991 .dev = dev, 8992 }, 8993 .flags_changed = changes, 8994 }; 8995 8996 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 8997 } 8998 } 8999 9000 /** 9001 * dev_change_flags - change device settings 9002 * @dev: device 9003 * @flags: device state flags 9004 * @extack: netlink extended ack 9005 * 9006 * Change settings on device based state flags. The flags are 9007 * in the userspace exported format. 9008 */ 9009 int dev_change_flags(struct net_device *dev, unsigned int flags, 9010 struct netlink_ext_ack *extack) 9011 { 9012 int ret; 9013 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9014 9015 ret = __dev_change_flags(dev, flags, extack); 9016 if (ret < 0) 9017 return ret; 9018 9019 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9020 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9021 return ret; 9022 } 9023 EXPORT_SYMBOL(dev_change_flags); 9024 9025 int __dev_set_mtu(struct net_device *dev, int new_mtu) 9026 { 9027 const struct net_device_ops *ops = dev->netdev_ops; 9028 9029 if (ops->ndo_change_mtu) 9030 return ops->ndo_change_mtu(dev, new_mtu); 9031 9032 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9033 WRITE_ONCE(dev->mtu, new_mtu); 9034 return 0; 9035 } 9036 EXPORT_SYMBOL(__dev_set_mtu); 9037 9038 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9039 struct netlink_ext_ack *extack) 9040 { 9041 /* MTU must be positive, and in range */ 9042 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9043 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9044 return -EINVAL; 9045 } 9046 9047 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9048 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9049 return -EINVAL; 9050 } 9051 return 0; 9052 } 9053 9054 /** 9055 * dev_set_mtu_ext - Change maximum transfer unit 9056 * @dev: device 9057 * @new_mtu: new transfer unit 9058 * @extack: netlink extended ack 9059 * 9060 * Change the maximum transfer size of the network device. 9061 */ 9062 int dev_set_mtu_ext(struct net_device *dev, int new_mtu, 9063 struct netlink_ext_ack *extack) 9064 { 9065 int err, orig_mtu; 9066 9067 if (new_mtu == dev->mtu) 9068 return 0; 9069 9070 err = dev_validate_mtu(dev, new_mtu, extack); 9071 if (err) 9072 return err; 9073 9074 if (!netif_device_present(dev)) 9075 return -ENODEV; 9076 9077 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9078 err = notifier_to_errno(err); 9079 if (err) 9080 return err; 9081 9082 orig_mtu = dev->mtu; 9083 err = __dev_set_mtu(dev, new_mtu); 9084 9085 if (!err) { 9086 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9087 orig_mtu); 9088 err = notifier_to_errno(err); 9089 if (err) { 9090 /* setting mtu back and notifying everyone again, 9091 * so that they have a chance to revert changes. 9092 */ 9093 __dev_set_mtu(dev, orig_mtu); 9094 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9095 new_mtu); 9096 } 9097 } 9098 return err; 9099 } 9100 9101 int dev_set_mtu(struct net_device *dev, int new_mtu) 9102 { 9103 struct netlink_ext_ack extack; 9104 int err; 9105 9106 memset(&extack, 0, sizeof(extack)); 9107 err = dev_set_mtu_ext(dev, new_mtu, &extack); 9108 if (err && extack._msg) 9109 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9110 return err; 9111 } 9112 EXPORT_SYMBOL(dev_set_mtu); 9113 9114 /** 9115 * dev_change_tx_queue_len - Change TX queue length of a netdevice 9116 * @dev: device 9117 * @new_len: new tx queue length 9118 */ 9119 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9120 { 9121 unsigned int orig_len = dev->tx_queue_len; 9122 int res; 9123 9124 if (new_len != (unsigned int)new_len) 9125 return -ERANGE; 9126 9127 if (new_len != orig_len) { 9128 WRITE_ONCE(dev->tx_queue_len, new_len); 9129 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9130 res = notifier_to_errno(res); 9131 if (res) 9132 goto err_rollback; 9133 res = dev_qdisc_change_tx_queue_len(dev); 9134 if (res) 9135 goto err_rollback; 9136 } 9137 9138 return 0; 9139 9140 err_rollback: 9141 netdev_err(dev, "refused to change device tx_queue_len\n"); 9142 WRITE_ONCE(dev->tx_queue_len, orig_len); 9143 return res; 9144 } 9145 9146 /** 9147 * dev_set_group - Change group this device belongs to 9148 * @dev: device 9149 * @new_group: group this device should belong to 9150 */ 9151 void dev_set_group(struct net_device *dev, int new_group) 9152 { 9153 dev->group = new_group; 9154 } 9155 9156 /** 9157 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 9158 * @dev: device 9159 * @addr: new address 9160 * @extack: netlink extended ack 9161 */ 9162 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9163 struct netlink_ext_ack *extack) 9164 { 9165 struct netdev_notifier_pre_changeaddr_info info = { 9166 .info.dev = dev, 9167 .info.extack = extack, 9168 .dev_addr = addr, 9169 }; 9170 int rc; 9171 9172 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9173 return notifier_to_errno(rc); 9174 } 9175 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 9176 9177 /** 9178 * dev_set_mac_address - Change Media Access Control Address 9179 * @dev: device 9180 * @sa: new address 9181 * @extack: netlink extended ack 9182 * 9183 * Change the hardware (MAC) address of the device 9184 */ 9185 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa, 9186 struct netlink_ext_ack *extack) 9187 { 9188 const struct net_device_ops *ops = dev->netdev_ops; 9189 int err; 9190 9191 if (!ops->ndo_set_mac_address) 9192 return -EOPNOTSUPP; 9193 if (sa->sa_family != dev->type) 9194 return -EINVAL; 9195 if (!netif_device_present(dev)) 9196 return -ENODEV; 9197 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 9198 if (err) 9199 return err; 9200 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 9201 err = ops->ndo_set_mac_address(dev, sa); 9202 if (err) 9203 return err; 9204 } 9205 dev->addr_assign_type = NET_ADDR_SET; 9206 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9207 add_device_randomness(dev->dev_addr, dev->addr_len); 9208 return 0; 9209 } 9210 EXPORT_SYMBOL(dev_set_mac_address); 9211 9212 DECLARE_RWSEM(dev_addr_sem); 9213 9214 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa, 9215 struct netlink_ext_ack *extack) 9216 { 9217 int ret; 9218 9219 down_write(&dev_addr_sem); 9220 ret = dev_set_mac_address(dev, sa, extack); 9221 up_write(&dev_addr_sem); 9222 return ret; 9223 } 9224 EXPORT_SYMBOL(dev_set_mac_address_user); 9225 9226 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9227 { 9228 size_t size = sizeof(sa->sa_data_min); 9229 struct net_device *dev; 9230 int ret = 0; 9231 9232 down_read(&dev_addr_sem); 9233 rcu_read_lock(); 9234 9235 dev = dev_get_by_name_rcu(net, dev_name); 9236 if (!dev) { 9237 ret = -ENODEV; 9238 goto unlock; 9239 } 9240 if (!dev->addr_len) 9241 memset(sa->sa_data, 0, size); 9242 else 9243 memcpy(sa->sa_data, dev->dev_addr, 9244 min_t(size_t, size, dev->addr_len)); 9245 sa->sa_family = dev->type; 9246 9247 unlock: 9248 rcu_read_unlock(); 9249 up_read(&dev_addr_sem); 9250 return ret; 9251 } 9252 EXPORT_SYMBOL(dev_get_mac_address); 9253 9254 /** 9255 * dev_change_carrier - Change device carrier 9256 * @dev: device 9257 * @new_carrier: new value 9258 * 9259 * Change device carrier 9260 */ 9261 int dev_change_carrier(struct net_device *dev, bool new_carrier) 9262 { 9263 const struct net_device_ops *ops = dev->netdev_ops; 9264 9265 if (!ops->ndo_change_carrier) 9266 return -EOPNOTSUPP; 9267 if (!netif_device_present(dev)) 9268 return -ENODEV; 9269 return ops->ndo_change_carrier(dev, new_carrier); 9270 } 9271 9272 /** 9273 * dev_get_phys_port_id - Get device physical port ID 9274 * @dev: device 9275 * @ppid: port ID 9276 * 9277 * Get device physical port ID 9278 */ 9279 int dev_get_phys_port_id(struct net_device *dev, 9280 struct netdev_phys_item_id *ppid) 9281 { 9282 const struct net_device_ops *ops = dev->netdev_ops; 9283 9284 if (!ops->ndo_get_phys_port_id) 9285 return -EOPNOTSUPP; 9286 return ops->ndo_get_phys_port_id(dev, ppid); 9287 } 9288 9289 /** 9290 * dev_get_phys_port_name - Get device physical port name 9291 * @dev: device 9292 * @name: port name 9293 * @len: limit of bytes to copy to name 9294 * 9295 * Get device physical port name 9296 */ 9297 int dev_get_phys_port_name(struct net_device *dev, 9298 char *name, size_t len) 9299 { 9300 const struct net_device_ops *ops = dev->netdev_ops; 9301 int err; 9302 9303 if (ops->ndo_get_phys_port_name) { 9304 err = ops->ndo_get_phys_port_name(dev, name, len); 9305 if (err != -EOPNOTSUPP) 9306 return err; 9307 } 9308 return devlink_compat_phys_port_name_get(dev, name, len); 9309 } 9310 9311 /** 9312 * dev_get_port_parent_id - Get the device's port parent identifier 9313 * @dev: network device 9314 * @ppid: pointer to a storage for the port's parent identifier 9315 * @recurse: allow/disallow recursion to lower devices 9316 * 9317 * Get the devices's port parent identifier 9318 */ 9319 int dev_get_port_parent_id(struct net_device *dev, 9320 struct netdev_phys_item_id *ppid, 9321 bool recurse) 9322 { 9323 const struct net_device_ops *ops = dev->netdev_ops; 9324 struct netdev_phys_item_id first = { }; 9325 struct net_device *lower_dev; 9326 struct list_head *iter; 9327 int err; 9328 9329 if (ops->ndo_get_port_parent_id) { 9330 err = ops->ndo_get_port_parent_id(dev, ppid); 9331 if (err != -EOPNOTSUPP) 9332 return err; 9333 } 9334 9335 err = devlink_compat_switch_id_get(dev, ppid); 9336 if (!recurse || err != -EOPNOTSUPP) 9337 return err; 9338 9339 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9340 err = dev_get_port_parent_id(lower_dev, ppid, true); 9341 if (err) 9342 break; 9343 if (!first.id_len) 9344 first = *ppid; 9345 else if (memcmp(&first, ppid, sizeof(*ppid))) 9346 return -EOPNOTSUPP; 9347 } 9348 9349 return err; 9350 } 9351 EXPORT_SYMBOL(dev_get_port_parent_id); 9352 9353 /** 9354 * netdev_port_same_parent_id - Indicate if two network devices have 9355 * the same port parent identifier 9356 * @a: first network device 9357 * @b: second network device 9358 */ 9359 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9360 { 9361 struct netdev_phys_item_id a_id = { }; 9362 struct netdev_phys_item_id b_id = { }; 9363 9364 if (dev_get_port_parent_id(a, &a_id, true) || 9365 dev_get_port_parent_id(b, &b_id, true)) 9366 return false; 9367 9368 return netdev_phys_item_id_same(&a_id, &b_id); 9369 } 9370 EXPORT_SYMBOL(netdev_port_same_parent_id); 9371 9372 /** 9373 * dev_change_proto_down - set carrier according to proto_down. 9374 * 9375 * @dev: device 9376 * @proto_down: new value 9377 */ 9378 int dev_change_proto_down(struct net_device *dev, bool proto_down) 9379 { 9380 if (!dev->change_proto_down) 9381 return -EOPNOTSUPP; 9382 if (!netif_device_present(dev)) 9383 return -ENODEV; 9384 if (proto_down) 9385 netif_carrier_off(dev); 9386 else 9387 netif_carrier_on(dev); 9388 WRITE_ONCE(dev->proto_down, proto_down); 9389 return 0; 9390 } 9391 9392 /** 9393 * dev_change_proto_down_reason - proto down reason 9394 * 9395 * @dev: device 9396 * @mask: proto down mask 9397 * @value: proto down value 9398 */ 9399 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask, 9400 u32 value) 9401 { 9402 u32 proto_down_reason; 9403 int b; 9404 9405 if (!mask) { 9406 proto_down_reason = value; 9407 } else { 9408 proto_down_reason = dev->proto_down_reason; 9409 for_each_set_bit(b, &mask, 32) { 9410 if (value & (1 << b)) 9411 proto_down_reason |= BIT(b); 9412 else 9413 proto_down_reason &= ~BIT(b); 9414 } 9415 } 9416 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 9417 } 9418 9419 struct bpf_xdp_link { 9420 struct bpf_link link; 9421 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9422 int flags; 9423 }; 9424 9425 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9426 { 9427 if (flags & XDP_FLAGS_HW_MODE) 9428 return XDP_MODE_HW; 9429 if (flags & XDP_FLAGS_DRV_MODE) 9430 return XDP_MODE_DRV; 9431 if (flags & XDP_FLAGS_SKB_MODE) 9432 return XDP_MODE_SKB; 9433 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9434 } 9435 9436 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9437 { 9438 switch (mode) { 9439 case XDP_MODE_SKB: 9440 return generic_xdp_install; 9441 case XDP_MODE_DRV: 9442 case XDP_MODE_HW: 9443 return dev->netdev_ops->ndo_bpf; 9444 default: 9445 return NULL; 9446 } 9447 } 9448 9449 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9450 enum bpf_xdp_mode mode) 9451 { 9452 return dev->xdp_state[mode].link; 9453 } 9454 9455 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9456 enum bpf_xdp_mode mode) 9457 { 9458 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9459 9460 if (link) 9461 return link->link.prog; 9462 return dev->xdp_state[mode].prog; 9463 } 9464 9465 u8 dev_xdp_prog_count(struct net_device *dev) 9466 { 9467 u8 count = 0; 9468 int i; 9469 9470 for (i = 0; i < __MAX_XDP_MODE; i++) 9471 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9472 count++; 9473 return count; 9474 } 9475 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9476 9477 int dev_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 9478 { 9479 if (!dev->netdev_ops->ndo_bpf) 9480 return -EOPNOTSUPP; 9481 9482 if (dev_get_min_mp_channel_count(dev)) { 9483 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 9484 return -EBUSY; 9485 } 9486 9487 return dev->netdev_ops->ndo_bpf(dev, bpf); 9488 } 9489 EXPORT_SYMBOL_GPL(dev_xdp_propagate); 9490 9491 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9492 { 9493 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9494 9495 return prog ? prog->aux->id : 0; 9496 } 9497 9498 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9499 struct bpf_xdp_link *link) 9500 { 9501 dev->xdp_state[mode].link = link; 9502 dev->xdp_state[mode].prog = NULL; 9503 } 9504 9505 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9506 struct bpf_prog *prog) 9507 { 9508 dev->xdp_state[mode].link = NULL; 9509 dev->xdp_state[mode].prog = prog; 9510 } 9511 9512 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9513 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9514 u32 flags, struct bpf_prog *prog) 9515 { 9516 struct netdev_bpf xdp; 9517 int err; 9518 9519 if (dev_get_min_mp_channel_count(dev)) { 9520 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 9521 return -EBUSY; 9522 } 9523 9524 memset(&xdp, 0, sizeof(xdp)); 9525 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9526 xdp.extack = extack; 9527 xdp.flags = flags; 9528 xdp.prog = prog; 9529 9530 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9531 * "moved" into driver), so they don't increment it on their own, but 9532 * they do decrement refcnt when program is detached or replaced. 9533 * Given net_device also owns link/prog, we need to bump refcnt here 9534 * to prevent drivers from underflowing it. 9535 */ 9536 if (prog) 9537 bpf_prog_inc(prog); 9538 err = bpf_op(dev, &xdp); 9539 if (err) { 9540 if (prog) 9541 bpf_prog_put(prog); 9542 return err; 9543 } 9544 9545 if (mode != XDP_MODE_HW) 9546 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9547 9548 return 0; 9549 } 9550 9551 static void dev_xdp_uninstall(struct net_device *dev) 9552 { 9553 struct bpf_xdp_link *link; 9554 struct bpf_prog *prog; 9555 enum bpf_xdp_mode mode; 9556 bpf_op_t bpf_op; 9557 9558 ASSERT_RTNL(); 9559 9560 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 9561 prog = dev_xdp_prog(dev, mode); 9562 if (!prog) 9563 continue; 9564 9565 bpf_op = dev_xdp_bpf_op(dev, mode); 9566 if (!bpf_op) 9567 continue; 9568 9569 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9570 9571 /* auto-detach link from net device */ 9572 link = dev_xdp_link(dev, mode); 9573 if (link) 9574 link->dev = NULL; 9575 else 9576 bpf_prog_put(prog); 9577 9578 dev_xdp_set_link(dev, mode, NULL); 9579 } 9580 } 9581 9582 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 9583 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 9584 struct bpf_prog *old_prog, u32 flags) 9585 { 9586 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 9587 struct bpf_prog *cur_prog; 9588 struct net_device *upper; 9589 struct list_head *iter; 9590 enum bpf_xdp_mode mode; 9591 bpf_op_t bpf_op; 9592 int err; 9593 9594 ASSERT_RTNL(); 9595 9596 /* either link or prog attachment, never both */ 9597 if (link && (new_prog || old_prog)) 9598 return -EINVAL; 9599 /* link supports only XDP mode flags */ 9600 if (link && (flags & ~XDP_FLAGS_MODES)) { 9601 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 9602 return -EINVAL; 9603 } 9604 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 9605 if (num_modes > 1) { 9606 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 9607 return -EINVAL; 9608 } 9609 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 9610 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 9611 NL_SET_ERR_MSG(extack, 9612 "More than one program loaded, unset mode is ambiguous"); 9613 return -EINVAL; 9614 } 9615 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 9616 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 9617 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 9618 return -EINVAL; 9619 } 9620 9621 mode = dev_xdp_mode(dev, flags); 9622 /* can't replace attached link */ 9623 if (dev_xdp_link(dev, mode)) { 9624 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 9625 return -EBUSY; 9626 } 9627 9628 /* don't allow if an upper device already has a program */ 9629 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 9630 if (dev_xdp_prog_count(upper) > 0) { 9631 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 9632 return -EEXIST; 9633 } 9634 } 9635 9636 cur_prog = dev_xdp_prog(dev, mode); 9637 /* can't replace attached prog with link */ 9638 if (link && cur_prog) { 9639 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 9640 return -EBUSY; 9641 } 9642 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 9643 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 9644 return -EEXIST; 9645 } 9646 9647 /* put effective new program into new_prog */ 9648 if (link) 9649 new_prog = link->link.prog; 9650 9651 if (new_prog) { 9652 bool offload = mode == XDP_MODE_HW; 9653 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 9654 ? XDP_MODE_DRV : XDP_MODE_SKB; 9655 9656 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 9657 NL_SET_ERR_MSG(extack, "XDP program already attached"); 9658 return -EBUSY; 9659 } 9660 if (!offload && dev_xdp_prog(dev, other_mode)) { 9661 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 9662 return -EEXIST; 9663 } 9664 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 9665 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 9666 return -EINVAL; 9667 } 9668 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 9669 NL_SET_ERR_MSG(extack, "Program bound to different device"); 9670 return -EINVAL; 9671 } 9672 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 9673 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 9674 return -EINVAL; 9675 } 9676 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 9677 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 9678 return -EINVAL; 9679 } 9680 } 9681 9682 /* don't call drivers if the effective program didn't change */ 9683 if (new_prog != cur_prog) { 9684 bpf_op = dev_xdp_bpf_op(dev, mode); 9685 if (!bpf_op) { 9686 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 9687 return -EOPNOTSUPP; 9688 } 9689 9690 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 9691 if (err) 9692 return err; 9693 } 9694 9695 if (link) 9696 dev_xdp_set_link(dev, mode, link); 9697 else 9698 dev_xdp_set_prog(dev, mode, new_prog); 9699 if (cur_prog) 9700 bpf_prog_put(cur_prog); 9701 9702 return 0; 9703 } 9704 9705 static int dev_xdp_attach_link(struct net_device *dev, 9706 struct netlink_ext_ack *extack, 9707 struct bpf_xdp_link *link) 9708 { 9709 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 9710 } 9711 9712 static int dev_xdp_detach_link(struct net_device *dev, 9713 struct netlink_ext_ack *extack, 9714 struct bpf_xdp_link *link) 9715 { 9716 enum bpf_xdp_mode mode; 9717 bpf_op_t bpf_op; 9718 9719 ASSERT_RTNL(); 9720 9721 mode = dev_xdp_mode(dev, link->flags); 9722 if (dev_xdp_link(dev, mode) != link) 9723 return -EINVAL; 9724 9725 bpf_op = dev_xdp_bpf_op(dev, mode); 9726 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9727 dev_xdp_set_link(dev, mode, NULL); 9728 return 0; 9729 } 9730 9731 static void bpf_xdp_link_release(struct bpf_link *link) 9732 { 9733 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9734 9735 rtnl_lock(); 9736 9737 /* if racing with net_device's tear down, xdp_link->dev might be 9738 * already NULL, in which case link was already auto-detached 9739 */ 9740 if (xdp_link->dev) { 9741 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 9742 xdp_link->dev = NULL; 9743 } 9744 9745 rtnl_unlock(); 9746 } 9747 9748 static int bpf_xdp_link_detach(struct bpf_link *link) 9749 { 9750 bpf_xdp_link_release(link); 9751 return 0; 9752 } 9753 9754 static void bpf_xdp_link_dealloc(struct bpf_link *link) 9755 { 9756 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9757 9758 kfree(xdp_link); 9759 } 9760 9761 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 9762 struct seq_file *seq) 9763 { 9764 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9765 u32 ifindex = 0; 9766 9767 rtnl_lock(); 9768 if (xdp_link->dev) 9769 ifindex = xdp_link->dev->ifindex; 9770 rtnl_unlock(); 9771 9772 seq_printf(seq, "ifindex:\t%u\n", ifindex); 9773 } 9774 9775 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 9776 struct bpf_link_info *info) 9777 { 9778 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9779 u32 ifindex = 0; 9780 9781 rtnl_lock(); 9782 if (xdp_link->dev) 9783 ifindex = xdp_link->dev->ifindex; 9784 rtnl_unlock(); 9785 9786 info->xdp.ifindex = ifindex; 9787 return 0; 9788 } 9789 9790 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 9791 struct bpf_prog *old_prog) 9792 { 9793 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9794 enum bpf_xdp_mode mode; 9795 bpf_op_t bpf_op; 9796 int err = 0; 9797 9798 rtnl_lock(); 9799 9800 /* link might have been auto-released already, so fail */ 9801 if (!xdp_link->dev) { 9802 err = -ENOLINK; 9803 goto out_unlock; 9804 } 9805 9806 if (old_prog && link->prog != old_prog) { 9807 err = -EPERM; 9808 goto out_unlock; 9809 } 9810 old_prog = link->prog; 9811 if (old_prog->type != new_prog->type || 9812 old_prog->expected_attach_type != new_prog->expected_attach_type) { 9813 err = -EINVAL; 9814 goto out_unlock; 9815 } 9816 9817 if (old_prog == new_prog) { 9818 /* no-op, don't disturb drivers */ 9819 bpf_prog_put(new_prog); 9820 goto out_unlock; 9821 } 9822 9823 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 9824 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 9825 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 9826 xdp_link->flags, new_prog); 9827 if (err) 9828 goto out_unlock; 9829 9830 old_prog = xchg(&link->prog, new_prog); 9831 bpf_prog_put(old_prog); 9832 9833 out_unlock: 9834 rtnl_unlock(); 9835 return err; 9836 } 9837 9838 static const struct bpf_link_ops bpf_xdp_link_lops = { 9839 .release = bpf_xdp_link_release, 9840 .dealloc = bpf_xdp_link_dealloc, 9841 .detach = bpf_xdp_link_detach, 9842 .show_fdinfo = bpf_xdp_link_show_fdinfo, 9843 .fill_link_info = bpf_xdp_link_fill_link_info, 9844 .update_prog = bpf_xdp_link_update, 9845 }; 9846 9847 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 9848 { 9849 struct net *net = current->nsproxy->net_ns; 9850 struct bpf_link_primer link_primer; 9851 struct netlink_ext_ack extack = {}; 9852 struct bpf_xdp_link *link; 9853 struct net_device *dev; 9854 int err, fd; 9855 9856 rtnl_lock(); 9857 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 9858 if (!dev) { 9859 rtnl_unlock(); 9860 return -EINVAL; 9861 } 9862 9863 link = kzalloc(sizeof(*link), GFP_USER); 9864 if (!link) { 9865 err = -ENOMEM; 9866 goto unlock; 9867 } 9868 9869 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 9870 link->dev = dev; 9871 link->flags = attr->link_create.flags; 9872 9873 err = bpf_link_prime(&link->link, &link_primer); 9874 if (err) { 9875 kfree(link); 9876 goto unlock; 9877 } 9878 9879 err = dev_xdp_attach_link(dev, &extack, link); 9880 rtnl_unlock(); 9881 9882 if (err) { 9883 link->dev = NULL; 9884 bpf_link_cleanup(&link_primer); 9885 trace_bpf_xdp_link_attach_failed(extack._msg); 9886 goto out_put_dev; 9887 } 9888 9889 fd = bpf_link_settle(&link_primer); 9890 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 9891 dev_put(dev); 9892 return fd; 9893 9894 unlock: 9895 rtnl_unlock(); 9896 9897 out_put_dev: 9898 dev_put(dev); 9899 return err; 9900 } 9901 9902 /** 9903 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 9904 * @dev: device 9905 * @extack: netlink extended ack 9906 * @fd: new program fd or negative value to clear 9907 * @expected_fd: old program fd that userspace expects to replace or clear 9908 * @flags: xdp-related flags 9909 * 9910 * Set or clear a bpf program for a device 9911 */ 9912 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 9913 int fd, int expected_fd, u32 flags) 9914 { 9915 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 9916 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 9917 int err; 9918 9919 ASSERT_RTNL(); 9920 9921 if (fd >= 0) { 9922 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 9923 mode != XDP_MODE_SKB); 9924 if (IS_ERR(new_prog)) 9925 return PTR_ERR(new_prog); 9926 } 9927 9928 if (expected_fd >= 0) { 9929 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 9930 mode != XDP_MODE_SKB); 9931 if (IS_ERR(old_prog)) { 9932 err = PTR_ERR(old_prog); 9933 old_prog = NULL; 9934 goto err_out; 9935 } 9936 } 9937 9938 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 9939 9940 err_out: 9941 if (err && new_prog) 9942 bpf_prog_put(new_prog); 9943 if (old_prog) 9944 bpf_prog_put(old_prog); 9945 return err; 9946 } 9947 9948 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 9949 { 9950 int i; 9951 9952 ASSERT_RTNL(); 9953 9954 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 9955 if (dev->_rx[i].mp_params.mp_priv) 9956 /* The channel count is the idx plus 1. */ 9957 return i + 1; 9958 9959 return 0; 9960 } 9961 9962 /** 9963 * dev_index_reserve() - allocate an ifindex in a namespace 9964 * @net: the applicable net namespace 9965 * @ifindex: requested ifindex, pass %0 to get one allocated 9966 * 9967 * Allocate a ifindex for a new device. Caller must either use the ifindex 9968 * to store the device (via list_netdevice()) or call dev_index_release() 9969 * to give the index up. 9970 * 9971 * Return: a suitable unique value for a new device interface number or -errno. 9972 */ 9973 static int dev_index_reserve(struct net *net, u32 ifindex) 9974 { 9975 int err; 9976 9977 if (ifindex > INT_MAX) { 9978 DEBUG_NET_WARN_ON_ONCE(1); 9979 return -EINVAL; 9980 } 9981 9982 if (!ifindex) 9983 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 9984 xa_limit_31b, &net->ifindex, GFP_KERNEL); 9985 else 9986 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 9987 if (err < 0) 9988 return err; 9989 9990 return ifindex; 9991 } 9992 9993 static void dev_index_release(struct net *net, int ifindex) 9994 { 9995 /* Expect only unused indexes, unlist_netdevice() removes the used */ 9996 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 9997 } 9998 9999 /* Delayed registration/unregisteration */ 10000 LIST_HEAD(net_todo_list); 10001 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10002 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10003 10004 static void net_set_todo(struct net_device *dev) 10005 { 10006 list_add_tail(&dev->todo_list, &net_todo_list); 10007 } 10008 10009 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10010 struct net_device *upper, netdev_features_t features) 10011 { 10012 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10013 netdev_features_t feature; 10014 int feature_bit; 10015 10016 for_each_netdev_feature(upper_disables, feature_bit) { 10017 feature = __NETIF_F_BIT(feature_bit); 10018 if (!(upper->wanted_features & feature) 10019 && (features & feature)) { 10020 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10021 &feature, upper->name); 10022 features &= ~feature; 10023 } 10024 } 10025 10026 return features; 10027 } 10028 10029 static void netdev_sync_lower_features(struct net_device *upper, 10030 struct net_device *lower, netdev_features_t features) 10031 { 10032 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10033 netdev_features_t feature; 10034 int feature_bit; 10035 10036 for_each_netdev_feature(upper_disables, feature_bit) { 10037 feature = __NETIF_F_BIT(feature_bit); 10038 if (!(features & feature) && (lower->features & feature)) { 10039 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10040 &feature, lower->name); 10041 lower->wanted_features &= ~feature; 10042 __netdev_update_features(lower); 10043 10044 if (unlikely(lower->features & feature)) 10045 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10046 &feature, lower->name); 10047 else 10048 netdev_features_change(lower); 10049 } 10050 } 10051 } 10052 10053 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10054 { 10055 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10056 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10057 bool hw_csum = features & NETIF_F_HW_CSUM; 10058 10059 return ip_csum || hw_csum; 10060 } 10061 10062 static netdev_features_t netdev_fix_features(struct net_device *dev, 10063 netdev_features_t features) 10064 { 10065 /* Fix illegal checksum combinations */ 10066 if ((features & NETIF_F_HW_CSUM) && 10067 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10068 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10069 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10070 } 10071 10072 /* TSO requires that SG is present as well. */ 10073 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10074 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10075 features &= ~NETIF_F_ALL_TSO; 10076 } 10077 10078 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10079 !(features & NETIF_F_IP_CSUM)) { 10080 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10081 features &= ~NETIF_F_TSO; 10082 features &= ~NETIF_F_TSO_ECN; 10083 } 10084 10085 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10086 !(features & NETIF_F_IPV6_CSUM)) { 10087 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10088 features &= ~NETIF_F_TSO6; 10089 } 10090 10091 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10092 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10093 features &= ~NETIF_F_TSO_MANGLEID; 10094 10095 /* TSO ECN requires that TSO is present as well. */ 10096 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10097 features &= ~NETIF_F_TSO_ECN; 10098 10099 /* Software GSO depends on SG. */ 10100 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10101 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10102 features &= ~NETIF_F_GSO; 10103 } 10104 10105 /* GSO partial features require GSO partial be set */ 10106 if ((features & dev->gso_partial_features) && 10107 !(features & NETIF_F_GSO_PARTIAL)) { 10108 netdev_dbg(dev, 10109 "Dropping partially supported GSO features since no GSO partial.\n"); 10110 features &= ~dev->gso_partial_features; 10111 } 10112 10113 if (!(features & NETIF_F_RXCSUM)) { 10114 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10115 * successfully merged by hardware must also have the 10116 * checksum verified by hardware. If the user does not 10117 * want to enable RXCSUM, logically, we should disable GRO_HW. 10118 */ 10119 if (features & NETIF_F_GRO_HW) { 10120 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10121 features &= ~NETIF_F_GRO_HW; 10122 } 10123 } 10124 10125 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10126 if (features & NETIF_F_RXFCS) { 10127 if (features & NETIF_F_LRO) { 10128 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10129 features &= ~NETIF_F_LRO; 10130 } 10131 10132 if (features & NETIF_F_GRO_HW) { 10133 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10134 features &= ~NETIF_F_GRO_HW; 10135 } 10136 } 10137 10138 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10139 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10140 features &= ~NETIF_F_LRO; 10141 } 10142 10143 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10144 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10145 features &= ~NETIF_F_HW_TLS_TX; 10146 } 10147 10148 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10149 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10150 features &= ~NETIF_F_HW_TLS_RX; 10151 } 10152 10153 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10154 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10155 features &= ~NETIF_F_GSO_UDP_L4; 10156 } 10157 10158 return features; 10159 } 10160 10161 int __netdev_update_features(struct net_device *dev) 10162 { 10163 struct net_device *upper, *lower; 10164 netdev_features_t features; 10165 struct list_head *iter; 10166 int err = -1; 10167 10168 ASSERT_RTNL(); 10169 10170 features = netdev_get_wanted_features(dev); 10171 10172 if (dev->netdev_ops->ndo_fix_features) 10173 features = dev->netdev_ops->ndo_fix_features(dev, features); 10174 10175 /* driver might be less strict about feature dependencies */ 10176 features = netdev_fix_features(dev, features); 10177 10178 /* some features can't be enabled if they're off on an upper device */ 10179 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10180 features = netdev_sync_upper_features(dev, upper, features); 10181 10182 if (dev->features == features) 10183 goto sync_lower; 10184 10185 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10186 &dev->features, &features); 10187 10188 if (dev->netdev_ops->ndo_set_features) 10189 err = dev->netdev_ops->ndo_set_features(dev, features); 10190 else 10191 err = 0; 10192 10193 if (unlikely(err < 0)) { 10194 netdev_err(dev, 10195 "set_features() failed (%d); wanted %pNF, left %pNF\n", 10196 err, &features, &dev->features); 10197 /* return non-0 since some features might have changed and 10198 * it's better to fire a spurious notification than miss it 10199 */ 10200 return -1; 10201 } 10202 10203 sync_lower: 10204 /* some features must be disabled on lower devices when disabled 10205 * on an upper device (think: bonding master or bridge) 10206 */ 10207 netdev_for_each_lower_dev(dev, lower, iter) 10208 netdev_sync_lower_features(dev, lower, features); 10209 10210 if (!err) { 10211 netdev_features_t diff = features ^ dev->features; 10212 10213 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 10214 /* udp_tunnel_{get,drop}_rx_info both need 10215 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 10216 * device, or they won't do anything. 10217 * Thus we need to update dev->features 10218 * *before* calling udp_tunnel_get_rx_info, 10219 * but *after* calling udp_tunnel_drop_rx_info. 10220 */ 10221 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 10222 dev->features = features; 10223 udp_tunnel_get_rx_info(dev); 10224 } else { 10225 udp_tunnel_drop_rx_info(dev); 10226 } 10227 } 10228 10229 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 10230 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 10231 dev->features = features; 10232 err |= vlan_get_rx_ctag_filter_info(dev); 10233 } else { 10234 vlan_drop_rx_ctag_filter_info(dev); 10235 } 10236 } 10237 10238 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 10239 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 10240 dev->features = features; 10241 err |= vlan_get_rx_stag_filter_info(dev); 10242 } else { 10243 vlan_drop_rx_stag_filter_info(dev); 10244 } 10245 } 10246 10247 dev->features = features; 10248 } 10249 10250 return err < 0 ? 0 : 1; 10251 } 10252 10253 /** 10254 * netdev_update_features - recalculate device features 10255 * @dev: the device to check 10256 * 10257 * Recalculate dev->features set and send notifications if it 10258 * has changed. Should be called after driver or hardware dependent 10259 * conditions might have changed that influence the features. 10260 */ 10261 void netdev_update_features(struct net_device *dev) 10262 { 10263 if (__netdev_update_features(dev)) 10264 netdev_features_change(dev); 10265 } 10266 EXPORT_SYMBOL(netdev_update_features); 10267 10268 /** 10269 * netdev_change_features - recalculate device features 10270 * @dev: the device to check 10271 * 10272 * Recalculate dev->features set and send notifications even 10273 * if they have not changed. Should be called instead of 10274 * netdev_update_features() if also dev->vlan_features might 10275 * have changed to allow the changes to be propagated to stacked 10276 * VLAN devices. 10277 */ 10278 void netdev_change_features(struct net_device *dev) 10279 { 10280 __netdev_update_features(dev); 10281 netdev_features_change(dev); 10282 } 10283 EXPORT_SYMBOL(netdev_change_features); 10284 10285 /** 10286 * netif_stacked_transfer_operstate - transfer operstate 10287 * @rootdev: the root or lower level device to transfer state from 10288 * @dev: the device to transfer operstate to 10289 * 10290 * Transfer operational state from root to device. This is normally 10291 * called when a stacking relationship exists between the root 10292 * device and the device(a leaf device). 10293 */ 10294 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 10295 struct net_device *dev) 10296 { 10297 if (rootdev->operstate == IF_OPER_DORMANT) 10298 netif_dormant_on(dev); 10299 else 10300 netif_dormant_off(dev); 10301 10302 if (rootdev->operstate == IF_OPER_TESTING) 10303 netif_testing_on(dev); 10304 else 10305 netif_testing_off(dev); 10306 10307 if (netif_carrier_ok(rootdev)) 10308 netif_carrier_on(dev); 10309 else 10310 netif_carrier_off(dev); 10311 } 10312 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10313 10314 static int netif_alloc_rx_queues(struct net_device *dev) 10315 { 10316 unsigned int i, count = dev->num_rx_queues; 10317 struct netdev_rx_queue *rx; 10318 size_t sz = count * sizeof(*rx); 10319 int err = 0; 10320 10321 BUG_ON(count < 1); 10322 10323 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10324 if (!rx) 10325 return -ENOMEM; 10326 10327 dev->_rx = rx; 10328 10329 for (i = 0; i < count; i++) { 10330 rx[i].dev = dev; 10331 10332 /* XDP RX-queue setup */ 10333 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10334 if (err < 0) 10335 goto err_rxq_info; 10336 } 10337 return 0; 10338 10339 err_rxq_info: 10340 /* Rollback successful reg's and free other resources */ 10341 while (i--) 10342 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10343 kvfree(dev->_rx); 10344 dev->_rx = NULL; 10345 return err; 10346 } 10347 10348 static void netif_free_rx_queues(struct net_device *dev) 10349 { 10350 unsigned int i, count = dev->num_rx_queues; 10351 10352 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10353 if (!dev->_rx) 10354 return; 10355 10356 for (i = 0; i < count; i++) 10357 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10358 10359 kvfree(dev->_rx); 10360 } 10361 10362 static void netdev_init_one_queue(struct net_device *dev, 10363 struct netdev_queue *queue, void *_unused) 10364 { 10365 /* Initialize queue lock */ 10366 spin_lock_init(&queue->_xmit_lock); 10367 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10368 queue->xmit_lock_owner = -1; 10369 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10370 queue->dev = dev; 10371 #ifdef CONFIG_BQL 10372 dql_init(&queue->dql, HZ); 10373 #endif 10374 } 10375 10376 static void netif_free_tx_queues(struct net_device *dev) 10377 { 10378 kvfree(dev->_tx); 10379 } 10380 10381 static int netif_alloc_netdev_queues(struct net_device *dev) 10382 { 10383 unsigned int count = dev->num_tx_queues; 10384 struct netdev_queue *tx; 10385 size_t sz = count * sizeof(*tx); 10386 10387 if (count < 1 || count > 0xffff) 10388 return -EINVAL; 10389 10390 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10391 if (!tx) 10392 return -ENOMEM; 10393 10394 dev->_tx = tx; 10395 10396 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10397 spin_lock_init(&dev->tx_global_lock); 10398 10399 return 0; 10400 } 10401 10402 void netif_tx_stop_all_queues(struct net_device *dev) 10403 { 10404 unsigned int i; 10405 10406 for (i = 0; i < dev->num_tx_queues; i++) { 10407 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10408 10409 netif_tx_stop_queue(txq); 10410 } 10411 } 10412 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10413 10414 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10415 { 10416 void __percpu *v; 10417 10418 /* Drivers implementing ndo_get_peer_dev must support tstat 10419 * accounting, so that skb_do_redirect() can bump the dev's 10420 * RX stats upon network namespace switch. 10421 */ 10422 if (dev->netdev_ops->ndo_get_peer_dev && 10423 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10424 return -EOPNOTSUPP; 10425 10426 switch (dev->pcpu_stat_type) { 10427 case NETDEV_PCPU_STAT_NONE: 10428 return 0; 10429 case NETDEV_PCPU_STAT_LSTATS: 10430 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10431 break; 10432 case NETDEV_PCPU_STAT_TSTATS: 10433 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10434 break; 10435 case NETDEV_PCPU_STAT_DSTATS: 10436 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10437 break; 10438 default: 10439 return -EINVAL; 10440 } 10441 10442 return v ? 0 : -ENOMEM; 10443 } 10444 10445 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10446 { 10447 switch (dev->pcpu_stat_type) { 10448 case NETDEV_PCPU_STAT_NONE: 10449 return; 10450 case NETDEV_PCPU_STAT_LSTATS: 10451 free_percpu(dev->lstats); 10452 break; 10453 case NETDEV_PCPU_STAT_TSTATS: 10454 free_percpu(dev->tstats); 10455 break; 10456 case NETDEV_PCPU_STAT_DSTATS: 10457 free_percpu(dev->dstats); 10458 break; 10459 } 10460 } 10461 10462 static void netdev_free_phy_link_topology(struct net_device *dev) 10463 { 10464 struct phy_link_topology *topo = dev->link_topo; 10465 10466 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 10467 xa_destroy(&topo->phys); 10468 kfree(topo); 10469 dev->link_topo = NULL; 10470 } 10471 } 10472 10473 /** 10474 * register_netdevice() - register a network device 10475 * @dev: device to register 10476 * 10477 * Take a prepared network device structure and make it externally accessible. 10478 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10479 * Callers must hold the rtnl lock - you may want register_netdev() 10480 * instead of this. 10481 */ 10482 int register_netdevice(struct net_device *dev) 10483 { 10484 int ret; 10485 struct net *net = dev_net(dev); 10486 10487 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10488 NETDEV_FEATURE_COUNT); 10489 BUG_ON(dev_boot_phase); 10490 ASSERT_RTNL(); 10491 10492 might_sleep(); 10493 10494 /* When net_device's are persistent, this will be fatal. */ 10495 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10496 BUG_ON(!net); 10497 10498 ret = ethtool_check_ops(dev->ethtool_ops); 10499 if (ret) 10500 return ret; 10501 10502 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 10503 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 10504 mutex_init(&dev->ethtool->rss_lock); 10505 10506 spin_lock_init(&dev->addr_list_lock); 10507 netdev_set_addr_lockdep_class(dev); 10508 10509 ret = dev_get_valid_name(net, dev, dev->name); 10510 if (ret < 0) 10511 goto out; 10512 10513 ret = -ENOMEM; 10514 dev->name_node = netdev_name_node_head_alloc(dev); 10515 if (!dev->name_node) 10516 goto out; 10517 10518 /* Init, if this function is available */ 10519 if (dev->netdev_ops->ndo_init) { 10520 ret = dev->netdev_ops->ndo_init(dev); 10521 if (ret) { 10522 if (ret > 0) 10523 ret = -EIO; 10524 goto err_free_name; 10525 } 10526 } 10527 10528 if (((dev->hw_features | dev->features) & 10529 NETIF_F_HW_VLAN_CTAG_FILTER) && 10530 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 10531 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 10532 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 10533 ret = -EINVAL; 10534 goto err_uninit; 10535 } 10536 10537 ret = netdev_do_alloc_pcpu_stats(dev); 10538 if (ret) 10539 goto err_uninit; 10540 10541 ret = dev_index_reserve(net, dev->ifindex); 10542 if (ret < 0) 10543 goto err_free_pcpu; 10544 dev->ifindex = ret; 10545 10546 /* Transfer changeable features to wanted_features and enable 10547 * software offloads (GSO and GRO). 10548 */ 10549 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 10550 dev->features |= NETIF_F_SOFT_FEATURES; 10551 10552 if (dev->udp_tunnel_nic_info) { 10553 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10554 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10555 } 10556 10557 dev->wanted_features = dev->features & dev->hw_features; 10558 10559 if (!(dev->flags & IFF_LOOPBACK)) 10560 dev->hw_features |= NETIF_F_NOCACHE_COPY; 10561 10562 /* If IPv4 TCP segmentation offload is supported we should also 10563 * allow the device to enable segmenting the frame with the option 10564 * of ignoring a static IP ID value. This doesn't enable the 10565 * feature itself but allows the user to enable it later. 10566 */ 10567 if (dev->hw_features & NETIF_F_TSO) 10568 dev->hw_features |= NETIF_F_TSO_MANGLEID; 10569 if (dev->vlan_features & NETIF_F_TSO) 10570 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 10571 if (dev->mpls_features & NETIF_F_TSO) 10572 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 10573 if (dev->hw_enc_features & NETIF_F_TSO) 10574 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 10575 10576 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 10577 */ 10578 dev->vlan_features |= NETIF_F_HIGHDMA; 10579 10580 /* Make NETIF_F_SG inheritable to tunnel devices. 10581 */ 10582 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 10583 10584 /* Make NETIF_F_SG inheritable to MPLS. 10585 */ 10586 dev->mpls_features |= NETIF_F_SG; 10587 10588 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 10589 ret = notifier_to_errno(ret); 10590 if (ret) 10591 goto err_ifindex_release; 10592 10593 ret = netdev_register_kobject(dev); 10594 10595 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 10596 10597 if (ret) 10598 goto err_uninit_notify; 10599 10600 __netdev_update_features(dev); 10601 10602 /* 10603 * Default initial state at registry is that the 10604 * device is present. 10605 */ 10606 10607 set_bit(__LINK_STATE_PRESENT, &dev->state); 10608 10609 linkwatch_init_dev(dev); 10610 10611 dev_init_scheduler(dev); 10612 10613 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 10614 list_netdevice(dev); 10615 10616 add_device_randomness(dev->dev_addr, dev->addr_len); 10617 10618 /* If the device has permanent device address, driver should 10619 * set dev_addr and also addr_assign_type should be set to 10620 * NET_ADDR_PERM (default value). 10621 */ 10622 if (dev->addr_assign_type == NET_ADDR_PERM) 10623 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 10624 10625 /* Notify protocols, that a new device appeared. */ 10626 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 10627 ret = notifier_to_errno(ret); 10628 if (ret) { 10629 /* Expect explicit free_netdev() on failure */ 10630 dev->needs_free_netdev = false; 10631 unregister_netdevice_queue(dev, NULL); 10632 goto out; 10633 } 10634 /* 10635 * Prevent userspace races by waiting until the network 10636 * device is fully setup before sending notifications. 10637 */ 10638 if (!dev->rtnl_link_ops || 10639 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 10640 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 10641 10642 out: 10643 return ret; 10644 10645 err_uninit_notify: 10646 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 10647 err_ifindex_release: 10648 dev_index_release(net, dev->ifindex); 10649 err_free_pcpu: 10650 netdev_do_free_pcpu_stats(dev); 10651 err_uninit: 10652 if (dev->netdev_ops->ndo_uninit) 10653 dev->netdev_ops->ndo_uninit(dev); 10654 if (dev->priv_destructor) 10655 dev->priv_destructor(dev); 10656 err_free_name: 10657 netdev_name_node_free(dev->name_node); 10658 goto out; 10659 } 10660 EXPORT_SYMBOL(register_netdevice); 10661 10662 /* Initialize the core of a dummy net device. 10663 * This is useful if you are calling this function after alloc_netdev(), 10664 * since it does not memset the net_device fields. 10665 */ 10666 static void init_dummy_netdev_core(struct net_device *dev) 10667 { 10668 /* make sure we BUG if trying to hit standard 10669 * register/unregister code path 10670 */ 10671 dev->reg_state = NETREG_DUMMY; 10672 10673 /* NAPI wants this */ 10674 INIT_LIST_HEAD(&dev->napi_list); 10675 10676 /* a dummy interface is started by default */ 10677 set_bit(__LINK_STATE_PRESENT, &dev->state); 10678 set_bit(__LINK_STATE_START, &dev->state); 10679 10680 /* napi_busy_loop stats accounting wants this */ 10681 dev_net_set(dev, &init_net); 10682 10683 /* Note : We dont allocate pcpu_refcnt for dummy devices, 10684 * because users of this 'device' dont need to change 10685 * its refcount. 10686 */ 10687 } 10688 10689 /** 10690 * init_dummy_netdev - init a dummy network device for NAPI 10691 * @dev: device to init 10692 * 10693 * This takes a network device structure and initializes the minimum 10694 * amount of fields so it can be used to schedule NAPI polls without 10695 * registering a full blown interface. This is to be used by drivers 10696 * that need to tie several hardware interfaces to a single NAPI 10697 * poll scheduler due to HW limitations. 10698 */ 10699 void init_dummy_netdev(struct net_device *dev) 10700 { 10701 /* Clear everything. Note we don't initialize spinlocks 10702 * as they aren't supposed to be taken by any of the 10703 * NAPI code and this dummy netdev is supposed to be 10704 * only ever used for NAPI polls 10705 */ 10706 memset(dev, 0, sizeof(struct net_device)); 10707 init_dummy_netdev_core(dev); 10708 } 10709 EXPORT_SYMBOL_GPL(init_dummy_netdev); 10710 10711 /** 10712 * register_netdev - register a network device 10713 * @dev: device to register 10714 * 10715 * Take a completed network device structure and add it to the kernel 10716 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 10717 * chain. 0 is returned on success. A negative errno code is returned 10718 * on a failure to set up the device, or if the name is a duplicate. 10719 * 10720 * This is a wrapper around register_netdevice that takes the rtnl semaphore 10721 * and expands the device name if you passed a format string to 10722 * alloc_netdev. 10723 */ 10724 int register_netdev(struct net_device *dev) 10725 { 10726 int err; 10727 10728 if (rtnl_lock_killable()) 10729 return -EINTR; 10730 err = register_netdevice(dev); 10731 rtnl_unlock(); 10732 return err; 10733 } 10734 EXPORT_SYMBOL(register_netdev); 10735 10736 int netdev_refcnt_read(const struct net_device *dev) 10737 { 10738 #ifdef CONFIG_PCPU_DEV_REFCNT 10739 int i, refcnt = 0; 10740 10741 for_each_possible_cpu(i) 10742 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 10743 return refcnt; 10744 #else 10745 return refcount_read(&dev->dev_refcnt); 10746 #endif 10747 } 10748 EXPORT_SYMBOL(netdev_refcnt_read); 10749 10750 int netdev_unregister_timeout_secs __read_mostly = 10; 10751 10752 #define WAIT_REFS_MIN_MSECS 1 10753 #define WAIT_REFS_MAX_MSECS 250 10754 /** 10755 * netdev_wait_allrefs_any - wait until all references are gone. 10756 * @list: list of net_devices to wait on 10757 * 10758 * This is called when unregistering network devices. 10759 * 10760 * Any protocol or device that holds a reference should register 10761 * for netdevice notification, and cleanup and put back the 10762 * reference if they receive an UNREGISTER event. 10763 * We can get stuck here if buggy protocols don't correctly 10764 * call dev_put. 10765 */ 10766 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 10767 { 10768 unsigned long rebroadcast_time, warning_time; 10769 struct net_device *dev; 10770 int wait = 0; 10771 10772 rebroadcast_time = warning_time = jiffies; 10773 10774 list_for_each_entry(dev, list, todo_list) 10775 if (netdev_refcnt_read(dev) == 1) 10776 return dev; 10777 10778 while (true) { 10779 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 10780 rtnl_lock(); 10781 10782 /* Rebroadcast unregister notification */ 10783 list_for_each_entry(dev, list, todo_list) 10784 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 10785 10786 __rtnl_unlock(); 10787 rcu_barrier(); 10788 rtnl_lock(); 10789 10790 list_for_each_entry(dev, list, todo_list) 10791 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 10792 &dev->state)) { 10793 /* We must not have linkwatch events 10794 * pending on unregister. If this 10795 * happens, we simply run the queue 10796 * unscheduled, resulting in a noop 10797 * for this device. 10798 */ 10799 linkwatch_run_queue(); 10800 break; 10801 } 10802 10803 __rtnl_unlock(); 10804 10805 rebroadcast_time = jiffies; 10806 } 10807 10808 rcu_barrier(); 10809 10810 if (!wait) { 10811 wait = WAIT_REFS_MIN_MSECS; 10812 } else { 10813 msleep(wait); 10814 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 10815 } 10816 10817 list_for_each_entry(dev, list, todo_list) 10818 if (netdev_refcnt_read(dev) == 1) 10819 return dev; 10820 10821 if (time_after(jiffies, warning_time + 10822 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 10823 list_for_each_entry(dev, list, todo_list) { 10824 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 10825 dev->name, netdev_refcnt_read(dev)); 10826 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 10827 } 10828 10829 warning_time = jiffies; 10830 } 10831 } 10832 } 10833 10834 /* The sequence is: 10835 * 10836 * rtnl_lock(); 10837 * ... 10838 * register_netdevice(x1); 10839 * register_netdevice(x2); 10840 * ... 10841 * unregister_netdevice(y1); 10842 * unregister_netdevice(y2); 10843 * ... 10844 * rtnl_unlock(); 10845 * free_netdev(y1); 10846 * free_netdev(y2); 10847 * 10848 * We are invoked by rtnl_unlock(). 10849 * This allows us to deal with problems: 10850 * 1) We can delete sysfs objects which invoke hotplug 10851 * without deadlocking with linkwatch via keventd. 10852 * 2) Since we run with the RTNL semaphore not held, we can sleep 10853 * safely in order to wait for the netdev refcnt to drop to zero. 10854 * 10855 * We must not return until all unregister events added during 10856 * the interval the lock was held have been completed. 10857 */ 10858 void netdev_run_todo(void) 10859 { 10860 struct net_device *dev, *tmp; 10861 struct list_head list; 10862 int cnt; 10863 #ifdef CONFIG_LOCKDEP 10864 struct list_head unlink_list; 10865 10866 list_replace_init(&net_unlink_list, &unlink_list); 10867 10868 while (!list_empty(&unlink_list)) { 10869 struct net_device *dev = list_first_entry(&unlink_list, 10870 struct net_device, 10871 unlink_list); 10872 list_del_init(&dev->unlink_list); 10873 dev->nested_level = dev->lower_level - 1; 10874 } 10875 #endif 10876 10877 /* Snapshot list, allow later requests */ 10878 list_replace_init(&net_todo_list, &list); 10879 10880 __rtnl_unlock(); 10881 10882 /* Wait for rcu callbacks to finish before next phase */ 10883 if (!list_empty(&list)) 10884 rcu_barrier(); 10885 10886 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 10887 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 10888 netdev_WARN(dev, "run_todo but not unregistering\n"); 10889 list_del(&dev->todo_list); 10890 continue; 10891 } 10892 10893 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 10894 linkwatch_sync_dev(dev); 10895 } 10896 10897 cnt = 0; 10898 while (!list_empty(&list)) { 10899 dev = netdev_wait_allrefs_any(&list); 10900 list_del(&dev->todo_list); 10901 10902 /* paranoia */ 10903 BUG_ON(netdev_refcnt_read(dev) != 1); 10904 BUG_ON(!list_empty(&dev->ptype_all)); 10905 BUG_ON(!list_empty(&dev->ptype_specific)); 10906 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 10907 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 10908 10909 netdev_do_free_pcpu_stats(dev); 10910 if (dev->priv_destructor) 10911 dev->priv_destructor(dev); 10912 if (dev->needs_free_netdev) 10913 free_netdev(dev); 10914 10915 cnt++; 10916 10917 /* Free network device */ 10918 kobject_put(&dev->dev.kobj); 10919 } 10920 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 10921 wake_up(&netdev_unregistering_wq); 10922 } 10923 10924 /* Collate per-cpu network dstats statistics 10925 * 10926 * Read per-cpu network statistics from dev->dstats and populate the related 10927 * fields in @s. 10928 */ 10929 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 10930 const struct pcpu_dstats __percpu *dstats) 10931 { 10932 int cpu; 10933 10934 for_each_possible_cpu(cpu) { 10935 u64 rx_packets, rx_bytes, rx_drops; 10936 u64 tx_packets, tx_bytes, tx_drops; 10937 const struct pcpu_dstats *stats; 10938 unsigned int start; 10939 10940 stats = per_cpu_ptr(dstats, cpu); 10941 do { 10942 start = u64_stats_fetch_begin(&stats->syncp); 10943 rx_packets = u64_stats_read(&stats->rx_packets); 10944 rx_bytes = u64_stats_read(&stats->rx_bytes); 10945 rx_drops = u64_stats_read(&stats->rx_drops); 10946 tx_packets = u64_stats_read(&stats->tx_packets); 10947 tx_bytes = u64_stats_read(&stats->tx_bytes); 10948 tx_drops = u64_stats_read(&stats->tx_drops); 10949 } while (u64_stats_fetch_retry(&stats->syncp, start)); 10950 10951 s->rx_packets += rx_packets; 10952 s->rx_bytes += rx_bytes; 10953 s->rx_dropped += rx_drops; 10954 s->tx_packets += tx_packets; 10955 s->tx_bytes += tx_bytes; 10956 s->tx_dropped += tx_drops; 10957 } 10958 } 10959 10960 /* ndo_get_stats64 implementation for dtstats-based accounting. 10961 * 10962 * Populate @s from dev->stats and dev->dstats. This is used internally by the 10963 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 10964 */ 10965 static void dev_get_dstats64(const struct net_device *dev, 10966 struct rtnl_link_stats64 *s) 10967 { 10968 netdev_stats_to_stats64(s, &dev->stats); 10969 dev_fetch_dstats(s, dev->dstats); 10970 } 10971 10972 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 10973 * all the same fields in the same order as net_device_stats, with only 10974 * the type differing, but rtnl_link_stats64 may have additional fields 10975 * at the end for newer counters. 10976 */ 10977 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 10978 const struct net_device_stats *netdev_stats) 10979 { 10980 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 10981 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 10982 u64 *dst = (u64 *)stats64; 10983 10984 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 10985 for (i = 0; i < n; i++) 10986 dst[i] = (unsigned long)atomic_long_read(&src[i]); 10987 /* zero out counters that only exist in rtnl_link_stats64 */ 10988 memset((char *)stats64 + n * sizeof(u64), 0, 10989 sizeof(*stats64) - n * sizeof(u64)); 10990 } 10991 EXPORT_SYMBOL(netdev_stats_to_stats64); 10992 10993 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 10994 struct net_device *dev) 10995 { 10996 struct net_device_core_stats __percpu *p; 10997 10998 p = alloc_percpu_gfp(struct net_device_core_stats, 10999 GFP_ATOMIC | __GFP_NOWARN); 11000 11001 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11002 free_percpu(p); 11003 11004 /* This READ_ONCE() pairs with the cmpxchg() above */ 11005 return READ_ONCE(dev->core_stats); 11006 } 11007 11008 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11009 { 11010 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11011 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11012 unsigned long __percpu *field; 11013 11014 if (unlikely(!p)) { 11015 p = netdev_core_stats_alloc(dev); 11016 if (!p) 11017 return; 11018 } 11019 11020 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11021 this_cpu_inc(*field); 11022 } 11023 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11024 11025 /** 11026 * dev_get_stats - get network device statistics 11027 * @dev: device to get statistics from 11028 * @storage: place to store stats 11029 * 11030 * Get network statistics from device. Return @storage. 11031 * The device driver may provide its own method by setting 11032 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11033 * otherwise the internal statistics structure is used. 11034 */ 11035 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11036 struct rtnl_link_stats64 *storage) 11037 { 11038 const struct net_device_ops *ops = dev->netdev_ops; 11039 const struct net_device_core_stats __percpu *p; 11040 11041 if (ops->ndo_get_stats64) { 11042 memset(storage, 0, sizeof(*storage)); 11043 ops->ndo_get_stats64(dev, storage); 11044 } else if (ops->ndo_get_stats) { 11045 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11046 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11047 dev_get_tstats64(dev, storage); 11048 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11049 dev_get_dstats64(dev, storage); 11050 } else { 11051 netdev_stats_to_stats64(storage, &dev->stats); 11052 } 11053 11054 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11055 p = READ_ONCE(dev->core_stats); 11056 if (p) { 11057 const struct net_device_core_stats *core_stats; 11058 int i; 11059 11060 for_each_possible_cpu(i) { 11061 core_stats = per_cpu_ptr(p, i); 11062 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11063 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11064 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11065 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11066 } 11067 } 11068 return storage; 11069 } 11070 EXPORT_SYMBOL(dev_get_stats); 11071 11072 /** 11073 * dev_fetch_sw_netstats - get per-cpu network device statistics 11074 * @s: place to store stats 11075 * @netstats: per-cpu network stats to read from 11076 * 11077 * Read per-cpu network statistics and populate the related fields in @s. 11078 */ 11079 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11080 const struct pcpu_sw_netstats __percpu *netstats) 11081 { 11082 int cpu; 11083 11084 for_each_possible_cpu(cpu) { 11085 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11086 const struct pcpu_sw_netstats *stats; 11087 unsigned int start; 11088 11089 stats = per_cpu_ptr(netstats, cpu); 11090 do { 11091 start = u64_stats_fetch_begin(&stats->syncp); 11092 rx_packets = u64_stats_read(&stats->rx_packets); 11093 rx_bytes = u64_stats_read(&stats->rx_bytes); 11094 tx_packets = u64_stats_read(&stats->tx_packets); 11095 tx_bytes = u64_stats_read(&stats->tx_bytes); 11096 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11097 11098 s->rx_packets += rx_packets; 11099 s->rx_bytes += rx_bytes; 11100 s->tx_packets += tx_packets; 11101 s->tx_bytes += tx_bytes; 11102 } 11103 } 11104 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11105 11106 /** 11107 * dev_get_tstats64 - ndo_get_stats64 implementation 11108 * @dev: device to get statistics from 11109 * @s: place to store stats 11110 * 11111 * Populate @s from dev->stats and dev->tstats. Can be used as 11112 * ndo_get_stats64() callback. 11113 */ 11114 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11115 { 11116 netdev_stats_to_stats64(s, &dev->stats); 11117 dev_fetch_sw_netstats(s, dev->tstats); 11118 } 11119 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11120 11121 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11122 { 11123 struct netdev_queue *queue = dev_ingress_queue(dev); 11124 11125 #ifdef CONFIG_NET_CLS_ACT 11126 if (queue) 11127 return queue; 11128 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11129 if (!queue) 11130 return NULL; 11131 netdev_init_one_queue(dev, queue, NULL); 11132 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11133 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11134 rcu_assign_pointer(dev->ingress_queue, queue); 11135 #endif 11136 return queue; 11137 } 11138 11139 static const struct ethtool_ops default_ethtool_ops; 11140 11141 void netdev_set_default_ethtool_ops(struct net_device *dev, 11142 const struct ethtool_ops *ops) 11143 { 11144 if (dev->ethtool_ops == &default_ethtool_ops) 11145 dev->ethtool_ops = ops; 11146 } 11147 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11148 11149 /** 11150 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11151 * @dev: netdev to enable the IRQ coalescing on 11152 * 11153 * Sets a conservative default for SW IRQ coalescing. Users can use 11154 * sysfs attributes to override the default values. 11155 */ 11156 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11157 { 11158 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11159 11160 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11161 netdev_set_gro_flush_timeout(dev, 20000); 11162 netdev_set_defer_hard_irqs(dev, 1); 11163 } 11164 } 11165 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11166 11167 /** 11168 * alloc_netdev_mqs - allocate network device 11169 * @sizeof_priv: size of private data to allocate space for 11170 * @name: device name format string 11171 * @name_assign_type: origin of device name 11172 * @setup: callback to initialize device 11173 * @txqs: the number of TX subqueues to allocate 11174 * @rxqs: the number of RX subqueues to allocate 11175 * 11176 * Allocates a struct net_device with private data area for driver use 11177 * and performs basic initialization. Also allocates subqueue structs 11178 * for each queue on the device. 11179 */ 11180 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11181 unsigned char name_assign_type, 11182 void (*setup)(struct net_device *), 11183 unsigned int txqs, unsigned int rxqs) 11184 { 11185 struct net_device *dev; 11186 size_t napi_config_sz; 11187 unsigned int maxqs; 11188 11189 BUG_ON(strlen(name) >= sizeof(dev->name)); 11190 11191 if (txqs < 1) { 11192 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 11193 return NULL; 11194 } 11195 11196 if (rxqs < 1) { 11197 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 11198 return NULL; 11199 } 11200 11201 maxqs = max(txqs, rxqs); 11202 11203 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 11204 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11205 if (!dev) 11206 return NULL; 11207 11208 dev->priv_len = sizeof_priv; 11209 11210 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 11211 #ifdef CONFIG_PCPU_DEV_REFCNT 11212 dev->pcpu_refcnt = alloc_percpu(int); 11213 if (!dev->pcpu_refcnt) 11214 goto free_dev; 11215 __dev_hold(dev); 11216 #else 11217 refcount_set(&dev->dev_refcnt, 1); 11218 #endif 11219 11220 if (dev_addr_init(dev)) 11221 goto free_pcpu; 11222 11223 dev_mc_init(dev); 11224 dev_uc_init(dev); 11225 11226 dev_net_set(dev, &init_net); 11227 11228 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 11229 dev->xdp_zc_max_segs = 1; 11230 dev->gso_max_segs = GSO_MAX_SEGS; 11231 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 11232 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 11233 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 11234 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 11235 dev->tso_max_segs = TSO_MAX_SEGS; 11236 dev->upper_level = 1; 11237 dev->lower_level = 1; 11238 #ifdef CONFIG_LOCKDEP 11239 dev->nested_level = 0; 11240 INIT_LIST_HEAD(&dev->unlink_list); 11241 #endif 11242 11243 INIT_LIST_HEAD(&dev->napi_list); 11244 INIT_LIST_HEAD(&dev->unreg_list); 11245 INIT_LIST_HEAD(&dev->close_list); 11246 INIT_LIST_HEAD(&dev->link_watch_list); 11247 INIT_LIST_HEAD(&dev->adj_list.upper); 11248 INIT_LIST_HEAD(&dev->adj_list.lower); 11249 INIT_LIST_HEAD(&dev->ptype_all); 11250 INIT_LIST_HEAD(&dev->ptype_specific); 11251 INIT_LIST_HEAD(&dev->net_notifier_list); 11252 #ifdef CONFIG_NET_SCHED 11253 hash_init(dev->qdisc_hash); 11254 #endif 11255 11256 mutex_init(&dev->lock); 11257 11258 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 11259 setup(dev); 11260 11261 if (!dev->tx_queue_len) { 11262 dev->priv_flags |= IFF_NO_QUEUE; 11263 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 11264 } 11265 11266 dev->num_tx_queues = txqs; 11267 dev->real_num_tx_queues = txqs; 11268 if (netif_alloc_netdev_queues(dev)) 11269 goto free_all; 11270 11271 dev->num_rx_queues = rxqs; 11272 dev->real_num_rx_queues = rxqs; 11273 if (netif_alloc_rx_queues(dev)) 11274 goto free_all; 11275 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 11276 if (!dev->ethtool) 11277 goto free_all; 11278 11279 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 11280 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 11281 if (!dev->napi_config) 11282 goto free_all; 11283 11284 strscpy(dev->name, name); 11285 dev->name_assign_type = name_assign_type; 11286 dev->group = INIT_NETDEV_GROUP; 11287 if (!dev->ethtool_ops) 11288 dev->ethtool_ops = &default_ethtool_ops; 11289 11290 nf_hook_netdev_init(dev); 11291 11292 return dev; 11293 11294 free_all: 11295 free_netdev(dev); 11296 return NULL; 11297 11298 free_pcpu: 11299 #ifdef CONFIG_PCPU_DEV_REFCNT 11300 free_percpu(dev->pcpu_refcnt); 11301 free_dev: 11302 #endif 11303 kvfree(dev); 11304 return NULL; 11305 } 11306 EXPORT_SYMBOL(alloc_netdev_mqs); 11307 11308 /** 11309 * free_netdev - free network device 11310 * @dev: device 11311 * 11312 * This function does the last stage of destroying an allocated device 11313 * interface. The reference to the device object is released. If this 11314 * is the last reference then it will be freed.Must be called in process 11315 * context. 11316 */ 11317 void free_netdev(struct net_device *dev) 11318 { 11319 struct napi_struct *p, *n; 11320 11321 might_sleep(); 11322 11323 /* When called immediately after register_netdevice() failed the unwind 11324 * handling may still be dismantling the device. Handle that case by 11325 * deferring the free. 11326 */ 11327 if (dev->reg_state == NETREG_UNREGISTERING) { 11328 ASSERT_RTNL(); 11329 dev->needs_free_netdev = true; 11330 return; 11331 } 11332 11333 mutex_destroy(&dev->lock); 11334 11335 kfree(dev->ethtool); 11336 netif_free_tx_queues(dev); 11337 netif_free_rx_queues(dev); 11338 11339 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 11340 11341 /* Flush device addresses */ 11342 dev_addr_flush(dev); 11343 11344 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 11345 netif_napi_del(p); 11346 11347 kvfree(dev->napi_config); 11348 11349 ref_tracker_dir_exit(&dev->refcnt_tracker); 11350 #ifdef CONFIG_PCPU_DEV_REFCNT 11351 free_percpu(dev->pcpu_refcnt); 11352 dev->pcpu_refcnt = NULL; 11353 #endif 11354 free_percpu(dev->core_stats); 11355 dev->core_stats = NULL; 11356 free_percpu(dev->xdp_bulkq); 11357 dev->xdp_bulkq = NULL; 11358 11359 netdev_free_phy_link_topology(dev); 11360 11361 /* Compatibility with error handling in drivers */ 11362 if (dev->reg_state == NETREG_UNINITIALIZED || 11363 dev->reg_state == NETREG_DUMMY) { 11364 kvfree(dev); 11365 return; 11366 } 11367 11368 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 11369 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 11370 11371 /* will free via device release */ 11372 put_device(&dev->dev); 11373 } 11374 EXPORT_SYMBOL(free_netdev); 11375 11376 /** 11377 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 11378 * @sizeof_priv: size of private data to allocate space for 11379 * 11380 * Return: the allocated net_device on success, NULL otherwise 11381 */ 11382 struct net_device *alloc_netdev_dummy(int sizeof_priv) 11383 { 11384 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 11385 init_dummy_netdev_core); 11386 } 11387 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 11388 11389 /** 11390 * synchronize_net - Synchronize with packet receive processing 11391 * 11392 * Wait for packets currently being received to be done. 11393 * Does not block later packets from starting. 11394 */ 11395 void synchronize_net(void) 11396 { 11397 might_sleep(); 11398 if (rtnl_is_locked()) 11399 synchronize_rcu_expedited(); 11400 else 11401 synchronize_rcu(); 11402 } 11403 EXPORT_SYMBOL(synchronize_net); 11404 11405 static void netdev_rss_contexts_free(struct net_device *dev) 11406 { 11407 struct ethtool_rxfh_context *ctx; 11408 unsigned long context; 11409 11410 mutex_lock(&dev->ethtool->rss_lock); 11411 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 11412 struct ethtool_rxfh_param rxfh; 11413 11414 rxfh.indir = ethtool_rxfh_context_indir(ctx); 11415 rxfh.key = ethtool_rxfh_context_key(ctx); 11416 rxfh.hfunc = ctx->hfunc; 11417 rxfh.input_xfrm = ctx->input_xfrm; 11418 rxfh.rss_context = context; 11419 rxfh.rss_delete = true; 11420 11421 xa_erase(&dev->ethtool->rss_ctx, context); 11422 if (dev->ethtool_ops->create_rxfh_context) 11423 dev->ethtool_ops->remove_rxfh_context(dev, ctx, 11424 context, NULL); 11425 else 11426 dev->ethtool_ops->set_rxfh(dev, &rxfh, NULL); 11427 kfree(ctx); 11428 } 11429 xa_destroy(&dev->ethtool->rss_ctx); 11430 mutex_unlock(&dev->ethtool->rss_lock); 11431 } 11432 11433 /** 11434 * unregister_netdevice_queue - remove device from the kernel 11435 * @dev: device 11436 * @head: list 11437 * 11438 * This function shuts down a device interface and removes it 11439 * from the kernel tables. 11440 * If head not NULL, device is queued to be unregistered later. 11441 * 11442 * Callers must hold the rtnl semaphore. You may want 11443 * unregister_netdev() instead of this. 11444 */ 11445 11446 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11447 { 11448 ASSERT_RTNL(); 11449 11450 if (head) { 11451 list_move_tail(&dev->unreg_list, head); 11452 } else { 11453 LIST_HEAD(single); 11454 11455 list_add(&dev->unreg_list, &single); 11456 unregister_netdevice_many(&single); 11457 } 11458 } 11459 EXPORT_SYMBOL(unregister_netdevice_queue); 11460 11461 void unregister_netdevice_many_notify(struct list_head *head, 11462 u32 portid, const struct nlmsghdr *nlh) 11463 { 11464 struct net_device *dev, *tmp; 11465 LIST_HEAD(close_head); 11466 int cnt = 0; 11467 11468 BUG_ON(dev_boot_phase); 11469 ASSERT_RTNL(); 11470 11471 if (list_empty(head)) 11472 return; 11473 11474 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11475 /* Some devices call without registering 11476 * for initialization unwind. Remove those 11477 * devices and proceed with the remaining. 11478 */ 11479 if (dev->reg_state == NETREG_UNINITIALIZED) { 11480 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11481 dev->name, dev); 11482 11483 WARN_ON(1); 11484 list_del(&dev->unreg_list); 11485 continue; 11486 } 11487 dev->dismantle = true; 11488 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11489 } 11490 11491 /* If device is running, close it first. */ 11492 list_for_each_entry(dev, head, unreg_list) 11493 list_add_tail(&dev->close_list, &close_head); 11494 dev_close_many(&close_head, true); 11495 11496 list_for_each_entry(dev, head, unreg_list) { 11497 /* And unlink it from device chain. */ 11498 unlist_netdevice(dev); 11499 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 11500 } 11501 flush_all_backlogs(); 11502 11503 synchronize_net(); 11504 11505 list_for_each_entry(dev, head, unreg_list) { 11506 struct sk_buff *skb = NULL; 11507 11508 /* Shutdown queueing discipline. */ 11509 dev_shutdown(dev); 11510 dev_tcx_uninstall(dev); 11511 dev_xdp_uninstall(dev); 11512 bpf_dev_bound_netdev_unregister(dev); 11513 dev_dmabuf_uninstall(dev); 11514 11515 netdev_offload_xstats_disable_all(dev); 11516 11517 /* Notify protocols, that we are about to destroy 11518 * this device. They should clean all the things. 11519 */ 11520 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11521 11522 if (!dev->rtnl_link_ops || 11523 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11524 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 11525 GFP_KERNEL, NULL, 0, 11526 portid, nlh); 11527 11528 /* 11529 * Flush the unicast and multicast chains 11530 */ 11531 dev_uc_flush(dev); 11532 dev_mc_flush(dev); 11533 11534 netdev_name_node_alt_flush(dev); 11535 netdev_name_node_free(dev->name_node); 11536 11537 netdev_rss_contexts_free(dev); 11538 11539 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11540 11541 if (dev->netdev_ops->ndo_uninit) 11542 dev->netdev_ops->ndo_uninit(dev); 11543 11544 mutex_destroy(&dev->ethtool->rss_lock); 11545 11546 net_shaper_flush_netdev(dev); 11547 11548 if (skb) 11549 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 11550 11551 /* Notifier chain MUST detach us all upper devices. */ 11552 WARN_ON(netdev_has_any_upper_dev(dev)); 11553 WARN_ON(netdev_has_any_lower_dev(dev)); 11554 11555 /* Remove entries from kobject tree */ 11556 netdev_unregister_kobject(dev); 11557 #ifdef CONFIG_XPS 11558 /* Remove XPS queueing entries */ 11559 netif_reset_xps_queues_gt(dev, 0); 11560 #endif 11561 } 11562 11563 synchronize_net(); 11564 11565 list_for_each_entry(dev, head, unreg_list) { 11566 netdev_put(dev, &dev->dev_registered_tracker); 11567 net_set_todo(dev); 11568 cnt++; 11569 } 11570 atomic_add(cnt, &dev_unreg_count); 11571 11572 list_del(head); 11573 } 11574 11575 /** 11576 * unregister_netdevice_many - unregister many devices 11577 * @head: list of devices 11578 * 11579 * Note: As most callers use a stack allocated list_head, 11580 * we force a list_del() to make sure stack won't be corrupted later. 11581 */ 11582 void unregister_netdevice_many(struct list_head *head) 11583 { 11584 unregister_netdevice_many_notify(head, 0, NULL); 11585 } 11586 EXPORT_SYMBOL(unregister_netdevice_many); 11587 11588 /** 11589 * unregister_netdev - remove device from the kernel 11590 * @dev: device 11591 * 11592 * This function shuts down a device interface and removes it 11593 * from the kernel tables. 11594 * 11595 * This is just a wrapper for unregister_netdevice that takes 11596 * the rtnl semaphore. In general you want to use this and not 11597 * unregister_netdevice. 11598 */ 11599 void unregister_netdev(struct net_device *dev) 11600 { 11601 rtnl_lock(); 11602 unregister_netdevice(dev); 11603 rtnl_unlock(); 11604 } 11605 EXPORT_SYMBOL(unregister_netdev); 11606 11607 /** 11608 * __dev_change_net_namespace - move device to different nethost namespace 11609 * @dev: device 11610 * @net: network namespace 11611 * @pat: If not NULL name pattern to try if the current device name 11612 * is already taken in the destination network namespace. 11613 * @new_ifindex: If not zero, specifies device index in the target 11614 * namespace. 11615 * 11616 * This function shuts down a device interface and moves it 11617 * to a new network namespace. On success 0 is returned, on 11618 * a failure a netagive errno code is returned. 11619 * 11620 * Callers must hold the rtnl semaphore. 11621 */ 11622 11623 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 11624 const char *pat, int new_ifindex) 11625 { 11626 struct netdev_name_node *name_node; 11627 struct net *net_old = dev_net(dev); 11628 char new_name[IFNAMSIZ] = {}; 11629 int err, new_nsid; 11630 11631 ASSERT_RTNL(); 11632 11633 /* Don't allow namespace local devices to be moved. */ 11634 err = -EINVAL; 11635 if (dev->netns_local) 11636 goto out; 11637 11638 /* Ensure the device has been registered */ 11639 if (dev->reg_state != NETREG_REGISTERED) 11640 goto out; 11641 11642 /* Get out if there is nothing todo */ 11643 err = 0; 11644 if (net_eq(net_old, net)) 11645 goto out; 11646 11647 /* Pick the destination device name, and ensure 11648 * we can use it in the destination network namespace. 11649 */ 11650 err = -EEXIST; 11651 if (netdev_name_in_use(net, dev->name)) { 11652 /* We get here if we can't use the current device name */ 11653 if (!pat) 11654 goto out; 11655 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 11656 if (err < 0) 11657 goto out; 11658 } 11659 /* Check that none of the altnames conflicts. */ 11660 err = -EEXIST; 11661 netdev_for_each_altname(dev, name_node) 11662 if (netdev_name_in_use(net, name_node->name)) 11663 goto out; 11664 11665 /* Check that new_ifindex isn't used yet. */ 11666 if (new_ifindex) { 11667 err = dev_index_reserve(net, new_ifindex); 11668 if (err < 0) 11669 goto out; 11670 } else { 11671 /* If there is an ifindex conflict assign a new one */ 11672 err = dev_index_reserve(net, dev->ifindex); 11673 if (err == -EBUSY) 11674 err = dev_index_reserve(net, 0); 11675 if (err < 0) 11676 goto out; 11677 new_ifindex = err; 11678 } 11679 11680 /* 11681 * And now a mini version of register_netdevice unregister_netdevice. 11682 */ 11683 11684 /* If device is running close it first. */ 11685 dev_close(dev); 11686 11687 /* And unlink it from device chain */ 11688 unlist_netdevice(dev); 11689 11690 synchronize_net(); 11691 11692 /* Shutdown queueing discipline. */ 11693 dev_shutdown(dev); 11694 11695 /* Notify protocols, that we are about to destroy 11696 * this device. They should clean all the things. 11697 * 11698 * Note that dev->reg_state stays at NETREG_REGISTERED. 11699 * This is wanted because this way 8021q and macvlan know 11700 * the device is just moving and can keep their slaves up. 11701 */ 11702 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11703 rcu_barrier(); 11704 11705 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 11706 11707 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 11708 new_ifindex); 11709 11710 /* 11711 * Flush the unicast and multicast chains 11712 */ 11713 dev_uc_flush(dev); 11714 dev_mc_flush(dev); 11715 11716 /* Send a netdev-removed uevent to the old namespace */ 11717 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 11718 netdev_adjacent_del_links(dev); 11719 11720 /* Move per-net netdevice notifiers that are following the netdevice */ 11721 move_netdevice_notifiers_dev_net(dev, net); 11722 11723 /* Actually switch the network namespace */ 11724 dev_net_set(dev, net); 11725 dev->ifindex = new_ifindex; 11726 11727 if (new_name[0]) { 11728 /* Rename the netdev to prepared name */ 11729 write_seqlock_bh(&netdev_rename_lock); 11730 strscpy(dev->name, new_name, IFNAMSIZ); 11731 write_sequnlock_bh(&netdev_rename_lock); 11732 } 11733 11734 /* Fixup kobjects */ 11735 dev_set_uevent_suppress(&dev->dev, 1); 11736 err = device_rename(&dev->dev, dev->name); 11737 dev_set_uevent_suppress(&dev->dev, 0); 11738 WARN_ON(err); 11739 11740 /* Send a netdev-add uevent to the new namespace */ 11741 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 11742 netdev_adjacent_add_links(dev); 11743 11744 /* Adapt owner in case owning user namespace of target network 11745 * namespace is different from the original one. 11746 */ 11747 err = netdev_change_owner(dev, net_old, net); 11748 WARN_ON(err); 11749 11750 /* Add the device back in the hashes */ 11751 list_netdevice(dev); 11752 11753 /* Notify protocols, that a new device appeared. */ 11754 call_netdevice_notifiers(NETDEV_REGISTER, dev); 11755 11756 /* 11757 * Prevent userspace races by waiting until the network 11758 * device is fully setup before sending notifications. 11759 */ 11760 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11761 11762 synchronize_net(); 11763 err = 0; 11764 out: 11765 return err; 11766 } 11767 EXPORT_SYMBOL_GPL(__dev_change_net_namespace); 11768 11769 static int dev_cpu_dead(unsigned int oldcpu) 11770 { 11771 struct sk_buff **list_skb; 11772 struct sk_buff *skb; 11773 unsigned int cpu; 11774 struct softnet_data *sd, *oldsd, *remsd = NULL; 11775 11776 local_irq_disable(); 11777 cpu = smp_processor_id(); 11778 sd = &per_cpu(softnet_data, cpu); 11779 oldsd = &per_cpu(softnet_data, oldcpu); 11780 11781 /* Find end of our completion_queue. */ 11782 list_skb = &sd->completion_queue; 11783 while (*list_skb) 11784 list_skb = &(*list_skb)->next; 11785 /* Append completion queue from offline CPU. */ 11786 *list_skb = oldsd->completion_queue; 11787 oldsd->completion_queue = NULL; 11788 11789 /* Append output queue from offline CPU. */ 11790 if (oldsd->output_queue) { 11791 *sd->output_queue_tailp = oldsd->output_queue; 11792 sd->output_queue_tailp = oldsd->output_queue_tailp; 11793 oldsd->output_queue = NULL; 11794 oldsd->output_queue_tailp = &oldsd->output_queue; 11795 } 11796 /* Append NAPI poll list from offline CPU, with one exception : 11797 * process_backlog() must be called by cpu owning percpu backlog. 11798 * We properly handle process_queue & input_pkt_queue later. 11799 */ 11800 while (!list_empty(&oldsd->poll_list)) { 11801 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 11802 struct napi_struct, 11803 poll_list); 11804 11805 list_del_init(&napi->poll_list); 11806 if (napi->poll == process_backlog) 11807 napi->state &= NAPIF_STATE_THREADED; 11808 else 11809 ____napi_schedule(sd, napi); 11810 } 11811 11812 raise_softirq_irqoff(NET_TX_SOFTIRQ); 11813 local_irq_enable(); 11814 11815 if (!use_backlog_threads()) { 11816 #ifdef CONFIG_RPS 11817 remsd = oldsd->rps_ipi_list; 11818 oldsd->rps_ipi_list = NULL; 11819 #endif 11820 /* send out pending IPI's on offline CPU */ 11821 net_rps_send_ipi(remsd); 11822 } 11823 11824 /* Process offline CPU's input_pkt_queue */ 11825 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 11826 netif_rx(skb); 11827 rps_input_queue_head_incr(oldsd); 11828 } 11829 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 11830 netif_rx(skb); 11831 rps_input_queue_head_incr(oldsd); 11832 } 11833 11834 return 0; 11835 } 11836 11837 /** 11838 * netdev_increment_features - increment feature set by one 11839 * @all: current feature set 11840 * @one: new feature set 11841 * @mask: mask feature set 11842 * 11843 * Computes a new feature set after adding a device with feature set 11844 * @one to the master device with current feature set @all. Will not 11845 * enable anything that is off in @mask. Returns the new feature set. 11846 */ 11847 netdev_features_t netdev_increment_features(netdev_features_t all, 11848 netdev_features_t one, netdev_features_t mask) 11849 { 11850 if (mask & NETIF_F_HW_CSUM) 11851 mask |= NETIF_F_CSUM_MASK; 11852 mask |= NETIF_F_VLAN_CHALLENGED; 11853 11854 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 11855 all &= one | ~NETIF_F_ALL_FOR_ALL; 11856 11857 /* If one device supports hw checksumming, set for all. */ 11858 if (all & NETIF_F_HW_CSUM) 11859 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 11860 11861 return all; 11862 } 11863 EXPORT_SYMBOL(netdev_increment_features); 11864 11865 static struct hlist_head * __net_init netdev_create_hash(void) 11866 { 11867 int i; 11868 struct hlist_head *hash; 11869 11870 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 11871 if (hash != NULL) 11872 for (i = 0; i < NETDEV_HASHENTRIES; i++) 11873 INIT_HLIST_HEAD(&hash[i]); 11874 11875 return hash; 11876 } 11877 11878 /* Initialize per network namespace state */ 11879 static int __net_init netdev_init(struct net *net) 11880 { 11881 BUILD_BUG_ON(GRO_HASH_BUCKETS > 11882 8 * sizeof_field(struct napi_struct, gro_bitmask)); 11883 11884 INIT_LIST_HEAD(&net->dev_base_head); 11885 11886 net->dev_name_head = netdev_create_hash(); 11887 if (net->dev_name_head == NULL) 11888 goto err_name; 11889 11890 net->dev_index_head = netdev_create_hash(); 11891 if (net->dev_index_head == NULL) 11892 goto err_idx; 11893 11894 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 11895 11896 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 11897 11898 return 0; 11899 11900 err_idx: 11901 kfree(net->dev_name_head); 11902 err_name: 11903 return -ENOMEM; 11904 } 11905 11906 /** 11907 * netdev_drivername - network driver for the device 11908 * @dev: network device 11909 * 11910 * Determine network driver for device. 11911 */ 11912 const char *netdev_drivername(const struct net_device *dev) 11913 { 11914 const struct device_driver *driver; 11915 const struct device *parent; 11916 const char *empty = ""; 11917 11918 parent = dev->dev.parent; 11919 if (!parent) 11920 return empty; 11921 11922 driver = parent->driver; 11923 if (driver && driver->name) 11924 return driver->name; 11925 return empty; 11926 } 11927 11928 static void __netdev_printk(const char *level, const struct net_device *dev, 11929 struct va_format *vaf) 11930 { 11931 if (dev && dev->dev.parent) { 11932 dev_printk_emit(level[1] - '0', 11933 dev->dev.parent, 11934 "%s %s %s%s: %pV", 11935 dev_driver_string(dev->dev.parent), 11936 dev_name(dev->dev.parent), 11937 netdev_name(dev), netdev_reg_state(dev), 11938 vaf); 11939 } else if (dev) { 11940 printk("%s%s%s: %pV", 11941 level, netdev_name(dev), netdev_reg_state(dev), vaf); 11942 } else { 11943 printk("%s(NULL net_device): %pV", level, vaf); 11944 } 11945 } 11946 11947 void netdev_printk(const char *level, const struct net_device *dev, 11948 const char *format, ...) 11949 { 11950 struct va_format vaf; 11951 va_list args; 11952 11953 va_start(args, format); 11954 11955 vaf.fmt = format; 11956 vaf.va = &args; 11957 11958 __netdev_printk(level, dev, &vaf); 11959 11960 va_end(args); 11961 } 11962 EXPORT_SYMBOL(netdev_printk); 11963 11964 #define define_netdev_printk_level(func, level) \ 11965 void func(const struct net_device *dev, const char *fmt, ...) \ 11966 { \ 11967 struct va_format vaf; \ 11968 va_list args; \ 11969 \ 11970 va_start(args, fmt); \ 11971 \ 11972 vaf.fmt = fmt; \ 11973 vaf.va = &args; \ 11974 \ 11975 __netdev_printk(level, dev, &vaf); \ 11976 \ 11977 va_end(args); \ 11978 } \ 11979 EXPORT_SYMBOL(func); 11980 11981 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 11982 define_netdev_printk_level(netdev_alert, KERN_ALERT); 11983 define_netdev_printk_level(netdev_crit, KERN_CRIT); 11984 define_netdev_printk_level(netdev_err, KERN_ERR); 11985 define_netdev_printk_level(netdev_warn, KERN_WARNING); 11986 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 11987 define_netdev_printk_level(netdev_info, KERN_INFO); 11988 11989 static void __net_exit netdev_exit(struct net *net) 11990 { 11991 kfree(net->dev_name_head); 11992 kfree(net->dev_index_head); 11993 xa_destroy(&net->dev_by_index); 11994 if (net != &init_net) 11995 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 11996 } 11997 11998 static struct pernet_operations __net_initdata netdev_net_ops = { 11999 .init = netdev_init, 12000 .exit = netdev_exit, 12001 }; 12002 12003 static void __net_exit default_device_exit_net(struct net *net) 12004 { 12005 struct netdev_name_node *name_node, *tmp; 12006 struct net_device *dev, *aux; 12007 /* 12008 * Push all migratable network devices back to the 12009 * initial network namespace 12010 */ 12011 ASSERT_RTNL(); 12012 for_each_netdev_safe(net, dev, aux) { 12013 int err; 12014 char fb_name[IFNAMSIZ]; 12015 12016 /* Ignore unmoveable devices (i.e. loopback) */ 12017 if (dev->netns_local) 12018 continue; 12019 12020 /* Leave virtual devices for the generic cleanup */ 12021 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 12022 continue; 12023 12024 /* Push remaining network devices to init_net */ 12025 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 12026 if (netdev_name_in_use(&init_net, fb_name)) 12027 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 12028 12029 netdev_for_each_altname_safe(dev, name_node, tmp) 12030 if (netdev_name_in_use(&init_net, name_node->name)) 12031 __netdev_name_node_alt_destroy(name_node); 12032 12033 err = dev_change_net_namespace(dev, &init_net, fb_name); 12034 if (err) { 12035 pr_emerg("%s: failed to move %s to init_net: %d\n", 12036 __func__, dev->name, err); 12037 BUG(); 12038 } 12039 } 12040 } 12041 12042 static void __net_exit default_device_exit_batch(struct list_head *net_list) 12043 { 12044 /* At exit all network devices most be removed from a network 12045 * namespace. Do this in the reverse order of registration. 12046 * Do this across as many network namespaces as possible to 12047 * improve batching efficiency. 12048 */ 12049 struct net_device *dev; 12050 struct net *net; 12051 LIST_HEAD(dev_kill_list); 12052 12053 rtnl_lock(); 12054 list_for_each_entry(net, net_list, exit_list) { 12055 default_device_exit_net(net); 12056 cond_resched(); 12057 } 12058 12059 list_for_each_entry(net, net_list, exit_list) { 12060 for_each_netdev_reverse(net, dev) { 12061 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 12062 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 12063 else 12064 unregister_netdevice_queue(dev, &dev_kill_list); 12065 } 12066 } 12067 unregister_netdevice_many(&dev_kill_list); 12068 rtnl_unlock(); 12069 } 12070 12071 static struct pernet_operations __net_initdata default_device_ops = { 12072 .exit_batch = default_device_exit_batch, 12073 }; 12074 12075 static void __init net_dev_struct_check(void) 12076 { 12077 /* TX read-mostly hotpath */ 12078 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 12079 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 12080 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 12081 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 12082 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 12083 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 12084 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 12085 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 12086 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 12087 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 12088 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 12089 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 12090 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 12091 #ifdef CONFIG_XPS 12092 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 12093 #endif 12094 #ifdef CONFIG_NETFILTER_EGRESS 12095 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 12096 #endif 12097 #ifdef CONFIG_NET_XGRESS 12098 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 12099 #endif 12100 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 12101 12102 /* TXRX read-mostly hotpath */ 12103 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 12104 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 12105 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 12106 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 12107 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 12108 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 12109 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 12110 12111 /* RX read-mostly hotpath */ 12112 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 12113 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 12114 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 12115 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 12116 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 12117 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 12118 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 12119 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 12120 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 12121 #ifdef CONFIG_NETPOLL 12122 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 12123 #endif 12124 #ifdef CONFIG_NET_XGRESS 12125 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 12126 #endif 12127 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 12128 } 12129 12130 /* 12131 * Initialize the DEV module. At boot time this walks the device list and 12132 * unhooks any devices that fail to initialise (normally hardware not 12133 * present) and leaves us with a valid list of present and active devices. 12134 * 12135 */ 12136 12137 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 12138 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 12139 12140 static int net_page_pool_create(int cpuid) 12141 { 12142 #if IS_ENABLED(CONFIG_PAGE_POOL) 12143 struct page_pool_params page_pool_params = { 12144 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 12145 .flags = PP_FLAG_SYSTEM_POOL, 12146 .nid = cpu_to_mem(cpuid), 12147 }; 12148 struct page_pool *pp_ptr; 12149 12150 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 12151 if (IS_ERR(pp_ptr)) 12152 return -ENOMEM; 12153 12154 per_cpu(system_page_pool, cpuid) = pp_ptr; 12155 #endif 12156 return 0; 12157 } 12158 12159 static int backlog_napi_should_run(unsigned int cpu) 12160 { 12161 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12162 struct napi_struct *napi = &sd->backlog; 12163 12164 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 12165 } 12166 12167 static void run_backlog_napi(unsigned int cpu) 12168 { 12169 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12170 12171 napi_threaded_poll_loop(&sd->backlog); 12172 } 12173 12174 static void backlog_napi_setup(unsigned int cpu) 12175 { 12176 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12177 struct napi_struct *napi = &sd->backlog; 12178 12179 napi->thread = this_cpu_read(backlog_napi); 12180 set_bit(NAPI_STATE_THREADED, &napi->state); 12181 } 12182 12183 static struct smp_hotplug_thread backlog_threads = { 12184 .store = &backlog_napi, 12185 .thread_should_run = backlog_napi_should_run, 12186 .thread_fn = run_backlog_napi, 12187 .thread_comm = "backlog_napi/%u", 12188 .setup = backlog_napi_setup, 12189 }; 12190 12191 /* 12192 * This is called single threaded during boot, so no need 12193 * to take the rtnl semaphore. 12194 */ 12195 static int __init net_dev_init(void) 12196 { 12197 int i, rc = -ENOMEM; 12198 12199 BUG_ON(!dev_boot_phase); 12200 12201 net_dev_struct_check(); 12202 12203 if (dev_proc_init()) 12204 goto out; 12205 12206 if (netdev_kobject_init()) 12207 goto out; 12208 12209 for (i = 0; i < PTYPE_HASH_SIZE; i++) 12210 INIT_LIST_HEAD(&ptype_base[i]); 12211 12212 if (register_pernet_subsys(&netdev_net_ops)) 12213 goto out; 12214 12215 /* 12216 * Initialise the packet receive queues. 12217 */ 12218 12219 for_each_possible_cpu(i) { 12220 struct work_struct *flush = per_cpu_ptr(&flush_works, i); 12221 struct softnet_data *sd = &per_cpu(softnet_data, i); 12222 12223 INIT_WORK(flush, flush_backlog); 12224 12225 skb_queue_head_init(&sd->input_pkt_queue); 12226 skb_queue_head_init(&sd->process_queue); 12227 #ifdef CONFIG_XFRM_OFFLOAD 12228 skb_queue_head_init(&sd->xfrm_backlog); 12229 #endif 12230 INIT_LIST_HEAD(&sd->poll_list); 12231 sd->output_queue_tailp = &sd->output_queue; 12232 #ifdef CONFIG_RPS 12233 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 12234 sd->cpu = i; 12235 #endif 12236 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 12237 spin_lock_init(&sd->defer_lock); 12238 12239 init_gro_hash(&sd->backlog); 12240 sd->backlog.poll = process_backlog; 12241 sd->backlog.weight = weight_p; 12242 INIT_LIST_HEAD(&sd->backlog.poll_list); 12243 12244 if (net_page_pool_create(i)) 12245 goto out; 12246 } 12247 if (use_backlog_threads()) 12248 smpboot_register_percpu_thread(&backlog_threads); 12249 12250 dev_boot_phase = 0; 12251 12252 /* The loopback device is special if any other network devices 12253 * is present in a network namespace the loopback device must 12254 * be present. Since we now dynamically allocate and free the 12255 * loopback device ensure this invariant is maintained by 12256 * keeping the loopback device as the first device on the 12257 * list of network devices. Ensuring the loopback devices 12258 * is the first device that appears and the last network device 12259 * that disappears. 12260 */ 12261 if (register_pernet_device(&loopback_net_ops)) 12262 goto out; 12263 12264 if (register_pernet_device(&default_device_ops)) 12265 goto out; 12266 12267 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 12268 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 12269 12270 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 12271 NULL, dev_cpu_dead); 12272 WARN_ON(rc < 0); 12273 rc = 0; 12274 12275 /* avoid static key IPIs to isolated CPUs */ 12276 if (housekeeping_enabled(HK_TYPE_MISC)) 12277 net_enable_timestamp(); 12278 out: 12279 if (rc < 0) { 12280 for_each_possible_cpu(i) { 12281 struct page_pool *pp_ptr; 12282 12283 pp_ptr = per_cpu(system_page_pool, i); 12284 if (!pp_ptr) 12285 continue; 12286 12287 page_pool_destroy(pp_ptr); 12288 per_cpu(system_page_pool, i) = NULL; 12289 } 12290 } 12291 12292 return rc; 12293 } 12294 12295 subsys_initcall(net_dev_init); 12296