1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * NET3 Protocol independent device support routines. 4 * 5 * Derived from the non IP parts of dev.c 1.0.19 6 * Authors: Ross Biro 7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 8 * Mark Evans, <evansmp@uhura.aston.ac.uk> 9 * 10 * Additional Authors: 11 * Florian la Roche <rzsfl@rz.uni-sb.de> 12 * Alan Cox <gw4pts@gw4pts.ampr.org> 13 * David Hinds <dahinds@users.sourceforge.net> 14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru> 15 * Adam Sulmicki <adam@cfar.umd.edu> 16 * Pekka Riikonen <priikone@poesidon.pspt.fi> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/skbuff.h> 96 #include <linux/kthread.h> 97 #include <linux/bpf.h> 98 #include <linux/bpf_trace.h> 99 #include <net/net_namespace.h> 100 #include <net/sock.h> 101 #include <net/busy_poll.h> 102 #include <linux/rtnetlink.h> 103 #include <linux/stat.h> 104 #include <net/dsa.h> 105 #include <net/dst.h> 106 #include <net/dst_metadata.h> 107 #include <net/gro.h> 108 #include <net/pkt_sched.h> 109 #include <net/pkt_cls.h> 110 #include <net/checksum.h> 111 #include <net/xfrm.h> 112 #include <net/tcx.h> 113 #include <linux/highmem.h> 114 #include <linux/init.h> 115 #include <linux/module.h> 116 #include <linux/netpoll.h> 117 #include <linux/rcupdate.h> 118 #include <linux/delay.h> 119 #include <net/iw_handler.h> 120 #include <asm/current.h> 121 #include <linux/audit.h> 122 #include <linux/dmaengine.h> 123 #include <linux/err.h> 124 #include <linux/ctype.h> 125 #include <linux/if_arp.h> 126 #include <linux/if_vlan.h> 127 #include <linux/ip.h> 128 #include <net/ip.h> 129 #include <net/mpls.h> 130 #include <linux/ipv6.h> 131 #include <linux/in.h> 132 #include <linux/jhash.h> 133 #include <linux/random.h> 134 #include <trace/events/napi.h> 135 #include <trace/events/net.h> 136 #include <trace/events/skb.h> 137 #include <trace/events/qdisc.h> 138 #include <trace/events/xdp.h> 139 #include <linux/inetdevice.h> 140 #include <linux/cpu_rmap.h> 141 #include <linux/static_key.h> 142 #include <linux/hashtable.h> 143 #include <linux/vmalloc.h> 144 #include <linux/if_macvlan.h> 145 #include <linux/errqueue.h> 146 #include <linux/hrtimer.h> 147 #include <linux/netfilter_netdev.h> 148 #include <linux/crash_dump.h> 149 #include <linux/sctp.h> 150 #include <net/udp_tunnel.h> 151 #include <linux/net_namespace.h> 152 #include <linux/indirect_call_wrapper.h> 153 #include <net/devlink.h> 154 #include <linux/pm_runtime.h> 155 #include <linux/prandom.h> 156 #include <linux/once_lite.h> 157 #include <net/netdev_rx_queue.h> 158 #include <net/page_pool/types.h> 159 #include <net/page_pool/helpers.h> 160 #include <net/rps.h> 161 #include <linux/phy_link_topology.h> 162 163 #include "dev.h" 164 #include "devmem.h" 165 #include "net-sysfs.h" 166 167 static DEFINE_SPINLOCK(ptype_lock); 168 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 169 170 static int netif_rx_internal(struct sk_buff *skb); 171 static int call_netdevice_notifiers_extack(unsigned long val, 172 struct net_device *dev, 173 struct netlink_ext_ack *extack); 174 175 static DEFINE_MUTEX(ifalias_mutex); 176 177 /* protects napi_hash addition/deletion and napi_gen_id */ 178 static DEFINE_SPINLOCK(napi_hash_lock); 179 180 static unsigned int napi_gen_id = NR_CPUS; 181 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 182 183 static DECLARE_RWSEM(devnet_rename_sem); 184 185 static inline void dev_base_seq_inc(struct net *net) 186 { 187 unsigned int val = net->dev_base_seq + 1; 188 189 WRITE_ONCE(net->dev_base_seq, val ?: 1); 190 } 191 192 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 193 { 194 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 195 196 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 197 } 198 199 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 200 { 201 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 202 } 203 204 #ifndef CONFIG_PREEMPT_RT 205 206 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 207 208 static int __init setup_backlog_napi_threads(char *arg) 209 { 210 static_branch_enable(&use_backlog_threads_key); 211 return 0; 212 } 213 early_param("thread_backlog_napi", setup_backlog_napi_threads); 214 215 static bool use_backlog_threads(void) 216 { 217 return static_branch_unlikely(&use_backlog_threads_key); 218 } 219 220 #else 221 222 static bool use_backlog_threads(void) 223 { 224 return true; 225 } 226 227 #endif 228 229 static inline void backlog_lock_irq_save(struct softnet_data *sd, 230 unsigned long *flags) 231 { 232 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 233 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 234 else 235 local_irq_save(*flags); 236 } 237 238 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 239 { 240 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 241 spin_lock_irq(&sd->input_pkt_queue.lock); 242 else 243 local_irq_disable(); 244 } 245 246 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 247 unsigned long *flags) 248 { 249 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 250 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags); 251 else 252 local_irq_restore(*flags); 253 } 254 255 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 256 { 257 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 258 spin_unlock_irq(&sd->input_pkt_queue.lock); 259 else 260 local_irq_enable(); 261 } 262 263 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 264 const char *name) 265 { 266 struct netdev_name_node *name_node; 267 268 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 269 if (!name_node) 270 return NULL; 271 INIT_HLIST_NODE(&name_node->hlist); 272 name_node->dev = dev; 273 name_node->name = name; 274 return name_node; 275 } 276 277 static struct netdev_name_node * 278 netdev_name_node_head_alloc(struct net_device *dev) 279 { 280 struct netdev_name_node *name_node; 281 282 name_node = netdev_name_node_alloc(dev, dev->name); 283 if (!name_node) 284 return NULL; 285 INIT_LIST_HEAD(&name_node->list); 286 return name_node; 287 } 288 289 static void netdev_name_node_free(struct netdev_name_node *name_node) 290 { 291 kfree(name_node); 292 } 293 294 static void netdev_name_node_add(struct net *net, 295 struct netdev_name_node *name_node) 296 { 297 hlist_add_head_rcu(&name_node->hlist, 298 dev_name_hash(net, name_node->name)); 299 } 300 301 static void netdev_name_node_del(struct netdev_name_node *name_node) 302 { 303 hlist_del_rcu(&name_node->hlist); 304 } 305 306 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 307 const char *name) 308 { 309 struct hlist_head *head = dev_name_hash(net, name); 310 struct netdev_name_node *name_node; 311 312 hlist_for_each_entry(name_node, head, hlist) 313 if (!strcmp(name_node->name, name)) 314 return name_node; 315 return NULL; 316 } 317 318 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 319 const char *name) 320 { 321 struct hlist_head *head = dev_name_hash(net, name); 322 struct netdev_name_node *name_node; 323 324 hlist_for_each_entry_rcu(name_node, head, hlist) 325 if (!strcmp(name_node->name, name)) 326 return name_node; 327 return NULL; 328 } 329 330 bool netdev_name_in_use(struct net *net, const char *name) 331 { 332 return netdev_name_node_lookup(net, name); 333 } 334 EXPORT_SYMBOL(netdev_name_in_use); 335 336 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 337 { 338 struct netdev_name_node *name_node; 339 struct net *net = dev_net(dev); 340 341 name_node = netdev_name_node_lookup(net, name); 342 if (name_node) 343 return -EEXIST; 344 name_node = netdev_name_node_alloc(dev, name); 345 if (!name_node) 346 return -ENOMEM; 347 netdev_name_node_add(net, name_node); 348 /* The node that holds dev->name acts as a head of per-device list. */ 349 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 350 351 return 0; 352 } 353 354 static void netdev_name_node_alt_free(struct rcu_head *head) 355 { 356 struct netdev_name_node *name_node = 357 container_of(head, struct netdev_name_node, rcu); 358 359 kfree(name_node->name); 360 netdev_name_node_free(name_node); 361 } 362 363 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 364 { 365 netdev_name_node_del(name_node); 366 list_del(&name_node->list); 367 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 368 } 369 370 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 371 { 372 struct netdev_name_node *name_node; 373 struct net *net = dev_net(dev); 374 375 name_node = netdev_name_node_lookup(net, name); 376 if (!name_node) 377 return -ENOENT; 378 /* lookup might have found our primary name or a name belonging 379 * to another device. 380 */ 381 if (name_node == dev->name_node || name_node->dev != dev) 382 return -EINVAL; 383 384 __netdev_name_node_alt_destroy(name_node); 385 return 0; 386 } 387 388 static void netdev_name_node_alt_flush(struct net_device *dev) 389 { 390 struct netdev_name_node *name_node, *tmp; 391 392 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 393 list_del(&name_node->list); 394 netdev_name_node_alt_free(&name_node->rcu); 395 } 396 } 397 398 /* Device list insertion */ 399 static void list_netdevice(struct net_device *dev) 400 { 401 struct netdev_name_node *name_node; 402 struct net *net = dev_net(dev); 403 404 ASSERT_RTNL(); 405 406 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 407 netdev_name_node_add(net, dev->name_node); 408 hlist_add_head_rcu(&dev->index_hlist, 409 dev_index_hash(net, dev->ifindex)); 410 411 netdev_for_each_altname(dev, name_node) 412 netdev_name_node_add(net, name_node); 413 414 /* We reserved the ifindex, this can't fail */ 415 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 416 417 dev_base_seq_inc(net); 418 } 419 420 /* Device list removal 421 * caller must respect a RCU grace period before freeing/reusing dev 422 */ 423 static void unlist_netdevice(struct net_device *dev) 424 { 425 struct netdev_name_node *name_node; 426 struct net *net = dev_net(dev); 427 428 ASSERT_RTNL(); 429 430 xa_erase(&net->dev_by_index, dev->ifindex); 431 432 netdev_for_each_altname(dev, name_node) 433 netdev_name_node_del(name_node); 434 435 /* Unlink dev from the device chain */ 436 list_del_rcu(&dev->dev_list); 437 netdev_name_node_del(dev->name_node); 438 hlist_del_rcu(&dev->index_hlist); 439 440 dev_base_seq_inc(dev_net(dev)); 441 } 442 443 /* 444 * Our notifier list 445 */ 446 447 static RAW_NOTIFIER_HEAD(netdev_chain); 448 449 /* 450 * Device drivers call our routines to queue packets here. We empty the 451 * queue in the local softnet handler. 452 */ 453 454 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 455 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 456 }; 457 EXPORT_PER_CPU_SYMBOL(softnet_data); 458 459 /* Page_pool has a lockless array/stack to alloc/recycle pages. 460 * PP consumers must pay attention to run APIs in the appropriate context 461 * (e.g. NAPI context). 462 */ 463 static DEFINE_PER_CPU(struct page_pool *, system_page_pool); 464 465 #ifdef CONFIG_LOCKDEP 466 /* 467 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 468 * according to dev->type 469 */ 470 static const unsigned short netdev_lock_type[] = { 471 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 472 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 473 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 474 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 475 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 476 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 477 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 478 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 479 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 480 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 481 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 482 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 483 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 484 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 485 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE}; 486 487 static const char *const netdev_lock_name[] = { 488 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 489 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 490 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 491 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 492 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 493 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 494 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 495 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 496 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 497 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 498 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 499 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 500 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 501 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE", 502 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"}; 503 504 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 505 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 506 507 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 508 { 509 int i; 510 511 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 512 if (netdev_lock_type[i] == dev_type) 513 return i; 514 /* the last key is used by default */ 515 return ARRAY_SIZE(netdev_lock_type) - 1; 516 } 517 518 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 519 unsigned short dev_type) 520 { 521 int i; 522 523 i = netdev_lock_pos(dev_type); 524 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 525 netdev_lock_name[i]); 526 } 527 528 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 529 { 530 int i; 531 532 i = netdev_lock_pos(dev->type); 533 lockdep_set_class_and_name(&dev->addr_list_lock, 534 &netdev_addr_lock_key[i], 535 netdev_lock_name[i]); 536 } 537 #else 538 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 539 unsigned short dev_type) 540 { 541 } 542 543 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 544 { 545 } 546 #endif 547 548 /******************************************************************************* 549 * 550 * Protocol management and registration routines 551 * 552 *******************************************************************************/ 553 554 555 /* 556 * Add a protocol ID to the list. Now that the input handler is 557 * smarter we can dispense with all the messy stuff that used to be 558 * here. 559 * 560 * BEWARE!!! Protocol handlers, mangling input packets, 561 * MUST BE last in hash buckets and checking protocol handlers 562 * MUST start from promiscuous ptype_all chain in net_bh. 563 * It is true now, do not change it. 564 * Explanation follows: if protocol handler, mangling packet, will 565 * be the first on list, it is not able to sense, that packet 566 * is cloned and should be copied-on-write, so that it will 567 * change it and subsequent readers will get broken packet. 568 * --ANK (980803) 569 */ 570 571 static inline struct list_head *ptype_head(const struct packet_type *pt) 572 { 573 if (pt->type == htons(ETH_P_ALL)) 574 return pt->dev ? &pt->dev->ptype_all : &net_hotdata.ptype_all; 575 else 576 return pt->dev ? &pt->dev->ptype_specific : 577 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 578 } 579 580 /** 581 * dev_add_pack - add packet handler 582 * @pt: packet type declaration 583 * 584 * Add a protocol handler to the networking stack. The passed &packet_type 585 * is linked into kernel lists and may not be freed until it has been 586 * removed from the kernel lists. 587 * 588 * This call does not sleep therefore it can not 589 * guarantee all CPU's that are in middle of receiving packets 590 * will see the new packet type (until the next received packet). 591 */ 592 593 void dev_add_pack(struct packet_type *pt) 594 { 595 struct list_head *head = ptype_head(pt); 596 597 spin_lock(&ptype_lock); 598 list_add_rcu(&pt->list, head); 599 spin_unlock(&ptype_lock); 600 } 601 EXPORT_SYMBOL(dev_add_pack); 602 603 /** 604 * __dev_remove_pack - remove packet handler 605 * @pt: packet type declaration 606 * 607 * Remove a protocol handler that was previously added to the kernel 608 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 609 * from the kernel lists and can be freed or reused once this function 610 * returns. 611 * 612 * The packet type might still be in use by receivers 613 * and must not be freed until after all the CPU's have gone 614 * through a quiescent state. 615 */ 616 void __dev_remove_pack(struct packet_type *pt) 617 { 618 struct list_head *head = ptype_head(pt); 619 struct packet_type *pt1; 620 621 spin_lock(&ptype_lock); 622 623 list_for_each_entry(pt1, head, list) { 624 if (pt == pt1) { 625 list_del_rcu(&pt->list); 626 goto out; 627 } 628 } 629 630 pr_warn("dev_remove_pack: %p not found\n", pt); 631 out: 632 spin_unlock(&ptype_lock); 633 } 634 EXPORT_SYMBOL(__dev_remove_pack); 635 636 /** 637 * dev_remove_pack - remove packet handler 638 * @pt: packet type declaration 639 * 640 * Remove a protocol handler that was previously added to the kernel 641 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 642 * from the kernel lists and can be freed or reused once this function 643 * returns. 644 * 645 * This call sleeps to guarantee that no CPU is looking at the packet 646 * type after return. 647 */ 648 void dev_remove_pack(struct packet_type *pt) 649 { 650 __dev_remove_pack(pt); 651 652 synchronize_net(); 653 } 654 EXPORT_SYMBOL(dev_remove_pack); 655 656 657 /******************************************************************************* 658 * 659 * Device Interface Subroutines 660 * 661 *******************************************************************************/ 662 663 /** 664 * dev_get_iflink - get 'iflink' value of a interface 665 * @dev: targeted interface 666 * 667 * Indicates the ifindex the interface is linked to. 668 * Physical interfaces have the same 'ifindex' and 'iflink' values. 669 */ 670 671 int dev_get_iflink(const struct net_device *dev) 672 { 673 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 674 return dev->netdev_ops->ndo_get_iflink(dev); 675 676 return READ_ONCE(dev->ifindex); 677 } 678 EXPORT_SYMBOL(dev_get_iflink); 679 680 /** 681 * dev_fill_metadata_dst - Retrieve tunnel egress information. 682 * @dev: targeted interface 683 * @skb: The packet. 684 * 685 * For better visibility of tunnel traffic OVS needs to retrieve 686 * egress tunnel information for a packet. Following API allows 687 * user to get this info. 688 */ 689 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 690 { 691 struct ip_tunnel_info *info; 692 693 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 694 return -EINVAL; 695 696 info = skb_tunnel_info_unclone(skb); 697 if (!info) 698 return -ENOMEM; 699 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 700 return -EINVAL; 701 702 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 703 } 704 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 705 706 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 707 { 708 int k = stack->num_paths++; 709 710 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 711 return NULL; 712 713 return &stack->path[k]; 714 } 715 716 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 717 struct net_device_path_stack *stack) 718 { 719 const struct net_device *last_dev; 720 struct net_device_path_ctx ctx = { 721 .dev = dev, 722 }; 723 struct net_device_path *path; 724 int ret = 0; 725 726 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 727 stack->num_paths = 0; 728 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 729 last_dev = ctx.dev; 730 path = dev_fwd_path(stack); 731 if (!path) 732 return -1; 733 734 memset(path, 0, sizeof(struct net_device_path)); 735 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 736 if (ret < 0) 737 return -1; 738 739 if (WARN_ON_ONCE(last_dev == ctx.dev)) 740 return -1; 741 } 742 743 if (!ctx.dev) 744 return ret; 745 746 path = dev_fwd_path(stack); 747 if (!path) 748 return -1; 749 path->type = DEV_PATH_ETHERNET; 750 path->dev = ctx.dev; 751 752 return ret; 753 } 754 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 755 756 /** 757 * __dev_get_by_name - find a device by its name 758 * @net: the applicable net namespace 759 * @name: name to find 760 * 761 * Find an interface by name. Must be called under RTNL semaphore. 762 * If the name is found a pointer to the device is returned. 763 * If the name is not found then %NULL is returned. The 764 * reference counters are not incremented so the caller must be 765 * careful with locks. 766 */ 767 768 struct net_device *__dev_get_by_name(struct net *net, const char *name) 769 { 770 struct netdev_name_node *node_name; 771 772 node_name = netdev_name_node_lookup(net, name); 773 return node_name ? node_name->dev : NULL; 774 } 775 EXPORT_SYMBOL(__dev_get_by_name); 776 777 /** 778 * dev_get_by_name_rcu - find a device by its name 779 * @net: the applicable net namespace 780 * @name: name to find 781 * 782 * Find an interface by name. 783 * If the name is found a pointer to the device is returned. 784 * If the name is not found then %NULL is returned. 785 * The reference counters are not incremented so the caller must be 786 * careful with locks. The caller must hold RCU lock. 787 */ 788 789 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 790 { 791 struct netdev_name_node *node_name; 792 793 node_name = netdev_name_node_lookup_rcu(net, name); 794 return node_name ? node_name->dev : NULL; 795 } 796 EXPORT_SYMBOL(dev_get_by_name_rcu); 797 798 /* Deprecated for new users, call netdev_get_by_name() instead */ 799 struct net_device *dev_get_by_name(struct net *net, const char *name) 800 { 801 struct net_device *dev; 802 803 rcu_read_lock(); 804 dev = dev_get_by_name_rcu(net, name); 805 dev_hold(dev); 806 rcu_read_unlock(); 807 return dev; 808 } 809 EXPORT_SYMBOL(dev_get_by_name); 810 811 /** 812 * netdev_get_by_name() - find a device by its name 813 * @net: the applicable net namespace 814 * @name: name to find 815 * @tracker: tracking object for the acquired reference 816 * @gfp: allocation flags for the tracker 817 * 818 * Find an interface by name. This can be called from any 819 * context and does its own locking. The returned handle has 820 * the usage count incremented and the caller must use netdev_put() to 821 * release it when it is no longer needed. %NULL is returned if no 822 * matching device is found. 823 */ 824 struct net_device *netdev_get_by_name(struct net *net, const char *name, 825 netdevice_tracker *tracker, gfp_t gfp) 826 { 827 struct net_device *dev; 828 829 dev = dev_get_by_name(net, name); 830 if (dev) 831 netdev_tracker_alloc(dev, tracker, gfp); 832 return dev; 833 } 834 EXPORT_SYMBOL(netdev_get_by_name); 835 836 /** 837 * __dev_get_by_index - find a device by its ifindex 838 * @net: the applicable net namespace 839 * @ifindex: index of device 840 * 841 * Search for an interface by index. Returns %NULL if the device 842 * is not found or a pointer to the device. The device has not 843 * had its reference counter increased so the caller must be careful 844 * about locking. The caller must hold the RTNL semaphore. 845 */ 846 847 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 848 { 849 struct net_device *dev; 850 struct hlist_head *head = dev_index_hash(net, ifindex); 851 852 hlist_for_each_entry(dev, head, index_hlist) 853 if (dev->ifindex == ifindex) 854 return dev; 855 856 return NULL; 857 } 858 EXPORT_SYMBOL(__dev_get_by_index); 859 860 /** 861 * dev_get_by_index_rcu - find a device by its ifindex 862 * @net: the applicable net namespace 863 * @ifindex: index of device 864 * 865 * Search for an interface by index. Returns %NULL if the device 866 * is not found or a pointer to the device. The device has not 867 * had its reference counter increased so the caller must be careful 868 * about locking. The caller must hold RCU lock. 869 */ 870 871 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 872 { 873 struct net_device *dev; 874 struct hlist_head *head = dev_index_hash(net, ifindex); 875 876 hlist_for_each_entry_rcu(dev, head, index_hlist) 877 if (dev->ifindex == ifindex) 878 return dev; 879 880 return NULL; 881 } 882 EXPORT_SYMBOL(dev_get_by_index_rcu); 883 884 /* Deprecated for new users, call netdev_get_by_index() instead */ 885 struct net_device *dev_get_by_index(struct net *net, int ifindex) 886 { 887 struct net_device *dev; 888 889 rcu_read_lock(); 890 dev = dev_get_by_index_rcu(net, ifindex); 891 dev_hold(dev); 892 rcu_read_unlock(); 893 return dev; 894 } 895 EXPORT_SYMBOL(dev_get_by_index); 896 897 /** 898 * netdev_get_by_index() - find a device by its ifindex 899 * @net: the applicable net namespace 900 * @ifindex: index of device 901 * @tracker: tracking object for the acquired reference 902 * @gfp: allocation flags for the tracker 903 * 904 * Search for an interface by index. Returns NULL if the device 905 * is not found or a pointer to the device. The device returned has 906 * had a reference added and the pointer is safe until the user calls 907 * netdev_put() to indicate they have finished with it. 908 */ 909 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 910 netdevice_tracker *tracker, gfp_t gfp) 911 { 912 struct net_device *dev; 913 914 dev = dev_get_by_index(net, ifindex); 915 if (dev) 916 netdev_tracker_alloc(dev, tracker, gfp); 917 return dev; 918 } 919 EXPORT_SYMBOL(netdev_get_by_index); 920 921 /** 922 * dev_get_by_napi_id - find a device by napi_id 923 * @napi_id: ID of the NAPI struct 924 * 925 * Search for an interface by NAPI ID. Returns %NULL if the device 926 * is not found or a pointer to the device. The device has not had 927 * its reference counter increased so the caller must be careful 928 * about locking. The caller must hold RCU lock. 929 */ 930 931 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 932 { 933 struct napi_struct *napi; 934 935 WARN_ON_ONCE(!rcu_read_lock_held()); 936 937 if (napi_id < MIN_NAPI_ID) 938 return NULL; 939 940 napi = napi_by_id(napi_id); 941 942 return napi ? napi->dev : NULL; 943 } 944 EXPORT_SYMBOL(dev_get_by_napi_id); 945 946 static DEFINE_SEQLOCK(netdev_rename_lock); 947 948 void netdev_copy_name(struct net_device *dev, char *name) 949 { 950 unsigned int seq; 951 952 do { 953 seq = read_seqbegin(&netdev_rename_lock); 954 strscpy(name, dev->name, IFNAMSIZ); 955 } while (read_seqretry(&netdev_rename_lock, seq)); 956 } 957 958 /** 959 * netdev_get_name - get a netdevice name, knowing its ifindex. 960 * @net: network namespace 961 * @name: a pointer to the buffer where the name will be stored. 962 * @ifindex: the ifindex of the interface to get the name from. 963 */ 964 int netdev_get_name(struct net *net, char *name, int ifindex) 965 { 966 struct net_device *dev; 967 int ret; 968 969 rcu_read_lock(); 970 971 dev = dev_get_by_index_rcu(net, ifindex); 972 if (!dev) { 973 ret = -ENODEV; 974 goto out; 975 } 976 977 netdev_copy_name(dev, name); 978 979 ret = 0; 980 out: 981 rcu_read_unlock(); 982 return ret; 983 } 984 985 /** 986 * dev_getbyhwaddr_rcu - find a device by its hardware address 987 * @net: the applicable net namespace 988 * @type: media type of device 989 * @ha: hardware address 990 * 991 * Search for an interface by MAC address. Returns NULL if the device 992 * is not found or a pointer to the device. 993 * The caller must hold RCU or RTNL. 994 * The returned device has not had its ref count increased 995 * and the caller must therefore be careful about locking 996 * 997 */ 998 999 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1000 const char *ha) 1001 { 1002 struct net_device *dev; 1003 1004 for_each_netdev_rcu(net, dev) 1005 if (dev->type == type && 1006 !memcmp(dev->dev_addr, ha, dev->addr_len)) 1007 return dev; 1008 1009 return NULL; 1010 } 1011 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1012 1013 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1014 { 1015 struct net_device *dev, *ret = NULL; 1016 1017 rcu_read_lock(); 1018 for_each_netdev_rcu(net, dev) 1019 if (dev->type == type) { 1020 dev_hold(dev); 1021 ret = dev; 1022 break; 1023 } 1024 rcu_read_unlock(); 1025 return ret; 1026 } 1027 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1028 1029 /** 1030 * __dev_get_by_flags - find any device with given flags 1031 * @net: the applicable net namespace 1032 * @if_flags: IFF_* values 1033 * @mask: bitmask of bits in if_flags to check 1034 * 1035 * Search for any interface with the given flags. Returns NULL if a device 1036 * is not found or a pointer to the device. Must be called inside 1037 * rtnl_lock(), and result refcount is unchanged. 1038 */ 1039 1040 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags, 1041 unsigned short mask) 1042 { 1043 struct net_device *dev, *ret; 1044 1045 ASSERT_RTNL(); 1046 1047 ret = NULL; 1048 for_each_netdev(net, dev) { 1049 if (((dev->flags ^ if_flags) & mask) == 0) { 1050 ret = dev; 1051 break; 1052 } 1053 } 1054 return ret; 1055 } 1056 EXPORT_SYMBOL(__dev_get_by_flags); 1057 1058 /** 1059 * dev_valid_name - check if name is okay for network device 1060 * @name: name string 1061 * 1062 * Network device names need to be valid file names to 1063 * allow sysfs to work. We also disallow any kind of 1064 * whitespace. 1065 */ 1066 bool dev_valid_name(const char *name) 1067 { 1068 if (*name == '\0') 1069 return false; 1070 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1071 return false; 1072 if (!strcmp(name, ".") || !strcmp(name, "..")) 1073 return false; 1074 1075 while (*name) { 1076 if (*name == '/' || *name == ':' || isspace(*name)) 1077 return false; 1078 name++; 1079 } 1080 return true; 1081 } 1082 EXPORT_SYMBOL(dev_valid_name); 1083 1084 /** 1085 * __dev_alloc_name - allocate a name for a device 1086 * @net: network namespace to allocate the device name in 1087 * @name: name format string 1088 * @res: result name string 1089 * 1090 * Passed a format string - eg "lt%d" it will try and find a suitable 1091 * id. It scans list of devices to build up a free map, then chooses 1092 * the first empty slot. The caller must hold the dev_base or rtnl lock 1093 * while allocating the name and adding the device in order to avoid 1094 * duplicates. 1095 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1096 * Returns the number of the unit assigned or a negative errno code. 1097 */ 1098 1099 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1100 { 1101 int i = 0; 1102 const char *p; 1103 const int max_netdevices = 8*PAGE_SIZE; 1104 unsigned long *inuse; 1105 struct net_device *d; 1106 char buf[IFNAMSIZ]; 1107 1108 /* Verify the string as this thing may have come from the user. 1109 * There must be one "%d" and no other "%" characters. 1110 */ 1111 p = strchr(name, '%'); 1112 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1113 return -EINVAL; 1114 1115 /* Use one page as a bit array of possible slots */ 1116 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1117 if (!inuse) 1118 return -ENOMEM; 1119 1120 for_each_netdev(net, d) { 1121 struct netdev_name_node *name_node; 1122 1123 netdev_for_each_altname(d, name_node) { 1124 if (!sscanf(name_node->name, name, &i)) 1125 continue; 1126 if (i < 0 || i >= max_netdevices) 1127 continue; 1128 1129 /* avoid cases where sscanf is not exact inverse of printf */ 1130 snprintf(buf, IFNAMSIZ, name, i); 1131 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1132 __set_bit(i, inuse); 1133 } 1134 if (!sscanf(d->name, name, &i)) 1135 continue; 1136 if (i < 0 || i >= max_netdevices) 1137 continue; 1138 1139 /* avoid cases where sscanf is not exact inverse of printf */ 1140 snprintf(buf, IFNAMSIZ, name, i); 1141 if (!strncmp(buf, d->name, IFNAMSIZ)) 1142 __set_bit(i, inuse); 1143 } 1144 1145 i = find_first_zero_bit(inuse, max_netdevices); 1146 bitmap_free(inuse); 1147 if (i == max_netdevices) 1148 return -ENFILE; 1149 1150 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1151 strscpy(buf, name, IFNAMSIZ); 1152 snprintf(res, IFNAMSIZ, buf, i); 1153 return i; 1154 } 1155 1156 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1157 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1158 const char *want_name, char *out_name, 1159 int dup_errno) 1160 { 1161 if (!dev_valid_name(want_name)) 1162 return -EINVAL; 1163 1164 if (strchr(want_name, '%')) 1165 return __dev_alloc_name(net, want_name, out_name); 1166 1167 if (netdev_name_in_use(net, want_name)) 1168 return -dup_errno; 1169 if (out_name != want_name) 1170 strscpy(out_name, want_name, IFNAMSIZ); 1171 return 0; 1172 } 1173 1174 /** 1175 * dev_alloc_name - allocate a name for a device 1176 * @dev: device 1177 * @name: name format string 1178 * 1179 * Passed a format string - eg "lt%d" it will try and find a suitable 1180 * id. It scans list of devices to build up a free map, then chooses 1181 * the first empty slot. The caller must hold the dev_base or rtnl lock 1182 * while allocating the name and adding the device in order to avoid 1183 * duplicates. 1184 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1185 * Returns the number of the unit assigned or a negative errno code. 1186 */ 1187 1188 int dev_alloc_name(struct net_device *dev, const char *name) 1189 { 1190 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1191 } 1192 EXPORT_SYMBOL(dev_alloc_name); 1193 1194 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1195 const char *name) 1196 { 1197 int ret; 1198 1199 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1200 return ret < 0 ? ret : 0; 1201 } 1202 1203 /** 1204 * dev_change_name - change name of a device 1205 * @dev: device 1206 * @newname: name (or format string) must be at least IFNAMSIZ 1207 * 1208 * Change name of a device, can pass format strings "eth%d". 1209 * for wildcarding. 1210 */ 1211 int dev_change_name(struct net_device *dev, const char *newname) 1212 { 1213 unsigned char old_assign_type; 1214 char oldname[IFNAMSIZ]; 1215 int err = 0; 1216 int ret; 1217 struct net *net; 1218 1219 ASSERT_RTNL(); 1220 BUG_ON(!dev_net(dev)); 1221 1222 net = dev_net(dev); 1223 1224 down_write(&devnet_rename_sem); 1225 1226 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) { 1227 up_write(&devnet_rename_sem); 1228 return 0; 1229 } 1230 1231 memcpy(oldname, dev->name, IFNAMSIZ); 1232 1233 write_seqlock_bh(&netdev_rename_lock); 1234 err = dev_get_valid_name(net, dev, newname); 1235 write_sequnlock_bh(&netdev_rename_lock); 1236 1237 if (err < 0) { 1238 up_write(&devnet_rename_sem); 1239 return err; 1240 } 1241 1242 if (oldname[0] && !strchr(oldname, '%')) 1243 netdev_info(dev, "renamed from %s%s\n", oldname, 1244 dev->flags & IFF_UP ? " (while UP)" : ""); 1245 1246 old_assign_type = dev->name_assign_type; 1247 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1248 1249 rollback: 1250 ret = device_rename(&dev->dev, dev->name); 1251 if (ret) { 1252 memcpy(dev->name, oldname, IFNAMSIZ); 1253 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1254 up_write(&devnet_rename_sem); 1255 return ret; 1256 } 1257 1258 up_write(&devnet_rename_sem); 1259 1260 netdev_adjacent_rename_links(dev, oldname); 1261 1262 netdev_name_node_del(dev->name_node); 1263 1264 synchronize_net(); 1265 1266 netdev_name_node_add(net, dev->name_node); 1267 1268 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1269 ret = notifier_to_errno(ret); 1270 1271 if (ret) { 1272 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1273 if (err >= 0) { 1274 err = ret; 1275 down_write(&devnet_rename_sem); 1276 write_seqlock_bh(&netdev_rename_lock); 1277 memcpy(dev->name, oldname, IFNAMSIZ); 1278 write_sequnlock_bh(&netdev_rename_lock); 1279 memcpy(oldname, newname, IFNAMSIZ); 1280 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1281 old_assign_type = NET_NAME_RENAMED; 1282 goto rollback; 1283 } else { 1284 netdev_err(dev, "name change rollback failed: %d\n", 1285 ret); 1286 } 1287 } 1288 1289 return err; 1290 } 1291 1292 /** 1293 * dev_set_alias - change ifalias of a device 1294 * @dev: device 1295 * @alias: name up to IFALIASZ 1296 * @len: limit of bytes to copy from info 1297 * 1298 * Set ifalias for a device, 1299 */ 1300 int dev_set_alias(struct net_device *dev, const char *alias, size_t len) 1301 { 1302 struct dev_ifalias *new_alias = NULL; 1303 1304 if (len >= IFALIASZ) 1305 return -EINVAL; 1306 1307 if (len) { 1308 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1309 if (!new_alias) 1310 return -ENOMEM; 1311 1312 memcpy(new_alias->ifalias, alias, len); 1313 new_alias->ifalias[len] = 0; 1314 } 1315 1316 mutex_lock(&ifalias_mutex); 1317 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1318 mutex_is_locked(&ifalias_mutex)); 1319 mutex_unlock(&ifalias_mutex); 1320 1321 if (new_alias) 1322 kfree_rcu(new_alias, rcuhead); 1323 1324 return len; 1325 } 1326 EXPORT_SYMBOL(dev_set_alias); 1327 1328 /** 1329 * dev_get_alias - get ifalias of a device 1330 * @dev: device 1331 * @name: buffer to store name of ifalias 1332 * @len: size of buffer 1333 * 1334 * get ifalias for a device. Caller must make sure dev cannot go 1335 * away, e.g. rcu read lock or own a reference count to device. 1336 */ 1337 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1338 { 1339 const struct dev_ifalias *alias; 1340 int ret = 0; 1341 1342 rcu_read_lock(); 1343 alias = rcu_dereference(dev->ifalias); 1344 if (alias) 1345 ret = snprintf(name, len, "%s", alias->ifalias); 1346 rcu_read_unlock(); 1347 1348 return ret; 1349 } 1350 1351 /** 1352 * netdev_features_change - device changes features 1353 * @dev: device to cause notification 1354 * 1355 * Called to indicate a device has changed features. 1356 */ 1357 void netdev_features_change(struct net_device *dev) 1358 { 1359 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1360 } 1361 EXPORT_SYMBOL(netdev_features_change); 1362 1363 /** 1364 * netdev_state_change - device changes state 1365 * @dev: device to cause notification 1366 * 1367 * Called to indicate a device has changed state. This function calls 1368 * the notifier chains for netdev_chain and sends a NEWLINK message 1369 * to the routing socket. 1370 */ 1371 void netdev_state_change(struct net_device *dev) 1372 { 1373 if (dev->flags & IFF_UP) { 1374 struct netdev_notifier_change_info change_info = { 1375 .info.dev = dev, 1376 }; 1377 1378 call_netdevice_notifiers_info(NETDEV_CHANGE, 1379 &change_info.info); 1380 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1381 } 1382 } 1383 EXPORT_SYMBOL(netdev_state_change); 1384 1385 /** 1386 * __netdev_notify_peers - notify network peers about existence of @dev, 1387 * to be called when rtnl lock is already held. 1388 * @dev: network device 1389 * 1390 * Generate traffic such that interested network peers are aware of 1391 * @dev, such as by generating a gratuitous ARP. This may be used when 1392 * a device wants to inform the rest of the network about some sort of 1393 * reconfiguration such as a failover event or virtual machine 1394 * migration. 1395 */ 1396 void __netdev_notify_peers(struct net_device *dev) 1397 { 1398 ASSERT_RTNL(); 1399 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1400 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1401 } 1402 EXPORT_SYMBOL(__netdev_notify_peers); 1403 1404 /** 1405 * netdev_notify_peers - notify network peers about existence of @dev 1406 * @dev: network device 1407 * 1408 * Generate traffic such that interested network peers are aware of 1409 * @dev, such as by generating a gratuitous ARP. This may be used when 1410 * a device wants to inform the rest of the network about some sort of 1411 * reconfiguration such as a failover event or virtual machine 1412 * migration. 1413 */ 1414 void netdev_notify_peers(struct net_device *dev) 1415 { 1416 rtnl_lock(); 1417 __netdev_notify_peers(dev); 1418 rtnl_unlock(); 1419 } 1420 EXPORT_SYMBOL(netdev_notify_peers); 1421 1422 static int napi_threaded_poll(void *data); 1423 1424 static int napi_kthread_create(struct napi_struct *n) 1425 { 1426 int err = 0; 1427 1428 /* Create and wake up the kthread once to put it in 1429 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1430 * warning and work with loadavg. 1431 */ 1432 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1433 n->dev->name, n->napi_id); 1434 if (IS_ERR(n->thread)) { 1435 err = PTR_ERR(n->thread); 1436 pr_err("kthread_run failed with err %d\n", err); 1437 n->thread = NULL; 1438 } 1439 1440 return err; 1441 } 1442 1443 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1444 { 1445 const struct net_device_ops *ops = dev->netdev_ops; 1446 int ret; 1447 1448 ASSERT_RTNL(); 1449 dev_addr_check(dev); 1450 1451 if (!netif_device_present(dev)) { 1452 /* may be detached because parent is runtime-suspended */ 1453 if (dev->dev.parent) 1454 pm_runtime_resume(dev->dev.parent); 1455 if (!netif_device_present(dev)) 1456 return -ENODEV; 1457 } 1458 1459 /* Block netpoll from trying to do any rx path servicing. 1460 * If we don't do this there is a chance ndo_poll_controller 1461 * or ndo_poll may be running while we open the device 1462 */ 1463 netpoll_poll_disable(dev); 1464 1465 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1466 ret = notifier_to_errno(ret); 1467 if (ret) 1468 return ret; 1469 1470 set_bit(__LINK_STATE_START, &dev->state); 1471 1472 if (ops->ndo_validate_addr) 1473 ret = ops->ndo_validate_addr(dev); 1474 1475 if (!ret && ops->ndo_open) 1476 ret = ops->ndo_open(dev); 1477 1478 netpoll_poll_enable(dev); 1479 1480 if (ret) 1481 clear_bit(__LINK_STATE_START, &dev->state); 1482 else { 1483 dev->flags |= IFF_UP; 1484 dev_set_rx_mode(dev); 1485 dev_activate(dev); 1486 add_device_randomness(dev->dev_addr, dev->addr_len); 1487 } 1488 1489 return ret; 1490 } 1491 1492 /** 1493 * dev_open - prepare an interface for use. 1494 * @dev: device to open 1495 * @extack: netlink extended ack 1496 * 1497 * Takes a device from down to up state. The device's private open 1498 * function is invoked and then the multicast lists are loaded. Finally 1499 * the device is moved into the up state and a %NETDEV_UP message is 1500 * sent to the netdev notifier chain. 1501 * 1502 * Calling this function on an active interface is a nop. On a failure 1503 * a negative errno code is returned. 1504 */ 1505 int dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1506 { 1507 int ret; 1508 1509 if (dev->flags & IFF_UP) 1510 return 0; 1511 1512 ret = __dev_open(dev, extack); 1513 if (ret < 0) 1514 return ret; 1515 1516 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1517 call_netdevice_notifiers(NETDEV_UP, dev); 1518 1519 return ret; 1520 } 1521 EXPORT_SYMBOL(dev_open); 1522 1523 static void __dev_close_many(struct list_head *head) 1524 { 1525 struct net_device *dev; 1526 1527 ASSERT_RTNL(); 1528 might_sleep(); 1529 1530 list_for_each_entry(dev, head, close_list) { 1531 /* Temporarily disable netpoll until the interface is down */ 1532 netpoll_poll_disable(dev); 1533 1534 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1535 1536 clear_bit(__LINK_STATE_START, &dev->state); 1537 1538 /* Synchronize to scheduled poll. We cannot touch poll list, it 1539 * can be even on different cpu. So just clear netif_running(). 1540 * 1541 * dev->stop() will invoke napi_disable() on all of it's 1542 * napi_struct instances on this device. 1543 */ 1544 smp_mb__after_atomic(); /* Commit netif_running(). */ 1545 } 1546 1547 dev_deactivate_many(head); 1548 1549 list_for_each_entry(dev, head, close_list) { 1550 const struct net_device_ops *ops = dev->netdev_ops; 1551 1552 /* 1553 * Call the device specific close. This cannot fail. 1554 * Only if device is UP 1555 * 1556 * We allow it to be called even after a DETACH hot-plug 1557 * event. 1558 */ 1559 if (ops->ndo_stop) 1560 ops->ndo_stop(dev); 1561 1562 dev->flags &= ~IFF_UP; 1563 netpoll_poll_enable(dev); 1564 } 1565 } 1566 1567 static void __dev_close(struct net_device *dev) 1568 { 1569 LIST_HEAD(single); 1570 1571 list_add(&dev->close_list, &single); 1572 __dev_close_many(&single); 1573 list_del(&single); 1574 } 1575 1576 void dev_close_many(struct list_head *head, bool unlink) 1577 { 1578 struct net_device *dev, *tmp; 1579 1580 /* Remove the devices that don't need to be closed */ 1581 list_for_each_entry_safe(dev, tmp, head, close_list) 1582 if (!(dev->flags & IFF_UP)) 1583 list_del_init(&dev->close_list); 1584 1585 __dev_close_many(head); 1586 1587 list_for_each_entry_safe(dev, tmp, head, close_list) { 1588 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1589 call_netdevice_notifiers(NETDEV_DOWN, dev); 1590 if (unlink) 1591 list_del_init(&dev->close_list); 1592 } 1593 } 1594 EXPORT_SYMBOL(dev_close_many); 1595 1596 /** 1597 * dev_close - shutdown an interface. 1598 * @dev: device to shutdown 1599 * 1600 * This function moves an active device into down state. A 1601 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device 1602 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier 1603 * chain. 1604 */ 1605 void dev_close(struct net_device *dev) 1606 { 1607 if (dev->flags & IFF_UP) { 1608 LIST_HEAD(single); 1609 1610 list_add(&dev->close_list, &single); 1611 dev_close_many(&single, true); 1612 list_del(&single); 1613 } 1614 } 1615 EXPORT_SYMBOL(dev_close); 1616 1617 1618 /** 1619 * dev_disable_lro - disable Large Receive Offload on a device 1620 * @dev: device 1621 * 1622 * Disable Large Receive Offload (LRO) on a net device. Must be 1623 * called under RTNL. This is needed if received packets may be 1624 * forwarded to another interface. 1625 */ 1626 void dev_disable_lro(struct net_device *dev) 1627 { 1628 struct net_device *lower_dev; 1629 struct list_head *iter; 1630 1631 dev->wanted_features &= ~NETIF_F_LRO; 1632 netdev_update_features(dev); 1633 1634 if (unlikely(dev->features & NETIF_F_LRO)) 1635 netdev_WARN(dev, "failed to disable LRO!\n"); 1636 1637 netdev_for_each_lower_dev(dev, lower_dev, iter) 1638 dev_disable_lro(lower_dev); 1639 } 1640 EXPORT_SYMBOL(dev_disable_lro); 1641 1642 /** 1643 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1644 * @dev: device 1645 * 1646 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1647 * called under RTNL. This is needed if Generic XDP is installed on 1648 * the device. 1649 */ 1650 static void dev_disable_gro_hw(struct net_device *dev) 1651 { 1652 dev->wanted_features &= ~NETIF_F_GRO_HW; 1653 netdev_update_features(dev); 1654 1655 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1656 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1657 } 1658 1659 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1660 { 1661 #define N(val) \ 1662 case NETDEV_##val: \ 1663 return "NETDEV_" __stringify(val); 1664 switch (cmd) { 1665 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1666 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1667 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1668 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1669 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1670 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1671 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1672 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1673 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1674 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1675 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1676 N(XDP_FEAT_CHANGE) 1677 } 1678 #undef N 1679 return "UNKNOWN_NETDEV_EVENT"; 1680 } 1681 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1682 1683 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1684 struct net_device *dev) 1685 { 1686 struct netdev_notifier_info info = { 1687 .dev = dev, 1688 }; 1689 1690 return nb->notifier_call(nb, val, &info); 1691 } 1692 1693 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1694 struct net_device *dev) 1695 { 1696 int err; 1697 1698 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1699 err = notifier_to_errno(err); 1700 if (err) 1701 return err; 1702 1703 if (!(dev->flags & IFF_UP)) 1704 return 0; 1705 1706 call_netdevice_notifier(nb, NETDEV_UP, dev); 1707 return 0; 1708 } 1709 1710 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1711 struct net_device *dev) 1712 { 1713 if (dev->flags & IFF_UP) { 1714 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1715 dev); 1716 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1717 } 1718 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1719 } 1720 1721 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1722 struct net *net) 1723 { 1724 struct net_device *dev; 1725 int err; 1726 1727 for_each_netdev(net, dev) { 1728 err = call_netdevice_register_notifiers(nb, dev); 1729 if (err) 1730 goto rollback; 1731 } 1732 return 0; 1733 1734 rollback: 1735 for_each_netdev_continue_reverse(net, dev) 1736 call_netdevice_unregister_notifiers(nb, dev); 1737 return err; 1738 } 1739 1740 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1741 struct net *net) 1742 { 1743 struct net_device *dev; 1744 1745 for_each_netdev(net, dev) 1746 call_netdevice_unregister_notifiers(nb, dev); 1747 } 1748 1749 static int dev_boot_phase = 1; 1750 1751 /** 1752 * register_netdevice_notifier - register a network notifier block 1753 * @nb: notifier 1754 * 1755 * Register a notifier to be called when network device events occur. 1756 * The notifier passed is linked into the kernel structures and must 1757 * not be reused until it has been unregistered. A negative errno code 1758 * is returned on a failure. 1759 * 1760 * When registered all registration and up events are replayed 1761 * to the new notifier to allow device to have a race free 1762 * view of the network device list. 1763 */ 1764 1765 int register_netdevice_notifier(struct notifier_block *nb) 1766 { 1767 struct net *net; 1768 int err; 1769 1770 /* Close race with setup_net() and cleanup_net() */ 1771 down_write(&pernet_ops_rwsem); 1772 rtnl_lock(); 1773 err = raw_notifier_chain_register(&netdev_chain, nb); 1774 if (err) 1775 goto unlock; 1776 if (dev_boot_phase) 1777 goto unlock; 1778 for_each_net(net) { 1779 err = call_netdevice_register_net_notifiers(nb, net); 1780 if (err) 1781 goto rollback; 1782 } 1783 1784 unlock: 1785 rtnl_unlock(); 1786 up_write(&pernet_ops_rwsem); 1787 return err; 1788 1789 rollback: 1790 for_each_net_continue_reverse(net) 1791 call_netdevice_unregister_net_notifiers(nb, net); 1792 1793 raw_notifier_chain_unregister(&netdev_chain, nb); 1794 goto unlock; 1795 } 1796 EXPORT_SYMBOL(register_netdevice_notifier); 1797 1798 /** 1799 * unregister_netdevice_notifier - unregister a network notifier block 1800 * @nb: notifier 1801 * 1802 * Unregister a notifier previously registered by 1803 * register_netdevice_notifier(). The notifier is unlinked into the 1804 * kernel structures and may then be reused. A negative errno code 1805 * is returned on a failure. 1806 * 1807 * After unregistering unregister and down device events are synthesized 1808 * for all devices on the device list to the removed notifier to remove 1809 * the need for special case cleanup code. 1810 */ 1811 1812 int unregister_netdevice_notifier(struct notifier_block *nb) 1813 { 1814 struct net *net; 1815 int err; 1816 1817 /* Close race with setup_net() and cleanup_net() */ 1818 down_write(&pernet_ops_rwsem); 1819 rtnl_lock(); 1820 err = raw_notifier_chain_unregister(&netdev_chain, nb); 1821 if (err) 1822 goto unlock; 1823 1824 for_each_net(net) 1825 call_netdevice_unregister_net_notifiers(nb, net); 1826 1827 unlock: 1828 rtnl_unlock(); 1829 up_write(&pernet_ops_rwsem); 1830 return err; 1831 } 1832 EXPORT_SYMBOL(unregister_netdevice_notifier); 1833 1834 static int __register_netdevice_notifier_net(struct net *net, 1835 struct notifier_block *nb, 1836 bool ignore_call_fail) 1837 { 1838 int err; 1839 1840 err = raw_notifier_chain_register(&net->netdev_chain, nb); 1841 if (err) 1842 return err; 1843 if (dev_boot_phase) 1844 return 0; 1845 1846 err = call_netdevice_register_net_notifiers(nb, net); 1847 if (err && !ignore_call_fail) 1848 goto chain_unregister; 1849 1850 return 0; 1851 1852 chain_unregister: 1853 raw_notifier_chain_unregister(&net->netdev_chain, nb); 1854 return err; 1855 } 1856 1857 static int __unregister_netdevice_notifier_net(struct net *net, 1858 struct notifier_block *nb) 1859 { 1860 int err; 1861 1862 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 1863 if (err) 1864 return err; 1865 1866 call_netdevice_unregister_net_notifiers(nb, net); 1867 return 0; 1868 } 1869 1870 /** 1871 * register_netdevice_notifier_net - register a per-netns network notifier block 1872 * @net: network namespace 1873 * @nb: notifier 1874 * 1875 * Register a notifier to be called when network device events occur. 1876 * The notifier passed is linked into the kernel structures and must 1877 * not be reused until it has been unregistered. A negative errno code 1878 * is returned on a failure. 1879 * 1880 * When registered all registration and up events are replayed 1881 * to the new notifier to allow device to have a race free 1882 * view of the network device list. 1883 */ 1884 1885 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 1886 { 1887 int err; 1888 1889 rtnl_lock(); 1890 err = __register_netdevice_notifier_net(net, nb, false); 1891 rtnl_unlock(); 1892 return err; 1893 } 1894 EXPORT_SYMBOL(register_netdevice_notifier_net); 1895 1896 /** 1897 * unregister_netdevice_notifier_net - unregister a per-netns 1898 * network notifier block 1899 * @net: network namespace 1900 * @nb: notifier 1901 * 1902 * Unregister a notifier previously registered by 1903 * register_netdevice_notifier_net(). The notifier is unlinked from the 1904 * kernel structures and may then be reused. A negative errno code 1905 * is returned on a failure. 1906 * 1907 * After unregistering unregister and down device events are synthesized 1908 * for all devices on the device list to the removed notifier to remove 1909 * the need for special case cleanup code. 1910 */ 1911 1912 int unregister_netdevice_notifier_net(struct net *net, 1913 struct notifier_block *nb) 1914 { 1915 int err; 1916 1917 rtnl_lock(); 1918 err = __unregister_netdevice_notifier_net(net, nb); 1919 rtnl_unlock(); 1920 return err; 1921 } 1922 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 1923 1924 static void __move_netdevice_notifier_net(struct net *src_net, 1925 struct net *dst_net, 1926 struct notifier_block *nb) 1927 { 1928 __unregister_netdevice_notifier_net(src_net, nb); 1929 __register_netdevice_notifier_net(dst_net, nb, true); 1930 } 1931 1932 int register_netdevice_notifier_dev_net(struct net_device *dev, 1933 struct notifier_block *nb, 1934 struct netdev_net_notifier *nn) 1935 { 1936 int err; 1937 1938 rtnl_lock(); 1939 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 1940 if (!err) { 1941 nn->nb = nb; 1942 list_add(&nn->list, &dev->net_notifier_list); 1943 } 1944 rtnl_unlock(); 1945 return err; 1946 } 1947 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 1948 1949 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 1950 struct notifier_block *nb, 1951 struct netdev_net_notifier *nn) 1952 { 1953 int err; 1954 1955 rtnl_lock(); 1956 list_del(&nn->list); 1957 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 1958 rtnl_unlock(); 1959 return err; 1960 } 1961 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 1962 1963 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 1964 struct net *net) 1965 { 1966 struct netdev_net_notifier *nn; 1967 1968 list_for_each_entry(nn, &dev->net_notifier_list, list) 1969 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 1970 } 1971 1972 /** 1973 * call_netdevice_notifiers_info - call all network notifier blocks 1974 * @val: value passed unmodified to notifier function 1975 * @info: notifier information data 1976 * 1977 * Call all network notifier blocks. Parameters and return value 1978 * are as for raw_notifier_call_chain(). 1979 */ 1980 1981 int call_netdevice_notifiers_info(unsigned long val, 1982 struct netdev_notifier_info *info) 1983 { 1984 struct net *net = dev_net(info->dev); 1985 int ret; 1986 1987 ASSERT_RTNL(); 1988 1989 /* Run per-netns notifier block chain first, then run the global one. 1990 * Hopefully, one day, the global one is going to be removed after 1991 * all notifier block registrators get converted to be per-netns. 1992 */ 1993 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 1994 if (ret & NOTIFY_STOP_MASK) 1995 return ret; 1996 return raw_notifier_call_chain(&netdev_chain, val, info); 1997 } 1998 1999 /** 2000 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2001 * for and rollback on error 2002 * @val_up: value passed unmodified to notifier function 2003 * @val_down: value passed unmodified to the notifier function when 2004 * recovering from an error on @val_up 2005 * @info: notifier information data 2006 * 2007 * Call all per-netns network notifier blocks, but not notifier blocks on 2008 * the global notifier chain. Parameters and return value are as for 2009 * raw_notifier_call_chain_robust(). 2010 */ 2011 2012 static int 2013 call_netdevice_notifiers_info_robust(unsigned long val_up, 2014 unsigned long val_down, 2015 struct netdev_notifier_info *info) 2016 { 2017 struct net *net = dev_net(info->dev); 2018 2019 ASSERT_RTNL(); 2020 2021 return raw_notifier_call_chain_robust(&net->netdev_chain, 2022 val_up, val_down, info); 2023 } 2024 2025 static int call_netdevice_notifiers_extack(unsigned long val, 2026 struct net_device *dev, 2027 struct netlink_ext_ack *extack) 2028 { 2029 struct netdev_notifier_info info = { 2030 .dev = dev, 2031 .extack = extack, 2032 }; 2033 2034 return call_netdevice_notifiers_info(val, &info); 2035 } 2036 2037 /** 2038 * call_netdevice_notifiers - call all network notifier blocks 2039 * @val: value passed unmodified to notifier function 2040 * @dev: net_device pointer passed unmodified to notifier function 2041 * 2042 * Call all network notifier blocks. Parameters and return value 2043 * are as for raw_notifier_call_chain(). 2044 */ 2045 2046 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2047 { 2048 return call_netdevice_notifiers_extack(val, dev, NULL); 2049 } 2050 EXPORT_SYMBOL(call_netdevice_notifiers); 2051 2052 /** 2053 * call_netdevice_notifiers_mtu - call all network notifier blocks 2054 * @val: value passed unmodified to notifier function 2055 * @dev: net_device pointer passed unmodified to notifier function 2056 * @arg: additional u32 argument passed to the notifier function 2057 * 2058 * Call all network notifier blocks. Parameters and return value 2059 * are as for raw_notifier_call_chain(). 2060 */ 2061 static int call_netdevice_notifiers_mtu(unsigned long val, 2062 struct net_device *dev, u32 arg) 2063 { 2064 struct netdev_notifier_info_ext info = { 2065 .info.dev = dev, 2066 .ext.mtu = arg, 2067 }; 2068 2069 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2070 2071 return call_netdevice_notifiers_info(val, &info.info); 2072 } 2073 2074 #ifdef CONFIG_NET_INGRESS 2075 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2076 2077 void net_inc_ingress_queue(void) 2078 { 2079 static_branch_inc(&ingress_needed_key); 2080 } 2081 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2082 2083 void net_dec_ingress_queue(void) 2084 { 2085 static_branch_dec(&ingress_needed_key); 2086 } 2087 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2088 #endif 2089 2090 #ifdef CONFIG_NET_EGRESS 2091 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2092 2093 void net_inc_egress_queue(void) 2094 { 2095 static_branch_inc(&egress_needed_key); 2096 } 2097 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2098 2099 void net_dec_egress_queue(void) 2100 { 2101 static_branch_dec(&egress_needed_key); 2102 } 2103 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2104 #endif 2105 2106 #ifdef CONFIG_NET_CLS_ACT 2107 DEFINE_STATIC_KEY_FALSE(tcf_bypass_check_needed_key); 2108 EXPORT_SYMBOL(tcf_bypass_check_needed_key); 2109 #endif 2110 2111 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2112 EXPORT_SYMBOL(netstamp_needed_key); 2113 #ifdef CONFIG_JUMP_LABEL 2114 static atomic_t netstamp_needed_deferred; 2115 static atomic_t netstamp_wanted; 2116 static void netstamp_clear(struct work_struct *work) 2117 { 2118 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2119 int wanted; 2120 2121 wanted = atomic_add_return(deferred, &netstamp_wanted); 2122 if (wanted > 0) 2123 static_branch_enable(&netstamp_needed_key); 2124 else 2125 static_branch_disable(&netstamp_needed_key); 2126 } 2127 static DECLARE_WORK(netstamp_work, netstamp_clear); 2128 #endif 2129 2130 void net_enable_timestamp(void) 2131 { 2132 #ifdef CONFIG_JUMP_LABEL 2133 int wanted = atomic_read(&netstamp_wanted); 2134 2135 while (wanted > 0) { 2136 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2137 return; 2138 } 2139 atomic_inc(&netstamp_needed_deferred); 2140 schedule_work(&netstamp_work); 2141 #else 2142 static_branch_inc(&netstamp_needed_key); 2143 #endif 2144 } 2145 EXPORT_SYMBOL(net_enable_timestamp); 2146 2147 void net_disable_timestamp(void) 2148 { 2149 #ifdef CONFIG_JUMP_LABEL 2150 int wanted = atomic_read(&netstamp_wanted); 2151 2152 while (wanted > 1) { 2153 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2154 return; 2155 } 2156 atomic_dec(&netstamp_needed_deferred); 2157 schedule_work(&netstamp_work); 2158 #else 2159 static_branch_dec(&netstamp_needed_key); 2160 #endif 2161 } 2162 EXPORT_SYMBOL(net_disable_timestamp); 2163 2164 static inline void net_timestamp_set(struct sk_buff *skb) 2165 { 2166 skb->tstamp = 0; 2167 skb->tstamp_type = SKB_CLOCK_REALTIME; 2168 if (static_branch_unlikely(&netstamp_needed_key)) 2169 skb->tstamp = ktime_get_real(); 2170 } 2171 2172 #define net_timestamp_check(COND, SKB) \ 2173 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2174 if ((COND) && !(SKB)->tstamp) \ 2175 (SKB)->tstamp = ktime_get_real(); \ 2176 } \ 2177 2178 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2179 { 2180 return __is_skb_forwardable(dev, skb, true); 2181 } 2182 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2183 2184 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2185 bool check_mtu) 2186 { 2187 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2188 2189 if (likely(!ret)) { 2190 skb->protocol = eth_type_trans(skb, dev); 2191 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2192 } 2193 2194 return ret; 2195 } 2196 2197 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2198 { 2199 return __dev_forward_skb2(dev, skb, true); 2200 } 2201 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2202 2203 /** 2204 * dev_forward_skb - loopback an skb to another netif 2205 * 2206 * @dev: destination network device 2207 * @skb: buffer to forward 2208 * 2209 * return values: 2210 * NET_RX_SUCCESS (no congestion) 2211 * NET_RX_DROP (packet was dropped, but freed) 2212 * 2213 * dev_forward_skb can be used for injecting an skb from the 2214 * start_xmit function of one device into the receive queue 2215 * of another device. 2216 * 2217 * The receiving device may be in another namespace, so 2218 * we have to clear all information in the skb that could 2219 * impact namespace isolation. 2220 */ 2221 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2222 { 2223 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2224 } 2225 EXPORT_SYMBOL_GPL(dev_forward_skb); 2226 2227 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2228 { 2229 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2230 } 2231 2232 static inline int deliver_skb(struct sk_buff *skb, 2233 struct packet_type *pt_prev, 2234 struct net_device *orig_dev) 2235 { 2236 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2237 return -ENOMEM; 2238 refcount_inc(&skb->users); 2239 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2240 } 2241 2242 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2243 struct packet_type **pt, 2244 struct net_device *orig_dev, 2245 __be16 type, 2246 struct list_head *ptype_list) 2247 { 2248 struct packet_type *ptype, *pt_prev = *pt; 2249 2250 list_for_each_entry_rcu(ptype, ptype_list, list) { 2251 if (ptype->type != type) 2252 continue; 2253 if (pt_prev) 2254 deliver_skb(skb, pt_prev, orig_dev); 2255 pt_prev = ptype; 2256 } 2257 *pt = pt_prev; 2258 } 2259 2260 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2261 { 2262 if (!ptype->af_packet_priv || !skb->sk) 2263 return false; 2264 2265 if (ptype->id_match) 2266 return ptype->id_match(ptype, skb->sk); 2267 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2268 return true; 2269 2270 return false; 2271 } 2272 2273 /** 2274 * dev_nit_active - return true if any network interface taps are in use 2275 * 2276 * @dev: network device to check for the presence of taps 2277 */ 2278 bool dev_nit_active(struct net_device *dev) 2279 { 2280 return !list_empty(&net_hotdata.ptype_all) || 2281 !list_empty(&dev->ptype_all); 2282 } 2283 EXPORT_SYMBOL_GPL(dev_nit_active); 2284 2285 /* 2286 * Support routine. Sends outgoing frames to any network 2287 * taps currently in use. 2288 */ 2289 2290 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2291 { 2292 struct list_head *ptype_list = &net_hotdata.ptype_all; 2293 struct packet_type *ptype, *pt_prev = NULL; 2294 struct sk_buff *skb2 = NULL; 2295 2296 rcu_read_lock(); 2297 again: 2298 list_for_each_entry_rcu(ptype, ptype_list, list) { 2299 if (READ_ONCE(ptype->ignore_outgoing)) 2300 continue; 2301 2302 /* Never send packets back to the socket 2303 * they originated from - MvS (miquels@drinkel.ow.org) 2304 */ 2305 if (skb_loop_sk(ptype, skb)) 2306 continue; 2307 2308 if (pt_prev) { 2309 deliver_skb(skb2, pt_prev, skb->dev); 2310 pt_prev = ptype; 2311 continue; 2312 } 2313 2314 /* need to clone skb, done only once */ 2315 skb2 = skb_clone(skb, GFP_ATOMIC); 2316 if (!skb2) 2317 goto out_unlock; 2318 2319 net_timestamp_set(skb2); 2320 2321 /* skb->nh should be correctly 2322 * set by sender, so that the second statement is 2323 * just protection against buggy protocols. 2324 */ 2325 skb_reset_mac_header(skb2); 2326 2327 if (skb_network_header(skb2) < skb2->data || 2328 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2329 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2330 ntohs(skb2->protocol), 2331 dev->name); 2332 skb_reset_network_header(skb2); 2333 } 2334 2335 skb2->transport_header = skb2->network_header; 2336 skb2->pkt_type = PACKET_OUTGOING; 2337 pt_prev = ptype; 2338 } 2339 2340 if (ptype_list == &net_hotdata.ptype_all) { 2341 ptype_list = &dev->ptype_all; 2342 goto again; 2343 } 2344 out_unlock: 2345 if (pt_prev) { 2346 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2347 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2348 else 2349 kfree_skb(skb2); 2350 } 2351 rcu_read_unlock(); 2352 } 2353 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2354 2355 /** 2356 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2357 * @dev: Network device 2358 * @txq: number of queues available 2359 * 2360 * If real_num_tx_queues is changed the tc mappings may no longer be 2361 * valid. To resolve this verify the tc mapping remains valid and if 2362 * not NULL the mapping. With no priorities mapping to this 2363 * offset/count pair it will no longer be used. In the worst case TC0 2364 * is invalid nothing can be done so disable priority mappings. If is 2365 * expected that drivers will fix this mapping if they can before 2366 * calling netif_set_real_num_tx_queues. 2367 */ 2368 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2369 { 2370 int i; 2371 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2372 2373 /* If TC0 is invalidated disable TC mapping */ 2374 if (tc->offset + tc->count > txq) { 2375 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2376 dev->num_tc = 0; 2377 return; 2378 } 2379 2380 /* Invalidated prio to tc mappings set to TC0 */ 2381 for (i = 1; i < TC_BITMASK + 1; i++) { 2382 int q = netdev_get_prio_tc_map(dev, i); 2383 2384 tc = &dev->tc_to_txq[q]; 2385 if (tc->offset + tc->count > txq) { 2386 netdev_warn(dev, "Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n", 2387 i, q); 2388 netdev_set_prio_tc_map(dev, i, 0); 2389 } 2390 } 2391 } 2392 2393 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2394 { 2395 if (dev->num_tc) { 2396 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2397 int i; 2398 2399 /* walk through the TCs and see if it falls into any of them */ 2400 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2401 if ((txq - tc->offset) < tc->count) 2402 return i; 2403 } 2404 2405 /* didn't find it, just return -1 to indicate no match */ 2406 return -1; 2407 } 2408 2409 return 0; 2410 } 2411 EXPORT_SYMBOL(netdev_txq_to_tc); 2412 2413 #ifdef CONFIG_XPS 2414 static struct static_key xps_needed __read_mostly; 2415 static struct static_key xps_rxqs_needed __read_mostly; 2416 static DEFINE_MUTEX(xps_map_mutex); 2417 #define xmap_dereference(P) \ 2418 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2419 2420 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2421 struct xps_dev_maps *old_maps, int tci, u16 index) 2422 { 2423 struct xps_map *map = NULL; 2424 int pos; 2425 2426 map = xmap_dereference(dev_maps->attr_map[tci]); 2427 if (!map) 2428 return false; 2429 2430 for (pos = map->len; pos--;) { 2431 if (map->queues[pos] != index) 2432 continue; 2433 2434 if (map->len > 1) { 2435 map->queues[pos] = map->queues[--map->len]; 2436 break; 2437 } 2438 2439 if (old_maps) 2440 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2441 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2442 kfree_rcu(map, rcu); 2443 return false; 2444 } 2445 2446 return true; 2447 } 2448 2449 static bool remove_xps_queue_cpu(struct net_device *dev, 2450 struct xps_dev_maps *dev_maps, 2451 int cpu, u16 offset, u16 count) 2452 { 2453 int num_tc = dev_maps->num_tc; 2454 bool active = false; 2455 int tci; 2456 2457 for (tci = cpu * num_tc; num_tc--; tci++) { 2458 int i, j; 2459 2460 for (i = count, j = offset; i--; j++) { 2461 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2462 break; 2463 } 2464 2465 active |= i < 0; 2466 } 2467 2468 return active; 2469 } 2470 2471 static void reset_xps_maps(struct net_device *dev, 2472 struct xps_dev_maps *dev_maps, 2473 enum xps_map_type type) 2474 { 2475 static_key_slow_dec_cpuslocked(&xps_needed); 2476 if (type == XPS_RXQS) 2477 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2478 2479 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2480 2481 kfree_rcu(dev_maps, rcu); 2482 } 2483 2484 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2485 u16 offset, u16 count) 2486 { 2487 struct xps_dev_maps *dev_maps; 2488 bool active = false; 2489 int i, j; 2490 2491 dev_maps = xmap_dereference(dev->xps_maps[type]); 2492 if (!dev_maps) 2493 return; 2494 2495 for (j = 0; j < dev_maps->nr_ids; j++) 2496 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2497 if (!active) 2498 reset_xps_maps(dev, dev_maps, type); 2499 2500 if (type == XPS_CPUS) { 2501 for (i = offset + (count - 1); count--; i--) 2502 netdev_queue_numa_node_write( 2503 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2504 } 2505 } 2506 2507 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2508 u16 count) 2509 { 2510 if (!static_key_false(&xps_needed)) 2511 return; 2512 2513 cpus_read_lock(); 2514 mutex_lock(&xps_map_mutex); 2515 2516 if (static_key_false(&xps_rxqs_needed)) 2517 clean_xps_maps(dev, XPS_RXQS, offset, count); 2518 2519 clean_xps_maps(dev, XPS_CPUS, offset, count); 2520 2521 mutex_unlock(&xps_map_mutex); 2522 cpus_read_unlock(); 2523 } 2524 2525 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2526 { 2527 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2528 } 2529 2530 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2531 u16 index, bool is_rxqs_map) 2532 { 2533 struct xps_map *new_map; 2534 int alloc_len = XPS_MIN_MAP_ALLOC; 2535 int i, pos; 2536 2537 for (pos = 0; map && pos < map->len; pos++) { 2538 if (map->queues[pos] != index) 2539 continue; 2540 return map; 2541 } 2542 2543 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2544 if (map) { 2545 if (pos < map->alloc_len) 2546 return map; 2547 2548 alloc_len = map->alloc_len * 2; 2549 } 2550 2551 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2552 * map 2553 */ 2554 if (is_rxqs_map) 2555 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2556 else 2557 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2558 cpu_to_node(attr_index)); 2559 if (!new_map) 2560 return NULL; 2561 2562 for (i = 0; i < pos; i++) 2563 new_map->queues[i] = map->queues[i]; 2564 new_map->alloc_len = alloc_len; 2565 new_map->len = pos; 2566 2567 return new_map; 2568 } 2569 2570 /* Copy xps maps at a given index */ 2571 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2572 struct xps_dev_maps *new_dev_maps, int index, 2573 int tc, bool skip_tc) 2574 { 2575 int i, tci = index * dev_maps->num_tc; 2576 struct xps_map *map; 2577 2578 /* copy maps belonging to foreign traffic classes */ 2579 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2580 if (i == tc && skip_tc) 2581 continue; 2582 2583 /* fill in the new device map from the old device map */ 2584 map = xmap_dereference(dev_maps->attr_map[tci]); 2585 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2586 } 2587 } 2588 2589 /* Must be called under cpus_read_lock */ 2590 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2591 u16 index, enum xps_map_type type) 2592 { 2593 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2594 const unsigned long *online_mask = NULL; 2595 bool active = false, copy = false; 2596 int i, j, tci, numa_node_id = -2; 2597 int maps_sz, num_tc = 1, tc = 0; 2598 struct xps_map *map, *new_map; 2599 unsigned int nr_ids; 2600 2601 WARN_ON_ONCE(index >= dev->num_tx_queues); 2602 2603 if (dev->num_tc) { 2604 /* Do not allow XPS on subordinate device directly */ 2605 num_tc = dev->num_tc; 2606 if (num_tc < 0) 2607 return -EINVAL; 2608 2609 /* If queue belongs to subordinate dev use its map */ 2610 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2611 2612 tc = netdev_txq_to_tc(dev, index); 2613 if (tc < 0) 2614 return -EINVAL; 2615 } 2616 2617 mutex_lock(&xps_map_mutex); 2618 2619 dev_maps = xmap_dereference(dev->xps_maps[type]); 2620 if (type == XPS_RXQS) { 2621 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2622 nr_ids = dev->num_rx_queues; 2623 } else { 2624 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2625 if (num_possible_cpus() > 1) 2626 online_mask = cpumask_bits(cpu_online_mask); 2627 nr_ids = nr_cpu_ids; 2628 } 2629 2630 if (maps_sz < L1_CACHE_BYTES) 2631 maps_sz = L1_CACHE_BYTES; 2632 2633 /* The old dev_maps could be larger or smaller than the one we're 2634 * setting up now, as dev->num_tc or nr_ids could have been updated in 2635 * between. We could try to be smart, but let's be safe instead and only 2636 * copy foreign traffic classes if the two map sizes match. 2637 */ 2638 if (dev_maps && 2639 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2640 copy = true; 2641 2642 /* allocate memory for queue storage */ 2643 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2644 j < nr_ids;) { 2645 if (!new_dev_maps) { 2646 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2647 if (!new_dev_maps) { 2648 mutex_unlock(&xps_map_mutex); 2649 return -ENOMEM; 2650 } 2651 2652 new_dev_maps->nr_ids = nr_ids; 2653 new_dev_maps->num_tc = num_tc; 2654 } 2655 2656 tci = j * num_tc + tc; 2657 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2658 2659 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2660 if (!map) 2661 goto error; 2662 2663 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2664 } 2665 2666 if (!new_dev_maps) 2667 goto out_no_new_maps; 2668 2669 if (!dev_maps) { 2670 /* Increment static keys at most once per type */ 2671 static_key_slow_inc_cpuslocked(&xps_needed); 2672 if (type == XPS_RXQS) 2673 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2674 } 2675 2676 for (j = 0; j < nr_ids; j++) { 2677 bool skip_tc = false; 2678 2679 tci = j * num_tc + tc; 2680 if (netif_attr_test_mask(j, mask, nr_ids) && 2681 netif_attr_test_online(j, online_mask, nr_ids)) { 2682 /* add tx-queue to CPU/rx-queue maps */ 2683 int pos = 0; 2684 2685 skip_tc = true; 2686 2687 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2688 while ((pos < map->len) && (map->queues[pos] != index)) 2689 pos++; 2690 2691 if (pos == map->len) 2692 map->queues[map->len++] = index; 2693 #ifdef CONFIG_NUMA 2694 if (type == XPS_CPUS) { 2695 if (numa_node_id == -2) 2696 numa_node_id = cpu_to_node(j); 2697 else if (numa_node_id != cpu_to_node(j)) 2698 numa_node_id = -1; 2699 } 2700 #endif 2701 } 2702 2703 if (copy) 2704 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2705 skip_tc); 2706 } 2707 2708 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2709 2710 /* Cleanup old maps */ 2711 if (!dev_maps) 2712 goto out_no_old_maps; 2713 2714 for (j = 0; j < dev_maps->nr_ids; j++) { 2715 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2716 map = xmap_dereference(dev_maps->attr_map[tci]); 2717 if (!map) 2718 continue; 2719 2720 if (copy) { 2721 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2722 if (map == new_map) 2723 continue; 2724 } 2725 2726 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2727 kfree_rcu(map, rcu); 2728 } 2729 } 2730 2731 old_dev_maps = dev_maps; 2732 2733 out_no_old_maps: 2734 dev_maps = new_dev_maps; 2735 active = true; 2736 2737 out_no_new_maps: 2738 if (type == XPS_CPUS) 2739 /* update Tx queue numa node */ 2740 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2741 (numa_node_id >= 0) ? 2742 numa_node_id : NUMA_NO_NODE); 2743 2744 if (!dev_maps) 2745 goto out_no_maps; 2746 2747 /* removes tx-queue from unused CPUs/rx-queues */ 2748 for (j = 0; j < dev_maps->nr_ids; j++) { 2749 tci = j * dev_maps->num_tc; 2750 2751 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2752 if (i == tc && 2753 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 2754 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 2755 continue; 2756 2757 active |= remove_xps_queue(dev_maps, 2758 copy ? old_dev_maps : NULL, 2759 tci, index); 2760 } 2761 } 2762 2763 if (old_dev_maps) 2764 kfree_rcu(old_dev_maps, rcu); 2765 2766 /* free map if not active */ 2767 if (!active) 2768 reset_xps_maps(dev, dev_maps, type); 2769 2770 out_no_maps: 2771 mutex_unlock(&xps_map_mutex); 2772 2773 return 0; 2774 error: 2775 /* remove any maps that we added */ 2776 for (j = 0; j < nr_ids; j++) { 2777 for (i = num_tc, tci = j * num_tc; i--; tci++) { 2778 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2779 map = copy ? 2780 xmap_dereference(dev_maps->attr_map[tci]) : 2781 NULL; 2782 if (new_map && new_map != map) 2783 kfree(new_map); 2784 } 2785 } 2786 2787 mutex_unlock(&xps_map_mutex); 2788 2789 kfree(new_dev_maps); 2790 return -ENOMEM; 2791 } 2792 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 2793 2794 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 2795 u16 index) 2796 { 2797 int ret; 2798 2799 cpus_read_lock(); 2800 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 2801 cpus_read_unlock(); 2802 2803 return ret; 2804 } 2805 EXPORT_SYMBOL(netif_set_xps_queue); 2806 2807 #endif 2808 static void netdev_unbind_all_sb_channels(struct net_device *dev) 2809 { 2810 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 2811 2812 /* Unbind any subordinate channels */ 2813 while (txq-- != &dev->_tx[0]) { 2814 if (txq->sb_dev) 2815 netdev_unbind_sb_channel(dev, txq->sb_dev); 2816 } 2817 } 2818 2819 void netdev_reset_tc(struct net_device *dev) 2820 { 2821 #ifdef CONFIG_XPS 2822 netif_reset_xps_queues_gt(dev, 0); 2823 #endif 2824 netdev_unbind_all_sb_channels(dev); 2825 2826 /* Reset TC configuration of device */ 2827 dev->num_tc = 0; 2828 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 2829 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 2830 } 2831 EXPORT_SYMBOL(netdev_reset_tc); 2832 2833 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 2834 { 2835 if (tc >= dev->num_tc) 2836 return -EINVAL; 2837 2838 #ifdef CONFIG_XPS 2839 netif_reset_xps_queues(dev, offset, count); 2840 #endif 2841 dev->tc_to_txq[tc].count = count; 2842 dev->tc_to_txq[tc].offset = offset; 2843 return 0; 2844 } 2845 EXPORT_SYMBOL(netdev_set_tc_queue); 2846 2847 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 2848 { 2849 if (num_tc > TC_MAX_QUEUE) 2850 return -EINVAL; 2851 2852 #ifdef CONFIG_XPS 2853 netif_reset_xps_queues_gt(dev, 0); 2854 #endif 2855 netdev_unbind_all_sb_channels(dev); 2856 2857 dev->num_tc = num_tc; 2858 return 0; 2859 } 2860 EXPORT_SYMBOL(netdev_set_num_tc); 2861 2862 void netdev_unbind_sb_channel(struct net_device *dev, 2863 struct net_device *sb_dev) 2864 { 2865 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 2866 2867 #ifdef CONFIG_XPS 2868 netif_reset_xps_queues_gt(sb_dev, 0); 2869 #endif 2870 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 2871 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 2872 2873 while (txq-- != &dev->_tx[0]) { 2874 if (txq->sb_dev == sb_dev) 2875 txq->sb_dev = NULL; 2876 } 2877 } 2878 EXPORT_SYMBOL(netdev_unbind_sb_channel); 2879 2880 int netdev_bind_sb_channel_queue(struct net_device *dev, 2881 struct net_device *sb_dev, 2882 u8 tc, u16 count, u16 offset) 2883 { 2884 /* Make certain the sb_dev and dev are already configured */ 2885 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 2886 return -EINVAL; 2887 2888 /* We cannot hand out queues we don't have */ 2889 if ((offset + count) > dev->real_num_tx_queues) 2890 return -EINVAL; 2891 2892 /* Record the mapping */ 2893 sb_dev->tc_to_txq[tc].count = count; 2894 sb_dev->tc_to_txq[tc].offset = offset; 2895 2896 /* Provide a way for Tx queue to find the tc_to_txq map or 2897 * XPS map for itself. 2898 */ 2899 while (count--) 2900 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 2901 2902 return 0; 2903 } 2904 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 2905 2906 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 2907 { 2908 /* Do not use a multiqueue device to represent a subordinate channel */ 2909 if (netif_is_multiqueue(dev)) 2910 return -ENODEV; 2911 2912 /* We allow channels 1 - 32767 to be used for subordinate channels. 2913 * Channel 0 is meant to be "native" mode and used only to represent 2914 * the main root device. We allow writing 0 to reset the device back 2915 * to normal mode after being used as a subordinate channel. 2916 */ 2917 if (channel > S16_MAX) 2918 return -EINVAL; 2919 2920 dev->num_tc = -channel; 2921 2922 return 0; 2923 } 2924 EXPORT_SYMBOL(netdev_set_sb_channel); 2925 2926 /* 2927 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 2928 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 2929 */ 2930 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 2931 { 2932 bool disabling; 2933 int rc; 2934 2935 disabling = txq < dev->real_num_tx_queues; 2936 2937 if (txq < 1 || txq > dev->num_tx_queues) 2938 return -EINVAL; 2939 2940 if (dev->reg_state == NETREG_REGISTERED || 2941 dev->reg_state == NETREG_UNREGISTERING) { 2942 ASSERT_RTNL(); 2943 2944 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 2945 txq); 2946 if (rc) 2947 return rc; 2948 2949 if (dev->num_tc) 2950 netif_setup_tc(dev, txq); 2951 2952 dev_qdisc_change_real_num_tx(dev, txq); 2953 2954 dev->real_num_tx_queues = txq; 2955 2956 if (disabling) { 2957 synchronize_net(); 2958 qdisc_reset_all_tx_gt(dev, txq); 2959 #ifdef CONFIG_XPS 2960 netif_reset_xps_queues_gt(dev, txq); 2961 #endif 2962 } 2963 } else { 2964 dev->real_num_tx_queues = txq; 2965 } 2966 2967 return 0; 2968 } 2969 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 2970 2971 #ifdef CONFIG_SYSFS 2972 /** 2973 * netif_set_real_num_rx_queues - set actual number of RX queues used 2974 * @dev: Network device 2975 * @rxq: Actual number of RX queues 2976 * 2977 * This must be called either with the rtnl_lock held or before 2978 * registration of the net device. Returns 0 on success, or a 2979 * negative error code. If called before registration, it always 2980 * succeeds. 2981 */ 2982 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 2983 { 2984 int rc; 2985 2986 if (rxq < 1 || rxq > dev->num_rx_queues) 2987 return -EINVAL; 2988 2989 if (dev->reg_state == NETREG_REGISTERED) { 2990 ASSERT_RTNL(); 2991 2992 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 2993 rxq); 2994 if (rc) 2995 return rc; 2996 } 2997 2998 dev->real_num_rx_queues = rxq; 2999 return 0; 3000 } 3001 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3002 #endif 3003 3004 /** 3005 * netif_set_real_num_queues - set actual number of RX and TX queues used 3006 * @dev: Network device 3007 * @txq: Actual number of TX queues 3008 * @rxq: Actual number of RX queues 3009 * 3010 * Set the real number of both TX and RX queues. 3011 * Does nothing if the number of queues is already correct. 3012 */ 3013 int netif_set_real_num_queues(struct net_device *dev, 3014 unsigned int txq, unsigned int rxq) 3015 { 3016 unsigned int old_rxq = dev->real_num_rx_queues; 3017 int err; 3018 3019 if (txq < 1 || txq > dev->num_tx_queues || 3020 rxq < 1 || rxq > dev->num_rx_queues) 3021 return -EINVAL; 3022 3023 /* Start from increases, so the error path only does decreases - 3024 * decreases can't fail. 3025 */ 3026 if (rxq > dev->real_num_rx_queues) { 3027 err = netif_set_real_num_rx_queues(dev, rxq); 3028 if (err) 3029 return err; 3030 } 3031 if (txq > dev->real_num_tx_queues) { 3032 err = netif_set_real_num_tx_queues(dev, txq); 3033 if (err) 3034 goto undo_rx; 3035 } 3036 if (rxq < dev->real_num_rx_queues) 3037 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3038 if (txq < dev->real_num_tx_queues) 3039 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3040 3041 return 0; 3042 undo_rx: 3043 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3044 return err; 3045 } 3046 EXPORT_SYMBOL(netif_set_real_num_queues); 3047 3048 /** 3049 * netif_set_tso_max_size() - set the max size of TSO frames supported 3050 * @dev: netdev to update 3051 * @size: max skb->len of a TSO frame 3052 * 3053 * Set the limit on the size of TSO super-frames the device can handle. 3054 * Unless explicitly set the stack will assume the value of 3055 * %GSO_LEGACY_MAX_SIZE. 3056 */ 3057 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3058 { 3059 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3060 if (size < READ_ONCE(dev->gso_max_size)) 3061 netif_set_gso_max_size(dev, size); 3062 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3063 netif_set_gso_ipv4_max_size(dev, size); 3064 } 3065 EXPORT_SYMBOL(netif_set_tso_max_size); 3066 3067 /** 3068 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3069 * @dev: netdev to update 3070 * @segs: max number of TCP segments 3071 * 3072 * Set the limit on the number of TCP segments the device can generate from 3073 * a single TSO super-frame. 3074 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3075 */ 3076 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3077 { 3078 dev->tso_max_segs = segs; 3079 if (segs < READ_ONCE(dev->gso_max_segs)) 3080 netif_set_gso_max_segs(dev, segs); 3081 } 3082 EXPORT_SYMBOL(netif_set_tso_max_segs); 3083 3084 /** 3085 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3086 * @to: netdev to update 3087 * @from: netdev from which to copy the limits 3088 */ 3089 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3090 { 3091 netif_set_tso_max_size(to, from->tso_max_size); 3092 netif_set_tso_max_segs(to, from->tso_max_segs); 3093 } 3094 EXPORT_SYMBOL(netif_inherit_tso_max); 3095 3096 /** 3097 * netif_get_num_default_rss_queues - default number of RSS queues 3098 * 3099 * Default value is the number of physical cores if there are only 1 or 2, or 3100 * divided by 2 if there are more. 3101 */ 3102 int netif_get_num_default_rss_queues(void) 3103 { 3104 cpumask_var_t cpus; 3105 int cpu, count = 0; 3106 3107 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3108 return 1; 3109 3110 cpumask_copy(cpus, cpu_online_mask); 3111 for_each_cpu(cpu, cpus) { 3112 ++count; 3113 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3114 } 3115 free_cpumask_var(cpus); 3116 3117 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3118 } 3119 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3120 3121 static void __netif_reschedule(struct Qdisc *q) 3122 { 3123 struct softnet_data *sd; 3124 unsigned long flags; 3125 3126 local_irq_save(flags); 3127 sd = this_cpu_ptr(&softnet_data); 3128 q->next_sched = NULL; 3129 *sd->output_queue_tailp = q; 3130 sd->output_queue_tailp = &q->next_sched; 3131 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3132 local_irq_restore(flags); 3133 } 3134 3135 void __netif_schedule(struct Qdisc *q) 3136 { 3137 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3138 __netif_reschedule(q); 3139 } 3140 EXPORT_SYMBOL(__netif_schedule); 3141 3142 struct dev_kfree_skb_cb { 3143 enum skb_drop_reason reason; 3144 }; 3145 3146 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3147 { 3148 return (struct dev_kfree_skb_cb *)skb->cb; 3149 } 3150 3151 void netif_schedule_queue(struct netdev_queue *txq) 3152 { 3153 rcu_read_lock(); 3154 if (!netif_xmit_stopped(txq)) { 3155 struct Qdisc *q = rcu_dereference(txq->qdisc); 3156 3157 __netif_schedule(q); 3158 } 3159 rcu_read_unlock(); 3160 } 3161 EXPORT_SYMBOL(netif_schedule_queue); 3162 3163 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3164 { 3165 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3166 struct Qdisc *q; 3167 3168 rcu_read_lock(); 3169 q = rcu_dereference(dev_queue->qdisc); 3170 __netif_schedule(q); 3171 rcu_read_unlock(); 3172 } 3173 } 3174 EXPORT_SYMBOL(netif_tx_wake_queue); 3175 3176 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3177 { 3178 unsigned long flags; 3179 3180 if (unlikely(!skb)) 3181 return; 3182 3183 if (likely(refcount_read(&skb->users) == 1)) { 3184 smp_rmb(); 3185 refcount_set(&skb->users, 0); 3186 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3187 return; 3188 } 3189 get_kfree_skb_cb(skb)->reason = reason; 3190 local_irq_save(flags); 3191 skb->next = __this_cpu_read(softnet_data.completion_queue); 3192 __this_cpu_write(softnet_data.completion_queue, skb); 3193 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3194 local_irq_restore(flags); 3195 } 3196 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3197 3198 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3199 { 3200 if (in_hardirq() || irqs_disabled()) 3201 dev_kfree_skb_irq_reason(skb, reason); 3202 else 3203 kfree_skb_reason(skb, reason); 3204 } 3205 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3206 3207 3208 /** 3209 * netif_device_detach - mark device as removed 3210 * @dev: network device 3211 * 3212 * Mark device as removed from system and therefore no longer available. 3213 */ 3214 void netif_device_detach(struct net_device *dev) 3215 { 3216 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3217 netif_running(dev)) { 3218 netif_tx_stop_all_queues(dev); 3219 } 3220 } 3221 EXPORT_SYMBOL(netif_device_detach); 3222 3223 /** 3224 * netif_device_attach - mark device as attached 3225 * @dev: network device 3226 * 3227 * Mark device as attached from system and restart if needed. 3228 */ 3229 void netif_device_attach(struct net_device *dev) 3230 { 3231 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3232 netif_running(dev)) { 3233 netif_tx_wake_all_queues(dev); 3234 __netdev_watchdog_up(dev); 3235 } 3236 } 3237 EXPORT_SYMBOL(netif_device_attach); 3238 3239 /* 3240 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3241 * to be used as a distribution range. 3242 */ 3243 static u16 skb_tx_hash(const struct net_device *dev, 3244 const struct net_device *sb_dev, 3245 struct sk_buff *skb) 3246 { 3247 u32 hash; 3248 u16 qoffset = 0; 3249 u16 qcount = dev->real_num_tx_queues; 3250 3251 if (dev->num_tc) { 3252 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3253 3254 qoffset = sb_dev->tc_to_txq[tc].offset; 3255 qcount = sb_dev->tc_to_txq[tc].count; 3256 if (unlikely(!qcount)) { 3257 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3258 sb_dev->name, qoffset, tc); 3259 qoffset = 0; 3260 qcount = dev->real_num_tx_queues; 3261 } 3262 } 3263 3264 if (skb_rx_queue_recorded(skb)) { 3265 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3266 hash = skb_get_rx_queue(skb); 3267 if (hash >= qoffset) 3268 hash -= qoffset; 3269 while (unlikely(hash >= qcount)) 3270 hash -= qcount; 3271 return hash + qoffset; 3272 } 3273 3274 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3275 } 3276 3277 void skb_warn_bad_offload(const struct sk_buff *skb) 3278 { 3279 static const netdev_features_t null_features; 3280 struct net_device *dev = skb->dev; 3281 const char *name = ""; 3282 3283 if (!net_ratelimit()) 3284 return; 3285 3286 if (dev) { 3287 if (dev->dev.parent) 3288 name = dev_driver_string(dev->dev.parent); 3289 else 3290 name = netdev_name(dev); 3291 } 3292 skb_dump(KERN_WARNING, skb, false); 3293 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3294 name, dev ? &dev->features : &null_features, 3295 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3296 } 3297 3298 /* 3299 * Invalidate hardware checksum when packet is to be mangled, and 3300 * complete checksum manually on outgoing path. 3301 */ 3302 int skb_checksum_help(struct sk_buff *skb) 3303 { 3304 __wsum csum; 3305 int ret = 0, offset; 3306 3307 if (skb->ip_summed == CHECKSUM_COMPLETE) 3308 goto out_set_summed; 3309 3310 if (unlikely(skb_is_gso(skb))) { 3311 skb_warn_bad_offload(skb); 3312 return -EINVAL; 3313 } 3314 3315 if (!skb_frags_readable(skb)) { 3316 return -EFAULT; 3317 } 3318 3319 /* Before computing a checksum, we should make sure no frag could 3320 * be modified by an external entity : checksum could be wrong. 3321 */ 3322 if (skb_has_shared_frag(skb)) { 3323 ret = __skb_linearize(skb); 3324 if (ret) 3325 goto out; 3326 } 3327 3328 offset = skb_checksum_start_offset(skb); 3329 ret = -EINVAL; 3330 if (unlikely(offset >= skb_headlen(skb))) { 3331 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3332 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3333 offset, skb_headlen(skb)); 3334 goto out; 3335 } 3336 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3337 3338 offset += skb->csum_offset; 3339 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3340 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3341 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3342 offset + sizeof(__sum16), skb_headlen(skb)); 3343 goto out; 3344 } 3345 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3346 if (ret) 3347 goto out; 3348 3349 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3350 out_set_summed: 3351 skb->ip_summed = CHECKSUM_NONE; 3352 out: 3353 return ret; 3354 } 3355 EXPORT_SYMBOL(skb_checksum_help); 3356 3357 int skb_crc32c_csum_help(struct sk_buff *skb) 3358 { 3359 __le32 crc32c_csum; 3360 int ret = 0, offset, start; 3361 3362 if (skb->ip_summed != CHECKSUM_PARTIAL) 3363 goto out; 3364 3365 if (unlikely(skb_is_gso(skb))) 3366 goto out; 3367 3368 /* Before computing a checksum, we should make sure no frag could 3369 * be modified by an external entity : checksum could be wrong. 3370 */ 3371 if (unlikely(skb_has_shared_frag(skb))) { 3372 ret = __skb_linearize(skb); 3373 if (ret) 3374 goto out; 3375 } 3376 start = skb_checksum_start_offset(skb); 3377 offset = start + offsetof(struct sctphdr, checksum); 3378 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3379 ret = -EINVAL; 3380 goto out; 3381 } 3382 3383 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3384 if (ret) 3385 goto out; 3386 3387 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start, 3388 skb->len - start, ~(__u32)0, 3389 crc32c_csum_stub)); 3390 *(__le32 *)(skb->data + offset) = crc32c_csum; 3391 skb_reset_csum_not_inet(skb); 3392 out: 3393 return ret; 3394 } 3395 EXPORT_SYMBOL(skb_crc32c_csum_help); 3396 3397 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3398 { 3399 __be16 type = skb->protocol; 3400 3401 /* Tunnel gso handlers can set protocol to ethernet. */ 3402 if (type == htons(ETH_P_TEB)) { 3403 struct ethhdr *eth; 3404 3405 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3406 return 0; 3407 3408 eth = (struct ethhdr *)skb->data; 3409 type = eth->h_proto; 3410 } 3411 3412 return vlan_get_protocol_and_depth(skb, type, depth); 3413 } 3414 3415 3416 /* Take action when hardware reception checksum errors are detected. */ 3417 #ifdef CONFIG_BUG 3418 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3419 { 3420 netdev_err(dev, "hw csum failure\n"); 3421 skb_dump(KERN_ERR, skb, true); 3422 dump_stack(); 3423 } 3424 3425 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3426 { 3427 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3428 } 3429 EXPORT_SYMBOL(netdev_rx_csum_fault); 3430 #endif 3431 3432 /* XXX: check that highmem exists at all on the given machine. */ 3433 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3434 { 3435 #ifdef CONFIG_HIGHMEM 3436 int i; 3437 3438 if (!(dev->features & NETIF_F_HIGHDMA)) { 3439 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3440 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3441 struct page *page = skb_frag_page(frag); 3442 3443 if (page && PageHighMem(page)) 3444 return 1; 3445 } 3446 } 3447 #endif 3448 return 0; 3449 } 3450 3451 /* If MPLS offload request, verify we are testing hardware MPLS features 3452 * instead of standard features for the netdev. 3453 */ 3454 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3455 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3456 netdev_features_t features, 3457 __be16 type) 3458 { 3459 if (eth_p_mpls(type)) 3460 features &= skb->dev->mpls_features; 3461 3462 return features; 3463 } 3464 #else 3465 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3466 netdev_features_t features, 3467 __be16 type) 3468 { 3469 return features; 3470 } 3471 #endif 3472 3473 static netdev_features_t harmonize_features(struct sk_buff *skb, 3474 netdev_features_t features) 3475 { 3476 __be16 type; 3477 3478 type = skb_network_protocol(skb, NULL); 3479 features = net_mpls_features(skb, features, type); 3480 3481 if (skb->ip_summed != CHECKSUM_NONE && 3482 !can_checksum_protocol(features, type)) { 3483 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3484 } 3485 if (illegal_highdma(skb->dev, skb)) 3486 features &= ~NETIF_F_SG; 3487 3488 return features; 3489 } 3490 3491 netdev_features_t passthru_features_check(struct sk_buff *skb, 3492 struct net_device *dev, 3493 netdev_features_t features) 3494 { 3495 return features; 3496 } 3497 EXPORT_SYMBOL(passthru_features_check); 3498 3499 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3500 struct net_device *dev, 3501 netdev_features_t features) 3502 { 3503 return vlan_features_check(skb, features); 3504 } 3505 3506 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3507 struct net_device *dev, 3508 netdev_features_t features) 3509 { 3510 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3511 3512 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3513 return features & ~NETIF_F_GSO_MASK; 3514 3515 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3516 return features & ~NETIF_F_GSO_MASK; 3517 3518 if (!skb_shinfo(skb)->gso_type) { 3519 skb_warn_bad_offload(skb); 3520 return features & ~NETIF_F_GSO_MASK; 3521 } 3522 3523 /* Support for GSO partial features requires software 3524 * intervention before we can actually process the packets 3525 * so we need to strip support for any partial features now 3526 * and we can pull them back in after we have partially 3527 * segmented the frame. 3528 */ 3529 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3530 features &= ~dev->gso_partial_features; 3531 3532 /* Make sure to clear the IPv4 ID mangling feature if the 3533 * IPv4 header has the potential to be fragmented. 3534 */ 3535 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3536 struct iphdr *iph = skb->encapsulation ? 3537 inner_ip_hdr(skb) : ip_hdr(skb); 3538 3539 if (!(iph->frag_off & htons(IP_DF))) 3540 features &= ~NETIF_F_TSO_MANGLEID; 3541 } 3542 3543 return features; 3544 } 3545 3546 netdev_features_t netif_skb_features(struct sk_buff *skb) 3547 { 3548 struct net_device *dev = skb->dev; 3549 netdev_features_t features = dev->features; 3550 3551 if (skb_is_gso(skb)) 3552 features = gso_features_check(skb, dev, features); 3553 3554 /* If encapsulation offload request, verify we are testing 3555 * hardware encapsulation features instead of standard 3556 * features for the netdev 3557 */ 3558 if (skb->encapsulation) 3559 features &= dev->hw_enc_features; 3560 3561 if (skb_vlan_tagged(skb)) 3562 features = netdev_intersect_features(features, 3563 dev->vlan_features | 3564 NETIF_F_HW_VLAN_CTAG_TX | 3565 NETIF_F_HW_VLAN_STAG_TX); 3566 3567 if (dev->netdev_ops->ndo_features_check) 3568 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3569 features); 3570 else 3571 features &= dflt_features_check(skb, dev, features); 3572 3573 return harmonize_features(skb, features); 3574 } 3575 EXPORT_SYMBOL(netif_skb_features); 3576 3577 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3578 struct netdev_queue *txq, bool more) 3579 { 3580 unsigned int len; 3581 int rc; 3582 3583 if (dev_nit_active(dev)) 3584 dev_queue_xmit_nit(skb, dev); 3585 3586 len = skb->len; 3587 trace_net_dev_start_xmit(skb, dev); 3588 rc = netdev_start_xmit(skb, dev, txq, more); 3589 trace_net_dev_xmit(skb, rc, dev, len); 3590 3591 return rc; 3592 } 3593 3594 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3595 struct netdev_queue *txq, int *ret) 3596 { 3597 struct sk_buff *skb = first; 3598 int rc = NETDEV_TX_OK; 3599 3600 while (skb) { 3601 struct sk_buff *next = skb->next; 3602 3603 skb_mark_not_on_list(skb); 3604 rc = xmit_one(skb, dev, txq, next != NULL); 3605 if (unlikely(!dev_xmit_complete(rc))) { 3606 skb->next = next; 3607 goto out; 3608 } 3609 3610 skb = next; 3611 if (netif_tx_queue_stopped(txq) && skb) { 3612 rc = NETDEV_TX_BUSY; 3613 break; 3614 } 3615 } 3616 3617 out: 3618 *ret = rc; 3619 return skb; 3620 } 3621 3622 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3623 netdev_features_t features) 3624 { 3625 if (skb_vlan_tag_present(skb) && 3626 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3627 skb = __vlan_hwaccel_push_inside(skb); 3628 return skb; 3629 } 3630 3631 int skb_csum_hwoffload_help(struct sk_buff *skb, 3632 const netdev_features_t features) 3633 { 3634 if (unlikely(skb_csum_is_sctp(skb))) 3635 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3636 skb_crc32c_csum_help(skb); 3637 3638 if (features & NETIF_F_HW_CSUM) 3639 return 0; 3640 3641 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3642 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3643 skb_network_header_len(skb) != sizeof(struct ipv6hdr)) 3644 goto sw_checksum; 3645 switch (skb->csum_offset) { 3646 case offsetof(struct tcphdr, check): 3647 case offsetof(struct udphdr, check): 3648 return 0; 3649 } 3650 } 3651 3652 sw_checksum: 3653 return skb_checksum_help(skb); 3654 } 3655 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3656 3657 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3658 { 3659 netdev_features_t features; 3660 3661 features = netif_skb_features(skb); 3662 skb = validate_xmit_vlan(skb, features); 3663 if (unlikely(!skb)) 3664 goto out_null; 3665 3666 skb = sk_validate_xmit_skb(skb, dev); 3667 if (unlikely(!skb)) 3668 goto out_null; 3669 3670 if (netif_needs_gso(skb, features)) { 3671 struct sk_buff *segs; 3672 3673 segs = skb_gso_segment(skb, features); 3674 if (IS_ERR(segs)) { 3675 goto out_kfree_skb; 3676 } else if (segs) { 3677 consume_skb(skb); 3678 skb = segs; 3679 } 3680 } else { 3681 if (skb_needs_linearize(skb, features) && 3682 __skb_linearize(skb)) 3683 goto out_kfree_skb; 3684 3685 /* If packet is not checksummed and device does not 3686 * support checksumming for this protocol, complete 3687 * checksumming here. 3688 */ 3689 if (skb->ip_summed == CHECKSUM_PARTIAL) { 3690 if (skb->encapsulation) 3691 skb_set_inner_transport_header(skb, 3692 skb_checksum_start_offset(skb)); 3693 else 3694 skb_set_transport_header(skb, 3695 skb_checksum_start_offset(skb)); 3696 if (skb_csum_hwoffload_help(skb, features)) 3697 goto out_kfree_skb; 3698 } 3699 } 3700 3701 skb = validate_xmit_xfrm(skb, features, again); 3702 3703 return skb; 3704 3705 out_kfree_skb: 3706 kfree_skb(skb); 3707 out_null: 3708 dev_core_stats_tx_dropped_inc(dev); 3709 return NULL; 3710 } 3711 3712 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 3713 { 3714 struct sk_buff *next, *head = NULL, *tail; 3715 3716 for (; skb != NULL; skb = next) { 3717 next = skb->next; 3718 skb_mark_not_on_list(skb); 3719 3720 /* in case skb won't be segmented, point to itself */ 3721 skb->prev = skb; 3722 3723 skb = validate_xmit_skb(skb, dev, again); 3724 if (!skb) 3725 continue; 3726 3727 if (!head) 3728 head = skb; 3729 else 3730 tail->next = skb; 3731 /* If skb was segmented, skb->prev points to 3732 * the last segment. If not, it still contains skb. 3733 */ 3734 tail = skb->prev; 3735 } 3736 return head; 3737 } 3738 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 3739 3740 static void qdisc_pkt_len_init(struct sk_buff *skb) 3741 { 3742 const struct skb_shared_info *shinfo = skb_shinfo(skb); 3743 3744 qdisc_skb_cb(skb)->pkt_len = skb->len; 3745 3746 /* To get more precise estimation of bytes sent on wire, 3747 * we add to pkt_len the headers size of all segments 3748 */ 3749 if (shinfo->gso_size && skb_transport_header_was_set(skb)) { 3750 u16 gso_segs = shinfo->gso_segs; 3751 unsigned int hdr_len; 3752 3753 /* mac layer + network layer */ 3754 hdr_len = skb_transport_offset(skb); 3755 3756 /* + transport layer */ 3757 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 3758 const struct tcphdr *th; 3759 struct tcphdr _tcphdr; 3760 3761 th = skb_header_pointer(skb, hdr_len, 3762 sizeof(_tcphdr), &_tcphdr); 3763 if (likely(th)) 3764 hdr_len += __tcp_hdrlen(th); 3765 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 3766 struct udphdr _udphdr; 3767 3768 if (skb_header_pointer(skb, hdr_len, 3769 sizeof(_udphdr), &_udphdr)) 3770 hdr_len += sizeof(struct udphdr); 3771 } 3772 3773 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 3774 int payload = skb->len - hdr_len; 3775 3776 /* Malicious packet. */ 3777 if (payload <= 0) 3778 return; 3779 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 3780 } 3781 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 3782 } 3783 } 3784 3785 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 3786 struct sk_buff **to_free, 3787 struct netdev_queue *txq) 3788 { 3789 int rc; 3790 3791 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 3792 if (rc == NET_XMIT_SUCCESS) 3793 trace_qdisc_enqueue(q, txq, skb); 3794 return rc; 3795 } 3796 3797 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 3798 struct net_device *dev, 3799 struct netdev_queue *txq) 3800 { 3801 spinlock_t *root_lock = qdisc_lock(q); 3802 struct sk_buff *to_free = NULL; 3803 bool contended; 3804 int rc; 3805 3806 qdisc_calculate_pkt_len(skb, q); 3807 3808 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 3809 3810 if (q->flags & TCQ_F_NOLOCK) { 3811 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 3812 qdisc_run_begin(q)) { 3813 /* Retest nolock_qdisc_is_empty() within the protection 3814 * of q->seqlock to protect from racing with requeuing. 3815 */ 3816 if (unlikely(!nolock_qdisc_is_empty(q))) { 3817 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3818 __qdisc_run(q); 3819 qdisc_run_end(q); 3820 3821 goto no_lock_out; 3822 } 3823 3824 qdisc_bstats_cpu_update(q, skb); 3825 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 3826 !nolock_qdisc_is_empty(q)) 3827 __qdisc_run(q); 3828 3829 qdisc_run_end(q); 3830 return NET_XMIT_SUCCESS; 3831 } 3832 3833 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3834 qdisc_run(q); 3835 3836 no_lock_out: 3837 if (unlikely(to_free)) 3838 kfree_skb_list_reason(to_free, 3839 tcf_get_drop_reason(to_free)); 3840 return rc; 3841 } 3842 3843 if (unlikely(READ_ONCE(q->owner) == smp_processor_id())) { 3844 kfree_skb_reason(skb, SKB_DROP_REASON_TC_RECLASSIFY_LOOP); 3845 return NET_XMIT_DROP; 3846 } 3847 /* 3848 * Heuristic to force contended enqueues to serialize on a 3849 * separate lock before trying to get qdisc main lock. 3850 * This permits qdisc->running owner to get the lock more 3851 * often and dequeue packets faster. 3852 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 3853 * and then other tasks will only enqueue packets. The packets will be 3854 * sent after the qdisc owner is scheduled again. To prevent this 3855 * scenario the task always serialize on the lock. 3856 */ 3857 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 3858 if (unlikely(contended)) 3859 spin_lock(&q->busylock); 3860 3861 spin_lock(root_lock); 3862 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 3863 __qdisc_drop(skb, &to_free); 3864 rc = NET_XMIT_DROP; 3865 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 3866 qdisc_run_begin(q)) { 3867 /* 3868 * This is a work-conserving queue; there are no old skbs 3869 * waiting to be sent out; and the qdisc is not running - 3870 * xmit the skb directly. 3871 */ 3872 3873 qdisc_bstats_update(q, skb); 3874 3875 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 3876 if (unlikely(contended)) { 3877 spin_unlock(&q->busylock); 3878 contended = false; 3879 } 3880 __qdisc_run(q); 3881 } 3882 3883 qdisc_run_end(q); 3884 rc = NET_XMIT_SUCCESS; 3885 } else { 3886 WRITE_ONCE(q->owner, smp_processor_id()); 3887 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3888 WRITE_ONCE(q->owner, -1); 3889 if (qdisc_run_begin(q)) { 3890 if (unlikely(contended)) { 3891 spin_unlock(&q->busylock); 3892 contended = false; 3893 } 3894 __qdisc_run(q); 3895 qdisc_run_end(q); 3896 } 3897 } 3898 spin_unlock(root_lock); 3899 if (unlikely(to_free)) 3900 kfree_skb_list_reason(to_free, 3901 tcf_get_drop_reason(to_free)); 3902 if (unlikely(contended)) 3903 spin_unlock(&q->busylock); 3904 return rc; 3905 } 3906 3907 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 3908 static void skb_update_prio(struct sk_buff *skb) 3909 { 3910 const struct netprio_map *map; 3911 const struct sock *sk; 3912 unsigned int prioidx; 3913 3914 if (skb->priority) 3915 return; 3916 map = rcu_dereference_bh(skb->dev->priomap); 3917 if (!map) 3918 return; 3919 sk = skb_to_full_sk(skb); 3920 if (!sk) 3921 return; 3922 3923 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 3924 3925 if (prioidx < map->priomap_len) 3926 skb->priority = map->priomap[prioidx]; 3927 } 3928 #else 3929 #define skb_update_prio(skb) 3930 #endif 3931 3932 /** 3933 * dev_loopback_xmit - loop back @skb 3934 * @net: network namespace this loopback is happening in 3935 * @sk: sk needed to be a netfilter okfn 3936 * @skb: buffer to transmit 3937 */ 3938 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 3939 { 3940 skb_reset_mac_header(skb); 3941 __skb_pull(skb, skb_network_offset(skb)); 3942 skb->pkt_type = PACKET_LOOPBACK; 3943 if (skb->ip_summed == CHECKSUM_NONE) 3944 skb->ip_summed = CHECKSUM_UNNECESSARY; 3945 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 3946 skb_dst_force(skb); 3947 netif_rx(skb); 3948 return 0; 3949 } 3950 EXPORT_SYMBOL(dev_loopback_xmit); 3951 3952 #ifdef CONFIG_NET_EGRESS 3953 static struct netdev_queue * 3954 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 3955 { 3956 int qm = skb_get_queue_mapping(skb); 3957 3958 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 3959 } 3960 3961 #ifndef CONFIG_PREEMPT_RT 3962 static bool netdev_xmit_txqueue_skipped(void) 3963 { 3964 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 3965 } 3966 3967 void netdev_xmit_skip_txqueue(bool skip) 3968 { 3969 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 3970 } 3971 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 3972 3973 #else 3974 static bool netdev_xmit_txqueue_skipped(void) 3975 { 3976 return current->net_xmit.skip_txqueue; 3977 } 3978 3979 void netdev_xmit_skip_txqueue(bool skip) 3980 { 3981 current->net_xmit.skip_txqueue = skip; 3982 } 3983 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 3984 #endif 3985 #endif /* CONFIG_NET_EGRESS */ 3986 3987 #ifdef CONFIG_NET_XGRESS 3988 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 3989 enum skb_drop_reason *drop_reason) 3990 { 3991 int ret = TC_ACT_UNSPEC; 3992 #ifdef CONFIG_NET_CLS_ACT 3993 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 3994 struct tcf_result res; 3995 3996 if (!miniq) 3997 return ret; 3998 3999 if (static_branch_unlikely(&tcf_bypass_check_needed_key)) { 4000 if (tcf_block_bypass_sw(miniq->block)) 4001 return ret; 4002 } 4003 4004 tc_skb_cb(skb)->mru = 0; 4005 tc_skb_cb(skb)->post_ct = false; 4006 tcf_set_drop_reason(skb, *drop_reason); 4007 4008 mini_qdisc_bstats_cpu_update(miniq, skb); 4009 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4010 /* Only tcf related quirks below. */ 4011 switch (ret) { 4012 case TC_ACT_SHOT: 4013 *drop_reason = tcf_get_drop_reason(skb); 4014 mini_qdisc_qstats_cpu_drop(miniq); 4015 break; 4016 case TC_ACT_OK: 4017 case TC_ACT_RECLASSIFY: 4018 skb->tc_index = TC_H_MIN(res.classid); 4019 break; 4020 } 4021 #endif /* CONFIG_NET_CLS_ACT */ 4022 return ret; 4023 } 4024 4025 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4026 4027 void tcx_inc(void) 4028 { 4029 static_branch_inc(&tcx_needed_key); 4030 } 4031 4032 void tcx_dec(void) 4033 { 4034 static_branch_dec(&tcx_needed_key); 4035 } 4036 4037 static __always_inline enum tcx_action_base 4038 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4039 const bool needs_mac) 4040 { 4041 const struct bpf_mprog_fp *fp; 4042 const struct bpf_prog *prog; 4043 int ret = TCX_NEXT; 4044 4045 if (needs_mac) 4046 __skb_push(skb, skb->mac_len); 4047 bpf_mprog_foreach_prog(entry, fp, prog) { 4048 bpf_compute_data_pointers(skb); 4049 ret = bpf_prog_run(prog, skb); 4050 if (ret != TCX_NEXT) 4051 break; 4052 } 4053 if (needs_mac) 4054 __skb_pull(skb, skb->mac_len); 4055 return tcx_action_code(skb, ret); 4056 } 4057 4058 static __always_inline struct sk_buff * 4059 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4060 struct net_device *orig_dev, bool *another) 4061 { 4062 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4063 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4064 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4065 int sch_ret; 4066 4067 if (!entry) 4068 return skb; 4069 4070 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4071 if (*pt_prev) { 4072 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4073 *pt_prev = NULL; 4074 } 4075 4076 qdisc_skb_cb(skb)->pkt_len = skb->len; 4077 tcx_set_ingress(skb, true); 4078 4079 if (static_branch_unlikely(&tcx_needed_key)) { 4080 sch_ret = tcx_run(entry, skb, true); 4081 if (sch_ret != TC_ACT_UNSPEC) 4082 goto ingress_verdict; 4083 } 4084 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4085 ingress_verdict: 4086 switch (sch_ret) { 4087 case TC_ACT_REDIRECT: 4088 /* skb_mac_header check was done by BPF, so we can safely 4089 * push the L2 header back before redirecting to another 4090 * netdev. 4091 */ 4092 __skb_push(skb, skb->mac_len); 4093 if (skb_do_redirect(skb) == -EAGAIN) { 4094 __skb_pull(skb, skb->mac_len); 4095 *another = true; 4096 break; 4097 } 4098 *ret = NET_RX_SUCCESS; 4099 bpf_net_ctx_clear(bpf_net_ctx); 4100 return NULL; 4101 case TC_ACT_SHOT: 4102 kfree_skb_reason(skb, drop_reason); 4103 *ret = NET_RX_DROP; 4104 bpf_net_ctx_clear(bpf_net_ctx); 4105 return NULL; 4106 /* used by tc_run */ 4107 case TC_ACT_STOLEN: 4108 case TC_ACT_QUEUED: 4109 case TC_ACT_TRAP: 4110 consume_skb(skb); 4111 fallthrough; 4112 case TC_ACT_CONSUMED: 4113 *ret = NET_RX_SUCCESS; 4114 bpf_net_ctx_clear(bpf_net_ctx); 4115 return NULL; 4116 } 4117 bpf_net_ctx_clear(bpf_net_ctx); 4118 4119 return skb; 4120 } 4121 4122 static __always_inline struct sk_buff * 4123 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4124 { 4125 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4126 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4127 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4128 int sch_ret; 4129 4130 if (!entry) 4131 return skb; 4132 4133 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4134 4135 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4136 * already set by the caller. 4137 */ 4138 if (static_branch_unlikely(&tcx_needed_key)) { 4139 sch_ret = tcx_run(entry, skb, false); 4140 if (sch_ret != TC_ACT_UNSPEC) 4141 goto egress_verdict; 4142 } 4143 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4144 egress_verdict: 4145 switch (sch_ret) { 4146 case TC_ACT_REDIRECT: 4147 /* No need to push/pop skb's mac_header here on egress! */ 4148 skb_do_redirect(skb); 4149 *ret = NET_XMIT_SUCCESS; 4150 bpf_net_ctx_clear(bpf_net_ctx); 4151 return NULL; 4152 case TC_ACT_SHOT: 4153 kfree_skb_reason(skb, drop_reason); 4154 *ret = NET_XMIT_DROP; 4155 bpf_net_ctx_clear(bpf_net_ctx); 4156 return NULL; 4157 /* used by tc_run */ 4158 case TC_ACT_STOLEN: 4159 case TC_ACT_QUEUED: 4160 case TC_ACT_TRAP: 4161 consume_skb(skb); 4162 fallthrough; 4163 case TC_ACT_CONSUMED: 4164 *ret = NET_XMIT_SUCCESS; 4165 bpf_net_ctx_clear(bpf_net_ctx); 4166 return NULL; 4167 } 4168 bpf_net_ctx_clear(bpf_net_ctx); 4169 4170 return skb; 4171 } 4172 #else 4173 static __always_inline struct sk_buff * 4174 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4175 struct net_device *orig_dev, bool *another) 4176 { 4177 return skb; 4178 } 4179 4180 static __always_inline struct sk_buff * 4181 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4182 { 4183 return skb; 4184 } 4185 #endif /* CONFIG_NET_XGRESS */ 4186 4187 #ifdef CONFIG_XPS 4188 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4189 struct xps_dev_maps *dev_maps, unsigned int tci) 4190 { 4191 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4192 struct xps_map *map; 4193 int queue_index = -1; 4194 4195 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4196 return queue_index; 4197 4198 tci *= dev_maps->num_tc; 4199 tci += tc; 4200 4201 map = rcu_dereference(dev_maps->attr_map[tci]); 4202 if (map) { 4203 if (map->len == 1) 4204 queue_index = map->queues[0]; 4205 else 4206 queue_index = map->queues[reciprocal_scale( 4207 skb_get_hash(skb), map->len)]; 4208 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4209 queue_index = -1; 4210 } 4211 return queue_index; 4212 } 4213 #endif 4214 4215 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4216 struct sk_buff *skb) 4217 { 4218 #ifdef CONFIG_XPS 4219 struct xps_dev_maps *dev_maps; 4220 struct sock *sk = skb->sk; 4221 int queue_index = -1; 4222 4223 if (!static_key_false(&xps_needed)) 4224 return -1; 4225 4226 rcu_read_lock(); 4227 if (!static_key_false(&xps_rxqs_needed)) 4228 goto get_cpus_map; 4229 4230 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4231 if (dev_maps) { 4232 int tci = sk_rx_queue_get(sk); 4233 4234 if (tci >= 0) 4235 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4236 tci); 4237 } 4238 4239 get_cpus_map: 4240 if (queue_index < 0) { 4241 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4242 if (dev_maps) { 4243 unsigned int tci = skb->sender_cpu - 1; 4244 4245 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4246 tci); 4247 } 4248 } 4249 rcu_read_unlock(); 4250 4251 return queue_index; 4252 #else 4253 return -1; 4254 #endif 4255 } 4256 4257 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4258 struct net_device *sb_dev) 4259 { 4260 return 0; 4261 } 4262 EXPORT_SYMBOL(dev_pick_tx_zero); 4263 4264 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4265 struct net_device *sb_dev) 4266 { 4267 struct sock *sk = skb->sk; 4268 int queue_index = sk_tx_queue_get(sk); 4269 4270 sb_dev = sb_dev ? : dev; 4271 4272 if (queue_index < 0 || skb->ooo_okay || 4273 queue_index >= dev->real_num_tx_queues) { 4274 int new_index = get_xps_queue(dev, sb_dev, skb); 4275 4276 if (new_index < 0) 4277 new_index = skb_tx_hash(dev, sb_dev, skb); 4278 4279 if (queue_index != new_index && sk && 4280 sk_fullsock(sk) && 4281 rcu_access_pointer(sk->sk_dst_cache)) 4282 sk_tx_queue_set(sk, new_index); 4283 4284 queue_index = new_index; 4285 } 4286 4287 return queue_index; 4288 } 4289 EXPORT_SYMBOL(netdev_pick_tx); 4290 4291 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4292 struct sk_buff *skb, 4293 struct net_device *sb_dev) 4294 { 4295 int queue_index = 0; 4296 4297 #ifdef CONFIG_XPS 4298 u32 sender_cpu = skb->sender_cpu - 1; 4299 4300 if (sender_cpu >= (u32)NR_CPUS) 4301 skb->sender_cpu = raw_smp_processor_id() + 1; 4302 #endif 4303 4304 if (dev->real_num_tx_queues != 1) { 4305 const struct net_device_ops *ops = dev->netdev_ops; 4306 4307 if (ops->ndo_select_queue) 4308 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4309 else 4310 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4311 4312 queue_index = netdev_cap_txqueue(dev, queue_index); 4313 } 4314 4315 skb_set_queue_mapping(skb, queue_index); 4316 return netdev_get_tx_queue(dev, queue_index); 4317 } 4318 4319 /** 4320 * __dev_queue_xmit() - transmit a buffer 4321 * @skb: buffer to transmit 4322 * @sb_dev: suboordinate device used for L2 forwarding offload 4323 * 4324 * Queue a buffer for transmission to a network device. The caller must 4325 * have set the device and priority and built the buffer before calling 4326 * this function. The function can be called from an interrupt. 4327 * 4328 * When calling this method, interrupts MUST be enabled. This is because 4329 * the BH enable code must have IRQs enabled so that it will not deadlock. 4330 * 4331 * Regardless of the return value, the skb is consumed, so it is currently 4332 * difficult to retry a send to this method. (You can bump the ref count 4333 * before sending to hold a reference for retry if you are careful.) 4334 * 4335 * Return: 4336 * * 0 - buffer successfully transmitted 4337 * * positive qdisc return code - NET_XMIT_DROP etc. 4338 * * negative errno - other errors 4339 */ 4340 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4341 { 4342 struct net_device *dev = skb->dev; 4343 struct netdev_queue *txq = NULL; 4344 struct Qdisc *q; 4345 int rc = -ENOMEM; 4346 bool again = false; 4347 4348 skb_reset_mac_header(skb); 4349 skb_assert_len(skb); 4350 4351 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP)) 4352 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4353 4354 /* Disable soft irqs for various locks below. Also 4355 * stops preemption for RCU. 4356 */ 4357 rcu_read_lock_bh(); 4358 4359 skb_update_prio(skb); 4360 4361 qdisc_pkt_len_init(skb); 4362 tcx_set_ingress(skb, false); 4363 #ifdef CONFIG_NET_EGRESS 4364 if (static_branch_unlikely(&egress_needed_key)) { 4365 if (nf_hook_egress_active()) { 4366 skb = nf_hook_egress(skb, &rc, dev); 4367 if (!skb) 4368 goto out; 4369 } 4370 4371 netdev_xmit_skip_txqueue(false); 4372 4373 nf_skip_egress(skb, true); 4374 skb = sch_handle_egress(skb, &rc, dev); 4375 if (!skb) 4376 goto out; 4377 nf_skip_egress(skb, false); 4378 4379 if (netdev_xmit_txqueue_skipped()) 4380 txq = netdev_tx_queue_mapping(dev, skb); 4381 } 4382 #endif 4383 /* If device/qdisc don't need skb->dst, release it right now while 4384 * its hot in this cpu cache. 4385 */ 4386 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4387 skb_dst_drop(skb); 4388 else 4389 skb_dst_force(skb); 4390 4391 if (!txq) 4392 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4393 4394 q = rcu_dereference_bh(txq->qdisc); 4395 4396 trace_net_dev_queue(skb); 4397 if (q->enqueue) { 4398 rc = __dev_xmit_skb(skb, q, dev, txq); 4399 goto out; 4400 } 4401 4402 /* The device has no queue. Common case for software devices: 4403 * loopback, all the sorts of tunnels... 4404 4405 * Really, it is unlikely that netif_tx_lock protection is necessary 4406 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4407 * counters.) 4408 * However, it is possible, that they rely on protection 4409 * made by us here. 4410 4411 * Check this and shot the lock. It is not prone from deadlocks. 4412 *Either shot noqueue qdisc, it is even simpler 8) 4413 */ 4414 if (dev->flags & IFF_UP) { 4415 int cpu = smp_processor_id(); /* ok because BHs are off */ 4416 4417 /* Other cpus might concurrently change txq->xmit_lock_owner 4418 * to -1 or to their cpu id, but not to our id. 4419 */ 4420 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4421 if (dev_xmit_recursion()) 4422 goto recursion_alert; 4423 4424 skb = validate_xmit_skb(skb, dev, &again); 4425 if (!skb) 4426 goto out; 4427 4428 HARD_TX_LOCK(dev, txq, cpu); 4429 4430 if (!netif_xmit_stopped(txq)) { 4431 dev_xmit_recursion_inc(); 4432 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4433 dev_xmit_recursion_dec(); 4434 if (dev_xmit_complete(rc)) { 4435 HARD_TX_UNLOCK(dev, txq); 4436 goto out; 4437 } 4438 } 4439 HARD_TX_UNLOCK(dev, txq); 4440 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4441 dev->name); 4442 } else { 4443 /* Recursion is detected! It is possible, 4444 * unfortunately 4445 */ 4446 recursion_alert: 4447 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4448 dev->name); 4449 } 4450 } 4451 4452 rc = -ENETDOWN; 4453 rcu_read_unlock_bh(); 4454 4455 dev_core_stats_tx_dropped_inc(dev); 4456 kfree_skb_list(skb); 4457 return rc; 4458 out: 4459 rcu_read_unlock_bh(); 4460 return rc; 4461 } 4462 EXPORT_SYMBOL(__dev_queue_xmit); 4463 4464 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4465 { 4466 struct net_device *dev = skb->dev; 4467 struct sk_buff *orig_skb = skb; 4468 struct netdev_queue *txq; 4469 int ret = NETDEV_TX_BUSY; 4470 bool again = false; 4471 4472 if (unlikely(!netif_running(dev) || 4473 !netif_carrier_ok(dev))) 4474 goto drop; 4475 4476 skb = validate_xmit_skb_list(skb, dev, &again); 4477 if (skb != orig_skb) 4478 goto drop; 4479 4480 skb_set_queue_mapping(skb, queue_id); 4481 txq = skb_get_tx_queue(dev, skb); 4482 4483 local_bh_disable(); 4484 4485 dev_xmit_recursion_inc(); 4486 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4487 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4488 ret = netdev_start_xmit(skb, dev, txq, false); 4489 HARD_TX_UNLOCK(dev, txq); 4490 dev_xmit_recursion_dec(); 4491 4492 local_bh_enable(); 4493 return ret; 4494 drop: 4495 dev_core_stats_tx_dropped_inc(dev); 4496 kfree_skb_list(skb); 4497 return NET_XMIT_DROP; 4498 } 4499 EXPORT_SYMBOL(__dev_direct_xmit); 4500 4501 /************************************************************************* 4502 * Receiver routines 4503 *************************************************************************/ 4504 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4505 4506 int weight_p __read_mostly = 64; /* old backlog weight */ 4507 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4508 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4509 4510 /* Called with irq disabled */ 4511 static inline void ____napi_schedule(struct softnet_data *sd, 4512 struct napi_struct *napi) 4513 { 4514 struct task_struct *thread; 4515 4516 lockdep_assert_irqs_disabled(); 4517 4518 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4519 /* Paired with smp_mb__before_atomic() in 4520 * napi_enable()/dev_set_threaded(). 4521 * Use READ_ONCE() to guarantee a complete 4522 * read on napi->thread. Only call 4523 * wake_up_process() when it's not NULL. 4524 */ 4525 thread = READ_ONCE(napi->thread); 4526 if (thread) { 4527 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4528 goto use_local_napi; 4529 4530 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4531 wake_up_process(thread); 4532 return; 4533 } 4534 } 4535 4536 use_local_napi: 4537 list_add_tail(&napi->poll_list, &sd->poll_list); 4538 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4539 /* If not called from net_rx_action() 4540 * we have to raise NET_RX_SOFTIRQ. 4541 */ 4542 if (!sd->in_net_rx_action) 4543 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4544 } 4545 4546 #ifdef CONFIG_RPS 4547 4548 struct static_key_false rps_needed __read_mostly; 4549 EXPORT_SYMBOL(rps_needed); 4550 struct static_key_false rfs_needed __read_mostly; 4551 EXPORT_SYMBOL(rfs_needed); 4552 4553 static struct rps_dev_flow * 4554 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4555 struct rps_dev_flow *rflow, u16 next_cpu) 4556 { 4557 if (next_cpu < nr_cpu_ids) { 4558 u32 head; 4559 #ifdef CONFIG_RFS_ACCEL 4560 struct netdev_rx_queue *rxqueue; 4561 struct rps_dev_flow_table *flow_table; 4562 struct rps_dev_flow *old_rflow; 4563 u16 rxq_index; 4564 u32 flow_id; 4565 int rc; 4566 4567 /* Should we steer this flow to a different hardware queue? */ 4568 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4569 !(dev->features & NETIF_F_NTUPLE)) 4570 goto out; 4571 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4572 if (rxq_index == skb_get_rx_queue(skb)) 4573 goto out; 4574 4575 rxqueue = dev->_rx + rxq_index; 4576 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4577 if (!flow_table) 4578 goto out; 4579 flow_id = skb_get_hash(skb) & flow_table->mask; 4580 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4581 rxq_index, flow_id); 4582 if (rc < 0) 4583 goto out; 4584 old_rflow = rflow; 4585 rflow = &flow_table->flows[flow_id]; 4586 WRITE_ONCE(rflow->filter, rc); 4587 if (old_rflow->filter == rc) 4588 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 4589 out: 4590 #endif 4591 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 4592 rps_input_queue_tail_save(&rflow->last_qtail, head); 4593 } 4594 4595 WRITE_ONCE(rflow->cpu, next_cpu); 4596 return rflow; 4597 } 4598 4599 /* 4600 * get_rps_cpu is called from netif_receive_skb and returns the target 4601 * CPU from the RPS map of the receiving queue for a given skb. 4602 * rcu_read_lock must be held on entry. 4603 */ 4604 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4605 struct rps_dev_flow **rflowp) 4606 { 4607 const struct rps_sock_flow_table *sock_flow_table; 4608 struct netdev_rx_queue *rxqueue = dev->_rx; 4609 struct rps_dev_flow_table *flow_table; 4610 struct rps_map *map; 4611 int cpu = -1; 4612 u32 tcpu; 4613 u32 hash; 4614 4615 if (skb_rx_queue_recorded(skb)) { 4616 u16 index = skb_get_rx_queue(skb); 4617 4618 if (unlikely(index >= dev->real_num_rx_queues)) { 4619 WARN_ONCE(dev->real_num_rx_queues > 1, 4620 "%s received packet on queue %u, but number " 4621 "of RX queues is %u\n", 4622 dev->name, index, dev->real_num_rx_queues); 4623 goto done; 4624 } 4625 rxqueue += index; 4626 } 4627 4628 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4629 4630 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4631 map = rcu_dereference(rxqueue->rps_map); 4632 if (!flow_table && !map) 4633 goto done; 4634 4635 skb_reset_network_header(skb); 4636 hash = skb_get_hash(skb); 4637 if (!hash) 4638 goto done; 4639 4640 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4641 if (flow_table && sock_flow_table) { 4642 struct rps_dev_flow *rflow; 4643 u32 next_cpu; 4644 u32 ident; 4645 4646 /* First check into global flow table if there is a match. 4647 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4648 */ 4649 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4650 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4651 goto try_rps; 4652 4653 next_cpu = ident & net_hotdata.rps_cpu_mask; 4654 4655 /* OK, now we know there is a match, 4656 * we can look at the local (per receive queue) flow table 4657 */ 4658 rflow = &flow_table->flows[hash & flow_table->mask]; 4659 tcpu = rflow->cpu; 4660 4661 /* 4662 * If the desired CPU (where last recvmsg was done) is 4663 * different from current CPU (one in the rx-queue flow 4664 * table entry), switch if one of the following holds: 4665 * - Current CPU is unset (>= nr_cpu_ids). 4666 * - Current CPU is offline. 4667 * - The current CPU's queue tail has advanced beyond the 4668 * last packet that was enqueued using this table entry. 4669 * This guarantees that all previous packets for the flow 4670 * have been dequeued, thus preserving in order delivery. 4671 */ 4672 if (unlikely(tcpu != next_cpu) && 4673 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4674 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 4675 rflow->last_qtail)) >= 0)) { 4676 tcpu = next_cpu; 4677 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4678 } 4679 4680 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4681 *rflowp = rflow; 4682 cpu = tcpu; 4683 goto done; 4684 } 4685 } 4686 4687 try_rps: 4688 4689 if (map) { 4690 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4691 if (cpu_online(tcpu)) { 4692 cpu = tcpu; 4693 goto done; 4694 } 4695 } 4696 4697 done: 4698 return cpu; 4699 } 4700 4701 #ifdef CONFIG_RFS_ACCEL 4702 4703 /** 4704 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4705 * @dev: Device on which the filter was set 4706 * @rxq_index: RX queue index 4707 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4708 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4709 * 4710 * Drivers that implement ndo_rx_flow_steer() should periodically call 4711 * this function for each installed filter and remove the filters for 4712 * which it returns %true. 4713 */ 4714 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4715 u32 flow_id, u16 filter_id) 4716 { 4717 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4718 struct rps_dev_flow_table *flow_table; 4719 struct rps_dev_flow *rflow; 4720 bool expire = true; 4721 unsigned int cpu; 4722 4723 rcu_read_lock(); 4724 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4725 if (flow_table && flow_id <= flow_table->mask) { 4726 rflow = &flow_table->flows[flow_id]; 4727 cpu = READ_ONCE(rflow->cpu); 4728 if (READ_ONCE(rflow->filter) == filter_id && cpu < nr_cpu_ids && 4729 ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - 4730 READ_ONCE(rflow->last_qtail)) < 4731 (int)(10 * flow_table->mask))) 4732 expire = false; 4733 } 4734 rcu_read_unlock(); 4735 return expire; 4736 } 4737 EXPORT_SYMBOL(rps_may_expire_flow); 4738 4739 #endif /* CONFIG_RFS_ACCEL */ 4740 4741 /* Called from hardirq (IPI) context */ 4742 static void rps_trigger_softirq(void *data) 4743 { 4744 struct softnet_data *sd = data; 4745 4746 ____napi_schedule(sd, &sd->backlog); 4747 sd->received_rps++; 4748 } 4749 4750 #endif /* CONFIG_RPS */ 4751 4752 /* Called from hardirq (IPI) context */ 4753 static void trigger_rx_softirq(void *data) 4754 { 4755 struct softnet_data *sd = data; 4756 4757 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4758 smp_store_release(&sd->defer_ipi_scheduled, 0); 4759 } 4760 4761 /* 4762 * After we queued a packet into sd->input_pkt_queue, 4763 * we need to make sure this queue is serviced soon. 4764 * 4765 * - If this is another cpu queue, link it to our rps_ipi_list, 4766 * and make sure we will process rps_ipi_list from net_rx_action(). 4767 * 4768 * - If this is our own queue, NAPI schedule our backlog. 4769 * Note that this also raises NET_RX_SOFTIRQ. 4770 */ 4771 static void napi_schedule_rps(struct softnet_data *sd) 4772 { 4773 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 4774 4775 #ifdef CONFIG_RPS 4776 if (sd != mysd) { 4777 if (use_backlog_threads()) { 4778 __napi_schedule_irqoff(&sd->backlog); 4779 return; 4780 } 4781 4782 sd->rps_ipi_next = mysd->rps_ipi_list; 4783 mysd->rps_ipi_list = sd; 4784 4785 /* If not called from net_rx_action() or napi_threaded_poll() 4786 * we have to raise NET_RX_SOFTIRQ. 4787 */ 4788 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 4789 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4790 return; 4791 } 4792 #endif /* CONFIG_RPS */ 4793 __napi_schedule_irqoff(&mysd->backlog); 4794 } 4795 4796 void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu) 4797 { 4798 unsigned long flags; 4799 4800 if (use_backlog_threads()) { 4801 backlog_lock_irq_save(sd, &flags); 4802 4803 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 4804 __napi_schedule_irqoff(&sd->backlog); 4805 4806 backlog_unlock_irq_restore(sd, &flags); 4807 4808 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 4809 smp_call_function_single_async(cpu, &sd->defer_csd); 4810 } 4811 } 4812 4813 #ifdef CONFIG_NET_FLOW_LIMIT 4814 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 4815 #endif 4816 4817 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 4818 { 4819 #ifdef CONFIG_NET_FLOW_LIMIT 4820 struct sd_flow_limit *fl; 4821 struct softnet_data *sd; 4822 unsigned int old_flow, new_flow; 4823 4824 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 4825 return false; 4826 4827 sd = this_cpu_ptr(&softnet_data); 4828 4829 rcu_read_lock(); 4830 fl = rcu_dereference(sd->flow_limit); 4831 if (fl) { 4832 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1); 4833 old_flow = fl->history[fl->history_head]; 4834 fl->history[fl->history_head] = new_flow; 4835 4836 fl->history_head++; 4837 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 4838 4839 if (likely(fl->buckets[old_flow])) 4840 fl->buckets[old_flow]--; 4841 4842 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 4843 fl->count++; 4844 rcu_read_unlock(); 4845 return true; 4846 } 4847 } 4848 rcu_read_unlock(); 4849 #endif 4850 return false; 4851 } 4852 4853 /* 4854 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 4855 * queue (may be a remote CPU queue). 4856 */ 4857 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 4858 unsigned int *qtail) 4859 { 4860 enum skb_drop_reason reason; 4861 struct softnet_data *sd; 4862 unsigned long flags; 4863 unsigned int qlen; 4864 int max_backlog; 4865 u32 tail; 4866 4867 reason = SKB_DROP_REASON_DEV_READY; 4868 if (!netif_running(skb->dev)) 4869 goto bad_dev; 4870 4871 reason = SKB_DROP_REASON_CPU_BACKLOG; 4872 sd = &per_cpu(softnet_data, cpu); 4873 4874 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 4875 max_backlog = READ_ONCE(net_hotdata.max_backlog); 4876 if (unlikely(qlen > max_backlog)) 4877 goto cpu_backlog_drop; 4878 backlog_lock_irq_save(sd, &flags); 4879 qlen = skb_queue_len(&sd->input_pkt_queue); 4880 if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) { 4881 if (!qlen) { 4882 /* Schedule NAPI for backlog device. We can use 4883 * non atomic operation as we own the queue lock. 4884 */ 4885 if (!__test_and_set_bit(NAPI_STATE_SCHED, 4886 &sd->backlog.state)) 4887 napi_schedule_rps(sd); 4888 } 4889 __skb_queue_tail(&sd->input_pkt_queue, skb); 4890 tail = rps_input_queue_tail_incr(sd); 4891 backlog_unlock_irq_restore(sd, &flags); 4892 4893 /* save the tail outside of the critical section */ 4894 rps_input_queue_tail_save(qtail, tail); 4895 return NET_RX_SUCCESS; 4896 } 4897 4898 backlog_unlock_irq_restore(sd, &flags); 4899 4900 cpu_backlog_drop: 4901 atomic_inc(&sd->dropped); 4902 bad_dev: 4903 dev_core_stats_rx_dropped_inc(skb->dev); 4904 kfree_skb_reason(skb, reason); 4905 return NET_RX_DROP; 4906 } 4907 4908 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 4909 { 4910 struct net_device *dev = skb->dev; 4911 struct netdev_rx_queue *rxqueue; 4912 4913 rxqueue = dev->_rx; 4914 4915 if (skb_rx_queue_recorded(skb)) { 4916 u16 index = skb_get_rx_queue(skb); 4917 4918 if (unlikely(index >= dev->real_num_rx_queues)) { 4919 WARN_ONCE(dev->real_num_rx_queues > 1, 4920 "%s received packet on queue %u, but number " 4921 "of RX queues is %u\n", 4922 dev->name, index, dev->real_num_rx_queues); 4923 4924 return rxqueue; /* Return first rxqueue */ 4925 } 4926 rxqueue += index; 4927 } 4928 return rxqueue; 4929 } 4930 4931 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 4932 struct bpf_prog *xdp_prog) 4933 { 4934 void *orig_data, *orig_data_end, *hard_start; 4935 struct netdev_rx_queue *rxqueue; 4936 bool orig_bcast, orig_host; 4937 u32 mac_len, frame_sz; 4938 __be16 orig_eth_type; 4939 struct ethhdr *eth; 4940 u32 metalen, act; 4941 int off; 4942 4943 /* The XDP program wants to see the packet starting at the MAC 4944 * header. 4945 */ 4946 mac_len = skb->data - skb_mac_header(skb); 4947 hard_start = skb->data - skb_headroom(skb); 4948 4949 /* SKB "head" area always have tailroom for skb_shared_info */ 4950 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 4951 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 4952 4953 rxqueue = netif_get_rxqueue(skb); 4954 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 4955 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 4956 skb_headlen(skb) + mac_len, true); 4957 if (skb_is_nonlinear(skb)) { 4958 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 4959 xdp_buff_set_frags_flag(xdp); 4960 } else { 4961 xdp_buff_clear_frags_flag(xdp); 4962 } 4963 4964 orig_data_end = xdp->data_end; 4965 orig_data = xdp->data; 4966 eth = (struct ethhdr *)xdp->data; 4967 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 4968 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 4969 orig_eth_type = eth->h_proto; 4970 4971 act = bpf_prog_run_xdp(xdp_prog, xdp); 4972 4973 /* check if bpf_xdp_adjust_head was used */ 4974 off = xdp->data - orig_data; 4975 if (off) { 4976 if (off > 0) 4977 __skb_pull(skb, off); 4978 else if (off < 0) 4979 __skb_push(skb, -off); 4980 4981 skb->mac_header += off; 4982 skb_reset_network_header(skb); 4983 } 4984 4985 /* check if bpf_xdp_adjust_tail was used */ 4986 off = xdp->data_end - orig_data_end; 4987 if (off != 0) { 4988 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 4989 skb->len += off; /* positive on grow, negative on shrink */ 4990 } 4991 4992 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 4993 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 4994 */ 4995 if (xdp_buff_has_frags(xdp)) 4996 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 4997 else 4998 skb->data_len = 0; 4999 5000 /* check if XDP changed eth hdr such SKB needs update */ 5001 eth = (struct ethhdr *)xdp->data; 5002 if ((orig_eth_type != eth->h_proto) || 5003 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5004 skb->dev->dev_addr)) || 5005 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5006 __skb_push(skb, ETH_HLEN); 5007 skb->pkt_type = PACKET_HOST; 5008 skb->protocol = eth_type_trans(skb, skb->dev); 5009 } 5010 5011 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5012 * before calling us again on redirect path. We do not call do_redirect 5013 * as we leave that up to the caller. 5014 * 5015 * Caller is responsible for managing lifetime of skb (i.e. calling 5016 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5017 */ 5018 switch (act) { 5019 case XDP_REDIRECT: 5020 case XDP_TX: 5021 __skb_push(skb, mac_len); 5022 break; 5023 case XDP_PASS: 5024 metalen = xdp->data - xdp->data_meta; 5025 if (metalen) 5026 skb_metadata_set(skb, metalen); 5027 break; 5028 } 5029 5030 return act; 5031 } 5032 5033 static int 5034 netif_skb_check_for_xdp(struct sk_buff **pskb, struct bpf_prog *prog) 5035 { 5036 struct sk_buff *skb = *pskb; 5037 int err, hroom, troom; 5038 5039 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 5040 return 0; 5041 5042 /* In case we have to go down the path and also linearize, 5043 * then lets do the pskb_expand_head() work just once here. 5044 */ 5045 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5046 troom = skb->tail + skb->data_len - skb->end; 5047 err = pskb_expand_head(skb, 5048 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5049 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5050 if (err) 5051 return err; 5052 5053 return skb_linearize(skb); 5054 } 5055 5056 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5057 struct xdp_buff *xdp, 5058 struct bpf_prog *xdp_prog) 5059 { 5060 struct sk_buff *skb = *pskb; 5061 u32 mac_len, act = XDP_DROP; 5062 5063 /* Reinjected packets coming from act_mirred or similar should 5064 * not get XDP generic processing. 5065 */ 5066 if (skb_is_redirected(skb)) 5067 return XDP_PASS; 5068 5069 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5070 * bytes. This is the guarantee that also native XDP provides, 5071 * thus we need to do it here as well. 5072 */ 5073 mac_len = skb->data - skb_mac_header(skb); 5074 __skb_push(skb, mac_len); 5075 5076 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5077 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5078 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5079 goto do_drop; 5080 } 5081 5082 __skb_pull(*pskb, mac_len); 5083 5084 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5085 switch (act) { 5086 case XDP_REDIRECT: 5087 case XDP_TX: 5088 case XDP_PASS: 5089 break; 5090 default: 5091 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5092 fallthrough; 5093 case XDP_ABORTED: 5094 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5095 fallthrough; 5096 case XDP_DROP: 5097 do_drop: 5098 kfree_skb(*pskb); 5099 break; 5100 } 5101 5102 return act; 5103 } 5104 5105 /* When doing generic XDP we have to bypass the qdisc layer and the 5106 * network taps in order to match in-driver-XDP behavior. This also means 5107 * that XDP packets are able to starve other packets going through a qdisc, 5108 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5109 * queues, so they do not have this starvation issue. 5110 */ 5111 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog) 5112 { 5113 struct net_device *dev = skb->dev; 5114 struct netdev_queue *txq; 5115 bool free_skb = true; 5116 int cpu, rc; 5117 5118 txq = netdev_core_pick_tx(dev, skb, NULL); 5119 cpu = smp_processor_id(); 5120 HARD_TX_LOCK(dev, txq, cpu); 5121 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5122 rc = netdev_start_xmit(skb, dev, txq, 0); 5123 if (dev_xmit_complete(rc)) 5124 free_skb = false; 5125 } 5126 HARD_TX_UNLOCK(dev, txq); 5127 if (free_skb) { 5128 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5129 dev_core_stats_tx_dropped_inc(dev); 5130 kfree_skb(skb); 5131 } 5132 } 5133 5134 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5135 5136 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5137 { 5138 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5139 5140 if (xdp_prog) { 5141 struct xdp_buff xdp; 5142 u32 act; 5143 int err; 5144 5145 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5146 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5147 if (act != XDP_PASS) { 5148 switch (act) { 5149 case XDP_REDIRECT: 5150 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5151 &xdp, xdp_prog); 5152 if (err) 5153 goto out_redir; 5154 break; 5155 case XDP_TX: 5156 generic_xdp_tx(*pskb, xdp_prog); 5157 break; 5158 } 5159 bpf_net_ctx_clear(bpf_net_ctx); 5160 return XDP_DROP; 5161 } 5162 bpf_net_ctx_clear(bpf_net_ctx); 5163 } 5164 return XDP_PASS; 5165 out_redir: 5166 bpf_net_ctx_clear(bpf_net_ctx); 5167 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5168 return XDP_DROP; 5169 } 5170 EXPORT_SYMBOL_GPL(do_xdp_generic); 5171 5172 static int netif_rx_internal(struct sk_buff *skb) 5173 { 5174 int ret; 5175 5176 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5177 5178 trace_netif_rx(skb); 5179 5180 #ifdef CONFIG_RPS 5181 if (static_branch_unlikely(&rps_needed)) { 5182 struct rps_dev_flow voidflow, *rflow = &voidflow; 5183 int cpu; 5184 5185 rcu_read_lock(); 5186 5187 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5188 if (cpu < 0) 5189 cpu = smp_processor_id(); 5190 5191 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5192 5193 rcu_read_unlock(); 5194 } else 5195 #endif 5196 { 5197 unsigned int qtail; 5198 5199 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5200 } 5201 return ret; 5202 } 5203 5204 /** 5205 * __netif_rx - Slightly optimized version of netif_rx 5206 * @skb: buffer to post 5207 * 5208 * This behaves as netif_rx except that it does not disable bottom halves. 5209 * As a result this function may only be invoked from the interrupt context 5210 * (either hard or soft interrupt). 5211 */ 5212 int __netif_rx(struct sk_buff *skb) 5213 { 5214 int ret; 5215 5216 lockdep_assert_once(hardirq_count() | softirq_count()); 5217 5218 trace_netif_rx_entry(skb); 5219 ret = netif_rx_internal(skb); 5220 trace_netif_rx_exit(ret); 5221 return ret; 5222 } 5223 EXPORT_SYMBOL(__netif_rx); 5224 5225 /** 5226 * netif_rx - post buffer to the network code 5227 * @skb: buffer to post 5228 * 5229 * This function receives a packet from a device driver and queues it for 5230 * the upper (protocol) levels to process via the backlog NAPI device. It 5231 * always succeeds. The buffer may be dropped during processing for 5232 * congestion control or by the protocol layers. 5233 * The network buffer is passed via the backlog NAPI device. Modern NIC 5234 * driver should use NAPI and GRO. 5235 * This function can used from interrupt and from process context. The 5236 * caller from process context must not disable interrupts before invoking 5237 * this function. 5238 * 5239 * return values: 5240 * NET_RX_SUCCESS (no congestion) 5241 * NET_RX_DROP (packet was dropped) 5242 * 5243 */ 5244 int netif_rx(struct sk_buff *skb) 5245 { 5246 bool need_bh_off = !(hardirq_count() | softirq_count()); 5247 int ret; 5248 5249 if (need_bh_off) 5250 local_bh_disable(); 5251 trace_netif_rx_entry(skb); 5252 ret = netif_rx_internal(skb); 5253 trace_netif_rx_exit(ret); 5254 if (need_bh_off) 5255 local_bh_enable(); 5256 return ret; 5257 } 5258 EXPORT_SYMBOL(netif_rx); 5259 5260 static __latent_entropy void net_tx_action(void) 5261 { 5262 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5263 5264 if (sd->completion_queue) { 5265 struct sk_buff *clist; 5266 5267 local_irq_disable(); 5268 clist = sd->completion_queue; 5269 sd->completion_queue = NULL; 5270 local_irq_enable(); 5271 5272 while (clist) { 5273 struct sk_buff *skb = clist; 5274 5275 clist = clist->next; 5276 5277 WARN_ON(refcount_read(&skb->users)); 5278 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5279 trace_consume_skb(skb, net_tx_action); 5280 else 5281 trace_kfree_skb(skb, net_tx_action, 5282 get_kfree_skb_cb(skb)->reason, NULL); 5283 5284 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5285 __kfree_skb(skb); 5286 else 5287 __napi_kfree_skb(skb, 5288 get_kfree_skb_cb(skb)->reason); 5289 } 5290 } 5291 5292 if (sd->output_queue) { 5293 struct Qdisc *head; 5294 5295 local_irq_disable(); 5296 head = sd->output_queue; 5297 sd->output_queue = NULL; 5298 sd->output_queue_tailp = &sd->output_queue; 5299 local_irq_enable(); 5300 5301 rcu_read_lock(); 5302 5303 while (head) { 5304 struct Qdisc *q = head; 5305 spinlock_t *root_lock = NULL; 5306 5307 head = head->next_sched; 5308 5309 /* We need to make sure head->next_sched is read 5310 * before clearing __QDISC_STATE_SCHED 5311 */ 5312 smp_mb__before_atomic(); 5313 5314 if (!(q->flags & TCQ_F_NOLOCK)) { 5315 root_lock = qdisc_lock(q); 5316 spin_lock(root_lock); 5317 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5318 &q->state))) { 5319 /* There is a synchronize_net() between 5320 * STATE_DEACTIVATED flag being set and 5321 * qdisc_reset()/some_qdisc_is_busy() in 5322 * dev_deactivate(), so we can safely bail out 5323 * early here to avoid data race between 5324 * qdisc_deactivate() and some_qdisc_is_busy() 5325 * for lockless qdisc. 5326 */ 5327 clear_bit(__QDISC_STATE_SCHED, &q->state); 5328 continue; 5329 } 5330 5331 clear_bit(__QDISC_STATE_SCHED, &q->state); 5332 qdisc_run(q); 5333 if (root_lock) 5334 spin_unlock(root_lock); 5335 } 5336 5337 rcu_read_unlock(); 5338 } 5339 5340 xfrm_dev_backlog(sd); 5341 } 5342 5343 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5344 /* This hook is defined here for ATM LANE */ 5345 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5346 unsigned char *addr) __read_mostly; 5347 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5348 #endif 5349 5350 /** 5351 * netdev_is_rx_handler_busy - check if receive handler is registered 5352 * @dev: device to check 5353 * 5354 * Check if a receive handler is already registered for a given device. 5355 * Return true if there one. 5356 * 5357 * The caller must hold the rtnl_mutex. 5358 */ 5359 bool netdev_is_rx_handler_busy(struct net_device *dev) 5360 { 5361 ASSERT_RTNL(); 5362 return dev && rtnl_dereference(dev->rx_handler); 5363 } 5364 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5365 5366 /** 5367 * netdev_rx_handler_register - register receive handler 5368 * @dev: device to register a handler for 5369 * @rx_handler: receive handler to register 5370 * @rx_handler_data: data pointer that is used by rx handler 5371 * 5372 * Register a receive handler for a device. This handler will then be 5373 * called from __netif_receive_skb. A negative errno code is returned 5374 * on a failure. 5375 * 5376 * The caller must hold the rtnl_mutex. 5377 * 5378 * For a general description of rx_handler, see enum rx_handler_result. 5379 */ 5380 int netdev_rx_handler_register(struct net_device *dev, 5381 rx_handler_func_t *rx_handler, 5382 void *rx_handler_data) 5383 { 5384 if (netdev_is_rx_handler_busy(dev)) 5385 return -EBUSY; 5386 5387 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5388 return -EINVAL; 5389 5390 /* Note: rx_handler_data must be set before rx_handler */ 5391 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5392 rcu_assign_pointer(dev->rx_handler, rx_handler); 5393 5394 return 0; 5395 } 5396 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5397 5398 /** 5399 * netdev_rx_handler_unregister - unregister receive handler 5400 * @dev: device to unregister a handler from 5401 * 5402 * Unregister a receive handler from a device. 5403 * 5404 * The caller must hold the rtnl_mutex. 5405 */ 5406 void netdev_rx_handler_unregister(struct net_device *dev) 5407 { 5408 5409 ASSERT_RTNL(); 5410 RCU_INIT_POINTER(dev->rx_handler, NULL); 5411 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5412 * section has a guarantee to see a non NULL rx_handler_data 5413 * as well. 5414 */ 5415 synchronize_net(); 5416 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5417 } 5418 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5419 5420 /* 5421 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5422 * the special handling of PFMEMALLOC skbs. 5423 */ 5424 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5425 { 5426 switch (skb->protocol) { 5427 case htons(ETH_P_ARP): 5428 case htons(ETH_P_IP): 5429 case htons(ETH_P_IPV6): 5430 case htons(ETH_P_8021Q): 5431 case htons(ETH_P_8021AD): 5432 return true; 5433 default: 5434 return false; 5435 } 5436 } 5437 5438 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5439 int *ret, struct net_device *orig_dev) 5440 { 5441 if (nf_hook_ingress_active(skb)) { 5442 int ingress_retval; 5443 5444 if (*pt_prev) { 5445 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5446 *pt_prev = NULL; 5447 } 5448 5449 rcu_read_lock(); 5450 ingress_retval = nf_hook_ingress(skb); 5451 rcu_read_unlock(); 5452 return ingress_retval; 5453 } 5454 return 0; 5455 } 5456 5457 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5458 struct packet_type **ppt_prev) 5459 { 5460 struct packet_type *ptype, *pt_prev; 5461 rx_handler_func_t *rx_handler; 5462 struct sk_buff *skb = *pskb; 5463 struct net_device *orig_dev; 5464 bool deliver_exact = false; 5465 int ret = NET_RX_DROP; 5466 __be16 type; 5467 5468 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5469 5470 trace_netif_receive_skb(skb); 5471 5472 orig_dev = skb->dev; 5473 5474 skb_reset_network_header(skb); 5475 if (!skb_transport_header_was_set(skb)) 5476 skb_reset_transport_header(skb); 5477 skb_reset_mac_len(skb); 5478 5479 pt_prev = NULL; 5480 5481 another_round: 5482 skb->skb_iif = skb->dev->ifindex; 5483 5484 __this_cpu_inc(softnet_data.processed); 5485 5486 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5487 int ret2; 5488 5489 migrate_disable(); 5490 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5491 &skb); 5492 migrate_enable(); 5493 5494 if (ret2 != XDP_PASS) { 5495 ret = NET_RX_DROP; 5496 goto out; 5497 } 5498 } 5499 5500 if (eth_type_vlan(skb->protocol)) { 5501 skb = skb_vlan_untag(skb); 5502 if (unlikely(!skb)) 5503 goto out; 5504 } 5505 5506 if (skb_skip_tc_classify(skb)) 5507 goto skip_classify; 5508 5509 if (pfmemalloc) 5510 goto skip_taps; 5511 5512 list_for_each_entry_rcu(ptype, &net_hotdata.ptype_all, list) { 5513 if (pt_prev) 5514 ret = deliver_skb(skb, pt_prev, orig_dev); 5515 pt_prev = ptype; 5516 } 5517 5518 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5519 if (pt_prev) 5520 ret = deliver_skb(skb, pt_prev, orig_dev); 5521 pt_prev = ptype; 5522 } 5523 5524 skip_taps: 5525 #ifdef CONFIG_NET_INGRESS 5526 if (static_branch_unlikely(&ingress_needed_key)) { 5527 bool another = false; 5528 5529 nf_skip_egress(skb, true); 5530 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5531 &another); 5532 if (another) 5533 goto another_round; 5534 if (!skb) 5535 goto out; 5536 5537 nf_skip_egress(skb, false); 5538 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5539 goto out; 5540 } 5541 #endif 5542 skb_reset_redirect(skb); 5543 skip_classify: 5544 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5545 goto drop; 5546 5547 if (skb_vlan_tag_present(skb)) { 5548 if (pt_prev) { 5549 ret = deliver_skb(skb, pt_prev, orig_dev); 5550 pt_prev = NULL; 5551 } 5552 if (vlan_do_receive(&skb)) 5553 goto another_round; 5554 else if (unlikely(!skb)) 5555 goto out; 5556 } 5557 5558 rx_handler = rcu_dereference(skb->dev->rx_handler); 5559 if (rx_handler) { 5560 if (pt_prev) { 5561 ret = deliver_skb(skb, pt_prev, orig_dev); 5562 pt_prev = NULL; 5563 } 5564 switch (rx_handler(&skb)) { 5565 case RX_HANDLER_CONSUMED: 5566 ret = NET_RX_SUCCESS; 5567 goto out; 5568 case RX_HANDLER_ANOTHER: 5569 goto another_round; 5570 case RX_HANDLER_EXACT: 5571 deliver_exact = true; 5572 break; 5573 case RX_HANDLER_PASS: 5574 break; 5575 default: 5576 BUG(); 5577 } 5578 } 5579 5580 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5581 check_vlan_id: 5582 if (skb_vlan_tag_get_id(skb)) { 5583 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5584 * find vlan device. 5585 */ 5586 skb->pkt_type = PACKET_OTHERHOST; 5587 } else if (eth_type_vlan(skb->protocol)) { 5588 /* Outer header is 802.1P with vlan 0, inner header is 5589 * 802.1Q or 802.1AD and vlan_do_receive() above could 5590 * not find vlan dev for vlan id 0. 5591 */ 5592 __vlan_hwaccel_clear_tag(skb); 5593 skb = skb_vlan_untag(skb); 5594 if (unlikely(!skb)) 5595 goto out; 5596 if (vlan_do_receive(&skb)) 5597 /* After stripping off 802.1P header with vlan 0 5598 * vlan dev is found for inner header. 5599 */ 5600 goto another_round; 5601 else if (unlikely(!skb)) 5602 goto out; 5603 else 5604 /* We have stripped outer 802.1P vlan 0 header. 5605 * But could not find vlan dev. 5606 * check again for vlan id to set OTHERHOST. 5607 */ 5608 goto check_vlan_id; 5609 } 5610 /* Note: we might in the future use prio bits 5611 * and set skb->priority like in vlan_do_receive() 5612 * For the time being, just ignore Priority Code Point 5613 */ 5614 __vlan_hwaccel_clear_tag(skb); 5615 } 5616 5617 type = skb->protocol; 5618 5619 /* deliver only exact match when indicated */ 5620 if (likely(!deliver_exact)) { 5621 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5622 &ptype_base[ntohs(type) & 5623 PTYPE_HASH_MASK]); 5624 } 5625 5626 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5627 &orig_dev->ptype_specific); 5628 5629 if (unlikely(skb->dev != orig_dev)) { 5630 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5631 &skb->dev->ptype_specific); 5632 } 5633 5634 if (pt_prev) { 5635 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5636 goto drop; 5637 *ppt_prev = pt_prev; 5638 } else { 5639 drop: 5640 if (!deliver_exact) 5641 dev_core_stats_rx_dropped_inc(skb->dev); 5642 else 5643 dev_core_stats_rx_nohandler_inc(skb->dev); 5644 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5645 /* Jamal, now you will not able to escape explaining 5646 * me how you were going to use this. :-) 5647 */ 5648 ret = NET_RX_DROP; 5649 } 5650 5651 out: 5652 /* The invariant here is that if *ppt_prev is not NULL 5653 * then skb should also be non-NULL. 5654 * 5655 * Apparently *ppt_prev assignment above holds this invariant due to 5656 * skb dereferencing near it. 5657 */ 5658 *pskb = skb; 5659 return ret; 5660 } 5661 5662 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5663 { 5664 struct net_device *orig_dev = skb->dev; 5665 struct packet_type *pt_prev = NULL; 5666 int ret; 5667 5668 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5669 if (pt_prev) 5670 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5671 skb->dev, pt_prev, orig_dev); 5672 return ret; 5673 } 5674 5675 /** 5676 * netif_receive_skb_core - special purpose version of netif_receive_skb 5677 * @skb: buffer to process 5678 * 5679 * More direct receive version of netif_receive_skb(). It should 5680 * only be used by callers that have a need to skip RPS and Generic XDP. 5681 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5682 * 5683 * This function may only be called from softirq context and interrupts 5684 * should be enabled. 5685 * 5686 * Return values (usually ignored): 5687 * NET_RX_SUCCESS: no congestion 5688 * NET_RX_DROP: packet was dropped 5689 */ 5690 int netif_receive_skb_core(struct sk_buff *skb) 5691 { 5692 int ret; 5693 5694 rcu_read_lock(); 5695 ret = __netif_receive_skb_one_core(skb, false); 5696 rcu_read_unlock(); 5697 5698 return ret; 5699 } 5700 EXPORT_SYMBOL(netif_receive_skb_core); 5701 5702 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5703 struct packet_type *pt_prev, 5704 struct net_device *orig_dev) 5705 { 5706 struct sk_buff *skb, *next; 5707 5708 if (!pt_prev) 5709 return; 5710 if (list_empty(head)) 5711 return; 5712 if (pt_prev->list_func != NULL) 5713 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5714 ip_list_rcv, head, pt_prev, orig_dev); 5715 else 5716 list_for_each_entry_safe(skb, next, head, list) { 5717 skb_list_del_init(skb); 5718 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5719 } 5720 } 5721 5722 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5723 { 5724 /* Fast-path assumptions: 5725 * - There is no RX handler. 5726 * - Only one packet_type matches. 5727 * If either of these fails, we will end up doing some per-packet 5728 * processing in-line, then handling the 'last ptype' for the whole 5729 * sublist. This can't cause out-of-order delivery to any single ptype, 5730 * because the 'last ptype' must be constant across the sublist, and all 5731 * other ptypes are handled per-packet. 5732 */ 5733 /* Current (common) ptype of sublist */ 5734 struct packet_type *pt_curr = NULL; 5735 /* Current (common) orig_dev of sublist */ 5736 struct net_device *od_curr = NULL; 5737 struct sk_buff *skb, *next; 5738 LIST_HEAD(sublist); 5739 5740 list_for_each_entry_safe(skb, next, head, list) { 5741 struct net_device *orig_dev = skb->dev; 5742 struct packet_type *pt_prev = NULL; 5743 5744 skb_list_del_init(skb); 5745 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5746 if (!pt_prev) 5747 continue; 5748 if (pt_curr != pt_prev || od_curr != orig_dev) { 5749 /* dispatch old sublist */ 5750 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5751 /* start new sublist */ 5752 INIT_LIST_HEAD(&sublist); 5753 pt_curr = pt_prev; 5754 od_curr = orig_dev; 5755 } 5756 list_add_tail(&skb->list, &sublist); 5757 } 5758 5759 /* dispatch final sublist */ 5760 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5761 } 5762 5763 static int __netif_receive_skb(struct sk_buff *skb) 5764 { 5765 int ret; 5766 5767 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 5768 unsigned int noreclaim_flag; 5769 5770 /* 5771 * PFMEMALLOC skbs are special, they should 5772 * - be delivered to SOCK_MEMALLOC sockets only 5773 * - stay away from userspace 5774 * - have bounded memory usage 5775 * 5776 * Use PF_MEMALLOC as this saves us from propagating the allocation 5777 * context down to all allocation sites. 5778 */ 5779 noreclaim_flag = memalloc_noreclaim_save(); 5780 ret = __netif_receive_skb_one_core(skb, true); 5781 memalloc_noreclaim_restore(noreclaim_flag); 5782 } else 5783 ret = __netif_receive_skb_one_core(skb, false); 5784 5785 return ret; 5786 } 5787 5788 static void __netif_receive_skb_list(struct list_head *head) 5789 { 5790 unsigned long noreclaim_flag = 0; 5791 struct sk_buff *skb, *next; 5792 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 5793 5794 list_for_each_entry_safe(skb, next, head, list) { 5795 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 5796 struct list_head sublist; 5797 5798 /* Handle the previous sublist */ 5799 list_cut_before(&sublist, head, &skb->list); 5800 if (!list_empty(&sublist)) 5801 __netif_receive_skb_list_core(&sublist, pfmemalloc); 5802 pfmemalloc = !pfmemalloc; 5803 /* See comments in __netif_receive_skb */ 5804 if (pfmemalloc) 5805 noreclaim_flag = memalloc_noreclaim_save(); 5806 else 5807 memalloc_noreclaim_restore(noreclaim_flag); 5808 } 5809 } 5810 /* Handle the remaining sublist */ 5811 if (!list_empty(head)) 5812 __netif_receive_skb_list_core(head, pfmemalloc); 5813 /* Restore pflags */ 5814 if (pfmemalloc) 5815 memalloc_noreclaim_restore(noreclaim_flag); 5816 } 5817 5818 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 5819 { 5820 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 5821 struct bpf_prog *new = xdp->prog; 5822 int ret = 0; 5823 5824 switch (xdp->command) { 5825 case XDP_SETUP_PROG: 5826 rcu_assign_pointer(dev->xdp_prog, new); 5827 if (old) 5828 bpf_prog_put(old); 5829 5830 if (old && !new) { 5831 static_branch_dec(&generic_xdp_needed_key); 5832 } else if (new && !old) { 5833 static_branch_inc(&generic_xdp_needed_key); 5834 dev_disable_lro(dev); 5835 dev_disable_gro_hw(dev); 5836 } 5837 break; 5838 5839 default: 5840 ret = -EINVAL; 5841 break; 5842 } 5843 5844 return ret; 5845 } 5846 5847 static int netif_receive_skb_internal(struct sk_buff *skb) 5848 { 5849 int ret; 5850 5851 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5852 5853 if (skb_defer_rx_timestamp(skb)) 5854 return NET_RX_SUCCESS; 5855 5856 rcu_read_lock(); 5857 #ifdef CONFIG_RPS 5858 if (static_branch_unlikely(&rps_needed)) { 5859 struct rps_dev_flow voidflow, *rflow = &voidflow; 5860 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 5861 5862 if (cpu >= 0) { 5863 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5864 rcu_read_unlock(); 5865 return ret; 5866 } 5867 } 5868 #endif 5869 ret = __netif_receive_skb(skb); 5870 rcu_read_unlock(); 5871 return ret; 5872 } 5873 5874 void netif_receive_skb_list_internal(struct list_head *head) 5875 { 5876 struct sk_buff *skb, *next; 5877 LIST_HEAD(sublist); 5878 5879 list_for_each_entry_safe(skb, next, head, list) { 5880 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 5881 skb); 5882 skb_list_del_init(skb); 5883 if (!skb_defer_rx_timestamp(skb)) 5884 list_add_tail(&skb->list, &sublist); 5885 } 5886 list_splice_init(&sublist, head); 5887 5888 rcu_read_lock(); 5889 #ifdef CONFIG_RPS 5890 if (static_branch_unlikely(&rps_needed)) { 5891 list_for_each_entry_safe(skb, next, head, list) { 5892 struct rps_dev_flow voidflow, *rflow = &voidflow; 5893 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 5894 5895 if (cpu >= 0) { 5896 /* Will be handled, remove from list */ 5897 skb_list_del_init(skb); 5898 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5899 } 5900 } 5901 } 5902 #endif 5903 __netif_receive_skb_list(head); 5904 rcu_read_unlock(); 5905 } 5906 5907 /** 5908 * netif_receive_skb - process receive buffer from network 5909 * @skb: buffer to process 5910 * 5911 * netif_receive_skb() is the main receive data processing function. 5912 * It always succeeds. The buffer may be dropped during processing 5913 * for congestion control or by the protocol layers. 5914 * 5915 * This function may only be called from softirq context and interrupts 5916 * should be enabled. 5917 * 5918 * Return values (usually ignored): 5919 * NET_RX_SUCCESS: no congestion 5920 * NET_RX_DROP: packet was dropped 5921 */ 5922 int netif_receive_skb(struct sk_buff *skb) 5923 { 5924 int ret; 5925 5926 trace_netif_receive_skb_entry(skb); 5927 5928 ret = netif_receive_skb_internal(skb); 5929 trace_netif_receive_skb_exit(ret); 5930 5931 return ret; 5932 } 5933 EXPORT_SYMBOL(netif_receive_skb); 5934 5935 /** 5936 * netif_receive_skb_list - process many receive buffers from network 5937 * @head: list of skbs to process. 5938 * 5939 * Since return value of netif_receive_skb() is normally ignored, and 5940 * wouldn't be meaningful for a list, this function returns void. 5941 * 5942 * This function may only be called from softirq context and interrupts 5943 * should be enabled. 5944 */ 5945 void netif_receive_skb_list(struct list_head *head) 5946 { 5947 struct sk_buff *skb; 5948 5949 if (list_empty(head)) 5950 return; 5951 if (trace_netif_receive_skb_list_entry_enabled()) { 5952 list_for_each_entry(skb, head, list) 5953 trace_netif_receive_skb_list_entry(skb); 5954 } 5955 netif_receive_skb_list_internal(head); 5956 trace_netif_receive_skb_list_exit(0); 5957 } 5958 EXPORT_SYMBOL(netif_receive_skb_list); 5959 5960 static DEFINE_PER_CPU(struct work_struct, flush_works); 5961 5962 /* Network device is going away, flush any packets still pending */ 5963 static void flush_backlog(struct work_struct *work) 5964 { 5965 struct sk_buff *skb, *tmp; 5966 struct softnet_data *sd; 5967 5968 local_bh_disable(); 5969 sd = this_cpu_ptr(&softnet_data); 5970 5971 backlog_lock_irq_disable(sd); 5972 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 5973 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 5974 __skb_unlink(skb, &sd->input_pkt_queue); 5975 dev_kfree_skb_irq(skb); 5976 rps_input_queue_head_incr(sd); 5977 } 5978 } 5979 backlog_unlock_irq_enable(sd); 5980 5981 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 5982 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 5983 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 5984 __skb_unlink(skb, &sd->process_queue); 5985 kfree_skb(skb); 5986 rps_input_queue_head_incr(sd); 5987 } 5988 } 5989 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 5990 local_bh_enable(); 5991 } 5992 5993 static bool flush_required(int cpu) 5994 { 5995 #if IS_ENABLED(CONFIG_RPS) 5996 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 5997 bool do_flush; 5998 5999 backlog_lock_irq_disable(sd); 6000 6001 /* as insertion into process_queue happens with the rps lock held, 6002 * process_queue access may race only with dequeue 6003 */ 6004 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6005 !skb_queue_empty_lockless(&sd->process_queue); 6006 backlog_unlock_irq_enable(sd); 6007 6008 return do_flush; 6009 #endif 6010 /* without RPS we can't safely check input_pkt_queue: during a 6011 * concurrent remote skb_queue_splice() we can detect as empty both 6012 * input_pkt_queue and process_queue even if the latter could end-up 6013 * containing a lot of packets. 6014 */ 6015 return true; 6016 } 6017 6018 static void flush_all_backlogs(void) 6019 { 6020 static cpumask_t flush_cpus; 6021 unsigned int cpu; 6022 6023 /* since we are under rtnl lock protection we can use static data 6024 * for the cpumask and avoid allocating on stack the possibly 6025 * large mask 6026 */ 6027 ASSERT_RTNL(); 6028 6029 cpus_read_lock(); 6030 6031 cpumask_clear(&flush_cpus); 6032 for_each_online_cpu(cpu) { 6033 if (flush_required(cpu)) { 6034 queue_work_on(cpu, system_highpri_wq, 6035 per_cpu_ptr(&flush_works, cpu)); 6036 cpumask_set_cpu(cpu, &flush_cpus); 6037 } 6038 } 6039 6040 /* we can have in flight packet[s] on the cpus we are not flushing, 6041 * synchronize_net() in unregister_netdevice_many() will take care of 6042 * them 6043 */ 6044 for_each_cpu(cpu, &flush_cpus) 6045 flush_work(per_cpu_ptr(&flush_works, cpu)); 6046 6047 cpus_read_unlock(); 6048 } 6049 6050 static void net_rps_send_ipi(struct softnet_data *remsd) 6051 { 6052 #ifdef CONFIG_RPS 6053 while (remsd) { 6054 struct softnet_data *next = remsd->rps_ipi_next; 6055 6056 if (cpu_online(remsd->cpu)) 6057 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6058 remsd = next; 6059 } 6060 #endif 6061 } 6062 6063 /* 6064 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6065 * Note: called with local irq disabled, but exits with local irq enabled. 6066 */ 6067 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6068 { 6069 #ifdef CONFIG_RPS 6070 struct softnet_data *remsd = sd->rps_ipi_list; 6071 6072 if (!use_backlog_threads() && remsd) { 6073 sd->rps_ipi_list = NULL; 6074 6075 local_irq_enable(); 6076 6077 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6078 net_rps_send_ipi(remsd); 6079 } else 6080 #endif 6081 local_irq_enable(); 6082 } 6083 6084 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6085 { 6086 #ifdef CONFIG_RPS 6087 return !use_backlog_threads() && sd->rps_ipi_list; 6088 #else 6089 return false; 6090 #endif 6091 } 6092 6093 static int process_backlog(struct napi_struct *napi, int quota) 6094 { 6095 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6096 bool again = true; 6097 int work = 0; 6098 6099 /* Check if we have pending ipi, its better to send them now, 6100 * not waiting net_rx_action() end. 6101 */ 6102 if (sd_has_rps_ipi_waiting(sd)) { 6103 local_irq_disable(); 6104 net_rps_action_and_irq_enable(sd); 6105 } 6106 6107 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6108 while (again) { 6109 struct sk_buff *skb; 6110 6111 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6112 while ((skb = __skb_dequeue(&sd->process_queue))) { 6113 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6114 rcu_read_lock(); 6115 __netif_receive_skb(skb); 6116 rcu_read_unlock(); 6117 if (++work >= quota) { 6118 rps_input_queue_head_add(sd, work); 6119 return work; 6120 } 6121 6122 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6123 } 6124 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6125 6126 backlog_lock_irq_disable(sd); 6127 if (skb_queue_empty(&sd->input_pkt_queue)) { 6128 /* 6129 * Inline a custom version of __napi_complete(). 6130 * only current cpu owns and manipulates this napi, 6131 * and NAPI_STATE_SCHED is the only possible flag set 6132 * on backlog. 6133 * We can use a plain write instead of clear_bit(), 6134 * and we dont need an smp_mb() memory barrier. 6135 */ 6136 napi->state &= NAPIF_STATE_THREADED; 6137 again = false; 6138 } else { 6139 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6140 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6141 &sd->process_queue); 6142 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6143 } 6144 backlog_unlock_irq_enable(sd); 6145 } 6146 6147 if (work) 6148 rps_input_queue_head_add(sd, work); 6149 return work; 6150 } 6151 6152 /** 6153 * __napi_schedule - schedule for receive 6154 * @n: entry to schedule 6155 * 6156 * The entry's receive function will be scheduled to run. 6157 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6158 */ 6159 void __napi_schedule(struct napi_struct *n) 6160 { 6161 unsigned long flags; 6162 6163 local_irq_save(flags); 6164 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6165 local_irq_restore(flags); 6166 } 6167 EXPORT_SYMBOL(__napi_schedule); 6168 6169 /** 6170 * napi_schedule_prep - check if napi can be scheduled 6171 * @n: napi context 6172 * 6173 * Test if NAPI routine is already running, and if not mark 6174 * it as running. This is used as a condition variable to 6175 * insure only one NAPI poll instance runs. We also make 6176 * sure there is no pending NAPI disable. 6177 */ 6178 bool napi_schedule_prep(struct napi_struct *n) 6179 { 6180 unsigned long new, val = READ_ONCE(n->state); 6181 6182 do { 6183 if (unlikely(val & NAPIF_STATE_DISABLE)) 6184 return false; 6185 new = val | NAPIF_STATE_SCHED; 6186 6187 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6188 * This was suggested by Alexander Duyck, as compiler 6189 * emits better code than : 6190 * if (val & NAPIF_STATE_SCHED) 6191 * new |= NAPIF_STATE_MISSED; 6192 */ 6193 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6194 NAPIF_STATE_MISSED; 6195 } while (!try_cmpxchg(&n->state, &val, new)); 6196 6197 return !(val & NAPIF_STATE_SCHED); 6198 } 6199 EXPORT_SYMBOL(napi_schedule_prep); 6200 6201 /** 6202 * __napi_schedule_irqoff - schedule for receive 6203 * @n: entry to schedule 6204 * 6205 * Variant of __napi_schedule() assuming hard irqs are masked. 6206 * 6207 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6208 * because the interrupt disabled assumption might not be true 6209 * due to force-threaded interrupts and spinlock substitution. 6210 */ 6211 void __napi_schedule_irqoff(struct napi_struct *n) 6212 { 6213 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6214 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6215 else 6216 __napi_schedule(n); 6217 } 6218 EXPORT_SYMBOL(__napi_schedule_irqoff); 6219 6220 bool napi_complete_done(struct napi_struct *n, int work_done) 6221 { 6222 unsigned long flags, val, new, timeout = 0; 6223 bool ret = true; 6224 6225 /* 6226 * 1) Don't let napi dequeue from the cpu poll list 6227 * just in case its running on a different cpu. 6228 * 2) If we are busy polling, do nothing here, we have 6229 * the guarantee we will be called later. 6230 */ 6231 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6232 NAPIF_STATE_IN_BUSY_POLL))) 6233 return false; 6234 6235 if (work_done) { 6236 if (n->gro_bitmask) 6237 timeout = READ_ONCE(n->dev->gro_flush_timeout); 6238 n->defer_hard_irqs_count = READ_ONCE(n->dev->napi_defer_hard_irqs); 6239 } 6240 if (n->defer_hard_irqs_count > 0) { 6241 n->defer_hard_irqs_count--; 6242 timeout = READ_ONCE(n->dev->gro_flush_timeout); 6243 if (timeout) 6244 ret = false; 6245 } 6246 if (n->gro_bitmask) { 6247 /* When the NAPI instance uses a timeout and keeps postponing 6248 * it, we need to bound somehow the time packets are kept in 6249 * the GRO layer 6250 */ 6251 napi_gro_flush(n, !!timeout); 6252 } 6253 6254 gro_normal_list(n); 6255 6256 if (unlikely(!list_empty(&n->poll_list))) { 6257 /* If n->poll_list is not empty, we need to mask irqs */ 6258 local_irq_save(flags); 6259 list_del_init(&n->poll_list); 6260 local_irq_restore(flags); 6261 } 6262 WRITE_ONCE(n->list_owner, -1); 6263 6264 val = READ_ONCE(n->state); 6265 do { 6266 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6267 6268 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6269 NAPIF_STATE_SCHED_THREADED | 6270 NAPIF_STATE_PREFER_BUSY_POLL); 6271 6272 /* If STATE_MISSED was set, leave STATE_SCHED set, 6273 * because we will call napi->poll() one more time. 6274 * This C code was suggested by Alexander Duyck to help gcc. 6275 */ 6276 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6277 NAPIF_STATE_SCHED; 6278 } while (!try_cmpxchg(&n->state, &val, new)); 6279 6280 if (unlikely(val & NAPIF_STATE_MISSED)) { 6281 __napi_schedule(n); 6282 return false; 6283 } 6284 6285 if (timeout) 6286 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6287 HRTIMER_MODE_REL_PINNED); 6288 return ret; 6289 } 6290 EXPORT_SYMBOL(napi_complete_done); 6291 6292 /* must be called under rcu_read_lock(), as we dont take a reference */ 6293 struct napi_struct *napi_by_id(unsigned int napi_id) 6294 { 6295 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 6296 struct napi_struct *napi; 6297 6298 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 6299 if (napi->napi_id == napi_id) 6300 return napi; 6301 6302 return NULL; 6303 } 6304 6305 static void skb_defer_free_flush(struct softnet_data *sd) 6306 { 6307 struct sk_buff *skb, *next; 6308 6309 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6310 if (!READ_ONCE(sd->defer_list)) 6311 return; 6312 6313 spin_lock(&sd->defer_lock); 6314 skb = sd->defer_list; 6315 sd->defer_list = NULL; 6316 sd->defer_count = 0; 6317 spin_unlock(&sd->defer_lock); 6318 6319 while (skb != NULL) { 6320 next = skb->next; 6321 napi_consume_skb(skb, 1); 6322 skb = next; 6323 } 6324 } 6325 6326 #if defined(CONFIG_NET_RX_BUSY_POLL) 6327 6328 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6329 { 6330 if (!skip_schedule) { 6331 gro_normal_list(napi); 6332 __napi_schedule(napi); 6333 return; 6334 } 6335 6336 if (napi->gro_bitmask) { 6337 /* flush too old packets 6338 * If HZ < 1000, flush all packets. 6339 */ 6340 napi_gro_flush(napi, HZ >= 1000); 6341 } 6342 6343 gro_normal_list(napi); 6344 clear_bit(NAPI_STATE_SCHED, &napi->state); 6345 } 6346 6347 enum { 6348 NAPI_F_PREFER_BUSY_POLL = 1, 6349 NAPI_F_END_ON_RESCHED = 2, 6350 }; 6351 6352 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6353 unsigned flags, u16 budget) 6354 { 6355 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6356 bool skip_schedule = false; 6357 unsigned long timeout; 6358 int rc; 6359 6360 /* Busy polling means there is a high chance device driver hard irq 6361 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6362 * set in napi_schedule_prep(). 6363 * Since we are about to call napi->poll() once more, we can safely 6364 * clear NAPI_STATE_MISSED. 6365 * 6366 * Note: x86 could use a single "lock and ..." instruction 6367 * to perform these two clear_bit() 6368 */ 6369 clear_bit(NAPI_STATE_MISSED, &napi->state); 6370 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6371 6372 local_bh_disable(); 6373 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6374 6375 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6376 napi->defer_hard_irqs_count = READ_ONCE(napi->dev->napi_defer_hard_irqs); 6377 timeout = READ_ONCE(napi->dev->gro_flush_timeout); 6378 if (napi->defer_hard_irqs_count && timeout) { 6379 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6380 skip_schedule = true; 6381 } 6382 } 6383 6384 /* All we really want here is to re-enable device interrupts. 6385 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6386 */ 6387 rc = napi->poll(napi, budget); 6388 /* We can't gro_normal_list() here, because napi->poll() might have 6389 * rearmed the napi (napi_complete_done()) in which case it could 6390 * already be running on another CPU. 6391 */ 6392 trace_napi_poll(napi, rc, budget); 6393 netpoll_poll_unlock(have_poll_lock); 6394 if (rc == budget) 6395 __busy_poll_stop(napi, skip_schedule); 6396 bpf_net_ctx_clear(bpf_net_ctx); 6397 local_bh_enable(); 6398 } 6399 6400 static void __napi_busy_loop(unsigned int napi_id, 6401 bool (*loop_end)(void *, unsigned long), 6402 void *loop_end_arg, unsigned flags, u16 budget) 6403 { 6404 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6405 int (*napi_poll)(struct napi_struct *napi, int budget); 6406 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6407 void *have_poll_lock = NULL; 6408 struct napi_struct *napi; 6409 6410 WARN_ON_ONCE(!rcu_read_lock_held()); 6411 6412 restart: 6413 napi_poll = NULL; 6414 6415 napi = napi_by_id(napi_id); 6416 if (!napi) 6417 return; 6418 6419 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6420 preempt_disable(); 6421 for (;;) { 6422 int work = 0; 6423 6424 local_bh_disable(); 6425 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6426 if (!napi_poll) { 6427 unsigned long val = READ_ONCE(napi->state); 6428 6429 /* If multiple threads are competing for this napi, 6430 * we avoid dirtying napi->state as much as we can. 6431 */ 6432 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6433 NAPIF_STATE_IN_BUSY_POLL)) { 6434 if (flags & NAPI_F_PREFER_BUSY_POLL) 6435 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6436 goto count; 6437 } 6438 if (cmpxchg(&napi->state, val, 6439 val | NAPIF_STATE_IN_BUSY_POLL | 6440 NAPIF_STATE_SCHED) != val) { 6441 if (flags & NAPI_F_PREFER_BUSY_POLL) 6442 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6443 goto count; 6444 } 6445 have_poll_lock = netpoll_poll_lock(napi); 6446 napi_poll = napi->poll; 6447 } 6448 work = napi_poll(napi, budget); 6449 trace_napi_poll(napi, work, budget); 6450 gro_normal_list(napi); 6451 count: 6452 if (work > 0) 6453 __NET_ADD_STATS(dev_net(napi->dev), 6454 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6455 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6456 bpf_net_ctx_clear(bpf_net_ctx); 6457 local_bh_enable(); 6458 6459 if (!loop_end || loop_end(loop_end_arg, start_time)) 6460 break; 6461 6462 if (unlikely(need_resched())) { 6463 if (flags & NAPI_F_END_ON_RESCHED) 6464 break; 6465 if (napi_poll) 6466 busy_poll_stop(napi, have_poll_lock, flags, budget); 6467 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6468 preempt_enable(); 6469 rcu_read_unlock(); 6470 cond_resched(); 6471 rcu_read_lock(); 6472 if (loop_end(loop_end_arg, start_time)) 6473 return; 6474 goto restart; 6475 } 6476 cpu_relax(); 6477 } 6478 if (napi_poll) 6479 busy_poll_stop(napi, have_poll_lock, flags, budget); 6480 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6481 preempt_enable(); 6482 } 6483 6484 void napi_busy_loop_rcu(unsigned int napi_id, 6485 bool (*loop_end)(void *, unsigned long), 6486 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6487 { 6488 unsigned flags = NAPI_F_END_ON_RESCHED; 6489 6490 if (prefer_busy_poll) 6491 flags |= NAPI_F_PREFER_BUSY_POLL; 6492 6493 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6494 } 6495 6496 void napi_busy_loop(unsigned int napi_id, 6497 bool (*loop_end)(void *, unsigned long), 6498 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6499 { 6500 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6501 6502 rcu_read_lock(); 6503 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6504 rcu_read_unlock(); 6505 } 6506 EXPORT_SYMBOL(napi_busy_loop); 6507 6508 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6509 6510 static void napi_hash_add(struct napi_struct *napi) 6511 { 6512 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6513 return; 6514 6515 spin_lock(&napi_hash_lock); 6516 6517 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6518 do { 6519 if (unlikely(++napi_gen_id < MIN_NAPI_ID)) 6520 napi_gen_id = MIN_NAPI_ID; 6521 } while (napi_by_id(napi_gen_id)); 6522 napi->napi_id = napi_gen_id; 6523 6524 hlist_add_head_rcu(&napi->napi_hash_node, 6525 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6526 6527 spin_unlock(&napi_hash_lock); 6528 } 6529 6530 /* Warning : caller is responsible to make sure rcu grace period 6531 * is respected before freeing memory containing @napi 6532 */ 6533 static void napi_hash_del(struct napi_struct *napi) 6534 { 6535 spin_lock(&napi_hash_lock); 6536 6537 hlist_del_init_rcu(&napi->napi_hash_node); 6538 6539 spin_unlock(&napi_hash_lock); 6540 } 6541 6542 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6543 { 6544 struct napi_struct *napi; 6545 6546 napi = container_of(timer, struct napi_struct, timer); 6547 6548 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6549 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6550 */ 6551 if (!napi_disable_pending(napi) && 6552 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6553 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6554 __napi_schedule_irqoff(napi); 6555 } 6556 6557 return HRTIMER_NORESTART; 6558 } 6559 6560 static void init_gro_hash(struct napi_struct *napi) 6561 { 6562 int i; 6563 6564 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6565 INIT_LIST_HEAD(&napi->gro_hash[i].list); 6566 napi->gro_hash[i].count = 0; 6567 } 6568 napi->gro_bitmask = 0; 6569 } 6570 6571 int dev_set_threaded(struct net_device *dev, bool threaded) 6572 { 6573 struct napi_struct *napi; 6574 int err = 0; 6575 6576 if (dev->threaded == threaded) 6577 return 0; 6578 6579 if (threaded) { 6580 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6581 if (!napi->thread) { 6582 err = napi_kthread_create(napi); 6583 if (err) { 6584 threaded = false; 6585 break; 6586 } 6587 } 6588 } 6589 } 6590 6591 WRITE_ONCE(dev->threaded, threaded); 6592 6593 /* Make sure kthread is created before THREADED bit 6594 * is set. 6595 */ 6596 smp_mb__before_atomic(); 6597 6598 /* Setting/unsetting threaded mode on a napi might not immediately 6599 * take effect, if the current napi instance is actively being 6600 * polled. In this case, the switch between threaded mode and 6601 * softirq mode will happen in the next round of napi_schedule(). 6602 * This should not cause hiccups/stalls to the live traffic. 6603 */ 6604 list_for_each_entry(napi, &dev->napi_list, dev_list) 6605 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6606 6607 return err; 6608 } 6609 EXPORT_SYMBOL(dev_set_threaded); 6610 6611 /** 6612 * netif_queue_set_napi - Associate queue with the napi 6613 * @dev: device to which NAPI and queue belong 6614 * @queue_index: Index of queue 6615 * @type: queue type as RX or TX 6616 * @napi: NAPI context, pass NULL to clear previously set NAPI 6617 * 6618 * Set queue with its corresponding napi context. This should be done after 6619 * registering the NAPI handler for the queue-vector and the queues have been 6620 * mapped to the corresponding interrupt vector. 6621 */ 6622 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6623 enum netdev_queue_type type, struct napi_struct *napi) 6624 { 6625 struct netdev_rx_queue *rxq; 6626 struct netdev_queue *txq; 6627 6628 if (WARN_ON_ONCE(napi && !napi->dev)) 6629 return; 6630 if (dev->reg_state >= NETREG_REGISTERED) 6631 ASSERT_RTNL(); 6632 6633 switch (type) { 6634 case NETDEV_QUEUE_TYPE_RX: 6635 rxq = __netif_get_rx_queue(dev, queue_index); 6636 rxq->napi = napi; 6637 return; 6638 case NETDEV_QUEUE_TYPE_TX: 6639 txq = netdev_get_tx_queue(dev, queue_index); 6640 txq->napi = napi; 6641 return; 6642 default: 6643 return; 6644 } 6645 } 6646 EXPORT_SYMBOL(netif_queue_set_napi); 6647 6648 void netif_napi_add_weight(struct net_device *dev, struct napi_struct *napi, 6649 int (*poll)(struct napi_struct *, int), int weight) 6650 { 6651 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 6652 return; 6653 6654 INIT_LIST_HEAD(&napi->poll_list); 6655 INIT_HLIST_NODE(&napi->napi_hash_node); 6656 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 6657 napi->timer.function = napi_watchdog; 6658 init_gro_hash(napi); 6659 napi->skb = NULL; 6660 INIT_LIST_HEAD(&napi->rx_list); 6661 napi->rx_count = 0; 6662 napi->poll = poll; 6663 if (weight > NAPI_POLL_WEIGHT) 6664 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 6665 weight); 6666 napi->weight = weight; 6667 napi->dev = dev; 6668 #ifdef CONFIG_NETPOLL 6669 napi->poll_owner = -1; 6670 #endif 6671 napi->list_owner = -1; 6672 set_bit(NAPI_STATE_SCHED, &napi->state); 6673 set_bit(NAPI_STATE_NPSVC, &napi->state); 6674 list_add_rcu(&napi->dev_list, &dev->napi_list); 6675 napi_hash_add(napi); 6676 napi_get_frags_check(napi); 6677 /* Create kthread for this napi if dev->threaded is set. 6678 * Clear dev->threaded if kthread creation failed so that 6679 * threaded mode will not be enabled in napi_enable(). 6680 */ 6681 if (dev->threaded && napi_kthread_create(napi)) 6682 dev->threaded = false; 6683 netif_napi_set_irq(napi, -1); 6684 } 6685 EXPORT_SYMBOL(netif_napi_add_weight); 6686 6687 void napi_disable(struct napi_struct *n) 6688 { 6689 unsigned long val, new; 6690 6691 might_sleep(); 6692 set_bit(NAPI_STATE_DISABLE, &n->state); 6693 6694 val = READ_ONCE(n->state); 6695 do { 6696 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 6697 usleep_range(20, 200); 6698 val = READ_ONCE(n->state); 6699 } 6700 6701 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 6702 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 6703 } while (!try_cmpxchg(&n->state, &val, new)); 6704 6705 hrtimer_cancel(&n->timer); 6706 6707 clear_bit(NAPI_STATE_DISABLE, &n->state); 6708 } 6709 EXPORT_SYMBOL(napi_disable); 6710 6711 /** 6712 * napi_enable - enable NAPI scheduling 6713 * @n: NAPI context 6714 * 6715 * Resume NAPI from being scheduled on this context. 6716 * Must be paired with napi_disable. 6717 */ 6718 void napi_enable(struct napi_struct *n) 6719 { 6720 unsigned long new, val = READ_ONCE(n->state); 6721 6722 do { 6723 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 6724 6725 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 6726 if (n->dev->threaded && n->thread) 6727 new |= NAPIF_STATE_THREADED; 6728 } while (!try_cmpxchg(&n->state, &val, new)); 6729 } 6730 EXPORT_SYMBOL(napi_enable); 6731 6732 static void flush_gro_hash(struct napi_struct *napi) 6733 { 6734 int i; 6735 6736 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6737 struct sk_buff *skb, *n; 6738 6739 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list) 6740 kfree_skb(skb); 6741 napi->gro_hash[i].count = 0; 6742 } 6743 } 6744 6745 /* Must be called in process context */ 6746 void __netif_napi_del(struct napi_struct *napi) 6747 { 6748 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 6749 return; 6750 6751 napi_hash_del(napi); 6752 list_del_rcu(&napi->dev_list); 6753 napi_free_frags(napi); 6754 6755 flush_gro_hash(napi); 6756 napi->gro_bitmask = 0; 6757 6758 if (napi->thread) { 6759 kthread_stop(napi->thread); 6760 napi->thread = NULL; 6761 } 6762 } 6763 EXPORT_SYMBOL(__netif_napi_del); 6764 6765 static int __napi_poll(struct napi_struct *n, bool *repoll) 6766 { 6767 int work, weight; 6768 6769 weight = n->weight; 6770 6771 /* This NAPI_STATE_SCHED test is for avoiding a race 6772 * with netpoll's poll_napi(). Only the entity which 6773 * obtains the lock and sees NAPI_STATE_SCHED set will 6774 * actually make the ->poll() call. Therefore we avoid 6775 * accidentally calling ->poll() when NAPI is not scheduled. 6776 */ 6777 work = 0; 6778 if (napi_is_scheduled(n)) { 6779 work = n->poll(n, weight); 6780 trace_napi_poll(n, work, weight); 6781 6782 xdp_do_check_flushed(n); 6783 } 6784 6785 if (unlikely(work > weight)) 6786 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 6787 n->poll, work, weight); 6788 6789 if (likely(work < weight)) 6790 return work; 6791 6792 /* Drivers must not modify the NAPI state if they 6793 * consume the entire weight. In such cases this code 6794 * still "owns" the NAPI instance and therefore can 6795 * move the instance around on the list at-will. 6796 */ 6797 if (unlikely(napi_disable_pending(n))) { 6798 napi_complete(n); 6799 return work; 6800 } 6801 6802 /* The NAPI context has more processing work, but busy-polling 6803 * is preferred. Exit early. 6804 */ 6805 if (napi_prefer_busy_poll(n)) { 6806 if (napi_complete_done(n, work)) { 6807 /* If timeout is not set, we need to make sure 6808 * that the NAPI is re-scheduled. 6809 */ 6810 napi_schedule(n); 6811 } 6812 return work; 6813 } 6814 6815 if (n->gro_bitmask) { 6816 /* flush too old packets 6817 * If HZ < 1000, flush all packets. 6818 */ 6819 napi_gro_flush(n, HZ >= 1000); 6820 } 6821 6822 gro_normal_list(n); 6823 6824 /* Some drivers may have called napi_schedule 6825 * prior to exhausting their budget. 6826 */ 6827 if (unlikely(!list_empty(&n->poll_list))) { 6828 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 6829 n->dev ? n->dev->name : "backlog"); 6830 return work; 6831 } 6832 6833 *repoll = true; 6834 6835 return work; 6836 } 6837 6838 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 6839 { 6840 bool do_repoll = false; 6841 void *have; 6842 int work; 6843 6844 list_del_init(&n->poll_list); 6845 6846 have = netpoll_poll_lock(n); 6847 6848 work = __napi_poll(n, &do_repoll); 6849 6850 if (do_repoll) 6851 list_add_tail(&n->poll_list, repoll); 6852 6853 netpoll_poll_unlock(have); 6854 6855 return work; 6856 } 6857 6858 static int napi_thread_wait(struct napi_struct *napi) 6859 { 6860 set_current_state(TASK_INTERRUPTIBLE); 6861 6862 while (!kthread_should_stop()) { 6863 /* Testing SCHED_THREADED bit here to make sure the current 6864 * kthread owns this napi and could poll on this napi. 6865 * Testing SCHED bit is not enough because SCHED bit might be 6866 * set by some other busy poll thread or by napi_disable(). 6867 */ 6868 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 6869 WARN_ON(!list_empty(&napi->poll_list)); 6870 __set_current_state(TASK_RUNNING); 6871 return 0; 6872 } 6873 6874 schedule(); 6875 set_current_state(TASK_INTERRUPTIBLE); 6876 } 6877 __set_current_state(TASK_RUNNING); 6878 6879 return -1; 6880 } 6881 6882 static void napi_threaded_poll_loop(struct napi_struct *napi) 6883 { 6884 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6885 struct softnet_data *sd; 6886 unsigned long last_qs = jiffies; 6887 6888 for (;;) { 6889 bool repoll = false; 6890 void *have; 6891 6892 local_bh_disable(); 6893 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6894 6895 sd = this_cpu_ptr(&softnet_data); 6896 sd->in_napi_threaded_poll = true; 6897 6898 have = netpoll_poll_lock(napi); 6899 __napi_poll(napi, &repoll); 6900 netpoll_poll_unlock(have); 6901 6902 sd->in_napi_threaded_poll = false; 6903 barrier(); 6904 6905 if (sd_has_rps_ipi_waiting(sd)) { 6906 local_irq_disable(); 6907 net_rps_action_and_irq_enable(sd); 6908 } 6909 skb_defer_free_flush(sd); 6910 bpf_net_ctx_clear(bpf_net_ctx); 6911 local_bh_enable(); 6912 6913 if (!repoll) 6914 break; 6915 6916 rcu_softirq_qs_periodic(last_qs); 6917 cond_resched(); 6918 } 6919 } 6920 6921 static int napi_threaded_poll(void *data) 6922 { 6923 struct napi_struct *napi = data; 6924 6925 while (!napi_thread_wait(napi)) 6926 napi_threaded_poll_loop(napi); 6927 6928 return 0; 6929 } 6930 6931 static __latent_entropy void net_rx_action(void) 6932 { 6933 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 6934 unsigned long time_limit = jiffies + 6935 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 6936 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6937 int budget = READ_ONCE(net_hotdata.netdev_budget); 6938 LIST_HEAD(list); 6939 LIST_HEAD(repoll); 6940 6941 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6942 start: 6943 sd->in_net_rx_action = true; 6944 local_irq_disable(); 6945 list_splice_init(&sd->poll_list, &list); 6946 local_irq_enable(); 6947 6948 for (;;) { 6949 struct napi_struct *n; 6950 6951 skb_defer_free_flush(sd); 6952 6953 if (list_empty(&list)) { 6954 if (list_empty(&repoll)) { 6955 sd->in_net_rx_action = false; 6956 barrier(); 6957 /* We need to check if ____napi_schedule() 6958 * had refilled poll_list while 6959 * sd->in_net_rx_action was true. 6960 */ 6961 if (!list_empty(&sd->poll_list)) 6962 goto start; 6963 if (!sd_has_rps_ipi_waiting(sd)) 6964 goto end; 6965 } 6966 break; 6967 } 6968 6969 n = list_first_entry(&list, struct napi_struct, poll_list); 6970 budget -= napi_poll(n, &repoll); 6971 6972 /* If softirq window is exhausted then punt. 6973 * Allow this to run for 2 jiffies since which will allow 6974 * an average latency of 1.5/HZ. 6975 */ 6976 if (unlikely(budget <= 0 || 6977 time_after_eq(jiffies, time_limit))) { 6978 sd->time_squeeze++; 6979 break; 6980 } 6981 } 6982 6983 local_irq_disable(); 6984 6985 list_splice_tail_init(&sd->poll_list, &list); 6986 list_splice_tail(&repoll, &list); 6987 list_splice(&list, &sd->poll_list); 6988 if (!list_empty(&sd->poll_list)) 6989 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 6990 else 6991 sd->in_net_rx_action = false; 6992 6993 net_rps_action_and_irq_enable(sd); 6994 end: 6995 bpf_net_ctx_clear(bpf_net_ctx); 6996 } 6997 6998 struct netdev_adjacent { 6999 struct net_device *dev; 7000 netdevice_tracker dev_tracker; 7001 7002 /* upper master flag, there can only be one master device per list */ 7003 bool master; 7004 7005 /* lookup ignore flag */ 7006 bool ignore; 7007 7008 /* counter for the number of times this device was added to us */ 7009 u16 ref_nr; 7010 7011 /* private field for the users */ 7012 void *private; 7013 7014 struct list_head list; 7015 struct rcu_head rcu; 7016 }; 7017 7018 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7019 struct list_head *adj_list) 7020 { 7021 struct netdev_adjacent *adj; 7022 7023 list_for_each_entry(adj, adj_list, list) { 7024 if (adj->dev == adj_dev) 7025 return adj; 7026 } 7027 return NULL; 7028 } 7029 7030 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7031 struct netdev_nested_priv *priv) 7032 { 7033 struct net_device *dev = (struct net_device *)priv->data; 7034 7035 return upper_dev == dev; 7036 } 7037 7038 /** 7039 * netdev_has_upper_dev - Check if device is linked to an upper device 7040 * @dev: device 7041 * @upper_dev: upper device to check 7042 * 7043 * Find out if a device is linked to specified upper device and return true 7044 * in case it is. Note that this checks only immediate upper device, 7045 * not through a complete stack of devices. The caller must hold the RTNL lock. 7046 */ 7047 bool netdev_has_upper_dev(struct net_device *dev, 7048 struct net_device *upper_dev) 7049 { 7050 struct netdev_nested_priv priv = { 7051 .data = (void *)upper_dev, 7052 }; 7053 7054 ASSERT_RTNL(); 7055 7056 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7057 &priv); 7058 } 7059 EXPORT_SYMBOL(netdev_has_upper_dev); 7060 7061 /** 7062 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7063 * @dev: device 7064 * @upper_dev: upper device to check 7065 * 7066 * Find out if a device is linked to specified upper device and return true 7067 * in case it is. Note that this checks the entire upper device chain. 7068 * The caller must hold rcu lock. 7069 */ 7070 7071 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 7072 struct net_device *upper_dev) 7073 { 7074 struct netdev_nested_priv priv = { 7075 .data = (void *)upper_dev, 7076 }; 7077 7078 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7079 &priv); 7080 } 7081 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 7082 7083 /** 7084 * netdev_has_any_upper_dev - Check if device is linked to some device 7085 * @dev: device 7086 * 7087 * Find out if a device is linked to an upper device and return true in case 7088 * it is. The caller must hold the RTNL lock. 7089 */ 7090 bool netdev_has_any_upper_dev(struct net_device *dev) 7091 { 7092 ASSERT_RTNL(); 7093 7094 return !list_empty(&dev->adj_list.upper); 7095 } 7096 EXPORT_SYMBOL(netdev_has_any_upper_dev); 7097 7098 /** 7099 * netdev_master_upper_dev_get - Get master upper device 7100 * @dev: device 7101 * 7102 * Find a master upper device and return pointer to it or NULL in case 7103 * it's not there. The caller must hold the RTNL lock. 7104 */ 7105 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 7106 { 7107 struct netdev_adjacent *upper; 7108 7109 ASSERT_RTNL(); 7110 7111 if (list_empty(&dev->adj_list.upper)) 7112 return NULL; 7113 7114 upper = list_first_entry(&dev->adj_list.upper, 7115 struct netdev_adjacent, list); 7116 if (likely(upper->master)) 7117 return upper->dev; 7118 return NULL; 7119 } 7120 EXPORT_SYMBOL(netdev_master_upper_dev_get); 7121 7122 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 7123 { 7124 struct netdev_adjacent *upper; 7125 7126 ASSERT_RTNL(); 7127 7128 if (list_empty(&dev->adj_list.upper)) 7129 return NULL; 7130 7131 upper = list_first_entry(&dev->adj_list.upper, 7132 struct netdev_adjacent, list); 7133 if (likely(upper->master) && !upper->ignore) 7134 return upper->dev; 7135 return NULL; 7136 } 7137 7138 /** 7139 * netdev_has_any_lower_dev - Check if device is linked to some device 7140 * @dev: device 7141 * 7142 * Find out if a device is linked to a lower device and return true in case 7143 * it is. The caller must hold the RTNL lock. 7144 */ 7145 static bool netdev_has_any_lower_dev(struct net_device *dev) 7146 { 7147 ASSERT_RTNL(); 7148 7149 return !list_empty(&dev->adj_list.lower); 7150 } 7151 7152 void *netdev_adjacent_get_private(struct list_head *adj_list) 7153 { 7154 struct netdev_adjacent *adj; 7155 7156 adj = list_entry(adj_list, struct netdev_adjacent, list); 7157 7158 return adj->private; 7159 } 7160 EXPORT_SYMBOL(netdev_adjacent_get_private); 7161 7162 /** 7163 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7164 * @dev: device 7165 * @iter: list_head ** of the current position 7166 * 7167 * Gets the next device from the dev's upper list, starting from iter 7168 * position. The caller must hold RCU read lock. 7169 */ 7170 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7171 struct list_head **iter) 7172 { 7173 struct netdev_adjacent *upper; 7174 7175 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7176 7177 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7178 7179 if (&upper->list == &dev->adj_list.upper) 7180 return NULL; 7181 7182 *iter = &upper->list; 7183 7184 return upper->dev; 7185 } 7186 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7187 7188 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7189 struct list_head **iter, 7190 bool *ignore) 7191 { 7192 struct netdev_adjacent *upper; 7193 7194 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7195 7196 if (&upper->list == &dev->adj_list.upper) 7197 return NULL; 7198 7199 *iter = &upper->list; 7200 *ignore = upper->ignore; 7201 7202 return upper->dev; 7203 } 7204 7205 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7206 struct list_head **iter) 7207 { 7208 struct netdev_adjacent *upper; 7209 7210 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7211 7212 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7213 7214 if (&upper->list == &dev->adj_list.upper) 7215 return NULL; 7216 7217 *iter = &upper->list; 7218 7219 return upper->dev; 7220 } 7221 7222 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7223 int (*fn)(struct net_device *dev, 7224 struct netdev_nested_priv *priv), 7225 struct netdev_nested_priv *priv) 7226 { 7227 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7228 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7229 int ret, cur = 0; 7230 bool ignore; 7231 7232 now = dev; 7233 iter = &dev->adj_list.upper; 7234 7235 while (1) { 7236 if (now != dev) { 7237 ret = fn(now, priv); 7238 if (ret) 7239 return ret; 7240 } 7241 7242 next = NULL; 7243 while (1) { 7244 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7245 if (!udev) 7246 break; 7247 if (ignore) 7248 continue; 7249 7250 next = udev; 7251 niter = &udev->adj_list.upper; 7252 dev_stack[cur] = now; 7253 iter_stack[cur++] = iter; 7254 break; 7255 } 7256 7257 if (!next) { 7258 if (!cur) 7259 return 0; 7260 next = dev_stack[--cur]; 7261 niter = iter_stack[cur]; 7262 } 7263 7264 now = next; 7265 iter = niter; 7266 } 7267 7268 return 0; 7269 } 7270 7271 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7272 int (*fn)(struct net_device *dev, 7273 struct netdev_nested_priv *priv), 7274 struct netdev_nested_priv *priv) 7275 { 7276 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7277 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7278 int ret, cur = 0; 7279 7280 now = dev; 7281 iter = &dev->adj_list.upper; 7282 7283 while (1) { 7284 if (now != dev) { 7285 ret = fn(now, priv); 7286 if (ret) 7287 return ret; 7288 } 7289 7290 next = NULL; 7291 while (1) { 7292 udev = netdev_next_upper_dev_rcu(now, &iter); 7293 if (!udev) 7294 break; 7295 7296 next = udev; 7297 niter = &udev->adj_list.upper; 7298 dev_stack[cur] = now; 7299 iter_stack[cur++] = iter; 7300 break; 7301 } 7302 7303 if (!next) { 7304 if (!cur) 7305 return 0; 7306 next = dev_stack[--cur]; 7307 niter = iter_stack[cur]; 7308 } 7309 7310 now = next; 7311 iter = niter; 7312 } 7313 7314 return 0; 7315 } 7316 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7317 7318 static bool __netdev_has_upper_dev(struct net_device *dev, 7319 struct net_device *upper_dev) 7320 { 7321 struct netdev_nested_priv priv = { 7322 .flags = 0, 7323 .data = (void *)upper_dev, 7324 }; 7325 7326 ASSERT_RTNL(); 7327 7328 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7329 &priv); 7330 } 7331 7332 /** 7333 * netdev_lower_get_next_private - Get the next ->private from the 7334 * lower neighbour list 7335 * @dev: device 7336 * @iter: list_head ** of the current position 7337 * 7338 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7339 * list, starting from iter position. The caller must hold either hold the 7340 * RTNL lock or its own locking that guarantees that the neighbour lower 7341 * list will remain unchanged. 7342 */ 7343 void *netdev_lower_get_next_private(struct net_device *dev, 7344 struct list_head **iter) 7345 { 7346 struct netdev_adjacent *lower; 7347 7348 lower = list_entry(*iter, struct netdev_adjacent, list); 7349 7350 if (&lower->list == &dev->adj_list.lower) 7351 return NULL; 7352 7353 *iter = lower->list.next; 7354 7355 return lower->private; 7356 } 7357 EXPORT_SYMBOL(netdev_lower_get_next_private); 7358 7359 /** 7360 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7361 * lower neighbour list, RCU 7362 * variant 7363 * @dev: device 7364 * @iter: list_head ** of the current position 7365 * 7366 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7367 * list, starting from iter position. The caller must hold RCU read lock. 7368 */ 7369 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7370 struct list_head **iter) 7371 { 7372 struct netdev_adjacent *lower; 7373 7374 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7375 7376 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7377 7378 if (&lower->list == &dev->adj_list.lower) 7379 return NULL; 7380 7381 *iter = &lower->list; 7382 7383 return lower->private; 7384 } 7385 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7386 7387 /** 7388 * netdev_lower_get_next - Get the next device from the lower neighbour 7389 * list 7390 * @dev: device 7391 * @iter: list_head ** of the current position 7392 * 7393 * Gets the next netdev_adjacent from the dev's lower neighbour 7394 * list, starting from iter position. The caller must hold RTNL lock or 7395 * its own locking that guarantees that the neighbour lower 7396 * list will remain unchanged. 7397 */ 7398 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7399 { 7400 struct netdev_adjacent *lower; 7401 7402 lower = list_entry(*iter, struct netdev_adjacent, list); 7403 7404 if (&lower->list == &dev->adj_list.lower) 7405 return NULL; 7406 7407 *iter = lower->list.next; 7408 7409 return lower->dev; 7410 } 7411 EXPORT_SYMBOL(netdev_lower_get_next); 7412 7413 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 7414 struct list_head **iter) 7415 { 7416 struct netdev_adjacent *lower; 7417 7418 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7419 7420 if (&lower->list == &dev->adj_list.lower) 7421 return NULL; 7422 7423 *iter = &lower->list; 7424 7425 return lower->dev; 7426 } 7427 7428 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 7429 struct list_head **iter, 7430 bool *ignore) 7431 { 7432 struct netdev_adjacent *lower; 7433 7434 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7435 7436 if (&lower->list == &dev->adj_list.lower) 7437 return NULL; 7438 7439 *iter = &lower->list; 7440 *ignore = lower->ignore; 7441 7442 return lower->dev; 7443 } 7444 7445 int netdev_walk_all_lower_dev(struct net_device *dev, 7446 int (*fn)(struct net_device *dev, 7447 struct netdev_nested_priv *priv), 7448 struct netdev_nested_priv *priv) 7449 { 7450 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7451 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7452 int ret, cur = 0; 7453 7454 now = dev; 7455 iter = &dev->adj_list.lower; 7456 7457 while (1) { 7458 if (now != dev) { 7459 ret = fn(now, priv); 7460 if (ret) 7461 return ret; 7462 } 7463 7464 next = NULL; 7465 while (1) { 7466 ldev = netdev_next_lower_dev(now, &iter); 7467 if (!ldev) 7468 break; 7469 7470 next = ldev; 7471 niter = &ldev->adj_list.lower; 7472 dev_stack[cur] = now; 7473 iter_stack[cur++] = iter; 7474 break; 7475 } 7476 7477 if (!next) { 7478 if (!cur) 7479 return 0; 7480 next = dev_stack[--cur]; 7481 niter = iter_stack[cur]; 7482 } 7483 7484 now = next; 7485 iter = niter; 7486 } 7487 7488 return 0; 7489 } 7490 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 7491 7492 static int __netdev_walk_all_lower_dev(struct net_device *dev, 7493 int (*fn)(struct net_device *dev, 7494 struct netdev_nested_priv *priv), 7495 struct netdev_nested_priv *priv) 7496 { 7497 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7498 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7499 int ret, cur = 0; 7500 bool ignore; 7501 7502 now = dev; 7503 iter = &dev->adj_list.lower; 7504 7505 while (1) { 7506 if (now != dev) { 7507 ret = fn(now, priv); 7508 if (ret) 7509 return ret; 7510 } 7511 7512 next = NULL; 7513 while (1) { 7514 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 7515 if (!ldev) 7516 break; 7517 if (ignore) 7518 continue; 7519 7520 next = ldev; 7521 niter = &ldev->adj_list.lower; 7522 dev_stack[cur] = now; 7523 iter_stack[cur++] = iter; 7524 break; 7525 } 7526 7527 if (!next) { 7528 if (!cur) 7529 return 0; 7530 next = dev_stack[--cur]; 7531 niter = iter_stack[cur]; 7532 } 7533 7534 now = next; 7535 iter = niter; 7536 } 7537 7538 return 0; 7539 } 7540 7541 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 7542 struct list_head **iter) 7543 { 7544 struct netdev_adjacent *lower; 7545 7546 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7547 if (&lower->list == &dev->adj_list.lower) 7548 return NULL; 7549 7550 *iter = &lower->list; 7551 7552 return lower->dev; 7553 } 7554 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 7555 7556 static u8 __netdev_upper_depth(struct net_device *dev) 7557 { 7558 struct net_device *udev; 7559 struct list_head *iter; 7560 u8 max_depth = 0; 7561 bool ignore; 7562 7563 for (iter = &dev->adj_list.upper, 7564 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 7565 udev; 7566 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 7567 if (ignore) 7568 continue; 7569 if (max_depth < udev->upper_level) 7570 max_depth = udev->upper_level; 7571 } 7572 7573 return max_depth; 7574 } 7575 7576 static u8 __netdev_lower_depth(struct net_device *dev) 7577 { 7578 struct net_device *ldev; 7579 struct list_head *iter; 7580 u8 max_depth = 0; 7581 bool ignore; 7582 7583 for (iter = &dev->adj_list.lower, 7584 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 7585 ldev; 7586 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 7587 if (ignore) 7588 continue; 7589 if (max_depth < ldev->lower_level) 7590 max_depth = ldev->lower_level; 7591 } 7592 7593 return max_depth; 7594 } 7595 7596 static int __netdev_update_upper_level(struct net_device *dev, 7597 struct netdev_nested_priv *__unused) 7598 { 7599 dev->upper_level = __netdev_upper_depth(dev) + 1; 7600 return 0; 7601 } 7602 7603 #ifdef CONFIG_LOCKDEP 7604 static LIST_HEAD(net_unlink_list); 7605 7606 static void net_unlink_todo(struct net_device *dev) 7607 { 7608 if (list_empty(&dev->unlink_list)) 7609 list_add_tail(&dev->unlink_list, &net_unlink_list); 7610 } 7611 #endif 7612 7613 static int __netdev_update_lower_level(struct net_device *dev, 7614 struct netdev_nested_priv *priv) 7615 { 7616 dev->lower_level = __netdev_lower_depth(dev) + 1; 7617 7618 #ifdef CONFIG_LOCKDEP 7619 if (!priv) 7620 return 0; 7621 7622 if (priv->flags & NESTED_SYNC_IMM) 7623 dev->nested_level = dev->lower_level - 1; 7624 if (priv->flags & NESTED_SYNC_TODO) 7625 net_unlink_todo(dev); 7626 #endif 7627 return 0; 7628 } 7629 7630 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 7631 int (*fn)(struct net_device *dev, 7632 struct netdev_nested_priv *priv), 7633 struct netdev_nested_priv *priv) 7634 { 7635 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7636 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7637 int ret, cur = 0; 7638 7639 now = dev; 7640 iter = &dev->adj_list.lower; 7641 7642 while (1) { 7643 if (now != dev) { 7644 ret = fn(now, priv); 7645 if (ret) 7646 return ret; 7647 } 7648 7649 next = NULL; 7650 while (1) { 7651 ldev = netdev_next_lower_dev_rcu(now, &iter); 7652 if (!ldev) 7653 break; 7654 7655 next = ldev; 7656 niter = &ldev->adj_list.lower; 7657 dev_stack[cur] = now; 7658 iter_stack[cur++] = iter; 7659 break; 7660 } 7661 7662 if (!next) { 7663 if (!cur) 7664 return 0; 7665 next = dev_stack[--cur]; 7666 niter = iter_stack[cur]; 7667 } 7668 7669 now = next; 7670 iter = niter; 7671 } 7672 7673 return 0; 7674 } 7675 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 7676 7677 /** 7678 * netdev_lower_get_first_private_rcu - Get the first ->private from the 7679 * lower neighbour list, RCU 7680 * variant 7681 * @dev: device 7682 * 7683 * Gets the first netdev_adjacent->private from the dev's lower neighbour 7684 * list. The caller must hold RCU read lock. 7685 */ 7686 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 7687 { 7688 struct netdev_adjacent *lower; 7689 7690 lower = list_first_or_null_rcu(&dev->adj_list.lower, 7691 struct netdev_adjacent, list); 7692 if (lower) 7693 return lower->private; 7694 return NULL; 7695 } 7696 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 7697 7698 /** 7699 * netdev_master_upper_dev_get_rcu - Get master upper device 7700 * @dev: device 7701 * 7702 * Find a master upper device and return pointer to it or NULL in case 7703 * it's not there. The caller must hold the RCU read lock. 7704 */ 7705 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 7706 { 7707 struct netdev_adjacent *upper; 7708 7709 upper = list_first_or_null_rcu(&dev->adj_list.upper, 7710 struct netdev_adjacent, list); 7711 if (upper && likely(upper->master)) 7712 return upper->dev; 7713 return NULL; 7714 } 7715 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 7716 7717 static int netdev_adjacent_sysfs_add(struct net_device *dev, 7718 struct net_device *adj_dev, 7719 struct list_head *dev_list) 7720 { 7721 char linkname[IFNAMSIZ+7]; 7722 7723 sprintf(linkname, dev_list == &dev->adj_list.upper ? 7724 "upper_%s" : "lower_%s", adj_dev->name); 7725 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 7726 linkname); 7727 } 7728 static void netdev_adjacent_sysfs_del(struct net_device *dev, 7729 char *name, 7730 struct list_head *dev_list) 7731 { 7732 char linkname[IFNAMSIZ+7]; 7733 7734 sprintf(linkname, dev_list == &dev->adj_list.upper ? 7735 "upper_%s" : "lower_%s", name); 7736 sysfs_remove_link(&(dev->dev.kobj), linkname); 7737 } 7738 7739 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 7740 struct net_device *adj_dev, 7741 struct list_head *dev_list) 7742 { 7743 return (dev_list == &dev->adj_list.upper || 7744 dev_list == &dev->adj_list.lower) && 7745 net_eq(dev_net(dev), dev_net(adj_dev)); 7746 } 7747 7748 static int __netdev_adjacent_dev_insert(struct net_device *dev, 7749 struct net_device *adj_dev, 7750 struct list_head *dev_list, 7751 void *private, bool master) 7752 { 7753 struct netdev_adjacent *adj; 7754 int ret; 7755 7756 adj = __netdev_find_adj(adj_dev, dev_list); 7757 7758 if (adj) { 7759 adj->ref_nr += 1; 7760 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 7761 dev->name, adj_dev->name, adj->ref_nr); 7762 7763 return 0; 7764 } 7765 7766 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 7767 if (!adj) 7768 return -ENOMEM; 7769 7770 adj->dev = adj_dev; 7771 adj->master = master; 7772 adj->ref_nr = 1; 7773 adj->private = private; 7774 adj->ignore = false; 7775 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 7776 7777 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 7778 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 7779 7780 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 7781 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 7782 if (ret) 7783 goto free_adj; 7784 } 7785 7786 /* Ensure that master link is always the first item in list. */ 7787 if (master) { 7788 ret = sysfs_create_link(&(dev->dev.kobj), 7789 &(adj_dev->dev.kobj), "master"); 7790 if (ret) 7791 goto remove_symlinks; 7792 7793 list_add_rcu(&adj->list, dev_list); 7794 } else { 7795 list_add_tail_rcu(&adj->list, dev_list); 7796 } 7797 7798 return 0; 7799 7800 remove_symlinks: 7801 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 7802 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 7803 free_adj: 7804 netdev_put(adj_dev, &adj->dev_tracker); 7805 kfree(adj); 7806 7807 return ret; 7808 } 7809 7810 static void __netdev_adjacent_dev_remove(struct net_device *dev, 7811 struct net_device *adj_dev, 7812 u16 ref_nr, 7813 struct list_head *dev_list) 7814 { 7815 struct netdev_adjacent *adj; 7816 7817 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 7818 dev->name, adj_dev->name, ref_nr); 7819 7820 adj = __netdev_find_adj(adj_dev, dev_list); 7821 7822 if (!adj) { 7823 pr_err("Adjacency does not exist for device %s from %s\n", 7824 dev->name, adj_dev->name); 7825 WARN_ON(1); 7826 return; 7827 } 7828 7829 if (adj->ref_nr > ref_nr) { 7830 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 7831 dev->name, adj_dev->name, ref_nr, 7832 adj->ref_nr - ref_nr); 7833 adj->ref_nr -= ref_nr; 7834 return; 7835 } 7836 7837 if (adj->master) 7838 sysfs_remove_link(&(dev->dev.kobj), "master"); 7839 7840 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 7841 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 7842 7843 list_del_rcu(&adj->list); 7844 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 7845 adj_dev->name, dev->name, adj_dev->name); 7846 netdev_put(adj_dev, &adj->dev_tracker); 7847 kfree_rcu(adj, rcu); 7848 } 7849 7850 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 7851 struct net_device *upper_dev, 7852 struct list_head *up_list, 7853 struct list_head *down_list, 7854 void *private, bool master) 7855 { 7856 int ret; 7857 7858 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 7859 private, master); 7860 if (ret) 7861 return ret; 7862 7863 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 7864 private, false); 7865 if (ret) { 7866 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 7867 return ret; 7868 } 7869 7870 return 0; 7871 } 7872 7873 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 7874 struct net_device *upper_dev, 7875 u16 ref_nr, 7876 struct list_head *up_list, 7877 struct list_head *down_list) 7878 { 7879 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 7880 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 7881 } 7882 7883 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 7884 struct net_device *upper_dev, 7885 void *private, bool master) 7886 { 7887 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 7888 &dev->adj_list.upper, 7889 &upper_dev->adj_list.lower, 7890 private, master); 7891 } 7892 7893 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 7894 struct net_device *upper_dev) 7895 { 7896 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 7897 &dev->adj_list.upper, 7898 &upper_dev->adj_list.lower); 7899 } 7900 7901 static int __netdev_upper_dev_link(struct net_device *dev, 7902 struct net_device *upper_dev, bool master, 7903 void *upper_priv, void *upper_info, 7904 struct netdev_nested_priv *priv, 7905 struct netlink_ext_ack *extack) 7906 { 7907 struct netdev_notifier_changeupper_info changeupper_info = { 7908 .info = { 7909 .dev = dev, 7910 .extack = extack, 7911 }, 7912 .upper_dev = upper_dev, 7913 .master = master, 7914 .linking = true, 7915 .upper_info = upper_info, 7916 }; 7917 struct net_device *master_dev; 7918 int ret = 0; 7919 7920 ASSERT_RTNL(); 7921 7922 if (dev == upper_dev) 7923 return -EBUSY; 7924 7925 /* To prevent loops, check if dev is not upper device to upper_dev. */ 7926 if (__netdev_has_upper_dev(upper_dev, dev)) 7927 return -EBUSY; 7928 7929 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 7930 return -EMLINK; 7931 7932 if (!master) { 7933 if (__netdev_has_upper_dev(dev, upper_dev)) 7934 return -EEXIST; 7935 } else { 7936 master_dev = __netdev_master_upper_dev_get(dev); 7937 if (master_dev) 7938 return master_dev == upper_dev ? -EEXIST : -EBUSY; 7939 } 7940 7941 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 7942 &changeupper_info.info); 7943 ret = notifier_to_errno(ret); 7944 if (ret) 7945 return ret; 7946 7947 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 7948 master); 7949 if (ret) 7950 return ret; 7951 7952 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 7953 &changeupper_info.info); 7954 ret = notifier_to_errno(ret); 7955 if (ret) 7956 goto rollback; 7957 7958 __netdev_update_upper_level(dev, NULL); 7959 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 7960 7961 __netdev_update_lower_level(upper_dev, priv); 7962 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 7963 priv); 7964 7965 return 0; 7966 7967 rollback: 7968 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 7969 7970 return ret; 7971 } 7972 7973 /** 7974 * netdev_upper_dev_link - Add a link to the upper device 7975 * @dev: device 7976 * @upper_dev: new upper device 7977 * @extack: netlink extended ack 7978 * 7979 * Adds a link to device which is upper to this one. The caller must hold 7980 * the RTNL lock. On a failure a negative errno code is returned. 7981 * On success the reference counts are adjusted and the function 7982 * returns zero. 7983 */ 7984 int netdev_upper_dev_link(struct net_device *dev, 7985 struct net_device *upper_dev, 7986 struct netlink_ext_ack *extack) 7987 { 7988 struct netdev_nested_priv priv = { 7989 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 7990 .data = NULL, 7991 }; 7992 7993 return __netdev_upper_dev_link(dev, upper_dev, false, 7994 NULL, NULL, &priv, extack); 7995 } 7996 EXPORT_SYMBOL(netdev_upper_dev_link); 7997 7998 /** 7999 * netdev_master_upper_dev_link - Add a master link to the upper device 8000 * @dev: device 8001 * @upper_dev: new upper device 8002 * @upper_priv: upper device private 8003 * @upper_info: upper info to be passed down via notifier 8004 * @extack: netlink extended ack 8005 * 8006 * Adds a link to device which is upper to this one. In this case, only 8007 * one master upper device can be linked, although other non-master devices 8008 * might be linked as well. The caller must hold the RTNL lock. 8009 * On a failure a negative errno code is returned. On success the reference 8010 * counts are adjusted and the function returns zero. 8011 */ 8012 int netdev_master_upper_dev_link(struct net_device *dev, 8013 struct net_device *upper_dev, 8014 void *upper_priv, void *upper_info, 8015 struct netlink_ext_ack *extack) 8016 { 8017 struct netdev_nested_priv priv = { 8018 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8019 .data = NULL, 8020 }; 8021 8022 return __netdev_upper_dev_link(dev, upper_dev, true, 8023 upper_priv, upper_info, &priv, extack); 8024 } 8025 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8026 8027 static void __netdev_upper_dev_unlink(struct net_device *dev, 8028 struct net_device *upper_dev, 8029 struct netdev_nested_priv *priv) 8030 { 8031 struct netdev_notifier_changeupper_info changeupper_info = { 8032 .info = { 8033 .dev = dev, 8034 }, 8035 .upper_dev = upper_dev, 8036 .linking = false, 8037 }; 8038 8039 ASSERT_RTNL(); 8040 8041 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8042 8043 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8044 &changeupper_info.info); 8045 8046 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8047 8048 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8049 &changeupper_info.info); 8050 8051 __netdev_update_upper_level(dev, NULL); 8052 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8053 8054 __netdev_update_lower_level(upper_dev, priv); 8055 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8056 priv); 8057 } 8058 8059 /** 8060 * netdev_upper_dev_unlink - Removes a link to upper device 8061 * @dev: device 8062 * @upper_dev: new upper device 8063 * 8064 * Removes a link to device which is upper to this one. The caller must hold 8065 * the RTNL lock. 8066 */ 8067 void netdev_upper_dev_unlink(struct net_device *dev, 8068 struct net_device *upper_dev) 8069 { 8070 struct netdev_nested_priv priv = { 8071 .flags = NESTED_SYNC_TODO, 8072 .data = NULL, 8073 }; 8074 8075 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 8076 } 8077 EXPORT_SYMBOL(netdev_upper_dev_unlink); 8078 8079 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 8080 struct net_device *lower_dev, 8081 bool val) 8082 { 8083 struct netdev_adjacent *adj; 8084 8085 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 8086 if (adj) 8087 adj->ignore = val; 8088 8089 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 8090 if (adj) 8091 adj->ignore = val; 8092 } 8093 8094 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 8095 struct net_device *lower_dev) 8096 { 8097 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 8098 } 8099 8100 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 8101 struct net_device *lower_dev) 8102 { 8103 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 8104 } 8105 8106 int netdev_adjacent_change_prepare(struct net_device *old_dev, 8107 struct net_device *new_dev, 8108 struct net_device *dev, 8109 struct netlink_ext_ack *extack) 8110 { 8111 struct netdev_nested_priv priv = { 8112 .flags = 0, 8113 .data = NULL, 8114 }; 8115 int err; 8116 8117 if (!new_dev) 8118 return 0; 8119 8120 if (old_dev && new_dev != old_dev) 8121 netdev_adjacent_dev_disable(dev, old_dev); 8122 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 8123 extack); 8124 if (err) { 8125 if (old_dev && new_dev != old_dev) 8126 netdev_adjacent_dev_enable(dev, old_dev); 8127 return err; 8128 } 8129 8130 return 0; 8131 } 8132 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 8133 8134 void netdev_adjacent_change_commit(struct net_device *old_dev, 8135 struct net_device *new_dev, 8136 struct net_device *dev) 8137 { 8138 struct netdev_nested_priv priv = { 8139 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8140 .data = NULL, 8141 }; 8142 8143 if (!new_dev || !old_dev) 8144 return; 8145 8146 if (new_dev == old_dev) 8147 return; 8148 8149 netdev_adjacent_dev_enable(dev, old_dev); 8150 __netdev_upper_dev_unlink(old_dev, dev, &priv); 8151 } 8152 EXPORT_SYMBOL(netdev_adjacent_change_commit); 8153 8154 void netdev_adjacent_change_abort(struct net_device *old_dev, 8155 struct net_device *new_dev, 8156 struct net_device *dev) 8157 { 8158 struct netdev_nested_priv priv = { 8159 .flags = 0, 8160 .data = NULL, 8161 }; 8162 8163 if (!new_dev) 8164 return; 8165 8166 if (old_dev && new_dev != old_dev) 8167 netdev_adjacent_dev_enable(dev, old_dev); 8168 8169 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8170 } 8171 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8172 8173 /** 8174 * netdev_bonding_info_change - Dispatch event about slave change 8175 * @dev: device 8176 * @bonding_info: info to dispatch 8177 * 8178 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8179 * The caller must hold the RTNL lock. 8180 */ 8181 void netdev_bonding_info_change(struct net_device *dev, 8182 struct netdev_bonding_info *bonding_info) 8183 { 8184 struct netdev_notifier_bonding_info info = { 8185 .info.dev = dev, 8186 }; 8187 8188 memcpy(&info.bonding_info, bonding_info, 8189 sizeof(struct netdev_bonding_info)); 8190 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8191 &info.info); 8192 } 8193 EXPORT_SYMBOL(netdev_bonding_info_change); 8194 8195 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8196 struct netlink_ext_ack *extack) 8197 { 8198 struct netdev_notifier_offload_xstats_info info = { 8199 .info.dev = dev, 8200 .info.extack = extack, 8201 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8202 }; 8203 int err; 8204 int rc; 8205 8206 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8207 GFP_KERNEL); 8208 if (!dev->offload_xstats_l3) 8209 return -ENOMEM; 8210 8211 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8212 NETDEV_OFFLOAD_XSTATS_DISABLE, 8213 &info.info); 8214 err = notifier_to_errno(rc); 8215 if (err) 8216 goto free_stats; 8217 8218 return 0; 8219 8220 free_stats: 8221 kfree(dev->offload_xstats_l3); 8222 dev->offload_xstats_l3 = NULL; 8223 return err; 8224 } 8225 8226 int netdev_offload_xstats_enable(struct net_device *dev, 8227 enum netdev_offload_xstats_type type, 8228 struct netlink_ext_ack *extack) 8229 { 8230 ASSERT_RTNL(); 8231 8232 if (netdev_offload_xstats_enabled(dev, type)) 8233 return -EALREADY; 8234 8235 switch (type) { 8236 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8237 return netdev_offload_xstats_enable_l3(dev, extack); 8238 } 8239 8240 WARN_ON(1); 8241 return -EINVAL; 8242 } 8243 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8244 8245 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8246 { 8247 struct netdev_notifier_offload_xstats_info info = { 8248 .info.dev = dev, 8249 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8250 }; 8251 8252 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8253 &info.info); 8254 kfree(dev->offload_xstats_l3); 8255 dev->offload_xstats_l3 = NULL; 8256 } 8257 8258 int netdev_offload_xstats_disable(struct net_device *dev, 8259 enum netdev_offload_xstats_type type) 8260 { 8261 ASSERT_RTNL(); 8262 8263 if (!netdev_offload_xstats_enabled(dev, type)) 8264 return -EALREADY; 8265 8266 switch (type) { 8267 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8268 netdev_offload_xstats_disable_l3(dev); 8269 return 0; 8270 } 8271 8272 WARN_ON(1); 8273 return -EINVAL; 8274 } 8275 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8276 8277 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8278 { 8279 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8280 } 8281 8282 static struct rtnl_hw_stats64 * 8283 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8284 enum netdev_offload_xstats_type type) 8285 { 8286 switch (type) { 8287 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8288 return dev->offload_xstats_l3; 8289 } 8290 8291 WARN_ON(1); 8292 return NULL; 8293 } 8294 8295 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8296 enum netdev_offload_xstats_type type) 8297 { 8298 ASSERT_RTNL(); 8299 8300 return netdev_offload_xstats_get_ptr(dev, type); 8301 } 8302 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8303 8304 struct netdev_notifier_offload_xstats_ru { 8305 bool used; 8306 }; 8307 8308 struct netdev_notifier_offload_xstats_rd { 8309 struct rtnl_hw_stats64 stats; 8310 bool used; 8311 }; 8312 8313 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8314 const struct rtnl_hw_stats64 *src) 8315 { 8316 dest->rx_packets += src->rx_packets; 8317 dest->tx_packets += src->tx_packets; 8318 dest->rx_bytes += src->rx_bytes; 8319 dest->tx_bytes += src->tx_bytes; 8320 dest->rx_errors += src->rx_errors; 8321 dest->tx_errors += src->tx_errors; 8322 dest->rx_dropped += src->rx_dropped; 8323 dest->tx_dropped += src->tx_dropped; 8324 dest->multicast += src->multicast; 8325 } 8326 8327 static int netdev_offload_xstats_get_used(struct net_device *dev, 8328 enum netdev_offload_xstats_type type, 8329 bool *p_used, 8330 struct netlink_ext_ack *extack) 8331 { 8332 struct netdev_notifier_offload_xstats_ru report_used = {}; 8333 struct netdev_notifier_offload_xstats_info info = { 8334 .info.dev = dev, 8335 .info.extack = extack, 8336 .type = type, 8337 .report_used = &report_used, 8338 }; 8339 int rc; 8340 8341 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8342 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8343 &info.info); 8344 *p_used = report_used.used; 8345 return notifier_to_errno(rc); 8346 } 8347 8348 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8349 enum netdev_offload_xstats_type type, 8350 struct rtnl_hw_stats64 *p_stats, 8351 bool *p_used, 8352 struct netlink_ext_ack *extack) 8353 { 8354 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8355 struct netdev_notifier_offload_xstats_info info = { 8356 .info.dev = dev, 8357 .info.extack = extack, 8358 .type = type, 8359 .report_delta = &report_delta, 8360 }; 8361 struct rtnl_hw_stats64 *stats; 8362 int rc; 8363 8364 stats = netdev_offload_xstats_get_ptr(dev, type); 8365 if (WARN_ON(!stats)) 8366 return -EINVAL; 8367 8368 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8369 &info.info); 8370 8371 /* Cache whatever we got, even if there was an error, otherwise the 8372 * successful stats retrievals would get lost. 8373 */ 8374 netdev_hw_stats64_add(stats, &report_delta.stats); 8375 8376 if (p_stats) 8377 *p_stats = *stats; 8378 *p_used = report_delta.used; 8379 8380 return notifier_to_errno(rc); 8381 } 8382 8383 int netdev_offload_xstats_get(struct net_device *dev, 8384 enum netdev_offload_xstats_type type, 8385 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8386 struct netlink_ext_ack *extack) 8387 { 8388 ASSERT_RTNL(); 8389 8390 if (p_stats) 8391 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8392 p_used, extack); 8393 else 8394 return netdev_offload_xstats_get_used(dev, type, p_used, 8395 extack); 8396 } 8397 EXPORT_SYMBOL(netdev_offload_xstats_get); 8398 8399 void 8400 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8401 const struct rtnl_hw_stats64 *stats) 8402 { 8403 report_delta->used = true; 8404 netdev_hw_stats64_add(&report_delta->stats, stats); 8405 } 8406 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 8407 8408 void 8409 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 8410 { 8411 report_used->used = true; 8412 } 8413 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 8414 8415 void netdev_offload_xstats_push_delta(struct net_device *dev, 8416 enum netdev_offload_xstats_type type, 8417 const struct rtnl_hw_stats64 *p_stats) 8418 { 8419 struct rtnl_hw_stats64 *stats; 8420 8421 ASSERT_RTNL(); 8422 8423 stats = netdev_offload_xstats_get_ptr(dev, type); 8424 if (WARN_ON(!stats)) 8425 return; 8426 8427 netdev_hw_stats64_add(stats, p_stats); 8428 } 8429 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 8430 8431 /** 8432 * netdev_get_xmit_slave - Get the xmit slave of master device 8433 * @dev: device 8434 * @skb: The packet 8435 * @all_slaves: assume all the slaves are active 8436 * 8437 * The reference counters are not incremented so the caller must be 8438 * careful with locks. The caller must hold RCU lock. 8439 * %NULL is returned if no slave is found. 8440 */ 8441 8442 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 8443 struct sk_buff *skb, 8444 bool all_slaves) 8445 { 8446 const struct net_device_ops *ops = dev->netdev_ops; 8447 8448 if (!ops->ndo_get_xmit_slave) 8449 return NULL; 8450 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 8451 } 8452 EXPORT_SYMBOL(netdev_get_xmit_slave); 8453 8454 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 8455 struct sock *sk) 8456 { 8457 const struct net_device_ops *ops = dev->netdev_ops; 8458 8459 if (!ops->ndo_sk_get_lower_dev) 8460 return NULL; 8461 return ops->ndo_sk_get_lower_dev(dev, sk); 8462 } 8463 8464 /** 8465 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 8466 * @dev: device 8467 * @sk: the socket 8468 * 8469 * %NULL is returned if no lower device is found. 8470 */ 8471 8472 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 8473 struct sock *sk) 8474 { 8475 struct net_device *lower; 8476 8477 lower = netdev_sk_get_lower_dev(dev, sk); 8478 while (lower) { 8479 dev = lower; 8480 lower = netdev_sk_get_lower_dev(dev, sk); 8481 } 8482 8483 return dev; 8484 } 8485 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 8486 8487 static void netdev_adjacent_add_links(struct net_device *dev) 8488 { 8489 struct netdev_adjacent *iter; 8490 8491 struct net *net = dev_net(dev); 8492 8493 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8494 if (!net_eq(net, dev_net(iter->dev))) 8495 continue; 8496 netdev_adjacent_sysfs_add(iter->dev, dev, 8497 &iter->dev->adj_list.lower); 8498 netdev_adjacent_sysfs_add(dev, iter->dev, 8499 &dev->adj_list.upper); 8500 } 8501 8502 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8503 if (!net_eq(net, dev_net(iter->dev))) 8504 continue; 8505 netdev_adjacent_sysfs_add(iter->dev, dev, 8506 &iter->dev->adj_list.upper); 8507 netdev_adjacent_sysfs_add(dev, iter->dev, 8508 &dev->adj_list.lower); 8509 } 8510 } 8511 8512 static void netdev_adjacent_del_links(struct net_device *dev) 8513 { 8514 struct netdev_adjacent *iter; 8515 8516 struct net *net = dev_net(dev); 8517 8518 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8519 if (!net_eq(net, dev_net(iter->dev))) 8520 continue; 8521 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8522 &iter->dev->adj_list.lower); 8523 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8524 &dev->adj_list.upper); 8525 } 8526 8527 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8528 if (!net_eq(net, dev_net(iter->dev))) 8529 continue; 8530 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8531 &iter->dev->adj_list.upper); 8532 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8533 &dev->adj_list.lower); 8534 } 8535 } 8536 8537 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 8538 { 8539 struct netdev_adjacent *iter; 8540 8541 struct net *net = dev_net(dev); 8542 8543 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8544 if (!net_eq(net, dev_net(iter->dev))) 8545 continue; 8546 netdev_adjacent_sysfs_del(iter->dev, oldname, 8547 &iter->dev->adj_list.lower); 8548 netdev_adjacent_sysfs_add(iter->dev, dev, 8549 &iter->dev->adj_list.lower); 8550 } 8551 8552 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8553 if (!net_eq(net, dev_net(iter->dev))) 8554 continue; 8555 netdev_adjacent_sysfs_del(iter->dev, oldname, 8556 &iter->dev->adj_list.upper); 8557 netdev_adjacent_sysfs_add(iter->dev, dev, 8558 &iter->dev->adj_list.upper); 8559 } 8560 } 8561 8562 void *netdev_lower_dev_get_private(struct net_device *dev, 8563 struct net_device *lower_dev) 8564 { 8565 struct netdev_adjacent *lower; 8566 8567 if (!lower_dev) 8568 return NULL; 8569 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 8570 if (!lower) 8571 return NULL; 8572 8573 return lower->private; 8574 } 8575 EXPORT_SYMBOL(netdev_lower_dev_get_private); 8576 8577 8578 /** 8579 * netdev_lower_state_changed - Dispatch event about lower device state change 8580 * @lower_dev: device 8581 * @lower_state_info: state to dispatch 8582 * 8583 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 8584 * The caller must hold the RTNL lock. 8585 */ 8586 void netdev_lower_state_changed(struct net_device *lower_dev, 8587 void *lower_state_info) 8588 { 8589 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 8590 .info.dev = lower_dev, 8591 }; 8592 8593 ASSERT_RTNL(); 8594 changelowerstate_info.lower_state_info = lower_state_info; 8595 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 8596 &changelowerstate_info.info); 8597 } 8598 EXPORT_SYMBOL(netdev_lower_state_changed); 8599 8600 static void dev_change_rx_flags(struct net_device *dev, int flags) 8601 { 8602 const struct net_device_ops *ops = dev->netdev_ops; 8603 8604 if (ops->ndo_change_rx_flags) 8605 ops->ndo_change_rx_flags(dev, flags); 8606 } 8607 8608 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 8609 { 8610 unsigned int old_flags = dev->flags; 8611 unsigned int promiscuity, flags; 8612 kuid_t uid; 8613 kgid_t gid; 8614 8615 ASSERT_RTNL(); 8616 8617 promiscuity = dev->promiscuity + inc; 8618 if (promiscuity == 0) { 8619 /* 8620 * Avoid overflow. 8621 * If inc causes overflow, untouch promisc and return error. 8622 */ 8623 if (unlikely(inc > 0)) { 8624 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 8625 return -EOVERFLOW; 8626 } 8627 flags = old_flags & ~IFF_PROMISC; 8628 } else { 8629 flags = old_flags | IFF_PROMISC; 8630 } 8631 WRITE_ONCE(dev->promiscuity, promiscuity); 8632 if (flags != old_flags) { 8633 WRITE_ONCE(dev->flags, flags); 8634 netdev_info(dev, "%s promiscuous mode\n", 8635 dev->flags & IFF_PROMISC ? "entered" : "left"); 8636 if (audit_enabled) { 8637 current_uid_gid(&uid, &gid); 8638 audit_log(audit_context(), GFP_ATOMIC, 8639 AUDIT_ANOM_PROMISCUOUS, 8640 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 8641 dev->name, (dev->flags & IFF_PROMISC), 8642 (old_flags & IFF_PROMISC), 8643 from_kuid(&init_user_ns, audit_get_loginuid(current)), 8644 from_kuid(&init_user_ns, uid), 8645 from_kgid(&init_user_ns, gid), 8646 audit_get_sessionid(current)); 8647 } 8648 8649 dev_change_rx_flags(dev, IFF_PROMISC); 8650 } 8651 if (notify) 8652 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 8653 return 0; 8654 } 8655 8656 /** 8657 * dev_set_promiscuity - update promiscuity count on a device 8658 * @dev: device 8659 * @inc: modifier 8660 * 8661 * Add or remove promiscuity from a device. While the count in the device 8662 * remains above zero the interface remains promiscuous. Once it hits zero 8663 * the device reverts back to normal filtering operation. A negative inc 8664 * value is used to drop promiscuity on the device. 8665 * Return 0 if successful or a negative errno code on error. 8666 */ 8667 int dev_set_promiscuity(struct net_device *dev, int inc) 8668 { 8669 unsigned int old_flags = dev->flags; 8670 int err; 8671 8672 err = __dev_set_promiscuity(dev, inc, true); 8673 if (err < 0) 8674 return err; 8675 if (dev->flags != old_flags) 8676 dev_set_rx_mode(dev); 8677 return err; 8678 } 8679 EXPORT_SYMBOL(dev_set_promiscuity); 8680 8681 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify) 8682 { 8683 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 8684 unsigned int allmulti, flags; 8685 8686 ASSERT_RTNL(); 8687 8688 allmulti = dev->allmulti + inc; 8689 if (allmulti == 0) { 8690 /* 8691 * Avoid overflow. 8692 * If inc causes overflow, untouch allmulti and return error. 8693 */ 8694 if (unlikely(inc > 0)) { 8695 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 8696 return -EOVERFLOW; 8697 } 8698 flags = old_flags & ~IFF_ALLMULTI; 8699 } else { 8700 flags = old_flags | IFF_ALLMULTI; 8701 } 8702 WRITE_ONCE(dev->allmulti, allmulti); 8703 if (flags != old_flags) { 8704 WRITE_ONCE(dev->flags, flags); 8705 netdev_info(dev, "%s allmulticast mode\n", 8706 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 8707 dev_change_rx_flags(dev, IFF_ALLMULTI); 8708 dev_set_rx_mode(dev); 8709 if (notify) 8710 __dev_notify_flags(dev, old_flags, 8711 dev->gflags ^ old_gflags, 0, NULL); 8712 } 8713 return 0; 8714 } 8715 8716 /** 8717 * dev_set_allmulti - update allmulti count on a device 8718 * @dev: device 8719 * @inc: modifier 8720 * 8721 * Add or remove reception of all multicast frames to a device. While the 8722 * count in the device remains above zero the interface remains listening 8723 * to all interfaces. Once it hits zero the device reverts back to normal 8724 * filtering operation. A negative @inc value is used to drop the counter 8725 * when releasing a resource needing all multicasts. 8726 * Return 0 if successful or a negative errno code on error. 8727 */ 8728 8729 int dev_set_allmulti(struct net_device *dev, int inc) 8730 { 8731 return __dev_set_allmulti(dev, inc, true); 8732 } 8733 EXPORT_SYMBOL(dev_set_allmulti); 8734 8735 /* 8736 * Upload unicast and multicast address lists to device and 8737 * configure RX filtering. When the device doesn't support unicast 8738 * filtering it is put in promiscuous mode while unicast addresses 8739 * are present. 8740 */ 8741 void __dev_set_rx_mode(struct net_device *dev) 8742 { 8743 const struct net_device_ops *ops = dev->netdev_ops; 8744 8745 /* dev_open will call this function so the list will stay sane. */ 8746 if (!(dev->flags&IFF_UP)) 8747 return; 8748 8749 if (!netif_device_present(dev)) 8750 return; 8751 8752 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 8753 /* Unicast addresses changes may only happen under the rtnl, 8754 * therefore calling __dev_set_promiscuity here is safe. 8755 */ 8756 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 8757 __dev_set_promiscuity(dev, 1, false); 8758 dev->uc_promisc = true; 8759 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 8760 __dev_set_promiscuity(dev, -1, false); 8761 dev->uc_promisc = false; 8762 } 8763 } 8764 8765 if (ops->ndo_set_rx_mode) 8766 ops->ndo_set_rx_mode(dev); 8767 } 8768 8769 void dev_set_rx_mode(struct net_device *dev) 8770 { 8771 netif_addr_lock_bh(dev); 8772 __dev_set_rx_mode(dev); 8773 netif_addr_unlock_bh(dev); 8774 } 8775 8776 /** 8777 * dev_get_flags - get flags reported to userspace 8778 * @dev: device 8779 * 8780 * Get the combination of flag bits exported through APIs to userspace. 8781 */ 8782 unsigned int dev_get_flags(const struct net_device *dev) 8783 { 8784 unsigned int flags; 8785 8786 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 8787 IFF_ALLMULTI | 8788 IFF_RUNNING | 8789 IFF_LOWER_UP | 8790 IFF_DORMANT)) | 8791 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 8792 IFF_ALLMULTI)); 8793 8794 if (netif_running(dev)) { 8795 if (netif_oper_up(dev)) 8796 flags |= IFF_RUNNING; 8797 if (netif_carrier_ok(dev)) 8798 flags |= IFF_LOWER_UP; 8799 if (netif_dormant(dev)) 8800 flags |= IFF_DORMANT; 8801 } 8802 8803 return flags; 8804 } 8805 EXPORT_SYMBOL(dev_get_flags); 8806 8807 int __dev_change_flags(struct net_device *dev, unsigned int flags, 8808 struct netlink_ext_ack *extack) 8809 { 8810 unsigned int old_flags = dev->flags; 8811 int ret; 8812 8813 ASSERT_RTNL(); 8814 8815 /* 8816 * Set the flags on our device. 8817 */ 8818 8819 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 8820 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 8821 IFF_AUTOMEDIA)) | 8822 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 8823 IFF_ALLMULTI)); 8824 8825 /* 8826 * Load in the correct multicast list now the flags have changed. 8827 */ 8828 8829 if ((old_flags ^ flags) & IFF_MULTICAST) 8830 dev_change_rx_flags(dev, IFF_MULTICAST); 8831 8832 dev_set_rx_mode(dev); 8833 8834 /* 8835 * Have we downed the interface. We handle IFF_UP ourselves 8836 * according to user attempts to set it, rather than blindly 8837 * setting it. 8838 */ 8839 8840 ret = 0; 8841 if ((old_flags ^ flags) & IFF_UP) { 8842 if (old_flags & IFF_UP) 8843 __dev_close(dev); 8844 else 8845 ret = __dev_open(dev, extack); 8846 } 8847 8848 if ((flags ^ dev->gflags) & IFF_PROMISC) { 8849 int inc = (flags & IFF_PROMISC) ? 1 : -1; 8850 unsigned int old_flags = dev->flags; 8851 8852 dev->gflags ^= IFF_PROMISC; 8853 8854 if (__dev_set_promiscuity(dev, inc, false) >= 0) 8855 if (dev->flags != old_flags) 8856 dev_set_rx_mode(dev); 8857 } 8858 8859 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 8860 * is important. Some (broken) drivers set IFF_PROMISC, when 8861 * IFF_ALLMULTI is requested not asking us and not reporting. 8862 */ 8863 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 8864 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 8865 8866 dev->gflags ^= IFF_ALLMULTI; 8867 __dev_set_allmulti(dev, inc, false); 8868 } 8869 8870 return ret; 8871 } 8872 8873 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 8874 unsigned int gchanges, u32 portid, 8875 const struct nlmsghdr *nlh) 8876 { 8877 unsigned int changes = dev->flags ^ old_flags; 8878 8879 if (gchanges) 8880 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 8881 8882 if (changes & IFF_UP) { 8883 if (dev->flags & IFF_UP) 8884 call_netdevice_notifiers(NETDEV_UP, dev); 8885 else 8886 call_netdevice_notifiers(NETDEV_DOWN, dev); 8887 } 8888 8889 if (dev->flags & IFF_UP && 8890 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 8891 struct netdev_notifier_change_info change_info = { 8892 .info = { 8893 .dev = dev, 8894 }, 8895 .flags_changed = changes, 8896 }; 8897 8898 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 8899 } 8900 } 8901 8902 /** 8903 * dev_change_flags - change device settings 8904 * @dev: device 8905 * @flags: device state flags 8906 * @extack: netlink extended ack 8907 * 8908 * Change settings on device based state flags. The flags are 8909 * in the userspace exported format. 8910 */ 8911 int dev_change_flags(struct net_device *dev, unsigned int flags, 8912 struct netlink_ext_ack *extack) 8913 { 8914 int ret; 8915 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 8916 8917 ret = __dev_change_flags(dev, flags, extack); 8918 if (ret < 0) 8919 return ret; 8920 8921 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 8922 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 8923 return ret; 8924 } 8925 EXPORT_SYMBOL(dev_change_flags); 8926 8927 int __dev_set_mtu(struct net_device *dev, int new_mtu) 8928 { 8929 const struct net_device_ops *ops = dev->netdev_ops; 8930 8931 if (ops->ndo_change_mtu) 8932 return ops->ndo_change_mtu(dev, new_mtu); 8933 8934 /* Pairs with all the lockless reads of dev->mtu in the stack */ 8935 WRITE_ONCE(dev->mtu, new_mtu); 8936 return 0; 8937 } 8938 EXPORT_SYMBOL(__dev_set_mtu); 8939 8940 int dev_validate_mtu(struct net_device *dev, int new_mtu, 8941 struct netlink_ext_ack *extack) 8942 { 8943 /* MTU must be positive, and in range */ 8944 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 8945 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 8946 return -EINVAL; 8947 } 8948 8949 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 8950 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 8951 return -EINVAL; 8952 } 8953 return 0; 8954 } 8955 8956 /** 8957 * dev_set_mtu_ext - Change maximum transfer unit 8958 * @dev: device 8959 * @new_mtu: new transfer unit 8960 * @extack: netlink extended ack 8961 * 8962 * Change the maximum transfer size of the network device. 8963 */ 8964 int dev_set_mtu_ext(struct net_device *dev, int new_mtu, 8965 struct netlink_ext_ack *extack) 8966 { 8967 int err, orig_mtu; 8968 8969 if (new_mtu == dev->mtu) 8970 return 0; 8971 8972 err = dev_validate_mtu(dev, new_mtu, extack); 8973 if (err) 8974 return err; 8975 8976 if (!netif_device_present(dev)) 8977 return -ENODEV; 8978 8979 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 8980 err = notifier_to_errno(err); 8981 if (err) 8982 return err; 8983 8984 orig_mtu = dev->mtu; 8985 err = __dev_set_mtu(dev, new_mtu); 8986 8987 if (!err) { 8988 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 8989 orig_mtu); 8990 err = notifier_to_errno(err); 8991 if (err) { 8992 /* setting mtu back and notifying everyone again, 8993 * so that they have a chance to revert changes. 8994 */ 8995 __dev_set_mtu(dev, orig_mtu); 8996 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 8997 new_mtu); 8998 } 8999 } 9000 return err; 9001 } 9002 9003 int dev_set_mtu(struct net_device *dev, int new_mtu) 9004 { 9005 struct netlink_ext_ack extack; 9006 int err; 9007 9008 memset(&extack, 0, sizeof(extack)); 9009 err = dev_set_mtu_ext(dev, new_mtu, &extack); 9010 if (err && extack._msg) 9011 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9012 return err; 9013 } 9014 EXPORT_SYMBOL(dev_set_mtu); 9015 9016 /** 9017 * dev_change_tx_queue_len - Change TX queue length of a netdevice 9018 * @dev: device 9019 * @new_len: new tx queue length 9020 */ 9021 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9022 { 9023 unsigned int orig_len = dev->tx_queue_len; 9024 int res; 9025 9026 if (new_len != (unsigned int)new_len) 9027 return -ERANGE; 9028 9029 if (new_len != orig_len) { 9030 WRITE_ONCE(dev->tx_queue_len, new_len); 9031 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9032 res = notifier_to_errno(res); 9033 if (res) 9034 goto err_rollback; 9035 res = dev_qdisc_change_tx_queue_len(dev); 9036 if (res) 9037 goto err_rollback; 9038 } 9039 9040 return 0; 9041 9042 err_rollback: 9043 netdev_err(dev, "refused to change device tx_queue_len\n"); 9044 WRITE_ONCE(dev->tx_queue_len, orig_len); 9045 return res; 9046 } 9047 9048 /** 9049 * dev_set_group - Change group this device belongs to 9050 * @dev: device 9051 * @new_group: group this device should belong to 9052 */ 9053 void dev_set_group(struct net_device *dev, int new_group) 9054 { 9055 dev->group = new_group; 9056 } 9057 9058 /** 9059 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 9060 * @dev: device 9061 * @addr: new address 9062 * @extack: netlink extended ack 9063 */ 9064 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9065 struct netlink_ext_ack *extack) 9066 { 9067 struct netdev_notifier_pre_changeaddr_info info = { 9068 .info.dev = dev, 9069 .info.extack = extack, 9070 .dev_addr = addr, 9071 }; 9072 int rc; 9073 9074 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9075 return notifier_to_errno(rc); 9076 } 9077 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 9078 9079 /** 9080 * dev_set_mac_address - Change Media Access Control Address 9081 * @dev: device 9082 * @sa: new address 9083 * @extack: netlink extended ack 9084 * 9085 * Change the hardware (MAC) address of the device 9086 */ 9087 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa, 9088 struct netlink_ext_ack *extack) 9089 { 9090 const struct net_device_ops *ops = dev->netdev_ops; 9091 int err; 9092 9093 if (!ops->ndo_set_mac_address) 9094 return -EOPNOTSUPP; 9095 if (sa->sa_family != dev->type) 9096 return -EINVAL; 9097 if (!netif_device_present(dev)) 9098 return -ENODEV; 9099 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 9100 if (err) 9101 return err; 9102 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 9103 err = ops->ndo_set_mac_address(dev, sa); 9104 if (err) 9105 return err; 9106 } 9107 dev->addr_assign_type = NET_ADDR_SET; 9108 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9109 add_device_randomness(dev->dev_addr, dev->addr_len); 9110 return 0; 9111 } 9112 EXPORT_SYMBOL(dev_set_mac_address); 9113 9114 DECLARE_RWSEM(dev_addr_sem); 9115 9116 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa, 9117 struct netlink_ext_ack *extack) 9118 { 9119 int ret; 9120 9121 down_write(&dev_addr_sem); 9122 ret = dev_set_mac_address(dev, sa, extack); 9123 up_write(&dev_addr_sem); 9124 return ret; 9125 } 9126 EXPORT_SYMBOL(dev_set_mac_address_user); 9127 9128 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9129 { 9130 size_t size = sizeof(sa->sa_data_min); 9131 struct net_device *dev; 9132 int ret = 0; 9133 9134 down_read(&dev_addr_sem); 9135 rcu_read_lock(); 9136 9137 dev = dev_get_by_name_rcu(net, dev_name); 9138 if (!dev) { 9139 ret = -ENODEV; 9140 goto unlock; 9141 } 9142 if (!dev->addr_len) 9143 memset(sa->sa_data, 0, size); 9144 else 9145 memcpy(sa->sa_data, dev->dev_addr, 9146 min_t(size_t, size, dev->addr_len)); 9147 sa->sa_family = dev->type; 9148 9149 unlock: 9150 rcu_read_unlock(); 9151 up_read(&dev_addr_sem); 9152 return ret; 9153 } 9154 EXPORT_SYMBOL(dev_get_mac_address); 9155 9156 /** 9157 * dev_change_carrier - Change device carrier 9158 * @dev: device 9159 * @new_carrier: new value 9160 * 9161 * Change device carrier 9162 */ 9163 int dev_change_carrier(struct net_device *dev, bool new_carrier) 9164 { 9165 const struct net_device_ops *ops = dev->netdev_ops; 9166 9167 if (!ops->ndo_change_carrier) 9168 return -EOPNOTSUPP; 9169 if (!netif_device_present(dev)) 9170 return -ENODEV; 9171 return ops->ndo_change_carrier(dev, new_carrier); 9172 } 9173 9174 /** 9175 * dev_get_phys_port_id - Get device physical port ID 9176 * @dev: device 9177 * @ppid: port ID 9178 * 9179 * Get device physical port ID 9180 */ 9181 int dev_get_phys_port_id(struct net_device *dev, 9182 struct netdev_phys_item_id *ppid) 9183 { 9184 const struct net_device_ops *ops = dev->netdev_ops; 9185 9186 if (!ops->ndo_get_phys_port_id) 9187 return -EOPNOTSUPP; 9188 return ops->ndo_get_phys_port_id(dev, ppid); 9189 } 9190 9191 /** 9192 * dev_get_phys_port_name - Get device physical port name 9193 * @dev: device 9194 * @name: port name 9195 * @len: limit of bytes to copy to name 9196 * 9197 * Get device physical port name 9198 */ 9199 int dev_get_phys_port_name(struct net_device *dev, 9200 char *name, size_t len) 9201 { 9202 const struct net_device_ops *ops = dev->netdev_ops; 9203 int err; 9204 9205 if (ops->ndo_get_phys_port_name) { 9206 err = ops->ndo_get_phys_port_name(dev, name, len); 9207 if (err != -EOPNOTSUPP) 9208 return err; 9209 } 9210 return devlink_compat_phys_port_name_get(dev, name, len); 9211 } 9212 9213 /** 9214 * dev_get_port_parent_id - Get the device's port parent identifier 9215 * @dev: network device 9216 * @ppid: pointer to a storage for the port's parent identifier 9217 * @recurse: allow/disallow recursion to lower devices 9218 * 9219 * Get the devices's port parent identifier 9220 */ 9221 int dev_get_port_parent_id(struct net_device *dev, 9222 struct netdev_phys_item_id *ppid, 9223 bool recurse) 9224 { 9225 const struct net_device_ops *ops = dev->netdev_ops; 9226 struct netdev_phys_item_id first = { }; 9227 struct net_device *lower_dev; 9228 struct list_head *iter; 9229 int err; 9230 9231 if (ops->ndo_get_port_parent_id) { 9232 err = ops->ndo_get_port_parent_id(dev, ppid); 9233 if (err != -EOPNOTSUPP) 9234 return err; 9235 } 9236 9237 err = devlink_compat_switch_id_get(dev, ppid); 9238 if (!recurse || err != -EOPNOTSUPP) 9239 return err; 9240 9241 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9242 err = dev_get_port_parent_id(lower_dev, ppid, true); 9243 if (err) 9244 break; 9245 if (!first.id_len) 9246 first = *ppid; 9247 else if (memcmp(&first, ppid, sizeof(*ppid))) 9248 return -EOPNOTSUPP; 9249 } 9250 9251 return err; 9252 } 9253 EXPORT_SYMBOL(dev_get_port_parent_id); 9254 9255 /** 9256 * netdev_port_same_parent_id - Indicate if two network devices have 9257 * the same port parent identifier 9258 * @a: first network device 9259 * @b: second network device 9260 */ 9261 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9262 { 9263 struct netdev_phys_item_id a_id = { }; 9264 struct netdev_phys_item_id b_id = { }; 9265 9266 if (dev_get_port_parent_id(a, &a_id, true) || 9267 dev_get_port_parent_id(b, &b_id, true)) 9268 return false; 9269 9270 return netdev_phys_item_id_same(&a_id, &b_id); 9271 } 9272 EXPORT_SYMBOL(netdev_port_same_parent_id); 9273 9274 /** 9275 * dev_change_proto_down - set carrier according to proto_down. 9276 * 9277 * @dev: device 9278 * @proto_down: new value 9279 */ 9280 int dev_change_proto_down(struct net_device *dev, bool proto_down) 9281 { 9282 if (!dev->change_proto_down) 9283 return -EOPNOTSUPP; 9284 if (!netif_device_present(dev)) 9285 return -ENODEV; 9286 if (proto_down) 9287 netif_carrier_off(dev); 9288 else 9289 netif_carrier_on(dev); 9290 WRITE_ONCE(dev->proto_down, proto_down); 9291 return 0; 9292 } 9293 9294 /** 9295 * dev_change_proto_down_reason - proto down reason 9296 * 9297 * @dev: device 9298 * @mask: proto down mask 9299 * @value: proto down value 9300 */ 9301 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask, 9302 u32 value) 9303 { 9304 u32 proto_down_reason; 9305 int b; 9306 9307 if (!mask) { 9308 proto_down_reason = value; 9309 } else { 9310 proto_down_reason = dev->proto_down_reason; 9311 for_each_set_bit(b, &mask, 32) { 9312 if (value & (1 << b)) 9313 proto_down_reason |= BIT(b); 9314 else 9315 proto_down_reason &= ~BIT(b); 9316 } 9317 } 9318 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 9319 } 9320 9321 struct bpf_xdp_link { 9322 struct bpf_link link; 9323 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9324 int flags; 9325 }; 9326 9327 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9328 { 9329 if (flags & XDP_FLAGS_HW_MODE) 9330 return XDP_MODE_HW; 9331 if (flags & XDP_FLAGS_DRV_MODE) 9332 return XDP_MODE_DRV; 9333 if (flags & XDP_FLAGS_SKB_MODE) 9334 return XDP_MODE_SKB; 9335 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9336 } 9337 9338 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9339 { 9340 switch (mode) { 9341 case XDP_MODE_SKB: 9342 return generic_xdp_install; 9343 case XDP_MODE_DRV: 9344 case XDP_MODE_HW: 9345 return dev->netdev_ops->ndo_bpf; 9346 default: 9347 return NULL; 9348 } 9349 } 9350 9351 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9352 enum bpf_xdp_mode mode) 9353 { 9354 return dev->xdp_state[mode].link; 9355 } 9356 9357 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9358 enum bpf_xdp_mode mode) 9359 { 9360 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9361 9362 if (link) 9363 return link->link.prog; 9364 return dev->xdp_state[mode].prog; 9365 } 9366 9367 u8 dev_xdp_prog_count(struct net_device *dev) 9368 { 9369 u8 count = 0; 9370 int i; 9371 9372 for (i = 0; i < __MAX_XDP_MODE; i++) 9373 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9374 count++; 9375 return count; 9376 } 9377 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9378 9379 int dev_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 9380 { 9381 if (!dev->netdev_ops->ndo_bpf) 9382 return -EOPNOTSUPP; 9383 9384 if (dev_get_min_mp_channel_count(dev)) { 9385 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 9386 return -EBUSY; 9387 } 9388 9389 return dev->netdev_ops->ndo_bpf(dev, bpf); 9390 } 9391 EXPORT_SYMBOL_GPL(dev_xdp_propagate); 9392 9393 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9394 { 9395 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9396 9397 return prog ? prog->aux->id : 0; 9398 } 9399 9400 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9401 struct bpf_xdp_link *link) 9402 { 9403 dev->xdp_state[mode].link = link; 9404 dev->xdp_state[mode].prog = NULL; 9405 } 9406 9407 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9408 struct bpf_prog *prog) 9409 { 9410 dev->xdp_state[mode].link = NULL; 9411 dev->xdp_state[mode].prog = prog; 9412 } 9413 9414 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9415 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9416 u32 flags, struct bpf_prog *prog) 9417 { 9418 struct netdev_bpf xdp; 9419 int err; 9420 9421 if (dev_get_min_mp_channel_count(dev)) { 9422 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 9423 return -EBUSY; 9424 } 9425 9426 memset(&xdp, 0, sizeof(xdp)); 9427 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9428 xdp.extack = extack; 9429 xdp.flags = flags; 9430 xdp.prog = prog; 9431 9432 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9433 * "moved" into driver), so they don't increment it on their own, but 9434 * they do decrement refcnt when program is detached or replaced. 9435 * Given net_device also owns link/prog, we need to bump refcnt here 9436 * to prevent drivers from underflowing it. 9437 */ 9438 if (prog) 9439 bpf_prog_inc(prog); 9440 err = bpf_op(dev, &xdp); 9441 if (err) { 9442 if (prog) 9443 bpf_prog_put(prog); 9444 return err; 9445 } 9446 9447 if (mode != XDP_MODE_HW) 9448 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9449 9450 return 0; 9451 } 9452 9453 static void dev_xdp_uninstall(struct net_device *dev) 9454 { 9455 struct bpf_xdp_link *link; 9456 struct bpf_prog *prog; 9457 enum bpf_xdp_mode mode; 9458 bpf_op_t bpf_op; 9459 9460 ASSERT_RTNL(); 9461 9462 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 9463 prog = dev_xdp_prog(dev, mode); 9464 if (!prog) 9465 continue; 9466 9467 bpf_op = dev_xdp_bpf_op(dev, mode); 9468 if (!bpf_op) 9469 continue; 9470 9471 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9472 9473 /* auto-detach link from net device */ 9474 link = dev_xdp_link(dev, mode); 9475 if (link) 9476 link->dev = NULL; 9477 else 9478 bpf_prog_put(prog); 9479 9480 dev_xdp_set_link(dev, mode, NULL); 9481 } 9482 } 9483 9484 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 9485 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 9486 struct bpf_prog *old_prog, u32 flags) 9487 { 9488 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 9489 struct bpf_prog *cur_prog; 9490 struct net_device *upper; 9491 struct list_head *iter; 9492 enum bpf_xdp_mode mode; 9493 bpf_op_t bpf_op; 9494 int err; 9495 9496 ASSERT_RTNL(); 9497 9498 /* either link or prog attachment, never both */ 9499 if (link && (new_prog || old_prog)) 9500 return -EINVAL; 9501 /* link supports only XDP mode flags */ 9502 if (link && (flags & ~XDP_FLAGS_MODES)) { 9503 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 9504 return -EINVAL; 9505 } 9506 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 9507 if (num_modes > 1) { 9508 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 9509 return -EINVAL; 9510 } 9511 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 9512 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 9513 NL_SET_ERR_MSG(extack, 9514 "More than one program loaded, unset mode is ambiguous"); 9515 return -EINVAL; 9516 } 9517 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 9518 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 9519 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 9520 return -EINVAL; 9521 } 9522 9523 mode = dev_xdp_mode(dev, flags); 9524 /* can't replace attached link */ 9525 if (dev_xdp_link(dev, mode)) { 9526 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 9527 return -EBUSY; 9528 } 9529 9530 /* don't allow if an upper device already has a program */ 9531 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 9532 if (dev_xdp_prog_count(upper) > 0) { 9533 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 9534 return -EEXIST; 9535 } 9536 } 9537 9538 cur_prog = dev_xdp_prog(dev, mode); 9539 /* can't replace attached prog with link */ 9540 if (link && cur_prog) { 9541 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 9542 return -EBUSY; 9543 } 9544 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 9545 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 9546 return -EEXIST; 9547 } 9548 9549 /* put effective new program into new_prog */ 9550 if (link) 9551 new_prog = link->link.prog; 9552 9553 if (new_prog) { 9554 bool offload = mode == XDP_MODE_HW; 9555 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 9556 ? XDP_MODE_DRV : XDP_MODE_SKB; 9557 9558 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 9559 NL_SET_ERR_MSG(extack, "XDP program already attached"); 9560 return -EBUSY; 9561 } 9562 if (!offload && dev_xdp_prog(dev, other_mode)) { 9563 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 9564 return -EEXIST; 9565 } 9566 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 9567 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 9568 return -EINVAL; 9569 } 9570 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 9571 NL_SET_ERR_MSG(extack, "Program bound to different device"); 9572 return -EINVAL; 9573 } 9574 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 9575 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 9576 return -EINVAL; 9577 } 9578 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 9579 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 9580 return -EINVAL; 9581 } 9582 } 9583 9584 /* don't call drivers if the effective program didn't change */ 9585 if (new_prog != cur_prog) { 9586 bpf_op = dev_xdp_bpf_op(dev, mode); 9587 if (!bpf_op) { 9588 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 9589 return -EOPNOTSUPP; 9590 } 9591 9592 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 9593 if (err) 9594 return err; 9595 } 9596 9597 if (link) 9598 dev_xdp_set_link(dev, mode, link); 9599 else 9600 dev_xdp_set_prog(dev, mode, new_prog); 9601 if (cur_prog) 9602 bpf_prog_put(cur_prog); 9603 9604 return 0; 9605 } 9606 9607 static int dev_xdp_attach_link(struct net_device *dev, 9608 struct netlink_ext_ack *extack, 9609 struct bpf_xdp_link *link) 9610 { 9611 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 9612 } 9613 9614 static int dev_xdp_detach_link(struct net_device *dev, 9615 struct netlink_ext_ack *extack, 9616 struct bpf_xdp_link *link) 9617 { 9618 enum bpf_xdp_mode mode; 9619 bpf_op_t bpf_op; 9620 9621 ASSERT_RTNL(); 9622 9623 mode = dev_xdp_mode(dev, link->flags); 9624 if (dev_xdp_link(dev, mode) != link) 9625 return -EINVAL; 9626 9627 bpf_op = dev_xdp_bpf_op(dev, mode); 9628 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9629 dev_xdp_set_link(dev, mode, NULL); 9630 return 0; 9631 } 9632 9633 static void bpf_xdp_link_release(struct bpf_link *link) 9634 { 9635 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9636 9637 rtnl_lock(); 9638 9639 /* if racing with net_device's tear down, xdp_link->dev might be 9640 * already NULL, in which case link was already auto-detached 9641 */ 9642 if (xdp_link->dev) { 9643 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 9644 xdp_link->dev = NULL; 9645 } 9646 9647 rtnl_unlock(); 9648 } 9649 9650 static int bpf_xdp_link_detach(struct bpf_link *link) 9651 { 9652 bpf_xdp_link_release(link); 9653 return 0; 9654 } 9655 9656 static void bpf_xdp_link_dealloc(struct bpf_link *link) 9657 { 9658 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9659 9660 kfree(xdp_link); 9661 } 9662 9663 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 9664 struct seq_file *seq) 9665 { 9666 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9667 u32 ifindex = 0; 9668 9669 rtnl_lock(); 9670 if (xdp_link->dev) 9671 ifindex = xdp_link->dev->ifindex; 9672 rtnl_unlock(); 9673 9674 seq_printf(seq, "ifindex:\t%u\n", ifindex); 9675 } 9676 9677 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 9678 struct bpf_link_info *info) 9679 { 9680 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9681 u32 ifindex = 0; 9682 9683 rtnl_lock(); 9684 if (xdp_link->dev) 9685 ifindex = xdp_link->dev->ifindex; 9686 rtnl_unlock(); 9687 9688 info->xdp.ifindex = ifindex; 9689 return 0; 9690 } 9691 9692 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 9693 struct bpf_prog *old_prog) 9694 { 9695 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9696 enum bpf_xdp_mode mode; 9697 bpf_op_t bpf_op; 9698 int err = 0; 9699 9700 rtnl_lock(); 9701 9702 /* link might have been auto-released already, so fail */ 9703 if (!xdp_link->dev) { 9704 err = -ENOLINK; 9705 goto out_unlock; 9706 } 9707 9708 if (old_prog && link->prog != old_prog) { 9709 err = -EPERM; 9710 goto out_unlock; 9711 } 9712 old_prog = link->prog; 9713 if (old_prog->type != new_prog->type || 9714 old_prog->expected_attach_type != new_prog->expected_attach_type) { 9715 err = -EINVAL; 9716 goto out_unlock; 9717 } 9718 9719 if (old_prog == new_prog) { 9720 /* no-op, don't disturb drivers */ 9721 bpf_prog_put(new_prog); 9722 goto out_unlock; 9723 } 9724 9725 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 9726 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 9727 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 9728 xdp_link->flags, new_prog); 9729 if (err) 9730 goto out_unlock; 9731 9732 old_prog = xchg(&link->prog, new_prog); 9733 bpf_prog_put(old_prog); 9734 9735 out_unlock: 9736 rtnl_unlock(); 9737 return err; 9738 } 9739 9740 static const struct bpf_link_ops bpf_xdp_link_lops = { 9741 .release = bpf_xdp_link_release, 9742 .dealloc = bpf_xdp_link_dealloc, 9743 .detach = bpf_xdp_link_detach, 9744 .show_fdinfo = bpf_xdp_link_show_fdinfo, 9745 .fill_link_info = bpf_xdp_link_fill_link_info, 9746 .update_prog = bpf_xdp_link_update, 9747 }; 9748 9749 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 9750 { 9751 struct net *net = current->nsproxy->net_ns; 9752 struct bpf_link_primer link_primer; 9753 struct netlink_ext_ack extack = {}; 9754 struct bpf_xdp_link *link; 9755 struct net_device *dev; 9756 int err, fd; 9757 9758 rtnl_lock(); 9759 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 9760 if (!dev) { 9761 rtnl_unlock(); 9762 return -EINVAL; 9763 } 9764 9765 link = kzalloc(sizeof(*link), GFP_USER); 9766 if (!link) { 9767 err = -ENOMEM; 9768 goto unlock; 9769 } 9770 9771 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 9772 link->dev = dev; 9773 link->flags = attr->link_create.flags; 9774 9775 err = bpf_link_prime(&link->link, &link_primer); 9776 if (err) { 9777 kfree(link); 9778 goto unlock; 9779 } 9780 9781 err = dev_xdp_attach_link(dev, &extack, link); 9782 rtnl_unlock(); 9783 9784 if (err) { 9785 link->dev = NULL; 9786 bpf_link_cleanup(&link_primer); 9787 trace_bpf_xdp_link_attach_failed(extack._msg); 9788 goto out_put_dev; 9789 } 9790 9791 fd = bpf_link_settle(&link_primer); 9792 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 9793 dev_put(dev); 9794 return fd; 9795 9796 unlock: 9797 rtnl_unlock(); 9798 9799 out_put_dev: 9800 dev_put(dev); 9801 return err; 9802 } 9803 9804 /** 9805 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 9806 * @dev: device 9807 * @extack: netlink extended ack 9808 * @fd: new program fd or negative value to clear 9809 * @expected_fd: old program fd that userspace expects to replace or clear 9810 * @flags: xdp-related flags 9811 * 9812 * Set or clear a bpf program for a device 9813 */ 9814 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 9815 int fd, int expected_fd, u32 flags) 9816 { 9817 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 9818 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 9819 int err; 9820 9821 ASSERT_RTNL(); 9822 9823 if (fd >= 0) { 9824 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 9825 mode != XDP_MODE_SKB); 9826 if (IS_ERR(new_prog)) 9827 return PTR_ERR(new_prog); 9828 } 9829 9830 if (expected_fd >= 0) { 9831 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 9832 mode != XDP_MODE_SKB); 9833 if (IS_ERR(old_prog)) { 9834 err = PTR_ERR(old_prog); 9835 old_prog = NULL; 9836 goto err_out; 9837 } 9838 } 9839 9840 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 9841 9842 err_out: 9843 if (err && new_prog) 9844 bpf_prog_put(new_prog); 9845 if (old_prog) 9846 bpf_prog_put(old_prog); 9847 return err; 9848 } 9849 9850 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 9851 { 9852 int i; 9853 9854 ASSERT_RTNL(); 9855 9856 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 9857 if (dev->_rx[i].mp_params.mp_priv) 9858 /* The channel count is the idx plus 1. */ 9859 return i + 1; 9860 9861 return 0; 9862 } 9863 9864 /** 9865 * dev_index_reserve() - allocate an ifindex in a namespace 9866 * @net: the applicable net namespace 9867 * @ifindex: requested ifindex, pass %0 to get one allocated 9868 * 9869 * Allocate a ifindex for a new device. Caller must either use the ifindex 9870 * to store the device (via list_netdevice()) or call dev_index_release() 9871 * to give the index up. 9872 * 9873 * Return: a suitable unique value for a new device interface number or -errno. 9874 */ 9875 static int dev_index_reserve(struct net *net, u32 ifindex) 9876 { 9877 int err; 9878 9879 if (ifindex > INT_MAX) { 9880 DEBUG_NET_WARN_ON_ONCE(1); 9881 return -EINVAL; 9882 } 9883 9884 if (!ifindex) 9885 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 9886 xa_limit_31b, &net->ifindex, GFP_KERNEL); 9887 else 9888 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 9889 if (err < 0) 9890 return err; 9891 9892 return ifindex; 9893 } 9894 9895 static void dev_index_release(struct net *net, int ifindex) 9896 { 9897 /* Expect only unused indexes, unlist_netdevice() removes the used */ 9898 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 9899 } 9900 9901 /* Delayed registration/unregisteration */ 9902 LIST_HEAD(net_todo_list); 9903 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 9904 atomic_t dev_unreg_count = ATOMIC_INIT(0); 9905 9906 static void net_set_todo(struct net_device *dev) 9907 { 9908 list_add_tail(&dev->todo_list, &net_todo_list); 9909 } 9910 9911 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 9912 struct net_device *upper, netdev_features_t features) 9913 { 9914 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 9915 netdev_features_t feature; 9916 int feature_bit; 9917 9918 for_each_netdev_feature(upper_disables, feature_bit) { 9919 feature = __NETIF_F_BIT(feature_bit); 9920 if (!(upper->wanted_features & feature) 9921 && (features & feature)) { 9922 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 9923 &feature, upper->name); 9924 features &= ~feature; 9925 } 9926 } 9927 9928 return features; 9929 } 9930 9931 static void netdev_sync_lower_features(struct net_device *upper, 9932 struct net_device *lower, netdev_features_t features) 9933 { 9934 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 9935 netdev_features_t feature; 9936 int feature_bit; 9937 9938 for_each_netdev_feature(upper_disables, feature_bit) { 9939 feature = __NETIF_F_BIT(feature_bit); 9940 if (!(features & feature) && (lower->features & feature)) { 9941 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 9942 &feature, lower->name); 9943 lower->wanted_features &= ~feature; 9944 __netdev_update_features(lower); 9945 9946 if (unlikely(lower->features & feature)) 9947 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 9948 &feature, lower->name); 9949 else 9950 netdev_features_change(lower); 9951 } 9952 } 9953 } 9954 9955 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 9956 { 9957 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 9958 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 9959 bool hw_csum = features & NETIF_F_HW_CSUM; 9960 9961 return ip_csum || hw_csum; 9962 } 9963 9964 static netdev_features_t netdev_fix_features(struct net_device *dev, 9965 netdev_features_t features) 9966 { 9967 /* Fix illegal checksum combinations */ 9968 if ((features & NETIF_F_HW_CSUM) && 9969 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 9970 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 9971 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 9972 } 9973 9974 /* TSO requires that SG is present as well. */ 9975 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 9976 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 9977 features &= ~NETIF_F_ALL_TSO; 9978 } 9979 9980 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 9981 !(features & NETIF_F_IP_CSUM)) { 9982 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 9983 features &= ~NETIF_F_TSO; 9984 features &= ~NETIF_F_TSO_ECN; 9985 } 9986 9987 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 9988 !(features & NETIF_F_IPV6_CSUM)) { 9989 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 9990 features &= ~NETIF_F_TSO6; 9991 } 9992 9993 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 9994 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 9995 features &= ~NETIF_F_TSO_MANGLEID; 9996 9997 /* TSO ECN requires that TSO is present as well. */ 9998 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 9999 features &= ~NETIF_F_TSO_ECN; 10000 10001 /* Software GSO depends on SG. */ 10002 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10003 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10004 features &= ~NETIF_F_GSO; 10005 } 10006 10007 /* GSO partial features require GSO partial be set */ 10008 if ((features & dev->gso_partial_features) && 10009 !(features & NETIF_F_GSO_PARTIAL)) { 10010 netdev_dbg(dev, 10011 "Dropping partially supported GSO features since no GSO partial.\n"); 10012 features &= ~dev->gso_partial_features; 10013 } 10014 10015 if (!(features & NETIF_F_RXCSUM)) { 10016 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10017 * successfully merged by hardware must also have the 10018 * checksum verified by hardware. If the user does not 10019 * want to enable RXCSUM, logically, we should disable GRO_HW. 10020 */ 10021 if (features & NETIF_F_GRO_HW) { 10022 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10023 features &= ~NETIF_F_GRO_HW; 10024 } 10025 } 10026 10027 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10028 if (features & NETIF_F_RXFCS) { 10029 if (features & NETIF_F_LRO) { 10030 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10031 features &= ~NETIF_F_LRO; 10032 } 10033 10034 if (features & NETIF_F_GRO_HW) { 10035 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10036 features &= ~NETIF_F_GRO_HW; 10037 } 10038 } 10039 10040 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10041 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10042 features &= ~NETIF_F_LRO; 10043 } 10044 10045 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10046 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10047 features &= ~NETIF_F_HW_TLS_TX; 10048 } 10049 10050 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10051 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10052 features &= ~NETIF_F_HW_TLS_RX; 10053 } 10054 10055 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10056 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10057 features &= ~NETIF_F_GSO_UDP_L4; 10058 } 10059 10060 return features; 10061 } 10062 10063 int __netdev_update_features(struct net_device *dev) 10064 { 10065 struct net_device *upper, *lower; 10066 netdev_features_t features; 10067 struct list_head *iter; 10068 int err = -1; 10069 10070 ASSERT_RTNL(); 10071 10072 features = netdev_get_wanted_features(dev); 10073 10074 if (dev->netdev_ops->ndo_fix_features) 10075 features = dev->netdev_ops->ndo_fix_features(dev, features); 10076 10077 /* driver might be less strict about feature dependencies */ 10078 features = netdev_fix_features(dev, features); 10079 10080 /* some features can't be enabled if they're off on an upper device */ 10081 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10082 features = netdev_sync_upper_features(dev, upper, features); 10083 10084 if (dev->features == features) 10085 goto sync_lower; 10086 10087 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10088 &dev->features, &features); 10089 10090 if (dev->netdev_ops->ndo_set_features) 10091 err = dev->netdev_ops->ndo_set_features(dev, features); 10092 else 10093 err = 0; 10094 10095 if (unlikely(err < 0)) { 10096 netdev_err(dev, 10097 "set_features() failed (%d); wanted %pNF, left %pNF\n", 10098 err, &features, &dev->features); 10099 /* return non-0 since some features might have changed and 10100 * it's better to fire a spurious notification than miss it 10101 */ 10102 return -1; 10103 } 10104 10105 sync_lower: 10106 /* some features must be disabled on lower devices when disabled 10107 * on an upper device (think: bonding master or bridge) 10108 */ 10109 netdev_for_each_lower_dev(dev, lower, iter) 10110 netdev_sync_lower_features(dev, lower, features); 10111 10112 if (!err) { 10113 netdev_features_t diff = features ^ dev->features; 10114 10115 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 10116 /* udp_tunnel_{get,drop}_rx_info both need 10117 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 10118 * device, or they won't do anything. 10119 * Thus we need to update dev->features 10120 * *before* calling udp_tunnel_get_rx_info, 10121 * but *after* calling udp_tunnel_drop_rx_info. 10122 */ 10123 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 10124 dev->features = features; 10125 udp_tunnel_get_rx_info(dev); 10126 } else { 10127 udp_tunnel_drop_rx_info(dev); 10128 } 10129 } 10130 10131 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 10132 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 10133 dev->features = features; 10134 err |= vlan_get_rx_ctag_filter_info(dev); 10135 } else { 10136 vlan_drop_rx_ctag_filter_info(dev); 10137 } 10138 } 10139 10140 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 10141 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 10142 dev->features = features; 10143 err |= vlan_get_rx_stag_filter_info(dev); 10144 } else { 10145 vlan_drop_rx_stag_filter_info(dev); 10146 } 10147 } 10148 10149 dev->features = features; 10150 } 10151 10152 return err < 0 ? 0 : 1; 10153 } 10154 10155 /** 10156 * netdev_update_features - recalculate device features 10157 * @dev: the device to check 10158 * 10159 * Recalculate dev->features set and send notifications if it 10160 * has changed. Should be called after driver or hardware dependent 10161 * conditions might have changed that influence the features. 10162 */ 10163 void netdev_update_features(struct net_device *dev) 10164 { 10165 if (__netdev_update_features(dev)) 10166 netdev_features_change(dev); 10167 } 10168 EXPORT_SYMBOL(netdev_update_features); 10169 10170 /** 10171 * netdev_change_features - recalculate device features 10172 * @dev: the device to check 10173 * 10174 * Recalculate dev->features set and send notifications even 10175 * if they have not changed. Should be called instead of 10176 * netdev_update_features() if also dev->vlan_features might 10177 * have changed to allow the changes to be propagated to stacked 10178 * VLAN devices. 10179 */ 10180 void netdev_change_features(struct net_device *dev) 10181 { 10182 __netdev_update_features(dev); 10183 netdev_features_change(dev); 10184 } 10185 EXPORT_SYMBOL(netdev_change_features); 10186 10187 /** 10188 * netif_stacked_transfer_operstate - transfer operstate 10189 * @rootdev: the root or lower level device to transfer state from 10190 * @dev: the device to transfer operstate to 10191 * 10192 * Transfer operational state from root to device. This is normally 10193 * called when a stacking relationship exists between the root 10194 * device and the device(a leaf device). 10195 */ 10196 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 10197 struct net_device *dev) 10198 { 10199 if (rootdev->operstate == IF_OPER_DORMANT) 10200 netif_dormant_on(dev); 10201 else 10202 netif_dormant_off(dev); 10203 10204 if (rootdev->operstate == IF_OPER_TESTING) 10205 netif_testing_on(dev); 10206 else 10207 netif_testing_off(dev); 10208 10209 if (netif_carrier_ok(rootdev)) 10210 netif_carrier_on(dev); 10211 else 10212 netif_carrier_off(dev); 10213 } 10214 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10215 10216 static int netif_alloc_rx_queues(struct net_device *dev) 10217 { 10218 unsigned int i, count = dev->num_rx_queues; 10219 struct netdev_rx_queue *rx; 10220 size_t sz = count * sizeof(*rx); 10221 int err = 0; 10222 10223 BUG_ON(count < 1); 10224 10225 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10226 if (!rx) 10227 return -ENOMEM; 10228 10229 dev->_rx = rx; 10230 10231 for (i = 0; i < count; i++) { 10232 rx[i].dev = dev; 10233 10234 /* XDP RX-queue setup */ 10235 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10236 if (err < 0) 10237 goto err_rxq_info; 10238 } 10239 return 0; 10240 10241 err_rxq_info: 10242 /* Rollback successful reg's and free other resources */ 10243 while (i--) 10244 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10245 kvfree(dev->_rx); 10246 dev->_rx = NULL; 10247 return err; 10248 } 10249 10250 static void netif_free_rx_queues(struct net_device *dev) 10251 { 10252 unsigned int i, count = dev->num_rx_queues; 10253 10254 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10255 if (!dev->_rx) 10256 return; 10257 10258 for (i = 0; i < count; i++) 10259 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10260 10261 kvfree(dev->_rx); 10262 } 10263 10264 static void netdev_init_one_queue(struct net_device *dev, 10265 struct netdev_queue *queue, void *_unused) 10266 { 10267 /* Initialize queue lock */ 10268 spin_lock_init(&queue->_xmit_lock); 10269 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10270 queue->xmit_lock_owner = -1; 10271 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10272 queue->dev = dev; 10273 #ifdef CONFIG_BQL 10274 dql_init(&queue->dql, HZ); 10275 #endif 10276 } 10277 10278 static void netif_free_tx_queues(struct net_device *dev) 10279 { 10280 kvfree(dev->_tx); 10281 } 10282 10283 static int netif_alloc_netdev_queues(struct net_device *dev) 10284 { 10285 unsigned int count = dev->num_tx_queues; 10286 struct netdev_queue *tx; 10287 size_t sz = count * sizeof(*tx); 10288 10289 if (count < 1 || count > 0xffff) 10290 return -EINVAL; 10291 10292 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10293 if (!tx) 10294 return -ENOMEM; 10295 10296 dev->_tx = tx; 10297 10298 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10299 spin_lock_init(&dev->tx_global_lock); 10300 10301 return 0; 10302 } 10303 10304 void netif_tx_stop_all_queues(struct net_device *dev) 10305 { 10306 unsigned int i; 10307 10308 for (i = 0; i < dev->num_tx_queues; i++) { 10309 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10310 10311 netif_tx_stop_queue(txq); 10312 } 10313 } 10314 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10315 10316 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10317 { 10318 void __percpu *v; 10319 10320 /* Drivers implementing ndo_get_peer_dev must support tstat 10321 * accounting, so that skb_do_redirect() can bump the dev's 10322 * RX stats upon network namespace switch. 10323 */ 10324 if (dev->netdev_ops->ndo_get_peer_dev && 10325 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10326 return -EOPNOTSUPP; 10327 10328 switch (dev->pcpu_stat_type) { 10329 case NETDEV_PCPU_STAT_NONE: 10330 return 0; 10331 case NETDEV_PCPU_STAT_LSTATS: 10332 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10333 break; 10334 case NETDEV_PCPU_STAT_TSTATS: 10335 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10336 break; 10337 case NETDEV_PCPU_STAT_DSTATS: 10338 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10339 break; 10340 default: 10341 return -EINVAL; 10342 } 10343 10344 return v ? 0 : -ENOMEM; 10345 } 10346 10347 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10348 { 10349 switch (dev->pcpu_stat_type) { 10350 case NETDEV_PCPU_STAT_NONE: 10351 return; 10352 case NETDEV_PCPU_STAT_LSTATS: 10353 free_percpu(dev->lstats); 10354 break; 10355 case NETDEV_PCPU_STAT_TSTATS: 10356 free_percpu(dev->tstats); 10357 break; 10358 case NETDEV_PCPU_STAT_DSTATS: 10359 free_percpu(dev->dstats); 10360 break; 10361 } 10362 } 10363 10364 static void netdev_free_phy_link_topology(struct net_device *dev) 10365 { 10366 struct phy_link_topology *topo = dev->link_topo; 10367 10368 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 10369 xa_destroy(&topo->phys); 10370 kfree(topo); 10371 dev->link_topo = NULL; 10372 } 10373 } 10374 10375 /** 10376 * register_netdevice() - register a network device 10377 * @dev: device to register 10378 * 10379 * Take a prepared network device structure and make it externally accessible. 10380 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10381 * Callers must hold the rtnl lock - you may want register_netdev() 10382 * instead of this. 10383 */ 10384 int register_netdevice(struct net_device *dev) 10385 { 10386 int ret; 10387 struct net *net = dev_net(dev); 10388 10389 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10390 NETDEV_FEATURE_COUNT); 10391 BUG_ON(dev_boot_phase); 10392 ASSERT_RTNL(); 10393 10394 might_sleep(); 10395 10396 /* When net_device's are persistent, this will be fatal. */ 10397 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10398 BUG_ON(!net); 10399 10400 ret = ethtool_check_ops(dev->ethtool_ops); 10401 if (ret) 10402 return ret; 10403 10404 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 10405 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 10406 mutex_init(&dev->ethtool->rss_lock); 10407 10408 spin_lock_init(&dev->addr_list_lock); 10409 netdev_set_addr_lockdep_class(dev); 10410 10411 ret = dev_get_valid_name(net, dev, dev->name); 10412 if (ret < 0) 10413 goto out; 10414 10415 ret = -ENOMEM; 10416 dev->name_node = netdev_name_node_head_alloc(dev); 10417 if (!dev->name_node) 10418 goto out; 10419 10420 /* Init, if this function is available */ 10421 if (dev->netdev_ops->ndo_init) { 10422 ret = dev->netdev_ops->ndo_init(dev); 10423 if (ret) { 10424 if (ret > 0) 10425 ret = -EIO; 10426 goto err_free_name; 10427 } 10428 } 10429 10430 if (((dev->hw_features | dev->features) & 10431 NETIF_F_HW_VLAN_CTAG_FILTER) && 10432 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 10433 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 10434 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 10435 ret = -EINVAL; 10436 goto err_uninit; 10437 } 10438 10439 ret = netdev_do_alloc_pcpu_stats(dev); 10440 if (ret) 10441 goto err_uninit; 10442 10443 ret = dev_index_reserve(net, dev->ifindex); 10444 if (ret < 0) 10445 goto err_free_pcpu; 10446 dev->ifindex = ret; 10447 10448 /* Transfer changeable features to wanted_features and enable 10449 * software offloads (GSO and GRO). 10450 */ 10451 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 10452 dev->features |= NETIF_F_SOFT_FEATURES; 10453 10454 if (dev->udp_tunnel_nic_info) { 10455 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10456 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10457 } 10458 10459 dev->wanted_features = dev->features & dev->hw_features; 10460 10461 if (!(dev->flags & IFF_LOOPBACK)) 10462 dev->hw_features |= NETIF_F_NOCACHE_COPY; 10463 10464 /* If IPv4 TCP segmentation offload is supported we should also 10465 * allow the device to enable segmenting the frame with the option 10466 * of ignoring a static IP ID value. This doesn't enable the 10467 * feature itself but allows the user to enable it later. 10468 */ 10469 if (dev->hw_features & NETIF_F_TSO) 10470 dev->hw_features |= NETIF_F_TSO_MANGLEID; 10471 if (dev->vlan_features & NETIF_F_TSO) 10472 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 10473 if (dev->mpls_features & NETIF_F_TSO) 10474 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 10475 if (dev->hw_enc_features & NETIF_F_TSO) 10476 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 10477 10478 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 10479 */ 10480 dev->vlan_features |= NETIF_F_HIGHDMA; 10481 10482 /* Make NETIF_F_SG inheritable to tunnel devices. 10483 */ 10484 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 10485 10486 /* Make NETIF_F_SG inheritable to MPLS. 10487 */ 10488 dev->mpls_features |= NETIF_F_SG; 10489 10490 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 10491 ret = notifier_to_errno(ret); 10492 if (ret) 10493 goto err_ifindex_release; 10494 10495 ret = netdev_register_kobject(dev); 10496 10497 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 10498 10499 if (ret) 10500 goto err_uninit_notify; 10501 10502 __netdev_update_features(dev); 10503 10504 /* 10505 * Default initial state at registry is that the 10506 * device is present. 10507 */ 10508 10509 set_bit(__LINK_STATE_PRESENT, &dev->state); 10510 10511 linkwatch_init_dev(dev); 10512 10513 dev_init_scheduler(dev); 10514 10515 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 10516 list_netdevice(dev); 10517 10518 add_device_randomness(dev->dev_addr, dev->addr_len); 10519 10520 /* If the device has permanent device address, driver should 10521 * set dev_addr and also addr_assign_type should be set to 10522 * NET_ADDR_PERM (default value). 10523 */ 10524 if (dev->addr_assign_type == NET_ADDR_PERM) 10525 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 10526 10527 /* Notify protocols, that a new device appeared. */ 10528 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 10529 ret = notifier_to_errno(ret); 10530 if (ret) { 10531 /* Expect explicit free_netdev() on failure */ 10532 dev->needs_free_netdev = false; 10533 unregister_netdevice_queue(dev, NULL); 10534 goto out; 10535 } 10536 /* 10537 * Prevent userspace races by waiting until the network 10538 * device is fully setup before sending notifications. 10539 */ 10540 if (!dev->rtnl_link_ops || 10541 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 10542 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 10543 10544 out: 10545 return ret; 10546 10547 err_uninit_notify: 10548 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 10549 err_ifindex_release: 10550 dev_index_release(net, dev->ifindex); 10551 err_free_pcpu: 10552 netdev_do_free_pcpu_stats(dev); 10553 err_uninit: 10554 if (dev->netdev_ops->ndo_uninit) 10555 dev->netdev_ops->ndo_uninit(dev); 10556 if (dev->priv_destructor) 10557 dev->priv_destructor(dev); 10558 err_free_name: 10559 netdev_name_node_free(dev->name_node); 10560 goto out; 10561 } 10562 EXPORT_SYMBOL(register_netdevice); 10563 10564 /* Initialize the core of a dummy net device. 10565 * This is useful if you are calling this function after alloc_netdev(), 10566 * since it does not memset the net_device fields. 10567 */ 10568 static void init_dummy_netdev_core(struct net_device *dev) 10569 { 10570 /* make sure we BUG if trying to hit standard 10571 * register/unregister code path 10572 */ 10573 dev->reg_state = NETREG_DUMMY; 10574 10575 /* NAPI wants this */ 10576 INIT_LIST_HEAD(&dev->napi_list); 10577 10578 /* a dummy interface is started by default */ 10579 set_bit(__LINK_STATE_PRESENT, &dev->state); 10580 set_bit(__LINK_STATE_START, &dev->state); 10581 10582 /* napi_busy_loop stats accounting wants this */ 10583 dev_net_set(dev, &init_net); 10584 10585 /* Note : We dont allocate pcpu_refcnt for dummy devices, 10586 * because users of this 'device' dont need to change 10587 * its refcount. 10588 */ 10589 } 10590 10591 /** 10592 * init_dummy_netdev - init a dummy network device for NAPI 10593 * @dev: device to init 10594 * 10595 * This takes a network device structure and initializes the minimum 10596 * amount of fields so it can be used to schedule NAPI polls without 10597 * registering a full blown interface. This is to be used by drivers 10598 * that need to tie several hardware interfaces to a single NAPI 10599 * poll scheduler due to HW limitations. 10600 */ 10601 void init_dummy_netdev(struct net_device *dev) 10602 { 10603 /* Clear everything. Note we don't initialize spinlocks 10604 * as they aren't supposed to be taken by any of the 10605 * NAPI code and this dummy netdev is supposed to be 10606 * only ever used for NAPI polls 10607 */ 10608 memset(dev, 0, sizeof(struct net_device)); 10609 init_dummy_netdev_core(dev); 10610 } 10611 EXPORT_SYMBOL_GPL(init_dummy_netdev); 10612 10613 /** 10614 * register_netdev - register a network device 10615 * @dev: device to register 10616 * 10617 * Take a completed network device structure and add it to the kernel 10618 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 10619 * chain. 0 is returned on success. A negative errno code is returned 10620 * on a failure to set up the device, or if the name is a duplicate. 10621 * 10622 * This is a wrapper around register_netdevice that takes the rtnl semaphore 10623 * and expands the device name if you passed a format string to 10624 * alloc_netdev. 10625 */ 10626 int register_netdev(struct net_device *dev) 10627 { 10628 int err; 10629 10630 if (rtnl_lock_killable()) 10631 return -EINTR; 10632 err = register_netdevice(dev); 10633 rtnl_unlock(); 10634 return err; 10635 } 10636 EXPORT_SYMBOL(register_netdev); 10637 10638 int netdev_refcnt_read(const struct net_device *dev) 10639 { 10640 #ifdef CONFIG_PCPU_DEV_REFCNT 10641 int i, refcnt = 0; 10642 10643 for_each_possible_cpu(i) 10644 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 10645 return refcnt; 10646 #else 10647 return refcount_read(&dev->dev_refcnt); 10648 #endif 10649 } 10650 EXPORT_SYMBOL(netdev_refcnt_read); 10651 10652 int netdev_unregister_timeout_secs __read_mostly = 10; 10653 10654 #define WAIT_REFS_MIN_MSECS 1 10655 #define WAIT_REFS_MAX_MSECS 250 10656 /** 10657 * netdev_wait_allrefs_any - wait until all references are gone. 10658 * @list: list of net_devices to wait on 10659 * 10660 * This is called when unregistering network devices. 10661 * 10662 * Any protocol or device that holds a reference should register 10663 * for netdevice notification, and cleanup and put back the 10664 * reference if they receive an UNREGISTER event. 10665 * We can get stuck here if buggy protocols don't correctly 10666 * call dev_put. 10667 */ 10668 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 10669 { 10670 unsigned long rebroadcast_time, warning_time; 10671 struct net_device *dev; 10672 int wait = 0; 10673 10674 rebroadcast_time = warning_time = jiffies; 10675 10676 list_for_each_entry(dev, list, todo_list) 10677 if (netdev_refcnt_read(dev) == 1) 10678 return dev; 10679 10680 while (true) { 10681 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 10682 rtnl_lock(); 10683 10684 /* Rebroadcast unregister notification */ 10685 list_for_each_entry(dev, list, todo_list) 10686 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 10687 10688 __rtnl_unlock(); 10689 rcu_barrier(); 10690 rtnl_lock(); 10691 10692 list_for_each_entry(dev, list, todo_list) 10693 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 10694 &dev->state)) { 10695 /* We must not have linkwatch events 10696 * pending on unregister. If this 10697 * happens, we simply run the queue 10698 * unscheduled, resulting in a noop 10699 * for this device. 10700 */ 10701 linkwatch_run_queue(); 10702 break; 10703 } 10704 10705 __rtnl_unlock(); 10706 10707 rebroadcast_time = jiffies; 10708 } 10709 10710 rcu_barrier(); 10711 10712 if (!wait) { 10713 wait = WAIT_REFS_MIN_MSECS; 10714 } else { 10715 msleep(wait); 10716 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 10717 } 10718 10719 list_for_each_entry(dev, list, todo_list) 10720 if (netdev_refcnt_read(dev) == 1) 10721 return dev; 10722 10723 if (time_after(jiffies, warning_time + 10724 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 10725 list_for_each_entry(dev, list, todo_list) { 10726 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 10727 dev->name, netdev_refcnt_read(dev)); 10728 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 10729 } 10730 10731 warning_time = jiffies; 10732 } 10733 } 10734 } 10735 10736 /* The sequence is: 10737 * 10738 * rtnl_lock(); 10739 * ... 10740 * register_netdevice(x1); 10741 * register_netdevice(x2); 10742 * ... 10743 * unregister_netdevice(y1); 10744 * unregister_netdevice(y2); 10745 * ... 10746 * rtnl_unlock(); 10747 * free_netdev(y1); 10748 * free_netdev(y2); 10749 * 10750 * We are invoked by rtnl_unlock(). 10751 * This allows us to deal with problems: 10752 * 1) We can delete sysfs objects which invoke hotplug 10753 * without deadlocking with linkwatch via keventd. 10754 * 2) Since we run with the RTNL semaphore not held, we can sleep 10755 * safely in order to wait for the netdev refcnt to drop to zero. 10756 * 10757 * We must not return until all unregister events added during 10758 * the interval the lock was held have been completed. 10759 */ 10760 void netdev_run_todo(void) 10761 { 10762 struct net_device *dev, *tmp; 10763 struct list_head list; 10764 int cnt; 10765 #ifdef CONFIG_LOCKDEP 10766 struct list_head unlink_list; 10767 10768 list_replace_init(&net_unlink_list, &unlink_list); 10769 10770 while (!list_empty(&unlink_list)) { 10771 struct net_device *dev = list_first_entry(&unlink_list, 10772 struct net_device, 10773 unlink_list); 10774 list_del_init(&dev->unlink_list); 10775 dev->nested_level = dev->lower_level - 1; 10776 } 10777 #endif 10778 10779 /* Snapshot list, allow later requests */ 10780 list_replace_init(&net_todo_list, &list); 10781 10782 __rtnl_unlock(); 10783 10784 /* Wait for rcu callbacks to finish before next phase */ 10785 if (!list_empty(&list)) 10786 rcu_barrier(); 10787 10788 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 10789 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 10790 netdev_WARN(dev, "run_todo but not unregistering\n"); 10791 list_del(&dev->todo_list); 10792 continue; 10793 } 10794 10795 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 10796 linkwatch_sync_dev(dev); 10797 } 10798 10799 cnt = 0; 10800 while (!list_empty(&list)) { 10801 dev = netdev_wait_allrefs_any(&list); 10802 list_del(&dev->todo_list); 10803 10804 /* paranoia */ 10805 BUG_ON(netdev_refcnt_read(dev) != 1); 10806 BUG_ON(!list_empty(&dev->ptype_all)); 10807 BUG_ON(!list_empty(&dev->ptype_specific)); 10808 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 10809 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 10810 10811 netdev_do_free_pcpu_stats(dev); 10812 if (dev->priv_destructor) 10813 dev->priv_destructor(dev); 10814 if (dev->needs_free_netdev) 10815 free_netdev(dev); 10816 10817 cnt++; 10818 10819 /* Free network device */ 10820 kobject_put(&dev->dev.kobj); 10821 } 10822 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 10823 wake_up(&netdev_unregistering_wq); 10824 } 10825 10826 /* Collate per-cpu network dstats statistics 10827 * 10828 * Read per-cpu network statistics from dev->dstats and populate the related 10829 * fields in @s. 10830 */ 10831 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 10832 const struct pcpu_dstats __percpu *dstats) 10833 { 10834 int cpu; 10835 10836 for_each_possible_cpu(cpu) { 10837 u64 rx_packets, rx_bytes, rx_drops; 10838 u64 tx_packets, tx_bytes, tx_drops; 10839 const struct pcpu_dstats *stats; 10840 unsigned int start; 10841 10842 stats = per_cpu_ptr(dstats, cpu); 10843 do { 10844 start = u64_stats_fetch_begin(&stats->syncp); 10845 rx_packets = u64_stats_read(&stats->rx_packets); 10846 rx_bytes = u64_stats_read(&stats->rx_bytes); 10847 rx_drops = u64_stats_read(&stats->rx_drops); 10848 tx_packets = u64_stats_read(&stats->tx_packets); 10849 tx_bytes = u64_stats_read(&stats->tx_bytes); 10850 tx_drops = u64_stats_read(&stats->tx_drops); 10851 } while (u64_stats_fetch_retry(&stats->syncp, start)); 10852 10853 s->rx_packets += rx_packets; 10854 s->rx_bytes += rx_bytes; 10855 s->rx_dropped += rx_drops; 10856 s->tx_packets += tx_packets; 10857 s->tx_bytes += tx_bytes; 10858 s->tx_dropped += tx_drops; 10859 } 10860 } 10861 10862 /* ndo_get_stats64 implementation for dtstats-based accounting. 10863 * 10864 * Populate @s from dev->stats and dev->dstats. This is used internally by the 10865 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 10866 */ 10867 static void dev_get_dstats64(const struct net_device *dev, 10868 struct rtnl_link_stats64 *s) 10869 { 10870 netdev_stats_to_stats64(s, &dev->stats); 10871 dev_fetch_dstats(s, dev->dstats); 10872 } 10873 10874 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 10875 * all the same fields in the same order as net_device_stats, with only 10876 * the type differing, but rtnl_link_stats64 may have additional fields 10877 * at the end for newer counters. 10878 */ 10879 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 10880 const struct net_device_stats *netdev_stats) 10881 { 10882 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 10883 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 10884 u64 *dst = (u64 *)stats64; 10885 10886 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 10887 for (i = 0; i < n; i++) 10888 dst[i] = (unsigned long)atomic_long_read(&src[i]); 10889 /* zero out counters that only exist in rtnl_link_stats64 */ 10890 memset((char *)stats64 + n * sizeof(u64), 0, 10891 sizeof(*stats64) - n * sizeof(u64)); 10892 } 10893 EXPORT_SYMBOL(netdev_stats_to_stats64); 10894 10895 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 10896 struct net_device *dev) 10897 { 10898 struct net_device_core_stats __percpu *p; 10899 10900 p = alloc_percpu_gfp(struct net_device_core_stats, 10901 GFP_ATOMIC | __GFP_NOWARN); 10902 10903 if (p && cmpxchg(&dev->core_stats, NULL, p)) 10904 free_percpu(p); 10905 10906 /* This READ_ONCE() pairs with the cmpxchg() above */ 10907 return READ_ONCE(dev->core_stats); 10908 } 10909 10910 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 10911 { 10912 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 10913 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 10914 unsigned long __percpu *field; 10915 10916 if (unlikely(!p)) { 10917 p = netdev_core_stats_alloc(dev); 10918 if (!p) 10919 return; 10920 } 10921 10922 field = (unsigned long __percpu *)((void __percpu *)p + offset); 10923 this_cpu_inc(*field); 10924 } 10925 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 10926 10927 /** 10928 * dev_get_stats - get network device statistics 10929 * @dev: device to get statistics from 10930 * @storage: place to store stats 10931 * 10932 * Get network statistics from device. Return @storage. 10933 * The device driver may provide its own method by setting 10934 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 10935 * otherwise the internal statistics structure is used. 10936 */ 10937 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 10938 struct rtnl_link_stats64 *storage) 10939 { 10940 const struct net_device_ops *ops = dev->netdev_ops; 10941 const struct net_device_core_stats __percpu *p; 10942 10943 if (ops->ndo_get_stats64) { 10944 memset(storage, 0, sizeof(*storage)); 10945 ops->ndo_get_stats64(dev, storage); 10946 } else if (ops->ndo_get_stats) { 10947 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 10948 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 10949 dev_get_tstats64(dev, storage); 10950 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 10951 dev_get_dstats64(dev, storage); 10952 } else { 10953 netdev_stats_to_stats64(storage, &dev->stats); 10954 } 10955 10956 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 10957 p = READ_ONCE(dev->core_stats); 10958 if (p) { 10959 const struct net_device_core_stats *core_stats; 10960 int i; 10961 10962 for_each_possible_cpu(i) { 10963 core_stats = per_cpu_ptr(p, i); 10964 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 10965 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 10966 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 10967 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 10968 } 10969 } 10970 return storage; 10971 } 10972 EXPORT_SYMBOL(dev_get_stats); 10973 10974 /** 10975 * dev_fetch_sw_netstats - get per-cpu network device statistics 10976 * @s: place to store stats 10977 * @netstats: per-cpu network stats to read from 10978 * 10979 * Read per-cpu network statistics and populate the related fields in @s. 10980 */ 10981 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 10982 const struct pcpu_sw_netstats __percpu *netstats) 10983 { 10984 int cpu; 10985 10986 for_each_possible_cpu(cpu) { 10987 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 10988 const struct pcpu_sw_netstats *stats; 10989 unsigned int start; 10990 10991 stats = per_cpu_ptr(netstats, cpu); 10992 do { 10993 start = u64_stats_fetch_begin(&stats->syncp); 10994 rx_packets = u64_stats_read(&stats->rx_packets); 10995 rx_bytes = u64_stats_read(&stats->rx_bytes); 10996 tx_packets = u64_stats_read(&stats->tx_packets); 10997 tx_bytes = u64_stats_read(&stats->tx_bytes); 10998 } while (u64_stats_fetch_retry(&stats->syncp, start)); 10999 11000 s->rx_packets += rx_packets; 11001 s->rx_bytes += rx_bytes; 11002 s->tx_packets += tx_packets; 11003 s->tx_bytes += tx_bytes; 11004 } 11005 } 11006 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11007 11008 /** 11009 * dev_get_tstats64 - ndo_get_stats64 implementation 11010 * @dev: device to get statistics from 11011 * @s: place to store stats 11012 * 11013 * Populate @s from dev->stats and dev->tstats. Can be used as 11014 * ndo_get_stats64() callback. 11015 */ 11016 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11017 { 11018 netdev_stats_to_stats64(s, &dev->stats); 11019 dev_fetch_sw_netstats(s, dev->tstats); 11020 } 11021 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11022 11023 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11024 { 11025 struct netdev_queue *queue = dev_ingress_queue(dev); 11026 11027 #ifdef CONFIG_NET_CLS_ACT 11028 if (queue) 11029 return queue; 11030 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11031 if (!queue) 11032 return NULL; 11033 netdev_init_one_queue(dev, queue, NULL); 11034 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11035 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11036 rcu_assign_pointer(dev->ingress_queue, queue); 11037 #endif 11038 return queue; 11039 } 11040 11041 static const struct ethtool_ops default_ethtool_ops; 11042 11043 void netdev_set_default_ethtool_ops(struct net_device *dev, 11044 const struct ethtool_ops *ops) 11045 { 11046 if (dev->ethtool_ops == &default_ethtool_ops) 11047 dev->ethtool_ops = ops; 11048 } 11049 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11050 11051 /** 11052 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11053 * @dev: netdev to enable the IRQ coalescing on 11054 * 11055 * Sets a conservative default for SW IRQ coalescing. Users can use 11056 * sysfs attributes to override the default values. 11057 */ 11058 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11059 { 11060 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11061 11062 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11063 dev->gro_flush_timeout = 20000; 11064 dev->napi_defer_hard_irqs = 1; 11065 } 11066 } 11067 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11068 11069 /** 11070 * alloc_netdev_mqs - allocate network device 11071 * @sizeof_priv: size of private data to allocate space for 11072 * @name: device name format string 11073 * @name_assign_type: origin of device name 11074 * @setup: callback to initialize device 11075 * @txqs: the number of TX subqueues to allocate 11076 * @rxqs: the number of RX subqueues to allocate 11077 * 11078 * Allocates a struct net_device with private data area for driver use 11079 * and performs basic initialization. Also allocates subqueue structs 11080 * for each queue on the device. 11081 */ 11082 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11083 unsigned char name_assign_type, 11084 void (*setup)(struct net_device *), 11085 unsigned int txqs, unsigned int rxqs) 11086 { 11087 struct net_device *dev; 11088 11089 BUG_ON(strlen(name) >= sizeof(dev->name)); 11090 11091 if (txqs < 1) { 11092 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 11093 return NULL; 11094 } 11095 11096 if (rxqs < 1) { 11097 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 11098 return NULL; 11099 } 11100 11101 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 11102 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11103 if (!dev) 11104 return NULL; 11105 11106 dev->priv_len = sizeof_priv; 11107 11108 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 11109 #ifdef CONFIG_PCPU_DEV_REFCNT 11110 dev->pcpu_refcnt = alloc_percpu(int); 11111 if (!dev->pcpu_refcnt) 11112 goto free_dev; 11113 __dev_hold(dev); 11114 #else 11115 refcount_set(&dev->dev_refcnt, 1); 11116 #endif 11117 11118 if (dev_addr_init(dev)) 11119 goto free_pcpu; 11120 11121 dev_mc_init(dev); 11122 dev_uc_init(dev); 11123 11124 dev_net_set(dev, &init_net); 11125 11126 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 11127 dev->xdp_zc_max_segs = 1; 11128 dev->gso_max_segs = GSO_MAX_SEGS; 11129 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 11130 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 11131 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 11132 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 11133 dev->tso_max_segs = TSO_MAX_SEGS; 11134 dev->upper_level = 1; 11135 dev->lower_level = 1; 11136 #ifdef CONFIG_LOCKDEP 11137 dev->nested_level = 0; 11138 INIT_LIST_HEAD(&dev->unlink_list); 11139 #endif 11140 11141 INIT_LIST_HEAD(&dev->napi_list); 11142 INIT_LIST_HEAD(&dev->unreg_list); 11143 INIT_LIST_HEAD(&dev->close_list); 11144 INIT_LIST_HEAD(&dev->link_watch_list); 11145 INIT_LIST_HEAD(&dev->adj_list.upper); 11146 INIT_LIST_HEAD(&dev->adj_list.lower); 11147 INIT_LIST_HEAD(&dev->ptype_all); 11148 INIT_LIST_HEAD(&dev->ptype_specific); 11149 INIT_LIST_HEAD(&dev->net_notifier_list); 11150 #ifdef CONFIG_NET_SCHED 11151 hash_init(dev->qdisc_hash); 11152 #endif 11153 11154 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 11155 setup(dev); 11156 11157 if (!dev->tx_queue_len) { 11158 dev->priv_flags |= IFF_NO_QUEUE; 11159 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 11160 } 11161 11162 dev->num_tx_queues = txqs; 11163 dev->real_num_tx_queues = txqs; 11164 if (netif_alloc_netdev_queues(dev)) 11165 goto free_all; 11166 11167 dev->num_rx_queues = rxqs; 11168 dev->real_num_rx_queues = rxqs; 11169 if (netif_alloc_rx_queues(dev)) 11170 goto free_all; 11171 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 11172 if (!dev->ethtool) 11173 goto free_all; 11174 11175 strscpy(dev->name, name); 11176 dev->name_assign_type = name_assign_type; 11177 dev->group = INIT_NETDEV_GROUP; 11178 if (!dev->ethtool_ops) 11179 dev->ethtool_ops = &default_ethtool_ops; 11180 11181 nf_hook_netdev_init(dev); 11182 11183 return dev; 11184 11185 free_all: 11186 free_netdev(dev); 11187 return NULL; 11188 11189 free_pcpu: 11190 #ifdef CONFIG_PCPU_DEV_REFCNT 11191 free_percpu(dev->pcpu_refcnt); 11192 free_dev: 11193 #endif 11194 kvfree(dev); 11195 return NULL; 11196 } 11197 EXPORT_SYMBOL(alloc_netdev_mqs); 11198 11199 /** 11200 * free_netdev - free network device 11201 * @dev: device 11202 * 11203 * This function does the last stage of destroying an allocated device 11204 * interface. The reference to the device object is released. If this 11205 * is the last reference then it will be freed.Must be called in process 11206 * context. 11207 */ 11208 void free_netdev(struct net_device *dev) 11209 { 11210 struct napi_struct *p, *n; 11211 11212 might_sleep(); 11213 11214 /* When called immediately after register_netdevice() failed the unwind 11215 * handling may still be dismantling the device. Handle that case by 11216 * deferring the free. 11217 */ 11218 if (dev->reg_state == NETREG_UNREGISTERING) { 11219 ASSERT_RTNL(); 11220 dev->needs_free_netdev = true; 11221 return; 11222 } 11223 11224 kfree(dev->ethtool); 11225 netif_free_tx_queues(dev); 11226 netif_free_rx_queues(dev); 11227 11228 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 11229 11230 /* Flush device addresses */ 11231 dev_addr_flush(dev); 11232 11233 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 11234 netif_napi_del(p); 11235 11236 ref_tracker_dir_exit(&dev->refcnt_tracker); 11237 #ifdef CONFIG_PCPU_DEV_REFCNT 11238 free_percpu(dev->pcpu_refcnt); 11239 dev->pcpu_refcnt = NULL; 11240 #endif 11241 free_percpu(dev->core_stats); 11242 dev->core_stats = NULL; 11243 free_percpu(dev->xdp_bulkq); 11244 dev->xdp_bulkq = NULL; 11245 11246 netdev_free_phy_link_topology(dev); 11247 11248 /* Compatibility with error handling in drivers */ 11249 if (dev->reg_state == NETREG_UNINITIALIZED || 11250 dev->reg_state == NETREG_DUMMY) { 11251 kvfree(dev); 11252 return; 11253 } 11254 11255 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 11256 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 11257 11258 /* will free via device release */ 11259 put_device(&dev->dev); 11260 } 11261 EXPORT_SYMBOL(free_netdev); 11262 11263 /** 11264 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 11265 * @sizeof_priv: size of private data to allocate space for 11266 * 11267 * Return: the allocated net_device on success, NULL otherwise 11268 */ 11269 struct net_device *alloc_netdev_dummy(int sizeof_priv) 11270 { 11271 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 11272 init_dummy_netdev_core); 11273 } 11274 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 11275 11276 /** 11277 * synchronize_net - Synchronize with packet receive processing 11278 * 11279 * Wait for packets currently being received to be done. 11280 * Does not block later packets from starting. 11281 */ 11282 void synchronize_net(void) 11283 { 11284 might_sleep(); 11285 if (rtnl_is_locked()) 11286 synchronize_rcu_expedited(); 11287 else 11288 synchronize_rcu(); 11289 } 11290 EXPORT_SYMBOL(synchronize_net); 11291 11292 static void netdev_rss_contexts_free(struct net_device *dev) 11293 { 11294 struct ethtool_rxfh_context *ctx; 11295 unsigned long context; 11296 11297 mutex_lock(&dev->ethtool->rss_lock); 11298 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 11299 struct ethtool_rxfh_param rxfh; 11300 11301 rxfh.indir = ethtool_rxfh_context_indir(ctx); 11302 rxfh.key = ethtool_rxfh_context_key(ctx); 11303 rxfh.hfunc = ctx->hfunc; 11304 rxfh.input_xfrm = ctx->input_xfrm; 11305 rxfh.rss_context = context; 11306 rxfh.rss_delete = true; 11307 11308 xa_erase(&dev->ethtool->rss_ctx, context); 11309 if (dev->ethtool_ops->create_rxfh_context) 11310 dev->ethtool_ops->remove_rxfh_context(dev, ctx, 11311 context, NULL); 11312 else 11313 dev->ethtool_ops->set_rxfh(dev, &rxfh, NULL); 11314 kfree(ctx); 11315 } 11316 xa_destroy(&dev->ethtool->rss_ctx); 11317 mutex_unlock(&dev->ethtool->rss_lock); 11318 } 11319 11320 /** 11321 * unregister_netdevice_queue - remove device from the kernel 11322 * @dev: device 11323 * @head: list 11324 * 11325 * This function shuts down a device interface and removes it 11326 * from the kernel tables. 11327 * If head not NULL, device is queued to be unregistered later. 11328 * 11329 * Callers must hold the rtnl semaphore. You may want 11330 * unregister_netdev() instead of this. 11331 */ 11332 11333 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11334 { 11335 ASSERT_RTNL(); 11336 11337 if (head) { 11338 list_move_tail(&dev->unreg_list, head); 11339 } else { 11340 LIST_HEAD(single); 11341 11342 list_add(&dev->unreg_list, &single); 11343 unregister_netdevice_many(&single); 11344 } 11345 } 11346 EXPORT_SYMBOL(unregister_netdevice_queue); 11347 11348 void unregister_netdevice_many_notify(struct list_head *head, 11349 u32 portid, const struct nlmsghdr *nlh) 11350 { 11351 struct net_device *dev, *tmp; 11352 LIST_HEAD(close_head); 11353 int cnt = 0; 11354 11355 BUG_ON(dev_boot_phase); 11356 ASSERT_RTNL(); 11357 11358 if (list_empty(head)) 11359 return; 11360 11361 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11362 /* Some devices call without registering 11363 * for initialization unwind. Remove those 11364 * devices and proceed with the remaining. 11365 */ 11366 if (dev->reg_state == NETREG_UNINITIALIZED) { 11367 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11368 dev->name, dev); 11369 11370 WARN_ON(1); 11371 list_del(&dev->unreg_list); 11372 continue; 11373 } 11374 dev->dismantle = true; 11375 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11376 } 11377 11378 /* If device is running, close it first. */ 11379 list_for_each_entry(dev, head, unreg_list) 11380 list_add_tail(&dev->close_list, &close_head); 11381 dev_close_many(&close_head, true); 11382 11383 list_for_each_entry(dev, head, unreg_list) { 11384 /* And unlink it from device chain. */ 11385 unlist_netdevice(dev); 11386 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 11387 } 11388 flush_all_backlogs(); 11389 11390 synchronize_net(); 11391 11392 list_for_each_entry(dev, head, unreg_list) { 11393 struct sk_buff *skb = NULL; 11394 11395 /* Shutdown queueing discipline. */ 11396 dev_shutdown(dev); 11397 dev_tcx_uninstall(dev); 11398 dev_xdp_uninstall(dev); 11399 bpf_dev_bound_netdev_unregister(dev); 11400 dev_dmabuf_uninstall(dev); 11401 11402 netdev_offload_xstats_disable_all(dev); 11403 11404 /* Notify protocols, that we are about to destroy 11405 * this device. They should clean all the things. 11406 */ 11407 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11408 11409 if (!dev->rtnl_link_ops || 11410 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11411 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 11412 GFP_KERNEL, NULL, 0, 11413 portid, nlh); 11414 11415 /* 11416 * Flush the unicast and multicast chains 11417 */ 11418 dev_uc_flush(dev); 11419 dev_mc_flush(dev); 11420 11421 netdev_name_node_alt_flush(dev); 11422 netdev_name_node_free(dev->name_node); 11423 11424 netdev_rss_contexts_free(dev); 11425 11426 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11427 11428 if (dev->netdev_ops->ndo_uninit) 11429 dev->netdev_ops->ndo_uninit(dev); 11430 11431 mutex_destroy(&dev->ethtool->rss_lock); 11432 11433 if (skb) 11434 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 11435 11436 /* Notifier chain MUST detach us all upper devices. */ 11437 WARN_ON(netdev_has_any_upper_dev(dev)); 11438 WARN_ON(netdev_has_any_lower_dev(dev)); 11439 11440 /* Remove entries from kobject tree */ 11441 netdev_unregister_kobject(dev); 11442 #ifdef CONFIG_XPS 11443 /* Remove XPS queueing entries */ 11444 netif_reset_xps_queues_gt(dev, 0); 11445 #endif 11446 } 11447 11448 synchronize_net(); 11449 11450 list_for_each_entry(dev, head, unreg_list) { 11451 netdev_put(dev, &dev->dev_registered_tracker); 11452 net_set_todo(dev); 11453 cnt++; 11454 } 11455 atomic_add(cnt, &dev_unreg_count); 11456 11457 list_del(head); 11458 } 11459 11460 /** 11461 * unregister_netdevice_many - unregister many devices 11462 * @head: list of devices 11463 * 11464 * Note: As most callers use a stack allocated list_head, 11465 * we force a list_del() to make sure stack won't be corrupted later. 11466 */ 11467 void unregister_netdevice_many(struct list_head *head) 11468 { 11469 unregister_netdevice_many_notify(head, 0, NULL); 11470 } 11471 EXPORT_SYMBOL(unregister_netdevice_many); 11472 11473 /** 11474 * unregister_netdev - remove device from the kernel 11475 * @dev: device 11476 * 11477 * This function shuts down a device interface and removes it 11478 * from the kernel tables. 11479 * 11480 * This is just a wrapper for unregister_netdevice that takes 11481 * the rtnl semaphore. In general you want to use this and not 11482 * unregister_netdevice. 11483 */ 11484 void unregister_netdev(struct net_device *dev) 11485 { 11486 rtnl_lock(); 11487 unregister_netdevice(dev); 11488 rtnl_unlock(); 11489 } 11490 EXPORT_SYMBOL(unregister_netdev); 11491 11492 /** 11493 * __dev_change_net_namespace - move device to different nethost namespace 11494 * @dev: device 11495 * @net: network namespace 11496 * @pat: If not NULL name pattern to try if the current device name 11497 * is already taken in the destination network namespace. 11498 * @new_ifindex: If not zero, specifies device index in the target 11499 * namespace. 11500 * 11501 * This function shuts down a device interface and moves it 11502 * to a new network namespace. On success 0 is returned, on 11503 * a failure a netagive errno code is returned. 11504 * 11505 * Callers must hold the rtnl semaphore. 11506 */ 11507 11508 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 11509 const char *pat, int new_ifindex) 11510 { 11511 struct netdev_name_node *name_node; 11512 struct net *net_old = dev_net(dev); 11513 char new_name[IFNAMSIZ] = {}; 11514 int err, new_nsid; 11515 11516 ASSERT_RTNL(); 11517 11518 /* Don't allow namespace local devices to be moved. */ 11519 err = -EINVAL; 11520 if (dev->netns_local) 11521 goto out; 11522 11523 /* Ensure the device has been registered */ 11524 if (dev->reg_state != NETREG_REGISTERED) 11525 goto out; 11526 11527 /* Get out if there is nothing todo */ 11528 err = 0; 11529 if (net_eq(net_old, net)) 11530 goto out; 11531 11532 /* Pick the destination device name, and ensure 11533 * we can use it in the destination network namespace. 11534 */ 11535 err = -EEXIST; 11536 if (netdev_name_in_use(net, dev->name)) { 11537 /* We get here if we can't use the current device name */ 11538 if (!pat) 11539 goto out; 11540 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 11541 if (err < 0) 11542 goto out; 11543 } 11544 /* Check that none of the altnames conflicts. */ 11545 err = -EEXIST; 11546 netdev_for_each_altname(dev, name_node) 11547 if (netdev_name_in_use(net, name_node->name)) 11548 goto out; 11549 11550 /* Check that new_ifindex isn't used yet. */ 11551 if (new_ifindex) { 11552 err = dev_index_reserve(net, new_ifindex); 11553 if (err < 0) 11554 goto out; 11555 } else { 11556 /* If there is an ifindex conflict assign a new one */ 11557 err = dev_index_reserve(net, dev->ifindex); 11558 if (err == -EBUSY) 11559 err = dev_index_reserve(net, 0); 11560 if (err < 0) 11561 goto out; 11562 new_ifindex = err; 11563 } 11564 11565 /* 11566 * And now a mini version of register_netdevice unregister_netdevice. 11567 */ 11568 11569 /* If device is running close it first. */ 11570 dev_close(dev); 11571 11572 /* And unlink it from device chain */ 11573 unlist_netdevice(dev); 11574 11575 synchronize_net(); 11576 11577 /* Shutdown queueing discipline. */ 11578 dev_shutdown(dev); 11579 11580 /* Notify protocols, that we are about to destroy 11581 * this device. They should clean all the things. 11582 * 11583 * Note that dev->reg_state stays at NETREG_REGISTERED. 11584 * This is wanted because this way 8021q and macvlan know 11585 * the device is just moving and can keep their slaves up. 11586 */ 11587 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11588 rcu_barrier(); 11589 11590 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 11591 11592 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 11593 new_ifindex); 11594 11595 /* 11596 * Flush the unicast and multicast chains 11597 */ 11598 dev_uc_flush(dev); 11599 dev_mc_flush(dev); 11600 11601 /* Send a netdev-removed uevent to the old namespace */ 11602 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 11603 netdev_adjacent_del_links(dev); 11604 11605 /* Move per-net netdevice notifiers that are following the netdevice */ 11606 move_netdevice_notifiers_dev_net(dev, net); 11607 11608 /* Actually switch the network namespace */ 11609 dev_net_set(dev, net); 11610 dev->ifindex = new_ifindex; 11611 11612 if (new_name[0]) { 11613 /* Rename the netdev to prepared name */ 11614 write_seqlock_bh(&netdev_rename_lock); 11615 strscpy(dev->name, new_name, IFNAMSIZ); 11616 write_sequnlock_bh(&netdev_rename_lock); 11617 } 11618 11619 /* Fixup kobjects */ 11620 dev_set_uevent_suppress(&dev->dev, 1); 11621 err = device_rename(&dev->dev, dev->name); 11622 dev_set_uevent_suppress(&dev->dev, 0); 11623 WARN_ON(err); 11624 11625 /* Send a netdev-add uevent to the new namespace */ 11626 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 11627 netdev_adjacent_add_links(dev); 11628 11629 /* Adapt owner in case owning user namespace of target network 11630 * namespace is different from the original one. 11631 */ 11632 err = netdev_change_owner(dev, net_old, net); 11633 WARN_ON(err); 11634 11635 /* Add the device back in the hashes */ 11636 list_netdevice(dev); 11637 11638 /* Notify protocols, that a new device appeared. */ 11639 call_netdevice_notifiers(NETDEV_REGISTER, dev); 11640 11641 /* 11642 * Prevent userspace races by waiting until the network 11643 * device is fully setup before sending notifications. 11644 */ 11645 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11646 11647 synchronize_net(); 11648 err = 0; 11649 out: 11650 return err; 11651 } 11652 EXPORT_SYMBOL_GPL(__dev_change_net_namespace); 11653 11654 static int dev_cpu_dead(unsigned int oldcpu) 11655 { 11656 struct sk_buff **list_skb; 11657 struct sk_buff *skb; 11658 unsigned int cpu; 11659 struct softnet_data *sd, *oldsd, *remsd = NULL; 11660 11661 local_irq_disable(); 11662 cpu = smp_processor_id(); 11663 sd = &per_cpu(softnet_data, cpu); 11664 oldsd = &per_cpu(softnet_data, oldcpu); 11665 11666 /* Find end of our completion_queue. */ 11667 list_skb = &sd->completion_queue; 11668 while (*list_skb) 11669 list_skb = &(*list_skb)->next; 11670 /* Append completion queue from offline CPU. */ 11671 *list_skb = oldsd->completion_queue; 11672 oldsd->completion_queue = NULL; 11673 11674 /* Append output queue from offline CPU. */ 11675 if (oldsd->output_queue) { 11676 *sd->output_queue_tailp = oldsd->output_queue; 11677 sd->output_queue_tailp = oldsd->output_queue_tailp; 11678 oldsd->output_queue = NULL; 11679 oldsd->output_queue_tailp = &oldsd->output_queue; 11680 } 11681 /* Append NAPI poll list from offline CPU, with one exception : 11682 * process_backlog() must be called by cpu owning percpu backlog. 11683 * We properly handle process_queue & input_pkt_queue later. 11684 */ 11685 while (!list_empty(&oldsd->poll_list)) { 11686 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 11687 struct napi_struct, 11688 poll_list); 11689 11690 list_del_init(&napi->poll_list); 11691 if (napi->poll == process_backlog) 11692 napi->state &= NAPIF_STATE_THREADED; 11693 else 11694 ____napi_schedule(sd, napi); 11695 } 11696 11697 raise_softirq_irqoff(NET_TX_SOFTIRQ); 11698 local_irq_enable(); 11699 11700 if (!use_backlog_threads()) { 11701 #ifdef CONFIG_RPS 11702 remsd = oldsd->rps_ipi_list; 11703 oldsd->rps_ipi_list = NULL; 11704 #endif 11705 /* send out pending IPI's on offline CPU */ 11706 net_rps_send_ipi(remsd); 11707 } 11708 11709 /* Process offline CPU's input_pkt_queue */ 11710 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 11711 netif_rx(skb); 11712 rps_input_queue_head_incr(oldsd); 11713 } 11714 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 11715 netif_rx(skb); 11716 rps_input_queue_head_incr(oldsd); 11717 } 11718 11719 return 0; 11720 } 11721 11722 /** 11723 * netdev_increment_features - increment feature set by one 11724 * @all: current feature set 11725 * @one: new feature set 11726 * @mask: mask feature set 11727 * 11728 * Computes a new feature set after adding a device with feature set 11729 * @one to the master device with current feature set @all. Will not 11730 * enable anything that is off in @mask. Returns the new feature set. 11731 */ 11732 netdev_features_t netdev_increment_features(netdev_features_t all, 11733 netdev_features_t one, netdev_features_t mask) 11734 { 11735 if (mask & NETIF_F_HW_CSUM) 11736 mask |= NETIF_F_CSUM_MASK; 11737 mask |= NETIF_F_VLAN_CHALLENGED; 11738 11739 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 11740 all &= one | ~NETIF_F_ALL_FOR_ALL; 11741 11742 /* If one device supports hw checksumming, set for all. */ 11743 if (all & NETIF_F_HW_CSUM) 11744 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 11745 11746 return all; 11747 } 11748 EXPORT_SYMBOL(netdev_increment_features); 11749 11750 static struct hlist_head * __net_init netdev_create_hash(void) 11751 { 11752 int i; 11753 struct hlist_head *hash; 11754 11755 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 11756 if (hash != NULL) 11757 for (i = 0; i < NETDEV_HASHENTRIES; i++) 11758 INIT_HLIST_HEAD(&hash[i]); 11759 11760 return hash; 11761 } 11762 11763 /* Initialize per network namespace state */ 11764 static int __net_init netdev_init(struct net *net) 11765 { 11766 BUILD_BUG_ON(GRO_HASH_BUCKETS > 11767 8 * sizeof_field(struct napi_struct, gro_bitmask)); 11768 11769 INIT_LIST_HEAD(&net->dev_base_head); 11770 11771 net->dev_name_head = netdev_create_hash(); 11772 if (net->dev_name_head == NULL) 11773 goto err_name; 11774 11775 net->dev_index_head = netdev_create_hash(); 11776 if (net->dev_index_head == NULL) 11777 goto err_idx; 11778 11779 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 11780 11781 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 11782 11783 return 0; 11784 11785 err_idx: 11786 kfree(net->dev_name_head); 11787 err_name: 11788 return -ENOMEM; 11789 } 11790 11791 /** 11792 * netdev_drivername - network driver for the device 11793 * @dev: network device 11794 * 11795 * Determine network driver for device. 11796 */ 11797 const char *netdev_drivername(const struct net_device *dev) 11798 { 11799 const struct device_driver *driver; 11800 const struct device *parent; 11801 const char *empty = ""; 11802 11803 parent = dev->dev.parent; 11804 if (!parent) 11805 return empty; 11806 11807 driver = parent->driver; 11808 if (driver && driver->name) 11809 return driver->name; 11810 return empty; 11811 } 11812 11813 static void __netdev_printk(const char *level, const struct net_device *dev, 11814 struct va_format *vaf) 11815 { 11816 if (dev && dev->dev.parent) { 11817 dev_printk_emit(level[1] - '0', 11818 dev->dev.parent, 11819 "%s %s %s%s: %pV", 11820 dev_driver_string(dev->dev.parent), 11821 dev_name(dev->dev.parent), 11822 netdev_name(dev), netdev_reg_state(dev), 11823 vaf); 11824 } else if (dev) { 11825 printk("%s%s%s: %pV", 11826 level, netdev_name(dev), netdev_reg_state(dev), vaf); 11827 } else { 11828 printk("%s(NULL net_device): %pV", level, vaf); 11829 } 11830 } 11831 11832 void netdev_printk(const char *level, const struct net_device *dev, 11833 const char *format, ...) 11834 { 11835 struct va_format vaf; 11836 va_list args; 11837 11838 va_start(args, format); 11839 11840 vaf.fmt = format; 11841 vaf.va = &args; 11842 11843 __netdev_printk(level, dev, &vaf); 11844 11845 va_end(args); 11846 } 11847 EXPORT_SYMBOL(netdev_printk); 11848 11849 #define define_netdev_printk_level(func, level) \ 11850 void func(const struct net_device *dev, const char *fmt, ...) \ 11851 { \ 11852 struct va_format vaf; \ 11853 va_list args; \ 11854 \ 11855 va_start(args, fmt); \ 11856 \ 11857 vaf.fmt = fmt; \ 11858 vaf.va = &args; \ 11859 \ 11860 __netdev_printk(level, dev, &vaf); \ 11861 \ 11862 va_end(args); \ 11863 } \ 11864 EXPORT_SYMBOL(func); 11865 11866 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 11867 define_netdev_printk_level(netdev_alert, KERN_ALERT); 11868 define_netdev_printk_level(netdev_crit, KERN_CRIT); 11869 define_netdev_printk_level(netdev_err, KERN_ERR); 11870 define_netdev_printk_level(netdev_warn, KERN_WARNING); 11871 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 11872 define_netdev_printk_level(netdev_info, KERN_INFO); 11873 11874 static void __net_exit netdev_exit(struct net *net) 11875 { 11876 kfree(net->dev_name_head); 11877 kfree(net->dev_index_head); 11878 xa_destroy(&net->dev_by_index); 11879 if (net != &init_net) 11880 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 11881 } 11882 11883 static struct pernet_operations __net_initdata netdev_net_ops = { 11884 .init = netdev_init, 11885 .exit = netdev_exit, 11886 }; 11887 11888 static void __net_exit default_device_exit_net(struct net *net) 11889 { 11890 struct netdev_name_node *name_node, *tmp; 11891 struct net_device *dev, *aux; 11892 /* 11893 * Push all migratable network devices back to the 11894 * initial network namespace 11895 */ 11896 ASSERT_RTNL(); 11897 for_each_netdev_safe(net, dev, aux) { 11898 int err; 11899 char fb_name[IFNAMSIZ]; 11900 11901 /* Ignore unmoveable devices (i.e. loopback) */ 11902 if (dev->netns_local) 11903 continue; 11904 11905 /* Leave virtual devices for the generic cleanup */ 11906 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 11907 continue; 11908 11909 /* Push remaining network devices to init_net */ 11910 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 11911 if (netdev_name_in_use(&init_net, fb_name)) 11912 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 11913 11914 netdev_for_each_altname_safe(dev, name_node, tmp) 11915 if (netdev_name_in_use(&init_net, name_node->name)) 11916 __netdev_name_node_alt_destroy(name_node); 11917 11918 err = dev_change_net_namespace(dev, &init_net, fb_name); 11919 if (err) { 11920 pr_emerg("%s: failed to move %s to init_net: %d\n", 11921 __func__, dev->name, err); 11922 BUG(); 11923 } 11924 } 11925 } 11926 11927 static void __net_exit default_device_exit_batch(struct list_head *net_list) 11928 { 11929 /* At exit all network devices most be removed from a network 11930 * namespace. Do this in the reverse order of registration. 11931 * Do this across as many network namespaces as possible to 11932 * improve batching efficiency. 11933 */ 11934 struct net_device *dev; 11935 struct net *net; 11936 LIST_HEAD(dev_kill_list); 11937 11938 rtnl_lock(); 11939 list_for_each_entry(net, net_list, exit_list) { 11940 default_device_exit_net(net); 11941 cond_resched(); 11942 } 11943 11944 list_for_each_entry(net, net_list, exit_list) { 11945 for_each_netdev_reverse(net, dev) { 11946 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 11947 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 11948 else 11949 unregister_netdevice_queue(dev, &dev_kill_list); 11950 } 11951 } 11952 unregister_netdevice_many(&dev_kill_list); 11953 rtnl_unlock(); 11954 } 11955 11956 static struct pernet_operations __net_initdata default_device_ops = { 11957 .exit_batch = default_device_exit_batch, 11958 }; 11959 11960 static void __init net_dev_struct_check(void) 11961 { 11962 /* TX read-mostly hotpath */ 11963 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 11964 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 11965 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 11966 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 11967 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 11968 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 11969 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 11970 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 11971 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 11972 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 11973 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 11974 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 11975 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 11976 #ifdef CONFIG_XPS 11977 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 11978 #endif 11979 #ifdef CONFIG_NETFILTER_EGRESS 11980 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 11981 #endif 11982 #ifdef CONFIG_NET_XGRESS 11983 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 11984 #endif 11985 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 11986 11987 /* TXRX read-mostly hotpath */ 11988 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 11989 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 11990 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 11991 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 11992 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 11993 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 11994 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 11995 11996 /* RX read-mostly hotpath */ 11997 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 11998 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 11999 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 12000 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 12001 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_flush_timeout); 12002 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, napi_defer_hard_irqs); 12003 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 12004 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 12005 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 12006 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 12007 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 12008 #ifdef CONFIG_NETPOLL 12009 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 12010 #endif 12011 #ifdef CONFIG_NET_XGRESS 12012 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 12013 #endif 12014 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 104); 12015 } 12016 12017 /* 12018 * Initialize the DEV module. At boot time this walks the device list and 12019 * unhooks any devices that fail to initialise (normally hardware not 12020 * present) and leaves us with a valid list of present and active devices. 12021 * 12022 */ 12023 12024 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 12025 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 12026 12027 static int net_page_pool_create(int cpuid) 12028 { 12029 #if IS_ENABLED(CONFIG_PAGE_POOL) 12030 struct page_pool_params page_pool_params = { 12031 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 12032 .flags = PP_FLAG_SYSTEM_POOL, 12033 .nid = cpu_to_mem(cpuid), 12034 }; 12035 struct page_pool *pp_ptr; 12036 12037 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 12038 if (IS_ERR(pp_ptr)) 12039 return -ENOMEM; 12040 12041 per_cpu(system_page_pool, cpuid) = pp_ptr; 12042 #endif 12043 return 0; 12044 } 12045 12046 static int backlog_napi_should_run(unsigned int cpu) 12047 { 12048 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12049 struct napi_struct *napi = &sd->backlog; 12050 12051 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 12052 } 12053 12054 static void run_backlog_napi(unsigned int cpu) 12055 { 12056 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12057 12058 napi_threaded_poll_loop(&sd->backlog); 12059 } 12060 12061 static void backlog_napi_setup(unsigned int cpu) 12062 { 12063 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12064 struct napi_struct *napi = &sd->backlog; 12065 12066 napi->thread = this_cpu_read(backlog_napi); 12067 set_bit(NAPI_STATE_THREADED, &napi->state); 12068 } 12069 12070 static struct smp_hotplug_thread backlog_threads = { 12071 .store = &backlog_napi, 12072 .thread_should_run = backlog_napi_should_run, 12073 .thread_fn = run_backlog_napi, 12074 .thread_comm = "backlog_napi/%u", 12075 .setup = backlog_napi_setup, 12076 }; 12077 12078 /* 12079 * This is called single threaded during boot, so no need 12080 * to take the rtnl semaphore. 12081 */ 12082 static int __init net_dev_init(void) 12083 { 12084 int i, rc = -ENOMEM; 12085 12086 BUG_ON(!dev_boot_phase); 12087 12088 net_dev_struct_check(); 12089 12090 if (dev_proc_init()) 12091 goto out; 12092 12093 if (netdev_kobject_init()) 12094 goto out; 12095 12096 for (i = 0; i < PTYPE_HASH_SIZE; i++) 12097 INIT_LIST_HEAD(&ptype_base[i]); 12098 12099 if (register_pernet_subsys(&netdev_net_ops)) 12100 goto out; 12101 12102 /* 12103 * Initialise the packet receive queues. 12104 */ 12105 12106 for_each_possible_cpu(i) { 12107 struct work_struct *flush = per_cpu_ptr(&flush_works, i); 12108 struct softnet_data *sd = &per_cpu(softnet_data, i); 12109 12110 INIT_WORK(flush, flush_backlog); 12111 12112 skb_queue_head_init(&sd->input_pkt_queue); 12113 skb_queue_head_init(&sd->process_queue); 12114 #ifdef CONFIG_XFRM_OFFLOAD 12115 skb_queue_head_init(&sd->xfrm_backlog); 12116 #endif 12117 INIT_LIST_HEAD(&sd->poll_list); 12118 sd->output_queue_tailp = &sd->output_queue; 12119 #ifdef CONFIG_RPS 12120 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 12121 sd->cpu = i; 12122 #endif 12123 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 12124 spin_lock_init(&sd->defer_lock); 12125 12126 init_gro_hash(&sd->backlog); 12127 sd->backlog.poll = process_backlog; 12128 sd->backlog.weight = weight_p; 12129 INIT_LIST_HEAD(&sd->backlog.poll_list); 12130 12131 if (net_page_pool_create(i)) 12132 goto out; 12133 } 12134 if (use_backlog_threads()) 12135 smpboot_register_percpu_thread(&backlog_threads); 12136 12137 dev_boot_phase = 0; 12138 12139 /* The loopback device is special if any other network devices 12140 * is present in a network namespace the loopback device must 12141 * be present. Since we now dynamically allocate and free the 12142 * loopback device ensure this invariant is maintained by 12143 * keeping the loopback device as the first device on the 12144 * list of network devices. Ensuring the loopback devices 12145 * is the first device that appears and the last network device 12146 * that disappears. 12147 */ 12148 if (register_pernet_device(&loopback_net_ops)) 12149 goto out; 12150 12151 if (register_pernet_device(&default_device_ops)) 12152 goto out; 12153 12154 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 12155 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 12156 12157 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 12158 NULL, dev_cpu_dead); 12159 WARN_ON(rc < 0); 12160 rc = 0; 12161 12162 /* avoid static key IPIs to isolated CPUs */ 12163 if (housekeeping_enabled(HK_TYPE_MISC)) 12164 net_enable_timestamp(); 12165 out: 12166 if (rc < 0) { 12167 for_each_possible_cpu(i) { 12168 struct page_pool *pp_ptr; 12169 12170 pp_ptr = per_cpu(system_page_pool, i); 12171 if (!pp_ptr) 12172 continue; 12173 12174 page_pool_destroy(pp_ptr); 12175 per_cpu(system_page_pool, i) = NULL; 12176 } 12177 } 12178 12179 return rc; 12180 } 12181 12182 subsys_initcall(net_dev_init); 12183