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, <[email protected]> 8 * Mark Evans, <[email protected]> 9 * 10 * Additional Authors: 11 * Florian la Roche <[email protected]> 12 * Alan Cox <[email protected]> 13 * David Hinds <[email protected]> 14 * Alexey Kuznetsov <[email protected]> 15 * Adam Sulmicki <[email protected]> 16 * Pekka Riikonen <[email protected]> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/skbuff.h> 96 #include <linux/kthread.h> 97 #include <linux/bpf.h> 98 #include <linux/bpf_trace.h> 99 #include <net/net_namespace.h> 100 #include <net/sock.h> 101 #include <net/busy_poll.h> 102 #include <linux/rtnetlink.h> 103 #include <linux/stat.h> 104 #include <net/dsa.h> 105 #include <net/dst.h> 106 #include <net/dst_metadata.h> 107 #include <net/gro.h> 108 #include <net/pkt_sched.h> 109 #include <net/pkt_cls.h> 110 #include <net/checksum.h> 111 #include <net/xfrm.h> 112 #include <net/tcx.h> 113 #include <linux/highmem.h> 114 #include <linux/init.h> 115 #include <linux/module.h> 116 #include <linux/netpoll.h> 117 #include <linux/rcupdate.h> 118 #include <linux/delay.h> 119 #include <net/iw_handler.h> 120 #include <asm/current.h> 121 #include <linux/audit.h> 122 #include <linux/dmaengine.h> 123 #include <linux/err.h> 124 #include <linux/ctype.h> 125 #include <linux/if_arp.h> 126 #include <linux/if_vlan.h> 127 #include <linux/ip.h> 128 #include <net/ip.h> 129 #include <net/mpls.h> 130 #include <linux/ipv6.h> 131 #include <linux/in.h> 132 #include <linux/jhash.h> 133 #include <linux/random.h> 134 #include <trace/events/napi.h> 135 #include <trace/events/net.h> 136 #include <trace/events/skb.h> 137 #include <trace/events/qdisc.h> 138 #include <trace/events/xdp.h> 139 #include <linux/inetdevice.h> 140 #include <linux/cpu_rmap.h> 141 #include <linux/static_key.h> 142 #include <linux/hashtable.h> 143 #include <linux/vmalloc.h> 144 #include <linux/if_macvlan.h> 145 #include <linux/errqueue.h> 146 #include <linux/hrtimer.h> 147 #include <linux/netfilter_netdev.h> 148 #include <linux/crash_dump.h> 149 #include <linux/sctp.h> 150 #include <net/udp_tunnel.h> 151 #include <linux/net_namespace.h> 152 #include <linux/indirect_call_wrapper.h> 153 #include <net/devlink.h> 154 #include <linux/pm_runtime.h> 155 #include <linux/prandom.h> 156 #include <linux/once_lite.h> 157 #include <net/netdev_rx_queue.h> 158 #include <net/page_pool/types.h> 159 #include <net/page_pool/helpers.h> 160 #include <net/rps.h> 161 162 #include "dev.h" 163 #include "net-sysfs.h" 164 165 static DEFINE_SPINLOCK(ptype_lock); 166 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 167 168 static int netif_rx_internal(struct sk_buff *skb); 169 static int call_netdevice_notifiers_extack(unsigned long val, 170 struct net_device *dev, 171 struct netlink_ext_ack *extack); 172 173 static DEFINE_MUTEX(ifalias_mutex); 174 175 /* protects napi_hash addition/deletion and napi_gen_id */ 176 static DEFINE_SPINLOCK(napi_hash_lock); 177 178 static unsigned int napi_gen_id = NR_CPUS; 179 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 180 181 static DECLARE_RWSEM(devnet_rename_sem); 182 183 static inline void dev_base_seq_inc(struct net *net) 184 { 185 unsigned int val = net->dev_base_seq + 1; 186 187 WRITE_ONCE(net->dev_base_seq, val ?: 1); 188 } 189 190 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 191 { 192 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 193 194 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 195 } 196 197 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 198 { 199 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 200 } 201 202 #ifndef CONFIG_PREEMPT_RT 203 204 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 205 206 static int __init setup_backlog_napi_threads(char *arg) 207 { 208 static_branch_enable(&use_backlog_threads_key); 209 return 0; 210 } 211 early_param("thread_backlog_napi", setup_backlog_napi_threads); 212 213 static bool use_backlog_threads(void) 214 { 215 return static_branch_unlikely(&use_backlog_threads_key); 216 } 217 218 #else 219 220 static bool use_backlog_threads(void) 221 { 222 return true; 223 } 224 225 #endif 226 227 static inline void backlog_lock_irq_save(struct softnet_data *sd, 228 unsigned long *flags) 229 { 230 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 231 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 232 else 233 local_irq_save(*flags); 234 } 235 236 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 237 { 238 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 239 spin_lock_irq(&sd->input_pkt_queue.lock); 240 else 241 local_irq_disable(); 242 } 243 244 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 245 unsigned long *flags) 246 { 247 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 248 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags); 249 else 250 local_irq_restore(*flags); 251 } 252 253 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 254 { 255 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 256 spin_unlock_irq(&sd->input_pkt_queue.lock); 257 else 258 local_irq_enable(); 259 } 260 261 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 262 const char *name) 263 { 264 struct netdev_name_node *name_node; 265 266 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 267 if (!name_node) 268 return NULL; 269 INIT_HLIST_NODE(&name_node->hlist); 270 name_node->dev = dev; 271 name_node->name = name; 272 return name_node; 273 } 274 275 static struct netdev_name_node * 276 netdev_name_node_head_alloc(struct net_device *dev) 277 { 278 struct netdev_name_node *name_node; 279 280 name_node = netdev_name_node_alloc(dev, dev->name); 281 if (!name_node) 282 return NULL; 283 INIT_LIST_HEAD(&name_node->list); 284 return name_node; 285 } 286 287 static void netdev_name_node_free(struct netdev_name_node *name_node) 288 { 289 kfree(name_node); 290 } 291 292 static void netdev_name_node_add(struct net *net, 293 struct netdev_name_node *name_node) 294 { 295 hlist_add_head_rcu(&name_node->hlist, 296 dev_name_hash(net, name_node->name)); 297 } 298 299 static void netdev_name_node_del(struct netdev_name_node *name_node) 300 { 301 hlist_del_rcu(&name_node->hlist); 302 } 303 304 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 305 const char *name) 306 { 307 struct hlist_head *head = dev_name_hash(net, name); 308 struct netdev_name_node *name_node; 309 310 hlist_for_each_entry(name_node, head, hlist) 311 if (!strcmp(name_node->name, name)) 312 return name_node; 313 return NULL; 314 } 315 316 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 317 const char *name) 318 { 319 struct hlist_head *head = dev_name_hash(net, name); 320 struct netdev_name_node *name_node; 321 322 hlist_for_each_entry_rcu(name_node, head, hlist) 323 if (!strcmp(name_node->name, name)) 324 return name_node; 325 return NULL; 326 } 327 328 bool netdev_name_in_use(struct net *net, const char *name) 329 { 330 return netdev_name_node_lookup(net, name); 331 } 332 EXPORT_SYMBOL(netdev_name_in_use); 333 334 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 335 { 336 struct netdev_name_node *name_node; 337 struct net *net = dev_net(dev); 338 339 name_node = netdev_name_node_lookup(net, name); 340 if (name_node) 341 return -EEXIST; 342 name_node = netdev_name_node_alloc(dev, name); 343 if (!name_node) 344 return -ENOMEM; 345 netdev_name_node_add(net, name_node); 346 /* The node that holds dev->name acts as a head of per-device list. */ 347 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 348 349 return 0; 350 } 351 352 static void netdev_name_node_alt_free(struct rcu_head *head) 353 { 354 struct netdev_name_node *name_node = 355 container_of(head, struct netdev_name_node, rcu); 356 357 kfree(name_node->name); 358 netdev_name_node_free(name_node); 359 } 360 361 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 362 { 363 netdev_name_node_del(name_node); 364 list_del(&name_node->list); 365 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 366 } 367 368 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 369 { 370 struct netdev_name_node *name_node; 371 struct net *net = dev_net(dev); 372 373 name_node = netdev_name_node_lookup(net, name); 374 if (!name_node) 375 return -ENOENT; 376 /* lookup might have found our primary name or a name belonging 377 * to another device. 378 */ 379 if (name_node == dev->name_node || name_node->dev != dev) 380 return -EINVAL; 381 382 __netdev_name_node_alt_destroy(name_node); 383 return 0; 384 } 385 386 static void netdev_name_node_alt_flush(struct net_device *dev) 387 { 388 struct netdev_name_node *name_node, *tmp; 389 390 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 391 list_del(&name_node->list); 392 netdev_name_node_alt_free(&name_node->rcu); 393 } 394 } 395 396 /* Device list insertion */ 397 static void list_netdevice(struct net_device *dev) 398 { 399 struct netdev_name_node *name_node; 400 struct net *net = dev_net(dev); 401 402 ASSERT_RTNL(); 403 404 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 405 netdev_name_node_add(net, dev->name_node); 406 hlist_add_head_rcu(&dev->index_hlist, 407 dev_index_hash(net, dev->ifindex)); 408 409 netdev_for_each_altname(dev, name_node) 410 netdev_name_node_add(net, name_node); 411 412 /* We reserved the ifindex, this can't fail */ 413 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 414 415 dev_base_seq_inc(net); 416 } 417 418 /* Device list removal 419 * caller must respect a RCU grace period before freeing/reusing dev 420 */ 421 static void unlist_netdevice(struct net_device *dev) 422 { 423 struct netdev_name_node *name_node; 424 struct net *net = dev_net(dev); 425 426 ASSERT_RTNL(); 427 428 xa_erase(&net->dev_by_index, dev->ifindex); 429 430 netdev_for_each_altname(dev, name_node) 431 netdev_name_node_del(name_node); 432 433 /* Unlink dev from the device chain */ 434 list_del_rcu(&dev->dev_list); 435 netdev_name_node_del(dev->name_node); 436 hlist_del_rcu(&dev->index_hlist); 437 438 dev_base_seq_inc(dev_net(dev)); 439 } 440 441 /* 442 * Our notifier list 443 */ 444 445 static RAW_NOTIFIER_HEAD(netdev_chain); 446 447 /* 448 * Device drivers call our routines to queue packets here. We empty the 449 * queue in the local softnet handler. 450 */ 451 452 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 453 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 454 }; 455 EXPORT_PER_CPU_SYMBOL(softnet_data); 456 457 /* Page_pool has a lockless array/stack to alloc/recycle pages. 458 * PP consumers must pay attention to run APIs in the appropriate context 459 * (e.g. NAPI context). 460 */ 461 static DEFINE_PER_CPU(struct page_pool *, system_page_pool); 462 463 #ifdef CONFIG_LOCKDEP 464 /* 465 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 466 * according to dev->type 467 */ 468 static const unsigned short netdev_lock_type[] = { 469 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 470 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 471 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 472 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 473 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 474 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 475 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 476 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 477 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 478 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 479 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 480 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 481 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 482 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 483 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE}; 484 485 static const char *const netdev_lock_name[] = { 486 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 487 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 488 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 489 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 490 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 491 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 492 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 493 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 494 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 495 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 496 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 497 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 498 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 499 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE", 500 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"}; 501 502 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 503 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 504 505 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 506 { 507 int i; 508 509 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 510 if (netdev_lock_type[i] == dev_type) 511 return i; 512 /* the last key is used by default */ 513 return ARRAY_SIZE(netdev_lock_type) - 1; 514 } 515 516 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 517 unsigned short dev_type) 518 { 519 int i; 520 521 i = netdev_lock_pos(dev_type); 522 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 523 netdev_lock_name[i]); 524 } 525 526 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 527 { 528 int i; 529 530 i = netdev_lock_pos(dev->type); 531 lockdep_set_class_and_name(&dev->addr_list_lock, 532 &netdev_addr_lock_key[i], 533 netdev_lock_name[i]); 534 } 535 #else 536 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 537 unsigned short dev_type) 538 { 539 } 540 541 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 542 { 543 } 544 #endif 545 546 /******************************************************************************* 547 * 548 * Protocol management and registration routines 549 * 550 *******************************************************************************/ 551 552 553 /* 554 * Add a protocol ID to the list. Now that the input handler is 555 * smarter we can dispense with all the messy stuff that used to be 556 * here. 557 * 558 * BEWARE!!! Protocol handlers, mangling input packets, 559 * MUST BE last in hash buckets and checking protocol handlers 560 * MUST start from promiscuous ptype_all chain in net_bh. 561 * It is true now, do not change it. 562 * Explanation follows: if protocol handler, mangling packet, will 563 * be the first on list, it is not able to sense, that packet 564 * is cloned and should be copied-on-write, so that it will 565 * change it and subsequent readers will get broken packet. 566 * --ANK (980803) 567 */ 568 569 static inline struct list_head *ptype_head(const struct packet_type *pt) 570 { 571 if (pt->type == htons(ETH_P_ALL)) 572 return pt->dev ? &pt->dev->ptype_all : &net_hotdata.ptype_all; 573 else 574 return pt->dev ? &pt->dev->ptype_specific : 575 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 576 } 577 578 /** 579 * dev_add_pack - add packet handler 580 * @pt: packet type declaration 581 * 582 * Add a protocol handler to the networking stack. The passed &packet_type 583 * is linked into kernel lists and may not be freed until it has been 584 * removed from the kernel lists. 585 * 586 * This call does not sleep therefore it can not 587 * guarantee all CPU's that are in middle of receiving packets 588 * will see the new packet type (until the next received packet). 589 */ 590 591 void dev_add_pack(struct packet_type *pt) 592 { 593 struct list_head *head = ptype_head(pt); 594 595 spin_lock(&ptype_lock); 596 list_add_rcu(&pt->list, head); 597 spin_unlock(&ptype_lock); 598 } 599 EXPORT_SYMBOL(dev_add_pack); 600 601 /** 602 * __dev_remove_pack - remove packet handler 603 * @pt: packet type declaration 604 * 605 * Remove a protocol handler that was previously added to the kernel 606 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 607 * from the kernel lists and can be freed or reused once this function 608 * returns. 609 * 610 * The packet type might still be in use by receivers 611 * and must not be freed until after all the CPU's have gone 612 * through a quiescent state. 613 */ 614 void __dev_remove_pack(struct packet_type *pt) 615 { 616 struct list_head *head = ptype_head(pt); 617 struct packet_type *pt1; 618 619 spin_lock(&ptype_lock); 620 621 list_for_each_entry(pt1, head, list) { 622 if (pt == pt1) { 623 list_del_rcu(&pt->list); 624 goto out; 625 } 626 } 627 628 pr_warn("dev_remove_pack: %p not found\n", pt); 629 out: 630 spin_unlock(&ptype_lock); 631 } 632 EXPORT_SYMBOL(__dev_remove_pack); 633 634 /** 635 * dev_remove_pack - remove packet handler 636 * @pt: packet type declaration 637 * 638 * Remove a protocol handler that was previously added to the kernel 639 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 640 * from the kernel lists and can be freed or reused once this function 641 * returns. 642 * 643 * This call sleeps to guarantee that no CPU is looking at the packet 644 * type after return. 645 */ 646 void dev_remove_pack(struct packet_type *pt) 647 { 648 __dev_remove_pack(pt); 649 650 synchronize_net(); 651 } 652 EXPORT_SYMBOL(dev_remove_pack); 653 654 655 /******************************************************************************* 656 * 657 * Device Interface Subroutines 658 * 659 *******************************************************************************/ 660 661 /** 662 * dev_get_iflink - get 'iflink' value of a interface 663 * @dev: targeted interface 664 * 665 * Indicates the ifindex the interface is linked to. 666 * Physical interfaces have the same 'ifindex' and 'iflink' values. 667 */ 668 669 int dev_get_iflink(const struct net_device *dev) 670 { 671 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 672 return dev->netdev_ops->ndo_get_iflink(dev); 673 674 return READ_ONCE(dev->ifindex); 675 } 676 EXPORT_SYMBOL(dev_get_iflink); 677 678 /** 679 * dev_fill_metadata_dst - Retrieve tunnel egress information. 680 * @dev: targeted interface 681 * @skb: The packet. 682 * 683 * For better visibility of tunnel traffic OVS needs to retrieve 684 * egress tunnel information for a packet. Following API allows 685 * user to get this info. 686 */ 687 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 688 { 689 struct ip_tunnel_info *info; 690 691 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 692 return -EINVAL; 693 694 info = skb_tunnel_info_unclone(skb); 695 if (!info) 696 return -ENOMEM; 697 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 698 return -EINVAL; 699 700 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 701 } 702 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 703 704 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 705 { 706 int k = stack->num_paths++; 707 708 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 709 return NULL; 710 711 return &stack->path[k]; 712 } 713 714 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 715 struct net_device_path_stack *stack) 716 { 717 const struct net_device *last_dev; 718 struct net_device_path_ctx ctx = { 719 .dev = dev, 720 }; 721 struct net_device_path *path; 722 int ret = 0; 723 724 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 725 stack->num_paths = 0; 726 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 727 last_dev = ctx.dev; 728 path = dev_fwd_path(stack); 729 if (!path) 730 return -1; 731 732 memset(path, 0, sizeof(struct net_device_path)); 733 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 734 if (ret < 0) 735 return -1; 736 737 if (WARN_ON_ONCE(last_dev == ctx.dev)) 738 return -1; 739 } 740 741 if (!ctx.dev) 742 return ret; 743 744 path = dev_fwd_path(stack); 745 if (!path) 746 return -1; 747 path->type = DEV_PATH_ETHERNET; 748 path->dev = ctx.dev; 749 750 return ret; 751 } 752 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 753 754 /** 755 * __dev_get_by_name - find a device by its name 756 * @net: the applicable net namespace 757 * @name: name to find 758 * 759 * Find an interface by name. Must be called under RTNL semaphore. 760 * If the name is found a pointer to the device is returned. 761 * If the name is not found then %NULL is returned. The 762 * reference counters are not incremented so the caller must be 763 * careful with locks. 764 */ 765 766 struct net_device *__dev_get_by_name(struct net *net, const char *name) 767 { 768 struct netdev_name_node *node_name; 769 770 node_name = netdev_name_node_lookup(net, name); 771 return node_name ? node_name->dev : NULL; 772 } 773 EXPORT_SYMBOL(__dev_get_by_name); 774 775 /** 776 * dev_get_by_name_rcu - find a device by its name 777 * @net: the applicable net namespace 778 * @name: name to find 779 * 780 * Find an interface by name. 781 * If the name is found a pointer to the device is returned. 782 * If the name is not found then %NULL is returned. 783 * The reference counters are not incremented so the caller must be 784 * careful with locks. The caller must hold RCU lock. 785 */ 786 787 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 788 { 789 struct netdev_name_node *node_name; 790 791 node_name = netdev_name_node_lookup_rcu(net, name); 792 return node_name ? node_name->dev : NULL; 793 } 794 EXPORT_SYMBOL(dev_get_by_name_rcu); 795 796 /* Deprecated for new users, call netdev_get_by_name() instead */ 797 struct net_device *dev_get_by_name(struct net *net, const char *name) 798 { 799 struct net_device *dev; 800 801 rcu_read_lock(); 802 dev = dev_get_by_name_rcu(net, name); 803 dev_hold(dev); 804 rcu_read_unlock(); 805 return dev; 806 } 807 EXPORT_SYMBOL(dev_get_by_name); 808 809 /** 810 * netdev_get_by_name() - find a device by its name 811 * @net: the applicable net namespace 812 * @name: name to find 813 * @tracker: tracking object for the acquired reference 814 * @gfp: allocation flags for the tracker 815 * 816 * Find an interface by name. This can be called from any 817 * context and does its own locking. The returned handle has 818 * the usage count incremented and the caller must use netdev_put() to 819 * release it when it is no longer needed. %NULL is returned if no 820 * matching device is found. 821 */ 822 struct net_device *netdev_get_by_name(struct net *net, const char *name, 823 netdevice_tracker *tracker, gfp_t gfp) 824 { 825 struct net_device *dev; 826 827 dev = dev_get_by_name(net, name); 828 if (dev) 829 netdev_tracker_alloc(dev, tracker, gfp); 830 return dev; 831 } 832 EXPORT_SYMBOL(netdev_get_by_name); 833 834 /** 835 * __dev_get_by_index - find a device by its ifindex 836 * @net: the applicable net namespace 837 * @ifindex: index of device 838 * 839 * Search for an interface by index. Returns %NULL if the device 840 * is not found or a pointer to the device. The device has not 841 * had its reference counter increased so the caller must be careful 842 * about locking. The caller must hold the RTNL semaphore. 843 */ 844 845 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 846 { 847 struct net_device *dev; 848 struct hlist_head *head = dev_index_hash(net, ifindex); 849 850 hlist_for_each_entry(dev, head, index_hlist) 851 if (dev->ifindex == ifindex) 852 return dev; 853 854 return NULL; 855 } 856 EXPORT_SYMBOL(__dev_get_by_index); 857 858 /** 859 * dev_get_by_index_rcu - find a device by its ifindex 860 * @net: the applicable net namespace 861 * @ifindex: index of device 862 * 863 * Search for an interface by index. Returns %NULL if the device 864 * is not found or a pointer to the device. The device has not 865 * had its reference counter increased so the caller must be careful 866 * about locking. The caller must hold RCU lock. 867 */ 868 869 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 870 { 871 struct net_device *dev; 872 struct hlist_head *head = dev_index_hash(net, ifindex); 873 874 hlist_for_each_entry_rcu(dev, head, index_hlist) 875 if (dev->ifindex == ifindex) 876 return dev; 877 878 return NULL; 879 } 880 EXPORT_SYMBOL(dev_get_by_index_rcu); 881 882 /* Deprecated for new users, call netdev_get_by_index() instead */ 883 struct net_device *dev_get_by_index(struct net *net, int ifindex) 884 { 885 struct net_device *dev; 886 887 rcu_read_lock(); 888 dev = dev_get_by_index_rcu(net, ifindex); 889 dev_hold(dev); 890 rcu_read_unlock(); 891 return dev; 892 } 893 EXPORT_SYMBOL(dev_get_by_index); 894 895 /** 896 * netdev_get_by_index() - find a device by its ifindex 897 * @net: the applicable net namespace 898 * @ifindex: index of device 899 * @tracker: tracking object for the acquired reference 900 * @gfp: allocation flags for the tracker 901 * 902 * Search for an interface by index. Returns NULL if the device 903 * is not found or a pointer to the device. The device returned has 904 * had a reference added and the pointer is safe until the user calls 905 * netdev_put() to indicate they have finished with it. 906 */ 907 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 908 netdevice_tracker *tracker, gfp_t gfp) 909 { 910 struct net_device *dev; 911 912 dev = dev_get_by_index(net, ifindex); 913 if (dev) 914 netdev_tracker_alloc(dev, tracker, gfp); 915 return dev; 916 } 917 EXPORT_SYMBOL(netdev_get_by_index); 918 919 /** 920 * dev_get_by_napi_id - find a device by napi_id 921 * @napi_id: ID of the NAPI struct 922 * 923 * Search for an interface by NAPI ID. Returns %NULL if the device 924 * is not found or a pointer to the device. The device has not had 925 * its reference counter increased so the caller must be careful 926 * about locking. The caller must hold RCU lock. 927 */ 928 929 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 930 { 931 struct napi_struct *napi; 932 933 WARN_ON_ONCE(!rcu_read_lock_held()); 934 935 if (napi_id < MIN_NAPI_ID) 936 return NULL; 937 938 napi = napi_by_id(napi_id); 939 940 return napi ? napi->dev : NULL; 941 } 942 EXPORT_SYMBOL(dev_get_by_napi_id); 943 944 static DEFINE_SEQLOCK(netdev_rename_lock); 945 946 void netdev_copy_name(struct net_device *dev, char *name) 947 { 948 unsigned int seq; 949 950 do { 951 seq = read_seqbegin(&netdev_rename_lock); 952 strscpy(name, dev->name, IFNAMSIZ); 953 } while (read_seqretry(&netdev_rename_lock, seq)); 954 } 955 956 /** 957 * netdev_get_name - get a netdevice name, knowing its ifindex. 958 * @net: network namespace 959 * @name: a pointer to the buffer where the name will be stored. 960 * @ifindex: the ifindex of the interface to get the name from. 961 */ 962 int netdev_get_name(struct net *net, char *name, int ifindex) 963 { 964 struct net_device *dev; 965 int ret; 966 967 rcu_read_lock(); 968 969 dev = dev_get_by_index_rcu(net, ifindex); 970 if (!dev) { 971 ret = -ENODEV; 972 goto out; 973 } 974 975 netdev_copy_name(dev, name); 976 977 ret = 0; 978 out: 979 rcu_read_unlock(); 980 return ret; 981 } 982 983 /** 984 * dev_getbyhwaddr_rcu - find a device by its hardware address 985 * @net: the applicable net namespace 986 * @type: media type of device 987 * @ha: hardware address 988 * 989 * Search for an interface by MAC address. Returns NULL if the device 990 * is not found or a pointer to the device. 991 * The caller must hold RCU or RTNL. 992 * The returned device has not had its ref count increased 993 * and the caller must therefore be careful about locking 994 * 995 */ 996 997 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 998 const char *ha) 999 { 1000 struct net_device *dev; 1001 1002 for_each_netdev_rcu(net, dev) 1003 if (dev->type == type && 1004 !memcmp(dev->dev_addr, ha, dev->addr_len)) 1005 return dev; 1006 1007 return NULL; 1008 } 1009 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1010 1011 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1012 { 1013 struct net_device *dev, *ret = NULL; 1014 1015 rcu_read_lock(); 1016 for_each_netdev_rcu(net, dev) 1017 if (dev->type == type) { 1018 dev_hold(dev); 1019 ret = dev; 1020 break; 1021 } 1022 rcu_read_unlock(); 1023 return ret; 1024 } 1025 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1026 1027 /** 1028 * __dev_get_by_flags - find any device with given flags 1029 * @net: the applicable net namespace 1030 * @if_flags: IFF_* values 1031 * @mask: bitmask of bits in if_flags to check 1032 * 1033 * Search for any interface with the given flags. Returns NULL if a device 1034 * is not found or a pointer to the device. Must be called inside 1035 * rtnl_lock(), and result refcount is unchanged. 1036 */ 1037 1038 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags, 1039 unsigned short mask) 1040 { 1041 struct net_device *dev, *ret; 1042 1043 ASSERT_RTNL(); 1044 1045 ret = NULL; 1046 for_each_netdev(net, dev) { 1047 if (((dev->flags ^ if_flags) & mask) == 0) { 1048 ret = dev; 1049 break; 1050 } 1051 } 1052 return ret; 1053 } 1054 EXPORT_SYMBOL(__dev_get_by_flags); 1055 1056 /** 1057 * dev_valid_name - check if name is okay for network device 1058 * @name: name string 1059 * 1060 * Network device names need to be valid file names to 1061 * allow sysfs to work. We also disallow any kind of 1062 * whitespace. 1063 */ 1064 bool dev_valid_name(const char *name) 1065 { 1066 if (*name == '\0') 1067 return false; 1068 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1069 return false; 1070 if (!strcmp(name, ".") || !strcmp(name, "..")) 1071 return false; 1072 1073 while (*name) { 1074 if (*name == '/' || *name == ':' || isspace(*name)) 1075 return false; 1076 name++; 1077 } 1078 return true; 1079 } 1080 EXPORT_SYMBOL(dev_valid_name); 1081 1082 /** 1083 * __dev_alloc_name - allocate a name for a device 1084 * @net: network namespace to allocate the device name in 1085 * @name: name format string 1086 * @res: result name string 1087 * 1088 * Passed a format string - eg "lt%d" it will try and find a suitable 1089 * id. It scans list of devices to build up a free map, then chooses 1090 * the first empty slot. The caller must hold the dev_base or rtnl lock 1091 * while allocating the name and adding the device in order to avoid 1092 * duplicates. 1093 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1094 * Returns the number of the unit assigned or a negative errno code. 1095 */ 1096 1097 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1098 { 1099 int i = 0; 1100 const char *p; 1101 const int max_netdevices = 8*PAGE_SIZE; 1102 unsigned long *inuse; 1103 struct net_device *d; 1104 char buf[IFNAMSIZ]; 1105 1106 /* Verify the string as this thing may have come from the user. 1107 * There must be one "%d" and no other "%" characters. 1108 */ 1109 p = strchr(name, '%'); 1110 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1111 return -EINVAL; 1112 1113 /* Use one page as a bit array of possible slots */ 1114 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1115 if (!inuse) 1116 return -ENOMEM; 1117 1118 for_each_netdev(net, d) { 1119 struct netdev_name_node *name_node; 1120 1121 netdev_for_each_altname(d, name_node) { 1122 if (!sscanf(name_node->name, name, &i)) 1123 continue; 1124 if (i < 0 || i >= max_netdevices) 1125 continue; 1126 1127 /* avoid cases where sscanf is not exact inverse of printf */ 1128 snprintf(buf, IFNAMSIZ, name, i); 1129 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1130 __set_bit(i, inuse); 1131 } 1132 if (!sscanf(d->name, name, &i)) 1133 continue; 1134 if (i < 0 || i >= max_netdevices) 1135 continue; 1136 1137 /* avoid cases where sscanf is not exact inverse of printf */ 1138 snprintf(buf, IFNAMSIZ, name, i); 1139 if (!strncmp(buf, d->name, IFNAMSIZ)) 1140 __set_bit(i, inuse); 1141 } 1142 1143 i = find_first_zero_bit(inuse, max_netdevices); 1144 bitmap_free(inuse); 1145 if (i == max_netdevices) 1146 return -ENFILE; 1147 1148 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1149 strscpy(buf, name, IFNAMSIZ); 1150 snprintf(res, IFNAMSIZ, buf, i); 1151 return i; 1152 } 1153 1154 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1155 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1156 const char *want_name, char *out_name, 1157 int dup_errno) 1158 { 1159 if (!dev_valid_name(want_name)) 1160 return -EINVAL; 1161 1162 if (strchr(want_name, '%')) 1163 return __dev_alloc_name(net, want_name, out_name); 1164 1165 if (netdev_name_in_use(net, want_name)) 1166 return -dup_errno; 1167 if (out_name != want_name) 1168 strscpy(out_name, want_name, IFNAMSIZ); 1169 return 0; 1170 } 1171 1172 /** 1173 * dev_alloc_name - allocate a name for a device 1174 * @dev: device 1175 * @name: name format string 1176 * 1177 * Passed a format string - eg "lt%d" it will try and find a suitable 1178 * id. It scans list of devices to build up a free map, then chooses 1179 * the first empty slot. The caller must hold the dev_base or rtnl lock 1180 * while allocating the name and adding the device in order to avoid 1181 * duplicates. 1182 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1183 * Returns the number of the unit assigned or a negative errno code. 1184 */ 1185 1186 int dev_alloc_name(struct net_device *dev, const char *name) 1187 { 1188 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1189 } 1190 EXPORT_SYMBOL(dev_alloc_name); 1191 1192 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1193 const char *name) 1194 { 1195 int ret; 1196 1197 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1198 return ret < 0 ? ret : 0; 1199 } 1200 1201 /** 1202 * dev_change_name - change name of a device 1203 * @dev: device 1204 * @newname: name (or format string) must be at least IFNAMSIZ 1205 * 1206 * Change name of a device, can pass format strings "eth%d". 1207 * for wildcarding. 1208 */ 1209 int dev_change_name(struct net_device *dev, const char *newname) 1210 { 1211 unsigned char old_assign_type; 1212 char oldname[IFNAMSIZ]; 1213 int err = 0; 1214 int ret; 1215 struct net *net; 1216 1217 ASSERT_RTNL(); 1218 BUG_ON(!dev_net(dev)); 1219 1220 net = dev_net(dev); 1221 1222 down_write(&devnet_rename_sem); 1223 1224 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) { 1225 up_write(&devnet_rename_sem); 1226 return 0; 1227 } 1228 1229 memcpy(oldname, dev->name, IFNAMSIZ); 1230 1231 write_seqlock_bh(&netdev_rename_lock); 1232 err = dev_get_valid_name(net, dev, newname); 1233 write_sequnlock_bh(&netdev_rename_lock); 1234 1235 if (err < 0) { 1236 up_write(&devnet_rename_sem); 1237 return err; 1238 } 1239 1240 if (oldname[0] && !strchr(oldname, '%')) 1241 netdev_info(dev, "renamed from %s%s\n", oldname, 1242 dev->flags & IFF_UP ? " (while UP)" : ""); 1243 1244 old_assign_type = dev->name_assign_type; 1245 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1246 1247 rollback: 1248 ret = device_rename(&dev->dev, dev->name); 1249 if (ret) { 1250 memcpy(dev->name, oldname, IFNAMSIZ); 1251 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1252 up_write(&devnet_rename_sem); 1253 return ret; 1254 } 1255 1256 up_write(&devnet_rename_sem); 1257 1258 netdev_adjacent_rename_links(dev, oldname); 1259 1260 netdev_name_node_del(dev->name_node); 1261 1262 synchronize_net(); 1263 1264 netdev_name_node_add(net, dev->name_node); 1265 1266 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1267 ret = notifier_to_errno(ret); 1268 1269 if (ret) { 1270 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1271 if (err >= 0) { 1272 err = ret; 1273 down_write(&devnet_rename_sem); 1274 write_seqlock_bh(&netdev_rename_lock); 1275 memcpy(dev->name, oldname, IFNAMSIZ); 1276 write_sequnlock_bh(&netdev_rename_lock); 1277 memcpy(oldname, newname, IFNAMSIZ); 1278 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1279 old_assign_type = NET_NAME_RENAMED; 1280 goto rollback; 1281 } else { 1282 netdev_err(dev, "name change rollback failed: %d\n", 1283 ret); 1284 } 1285 } 1286 1287 return err; 1288 } 1289 1290 /** 1291 * dev_set_alias - change ifalias of a device 1292 * @dev: device 1293 * @alias: name up to IFALIASZ 1294 * @len: limit of bytes to copy from info 1295 * 1296 * Set ifalias for a device, 1297 */ 1298 int dev_set_alias(struct net_device *dev, const char *alias, size_t len) 1299 { 1300 struct dev_ifalias *new_alias = NULL; 1301 1302 if (len >= IFALIASZ) 1303 return -EINVAL; 1304 1305 if (len) { 1306 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1307 if (!new_alias) 1308 return -ENOMEM; 1309 1310 memcpy(new_alias->ifalias, alias, len); 1311 new_alias->ifalias[len] = 0; 1312 } 1313 1314 mutex_lock(&ifalias_mutex); 1315 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1316 mutex_is_locked(&ifalias_mutex)); 1317 mutex_unlock(&ifalias_mutex); 1318 1319 if (new_alias) 1320 kfree_rcu(new_alias, rcuhead); 1321 1322 return len; 1323 } 1324 EXPORT_SYMBOL(dev_set_alias); 1325 1326 /** 1327 * dev_get_alias - get ifalias of a device 1328 * @dev: device 1329 * @name: buffer to store name of ifalias 1330 * @len: size of buffer 1331 * 1332 * get ifalias for a device. Caller must make sure dev cannot go 1333 * away, e.g. rcu read lock or own a reference count to device. 1334 */ 1335 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1336 { 1337 const struct dev_ifalias *alias; 1338 int ret = 0; 1339 1340 rcu_read_lock(); 1341 alias = rcu_dereference(dev->ifalias); 1342 if (alias) 1343 ret = snprintf(name, len, "%s", alias->ifalias); 1344 rcu_read_unlock(); 1345 1346 return ret; 1347 } 1348 1349 /** 1350 * netdev_features_change - device changes features 1351 * @dev: device to cause notification 1352 * 1353 * Called to indicate a device has changed features. 1354 */ 1355 void netdev_features_change(struct net_device *dev) 1356 { 1357 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1358 } 1359 EXPORT_SYMBOL(netdev_features_change); 1360 1361 /** 1362 * netdev_state_change - device changes state 1363 * @dev: device to cause notification 1364 * 1365 * Called to indicate a device has changed state. This function calls 1366 * the notifier chains for netdev_chain and sends a NEWLINK message 1367 * to the routing socket. 1368 */ 1369 void netdev_state_change(struct net_device *dev) 1370 { 1371 if (dev->flags & IFF_UP) { 1372 struct netdev_notifier_change_info change_info = { 1373 .info.dev = dev, 1374 }; 1375 1376 call_netdevice_notifiers_info(NETDEV_CHANGE, 1377 &change_info.info); 1378 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1379 } 1380 } 1381 EXPORT_SYMBOL(netdev_state_change); 1382 1383 /** 1384 * __netdev_notify_peers - notify network peers about existence of @dev, 1385 * to be called when rtnl lock is already held. 1386 * @dev: network device 1387 * 1388 * Generate traffic such that interested network peers are aware of 1389 * @dev, such as by generating a gratuitous ARP. This may be used when 1390 * a device wants to inform the rest of the network about some sort of 1391 * reconfiguration such as a failover event or virtual machine 1392 * migration. 1393 */ 1394 void __netdev_notify_peers(struct net_device *dev) 1395 { 1396 ASSERT_RTNL(); 1397 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1398 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1399 } 1400 EXPORT_SYMBOL(__netdev_notify_peers); 1401 1402 /** 1403 * netdev_notify_peers - notify network peers about existence of @dev 1404 * @dev: network device 1405 * 1406 * Generate traffic such that interested network peers are aware of 1407 * @dev, such as by generating a gratuitous ARP. This may be used when 1408 * a device wants to inform the rest of the network about some sort of 1409 * reconfiguration such as a failover event or virtual machine 1410 * migration. 1411 */ 1412 void netdev_notify_peers(struct net_device *dev) 1413 { 1414 rtnl_lock(); 1415 __netdev_notify_peers(dev); 1416 rtnl_unlock(); 1417 } 1418 EXPORT_SYMBOL(netdev_notify_peers); 1419 1420 static int napi_threaded_poll(void *data); 1421 1422 static int napi_kthread_create(struct napi_struct *n) 1423 { 1424 int err = 0; 1425 1426 /* Create and wake up the kthread once to put it in 1427 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1428 * warning and work with loadavg. 1429 */ 1430 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1431 n->dev->name, n->napi_id); 1432 if (IS_ERR(n->thread)) { 1433 err = PTR_ERR(n->thread); 1434 pr_err("kthread_run failed with err %d\n", err); 1435 n->thread = NULL; 1436 } 1437 1438 return err; 1439 } 1440 1441 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1442 { 1443 const struct net_device_ops *ops = dev->netdev_ops; 1444 int ret; 1445 1446 ASSERT_RTNL(); 1447 dev_addr_check(dev); 1448 1449 if (!netif_device_present(dev)) { 1450 /* may be detached because parent is runtime-suspended */ 1451 if (dev->dev.parent) 1452 pm_runtime_resume(dev->dev.parent); 1453 if (!netif_device_present(dev)) 1454 return -ENODEV; 1455 } 1456 1457 /* Block netpoll from trying to do any rx path servicing. 1458 * If we don't do this there is a chance ndo_poll_controller 1459 * or ndo_poll may be running while we open the device 1460 */ 1461 netpoll_poll_disable(dev); 1462 1463 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1464 ret = notifier_to_errno(ret); 1465 if (ret) 1466 return ret; 1467 1468 set_bit(__LINK_STATE_START, &dev->state); 1469 1470 if (ops->ndo_validate_addr) 1471 ret = ops->ndo_validate_addr(dev); 1472 1473 if (!ret && ops->ndo_open) 1474 ret = ops->ndo_open(dev); 1475 1476 netpoll_poll_enable(dev); 1477 1478 if (ret) 1479 clear_bit(__LINK_STATE_START, &dev->state); 1480 else { 1481 dev->flags |= IFF_UP; 1482 dev_set_rx_mode(dev); 1483 dev_activate(dev); 1484 add_device_randomness(dev->dev_addr, dev->addr_len); 1485 } 1486 1487 return ret; 1488 } 1489 1490 /** 1491 * dev_open - prepare an interface for use. 1492 * @dev: device to open 1493 * @extack: netlink extended ack 1494 * 1495 * Takes a device from down to up state. The device's private open 1496 * function is invoked and then the multicast lists are loaded. Finally 1497 * the device is moved into the up state and a %NETDEV_UP message is 1498 * sent to the netdev notifier chain. 1499 * 1500 * Calling this function on an active interface is a nop. On a failure 1501 * a negative errno code is returned. 1502 */ 1503 int dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1504 { 1505 int ret; 1506 1507 if (dev->flags & IFF_UP) 1508 return 0; 1509 1510 ret = __dev_open(dev, extack); 1511 if (ret < 0) 1512 return ret; 1513 1514 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1515 call_netdevice_notifiers(NETDEV_UP, dev); 1516 1517 return ret; 1518 } 1519 EXPORT_SYMBOL(dev_open); 1520 1521 static void __dev_close_many(struct list_head *head) 1522 { 1523 struct net_device *dev; 1524 1525 ASSERT_RTNL(); 1526 might_sleep(); 1527 1528 list_for_each_entry(dev, head, close_list) { 1529 /* Temporarily disable netpoll until the interface is down */ 1530 netpoll_poll_disable(dev); 1531 1532 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1533 1534 clear_bit(__LINK_STATE_START, &dev->state); 1535 1536 /* Synchronize to scheduled poll. We cannot touch poll list, it 1537 * can be even on different cpu. So just clear netif_running(). 1538 * 1539 * dev->stop() will invoke napi_disable() on all of it's 1540 * napi_struct instances on this device. 1541 */ 1542 smp_mb__after_atomic(); /* Commit netif_running(). */ 1543 } 1544 1545 dev_deactivate_many(head); 1546 1547 list_for_each_entry(dev, head, close_list) { 1548 const struct net_device_ops *ops = dev->netdev_ops; 1549 1550 /* 1551 * Call the device specific close. This cannot fail. 1552 * Only if device is UP 1553 * 1554 * We allow it to be called even after a DETACH hot-plug 1555 * event. 1556 */ 1557 if (ops->ndo_stop) 1558 ops->ndo_stop(dev); 1559 1560 dev->flags &= ~IFF_UP; 1561 netpoll_poll_enable(dev); 1562 } 1563 } 1564 1565 static void __dev_close(struct net_device *dev) 1566 { 1567 LIST_HEAD(single); 1568 1569 list_add(&dev->close_list, &single); 1570 __dev_close_many(&single); 1571 list_del(&single); 1572 } 1573 1574 void dev_close_many(struct list_head *head, bool unlink) 1575 { 1576 struct net_device *dev, *tmp; 1577 1578 /* Remove the devices that don't need to be closed */ 1579 list_for_each_entry_safe(dev, tmp, head, close_list) 1580 if (!(dev->flags & IFF_UP)) 1581 list_del_init(&dev->close_list); 1582 1583 __dev_close_many(head); 1584 1585 list_for_each_entry_safe(dev, tmp, head, close_list) { 1586 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1587 call_netdevice_notifiers(NETDEV_DOWN, dev); 1588 if (unlink) 1589 list_del_init(&dev->close_list); 1590 } 1591 } 1592 EXPORT_SYMBOL(dev_close_many); 1593 1594 /** 1595 * dev_close - shutdown an interface. 1596 * @dev: device to shutdown 1597 * 1598 * This function moves an active device into down state. A 1599 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device 1600 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier 1601 * chain. 1602 */ 1603 void dev_close(struct net_device *dev) 1604 { 1605 if (dev->flags & IFF_UP) { 1606 LIST_HEAD(single); 1607 1608 list_add(&dev->close_list, &single); 1609 dev_close_many(&single, true); 1610 list_del(&single); 1611 } 1612 } 1613 EXPORT_SYMBOL(dev_close); 1614 1615 1616 /** 1617 * dev_disable_lro - disable Large Receive Offload on a device 1618 * @dev: device 1619 * 1620 * Disable Large Receive Offload (LRO) on a net device. Must be 1621 * called under RTNL. This is needed if received packets may be 1622 * forwarded to another interface. 1623 */ 1624 void dev_disable_lro(struct net_device *dev) 1625 { 1626 struct net_device *lower_dev; 1627 struct list_head *iter; 1628 1629 dev->wanted_features &= ~NETIF_F_LRO; 1630 netdev_update_features(dev); 1631 1632 if (unlikely(dev->features & NETIF_F_LRO)) 1633 netdev_WARN(dev, "failed to disable LRO!\n"); 1634 1635 netdev_for_each_lower_dev(dev, lower_dev, iter) 1636 dev_disable_lro(lower_dev); 1637 } 1638 EXPORT_SYMBOL(dev_disable_lro); 1639 1640 /** 1641 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1642 * @dev: device 1643 * 1644 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1645 * called under RTNL. This is needed if Generic XDP is installed on 1646 * the device. 1647 */ 1648 static void dev_disable_gro_hw(struct net_device *dev) 1649 { 1650 dev->wanted_features &= ~NETIF_F_GRO_HW; 1651 netdev_update_features(dev); 1652 1653 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1654 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1655 } 1656 1657 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1658 { 1659 #define N(val) \ 1660 case NETDEV_##val: \ 1661 return "NETDEV_" __stringify(val); 1662 switch (cmd) { 1663 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1664 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1665 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1666 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1667 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1668 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1669 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1670 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1671 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1672 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1673 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1674 N(XDP_FEAT_CHANGE) 1675 } 1676 #undef N 1677 return "UNKNOWN_NETDEV_EVENT"; 1678 } 1679 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1680 1681 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1682 struct net_device *dev) 1683 { 1684 struct netdev_notifier_info info = { 1685 .dev = dev, 1686 }; 1687 1688 return nb->notifier_call(nb, val, &info); 1689 } 1690 1691 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1692 struct net_device *dev) 1693 { 1694 int err; 1695 1696 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1697 err = notifier_to_errno(err); 1698 if (err) 1699 return err; 1700 1701 if (!(dev->flags & IFF_UP)) 1702 return 0; 1703 1704 call_netdevice_notifier(nb, NETDEV_UP, dev); 1705 return 0; 1706 } 1707 1708 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1709 struct net_device *dev) 1710 { 1711 if (dev->flags & IFF_UP) { 1712 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1713 dev); 1714 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1715 } 1716 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1717 } 1718 1719 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1720 struct net *net) 1721 { 1722 struct net_device *dev; 1723 int err; 1724 1725 for_each_netdev(net, dev) { 1726 err = call_netdevice_register_notifiers(nb, dev); 1727 if (err) 1728 goto rollback; 1729 } 1730 return 0; 1731 1732 rollback: 1733 for_each_netdev_continue_reverse(net, dev) 1734 call_netdevice_unregister_notifiers(nb, dev); 1735 return err; 1736 } 1737 1738 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1739 struct net *net) 1740 { 1741 struct net_device *dev; 1742 1743 for_each_netdev(net, dev) 1744 call_netdevice_unregister_notifiers(nb, dev); 1745 } 1746 1747 static int dev_boot_phase = 1; 1748 1749 /** 1750 * register_netdevice_notifier - register a network notifier block 1751 * @nb: notifier 1752 * 1753 * Register a notifier to be called when network device events occur. 1754 * The notifier passed is linked into the kernel structures and must 1755 * not be reused until it has been unregistered. A negative errno code 1756 * is returned on a failure. 1757 * 1758 * When registered all registration and up events are replayed 1759 * to the new notifier to allow device to have a race free 1760 * view of the network device list. 1761 */ 1762 1763 int register_netdevice_notifier(struct notifier_block *nb) 1764 { 1765 struct net *net; 1766 int err; 1767 1768 /* Close race with setup_net() and cleanup_net() */ 1769 down_write(&pernet_ops_rwsem); 1770 rtnl_lock(); 1771 err = raw_notifier_chain_register(&netdev_chain, nb); 1772 if (err) 1773 goto unlock; 1774 if (dev_boot_phase) 1775 goto unlock; 1776 for_each_net(net) { 1777 err = call_netdevice_register_net_notifiers(nb, net); 1778 if (err) 1779 goto rollback; 1780 } 1781 1782 unlock: 1783 rtnl_unlock(); 1784 up_write(&pernet_ops_rwsem); 1785 return err; 1786 1787 rollback: 1788 for_each_net_continue_reverse(net) 1789 call_netdevice_unregister_net_notifiers(nb, net); 1790 1791 raw_notifier_chain_unregister(&netdev_chain, nb); 1792 goto unlock; 1793 } 1794 EXPORT_SYMBOL(register_netdevice_notifier); 1795 1796 /** 1797 * unregister_netdevice_notifier - unregister a network notifier block 1798 * @nb: notifier 1799 * 1800 * Unregister a notifier previously registered by 1801 * register_netdevice_notifier(). The notifier is unlinked into the 1802 * kernel structures and may then be reused. A negative errno code 1803 * is returned on a failure. 1804 * 1805 * After unregistering unregister and down device events are synthesized 1806 * for all devices on the device list to the removed notifier to remove 1807 * the need for special case cleanup code. 1808 */ 1809 1810 int unregister_netdevice_notifier(struct notifier_block *nb) 1811 { 1812 struct net *net; 1813 int err; 1814 1815 /* Close race with setup_net() and cleanup_net() */ 1816 down_write(&pernet_ops_rwsem); 1817 rtnl_lock(); 1818 err = raw_notifier_chain_unregister(&netdev_chain, nb); 1819 if (err) 1820 goto unlock; 1821 1822 for_each_net(net) 1823 call_netdevice_unregister_net_notifiers(nb, net); 1824 1825 unlock: 1826 rtnl_unlock(); 1827 up_write(&pernet_ops_rwsem); 1828 return err; 1829 } 1830 EXPORT_SYMBOL(unregister_netdevice_notifier); 1831 1832 static int __register_netdevice_notifier_net(struct net *net, 1833 struct notifier_block *nb, 1834 bool ignore_call_fail) 1835 { 1836 int err; 1837 1838 err = raw_notifier_chain_register(&net->netdev_chain, nb); 1839 if (err) 1840 return err; 1841 if (dev_boot_phase) 1842 return 0; 1843 1844 err = call_netdevice_register_net_notifiers(nb, net); 1845 if (err && !ignore_call_fail) 1846 goto chain_unregister; 1847 1848 return 0; 1849 1850 chain_unregister: 1851 raw_notifier_chain_unregister(&net->netdev_chain, nb); 1852 return err; 1853 } 1854 1855 static int __unregister_netdevice_notifier_net(struct net *net, 1856 struct notifier_block *nb) 1857 { 1858 int err; 1859 1860 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 1861 if (err) 1862 return err; 1863 1864 call_netdevice_unregister_net_notifiers(nb, net); 1865 return 0; 1866 } 1867 1868 /** 1869 * register_netdevice_notifier_net - register a per-netns network notifier block 1870 * @net: network namespace 1871 * @nb: notifier 1872 * 1873 * Register a notifier to be called when network device events occur. 1874 * The notifier passed is linked into the kernel structures and must 1875 * not be reused until it has been unregistered. A negative errno code 1876 * is returned on a failure. 1877 * 1878 * When registered all registration and up events are replayed 1879 * to the new notifier to allow device to have a race free 1880 * view of the network device list. 1881 */ 1882 1883 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 1884 { 1885 int err; 1886 1887 rtnl_lock(); 1888 err = __register_netdevice_notifier_net(net, nb, false); 1889 rtnl_unlock(); 1890 return err; 1891 } 1892 EXPORT_SYMBOL(register_netdevice_notifier_net); 1893 1894 /** 1895 * unregister_netdevice_notifier_net - unregister a per-netns 1896 * network notifier block 1897 * @net: network namespace 1898 * @nb: notifier 1899 * 1900 * Unregister a notifier previously registered by 1901 * register_netdevice_notifier_net(). The notifier is unlinked from the 1902 * kernel structures and may then be reused. A negative errno code 1903 * is returned on a failure. 1904 * 1905 * After unregistering unregister and down device events are synthesized 1906 * for all devices on the device list to the removed notifier to remove 1907 * the need for special case cleanup code. 1908 */ 1909 1910 int unregister_netdevice_notifier_net(struct net *net, 1911 struct notifier_block *nb) 1912 { 1913 int err; 1914 1915 rtnl_lock(); 1916 err = __unregister_netdevice_notifier_net(net, nb); 1917 rtnl_unlock(); 1918 return err; 1919 } 1920 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 1921 1922 static void __move_netdevice_notifier_net(struct net *src_net, 1923 struct net *dst_net, 1924 struct notifier_block *nb) 1925 { 1926 __unregister_netdevice_notifier_net(src_net, nb); 1927 __register_netdevice_notifier_net(dst_net, nb, true); 1928 } 1929 1930 int register_netdevice_notifier_dev_net(struct net_device *dev, 1931 struct notifier_block *nb, 1932 struct netdev_net_notifier *nn) 1933 { 1934 int err; 1935 1936 rtnl_lock(); 1937 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 1938 if (!err) { 1939 nn->nb = nb; 1940 list_add(&nn->list, &dev->net_notifier_list); 1941 } 1942 rtnl_unlock(); 1943 return err; 1944 } 1945 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 1946 1947 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 1948 struct notifier_block *nb, 1949 struct netdev_net_notifier *nn) 1950 { 1951 int err; 1952 1953 rtnl_lock(); 1954 list_del(&nn->list); 1955 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 1956 rtnl_unlock(); 1957 return err; 1958 } 1959 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 1960 1961 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 1962 struct net *net) 1963 { 1964 struct netdev_net_notifier *nn; 1965 1966 list_for_each_entry(nn, &dev->net_notifier_list, list) 1967 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 1968 } 1969 1970 /** 1971 * call_netdevice_notifiers_info - call all network notifier blocks 1972 * @val: value passed unmodified to notifier function 1973 * @info: notifier information data 1974 * 1975 * Call all network notifier blocks. Parameters and return value 1976 * are as for raw_notifier_call_chain(). 1977 */ 1978 1979 int call_netdevice_notifiers_info(unsigned long val, 1980 struct netdev_notifier_info *info) 1981 { 1982 struct net *net = dev_net(info->dev); 1983 int ret; 1984 1985 ASSERT_RTNL(); 1986 1987 /* Run per-netns notifier block chain first, then run the global one. 1988 * Hopefully, one day, the global one is going to be removed after 1989 * all notifier block registrators get converted to be per-netns. 1990 */ 1991 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 1992 if (ret & NOTIFY_STOP_MASK) 1993 return ret; 1994 return raw_notifier_call_chain(&netdev_chain, val, info); 1995 } 1996 1997 /** 1998 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 1999 * for and rollback on error 2000 * @val_up: value passed unmodified to notifier function 2001 * @val_down: value passed unmodified to the notifier function when 2002 * recovering from an error on @val_up 2003 * @info: notifier information data 2004 * 2005 * Call all per-netns network notifier blocks, but not notifier blocks on 2006 * the global notifier chain. Parameters and return value are as for 2007 * raw_notifier_call_chain_robust(). 2008 */ 2009 2010 static int 2011 call_netdevice_notifiers_info_robust(unsigned long val_up, 2012 unsigned long val_down, 2013 struct netdev_notifier_info *info) 2014 { 2015 struct net *net = dev_net(info->dev); 2016 2017 ASSERT_RTNL(); 2018 2019 return raw_notifier_call_chain_robust(&net->netdev_chain, 2020 val_up, val_down, info); 2021 } 2022 2023 static int call_netdevice_notifiers_extack(unsigned long val, 2024 struct net_device *dev, 2025 struct netlink_ext_ack *extack) 2026 { 2027 struct netdev_notifier_info info = { 2028 .dev = dev, 2029 .extack = extack, 2030 }; 2031 2032 return call_netdevice_notifiers_info(val, &info); 2033 } 2034 2035 /** 2036 * call_netdevice_notifiers - call all network notifier blocks 2037 * @val: value passed unmodified to notifier function 2038 * @dev: net_device pointer passed unmodified to notifier function 2039 * 2040 * Call all network notifier blocks. Parameters and return value 2041 * are as for raw_notifier_call_chain(). 2042 */ 2043 2044 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2045 { 2046 return call_netdevice_notifiers_extack(val, dev, NULL); 2047 } 2048 EXPORT_SYMBOL(call_netdevice_notifiers); 2049 2050 /** 2051 * call_netdevice_notifiers_mtu - call all network notifier blocks 2052 * @val: value passed unmodified to notifier function 2053 * @dev: net_device pointer passed unmodified to notifier function 2054 * @arg: additional u32 argument passed to the notifier function 2055 * 2056 * Call all network notifier blocks. Parameters and return value 2057 * are as for raw_notifier_call_chain(). 2058 */ 2059 static int call_netdevice_notifiers_mtu(unsigned long val, 2060 struct net_device *dev, u32 arg) 2061 { 2062 struct netdev_notifier_info_ext info = { 2063 .info.dev = dev, 2064 .ext.mtu = arg, 2065 }; 2066 2067 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2068 2069 return call_netdevice_notifiers_info(val, &info.info); 2070 } 2071 2072 #ifdef CONFIG_NET_INGRESS 2073 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2074 2075 void net_inc_ingress_queue(void) 2076 { 2077 static_branch_inc(&ingress_needed_key); 2078 } 2079 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2080 2081 void net_dec_ingress_queue(void) 2082 { 2083 static_branch_dec(&ingress_needed_key); 2084 } 2085 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2086 #endif 2087 2088 #ifdef CONFIG_NET_EGRESS 2089 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2090 2091 void net_inc_egress_queue(void) 2092 { 2093 static_branch_inc(&egress_needed_key); 2094 } 2095 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2096 2097 void net_dec_egress_queue(void) 2098 { 2099 static_branch_dec(&egress_needed_key); 2100 } 2101 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2102 #endif 2103 2104 #ifdef CONFIG_NET_CLS_ACT 2105 DEFINE_STATIC_KEY_FALSE(tcf_bypass_check_needed_key); 2106 EXPORT_SYMBOL(tcf_bypass_check_needed_key); 2107 #endif 2108 2109 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2110 EXPORT_SYMBOL(netstamp_needed_key); 2111 #ifdef CONFIG_JUMP_LABEL 2112 static atomic_t netstamp_needed_deferred; 2113 static atomic_t netstamp_wanted; 2114 static void netstamp_clear(struct work_struct *work) 2115 { 2116 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2117 int wanted; 2118 2119 wanted = atomic_add_return(deferred, &netstamp_wanted); 2120 if (wanted > 0) 2121 static_branch_enable(&netstamp_needed_key); 2122 else 2123 static_branch_disable(&netstamp_needed_key); 2124 } 2125 static DECLARE_WORK(netstamp_work, netstamp_clear); 2126 #endif 2127 2128 void net_enable_timestamp(void) 2129 { 2130 #ifdef CONFIG_JUMP_LABEL 2131 int wanted = atomic_read(&netstamp_wanted); 2132 2133 while (wanted > 0) { 2134 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2135 return; 2136 } 2137 atomic_inc(&netstamp_needed_deferred); 2138 schedule_work(&netstamp_work); 2139 #else 2140 static_branch_inc(&netstamp_needed_key); 2141 #endif 2142 } 2143 EXPORT_SYMBOL(net_enable_timestamp); 2144 2145 void net_disable_timestamp(void) 2146 { 2147 #ifdef CONFIG_JUMP_LABEL 2148 int wanted = atomic_read(&netstamp_wanted); 2149 2150 while (wanted > 1) { 2151 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2152 return; 2153 } 2154 atomic_dec(&netstamp_needed_deferred); 2155 schedule_work(&netstamp_work); 2156 #else 2157 static_branch_dec(&netstamp_needed_key); 2158 #endif 2159 } 2160 EXPORT_SYMBOL(net_disable_timestamp); 2161 2162 static inline void net_timestamp_set(struct sk_buff *skb) 2163 { 2164 skb->tstamp = 0; 2165 skb->tstamp_type = SKB_CLOCK_REALTIME; 2166 if (static_branch_unlikely(&netstamp_needed_key)) 2167 skb->tstamp = ktime_get_real(); 2168 } 2169 2170 #define net_timestamp_check(COND, SKB) \ 2171 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2172 if ((COND) && !(SKB)->tstamp) \ 2173 (SKB)->tstamp = ktime_get_real(); \ 2174 } \ 2175 2176 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2177 { 2178 return __is_skb_forwardable(dev, skb, true); 2179 } 2180 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2181 2182 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2183 bool check_mtu) 2184 { 2185 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2186 2187 if (likely(!ret)) { 2188 skb->protocol = eth_type_trans(skb, dev); 2189 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2190 } 2191 2192 return ret; 2193 } 2194 2195 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2196 { 2197 return __dev_forward_skb2(dev, skb, true); 2198 } 2199 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2200 2201 /** 2202 * dev_forward_skb - loopback an skb to another netif 2203 * 2204 * @dev: destination network device 2205 * @skb: buffer to forward 2206 * 2207 * return values: 2208 * NET_RX_SUCCESS (no congestion) 2209 * NET_RX_DROP (packet was dropped, but freed) 2210 * 2211 * dev_forward_skb can be used for injecting an skb from the 2212 * start_xmit function of one device into the receive queue 2213 * of another device. 2214 * 2215 * The receiving device may be in another namespace, so 2216 * we have to clear all information in the skb that could 2217 * impact namespace isolation. 2218 */ 2219 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2220 { 2221 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2222 } 2223 EXPORT_SYMBOL_GPL(dev_forward_skb); 2224 2225 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2226 { 2227 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2228 } 2229 2230 static inline int deliver_skb(struct sk_buff *skb, 2231 struct packet_type *pt_prev, 2232 struct net_device *orig_dev) 2233 { 2234 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2235 return -ENOMEM; 2236 refcount_inc(&skb->users); 2237 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2238 } 2239 2240 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2241 struct packet_type **pt, 2242 struct net_device *orig_dev, 2243 __be16 type, 2244 struct list_head *ptype_list) 2245 { 2246 struct packet_type *ptype, *pt_prev = *pt; 2247 2248 list_for_each_entry_rcu(ptype, ptype_list, list) { 2249 if (ptype->type != type) 2250 continue; 2251 if (pt_prev) 2252 deliver_skb(skb, pt_prev, orig_dev); 2253 pt_prev = ptype; 2254 } 2255 *pt = pt_prev; 2256 } 2257 2258 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2259 { 2260 if (!ptype->af_packet_priv || !skb->sk) 2261 return false; 2262 2263 if (ptype->id_match) 2264 return ptype->id_match(ptype, skb->sk); 2265 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2266 return true; 2267 2268 return false; 2269 } 2270 2271 /** 2272 * dev_nit_active - return true if any network interface taps are in use 2273 * 2274 * @dev: network device to check for the presence of taps 2275 */ 2276 bool dev_nit_active(struct net_device *dev) 2277 { 2278 return !list_empty(&net_hotdata.ptype_all) || 2279 !list_empty(&dev->ptype_all); 2280 } 2281 EXPORT_SYMBOL_GPL(dev_nit_active); 2282 2283 /* 2284 * Support routine. Sends outgoing frames to any network 2285 * taps currently in use. 2286 */ 2287 2288 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2289 { 2290 struct list_head *ptype_list = &net_hotdata.ptype_all; 2291 struct packet_type *ptype, *pt_prev = NULL; 2292 struct sk_buff *skb2 = NULL; 2293 2294 rcu_read_lock(); 2295 again: 2296 list_for_each_entry_rcu(ptype, ptype_list, list) { 2297 if (READ_ONCE(ptype->ignore_outgoing)) 2298 continue; 2299 2300 /* Never send packets back to the socket 2301 * they originated from - MvS ([email protected]) 2302 */ 2303 if (skb_loop_sk(ptype, skb)) 2304 continue; 2305 2306 if (pt_prev) { 2307 deliver_skb(skb2, pt_prev, skb->dev); 2308 pt_prev = ptype; 2309 continue; 2310 } 2311 2312 /* need to clone skb, done only once */ 2313 skb2 = skb_clone(skb, GFP_ATOMIC); 2314 if (!skb2) 2315 goto out_unlock; 2316 2317 net_timestamp_set(skb2); 2318 2319 /* skb->nh should be correctly 2320 * set by sender, so that the second statement is 2321 * just protection against buggy protocols. 2322 */ 2323 skb_reset_mac_header(skb2); 2324 2325 if (skb_network_header(skb2) < skb2->data || 2326 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2327 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2328 ntohs(skb2->protocol), 2329 dev->name); 2330 skb_reset_network_header(skb2); 2331 } 2332 2333 skb2->transport_header = skb2->network_header; 2334 skb2->pkt_type = PACKET_OUTGOING; 2335 pt_prev = ptype; 2336 } 2337 2338 if (ptype_list == &net_hotdata.ptype_all) { 2339 ptype_list = &dev->ptype_all; 2340 goto again; 2341 } 2342 out_unlock: 2343 if (pt_prev) { 2344 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2345 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2346 else 2347 kfree_skb(skb2); 2348 } 2349 rcu_read_unlock(); 2350 } 2351 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2352 2353 /** 2354 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2355 * @dev: Network device 2356 * @txq: number of queues available 2357 * 2358 * If real_num_tx_queues is changed the tc mappings may no longer be 2359 * valid. To resolve this verify the tc mapping remains valid and if 2360 * not NULL the mapping. With no priorities mapping to this 2361 * offset/count pair it will no longer be used. In the worst case TC0 2362 * is invalid nothing can be done so disable priority mappings. If is 2363 * expected that drivers will fix this mapping if they can before 2364 * calling netif_set_real_num_tx_queues. 2365 */ 2366 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2367 { 2368 int i; 2369 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2370 2371 /* If TC0 is invalidated disable TC mapping */ 2372 if (tc->offset + tc->count > txq) { 2373 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2374 dev->num_tc = 0; 2375 return; 2376 } 2377 2378 /* Invalidated prio to tc mappings set to TC0 */ 2379 for (i = 1; i < TC_BITMASK + 1; i++) { 2380 int q = netdev_get_prio_tc_map(dev, i); 2381 2382 tc = &dev->tc_to_txq[q]; 2383 if (tc->offset + tc->count > txq) { 2384 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", 2385 i, q); 2386 netdev_set_prio_tc_map(dev, i, 0); 2387 } 2388 } 2389 } 2390 2391 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2392 { 2393 if (dev->num_tc) { 2394 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2395 int i; 2396 2397 /* walk through the TCs and see if it falls into any of them */ 2398 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2399 if ((txq - tc->offset) < tc->count) 2400 return i; 2401 } 2402 2403 /* didn't find it, just return -1 to indicate no match */ 2404 return -1; 2405 } 2406 2407 return 0; 2408 } 2409 EXPORT_SYMBOL(netdev_txq_to_tc); 2410 2411 #ifdef CONFIG_XPS 2412 static struct static_key xps_needed __read_mostly; 2413 static struct static_key xps_rxqs_needed __read_mostly; 2414 static DEFINE_MUTEX(xps_map_mutex); 2415 #define xmap_dereference(P) \ 2416 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2417 2418 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2419 struct xps_dev_maps *old_maps, int tci, u16 index) 2420 { 2421 struct xps_map *map = NULL; 2422 int pos; 2423 2424 map = xmap_dereference(dev_maps->attr_map[tci]); 2425 if (!map) 2426 return false; 2427 2428 for (pos = map->len; pos--;) { 2429 if (map->queues[pos] != index) 2430 continue; 2431 2432 if (map->len > 1) { 2433 map->queues[pos] = map->queues[--map->len]; 2434 break; 2435 } 2436 2437 if (old_maps) 2438 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2439 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2440 kfree_rcu(map, rcu); 2441 return false; 2442 } 2443 2444 return true; 2445 } 2446 2447 static bool remove_xps_queue_cpu(struct net_device *dev, 2448 struct xps_dev_maps *dev_maps, 2449 int cpu, u16 offset, u16 count) 2450 { 2451 int num_tc = dev_maps->num_tc; 2452 bool active = false; 2453 int tci; 2454 2455 for (tci = cpu * num_tc; num_tc--; tci++) { 2456 int i, j; 2457 2458 for (i = count, j = offset; i--; j++) { 2459 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2460 break; 2461 } 2462 2463 active |= i < 0; 2464 } 2465 2466 return active; 2467 } 2468 2469 static void reset_xps_maps(struct net_device *dev, 2470 struct xps_dev_maps *dev_maps, 2471 enum xps_map_type type) 2472 { 2473 static_key_slow_dec_cpuslocked(&xps_needed); 2474 if (type == XPS_RXQS) 2475 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2476 2477 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2478 2479 kfree_rcu(dev_maps, rcu); 2480 } 2481 2482 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2483 u16 offset, u16 count) 2484 { 2485 struct xps_dev_maps *dev_maps; 2486 bool active = false; 2487 int i, j; 2488 2489 dev_maps = xmap_dereference(dev->xps_maps[type]); 2490 if (!dev_maps) 2491 return; 2492 2493 for (j = 0; j < dev_maps->nr_ids; j++) 2494 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2495 if (!active) 2496 reset_xps_maps(dev, dev_maps, type); 2497 2498 if (type == XPS_CPUS) { 2499 for (i = offset + (count - 1); count--; i--) 2500 netdev_queue_numa_node_write( 2501 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2502 } 2503 } 2504 2505 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2506 u16 count) 2507 { 2508 if (!static_key_false(&xps_needed)) 2509 return; 2510 2511 cpus_read_lock(); 2512 mutex_lock(&xps_map_mutex); 2513 2514 if (static_key_false(&xps_rxqs_needed)) 2515 clean_xps_maps(dev, XPS_RXQS, offset, count); 2516 2517 clean_xps_maps(dev, XPS_CPUS, offset, count); 2518 2519 mutex_unlock(&xps_map_mutex); 2520 cpus_read_unlock(); 2521 } 2522 2523 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2524 { 2525 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2526 } 2527 2528 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2529 u16 index, bool is_rxqs_map) 2530 { 2531 struct xps_map *new_map; 2532 int alloc_len = XPS_MIN_MAP_ALLOC; 2533 int i, pos; 2534 2535 for (pos = 0; map && pos < map->len; pos++) { 2536 if (map->queues[pos] != index) 2537 continue; 2538 return map; 2539 } 2540 2541 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2542 if (map) { 2543 if (pos < map->alloc_len) 2544 return map; 2545 2546 alloc_len = map->alloc_len * 2; 2547 } 2548 2549 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2550 * map 2551 */ 2552 if (is_rxqs_map) 2553 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2554 else 2555 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2556 cpu_to_node(attr_index)); 2557 if (!new_map) 2558 return NULL; 2559 2560 for (i = 0; i < pos; i++) 2561 new_map->queues[i] = map->queues[i]; 2562 new_map->alloc_len = alloc_len; 2563 new_map->len = pos; 2564 2565 return new_map; 2566 } 2567 2568 /* Copy xps maps at a given index */ 2569 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2570 struct xps_dev_maps *new_dev_maps, int index, 2571 int tc, bool skip_tc) 2572 { 2573 int i, tci = index * dev_maps->num_tc; 2574 struct xps_map *map; 2575 2576 /* copy maps belonging to foreign traffic classes */ 2577 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2578 if (i == tc && skip_tc) 2579 continue; 2580 2581 /* fill in the new device map from the old device map */ 2582 map = xmap_dereference(dev_maps->attr_map[tci]); 2583 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2584 } 2585 } 2586 2587 /* Must be called under cpus_read_lock */ 2588 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2589 u16 index, enum xps_map_type type) 2590 { 2591 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2592 const unsigned long *online_mask = NULL; 2593 bool active = false, copy = false; 2594 int i, j, tci, numa_node_id = -2; 2595 int maps_sz, num_tc = 1, tc = 0; 2596 struct xps_map *map, *new_map; 2597 unsigned int nr_ids; 2598 2599 WARN_ON_ONCE(index >= dev->num_tx_queues); 2600 2601 if (dev->num_tc) { 2602 /* Do not allow XPS on subordinate device directly */ 2603 num_tc = dev->num_tc; 2604 if (num_tc < 0) 2605 return -EINVAL; 2606 2607 /* If queue belongs to subordinate dev use its map */ 2608 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2609 2610 tc = netdev_txq_to_tc(dev, index); 2611 if (tc < 0) 2612 return -EINVAL; 2613 } 2614 2615 mutex_lock(&xps_map_mutex); 2616 2617 dev_maps = xmap_dereference(dev->xps_maps[type]); 2618 if (type == XPS_RXQS) { 2619 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2620 nr_ids = dev->num_rx_queues; 2621 } else { 2622 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2623 if (num_possible_cpus() > 1) 2624 online_mask = cpumask_bits(cpu_online_mask); 2625 nr_ids = nr_cpu_ids; 2626 } 2627 2628 if (maps_sz < L1_CACHE_BYTES) 2629 maps_sz = L1_CACHE_BYTES; 2630 2631 /* The old dev_maps could be larger or smaller than the one we're 2632 * setting up now, as dev->num_tc or nr_ids could have been updated in 2633 * between. We could try to be smart, but let's be safe instead and only 2634 * copy foreign traffic classes if the two map sizes match. 2635 */ 2636 if (dev_maps && 2637 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2638 copy = true; 2639 2640 /* allocate memory for queue storage */ 2641 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2642 j < nr_ids;) { 2643 if (!new_dev_maps) { 2644 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2645 if (!new_dev_maps) { 2646 mutex_unlock(&xps_map_mutex); 2647 return -ENOMEM; 2648 } 2649 2650 new_dev_maps->nr_ids = nr_ids; 2651 new_dev_maps->num_tc = num_tc; 2652 } 2653 2654 tci = j * num_tc + tc; 2655 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2656 2657 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2658 if (!map) 2659 goto error; 2660 2661 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2662 } 2663 2664 if (!new_dev_maps) 2665 goto out_no_new_maps; 2666 2667 if (!dev_maps) { 2668 /* Increment static keys at most once per type */ 2669 static_key_slow_inc_cpuslocked(&xps_needed); 2670 if (type == XPS_RXQS) 2671 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2672 } 2673 2674 for (j = 0; j < nr_ids; j++) { 2675 bool skip_tc = false; 2676 2677 tci = j * num_tc + tc; 2678 if (netif_attr_test_mask(j, mask, nr_ids) && 2679 netif_attr_test_online(j, online_mask, nr_ids)) { 2680 /* add tx-queue to CPU/rx-queue maps */ 2681 int pos = 0; 2682 2683 skip_tc = true; 2684 2685 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2686 while ((pos < map->len) && (map->queues[pos] != index)) 2687 pos++; 2688 2689 if (pos == map->len) 2690 map->queues[map->len++] = index; 2691 #ifdef CONFIG_NUMA 2692 if (type == XPS_CPUS) { 2693 if (numa_node_id == -2) 2694 numa_node_id = cpu_to_node(j); 2695 else if (numa_node_id != cpu_to_node(j)) 2696 numa_node_id = -1; 2697 } 2698 #endif 2699 } 2700 2701 if (copy) 2702 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2703 skip_tc); 2704 } 2705 2706 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2707 2708 /* Cleanup old maps */ 2709 if (!dev_maps) 2710 goto out_no_old_maps; 2711 2712 for (j = 0; j < dev_maps->nr_ids; j++) { 2713 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2714 map = xmap_dereference(dev_maps->attr_map[tci]); 2715 if (!map) 2716 continue; 2717 2718 if (copy) { 2719 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2720 if (map == new_map) 2721 continue; 2722 } 2723 2724 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2725 kfree_rcu(map, rcu); 2726 } 2727 } 2728 2729 old_dev_maps = dev_maps; 2730 2731 out_no_old_maps: 2732 dev_maps = new_dev_maps; 2733 active = true; 2734 2735 out_no_new_maps: 2736 if (type == XPS_CPUS) 2737 /* update Tx queue numa node */ 2738 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2739 (numa_node_id >= 0) ? 2740 numa_node_id : NUMA_NO_NODE); 2741 2742 if (!dev_maps) 2743 goto out_no_maps; 2744 2745 /* removes tx-queue from unused CPUs/rx-queues */ 2746 for (j = 0; j < dev_maps->nr_ids; j++) { 2747 tci = j * dev_maps->num_tc; 2748 2749 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2750 if (i == tc && 2751 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 2752 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 2753 continue; 2754 2755 active |= remove_xps_queue(dev_maps, 2756 copy ? old_dev_maps : NULL, 2757 tci, index); 2758 } 2759 } 2760 2761 if (old_dev_maps) 2762 kfree_rcu(old_dev_maps, rcu); 2763 2764 /* free map if not active */ 2765 if (!active) 2766 reset_xps_maps(dev, dev_maps, type); 2767 2768 out_no_maps: 2769 mutex_unlock(&xps_map_mutex); 2770 2771 return 0; 2772 error: 2773 /* remove any maps that we added */ 2774 for (j = 0; j < nr_ids; j++) { 2775 for (i = num_tc, tci = j * num_tc; i--; tci++) { 2776 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2777 map = copy ? 2778 xmap_dereference(dev_maps->attr_map[tci]) : 2779 NULL; 2780 if (new_map && new_map != map) 2781 kfree(new_map); 2782 } 2783 } 2784 2785 mutex_unlock(&xps_map_mutex); 2786 2787 kfree(new_dev_maps); 2788 return -ENOMEM; 2789 } 2790 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 2791 2792 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 2793 u16 index) 2794 { 2795 int ret; 2796 2797 cpus_read_lock(); 2798 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 2799 cpus_read_unlock(); 2800 2801 return ret; 2802 } 2803 EXPORT_SYMBOL(netif_set_xps_queue); 2804 2805 #endif 2806 static void netdev_unbind_all_sb_channels(struct net_device *dev) 2807 { 2808 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 2809 2810 /* Unbind any subordinate channels */ 2811 while (txq-- != &dev->_tx[0]) { 2812 if (txq->sb_dev) 2813 netdev_unbind_sb_channel(dev, txq->sb_dev); 2814 } 2815 } 2816 2817 void netdev_reset_tc(struct net_device *dev) 2818 { 2819 #ifdef CONFIG_XPS 2820 netif_reset_xps_queues_gt(dev, 0); 2821 #endif 2822 netdev_unbind_all_sb_channels(dev); 2823 2824 /* Reset TC configuration of device */ 2825 dev->num_tc = 0; 2826 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 2827 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 2828 } 2829 EXPORT_SYMBOL(netdev_reset_tc); 2830 2831 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 2832 { 2833 if (tc >= dev->num_tc) 2834 return -EINVAL; 2835 2836 #ifdef CONFIG_XPS 2837 netif_reset_xps_queues(dev, offset, count); 2838 #endif 2839 dev->tc_to_txq[tc].count = count; 2840 dev->tc_to_txq[tc].offset = offset; 2841 return 0; 2842 } 2843 EXPORT_SYMBOL(netdev_set_tc_queue); 2844 2845 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 2846 { 2847 if (num_tc > TC_MAX_QUEUE) 2848 return -EINVAL; 2849 2850 #ifdef CONFIG_XPS 2851 netif_reset_xps_queues_gt(dev, 0); 2852 #endif 2853 netdev_unbind_all_sb_channels(dev); 2854 2855 dev->num_tc = num_tc; 2856 return 0; 2857 } 2858 EXPORT_SYMBOL(netdev_set_num_tc); 2859 2860 void netdev_unbind_sb_channel(struct net_device *dev, 2861 struct net_device *sb_dev) 2862 { 2863 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 2864 2865 #ifdef CONFIG_XPS 2866 netif_reset_xps_queues_gt(sb_dev, 0); 2867 #endif 2868 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 2869 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 2870 2871 while (txq-- != &dev->_tx[0]) { 2872 if (txq->sb_dev == sb_dev) 2873 txq->sb_dev = NULL; 2874 } 2875 } 2876 EXPORT_SYMBOL(netdev_unbind_sb_channel); 2877 2878 int netdev_bind_sb_channel_queue(struct net_device *dev, 2879 struct net_device *sb_dev, 2880 u8 tc, u16 count, u16 offset) 2881 { 2882 /* Make certain the sb_dev and dev are already configured */ 2883 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 2884 return -EINVAL; 2885 2886 /* We cannot hand out queues we don't have */ 2887 if ((offset + count) > dev->real_num_tx_queues) 2888 return -EINVAL; 2889 2890 /* Record the mapping */ 2891 sb_dev->tc_to_txq[tc].count = count; 2892 sb_dev->tc_to_txq[tc].offset = offset; 2893 2894 /* Provide a way for Tx queue to find the tc_to_txq map or 2895 * XPS map for itself. 2896 */ 2897 while (count--) 2898 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 2899 2900 return 0; 2901 } 2902 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 2903 2904 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 2905 { 2906 /* Do not use a multiqueue device to represent a subordinate channel */ 2907 if (netif_is_multiqueue(dev)) 2908 return -ENODEV; 2909 2910 /* We allow channels 1 - 32767 to be used for subordinate channels. 2911 * Channel 0 is meant to be "native" mode and used only to represent 2912 * the main root device. We allow writing 0 to reset the device back 2913 * to normal mode after being used as a subordinate channel. 2914 */ 2915 if (channel > S16_MAX) 2916 return -EINVAL; 2917 2918 dev->num_tc = -channel; 2919 2920 return 0; 2921 } 2922 EXPORT_SYMBOL(netdev_set_sb_channel); 2923 2924 /* 2925 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 2926 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 2927 */ 2928 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 2929 { 2930 bool disabling; 2931 int rc; 2932 2933 disabling = txq < dev->real_num_tx_queues; 2934 2935 if (txq < 1 || txq > dev->num_tx_queues) 2936 return -EINVAL; 2937 2938 if (dev->reg_state == NETREG_REGISTERED || 2939 dev->reg_state == NETREG_UNREGISTERING) { 2940 ASSERT_RTNL(); 2941 2942 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 2943 txq); 2944 if (rc) 2945 return rc; 2946 2947 if (dev->num_tc) 2948 netif_setup_tc(dev, txq); 2949 2950 dev_qdisc_change_real_num_tx(dev, txq); 2951 2952 dev->real_num_tx_queues = txq; 2953 2954 if (disabling) { 2955 synchronize_net(); 2956 qdisc_reset_all_tx_gt(dev, txq); 2957 #ifdef CONFIG_XPS 2958 netif_reset_xps_queues_gt(dev, txq); 2959 #endif 2960 } 2961 } else { 2962 dev->real_num_tx_queues = txq; 2963 } 2964 2965 return 0; 2966 } 2967 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 2968 2969 #ifdef CONFIG_SYSFS 2970 /** 2971 * netif_set_real_num_rx_queues - set actual number of RX queues used 2972 * @dev: Network device 2973 * @rxq: Actual number of RX queues 2974 * 2975 * This must be called either with the rtnl_lock held or before 2976 * registration of the net device. Returns 0 on success, or a 2977 * negative error code. If called before registration, it always 2978 * succeeds. 2979 */ 2980 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 2981 { 2982 int rc; 2983 2984 if (rxq < 1 || rxq > dev->num_rx_queues) 2985 return -EINVAL; 2986 2987 if (dev->reg_state == NETREG_REGISTERED) { 2988 ASSERT_RTNL(); 2989 2990 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 2991 rxq); 2992 if (rc) 2993 return rc; 2994 } 2995 2996 dev->real_num_rx_queues = rxq; 2997 return 0; 2998 } 2999 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3000 #endif 3001 3002 /** 3003 * netif_set_real_num_queues - set actual number of RX and TX queues used 3004 * @dev: Network device 3005 * @txq: Actual number of TX queues 3006 * @rxq: Actual number of RX queues 3007 * 3008 * Set the real number of both TX and RX queues. 3009 * Does nothing if the number of queues is already correct. 3010 */ 3011 int netif_set_real_num_queues(struct net_device *dev, 3012 unsigned int txq, unsigned int rxq) 3013 { 3014 unsigned int old_rxq = dev->real_num_rx_queues; 3015 int err; 3016 3017 if (txq < 1 || txq > dev->num_tx_queues || 3018 rxq < 1 || rxq > dev->num_rx_queues) 3019 return -EINVAL; 3020 3021 /* Start from increases, so the error path only does decreases - 3022 * decreases can't fail. 3023 */ 3024 if (rxq > dev->real_num_rx_queues) { 3025 err = netif_set_real_num_rx_queues(dev, rxq); 3026 if (err) 3027 return err; 3028 } 3029 if (txq > dev->real_num_tx_queues) { 3030 err = netif_set_real_num_tx_queues(dev, txq); 3031 if (err) 3032 goto undo_rx; 3033 } 3034 if (rxq < dev->real_num_rx_queues) 3035 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3036 if (txq < dev->real_num_tx_queues) 3037 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3038 3039 return 0; 3040 undo_rx: 3041 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3042 return err; 3043 } 3044 EXPORT_SYMBOL(netif_set_real_num_queues); 3045 3046 /** 3047 * netif_set_tso_max_size() - set the max size of TSO frames supported 3048 * @dev: netdev to update 3049 * @size: max skb->len of a TSO frame 3050 * 3051 * Set the limit on the size of TSO super-frames the device can handle. 3052 * Unless explicitly set the stack will assume the value of 3053 * %GSO_LEGACY_MAX_SIZE. 3054 */ 3055 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3056 { 3057 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3058 if (size < READ_ONCE(dev->gso_max_size)) 3059 netif_set_gso_max_size(dev, size); 3060 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3061 netif_set_gso_ipv4_max_size(dev, size); 3062 } 3063 EXPORT_SYMBOL(netif_set_tso_max_size); 3064 3065 /** 3066 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3067 * @dev: netdev to update 3068 * @segs: max number of TCP segments 3069 * 3070 * Set the limit on the number of TCP segments the device can generate from 3071 * a single TSO super-frame. 3072 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3073 */ 3074 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3075 { 3076 dev->tso_max_segs = segs; 3077 if (segs < READ_ONCE(dev->gso_max_segs)) 3078 netif_set_gso_max_segs(dev, segs); 3079 } 3080 EXPORT_SYMBOL(netif_set_tso_max_segs); 3081 3082 /** 3083 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3084 * @to: netdev to update 3085 * @from: netdev from which to copy the limits 3086 */ 3087 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3088 { 3089 netif_set_tso_max_size(to, from->tso_max_size); 3090 netif_set_tso_max_segs(to, from->tso_max_segs); 3091 } 3092 EXPORT_SYMBOL(netif_inherit_tso_max); 3093 3094 /** 3095 * netif_get_num_default_rss_queues - default number of RSS queues 3096 * 3097 * Default value is the number of physical cores if there are only 1 or 2, or 3098 * divided by 2 if there are more. 3099 */ 3100 int netif_get_num_default_rss_queues(void) 3101 { 3102 cpumask_var_t cpus; 3103 int cpu, count = 0; 3104 3105 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3106 return 1; 3107 3108 cpumask_copy(cpus, cpu_online_mask); 3109 for_each_cpu(cpu, cpus) { 3110 ++count; 3111 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3112 } 3113 free_cpumask_var(cpus); 3114 3115 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3116 } 3117 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3118 3119 static void __netif_reschedule(struct Qdisc *q) 3120 { 3121 struct softnet_data *sd; 3122 unsigned long flags; 3123 3124 local_irq_save(flags); 3125 sd = this_cpu_ptr(&softnet_data); 3126 q->next_sched = NULL; 3127 *sd->output_queue_tailp = q; 3128 sd->output_queue_tailp = &q->next_sched; 3129 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3130 local_irq_restore(flags); 3131 } 3132 3133 void __netif_schedule(struct Qdisc *q) 3134 { 3135 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3136 __netif_reschedule(q); 3137 } 3138 EXPORT_SYMBOL(__netif_schedule); 3139 3140 struct dev_kfree_skb_cb { 3141 enum skb_drop_reason reason; 3142 }; 3143 3144 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3145 { 3146 return (struct dev_kfree_skb_cb *)skb->cb; 3147 } 3148 3149 void netif_schedule_queue(struct netdev_queue *txq) 3150 { 3151 rcu_read_lock(); 3152 if (!netif_xmit_stopped(txq)) { 3153 struct Qdisc *q = rcu_dereference(txq->qdisc); 3154 3155 __netif_schedule(q); 3156 } 3157 rcu_read_unlock(); 3158 } 3159 EXPORT_SYMBOL(netif_schedule_queue); 3160 3161 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3162 { 3163 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3164 struct Qdisc *q; 3165 3166 rcu_read_lock(); 3167 q = rcu_dereference(dev_queue->qdisc); 3168 __netif_schedule(q); 3169 rcu_read_unlock(); 3170 } 3171 } 3172 EXPORT_SYMBOL(netif_tx_wake_queue); 3173 3174 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3175 { 3176 unsigned long flags; 3177 3178 if (unlikely(!skb)) 3179 return; 3180 3181 if (likely(refcount_read(&skb->users) == 1)) { 3182 smp_rmb(); 3183 refcount_set(&skb->users, 0); 3184 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3185 return; 3186 } 3187 get_kfree_skb_cb(skb)->reason = reason; 3188 local_irq_save(flags); 3189 skb->next = __this_cpu_read(softnet_data.completion_queue); 3190 __this_cpu_write(softnet_data.completion_queue, skb); 3191 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3192 local_irq_restore(flags); 3193 } 3194 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3195 3196 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3197 { 3198 if (in_hardirq() || irqs_disabled()) 3199 dev_kfree_skb_irq_reason(skb, reason); 3200 else 3201 kfree_skb_reason(skb, reason); 3202 } 3203 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3204 3205 3206 /** 3207 * netif_device_detach - mark device as removed 3208 * @dev: network device 3209 * 3210 * Mark device as removed from system and therefore no longer available. 3211 */ 3212 void netif_device_detach(struct net_device *dev) 3213 { 3214 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3215 netif_running(dev)) { 3216 netif_tx_stop_all_queues(dev); 3217 } 3218 } 3219 EXPORT_SYMBOL(netif_device_detach); 3220 3221 /** 3222 * netif_device_attach - mark device as attached 3223 * @dev: network device 3224 * 3225 * Mark device as attached from system and restart if needed. 3226 */ 3227 void netif_device_attach(struct net_device *dev) 3228 { 3229 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3230 netif_running(dev)) { 3231 netif_tx_wake_all_queues(dev); 3232 __netdev_watchdog_up(dev); 3233 } 3234 } 3235 EXPORT_SYMBOL(netif_device_attach); 3236 3237 /* 3238 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3239 * to be used as a distribution range. 3240 */ 3241 static u16 skb_tx_hash(const struct net_device *dev, 3242 const struct net_device *sb_dev, 3243 struct sk_buff *skb) 3244 { 3245 u32 hash; 3246 u16 qoffset = 0; 3247 u16 qcount = dev->real_num_tx_queues; 3248 3249 if (dev->num_tc) { 3250 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3251 3252 qoffset = sb_dev->tc_to_txq[tc].offset; 3253 qcount = sb_dev->tc_to_txq[tc].count; 3254 if (unlikely(!qcount)) { 3255 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3256 sb_dev->name, qoffset, tc); 3257 qoffset = 0; 3258 qcount = dev->real_num_tx_queues; 3259 } 3260 } 3261 3262 if (skb_rx_queue_recorded(skb)) { 3263 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3264 hash = skb_get_rx_queue(skb); 3265 if (hash >= qoffset) 3266 hash -= qoffset; 3267 while (unlikely(hash >= qcount)) 3268 hash -= qcount; 3269 return hash + qoffset; 3270 } 3271 3272 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3273 } 3274 3275 void skb_warn_bad_offload(const struct sk_buff *skb) 3276 { 3277 static const netdev_features_t null_features; 3278 struct net_device *dev = skb->dev; 3279 const char *name = ""; 3280 3281 if (!net_ratelimit()) 3282 return; 3283 3284 if (dev) { 3285 if (dev->dev.parent) 3286 name = dev_driver_string(dev->dev.parent); 3287 else 3288 name = netdev_name(dev); 3289 } 3290 skb_dump(KERN_WARNING, skb, false); 3291 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3292 name, dev ? &dev->features : &null_features, 3293 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3294 } 3295 3296 /* 3297 * Invalidate hardware checksum when packet is to be mangled, and 3298 * complete checksum manually on outgoing path. 3299 */ 3300 int skb_checksum_help(struct sk_buff *skb) 3301 { 3302 __wsum csum; 3303 int ret = 0, offset; 3304 3305 if (skb->ip_summed == CHECKSUM_COMPLETE) 3306 goto out_set_summed; 3307 3308 if (unlikely(skb_is_gso(skb))) { 3309 skb_warn_bad_offload(skb); 3310 return -EINVAL; 3311 } 3312 3313 /* Before computing a checksum, we should make sure no frag could 3314 * be modified by an external entity : checksum could be wrong. 3315 */ 3316 if (skb_has_shared_frag(skb)) { 3317 ret = __skb_linearize(skb); 3318 if (ret) 3319 goto out; 3320 } 3321 3322 offset = skb_checksum_start_offset(skb); 3323 ret = -EINVAL; 3324 if (unlikely(offset >= skb_headlen(skb))) { 3325 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3326 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3327 offset, skb_headlen(skb)); 3328 goto out; 3329 } 3330 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3331 3332 offset += skb->csum_offset; 3333 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3334 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3335 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3336 offset + sizeof(__sum16), skb_headlen(skb)); 3337 goto out; 3338 } 3339 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3340 if (ret) 3341 goto out; 3342 3343 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3344 out_set_summed: 3345 skb->ip_summed = CHECKSUM_NONE; 3346 out: 3347 return ret; 3348 } 3349 EXPORT_SYMBOL(skb_checksum_help); 3350 3351 int skb_crc32c_csum_help(struct sk_buff *skb) 3352 { 3353 __le32 crc32c_csum; 3354 int ret = 0, offset, start; 3355 3356 if (skb->ip_summed != CHECKSUM_PARTIAL) 3357 goto out; 3358 3359 if (unlikely(skb_is_gso(skb))) 3360 goto out; 3361 3362 /* Before computing a checksum, we should make sure no frag could 3363 * be modified by an external entity : checksum could be wrong. 3364 */ 3365 if (unlikely(skb_has_shared_frag(skb))) { 3366 ret = __skb_linearize(skb); 3367 if (ret) 3368 goto out; 3369 } 3370 start = skb_checksum_start_offset(skb); 3371 offset = start + offsetof(struct sctphdr, checksum); 3372 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3373 ret = -EINVAL; 3374 goto out; 3375 } 3376 3377 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3378 if (ret) 3379 goto out; 3380 3381 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start, 3382 skb->len - start, ~(__u32)0, 3383 crc32c_csum_stub)); 3384 *(__le32 *)(skb->data + offset) = crc32c_csum; 3385 skb_reset_csum_not_inet(skb); 3386 out: 3387 return ret; 3388 } 3389 EXPORT_SYMBOL(skb_crc32c_csum_help); 3390 3391 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3392 { 3393 __be16 type = skb->protocol; 3394 3395 /* Tunnel gso handlers can set protocol to ethernet. */ 3396 if (type == htons(ETH_P_TEB)) { 3397 struct ethhdr *eth; 3398 3399 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3400 return 0; 3401 3402 eth = (struct ethhdr *)skb->data; 3403 type = eth->h_proto; 3404 } 3405 3406 return vlan_get_protocol_and_depth(skb, type, depth); 3407 } 3408 3409 3410 /* Take action when hardware reception checksum errors are detected. */ 3411 #ifdef CONFIG_BUG 3412 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3413 { 3414 netdev_err(dev, "hw csum failure\n"); 3415 skb_dump(KERN_ERR, skb, true); 3416 dump_stack(); 3417 } 3418 3419 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3420 { 3421 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3422 } 3423 EXPORT_SYMBOL(netdev_rx_csum_fault); 3424 #endif 3425 3426 /* XXX: check that highmem exists at all on the given machine. */ 3427 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3428 { 3429 #ifdef CONFIG_HIGHMEM 3430 int i; 3431 3432 if (!(dev->features & NETIF_F_HIGHDMA)) { 3433 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3434 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3435 3436 if (PageHighMem(skb_frag_page(frag))) 3437 return 1; 3438 } 3439 } 3440 #endif 3441 return 0; 3442 } 3443 3444 /* If MPLS offload request, verify we are testing hardware MPLS features 3445 * instead of standard features for the netdev. 3446 */ 3447 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3448 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3449 netdev_features_t features, 3450 __be16 type) 3451 { 3452 if (eth_p_mpls(type)) 3453 features &= skb->dev->mpls_features; 3454 3455 return features; 3456 } 3457 #else 3458 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3459 netdev_features_t features, 3460 __be16 type) 3461 { 3462 return features; 3463 } 3464 #endif 3465 3466 static netdev_features_t harmonize_features(struct sk_buff *skb, 3467 netdev_features_t features) 3468 { 3469 __be16 type; 3470 3471 type = skb_network_protocol(skb, NULL); 3472 features = net_mpls_features(skb, features, type); 3473 3474 if (skb->ip_summed != CHECKSUM_NONE && 3475 !can_checksum_protocol(features, type)) { 3476 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3477 } 3478 if (illegal_highdma(skb->dev, skb)) 3479 features &= ~NETIF_F_SG; 3480 3481 return features; 3482 } 3483 3484 netdev_features_t passthru_features_check(struct sk_buff *skb, 3485 struct net_device *dev, 3486 netdev_features_t features) 3487 { 3488 return features; 3489 } 3490 EXPORT_SYMBOL(passthru_features_check); 3491 3492 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3493 struct net_device *dev, 3494 netdev_features_t features) 3495 { 3496 return vlan_features_check(skb, features); 3497 } 3498 3499 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3500 struct net_device *dev, 3501 netdev_features_t features) 3502 { 3503 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3504 3505 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3506 return features & ~NETIF_F_GSO_MASK; 3507 3508 if (unlikely(skb->len >= READ_ONCE(dev->gso_max_size))) 3509 return features & ~NETIF_F_GSO_MASK; 3510 3511 if (!skb_shinfo(skb)->gso_type) { 3512 skb_warn_bad_offload(skb); 3513 return features & ~NETIF_F_GSO_MASK; 3514 } 3515 3516 /* Support for GSO partial features requires software 3517 * intervention before we can actually process the packets 3518 * so we need to strip support for any partial features now 3519 * and we can pull them back in after we have partially 3520 * segmented the frame. 3521 */ 3522 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3523 features &= ~dev->gso_partial_features; 3524 3525 /* Make sure to clear the IPv4 ID mangling feature if the 3526 * IPv4 header has the potential to be fragmented. 3527 */ 3528 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3529 struct iphdr *iph = skb->encapsulation ? 3530 inner_ip_hdr(skb) : ip_hdr(skb); 3531 3532 if (!(iph->frag_off & htons(IP_DF))) 3533 features &= ~NETIF_F_TSO_MANGLEID; 3534 } 3535 3536 return features; 3537 } 3538 3539 netdev_features_t netif_skb_features(struct sk_buff *skb) 3540 { 3541 struct net_device *dev = skb->dev; 3542 netdev_features_t features = dev->features; 3543 3544 if (skb_is_gso(skb)) 3545 features = gso_features_check(skb, dev, features); 3546 3547 /* If encapsulation offload request, verify we are testing 3548 * hardware encapsulation features instead of standard 3549 * features for the netdev 3550 */ 3551 if (skb->encapsulation) 3552 features &= dev->hw_enc_features; 3553 3554 if (skb_vlan_tagged(skb)) 3555 features = netdev_intersect_features(features, 3556 dev->vlan_features | 3557 NETIF_F_HW_VLAN_CTAG_TX | 3558 NETIF_F_HW_VLAN_STAG_TX); 3559 3560 if (dev->netdev_ops->ndo_features_check) 3561 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3562 features); 3563 else 3564 features &= dflt_features_check(skb, dev, features); 3565 3566 return harmonize_features(skb, features); 3567 } 3568 EXPORT_SYMBOL(netif_skb_features); 3569 3570 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3571 struct netdev_queue *txq, bool more) 3572 { 3573 unsigned int len; 3574 int rc; 3575 3576 if (dev_nit_active(dev)) 3577 dev_queue_xmit_nit(skb, dev); 3578 3579 len = skb->len; 3580 trace_net_dev_start_xmit(skb, dev); 3581 rc = netdev_start_xmit(skb, dev, txq, more); 3582 trace_net_dev_xmit(skb, rc, dev, len); 3583 3584 return rc; 3585 } 3586 3587 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3588 struct netdev_queue *txq, int *ret) 3589 { 3590 struct sk_buff *skb = first; 3591 int rc = NETDEV_TX_OK; 3592 3593 while (skb) { 3594 struct sk_buff *next = skb->next; 3595 3596 skb_mark_not_on_list(skb); 3597 rc = xmit_one(skb, dev, txq, next != NULL); 3598 if (unlikely(!dev_xmit_complete(rc))) { 3599 skb->next = next; 3600 goto out; 3601 } 3602 3603 skb = next; 3604 if (netif_tx_queue_stopped(txq) && skb) { 3605 rc = NETDEV_TX_BUSY; 3606 break; 3607 } 3608 } 3609 3610 out: 3611 *ret = rc; 3612 return skb; 3613 } 3614 3615 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3616 netdev_features_t features) 3617 { 3618 if (skb_vlan_tag_present(skb) && 3619 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3620 skb = __vlan_hwaccel_push_inside(skb); 3621 return skb; 3622 } 3623 3624 int skb_csum_hwoffload_help(struct sk_buff *skb, 3625 const netdev_features_t features) 3626 { 3627 if (unlikely(skb_csum_is_sctp(skb))) 3628 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3629 skb_crc32c_csum_help(skb); 3630 3631 if (features & NETIF_F_HW_CSUM) 3632 return 0; 3633 3634 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3635 switch (skb->csum_offset) { 3636 case offsetof(struct tcphdr, check): 3637 case offsetof(struct udphdr, check): 3638 return 0; 3639 } 3640 } 3641 3642 return skb_checksum_help(skb); 3643 } 3644 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3645 3646 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3647 { 3648 netdev_features_t features; 3649 3650 features = netif_skb_features(skb); 3651 skb = validate_xmit_vlan(skb, features); 3652 if (unlikely(!skb)) 3653 goto out_null; 3654 3655 skb = sk_validate_xmit_skb(skb, dev); 3656 if (unlikely(!skb)) 3657 goto out_null; 3658 3659 if (netif_needs_gso(skb, features)) { 3660 struct sk_buff *segs; 3661 3662 segs = skb_gso_segment(skb, features); 3663 if (IS_ERR(segs)) { 3664 goto out_kfree_skb; 3665 } else if (segs) { 3666 consume_skb(skb); 3667 skb = segs; 3668 } 3669 } else { 3670 if (skb_needs_linearize(skb, features) && 3671 __skb_linearize(skb)) 3672 goto out_kfree_skb; 3673 3674 /* If packet is not checksummed and device does not 3675 * support checksumming for this protocol, complete 3676 * checksumming here. 3677 */ 3678 if (skb->ip_summed == CHECKSUM_PARTIAL) { 3679 if (skb->encapsulation) 3680 skb_set_inner_transport_header(skb, 3681 skb_checksum_start_offset(skb)); 3682 else 3683 skb_set_transport_header(skb, 3684 skb_checksum_start_offset(skb)); 3685 if (skb_csum_hwoffload_help(skb, features)) 3686 goto out_kfree_skb; 3687 } 3688 } 3689 3690 skb = validate_xmit_xfrm(skb, features, again); 3691 3692 return skb; 3693 3694 out_kfree_skb: 3695 kfree_skb(skb); 3696 out_null: 3697 dev_core_stats_tx_dropped_inc(dev); 3698 return NULL; 3699 } 3700 3701 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 3702 { 3703 struct sk_buff *next, *head = NULL, *tail; 3704 3705 for (; skb != NULL; skb = next) { 3706 next = skb->next; 3707 skb_mark_not_on_list(skb); 3708 3709 /* in case skb wont be segmented, point to itself */ 3710 skb->prev = skb; 3711 3712 skb = validate_xmit_skb(skb, dev, again); 3713 if (!skb) 3714 continue; 3715 3716 if (!head) 3717 head = skb; 3718 else 3719 tail->next = skb; 3720 /* If skb was segmented, skb->prev points to 3721 * the last segment. If not, it still contains skb. 3722 */ 3723 tail = skb->prev; 3724 } 3725 return head; 3726 } 3727 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 3728 3729 static void qdisc_pkt_len_init(struct sk_buff *skb) 3730 { 3731 const struct skb_shared_info *shinfo = skb_shinfo(skb); 3732 3733 qdisc_skb_cb(skb)->pkt_len = skb->len; 3734 3735 /* To get more precise estimation of bytes sent on wire, 3736 * we add to pkt_len the headers size of all segments 3737 */ 3738 if (shinfo->gso_size && skb_transport_header_was_set(skb)) { 3739 u16 gso_segs = shinfo->gso_segs; 3740 unsigned int hdr_len; 3741 3742 /* mac layer + network layer */ 3743 hdr_len = skb_transport_offset(skb); 3744 3745 /* + transport layer */ 3746 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 3747 const struct tcphdr *th; 3748 struct tcphdr _tcphdr; 3749 3750 th = skb_header_pointer(skb, hdr_len, 3751 sizeof(_tcphdr), &_tcphdr); 3752 if (likely(th)) 3753 hdr_len += __tcp_hdrlen(th); 3754 } else { 3755 struct udphdr _udphdr; 3756 3757 if (skb_header_pointer(skb, hdr_len, 3758 sizeof(_udphdr), &_udphdr)) 3759 hdr_len += sizeof(struct udphdr); 3760 } 3761 3762 if (shinfo->gso_type & SKB_GSO_DODGY) 3763 gso_segs = DIV_ROUND_UP(skb->len - hdr_len, 3764 shinfo->gso_size); 3765 3766 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 3767 } 3768 } 3769 3770 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 3771 struct sk_buff **to_free, 3772 struct netdev_queue *txq) 3773 { 3774 int rc; 3775 3776 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 3777 if (rc == NET_XMIT_SUCCESS) 3778 trace_qdisc_enqueue(q, txq, skb); 3779 return rc; 3780 } 3781 3782 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 3783 struct net_device *dev, 3784 struct netdev_queue *txq) 3785 { 3786 spinlock_t *root_lock = qdisc_lock(q); 3787 struct sk_buff *to_free = NULL; 3788 bool contended; 3789 int rc; 3790 3791 qdisc_calculate_pkt_len(skb, q); 3792 3793 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 3794 3795 if (q->flags & TCQ_F_NOLOCK) { 3796 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 3797 qdisc_run_begin(q)) { 3798 /* Retest nolock_qdisc_is_empty() within the protection 3799 * of q->seqlock to protect from racing with requeuing. 3800 */ 3801 if (unlikely(!nolock_qdisc_is_empty(q))) { 3802 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3803 __qdisc_run(q); 3804 qdisc_run_end(q); 3805 3806 goto no_lock_out; 3807 } 3808 3809 qdisc_bstats_cpu_update(q, skb); 3810 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 3811 !nolock_qdisc_is_empty(q)) 3812 __qdisc_run(q); 3813 3814 qdisc_run_end(q); 3815 return NET_XMIT_SUCCESS; 3816 } 3817 3818 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3819 qdisc_run(q); 3820 3821 no_lock_out: 3822 if (unlikely(to_free)) 3823 kfree_skb_list_reason(to_free, 3824 tcf_get_drop_reason(to_free)); 3825 return rc; 3826 } 3827 3828 if (unlikely(READ_ONCE(q->owner) == smp_processor_id())) { 3829 kfree_skb_reason(skb, SKB_DROP_REASON_TC_RECLASSIFY_LOOP); 3830 return NET_XMIT_DROP; 3831 } 3832 /* 3833 * Heuristic to force contended enqueues to serialize on a 3834 * separate lock before trying to get qdisc main lock. 3835 * This permits qdisc->running owner to get the lock more 3836 * often and dequeue packets faster. 3837 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 3838 * and then other tasks will only enqueue packets. The packets will be 3839 * sent after the qdisc owner is scheduled again. To prevent this 3840 * scenario the task always serialize on the lock. 3841 */ 3842 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 3843 if (unlikely(contended)) 3844 spin_lock(&q->busylock); 3845 3846 spin_lock(root_lock); 3847 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 3848 __qdisc_drop(skb, &to_free); 3849 rc = NET_XMIT_DROP; 3850 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 3851 qdisc_run_begin(q)) { 3852 /* 3853 * This is a work-conserving queue; there are no old skbs 3854 * waiting to be sent out; and the qdisc is not running - 3855 * xmit the skb directly. 3856 */ 3857 3858 qdisc_bstats_update(q, skb); 3859 3860 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 3861 if (unlikely(contended)) { 3862 spin_unlock(&q->busylock); 3863 contended = false; 3864 } 3865 __qdisc_run(q); 3866 } 3867 3868 qdisc_run_end(q); 3869 rc = NET_XMIT_SUCCESS; 3870 } else { 3871 WRITE_ONCE(q->owner, smp_processor_id()); 3872 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3873 WRITE_ONCE(q->owner, -1); 3874 if (qdisc_run_begin(q)) { 3875 if (unlikely(contended)) { 3876 spin_unlock(&q->busylock); 3877 contended = false; 3878 } 3879 __qdisc_run(q); 3880 qdisc_run_end(q); 3881 } 3882 } 3883 spin_unlock(root_lock); 3884 if (unlikely(to_free)) 3885 kfree_skb_list_reason(to_free, 3886 tcf_get_drop_reason(to_free)); 3887 if (unlikely(contended)) 3888 spin_unlock(&q->busylock); 3889 return rc; 3890 } 3891 3892 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 3893 static void skb_update_prio(struct sk_buff *skb) 3894 { 3895 const struct netprio_map *map; 3896 const struct sock *sk; 3897 unsigned int prioidx; 3898 3899 if (skb->priority) 3900 return; 3901 map = rcu_dereference_bh(skb->dev->priomap); 3902 if (!map) 3903 return; 3904 sk = skb_to_full_sk(skb); 3905 if (!sk) 3906 return; 3907 3908 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 3909 3910 if (prioidx < map->priomap_len) 3911 skb->priority = map->priomap[prioidx]; 3912 } 3913 #else 3914 #define skb_update_prio(skb) 3915 #endif 3916 3917 /** 3918 * dev_loopback_xmit - loop back @skb 3919 * @net: network namespace this loopback is happening in 3920 * @sk: sk needed to be a netfilter okfn 3921 * @skb: buffer to transmit 3922 */ 3923 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 3924 { 3925 skb_reset_mac_header(skb); 3926 __skb_pull(skb, skb_network_offset(skb)); 3927 skb->pkt_type = PACKET_LOOPBACK; 3928 if (skb->ip_summed == CHECKSUM_NONE) 3929 skb->ip_summed = CHECKSUM_UNNECESSARY; 3930 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 3931 skb_dst_force(skb); 3932 netif_rx(skb); 3933 return 0; 3934 } 3935 EXPORT_SYMBOL(dev_loopback_xmit); 3936 3937 #ifdef CONFIG_NET_EGRESS 3938 static struct netdev_queue * 3939 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 3940 { 3941 int qm = skb_get_queue_mapping(skb); 3942 3943 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 3944 } 3945 3946 #ifndef CONFIG_PREEMPT_RT 3947 static bool netdev_xmit_txqueue_skipped(void) 3948 { 3949 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 3950 } 3951 3952 void netdev_xmit_skip_txqueue(bool skip) 3953 { 3954 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 3955 } 3956 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 3957 3958 #else 3959 static bool netdev_xmit_txqueue_skipped(void) 3960 { 3961 return current->net_xmit.skip_txqueue; 3962 } 3963 3964 void netdev_xmit_skip_txqueue(bool skip) 3965 { 3966 current->net_xmit.skip_txqueue = skip; 3967 } 3968 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 3969 #endif 3970 #endif /* CONFIG_NET_EGRESS */ 3971 3972 #ifdef CONFIG_NET_XGRESS 3973 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 3974 enum skb_drop_reason *drop_reason) 3975 { 3976 int ret = TC_ACT_UNSPEC; 3977 #ifdef CONFIG_NET_CLS_ACT 3978 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 3979 struct tcf_result res; 3980 3981 if (!miniq) 3982 return ret; 3983 3984 if (static_branch_unlikely(&tcf_bypass_check_needed_key)) { 3985 if (tcf_block_bypass_sw(miniq->block)) 3986 return ret; 3987 } 3988 3989 tc_skb_cb(skb)->mru = 0; 3990 tc_skb_cb(skb)->post_ct = false; 3991 tcf_set_drop_reason(skb, *drop_reason); 3992 3993 mini_qdisc_bstats_cpu_update(miniq, skb); 3994 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 3995 /* Only tcf related quirks below. */ 3996 switch (ret) { 3997 case TC_ACT_SHOT: 3998 *drop_reason = tcf_get_drop_reason(skb); 3999 mini_qdisc_qstats_cpu_drop(miniq); 4000 break; 4001 case TC_ACT_OK: 4002 case TC_ACT_RECLASSIFY: 4003 skb->tc_index = TC_H_MIN(res.classid); 4004 break; 4005 } 4006 #endif /* CONFIG_NET_CLS_ACT */ 4007 return ret; 4008 } 4009 4010 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4011 4012 void tcx_inc(void) 4013 { 4014 static_branch_inc(&tcx_needed_key); 4015 } 4016 4017 void tcx_dec(void) 4018 { 4019 static_branch_dec(&tcx_needed_key); 4020 } 4021 4022 static __always_inline enum tcx_action_base 4023 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4024 const bool needs_mac) 4025 { 4026 const struct bpf_mprog_fp *fp; 4027 const struct bpf_prog *prog; 4028 int ret = TCX_NEXT; 4029 4030 if (needs_mac) 4031 __skb_push(skb, skb->mac_len); 4032 bpf_mprog_foreach_prog(entry, fp, prog) { 4033 bpf_compute_data_pointers(skb); 4034 ret = bpf_prog_run(prog, skb); 4035 if (ret != TCX_NEXT) 4036 break; 4037 } 4038 if (needs_mac) 4039 __skb_pull(skb, skb->mac_len); 4040 return tcx_action_code(skb, ret); 4041 } 4042 4043 static __always_inline struct sk_buff * 4044 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4045 struct net_device *orig_dev, bool *another) 4046 { 4047 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4048 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4049 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4050 int sch_ret; 4051 4052 if (!entry) 4053 return skb; 4054 4055 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4056 if (*pt_prev) { 4057 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4058 *pt_prev = NULL; 4059 } 4060 4061 qdisc_skb_cb(skb)->pkt_len = skb->len; 4062 tcx_set_ingress(skb, true); 4063 4064 if (static_branch_unlikely(&tcx_needed_key)) { 4065 sch_ret = tcx_run(entry, skb, true); 4066 if (sch_ret != TC_ACT_UNSPEC) 4067 goto ingress_verdict; 4068 } 4069 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4070 ingress_verdict: 4071 switch (sch_ret) { 4072 case TC_ACT_REDIRECT: 4073 /* skb_mac_header check was done by BPF, so we can safely 4074 * push the L2 header back before redirecting to another 4075 * netdev. 4076 */ 4077 __skb_push(skb, skb->mac_len); 4078 if (skb_do_redirect(skb) == -EAGAIN) { 4079 __skb_pull(skb, skb->mac_len); 4080 *another = true; 4081 break; 4082 } 4083 *ret = NET_RX_SUCCESS; 4084 bpf_net_ctx_clear(bpf_net_ctx); 4085 return NULL; 4086 case TC_ACT_SHOT: 4087 kfree_skb_reason(skb, drop_reason); 4088 *ret = NET_RX_DROP; 4089 bpf_net_ctx_clear(bpf_net_ctx); 4090 return NULL; 4091 /* used by tc_run */ 4092 case TC_ACT_STOLEN: 4093 case TC_ACT_QUEUED: 4094 case TC_ACT_TRAP: 4095 consume_skb(skb); 4096 fallthrough; 4097 case TC_ACT_CONSUMED: 4098 *ret = NET_RX_SUCCESS; 4099 bpf_net_ctx_clear(bpf_net_ctx); 4100 return NULL; 4101 } 4102 bpf_net_ctx_clear(bpf_net_ctx); 4103 4104 return skb; 4105 } 4106 4107 static __always_inline struct sk_buff * 4108 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4109 { 4110 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4111 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4112 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4113 int sch_ret; 4114 4115 if (!entry) 4116 return skb; 4117 4118 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4119 4120 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4121 * already set by the caller. 4122 */ 4123 if (static_branch_unlikely(&tcx_needed_key)) { 4124 sch_ret = tcx_run(entry, skb, false); 4125 if (sch_ret != TC_ACT_UNSPEC) 4126 goto egress_verdict; 4127 } 4128 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4129 egress_verdict: 4130 switch (sch_ret) { 4131 case TC_ACT_REDIRECT: 4132 /* No need to push/pop skb's mac_header here on egress! */ 4133 skb_do_redirect(skb); 4134 *ret = NET_XMIT_SUCCESS; 4135 bpf_net_ctx_clear(bpf_net_ctx); 4136 return NULL; 4137 case TC_ACT_SHOT: 4138 kfree_skb_reason(skb, drop_reason); 4139 *ret = NET_XMIT_DROP; 4140 bpf_net_ctx_clear(bpf_net_ctx); 4141 return NULL; 4142 /* used by tc_run */ 4143 case TC_ACT_STOLEN: 4144 case TC_ACT_QUEUED: 4145 case TC_ACT_TRAP: 4146 consume_skb(skb); 4147 fallthrough; 4148 case TC_ACT_CONSUMED: 4149 *ret = NET_XMIT_SUCCESS; 4150 bpf_net_ctx_clear(bpf_net_ctx); 4151 return NULL; 4152 } 4153 bpf_net_ctx_clear(bpf_net_ctx); 4154 4155 return skb; 4156 } 4157 #else 4158 static __always_inline struct sk_buff * 4159 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4160 struct net_device *orig_dev, bool *another) 4161 { 4162 return skb; 4163 } 4164 4165 static __always_inline struct sk_buff * 4166 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4167 { 4168 return skb; 4169 } 4170 #endif /* CONFIG_NET_XGRESS */ 4171 4172 #ifdef CONFIG_XPS 4173 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4174 struct xps_dev_maps *dev_maps, unsigned int tci) 4175 { 4176 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4177 struct xps_map *map; 4178 int queue_index = -1; 4179 4180 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4181 return queue_index; 4182 4183 tci *= dev_maps->num_tc; 4184 tci += tc; 4185 4186 map = rcu_dereference(dev_maps->attr_map[tci]); 4187 if (map) { 4188 if (map->len == 1) 4189 queue_index = map->queues[0]; 4190 else 4191 queue_index = map->queues[reciprocal_scale( 4192 skb_get_hash(skb), map->len)]; 4193 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4194 queue_index = -1; 4195 } 4196 return queue_index; 4197 } 4198 #endif 4199 4200 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4201 struct sk_buff *skb) 4202 { 4203 #ifdef CONFIG_XPS 4204 struct xps_dev_maps *dev_maps; 4205 struct sock *sk = skb->sk; 4206 int queue_index = -1; 4207 4208 if (!static_key_false(&xps_needed)) 4209 return -1; 4210 4211 rcu_read_lock(); 4212 if (!static_key_false(&xps_rxqs_needed)) 4213 goto get_cpus_map; 4214 4215 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4216 if (dev_maps) { 4217 int tci = sk_rx_queue_get(sk); 4218 4219 if (tci >= 0) 4220 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4221 tci); 4222 } 4223 4224 get_cpus_map: 4225 if (queue_index < 0) { 4226 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4227 if (dev_maps) { 4228 unsigned int tci = skb->sender_cpu - 1; 4229 4230 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4231 tci); 4232 } 4233 } 4234 rcu_read_unlock(); 4235 4236 return queue_index; 4237 #else 4238 return -1; 4239 #endif 4240 } 4241 4242 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4243 struct net_device *sb_dev) 4244 { 4245 return 0; 4246 } 4247 EXPORT_SYMBOL(dev_pick_tx_zero); 4248 4249 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb, 4250 struct net_device *sb_dev) 4251 { 4252 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues; 4253 } 4254 EXPORT_SYMBOL(dev_pick_tx_cpu_id); 4255 4256 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4257 struct net_device *sb_dev) 4258 { 4259 struct sock *sk = skb->sk; 4260 int queue_index = sk_tx_queue_get(sk); 4261 4262 sb_dev = sb_dev ? : dev; 4263 4264 if (queue_index < 0 || skb->ooo_okay || 4265 queue_index >= dev->real_num_tx_queues) { 4266 int new_index = get_xps_queue(dev, sb_dev, skb); 4267 4268 if (new_index < 0) 4269 new_index = skb_tx_hash(dev, sb_dev, skb); 4270 4271 if (queue_index != new_index && sk && 4272 sk_fullsock(sk) && 4273 rcu_access_pointer(sk->sk_dst_cache)) 4274 sk_tx_queue_set(sk, new_index); 4275 4276 queue_index = new_index; 4277 } 4278 4279 return queue_index; 4280 } 4281 EXPORT_SYMBOL(netdev_pick_tx); 4282 4283 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4284 struct sk_buff *skb, 4285 struct net_device *sb_dev) 4286 { 4287 int queue_index = 0; 4288 4289 #ifdef CONFIG_XPS 4290 u32 sender_cpu = skb->sender_cpu - 1; 4291 4292 if (sender_cpu >= (u32)NR_CPUS) 4293 skb->sender_cpu = raw_smp_processor_id() + 1; 4294 #endif 4295 4296 if (dev->real_num_tx_queues != 1) { 4297 const struct net_device_ops *ops = dev->netdev_ops; 4298 4299 if (ops->ndo_select_queue) 4300 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4301 else 4302 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4303 4304 queue_index = netdev_cap_txqueue(dev, queue_index); 4305 } 4306 4307 skb_set_queue_mapping(skb, queue_index); 4308 return netdev_get_tx_queue(dev, queue_index); 4309 } 4310 4311 /** 4312 * __dev_queue_xmit() - transmit a buffer 4313 * @skb: buffer to transmit 4314 * @sb_dev: suboordinate device used for L2 forwarding offload 4315 * 4316 * Queue a buffer for transmission to a network device. The caller must 4317 * have set the device and priority and built the buffer before calling 4318 * this function. The function can be called from an interrupt. 4319 * 4320 * When calling this method, interrupts MUST be enabled. This is because 4321 * the BH enable code must have IRQs enabled so that it will not deadlock. 4322 * 4323 * Regardless of the return value, the skb is consumed, so it is currently 4324 * difficult to retry a send to this method. (You can bump the ref count 4325 * before sending to hold a reference for retry if you are careful.) 4326 * 4327 * Return: 4328 * * 0 - buffer successfully transmitted 4329 * * positive qdisc return code - NET_XMIT_DROP etc. 4330 * * negative errno - other errors 4331 */ 4332 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4333 { 4334 struct net_device *dev = skb->dev; 4335 struct netdev_queue *txq = NULL; 4336 struct Qdisc *q; 4337 int rc = -ENOMEM; 4338 bool again = false; 4339 4340 skb_reset_mac_header(skb); 4341 skb_assert_len(skb); 4342 4343 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP)) 4344 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4345 4346 /* Disable soft irqs for various locks below. Also 4347 * stops preemption for RCU. 4348 */ 4349 rcu_read_lock_bh(); 4350 4351 skb_update_prio(skb); 4352 4353 qdisc_pkt_len_init(skb); 4354 tcx_set_ingress(skb, false); 4355 #ifdef CONFIG_NET_EGRESS 4356 if (static_branch_unlikely(&egress_needed_key)) { 4357 if (nf_hook_egress_active()) { 4358 skb = nf_hook_egress(skb, &rc, dev); 4359 if (!skb) 4360 goto out; 4361 } 4362 4363 netdev_xmit_skip_txqueue(false); 4364 4365 nf_skip_egress(skb, true); 4366 skb = sch_handle_egress(skb, &rc, dev); 4367 if (!skb) 4368 goto out; 4369 nf_skip_egress(skb, false); 4370 4371 if (netdev_xmit_txqueue_skipped()) 4372 txq = netdev_tx_queue_mapping(dev, skb); 4373 } 4374 #endif 4375 /* If device/qdisc don't need skb->dst, release it right now while 4376 * its hot in this cpu cache. 4377 */ 4378 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4379 skb_dst_drop(skb); 4380 else 4381 skb_dst_force(skb); 4382 4383 if (!txq) 4384 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4385 4386 q = rcu_dereference_bh(txq->qdisc); 4387 4388 trace_net_dev_queue(skb); 4389 if (q->enqueue) { 4390 rc = __dev_xmit_skb(skb, q, dev, txq); 4391 goto out; 4392 } 4393 4394 /* The device has no queue. Common case for software devices: 4395 * loopback, all the sorts of tunnels... 4396 4397 * Really, it is unlikely that netif_tx_lock protection is necessary 4398 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4399 * counters.) 4400 * However, it is possible, that they rely on protection 4401 * made by us here. 4402 4403 * Check this and shot the lock. It is not prone from deadlocks. 4404 *Either shot noqueue qdisc, it is even simpler 8) 4405 */ 4406 if (dev->flags & IFF_UP) { 4407 int cpu = smp_processor_id(); /* ok because BHs are off */ 4408 4409 /* Other cpus might concurrently change txq->xmit_lock_owner 4410 * to -1 or to their cpu id, but not to our id. 4411 */ 4412 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4413 if (dev_xmit_recursion()) 4414 goto recursion_alert; 4415 4416 skb = validate_xmit_skb(skb, dev, &again); 4417 if (!skb) 4418 goto out; 4419 4420 HARD_TX_LOCK(dev, txq, cpu); 4421 4422 if (!netif_xmit_stopped(txq)) { 4423 dev_xmit_recursion_inc(); 4424 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4425 dev_xmit_recursion_dec(); 4426 if (dev_xmit_complete(rc)) { 4427 HARD_TX_UNLOCK(dev, txq); 4428 goto out; 4429 } 4430 } 4431 HARD_TX_UNLOCK(dev, txq); 4432 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4433 dev->name); 4434 } else { 4435 /* Recursion is detected! It is possible, 4436 * unfortunately 4437 */ 4438 recursion_alert: 4439 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4440 dev->name); 4441 } 4442 } 4443 4444 rc = -ENETDOWN; 4445 rcu_read_unlock_bh(); 4446 4447 dev_core_stats_tx_dropped_inc(dev); 4448 kfree_skb_list(skb); 4449 return rc; 4450 out: 4451 rcu_read_unlock_bh(); 4452 return rc; 4453 } 4454 EXPORT_SYMBOL(__dev_queue_xmit); 4455 4456 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4457 { 4458 struct net_device *dev = skb->dev; 4459 struct sk_buff *orig_skb = skb; 4460 struct netdev_queue *txq; 4461 int ret = NETDEV_TX_BUSY; 4462 bool again = false; 4463 4464 if (unlikely(!netif_running(dev) || 4465 !netif_carrier_ok(dev))) 4466 goto drop; 4467 4468 skb = validate_xmit_skb_list(skb, dev, &again); 4469 if (skb != orig_skb) 4470 goto drop; 4471 4472 skb_set_queue_mapping(skb, queue_id); 4473 txq = skb_get_tx_queue(dev, skb); 4474 4475 local_bh_disable(); 4476 4477 dev_xmit_recursion_inc(); 4478 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4479 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4480 ret = netdev_start_xmit(skb, dev, txq, false); 4481 HARD_TX_UNLOCK(dev, txq); 4482 dev_xmit_recursion_dec(); 4483 4484 local_bh_enable(); 4485 return ret; 4486 drop: 4487 dev_core_stats_tx_dropped_inc(dev); 4488 kfree_skb_list(skb); 4489 return NET_XMIT_DROP; 4490 } 4491 EXPORT_SYMBOL(__dev_direct_xmit); 4492 4493 /************************************************************************* 4494 * Receiver routines 4495 *************************************************************************/ 4496 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4497 4498 int weight_p __read_mostly = 64; /* old backlog weight */ 4499 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4500 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4501 4502 /* Called with irq disabled */ 4503 static inline void ____napi_schedule(struct softnet_data *sd, 4504 struct napi_struct *napi) 4505 { 4506 struct task_struct *thread; 4507 4508 lockdep_assert_irqs_disabled(); 4509 4510 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4511 /* Paired with smp_mb__before_atomic() in 4512 * napi_enable()/dev_set_threaded(). 4513 * Use READ_ONCE() to guarantee a complete 4514 * read on napi->thread. Only call 4515 * wake_up_process() when it's not NULL. 4516 */ 4517 thread = READ_ONCE(napi->thread); 4518 if (thread) { 4519 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4520 goto use_local_napi; 4521 4522 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4523 wake_up_process(thread); 4524 return; 4525 } 4526 } 4527 4528 use_local_napi: 4529 list_add_tail(&napi->poll_list, &sd->poll_list); 4530 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4531 /* If not called from net_rx_action() 4532 * we have to raise NET_RX_SOFTIRQ. 4533 */ 4534 if (!sd->in_net_rx_action) 4535 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4536 } 4537 4538 #ifdef CONFIG_RPS 4539 4540 struct static_key_false rps_needed __read_mostly; 4541 EXPORT_SYMBOL(rps_needed); 4542 struct static_key_false rfs_needed __read_mostly; 4543 EXPORT_SYMBOL(rfs_needed); 4544 4545 static struct rps_dev_flow * 4546 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4547 struct rps_dev_flow *rflow, u16 next_cpu) 4548 { 4549 if (next_cpu < nr_cpu_ids) { 4550 u32 head; 4551 #ifdef CONFIG_RFS_ACCEL 4552 struct netdev_rx_queue *rxqueue; 4553 struct rps_dev_flow_table *flow_table; 4554 struct rps_dev_flow *old_rflow; 4555 u16 rxq_index; 4556 u32 flow_id; 4557 int rc; 4558 4559 /* Should we steer this flow to a different hardware queue? */ 4560 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4561 !(dev->features & NETIF_F_NTUPLE)) 4562 goto out; 4563 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4564 if (rxq_index == skb_get_rx_queue(skb)) 4565 goto out; 4566 4567 rxqueue = dev->_rx + rxq_index; 4568 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4569 if (!flow_table) 4570 goto out; 4571 flow_id = skb_get_hash(skb) & flow_table->mask; 4572 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4573 rxq_index, flow_id); 4574 if (rc < 0) 4575 goto out; 4576 old_rflow = rflow; 4577 rflow = &flow_table->flows[flow_id]; 4578 WRITE_ONCE(rflow->filter, rc); 4579 if (old_rflow->filter == rc) 4580 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 4581 out: 4582 #endif 4583 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 4584 rps_input_queue_tail_save(&rflow->last_qtail, head); 4585 } 4586 4587 WRITE_ONCE(rflow->cpu, next_cpu); 4588 return rflow; 4589 } 4590 4591 /* 4592 * get_rps_cpu is called from netif_receive_skb and returns the target 4593 * CPU from the RPS map of the receiving queue for a given skb. 4594 * rcu_read_lock must be held on entry. 4595 */ 4596 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4597 struct rps_dev_flow **rflowp) 4598 { 4599 const struct rps_sock_flow_table *sock_flow_table; 4600 struct netdev_rx_queue *rxqueue = dev->_rx; 4601 struct rps_dev_flow_table *flow_table; 4602 struct rps_map *map; 4603 int cpu = -1; 4604 u32 tcpu; 4605 u32 hash; 4606 4607 if (skb_rx_queue_recorded(skb)) { 4608 u16 index = skb_get_rx_queue(skb); 4609 4610 if (unlikely(index >= dev->real_num_rx_queues)) { 4611 WARN_ONCE(dev->real_num_rx_queues > 1, 4612 "%s received packet on queue %u, but number " 4613 "of RX queues is %u\n", 4614 dev->name, index, dev->real_num_rx_queues); 4615 goto done; 4616 } 4617 rxqueue += index; 4618 } 4619 4620 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4621 4622 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4623 map = rcu_dereference(rxqueue->rps_map); 4624 if (!flow_table && !map) 4625 goto done; 4626 4627 skb_reset_network_header(skb); 4628 hash = skb_get_hash(skb); 4629 if (!hash) 4630 goto done; 4631 4632 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4633 if (flow_table && sock_flow_table) { 4634 struct rps_dev_flow *rflow; 4635 u32 next_cpu; 4636 u32 ident; 4637 4638 /* First check into global flow table if there is a match. 4639 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4640 */ 4641 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4642 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4643 goto try_rps; 4644 4645 next_cpu = ident & net_hotdata.rps_cpu_mask; 4646 4647 /* OK, now we know there is a match, 4648 * we can look at the local (per receive queue) flow table 4649 */ 4650 rflow = &flow_table->flows[hash & flow_table->mask]; 4651 tcpu = rflow->cpu; 4652 4653 /* 4654 * If the desired CPU (where last recvmsg was done) is 4655 * different from current CPU (one in the rx-queue flow 4656 * table entry), switch if one of the following holds: 4657 * - Current CPU is unset (>= nr_cpu_ids). 4658 * - Current CPU is offline. 4659 * - The current CPU's queue tail has advanced beyond the 4660 * last packet that was enqueued using this table entry. 4661 * This guarantees that all previous packets for the flow 4662 * have been dequeued, thus preserving in order delivery. 4663 */ 4664 if (unlikely(tcpu != next_cpu) && 4665 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4666 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 4667 rflow->last_qtail)) >= 0)) { 4668 tcpu = next_cpu; 4669 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4670 } 4671 4672 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4673 *rflowp = rflow; 4674 cpu = tcpu; 4675 goto done; 4676 } 4677 } 4678 4679 try_rps: 4680 4681 if (map) { 4682 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4683 if (cpu_online(tcpu)) { 4684 cpu = tcpu; 4685 goto done; 4686 } 4687 } 4688 4689 done: 4690 return cpu; 4691 } 4692 4693 #ifdef CONFIG_RFS_ACCEL 4694 4695 /** 4696 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4697 * @dev: Device on which the filter was set 4698 * @rxq_index: RX queue index 4699 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4700 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4701 * 4702 * Drivers that implement ndo_rx_flow_steer() should periodically call 4703 * this function for each installed filter and remove the filters for 4704 * which it returns %true. 4705 */ 4706 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4707 u32 flow_id, u16 filter_id) 4708 { 4709 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4710 struct rps_dev_flow_table *flow_table; 4711 struct rps_dev_flow *rflow; 4712 bool expire = true; 4713 unsigned int cpu; 4714 4715 rcu_read_lock(); 4716 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4717 if (flow_table && flow_id <= flow_table->mask) { 4718 rflow = &flow_table->flows[flow_id]; 4719 cpu = READ_ONCE(rflow->cpu); 4720 if (READ_ONCE(rflow->filter) == filter_id && cpu < nr_cpu_ids && 4721 ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - 4722 READ_ONCE(rflow->last_qtail)) < 4723 (int)(10 * flow_table->mask))) 4724 expire = false; 4725 } 4726 rcu_read_unlock(); 4727 return expire; 4728 } 4729 EXPORT_SYMBOL(rps_may_expire_flow); 4730 4731 #endif /* CONFIG_RFS_ACCEL */ 4732 4733 /* Called from hardirq (IPI) context */ 4734 static void rps_trigger_softirq(void *data) 4735 { 4736 struct softnet_data *sd = data; 4737 4738 ____napi_schedule(sd, &sd->backlog); 4739 sd->received_rps++; 4740 } 4741 4742 #endif /* CONFIG_RPS */ 4743 4744 /* Called from hardirq (IPI) context */ 4745 static void trigger_rx_softirq(void *data) 4746 { 4747 struct softnet_data *sd = data; 4748 4749 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4750 smp_store_release(&sd->defer_ipi_scheduled, 0); 4751 } 4752 4753 /* 4754 * After we queued a packet into sd->input_pkt_queue, 4755 * we need to make sure this queue is serviced soon. 4756 * 4757 * - If this is another cpu queue, link it to our rps_ipi_list, 4758 * and make sure we will process rps_ipi_list from net_rx_action(). 4759 * 4760 * - If this is our own queue, NAPI schedule our backlog. 4761 * Note that this also raises NET_RX_SOFTIRQ. 4762 */ 4763 static void napi_schedule_rps(struct softnet_data *sd) 4764 { 4765 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 4766 4767 #ifdef CONFIG_RPS 4768 if (sd != mysd) { 4769 if (use_backlog_threads()) { 4770 __napi_schedule_irqoff(&sd->backlog); 4771 return; 4772 } 4773 4774 sd->rps_ipi_next = mysd->rps_ipi_list; 4775 mysd->rps_ipi_list = sd; 4776 4777 /* If not called from net_rx_action() or napi_threaded_poll() 4778 * we have to raise NET_RX_SOFTIRQ. 4779 */ 4780 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 4781 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4782 return; 4783 } 4784 #endif /* CONFIG_RPS */ 4785 __napi_schedule_irqoff(&mysd->backlog); 4786 } 4787 4788 void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu) 4789 { 4790 unsigned long flags; 4791 4792 if (use_backlog_threads()) { 4793 backlog_lock_irq_save(sd, &flags); 4794 4795 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 4796 __napi_schedule_irqoff(&sd->backlog); 4797 4798 backlog_unlock_irq_restore(sd, &flags); 4799 4800 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 4801 smp_call_function_single_async(cpu, &sd->defer_csd); 4802 } 4803 } 4804 4805 #ifdef CONFIG_NET_FLOW_LIMIT 4806 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 4807 #endif 4808 4809 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 4810 { 4811 #ifdef CONFIG_NET_FLOW_LIMIT 4812 struct sd_flow_limit *fl; 4813 struct softnet_data *sd; 4814 unsigned int old_flow, new_flow; 4815 4816 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 4817 return false; 4818 4819 sd = this_cpu_ptr(&softnet_data); 4820 4821 rcu_read_lock(); 4822 fl = rcu_dereference(sd->flow_limit); 4823 if (fl) { 4824 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1); 4825 old_flow = fl->history[fl->history_head]; 4826 fl->history[fl->history_head] = new_flow; 4827 4828 fl->history_head++; 4829 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 4830 4831 if (likely(fl->buckets[old_flow])) 4832 fl->buckets[old_flow]--; 4833 4834 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 4835 fl->count++; 4836 rcu_read_unlock(); 4837 return true; 4838 } 4839 } 4840 rcu_read_unlock(); 4841 #endif 4842 return false; 4843 } 4844 4845 /* 4846 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 4847 * queue (may be a remote CPU queue). 4848 */ 4849 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 4850 unsigned int *qtail) 4851 { 4852 enum skb_drop_reason reason; 4853 struct softnet_data *sd; 4854 unsigned long flags; 4855 unsigned int qlen; 4856 int max_backlog; 4857 u32 tail; 4858 4859 reason = SKB_DROP_REASON_DEV_READY; 4860 if (!netif_running(skb->dev)) 4861 goto bad_dev; 4862 4863 reason = SKB_DROP_REASON_CPU_BACKLOG; 4864 sd = &per_cpu(softnet_data, cpu); 4865 4866 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 4867 max_backlog = READ_ONCE(net_hotdata.max_backlog); 4868 if (unlikely(qlen > max_backlog)) 4869 goto cpu_backlog_drop; 4870 backlog_lock_irq_save(sd, &flags); 4871 qlen = skb_queue_len(&sd->input_pkt_queue); 4872 if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) { 4873 if (!qlen) { 4874 /* Schedule NAPI for backlog device. We can use 4875 * non atomic operation as we own the queue lock. 4876 */ 4877 if (!__test_and_set_bit(NAPI_STATE_SCHED, 4878 &sd->backlog.state)) 4879 napi_schedule_rps(sd); 4880 } 4881 __skb_queue_tail(&sd->input_pkt_queue, skb); 4882 tail = rps_input_queue_tail_incr(sd); 4883 backlog_unlock_irq_restore(sd, &flags); 4884 4885 /* save the tail outside of the critical section */ 4886 rps_input_queue_tail_save(qtail, tail); 4887 return NET_RX_SUCCESS; 4888 } 4889 4890 backlog_unlock_irq_restore(sd, &flags); 4891 4892 cpu_backlog_drop: 4893 atomic_inc(&sd->dropped); 4894 bad_dev: 4895 dev_core_stats_rx_dropped_inc(skb->dev); 4896 kfree_skb_reason(skb, reason); 4897 return NET_RX_DROP; 4898 } 4899 4900 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 4901 { 4902 struct net_device *dev = skb->dev; 4903 struct netdev_rx_queue *rxqueue; 4904 4905 rxqueue = dev->_rx; 4906 4907 if (skb_rx_queue_recorded(skb)) { 4908 u16 index = skb_get_rx_queue(skb); 4909 4910 if (unlikely(index >= dev->real_num_rx_queues)) { 4911 WARN_ONCE(dev->real_num_rx_queues > 1, 4912 "%s received packet on queue %u, but number " 4913 "of RX queues is %u\n", 4914 dev->name, index, dev->real_num_rx_queues); 4915 4916 return rxqueue; /* Return first rxqueue */ 4917 } 4918 rxqueue += index; 4919 } 4920 return rxqueue; 4921 } 4922 4923 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 4924 struct bpf_prog *xdp_prog) 4925 { 4926 void *orig_data, *orig_data_end, *hard_start; 4927 struct netdev_rx_queue *rxqueue; 4928 bool orig_bcast, orig_host; 4929 u32 mac_len, frame_sz; 4930 __be16 orig_eth_type; 4931 struct ethhdr *eth; 4932 u32 metalen, act; 4933 int off; 4934 4935 /* The XDP program wants to see the packet starting at the MAC 4936 * header. 4937 */ 4938 mac_len = skb->data - skb_mac_header(skb); 4939 hard_start = skb->data - skb_headroom(skb); 4940 4941 /* SKB "head" area always have tailroom for skb_shared_info */ 4942 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 4943 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 4944 4945 rxqueue = netif_get_rxqueue(skb); 4946 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 4947 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 4948 skb_headlen(skb) + mac_len, true); 4949 if (skb_is_nonlinear(skb)) { 4950 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 4951 xdp_buff_set_frags_flag(xdp); 4952 } else { 4953 xdp_buff_clear_frags_flag(xdp); 4954 } 4955 4956 orig_data_end = xdp->data_end; 4957 orig_data = xdp->data; 4958 eth = (struct ethhdr *)xdp->data; 4959 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 4960 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 4961 orig_eth_type = eth->h_proto; 4962 4963 act = bpf_prog_run_xdp(xdp_prog, xdp); 4964 4965 /* check if bpf_xdp_adjust_head was used */ 4966 off = xdp->data - orig_data; 4967 if (off) { 4968 if (off > 0) 4969 __skb_pull(skb, off); 4970 else if (off < 0) 4971 __skb_push(skb, -off); 4972 4973 skb->mac_header += off; 4974 skb_reset_network_header(skb); 4975 } 4976 4977 /* check if bpf_xdp_adjust_tail was used */ 4978 off = xdp->data_end - orig_data_end; 4979 if (off != 0) { 4980 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 4981 skb->len += off; /* positive on grow, negative on shrink */ 4982 } 4983 4984 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 4985 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 4986 */ 4987 if (xdp_buff_has_frags(xdp)) 4988 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 4989 else 4990 skb->data_len = 0; 4991 4992 /* check if XDP changed eth hdr such SKB needs update */ 4993 eth = (struct ethhdr *)xdp->data; 4994 if ((orig_eth_type != eth->h_proto) || 4995 (orig_host != ether_addr_equal_64bits(eth->h_dest, 4996 skb->dev->dev_addr)) || 4997 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 4998 __skb_push(skb, ETH_HLEN); 4999 skb->pkt_type = PACKET_HOST; 5000 skb->protocol = eth_type_trans(skb, skb->dev); 5001 } 5002 5003 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5004 * before calling us again on redirect path. We do not call do_redirect 5005 * as we leave that up to the caller. 5006 * 5007 * Caller is responsible for managing lifetime of skb (i.e. calling 5008 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5009 */ 5010 switch (act) { 5011 case XDP_REDIRECT: 5012 case XDP_TX: 5013 __skb_push(skb, mac_len); 5014 break; 5015 case XDP_PASS: 5016 metalen = xdp->data - xdp->data_meta; 5017 if (metalen) 5018 skb_metadata_set(skb, metalen); 5019 break; 5020 } 5021 5022 return act; 5023 } 5024 5025 static int 5026 netif_skb_check_for_xdp(struct sk_buff **pskb, struct bpf_prog *prog) 5027 { 5028 struct sk_buff *skb = *pskb; 5029 int err, hroom, troom; 5030 5031 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 5032 return 0; 5033 5034 /* In case we have to go down the path and also linearize, 5035 * then lets do the pskb_expand_head() work just once here. 5036 */ 5037 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5038 troom = skb->tail + skb->data_len - skb->end; 5039 err = pskb_expand_head(skb, 5040 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5041 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5042 if (err) 5043 return err; 5044 5045 return skb_linearize(skb); 5046 } 5047 5048 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5049 struct xdp_buff *xdp, 5050 struct bpf_prog *xdp_prog) 5051 { 5052 struct sk_buff *skb = *pskb; 5053 u32 mac_len, act = XDP_DROP; 5054 5055 /* Reinjected packets coming from act_mirred or similar should 5056 * not get XDP generic processing. 5057 */ 5058 if (skb_is_redirected(skb)) 5059 return XDP_PASS; 5060 5061 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5062 * bytes. This is the guarantee that also native XDP provides, 5063 * thus we need to do it here as well. 5064 */ 5065 mac_len = skb->data - skb_mac_header(skb); 5066 __skb_push(skb, mac_len); 5067 5068 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5069 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5070 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5071 goto do_drop; 5072 } 5073 5074 __skb_pull(*pskb, mac_len); 5075 5076 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5077 switch (act) { 5078 case XDP_REDIRECT: 5079 case XDP_TX: 5080 case XDP_PASS: 5081 break; 5082 default: 5083 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5084 fallthrough; 5085 case XDP_ABORTED: 5086 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5087 fallthrough; 5088 case XDP_DROP: 5089 do_drop: 5090 kfree_skb(*pskb); 5091 break; 5092 } 5093 5094 return act; 5095 } 5096 5097 /* When doing generic XDP we have to bypass the qdisc layer and the 5098 * network taps in order to match in-driver-XDP behavior. This also means 5099 * that XDP packets are able to starve other packets going through a qdisc, 5100 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5101 * queues, so they do not have this starvation issue. 5102 */ 5103 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog) 5104 { 5105 struct net_device *dev = skb->dev; 5106 struct netdev_queue *txq; 5107 bool free_skb = true; 5108 int cpu, rc; 5109 5110 txq = netdev_core_pick_tx(dev, skb, NULL); 5111 cpu = smp_processor_id(); 5112 HARD_TX_LOCK(dev, txq, cpu); 5113 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5114 rc = netdev_start_xmit(skb, dev, txq, 0); 5115 if (dev_xmit_complete(rc)) 5116 free_skb = false; 5117 } 5118 HARD_TX_UNLOCK(dev, txq); 5119 if (free_skb) { 5120 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5121 dev_core_stats_tx_dropped_inc(dev); 5122 kfree_skb(skb); 5123 } 5124 } 5125 5126 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5127 5128 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5129 { 5130 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5131 5132 if (xdp_prog) { 5133 struct xdp_buff xdp; 5134 u32 act; 5135 int err; 5136 5137 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5138 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5139 if (act != XDP_PASS) { 5140 switch (act) { 5141 case XDP_REDIRECT: 5142 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5143 &xdp, xdp_prog); 5144 if (err) 5145 goto out_redir; 5146 break; 5147 case XDP_TX: 5148 generic_xdp_tx(*pskb, xdp_prog); 5149 break; 5150 } 5151 bpf_net_ctx_clear(bpf_net_ctx); 5152 return XDP_DROP; 5153 } 5154 bpf_net_ctx_clear(bpf_net_ctx); 5155 } 5156 return XDP_PASS; 5157 out_redir: 5158 bpf_net_ctx_clear(bpf_net_ctx); 5159 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5160 return XDP_DROP; 5161 } 5162 EXPORT_SYMBOL_GPL(do_xdp_generic); 5163 5164 static int netif_rx_internal(struct sk_buff *skb) 5165 { 5166 int ret; 5167 5168 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5169 5170 trace_netif_rx(skb); 5171 5172 #ifdef CONFIG_RPS 5173 if (static_branch_unlikely(&rps_needed)) { 5174 struct rps_dev_flow voidflow, *rflow = &voidflow; 5175 int cpu; 5176 5177 rcu_read_lock(); 5178 5179 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5180 if (cpu < 0) 5181 cpu = smp_processor_id(); 5182 5183 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5184 5185 rcu_read_unlock(); 5186 } else 5187 #endif 5188 { 5189 unsigned int qtail; 5190 5191 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5192 } 5193 return ret; 5194 } 5195 5196 /** 5197 * __netif_rx - Slightly optimized version of netif_rx 5198 * @skb: buffer to post 5199 * 5200 * This behaves as netif_rx except that it does not disable bottom halves. 5201 * As a result this function may only be invoked from the interrupt context 5202 * (either hard or soft interrupt). 5203 */ 5204 int __netif_rx(struct sk_buff *skb) 5205 { 5206 int ret; 5207 5208 lockdep_assert_once(hardirq_count() | softirq_count()); 5209 5210 trace_netif_rx_entry(skb); 5211 ret = netif_rx_internal(skb); 5212 trace_netif_rx_exit(ret); 5213 return ret; 5214 } 5215 EXPORT_SYMBOL(__netif_rx); 5216 5217 /** 5218 * netif_rx - post buffer to the network code 5219 * @skb: buffer to post 5220 * 5221 * This function receives a packet from a device driver and queues it for 5222 * the upper (protocol) levels to process via the backlog NAPI device. It 5223 * always succeeds. The buffer may be dropped during processing for 5224 * congestion control or by the protocol layers. 5225 * The network buffer is passed via the backlog NAPI device. Modern NIC 5226 * driver should use NAPI and GRO. 5227 * This function can used from interrupt and from process context. The 5228 * caller from process context must not disable interrupts before invoking 5229 * this function. 5230 * 5231 * return values: 5232 * NET_RX_SUCCESS (no congestion) 5233 * NET_RX_DROP (packet was dropped) 5234 * 5235 */ 5236 int netif_rx(struct sk_buff *skb) 5237 { 5238 bool need_bh_off = !(hardirq_count() | softirq_count()); 5239 int ret; 5240 5241 if (need_bh_off) 5242 local_bh_disable(); 5243 trace_netif_rx_entry(skb); 5244 ret = netif_rx_internal(skb); 5245 trace_netif_rx_exit(ret); 5246 if (need_bh_off) 5247 local_bh_enable(); 5248 return ret; 5249 } 5250 EXPORT_SYMBOL(netif_rx); 5251 5252 static __latent_entropy void net_tx_action(struct softirq_action *h) 5253 { 5254 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5255 5256 if (sd->completion_queue) { 5257 struct sk_buff *clist; 5258 5259 local_irq_disable(); 5260 clist = sd->completion_queue; 5261 sd->completion_queue = NULL; 5262 local_irq_enable(); 5263 5264 while (clist) { 5265 struct sk_buff *skb = clist; 5266 5267 clist = clist->next; 5268 5269 WARN_ON(refcount_read(&skb->users)); 5270 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5271 trace_consume_skb(skb, net_tx_action); 5272 else 5273 trace_kfree_skb(skb, net_tx_action, 5274 get_kfree_skb_cb(skb)->reason, NULL); 5275 5276 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5277 __kfree_skb(skb); 5278 else 5279 __napi_kfree_skb(skb, 5280 get_kfree_skb_cb(skb)->reason); 5281 } 5282 } 5283 5284 if (sd->output_queue) { 5285 struct Qdisc *head; 5286 5287 local_irq_disable(); 5288 head = sd->output_queue; 5289 sd->output_queue = NULL; 5290 sd->output_queue_tailp = &sd->output_queue; 5291 local_irq_enable(); 5292 5293 rcu_read_lock(); 5294 5295 while (head) { 5296 struct Qdisc *q = head; 5297 spinlock_t *root_lock = NULL; 5298 5299 head = head->next_sched; 5300 5301 /* We need to make sure head->next_sched is read 5302 * before clearing __QDISC_STATE_SCHED 5303 */ 5304 smp_mb__before_atomic(); 5305 5306 if (!(q->flags & TCQ_F_NOLOCK)) { 5307 root_lock = qdisc_lock(q); 5308 spin_lock(root_lock); 5309 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5310 &q->state))) { 5311 /* There is a synchronize_net() between 5312 * STATE_DEACTIVATED flag being set and 5313 * qdisc_reset()/some_qdisc_is_busy() in 5314 * dev_deactivate(), so we can safely bail out 5315 * early here to avoid data race between 5316 * qdisc_deactivate() and some_qdisc_is_busy() 5317 * for lockless qdisc. 5318 */ 5319 clear_bit(__QDISC_STATE_SCHED, &q->state); 5320 continue; 5321 } 5322 5323 clear_bit(__QDISC_STATE_SCHED, &q->state); 5324 qdisc_run(q); 5325 if (root_lock) 5326 spin_unlock(root_lock); 5327 } 5328 5329 rcu_read_unlock(); 5330 } 5331 5332 xfrm_dev_backlog(sd); 5333 } 5334 5335 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5336 /* This hook is defined here for ATM LANE */ 5337 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5338 unsigned char *addr) __read_mostly; 5339 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5340 #endif 5341 5342 /** 5343 * netdev_is_rx_handler_busy - check if receive handler is registered 5344 * @dev: device to check 5345 * 5346 * Check if a receive handler is already registered for a given device. 5347 * Return true if there one. 5348 * 5349 * The caller must hold the rtnl_mutex. 5350 */ 5351 bool netdev_is_rx_handler_busy(struct net_device *dev) 5352 { 5353 ASSERT_RTNL(); 5354 return dev && rtnl_dereference(dev->rx_handler); 5355 } 5356 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5357 5358 /** 5359 * netdev_rx_handler_register - register receive handler 5360 * @dev: device to register a handler for 5361 * @rx_handler: receive handler to register 5362 * @rx_handler_data: data pointer that is used by rx handler 5363 * 5364 * Register a receive handler for a device. This handler will then be 5365 * called from __netif_receive_skb. A negative errno code is returned 5366 * on a failure. 5367 * 5368 * The caller must hold the rtnl_mutex. 5369 * 5370 * For a general description of rx_handler, see enum rx_handler_result. 5371 */ 5372 int netdev_rx_handler_register(struct net_device *dev, 5373 rx_handler_func_t *rx_handler, 5374 void *rx_handler_data) 5375 { 5376 if (netdev_is_rx_handler_busy(dev)) 5377 return -EBUSY; 5378 5379 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5380 return -EINVAL; 5381 5382 /* Note: rx_handler_data must be set before rx_handler */ 5383 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5384 rcu_assign_pointer(dev->rx_handler, rx_handler); 5385 5386 return 0; 5387 } 5388 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5389 5390 /** 5391 * netdev_rx_handler_unregister - unregister receive handler 5392 * @dev: device to unregister a handler from 5393 * 5394 * Unregister a receive handler from a device. 5395 * 5396 * The caller must hold the rtnl_mutex. 5397 */ 5398 void netdev_rx_handler_unregister(struct net_device *dev) 5399 { 5400 5401 ASSERT_RTNL(); 5402 RCU_INIT_POINTER(dev->rx_handler, NULL); 5403 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5404 * section has a guarantee to see a non NULL rx_handler_data 5405 * as well. 5406 */ 5407 synchronize_net(); 5408 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5409 } 5410 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5411 5412 /* 5413 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5414 * the special handling of PFMEMALLOC skbs. 5415 */ 5416 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5417 { 5418 switch (skb->protocol) { 5419 case htons(ETH_P_ARP): 5420 case htons(ETH_P_IP): 5421 case htons(ETH_P_IPV6): 5422 case htons(ETH_P_8021Q): 5423 case htons(ETH_P_8021AD): 5424 return true; 5425 default: 5426 return false; 5427 } 5428 } 5429 5430 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5431 int *ret, struct net_device *orig_dev) 5432 { 5433 if (nf_hook_ingress_active(skb)) { 5434 int ingress_retval; 5435 5436 if (*pt_prev) { 5437 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5438 *pt_prev = NULL; 5439 } 5440 5441 rcu_read_lock(); 5442 ingress_retval = nf_hook_ingress(skb); 5443 rcu_read_unlock(); 5444 return ingress_retval; 5445 } 5446 return 0; 5447 } 5448 5449 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5450 struct packet_type **ppt_prev) 5451 { 5452 struct packet_type *ptype, *pt_prev; 5453 rx_handler_func_t *rx_handler; 5454 struct sk_buff *skb = *pskb; 5455 struct net_device *orig_dev; 5456 bool deliver_exact = false; 5457 int ret = NET_RX_DROP; 5458 __be16 type; 5459 5460 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5461 5462 trace_netif_receive_skb(skb); 5463 5464 orig_dev = skb->dev; 5465 5466 skb_reset_network_header(skb); 5467 if (!skb_transport_header_was_set(skb)) 5468 skb_reset_transport_header(skb); 5469 skb_reset_mac_len(skb); 5470 5471 pt_prev = NULL; 5472 5473 another_round: 5474 skb->skb_iif = skb->dev->ifindex; 5475 5476 __this_cpu_inc(softnet_data.processed); 5477 5478 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5479 int ret2; 5480 5481 migrate_disable(); 5482 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5483 &skb); 5484 migrate_enable(); 5485 5486 if (ret2 != XDP_PASS) { 5487 ret = NET_RX_DROP; 5488 goto out; 5489 } 5490 } 5491 5492 if (eth_type_vlan(skb->protocol)) { 5493 skb = skb_vlan_untag(skb); 5494 if (unlikely(!skb)) 5495 goto out; 5496 } 5497 5498 if (skb_skip_tc_classify(skb)) 5499 goto skip_classify; 5500 5501 if (pfmemalloc) 5502 goto skip_taps; 5503 5504 list_for_each_entry_rcu(ptype, &net_hotdata.ptype_all, list) { 5505 if (pt_prev) 5506 ret = deliver_skb(skb, pt_prev, orig_dev); 5507 pt_prev = ptype; 5508 } 5509 5510 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5511 if (pt_prev) 5512 ret = deliver_skb(skb, pt_prev, orig_dev); 5513 pt_prev = ptype; 5514 } 5515 5516 skip_taps: 5517 #ifdef CONFIG_NET_INGRESS 5518 if (static_branch_unlikely(&ingress_needed_key)) { 5519 bool another = false; 5520 5521 nf_skip_egress(skb, true); 5522 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5523 &another); 5524 if (another) 5525 goto another_round; 5526 if (!skb) 5527 goto out; 5528 5529 nf_skip_egress(skb, false); 5530 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5531 goto out; 5532 } 5533 #endif 5534 skb_reset_redirect(skb); 5535 skip_classify: 5536 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5537 goto drop; 5538 5539 if (skb_vlan_tag_present(skb)) { 5540 if (pt_prev) { 5541 ret = deliver_skb(skb, pt_prev, orig_dev); 5542 pt_prev = NULL; 5543 } 5544 if (vlan_do_receive(&skb)) 5545 goto another_round; 5546 else if (unlikely(!skb)) 5547 goto out; 5548 } 5549 5550 rx_handler = rcu_dereference(skb->dev->rx_handler); 5551 if (rx_handler) { 5552 if (pt_prev) { 5553 ret = deliver_skb(skb, pt_prev, orig_dev); 5554 pt_prev = NULL; 5555 } 5556 switch (rx_handler(&skb)) { 5557 case RX_HANDLER_CONSUMED: 5558 ret = NET_RX_SUCCESS; 5559 goto out; 5560 case RX_HANDLER_ANOTHER: 5561 goto another_round; 5562 case RX_HANDLER_EXACT: 5563 deliver_exact = true; 5564 break; 5565 case RX_HANDLER_PASS: 5566 break; 5567 default: 5568 BUG(); 5569 } 5570 } 5571 5572 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5573 check_vlan_id: 5574 if (skb_vlan_tag_get_id(skb)) { 5575 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5576 * find vlan device. 5577 */ 5578 skb->pkt_type = PACKET_OTHERHOST; 5579 } else if (eth_type_vlan(skb->protocol)) { 5580 /* Outer header is 802.1P with vlan 0, inner header is 5581 * 802.1Q or 802.1AD and vlan_do_receive() above could 5582 * not find vlan dev for vlan id 0. 5583 */ 5584 __vlan_hwaccel_clear_tag(skb); 5585 skb = skb_vlan_untag(skb); 5586 if (unlikely(!skb)) 5587 goto out; 5588 if (vlan_do_receive(&skb)) 5589 /* After stripping off 802.1P header with vlan 0 5590 * vlan dev is found for inner header. 5591 */ 5592 goto another_round; 5593 else if (unlikely(!skb)) 5594 goto out; 5595 else 5596 /* We have stripped outer 802.1P vlan 0 header. 5597 * But could not find vlan dev. 5598 * check again for vlan id to set OTHERHOST. 5599 */ 5600 goto check_vlan_id; 5601 } 5602 /* Note: we might in the future use prio bits 5603 * and set skb->priority like in vlan_do_receive() 5604 * For the time being, just ignore Priority Code Point 5605 */ 5606 __vlan_hwaccel_clear_tag(skb); 5607 } 5608 5609 type = skb->protocol; 5610 5611 /* deliver only exact match when indicated */ 5612 if (likely(!deliver_exact)) { 5613 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5614 &ptype_base[ntohs(type) & 5615 PTYPE_HASH_MASK]); 5616 } 5617 5618 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5619 &orig_dev->ptype_specific); 5620 5621 if (unlikely(skb->dev != orig_dev)) { 5622 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5623 &skb->dev->ptype_specific); 5624 } 5625 5626 if (pt_prev) { 5627 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5628 goto drop; 5629 *ppt_prev = pt_prev; 5630 } else { 5631 drop: 5632 if (!deliver_exact) 5633 dev_core_stats_rx_dropped_inc(skb->dev); 5634 else 5635 dev_core_stats_rx_nohandler_inc(skb->dev); 5636 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5637 /* Jamal, now you will not able to escape explaining 5638 * me how you were going to use this. :-) 5639 */ 5640 ret = NET_RX_DROP; 5641 } 5642 5643 out: 5644 /* The invariant here is that if *ppt_prev is not NULL 5645 * then skb should also be non-NULL. 5646 * 5647 * Apparently *ppt_prev assignment above holds this invariant due to 5648 * skb dereferencing near it. 5649 */ 5650 *pskb = skb; 5651 return ret; 5652 } 5653 5654 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5655 { 5656 struct net_device *orig_dev = skb->dev; 5657 struct packet_type *pt_prev = NULL; 5658 int ret; 5659 5660 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5661 if (pt_prev) 5662 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5663 skb->dev, pt_prev, orig_dev); 5664 return ret; 5665 } 5666 5667 /** 5668 * netif_receive_skb_core - special purpose version of netif_receive_skb 5669 * @skb: buffer to process 5670 * 5671 * More direct receive version of netif_receive_skb(). It should 5672 * only be used by callers that have a need to skip RPS and Generic XDP. 5673 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5674 * 5675 * This function may only be called from softirq context and interrupts 5676 * should be enabled. 5677 * 5678 * Return values (usually ignored): 5679 * NET_RX_SUCCESS: no congestion 5680 * NET_RX_DROP: packet was dropped 5681 */ 5682 int netif_receive_skb_core(struct sk_buff *skb) 5683 { 5684 int ret; 5685 5686 rcu_read_lock(); 5687 ret = __netif_receive_skb_one_core(skb, false); 5688 rcu_read_unlock(); 5689 5690 return ret; 5691 } 5692 EXPORT_SYMBOL(netif_receive_skb_core); 5693 5694 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5695 struct packet_type *pt_prev, 5696 struct net_device *orig_dev) 5697 { 5698 struct sk_buff *skb, *next; 5699 5700 if (!pt_prev) 5701 return; 5702 if (list_empty(head)) 5703 return; 5704 if (pt_prev->list_func != NULL) 5705 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5706 ip_list_rcv, head, pt_prev, orig_dev); 5707 else 5708 list_for_each_entry_safe(skb, next, head, list) { 5709 skb_list_del_init(skb); 5710 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5711 } 5712 } 5713 5714 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5715 { 5716 /* Fast-path assumptions: 5717 * - There is no RX handler. 5718 * - Only one packet_type matches. 5719 * If either of these fails, we will end up doing some per-packet 5720 * processing in-line, then handling the 'last ptype' for the whole 5721 * sublist. This can't cause out-of-order delivery to any single ptype, 5722 * because the 'last ptype' must be constant across the sublist, and all 5723 * other ptypes are handled per-packet. 5724 */ 5725 /* Current (common) ptype of sublist */ 5726 struct packet_type *pt_curr = NULL; 5727 /* Current (common) orig_dev of sublist */ 5728 struct net_device *od_curr = NULL; 5729 struct list_head sublist; 5730 struct sk_buff *skb, *next; 5731 5732 INIT_LIST_HEAD(&sublist); 5733 list_for_each_entry_safe(skb, next, head, list) { 5734 struct net_device *orig_dev = skb->dev; 5735 struct packet_type *pt_prev = NULL; 5736 5737 skb_list_del_init(skb); 5738 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5739 if (!pt_prev) 5740 continue; 5741 if (pt_curr != pt_prev || od_curr != orig_dev) { 5742 /* dispatch old sublist */ 5743 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5744 /* start new sublist */ 5745 INIT_LIST_HEAD(&sublist); 5746 pt_curr = pt_prev; 5747 od_curr = orig_dev; 5748 } 5749 list_add_tail(&skb->list, &sublist); 5750 } 5751 5752 /* dispatch final sublist */ 5753 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5754 } 5755 5756 static int __netif_receive_skb(struct sk_buff *skb) 5757 { 5758 int ret; 5759 5760 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 5761 unsigned int noreclaim_flag; 5762 5763 /* 5764 * PFMEMALLOC skbs are special, they should 5765 * - be delivered to SOCK_MEMALLOC sockets only 5766 * - stay away from userspace 5767 * - have bounded memory usage 5768 * 5769 * Use PF_MEMALLOC as this saves us from propagating the allocation 5770 * context down to all allocation sites. 5771 */ 5772 noreclaim_flag = memalloc_noreclaim_save(); 5773 ret = __netif_receive_skb_one_core(skb, true); 5774 memalloc_noreclaim_restore(noreclaim_flag); 5775 } else 5776 ret = __netif_receive_skb_one_core(skb, false); 5777 5778 return ret; 5779 } 5780 5781 static void __netif_receive_skb_list(struct list_head *head) 5782 { 5783 unsigned long noreclaim_flag = 0; 5784 struct sk_buff *skb, *next; 5785 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 5786 5787 list_for_each_entry_safe(skb, next, head, list) { 5788 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 5789 struct list_head sublist; 5790 5791 /* Handle the previous sublist */ 5792 list_cut_before(&sublist, head, &skb->list); 5793 if (!list_empty(&sublist)) 5794 __netif_receive_skb_list_core(&sublist, pfmemalloc); 5795 pfmemalloc = !pfmemalloc; 5796 /* See comments in __netif_receive_skb */ 5797 if (pfmemalloc) 5798 noreclaim_flag = memalloc_noreclaim_save(); 5799 else 5800 memalloc_noreclaim_restore(noreclaim_flag); 5801 } 5802 } 5803 /* Handle the remaining sublist */ 5804 if (!list_empty(head)) 5805 __netif_receive_skb_list_core(head, pfmemalloc); 5806 /* Restore pflags */ 5807 if (pfmemalloc) 5808 memalloc_noreclaim_restore(noreclaim_flag); 5809 } 5810 5811 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 5812 { 5813 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 5814 struct bpf_prog *new = xdp->prog; 5815 int ret = 0; 5816 5817 switch (xdp->command) { 5818 case XDP_SETUP_PROG: 5819 rcu_assign_pointer(dev->xdp_prog, new); 5820 if (old) 5821 bpf_prog_put(old); 5822 5823 if (old && !new) { 5824 static_branch_dec(&generic_xdp_needed_key); 5825 } else if (new && !old) { 5826 static_branch_inc(&generic_xdp_needed_key); 5827 dev_disable_lro(dev); 5828 dev_disable_gro_hw(dev); 5829 } 5830 break; 5831 5832 default: 5833 ret = -EINVAL; 5834 break; 5835 } 5836 5837 return ret; 5838 } 5839 5840 static int netif_receive_skb_internal(struct sk_buff *skb) 5841 { 5842 int ret; 5843 5844 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5845 5846 if (skb_defer_rx_timestamp(skb)) 5847 return NET_RX_SUCCESS; 5848 5849 rcu_read_lock(); 5850 #ifdef CONFIG_RPS 5851 if (static_branch_unlikely(&rps_needed)) { 5852 struct rps_dev_flow voidflow, *rflow = &voidflow; 5853 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 5854 5855 if (cpu >= 0) { 5856 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5857 rcu_read_unlock(); 5858 return ret; 5859 } 5860 } 5861 #endif 5862 ret = __netif_receive_skb(skb); 5863 rcu_read_unlock(); 5864 return ret; 5865 } 5866 5867 void netif_receive_skb_list_internal(struct list_head *head) 5868 { 5869 struct sk_buff *skb, *next; 5870 struct list_head sublist; 5871 5872 INIT_LIST_HEAD(&sublist); 5873 list_for_each_entry_safe(skb, next, head, list) { 5874 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 5875 skb); 5876 skb_list_del_init(skb); 5877 if (!skb_defer_rx_timestamp(skb)) 5878 list_add_tail(&skb->list, &sublist); 5879 } 5880 list_splice_init(&sublist, head); 5881 5882 rcu_read_lock(); 5883 #ifdef CONFIG_RPS 5884 if (static_branch_unlikely(&rps_needed)) { 5885 list_for_each_entry_safe(skb, next, head, list) { 5886 struct rps_dev_flow voidflow, *rflow = &voidflow; 5887 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 5888 5889 if (cpu >= 0) { 5890 /* Will be handled, remove from list */ 5891 skb_list_del_init(skb); 5892 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5893 } 5894 } 5895 } 5896 #endif 5897 __netif_receive_skb_list(head); 5898 rcu_read_unlock(); 5899 } 5900 5901 /** 5902 * netif_receive_skb - process receive buffer from network 5903 * @skb: buffer to process 5904 * 5905 * netif_receive_skb() is the main receive data processing function. 5906 * It always succeeds. The buffer may be dropped during processing 5907 * for congestion control or by the protocol layers. 5908 * 5909 * This function may only be called from softirq context and interrupts 5910 * should be enabled. 5911 * 5912 * Return values (usually ignored): 5913 * NET_RX_SUCCESS: no congestion 5914 * NET_RX_DROP: packet was dropped 5915 */ 5916 int netif_receive_skb(struct sk_buff *skb) 5917 { 5918 int ret; 5919 5920 trace_netif_receive_skb_entry(skb); 5921 5922 ret = netif_receive_skb_internal(skb); 5923 trace_netif_receive_skb_exit(ret); 5924 5925 return ret; 5926 } 5927 EXPORT_SYMBOL(netif_receive_skb); 5928 5929 /** 5930 * netif_receive_skb_list - process many receive buffers from network 5931 * @head: list of skbs to process. 5932 * 5933 * Since return value of netif_receive_skb() is normally ignored, and 5934 * wouldn't be meaningful for a list, this function returns void. 5935 * 5936 * This function may only be called from softirq context and interrupts 5937 * should be enabled. 5938 */ 5939 void netif_receive_skb_list(struct list_head *head) 5940 { 5941 struct sk_buff *skb; 5942 5943 if (list_empty(head)) 5944 return; 5945 if (trace_netif_receive_skb_list_entry_enabled()) { 5946 list_for_each_entry(skb, head, list) 5947 trace_netif_receive_skb_list_entry(skb); 5948 } 5949 netif_receive_skb_list_internal(head); 5950 trace_netif_receive_skb_list_exit(0); 5951 } 5952 EXPORT_SYMBOL(netif_receive_skb_list); 5953 5954 static DEFINE_PER_CPU(struct work_struct, flush_works); 5955 5956 /* Network device is going away, flush any packets still pending */ 5957 static void flush_backlog(struct work_struct *work) 5958 { 5959 struct sk_buff *skb, *tmp; 5960 struct softnet_data *sd; 5961 5962 local_bh_disable(); 5963 sd = this_cpu_ptr(&softnet_data); 5964 5965 backlog_lock_irq_disable(sd); 5966 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 5967 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 5968 __skb_unlink(skb, &sd->input_pkt_queue); 5969 dev_kfree_skb_irq(skb); 5970 rps_input_queue_head_incr(sd); 5971 } 5972 } 5973 backlog_unlock_irq_enable(sd); 5974 5975 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 5976 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 5977 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 5978 __skb_unlink(skb, &sd->process_queue); 5979 kfree_skb(skb); 5980 rps_input_queue_head_incr(sd); 5981 } 5982 } 5983 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 5984 local_bh_enable(); 5985 } 5986 5987 static bool flush_required(int cpu) 5988 { 5989 #if IS_ENABLED(CONFIG_RPS) 5990 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 5991 bool do_flush; 5992 5993 backlog_lock_irq_disable(sd); 5994 5995 /* as insertion into process_queue happens with the rps lock held, 5996 * process_queue access may race only with dequeue 5997 */ 5998 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 5999 !skb_queue_empty_lockless(&sd->process_queue); 6000 backlog_unlock_irq_enable(sd); 6001 6002 return do_flush; 6003 #endif 6004 /* without RPS we can't safely check input_pkt_queue: during a 6005 * concurrent remote skb_queue_splice() we can detect as empty both 6006 * input_pkt_queue and process_queue even if the latter could end-up 6007 * containing a lot of packets. 6008 */ 6009 return true; 6010 } 6011 6012 static void flush_all_backlogs(void) 6013 { 6014 static cpumask_t flush_cpus; 6015 unsigned int cpu; 6016 6017 /* since we are under rtnl lock protection we can use static data 6018 * for the cpumask and avoid allocating on stack the possibly 6019 * large mask 6020 */ 6021 ASSERT_RTNL(); 6022 6023 cpus_read_lock(); 6024 6025 cpumask_clear(&flush_cpus); 6026 for_each_online_cpu(cpu) { 6027 if (flush_required(cpu)) { 6028 queue_work_on(cpu, system_highpri_wq, 6029 per_cpu_ptr(&flush_works, cpu)); 6030 cpumask_set_cpu(cpu, &flush_cpus); 6031 } 6032 } 6033 6034 /* we can have in flight packet[s] on the cpus we are not flushing, 6035 * synchronize_net() in unregister_netdevice_many() will take care of 6036 * them 6037 */ 6038 for_each_cpu(cpu, &flush_cpus) 6039 flush_work(per_cpu_ptr(&flush_works, cpu)); 6040 6041 cpus_read_unlock(); 6042 } 6043 6044 static void net_rps_send_ipi(struct softnet_data *remsd) 6045 { 6046 #ifdef CONFIG_RPS 6047 while (remsd) { 6048 struct softnet_data *next = remsd->rps_ipi_next; 6049 6050 if (cpu_online(remsd->cpu)) 6051 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6052 remsd = next; 6053 } 6054 #endif 6055 } 6056 6057 /* 6058 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6059 * Note: called with local irq disabled, but exits with local irq enabled. 6060 */ 6061 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6062 { 6063 #ifdef CONFIG_RPS 6064 struct softnet_data *remsd = sd->rps_ipi_list; 6065 6066 if (!use_backlog_threads() && remsd) { 6067 sd->rps_ipi_list = NULL; 6068 6069 local_irq_enable(); 6070 6071 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6072 net_rps_send_ipi(remsd); 6073 } else 6074 #endif 6075 local_irq_enable(); 6076 } 6077 6078 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6079 { 6080 #ifdef CONFIG_RPS 6081 return !use_backlog_threads() && sd->rps_ipi_list; 6082 #else 6083 return false; 6084 #endif 6085 } 6086 6087 static int process_backlog(struct napi_struct *napi, int quota) 6088 { 6089 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6090 bool again = true; 6091 int work = 0; 6092 6093 /* Check if we have pending ipi, its better to send them now, 6094 * not waiting net_rx_action() end. 6095 */ 6096 if (sd_has_rps_ipi_waiting(sd)) { 6097 local_irq_disable(); 6098 net_rps_action_and_irq_enable(sd); 6099 } 6100 6101 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6102 while (again) { 6103 struct sk_buff *skb; 6104 6105 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6106 while ((skb = __skb_dequeue(&sd->process_queue))) { 6107 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6108 rcu_read_lock(); 6109 __netif_receive_skb(skb); 6110 rcu_read_unlock(); 6111 if (++work >= quota) { 6112 rps_input_queue_head_add(sd, work); 6113 return work; 6114 } 6115 6116 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6117 } 6118 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6119 6120 backlog_lock_irq_disable(sd); 6121 if (skb_queue_empty(&sd->input_pkt_queue)) { 6122 /* 6123 * Inline a custom version of __napi_complete(). 6124 * only current cpu owns and manipulates this napi, 6125 * and NAPI_STATE_SCHED is the only possible flag set 6126 * on backlog. 6127 * We can use a plain write instead of clear_bit(), 6128 * and we dont need an smp_mb() memory barrier. 6129 */ 6130 napi->state &= NAPIF_STATE_THREADED; 6131 again = false; 6132 } else { 6133 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6134 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6135 &sd->process_queue); 6136 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6137 } 6138 backlog_unlock_irq_enable(sd); 6139 } 6140 6141 if (work) 6142 rps_input_queue_head_add(sd, work); 6143 return work; 6144 } 6145 6146 /** 6147 * __napi_schedule - schedule for receive 6148 * @n: entry to schedule 6149 * 6150 * The entry's receive function will be scheduled to run. 6151 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6152 */ 6153 void __napi_schedule(struct napi_struct *n) 6154 { 6155 unsigned long flags; 6156 6157 local_irq_save(flags); 6158 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6159 local_irq_restore(flags); 6160 } 6161 EXPORT_SYMBOL(__napi_schedule); 6162 6163 /** 6164 * napi_schedule_prep - check if napi can be scheduled 6165 * @n: napi context 6166 * 6167 * Test if NAPI routine is already running, and if not mark 6168 * it as running. This is used as a condition variable to 6169 * insure only one NAPI poll instance runs. We also make 6170 * sure there is no pending NAPI disable. 6171 */ 6172 bool napi_schedule_prep(struct napi_struct *n) 6173 { 6174 unsigned long new, val = READ_ONCE(n->state); 6175 6176 do { 6177 if (unlikely(val & NAPIF_STATE_DISABLE)) 6178 return false; 6179 new = val | NAPIF_STATE_SCHED; 6180 6181 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6182 * This was suggested by Alexander Duyck, as compiler 6183 * emits better code than : 6184 * if (val & NAPIF_STATE_SCHED) 6185 * new |= NAPIF_STATE_MISSED; 6186 */ 6187 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6188 NAPIF_STATE_MISSED; 6189 } while (!try_cmpxchg(&n->state, &val, new)); 6190 6191 return !(val & NAPIF_STATE_SCHED); 6192 } 6193 EXPORT_SYMBOL(napi_schedule_prep); 6194 6195 /** 6196 * __napi_schedule_irqoff - schedule for receive 6197 * @n: entry to schedule 6198 * 6199 * Variant of __napi_schedule() assuming hard irqs are masked. 6200 * 6201 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6202 * because the interrupt disabled assumption might not be true 6203 * due to force-threaded interrupts and spinlock substitution. 6204 */ 6205 void __napi_schedule_irqoff(struct napi_struct *n) 6206 { 6207 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6208 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6209 else 6210 __napi_schedule(n); 6211 } 6212 EXPORT_SYMBOL(__napi_schedule_irqoff); 6213 6214 bool napi_complete_done(struct napi_struct *n, int work_done) 6215 { 6216 unsigned long flags, val, new, timeout = 0; 6217 bool ret = true; 6218 6219 /* 6220 * 1) Don't let napi dequeue from the cpu poll list 6221 * just in case its running on a different cpu. 6222 * 2) If we are busy polling, do nothing here, we have 6223 * the guarantee we will be called later. 6224 */ 6225 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6226 NAPIF_STATE_IN_BUSY_POLL))) 6227 return false; 6228 6229 if (work_done) { 6230 if (n->gro_bitmask) 6231 timeout = READ_ONCE(n->dev->gro_flush_timeout); 6232 n->defer_hard_irqs_count = READ_ONCE(n->dev->napi_defer_hard_irqs); 6233 } 6234 if (n->defer_hard_irqs_count > 0) { 6235 n->defer_hard_irqs_count--; 6236 timeout = READ_ONCE(n->dev->gro_flush_timeout); 6237 if (timeout) 6238 ret = false; 6239 } 6240 if (n->gro_bitmask) { 6241 /* When the NAPI instance uses a timeout and keeps postponing 6242 * it, we need to bound somehow the time packets are kept in 6243 * the GRO layer 6244 */ 6245 napi_gro_flush(n, !!timeout); 6246 } 6247 6248 gro_normal_list(n); 6249 6250 if (unlikely(!list_empty(&n->poll_list))) { 6251 /* If n->poll_list is not empty, we need to mask irqs */ 6252 local_irq_save(flags); 6253 list_del_init(&n->poll_list); 6254 local_irq_restore(flags); 6255 } 6256 WRITE_ONCE(n->list_owner, -1); 6257 6258 val = READ_ONCE(n->state); 6259 do { 6260 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6261 6262 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6263 NAPIF_STATE_SCHED_THREADED | 6264 NAPIF_STATE_PREFER_BUSY_POLL); 6265 6266 /* If STATE_MISSED was set, leave STATE_SCHED set, 6267 * because we will call napi->poll() one more time. 6268 * This C code was suggested by Alexander Duyck to help gcc. 6269 */ 6270 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6271 NAPIF_STATE_SCHED; 6272 } while (!try_cmpxchg(&n->state, &val, new)); 6273 6274 if (unlikely(val & NAPIF_STATE_MISSED)) { 6275 __napi_schedule(n); 6276 return false; 6277 } 6278 6279 if (timeout) 6280 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6281 HRTIMER_MODE_REL_PINNED); 6282 return ret; 6283 } 6284 EXPORT_SYMBOL(napi_complete_done); 6285 6286 /* must be called under rcu_read_lock(), as we dont take a reference */ 6287 struct napi_struct *napi_by_id(unsigned int napi_id) 6288 { 6289 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 6290 struct napi_struct *napi; 6291 6292 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 6293 if (napi->napi_id == napi_id) 6294 return napi; 6295 6296 return NULL; 6297 } 6298 6299 static void skb_defer_free_flush(struct softnet_data *sd) 6300 { 6301 struct sk_buff *skb, *next; 6302 6303 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6304 if (!READ_ONCE(sd->defer_list)) 6305 return; 6306 6307 spin_lock(&sd->defer_lock); 6308 skb = sd->defer_list; 6309 sd->defer_list = NULL; 6310 sd->defer_count = 0; 6311 spin_unlock(&sd->defer_lock); 6312 6313 while (skb != NULL) { 6314 next = skb->next; 6315 napi_consume_skb(skb, 1); 6316 skb = next; 6317 } 6318 } 6319 6320 #if defined(CONFIG_NET_RX_BUSY_POLL) 6321 6322 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6323 { 6324 if (!skip_schedule) { 6325 gro_normal_list(napi); 6326 __napi_schedule(napi); 6327 return; 6328 } 6329 6330 if (napi->gro_bitmask) { 6331 /* flush too old packets 6332 * If HZ < 1000, flush all packets. 6333 */ 6334 napi_gro_flush(napi, HZ >= 1000); 6335 } 6336 6337 gro_normal_list(napi); 6338 clear_bit(NAPI_STATE_SCHED, &napi->state); 6339 } 6340 6341 enum { 6342 NAPI_F_PREFER_BUSY_POLL = 1, 6343 NAPI_F_END_ON_RESCHED = 2, 6344 }; 6345 6346 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6347 unsigned flags, u16 budget) 6348 { 6349 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6350 bool skip_schedule = false; 6351 unsigned long timeout; 6352 int rc; 6353 6354 /* Busy polling means there is a high chance device driver hard irq 6355 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6356 * set in napi_schedule_prep(). 6357 * Since we are about to call napi->poll() once more, we can safely 6358 * clear NAPI_STATE_MISSED. 6359 * 6360 * Note: x86 could use a single "lock and ..." instruction 6361 * to perform these two clear_bit() 6362 */ 6363 clear_bit(NAPI_STATE_MISSED, &napi->state); 6364 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6365 6366 local_bh_disable(); 6367 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6368 6369 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6370 napi->defer_hard_irqs_count = READ_ONCE(napi->dev->napi_defer_hard_irqs); 6371 timeout = READ_ONCE(napi->dev->gro_flush_timeout); 6372 if (napi->defer_hard_irqs_count && timeout) { 6373 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6374 skip_schedule = true; 6375 } 6376 } 6377 6378 /* All we really want here is to re-enable device interrupts. 6379 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6380 */ 6381 rc = napi->poll(napi, budget); 6382 /* We can't gro_normal_list() here, because napi->poll() might have 6383 * rearmed the napi (napi_complete_done()) in which case it could 6384 * already be running on another CPU. 6385 */ 6386 trace_napi_poll(napi, rc, budget); 6387 netpoll_poll_unlock(have_poll_lock); 6388 if (rc == budget) 6389 __busy_poll_stop(napi, skip_schedule); 6390 bpf_net_ctx_clear(bpf_net_ctx); 6391 local_bh_enable(); 6392 } 6393 6394 static void __napi_busy_loop(unsigned int napi_id, 6395 bool (*loop_end)(void *, unsigned long), 6396 void *loop_end_arg, unsigned flags, u16 budget) 6397 { 6398 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6399 int (*napi_poll)(struct napi_struct *napi, int budget); 6400 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6401 void *have_poll_lock = NULL; 6402 struct napi_struct *napi; 6403 6404 WARN_ON_ONCE(!rcu_read_lock_held()); 6405 6406 restart: 6407 napi_poll = NULL; 6408 6409 napi = napi_by_id(napi_id); 6410 if (!napi) 6411 return; 6412 6413 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6414 preempt_disable(); 6415 for (;;) { 6416 int work = 0; 6417 6418 local_bh_disable(); 6419 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6420 if (!napi_poll) { 6421 unsigned long val = READ_ONCE(napi->state); 6422 6423 /* If multiple threads are competing for this napi, 6424 * we avoid dirtying napi->state as much as we can. 6425 */ 6426 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6427 NAPIF_STATE_IN_BUSY_POLL)) { 6428 if (flags & NAPI_F_PREFER_BUSY_POLL) 6429 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6430 goto count; 6431 } 6432 if (cmpxchg(&napi->state, val, 6433 val | NAPIF_STATE_IN_BUSY_POLL | 6434 NAPIF_STATE_SCHED) != val) { 6435 if (flags & NAPI_F_PREFER_BUSY_POLL) 6436 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6437 goto count; 6438 } 6439 have_poll_lock = netpoll_poll_lock(napi); 6440 napi_poll = napi->poll; 6441 } 6442 work = napi_poll(napi, budget); 6443 trace_napi_poll(napi, work, budget); 6444 gro_normal_list(napi); 6445 count: 6446 if (work > 0) 6447 __NET_ADD_STATS(dev_net(napi->dev), 6448 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6449 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6450 bpf_net_ctx_clear(bpf_net_ctx); 6451 local_bh_enable(); 6452 6453 if (!loop_end || loop_end(loop_end_arg, start_time)) 6454 break; 6455 6456 if (unlikely(need_resched())) { 6457 if (flags & NAPI_F_END_ON_RESCHED) 6458 break; 6459 if (napi_poll) 6460 busy_poll_stop(napi, have_poll_lock, flags, budget); 6461 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6462 preempt_enable(); 6463 rcu_read_unlock(); 6464 cond_resched(); 6465 rcu_read_lock(); 6466 if (loop_end(loop_end_arg, start_time)) 6467 return; 6468 goto restart; 6469 } 6470 cpu_relax(); 6471 } 6472 if (napi_poll) 6473 busy_poll_stop(napi, have_poll_lock, flags, budget); 6474 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6475 preempt_enable(); 6476 } 6477 6478 void napi_busy_loop_rcu(unsigned int napi_id, 6479 bool (*loop_end)(void *, unsigned long), 6480 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6481 { 6482 unsigned flags = NAPI_F_END_ON_RESCHED; 6483 6484 if (prefer_busy_poll) 6485 flags |= NAPI_F_PREFER_BUSY_POLL; 6486 6487 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6488 } 6489 6490 void napi_busy_loop(unsigned int napi_id, 6491 bool (*loop_end)(void *, unsigned long), 6492 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6493 { 6494 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6495 6496 rcu_read_lock(); 6497 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6498 rcu_read_unlock(); 6499 } 6500 EXPORT_SYMBOL(napi_busy_loop); 6501 6502 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6503 6504 static void napi_hash_add(struct napi_struct *napi) 6505 { 6506 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6507 return; 6508 6509 spin_lock(&napi_hash_lock); 6510 6511 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6512 do { 6513 if (unlikely(++napi_gen_id < MIN_NAPI_ID)) 6514 napi_gen_id = MIN_NAPI_ID; 6515 } while (napi_by_id(napi_gen_id)); 6516 napi->napi_id = napi_gen_id; 6517 6518 hlist_add_head_rcu(&napi->napi_hash_node, 6519 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6520 6521 spin_unlock(&napi_hash_lock); 6522 } 6523 6524 /* Warning : caller is responsible to make sure rcu grace period 6525 * is respected before freeing memory containing @napi 6526 */ 6527 static void napi_hash_del(struct napi_struct *napi) 6528 { 6529 spin_lock(&napi_hash_lock); 6530 6531 hlist_del_init_rcu(&napi->napi_hash_node); 6532 6533 spin_unlock(&napi_hash_lock); 6534 } 6535 6536 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6537 { 6538 struct napi_struct *napi; 6539 6540 napi = container_of(timer, struct napi_struct, timer); 6541 6542 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6543 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6544 */ 6545 if (!napi_disable_pending(napi) && 6546 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6547 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6548 __napi_schedule_irqoff(napi); 6549 } 6550 6551 return HRTIMER_NORESTART; 6552 } 6553 6554 static void init_gro_hash(struct napi_struct *napi) 6555 { 6556 int i; 6557 6558 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6559 INIT_LIST_HEAD(&napi->gro_hash[i].list); 6560 napi->gro_hash[i].count = 0; 6561 } 6562 napi->gro_bitmask = 0; 6563 } 6564 6565 int dev_set_threaded(struct net_device *dev, bool threaded) 6566 { 6567 struct napi_struct *napi; 6568 int err = 0; 6569 6570 if (dev->threaded == threaded) 6571 return 0; 6572 6573 if (threaded) { 6574 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6575 if (!napi->thread) { 6576 err = napi_kthread_create(napi); 6577 if (err) { 6578 threaded = false; 6579 break; 6580 } 6581 } 6582 } 6583 } 6584 6585 WRITE_ONCE(dev->threaded, threaded); 6586 6587 /* Make sure kthread is created before THREADED bit 6588 * is set. 6589 */ 6590 smp_mb__before_atomic(); 6591 6592 /* Setting/unsetting threaded mode on a napi might not immediately 6593 * take effect, if the current napi instance is actively being 6594 * polled. In this case, the switch between threaded mode and 6595 * softirq mode will happen in the next round of napi_schedule(). 6596 * This should not cause hiccups/stalls to the live traffic. 6597 */ 6598 list_for_each_entry(napi, &dev->napi_list, dev_list) 6599 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6600 6601 return err; 6602 } 6603 EXPORT_SYMBOL(dev_set_threaded); 6604 6605 /** 6606 * netif_queue_set_napi - Associate queue with the napi 6607 * @dev: device to which NAPI and queue belong 6608 * @queue_index: Index of queue 6609 * @type: queue type as RX or TX 6610 * @napi: NAPI context, pass NULL to clear previously set NAPI 6611 * 6612 * Set queue with its corresponding napi context. This should be done after 6613 * registering the NAPI handler for the queue-vector and the queues have been 6614 * mapped to the corresponding interrupt vector. 6615 */ 6616 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6617 enum netdev_queue_type type, struct napi_struct *napi) 6618 { 6619 struct netdev_rx_queue *rxq; 6620 struct netdev_queue *txq; 6621 6622 if (WARN_ON_ONCE(napi && !napi->dev)) 6623 return; 6624 if (dev->reg_state >= NETREG_REGISTERED) 6625 ASSERT_RTNL(); 6626 6627 switch (type) { 6628 case NETDEV_QUEUE_TYPE_RX: 6629 rxq = __netif_get_rx_queue(dev, queue_index); 6630 rxq->napi = napi; 6631 return; 6632 case NETDEV_QUEUE_TYPE_TX: 6633 txq = netdev_get_tx_queue(dev, queue_index); 6634 txq->napi = napi; 6635 return; 6636 default: 6637 return; 6638 } 6639 } 6640 EXPORT_SYMBOL(netif_queue_set_napi); 6641 6642 void netif_napi_add_weight(struct net_device *dev, struct napi_struct *napi, 6643 int (*poll)(struct napi_struct *, int), int weight) 6644 { 6645 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 6646 return; 6647 6648 INIT_LIST_HEAD(&napi->poll_list); 6649 INIT_HLIST_NODE(&napi->napi_hash_node); 6650 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 6651 napi->timer.function = napi_watchdog; 6652 init_gro_hash(napi); 6653 napi->skb = NULL; 6654 INIT_LIST_HEAD(&napi->rx_list); 6655 napi->rx_count = 0; 6656 napi->poll = poll; 6657 if (weight > NAPI_POLL_WEIGHT) 6658 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 6659 weight); 6660 napi->weight = weight; 6661 napi->dev = dev; 6662 #ifdef CONFIG_NETPOLL 6663 napi->poll_owner = -1; 6664 #endif 6665 napi->list_owner = -1; 6666 set_bit(NAPI_STATE_SCHED, &napi->state); 6667 set_bit(NAPI_STATE_NPSVC, &napi->state); 6668 list_add_rcu(&napi->dev_list, &dev->napi_list); 6669 napi_hash_add(napi); 6670 napi_get_frags_check(napi); 6671 /* Create kthread for this napi if dev->threaded is set. 6672 * Clear dev->threaded if kthread creation failed so that 6673 * threaded mode will not be enabled in napi_enable(). 6674 */ 6675 if (dev->threaded && napi_kthread_create(napi)) 6676 dev->threaded = false; 6677 netif_napi_set_irq(napi, -1); 6678 } 6679 EXPORT_SYMBOL(netif_napi_add_weight); 6680 6681 void napi_disable(struct napi_struct *n) 6682 { 6683 unsigned long val, new; 6684 6685 might_sleep(); 6686 set_bit(NAPI_STATE_DISABLE, &n->state); 6687 6688 val = READ_ONCE(n->state); 6689 do { 6690 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 6691 usleep_range(20, 200); 6692 val = READ_ONCE(n->state); 6693 } 6694 6695 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 6696 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 6697 } while (!try_cmpxchg(&n->state, &val, new)); 6698 6699 hrtimer_cancel(&n->timer); 6700 6701 clear_bit(NAPI_STATE_DISABLE, &n->state); 6702 } 6703 EXPORT_SYMBOL(napi_disable); 6704 6705 /** 6706 * napi_enable - enable NAPI scheduling 6707 * @n: NAPI context 6708 * 6709 * Resume NAPI from being scheduled on this context. 6710 * Must be paired with napi_disable. 6711 */ 6712 void napi_enable(struct napi_struct *n) 6713 { 6714 unsigned long new, val = READ_ONCE(n->state); 6715 6716 do { 6717 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 6718 6719 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 6720 if (n->dev->threaded && n->thread) 6721 new |= NAPIF_STATE_THREADED; 6722 } while (!try_cmpxchg(&n->state, &val, new)); 6723 } 6724 EXPORT_SYMBOL(napi_enable); 6725 6726 static void flush_gro_hash(struct napi_struct *napi) 6727 { 6728 int i; 6729 6730 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6731 struct sk_buff *skb, *n; 6732 6733 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list) 6734 kfree_skb(skb); 6735 napi->gro_hash[i].count = 0; 6736 } 6737 } 6738 6739 /* Must be called in process context */ 6740 void __netif_napi_del(struct napi_struct *napi) 6741 { 6742 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 6743 return; 6744 6745 napi_hash_del(napi); 6746 list_del_rcu(&napi->dev_list); 6747 napi_free_frags(napi); 6748 6749 flush_gro_hash(napi); 6750 napi->gro_bitmask = 0; 6751 6752 if (napi->thread) { 6753 kthread_stop(napi->thread); 6754 napi->thread = NULL; 6755 } 6756 } 6757 EXPORT_SYMBOL(__netif_napi_del); 6758 6759 static int __napi_poll(struct napi_struct *n, bool *repoll) 6760 { 6761 int work, weight; 6762 6763 weight = n->weight; 6764 6765 /* This NAPI_STATE_SCHED test is for avoiding a race 6766 * with netpoll's poll_napi(). Only the entity which 6767 * obtains the lock and sees NAPI_STATE_SCHED set will 6768 * actually make the ->poll() call. Therefore we avoid 6769 * accidentally calling ->poll() when NAPI is not scheduled. 6770 */ 6771 work = 0; 6772 if (napi_is_scheduled(n)) { 6773 work = n->poll(n, weight); 6774 trace_napi_poll(n, work, weight); 6775 6776 xdp_do_check_flushed(n); 6777 } 6778 6779 if (unlikely(work > weight)) 6780 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 6781 n->poll, work, weight); 6782 6783 if (likely(work < weight)) 6784 return work; 6785 6786 /* Drivers must not modify the NAPI state if they 6787 * consume the entire weight. In such cases this code 6788 * still "owns" the NAPI instance and therefore can 6789 * move the instance around on the list at-will. 6790 */ 6791 if (unlikely(napi_disable_pending(n))) { 6792 napi_complete(n); 6793 return work; 6794 } 6795 6796 /* The NAPI context has more processing work, but busy-polling 6797 * is preferred. Exit early. 6798 */ 6799 if (napi_prefer_busy_poll(n)) { 6800 if (napi_complete_done(n, work)) { 6801 /* If timeout is not set, we need to make sure 6802 * that the NAPI is re-scheduled. 6803 */ 6804 napi_schedule(n); 6805 } 6806 return work; 6807 } 6808 6809 if (n->gro_bitmask) { 6810 /* flush too old packets 6811 * If HZ < 1000, flush all packets. 6812 */ 6813 napi_gro_flush(n, HZ >= 1000); 6814 } 6815 6816 gro_normal_list(n); 6817 6818 /* Some drivers may have called napi_schedule 6819 * prior to exhausting their budget. 6820 */ 6821 if (unlikely(!list_empty(&n->poll_list))) { 6822 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 6823 n->dev ? n->dev->name : "backlog"); 6824 return work; 6825 } 6826 6827 *repoll = true; 6828 6829 return work; 6830 } 6831 6832 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 6833 { 6834 bool do_repoll = false; 6835 void *have; 6836 int work; 6837 6838 list_del_init(&n->poll_list); 6839 6840 have = netpoll_poll_lock(n); 6841 6842 work = __napi_poll(n, &do_repoll); 6843 6844 if (do_repoll) 6845 list_add_tail(&n->poll_list, repoll); 6846 6847 netpoll_poll_unlock(have); 6848 6849 return work; 6850 } 6851 6852 static int napi_thread_wait(struct napi_struct *napi) 6853 { 6854 set_current_state(TASK_INTERRUPTIBLE); 6855 6856 while (!kthread_should_stop()) { 6857 /* Testing SCHED_THREADED bit here to make sure the current 6858 * kthread owns this napi and could poll on this napi. 6859 * Testing SCHED bit is not enough because SCHED bit might be 6860 * set by some other busy poll thread or by napi_disable(). 6861 */ 6862 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 6863 WARN_ON(!list_empty(&napi->poll_list)); 6864 __set_current_state(TASK_RUNNING); 6865 return 0; 6866 } 6867 6868 schedule(); 6869 set_current_state(TASK_INTERRUPTIBLE); 6870 } 6871 __set_current_state(TASK_RUNNING); 6872 6873 return -1; 6874 } 6875 6876 static void napi_threaded_poll_loop(struct napi_struct *napi) 6877 { 6878 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6879 struct softnet_data *sd; 6880 unsigned long last_qs = jiffies; 6881 6882 for (;;) { 6883 bool repoll = false; 6884 void *have; 6885 6886 local_bh_disable(); 6887 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6888 6889 sd = this_cpu_ptr(&softnet_data); 6890 sd->in_napi_threaded_poll = true; 6891 6892 have = netpoll_poll_lock(napi); 6893 __napi_poll(napi, &repoll); 6894 netpoll_poll_unlock(have); 6895 6896 sd->in_napi_threaded_poll = false; 6897 barrier(); 6898 6899 if (sd_has_rps_ipi_waiting(sd)) { 6900 local_irq_disable(); 6901 net_rps_action_and_irq_enable(sd); 6902 } 6903 skb_defer_free_flush(sd); 6904 bpf_net_ctx_clear(bpf_net_ctx); 6905 local_bh_enable(); 6906 6907 if (!repoll) 6908 break; 6909 6910 rcu_softirq_qs_periodic(last_qs); 6911 cond_resched(); 6912 } 6913 } 6914 6915 static int napi_threaded_poll(void *data) 6916 { 6917 struct napi_struct *napi = data; 6918 6919 while (!napi_thread_wait(napi)) 6920 napi_threaded_poll_loop(napi); 6921 6922 return 0; 6923 } 6924 6925 static __latent_entropy void net_rx_action(struct softirq_action *h) 6926 { 6927 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 6928 unsigned long time_limit = jiffies + 6929 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 6930 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6931 int budget = READ_ONCE(net_hotdata.netdev_budget); 6932 LIST_HEAD(list); 6933 LIST_HEAD(repoll); 6934 6935 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6936 start: 6937 sd->in_net_rx_action = true; 6938 local_irq_disable(); 6939 list_splice_init(&sd->poll_list, &list); 6940 local_irq_enable(); 6941 6942 for (;;) { 6943 struct napi_struct *n; 6944 6945 skb_defer_free_flush(sd); 6946 6947 if (list_empty(&list)) { 6948 if (list_empty(&repoll)) { 6949 sd->in_net_rx_action = false; 6950 barrier(); 6951 /* We need to check if ____napi_schedule() 6952 * had refilled poll_list while 6953 * sd->in_net_rx_action was true. 6954 */ 6955 if (!list_empty(&sd->poll_list)) 6956 goto start; 6957 if (!sd_has_rps_ipi_waiting(sd)) 6958 goto end; 6959 } 6960 break; 6961 } 6962 6963 n = list_first_entry(&list, struct napi_struct, poll_list); 6964 budget -= napi_poll(n, &repoll); 6965 6966 /* If softirq window is exhausted then punt. 6967 * Allow this to run for 2 jiffies since which will allow 6968 * an average latency of 1.5/HZ. 6969 */ 6970 if (unlikely(budget <= 0 || 6971 time_after_eq(jiffies, time_limit))) { 6972 sd->time_squeeze++; 6973 break; 6974 } 6975 } 6976 6977 local_irq_disable(); 6978 6979 list_splice_tail_init(&sd->poll_list, &list); 6980 list_splice_tail(&repoll, &list); 6981 list_splice(&list, &sd->poll_list); 6982 if (!list_empty(&sd->poll_list)) 6983 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 6984 else 6985 sd->in_net_rx_action = false; 6986 6987 net_rps_action_and_irq_enable(sd); 6988 end: 6989 bpf_net_ctx_clear(bpf_net_ctx); 6990 } 6991 6992 struct netdev_adjacent { 6993 struct net_device *dev; 6994 netdevice_tracker dev_tracker; 6995 6996 /* upper master flag, there can only be one master device per list */ 6997 bool master; 6998 6999 /* lookup ignore flag */ 7000 bool ignore; 7001 7002 /* counter for the number of times this device was added to us */ 7003 u16 ref_nr; 7004 7005 /* private field for the users */ 7006 void *private; 7007 7008 struct list_head list; 7009 struct rcu_head rcu; 7010 }; 7011 7012 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7013 struct list_head *adj_list) 7014 { 7015 struct netdev_adjacent *adj; 7016 7017 list_for_each_entry(adj, adj_list, list) { 7018 if (adj->dev == adj_dev) 7019 return adj; 7020 } 7021 return NULL; 7022 } 7023 7024 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7025 struct netdev_nested_priv *priv) 7026 { 7027 struct net_device *dev = (struct net_device *)priv->data; 7028 7029 return upper_dev == dev; 7030 } 7031 7032 /** 7033 * netdev_has_upper_dev - Check if device is linked to an upper device 7034 * @dev: device 7035 * @upper_dev: upper device to check 7036 * 7037 * Find out if a device is linked to specified upper device and return true 7038 * in case it is. Note that this checks only immediate upper device, 7039 * not through a complete stack of devices. The caller must hold the RTNL lock. 7040 */ 7041 bool netdev_has_upper_dev(struct net_device *dev, 7042 struct net_device *upper_dev) 7043 { 7044 struct netdev_nested_priv priv = { 7045 .data = (void *)upper_dev, 7046 }; 7047 7048 ASSERT_RTNL(); 7049 7050 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7051 &priv); 7052 } 7053 EXPORT_SYMBOL(netdev_has_upper_dev); 7054 7055 /** 7056 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7057 * @dev: device 7058 * @upper_dev: upper device to check 7059 * 7060 * Find out if a device is linked to specified upper device and return true 7061 * in case it is. Note that this checks the entire upper device chain. 7062 * The caller must hold rcu lock. 7063 */ 7064 7065 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 7066 struct net_device *upper_dev) 7067 { 7068 struct netdev_nested_priv priv = { 7069 .data = (void *)upper_dev, 7070 }; 7071 7072 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7073 &priv); 7074 } 7075 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 7076 7077 /** 7078 * netdev_has_any_upper_dev - Check if device is linked to some device 7079 * @dev: device 7080 * 7081 * Find out if a device is linked to an upper device and return true in case 7082 * it is. The caller must hold the RTNL lock. 7083 */ 7084 bool netdev_has_any_upper_dev(struct net_device *dev) 7085 { 7086 ASSERT_RTNL(); 7087 7088 return !list_empty(&dev->adj_list.upper); 7089 } 7090 EXPORT_SYMBOL(netdev_has_any_upper_dev); 7091 7092 /** 7093 * netdev_master_upper_dev_get - Get master upper device 7094 * @dev: device 7095 * 7096 * Find a master upper device and return pointer to it or NULL in case 7097 * it's not there. The caller must hold the RTNL lock. 7098 */ 7099 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 7100 { 7101 struct netdev_adjacent *upper; 7102 7103 ASSERT_RTNL(); 7104 7105 if (list_empty(&dev->adj_list.upper)) 7106 return NULL; 7107 7108 upper = list_first_entry(&dev->adj_list.upper, 7109 struct netdev_adjacent, list); 7110 if (likely(upper->master)) 7111 return upper->dev; 7112 return NULL; 7113 } 7114 EXPORT_SYMBOL(netdev_master_upper_dev_get); 7115 7116 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 7117 { 7118 struct netdev_adjacent *upper; 7119 7120 ASSERT_RTNL(); 7121 7122 if (list_empty(&dev->adj_list.upper)) 7123 return NULL; 7124 7125 upper = list_first_entry(&dev->adj_list.upper, 7126 struct netdev_adjacent, list); 7127 if (likely(upper->master) && !upper->ignore) 7128 return upper->dev; 7129 return NULL; 7130 } 7131 7132 /** 7133 * netdev_has_any_lower_dev - Check if device is linked to some device 7134 * @dev: device 7135 * 7136 * Find out if a device is linked to a lower device and return true in case 7137 * it is. The caller must hold the RTNL lock. 7138 */ 7139 static bool netdev_has_any_lower_dev(struct net_device *dev) 7140 { 7141 ASSERT_RTNL(); 7142 7143 return !list_empty(&dev->adj_list.lower); 7144 } 7145 7146 void *netdev_adjacent_get_private(struct list_head *adj_list) 7147 { 7148 struct netdev_adjacent *adj; 7149 7150 adj = list_entry(adj_list, struct netdev_adjacent, list); 7151 7152 return adj->private; 7153 } 7154 EXPORT_SYMBOL(netdev_adjacent_get_private); 7155 7156 /** 7157 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7158 * @dev: device 7159 * @iter: list_head ** of the current position 7160 * 7161 * Gets the next device from the dev's upper list, starting from iter 7162 * position. The caller must hold RCU read lock. 7163 */ 7164 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7165 struct list_head **iter) 7166 { 7167 struct netdev_adjacent *upper; 7168 7169 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7170 7171 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7172 7173 if (&upper->list == &dev->adj_list.upper) 7174 return NULL; 7175 7176 *iter = &upper->list; 7177 7178 return upper->dev; 7179 } 7180 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7181 7182 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7183 struct list_head **iter, 7184 bool *ignore) 7185 { 7186 struct netdev_adjacent *upper; 7187 7188 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7189 7190 if (&upper->list == &dev->adj_list.upper) 7191 return NULL; 7192 7193 *iter = &upper->list; 7194 *ignore = upper->ignore; 7195 7196 return upper->dev; 7197 } 7198 7199 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7200 struct list_head **iter) 7201 { 7202 struct netdev_adjacent *upper; 7203 7204 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7205 7206 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7207 7208 if (&upper->list == &dev->adj_list.upper) 7209 return NULL; 7210 7211 *iter = &upper->list; 7212 7213 return upper->dev; 7214 } 7215 7216 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7217 int (*fn)(struct net_device *dev, 7218 struct netdev_nested_priv *priv), 7219 struct netdev_nested_priv *priv) 7220 { 7221 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7222 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7223 int ret, cur = 0; 7224 bool ignore; 7225 7226 now = dev; 7227 iter = &dev->adj_list.upper; 7228 7229 while (1) { 7230 if (now != dev) { 7231 ret = fn(now, priv); 7232 if (ret) 7233 return ret; 7234 } 7235 7236 next = NULL; 7237 while (1) { 7238 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7239 if (!udev) 7240 break; 7241 if (ignore) 7242 continue; 7243 7244 next = udev; 7245 niter = &udev->adj_list.upper; 7246 dev_stack[cur] = now; 7247 iter_stack[cur++] = iter; 7248 break; 7249 } 7250 7251 if (!next) { 7252 if (!cur) 7253 return 0; 7254 next = dev_stack[--cur]; 7255 niter = iter_stack[cur]; 7256 } 7257 7258 now = next; 7259 iter = niter; 7260 } 7261 7262 return 0; 7263 } 7264 7265 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7266 int (*fn)(struct net_device *dev, 7267 struct netdev_nested_priv *priv), 7268 struct netdev_nested_priv *priv) 7269 { 7270 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7271 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7272 int ret, cur = 0; 7273 7274 now = dev; 7275 iter = &dev->adj_list.upper; 7276 7277 while (1) { 7278 if (now != dev) { 7279 ret = fn(now, priv); 7280 if (ret) 7281 return ret; 7282 } 7283 7284 next = NULL; 7285 while (1) { 7286 udev = netdev_next_upper_dev_rcu(now, &iter); 7287 if (!udev) 7288 break; 7289 7290 next = udev; 7291 niter = &udev->adj_list.upper; 7292 dev_stack[cur] = now; 7293 iter_stack[cur++] = iter; 7294 break; 7295 } 7296 7297 if (!next) { 7298 if (!cur) 7299 return 0; 7300 next = dev_stack[--cur]; 7301 niter = iter_stack[cur]; 7302 } 7303 7304 now = next; 7305 iter = niter; 7306 } 7307 7308 return 0; 7309 } 7310 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7311 7312 static bool __netdev_has_upper_dev(struct net_device *dev, 7313 struct net_device *upper_dev) 7314 { 7315 struct netdev_nested_priv priv = { 7316 .flags = 0, 7317 .data = (void *)upper_dev, 7318 }; 7319 7320 ASSERT_RTNL(); 7321 7322 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7323 &priv); 7324 } 7325 7326 /** 7327 * netdev_lower_get_next_private - Get the next ->private from the 7328 * lower neighbour list 7329 * @dev: device 7330 * @iter: list_head ** of the current position 7331 * 7332 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7333 * list, starting from iter position. The caller must hold either hold the 7334 * RTNL lock or its own locking that guarantees that the neighbour lower 7335 * list will remain unchanged. 7336 */ 7337 void *netdev_lower_get_next_private(struct net_device *dev, 7338 struct list_head **iter) 7339 { 7340 struct netdev_adjacent *lower; 7341 7342 lower = list_entry(*iter, struct netdev_adjacent, list); 7343 7344 if (&lower->list == &dev->adj_list.lower) 7345 return NULL; 7346 7347 *iter = lower->list.next; 7348 7349 return lower->private; 7350 } 7351 EXPORT_SYMBOL(netdev_lower_get_next_private); 7352 7353 /** 7354 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7355 * lower neighbour list, RCU 7356 * variant 7357 * @dev: device 7358 * @iter: list_head ** of the current position 7359 * 7360 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7361 * list, starting from iter position. The caller must hold RCU read lock. 7362 */ 7363 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7364 struct list_head **iter) 7365 { 7366 struct netdev_adjacent *lower; 7367 7368 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7369 7370 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7371 7372 if (&lower->list == &dev->adj_list.lower) 7373 return NULL; 7374 7375 *iter = &lower->list; 7376 7377 return lower->private; 7378 } 7379 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7380 7381 /** 7382 * netdev_lower_get_next - Get the next device from the lower neighbour 7383 * list 7384 * @dev: device 7385 * @iter: list_head ** of the current position 7386 * 7387 * Gets the next netdev_adjacent from the dev's lower neighbour 7388 * list, starting from iter position. The caller must hold RTNL lock or 7389 * its own locking that guarantees that the neighbour lower 7390 * list will remain unchanged. 7391 */ 7392 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7393 { 7394 struct netdev_adjacent *lower; 7395 7396 lower = list_entry(*iter, struct netdev_adjacent, list); 7397 7398 if (&lower->list == &dev->adj_list.lower) 7399 return NULL; 7400 7401 *iter = lower->list.next; 7402 7403 return lower->dev; 7404 } 7405 EXPORT_SYMBOL(netdev_lower_get_next); 7406 7407 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 7408 struct list_head **iter) 7409 { 7410 struct netdev_adjacent *lower; 7411 7412 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7413 7414 if (&lower->list == &dev->adj_list.lower) 7415 return NULL; 7416 7417 *iter = &lower->list; 7418 7419 return lower->dev; 7420 } 7421 7422 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 7423 struct list_head **iter, 7424 bool *ignore) 7425 { 7426 struct netdev_adjacent *lower; 7427 7428 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7429 7430 if (&lower->list == &dev->adj_list.lower) 7431 return NULL; 7432 7433 *iter = &lower->list; 7434 *ignore = lower->ignore; 7435 7436 return lower->dev; 7437 } 7438 7439 int netdev_walk_all_lower_dev(struct net_device *dev, 7440 int (*fn)(struct net_device *dev, 7441 struct netdev_nested_priv *priv), 7442 struct netdev_nested_priv *priv) 7443 { 7444 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7445 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7446 int ret, cur = 0; 7447 7448 now = dev; 7449 iter = &dev->adj_list.lower; 7450 7451 while (1) { 7452 if (now != dev) { 7453 ret = fn(now, priv); 7454 if (ret) 7455 return ret; 7456 } 7457 7458 next = NULL; 7459 while (1) { 7460 ldev = netdev_next_lower_dev(now, &iter); 7461 if (!ldev) 7462 break; 7463 7464 next = ldev; 7465 niter = &ldev->adj_list.lower; 7466 dev_stack[cur] = now; 7467 iter_stack[cur++] = iter; 7468 break; 7469 } 7470 7471 if (!next) { 7472 if (!cur) 7473 return 0; 7474 next = dev_stack[--cur]; 7475 niter = iter_stack[cur]; 7476 } 7477 7478 now = next; 7479 iter = niter; 7480 } 7481 7482 return 0; 7483 } 7484 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 7485 7486 static int __netdev_walk_all_lower_dev(struct net_device *dev, 7487 int (*fn)(struct net_device *dev, 7488 struct netdev_nested_priv *priv), 7489 struct netdev_nested_priv *priv) 7490 { 7491 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7492 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7493 int ret, cur = 0; 7494 bool ignore; 7495 7496 now = dev; 7497 iter = &dev->adj_list.lower; 7498 7499 while (1) { 7500 if (now != dev) { 7501 ret = fn(now, priv); 7502 if (ret) 7503 return ret; 7504 } 7505 7506 next = NULL; 7507 while (1) { 7508 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 7509 if (!ldev) 7510 break; 7511 if (ignore) 7512 continue; 7513 7514 next = ldev; 7515 niter = &ldev->adj_list.lower; 7516 dev_stack[cur] = now; 7517 iter_stack[cur++] = iter; 7518 break; 7519 } 7520 7521 if (!next) { 7522 if (!cur) 7523 return 0; 7524 next = dev_stack[--cur]; 7525 niter = iter_stack[cur]; 7526 } 7527 7528 now = next; 7529 iter = niter; 7530 } 7531 7532 return 0; 7533 } 7534 7535 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 7536 struct list_head **iter) 7537 { 7538 struct netdev_adjacent *lower; 7539 7540 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7541 if (&lower->list == &dev->adj_list.lower) 7542 return NULL; 7543 7544 *iter = &lower->list; 7545 7546 return lower->dev; 7547 } 7548 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 7549 7550 static u8 __netdev_upper_depth(struct net_device *dev) 7551 { 7552 struct net_device *udev; 7553 struct list_head *iter; 7554 u8 max_depth = 0; 7555 bool ignore; 7556 7557 for (iter = &dev->adj_list.upper, 7558 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 7559 udev; 7560 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 7561 if (ignore) 7562 continue; 7563 if (max_depth < udev->upper_level) 7564 max_depth = udev->upper_level; 7565 } 7566 7567 return max_depth; 7568 } 7569 7570 static u8 __netdev_lower_depth(struct net_device *dev) 7571 { 7572 struct net_device *ldev; 7573 struct list_head *iter; 7574 u8 max_depth = 0; 7575 bool ignore; 7576 7577 for (iter = &dev->adj_list.lower, 7578 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 7579 ldev; 7580 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 7581 if (ignore) 7582 continue; 7583 if (max_depth < ldev->lower_level) 7584 max_depth = ldev->lower_level; 7585 } 7586 7587 return max_depth; 7588 } 7589 7590 static int __netdev_update_upper_level(struct net_device *dev, 7591 struct netdev_nested_priv *__unused) 7592 { 7593 dev->upper_level = __netdev_upper_depth(dev) + 1; 7594 return 0; 7595 } 7596 7597 #ifdef CONFIG_LOCKDEP 7598 static LIST_HEAD(net_unlink_list); 7599 7600 static void net_unlink_todo(struct net_device *dev) 7601 { 7602 if (list_empty(&dev->unlink_list)) 7603 list_add_tail(&dev->unlink_list, &net_unlink_list); 7604 } 7605 #endif 7606 7607 static int __netdev_update_lower_level(struct net_device *dev, 7608 struct netdev_nested_priv *priv) 7609 { 7610 dev->lower_level = __netdev_lower_depth(dev) + 1; 7611 7612 #ifdef CONFIG_LOCKDEP 7613 if (!priv) 7614 return 0; 7615 7616 if (priv->flags & NESTED_SYNC_IMM) 7617 dev->nested_level = dev->lower_level - 1; 7618 if (priv->flags & NESTED_SYNC_TODO) 7619 net_unlink_todo(dev); 7620 #endif 7621 return 0; 7622 } 7623 7624 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 7625 int (*fn)(struct net_device *dev, 7626 struct netdev_nested_priv *priv), 7627 struct netdev_nested_priv *priv) 7628 { 7629 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7630 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7631 int ret, cur = 0; 7632 7633 now = dev; 7634 iter = &dev->adj_list.lower; 7635 7636 while (1) { 7637 if (now != dev) { 7638 ret = fn(now, priv); 7639 if (ret) 7640 return ret; 7641 } 7642 7643 next = NULL; 7644 while (1) { 7645 ldev = netdev_next_lower_dev_rcu(now, &iter); 7646 if (!ldev) 7647 break; 7648 7649 next = ldev; 7650 niter = &ldev->adj_list.lower; 7651 dev_stack[cur] = now; 7652 iter_stack[cur++] = iter; 7653 break; 7654 } 7655 7656 if (!next) { 7657 if (!cur) 7658 return 0; 7659 next = dev_stack[--cur]; 7660 niter = iter_stack[cur]; 7661 } 7662 7663 now = next; 7664 iter = niter; 7665 } 7666 7667 return 0; 7668 } 7669 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 7670 7671 /** 7672 * netdev_lower_get_first_private_rcu - Get the first ->private from the 7673 * lower neighbour list, RCU 7674 * variant 7675 * @dev: device 7676 * 7677 * Gets the first netdev_adjacent->private from the dev's lower neighbour 7678 * list. The caller must hold RCU read lock. 7679 */ 7680 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 7681 { 7682 struct netdev_adjacent *lower; 7683 7684 lower = list_first_or_null_rcu(&dev->adj_list.lower, 7685 struct netdev_adjacent, list); 7686 if (lower) 7687 return lower->private; 7688 return NULL; 7689 } 7690 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 7691 7692 /** 7693 * netdev_master_upper_dev_get_rcu - Get master upper device 7694 * @dev: device 7695 * 7696 * Find a master upper device and return pointer to it or NULL in case 7697 * it's not there. The caller must hold the RCU read lock. 7698 */ 7699 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 7700 { 7701 struct netdev_adjacent *upper; 7702 7703 upper = list_first_or_null_rcu(&dev->adj_list.upper, 7704 struct netdev_adjacent, list); 7705 if (upper && likely(upper->master)) 7706 return upper->dev; 7707 return NULL; 7708 } 7709 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 7710 7711 static int netdev_adjacent_sysfs_add(struct net_device *dev, 7712 struct net_device *adj_dev, 7713 struct list_head *dev_list) 7714 { 7715 char linkname[IFNAMSIZ+7]; 7716 7717 sprintf(linkname, dev_list == &dev->adj_list.upper ? 7718 "upper_%s" : "lower_%s", adj_dev->name); 7719 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 7720 linkname); 7721 } 7722 static void netdev_adjacent_sysfs_del(struct net_device *dev, 7723 char *name, 7724 struct list_head *dev_list) 7725 { 7726 char linkname[IFNAMSIZ+7]; 7727 7728 sprintf(linkname, dev_list == &dev->adj_list.upper ? 7729 "upper_%s" : "lower_%s", name); 7730 sysfs_remove_link(&(dev->dev.kobj), linkname); 7731 } 7732 7733 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 7734 struct net_device *adj_dev, 7735 struct list_head *dev_list) 7736 { 7737 return (dev_list == &dev->adj_list.upper || 7738 dev_list == &dev->adj_list.lower) && 7739 net_eq(dev_net(dev), dev_net(adj_dev)); 7740 } 7741 7742 static int __netdev_adjacent_dev_insert(struct net_device *dev, 7743 struct net_device *adj_dev, 7744 struct list_head *dev_list, 7745 void *private, bool master) 7746 { 7747 struct netdev_adjacent *adj; 7748 int ret; 7749 7750 adj = __netdev_find_adj(adj_dev, dev_list); 7751 7752 if (adj) { 7753 adj->ref_nr += 1; 7754 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 7755 dev->name, adj_dev->name, adj->ref_nr); 7756 7757 return 0; 7758 } 7759 7760 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 7761 if (!adj) 7762 return -ENOMEM; 7763 7764 adj->dev = adj_dev; 7765 adj->master = master; 7766 adj->ref_nr = 1; 7767 adj->private = private; 7768 adj->ignore = false; 7769 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 7770 7771 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 7772 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 7773 7774 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 7775 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 7776 if (ret) 7777 goto free_adj; 7778 } 7779 7780 /* Ensure that master link is always the first item in list. */ 7781 if (master) { 7782 ret = sysfs_create_link(&(dev->dev.kobj), 7783 &(adj_dev->dev.kobj), "master"); 7784 if (ret) 7785 goto remove_symlinks; 7786 7787 list_add_rcu(&adj->list, dev_list); 7788 } else { 7789 list_add_tail_rcu(&adj->list, dev_list); 7790 } 7791 7792 return 0; 7793 7794 remove_symlinks: 7795 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 7796 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 7797 free_adj: 7798 netdev_put(adj_dev, &adj->dev_tracker); 7799 kfree(adj); 7800 7801 return ret; 7802 } 7803 7804 static void __netdev_adjacent_dev_remove(struct net_device *dev, 7805 struct net_device *adj_dev, 7806 u16 ref_nr, 7807 struct list_head *dev_list) 7808 { 7809 struct netdev_adjacent *adj; 7810 7811 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 7812 dev->name, adj_dev->name, ref_nr); 7813 7814 adj = __netdev_find_adj(adj_dev, dev_list); 7815 7816 if (!adj) { 7817 pr_err("Adjacency does not exist for device %s from %s\n", 7818 dev->name, adj_dev->name); 7819 WARN_ON(1); 7820 return; 7821 } 7822 7823 if (adj->ref_nr > ref_nr) { 7824 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 7825 dev->name, adj_dev->name, ref_nr, 7826 adj->ref_nr - ref_nr); 7827 adj->ref_nr -= ref_nr; 7828 return; 7829 } 7830 7831 if (adj->master) 7832 sysfs_remove_link(&(dev->dev.kobj), "master"); 7833 7834 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 7835 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 7836 7837 list_del_rcu(&adj->list); 7838 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 7839 adj_dev->name, dev->name, adj_dev->name); 7840 netdev_put(adj_dev, &adj->dev_tracker); 7841 kfree_rcu(adj, rcu); 7842 } 7843 7844 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 7845 struct net_device *upper_dev, 7846 struct list_head *up_list, 7847 struct list_head *down_list, 7848 void *private, bool master) 7849 { 7850 int ret; 7851 7852 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 7853 private, master); 7854 if (ret) 7855 return ret; 7856 7857 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 7858 private, false); 7859 if (ret) { 7860 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 7861 return ret; 7862 } 7863 7864 return 0; 7865 } 7866 7867 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 7868 struct net_device *upper_dev, 7869 u16 ref_nr, 7870 struct list_head *up_list, 7871 struct list_head *down_list) 7872 { 7873 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 7874 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 7875 } 7876 7877 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 7878 struct net_device *upper_dev, 7879 void *private, bool master) 7880 { 7881 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 7882 &dev->adj_list.upper, 7883 &upper_dev->adj_list.lower, 7884 private, master); 7885 } 7886 7887 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 7888 struct net_device *upper_dev) 7889 { 7890 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 7891 &dev->adj_list.upper, 7892 &upper_dev->adj_list.lower); 7893 } 7894 7895 static int __netdev_upper_dev_link(struct net_device *dev, 7896 struct net_device *upper_dev, bool master, 7897 void *upper_priv, void *upper_info, 7898 struct netdev_nested_priv *priv, 7899 struct netlink_ext_ack *extack) 7900 { 7901 struct netdev_notifier_changeupper_info changeupper_info = { 7902 .info = { 7903 .dev = dev, 7904 .extack = extack, 7905 }, 7906 .upper_dev = upper_dev, 7907 .master = master, 7908 .linking = true, 7909 .upper_info = upper_info, 7910 }; 7911 struct net_device *master_dev; 7912 int ret = 0; 7913 7914 ASSERT_RTNL(); 7915 7916 if (dev == upper_dev) 7917 return -EBUSY; 7918 7919 /* To prevent loops, check if dev is not upper device to upper_dev. */ 7920 if (__netdev_has_upper_dev(upper_dev, dev)) 7921 return -EBUSY; 7922 7923 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 7924 return -EMLINK; 7925 7926 if (!master) { 7927 if (__netdev_has_upper_dev(dev, upper_dev)) 7928 return -EEXIST; 7929 } else { 7930 master_dev = __netdev_master_upper_dev_get(dev); 7931 if (master_dev) 7932 return master_dev == upper_dev ? -EEXIST : -EBUSY; 7933 } 7934 7935 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 7936 &changeupper_info.info); 7937 ret = notifier_to_errno(ret); 7938 if (ret) 7939 return ret; 7940 7941 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 7942 master); 7943 if (ret) 7944 return ret; 7945 7946 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 7947 &changeupper_info.info); 7948 ret = notifier_to_errno(ret); 7949 if (ret) 7950 goto rollback; 7951 7952 __netdev_update_upper_level(dev, NULL); 7953 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 7954 7955 __netdev_update_lower_level(upper_dev, priv); 7956 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 7957 priv); 7958 7959 return 0; 7960 7961 rollback: 7962 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 7963 7964 return ret; 7965 } 7966 7967 /** 7968 * netdev_upper_dev_link - Add a link to the upper device 7969 * @dev: device 7970 * @upper_dev: new upper device 7971 * @extack: netlink extended ack 7972 * 7973 * Adds a link to device which is upper to this one. The caller must hold 7974 * the RTNL lock. On a failure a negative errno code is returned. 7975 * On success the reference counts are adjusted and the function 7976 * returns zero. 7977 */ 7978 int netdev_upper_dev_link(struct net_device *dev, 7979 struct net_device *upper_dev, 7980 struct netlink_ext_ack *extack) 7981 { 7982 struct netdev_nested_priv priv = { 7983 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 7984 .data = NULL, 7985 }; 7986 7987 return __netdev_upper_dev_link(dev, upper_dev, false, 7988 NULL, NULL, &priv, extack); 7989 } 7990 EXPORT_SYMBOL(netdev_upper_dev_link); 7991 7992 /** 7993 * netdev_master_upper_dev_link - Add a master link to the upper device 7994 * @dev: device 7995 * @upper_dev: new upper device 7996 * @upper_priv: upper device private 7997 * @upper_info: upper info to be passed down via notifier 7998 * @extack: netlink extended ack 7999 * 8000 * Adds a link to device which is upper to this one. In this case, only 8001 * one master upper device can be linked, although other non-master devices 8002 * might be linked as well. The caller must hold the RTNL lock. 8003 * On a failure a negative errno code is returned. On success the reference 8004 * counts are adjusted and the function returns zero. 8005 */ 8006 int netdev_master_upper_dev_link(struct net_device *dev, 8007 struct net_device *upper_dev, 8008 void *upper_priv, void *upper_info, 8009 struct netlink_ext_ack *extack) 8010 { 8011 struct netdev_nested_priv priv = { 8012 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8013 .data = NULL, 8014 }; 8015 8016 return __netdev_upper_dev_link(dev, upper_dev, true, 8017 upper_priv, upper_info, &priv, extack); 8018 } 8019 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8020 8021 static void __netdev_upper_dev_unlink(struct net_device *dev, 8022 struct net_device *upper_dev, 8023 struct netdev_nested_priv *priv) 8024 { 8025 struct netdev_notifier_changeupper_info changeupper_info = { 8026 .info = { 8027 .dev = dev, 8028 }, 8029 .upper_dev = upper_dev, 8030 .linking = false, 8031 }; 8032 8033 ASSERT_RTNL(); 8034 8035 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8036 8037 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8038 &changeupper_info.info); 8039 8040 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8041 8042 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8043 &changeupper_info.info); 8044 8045 __netdev_update_upper_level(dev, NULL); 8046 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8047 8048 __netdev_update_lower_level(upper_dev, priv); 8049 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8050 priv); 8051 } 8052 8053 /** 8054 * netdev_upper_dev_unlink - Removes a link to upper device 8055 * @dev: device 8056 * @upper_dev: new upper device 8057 * 8058 * Removes a link to device which is upper to this one. The caller must hold 8059 * the RTNL lock. 8060 */ 8061 void netdev_upper_dev_unlink(struct net_device *dev, 8062 struct net_device *upper_dev) 8063 { 8064 struct netdev_nested_priv priv = { 8065 .flags = NESTED_SYNC_TODO, 8066 .data = NULL, 8067 }; 8068 8069 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 8070 } 8071 EXPORT_SYMBOL(netdev_upper_dev_unlink); 8072 8073 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 8074 struct net_device *lower_dev, 8075 bool val) 8076 { 8077 struct netdev_adjacent *adj; 8078 8079 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 8080 if (adj) 8081 adj->ignore = val; 8082 8083 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 8084 if (adj) 8085 adj->ignore = val; 8086 } 8087 8088 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 8089 struct net_device *lower_dev) 8090 { 8091 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 8092 } 8093 8094 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 8095 struct net_device *lower_dev) 8096 { 8097 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 8098 } 8099 8100 int netdev_adjacent_change_prepare(struct net_device *old_dev, 8101 struct net_device *new_dev, 8102 struct net_device *dev, 8103 struct netlink_ext_ack *extack) 8104 { 8105 struct netdev_nested_priv priv = { 8106 .flags = 0, 8107 .data = NULL, 8108 }; 8109 int err; 8110 8111 if (!new_dev) 8112 return 0; 8113 8114 if (old_dev && new_dev != old_dev) 8115 netdev_adjacent_dev_disable(dev, old_dev); 8116 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 8117 extack); 8118 if (err) { 8119 if (old_dev && new_dev != old_dev) 8120 netdev_adjacent_dev_enable(dev, old_dev); 8121 return err; 8122 } 8123 8124 return 0; 8125 } 8126 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 8127 8128 void netdev_adjacent_change_commit(struct net_device *old_dev, 8129 struct net_device *new_dev, 8130 struct net_device *dev) 8131 { 8132 struct netdev_nested_priv priv = { 8133 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8134 .data = NULL, 8135 }; 8136 8137 if (!new_dev || !old_dev) 8138 return; 8139 8140 if (new_dev == old_dev) 8141 return; 8142 8143 netdev_adjacent_dev_enable(dev, old_dev); 8144 __netdev_upper_dev_unlink(old_dev, dev, &priv); 8145 } 8146 EXPORT_SYMBOL(netdev_adjacent_change_commit); 8147 8148 void netdev_adjacent_change_abort(struct net_device *old_dev, 8149 struct net_device *new_dev, 8150 struct net_device *dev) 8151 { 8152 struct netdev_nested_priv priv = { 8153 .flags = 0, 8154 .data = NULL, 8155 }; 8156 8157 if (!new_dev) 8158 return; 8159 8160 if (old_dev && new_dev != old_dev) 8161 netdev_adjacent_dev_enable(dev, old_dev); 8162 8163 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8164 } 8165 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8166 8167 /** 8168 * netdev_bonding_info_change - Dispatch event about slave change 8169 * @dev: device 8170 * @bonding_info: info to dispatch 8171 * 8172 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8173 * The caller must hold the RTNL lock. 8174 */ 8175 void netdev_bonding_info_change(struct net_device *dev, 8176 struct netdev_bonding_info *bonding_info) 8177 { 8178 struct netdev_notifier_bonding_info info = { 8179 .info.dev = dev, 8180 }; 8181 8182 memcpy(&info.bonding_info, bonding_info, 8183 sizeof(struct netdev_bonding_info)); 8184 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8185 &info.info); 8186 } 8187 EXPORT_SYMBOL(netdev_bonding_info_change); 8188 8189 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8190 struct netlink_ext_ack *extack) 8191 { 8192 struct netdev_notifier_offload_xstats_info info = { 8193 .info.dev = dev, 8194 .info.extack = extack, 8195 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8196 }; 8197 int err; 8198 int rc; 8199 8200 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8201 GFP_KERNEL); 8202 if (!dev->offload_xstats_l3) 8203 return -ENOMEM; 8204 8205 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8206 NETDEV_OFFLOAD_XSTATS_DISABLE, 8207 &info.info); 8208 err = notifier_to_errno(rc); 8209 if (err) 8210 goto free_stats; 8211 8212 return 0; 8213 8214 free_stats: 8215 kfree(dev->offload_xstats_l3); 8216 dev->offload_xstats_l3 = NULL; 8217 return err; 8218 } 8219 8220 int netdev_offload_xstats_enable(struct net_device *dev, 8221 enum netdev_offload_xstats_type type, 8222 struct netlink_ext_ack *extack) 8223 { 8224 ASSERT_RTNL(); 8225 8226 if (netdev_offload_xstats_enabled(dev, type)) 8227 return -EALREADY; 8228 8229 switch (type) { 8230 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8231 return netdev_offload_xstats_enable_l3(dev, extack); 8232 } 8233 8234 WARN_ON(1); 8235 return -EINVAL; 8236 } 8237 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8238 8239 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8240 { 8241 struct netdev_notifier_offload_xstats_info info = { 8242 .info.dev = dev, 8243 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8244 }; 8245 8246 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8247 &info.info); 8248 kfree(dev->offload_xstats_l3); 8249 dev->offload_xstats_l3 = NULL; 8250 } 8251 8252 int netdev_offload_xstats_disable(struct net_device *dev, 8253 enum netdev_offload_xstats_type type) 8254 { 8255 ASSERT_RTNL(); 8256 8257 if (!netdev_offload_xstats_enabled(dev, type)) 8258 return -EALREADY; 8259 8260 switch (type) { 8261 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8262 netdev_offload_xstats_disable_l3(dev); 8263 return 0; 8264 } 8265 8266 WARN_ON(1); 8267 return -EINVAL; 8268 } 8269 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8270 8271 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8272 { 8273 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8274 } 8275 8276 static struct rtnl_hw_stats64 * 8277 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8278 enum netdev_offload_xstats_type type) 8279 { 8280 switch (type) { 8281 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8282 return dev->offload_xstats_l3; 8283 } 8284 8285 WARN_ON(1); 8286 return NULL; 8287 } 8288 8289 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8290 enum netdev_offload_xstats_type type) 8291 { 8292 ASSERT_RTNL(); 8293 8294 return netdev_offload_xstats_get_ptr(dev, type); 8295 } 8296 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8297 8298 struct netdev_notifier_offload_xstats_ru { 8299 bool used; 8300 }; 8301 8302 struct netdev_notifier_offload_xstats_rd { 8303 struct rtnl_hw_stats64 stats; 8304 bool used; 8305 }; 8306 8307 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8308 const struct rtnl_hw_stats64 *src) 8309 { 8310 dest->rx_packets += src->rx_packets; 8311 dest->tx_packets += src->tx_packets; 8312 dest->rx_bytes += src->rx_bytes; 8313 dest->tx_bytes += src->tx_bytes; 8314 dest->rx_errors += src->rx_errors; 8315 dest->tx_errors += src->tx_errors; 8316 dest->rx_dropped += src->rx_dropped; 8317 dest->tx_dropped += src->tx_dropped; 8318 dest->multicast += src->multicast; 8319 } 8320 8321 static int netdev_offload_xstats_get_used(struct net_device *dev, 8322 enum netdev_offload_xstats_type type, 8323 bool *p_used, 8324 struct netlink_ext_ack *extack) 8325 { 8326 struct netdev_notifier_offload_xstats_ru report_used = {}; 8327 struct netdev_notifier_offload_xstats_info info = { 8328 .info.dev = dev, 8329 .info.extack = extack, 8330 .type = type, 8331 .report_used = &report_used, 8332 }; 8333 int rc; 8334 8335 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8336 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8337 &info.info); 8338 *p_used = report_used.used; 8339 return notifier_to_errno(rc); 8340 } 8341 8342 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8343 enum netdev_offload_xstats_type type, 8344 struct rtnl_hw_stats64 *p_stats, 8345 bool *p_used, 8346 struct netlink_ext_ack *extack) 8347 { 8348 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8349 struct netdev_notifier_offload_xstats_info info = { 8350 .info.dev = dev, 8351 .info.extack = extack, 8352 .type = type, 8353 .report_delta = &report_delta, 8354 }; 8355 struct rtnl_hw_stats64 *stats; 8356 int rc; 8357 8358 stats = netdev_offload_xstats_get_ptr(dev, type); 8359 if (WARN_ON(!stats)) 8360 return -EINVAL; 8361 8362 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8363 &info.info); 8364 8365 /* Cache whatever we got, even if there was an error, otherwise the 8366 * successful stats retrievals would get lost. 8367 */ 8368 netdev_hw_stats64_add(stats, &report_delta.stats); 8369 8370 if (p_stats) 8371 *p_stats = *stats; 8372 *p_used = report_delta.used; 8373 8374 return notifier_to_errno(rc); 8375 } 8376 8377 int netdev_offload_xstats_get(struct net_device *dev, 8378 enum netdev_offload_xstats_type type, 8379 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8380 struct netlink_ext_ack *extack) 8381 { 8382 ASSERT_RTNL(); 8383 8384 if (p_stats) 8385 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8386 p_used, extack); 8387 else 8388 return netdev_offload_xstats_get_used(dev, type, p_used, 8389 extack); 8390 } 8391 EXPORT_SYMBOL(netdev_offload_xstats_get); 8392 8393 void 8394 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8395 const struct rtnl_hw_stats64 *stats) 8396 { 8397 report_delta->used = true; 8398 netdev_hw_stats64_add(&report_delta->stats, stats); 8399 } 8400 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 8401 8402 void 8403 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 8404 { 8405 report_used->used = true; 8406 } 8407 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 8408 8409 void netdev_offload_xstats_push_delta(struct net_device *dev, 8410 enum netdev_offload_xstats_type type, 8411 const struct rtnl_hw_stats64 *p_stats) 8412 { 8413 struct rtnl_hw_stats64 *stats; 8414 8415 ASSERT_RTNL(); 8416 8417 stats = netdev_offload_xstats_get_ptr(dev, type); 8418 if (WARN_ON(!stats)) 8419 return; 8420 8421 netdev_hw_stats64_add(stats, p_stats); 8422 } 8423 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 8424 8425 /** 8426 * netdev_get_xmit_slave - Get the xmit slave of master device 8427 * @dev: device 8428 * @skb: The packet 8429 * @all_slaves: assume all the slaves are active 8430 * 8431 * The reference counters are not incremented so the caller must be 8432 * careful with locks. The caller must hold RCU lock. 8433 * %NULL is returned if no slave is found. 8434 */ 8435 8436 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 8437 struct sk_buff *skb, 8438 bool all_slaves) 8439 { 8440 const struct net_device_ops *ops = dev->netdev_ops; 8441 8442 if (!ops->ndo_get_xmit_slave) 8443 return NULL; 8444 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 8445 } 8446 EXPORT_SYMBOL(netdev_get_xmit_slave); 8447 8448 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 8449 struct sock *sk) 8450 { 8451 const struct net_device_ops *ops = dev->netdev_ops; 8452 8453 if (!ops->ndo_sk_get_lower_dev) 8454 return NULL; 8455 return ops->ndo_sk_get_lower_dev(dev, sk); 8456 } 8457 8458 /** 8459 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 8460 * @dev: device 8461 * @sk: the socket 8462 * 8463 * %NULL is returned if no lower device is found. 8464 */ 8465 8466 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 8467 struct sock *sk) 8468 { 8469 struct net_device *lower; 8470 8471 lower = netdev_sk_get_lower_dev(dev, sk); 8472 while (lower) { 8473 dev = lower; 8474 lower = netdev_sk_get_lower_dev(dev, sk); 8475 } 8476 8477 return dev; 8478 } 8479 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 8480 8481 static void netdev_adjacent_add_links(struct net_device *dev) 8482 { 8483 struct netdev_adjacent *iter; 8484 8485 struct net *net = dev_net(dev); 8486 8487 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8488 if (!net_eq(net, dev_net(iter->dev))) 8489 continue; 8490 netdev_adjacent_sysfs_add(iter->dev, dev, 8491 &iter->dev->adj_list.lower); 8492 netdev_adjacent_sysfs_add(dev, iter->dev, 8493 &dev->adj_list.upper); 8494 } 8495 8496 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8497 if (!net_eq(net, dev_net(iter->dev))) 8498 continue; 8499 netdev_adjacent_sysfs_add(iter->dev, dev, 8500 &iter->dev->adj_list.upper); 8501 netdev_adjacent_sysfs_add(dev, iter->dev, 8502 &dev->adj_list.lower); 8503 } 8504 } 8505 8506 static void netdev_adjacent_del_links(struct net_device *dev) 8507 { 8508 struct netdev_adjacent *iter; 8509 8510 struct net *net = dev_net(dev); 8511 8512 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8513 if (!net_eq(net, dev_net(iter->dev))) 8514 continue; 8515 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8516 &iter->dev->adj_list.lower); 8517 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8518 &dev->adj_list.upper); 8519 } 8520 8521 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8522 if (!net_eq(net, dev_net(iter->dev))) 8523 continue; 8524 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8525 &iter->dev->adj_list.upper); 8526 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8527 &dev->adj_list.lower); 8528 } 8529 } 8530 8531 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 8532 { 8533 struct netdev_adjacent *iter; 8534 8535 struct net *net = dev_net(dev); 8536 8537 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8538 if (!net_eq(net, dev_net(iter->dev))) 8539 continue; 8540 netdev_adjacent_sysfs_del(iter->dev, oldname, 8541 &iter->dev->adj_list.lower); 8542 netdev_adjacent_sysfs_add(iter->dev, dev, 8543 &iter->dev->adj_list.lower); 8544 } 8545 8546 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8547 if (!net_eq(net, dev_net(iter->dev))) 8548 continue; 8549 netdev_adjacent_sysfs_del(iter->dev, oldname, 8550 &iter->dev->adj_list.upper); 8551 netdev_adjacent_sysfs_add(iter->dev, dev, 8552 &iter->dev->adj_list.upper); 8553 } 8554 } 8555 8556 void *netdev_lower_dev_get_private(struct net_device *dev, 8557 struct net_device *lower_dev) 8558 { 8559 struct netdev_adjacent *lower; 8560 8561 if (!lower_dev) 8562 return NULL; 8563 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 8564 if (!lower) 8565 return NULL; 8566 8567 return lower->private; 8568 } 8569 EXPORT_SYMBOL(netdev_lower_dev_get_private); 8570 8571 8572 /** 8573 * netdev_lower_state_changed - Dispatch event about lower device state change 8574 * @lower_dev: device 8575 * @lower_state_info: state to dispatch 8576 * 8577 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 8578 * The caller must hold the RTNL lock. 8579 */ 8580 void netdev_lower_state_changed(struct net_device *lower_dev, 8581 void *lower_state_info) 8582 { 8583 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 8584 .info.dev = lower_dev, 8585 }; 8586 8587 ASSERT_RTNL(); 8588 changelowerstate_info.lower_state_info = lower_state_info; 8589 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 8590 &changelowerstate_info.info); 8591 } 8592 EXPORT_SYMBOL(netdev_lower_state_changed); 8593 8594 static void dev_change_rx_flags(struct net_device *dev, int flags) 8595 { 8596 const struct net_device_ops *ops = dev->netdev_ops; 8597 8598 if (ops->ndo_change_rx_flags) 8599 ops->ndo_change_rx_flags(dev, flags); 8600 } 8601 8602 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 8603 { 8604 unsigned int old_flags = dev->flags; 8605 unsigned int promiscuity, flags; 8606 kuid_t uid; 8607 kgid_t gid; 8608 8609 ASSERT_RTNL(); 8610 8611 promiscuity = dev->promiscuity + inc; 8612 if (promiscuity == 0) { 8613 /* 8614 * Avoid overflow. 8615 * If inc causes overflow, untouch promisc and return error. 8616 */ 8617 if (unlikely(inc > 0)) { 8618 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 8619 return -EOVERFLOW; 8620 } 8621 flags = old_flags & ~IFF_PROMISC; 8622 } else { 8623 flags = old_flags | IFF_PROMISC; 8624 } 8625 WRITE_ONCE(dev->promiscuity, promiscuity); 8626 if (flags != old_flags) { 8627 WRITE_ONCE(dev->flags, flags); 8628 netdev_info(dev, "%s promiscuous mode\n", 8629 dev->flags & IFF_PROMISC ? "entered" : "left"); 8630 if (audit_enabled) { 8631 current_uid_gid(&uid, &gid); 8632 audit_log(audit_context(), GFP_ATOMIC, 8633 AUDIT_ANOM_PROMISCUOUS, 8634 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 8635 dev->name, (dev->flags & IFF_PROMISC), 8636 (old_flags & IFF_PROMISC), 8637 from_kuid(&init_user_ns, audit_get_loginuid(current)), 8638 from_kuid(&init_user_ns, uid), 8639 from_kgid(&init_user_ns, gid), 8640 audit_get_sessionid(current)); 8641 } 8642 8643 dev_change_rx_flags(dev, IFF_PROMISC); 8644 } 8645 if (notify) 8646 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 8647 return 0; 8648 } 8649 8650 /** 8651 * dev_set_promiscuity - update promiscuity count on a device 8652 * @dev: device 8653 * @inc: modifier 8654 * 8655 * Add or remove promiscuity from a device. While the count in the device 8656 * remains above zero the interface remains promiscuous. Once it hits zero 8657 * the device reverts back to normal filtering operation. A negative inc 8658 * value is used to drop promiscuity on the device. 8659 * Return 0 if successful or a negative errno code on error. 8660 */ 8661 int dev_set_promiscuity(struct net_device *dev, int inc) 8662 { 8663 unsigned int old_flags = dev->flags; 8664 int err; 8665 8666 err = __dev_set_promiscuity(dev, inc, true); 8667 if (err < 0) 8668 return err; 8669 if (dev->flags != old_flags) 8670 dev_set_rx_mode(dev); 8671 return err; 8672 } 8673 EXPORT_SYMBOL(dev_set_promiscuity); 8674 8675 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify) 8676 { 8677 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 8678 unsigned int allmulti, flags; 8679 8680 ASSERT_RTNL(); 8681 8682 allmulti = dev->allmulti + inc; 8683 if (allmulti == 0) { 8684 /* 8685 * Avoid overflow. 8686 * If inc causes overflow, untouch allmulti and return error. 8687 */ 8688 if (unlikely(inc > 0)) { 8689 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 8690 return -EOVERFLOW; 8691 } 8692 flags = old_flags & ~IFF_ALLMULTI; 8693 } else { 8694 flags = old_flags | IFF_ALLMULTI; 8695 } 8696 WRITE_ONCE(dev->allmulti, allmulti); 8697 if (flags != old_flags) { 8698 WRITE_ONCE(dev->flags, flags); 8699 netdev_info(dev, "%s allmulticast mode\n", 8700 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 8701 dev_change_rx_flags(dev, IFF_ALLMULTI); 8702 dev_set_rx_mode(dev); 8703 if (notify) 8704 __dev_notify_flags(dev, old_flags, 8705 dev->gflags ^ old_gflags, 0, NULL); 8706 } 8707 return 0; 8708 } 8709 8710 /** 8711 * dev_set_allmulti - update allmulti count on a device 8712 * @dev: device 8713 * @inc: modifier 8714 * 8715 * Add or remove reception of all multicast frames to a device. While the 8716 * count in the device remains above zero the interface remains listening 8717 * to all interfaces. Once it hits zero the device reverts back to normal 8718 * filtering operation. A negative @inc value is used to drop the counter 8719 * when releasing a resource needing all multicasts. 8720 * Return 0 if successful or a negative errno code on error. 8721 */ 8722 8723 int dev_set_allmulti(struct net_device *dev, int inc) 8724 { 8725 return __dev_set_allmulti(dev, inc, true); 8726 } 8727 EXPORT_SYMBOL(dev_set_allmulti); 8728 8729 /* 8730 * Upload unicast and multicast address lists to device and 8731 * configure RX filtering. When the device doesn't support unicast 8732 * filtering it is put in promiscuous mode while unicast addresses 8733 * are present. 8734 */ 8735 void __dev_set_rx_mode(struct net_device *dev) 8736 { 8737 const struct net_device_ops *ops = dev->netdev_ops; 8738 8739 /* dev_open will call this function so the list will stay sane. */ 8740 if (!(dev->flags&IFF_UP)) 8741 return; 8742 8743 if (!netif_device_present(dev)) 8744 return; 8745 8746 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 8747 /* Unicast addresses changes may only happen under the rtnl, 8748 * therefore calling __dev_set_promiscuity here is safe. 8749 */ 8750 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 8751 __dev_set_promiscuity(dev, 1, false); 8752 dev->uc_promisc = true; 8753 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 8754 __dev_set_promiscuity(dev, -1, false); 8755 dev->uc_promisc = false; 8756 } 8757 } 8758 8759 if (ops->ndo_set_rx_mode) 8760 ops->ndo_set_rx_mode(dev); 8761 } 8762 8763 void dev_set_rx_mode(struct net_device *dev) 8764 { 8765 netif_addr_lock_bh(dev); 8766 __dev_set_rx_mode(dev); 8767 netif_addr_unlock_bh(dev); 8768 } 8769 8770 /** 8771 * dev_get_flags - get flags reported to userspace 8772 * @dev: device 8773 * 8774 * Get the combination of flag bits exported through APIs to userspace. 8775 */ 8776 unsigned int dev_get_flags(const struct net_device *dev) 8777 { 8778 unsigned int flags; 8779 8780 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 8781 IFF_ALLMULTI | 8782 IFF_RUNNING | 8783 IFF_LOWER_UP | 8784 IFF_DORMANT)) | 8785 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 8786 IFF_ALLMULTI)); 8787 8788 if (netif_running(dev)) { 8789 if (netif_oper_up(dev)) 8790 flags |= IFF_RUNNING; 8791 if (netif_carrier_ok(dev)) 8792 flags |= IFF_LOWER_UP; 8793 if (netif_dormant(dev)) 8794 flags |= IFF_DORMANT; 8795 } 8796 8797 return flags; 8798 } 8799 EXPORT_SYMBOL(dev_get_flags); 8800 8801 int __dev_change_flags(struct net_device *dev, unsigned int flags, 8802 struct netlink_ext_ack *extack) 8803 { 8804 unsigned int old_flags = dev->flags; 8805 int ret; 8806 8807 ASSERT_RTNL(); 8808 8809 /* 8810 * Set the flags on our device. 8811 */ 8812 8813 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 8814 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 8815 IFF_AUTOMEDIA)) | 8816 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 8817 IFF_ALLMULTI)); 8818 8819 /* 8820 * Load in the correct multicast list now the flags have changed. 8821 */ 8822 8823 if ((old_flags ^ flags) & IFF_MULTICAST) 8824 dev_change_rx_flags(dev, IFF_MULTICAST); 8825 8826 dev_set_rx_mode(dev); 8827 8828 /* 8829 * Have we downed the interface. We handle IFF_UP ourselves 8830 * according to user attempts to set it, rather than blindly 8831 * setting it. 8832 */ 8833 8834 ret = 0; 8835 if ((old_flags ^ flags) & IFF_UP) { 8836 if (old_flags & IFF_UP) 8837 __dev_close(dev); 8838 else 8839 ret = __dev_open(dev, extack); 8840 } 8841 8842 if ((flags ^ dev->gflags) & IFF_PROMISC) { 8843 int inc = (flags & IFF_PROMISC) ? 1 : -1; 8844 unsigned int old_flags = dev->flags; 8845 8846 dev->gflags ^= IFF_PROMISC; 8847 8848 if (__dev_set_promiscuity(dev, inc, false) >= 0) 8849 if (dev->flags != old_flags) 8850 dev_set_rx_mode(dev); 8851 } 8852 8853 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 8854 * is important. Some (broken) drivers set IFF_PROMISC, when 8855 * IFF_ALLMULTI is requested not asking us and not reporting. 8856 */ 8857 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 8858 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 8859 8860 dev->gflags ^= IFF_ALLMULTI; 8861 __dev_set_allmulti(dev, inc, false); 8862 } 8863 8864 return ret; 8865 } 8866 8867 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 8868 unsigned int gchanges, u32 portid, 8869 const struct nlmsghdr *nlh) 8870 { 8871 unsigned int changes = dev->flags ^ old_flags; 8872 8873 if (gchanges) 8874 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 8875 8876 if (changes & IFF_UP) { 8877 if (dev->flags & IFF_UP) 8878 call_netdevice_notifiers(NETDEV_UP, dev); 8879 else 8880 call_netdevice_notifiers(NETDEV_DOWN, dev); 8881 } 8882 8883 if (dev->flags & IFF_UP && 8884 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 8885 struct netdev_notifier_change_info change_info = { 8886 .info = { 8887 .dev = dev, 8888 }, 8889 .flags_changed = changes, 8890 }; 8891 8892 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 8893 } 8894 } 8895 8896 /** 8897 * dev_change_flags - change device settings 8898 * @dev: device 8899 * @flags: device state flags 8900 * @extack: netlink extended ack 8901 * 8902 * Change settings on device based state flags. The flags are 8903 * in the userspace exported format. 8904 */ 8905 int dev_change_flags(struct net_device *dev, unsigned int flags, 8906 struct netlink_ext_ack *extack) 8907 { 8908 int ret; 8909 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 8910 8911 ret = __dev_change_flags(dev, flags, extack); 8912 if (ret < 0) 8913 return ret; 8914 8915 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 8916 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 8917 return ret; 8918 } 8919 EXPORT_SYMBOL(dev_change_flags); 8920 8921 int __dev_set_mtu(struct net_device *dev, int new_mtu) 8922 { 8923 const struct net_device_ops *ops = dev->netdev_ops; 8924 8925 if (ops->ndo_change_mtu) 8926 return ops->ndo_change_mtu(dev, new_mtu); 8927 8928 /* Pairs with all the lockless reads of dev->mtu in the stack */ 8929 WRITE_ONCE(dev->mtu, new_mtu); 8930 return 0; 8931 } 8932 EXPORT_SYMBOL(__dev_set_mtu); 8933 8934 int dev_validate_mtu(struct net_device *dev, int new_mtu, 8935 struct netlink_ext_ack *extack) 8936 { 8937 /* MTU must be positive, and in range */ 8938 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 8939 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 8940 return -EINVAL; 8941 } 8942 8943 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 8944 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 8945 return -EINVAL; 8946 } 8947 return 0; 8948 } 8949 8950 /** 8951 * dev_set_mtu_ext - Change maximum transfer unit 8952 * @dev: device 8953 * @new_mtu: new transfer unit 8954 * @extack: netlink extended ack 8955 * 8956 * Change the maximum transfer size of the network device. 8957 */ 8958 int dev_set_mtu_ext(struct net_device *dev, int new_mtu, 8959 struct netlink_ext_ack *extack) 8960 { 8961 int err, orig_mtu; 8962 8963 if (new_mtu == dev->mtu) 8964 return 0; 8965 8966 err = dev_validate_mtu(dev, new_mtu, extack); 8967 if (err) 8968 return err; 8969 8970 if (!netif_device_present(dev)) 8971 return -ENODEV; 8972 8973 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 8974 err = notifier_to_errno(err); 8975 if (err) 8976 return err; 8977 8978 orig_mtu = dev->mtu; 8979 err = __dev_set_mtu(dev, new_mtu); 8980 8981 if (!err) { 8982 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 8983 orig_mtu); 8984 err = notifier_to_errno(err); 8985 if (err) { 8986 /* setting mtu back and notifying everyone again, 8987 * so that they have a chance to revert changes. 8988 */ 8989 __dev_set_mtu(dev, orig_mtu); 8990 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 8991 new_mtu); 8992 } 8993 } 8994 return err; 8995 } 8996 8997 int dev_set_mtu(struct net_device *dev, int new_mtu) 8998 { 8999 struct netlink_ext_ack extack; 9000 int err; 9001 9002 memset(&extack, 0, sizeof(extack)); 9003 err = dev_set_mtu_ext(dev, new_mtu, &extack); 9004 if (err && extack._msg) 9005 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9006 return err; 9007 } 9008 EXPORT_SYMBOL(dev_set_mtu); 9009 9010 /** 9011 * dev_change_tx_queue_len - Change TX queue length of a netdevice 9012 * @dev: device 9013 * @new_len: new tx queue length 9014 */ 9015 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9016 { 9017 unsigned int orig_len = dev->tx_queue_len; 9018 int res; 9019 9020 if (new_len != (unsigned int)new_len) 9021 return -ERANGE; 9022 9023 if (new_len != orig_len) { 9024 WRITE_ONCE(dev->tx_queue_len, new_len); 9025 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9026 res = notifier_to_errno(res); 9027 if (res) 9028 goto err_rollback; 9029 res = dev_qdisc_change_tx_queue_len(dev); 9030 if (res) 9031 goto err_rollback; 9032 } 9033 9034 return 0; 9035 9036 err_rollback: 9037 netdev_err(dev, "refused to change device tx_queue_len\n"); 9038 WRITE_ONCE(dev->tx_queue_len, orig_len); 9039 return res; 9040 } 9041 9042 /** 9043 * dev_set_group - Change group this device belongs to 9044 * @dev: device 9045 * @new_group: group this device should belong to 9046 */ 9047 void dev_set_group(struct net_device *dev, int new_group) 9048 { 9049 dev->group = new_group; 9050 } 9051 9052 /** 9053 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 9054 * @dev: device 9055 * @addr: new address 9056 * @extack: netlink extended ack 9057 */ 9058 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9059 struct netlink_ext_ack *extack) 9060 { 9061 struct netdev_notifier_pre_changeaddr_info info = { 9062 .info.dev = dev, 9063 .info.extack = extack, 9064 .dev_addr = addr, 9065 }; 9066 int rc; 9067 9068 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9069 return notifier_to_errno(rc); 9070 } 9071 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 9072 9073 /** 9074 * dev_set_mac_address - Change Media Access Control Address 9075 * @dev: device 9076 * @sa: new address 9077 * @extack: netlink extended ack 9078 * 9079 * Change the hardware (MAC) address of the device 9080 */ 9081 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa, 9082 struct netlink_ext_ack *extack) 9083 { 9084 const struct net_device_ops *ops = dev->netdev_ops; 9085 int err; 9086 9087 if (!ops->ndo_set_mac_address) 9088 return -EOPNOTSUPP; 9089 if (sa->sa_family != dev->type) 9090 return -EINVAL; 9091 if (!netif_device_present(dev)) 9092 return -ENODEV; 9093 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 9094 if (err) 9095 return err; 9096 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 9097 err = ops->ndo_set_mac_address(dev, sa); 9098 if (err) 9099 return err; 9100 } 9101 dev->addr_assign_type = NET_ADDR_SET; 9102 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9103 add_device_randomness(dev->dev_addr, dev->addr_len); 9104 return 0; 9105 } 9106 EXPORT_SYMBOL(dev_set_mac_address); 9107 9108 DECLARE_RWSEM(dev_addr_sem); 9109 9110 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa, 9111 struct netlink_ext_ack *extack) 9112 { 9113 int ret; 9114 9115 down_write(&dev_addr_sem); 9116 ret = dev_set_mac_address(dev, sa, extack); 9117 up_write(&dev_addr_sem); 9118 return ret; 9119 } 9120 EXPORT_SYMBOL(dev_set_mac_address_user); 9121 9122 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9123 { 9124 size_t size = sizeof(sa->sa_data_min); 9125 struct net_device *dev; 9126 int ret = 0; 9127 9128 down_read(&dev_addr_sem); 9129 rcu_read_lock(); 9130 9131 dev = dev_get_by_name_rcu(net, dev_name); 9132 if (!dev) { 9133 ret = -ENODEV; 9134 goto unlock; 9135 } 9136 if (!dev->addr_len) 9137 memset(sa->sa_data, 0, size); 9138 else 9139 memcpy(sa->sa_data, dev->dev_addr, 9140 min_t(size_t, size, dev->addr_len)); 9141 sa->sa_family = dev->type; 9142 9143 unlock: 9144 rcu_read_unlock(); 9145 up_read(&dev_addr_sem); 9146 return ret; 9147 } 9148 EXPORT_SYMBOL(dev_get_mac_address); 9149 9150 /** 9151 * dev_change_carrier - Change device carrier 9152 * @dev: device 9153 * @new_carrier: new value 9154 * 9155 * Change device carrier 9156 */ 9157 int dev_change_carrier(struct net_device *dev, bool new_carrier) 9158 { 9159 const struct net_device_ops *ops = dev->netdev_ops; 9160 9161 if (!ops->ndo_change_carrier) 9162 return -EOPNOTSUPP; 9163 if (!netif_device_present(dev)) 9164 return -ENODEV; 9165 return ops->ndo_change_carrier(dev, new_carrier); 9166 } 9167 9168 /** 9169 * dev_get_phys_port_id - Get device physical port ID 9170 * @dev: device 9171 * @ppid: port ID 9172 * 9173 * Get device physical port ID 9174 */ 9175 int dev_get_phys_port_id(struct net_device *dev, 9176 struct netdev_phys_item_id *ppid) 9177 { 9178 const struct net_device_ops *ops = dev->netdev_ops; 9179 9180 if (!ops->ndo_get_phys_port_id) 9181 return -EOPNOTSUPP; 9182 return ops->ndo_get_phys_port_id(dev, ppid); 9183 } 9184 9185 /** 9186 * dev_get_phys_port_name - Get device physical port name 9187 * @dev: device 9188 * @name: port name 9189 * @len: limit of bytes to copy to name 9190 * 9191 * Get device physical port name 9192 */ 9193 int dev_get_phys_port_name(struct net_device *dev, 9194 char *name, size_t len) 9195 { 9196 const struct net_device_ops *ops = dev->netdev_ops; 9197 int err; 9198 9199 if (ops->ndo_get_phys_port_name) { 9200 err = ops->ndo_get_phys_port_name(dev, name, len); 9201 if (err != -EOPNOTSUPP) 9202 return err; 9203 } 9204 return devlink_compat_phys_port_name_get(dev, name, len); 9205 } 9206 9207 /** 9208 * dev_get_port_parent_id - Get the device's port parent identifier 9209 * @dev: network device 9210 * @ppid: pointer to a storage for the port's parent identifier 9211 * @recurse: allow/disallow recursion to lower devices 9212 * 9213 * Get the devices's port parent identifier 9214 */ 9215 int dev_get_port_parent_id(struct net_device *dev, 9216 struct netdev_phys_item_id *ppid, 9217 bool recurse) 9218 { 9219 const struct net_device_ops *ops = dev->netdev_ops; 9220 struct netdev_phys_item_id first = { }; 9221 struct net_device *lower_dev; 9222 struct list_head *iter; 9223 int err; 9224 9225 if (ops->ndo_get_port_parent_id) { 9226 err = ops->ndo_get_port_parent_id(dev, ppid); 9227 if (err != -EOPNOTSUPP) 9228 return err; 9229 } 9230 9231 err = devlink_compat_switch_id_get(dev, ppid); 9232 if (!recurse || err != -EOPNOTSUPP) 9233 return err; 9234 9235 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9236 err = dev_get_port_parent_id(lower_dev, ppid, true); 9237 if (err) 9238 break; 9239 if (!first.id_len) 9240 first = *ppid; 9241 else if (memcmp(&first, ppid, sizeof(*ppid))) 9242 return -EOPNOTSUPP; 9243 } 9244 9245 return err; 9246 } 9247 EXPORT_SYMBOL(dev_get_port_parent_id); 9248 9249 /** 9250 * netdev_port_same_parent_id - Indicate if two network devices have 9251 * the same port parent identifier 9252 * @a: first network device 9253 * @b: second network device 9254 */ 9255 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9256 { 9257 struct netdev_phys_item_id a_id = { }; 9258 struct netdev_phys_item_id b_id = { }; 9259 9260 if (dev_get_port_parent_id(a, &a_id, true) || 9261 dev_get_port_parent_id(b, &b_id, true)) 9262 return false; 9263 9264 return netdev_phys_item_id_same(&a_id, &b_id); 9265 } 9266 EXPORT_SYMBOL(netdev_port_same_parent_id); 9267 9268 /** 9269 * dev_change_proto_down - set carrier according to proto_down. 9270 * 9271 * @dev: device 9272 * @proto_down: new value 9273 */ 9274 int dev_change_proto_down(struct net_device *dev, bool proto_down) 9275 { 9276 if (!(dev->priv_flags & IFF_CHANGE_PROTO_DOWN)) 9277 return -EOPNOTSUPP; 9278 if (!netif_device_present(dev)) 9279 return -ENODEV; 9280 if (proto_down) 9281 netif_carrier_off(dev); 9282 else 9283 netif_carrier_on(dev); 9284 WRITE_ONCE(dev->proto_down, proto_down); 9285 return 0; 9286 } 9287 9288 /** 9289 * dev_change_proto_down_reason - proto down reason 9290 * 9291 * @dev: device 9292 * @mask: proto down mask 9293 * @value: proto down value 9294 */ 9295 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask, 9296 u32 value) 9297 { 9298 u32 proto_down_reason; 9299 int b; 9300 9301 if (!mask) { 9302 proto_down_reason = value; 9303 } else { 9304 proto_down_reason = dev->proto_down_reason; 9305 for_each_set_bit(b, &mask, 32) { 9306 if (value & (1 << b)) 9307 proto_down_reason |= BIT(b); 9308 else 9309 proto_down_reason &= ~BIT(b); 9310 } 9311 } 9312 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 9313 } 9314 9315 struct bpf_xdp_link { 9316 struct bpf_link link; 9317 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9318 int flags; 9319 }; 9320 9321 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9322 { 9323 if (flags & XDP_FLAGS_HW_MODE) 9324 return XDP_MODE_HW; 9325 if (flags & XDP_FLAGS_DRV_MODE) 9326 return XDP_MODE_DRV; 9327 if (flags & XDP_FLAGS_SKB_MODE) 9328 return XDP_MODE_SKB; 9329 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9330 } 9331 9332 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9333 { 9334 switch (mode) { 9335 case XDP_MODE_SKB: 9336 return generic_xdp_install; 9337 case XDP_MODE_DRV: 9338 case XDP_MODE_HW: 9339 return dev->netdev_ops->ndo_bpf; 9340 default: 9341 return NULL; 9342 } 9343 } 9344 9345 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9346 enum bpf_xdp_mode mode) 9347 { 9348 return dev->xdp_state[mode].link; 9349 } 9350 9351 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9352 enum bpf_xdp_mode mode) 9353 { 9354 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9355 9356 if (link) 9357 return link->link.prog; 9358 return dev->xdp_state[mode].prog; 9359 } 9360 9361 u8 dev_xdp_prog_count(struct net_device *dev) 9362 { 9363 u8 count = 0; 9364 int i; 9365 9366 for (i = 0; i < __MAX_XDP_MODE; i++) 9367 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9368 count++; 9369 return count; 9370 } 9371 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9372 9373 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9374 { 9375 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9376 9377 return prog ? prog->aux->id : 0; 9378 } 9379 9380 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9381 struct bpf_xdp_link *link) 9382 { 9383 dev->xdp_state[mode].link = link; 9384 dev->xdp_state[mode].prog = NULL; 9385 } 9386 9387 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9388 struct bpf_prog *prog) 9389 { 9390 dev->xdp_state[mode].link = NULL; 9391 dev->xdp_state[mode].prog = prog; 9392 } 9393 9394 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9395 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9396 u32 flags, struct bpf_prog *prog) 9397 { 9398 struct netdev_bpf xdp; 9399 int err; 9400 9401 memset(&xdp, 0, sizeof(xdp)); 9402 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9403 xdp.extack = extack; 9404 xdp.flags = flags; 9405 xdp.prog = prog; 9406 9407 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9408 * "moved" into driver), so they don't increment it on their own, but 9409 * they do decrement refcnt when program is detached or replaced. 9410 * Given net_device also owns link/prog, we need to bump refcnt here 9411 * to prevent drivers from underflowing it. 9412 */ 9413 if (prog) 9414 bpf_prog_inc(prog); 9415 err = bpf_op(dev, &xdp); 9416 if (err) { 9417 if (prog) 9418 bpf_prog_put(prog); 9419 return err; 9420 } 9421 9422 if (mode != XDP_MODE_HW) 9423 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9424 9425 return 0; 9426 } 9427 9428 static void dev_xdp_uninstall(struct net_device *dev) 9429 { 9430 struct bpf_xdp_link *link; 9431 struct bpf_prog *prog; 9432 enum bpf_xdp_mode mode; 9433 bpf_op_t bpf_op; 9434 9435 ASSERT_RTNL(); 9436 9437 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 9438 prog = dev_xdp_prog(dev, mode); 9439 if (!prog) 9440 continue; 9441 9442 bpf_op = dev_xdp_bpf_op(dev, mode); 9443 if (!bpf_op) 9444 continue; 9445 9446 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9447 9448 /* auto-detach link from net device */ 9449 link = dev_xdp_link(dev, mode); 9450 if (link) 9451 link->dev = NULL; 9452 else 9453 bpf_prog_put(prog); 9454 9455 dev_xdp_set_link(dev, mode, NULL); 9456 } 9457 } 9458 9459 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 9460 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 9461 struct bpf_prog *old_prog, u32 flags) 9462 { 9463 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 9464 struct bpf_prog *cur_prog; 9465 struct net_device *upper; 9466 struct list_head *iter; 9467 enum bpf_xdp_mode mode; 9468 bpf_op_t bpf_op; 9469 int err; 9470 9471 ASSERT_RTNL(); 9472 9473 /* either link or prog attachment, never both */ 9474 if (link && (new_prog || old_prog)) 9475 return -EINVAL; 9476 /* link supports only XDP mode flags */ 9477 if (link && (flags & ~XDP_FLAGS_MODES)) { 9478 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 9479 return -EINVAL; 9480 } 9481 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 9482 if (num_modes > 1) { 9483 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 9484 return -EINVAL; 9485 } 9486 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 9487 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 9488 NL_SET_ERR_MSG(extack, 9489 "More than one program loaded, unset mode is ambiguous"); 9490 return -EINVAL; 9491 } 9492 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 9493 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 9494 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 9495 return -EINVAL; 9496 } 9497 9498 mode = dev_xdp_mode(dev, flags); 9499 /* can't replace attached link */ 9500 if (dev_xdp_link(dev, mode)) { 9501 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 9502 return -EBUSY; 9503 } 9504 9505 /* don't allow if an upper device already has a program */ 9506 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 9507 if (dev_xdp_prog_count(upper) > 0) { 9508 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 9509 return -EEXIST; 9510 } 9511 } 9512 9513 cur_prog = dev_xdp_prog(dev, mode); 9514 /* can't replace attached prog with link */ 9515 if (link && cur_prog) { 9516 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 9517 return -EBUSY; 9518 } 9519 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 9520 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 9521 return -EEXIST; 9522 } 9523 9524 /* put effective new program into new_prog */ 9525 if (link) 9526 new_prog = link->link.prog; 9527 9528 if (new_prog) { 9529 bool offload = mode == XDP_MODE_HW; 9530 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 9531 ? XDP_MODE_DRV : XDP_MODE_SKB; 9532 9533 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 9534 NL_SET_ERR_MSG(extack, "XDP program already attached"); 9535 return -EBUSY; 9536 } 9537 if (!offload && dev_xdp_prog(dev, other_mode)) { 9538 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 9539 return -EEXIST; 9540 } 9541 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 9542 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 9543 return -EINVAL; 9544 } 9545 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 9546 NL_SET_ERR_MSG(extack, "Program bound to different device"); 9547 return -EINVAL; 9548 } 9549 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 9550 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 9551 return -EINVAL; 9552 } 9553 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 9554 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 9555 return -EINVAL; 9556 } 9557 } 9558 9559 /* don't call drivers if the effective program didn't change */ 9560 if (new_prog != cur_prog) { 9561 bpf_op = dev_xdp_bpf_op(dev, mode); 9562 if (!bpf_op) { 9563 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 9564 return -EOPNOTSUPP; 9565 } 9566 9567 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 9568 if (err) 9569 return err; 9570 } 9571 9572 if (link) 9573 dev_xdp_set_link(dev, mode, link); 9574 else 9575 dev_xdp_set_prog(dev, mode, new_prog); 9576 if (cur_prog) 9577 bpf_prog_put(cur_prog); 9578 9579 return 0; 9580 } 9581 9582 static int dev_xdp_attach_link(struct net_device *dev, 9583 struct netlink_ext_ack *extack, 9584 struct bpf_xdp_link *link) 9585 { 9586 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 9587 } 9588 9589 static int dev_xdp_detach_link(struct net_device *dev, 9590 struct netlink_ext_ack *extack, 9591 struct bpf_xdp_link *link) 9592 { 9593 enum bpf_xdp_mode mode; 9594 bpf_op_t bpf_op; 9595 9596 ASSERT_RTNL(); 9597 9598 mode = dev_xdp_mode(dev, link->flags); 9599 if (dev_xdp_link(dev, mode) != link) 9600 return -EINVAL; 9601 9602 bpf_op = dev_xdp_bpf_op(dev, mode); 9603 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9604 dev_xdp_set_link(dev, mode, NULL); 9605 return 0; 9606 } 9607 9608 static void bpf_xdp_link_release(struct bpf_link *link) 9609 { 9610 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9611 9612 rtnl_lock(); 9613 9614 /* if racing with net_device's tear down, xdp_link->dev might be 9615 * already NULL, in which case link was already auto-detached 9616 */ 9617 if (xdp_link->dev) { 9618 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 9619 xdp_link->dev = NULL; 9620 } 9621 9622 rtnl_unlock(); 9623 } 9624 9625 static int bpf_xdp_link_detach(struct bpf_link *link) 9626 { 9627 bpf_xdp_link_release(link); 9628 return 0; 9629 } 9630 9631 static void bpf_xdp_link_dealloc(struct bpf_link *link) 9632 { 9633 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9634 9635 kfree(xdp_link); 9636 } 9637 9638 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 9639 struct seq_file *seq) 9640 { 9641 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9642 u32 ifindex = 0; 9643 9644 rtnl_lock(); 9645 if (xdp_link->dev) 9646 ifindex = xdp_link->dev->ifindex; 9647 rtnl_unlock(); 9648 9649 seq_printf(seq, "ifindex:\t%u\n", ifindex); 9650 } 9651 9652 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 9653 struct bpf_link_info *info) 9654 { 9655 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9656 u32 ifindex = 0; 9657 9658 rtnl_lock(); 9659 if (xdp_link->dev) 9660 ifindex = xdp_link->dev->ifindex; 9661 rtnl_unlock(); 9662 9663 info->xdp.ifindex = ifindex; 9664 return 0; 9665 } 9666 9667 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 9668 struct bpf_prog *old_prog) 9669 { 9670 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9671 enum bpf_xdp_mode mode; 9672 bpf_op_t bpf_op; 9673 int err = 0; 9674 9675 rtnl_lock(); 9676 9677 /* link might have been auto-released already, so fail */ 9678 if (!xdp_link->dev) { 9679 err = -ENOLINK; 9680 goto out_unlock; 9681 } 9682 9683 if (old_prog && link->prog != old_prog) { 9684 err = -EPERM; 9685 goto out_unlock; 9686 } 9687 old_prog = link->prog; 9688 if (old_prog->type != new_prog->type || 9689 old_prog->expected_attach_type != new_prog->expected_attach_type) { 9690 err = -EINVAL; 9691 goto out_unlock; 9692 } 9693 9694 if (old_prog == new_prog) { 9695 /* no-op, don't disturb drivers */ 9696 bpf_prog_put(new_prog); 9697 goto out_unlock; 9698 } 9699 9700 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 9701 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 9702 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 9703 xdp_link->flags, new_prog); 9704 if (err) 9705 goto out_unlock; 9706 9707 old_prog = xchg(&link->prog, new_prog); 9708 bpf_prog_put(old_prog); 9709 9710 out_unlock: 9711 rtnl_unlock(); 9712 return err; 9713 } 9714 9715 static const struct bpf_link_ops bpf_xdp_link_lops = { 9716 .release = bpf_xdp_link_release, 9717 .dealloc = bpf_xdp_link_dealloc, 9718 .detach = bpf_xdp_link_detach, 9719 .show_fdinfo = bpf_xdp_link_show_fdinfo, 9720 .fill_link_info = bpf_xdp_link_fill_link_info, 9721 .update_prog = bpf_xdp_link_update, 9722 }; 9723 9724 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 9725 { 9726 struct net *net = current->nsproxy->net_ns; 9727 struct bpf_link_primer link_primer; 9728 struct netlink_ext_ack extack = {}; 9729 struct bpf_xdp_link *link; 9730 struct net_device *dev; 9731 int err, fd; 9732 9733 rtnl_lock(); 9734 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 9735 if (!dev) { 9736 rtnl_unlock(); 9737 return -EINVAL; 9738 } 9739 9740 link = kzalloc(sizeof(*link), GFP_USER); 9741 if (!link) { 9742 err = -ENOMEM; 9743 goto unlock; 9744 } 9745 9746 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 9747 link->dev = dev; 9748 link->flags = attr->link_create.flags; 9749 9750 err = bpf_link_prime(&link->link, &link_primer); 9751 if (err) { 9752 kfree(link); 9753 goto unlock; 9754 } 9755 9756 err = dev_xdp_attach_link(dev, &extack, link); 9757 rtnl_unlock(); 9758 9759 if (err) { 9760 link->dev = NULL; 9761 bpf_link_cleanup(&link_primer); 9762 trace_bpf_xdp_link_attach_failed(extack._msg); 9763 goto out_put_dev; 9764 } 9765 9766 fd = bpf_link_settle(&link_primer); 9767 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 9768 dev_put(dev); 9769 return fd; 9770 9771 unlock: 9772 rtnl_unlock(); 9773 9774 out_put_dev: 9775 dev_put(dev); 9776 return err; 9777 } 9778 9779 /** 9780 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 9781 * @dev: device 9782 * @extack: netlink extended ack 9783 * @fd: new program fd or negative value to clear 9784 * @expected_fd: old program fd that userspace expects to replace or clear 9785 * @flags: xdp-related flags 9786 * 9787 * Set or clear a bpf program for a device 9788 */ 9789 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 9790 int fd, int expected_fd, u32 flags) 9791 { 9792 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 9793 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 9794 int err; 9795 9796 ASSERT_RTNL(); 9797 9798 if (fd >= 0) { 9799 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 9800 mode != XDP_MODE_SKB); 9801 if (IS_ERR(new_prog)) 9802 return PTR_ERR(new_prog); 9803 } 9804 9805 if (expected_fd >= 0) { 9806 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 9807 mode != XDP_MODE_SKB); 9808 if (IS_ERR(old_prog)) { 9809 err = PTR_ERR(old_prog); 9810 old_prog = NULL; 9811 goto err_out; 9812 } 9813 } 9814 9815 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 9816 9817 err_out: 9818 if (err && new_prog) 9819 bpf_prog_put(new_prog); 9820 if (old_prog) 9821 bpf_prog_put(old_prog); 9822 return err; 9823 } 9824 9825 /** 9826 * dev_index_reserve() - allocate an ifindex in a namespace 9827 * @net: the applicable net namespace 9828 * @ifindex: requested ifindex, pass %0 to get one allocated 9829 * 9830 * Allocate a ifindex for a new device. Caller must either use the ifindex 9831 * to store the device (via list_netdevice()) or call dev_index_release() 9832 * to give the index up. 9833 * 9834 * Return: a suitable unique value for a new device interface number or -errno. 9835 */ 9836 static int dev_index_reserve(struct net *net, u32 ifindex) 9837 { 9838 int err; 9839 9840 if (ifindex > INT_MAX) { 9841 DEBUG_NET_WARN_ON_ONCE(1); 9842 return -EINVAL; 9843 } 9844 9845 if (!ifindex) 9846 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 9847 xa_limit_31b, &net->ifindex, GFP_KERNEL); 9848 else 9849 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 9850 if (err < 0) 9851 return err; 9852 9853 return ifindex; 9854 } 9855 9856 static void dev_index_release(struct net *net, int ifindex) 9857 { 9858 /* Expect only unused indexes, unlist_netdevice() removes the used */ 9859 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 9860 } 9861 9862 /* Delayed registration/unregisteration */ 9863 LIST_HEAD(net_todo_list); 9864 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 9865 atomic_t dev_unreg_count = ATOMIC_INIT(0); 9866 9867 static void net_set_todo(struct net_device *dev) 9868 { 9869 list_add_tail(&dev->todo_list, &net_todo_list); 9870 } 9871 9872 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 9873 struct net_device *upper, netdev_features_t features) 9874 { 9875 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 9876 netdev_features_t feature; 9877 int feature_bit; 9878 9879 for_each_netdev_feature(upper_disables, feature_bit) { 9880 feature = __NETIF_F_BIT(feature_bit); 9881 if (!(upper->wanted_features & feature) 9882 && (features & feature)) { 9883 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 9884 &feature, upper->name); 9885 features &= ~feature; 9886 } 9887 } 9888 9889 return features; 9890 } 9891 9892 static void netdev_sync_lower_features(struct net_device *upper, 9893 struct net_device *lower, netdev_features_t features) 9894 { 9895 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 9896 netdev_features_t feature; 9897 int feature_bit; 9898 9899 for_each_netdev_feature(upper_disables, feature_bit) { 9900 feature = __NETIF_F_BIT(feature_bit); 9901 if (!(features & feature) && (lower->features & feature)) { 9902 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 9903 &feature, lower->name); 9904 lower->wanted_features &= ~feature; 9905 __netdev_update_features(lower); 9906 9907 if (unlikely(lower->features & feature)) 9908 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 9909 &feature, lower->name); 9910 else 9911 netdev_features_change(lower); 9912 } 9913 } 9914 } 9915 9916 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 9917 { 9918 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 9919 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 9920 bool hw_csum = features & NETIF_F_HW_CSUM; 9921 9922 return ip_csum || hw_csum; 9923 } 9924 9925 static netdev_features_t netdev_fix_features(struct net_device *dev, 9926 netdev_features_t features) 9927 { 9928 /* Fix illegal checksum combinations */ 9929 if ((features & NETIF_F_HW_CSUM) && 9930 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 9931 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 9932 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 9933 } 9934 9935 /* TSO requires that SG is present as well. */ 9936 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 9937 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 9938 features &= ~NETIF_F_ALL_TSO; 9939 } 9940 9941 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 9942 !(features & NETIF_F_IP_CSUM)) { 9943 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 9944 features &= ~NETIF_F_TSO; 9945 features &= ~NETIF_F_TSO_ECN; 9946 } 9947 9948 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 9949 !(features & NETIF_F_IPV6_CSUM)) { 9950 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 9951 features &= ~NETIF_F_TSO6; 9952 } 9953 9954 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 9955 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 9956 features &= ~NETIF_F_TSO_MANGLEID; 9957 9958 /* TSO ECN requires that TSO is present as well. */ 9959 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 9960 features &= ~NETIF_F_TSO_ECN; 9961 9962 /* Software GSO depends on SG. */ 9963 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 9964 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 9965 features &= ~NETIF_F_GSO; 9966 } 9967 9968 /* GSO partial features require GSO partial be set */ 9969 if ((features & dev->gso_partial_features) && 9970 !(features & NETIF_F_GSO_PARTIAL)) { 9971 netdev_dbg(dev, 9972 "Dropping partially supported GSO features since no GSO partial.\n"); 9973 features &= ~dev->gso_partial_features; 9974 } 9975 9976 if (!(features & NETIF_F_RXCSUM)) { 9977 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 9978 * successfully merged by hardware must also have the 9979 * checksum verified by hardware. If the user does not 9980 * want to enable RXCSUM, logically, we should disable GRO_HW. 9981 */ 9982 if (features & NETIF_F_GRO_HW) { 9983 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 9984 features &= ~NETIF_F_GRO_HW; 9985 } 9986 } 9987 9988 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 9989 if (features & NETIF_F_RXFCS) { 9990 if (features & NETIF_F_LRO) { 9991 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 9992 features &= ~NETIF_F_LRO; 9993 } 9994 9995 if (features & NETIF_F_GRO_HW) { 9996 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 9997 features &= ~NETIF_F_GRO_HW; 9998 } 9999 } 10000 10001 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10002 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10003 features &= ~NETIF_F_LRO; 10004 } 10005 10006 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10007 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10008 features &= ~NETIF_F_HW_TLS_TX; 10009 } 10010 10011 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10012 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10013 features &= ~NETIF_F_HW_TLS_RX; 10014 } 10015 10016 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10017 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10018 features &= ~NETIF_F_GSO_UDP_L4; 10019 } 10020 10021 return features; 10022 } 10023 10024 int __netdev_update_features(struct net_device *dev) 10025 { 10026 struct net_device *upper, *lower; 10027 netdev_features_t features; 10028 struct list_head *iter; 10029 int err = -1; 10030 10031 ASSERT_RTNL(); 10032 10033 features = netdev_get_wanted_features(dev); 10034 10035 if (dev->netdev_ops->ndo_fix_features) 10036 features = dev->netdev_ops->ndo_fix_features(dev, features); 10037 10038 /* driver might be less strict about feature dependencies */ 10039 features = netdev_fix_features(dev, features); 10040 10041 /* some features can't be enabled if they're off on an upper device */ 10042 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10043 features = netdev_sync_upper_features(dev, upper, features); 10044 10045 if (dev->features == features) 10046 goto sync_lower; 10047 10048 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10049 &dev->features, &features); 10050 10051 if (dev->netdev_ops->ndo_set_features) 10052 err = dev->netdev_ops->ndo_set_features(dev, features); 10053 else 10054 err = 0; 10055 10056 if (unlikely(err < 0)) { 10057 netdev_err(dev, 10058 "set_features() failed (%d); wanted %pNF, left %pNF\n", 10059 err, &features, &dev->features); 10060 /* return non-0 since some features might have changed and 10061 * it's better to fire a spurious notification than miss it 10062 */ 10063 return -1; 10064 } 10065 10066 sync_lower: 10067 /* some features must be disabled on lower devices when disabled 10068 * on an upper device (think: bonding master or bridge) 10069 */ 10070 netdev_for_each_lower_dev(dev, lower, iter) 10071 netdev_sync_lower_features(dev, lower, features); 10072 10073 if (!err) { 10074 netdev_features_t diff = features ^ dev->features; 10075 10076 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 10077 /* udp_tunnel_{get,drop}_rx_info both need 10078 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 10079 * device, or they won't do anything. 10080 * Thus we need to update dev->features 10081 * *before* calling udp_tunnel_get_rx_info, 10082 * but *after* calling udp_tunnel_drop_rx_info. 10083 */ 10084 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 10085 dev->features = features; 10086 udp_tunnel_get_rx_info(dev); 10087 } else { 10088 udp_tunnel_drop_rx_info(dev); 10089 } 10090 } 10091 10092 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 10093 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 10094 dev->features = features; 10095 err |= vlan_get_rx_ctag_filter_info(dev); 10096 } else { 10097 vlan_drop_rx_ctag_filter_info(dev); 10098 } 10099 } 10100 10101 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 10102 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 10103 dev->features = features; 10104 err |= vlan_get_rx_stag_filter_info(dev); 10105 } else { 10106 vlan_drop_rx_stag_filter_info(dev); 10107 } 10108 } 10109 10110 dev->features = features; 10111 } 10112 10113 return err < 0 ? 0 : 1; 10114 } 10115 10116 /** 10117 * netdev_update_features - recalculate device features 10118 * @dev: the device to check 10119 * 10120 * Recalculate dev->features set and send notifications if it 10121 * has changed. Should be called after driver or hardware dependent 10122 * conditions might have changed that influence the features. 10123 */ 10124 void netdev_update_features(struct net_device *dev) 10125 { 10126 if (__netdev_update_features(dev)) 10127 netdev_features_change(dev); 10128 } 10129 EXPORT_SYMBOL(netdev_update_features); 10130 10131 /** 10132 * netdev_change_features - recalculate device features 10133 * @dev: the device to check 10134 * 10135 * Recalculate dev->features set and send notifications even 10136 * if they have not changed. Should be called instead of 10137 * netdev_update_features() if also dev->vlan_features might 10138 * have changed to allow the changes to be propagated to stacked 10139 * VLAN devices. 10140 */ 10141 void netdev_change_features(struct net_device *dev) 10142 { 10143 __netdev_update_features(dev); 10144 netdev_features_change(dev); 10145 } 10146 EXPORT_SYMBOL(netdev_change_features); 10147 10148 /** 10149 * netif_stacked_transfer_operstate - transfer operstate 10150 * @rootdev: the root or lower level device to transfer state from 10151 * @dev: the device to transfer operstate to 10152 * 10153 * Transfer operational state from root to device. This is normally 10154 * called when a stacking relationship exists between the root 10155 * device and the device(a leaf device). 10156 */ 10157 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 10158 struct net_device *dev) 10159 { 10160 if (rootdev->operstate == IF_OPER_DORMANT) 10161 netif_dormant_on(dev); 10162 else 10163 netif_dormant_off(dev); 10164 10165 if (rootdev->operstate == IF_OPER_TESTING) 10166 netif_testing_on(dev); 10167 else 10168 netif_testing_off(dev); 10169 10170 if (netif_carrier_ok(rootdev)) 10171 netif_carrier_on(dev); 10172 else 10173 netif_carrier_off(dev); 10174 } 10175 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10176 10177 static int netif_alloc_rx_queues(struct net_device *dev) 10178 { 10179 unsigned int i, count = dev->num_rx_queues; 10180 struct netdev_rx_queue *rx; 10181 size_t sz = count * sizeof(*rx); 10182 int err = 0; 10183 10184 BUG_ON(count < 1); 10185 10186 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10187 if (!rx) 10188 return -ENOMEM; 10189 10190 dev->_rx = rx; 10191 10192 for (i = 0; i < count; i++) { 10193 rx[i].dev = dev; 10194 10195 /* XDP RX-queue setup */ 10196 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10197 if (err < 0) 10198 goto err_rxq_info; 10199 } 10200 return 0; 10201 10202 err_rxq_info: 10203 /* Rollback successful reg's and free other resources */ 10204 while (i--) 10205 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10206 kvfree(dev->_rx); 10207 dev->_rx = NULL; 10208 return err; 10209 } 10210 10211 static void netif_free_rx_queues(struct net_device *dev) 10212 { 10213 unsigned int i, count = dev->num_rx_queues; 10214 10215 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10216 if (!dev->_rx) 10217 return; 10218 10219 for (i = 0; i < count; i++) 10220 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10221 10222 kvfree(dev->_rx); 10223 } 10224 10225 static void netdev_init_one_queue(struct net_device *dev, 10226 struct netdev_queue *queue, void *_unused) 10227 { 10228 /* Initialize queue lock */ 10229 spin_lock_init(&queue->_xmit_lock); 10230 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10231 queue->xmit_lock_owner = -1; 10232 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10233 queue->dev = dev; 10234 #ifdef CONFIG_BQL 10235 dql_init(&queue->dql, HZ); 10236 #endif 10237 } 10238 10239 static void netif_free_tx_queues(struct net_device *dev) 10240 { 10241 kvfree(dev->_tx); 10242 } 10243 10244 static int netif_alloc_netdev_queues(struct net_device *dev) 10245 { 10246 unsigned int count = dev->num_tx_queues; 10247 struct netdev_queue *tx; 10248 size_t sz = count * sizeof(*tx); 10249 10250 if (count < 1 || count > 0xffff) 10251 return -EINVAL; 10252 10253 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10254 if (!tx) 10255 return -ENOMEM; 10256 10257 dev->_tx = tx; 10258 10259 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10260 spin_lock_init(&dev->tx_global_lock); 10261 10262 return 0; 10263 } 10264 10265 void netif_tx_stop_all_queues(struct net_device *dev) 10266 { 10267 unsigned int i; 10268 10269 for (i = 0; i < dev->num_tx_queues; i++) { 10270 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10271 10272 netif_tx_stop_queue(txq); 10273 } 10274 } 10275 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10276 10277 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10278 { 10279 void __percpu *v; 10280 10281 /* Drivers implementing ndo_get_peer_dev must support tstat 10282 * accounting, so that skb_do_redirect() can bump the dev's 10283 * RX stats upon network namespace switch. 10284 */ 10285 if (dev->netdev_ops->ndo_get_peer_dev && 10286 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10287 return -EOPNOTSUPP; 10288 10289 switch (dev->pcpu_stat_type) { 10290 case NETDEV_PCPU_STAT_NONE: 10291 return 0; 10292 case NETDEV_PCPU_STAT_LSTATS: 10293 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10294 break; 10295 case NETDEV_PCPU_STAT_TSTATS: 10296 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10297 break; 10298 case NETDEV_PCPU_STAT_DSTATS: 10299 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10300 break; 10301 default: 10302 return -EINVAL; 10303 } 10304 10305 return v ? 0 : -ENOMEM; 10306 } 10307 10308 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10309 { 10310 switch (dev->pcpu_stat_type) { 10311 case NETDEV_PCPU_STAT_NONE: 10312 return; 10313 case NETDEV_PCPU_STAT_LSTATS: 10314 free_percpu(dev->lstats); 10315 break; 10316 case NETDEV_PCPU_STAT_TSTATS: 10317 free_percpu(dev->tstats); 10318 break; 10319 case NETDEV_PCPU_STAT_DSTATS: 10320 free_percpu(dev->dstats); 10321 break; 10322 } 10323 } 10324 10325 /** 10326 * register_netdevice() - register a network device 10327 * @dev: device to register 10328 * 10329 * Take a prepared network device structure and make it externally accessible. 10330 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10331 * Callers must hold the rtnl lock - you may want register_netdev() 10332 * instead of this. 10333 */ 10334 int register_netdevice(struct net_device *dev) 10335 { 10336 int ret; 10337 struct net *net = dev_net(dev); 10338 10339 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10340 NETDEV_FEATURE_COUNT); 10341 BUG_ON(dev_boot_phase); 10342 ASSERT_RTNL(); 10343 10344 might_sleep(); 10345 10346 /* When net_device's are persistent, this will be fatal. */ 10347 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10348 BUG_ON(!net); 10349 10350 ret = ethtool_check_ops(dev->ethtool_ops); 10351 if (ret) 10352 return ret; 10353 10354 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 10355 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 10356 mutex_init(&dev->ethtool->rss_lock); 10357 10358 spin_lock_init(&dev->addr_list_lock); 10359 netdev_set_addr_lockdep_class(dev); 10360 10361 ret = dev_get_valid_name(net, dev, dev->name); 10362 if (ret < 0) 10363 goto out; 10364 10365 ret = -ENOMEM; 10366 dev->name_node = netdev_name_node_head_alloc(dev); 10367 if (!dev->name_node) 10368 goto out; 10369 10370 /* Init, if this function is available */ 10371 if (dev->netdev_ops->ndo_init) { 10372 ret = dev->netdev_ops->ndo_init(dev); 10373 if (ret) { 10374 if (ret > 0) 10375 ret = -EIO; 10376 goto err_free_name; 10377 } 10378 } 10379 10380 if (((dev->hw_features | dev->features) & 10381 NETIF_F_HW_VLAN_CTAG_FILTER) && 10382 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 10383 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 10384 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 10385 ret = -EINVAL; 10386 goto err_uninit; 10387 } 10388 10389 ret = netdev_do_alloc_pcpu_stats(dev); 10390 if (ret) 10391 goto err_uninit; 10392 10393 ret = dev_index_reserve(net, dev->ifindex); 10394 if (ret < 0) 10395 goto err_free_pcpu; 10396 dev->ifindex = ret; 10397 10398 /* Transfer changeable features to wanted_features and enable 10399 * software offloads (GSO and GRO). 10400 */ 10401 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 10402 dev->features |= NETIF_F_SOFT_FEATURES; 10403 10404 if (dev->udp_tunnel_nic_info) { 10405 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10406 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10407 } 10408 10409 dev->wanted_features = dev->features & dev->hw_features; 10410 10411 if (!(dev->flags & IFF_LOOPBACK)) 10412 dev->hw_features |= NETIF_F_NOCACHE_COPY; 10413 10414 /* If IPv4 TCP segmentation offload is supported we should also 10415 * allow the device to enable segmenting the frame with the option 10416 * of ignoring a static IP ID value. This doesn't enable the 10417 * feature itself but allows the user to enable it later. 10418 */ 10419 if (dev->hw_features & NETIF_F_TSO) 10420 dev->hw_features |= NETIF_F_TSO_MANGLEID; 10421 if (dev->vlan_features & NETIF_F_TSO) 10422 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 10423 if (dev->mpls_features & NETIF_F_TSO) 10424 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 10425 if (dev->hw_enc_features & NETIF_F_TSO) 10426 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 10427 10428 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 10429 */ 10430 dev->vlan_features |= NETIF_F_HIGHDMA; 10431 10432 /* Make NETIF_F_SG inheritable to tunnel devices. 10433 */ 10434 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 10435 10436 /* Make NETIF_F_SG inheritable to MPLS. 10437 */ 10438 dev->mpls_features |= NETIF_F_SG; 10439 10440 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 10441 ret = notifier_to_errno(ret); 10442 if (ret) 10443 goto err_ifindex_release; 10444 10445 ret = netdev_register_kobject(dev); 10446 10447 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 10448 10449 if (ret) 10450 goto err_uninit_notify; 10451 10452 __netdev_update_features(dev); 10453 10454 /* 10455 * Default initial state at registry is that the 10456 * device is present. 10457 */ 10458 10459 set_bit(__LINK_STATE_PRESENT, &dev->state); 10460 10461 linkwatch_init_dev(dev); 10462 10463 dev_init_scheduler(dev); 10464 10465 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 10466 list_netdevice(dev); 10467 10468 add_device_randomness(dev->dev_addr, dev->addr_len); 10469 10470 /* If the device has permanent device address, driver should 10471 * set dev_addr and also addr_assign_type should be set to 10472 * NET_ADDR_PERM (default value). 10473 */ 10474 if (dev->addr_assign_type == NET_ADDR_PERM) 10475 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 10476 10477 /* Notify protocols, that a new device appeared. */ 10478 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 10479 ret = notifier_to_errno(ret); 10480 if (ret) { 10481 /* Expect explicit free_netdev() on failure */ 10482 dev->needs_free_netdev = false; 10483 unregister_netdevice_queue(dev, NULL); 10484 goto out; 10485 } 10486 /* 10487 * Prevent userspace races by waiting until the network 10488 * device is fully setup before sending notifications. 10489 */ 10490 if (!dev->rtnl_link_ops || 10491 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 10492 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 10493 10494 out: 10495 return ret; 10496 10497 err_uninit_notify: 10498 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 10499 err_ifindex_release: 10500 dev_index_release(net, dev->ifindex); 10501 err_free_pcpu: 10502 netdev_do_free_pcpu_stats(dev); 10503 err_uninit: 10504 if (dev->netdev_ops->ndo_uninit) 10505 dev->netdev_ops->ndo_uninit(dev); 10506 if (dev->priv_destructor) 10507 dev->priv_destructor(dev); 10508 err_free_name: 10509 netdev_name_node_free(dev->name_node); 10510 goto out; 10511 } 10512 EXPORT_SYMBOL(register_netdevice); 10513 10514 /* Initialize the core of a dummy net device. 10515 * This is useful if you are calling this function after alloc_netdev(), 10516 * since it does not memset the net_device fields. 10517 */ 10518 static void init_dummy_netdev_core(struct net_device *dev) 10519 { 10520 /* make sure we BUG if trying to hit standard 10521 * register/unregister code path 10522 */ 10523 dev->reg_state = NETREG_DUMMY; 10524 10525 /* NAPI wants this */ 10526 INIT_LIST_HEAD(&dev->napi_list); 10527 10528 /* a dummy interface is started by default */ 10529 set_bit(__LINK_STATE_PRESENT, &dev->state); 10530 set_bit(__LINK_STATE_START, &dev->state); 10531 10532 /* napi_busy_loop stats accounting wants this */ 10533 dev_net_set(dev, &init_net); 10534 10535 /* Note : We dont allocate pcpu_refcnt for dummy devices, 10536 * because users of this 'device' dont need to change 10537 * its refcount. 10538 */ 10539 } 10540 10541 /** 10542 * init_dummy_netdev - init a dummy network device for NAPI 10543 * @dev: device to init 10544 * 10545 * This takes a network device structure and initializes the minimum 10546 * amount of fields so it can be used to schedule NAPI polls without 10547 * registering a full blown interface. This is to be used by drivers 10548 * that need to tie several hardware interfaces to a single NAPI 10549 * poll scheduler due to HW limitations. 10550 */ 10551 void init_dummy_netdev(struct net_device *dev) 10552 { 10553 /* Clear everything. Note we don't initialize spinlocks 10554 * as they aren't supposed to be taken by any of the 10555 * NAPI code and this dummy netdev is supposed to be 10556 * only ever used for NAPI polls 10557 */ 10558 memset(dev, 0, sizeof(struct net_device)); 10559 init_dummy_netdev_core(dev); 10560 } 10561 EXPORT_SYMBOL_GPL(init_dummy_netdev); 10562 10563 /** 10564 * register_netdev - register a network device 10565 * @dev: device to register 10566 * 10567 * Take a completed network device structure and add it to the kernel 10568 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 10569 * chain. 0 is returned on success. A negative errno code is returned 10570 * on a failure to set up the device, or if the name is a duplicate. 10571 * 10572 * This is a wrapper around register_netdevice that takes the rtnl semaphore 10573 * and expands the device name if you passed a format string to 10574 * alloc_netdev. 10575 */ 10576 int register_netdev(struct net_device *dev) 10577 { 10578 int err; 10579 10580 if (rtnl_lock_killable()) 10581 return -EINTR; 10582 err = register_netdevice(dev); 10583 rtnl_unlock(); 10584 return err; 10585 } 10586 EXPORT_SYMBOL(register_netdev); 10587 10588 int netdev_refcnt_read(const struct net_device *dev) 10589 { 10590 #ifdef CONFIG_PCPU_DEV_REFCNT 10591 int i, refcnt = 0; 10592 10593 for_each_possible_cpu(i) 10594 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 10595 return refcnt; 10596 #else 10597 return refcount_read(&dev->dev_refcnt); 10598 #endif 10599 } 10600 EXPORT_SYMBOL(netdev_refcnt_read); 10601 10602 int netdev_unregister_timeout_secs __read_mostly = 10; 10603 10604 #define WAIT_REFS_MIN_MSECS 1 10605 #define WAIT_REFS_MAX_MSECS 250 10606 /** 10607 * netdev_wait_allrefs_any - wait until all references are gone. 10608 * @list: list of net_devices to wait on 10609 * 10610 * This is called when unregistering network devices. 10611 * 10612 * Any protocol or device that holds a reference should register 10613 * for netdevice notification, and cleanup and put back the 10614 * reference if they receive an UNREGISTER event. 10615 * We can get stuck here if buggy protocols don't correctly 10616 * call dev_put. 10617 */ 10618 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 10619 { 10620 unsigned long rebroadcast_time, warning_time; 10621 struct net_device *dev; 10622 int wait = 0; 10623 10624 rebroadcast_time = warning_time = jiffies; 10625 10626 list_for_each_entry(dev, list, todo_list) 10627 if (netdev_refcnt_read(dev) == 1) 10628 return dev; 10629 10630 while (true) { 10631 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 10632 rtnl_lock(); 10633 10634 /* Rebroadcast unregister notification */ 10635 list_for_each_entry(dev, list, todo_list) 10636 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 10637 10638 __rtnl_unlock(); 10639 rcu_barrier(); 10640 rtnl_lock(); 10641 10642 list_for_each_entry(dev, list, todo_list) 10643 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 10644 &dev->state)) { 10645 /* We must not have linkwatch events 10646 * pending on unregister. If this 10647 * happens, we simply run the queue 10648 * unscheduled, resulting in a noop 10649 * for this device. 10650 */ 10651 linkwatch_run_queue(); 10652 break; 10653 } 10654 10655 __rtnl_unlock(); 10656 10657 rebroadcast_time = jiffies; 10658 } 10659 10660 rcu_barrier(); 10661 10662 if (!wait) { 10663 wait = WAIT_REFS_MIN_MSECS; 10664 } else { 10665 msleep(wait); 10666 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 10667 } 10668 10669 list_for_each_entry(dev, list, todo_list) 10670 if (netdev_refcnt_read(dev) == 1) 10671 return dev; 10672 10673 if (time_after(jiffies, warning_time + 10674 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 10675 list_for_each_entry(dev, list, todo_list) { 10676 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 10677 dev->name, netdev_refcnt_read(dev)); 10678 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 10679 } 10680 10681 warning_time = jiffies; 10682 } 10683 } 10684 } 10685 10686 /* The sequence is: 10687 * 10688 * rtnl_lock(); 10689 * ... 10690 * register_netdevice(x1); 10691 * register_netdevice(x2); 10692 * ... 10693 * unregister_netdevice(y1); 10694 * unregister_netdevice(y2); 10695 * ... 10696 * rtnl_unlock(); 10697 * free_netdev(y1); 10698 * free_netdev(y2); 10699 * 10700 * We are invoked by rtnl_unlock(). 10701 * This allows us to deal with problems: 10702 * 1) We can delete sysfs objects which invoke hotplug 10703 * without deadlocking with linkwatch via keventd. 10704 * 2) Since we run with the RTNL semaphore not held, we can sleep 10705 * safely in order to wait for the netdev refcnt to drop to zero. 10706 * 10707 * We must not return until all unregister events added during 10708 * the interval the lock was held have been completed. 10709 */ 10710 void netdev_run_todo(void) 10711 { 10712 struct net_device *dev, *tmp; 10713 struct list_head list; 10714 int cnt; 10715 #ifdef CONFIG_LOCKDEP 10716 struct list_head unlink_list; 10717 10718 list_replace_init(&net_unlink_list, &unlink_list); 10719 10720 while (!list_empty(&unlink_list)) { 10721 struct net_device *dev = list_first_entry(&unlink_list, 10722 struct net_device, 10723 unlink_list); 10724 list_del_init(&dev->unlink_list); 10725 dev->nested_level = dev->lower_level - 1; 10726 } 10727 #endif 10728 10729 /* Snapshot list, allow later requests */ 10730 list_replace_init(&net_todo_list, &list); 10731 10732 __rtnl_unlock(); 10733 10734 /* Wait for rcu callbacks to finish before next phase */ 10735 if (!list_empty(&list)) 10736 rcu_barrier(); 10737 10738 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 10739 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 10740 netdev_WARN(dev, "run_todo but not unregistering\n"); 10741 list_del(&dev->todo_list); 10742 continue; 10743 } 10744 10745 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 10746 linkwatch_sync_dev(dev); 10747 } 10748 10749 cnt = 0; 10750 while (!list_empty(&list)) { 10751 dev = netdev_wait_allrefs_any(&list); 10752 list_del(&dev->todo_list); 10753 10754 /* paranoia */ 10755 BUG_ON(netdev_refcnt_read(dev) != 1); 10756 BUG_ON(!list_empty(&dev->ptype_all)); 10757 BUG_ON(!list_empty(&dev->ptype_specific)); 10758 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 10759 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 10760 10761 netdev_do_free_pcpu_stats(dev); 10762 if (dev->priv_destructor) 10763 dev->priv_destructor(dev); 10764 if (dev->needs_free_netdev) 10765 free_netdev(dev); 10766 10767 cnt++; 10768 10769 /* Free network device */ 10770 kobject_put(&dev->dev.kobj); 10771 } 10772 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 10773 wake_up(&netdev_unregistering_wq); 10774 } 10775 10776 /* Collate per-cpu network dstats statistics 10777 * 10778 * Read per-cpu network statistics from dev->dstats and populate the related 10779 * fields in @s. 10780 */ 10781 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 10782 const struct pcpu_dstats __percpu *dstats) 10783 { 10784 int cpu; 10785 10786 for_each_possible_cpu(cpu) { 10787 u64 rx_packets, rx_bytes, rx_drops; 10788 u64 tx_packets, tx_bytes, tx_drops; 10789 const struct pcpu_dstats *stats; 10790 unsigned int start; 10791 10792 stats = per_cpu_ptr(dstats, cpu); 10793 do { 10794 start = u64_stats_fetch_begin(&stats->syncp); 10795 rx_packets = u64_stats_read(&stats->rx_packets); 10796 rx_bytes = u64_stats_read(&stats->rx_bytes); 10797 rx_drops = u64_stats_read(&stats->rx_drops); 10798 tx_packets = u64_stats_read(&stats->tx_packets); 10799 tx_bytes = u64_stats_read(&stats->tx_bytes); 10800 tx_drops = u64_stats_read(&stats->tx_drops); 10801 } while (u64_stats_fetch_retry(&stats->syncp, start)); 10802 10803 s->rx_packets += rx_packets; 10804 s->rx_bytes += rx_bytes; 10805 s->rx_dropped += rx_drops; 10806 s->tx_packets += tx_packets; 10807 s->tx_bytes += tx_bytes; 10808 s->tx_dropped += tx_drops; 10809 } 10810 } 10811 10812 /* ndo_get_stats64 implementation for dtstats-based accounting. 10813 * 10814 * Populate @s from dev->stats and dev->dstats. This is used internally by the 10815 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 10816 */ 10817 static void dev_get_dstats64(const struct net_device *dev, 10818 struct rtnl_link_stats64 *s) 10819 { 10820 netdev_stats_to_stats64(s, &dev->stats); 10821 dev_fetch_dstats(s, dev->dstats); 10822 } 10823 10824 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 10825 * all the same fields in the same order as net_device_stats, with only 10826 * the type differing, but rtnl_link_stats64 may have additional fields 10827 * at the end for newer counters. 10828 */ 10829 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 10830 const struct net_device_stats *netdev_stats) 10831 { 10832 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 10833 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 10834 u64 *dst = (u64 *)stats64; 10835 10836 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 10837 for (i = 0; i < n; i++) 10838 dst[i] = (unsigned long)atomic_long_read(&src[i]); 10839 /* zero out counters that only exist in rtnl_link_stats64 */ 10840 memset((char *)stats64 + n * sizeof(u64), 0, 10841 sizeof(*stats64) - n * sizeof(u64)); 10842 } 10843 EXPORT_SYMBOL(netdev_stats_to_stats64); 10844 10845 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 10846 struct net_device *dev) 10847 { 10848 struct net_device_core_stats __percpu *p; 10849 10850 p = alloc_percpu_gfp(struct net_device_core_stats, 10851 GFP_ATOMIC | __GFP_NOWARN); 10852 10853 if (p && cmpxchg(&dev->core_stats, NULL, p)) 10854 free_percpu(p); 10855 10856 /* This READ_ONCE() pairs with the cmpxchg() above */ 10857 return READ_ONCE(dev->core_stats); 10858 } 10859 10860 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 10861 { 10862 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 10863 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 10864 unsigned long __percpu *field; 10865 10866 if (unlikely(!p)) { 10867 p = netdev_core_stats_alloc(dev); 10868 if (!p) 10869 return; 10870 } 10871 10872 field = (unsigned long __percpu *)((void __percpu *)p + offset); 10873 this_cpu_inc(*field); 10874 } 10875 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 10876 10877 /** 10878 * dev_get_stats - get network device statistics 10879 * @dev: device to get statistics from 10880 * @storage: place to store stats 10881 * 10882 * Get network statistics from device. Return @storage. 10883 * The device driver may provide its own method by setting 10884 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 10885 * otherwise the internal statistics structure is used. 10886 */ 10887 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 10888 struct rtnl_link_stats64 *storage) 10889 { 10890 const struct net_device_ops *ops = dev->netdev_ops; 10891 const struct net_device_core_stats __percpu *p; 10892 10893 if (ops->ndo_get_stats64) { 10894 memset(storage, 0, sizeof(*storage)); 10895 ops->ndo_get_stats64(dev, storage); 10896 } else if (ops->ndo_get_stats) { 10897 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 10898 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 10899 dev_get_tstats64(dev, storage); 10900 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 10901 dev_get_dstats64(dev, storage); 10902 } else { 10903 netdev_stats_to_stats64(storage, &dev->stats); 10904 } 10905 10906 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 10907 p = READ_ONCE(dev->core_stats); 10908 if (p) { 10909 const struct net_device_core_stats *core_stats; 10910 int i; 10911 10912 for_each_possible_cpu(i) { 10913 core_stats = per_cpu_ptr(p, i); 10914 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 10915 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 10916 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 10917 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 10918 } 10919 } 10920 return storage; 10921 } 10922 EXPORT_SYMBOL(dev_get_stats); 10923 10924 /** 10925 * dev_fetch_sw_netstats - get per-cpu network device statistics 10926 * @s: place to store stats 10927 * @netstats: per-cpu network stats to read from 10928 * 10929 * Read per-cpu network statistics and populate the related fields in @s. 10930 */ 10931 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 10932 const struct pcpu_sw_netstats __percpu *netstats) 10933 { 10934 int cpu; 10935 10936 for_each_possible_cpu(cpu) { 10937 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 10938 const struct pcpu_sw_netstats *stats; 10939 unsigned int start; 10940 10941 stats = per_cpu_ptr(netstats, cpu); 10942 do { 10943 start = u64_stats_fetch_begin(&stats->syncp); 10944 rx_packets = u64_stats_read(&stats->rx_packets); 10945 rx_bytes = u64_stats_read(&stats->rx_bytes); 10946 tx_packets = u64_stats_read(&stats->tx_packets); 10947 tx_bytes = u64_stats_read(&stats->tx_bytes); 10948 } while (u64_stats_fetch_retry(&stats->syncp, start)); 10949 10950 s->rx_packets += rx_packets; 10951 s->rx_bytes += rx_bytes; 10952 s->tx_packets += tx_packets; 10953 s->tx_bytes += tx_bytes; 10954 } 10955 } 10956 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 10957 10958 /** 10959 * dev_get_tstats64 - ndo_get_stats64 implementation 10960 * @dev: device to get statistics from 10961 * @s: place to store stats 10962 * 10963 * Populate @s from dev->stats and dev->tstats. Can be used as 10964 * ndo_get_stats64() callback. 10965 */ 10966 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 10967 { 10968 netdev_stats_to_stats64(s, &dev->stats); 10969 dev_fetch_sw_netstats(s, dev->tstats); 10970 } 10971 EXPORT_SYMBOL_GPL(dev_get_tstats64); 10972 10973 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 10974 { 10975 struct netdev_queue *queue = dev_ingress_queue(dev); 10976 10977 #ifdef CONFIG_NET_CLS_ACT 10978 if (queue) 10979 return queue; 10980 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 10981 if (!queue) 10982 return NULL; 10983 netdev_init_one_queue(dev, queue, NULL); 10984 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 10985 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 10986 rcu_assign_pointer(dev->ingress_queue, queue); 10987 #endif 10988 return queue; 10989 } 10990 10991 static const struct ethtool_ops default_ethtool_ops; 10992 10993 void netdev_set_default_ethtool_ops(struct net_device *dev, 10994 const struct ethtool_ops *ops) 10995 { 10996 if (dev->ethtool_ops == &default_ethtool_ops) 10997 dev->ethtool_ops = ops; 10998 } 10999 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11000 11001 /** 11002 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11003 * @dev: netdev to enable the IRQ coalescing on 11004 * 11005 * Sets a conservative default for SW IRQ coalescing. Users can use 11006 * sysfs attributes to override the default values. 11007 */ 11008 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11009 { 11010 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11011 11012 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11013 dev->gro_flush_timeout = 20000; 11014 dev->napi_defer_hard_irqs = 1; 11015 } 11016 } 11017 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11018 11019 /** 11020 * alloc_netdev_mqs - allocate network device 11021 * @sizeof_priv: size of private data to allocate space for 11022 * @name: device name format string 11023 * @name_assign_type: origin of device name 11024 * @setup: callback to initialize device 11025 * @txqs: the number of TX subqueues to allocate 11026 * @rxqs: the number of RX subqueues to allocate 11027 * 11028 * Allocates a struct net_device with private data area for driver use 11029 * and performs basic initialization. Also allocates subqueue structs 11030 * for each queue on the device. 11031 */ 11032 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11033 unsigned char name_assign_type, 11034 void (*setup)(struct net_device *), 11035 unsigned int txqs, unsigned int rxqs) 11036 { 11037 struct net_device *dev; 11038 11039 BUG_ON(strlen(name) >= sizeof(dev->name)); 11040 11041 if (txqs < 1) { 11042 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 11043 return NULL; 11044 } 11045 11046 if (rxqs < 1) { 11047 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 11048 return NULL; 11049 } 11050 11051 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 11052 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11053 if (!dev) 11054 return NULL; 11055 11056 dev->priv_len = sizeof_priv; 11057 11058 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 11059 #ifdef CONFIG_PCPU_DEV_REFCNT 11060 dev->pcpu_refcnt = alloc_percpu(int); 11061 if (!dev->pcpu_refcnt) 11062 goto free_dev; 11063 __dev_hold(dev); 11064 #else 11065 refcount_set(&dev->dev_refcnt, 1); 11066 #endif 11067 11068 if (dev_addr_init(dev)) 11069 goto free_pcpu; 11070 11071 dev_mc_init(dev); 11072 dev_uc_init(dev); 11073 11074 dev_net_set(dev, &init_net); 11075 11076 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 11077 dev->xdp_zc_max_segs = 1; 11078 dev->gso_max_segs = GSO_MAX_SEGS; 11079 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 11080 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 11081 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 11082 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 11083 dev->tso_max_segs = TSO_MAX_SEGS; 11084 dev->upper_level = 1; 11085 dev->lower_level = 1; 11086 #ifdef CONFIG_LOCKDEP 11087 dev->nested_level = 0; 11088 INIT_LIST_HEAD(&dev->unlink_list); 11089 #endif 11090 11091 INIT_LIST_HEAD(&dev->napi_list); 11092 INIT_LIST_HEAD(&dev->unreg_list); 11093 INIT_LIST_HEAD(&dev->close_list); 11094 INIT_LIST_HEAD(&dev->link_watch_list); 11095 INIT_LIST_HEAD(&dev->adj_list.upper); 11096 INIT_LIST_HEAD(&dev->adj_list.lower); 11097 INIT_LIST_HEAD(&dev->ptype_all); 11098 INIT_LIST_HEAD(&dev->ptype_specific); 11099 INIT_LIST_HEAD(&dev->net_notifier_list); 11100 #ifdef CONFIG_NET_SCHED 11101 hash_init(dev->qdisc_hash); 11102 #endif 11103 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 11104 setup(dev); 11105 11106 if (!dev->tx_queue_len) { 11107 dev->priv_flags |= IFF_NO_QUEUE; 11108 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 11109 } 11110 11111 dev->num_tx_queues = txqs; 11112 dev->real_num_tx_queues = txqs; 11113 if (netif_alloc_netdev_queues(dev)) 11114 goto free_all; 11115 11116 dev->num_rx_queues = rxqs; 11117 dev->real_num_rx_queues = rxqs; 11118 if (netif_alloc_rx_queues(dev)) 11119 goto free_all; 11120 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 11121 if (!dev->ethtool) 11122 goto free_all; 11123 11124 strcpy(dev->name, name); 11125 dev->name_assign_type = name_assign_type; 11126 dev->group = INIT_NETDEV_GROUP; 11127 if (!dev->ethtool_ops) 11128 dev->ethtool_ops = &default_ethtool_ops; 11129 11130 nf_hook_netdev_init(dev); 11131 11132 return dev; 11133 11134 free_all: 11135 free_netdev(dev); 11136 return NULL; 11137 11138 free_pcpu: 11139 #ifdef CONFIG_PCPU_DEV_REFCNT 11140 free_percpu(dev->pcpu_refcnt); 11141 free_dev: 11142 #endif 11143 kvfree(dev); 11144 return NULL; 11145 } 11146 EXPORT_SYMBOL(alloc_netdev_mqs); 11147 11148 /** 11149 * free_netdev - free network device 11150 * @dev: device 11151 * 11152 * This function does the last stage of destroying an allocated device 11153 * interface. The reference to the device object is released. If this 11154 * is the last reference then it will be freed.Must be called in process 11155 * context. 11156 */ 11157 void free_netdev(struct net_device *dev) 11158 { 11159 struct napi_struct *p, *n; 11160 11161 might_sleep(); 11162 11163 /* When called immediately after register_netdevice() failed the unwind 11164 * handling may still be dismantling the device. Handle that case by 11165 * deferring the free. 11166 */ 11167 if (dev->reg_state == NETREG_UNREGISTERING) { 11168 ASSERT_RTNL(); 11169 dev->needs_free_netdev = true; 11170 return; 11171 } 11172 11173 kfree(dev->ethtool); 11174 netif_free_tx_queues(dev); 11175 netif_free_rx_queues(dev); 11176 11177 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 11178 11179 /* Flush device addresses */ 11180 dev_addr_flush(dev); 11181 11182 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 11183 netif_napi_del(p); 11184 11185 ref_tracker_dir_exit(&dev->refcnt_tracker); 11186 #ifdef CONFIG_PCPU_DEV_REFCNT 11187 free_percpu(dev->pcpu_refcnt); 11188 dev->pcpu_refcnt = NULL; 11189 #endif 11190 free_percpu(dev->core_stats); 11191 dev->core_stats = NULL; 11192 free_percpu(dev->xdp_bulkq); 11193 dev->xdp_bulkq = NULL; 11194 11195 /* Compatibility with error handling in drivers */ 11196 if (dev->reg_state == NETREG_UNINITIALIZED || 11197 dev->reg_state == NETREG_DUMMY) { 11198 kvfree(dev); 11199 return; 11200 } 11201 11202 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 11203 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 11204 11205 /* will free via device release */ 11206 put_device(&dev->dev); 11207 } 11208 EXPORT_SYMBOL(free_netdev); 11209 11210 /** 11211 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 11212 * @sizeof_priv: size of private data to allocate space for 11213 * 11214 * Return: the allocated net_device on success, NULL otherwise 11215 */ 11216 struct net_device *alloc_netdev_dummy(int sizeof_priv) 11217 { 11218 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 11219 init_dummy_netdev_core); 11220 } 11221 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 11222 11223 /** 11224 * synchronize_net - Synchronize with packet receive processing 11225 * 11226 * Wait for packets currently being received to be done. 11227 * Does not block later packets from starting. 11228 */ 11229 void synchronize_net(void) 11230 { 11231 might_sleep(); 11232 if (rtnl_is_locked()) 11233 synchronize_rcu_expedited(); 11234 else 11235 synchronize_rcu(); 11236 } 11237 EXPORT_SYMBOL(synchronize_net); 11238 11239 static void netdev_rss_contexts_free(struct net_device *dev) 11240 { 11241 struct ethtool_rxfh_context *ctx; 11242 unsigned long context; 11243 11244 mutex_lock(&dev->ethtool->rss_lock); 11245 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 11246 struct ethtool_rxfh_param rxfh; 11247 11248 rxfh.indir = ethtool_rxfh_context_indir(ctx); 11249 rxfh.key = ethtool_rxfh_context_key(ctx); 11250 rxfh.hfunc = ctx->hfunc; 11251 rxfh.input_xfrm = ctx->input_xfrm; 11252 rxfh.rss_context = context; 11253 rxfh.rss_delete = true; 11254 11255 xa_erase(&dev->ethtool->rss_ctx, context); 11256 if (dev->ethtool_ops->create_rxfh_context) 11257 dev->ethtool_ops->remove_rxfh_context(dev, ctx, 11258 context, NULL); 11259 else 11260 dev->ethtool_ops->set_rxfh(dev, &rxfh, NULL); 11261 kfree(ctx); 11262 } 11263 xa_destroy(&dev->ethtool->rss_ctx); 11264 mutex_unlock(&dev->ethtool->rss_lock); 11265 } 11266 11267 /** 11268 * unregister_netdevice_queue - remove device from the kernel 11269 * @dev: device 11270 * @head: list 11271 * 11272 * This function shuts down a device interface and removes it 11273 * from the kernel tables. 11274 * If head not NULL, device is queued to be unregistered later. 11275 * 11276 * Callers must hold the rtnl semaphore. You may want 11277 * unregister_netdev() instead of this. 11278 */ 11279 11280 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11281 { 11282 ASSERT_RTNL(); 11283 11284 if (head) { 11285 list_move_tail(&dev->unreg_list, head); 11286 } else { 11287 LIST_HEAD(single); 11288 11289 list_add(&dev->unreg_list, &single); 11290 unregister_netdevice_many(&single); 11291 } 11292 } 11293 EXPORT_SYMBOL(unregister_netdevice_queue); 11294 11295 void unregister_netdevice_many_notify(struct list_head *head, 11296 u32 portid, const struct nlmsghdr *nlh) 11297 { 11298 struct net_device *dev, *tmp; 11299 LIST_HEAD(close_head); 11300 int cnt = 0; 11301 11302 BUG_ON(dev_boot_phase); 11303 ASSERT_RTNL(); 11304 11305 if (list_empty(head)) 11306 return; 11307 11308 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11309 /* Some devices call without registering 11310 * for initialization unwind. Remove those 11311 * devices and proceed with the remaining. 11312 */ 11313 if (dev->reg_state == NETREG_UNINITIALIZED) { 11314 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11315 dev->name, dev); 11316 11317 WARN_ON(1); 11318 list_del(&dev->unreg_list); 11319 continue; 11320 } 11321 dev->dismantle = true; 11322 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11323 } 11324 11325 /* If device is running, close it first. */ 11326 list_for_each_entry(dev, head, unreg_list) 11327 list_add_tail(&dev->close_list, &close_head); 11328 dev_close_many(&close_head, true); 11329 11330 list_for_each_entry(dev, head, unreg_list) { 11331 /* And unlink it from device chain. */ 11332 unlist_netdevice(dev); 11333 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 11334 } 11335 flush_all_backlogs(); 11336 11337 synchronize_net(); 11338 11339 list_for_each_entry(dev, head, unreg_list) { 11340 struct sk_buff *skb = NULL; 11341 11342 /* Shutdown queueing discipline. */ 11343 dev_shutdown(dev); 11344 dev_tcx_uninstall(dev); 11345 dev_xdp_uninstall(dev); 11346 bpf_dev_bound_netdev_unregister(dev); 11347 11348 netdev_offload_xstats_disable_all(dev); 11349 11350 /* Notify protocols, that we are about to destroy 11351 * this device. They should clean all the things. 11352 */ 11353 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11354 11355 if (!dev->rtnl_link_ops || 11356 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11357 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 11358 GFP_KERNEL, NULL, 0, 11359 portid, nlh); 11360 11361 /* 11362 * Flush the unicast and multicast chains 11363 */ 11364 dev_uc_flush(dev); 11365 dev_mc_flush(dev); 11366 11367 netdev_name_node_alt_flush(dev); 11368 netdev_name_node_free(dev->name_node); 11369 11370 netdev_rss_contexts_free(dev); 11371 11372 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11373 11374 if (dev->netdev_ops->ndo_uninit) 11375 dev->netdev_ops->ndo_uninit(dev); 11376 11377 mutex_destroy(&dev->ethtool->rss_lock); 11378 11379 if (skb) 11380 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 11381 11382 /* Notifier chain MUST detach us all upper devices. */ 11383 WARN_ON(netdev_has_any_upper_dev(dev)); 11384 WARN_ON(netdev_has_any_lower_dev(dev)); 11385 11386 /* Remove entries from kobject tree */ 11387 netdev_unregister_kobject(dev); 11388 #ifdef CONFIG_XPS 11389 /* Remove XPS queueing entries */ 11390 netif_reset_xps_queues_gt(dev, 0); 11391 #endif 11392 } 11393 11394 synchronize_net(); 11395 11396 list_for_each_entry(dev, head, unreg_list) { 11397 netdev_put(dev, &dev->dev_registered_tracker); 11398 net_set_todo(dev); 11399 cnt++; 11400 } 11401 atomic_add(cnt, &dev_unreg_count); 11402 11403 list_del(head); 11404 } 11405 11406 /** 11407 * unregister_netdevice_many - unregister many devices 11408 * @head: list of devices 11409 * 11410 * Note: As most callers use a stack allocated list_head, 11411 * we force a list_del() to make sure stack wont be corrupted later. 11412 */ 11413 void unregister_netdevice_many(struct list_head *head) 11414 { 11415 unregister_netdevice_many_notify(head, 0, NULL); 11416 } 11417 EXPORT_SYMBOL(unregister_netdevice_many); 11418 11419 /** 11420 * unregister_netdev - remove device from the kernel 11421 * @dev: device 11422 * 11423 * This function shuts down a device interface and removes it 11424 * from the kernel tables. 11425 * 11426 * This is just a wrapper for unregister_netdevice that takes 11427 * the rtnl semaphore. In general you want to use this and not 11428 * unregister_netdevice. 11429 */ 11430 void unregister_netdev(struct net_device *dev) 11431 { 11432 rtnl_lock(); 11433 unregister_netdevice(dev); 11434 rtnl_unlock(); 11435 } 11436 EXPORT_SYMBOL(unregister_netdev); 11437 11438 /** 11439 * __dev_change_net_namespace - move device to different nethost namespace 11440 * @dev: device 11441 * @net: network namespace 11442 * @pat: If not NULL name pattern to try if the current device name 11443 * is already taken in the destination network namespace. 11444 * @new_ifindex: If not zero, specifies device index in the target 11445 * namespace. 11446 * 11447 * This function shuts down a device interface and moves it 11448 * to a new network namespace. On success 0 is returned, on 11449 * a failure a netagive errno code is returned. 11450 * 11451 * Callers must hold the rtnl semaphore. 11452 */ 11453 11454 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 11455 const char *pat, int new_ifindex) 11456 { 11457 struct netdev_name_node *name_node; 11458 struct net *net_old = dev_net(dev); 11459 char new_name[IFNAMSIZ] = {}; 11460 int err, new_nsid; 11461 11462 ASSERT_RTNL(); 11463 11464 /* Don't allow namespace local devices to be moved. */ 11465 err = -EINVAL; 11466 if (dev->features & NETIF_F_NETNS_LOCAL) 11467 goto out; 11468 11469 /* Ensure the device has been registrered */ 11470 if (dev->reg_state != NETREG_REGISTERED) 11471 goto out; 11472 11473 /* Get out if there is nothing todo */ 11474 err = 0; 11475 if (net_eq(net_old, net)) 11476 goto out; 11477 11478 /* Pick the destination device name, and ensure 11479 * we can use it in the destination network namespace. 11480 */ 11481 err = -EEXIST; 11482 if (netdev_name_in_use(net, dev->name)) { 11483 /* We get here if we can't use the current device name */ 11484 if (!pat) 11485 goto out; 11486 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 11487 if (err < 0) 11488 goto out; 11489 } 11490 /* Check that none of the altnames conflicts. */ 11491 err = -EEXIST; 11492 netdev_for_each_altname(dev, name_node) 11493 if (netdev_name_in_use(net, name_node->name)) 11494 goto out; 11495 11496 /* Check that new_ifindex isn't used yet. */ 11497 if (new_ifindex) { 11498 err = dev_index_reserve(net, new_ifindex); 11499 if (err < 0) 11500 goto out; 11501 } else { 11502 /* If there is an ifindex conflict assign a new one */ 11503 err = dev_index_reserve(net, dev->ifindex); 11504 if (err == -EBUSY) 11505 err = dev_index_reserve(net, 0); 11506 if (err < 0) 11507 goto out; 11508 new_ifindex = err; 11509 } 11510 11511 /* 11512 * And now a mini version of register_netdevice unregister_netdevice. 11513 */ 11514 11515 /* If device is running close it first. */ 11516 dev_close(dev); 11517 11518 /* And unlink it from device chain */ 11519 unlist_netdevice(dev); 11520 11521 synchronize_net(); 11522 11523 /* Shutdown queueing discipline. */ 11524 dev_shutdown(dev); 11525 11526 /* Notify protocols, that we are about to destroy 11527 * this device. They should clean all the things. 11528 * 11529 * Note that dev->reg_state stays at NETREG_REGISTERED. 11530 * This is wanted because this way 8021q and macvlan know 11531 * the device is just moving and can keep their slaves up. 11532 */ 11533 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11534 rcu_barrier(); 11535 11536 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 11537 11538 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 11539 new_ifindex); 11540 11541 /* 11542 * Flush the unicast and multicast chains 11543 */ 11544 dev_uc_flush(dev); 11545 dev_mc_flush(dev); 11546 11547 /* Send a netdev-removed uevent to the old namespace */ 11548 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 11549 netdev_adjacent_del_links(dev); 11550 11551 /* Move per-net netdevice notifiers that are following the netdevice */ 11552 move_netdevice_notifiers_dev_net(dev, net); 11553 11554 /* Actually switch the network namespace */ 11555 dev_net_set(dev, net); 11556 dev->ifindex = new_ifindex; 11557 11558 if (new_name[0]) { 11559 /* Rename the netdev to prepared name */ 11560 write_seqlock_bh(&netdev_rename_lock); 11561 strscpy(dev->name, new_name, IFNAMSIZ); 11562 write_sequnlock_bh(&netdev_rename_lock); 11563 } 11564 11565 /* Fixup kobjects */ 11566 dev_set_uevent_suppress(&dev->dev, 1); 11567 err = device_rename(&dev->dev, dev->name); 11568 dev_set_uevent_suppress(&dev->dev, 0); 11569 WARN_ON(err); 11570 11571 /* Send a netdev-add uevent to the new namespace */ 11572 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 11573 netdev_adjacent_add_links(dev); 11574 11575 /* Adapt owner in case owning user namespace of target network 11576 * namespace is different from the original one. 11577 */ 11578 err = netdev_change_owner(dev, net_old, net); 11579 WARN_ON(err); 11580 11581 /* Add the device back in the hashes */ 11582 list_netdevice(dev); 11583 11584 /* Notify protocols, that a new device appeared. */ 11585 call_netdevice_notifiers(NETDEV_REGISTER, dev); 11586 11587 /* 11588 * Prevent userspace races by waiting until the network 11589 * device is fully setup before sending notifications. 11590 */ 11591 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11592 11593 synchronize_net(); 11594 err = 0; 11595 out: 11596 return err; 11597 } 11598 EXPORT_SYMBOL_GPL(__dev_change_net_namespace); 11599 11600 static int dev_cpu_dead(unsigned int oldcpu) 11601 { 11602 struct sk_buff **list_skb; 11603 struct sk_buff *skb; 11604 unsigned int cpu; 11605 struct softnet_data *sd, *oldsd, *remsd = NULL; 11606 11607 local_irq_disable(); 11608 cpu = smp_processor_id(); 11609 sd = &per_cpu(softnet_data, cpu); 11610 oldsd = &per_cpu(softnet_data, oldcpu); 11611 11612 /* Find end of our completion_queue. */ 11613 list_skb = &sd->completion_queue; 11614 while (*list_skb) 11615 list_skb = &(*list_skb)->next; 11616 /* Append completion queue from offline CPU. */ 11617 *list_skb = oldsd->completion_queue; 11618 oldsd->completion_queue = NULL; 11619 11620 /* Append output queue from offline CPU. */ 11621 if (oldsd->output_queue) { 11622 *sd->output_queue_tailp = oldsd->output_queue; 11623 sd->output_queue_tailp = oldsd->output_queue_tailp; 11624 oldsd->output_queue = NULL; 11625 oldsd->output_queue_tailp = &oldsd->output_queue; 11626 } 11627 /* Append NAPI poll list from offline CPU, with one exception : 11628 * process_backlog() must be called by cpu owning percpu backlog. 11629 * We properly handle process_queue & input_pkt_queue later. 11630 */ 11631 while (!list_empty(&oldsd->poll_list)) { 11632 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 11633 struct napi_struct, 11634 poll_list); 11635 11636 list_del_init(&napi->poll_list); 11637 if (napi->poll == process_backlog) 11638 napi->state &= NAPIF_STATE_THREADED; 11639 else 11640 ____napi_schedule(sd, napi); 11641 } 11642 11643 raise_softirq_irqoff(NET_TX_SOFTIRQ); 11644 local_irq_enable(); 11645 11646 if (!use_backlog_threads()) { 11647 #ifdef CONFIG_RPS 11648 remsd = oldsd->rps_ipi_list; 11649 oldsd->rps_ipi_list = NULL; 11650 #endif 11651 /* send out pending IPI's on offline CPU */ 11652 net_rps_send_ipi(remsd); 11653 } 11654 11655 /* Process offline CPU's input_pkt_queue */ 11656 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 11657 netif_rx(skb); 11658 rps_input_queue_head_incr(oldsd); 11659 } 11660 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 11661 netif_rx(skb); 11662 rps_input_queue_head_incr(oldsd); 11663 } 11664 11665 return 0; 11666 } 11667 11668 /** 11669 * netdev_increment_features - increment feature set by one 11670 * @all: current feature set 11671 * @one: new feature set 11672 * @mask: mask feature set 11673 * 11674 * Computes a new feature set after adding a device with feature set 11675 * @one to the master device with current feature set @all. Will not 11676 * enable anything that is off in @mask. Returns the new feature set. 11677 */ 11678 netdev_features_t netdev_increment_features(netdev_features_t all, 11679 netdev_features_t one, netdev_features_t mask) 11680 { 11681 if (mask & NETIF_F_HW_CSUM) 11682 mask |= NETIF_F_CSUM_MASK; 11683 mask |= NETIF_F_VLAN_CHALLENGED; 11684 11685 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 11686 all &= one | ~NETIF_F_ALL_FOR_ALL; 11687 11688 /* If one device supports hw checksumming, set for all. */ 11689 if (all & NETIF_F_HW_CSUM) 11690 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 11691 11692 return all; 11693 } 11694 EXPORT_SYMBOL(netdev_increment_features); 11695 11696 static struct hlist_head * __net_init netdev_create_hash(void) 11697 { 11698 int i; 11699 struct hlist_head *hash; 11700 11701 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 11702 if (hash != NULL) 11703 for (i = 0; i < NETDEV_HASHENTRIES; i++) 11704 INIT_HLIST_HEAD(&hash[i]); 11705 11706 return hash; 11707 } 11708 11709 /* Initialize per network namespace state */ 11710 static int __net_init netdev_init(struct net *net) 11711 { 11712 BUILD_BUG_ON(GRO_HASH_BUCKETS > 11713 8 * sizeof_field(struct napi_struct, gro_bitmask)); 11714 11715 INIT_LIST_HEAD(&net->dev_base_head); 11716 11717 net->dev_name_head = netdev_create_hash(); 11718 if (net->dev_name_head == NULL) 11719 goto err_name; 11720 11721 net->dev_index_head = netdev_create_hash(); 11722 if (net->dev_index_head == NULL) 11723 goto err_idx; 11724 11725 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 11726 11727 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 11728 11729 return 0; 11730 11731 err_idx: 11732 kfree(net->dev_name_head); 11733 err_name: 11734 return -ENOMEM; 11735 } 11736 11737 /** 11738 * netdev_drivername - network driver for the device 11739 * @dev: network device 11740 * 11741 * Determine network driver for device. 11742 */ 11743 const char *netdev_drivername(const struct net_device *dev) 11744 { 11745 const struct device_driver *driver; 11746 const struct device *parent; 11747 const char *empty = ""; 11748 11749 parent = dev->dev.parent; 11750 if (!parent) 11751 return empty; 11752 11753 driver = parent->driver; 11754 if (driver && driver->name) 11755 return driver->name; 11756 return empty; 11757 } 11758 11759 static void __netdev_printk(const char *level, const struct net_device *dev, 11760 struct va_format *vaf) 11761 { 11762 if (dev && dev->dev.parent) { 11763 dev_printk_emit(level[1] - '0', 11764 dev->dev.parent, 11765 "%s %s %s%s: %pV", 11766 dev_driver_string(dev->dev.parent), 11767 dev_name(dev->dev.parent), 11768 netdev_name(dev), netdev_reg_state(dev), 11769 vaf); 11770 } else if (dev) { 11771 printk("%s%s%s: %pV", 11772 level, netdev_name(dev), netdev_reg_state(dev), vaf); 11773 } else { 11774 printk("%s(NULL net_device): %pV", level, vaf); 11775 } 11776 } 11777 11778 void netdev_printk(const char *level, const struct net_device *dev, 11779 const char *format, ...) 11780 { 11781 struct va_format vaf; 11782 va_list args; 11783 11784 va_start(args, format); 11785 11786 vaf.fmt = format; 11787 vaf.va = &args; 11788 11789 __netdev_printk(level, dev, &vaf); 11790 11791 va_end(args); 11792 } 11793 EXPORT_SYMBOL(netdev_printk); 11794 11795 #define define_netdev_printk_level(func, level) \ 11796 void func(const struct net_device *dev, const char *fmt, ...) \ 11797 { \ 11798 struct va_format vaf; \ 11799 va_list args; \ 11800 \ 11801 va_start(args, fmt); \ 11802 \ 11803 vaf.fmt = fmt; \ 11804 vaf.va = &args; \ 11805 \ 11806 __netdev_printk(level, dev, &vaf); \ 11807 \ 11808 va_end(args); \ 11809 } \ 11810 EXPORT_SYMBOL(func); 11811 11812 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 11813 define_netdev_printk_level(netdev_alert, KERN_ALERT); 11814 define_netdev_printk_level(netdev_crit, KERN_CRIT); 11815 define_netdev_printk_level(netdev_err, KERN_ERR); 11816 define_netdev_printk_level(netdev_warn, KERN_WARNING); 11817 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 11818 define_netdev_printk_level(netdev_info, KERN_INFO); 11819 11820 static void __net_exit netdev_exit(struct net *net) 11821 { 11822 kfree(net->dev_name_head); 11823 kfree(net->dev_index_head); 11824 xa_destroy(&net->dev_by_index); 11825 if (net != &init_net) 11826 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 11827 } 11828 11829 static struct pernet_operations __net_initdata netdev_net_ops = { 11830 .init = netdev_init, 11831 .exit = netdev_exit, 11832 }; 11833 11834 static void __net_exit default_device_exit_net(struct net *net) 11835 { 11836 struct netdev_name_node *name_node, *tmp; 11837 struct net_device *dev, *aux; 11838 /* 11839 * Push all migratable network devices back to the 11840 * initial network namespace 11841 */ 11842 ASSERT_RTNL(); 11843 for_each_netdev_safe(net, dev, aux) { 11844 int err; 11845 char fb_name[IFNAMSIZ]; 11846 11847 /* Ignore unmoveable devices (i.e. loopback) */ 11848 if (dev->features & NETIF_F_NETNS_LOCAL) 11849 continue; 11850 11851 /* Leave virtual devices for the generic cleanup */ 11852 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 11853 continue; 11854 11855 /* Push remaining network devices to init_net */ 11856 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 11857 if (netdev_name_in_use(&init_net, fb_name)) 11858 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 11859 11860 netdev_for_each_altname_safe(dev, name_node, tmp) 11861 if (netdev_name_in_use(&init_net, name_node->name)) 11862 __netdev_name_node_alt_destroy(name_node); 11863 11864 err = dev_change_net_namespace(dev, &init_net, fb_name); 11865 if (err) { 11866 pr_emerg("%s: failed to move %s to init_net: %d\n", 11867 __func__, dev->name, err); 11868 BUG(); 11869 } 11870 } 11871 } 11872 11873 static void __net_exit default_device_exit_batch(struct list_head *net_list) 11874 { 11875 /* At exit all network devices most be removed from a network 11876 * namespace. Do this in the reverse order of registration. 11877 * Do this across as many network namespaces as possible to 11878 * improve batching efficiency. 11879 */ 11880 struct net_device *dev; 11881 struct net *net; 11882 LIST_HEAD(dev_kill_list); 11883 11884 rtnl_lock(); 11885 list_for_each_entry(net, net_list, exit_list) { 11886 default_device_exit_net(net); 11887 cond_resched(); 11888 } 11889 11890 list_for_each_entry(net, net_list, exit_list) { 11891 for_each_netdev_reverse(net, dev) { 11892 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 11893 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 11894 else 11895 unregister_netdevice_queue(dev, &dev_kill_list); 11896 } 11897 } 11898 unregister_netdevice_many(&dev_kill_list); 11899 rtnl_unlock(); 11900 } 11901 11902 static struct pernet_operations __net_initdata default_device_ops = { 11903 .exit_batch = default_device_exit_batch, 11904 }; 11905 11906 static void __init net_dev_struct_check(void) 11907 { 11908 /* TX read-mostly hotpath */ 11909 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags); 11910 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 11911 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 11912 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 11913 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 11914 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 11915 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 11916 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 11917 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 11918 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 11919 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 11920 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 11921 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 11922 #ifdef CONFIG_XPS 11923 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 11924 #endif 11925 #ifdef CONFIG_NETFILTER_EGRESS 11926 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 11927 #endif 11928 #ifdef CONFIG_NET_XGRESS 11929 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 11930 #endif 11931 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 11932 11933 /* TXRX read-mostly hotpath */ 11934 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 11935 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 11936 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 11937 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 11938 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 11939 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 11940 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 11941 11942 /* RX read-mostly hotpath */ 11943 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 11944 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 11945 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 11946 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 11947 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_flush_timeout); 11948 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, napi_defer_hard_irqs); 11949 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 11950 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 11951 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 11952 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 11953 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 11954 #ifdef CONFIG_NETPOLL 11955 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 11956 #endif 11957 #ifdef CONFIG_NET_XGRESS 11958 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 11959 #endif 11960 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 104); 11961 } 11962 11963 /* 11964 * Initialize the DEV module. At boot time this walks the device list and 11965 * unhooks any devices that fail to initialise (normally hardware not 11966 * present) and leaves us with a valid list of present and active devices. 11967 * 11968 */ 11969 11970 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 11971 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 11972 11973 static int net_page_pool_create(int cpuid) 11974 { 11975 #if IS_ENABLED(CONFIG_PAGE_POOL) 11976 struct page_pool_params page_pool_params = { 11977 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 11978 .flags = PP_FLAG_SYSTEM_POOL, 11979 .nid = cpu_to_mem(cpuid), 11980 }; 11981 struct page_pool *pp_ptr; 11982 11983 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 11984 if (IS_ERR(pp_ptr)) 11985 return -ENOMEM; 11986 11987 per_cpu(system_page_pool, cpuid) = pp_ptr; 11988 #endif 11989 return 0; 11990 } 11991 11992 static int backlog_napi_should_run(unsigned int cpu) 11993 { 11994 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 11995 struct napi_struct *napi = &sd->backlog; 11996 11997 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 11998 } 11999 12000 static void run_backlog_napi(unsigned int cpu) 12001 { 12002 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12003 12004 napi_threaded_poll_loop(&sd->backlog); 12005 } 12006 12007 static void backlog_napi_setup(unsigned int cpu) 12008 { 12009 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12010 struct napi_struct *napi = &sd->backlog; 12011 12012 napi->thread = this_cpu_read(backlog_napi); 12013 set_bit(NAPI_STATE_THREADED, &napi->state); 12014 } 12015 12016 static struct smp_hotplug_thread backlog_threads = { 12017 .store = &backlog_napi, 12018 .thread_should_run = backlog_napi_should_run, 12019 .thread_fn = run_backlog_napi, 12020 .thread_comm = "backlog_napi/%u", 12021 .setup = backlog_napi_setup, 12022 }; 12023 12024 /* 12025 * This is called single threaded during boot, so no need 12026 * to take the rtnl semaphore. 12027 */ 12028 static int __init net_dev_init(void) 12029 { 12030 int i, rc = -ENOMEM; 12031 12032 BUG_ON(!dev_boot_phase); 12033 12034 net_dev_struct_check(); 12035 12036 if (dev_proc_init()) 12037 goto out; 12038 12039 if (netdev_kobject_init()) 12040 goto out; 12041 12042 for (i = 0; i < PTYPE_HASH_SIZE; i++) 12043 INIT_LIST_HEAD(&ptype_base[i]); 12044 12045 if (register_pernet_subsys(&netdev_net_ops)) 12046 goto out; 12047 12048 /* 12049 * Initialise the packet receive queues. 12050 */ 12051 12052 for_each_possible_cpu(i) { 12053 struct work_struct *flush = per_cpu_ptr(&flush_works, i); 12054 struct softnet_data *sd = &per_cpu(softnet_data, i); 12055 12056 INIT_WORK(flush, flush_backlog); 12057 12058 skb_queue_head_init(&sd->input_pkt_queue); 12059 skb_queue_head_init(&sd->process_queue); 12060 #ifdef CONFIG_XFRM_OFFLOAD 12061 skb_queue_head_init(&sd->xfrm_backlog); 12062 #endif 12063 INIT_LIST_HEAD(&sd->poll_list); 12064 sd->output_queue_tailp = &sd->output_queue; 12065 #ifdef CONFIG_RPS 12066 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 12067 sd->cpu = i; 12068 #endif 12069 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 12070 spin_lock_init(&sd->defer_lock); 12071 12072 init_gro_hash(&sd->backlog); 12073 sd->backlog.poll = process_backlog; 12074 sd->backlog.weight = weight_p; 12075 INIT_LIST_HEAD(&sd->backlog.poll_list); 12076 12077 if (net_page_pool_create(i)) 12078 goto out; 12079 } 12080 if (use_backlog_threads()) 12081 smpboot_register_percpu_thread(&backlog_threads); 12082 12083 dev_boot_phase = 0; 12084 12085 /* The loopback device is special if any other network devices 12086 * is present in a network namespace the loopback device must 12087 * be present. Since we now dynamically allocate and free the 12088 * loopback device ensure this invariant is maintained by 12089 * keeping the loopback device as the first device on the 12090 * list of network devices. Ensuring the loopback devices 12091 * is the first device that appears and the last network device 12092 * that disappears. 12093 */ 12094 if (register_pernet_device(&loopback_net_ops)) 12095 goto out; 12096 12097 if (register_pernet_device(&default_device_ops)) 12098 goto out; 12099 12100 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 12101 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 12102 12103 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 12104 NULL, dev_cpu_dead); 12105 WARN_ON(rc < 0); 12106 rc = 0; 12107 12108 /* avoid static key IPIs to isolated CPUs */ 12109 if (housekeeping_enabled(HK_TYPE_MISC)) 12110 net_enable_timestamp(); 12111 out: 12112 if (rc < 0) { 12113 for_each_possible_cpu(i) { 12114 struct page_pool *pp_ptr; 12115 12116 pp_ptr = per_cpu(system_page_pool, i); 12117 if (!pp_ptr) 12118 continue; 12119 12120 page_pool_destroy(pp_ptr); 12121 per_cpu(system_page_pool, i) = NULL; 12122 } 12123 } 12124 12125 return rc; 12126 } 12127 12128 subsys_initcall(net_dev_init); 12129