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